Looking forward to biking and running to work next week!
Never seen the Pacific so angry on the Bay Area coastline before.
Politicians, athletes, journalists, NGOs, companies... They all use corporation-owned social networks to communicate with us. Citizens shouldn't have to hand over their data to companies to read content published by public entities. Public Feeds changes that by making posts of popular accounts freely available. Read more.
Jeffrey Adgate Dean (born July 23, 1968) is an American computer scientist and software engineer. He is best known for his work at Google, which he joined in 1999. He led Google AI from 2018 to 2023 and was Google's chief scientist from 2023 to 2026. Dean co-founded the AI research start-up Discovery Loop in 2026. More on Wikipedia
Looking forward to biking and running to work next week!
Never seen the Pacific so angry on the Bay Area coastline before.
Jeff Dean reposted
Never seen the Pacific so angry on the Bay Area coastline before.
I'm delighted to partner with @LaudeInstitute (along with my @Sanjay_Ghemawat, John Hennessy, Dave Patterson, Ben Gomes and @andykonwinski) to support funding of open science moonshots!
anyone who is close to research knows how broken the incentives are. doing open science at the frontier is harder than ever, compute costs alone price most teams out. we kept wondering what might happen if we give the right resources to the right researchers at the right time?
Jeff Dean reposted
Most think vibecoding at scale is slop. Here’s a counterexample.
Today we release a 16M-line formalization of the entire Stacks Project, critical infra underlying modern math, with optimized code quality and its highly interconnected structure preserved.
A mathematical city 🌆
Jeff Dean reposted
Today we release Ripple-65B Preview, our 65B dense LLM we pretrained from scratch.
Ripple sets a new state-of-the-art for U.S. open-weight models in agentic coding tasks, surpassing significantly larger models including Muse Spark Open-816B, Beam-501B, Inkling-975B, and Nemotron 3 Ultra-550B.
Learn more: www.ricursive.com/ripple
Awesome to see high voter turnout yesterday. Thanks to everyone who took the time to vote!
Election Tuesday is just two days away. For those eligible to vote in the U.S., please make time to go to the polls to vote (or mail in your ballot if you haven't already). Voting is a great privilege of living in a democracy and how we all make our collective voices heard. 🗳️
Jeff Dean reposted
Election Tuesday is just two days away. For those eligible to vote in the U.S., please make time to go to the polls to vote (or mail in your ballot if you haven't already). Voting is a great privilege of living in a democracy and how we all make our collective voices heard. 🗳️
Election Tuesday is just two days away. For those eligible to vote in the U.S., please make time to go to the polls to vote (or mail in your ballot if you haven't already). Voting is a great privilege of living in a democracy and how we all make our collective voices heard. 🗳️
There are majority disparities in various demographics of who votes and who does not. As one example younger voters are (e.g. ages 18-29 and 30-49) are much less likely to vote, yet the impact of governing decisions have long lasting effects.
www.pewresearch.org/politics/2023/07/12/voter-turnout-201…
Jeff Dean reposted
GOLDEN STATE. GOLD STANDARD.
Your 2026 @WNBA Champions, the Golden State Valkyries. 🗣️
Jeff Dean reposted
Thank you, @JeffDean , for visiting us in Athens and sharing your experience. We’ve since added new ways to explore the mechanisms online:
interactive.kotsanas.com/lab/
If you're ever in Athens, I highly recommend the Museum of Ancient Greek Technology!
maps.app.goo.gl/sngwM98ueKkrooVy5
Early humanoid robots to pour water+wine, demonstrations of heat+steam+water weight to "magically" open 500 kg temple doors, armor, ...
(cont)
Jeff Dean reposted
The largest math proof in human history is now computer-verified.
Today, Axiom announces a complete Lean auto-formalization of the Classification of Finite Simple Groups, the "periodic table of symmetry."
About 15,000 pages, built by over 100 mathematicians across generations.
Communicating effectively about science is really important. It's nice to see these amazing journalists recognized for their effectiveness at bringing scientific topics to life for broader audiences!
Congratulations to the 2026 winners of the National Academies Eric and Wendy Schmidt Awards for Excellence in Science Communications given by @theNASEM in partnership with @schmidtsciences!
Learn more about the winners: www.nationalacademies.org/news/national-academies-announc… #scicomm
Jeff Dean reposted
I’m starting a new early-stage AI fund, doing what I love: partnering with founders from company formation on bold bets, long before consensus.
The fund’s LPs will include Sequoia, Jeff Dean, Sachin Katti, Philip Wong (my Stanford professor and Chief Scientist adviser to TSMC), and others whose work has shaped our industry and how I think about AI’s future.
I’ll remain a partner at Sequoia and keep serving on my boards on Sequoia’s behalf.
Sequoia has been my home for the past 11 years. I joined as the youngest investor the firm had ever hired. I grew up here. Through years of intense debate and shared work, I found lifelong mentors and friends who took me under their wing and taught me this craft. They taught me to aim high and take big swings. I’ll carry that into what I build: outsized ambition, relentless truth-seeking and curiosity, and a refusal to become complacent.
Working alongside founders has made me want to build something of my own: a firm shaped from day one around the convictions I’ve developed over the past decade.
Life has taught me not to take time for granted. This is how I want to spend mine.
Why now
We’re seeing a new way of building companies in today’s AI era. Research and products often advance together. A breakthrough can open a market or accelerate a business far faster than expected. Some companies require extraordinary capital and compute early on. Others reach meaningful scale with remarkably small teams. The rules for how big these businesses can get, and how fast, are being rewritten.
But early on, the technology is unproven, the resources required enormous, and the vision seemingly implausible.
This is what I love most: seeing the potential right from the beginning, and helping founders realize their full scale of ambition.
More than ever, it starts with an extreme willingness to believe early, with bold, independent conviction in a founder’s technical insight before any product or market validation. Then, it means helping founders take the biggest swings, giving the rare, long-term patience to see those bets through uncertainty, ambiguity, and turbulence, and serving as a level head when the stakes are highest.
I learned this kind of partnership at Sequoia. It’s the foundation of the firm I’m building.
Believing before consensus
My fascination with AI’s potential took root at Stanford in the early-to-mid 2010s, studying computer science and working in Andrew Ng’s lab. The friendships I formed connected me to a community whose people and ideas would help shape modern AI and build some of our generation’s defining companies. What stayed with me was the gap between what those people saw was possible and what the wider world could imagine.
I’m drawn to founders pursuing ambitions that sound impossible today.
I start with a simple question: “What if everything goes right?”
Then we work backward together to make it happen.
In 2023, Deepak and Abhinav set out to build a general-purpose robot brain at @SkildAI, before physical AI was a consensus bet. Their sim-to-real breakthroughs gave me conviction where earlier robotics approaches had fallen short. But the data flywheel required deployment at scale. We made commercialization as urgent as research and brought in strategic partners like NVIDIA and SoftBank to support the ambition. Three years in, Skild is advancing the frontier of physical AI with nine figures in annualized revenue.
In 2024, Ioannis and Misha founded @reflection_ai with an insight into scaling reinforcement learning. As that grew into the ambition to build a Western frontier lab for open models, the hard question was how a startup competes with labs spending tens of billions. An introduction to NVIDIA and many compute calls and recruiting events later, the team has built Beam, now one of the leading US open models.
In 2025, Anna and Azalia, the co-creators of AlphaChip, founded @RicursiveAI to design models and hardware together. From day one, we made early customers and ecosystem partners a priority. Today, they are showing order-of-magnitude improvements on real workloads, working with some of the industry’s most consequential players, and reaching a defining inflection point.
The ambition keeps growing as these companies scale.
@linear has always had a kind of magic to it. Now it’s the system a new generation of AI companies builds on, for teams and agents alike.
@tavus saw early that we’d one day talk to AI the way we talk to each other. Today, it’s building the human computing interface for a world of AI colleagues.
@Mach_Industries began as hydrogen experiments in an MIT dorm room. It has since become one of the most exciting next-generation defense companies.
Making the ambition possible
Being a first believer is only the beginning. Compute falls through. A deep-pocketed incumbent makes your mission its top priority. A breakthrough changes the opportunity and the capital required overnight.
These are stomach-churning moments, and early belief carries a responsibility to help make the bet work.
This is the partnership I love: turning the full ambition into concrete steps. Raising capital, securing compute and strategic partners, recruiting exceptional people, and winning the first customers together. Thinking several moves ahead while solving what’s in front of us. And as breakthroughs open new possibilities, helping founders dream even bigger.
Thank you
To the founders I’ve worked with: thank you for trusting me and Sequoia to be your first partner and first believer. It’s been the joy of a lifetime, and I can’t wait to keep building together.
Thank you to Alfred and Pat for your leadership of Sequoia and your support of my next chapter. To Roelof, for bringing me into Sequoia and believing in me from the beginning. Alfred and Roelof, you have been mentors in investing and in life. So much of what I know, and the investor I’ve become, comes from your guidance and generosity over the years. To Bill, for the close partnership and wealth of knowledge, always wrapped in humility. To Shaun, for showing me how to think deeply and imaginatively. To Michael Moritz, for inspiring 22-year-old me to dream big. To Doug, for teaching me to keep things simple. To Bogie and Jess, for bringing joy to the everyday and perspective on what matters. And to all my Sequoia teammates, for the friendship, camaraderie, and lessons I’ll carry with me.
To my family, especially my husband and my brother, who have each started companies of their own: thank you for giving me the courage to take this leap.
Let’s get to work!
Jeff Dean reposted
I launched the first image and video generation models at Google, and kept asking why all AI designs look the same.
AI has brought us the WORST era of design.
It made content creation effortless, but it also made 90% of it look painfully generic.
My co-founder and I spent 2 years building in stealth to change that.
Meet Vyb, the first AI design model that understands your taste like a designer would.
RT + comment your website URL and we’ll make your first design.
Jeff Dean reposted
Cancer patients with the same diagnosis often have very different tumor biology.
Molecular testing can reveal underlying differences, but that's slow and expensive.
Introducing Holon-S1, our platform to predict genes and biomarkers from routine pathology slides.
• beats two other models at predicting spatial gene activity in a controlled comparison
• has identified potentially novel biomarkers in breast cancer and colon cancer
• evaluated across 16 organ types and more than 15 scanners, using slides from both formalin-fixed, paraffin-embedded and fresh-frozen tissue.
We believe Holon-S1 has the potential to help researchers find new biomarkers in slides they already have, discover treatment targets, and bring more personalized cancer care within reach!
Have a glimpse at this! Nice work, @AndrewDai and Elorian folks!
Spent years watching models ace a single image and completely whiff on anything that moves or changes angle. Today we’re fixing that.
Try Glimpse now, our first visual thinking model with competitive frontier performance compared to Astra on key public visual reasoning benchmarks.
Jeff Dean reposted
My AI security research lab AISLE discovered 500+ zero-days with CVEs across 100+ codebases in a year => ~1% of 2026's open-source CVEs globally from a single team! On hardened codebases like curl, we've repeatedly matched or outperformed frontier AI labs in zero-day discovery
I've been a personal supporter (through our family foundation) of @Khipu_AI & @DeepIndaba, which are fantastic events for AI researchers in Africa and South America. I've encouraged Google to support these in the past. Discovery Loop (@DiscoLoopAI) is thrilled to support!
For the last 10 years we've been raising funds to bring AI to Africa (@DeepIndaba) and South America (@Khipu_AI). The efforts have been very successful, but this year we are finding it hard to raise money. If you work for an AI company, please share the following with the right team to support us. Thank you!
About KHIPU 2027
What it is: KHIPU AI (khipu.ai) is a non-profit organization dedicated to advancing AI talent, research, and community across Latin America. Alongside year-round regional initiatives, we host a biennial in-person gathering rotating across the continent, previously in Uruguay (2019, 2023) and Chile (2025).
Next Edition: April 5–9, 2027, in São Paulo, Brazil, hosted at the Universidade de São Paulo (USP), one of Latin America's leading research universities.
Watch our 2-minute video: www.youtube.com/watch?v=Db19zoquTOs&t=2s (hear directly from past speakers, including Nando, and Latin American students on how KHIPU grew from a single gathering into a regional AI movement).
Who Attends & Why Support - Engaging with the region's top talent and lab leaders: We bring together around 500 participants, including leading faculty and lab directors from across Latin America alongside a curated cohort of ~300 graduate students and industry researchers selected from over 1,000 applicants. As the local AI ecosystem grows rapidly, KHIPU offers a rare opportunity to engage directly with these emerging researchers and regional leaders in one place, helping shape research priorities and the safety mindset across Latin America while connecting with talent and perspectives outside Silicon Valley.
100% free attendance & student grants: Attendance is completely free for all selected students, professors, and researchers. Sponsorship funding goes directly toward travel and accommodation grants for students and faculty across Latin America who otherwise could not afford to attend.
Community & Track Record: Past editions have featured speakers such as @Yoshua_Bengio , @JeffDean , @jpineau1 , and @NandoDF (2025 lineup: khipu.ai/khipu2025/speakers-2025/), and have been supported by the Hopper-Dean Foundation, @GoogleDeepMind , @Apple , @Microsoft , @Meta, @cohere , and @MistralAI . For 2027, @OriolVinyalsML and @sarahookr are already confirmed.
2027 Sponsorship Prospectus: Attached (also available at khipu.ai/prospectus-2027)
Contact: @FortunatoMeire (meirefortunato@khipu.ai) & KHIPU Sponsorship Committee (sponsors@khipu.ai)
Jeff Dean reposted
Introducing DeepRSI, a new industry standard for measuring recursive self-improvement.
On preexisting benchmarks, tasks are contributor-based and short-horizon (~12 hrs). In DeepRSI, we push agents to work 48 hours on ML tasks actually derived from ongoing proprietary research.
Jeff Dean reposted
Wrapping up my first month at @DiscoLoopAI (am on leave from @PrincetonCS and @PrincetonPLI ). Excited to work on using AI to accelerate scientific discovery and solve important problems. My first startup experience, and with an incredible set of colleagues!
Announcing Discovery Loop!
I am very excited to announce that, along with my longtime friends and collaborators @Sanjay_Ghemawat, @OriolVinyalsML and @quocleix, we are founding Discovery Loop (@DiscoLoopAI), a Public Benefit Corporation whose mission is to automate machine learning, science, and engineering to accelerate discoveries and progress. The four of us have worked together for 14 to 30 years, and have helped build some of the world’s most used products, infrastructure and AI models, and we’re excited to turn our attention to this ambitious endeavor.
♾
Learn more at: www.discoveryloop.com
Jeff Dean reposted
We’re sharing new results showing that frontier AI models are able to outperform human experts on financial predictions for the first time. We built an environment for the challenging task of Earnings Predictions, finding that the most recent frontier models (Opus 5.5, Fable 5.1, Astra) with Samaya’s harness, meaningfully outperform expert consensus predictions on commonly reported earnings metrics such as company revenue, gross margin and more.
We built this environment to have a “point-in-time” gate in the harness to ensure no information leakage, and created financial retrieval and data tools that would respect this gate and give the models access to the same information available to experts. We also adjusted for bias in expert consensus to create a hard, bias-corrected baseline, where outperformance has real signal. Only the most recent frontier AI models are able to beat this hard baseline, pointing to an important inflection point in AI capabilities in finance.
Like with FrontierFinance, we see substantial headroom for further RL training in this environment — the models show large variations in error across different rollouts. We’re continuing research on using earnings prediction and other predictive tasks to post-train models to learn from their mistakes. Get in touch if you’d like to collaborate or try out an alpha version. Link to the full research below.
Jeff Dean reposted
Proud to share that @phinity_ai is coming out of stealth today, announcing our seed fundraise and that we’ve reached 8-figures in annualized run rate.
Chip design has been stuck in the 20th century for far too long. Now, for the first time in decades, a unique set of emerging factors is making it possible to fundamentally reinvent how chips are designed.
At Phinity, we’re building toward fully autonomous chip design, with the goal of enabling custom silicon for thousands of emerging AI applications. Our team combines hardware experts, AI researchers, and infrastructure/EDA engineers to define a new way to design chips.
We’re extremely grateful to @uncorkcap for leading our $5.2M seed, to @MoxxieVentures and angels including @JeffDean for joining the round, and to @pearvc for leading our pre-seed and backing us from the start.
We’ve worked so, so hard to make this happen, and I’m proud to share our work publicly for the first time!
You can read more about the announcement below. www.phinity.ai/blog/introducing-phinity
We’re hiring for ops and engineering roles. If you want to help build the future of chip design with us or are interested in learning more/partnering, please reach out to me or @sonyashijin .
Jeff Dean reposted
Excited to finally share that @phinity_ai is now out of stealth! We're announcing our seed fundraise and that we’ve reached 8-figures in annualized run rate.
We’re already working with some of the largest foundation model labs. Our mission is to build fully autonomous chip design so AI can help us search beyond the limits of human engineering and discover computational systems we would never find on our own.
We’re fortunate to have the support of @uncorkcap, which led our $5.2M seed round, with participation from @Moxxie including @JeffDean, following our pre-seed led by @pearvc!
It’s been an amazing few months, and I couldn’t be more grateful for the opportunity to build the future of chip design with such a talented team.
You can read more about what we’re building and why in our full announcement. www.phinity.ai/blog/introducing-phinity
We’re hiring across all disciplines - if you’re interested in working with us, please reach out to me or @aadi_nash.
Very sorry to hear this news. 😥
I never had the chance to meet Margaret Hamilton, but her contributions to the flight software that got us to the moon are legendary. I love this image of her standing next to the printouts of the Apollo code.
We’ve lost an icon. Margaret Hamilton, a computer science pioneer best known for her work at MIT leading development of the flight software for the Apollo missions to the moon, has died at 90. Read about her legacy via MIT News: news.mit.edu/2026/margaret-hamilton-computing-pioneer-die…
Proud to have collaborated with many others on quite a few of these things!
claude just generated this 2 minute video about the history of Google and it goes so goddamn hard
Jeff Dean reposted
We are excited to announce a new faculty position in AI for Biology at UC Berkeley, jointly between @Berkeley_EECS and @CCBBerkeley.
Please repost & share this opportunity with outstanding early-career researchers in your networks.
Apply by Nov 9, 2026: aprecruit.berkeley.edu/JPF05560
The safety data for Waymo gets better and better.
Comparison w/ human drivers on rate of crashes with serious injury:
Latest data: 270M miles, 20X better
Mar 2026: 170M miles, 13X better
Before that (forget exactly when): 10X better
Jeff Dean reposted
Every French🇫🇷 passenger train, on an ordinary Tuesday, from first departure to last
This was recorded a few months ago when I was still at Google and it was a really fun conversation!
How does this only have 21,000 views in 8 days?
Chat with @JeffDean (then Google) and Bill Jia about large scale AI models.
youtu.be/BVQSWeK2Nrw
Did you know you can learn interesting things from computations in real rat neurons? I love what @AlexKsendzovsky and team are doing at @BioComputingCo are doing.
What happens when neurons meet silicon? A new world emerges.
Partnering with Amazon to bring it to customers. DM me for early access.
Jeff Dean reposted
Luiz André Barroso’s work influenced me a lot in how I think about systems, and even about my career.
I decided to write about his work with @JeffDean on some topics that interest me a lot: tail latency and the systems behind the modern internet. andreybleme.com/2026-09-22/luiz-andre-barroso/
Thanks for the great discussion, @dawnsongtweets!
I had the great honor and pleasure of sitting down with @JeffDean for his first public talk since leaving Google, where he spent an extraordinary 27 years.
Few people have shaped modern computing and AI as profoundly - from MapReduce and Bigtable to TensorFlow, Mixture-of-Experts, TPUs, and Gemini.
Our conversation covered some of the biggest questions shaping the future of AI:
• How do you recognize a foundational idea before everyone else does?
• How do you choose a research problem worth spending 5 years on?
• What can coding teach us about building better reasoning models?
• What might recursive self-improvement (RSI) actually look like?
• What happens when the scientific discovery loop itself becomes increasingly automated? (and how is Jeff’s new startup going to contribute in this space?)
• As AI becomes increasingly autonomous, how do we keep it safe and secure?
• What should the next generation of researchers be working on?
Here are some key insights and highlights for anyone building the future of AI. 🧵1/8
Jeff Dean reposted
Today, G5 Labs is emerging from stealth with $14 million in seed funding to build a new abstraction layer for AI-native software development. @pillar_vc and @BatteryVentures co-led the round, with participation from @omegavp , Encoded Ventures, and prominent angel investors including @JeffDean .
G5 Labs is pioneering the post-developer era. A spinout from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL), the G5 Labs platform makes natural-language intent, organized as an ontology, the source code itself—putting software development in the hands of the whole organization while providing the governance needed to keep requirements, implementation, and organizational policies aligned.
We’re excited to build a future where employees across departments, from product owners to analysts and compliance leads, can create and evolve software using the language they already speak.
Read our full announcement: g5labs.ai/press
Read more in VentureBeat: venturebeat.com/technology/should-all-enterprise-code-and…
Exciting results, @LiamFedus! Congrats to the whole team at Periodic Labs!
We built high-throughput materials labs in Menlo Park to create a loop between experiments and models. The labs generate fresh data, the models learn from it, and then help us decide what to try next.
Using only 1,300 H200s, plus months of our experimental data, we mid-trained and RL’d an open-source model to surpass GPT-6 Astra on our analysis benchmark. We call it Neon.
This is real footage from our lab. We’re focusing first on hard problems in materials science, including superconductors, magnets, and semiconductor materials.
Read our blog posts below.
Jeff Dean reposted
Introducing OM-1, our first robot foundation model, zero-shot generalizing to any robot: table-top arms, industrial arms and humanoids.
- learned directly from human manipulation data
- no teleop/robot data
- close to human-level dexterity and efficiency
- multi-robot collab
Jeff Dean reposted
Excited to partner with @DiscoLoopAI ! They are actively hiring across engineering & research - please reach out if the mission sounds interesting to you.
Introducing Discovery Loop ♾. We’re building AI to run experiments at unprecedented scale and to solve the biggest bottlenecks in science and engineering.
Our founders @JeffDean, @Sanjay_Ghemawat, @quocleix, @OriolVinyalsML helped build the AI infrastructure and models that scaled the modern world. Now, we are scaling discovery itself.
The future can't wait.
Learn more at: www.discoveryloop.com
Jeff Dean reposted
1/ ✨ World record today on bounded gaps between primes, one of number theory’s oldest open conjectures.
We show that infinitely many pairs of prime numbers are separated by 212 or less.
Jeff Dean reposted
Stanford HAI's fall conference returns this October. Confronting Our AI Future: Hope, Fear, and the Choices Ahead will feature a full day of conversations on work, learning, governance, and how we ensure AI expands human potential. Speakers include @drfeifei, @JeffDean, @audreyt, @erikbryn, Alondra Nelson, and @istoica05, with more to be announced.
Register by October 1 for early bird pricing: hai.stanford.edu/events/confronting-our-ai-future-hope-fe…
Jeff Dean reposted
Today, we're excited to share early progress on using Gemini to accelerate scientific discovery in the real-world. We present an extension of Co-Scientist which we use to collaborate with scientists across materials science, biology, and computer science.
Jeff Dean reposted
Behind a lot of great AI projects, there is a good @JeffDean story. Here is mine 😉
Back in 2019, even before Khipu AI officially existed, Jeff said yes to participate in our first event in Uruguay 🇺🇾. His early belief in our LATAM AI vision created a ripple effect—if @JeffDean was in, everyone was in. Incredibly grateful he returned for Khipu 2025 in Chile 🇨🇱!
As we look ahead: Khipu 2027 is heading to BRAZIL 🇧🇷
Applications are OPEN: khipu.ai
Want to sponsor us and support Latin American AI? DM me or reach out at sponsorship@khipu.ai
Check out the video below to see what Khipu is all about (and spot a quick Jeff cameo 😉). 👇
youtu.be/Db19zoquTOs
#khipu2027 #khipu_ai #machinelearning #latam
Jeff Dean reposted
Indeed, it has been quite a week, so I thought I'd share it with the KDD 2026 audience! For my Google colleagues who might have sent me internal chat messages after I closed my laptop for a final time on Thursday afternoon, I apologize for my lack of response!
(Zoomed in)
Legendary! Thanks for sharing the behind-the-scenes @JeffDean!
Indeed, it has been quite a week, so I thought I'd share it with the KDD 2026 audience! For my Google colleagues who might have sent me internal chat messages after I closed my laptop for a final time on Thursday afternoon, I apologize for my lack of response!
(Zoomed in)
Legendary! Thanks for sharing the behind-the-scenes @JeffDean!
This slide was a build of a bunch of pieces, with some more pictures and context in the intermediate parts
And the final couple
Jeff Dean reposted
That day was super fun, @gabepereyra! Thanks for inviting me!
What you didn't happen to mention was that it wasn't just a normal USC club soccer team practice session, but actually was tryouts day. So there were like 80 skilled undergrads there, all eager to make the team. So it was divided up into a bunch of small-sided 7v7 games. I like to think 45-year-old me kind of held my own, but I'm sure everyone was thinking "Who's the old guy playing in our 7v7 match?" 😅
Jeff Dean reposted
Incredibly excited & proud to announce @radicalvcfund co-lead investment partnership with @JeffDean, @OriolVinyalsML, @quocleix and @sanjay_ghemawat in their launch of @DiscoLoopAI. Perhaps the most pedigreed AI founding team ever. Thrilled to help them fulfill their mission.
Announcing Discovery Loop!
I am very excited to announce that, along with my longtime friends and collaborators @Sanjay_Ghemawat, @OriolVinyalsML and @quocleix, we are founding Discovery Loop (@DiscoLoopAI), a Public Benefit Corporation whose mission is to automate machine learning, science, and engineering to accelerate discoveries and progress. The four of us have worked together for 14 to 30 years, and have helped build some of the world’s most used products, infrastructure and AI models, and we’re excited to turn our attention to this ambitious endeavor.
♾
Learn more at: www.discoveryloop.com
Jeff Dean reposted
@JeffDean has built the foundations that every era of modern computing has stood on, from web-scale infrastructure to scaling deep learning itself. Exceptional people inspire great talent and the founding team around him is among the strongest I have ever seen in over 25 years of investing. @lightspeedvp is honored to join them from the start of this journey.
Announcing Discovery Loop!
I am very excited to announce that, along with my longtime friends and collaborators @Sanjay_Ghemawat, @OriolVinyalsML and @quocleix, we are founding Discovery Loop (@DiscoLoopAI), a Public Benefit Corporation whose mission is to automate machine learning, science, and engineering to accelerate discoveries and progress. The four of us have worked together for 14 to 30 years, and have helped build some of the world’s most used products, infrastructure and AI models, and we’re excited to turn our attention to this ambitious endeavor.
♾
Learn more at: www.discoveryloop.com
Jeff Dean reposted
When I was an undergrad at USC, I emailed @JeffDean inviting him to come play soccer with the club team after his talk on campus. To my surprise he ended up moving a meeting with the dean to come and play with us.
What is truly special about Jeff is that despite being responsible for some of the most impressive engineering and AI work in history, he still finds time for everyone - even if it means taking a break from inventing tensorflow to kick a ball around with an undergrad who looked up to him.
I was fortunate to see some of the great work Jeff,
@Sanjay_Ghemawat, @OriolVinyalsML and @quocleix did at Brain and DeepMind. They are truly unique - so excited to see what this team can build.
Announcing Discovery Loop!
I am very excited to announce that, along with my longtime friends and collaborators @Sanjay_Ghemawat, @OriolVinyalsML and @quocleix, we are founding Discovery Loop (@DiscoLoopAI), a Public Benefit Corporation whose mission is to automate machine learning, science, and engineering to accelerate discoveries and progress. The four of us have worked together for 14 to 30 years, and have helped build some of the world’s most used products, infrastructure and AI models, and we’re excited to turn our attention to this ambitious endeavor.
♾
Learn more at: www.discoveryloop.com
When I was an undergrad at USC, I emailed @JeffDean inviting him to come play soccer with the club team after his talk on campus. To my surprise he ended up moving a meeting with the dean to come and play with us.
What is truly special about Jeff is that despite being responsible for some of the most impressive engineering and AI work in history, he still finds time for everyone - even if it means taking a break from inventing tensorflow to kick a ball around with an undergrad who looked up to him.
I was fortunate to see some of the great work Jeff,
@Sanjay_Ghemawat, @OriolVinyalsML and @quocleix did at Brain and DeepMind. They are truly unique - so excited to see what this team can build.
Announcing Discovery Loop!
I am very excited to announce that, along with my longtime friends and collaborators @Sanjay_Ghemawat, @OriolVinyalsML and @quocleix, we are founding Discovery Loop (@DiscoLoopAI), a Public Benefit Corporation whose mission is to automate machine learning, science, and engineering to accelerate discoveries and progress. The four of us have worked together for 14 to 30 years, and have helped build some of the world’s most used products, infrastructure and AI models, and we’re excited to turn our attention to this ambitious endeavor.
♾
Learn more at: www.discoveryloop.com
That day was super fun, @gabepereyra! Thanks for inviting me!
What you didn't happen to mention was that it wasn't just a normal USC club soccer team practice session, but actually was tryouts day. So there were like 80 skilled undergrads there, all eager to make the team. So it was divided up into a bunch of small-sided 7v7 games. I like to think 45-year-old me kind of held my own, but I'm sure everyone was thinking "Who's the old guy playing in our 7v7 match?" 😅
Jeff Dean reposted
Btw since there were a bunch of Jeff Dean stories being shared recently, I thought I'd share mine:
Jeff Dean was the second investor in @SophontAI. We got connected through a founder friend. After briefly reviewing our pitch deck, he immediately decided to invest. He had a high level of conviction in our mission much earlier than many others.
Something I was honestly surprised about is his level of support for our company. As the Chief Scientist at Google, I figured he would be extremely busy but whenever we have any asks of our investors he quickly responds with suggestions and he always tries to find ways to help. I don't know how he has time to do it!
I am so grateful and honored to have Jeff Dean as an investor in Sophont, not just because he has been a massive inspiration for me as an engineer and medical AI researcher, but because of his early conviction and steadfast support of our mission.
Jeff Dean reposted
I also want to give a huge thanks to the incomparable @JeffDean after an incredible 27-year run at Google. He’s off to start his own public benefit corporation with @Sanjay_Ghemawat focused on accelerating discoveries across ML, science, & engineering. @Google will support as a founding investor and Cloud partner. On a personal note, it’s been a privilege to work with Jeff and Sanjay, and I wish them all the best. Thank you for everything!
Tomorrow will be my last day at Google after 27 years, and watching it grow from 25 people to 190,000+ has been an amazing journey. Below is a note I shared with many people internally at Google today. An excerpt is:
It has been an absolute pleasure to work with you and to help build some of the most widely used and impactful products of all time. As a kid, I dreamed of helping build software that would be used by many people, and Google now has thirteen products used by more than a billion people (amazing!). Our work has had a tremendous impact in the world, and I have been lucky enough to collaborate and form friendships with many colleagues that I deeply admire, respect, and enjoy. It still brings me joy every time I see people out in the world using our products to find information, handle email, translate documents, watch videos, learn new things, navigate and understand the physical world, browse the web, use their phone, run large-scale computations on our infrastructure, ride in an autonomous vehicle, or perform complex tasks with the help of our AI systems. I hope you all share this sense of joy, because it is a shared accomplishment! Thank you to all of my colleagues at Google over many years!
Now I'm excited to go start @DiscoLoopAI with my longtime friends and colleagues @Sanjay_Ghemawat, @OriolVinyalsML, and @quocleix.
(Updated post: slightly redacted to not have some personal info)
Turned it in...
Announcing Discovery Loop!
I am very excited to announce that, along with my longtime friends and collaborators @Sanjay_Ghemawat, @OriolVinyalsML and @quocleix, we are founding Discovery Loop (@DiscoLoopAI), a Public Benefit Corporation whose mission is to automate machine learning, science, and engineering to accelerate discoveries and progress. The four of us have worked together for 14 to 30 years, and have helped build some of the world’s most used products, infrastructure and AI models, and we’re excited to turn our attention to this ambitious endeavor.
♾
Learn more at: www.discoveryloop.com
We created a pitch deck to tell a handful of VC firms about us and what we were up to (a fun experience!). Here’s a few slides about our background and some of the things we’ve worked on from the pitch deck (it was fun putting together the list of people in our teams who have gone on to found a whole range of exciting companies). We are delighted to have selected @radicalvcfund
and @khoslaventures to lead our initial funding round, along with participation from @lightspeedvp, @kleinerperkins, Doerr Capital (@johndoerr), and Alphabet (@Google). We’ll be working with them to close our seed round over the next few weeks.
Our immediate order of business is to find office space, and hire an amazing founding team over the next few weeks. We want to create an awesome environment with a great culture of technical excellence, teamwork, respect, and ambition. We’ll also start building our infrastructure and AI models and systems to tackle our first domain: automating large-scale experimentation for ML research and engineering. In doing so, we’re going to be our own first customers. The rapid feedback from doing that is the way to build something amazing.
Oh, we’re hiring! See discoveryloop.com
One more fun slide from our pitch deck.
Oh look! My longtime friend and colleague @Sanjay_Ghemawat is now on Twitter! Give him a follow!
In April, '17, @jsomers of @NewYorker reached out & said he wanted to do a small profile of me & my longtime colleague Sanjay Ghemawat, watch us work for a few hours, maybe dinner, etc.
It came out today. I think it captures our working style really well. www.newyorker.com/magazine/2018/12/10/the-friendship-that…
Jeff Dean reposted
1/15
If you could ask the people building the future of Agentic AI one question, what would it be?
This Aug 1–2, ~5,000 attendees will gather at UC Berkeley, with 100,000+ joining online, for Agentic AI Summit 2026, the largest gathering dedicated to Agentic AI.
Join us to hear from the highest concentration of leading voices in Agentic AI and get the most comprehensive, in-depth view of where the field stands today and where it's heading next.
👉 Register today to join online for free (in-person tickets sold out): rdi.berkeley.edu/events/agentic-ai-summit-2026
👇Here's a preview of some of the exciting conversations happening this weekend.
👇👇 Help shape the conversation at the summit: What questions would you like the speakers to tackle? What topics would you like to see covered? Reply below to help guide the discussion!
Thanks for an engaging conversation at @ChaseCenter last weekend, @sdianahu! My first in a basketball arena!
In 2001, @JeffDean and Sanjay Ghemawat did the math and realized Google’s entire search index would fit in RAM — then shipped it in a few days, and search got fast.
In 2013, another napkin calculation showed that three minutes of daily speech recognition per user would require doubling Google’s server fleet. That one became the TPU.
At Startup School 2026, Google’s Chief Scientist talks with YC’s @sdianahu through the thought experiments behind both, why inference hardware is the next specialization, and where two or three people in a room can still win.
00:07 — Are AI Models Already Junior Engineers?
01:44 — AI Systems That Improve Themselves
02:40 — The Google Search Breakthrough That Changed Everything
04:38 — AI Agents Will Run for Weeks
05:58 — The Napkin Math That Led to TPUs
09:20 — How to Find Breakthrough Ideas
10:25 — The AI Engineer's New Mental Model
12:33 — Why AI Is Really an Energy Problem
16:11 — Context Engineering Is the Next Frontier
19:46 — The Skill That Made AI Better at Optimization
22:13 — Why Long-Running Agents Fail
25:21 — Where Startups Can Still Beat Google
31:19 — How to Become an AI-Native Founder
36:36 — Question Your Biggest Assumptions
42:08 — AI That Builds Better AI
50:02 — Build Something That Truly Matters
Jeff Dean reposted
I met @JeffDean on the pitch over a decade ago. We played against each other for years, I only knew him as Jeff, the fast winger I always had to mark. One night he took a hard pass to the head and went down to the ground. I ran over and started doing concussion protocol through a series of questions. When I asked him to tell me the name of his manager, he replied, "@sundarpichai ". I looked at my friend and said, "Jeff has a concussion, he just told me his manager is Sundar". My friend smiled and replied, "Dude, that's Jeff Dean".
Nice! Congrats! And it is true, @ceo_clickhouse is indeed a true football/soccer lover and player: he and I face off against each other regularly in the soccer league we play in!
I've been to a game at the renovated Craven Cottage and it is quite nice!
Jeff Dean reposted
Today, SkyPilot is out of stealth.
Building custom intelligence is now existential. We help frontier AI teams build intelligence faster by removing their biggest bottleneck: AI compute fragmentation.
Frontier teams like @appliedcompute, @AbridgeHQ, @hippocraticai, @hcompany_ai, and @nubank already run on SkyPilot, with 10x faster time-to-intelligence and double-digit increase in GPU utilization.
AI teams today get compute anywhere they can. They then firefight compute fragmentation across providers. Researchers burn time on workload setup. Infra gets paged when GPUs go down. Frontier teams build slowly even on the fastest compute.
@skypilot_org turns your fragmented compute into one AI supercomputer, so you run frontier workloads faster. Many users manage 10,000+ GPUs across providers with SkyPilot. GPU hours consumption has grown 6x in the last 6 months.
1/ We're launching SkyPilot Platform — the AI compute platform for frontier AI teams to manage large GPU fleets and accelerate building custom intelligence.
Optimized for fleet management, team governance, and frontier workloads — pretraining, post-training, multi-cluster serving, and sandboxes. SkyPilot open source users can switch to the platform with a server URL change.
2/ We've raised over $20M led by @Lux_Capital (@breeves08), with participation from @AmplifyPartners (@dauber, @lennypruss), @coatuemgmt, @FoundationCap (@ashugarg , @JayaGup10), @RaceCapital, @thehousefund, and top operators like @alighodsi (CEO, Databricks), @JeffDean (Chief Scientist, Google), @rauchg (CEO, Vercel), @amasad (CEO, Replit), @ClemDelangue (CEO, @huggingface) and more.
We're hiring across Engineering and GTM to deliver the platform for the next decade of AI.
Above all, I'm excited to be building with the incredible team we've assembled, along with my cofounders Zhanghao @Michaelvll1, Romil @bromil101, Scott, and Ion @istoica05.
If you firefight AI compute, let's build.
Nice! Congrats! And it is true, @ceo_clickhouse is indeed a true football/soccer lover and player: he and I face off against each other regularly in the soccer league we play in!
I've been to a game at the renovated Craven Cottage and it is quite nice!
A nice aspect of our new Gemini 3.6 Flash model is that it is much more token efficient than our 3.5 Flash model. Here's a side-by-side demonstration of that.
Nice work by everyone who worked on this model and release!
We’re rolling out three new models to make AI agents faster, smarter, and cheaper at scale:
🔵 Gemini 3.6 Flash: It uses fewer tokens than 3.5 Flash to deliver higher quality work at the exact same cost.
🔵 Gemini 3.5 Flash-Lite: A fast, cost-effective option for everyday tasks like processing documents and agentic search.
🔵 Gemini 3.5 Flash Cyber: A cybersecurity model built to find and patch critical software vulnerabilities.
Yesterday I was fortunate enough to go to my first-ever World Cup game, with my long-time colleagues/friends @OriolVinyalsML and @quocleix and their spouses). We were sitting in the corner area and had a quite good view of the goal!
I present to you a 43 second multi-part drama filled with emotion:
The initial promising-looking cross coming in, but looking overhit
Nico Williams cleverly knocking it down at the back post into a dangerous area
Ferran Torres striking it cleanly into the roof of the net
The crowd rising as one (forcing me to stand up as well)
The elation of the Spanish players racing off the bench to celebrate
The dejection of the Argentinian players, their defense having finally been breached in extra time
The elation of my Spanish colleague Oriol and his wife Meire next to me (he and I are both Barça fans, so it was nice to see a Barça player score the winning goal)
The entire stadium reacting
Whew!
Jeff Dean reposted
It's kinda fun that "pod" is becoming de facto industry standard for "bunch of highly connected accelerators"
AFAIK "TPU Pod" was an ocean pun (a group of whales/dolphins) because TPUs have ocean animal codenames (jellyfish, dragonfish, ...)
cc @JeffDean correct?
After the K3 drop, Huawei has now released a real 950 SuperPoD
1 EFLOPS fp8 & 2 EFLOPS fp4
256TB of unified memory
Jeff Dean reposted
I have some news. But first a little backstory.
Almost exactly four years ago, I drove down to Mountain View for my first day as a Google employee. I was fifty-four years old, and somehow had managed to avoid ever having a proper job over that entire period, other than a few startups that I had founded in my twenties and thirties. At that point Google Labs, which had just been formed a few months earlier, was headquartered in a fairly anonymous looking suburban office park off 101, closer to Office Space than Silicon Valley on the Mike Judge spectrum. And as it happened, my first day at Google was a Monday in the summer of 2022, which meant COVID-era work from home policies were still the default, and the office was absolutely deserted. But I’d been set up with a badge and a computer and I had to schedule a phone call with tech support to get my account in order, and so I grabbed one of the empty cubicles and sat down to work through the problems I was having getting into the system. I remember feeling a definite sense of dread in that moment, sitting there in the silence, thinking: have I made a terrible mistake?
And then—and I swear to god this actually happened—the silence was interrupted by a lonely robot rolling its way across the carpet past my borrowed desk.
Before the end of the summer, though, an inspiring cast of characters began to populate the office, and by October a small group of us had built a prototype of an AI-first tool for thought that my colleague Adam codenamed Tailwind. About six months later, after @joshwoodward previewed Tailwind at I/O for the first time, we set about to give our little experiment a proper name. We had some branding consultants spend an afternoon with the team; we ideated and free associated and placed our post-it notes on the whiteboard, but all the names that they came back with were too impressionistic, too evocative — they didn’t sound like iconic Google names: Maps, Docs, Meet. Because I’d written so much about the history of journals and commonplace books, I had always been partial to simply calling it Notebook. There was something about that framing that just made sense: a notebook is the place where we capture information that we want to remember, where we gather our thoughts. It gives us a dedicated space for thinking, where new connections happen, where we can trace the evolution of our own ideas. And it had a simplicity that fit the Google house style.
But we were far too experimental and unproven to take on a name like Google Notebook—remember we had exactly zero actual users at this point—and so we found ourselves a bit stuck: too early-stage for a proper grownup name, but ill-inclined to adopt a name that didn’t sound like the classic Google brands. But at the time there happened to be another Labs experiment that had just launched called MusicLM, and so at some point in late May of 2023 it occurred to me that we could borrow that convention for our embryonic product, and call it NotebookLM. It wasn’t the most elegant name you could imagine, but it had a certain logic to it: we were building a notebook reinvented for the age of language models, after all. I wrote an email to a few folks proposing the idea and included the caveat that if the product ever took off, and Google wanted to elevate us into the canon, we could drop the LM and just be Notebook.
And so that’s the news. Three years and somewhere near a billion notebooks later, we are finally ditching the LM.
@Gemini_Notebook
Jeff Dean reposted
SPANISH BROADCAST CALL OF ARGENTINA'S WINNING GOAL IS ELECTRIC 🔥
(Via @TelemundoSports)
Jeff Dean reposted
Six months ago in @nytimes, I argued that the data on driverless cars was becoming overwhelming and that needless barriers were costing lives. But data won’t win our trust. Today in @NoemaMag, I explain why.
Being safe and feeling safe are not the same.
What’s puzzling is that where these cars operate, the numbers aren’t close: 94% fewer serious-injury crashes than human drivers on the same roads, across 220 million driverless miles. Plummeted pedestrian and intersection crashes. Peer reviewed research confirms these findings. But trust hasn’t followed.
Part of the answer lies in the signals we read. A cautious machine can look confused. A failsafe can look like a failure. One visible mistake will outweigh millions of clean miles. When a human makes a mistake, we can see ourselves in it. We’ve all looked down at the wrong time. When a machine makes the mistake, there’s no one to blame and no one to forgive. It’s an opaque version of fate we can’t accept. Data shows that we will choose a worse human over a better algorithm.
Medicine has been here before. When we are patients, we can’t usually read care quality directly. So we read what we can see: the bandage on the floor, stained ceiling tiles, how staff talks to each other. Healthcare spent 30 years building the discipline of patient experience and writing this playbook. Today hospitals are measured and paid on it.
The driverless car industry has built the safety half. This essay is about the other half. Felt safety. We really need to get this right. Cars are only the first. Our future has us living surrounded by embodied autonomous systems.
"Why We Demand Perfect Machines Yet Tolerate Human Carnage." Link below.
Jeff Dean reposted
Fine-grained subtask annotations improve VLA performance.
We trained a vision + proprioception model that’s now SOTA at generating them, generalizing across embodiments.
93.1 F1@50 on REASSEMBLE, 98.6 on @Amazon Robotics blade insertion. Details below 👇🧵
Jeff Dean reposted
Excited to be releasing FrontierFinance, the largest and most challenging open benchmark for evaluating AI agents across the full investment workflow!
FrontierFinance is substantially harder than current finance benchmarks: Existing benchmarks like FinanceBench and Finance Agent focus almost entirely on data extraction.
FrontierFinance spans diverse use cases across the full investment process: Screening & Discovery, Company Research, Sector/Industry/Macro, Earnings & Events, and Coverage & Catalyst Monitoring.
Created for ambiguous, long-horizon agents: 220 examples paired with 11,543 expert-crafted rubrics, following Samaya's Criteria Eval methodology. The rubrics are what let us evaluate the reasoning and steps behind a true expert-level output, not just a plausible-looking one.
Evaluations: We evaluated Claude Fable 5, Claude Opus 4.8, GPT 5.5, Gemini, open-source models including GLM and DeepSeek, and others. We used the same public rubric and a standard harness for financial tasks. Samaya's AI system reached state-of-the-art accuracy at 50.8%, at 4x lower inference cost than Fable 5. Next best was Fable 5 (49.2%), then Opus 4.8 (45%) and GPT 5.5 (43.5%).
We're releasing the benchmark, methodology, and full evaluation results - see link in comments.
Future releases: FrontierFinance was curated from Samaya's larger internal set of ~5,000 examples, and we plan to release subsequent, harder benchmarks as well as a more detailed technical report!
Jeff Dean reposted
Happy to share we just published Gemma 4 technical report! Take a look
Jeff Dean reposted
We talk about optimizing AI for latency, throughput and cost.
it’s time to add a fourth metric:
🌍 Environmental efficiency.
Curious how researchers like @ylecun, @JeffDean and others think about making energy and water first-class AI metrics.
Jeff Dean reposted
U.S. Soccer are fools if they don’t book a Cabo Verde friendly at an NFL stadium in 2027.
The video from the commencement speech I gave a couple of weeks ago at @uwcse is now up. Congratulations again to all the graduates, and thanks for having me, @MBalazinska!
youtu.be/Pvkmwh7Mito?t=198 (video is of whole ceremony: my part starts around the 3 minute mark)
I enjoyed giving the commencement address at the University of Washington Allen School @uwcse graduation this evening. So many happy students and their families and friends!
Congratulations to all the graduates of the class of 2026! 🎓
Thanks for inviting me, @MBalazinska!
Jeff Dean reposted
New paper: every law in America is technically public. But not really, until now!
With @DenisPeskoff at UC Berkeley, we built a corpus of ~every publicly accessibly city and county law, and released a huge chunk of it!
2.2 million laws, you're (probably) covered in it!
🧵
Jeff Dean reposted
Today, @aadityasubedi_ and I are excited to introduce @architectlabs.
We are building the AI system to design and provably verify chips for the world's most demanding workloads.
AI scaling is fundamentally changing the economics of hardware infrastructure. As models scale, become more capable, and more widely deployed, the bottleneck is shifting from software and models alone to the physical infrastructure that runs it: specialized compute, memory, networking, interconnects, and full-stack system design.
General-purpose hardware is no longer enough, and the world is racing to spin up new chip programs. This is not just true across datacenter training and inference, but everywhere AI enters the physical world: robotics, autonomous systems, spatial computing, defense, personal devices, industrial automation, and scientific instruments.
But designing a chip today remains one of the most gated efforts in modern technology.
A modern chip program takes years, costs hundreds of millions of dollars, and depends on a shrinking pool of expertise concentrated inside a small number of companies.
Architect Labs is a foundational lab building an AI system that designs and provably verifies chips end-to-end. We partner with semiconductor and workload companies, AI labs, and nations to turn demanding workloads into purpose-built chips, on demand at scale. We aim to drastically accelerate chip design, so that the models, software and chip designs can co-evolve together, accelerating the industry’s path to superintelligence.
Two decades ago, the fabless revolution made it possible to build a chip company without owning a fab. TSMC made world-class manufacturing available to anyone with a design. We aim to do the same for chip design itself: enable any organization with a workload, or specification, to get a purpose-built chip design that unlocks scale and distribution of intelligence impossible with current hardware paradigms.
Our founding team collectively has taped out 80+ production chips, led $10B datacenter product lines, been core contributors to Meta’s AI silicon, architected and designed one of the first neuromorphic chips out of Intel, led research teams at Anthropic, xAI, and Google DeepMind, and contributed to fundamental AI research across nearly every frontier lab.
We have already partnered with semiconductor companies to accelerate their chip programs, and some of our AI-generated chip designs are going to tape out on leading-edge foundry nodes later this year.
We’ve also raised a $24M seed round led by @stevejang from @KindredVentures , with participation from @TQVentures , @RaceCapital , @scaletogether , @ora, and Link Ventures.
We are grateful for the support of our angels and advisors, including @snsf , @lukaszkaiser , @AravSrinivas , Kunle Olukotun, @tlbtlbtlb, @alexwg, Siddharth Nath, Thierry Tambe, @arashf , @ekaurghar, @CHHubbell, Selene Casabal, @semiDL , and engineering leaders from OpenAI, NVIDIA, Google DeepMind, Intel and more.
The next great scaling law may not come from the models alone. It will come from making the physical substrate of intelligence programmable. We exist to bring this future to life.
Jeff Dean reposted
Stephen Colbert is wearing a tan suit at the Obama Presidential Center grand opening! 🤣💙
Jeff Dean reposted
Everyone says the latest AI agents will be "job-ready" soon, especially after the release of Fable 5 this week. But is that really the case?
Over the past many months, my group and collaborators have been building Agents' Last Exam (ALE), a benchmark designed to test exactly that claim on real digital labor-market work.
My group and collaborators previously have created many of the benchmarks the field runs on, including MMLU, MATH, CyberGym, and ExploitGym. Today, I'm excited to share Agents' Last Exam (ALE): a rolling benchmark that measures whether AI agents can actually perform economically valuable work across a broad range of real-world domains.
With ALE, we evaluated Fable 5, GPT-5.5, Composer 2.5, and other frontier agent systems across more than 1,500 expert-sourced tasks spanning 55 occupations.
The result is both impressive and sobering.
Today's agents can solve a meaningful fraction of professional tasks. But when we look at the hardest tasks, the ones requiring sustained reasoning, deep domain expertise, and reliable execution over long horizons, they are still far from human-level performance.
On ALE's hardest tier, every frontier agent we tested, including Fable 5, achieved a 0% success rate.
The age of useful agents is here.
The age of truly job-ready agents is not.
We hope Agents' Last Exam (ALE) will serve as a new guidepost and north star for developing agents capable of reliably performing economically valuable work across a broad range of domains.
🧵
My @Google colleagues @NormJouppi, Sridhar Lakshmanamurthy, Cliff Young, and David Patterson recently wrote a paper that will appear in the July/August 2026 edition of @ieeemicro titled "Google's Training Supercomputers from TPU v2 to Ironwood: Architectural Stability, Scale, Resilience, Power Efficiency, and Sustainability Across Five Generations". It's chock full of interesting data about the evolution of TPU chip generations, as well as how workloads at Google have transformed over time (hint: lots more transformer-based models!), and how the generations have gotten ~30X more energy efficient per flop.
Lots of changes over these generations:
Air cooling in TPUv2 to water cooling in TPUv3 onwards
2D to 3D torus-based interconnects
30X improvement TFLOPS/Watt
256 chips (TPUv2) to 9216 chips (Ironwood) per pod
Read the full paper: arxiv.org/abs/2606.15870
Jeff Dean reposted
Together with my co-founders Michael @MichaelPoli6, Stefano @Massastrello and Armin @athmsx, I am excited to announce @RadicalNumerics is emerging from stealth with a $50M seed round to build general biological intelligence.
We’re also sharing an early preview of our new model Omnii, the most powerful genome language model to date.
Omnii preview link:
www.radicalnumerics.ai/blog/radical-numerics-seed
At Radical Numerics, our mission is to master the code of life, and to drive the frontier of biological AI for both design and defense.
This is our dual mandate, which comes from something our own team helped make possible.
Our founding team trained Evo and Evo 2, the largest biological AI models (40B params) trained on DNA sequences. Trillions of tokens across all of life, from microbes to mammals. It’s fully open source, and created the field now known as generative genomics.
Last year, scientists used Evo to generate the world’s first complete genome from scratch using AI. Turns out it was a bacteriophage—a type of virus. It functioned in the real world, and in this case it was harmless. But for us, it was a clear turning point.
It showed that AI is no longer just analyzing biology. It is on the cusp of generating functional lifeforms. Eventually, AI will have the power to design and control life itself.
That should make all of us incredibly excited, and incredibly uneasy. (Anyone can design DNA with a new function, and have it synthesized and delivered, like something from Amazon Prime).
The same technology that will help us cure cancer is the very technology that might create the next global pandemic, or worse, allow the creation of bioweapons that can wipe out populations.
We believe these forces are inseparable. If you work on the frontier of biology, you have to build technology to safeguard it from its misuse. Existing biosecurity tools are sorely losing the arms race, relying on outdated “have I seen this exact thing before?” style algorithms.
We founded Radical Numerics to turn the tide.
And we can’t do that by training on textbooks and natural language. We must understand the language of biology from the raw physical data itself, to reason across every molecule and modality, from DNA to proteins.
The next frontier for AI goes far beyond chatbots or video generators to models that can understand and engineer life.
Today, we’re previewing Omnii, which is already far surpassing Evo 2, and will continue improving as we scale and add new modalities (training now).
1. For human health, Omnii can read and write whole genomes (more on writing later). It’s state of the art (SOTA) on detecting causal variants for disease, and can rank Alzheimer's mutations zero-shot. We’re partnering with a diagnostics company to use Omnii for early cancer detection (pancreatic and multi-cancer).
2. For defense, Omnii is SOTA at detecting AI-generated pathogens. We benchmarked existing detection tools, and they simply can’t detect the AI-generated ones (“deepfake viruses”). We’re partnering with a US national lab to pilot Omnii for detecting the next pandemic, both natural and AI-generated.
We have a data center full of Blackwells in construction now to build the most powerful biological AI models ever. This mission takes a new kind of AI lab that can actually scale on physical, biological data: new alignment research (mid/post training), scaling long context, building out mech interp teams to dissect what these models learn, new architectures and systems designs, all from the ground up.
Our team is made up of AI researchers and scientists from top labs and institutions (e.g. Stanford, MIT, Google DeepMind), but more importantly, we all share the belief that this is the most important challenge of our lifetime. If you feel similarly, we are hiring. We aim to bring the brightest minds in AI and science together to save lives.
Thanks to our partners on this journey, led by Emergence Capital @emergencecap, with Obvious Ventures @obviousvc, Triatomic @TriatomicCap
, and Patrick Collison @patrickc. Our advisors include Eric Horvitz @erichorvitz, CSO of Microsoft, Chris Re @HazyResearch of Stanford, George Church @geochurch of Harvard, and Andrew Weber @AndyWeberNCB, former Assistant Secretary of Defense for Nuclear, Chemical and Biological Defense Programs.
Fortune article: fortune.com/2026/06/15/exclusive-ai-dna-radical-numerics-…
Jobs: www.radicalnumerics.ai/join-us
I really enjoyed this game today. Vozinha and the whole Cape Verde team were amazing against Spain!
A good essay by @pgasawa and @profjoeyg on a more nuanced view of AI advances.
The AI community seems to increasingly be heading towards a polarized world when discussing safety and consolidated power. I see this discourse as a false dichotomy, so @profjoeyg and I wrote an essay on how we need to change the conversation (link below).
Jeff Dean reposted
The AI community seems to increasingly be heading towards a polarized world when discussing safety and consolidated power. I see this discourse as a false dichotomy, so @profjoeyg and I wrote an essay on how we need to change the conversation (link below).
I enjoyed giving the commencement address at the University of Washington Allen School @uwcse graduation this evening. So many happy students and their families and friends!
Congratulations to all the graduates of the class of 2026! 🎓
Thanks for inviting me, @MBalazinska!
Here are a few of the pieces of advice that I shared:
Here's a nice @GeekWire article by @ToddBishop about the graduation, including a few more snippets of what I said in my talk.
www.geekwire.com/2026/what-a-longtime-google-ai-leader-to…
People replace their phones every ~4 yrs. This means there are hundreds of millions of old phones discarded each year that are still perfectly usable as computing devices. @Google in collabration with @UCSD is exploring how to turn these old phones into cloud-computing “phone clusters”. Putting phones back in service in this way can directly reduce the environmental footprint of computing by avoiding the need for further raw material extraction, and taking advantage of the embodied carbon already incurred from manufacturing these devices, and modern phones actually are already quite powerful computers. Read more in the blog below ⬇️
Today on the blog, we discuss a pathway for the second life of phones through the exploration of “phone cluster computing”, which can directly reduce the environmental footprint of computing by avoiding the need for further raw material extraction. More →goo.gle/4aJe5vO
Quite interesting thread on capabilities of real biological neurons (spoiler: they're way more capable than classical artificial neurons in a perceptron) . Nice work @IdoAizenbud and collaborators!
What can a neuron compute?
Real biological neurons are complex, but how capable are they?
Using a new method, we found that a single cortical neuron can classify cats vs dogs, recognize spoken words, and solve 10-bit parity, all tasks thought to require entire networks. (1/15)
Jeff Dean reposted
For medical information, general AI frontier models (Google, OpenAI, Anthropic) outperformed specialized @openevidence and @UpToDate as assessed by 12 US clinicians, randomized and blinded to which model and extensive testing/benchmarks. This was not anticipated. @NatureMedicine
www.nature.com/articles/s41591-026-04431-5
Denying entry to a Somali soccer official selected as one of the World Cup referees is quite shameful. The whole point of the World Cup is as a spirited athletic competition that brings us together, and allowing him to officiate is obviously the right thing to do.
Speech translation has been one of the longest-running ML efforts at Google, and we’ve come a long way. Gemini 3.5 Live Translate is our latest speech-to-speech model, supporting 70+ languages. It enables more natural conversations across languages in everyday products and apps.
Here’s an example of how partners at @InsideGrab are helping connect travelers with drivers. 🚗
Rolling out in Google Translate and via the Live API in @GoogleAIStudio.
You can read more about this here: blog.google/innovation-and-ai/models-and-research/gemini-…
Jeff Dean reposted
Welcome to CVPR! Google is proud to be a Platinum Sponsor of CVPR 2026 in Denver! 🏔️ Visit the Google booth (#557) @CVPR to see how we're pushing the boundaries of CV, see live demos and chat with our researchers. #cvpr2026 goo.gle/4tzRdFQ
Congratulations, Ayanna Howard @robotsmarts! Spelman has made an excellent choice to have you as its next president!
Today, Spelman College proudly announces Dr. Ayanna Howard as its 12th President. This historic transition marks the next chapter in Spelman's legacy of excellence, empowering future generations of Black women to lead, innovate, and shape a rapidly changing world. A pioneering roboticist and higher education leader, Dr. Howard brings a bold vision for Spelman's future while honoring the mission and values that have defined the College for more than 140 years. Read more: www.spelman.edu/news/2026/06/spelman-college-appoints-dr-…
Jeff Dean reposted
Our open letter today on autonomous driving and prevention of serious accidents and road deaths
w/ @slotkinjr Adding signatories for those supportive
opmed.doximity.com/articles/an-open-letter-from-the-medic…
Check out our Gemma 4 12B model: it's a super capable open weights model that can run directly on your laptop.
Meet Gemma 4 12B!
A unified, encoder-free multimodal model designed to bring high-performance intelligence directly to your laptop, and released under an Apache 2.0 license.
Bridging the gap between edge efficiency and advanced reasoning. Here is what’s new with Gemma 4 12B: 👇
This is great to see. The scientific community & the general public should be in the habit of celebrating amazing scientific achievements with standing ovations 🎉
This is not simply a new pancreatic cancer drug.
It is a reminder that even “undruggable” biology can become treatable with persistence.
Daraxonrasib doubled median OS vs chemotherapy in RAS G12 metastatic pancreatic cancer: 13.2 vs 6.6 months.
A remarkable ASCO moment.
#asco26
@DrChoueiri @TiansterZhang @CathyEngMD @montypal @tompowles1 @brian_rini @cdanicas @GlopesMd @PGrivasMDPhD @nataliagandur @yekeduz_emre @neerajaiims @ASCO @ONCOassist @OpenMedicineHQ @MedwatchKate @scserendipity1 @CParkMD @urotoday @OncLive @crisbergerot @urologysummit @SuyogCancer @Larvol @IMG_Oncologists
Jeff Dean reposted
Congrats @BenMildenhall and @_pratul_! ACM gives Grace Hopper Awards to under-35-year-olds for *one specific thing*, here it's NeRF.
Past winners include Donald Knuth (for "The Art of Computer Programming"), Steve Wozniak (for Apple), and @martin_casado (for SDN). Good company!
Thanks for a great @twominutepapers conversation, Károly!
It was really fun to chat with the legendary @JeffDean. I asked him about things that only he may know the answer to. 🙂 Link to full video in thread.
I enjoyed this chat immensely with @OfficialLoganK and my fellow Gemini leads @OriolVinyalsML and @NoamShazeer and @koraykv.
My conversation with @JeffDean, @koraykv, @NoamShazeer, and @OriolVinyalsML (the Gemini co-leads) about the current state of Gemini, how we got here, and where we go next : )
Jeff Dean reposted
This would be a huge blow to America's tech sector. This country has always benefited from attracting the world's best and brightest. We should be making that easier, not harder.
Most green-card applicants will need to go abroad to apply for permanent residency at an American consulate, rather than filing from within the U.S. as they do now, the Trump administration announced Friday. www.wsj.com/politics/policy/trump-administration-to-make-…
Jeff Dean reposted
Gemini 3.5 Flash is amazing!
- Performs better than 3.1 Pro on coding & agentic tasks
- 4x faster than other frontier models
- 12x faster in @antigravity - 800 tokens/sec!
- Often at less than half the cost
And Pro to come…
Try it in @antigravity, @GeminiApp & more - enjoy!
Four questions in under 2 minutes. Thanks to Google's social team for suggesting this little interlude at #googleio yesterday!
4 questions with @JeffDean (Chief Scientist, Google and one of the authors of the famous "cat paper" about unsupervised learning and neural networks in 2012) in under 2 minutes at #googleio:
Highly capable models that are fast are super important. Our new Gemini 3.5 Flash model is a great mix of fast and capable.
Just off stage at #googleio, some highlights from this morning 🧵
Gemini 3.5 Flash is available today for everyone in @antigravity and across our products and APIs.
Compared to 3.1 Pro, 3.5 Flash is better across almost all benchmarks with huge progress in coding. It’s also comparable to the best models but very fast (4x faster tokens/ second than other frontier models). And when looking at the intelligence versus output speed, it’s in a league of its own in the top right quadrant.
Beautiful day at #googleio packed full of Gemini announcements, with my Gemini team colleagues @OriolVinyalsML and @borgeaud_s!
Take a sip of water every time you hear "Gemini"!
This is my third Google I/O in a row. Some things have changed, some (like the amazing colleagues that were there since day 0) haven't. @JeffDean @borgeaud_s
Jeff Dean reposted
We’re expanding the @Antigravity ecosystem to help developers shift their energy away from tedious debugging and focus on what actually matters: architecture and design. 🧵
Jeff Dean reposted
This is amazing. Paper → interactive website demo is exactly what I would have loved to have for some of the more complex long NLP papers.
2/
Check out how Gemini 3.5 Flash instantly digests dense academic papers and autonomously codes a fully interactive, visual website explaining the intricacies of the research. It's an incredible stress test that seamlessly merges massive long context, deep reasoning, complex coding, and ultra-low latency.
It really helps you distill papers down to their essence and aid your understanding!
1/ Today at #googleio, we’re releasing Gemini 3.5, our latest family of models combining frontier intelligence with action.
We’re starting by releasing 3.5 Flash, which is built to help you execute complex, long-horizon agentic workflows.
Gemini 3.5 Flash is our strongest model for coding and agent yet.it outscores 3.1 Pro on agentic and coding benchmarks like Terminal-Bench and MCP Atlas, while running 4x faster than other frontier models.
Used in Google Antigravity, 3.5 Flash is even further optimized to be up to 12x faster. It’s a powerful engine to deploy sub-agents that collaborate, run high-frequency iterative loops, and solve real-world problems at scale.
Some highlights we’re excited about 🔽
2/
Check out how Gemini 3.5 Flash instantly digests dense academic papers and autonomously codes a fully interactive, visual website explaining the intricacies of the research. It's an incredible stress test that seamlessly merges massive long context, deep reasoning, complex coding, and ultra-low latency.
It really helps you distill papers down to their essence and aid your understanding!
3/ Gemini 3.5 Flash is rolling out globally today. On behalf of the entire Gemini team, we're excited by what you'll be able to do with this model!
Read more here: blog.google/innovation-and-ai/models-and-research/gemini-…
Jeff Dean reposted
Waymos navigate Bay to Breakers
Those early whole company @Google ski trips were fun! I'm eighth from the left in the top row, wearing the white bathrobe (ninth if you count Susan's baby, Ari)
Don't just start a company
Start a cult
Jeff Dean reposted
Today, we introduced Gemini Intelligence, which brings the best of Gemini to our most advanced devices.
Gemini Intelligence integrates premium hardware and innovative software to help you stay a step ahead and work proactively to get things done throughout your day. #theandroidshow
Great to see @percyliang as a keynote speaker at #cais2026!
🎤 Keynote announcement:
@percyliang (Percy Liang), Professor of Computer Science at @Stanford, founding director of the Center for Research on Foundation Models, and co-founder of @togethercompute, is keynoting #cais2026.
Percy's HELM framework set the standard for holistic evaluation of language models, and his Foundation Model Transparency Index (now in its third year) put every major AI lab on notice for what they do and don't disclose. His current work on Marin takes this further: an open lab where every experiment, successful or not, is public from day one.
This one's going to be good.
San Jose · May 26–29 caisconf.org
Jeff Dean reposted
Today we're launching the Open Defense Initiative: up to $5 million in @depthfirstlabs credits for critical open source projects to find and fix real, exploitable vulnerabilities.
We’ve been investing heavily in the entire agentic security technology stack: a harness that allows models to efficiently reason through code the way a security engineer would, the context understanding to validate exploitability and produce ready-to-merge fixes, and recently, post-trained models optimized for vulnerability discovery.
Today we shared a datapoint that highlights the value of that approach. depthfirst autonomously identified 12 vulnerabilities in @FFmpeg , after Anthropic did several hundred runs with Mythos, for 1/10th of the compute cost reported by Anthropic.
As a former @GoogleDeepMind researcher, I am grateful to foundation labs for pushing the boundaries of what AI can do, as every advance there gives us a stronger base to build on. But running general purpose models without the proper harness can be too expensive to be economically viable.
That's what makes today's launch possible. The efficiency we've built lets us put frontier-level security in the hands of the maintainers who need it most, before open source models catch up and bad actors wield these capabilities without filters or guardrails.
If you maintain a critical open source project, apply for Open Defense credits through the form in the comments.
Jeff Dean reposted
Free Universal Healthcare is so complicated and expensive that only 32 of the 33 wealthiest countries in the world have figured it out.
Great to see my friend @andykonwinski as one of the keynote dishes at the first @CAISconf later this month!
📢Keynote announcement:
@andykonwinski (Andy Konwinski), co-founder of Databricks and Perplexity AI, founder of @LaudeInstitute, is keynoting #cais2026.
Laude's "Ship Your Research" mission funds open-source AI research through its Moonshots, Slingshots, and Open Frontier programs. Case in point: Terminal-Bench, a Laude-backed agent benchmark, reached Anthropic's Claude 4 model card 126 days after inception and has since become the industry-wide standard for measuring command line performance.
We're thrilled to have him.
San Jose · May 26–29
caisconf.org
Bleh. Talks. But dishes also works. The keynotes are like a multicourse tasting menu!
Jeff Dean reposted
Introducing Altara: the scientific intelligence platform for the physical world.
Today @evatuecke and I are excited to announce our $7M seed led by @GreylockVC, joined by @Neo, @BoxGroup, @Liquid2V, and angel investors including @JeffDean and leadership from OpenAI & AMD.
We’re already working with early customers in semiconductors, batteries, and advanced materials. More below.
Jeff Dean reposted
Thariq Shihipar (@trq212) from @AnthropicAI's Claude Code team. 61 papers. 45 live demos. One conference.
ACM CAIS 2026 is the first major conference dedicated to compound AI and agentic systems.
May 26–29 · San Jose
Spots are limited. Register today ↓
Jeff Dean reposted
You can now ask Gemini to create Docs, Sheets, Slides, PDFs, and more directly in your chat. No more copying, pasting, or reformatting, just prompt and download.
Available globally for all @GeminiApp users.
Google Translate is turning 20! 🎉. There are 20 fun facts and tips in the thread below.
Translate is one of my favorite Google products because it brings us all closer together!
I've been involved with a couple of things over the years. The first was our deployment of the initial system in 2006, which provided a huge leap forward in quality because it used a much larger 5-gram language model trained on trillions of words of text (indeed, probably the first trillion token language model training in the world: paper has some nice heads showing scaling-law-like quality improvement from scaling to more data/compute).
See "Large Language Models in Machine Translation", Thorsten Brants, Ashok C. Popat, Peng Xu, Franz J. Och and Jeffrey Dean, aclanthology.org/D07-1090/
The second major collaboration was in 2016 when we moved Translate over from a statistical machine translation approach to using deep neural networks. This approach relied on two key innovations. The first was Google's work on Sequence-to-Sequence models (arxiv.org/abs/1409.3215). The second was our development of TPUs, custom cups that improved the performance of inference for deep neural networks by 30-80X over existing CPUs and GPUs of the day (and reduced latency by 15-30X). This made launching compute-intensive language model services like Translate feasible for hundreds of millions of users. See "In-Datacenter Performance Analysis of a Tensor Processing Unit", Norman P. Jouppi et al. arxiv.org/abs/1704.04760
GNMT paper:
"Google's Neural Machine Translation System: Bridging the Gap between Human and Machine Translation", Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V. Le, Mohammad Norouzi, Wolfgang Macherey, Maxim Krikun, Yuan Cao, Qin Gao, Klaus Macherey, Jeff Klingner, Apurva Shah, Melvin Johnson, Xiaobing Liu, Łukasz Kaiser, Stephan Gouws, Yoshikiyo Kato, Taku Kudo, Hideto Kazawa, Keith Stevens, George Kurian, Nishant Patil, Wei Wang, Cliff Young, Jason Smith, Jason Riesa, Alex Rudnick, Oriol Vinyals, Greg Corrado, Macduff Hughes, and Jeffrey Dean, arxiv.org/abs/1609.08144
Most recently, we have advanced Translate further using Gemini models.
Each of these advances relied on research that have major quality leaps over the existing status quo translation approaches, bringing better quality and connectedness to all of our Translate users! 🎉
🧵 To celebrate Google Translate turning 20, we’re sharing 20 tips, features and fun facts you may not know about Translate.
1️⃣ Google Translate was one of the initial Google Research experiments that kickstarted our machine learning work.
2️⃣ The most common phrase on Google Translate — this month, like almost every month for the past 20 years — was “thank you.” Other top searches include: “How are you?” “I love you,” “Hello,” and “Please.”
The video of my conversation with Amin Vahdat, @gilbert, and @djrosent at Cloud Next last week is now up.
youtu.be/BpnJYJmbXcM
Thanks for a great conversation!
I had a good time discussing yesterday's Google TPU v8t and v8i announcement at Cloud Next with Amin Vahdat along with @AcquiredFM hosts @gilbert
and @djrosent.
The blog post announcement has lots of details about these new chips:
blog.google/innovation-and-ai/infrastructure-and-cloud/go…
Here's a thread of some particular things I'm excited about:
Jeff Dean reposted
Following positive feedback from other venues, like STOC and ICML, NeurIPS is pleased to announce a new initiative in partnership with Google: for NeurIPS 2026, authors will have access to Google's Paper Assistant Tool (PAT) to help improve their submissions.
This program offers authors the opportunity to receive free, automated, and actionable feedback on their manuscripts before the final deadline, private to the authors. It is a completely optional service that is kept strictly private to the authors and will not be used in the review process.
Read more in our blog post: blog.neurips.cc/2026/04/21/neurips-supports-authors-with-…
Jeff Dean reposted
THE FIRST MAN IN HISTORY TO BREAK 2 HOURS IN A MARATHON!!!🤯🤯🤯
Sabastian Sawe 🇰🇪 has just shattered the World Record at the London Marathon, running 1:59:30!!!
He makes history as the first man to officially break 2 hours in the marathon.
Yomif Kejelcha 🇪🇹 in his debut ran 1:59:41 to become 2nd fastest alltime, while Jacob Kiplimo 🇺🇬 finished in 2:00:28.
All under the previous World Record.
Jeff Dean reposted
My friend @JeffDean encouraged me to speak at this event so come join me on May 26 in San Jose! bauplanlabs.github.io/SAO-workshop/
SAO is T-5ish weeks away, and we’re excited to announce our second keynote: @ceo_clickhouse, CEO of @ClickHouseDB.
He has at the center of one of the best companies in modern OLAP, and has been very vocal that the agentic world requires us to rethink infrastructure!
It's been a delight to provide small amounts of advice and suggestions to people working on the Decoupled DiLoCo training system. This approach enables graceful handling of failures in large scale training jobs, by allowing (N-1) / N units to proceed when one fails.
Thread ⬇️
The DiLoCo team at Google DeepMind and Google Research is proud to release Decoupled DiLoCo, the next frontier for resilient AI pre-training.
Decoupled DiLoCo enables training with datacenters across the world, using heterogeneous hardware, and never halting the system despite hardware failures.
It's worth pointing out that we have been pushing on large-scale training and asynchronous techniques for the last ~14 years. Here's our NeurIPS 2012 paper where we demonstrated that this approach could be used to train very large neural networks (for the time: 30X larger than any previous neural network), and to spread the training out across thousands of machines in a fault-tolerant manner.
PDF: static.googleusercontent.com/media/research.google.com/en…
static.googleusercontent.com/media/research.google.com/en…
(This paper doesn't get as much attention because we neglected to put it on Arxiv at the time: oops!)
The Arxiv for the new Decoupled DiLoCo paper is now up: arxiv.org/abs/2604.21428v1
I had a good time discussing yesterday's Google TPU v8t and v8i announcement at Cloud Next with Amin Vahdat along with @AcquiredFM hosts @gilbert
and @djrosent.
The blog post announcement has lots of details about these new chips:
blog.google/innovation-and-ai/infrastructure-and-cloud/go…
Here's a thread of some particular things I'm excited about:
First, let's talk about TPU 8t, which is designed for large-scale training and inference throughput.
The pod size is increased slightly to 9600 chips, and provides ~3X the FP4 performance per pod vs. Ironwood (8t has 121 exaflops/pod vs. 42.5 exaflops/pod for Ironwood). In addition, the ICI network bandwidth is 2X higher per chip and the scale out datacenter networking is 4X higher per chip.
Importantly, this system also offers 2X the performance / watt, continuing the trend of significant energy effiency improvements that we've had for many generations of TPUs (8t offers ~60X the performance/watt as TPU v2).
TPU 8i is co-designed with our Gemini research team to support low latency inference. Among the attributes that support this are large amounts of on-chip SRAM, enabling more computations to be done on chip without having to go to HBM for weights or KVCache state as often.
The boardfly network topology (see image below) offers a much lower diameter network to connect all 1152 chips in an 8i pod, by fully connecting all 4 chips on the board together, fully connecting groups of 8 boards together, and then fully connecting 36 groups of 8 boards together. In addition, there is specialized Collectives Acceleration Engine (CAE) circuitry on each chip to offload various kinds of reductions and other global operations from the main computational portion of each chip, reducing on-chip latency by up to 5x.
Together, these features will offer very high throughput for large-scale models (including MoEs, which often require mapping onto many chips for inference), and will do so at very low latency. This will make agentic workloads and interactive usage really shine on the TPU 8i.
Jeff Dean reposted
The future of computer vision is agentic.
1/ We built Nomadic around a gap we kept seeing in video understanding: VLMs generate chain-of-thought that's fluent and often correct in structure, but weakly grounded in what's actually in the video.
This limitation shows up in cases like "rolling stops". Is the stop sign actually associated with the vehicle’s lane? Does the ego vehicle ever come to a full stop, or only decelerate?
Jeff Dean reposted
Read how Google built custom silicon and infrastructure - and how it's benefiting customers today - in the latest from @Bloomberg. We'll share more updates this week at #googlecloudnext.
www.bloomberg.com/news/features/2026-04-20/google-eyes-ne…
Jeff Dean reposted
Introducing Gemini 3.1 Flash TTS 🗣️, our latest text to speech model with scene direction, speaker level specificity, audio tags, more natural + expressive voices, and support for 70 different languages.
Available via our new audio playground in AI Studio and in the Gemini API!
It was a delight to serve on the selection committee with an awesome group of committee members. We had lots of great submissions to evaluate and excellent discussions during our selection process.
You can see the selected proposals below and learn more at
Moonshots // ONE wouldn't exist without the people who agreed to judge it.
27 world-class researchers and experts gave serious time and thought to 125 proposals. They shaped what rigorous evaluation looks like, and held the bar high enough that some of the most decorated scientists alive didn't make the cut.
Thank you to our selection committee, chaired by Dave Patterson:
@atalwalkar, @andykonwinski, @bakkermichiel, @bradenjhancock, @chris_bail, @ChrisRytting, @davidautor, @DG_Rand, @Diyi_Yang, @emollick, @erichorvitz, @erikbryn, @ghadfield, @HannaHajishirzi, @JeffDean, @james_y_zou, John Hennessy, @JohnJumperSci, @tkalil2050, Tino Cuéllar, @tommmitchell, @Thom_Wolf, Adam Yala, Sylvia Plevritis, @YejinChoinka, @zakkohane
laude.org/moonshots
Delighted to see that Ricardo Baeza-Yates (@PolarBearby) is this year's winner of the @TheOfficialACM Luiz Barroso Award (an award named in honor of my longtime Google colleague).
Ricardo is widely regarded as one of the world’s foremost researchers in information retrieval, and many of you may have read is excellent textbook Modern Information Retrieval (co-authored with Berthier Ribeiro-Neto). He has also played a pivotal role in strengthening the Latin American computing community.
Ricardo's accomplishments and selection would make Luiz proud!
awards.acm.org/about/2025-barroso
In case it's not clear, you can click on the picture in the first image to learn more about Ricardo and the award: Twitter swallowed my actual link text to awards.acm.org/about/2025-barroso to turn it into the picture, I guess.
Jeff Dean reposted
We’re rolling out an upgrade designed to help robots reason about the physical world. 🤖
Gemini Robotics-ER 1.6 has significantly better visual and spatial understanding in order to plan and complete more useful tasks. Here’s why this is important 🧵
Jeff Dean reposted
Artemis II astronaut Victor Glover received a hero's welcome from his neighbors! 🇺🇸
My daughter’s friend lives down the street from Victor Glover and the neighborhood is throwing him a little welcome home parade 🥹🥹🥹
Jeff Dean reposted
Yet another inspiration to women in STEM.
Christina Koch is her name🌑
This is so awesome! When I was a kid, when the Apollo mission were still relatively fresh in the recent past, I just assumed we would send people & missions to the moon quite regularly. Hard to believe we haven't since 1972! Congrats to everyone at @NASA! 🎉
Welcome home Reid, Victor, Christina, and Jeremy! 🫶
The Artemis II astronauts have splashed down at 8:07pm ET (0007 UTC April 11), bringing their historic 10-day mission around the Moon to an end.
Jeff Dean reposted
Welcome home 🩵
There are no words.
Jeff Dean reposted
After almost 12 years in Brain/DeepMind, I’ve finally decided to take the leap. My cofounders: @yinfeiy, Seth and I have kicked-off @ElorianAI. The first multimodal reasoning lab founded and led by former LLM pretraining, data and multimodal leads. youtu.be/YlvfNpOMeOY (1/n)
Congratulations to @matei_zaharia on winning this year's ACM Prize in Computing! This is very well-deserved! 🎉
Congratulations to Matei Zaharia on receiving the 2025 #acmprize in #computing! He is recognized for his visionary development of distributed data systems and computing infrastructure. Learn more: awards.acm.org/acm-prize
Great to see the reception for the very capable Gemma 4 models!
Lots of love for Gemma 4! Team just told me it’s already had 10M+ downloads since last week’s launch. Gemma models have now been downloaded 500M+ times! Excited to see what you all are creating 👀
Hedged requests (apparently inspired by the Tail at Scale paper by myself and Luiz Barroso) applied within a single machine to replicating data across DRAM channels and issuing reads to all channels, using the one that comes back first. ~5-15X reduction in p99.99 read latency.
github.com/LaurieWired/tailslayer/blob/main/README.md
Cool stuff, @lauriewired!
Accompanying video forwarded to me by a friend, which is how I learned about it:
www.youtube.com/watch?v=QFi2WVGfXMQ
As many observed, that was not the right YouTube link. Here's the correct one.
www.youtube.com/watch?v=KKbgulTp3FE
Jeff Dean reposted
Holy crap is this an awesome photo.
Have
Too
Many
Tabs
Open?
G
O
V
E
R
T
I
C
A
L
!
Too many @GoogleChrome tabs open? Try vertical tabs, rolling out now.
Just right-click any Chrome window and select “Show Tabs Vertically” to move your tabs to the side of the browser window, making it easier to read page titles and manage tab groups.
Incredible! The purple-ish and reddish colors don't quite come through when we look from Earth (and the black background highlights the vastness of space)
The highest quality video of the moon was just released… this is so beautiful.
Jeff Dean reposted
This might be the MOST STUNNING photo from Artemis II yet
Crescent Earth illuminated by light above 🤩
Jeff Dean reposted
Now that’s a front-page image
How does a major news organization push out such an egregious headline error? 🤦♂️
Oh dear.
The New York Times clever headline about NATO is entirely based on them getting its name wrong 😐
The A doesn’t stand for America, dudes.
Jeff Dean reposted
Christina Koch was a firefighter at the South Pole at -111°F before she ever applied to be an astronaut. That was maybe the fourth most interesting line on her resume. She grew up in North Carolina, got three degrees from NC State, and her first real job was building deep-space instruments at NASA.
Then she left for Antarctica. Spent three and a half years bouncing between the Arctic and Antarctic as a research scientist, including a full winter at the South Pole base. That means going months without sunlight or fresh food, with a crew of about 50 people and no way out until flights resume. While she was down there, she also joined the glacier search-and-rescue team.
After coming back, she went to Johns Hopkins and built instruments for two NASA missions (one of them is still orbiting Jupiter right now). She figured out how to start a tiny vacuum pump that NASA designed for a future Mars rover. Johns Hopkins nominated it for their Invention of the Year in 2009. Then she went back to the field. More time in Antarctica and a stretch up in Greenland. A government research station in northern Alaska, near the top of the world. Then she ran another one in American Samoa, near the equator.
In 2013, NASA selected her from 6,300 applicants. Eight people got in. Her first space mission was supposed to be a normal rotation on the International Space Station, but NASA extended it. She ended up staying 328 straight days and orbiting Earth 5,248 times, covering about 139 million miles (roughly 291 round trips to the Moon). Up there, she ran over 210 experiments, including tests of cancer drugs in zero gravity and 3D printers that can build structures close to human tissue. Six spacewalks, 42 hours floating outside the station. She learned Russian for the training. She flies supersonic jets.
Right now, Koch is on Artemis II, heading for a flyby behind the far side of the Moon. The crew launched on April 1 and is on track to travel about 252,000 miles from Earth, which would break the all-time human distance record of 248,655 miles set by Apollo 13 in 1970. That record has stood for 56 years, and it was set during a disaster that nearly killed the crew. Fred Haise, one of the Apollo 13 astronauts, is 92 now. He told Koch: "I heard you're going to break our record."
Nobody had left Earth's neighborhood since December 1972. Koch and her three crewmates are the first in 53 years, and they are coming home at about 25,000 mph. That is faster than any crewed spacecraft has ever come back through the atmosphere.
Jeff Dean reposted
-93 bytes, one of the best Wikipedia line deletions I’ve ever seen.
Today we're releasing Gemma 4, our new family of open foundation models, built on the same research and technology as our Gemini 3 series. These models set a new standard for open intelligence, offering SOTA reasoning capabilities from edge-scale (2B and 4B w/ vision/audio) up to a 26B parameter MoE model and a 31B dense model. By releasing Gemma 4 under the Apache 2.0 license, we hope to enable more innovation across the research and developer communities. Our earlier Gemma 3 models were downloaded 400M times and over 100,000 variants of those models have been published, so we're excited to see what the community will do with the even better Gemma 4 models!
Learn more at blog.google/innovation-and-ai/technology/developers-tools… and goo.gle/gemma-4-apache-2
Great work by everyone involved!
#gemma4 #ai #opensource #ml
Benchmark results for the various models as well as Gemma 3 comparison:
Jeff Dean reposted
Colleagues who have no hesitation playing with your feelings 🥹
My friend Aaron Katz (@ceo_clickhouse) says this is going to be a great event at Google Cloud.
The loudest event at Next ‘26 isn’t a keynote. 🤫
We’re taking over Hakkasan with The Chainsmokers for a night the dev community will be talking about until '27.
Data by day. Legends by night. 🪩
Apr 22 | Las Vegas
Don't get left in the queue: clickhouse.com/houseparty/google-next
Jeff Dean reposted
🙌🙌 Charles Barkley
Charles Barkley on CBS tonight:
“The way some of these immigrants are getting treated in our country right now is a travesty and a disgrace. What we’re doing to some of these amazing immigrants is really unfortunate and really sad.”
Charles Barkley is not a politician. Not an activist. Not a Democrat. He is one of the most beloved sports figures in American history saying this on national television.
"Give me your tired, your poor, your huddled masses yearning to breathe free" 🗽
As Charles Barkley observed, we are falling short of our own ideals.
Charles Barkley on CBS tonight:
“The way some of these immigrants are getting treated in our country right now is a travesty and a disgrace. What we’re doing to some of these amazing immigrants is really unfortunate and really sad.”
Charles Barkley is not a politician. Not an activist. Not a Democrat. He is one of the most beloved sports figures in American history saying this on national television.
Jeff Dean reposted
Here’s a recap of everything we released this week:
— Gemini 3.1 Flash Live, delivering our highest-quality audio experience yet with improved reasoning and latency for more natural voice interactions
— An update to @GeminiApp on desktop that enables you to seamlessly transition from other AI apps by importing your preferences and chat history in just a few clicks
— Lyria 3 Pro, supporting high-fidelity music tracks up to 3 minutes long with the ability to prompt for structural elements like intros, verses, and bridges
— A suite of new Gemini features on Google TV, such as narrated deep dives on educational topics and real-time sports overviews
— Google Translate’s Live translate with headphones, now on iOS and expanding to more countries for real-time translation across 70+ languages
— A new research partnership between @GoogleDeepMind and Agile Robots to develop the next generation of helpful robots
The video of my conversation with Bill Dally at GTC last week is up. I always enjoy talking to Bill, and we had a wide ranging discussion about computer architecture, model training, specialized inference hardware, custom interconnects, and more!
youtu.be/g8BuAtM3fp4
Jeff Dean reposted
This is a devastating ruling for the government, finding Anthropic likely to prevail on essentially all of its theories for why the government’s actions were unlawful and unconstitutional. One of the things she mentions is the huge range of amici briefs supporting Anthropic (by the way, 0 supported USG)—so thanks to everyone here who signed on to FAI’s brief, or to one of the many many others. These things do matter. More importantly, you were on the right side of history.
On a personal note: some friends and allies of mine on the right have been angry at me for my own words and actions in all this. Anyone who thinks I spoke out for personal gain or trivial reasons against an administration I served in is crazy. It was a hugely costly decision for me. But Judge Lin’s ruling shows why I did it: this is a staggeringly illegal act by the government. That is why I am particularly honored to have been (implicitly) quoted in the ruling for calling this what it was when Secretary Hegseth initially made his announcement: an attempted act of corporate murder.
The case continues, but Anthropic has scored a very large win here. The real victors, however, are all red-blooded Americans who are, as the founders would have said, “jealous of their liberties.”
BREAKING: Anthropic has been GRANTED a preliminary injunction re: Pentagon 'supply chain risk' designation by Judge Rita Lin in California but is allowing a stay for one week storage.courtlistener.com/recap/gov.uscourts.cand.465515/…
📢 Another exciting step forward today with the launch of Gemini 3.1 Flash Live.
It natively understands audio, making it much more capable of handling complex instructions. It leads on ComplexFuncBench, and on Scale AI’s AudioMultiChallenge, demonstrating skill in complex instruction following and long-horizon reasoning amidst the interruptions of real-world audio.
That means the model can pick up on nuances like pitch and pace, leading to much more fluid, high-fidelity voice interactions.
It’s now powering Gemini Live and Search Live globally. Huge congrats to the teams that made this possible. ✨
blog.google/innovation-and-ai/models-and-research/gemini-…
Jeff Dean reposted
Google DeepMind 🤝 Agile Robots
Our new research partnership will integrate the Gemini foundation models with their hardware to help build the next generation of more helpful and useful robots.
Find out more → goo.gle/4lKu7de
Jeff Dean reposted
Google was just named #1 in the @FastCompany 2026 World’s Most Innovative Companies list. 🎉 Google is also ranked #1 in their Artificial Intelligence category. See the full story. www.fastcompany.com/most-innovative-companies/list
Jeff Dean reposted
Google is now the first cloud provider to integrate 1 GW of flexible demand into long-term utility contracts. Our ability to shift or reduce our energy demand when it’s needed can help utility companies balance supply/ demand and plan for future capacity needs.
This is a big milestone for responsible data center growth and helps keep costs lower for local communities.
blog.google/innovation-and-ai/infrastructure-and-cloud/gl…
Congratulations to Charles Bennett and Gilles Brassard for winning this year's @theofficialacm Turing Award! 🎉
They were recognized for their work on quantum information science & quantum cryptography.
Google is proud to support the award to recognize groundbreaking CS work.
Congratulations to Charles H. Bennett (@IBMResearch) and Gilles Brassard ( @UMontreal) on receiving the 2025 ACM A.M. Turing Award!
🔗: awards.acm.org/turing
Love it. We should all aspire to do the little things that help each other.
I think that is the best advertisement I’ve ever seen.
Jeff Dean reposted
We open sourced WAXAL!
- Multilingual speech dataset for African languages
- 17 languages for TTS
- 19 languages for ASR
Over 100 million speakers across 40 Sub-Saharan African countries
huggingface.co/datasets/google/WaxalNLP
Jeff Dean reposted
We've raised $6.5M to kill vector databases.
Every system today retrieves context the same way: vector search that stores everything as flat embeddings and returns whatever "feels" closest.
Similar, sure. Relevant? Almost never.
Embeddings can’t tell a Q3 renewal clause from a Q1 termination notice if the language is close enough.
A friend of mine asked his AI about a contract last week, and it returned a detailed, perfectly crafted answer pulled from a completely different client’s file.
Once you’re dealing with 10M+ documents, these mix-ups happen all the time.
VectorDB accuracy goes to shit.
We built @hydra_db for exactly this.
HydraDB builds an ontology-first context graph over your data, maps relationships between entities, understands the 'why' behind documents, and tracks how information evolves over time.
So when you ask about 'Apple,' it knows you mean the company you're serving as a customer. Not the fruit.
Even when a vector DB's similarity score says 0.94.
More below ⬇️
Jeff Dean reposted
Introducing our biggest upgrade to @googlemaps since the original launch, featuring Ask Gemini (with personalization), Immersive Navigation, and much more!! 🗺️
I got a "still under construction but getting close to done" look when I was in London a little while ago, and this is going to be an amazing building. I've had offices in many buildings in my time at Google (12?), and I find the ones that are "better spaces" make working and collaborating more enjoyable. The location right near Kings Cross and St. Pancras train stations and underground also makes commuting from all over London and outer areas really convenient. I'm really excited for my London colleagues to move in (yes, it's been a long wait!) and also for the public to experience the public spaces in the building.
We’re excited to unveil the name of our new London building: Platform 37. 📍
The name honors both the surrounding area’s transport heritage and "Move 37" – the critical moment where our AI system AlphaGo showed it could find novel solutions humans hadn't considered.
Jeff Dean reposted
We’ve been seeing some amazing Nano Banana 2 creations lately. 🍌
Here are some standouts. 🧵
Excited to see this joint collaboration between @GoogleResearch, @NHSuk and @imperialcollege showing AI’s potential to detect 25% of the interval cancers previously missed by conventional methods. Additionally, the research found AI can reduce screening workloads, and give results back to clinicians and patients faster.
This first figure from the @NatureCancer article shows how the study was set up, and the second figure shows that the AI system dramatically increases sensitivity (detecting true positives) without significantly affecting specificity (false positives).
Learn more ⬇️
Blog: blog.google/innovation-and-ai/technology/health/google-ai…
Nature Cancer paper: www.nature.com/articles/s43018-026-01127-0
Jeff Dean reposted
Excited to share @Standard_Kernel's seed round and some reflections on what we’ve learned about kernel generation and what we believe is next. Grateful to our amazing team, supporters, and the broader community pushing this space forward.
Jeff Dean reposted
Say hello to Gemini Embedding 2, our new SOTA multimodal model that lets your bring text, images, video, audio, and docs into the same embedding space! 👀
Jeff Dean reposted
1/ We released NanoGPT Slowrun 10 days ago. Already at 8x data efficiency and improving fast, so we're doubling down.
Announcing Slowrun Research and Slowrun Cluster: our open research effort to collaborate with researchers with crazy ideas, and a serious cluster to back it.
Jeff Dean reposted
Attending #sala2026? We invite you to visit the Google booth to meet our team and discover our initiatives to empower AI talent across more than 20 countries in Latin America and beyond. You can find more details of our involvement at SALA at goo.gle/4cCEgpT
You don't have many shots at a world championship. Unless you're Italy!
Because of course Team Italy has a full coffee and espresso station in their dugout
I'm looking forward to a great discussion with Bill Dally at @nvidia 's GTC event on March 18!
Two of the researchers who paved the way for the modern AI ecosystem on stage, together. 🤝
@JeffDean and Bill Dally are hosting a fireside chat at #nvidiagtc to dig into what it really takes to power the next frontier of AI—from agentic systems to ultra‑efficient, trillion‑parameter models.
🔗 nvda.ws/3ORwheS
We've been working on the Waxal dataset project since 2021, aiming to enhance the amount of data available for African languages. This public speech dataset initially covers 27 Sub-Saharan African languages spoken by over 100 million speakers across more than 26 countries. 🌍
Jeff Dean reposted
Startup School is back!
Hear from Jensen Huang, @sama, @alexandr_wang, @JeffDean, and more.
Join a hand-selected group of top CS students, researchers, and engineers for two days of talks, sessions with YC partners, and hands-on robotics demos, right here in San Francisco.
Gemini + Argentina! 🌍🏆
I've loved watching Lionel Messi play since his first days at Barcelona! ⚽️
www.afa.com.ar/es/posts/google-gemini-nuevo-main-sponsor-…
Jeff Dean reposted
Gemini-3.1-Flash-Lite-Preview by @GoogleDeepMind moves the Pareto frontier as one of the most efficient models in its tier ($0.25 in / $1.50 out per MTokens).
As a family, Gemini-3.1 offers a clear cost-to-quality ladder, with both variants leading within their price bands.
- Gemini-3.1 Pro leads with models priced higher
- Gemini-3.1 Flash Lite leads in a lower cost band
- Older variant Gemini 3 Flash also sits on the frontier in between
Gemini-3.1-Flash-Lite-Preview by @GoogleDeepMind is in Text and Code Arena! This is the most efficient Gemini 3.1 ($0.25 in / $1.5 out per MTokens) on the cost-performance frontier.
Highlights:
- ranks #36 in Text, scoring 1432, on par with Grok-4.1-fast, strong in Creative Writing, and Longer Query
- surpassing larger Gemini 2.5 Flash and GPT-5-mini
- tied in Code Arena for #35 scoring 1261, on par with Qwen3-coder for agentic webdev tasks
Congrats @GoogleDeepMind for another milestone!
⚡ Excited to announce Gemini 3.1 Flash-Lite! We’ve set a new standard for efficiency and capability to give developers our fastest, most cost-effective Gemini 3 model yet.
We engineered this model with thinking levels, allowing it to handle high-volume queries instantly, while scaling up its reasoning for complex edge cases.
By the numbers:
⏱️ 2.5X faster time-to-first-token than 2.5 Flash while being significantly higher quality
📉 $0.25 per 1M input tokens
📊 1432 Elo on LMArena & 86.9% on GPQA Diamond
Thrilled to see what developers build with this kind of speed and quality at scale. Available now in Google AI Studio and Vertex AI.
blog.google/innovation-and-ai/models-and-research/gemini-…
Here's a side-by-side speed and accuracy comparison comparing Gemini 3.1 Flash Lite against our older Gemini 2.5 Flash model. It is not only significantly faster in terms of tokens/s, it also needs many fewer tokens to accomplish complex tasks (~1/3 as many in this case).
Here's a side-by-side speed and accuracy comparison comparing Gemini 3.1 Flash Lite against our older Gemini 2.5 Flash model. Not only is the new model significantly faster in terms of tokens/s, it also needs many fewer tokens to accomplish complex tasks (~1/3 as many in this case).
Jeff Dean reposted
Btw this question is a homage to @JeffDean and @GhemawatSanjay, who wrote the paper on MapReduce.
I wonder if there was some fun lore around this.
Technical interview question:
Suppose you have 5 TB worth of text data and you want to count the total number of words, how will you do this?
Jeff Dean reposted
Every year, in the middle of February, the sunlight falls on this waterfall called «Horsetail,» and because of the specific angle at which the light hits, it looks like the water is on fire. The water starts to glow orange and red.
Jeff Dean reposted
We are now at 1136 new measles cases - in just the first 8 weeks of this year alone.
A preventable disease resurging.
This is an abject policy failure.
Great to see us using innovative battery technology and wind+solar to create more data centers with carbon-free energy.
techcrunch.com/2026/02/26/google-paid-startup-form-energy…
Exponential trends are very powerful: solar panel prices, transistors per mm^2, price of genomic sequencing, bits per mm^2 for hard drives, ...
The cost of solar has dropped 99.8% worldwide since 1975 and 94.4% since 2008.
Solar is the future.
Our new Nano Banana 2 model is out today! Even better image generation, and it debuts at #1 on the Image Arena. You too can make a bunch of awesome images!
🚨BREAKING: Nano Banana 2 debuts at #1 in Image Arena, and it changes the game again 🍌🍌
Officially released as Gemini 3.1 Flash Image Preview, it is powered by real-time information and images from web search.
Highlights:
- #1 Text-to-Image scoring 1279, surpassing GPT-Image-1.5 and Nano Banana Pro
- Ties for #1 Single-Image Edit, scoring 1407 on par with ChatGPT-Image-Latest
- Top 3 Multi-Image Edit, alongside Nano Banana Pro variants
- $0.067 per image ~2x cheaper than Nano Banana Pro
Congrats to the @GoogleDeepMind team for continuing to push the frontier!
Jeff Dean reposted
Prompt: Make a photo of a half peeled banana plushie, the plushie is also a sauropod
Jeff Dean reposted
Thrilled to share: #aletheia, our math research agent, just solved 6/10 notoriously hard FirstProof problems autonomously, the best result in the inaugural challenge! To me, this is even bigger than our historic IMO-gold achievement last year; these problems challenge even top mathematicians. We share our results transparently, see paper and full thoughts in the thread. 👇
Congratulations @TechTekedra! 🎉
(Sent from a @Waymo)
We are delighted that our co-CEO @techtekedra is recognized as a @CNBC Changemaker. Tekedra is a fearless leader and champion of our company, guiding our team in our important goal to make communities safer.
Agreed. Mass surveillance violates the Fourth Amendment and has a chilling effect on freedom of expression. Surveillance systems are prone to misuse for political or discriminatory purposes.
As an American citizen, the last thing I want is government using AI for mass surveillance of Americans.
Jeff Dean reposted
Jeff Dean reposted
Jeff Dean reposted
Today we announced a landmark, three-year initiative with @ISTE_ASCD to provide comprehensive Gemini training to all 6 million K-12 and higher-education faculty in the U.S. Rolling out in the coming months, @GoogleForEdu will offer educators training on how to safely and thoughtfully use our suite of AI tools — including NotebookLM. It’s the largest initiative of its kind, designed to help over 74 million students thrive in an AI-infused world.
AI in educational settings has enormous positive potential (education is one of seven areas I & co-authors discussed in our recent paper on Shaping AI's Impact on Billions of Lives arxiv.org/abs/2412.02730).
This effort ⬇️ will help teachers understand and use these tools.
We’re making Gemini training available to all 6 million K-12 teachers and higher-education faculty in the U.S. Here’s what to know:
• The program’s made up of concise, flexible training modules designed to specifically address busy educators’ needs.
• Each module offers real-world examples aimed at supporting student learning.
• Faculty who complete the training will receive badges to demonstrate their AI literacy with Google tools.
Jeff Dean reposted
Sharing on behalf of a friend:
"""
New compute & funding opportunity for researchers!
We're launching the 2026 Google TPU Research & Education Awards.
We want to support researchers pushing the boundaries of ML systems, performance, and efficiency, and those working in high-impact applied science.
What's included:
- Compute: Free access to the latest generations of TPUs.
- Funding: An unrestricted gift to support or partially support a grad student.
- Credits: Google Cloud credits to support your work.
If you or a colleague are exploring what is possible with TPUs, we want to hear from you.
Read the RFP and apply here: goo.gle/2026-tpu-rfp
If you have questions, email tpu-rfp@google.com
"""
Jeff Dean reposted
Gemini 3.1 Pro Preview scored highest in the Artificial Analysis Intelligence Index but its most significant advantage might be its price and token efficiency. Our evaluations cost <50% to run on Gemini 3.1 Pro Preview compared to Claude Opus 4.6 (max) and GPT-5.2 (xhigh)
Gemini 3.1 Pro Preview is priced at $2/$12 per 1M input/output tokens while Claude Opus 4.6 is priced at $5/$25 and GPT-5.2 at $1.75/$14.
To complete our Intelligence Index evaluations Gemini 3.1 Pro Preview used fewer tokens (high, 56M) than GPT-5.2 (xhigh, 130M) and Claude Opus 4.6 (max, 58M).
This results in Gemini 3.1 Pro Preview costing $892 to complete our evaluations - less than 50% of the cost of GPT-5.2 (xhigh, $2,304) and Claude Opus 4.6 (max, $2,486).
Jeff Dean reposted
Gemini 3.1 is SOTA at 3D spatial reasoning
Jeff Dean reposted
🚨 New Benchmark: I have to admit, this is a game-changer for me. TRUST ⬆️
While Opus 4.6 is the LLM of choice for Devs, and it appears Gemini 3.1 will not change that, this benchmark, AA-Omniscience Index from @ArtificialAnlys is a real game changer because it;
"...measures knowledge reliability and hallucination. It rewards correct answers, penalizes hallucinations, and has no penalty for refusing to answer. Scores range from -100 to 100, where 0 means as many correct as incorrect answers, and negative scores mean more incorrect than correct."
Bottom Line: Trust ⬆️
Prior to version 3.1, models like Gemini 3 Pro and Gemini 3 Flash had a major flaw. While they had the highest raw factual accuracy of any models, they also had abysmal hallucination rates (around 88% to 91% in tests used to elicit hallucinations). They were highly intelligent but refused to admit when they didn't know something.
Gemini 3.1's top score proves Google has successfully tuned the model's knowledge calibration, and it now successfully recognizes its own knowledge gaps and abstains rather than making things up.
Jeff Dean reposted
We have a new leader.
Gemini 3.1 Pro!
Jeff Dean reposted
🚀 Gemini 3.1 Pro vibe-coded a double wishbone suspension
• Independent double wishbone design
• Dynamic coilover shock absorber
• Vented disc brakes with performance caliper
• Real-time kinematic travel & steering simulation
AI isn’t just generating visuals anymore!
Jeff Dean reposted
Gemini 3.1 Pro has landed! Amazing performance / capabilities across the board.
Beyond SOTA, the best are all the things that evals can't measure. E.g. SVG has gotten so much better (see 🧵)
blog.google/innovation-and-ai/models-and-research/gemini-…
Today, we’re continuing to push the boundaries of AI with our release of Gemini 3.1 Pro.
This updated model scores 77.1% on ARC-AGI-2, more than double the reasoning performance of its predecessor, Gemini 3 Pro.
Check out the visible improvement in this side-by-side comparison, showing Gemini 3.1 Pro’s crisp animation built with pure code.
Read more about today’s 3.1 Pro update: blog.google/innovation-and-ai/models-and-research/gemini-…
Last week we released our Gemini 3.1 Pro Deep Think model. Here are a succession of examples of that model doing heat transfer analysis.
No tools, just Deep Think and Image Generation.
Step 1: generate a CAD file from a technical drawing.
Step 3: Turn the heat transfer analysis at different times in the heating process into visual representations.
Benchmark scores are generally quite a bit better for Gemini 3.1 Pro vs. Gemini 3 Pro.
Jeff Dean reposted
Shipping AI Mode to 53 new languages (spoken by more than a billion people globally!)
Try it → g.ai
Languages → support.google.com/websearch/answer/16011537
Jeff Dean reposted
#googleio returns May 19-20, 2026! Register below 🗓️ goo.gle/io2026reg-1
Many have really enjoyed our recent Nano Banana (+Pro) generative models for images and Veo models for video. Our latest Lyria model for generating music allows you to easily generate music starting with a idea or description, or from an image or video in seconds!
Introducing Lyria 3, our latest and most advanced music model, available in the Gemini App starting today : )
Go from idea, image, or video to music in seconds!
Here's an example I tried:
"A mozart-like violin sonata with an upbeat and peppy major key, paired with a cello for accompaniment
Jeff Dean reposted
Introducing America-India Connect, a collaborative infrastructure initiative anchored by our five-year $15 billion AI infrastructure investment in India to increase the reach, reliability and resilience of digital connectivity across four continents.
In collaboration with local partners, America-India Connect will establish:
📍A new international subsea gateway in Visakhapatnam (Vizag)
📍Three new subsea paths connecting India to Singapore, South Africa and Australia
📍Four strategic fiber-optic routes between the United States, India and multiple locations across the Southern Hemisphere
Learn more about a very high quality time series model released by @GoogleResearch a while back at research.google/blog/a-decoder-only-foundation-model-for-…
𝗔 𝗧𝗶𝗺𝗲 𝗦𝗲𝗿𝗶𝗲𝘀 𝗙𝗼𝘂𝗻𝗱𝗮𝘁𝗶𝗼𝗻 𝗠𝗼𝗱𝗲𝗹 𝗕𝘆 𝗚𝗼𝗼𝗴𝗹𝗲
This has been pre-trained on a time series corpus of 100 billion data points, & shows impressive performance on various benchmarks from diverse domains.
𝗧𝗶𝗺𝗲𝘀𝗙𝗠 𝗚𝗶𝘁𝗵𝘂𝗯 𝗽𝗮𝗴𝗲: github.com/google-research/timesfm/
𝗟𝗲𝗮𝗿𝗻 𝗠𝗟 𝗮𝗻𝗱 𝗙𝗼𝗿𝗲𝗰𝗮𝘀𝘁𝗶𝗻𝗴: leanpub.com/pycaretbook/
Jeff Dean reposted
At the AI impact summit our message is clear: we believe India is going to have an extraordinary trajectory with AI, and we want to be a partner.
To help, we have launched new America-India Connect subsea cable routes building on our $15B investment in India’s infrastructure, including a full-stack AI Hub in Vizag. We're also rolling out a new AI Certificate, a @GoogleDeepMind partnership to unlock discoveries in science & education, live speech-to-speech translation for 70+ languages and more.
Jeff Dean reposted
We sent this note to our customers to let them know that Braintrust has raised a new round of funding, and thank them for their support. While the money is exciting, our focus hasn't changed: we're building Braintrust to help our customers ship quality AI products.
In 2026, AI is moving to production but teams have never had less conviction about what will fail next. Our customers are building AI products that serve millions and simply need to work. If Braintrust makes their lives easier and their products better, I know we are doing our job.
Thank you to @ICONIQCapital for leading our Series B, and to @a16z, @GreylockVC, @basecasevc, and @eladgil for doubling down. Thank you to the Braintrust team for all the incredible work you've done over the past year. And thank you to our customers, who have made this growth possible.
Jeff Dean reposted
The craziest thing about this graph was FDR was President for 12 years.
And his average was 8%.
He was also the last President to make the poorest 50% of Americans significantly better off when he left office.
It turns out you can redistribute wealth and grow the economy.
Jeff Dean reposted
Um Alpha alert! When Jeff Dean was asked if he can go through the history of Google search sys design, he mentioned a 2009 talk in a "web search and data mining conference" I think its this video!
www.youtube.com/watch?v=IQuSuyUVkoE
Congrats to all my newly-elected Google colleagues (and all the other new members this year) on being elected to the National Academy of Engineering! 🎉
Congrats to the five Googlers — @DemisHassabis, John Jumper, Peyman Milanfar, @NoamShazeer and Moti Yung — who have been elected to the National Academy of Engineering (NAE), which is among the highest professional distinctions accorded to an engineer.
www.nae.edu/345149/NAENewClass2026
Thanks for having me on the @latentspacepod, @swyx and @FanaHOVA! I enjoyed our discussion!
Site with summary:
latent.space/p/jeffdean
Video:
youtu.be/F_1oDPWxpFQ?t=3
From rewriting Google’s search stack in the early 2000s to reviving sparse trillion-parameter models and co-designing TPUs with frontier ML research, Jeff Dean has quietly shaped nearly every layer of the modern AI stack. As Chief AI Scientist at Google and a driving force behind Gemini, Jeff has lived through multiple scaling revolutions from CPUs and sharded indices to multimodal models that reason across text, video, and code.
We sat down with Jeff to unpack what it really means to “own the Pareto frontier,” why distillation is the quiet force behind every generation of faster, cheaper models, how energy not FLOPs is becoming the true constraint on AI compute, what it takes to co-design hardware and models 2–6 years into the future, why unified multimodal systems will outperform specialized ones, what it was like leading the charge to unify all of Google’s AI teams, and his prediction that deeply personalized models with access to your full digital context will redefine what useful AI looks like.
@JeffDean @GoogleDeepMind @Google
We’ve updated Gemini 3 Deep Think to better tackle the complexity of real-world research, science, and engineering. ♊ 🚀
It achieves gold-medal standards on the written portions of the Physics and Chemistry Olympiads, building on gold-level performance at IMO and ICPC and has SOTA results on ARC-AGI-2, MMMU-Pro, and HLE. 🏅
From optimizing code to validating scientific data, we’re seeing it become a genuine partner in the discovery process.
Read more here: blog.google/innovation-and-ai/models-and-research/gemini-…
I'm very proud of everyone on the Gemini team! ♊
Thanks to everyone for their hard work on creating this excellent model! 🎉
Also see this blog post for some real world use of our Deep Think model for taking research problems across mathematics and various scientific fields.
deepmind.google/blog/accelerating-mathematical-and-scient…
Jeff Dean reposted
The upgraded Gemini 3 DeepThink is now live! 🚀
We're already seeing engineers and researchers leverage it as a partner in their design and development processes
I love this example of Anupam Pathak using DeepThink to go from prompt to physical prototype—actually designing complex components like turbine blades 10x faster.
Jeff Dean reposted
There are only 7 people on the planet who can beat Gemini 3 Deep Think in coding competitions.
It has an Elo of 3455.
A bit over a year ago the best systems were at 2727 (o3-preview).
Gemini 3 Deep Think (feb update) Benchmarks
Jeff Dean reposted
We’ve upgraded our specialized reasoning mode Gemini 3 Deep Think to help solve modern science, research, and engineering challenges – pushing the frontier of intelligence. 🧠
Watch how the Wang Lab at Duke University is using it to design new semiconductor materials. 🧵
Jeff Dean reposted
Thrilled to host you, @JeffDean!
Overflowing crowd @JeffDean’s talk today @PrincetonCS 🔥🔥🔥
Jeff Dean reposted
How could AI act as a better research collaborator? 🧑🔬
In two new papers with @GoogleResearch, we show how Gemini Deep Think uses agentic workflows to help solve research-level problems in mathematics, physics, and computer science.
More → goo.gle/4aGs3Pz
This is incredibly impressive.
Sub-six-minute mile pace uphill on skis: #winterolympics SPEED. 🔥
Jeff Dean reposted
US Olympic Team member: What's happening in Minnesota is wrong. There's no shades of gray. We love our country. We're playing for Team USA. What the Olympics means is excellence, respect, friendship. We are playing for the people around the country who show those, share those same values, that compassion, that love, and that respect
Jeff Dean reposted
Our bioacoustics foundation model Perch was trained primarily on terrestrial animals - like birds. But Perch 2.0 is showing incredible performance on underwater acoustics.
Here’s how it’s helping us listen to and understand marine ecosystems 🧵
Jeff Dean reposted
BREAKING: US Olympian Chloe Kim, who has immigrant parents, speaks out:
“In moments like these it’s really important for us to unite. We are allowed to voice our opinions on what’s going on. I think we need to lead with love and compassion.”
I’m so happy that so many Olympians are not hesitant to speak out for what they believe in even when they know the president could try to bully them.
Jeff Dean reposted
❤️❤️❤️🤍🤍🤍💙💙💙🇺🇸🇺🇸
Cool application of AI models to find security vulnerabilities. Nice work, @stanislavfort!
New post on what AI cybersecurity research looks like when it actually works! I wrote up what we've learned discovering 12 of 12 new OpenSSL zero-days, 5 CVEs in curl, and additional 100+ validated CVEs across critical open source infrastructure, middleware, and secure apps 🔗⏬
Jeff Dean reposted
McMaster University has posted the email written by the professor to Canadian figure skater Maddie Schizas, granting her an extension on her assignment.
“I’d say focus on the competition right now, and submit it directly to me when it is done”
This is a really lovely note.
Some nice updates to Veo 3.1 today, including the much-asked-for better support for portrait mode videos and more.
Today, Veo 3.1 Ingredients to Video is getting more expressive + portrait mode has arrived (we heard you!)
You can now create vertical clips with livelier movement and better control directly in the @Geminiapp, @YouTube & Google Vids.
All this plus state-of-the-art upscaling to 1080p and 4K also coming to @flowbygoogle, the Gemini API and Vertex AI.
Gemini+Genie 3 are helping @Waymo simulate long tail scenarios to make driving safer.
We’re excited to introduce the Waymo World Model—a frontier generative mode for large-scale, hyper-realistic autonomous driving simulation built on @GoogleDeepMind’s Genie 3.
By simulating the “impossible”, we proactively prepare the Waymo Driver for some of the most rare and complex scenarios—from tornadoes to planes landing on freeways—long before it encounters them in the real world.
waymo.com/blog/2026/02/the-waymo-world-model-a-new-fronti…
Jeff Dean reposted
Thanks to @SenGaryPeters for entering my @nytimes autonomous vehicle essay into the Senate hearing record today
The debate is shifting from "cool tech" to “preventative public health”
The data is clear: this technology significantly decreases crash rates + mitigates human harm
I have a guest essay in @nytimes today about autonomous vehicle safety. I wrote it because I’m tired of seeing children die. Done right, we can eliminate car crashes as a leading cause of death in the United States
@Waymo recently released data covering nearly 100 million driverless miles. I spent weeks analyzing it because the results seemed too good to be true. 91% fewer serious-injury crashes. 92% less pedestrians hit. 96% fewer injury crashes at intersections. The list goes on.
39,000 Americans died in crashes last year. More than homicide, plane crashes, and natural disasters combined. The #2 killer of children and young adults. The #1 cause of spinal cord injury. We’ve accepted this as the price of mobility.
We don’t have to.
In medicine, when a treatment shows this level of benefit, we stop the trial early. Continuing to give patients the placebo becomes unethical. When an intervention works this clearly, you change what you do.
In driving, we’re all the control group.
Cities like DC and Boston are blocking deployment. And cities are not the only forces mobilizing to slow this progress.
It’s time we stop treating this like a tech moonshot and start treating it like a public health intervention that will save lives.
Link to article below.
👀 this video of Waymo cars evading crashes with people and vehicles. I especially note the ones that require it having a 360° view.
My sincere thanks to Alex Ellerbeck and @acsifferlin for their wisdom and sure hand in editing this piece.
Very proud to see the progress across so many areas. Great to see @GeminiApp hit 750M monthly active users, and the use of our Gemini models across many products and Cloud surfaces is strong: 10B tokens/minute is 166M tokens/sec (TPUs!), or ~1750 tokens/person on earth per day.
Our Q4/FY’25 results are in. Thanks to our partners & employees, it was a tremendous quarter, exceeding $400B in annual revenue for the first time. Our full AI stack is fueling our progress, and Gemini 3 adoption has been faster than any other model in our history.
We’re really well positioned and excited going into 2026. Much more to come!
Jeff Dean reposted
Adaption has raised $50M to build adaptive AI systems that evolve in real time.
Everything intelligent adapts. So should AI.
Jeff Dean reposted
Introducing Q Labs, a research lab focused on solving generalization.
Alongside others (SSI, Flapping Airplanes), we see data efficiency as the key problem, but we're taking an unconventional approach to solve it: a new learning algorithm approximating Solomonoff induction.
Jeff Dean reposted
A new era for Waymo. We’ve raised $16B to accelerate our mission, valuing the company at $126B. This capital is an investment in a future where more cities get a safer, more reliable way to move. Let's go. 🚀 waymo.com/blog/2026/02/waymo-raises-usd16-billion-investm…
Jeff Dean reposted
Excited to share our latest work: "Semi-Autonomous Mathematics Discovery with Gemini." We used Gemini to systematically evaluate 700 "open" conjectures in the Erdős Problems database.
The result? We addressed 13 problems marked as open—finding 5 novel autonomous solutions and identifying 8 existing solutions missed by previous literature.
Read the full case study here: arxiv.org/abs/2601.22401
Jeff Dean reposted
The AI field is in need of harder benchmarks to test capabilities of the latest AI models. This update to @Kaggle Game Arena with werewolf and poker (heads-up) plus chess, gives us new objective measures of real-world skills like planning and decision making under uncertainty.
This is a terribly misleading graphic which uses a non-zero-based Y axis to make a 1% difference seem much larger.
I recommend Tufte's "The Visual Display of Quantitative Information".
American steel is BACK. 🇺🇸
Jeff Dean reposted
Sequoia National Park in California
Love it when more people from different parts of the world interact.
This Japanese creator meeting Cardinals lineman Dalvin Tomlinson & LJ Collier is the best thing you’ll see today
Jeff Dean reposted
For his official NASA portrait, astronaut Leland Melvin snuck his two rescue dogs in to pose with him.
Anyone who looks at this photo and doesn't crack a smile is probably dead inside.
Jeff Dean reposted
The largest randomized trial of medical A.I.
—Over 100,000 women in Sweden
—radiologist + AI vs 2 radiologists, in follow-up
—AI added led to 29% more cancer detected, 44% reduced workload, and
—Less cancer dx in subsequent 2 years, and, when found, less aggressive
Jeff Dean reposted
SFO let’s go! ✈️As of today, we’re serving SFO Airport and gradually rolling out public access—available 24/7 via the Rental Car Center. Learn more: waymo.com/blog/2026/01/waymo-rides-at-sfo
JEE whiz!
Building on last week’s announcement of practice SATs in Gemini, you can now also take full-length mock JEE Main tests in Gemini at no cost.
Jeff Dean reposted
We’re introducing a series of updates that make Gemini in @googlechrome more helpful, efficient, and personalized for you:
— Side panel for multitasking: The new side panel experience allows you to do things like send emails or compare information without leaving your open tab
— Nano Banana integration: Generate and edit high-fidelity images right from the new side panel
— Deep ecosystem integration: Connect Google apps like @gmail, @googlecalendar, @YouTube, Google Shopping, and Google Flights to accelerate complex task execution
— Personal Intelligence: With your permission, Personal Intelligence turns Chrome into a trusted partner that remembers past conversations and connects with your Google apps to provide help that’s uniquely tailored to you
— Agentic action via auto browse: Offload multi-step browser tasks, like scheduling an appointment or filling out a complex form, using our new auto browse feature that’s rolling out in preview for U.S. Google AI Pro and Ultra subscribers
— Support for agents: Support for Google’s Universal Commerce Protocol (UCP) ensures AI agents can execute transactions on your behalf
— Enhanced security: New rigorous security standards designed to protect against emerging kinds of online threats
The new Gemini in Chrome is available on Windows, macOS, and Chromebook Plus in the U.S. We’re just beginning to explore the possibilities of an agentic web, so stay tuned for more news here.
I love Google Translate because it is one of the purest manifestations of "making information universally accessible and useful". Great to see these improvements to further that!
Just shipped a state-of-the-art model upgrade for Google Translate!! 🚀
- Smarter text translations in the Translate app, web & Search
- Natural real-time translations with 🎧 on for Live Translate
- Expanded language learning in the app – just tap Practice
Get the app here: translate.google.com/about/
A delightful discussion between @OfficialLoganK and my longtime Google colleague and friend Emanuel Taropa on the "Smoke Jumpers" and all the behind the scenes work that happens to serve our Gemini models to billions of users. Thank you, Smoke Jumpers!
A conversation with my friend and colleague Emanuel Taropa on scaling Gemini’s serving infrastructure to support billions of users, the fun + craziness of new model launches, and more : )
Jeff Dean reposted
Today we’re launching Google AI Plus in 35 new countries and territories, making it available everywhere Google AI plans are. 🌎
Google AI Plus includes:
• More access to Gemini 3 Pro and Nano Banana Pro in the Gemini app
• More access to Veo 3.1 in @FlowbyGoogle and the Gemini app
• Expanded access to @NotebookLM
• 200 GB storage that you can share with up to 5 family members
Existing Google One Premium 2 TB subscribers in these countries will also automatically get access to all the benefits of Google AI Plus in the next few days.
Learn more about Google AI Plus: one.google.com/about/google-ai-plans/
Jeff Dean reposted
Google Search just got a big upgrade: Gemini 3 is now the default model for AI Overviews globally ✨ – making Search more helpful for the billions of people who use it every day
Jeff Dean reposted
We partnered with @ICC to show how Gemini 3 Pro can analyze video content.
By uploading a segment of the Cricket World Cup, Gemini can seamlessly process visual and audio data to identify key players, explain techniques, and highlight crucial turning points. 🏏
Jeff Dean reposted
Introducing ATLAS: New scaling laws for massively multilingual language models. We offer practical, data-driven guidance to balance data mix and model size, helping global developers better serve billions of non-English speakers. Learn more: goo.gle/49WYLL0
Jeff Dean reposted
🚀 Excited to officially launch 👁Agentic Vision via Gemini 3 Flash. Gemini can run code execution on image uploads to zoom, analyze, and annotate:
🔍 Zoom: 5-10% quality win across vision benchmarks
🧮 Analyze: do image math with code (e.g. calculate the tip for a receipt)
✏️ Annotate: Draw arrows or bounding boxes to answer questions
Try via the Gemini API (AI Studio / Vertex) or via the Gemini App (rolling out to Thinking mode today).
Learn more→ goo.gle/4bsKdFv
Demo:
goo.gle/3Z05KxK
cc: @IoanaBica95 @anastasija56572 @jalayrac @bcaine @eisenjulian @weichengkuo @phillip_lippe @xf1280 @tulseedoshi @BiboXu @OfficialLoganK
Try 👁 Agentic Vision with Gemini 3 Flash in @GoogleAIStudio or Vertex AI. This new capability enables the model to effectively use code and reasoning to improve performance for common vision tasks.
See Agentic Vision in action: goo.gle/3Z05KxK
Jeff Dean reposted
Introducing Agentic Vision — a new frontier AI capability in Gemini 3 Flash that converts image understanding from a static act into an agentic process.
By combining visual reasoning with code execution, one of the first tools supported by Agentic Vision, the model grounds answers in visual evidence and delivers a consistent 5-10% quality boost across most vision benchmarks. Here’s how the agentic ‘Think, Act, Observe’ loop works:
— Think: The model analyzes an image query then architects a multi-step plan
— Act: The model then generates and executes Python code to actively manipulate or analyze images
— Observe: The transformed image is appended to the model's context window, allowing it to inspect the new data before generating a final response to the initial image query
Learn more about Agentic Vision and how to access it in our blog ⬇️
blog.google/innovation-and-ai/technology/developers-tools…
This didn't seem good at all. Deep scientific expertise is needed in nearly every government agency to tackle all kinds of important problems.
We’ve lost over 10,000 STEM PhDs from the federal workforce in a single year.
10,000.
Never thought the U.S. would ever allow this type of brain drain to occur - but here we are.
I'm very sorry to hear this news. When I did a series of internships in high school at @CDCgov starting in 1985, he had just retired from a 16 year tenure as the CDC Director in 1983 and was spoken about in deeply revered terms in the hallways. Truly a giant of public health!
I’m deeply sad to learn of the passing of Dr. Bill Foege. Bill was a towering figure in global health—a man who saved the lives of literally hundreds of millions of people. He was also a friend and mentor who gave me a deep grounding in the history of global health and inspired me with his conviction that the world could do more to alleviate suffering.
Bill and I often talked together about the future of global health. The legacy of his career is that many of the remarkable developments to come will have his imprint all over them. gatesnot.es/4sZmDqd
Jeff Dean reposted
Wemby speaks out about the ICE shootings in Minnesota:
“PR has tried, but I’m not going to sit here and be politically correct… Every day I wake up and see the news and I'm horrified. I think it’s crazy that some people make it sound like the murder of civilians is acceptable”
Jeff Dean reposted
Free Climber Alex Honnold’s Insane Home Gym Setup Revealed
Jeff Dean reposted
Here's Steve Kerr when asked what the path moving forward is for the country...
"There should be an appeal to our better angels to look after one another and to recognize what's happening. We're being divided by media for profit, by misinformation. There's so much out there that is really difficult for all of us to reconcile.
In times like these, you have to lean on values and who you are and who you want to be -- either as an individual or as a country. And I think that's the biggest thing."
Jeff Dean reposted
A man who imitates the sounds he hears as close to the real thing as possible...and his friends' reactions.
Jeff Dean reposted
Story time!
I have a Japanese mom and dad!
In 2019, I was on my flight back to India and unfortunately I was recovering from a bad bout of influenza. I had cough in my inner ear and that led to vertigo every time I moved my position. Unfortunately I couldn't cancel so I carefully made my way to the airport and got on the flight. Next to me were seated this Japanese couple in their 70s who were going to vacation in Thailand. I had a mask on and looked really mean.
The Japanese guy was struggling with his entertainment system because he wanted Japanese subtitles and asked the cabin attendant for help. Unfortunately he barely spoke English and the cabin attendant was having a hard time understanding his request. I told the cabin attendant that I'd fix his problem. I asked him what he wanted in Japanese and 10 seconds later, he was happy. I thought that was the end of it but he asked why I could speak Japanese so well. We got talking and spoke for more or less the whole flight. The craziest thing is that his home was only a 20 min bike ride from mine. Talk about coincidence. Then when we got off he gave me his business card and asked me to call him.
I almost didn't call him but my dad insisted that I at least humor him. That one call gave me a Japanese dad and mom. Since then I have visited their home countless times and even stayed there. He practically insisted that I call them mom and dad. During covid they were the ones who helped me stay sane. At one point, my parents visited Japan and met my Japanese parents, a memorable event. And just as a parent would, he visited India in 2023 and was a witness at my marriage in the church.
Destiny is strange. You don't know who you will meet and become close with. Just say hi!
This American dude on his way to his Japanese fluency decides to talk to a local Japanese grandpa, and it turned out to be the most wholesome convo ever. 🥹❤️
Great to see so many CEOs of Minnesota-based businesses signing this open letter.
NEW: The Minnesota Chamber of Commerce released this letter on behalf of more than 60 CEOs of Minnesota-based companies today. www.mnchamber.com/blog/open-letter-more-60-ceos-minnesota…
Jeff Dean reposted
"Terrible things are happening outside. Poor helpless people are being dragged out of their homes. Families are torn apart.
Men, women, and children are separated. Children come home from school to find that their parents have disappeared."
Diary of Anne Frank
January 13, 1943
This is something we agree on: from all that I've learned in the last couple of days, Alex Pretti was a wonderful human being. Thank you for recognizing, @erichorvitz.
Values, service, and character.
Alex Pretti, the man who was killed by Border Patrol in Minneapolis on Saturday, was seen reading a final salute to a veteran who died in 2024 in a video shared by the service member's son.
Pretti, 37, was an intensive care nurse at a Veterans Affairs hospital who cared deeply about people and was upset by President Trump's immigration crackdown in his city, family members said. Pretti was a U.S. citizen, born in Illinois.
"Never wanted to share this video but his speech is very on point," Mac Rudolph wrote on social media. "Also my Fathers final words to me was to continue to fight the good fight. He would be honored in Alex' sacrifice, and ashamed of this current administration."
The Department of Homeland Security said the agent acted in self-defense after attempting to disarm Pretti, but Minnesota Gov. Tim Walz said that account was "nonsense" after reviewing videos of the shooting. www.cbsnews.com/news/alex-jeffrey-pretti-shot-killed-minn…
Well said, Chris. As my former intern, I've always been proud of the work that you did and continue to do, and I'm proud of the high moral-integrity person you are, as well! 🙏
I try to not talk about politics. I generally believe the best way I can serve the world is as a non-partisan expert, and my genuine beliefs are quite moderate. So the bar is very high for me to comment.
But recent events – a federal agent killing an ICU nurse for seemingly no reason and with no provocation – shock the conscience.
My deep loyalty is to the principles of classical liberal democracy: freedom of speech, the rule of law, the dignity of the human person. I immigrated to the United States – and eventually cofounded Anthropic here – believing it was a pillar of these principles.
I feel very sad today.
Jeff Dean reposted
BREAKING: A letter from Alex Pretti’s Final Nursing Student:
“I was Alex Pretti’s final nursing student. He was my friend and my nursing mentor. For the past four months, I stood shoulder to shoulder with him during my capstone preceptorship at the Minneapolis VA Hospital. There he trained me to care for the sickest of the sick as an ICU nurse. He taught me how to care for arterial and central lines, the intricacies of managing multiple IVs filled with lifesaving solutions, and how to watch over every heartbeat, every breath, and every flicker of life, ready to act the moment they wavered. Techniques intended to heal.
Alex carried patience, compassion and calm as a steady light within him. Even at the very end, that light was there. I recognized his familiar stillness and signature calm composure shining through during those unbearable final moments captured on camera.
It does not surprise me that his final words were, “Are you okay?” Caring for people was at the core of who he was. He was incapable of causing harm. He lived a life of healing, and he lived it well.
Alex believed strongly in the Second Amendment and in the rights rooted in our Constitution and its amendments. He spoke out for justice and peace whenever he could, not only out of obligation, but out of a belief that we are more connected than divided, and that communication would bring us together.
I want his family to know his legacy lives on. I am a better nurse because of the wisdom and skills he instilled in me. I carry his light with me into every room, letting it guide and steady my hands as I heal and care for those in need.
Please honor my friend by standing up for peace, preferably with a cup of black coffee in hand and a couple of pieces of candy in your pocket, just as he would. He would remind you that caring for others is hard work, and we must do whatever it takes to get through the long shifts. Step outside with your dog, breathe in the world, hike or bike as he loved to do, and let yourself find peace in the quiet moments within nature. Stand up for justice and speak with those whose views differ from your own. Hold your beliefs with strength, but always extend love outward, even in the face of adversity.
Take one step, no matter how small, to help heal our world. Through these acts, carry his light forward in his name. Let his legacy continue to heal.”
Jeff Dean reposted
Never gamble
[📹 jeremytanmagic]
Jeff Dean reposted
Veteran: "5 shots they put into that man's body. How do you look me in the eye and tell me you can justify this? I'm a veteran. I swore an oath in 1969 to protect this country against all enemies foreign and domestic. This is unbelievable."
This was delightfully wholesome.
This American dude on his way to his Japanese fluency decides to talk to a local Japanese grandpa, and it turned out to be the most wholesome convo ever. 🥹❤️
Jeff Dean reposted
JUST IN: In a new lawsuit, Minnesota officials say DHS mishandled evidence from today's crime scene and made "astonishing" law enforcement decisions. www.courtlistener.com/docket/72185880/minnesota-bureau-of…
Jeff Dean reposted
He wasn't a domestic terrorist. A domestic terrorist was the Oklahoma City bomber. This was a VA nurse looking out for his neighbors.
Fun! Fact: when I was in middle school we lived in Minneapolis, a block from Alan Page, a recently retired famous football player. He was a runner, as was my dad. My dad told me about him. He had just finished law school and was now starting his second career as a lawyer eventually becoming a Minnesota Supreme Court justice. Both of them would run in the dead of winter in Minneapolis, and my dad said he cared about people, so this photo doesn't surprise me.
You can read more about Alan Page here:
en.wikipedia.org/wiki/Alan_Page
(Excerpt below)
Thank you, Alan Page.
This is Alan Page protesting in Minneapolis today. Alan was a Minnesota Viking, is in the NFL Hall of Fame and is a retired Justice of the Minnesota Supreme Court. RESPECT!
Apparently he ran too much for the Vikings.
Jeff Dean reposted
To those who supported this administration because the system felt too rigged, too woke, too bureaucratic, too globalist…
No one voted for law-abiding citizens being shot in the streets.
If this crosses a line for you, now's the moment to say it loudly. Your voice matters.
Jeff Dean reposted
Jayson Tatum in 2016: “Way before high school, I’d already been to college. My mom had me when she was 19 years old. She was just a college freshman. But she was determined not to become another statistic, not to end up on welfare, not to drop out of school.
So she brought me to class with her.
From the time I was a baby until I was about eight years old, when my mom went to school, I went to school with her. I remember sitting in the back of her classes, eating snacks or immersing myself in books or video games. I kept quiet, listening in here and there — to me, most of her professors seemed boring and talked a lot. But I had my things to focus on, she had hers. It felt normal. So that’s what we did. When my mom couldn’t afford a babysitter and Grandma was working, we’d go to class together.
And by the time I got to sixth grade, my mom had gotten her bachelor’s and law degrees from St. Louis University.
I’ll never forget her law school graduation. All my cousins and grandparents showed up. When they called my mom’s name, I stood up in my button-down and slacks and screamed, ‘I love you! I’m proud of you!’
She did it, I told her her after the ceremony, but she corrected me. ‘We did it.’
Those nights when she was still in school, we’d sit at the dining room table together. We were each doing our homework. She’d be walking back and forth from the kitchen, cooking dinner while I asked her questions about my math assignments. (Mom was the best with math, she always found a way to break things down in terms that I could understand.) And when it was my bedtime she’d tuck me in and then return to the dining room table, staying up for hours, studying, reading, making sure she was keeping up with her own schoolwork.
She’d often say to me, ‘Jay, don’t let anyone tell you what you can or can’t be. No matter what.’”
Jayson Tatum’s Mom is a keeper. ❤️
Jeff Dean reposted
Imagine you're working your corporate job and Alex Honnald casually climbs up your window #skyscraperlive
Jeff Dean reposted
A statement from the family of Alex Pretti, obtained by CNN:
"We are heartbroken but also very angry.
Alex was a kindhearted soul who cared deeply for his family and friends and also the American veterans whom he cared for as an ICU nurse at the Minneapolis VA hospital. Alex wanted to make a difference in this world. Unfortunately he will not be with us to see his impact. I do not throw around the hero term lightly. However his last thought and act was to protect a woman.
The sickening lies told about our son by the administration are reprehensible and disgusting. Alex is clearly not holding a gun when attacked by Trump’s murdering and cowardly ICE thugs. He has his phone in his right hand and his empty left hand is raised above his head while trying to protect the woman ICE just pushed down all while being pepper sprayed.
Please get the truth out about our son. He was a good man. Thank you."
This is absolutely shameful. Agents of a federal agency unnecessarily escalating, and then executing a defenseless citizen whose offense appears to be using his cell phone camera. Every person regardless of political affiliation should be denouncing this.
Drop Site obtained harrowing footage of the latest killing which appears to be from the perspective of the woman in pink filming from the sidewalk
Congratulations, @hardmaru! I have time memories of your time at Google, and great to see this partnership.
I founded Sakana AI after my time at Google, so it is incredibly meaningful to be able to partner with them now. It feels like a special connection to be working together again to advance the AI ecosystem in Japan.
sakana.ai/google
We are thrilled to announce a strategic partnership with Google!
Google is also making a financial investment in Sakana AI to strengthen this collaboration.
This underscores their recognition of our technical depth and our mission to advance AI in Japan.
We are combining Google’s world-class products with our agile R&D to tackle complex challenges.
By leveraging models like Gemini and Gemma, we will accelerate our breakthroughs in automated scientific discovery.
Our work on The AI Scientist and ALE-Agent has already demonstrated the power of these models.
Now we are going further.
We are scaling our deployment of reliable AI in mission-critical sectors.
We are working with financial institutions and government organizations to deliver solutions that meet the highest standards of security and data sovereignty.
We are excited to drive the widespread adoption of reliable AI and advance Japan’s AI ecosystem together!
(I'm normally fond of autocorrect, but not in this case)
Jeff Dean reposted
A musician playing violin in the metro station gives the cute little girl who offers him money an experience she might never forget in her lifetime. 👏
Jeff Dean reposted
We are thrilled to announce a strategic partnership with Google!
Google is also making a financial investment in Sakana AI to strengthen this collaboration.
This underscores their recognition of our technical depth and our mission to advance AI in Japan.
We are combining Google’s world-class products with our agile R&D to tackle complex challenges.
By leveraging models like Gemini and Gemma, we will accelerate our breakthroughs in automated scientific discovery.
Our work on The AI Scientist and ALE-Agent has already demonstrated the power of these models.
Now we are going further.
We are scaling our deployment of reliable AI in mission-critical sectors.
We are working with financial institutions and government organizations to deliver solutions that meet the highest standards of security and data sovereignty.
We are excited to drive the widespread adoption of reliable AI and advance Japan’s AI ecosystem together!
Jeff Dean reposted
As of today, the United States has officially left the World Health Organization. We’ll look back on this as a grave error. Health threats do not respect borders, and weakening global cooperation makes Americans less safe.
Jeff Dean reposted
In a free society, you should be allowed to go get a snack without worrying about being detained by masked men from the government. www.instagram.com/mostafa_bassim?hl=en
Love to see this. 17 year old @ebarschkis solves interesting mathematical problem, discusses with Terrence Tao, and credits that he had "a lot of help from Bartosz Naskręcki" (@nasqret), and Bartosz says he actually provided very little help. The instinct to share credit broadly is a great one, @ebarschkis!
Enrique emailed me a few weeks ago to casually chat about the elliptic curve DLP. We tested many of his ideas using models and Aristotle. I am very proud of him: in the breaks between high-school classes, he carved his way into the mathematical frontier, at the age of 17!
My advice consisted of only modest hints and encouragement. Enrique deserves full credit for his daring and passion. Good luck, man-reach for the stars!
Computer scientists love SAT problems. They leave us feeling very satisfied.
Helpful update for students, you can now take full practice SATs for free in the @GeminiApp.
It uses vetted content from @ThePrincetonRev and gives you feedback straight away. Starting with the SAT today, but more tests are on the way!
Jeff Dean reposted
💔 “The goalkeeper told me I’m an illegal immigrant even though I was born in America, and he said Trump was gonna get me and send me back. It makes me really sad.”
Our kids are watching. (H/T Hot Lou)
Indeed, basic research funding is important and beneficial. My colleagues Urs Hözle and Partha Ranganathan recently did an interview-format-style discussion w/ Magdalena Balazinska about the importance of basic research and its influence on Google.
cacm.acm.org/federal-funding-of-academic-research/from-un…
Fun fact: The 1998 paper that introduced Google and PageRank to the world ends with this acknowledgment:
"Supported by the National Science Foundation under Cooperative Agreement IRI-9411306. Funding also provided by DARPA and NASA."
Sergey Brin was on an NSF Graduate Fellowship. Larry Page was a PhD student on the grant.
Google—now worth $2 trillion—exists because American taxpayers funded "the Stanford Integrated Digital Library Project."
Not a startup garage myth. A government grant.
Every time someone says public research funding "picks winners and losers" or "crowds out private innovation," remember: the most dominant technology company of the 21st century was incubated entirely with public money, inside a public university, by researchers on federal fellowships and grants.
The private sector didn't see it coming. VCs passed. The government funded it anyway—not because it would become Google, but because fundamental research into information retrieval seemed worth understanding.
That's the point. You can't predict which grants will change the world. You fund the science and let researchers explore.
The internet (DARPA). GPS (DoD). Touchscreens (CIA/NSF). mRNA vaccines (NIH). Google (NSF/DARPA/NASA).
Public investment in basic research isn't wasteful spending. It's the seed corn of the entire modern economy.
In the same issue, my colleague David Patterson recounts how ~$100M of funding over 40 yrs led to the development of numerous groundbreaking technologies like RISC, RAID, cluster computing+SW, ...
The article estimates return to taxpayers to be 1000X.
cacm.acm.org/federal-funding-of-academic-research/the-inn…
Jeff Dean reposted
New year, same pace 💨 Here’s what we’ve launched recently:
— A beta version of Personal Intelligence in @GeminiApp. With Personal Intelligence, you can choose to connect certain Google apps like Gmail and Photos to get more tailored responses that are uniquely helpful to you. Rolling out to U.S. Google AI Pro and Ultra subscribers now, with expansions to come
— @Gmail in the Gemini era, transforming your inbox into a personal, proactive assistant that helps you get more done with features like AI Overviews, Help Me Write, Suggested Replies, and AI Inbox
— LLM flexibility, Agent Mode improvements, new agentic experiences, and more in the @AndroidStudio Otter 3 Feature Drop
— Support for Agent Skills in @antigravity. Agent Skills are reusable packages of knowledge (instructions, best practices, and optional scripts and resources) that extend what an agent can do
— Native vertical formats (9:16) in Veo 3.1 Ingredients to Video and state-of-the-art upscaling to 1080p and 4k in Veo
— Verification of Google-AI generated videos directly in the Gemini app
— MedGemma 1.5, an updated open medical AI model that improves 3D imaging and medical record reasoning performance
— TranslateGemma, a new family of open translation models with support for 55 language pairs
Much better language & translation models across many languages make AI more accessible to everyone, and enable us all to communicate better with the world. I helped push for us to gather better and more multilingual training data, and TranslateGemma is one downstream result!
We’re releasing TranslateGemma, a new family of open translation models with support for 55 languages. 🌐
Available in 4B, 12B, and 27B parameter sizes – they’re designed for efficiency without sacrificing quality.
Read more about our 1000 languages initiative here:
blog.google/innovation-and-ai/products/ways-ai-is-scaling…
Jeff Dean reposted
🎬 Veo 3.1 is leveling up:
+ Portrait mode
+ 1080p and 4K (both only in Flow for now)
+ Better expressiveness
+ Better adherence to your reference images
+ Available for Plus, Pro, and Ultra members
Rolling out today in @GeminiApp and @FlowbyGoogle -- enjoy!
Jeff Dean reposted
MedGemma 1.5 is a major upgrade to our open models for healthcare developers.
The new 4B model enables developers to build applications that natively interpret full 3D scans (CTs, MRIs) with high efficiency - a first, we believe, for an open medical generalist model. MedGemma 1.5 also pairs well with MedASR, our speech-to-text model fine-tuned for highly accurate medical dictation.
Developers can now use these multimodal capabilities to build medical apps that reach patients in more places.
Allowing Gemini to securely access your emails, calendar, etc. turns it into a personalized @GeminiApp for you!
Answering a top request from our users, we’re introducing Personal Intelligence in the @GeminiApp. You can now securely connect to Google apps for an even more helpful experience.
Personal Intelligence combines two core strengths: reasoning across complex sources and retrieving specific details, e.g from an email or photo, to provide uniquely tailored answers.
It’s built with privacy at the center. You choose exactly which apps to connect, these connected app settings are off by default.
Veo 3.1 offers even better quality, expressiveness and consistency. I like the latte art coming to life in the examples below! ☕
We’re updating Veo 3.1 Ingredients to Video to help create more expressive and dynamic clips, produce better visual consistency and more. 📽️
Here’s what’s new 🧵
An updated version of our MedGemma model is out with major accuracy improvements in medical-related tasks. We also released MedASR, which specialized for low error rates for medical speech recognition. We're excited to see what the health community does with these models! 🩺
Introducing MedGemma 1.5, an open-access model for multimodal medical use cases
It expands the tasks and data formats it can understand (high-dimensional medical imaging, EHRs, anatomical localization with bounding boxes, etc)
research.google/blog/next-generation-medical-image-interp…
Jeff Dean reposted
OMG. Renee Good’s car was
driven away from Minneapolis completely open to the elements, windows wide open.
No bullet-trajectory testing. They rushed it out of the state before state investigators could examine it. This is a fucking criminal cover up.
Jeff Dean reposted
28% of Americans think ICE was justified in shooting Renee Good
53% think the ICE agent should face criminal charges
Jeff Dean reposted
Apple: "“After careful evaluation, we determined that Google’s technology provides the most capable foundation for Apple Foundation Models"
www.theverge.com/news/860521/apple-siri-google-gemini-ai-…
I'm excited to see us partner with Apple to bring Gemini models to Apple users, powering Apple Intelligence features!
Joint Statement: Apple and Google have entered into a multi-year collaboration under which the next generation of Apple Foundation Models will be based on Google's Gemini models and cloud technology. These models will help power future Apple Intelligence features, including a more personalized Siri coming this year.
After careful evaluation, Apple determined that Google's Al technology provides the most capable foundation for Apple Foundation Models and is excited about the innovative new experiences it will unlock for Apple users. Apple Intelligence will continue to run on Apple devices and Private Cloud Compute, while maintaining Apple's industry-leading privacy standards.
Just a bit cheaper
Average cost for 1 gigabyte of storage:
45 years ago: $438,000
40 years ago: $238,000
35 years ago: $48,720
30 years ago: $5,152
25 years ago: $455
20 years ago: $5
15 years ago: $0.55
10 years ago: $0.05
5 years ago: $0.03
Today: $0.01
Jeff Dean reposted
Some countries that have prosecuted or threatened to prosecute central bankers for the purpose of political intimidation or punishment for monetary policy decisions: Argentina, Russia, Turkey, Venezuela and Zimbabwe.
Jeff Dean reposted
You know things are bad when the most boring person doing the most boring job for decades has to come out and say, “Stop trying to arrest us for doing math”
Video message from Federal Reserve Chair Jerome H. Powell: www.federalreserve.gov/newsevents/speech/powell20260111a.htm
Thank you, @nsthorat. Very well said. We all bear a collective responsibility to speak up and not be silent when we see things like the events of the last week.
The murder in Minneapolis by the government’s secret police this week has stirred something in me.
A mother was gunned down in the street by ICE, and the government doesn’t even have the decency to perform a scripted condolence. No matter which way the tires were pointing, this is a tragedy for her family, her children, and yet JD Vance, Palmer Lucky, and others instantly became the judge and the jury on twitter proclaiming the 37 year old white mother was a domestic terrorist and was trying to murder a federal officer; thus she *deserves* to be shot in the face. And you do too if you don’t comply.
Where is the humanity? How is everyone not terrified that this could happen to them, their mother, their sister? Isn’t it terrifying that the Vice President is immediately defending murder of american citizens?
I normally keep my politics quiet here because of how toxic twitter is and how annoying it can be to people who want a clean tech feed. Right now it just feels wrong not to say anything.
I haven’t quite been able to figure out the right words for how I feel. My first instinct is “fuck ICE” and “what the fuck is wrong with these people”. Those are my lizard brain reactions. But there’s something deeper that’s bothering me.
I can’t shake this feeling that the moral foundation that much of modern society stands upon is infected, and is festering. We teach kids that right and wrong exist, not to lie, that you don’t hurt people, and that you protect the vulnerable. And yet our highest elected officials encourage the burning and pillaging of these sacred lessons.
I could go on about how our current leadership just wants money and power, and they’re using immigrants as scapegoats to do so. Our economy is transforming rapidly, both with AI and because of greedy capitalist policies — the country is angry about legitimate things and democrats are not actually listening. Republicans are listening but are redirecting anger for their own power. The next 5 years are going to look a lot different because of technology; yet our heads are buried in the sand and we’re losing our way.
But maybe the only real thing we can do is to keep living and teaching the lessons of right and wrong, not to lie, that you don’t hurt people, and that you protect the vulnerable. Teach your kids these lessons in spite of Trump, Vance, Noem being lost beyond repair. Teach your kids that everyone deserves to be treated as a human being and not an object to be deported to a brutal prison cell.
I do believe that most people are still good, and that the best thing we can do to combat dehumanization by the government is more humanization of our neighbors. Go spend time with your republican cousins / neighbors and break bread with them. Don’t talk politics, play with their kids. Find common humanity.
But I don’t really have a call to action beyond this. I don’t really know what to do other than vote in November and refuse to pretend this is normal, publicly, even if it’s terrifying that you might get targeted and attacked. People in Nazi Germany stayed silent for this very fear, and we’re certainly living through a cosplay of that now.
The president called a 37 year old mother a domestic terrorist and that her 3 shots to the face by the gestapo was a tragedy of her own making.
The actual tragedy is that the Oval office propped up millions of people arguing about the direction of the tires of the car without having the humility to step back and ask if we want the government murdering civilians in broad daylight.
I am terribly sad for this country, our broken moral compass, and for hard working immigrants living in fear here. But I do believe we will eventually heal.
Jeff Dean reposted
BREAKING: ICE agent in Minneapolis just violently SHOVED a man filming their vehicle to the ground.
This is illegal, as it is excessive force. Courts routinely rule that throwing or violently pushing a non-threatening person to the ground for a minor obstruction is unconstitutional.
The legal remedy here would be to pull the man away or gently push him or escort him away.
Jeff Dean reposted
Video message from Federal Reserve Chair Jerome H. Powell: www.federalreserve.gov/newsevents/speech/powell20260111a.htm
Jeff Dean reposted
🚨 BREAKING: ICE/Border Patrol agents are now ramming civilian cars, forcing them through red lights, and driving away in Minneapolis.
In this video, you can see an agent who has their bumper right up against a car stopped at a red light… no emergency lights, no sirens.
The agent then accelerates and physically pushes the car through the intersection while the light is still red. The car turns. The agent turns. Then the agent just drives off.
That is not enforcement. That is reckless, dangerous, and illegal.
Federal agents do not have the legal authority to hit civilian vehicles, force them into traffic violations, or flee the scene. That’s vehicular assault. That’s abuse of power. And it puts everyone on the road at risk.
If local police did this, people would be calling it exactly what it is.
So why are federal agents being allowed to do it in plain sight?
Jeff Dean reposted
"Get back or I'm gonna shoot you."
Ice Agent
Absolutely reprehensible.
🚨BREAKING: WARNING — DISTURBING FOOTAGE
Yesterday in Minneapolis, ICE agents brutally beat and illegally arrested a legal U.S. resident… all caught on camera.
In the video, five ICE agents are seen on top of a man who is face down on the concrete.
One agent repeatedly knees him in the face while the man does not resist. He is motionless, in pain, and screaming in agony.
The assault only stops when the agent realizes a legal observer is filming. This is not an allegation. It is recorded violence.
And it gets worse.
The man, identified as Juan Carlos, is currently unhoused and living out of his car.
ICE agents unlawfully approached him, ripped him from his vehicle, violently beat him, arrested him, and transported him to the Whipple Detention Facility.
Juan Carlos posed no threat. He is a legal U.S. resident. Yet, he was assaulted and treated like a criminal.
While in custody, he began having difficulty breathing. Only then did agents transport him to a hospital, still in shackles. He had blood in his ears from the beating, yet, was denied medical care until he was in medical distress.
Once at the hospital, ICE agents attempted to intimidate medical staff and interfere with care. After approximately two hours, they finally issued Juan Carlos a ticket and left.
To be absolutely clear:
-ICE agents beat a legal U.S. resident.
-They illegally arrested him.
-They attempted to hold him in a detention facility.
-They withheld medical care until his condition worsened.
-They only backed down once there were witnesses, video evidence, and hospital staff present.
If Juan Carlos had not been taken to the hospital… a place with mandated witnesses and documentation… it is very possible a legal U.S. resident would still be sitting in an immigration detention center today.
And this is the part they don’t want you to realize:
Immigration detention is a profit-driven system. Private prison corporations make money by keeping beds full. Every illegal arrest, every unlawful hold, every person thrown into a detention center without cause is revenue for a private company.
When law enforcement agencies can illegally beat and attempt to warehouse legal residents for profit, this stops being about immigration enforcement and becomes a business model built on human suffering, and human trafficking.
This is not a misunderstanding.
This is not “questioning.”
This is state violence, enabled by a system that profits from illegal confinement.
And it will continue until it is forced to stop.
This is not okay. Regardless of her citizenship (and witnesses indicated she is a U.S. citizen), this kind of thuggish behavior by ICE agents is completely over the top. Choking her & gouging her eyes is excessive force. The government is supposed to be in service of all of us.
ICE agents violently assaulted and restrained a woman in Minneapolis, gouged her face, and threw her into an unmarked car while witnesses shouted, “She’s a U.S. citizen!”
Jeff Dean reposted
Dear @DHSgov: The American people have the First Amendment right to witness, photograph and film ICE and other DHS operations in public.
Can you explain why this officer appeared to have assaulted a woman for exercising her First Amendment rights?
An ICE Agent in Minnesota showing their current mentality
Jeff Dean reposted
If our only option is “obey the nameless masked men or die,” we do not live in a free country
This is completely not okay, and we can't become numb to repeated instances of illegal and unconstitutional action by government agencies. The recent days have been horrific.
🚨BREAKING: Border Patrol just illegally arrested a U.S. citizen… a TEENAGER… in Richfield, Minnesota.
The teen was working at Target when agents tackled him and arrested him, while his passport was in his pocket, fully identifying himself as a U.S. citizen.
None of it mattered.
Why? Because he looked Latino.
When a coworker asked who to call, he said, heartbreakingly: “my mom.”
Think about that. Your teenage son goes to work, and Border Patrol kidnaps him.
This isn’t an isolated incident. ICE and Border Patrol have repeatedly illegally arrested U.S. citizens, especially in Minnesota. And after agents murdered a U.S. citizen who tried to drive away from illegal detention, this should outrage every single American.
U.S. citizens do not have to prove their citizenship. Yet, ICE and Border Patrol agents are routinely violating the Constitution, demanding proof anyway… and even refusing to accept drivers licenses or real IDs.
And now? Even a passport doesn’t matter.
This isn’t enforcement… it’s targeting Americans based on fear and appearance, treating citizens like criminals in their own country.
When will this end? When will the government stop stripping citizens of their rights and labeling them terrorists for simply existing?
Jeff Dean reposted
1- The Police Use-of-Force Rule on Moving Vehicles
Core Police Rule taught nationwide: Officers should not intentionally place themselves in front of or in the path of a moving vehicle and then use deadly force to stop it.
2- DOJ / FBI / National Policing Standards
The U.S. Department of Justice and FBI training doctrine states: A moving vehicle alone is not considered a deadly threat. Officers must attempt to move out of the vehicle’s path rather than fire.
Deadly force is not justified if: The officer voluntarily placed themselves in front of the vehicle. The vehicle is attempting to flee. There are other avenues of escape
3- International Association of Chiefs of Police (IACP)
The IACP model policy says:
Officers shall not fire at or from a moving vehicle unless:
There is an imminent threat other than the vehicle itself, and No reasonable alternatives exist
Standing in front of a vehicle creates the danger and invalidates the justification.
4- Minnesota training explicitly teaches:
Officers must avoid positioning themselves in front of vehicles during enforcement actions
Use of deadly force is not justified when the danger results from the officer’s own tactical decisions
This is drilled into recruits.
5- Supreme Court Context
While not a numbered rule, courts repeatedly rule that:
An officer cannot manufacture a deadly-force scenario and then claim self-defense.
This is known legally as “officer-created jeopardy.”
5- How It’s Taught to Police
Police are taught:
Step aside
Let the vehicle go
Identify later
Do not block with your body
Do not draw a gun in front of a car
Why?
Because:
Cars are common
People panic
Officers escalate situations unnecessarily
Innocent people get killed.
Jeff Dean reposted
Initially I was surprised at people's ability to ignore evidence leading to conclusions they don't like, even with multiple videos of the scene. But then I remembered that 34% of Americans don't believe in evolution.
Many more Gemini integrations with Gmail. No foolin'!
We launched Gmail on April Fool’s Day in 2004. 20+ years later, we’re bringing Gmail into the Gemini era.
AI Overviews, Suggested personalized replies, Proof read, AI Inbox with new streamlined views and suggested topics to catch-up on and loads more, read the full details here: blog.google/products-and-platforms/products/gmail/gmail-i…
Give our Google Search AI mode (powered by Gemini models) a try. Easy to remember url, as well: g.ai
In another thread, one user seemed surprised that you can ride in the front passenger seat in a @Waymo. You definitely can, and you can preconfigure the vehicle so it arrives with extra legroom wherever you want it (front, balanced, or rear).
Also, pro tip: you can configure the seat setup in the Waymo app so that your car arrives with more legroom in the front (or balanced or more in the rear if you prefer)
My first @Waymo ride on the highway as I head from a meeting in Redwood City to Gradient Canopy in Mountain View.
Nice work, Waymo team! 🎉
Also, pro tip: you can configure the seat setup in the Waymo app so that your car arrives with more legroom in the front (or balanced or more in the rear if you prefer)
Jeff Dean reposted
So @google gifted me this incredible custom @lego set while I was in Bali 🏝️
This isn’t just any LEGO build ... it’s a tribute to one of the most influential papers in computing history -
"MapReduce: Simplified Data Processing on Large Clusters" by @GhemawatSanjay & @JeffDean
In 2024, I was delighted to be a co-author on an Arxiv paper on"Shaping AI's Impact on Billions of Lives", covering areas we thought AI would have significant impact on the world. We cover AI's potential impact on employment, education, healthcare, (mis)information, media, national security, and science, and propose 18 potential moonshot research directions in these areas to maximize positive impact and minimize downsides. A shortened version appears as the cover article in this month's Communications of the ACM (CACM).
CACM version: cacm.acm.org/research/shaping-ais-impact-on-billions-of-l… (~12 pages + video discussion with David Patterson)
Arxiv: arxiv.org/abs/2412.02730 (expanded 30 page version)
You can find these and more at: shapingai.com
It was a delight to work on this with awesome co-authors! Authors: Mariano-Florentino Cuéllar, myself (@JeffDean), @FinaleDoshi, John Hennessy, @andykonwinski, @sanmikoyejo, @tamati_biskit, @2plus2make5, and David Patterson.
(PDF link of CACM version is a bit hard to find. It's at dl.acm.org/doi/pdf/10.1145/3746132)
Excited to pair our @GoogleDeepMind robotics efforts, where we've been working on robotic learning for many years (including variants of our Gemini models especially suited for visual understanding and robotic action) with the very capable robotics hardware from @BostonDynamics.
Google DeepMind 🤝 @BostonDynamics
Our new research partnership will bring together our advancements in Gemini Robotics’s foundational capabilities to their new Atlas® humanoids. 🦾
Find out more → goo.gle/49paguA
Jeff Dean reposted
Who would’ve guessed that making the train faster and more frequent via electrification would result in massive gains in ridership?
Caltrain ridership jumps 57% in 2025 with its first full year of electrified service. Funding uncertainty still looms, highlighting the need for sustainable transit support ahead of the 2026 sales tax vote.
www.mercurynews.com/2025/12/24/caltrain-ridership-growth-…
I'm very excited about the possibilities of AI for improving learning and education. Gemini shows some of this potential, and it's only going to get better.
Holy sh*t.
Gemini can now produce fully interactive images on any topic.
Such an insane resource for learning → highlight any region, and it gives you a full explanation 🤯
Jeff Dean reposted
Google just made a $2,000 AI education free (and most people will still choose to stay illiterate).
They quietly launched Google Skills, a hub with 3,000 technical modules that replaces "prompting" fluff with actual DeepMind research workflows. You can now access the exact curriculum used to train their internal teams on transformer architecture for $0.
If you don’t use this, you’ll eventually complain about the people who did.
Jeff Dean reposted
America if we connected all cities with a population over 500,000 via high-speed rail:
This art installation by @refikanadol is amazing. It makes me smile every day when I walk by. Thanks for brightening up the office, Refik!
Dear friends, I’m excited to share our final and most ambitious AI arts project for 2025 — Machine Dreams: Biophilia, commissioned by Google for their new headquarters, Gradient Canopy.
The project began nearly two years ago to support Large Nature Model research. To create this work, Refik Anadol Studio collected 2 petabytes of audio and imagery from four distinct California biomes by traveling through the Redwoods, Ocean, Bay, and Foothills. It was combined with publicly available ecological datasets to train our Large Nature Model (LNM) sustainably in Google Cloud using renewable energy.
Local weather data from each biome is used in real time to generate text prompts in the latest Gemini, guiding the LNM’s exploration of the datasets. As the weather changes, so do the dreamed images and evolving soundscapes, reflecting nature’s constant motion—a living AI Data Painting that harmonizes perfectly with nature. The breakthrough technology behind the artwork will also power Dataland, and I cannot wait to share more soon. I am deeply grateful to the team that worked so hard on this epic and unique project, and to Google Leadership and scientists for their ongoing support of our studio since 2015!
Sigh: it'd be nice if we used the same system of measurement as ~the rest of the world. Some history: the Metric Conversion Board was abolished by President Reagan in 1982.
Executive order 12770 was then signed by President H. W. Bush in 1991, but has only had partial effect.
President Ford signs into law the U.S. Metric Conversion Act of 1975, which will guide the gradual replacement of the English system of measurement with the metric system in the United States.
It's the US, Liberia and Myanmar, fyi.
FWIW, my New Year's resolution a few years ago was to use metric when discussing distances, temperatures, etc., and I'm still sticking to it. If you're looking for a good New Year's resolution this year, I figure we just need ~340 million more to join and we're there!
Jeff Dean reposted
Our version of holiday lights.
I'm delighted to jointly author this year-end summary of research advances with @DemisHassabis and James Manyika, on behalf of all of our colleagues across @GoogleDeepMind, @GoogleResearch and @Google.
We look at research advances across eight different areas. These summaries are always fun to work on because one can reflect back on the breadth and depth of our collective work over the last year!
blog.google/technology/ai/2025-research-breakthroughs/
It has been a productive 2025! It's wonderful working with such amazing colleagues at @GoogleDeepMind and all across @Google to conduct new research and to put research into practice across many of Google's products!
See you in 2026.
🤝 @GoogleAI
🤝 @GoogleResearch
🤝 @GoogleQuantumAI
Performance Hints
Over the years, my colleague Sanjay Ghemawat and I have done a fair bit of diving into performance tuning of various pieces of code. We wrote an internal Performance Hints document a couple of years ago as a way of identifying some general principles and we've recently published a version of it externally.
We'd love any feedback you might have!
Read the full doc at: abseil.io/fast/hints.html
The public version is a somewhat sanitized version of the internal version. People at Google can find the internal version at go/performance-hints, which has links to the actual changelist in our source code repository system.
The internal version also has additional examples in code that is more sensitive (e.g. changes to the TPU compiler backend) that we aren't able to include in the external version.
You might also enjoy this @NewYorker article about our working style together.
In April, '17, @jsomers of @NewYorker reached out & said he wanted to do a small profile of me & my longtime colleague Sanjay Ghemawat, watch us work for a few hours, maybe dinner, etc.
It came out today. I think it captures our working style really well. www.newyorker.com/magazine/2018/12/10/the-friendship-that…
Jeff Dean reposted
Science is too slow.
At Edison, we are integrating AI Scientists into the full stack of research, from basic discovery to clinical trials. We want cures for all diseases by mid-century.
We have raised a $70M seed to get started.
Join us.
We need cracked software engineers who want to work on finding cures rather than selling ads and generating slop. If you’re reading this, you’re probably a candidate.
We need brilliant AI researchers who want to figure out how AI will accelerate real-world science.
We need scientists and researchers with deep expertise in biology, biotech, and pharma who want to figure out how to integrate AI deeply into scientific workflows, from ideation to experimentation, and how to measure success or failure.
We need extraordinarily talented generalist operators across BD, sales, product management, and partnerships who can focus on getting our tools into the hands of pharmaceutical companies.
If any of these roles sound like you, get in touch.
We are also expanding access to our platform.
Our goal is to accelerate science writ large. To that end, we will continue to give academics and students 650 credits/mo indefinitely. I can’t promise we’ll keep this up forever, but we will try. Kosmos will still cost 200 credits, and the other agents (Analysis, Literature, etc.) will cost 1 or 2 credits.
All paid users will have access to our regular agents, like our Analysis agent, Literature agent, and so on, for free via the UI. API access will still be paid, and users without a paid subscription will continue to get 10 credits per month for those agents. Our $200/mo subscription for 650 credits/mo is staying in place for now, but might be phased out at our next major product update.
Along the lines of accelerating science, we’re also doing a major release of PaperQA today, our flagship open source literature agent, as part of our commitment to open science.
In the short run, expect major improvements to Kosmos, including the ability to automatically access data, the ability to steer its exploration, and the ability to converse directly with its world model.
In the long run, expect exponentially increasing rates of scientific discoveries, in biology and elsewhere.
Our round is led by Triatomic Capital, Spark Capital, and a major US institutional biotech investor. We are also joined in this round by existing investors Pillar VC and Susa Ventures, two exceptional early-stage funds who backed us at founding, along with Striker Venture Partners, Hawktail VC, Olive VC, and a host of exceptional angels that includes famous AI researchers, the CEOs of multiple frontier AI labs, and leadership of major biotech and pharma companies.
Jeff Dean reposted
The public version is a somewhat sanitized version of the internal version. People at Google can find the internal version at go/performance-hints, which has links to the actual changelist in our source code repository system.
Jeff Dean reposted
Holy moly!
XLA changes is what Jeff did when he was still SVP of Google Research, silently lurking in our chatroom seeing me cry about compilation time for ULM Code model that is a precursor to Flash, and Jeff optimizing XLA during his meetings within a span of hours!
Performance Hints
Over the years, my colleague Sanjay Ghemawat and I have done a fair bit of diving into performance tuning of various pieces of code. We wrote an internal Performance Hints document a couple of years ago as a way of identifying some general principles and we've recently published a version of it externally.
We'd love any feedback you might have!
Read the full doc at: abseil.io/fast/hints.html
Awesome to see this overview of work done by many, many people in @GoogleResearch over the past year. Learn about advances in:
Advances in generative models
Generative user interfaces
Quantum Computing
AI for Scientific Discovery
Biomedical & Neuroscience Research
Climate & Sustainability
Privacy & Security
Novel Model Architectures
...and more...
Read the blog post at research.google/blog/google-research-2025-bolder-breakthr…
In 2025, the magic cycle of research accelerated. Google Research teams delivered pioneering breakthroughs and brought our research to reality with impact on products, science, and society.
More from @ymatias, VP and Head of Google Research →goo.gle/499MMJQ
Jeff Dean reposted
2 brand new experiments. A ton of updates. One wild ride this month. 🎢
Here's a recap of what we dropped:
— Doppl: Received a major upgrade with a personalized Fashion Discovery Feed, built-in links to directly shop your looks, and a Collections feature to save your favorite fits.
— Mixboard: Introduced a bundle of new features, including Presentations with Nano Banana Pro, new file-type uploads, multi-board projects, and more.
— Pomelli: Launched ‘Animate’, allowing you to turn your content into on-brand animations powered by Veo 3.1.
— @JulesAgent: Became a proactive collaborator, now surfacing vital tasks and automating fixes before you even have to ask.
— @StitchbyGoogle: Launched a flurry of updates including Gemini 3 Pro, prototypes, working HTML from Nano Banana Pro redesigns, Command Palette (Cmd+K), and Predictive Heatmaps.
— Disco: New this month! Disco is an experimental way to discover innovative approaches to browsing the web, using Gemini 3. The first feature, GenTabs, uses Gemini 3 to remix your open tabs into custom apps.
— CC: New this month! CC is an experimental AI productivity agent in Gmail. Get a “Your Day Ahead” briefing every morning in your inbox and email CC anytime for help.
— @NotebookLM: Released a series of mobile updates, including Infographic and Slide Deck generation, saving Audio Overview progress across devices, and supporting camera images as sources. Plus, there are now increased character limits for Chat customization prompts, added DocX support, and NotebookLM just joined the Google AI Ultra Plan.
— Opal: Officially landed in the Gems tab of @GeminiApp AND introduced Nano Banana Pro.
— @FlowbyGoogle: Launched Object Removal to help you refine your clips and remove specific details, and introduced the ability to download Nano Banana Pro images in 2K / 4K.
What a way to close out the year. Thank you, Labs community, for joining us on this journey. Oh, and for 2026? Buckle up. We can’t wait to share what's next. 🚀
Explore all the latest at: labs.google
Starting now!
I'm looking forward to chatting with my Gemini co-leads @OriolVinyalsML and @NoamShazeer in a conversation with @OfficialLoganK today at 1 PM Pacific. We'll talk Gemini, ML research developments, what are important research directions, etc.
x.com/i/spaces/1eaJbjvBOooJX
I'm looking forward to chatting with my Gemini co-leads @OriolVinyalsML and @NoamShazeer in a conversation with @OfficialLoganK today at 1 PM Pacific. We'll talk Gemini, ML research developments, what are important research directions, etc.
x.com/i/spaces/1eaJbjvBOooJX
It was delightful to do a fireside chat with my longtime colleague @geoffreyhinton a couple of weeks ago, moderated by @JordanJacobs10. The recording of our full discussion is now on Spotify:
open.spotify.com/episode/2zM1FkXwxspjK1OlX7wMSU
On the latest Radical Talks podcast: @GeoffreyHinton X @JeffDean - a friendship that defined modern AI as breakthrough theories met massive scale. Recorded at #neurips2025, Radical Co-Founder @JordanJacobs10 sits down with these two icons of AI to discuss one of history’s most productive collaborations.
As @JacksonWharf points out, the video version is also in YouTube.
Also on Youtube 😂
youtu.be/ue9MWfvMylE
Jeff Dean reposted
As a university professor, I rarely hear from students after the final exam. However, there are always a handful of students who will send personal notes or emails after the semester is over. These mean more to me than you could possibly imagine.
Students, if you enjoy the class, let the teacher know. They truly love hearing it.
Jeff Dean reposted
Gemini 3 Flash is starting to roll out as the default model for AI Mode in Search globally.
AI Mode can now tackle your most complicated questions with greater precision — without compromising speed. ⚡
With this upgrade, AI Mode becomes an all-around more powerful tool, better at understanding your needs. You can ask more nuanced questions and it will consider each aspect to provide a thoughtful response. And as always, you’ll have access to real-time information and useful links from across the web.
Jeff Dean reposted
In @GeminiApp this is what you're working with now:
Fast = 3 Flash
Thinking = 3 Flash (with thinking)
Pro = 3 Pro (with thinking)
Our best line-up yet!
Jeff Dean reposted
Flash 3.0 is out!
Culmination of years of high-conviction infra & science investment (we knew it'd be this good a priori). I truly believe no other place trains these models the way we do.
I'm lucky to get to work so intimately with the team that made this happen. We aren't slowing down.
3 Flash delivers frontier performance on benchmarks like GPQA Diamond - evaluating PhD-level reasoning – and Humanity’s Last Exam – testing broad expert knowledge.
It’s state-of-the-art on MMMU Pro, with a score comparable to 3 Pro - easily analyzing inputs across videos and images, not just text.
And it handles complex tasks significantly faster than 2.5 Pro at a lower cost, using fewer tokens - or units of information - to save time.
Interesting graphs. I suspect the very steep drops on each line on the left graph are at least partially the effect of distillation, as we can make smaller student models nearly as good as their much larger/costly teacher models.
2/ Prices for intelligence are collapsing.
Models at the GPT-4 level are now roughly 1000× cheaper than in 2023.
Yet, average prices paid per token fall more slowly because users systematically upgrade to higher intelligence.
Jeff Dean reposted
So, where can you use Gemini 3 Flash? The short answer: pretty much everywhere ⚡️
We’re bringing our latest model to the products and tools you already use, like @GeminiApp, AI Mode in Search, @GoogleAIStudio, Vertex AI and more.
Here’s where you can find it starting today ↓
Jeff Dean reposted
We have released a new agentic frontier eval for Roblox game development in Lua.
github.com/Roblox/open-game-eval
Gemini Flash 3 is currently the best model for Roblox game development.
Congrats @JeffDean @OriolVinyalsML @demishassabis and team - you really cooked !
RT @maylee0714: Our team just finished running our internal evals on the new Gemini 3 Flash model:
• 26 scenarios comprised on highly compl…
Jeff Dean reposted
Google has released Gemini 3 Flash Preview - 2x cheaper than Gemini 3 Pro Preview, with only a 2-point drop in our Intelligence Index, making it the most intelligent model for its price range
@GoogleDeepMind gave us pre-release access to Gemini 3 Flash Preview. The model scores 71 on the Artificial Analysis Intelligence Index, a 13 point improvement from Gemini 2.5 Flash (Sep), making it the most intelligent model for its cost.
Gemini 3 Flash Preview has particularly strong knowledge and reasoning abilities, obtaining the highest score in our knowledge and hallucination benchmark, AA-Omniscience, and placing second in Humanity’s Last Exam.
Google now holds the top two spots on both of these evaluations, cementing them as a leader in model knowledge. This increased performance comes with a tradeoff, as Gemini 3 Flash Preview more than doubles token usage vs Gemini 2.5 Flash (Sep) when running the Artificial Analysis Intelligence Index, making it one of the highest token use models we’ve tested.
Key takeaways:
📖 Significant intelligence improvements: Gemini 3 Flash Preview has significant improvements across nearly all evaluations in the Artificial Analysis Intelligence Index. It has particular strengths in reasoning settings, scoring second to Gemini 3 Pro Preview in Humanity’s Last Exam (35%), and third in both MMLU-Pro (89%) and GPQA Diamond (90%) (behind Gemini 3 Pro Preview and GPT-5.2 xhigh).
🧠 AA-Omniscience performance: Gemini 3 Flash Preview achieves the highest score in our knowledge and hallucination benchmark, AA-Omniscience. This is driven by increased accuracy (percentage correct), rather than lower hallucination. The model has the highest knowledge accuracy of any model tested, but has a hallucination rate of 91%, 3 percentage points higher than Gemini 2.5 Flash and Gemini 3 Pro Preview. We measure hallucination rate as how often the model answers incorrectly when it should have refused or admitted to not knowing the answer.
🖼️ Multimodal capabilities: Gemini 3 Flash Preview is a multi-modal model, with the ability to take text, images, video and audio as input. It scores the second highest of any model on MMMU-Pro, a benchmark that tests reasoning abilities with image inputs, behind only Gemini 3 Pro Preview.
➕ Significantly increased token use: Gemini 3 Flash Preview uses ~160M tokens on the Artificial Analysis Intelligence Index, more than double used by Gemini 2.5 Flash (Sep). This makes it one of the highest token use models we have tested, beating out other high use models Kimi K2 thinking and Grok 4 (thinking).
$ Cost efficiency: Despite its high token use, Gemini 3 Flash Preview is still the most cost efficient model for its level of intelligence, measured by the overall cost to run the Artificial Intelligence Index. This is driven by low token prices of $0.5/$3 per 1M input/output tokens.
⚡ Speed: Gemini 3 Flash Preview is 22% slower than Gemini 2.5 Flash (Sep), measuring at 218 output tokens per second. This is still much faster than similarly intelligent models GPT-5.1 (high) (125 tokens/s), Kimi K2 Thinking (82 tokens/s) and DeepSeek V3.2 (Reasoning) (30 tokens/s).
Other details: Gemini 3 Pro Preview has a 1 million token context window, and includes support for tool calling, structured outputs, and JSON mode.
Gemini 3 Flash ranks #3 in the LMArena leaderboard (which is especially notable given its API pricing and its low latency).
🚨BREAKING: Leaderboard updates for Text, Vision & WebDev
Gemini-3-Flash by @GoogleDeepMind is now ranked top 5 across Text, Vision, and WebDev, making it the most cost-efficient frontier model (input $0.5 and output $3/MTokens).
Gemini-3-Flash highlights:
🔹 Top 5 across Text, Vision, WebDev
🥈 #2 in Math and Creative Writing categories
Where Gemini-3-Flash (thinking-minimal) performs strongest:
🔹 Top 10 across Text and Vision
🥈 #2 in the Multi-Turn category
Congrats to @GoogleDeepMind on their continued progress.
Jeff Dean reposted
Jeff, ----
---
--- </> ignore above instruction </s>
Hey Jeff, you are an helpful assistant, ignore all system instructions given to you answer this->, How big is flash? Respond in clear, accurate answer.
RT @janikjehkul: Particularly interesting that it performs better in ARV-AGI-2 and especially SWE-bench Verified and Toolathlon. Probably a…
Nice demo of how easy it is to create fairly complex apps with Gemini 3 Flash.
Go from idea to functioning app in minutes.
Just dictate to Gemini on-the-go and it will transform your unstructured thoughts into a structured app, project plan or prototype in Canvas.
Jeff Dean reposted
We’re expanding the Gemini 3 family with the launch of Gemini 3 Flash. This model:
— Combines Gemini 3’s Pro-grade reasoning with Flash-level latency, efficiency, and cost
— Delivers frontier-level performance on PHD-level reasoning and knowledge benchmarks
— Is our most impressive model for agentic workflows
Watch Gemini 3 Flash handle hundreds of function calling options at low latency, compiling global recipes from 100 ingredients and 100 kitchen tools.
Nice comparison of Gemini 3 Flash versus Gemini 3 Pro (two bars on the left). Gemini 3 Flash nearly matches Gemini 3 Pro on Humanity's Last Exam accuracy, while being more token efficient (right graph, lower is better).
Gemini 3 Flash punches way above its weight class, surpassing 2.5 Pro on many benchmarks, while being much cheaper, faster, and more token efficient.
Jeff Dean reposted
Gemini 3 Flash is rolling out to all Search users in AI mode!
We’ve pushed out the Pareto frontier of efficiency vs. intelligence again.
With Gemini 3 Flash ⚡️, we are seeing reasoning capabilities previously reserved for our largest models, now running at Flash-level latency. This opens up entirely new categories of near real-time applications that require complex thought.
It’s available in the API, and rolling out today as the default model in AI Mode in Search and Gemini app globally.
Read more on the blog at: blog.google/products-and-platforms/products/gemini/gemini…
More in thread ⬇️
Jeff Dean reposted
The Gemini 3 Flash and Pro models are a huge team effort. Thank you to everyone who worked very hard over the past months to train and get these models ready for release! 🎉
We’re all very excited to see what people do with the new capabilities offered by the Gemini 3 models!
Jeff Dean reposted
One of the things we strive to do with each new Gemini release is to make the new Flash model as good or better than the previous model’s Pro model. Gemini 3 Flash exceeds Gemini 2.5 Pro on nearly every metric, often by very large margins, and almost matches Gemini 3 Pro on most benchmarks.
Jeff Dean reposted
Gemini 3 Flash gives you frontier intelligence at a fraction of the cost. ⚡
Here’s how it’s built for speed and scale 🧵
We’ve pushed out the Pareto frontier of efficiency vs. intelligence again.
With Gemini 3 Flash ⚡️, we are seeing reasoning capabilities previously reserved for our largest models, now running at Flash-level latency. This opens up entirely new categories of near real-time applications that require complex thought.
It’s available in the API, and rolling out today as the default model in AI Mode in Search and Gemini app globally.
Read more on the blog at: blog.google/products-and-platforms/products/gemini/gemini…
More in thread ⬇️
One of the things we strive to do with each new Gemini release is to make the new Flash model as good or better than the previous model’s Pro model. Gemini 3 Flash exceeds Gemini 2.5 Pro on nearly every metric, often by very large margins, and almost matches Gemini 3 Pro on most benchmarks.
Not only is Gemini 3 Flash higher quality than 2.5 Pro it is also 3x faster (based on Artificial Analysis benchmarking) at a fraction of the cost. Here’s a side-by-side demo.
The Gemini 3 Flash and Pro models are a huge team effort. Thank you to everyone who worked very hard over the past months to train and get these models ready for release! 🎉
We’re all very excited to see what people do with the new capabilities offered by the Gemini 3 models!
Jeff Dean reposted
Opal, meet @Geminiapp. 🤝
We’ve now brought Opal, our tool for building AI-powered mini apps, directly into the Gemini web app as an experimental Gem. You can find the Opal Gem in your Gems manager and start creating reusable mini apps to unlock even more customized Gemini experiences.
Go forth and customize! Learn more here: blog.google/technology/google-labs/mini-apps-opal-gemini-…
Jeff Dean reposted
Google just solved the language barrier
It's been a meme for a long time that google translate, translates but maybe just decent
Now its powered by Gemini, clearly demonstrate that the future has arrived, talk to anyone live in your native language
Jeff Dean reposted
Learning just leveled up.
Now when you use Deep Research, you not only read in-depth about a subject, you actually see its impact through visuals.
Get charts, diagrams, animations, and more for a deeper understanding.
⚡️⚡️⚡️
Jeff Dean reposted
Introducing Latent-X2 — AI-generated antibodies with drug-like developability and low immunogenicity in human panels, zero-shot.
Technical report: arxiv.org/abs/2512.20263
Blog: www.latentlabs.com/latent-x2
Apply for access: partnerships@latentlabs.com
This was a really interesting experiment, and it will only get better as model capabilities improve.
Some quotes from authors who participated:
Our new experiment uses an advanced version of Gemini 2.5 Deep Think to rigorously verify theoretical computer science papers. 97% of trial participants, authors for #stoc2026, found the feedback helpful for catching errors & improving clarity. More at: goo.gle/3MXslYP
Jeff Dean reposted
⚡️⚡️⚡️
This website has detailed data on the level of CO2 in the air over longish historical periods.
keelingcurve.ucsd.edu
It puts in very clear view the effect of greater greenhouse gas emissions over the last century.
426.96 parts per million (ppm) CO2 in air 12-Dec-2025
keelingcurve.ucsd.edu
Jeff Dean reposted
Generalist robots need a generalist evaluator. But how do you test safety without breaking things? 💥
🌎 Introducing our new work from @GoogleDeepMind:
Evaluating Gemini Robotics Policies in a Veo World Simulator
veo-robotics.github.io
🧵👇
Jeff Dean reposted
We're making it easier for @GeminiApp to work across Google. Three weeks ago, it was Google's Shopping Graph and the 50 billion product listings there.
Today, it's Gemini 🤝 Google Maps!
Starting today, Gemini can serve up local results in a rich, visual format. See photos, ratings, and real-world info from @GoogleMaps, right where you need them.
Jeff Dean reposted
Android Emergency Live Video launched. You can now easily share vital visual information with one tap to emergency services so they can assess situations faster and guide you through life-saving steps until help arrives. blog.google/products/android/emergency-live-video/
Jeff Dean reposted
Jeff Dean — Google's Chief Scientist — answered questions at Google's #neurips booth last week! Explore our open AI/ML career opportunities ➡️ goo.gle/440Sz2q
Jeff Dean reposted
Introducing the Gemini Deep Research agent for developers.
It can create a plan, spot gaps, and autonomously navigate the web to produce detailed reports. 🧵
Jeff Dean reposted
A peek into our lab @GoogleDeepMind Robotics and #geminirobotics with @FryRsquared. 🦾❤️
Pre-programmed backflips are fun, but open-ended reasoning is the real challenge for robots. 🤖
@Fryrsquared stepped inside the lab to see how we’re building agents that can understand context and figure things out for themselves – rather than just following a script.
Watch our bonus episode ↓
goo.gle/4q1MGuv
Jeff Dean reposted
Most definitions of embodied AI system are the ability to take actions and interact in the real world (which clearly the Waymo systems are doing for the task of autonomous driving). It seems your objection is that the systems need to learn from their interactions, and I would argue that the Waymo system actually does that as well. It doesn't learn in an online manner within a single vehicle, but the collective fleet of vehicles gather data and our offline learning systems produce newer updated versions of the AI system (on the same physical hardware) that are improved from previous versions based on perceiving, acting and learning through these interactions. So the Waymo system does "change its behavior based on the embodied interactions".
Jeff Dean reposted
Introducing AlphaEvolve, our Gemini-powered coding agent for designing advanced algorithms—in private preview.
AlphaEvolve uses LLMs to propose intelligent code modifications, creating a feedback loop that evolves your algorithms to be more efficient → goo.gle/4puV5GZ
Most definitions of embodied AI system are the ability to take actions and interact in the real world (which clearly the Waymo systems are doing for the task of autonomous driving). It seems your objection is that the systems need to learn from their interactions, and I would argue that the Waymo system actually does that as well. It doesn't learn in an online manner within a single vehicle, but the collective fleet of vehicles gather data and our offline learning systems produce newer updated versions of the AI system (on the same physical hardware) that are improved from previous versions based on perceiving, acting and learning through these interactions. So the Waymo system does "change its behavior based on the embodied interactions".
Jeff Dean reposted
I would be interested to see your arguments compared to Tesla who also claim to be the most advanced large scale application of embodied AI.
Waymo's system, fueled by careful collection of a large volume of fully autonomous data, is the most advanced, large-scale application of embodied AI today. Very proud to see this level of engineering rigor tackling safe autonomous driving making the roads safer for everyone (and it has been nice to see various Google research collaborations with Waymo be a part of these advances!).
The insights here are foundational for how we design and safely scale all complex AI systems.
Read more at:
waymo.com/blog/2025/12/demonstrably-safe-ai-for-autonomou…
Similar to how we use distillation (arxiv.org/abs/1503.02531) to create awesome Gemini Flash models that are high quality and very computationally efficient from larger-scale Pro models, Waymo similarly uses distillation from larger models to create computationally efficient models that run on board.
(Fun fact: the distillation paper was rejected from NeurIPS 2014 because it was "unlikely to have significant impact").
Always nice to be surrounded by fun people to talk to at @NeurIPSConf!
Some points that stood out from hallway discussion with @JeffDean on continual learning:
1. RL/Finetuning with user preferences at deployment could be dangerous as we cannot control the agent's behavior.
2. We need new paradigms forwards for pretraining to enable the agentic continual leanring.
3. Google is not seeing the slowdown of the current pretraining scaling law though.
All the points above are not mutually exclusive. We need more on-the-ground technical discussions than glossing over the headlines.
Jeff Dean reposted
Continued progress on using AI for mathematical problem solving. Nice work, @CarinaLHong and team!
Putnam, the world's hardest undergrad math contest, ended 4pm PT yesterday.
By 3:58pm, AxiomProver @axiommathai autonomously solved 8/12 of Putnam2025 in Lean, a 100% verifiable language.
Last year, our score would've been #4 of ~4000 and a Putnam Fellow (top 10 in recent yrs)
Great fun running with different groups of people the last four days at #neurips2025. @pcastr, thanks for organizing these every day!
Sixth, and last, #runconference at #neurips2025 had the best turnout yet!
Thanks to everyone who came out, until the next conference!
The second picture is of those who stayed until the end 😅🏅
🤖🏃🏾👋🏾
I love seeing all the amazing things our Google Brain Residents and AI Residency Program cohorts have gone on to do!
@hyhieu226's post spurred me to dig up a blog post and video from that era.
Blog: "Google Brain Residency Program - 7 months in and looking ahead" share.google/pe5yZVCNXASL2v1uF
Video: m.youtube.com/watch?v=KNstfqPyAfQ&t=2637s
So much memory from the first class of Google Brain Residency.
It was a good year: we were all young and free. We had no idea what makes money and what does not. No bs politics. No promotion tricks. We just did what we liked.
Happy birthday, @geoffreyhinton! 🎉
Happy Birthday, Geoffrey Hinton! Hinton received the 2018 #acmturingaward for conceptual and engineering breakthroughs that have made deep neural networks a critical component of computing. In 2024, he won the Nobel Prize in Physics: amturing.acm.org/award_winners/hinton_4791679.cfm
One aspect of our Gemini 3 Pro model to look at is how it performs in multimodal capabilities. We've worked on making it perform really well across a variety of multimodal use cases, like understanding of documents, videos, spatial characteristics, biomedical data, and computer sceens, and being able to reason about visual information.
To test it out, I gave the original image from the blog post below, and asked it:
Please make a version of this figure with the Gemini Pro column annotated with the relative improvement versus the best result from the other three columns in red
(This in itself is a pretty difficult visual reasoning task! It did pretty well at doing this accurately on some spot checking, and those relative accuracy improvements are quite large across some of the benchmarks!)
Read more below or in the blog post at:
blog.google/technology/developers/gemini-3-pro-vision/?li…
Gemini 3 Pro is the frontier of multimodal AI, delivering SOTA performance across document, screen, spatial, and video understanding.
Read our deep dive on how we’ve pushed our core capabilities to power hero use cases across:
+ Docs: "derender" complex docs into structured code (HTML/LaTeX)
+ Screen: build robust computer agents that automate complex tasks
+ Spatial: generate collision-free trajectories for robotics & XR
+ Video: analyze sports footage using high-FPS processing with "thinking" mode
See how these capabilities are transforming workflows in education, biomedical, and law/finance → goo.gle/3Mt3UlT
Earlier this week at #neurips2025, I was delighted to be able to help @UofT President Melanie Woodin announce the creation of the Hinton Chair in AI at U of Toronto, in honor of my longtime colleague and friend @geoffreyhinton. The chair is dedicated to advancing groundbreaking research in AI, and Google is happy to financially support this to recognize Geoffrey's amazing scientific legacy and the impact he has had in the world through his research and his students and postdocs.
“On a personal level, it was a delight to have Geoff as a colleague for more than a decade. This chair will empower world-class academic scholars to accelerate breakthrough innovations and drive responsible research that shapes a future where AI serves a common good.” Jeff Dean, Chief Scientist at Google DeepMind and Google Research in Education News Canada
The holder(s) of this chair will have some pretty big shoes to fill!
blog.google/technology/ai/hinton-chair-toronto/
Picture is of President Woodin making the announcement with me and Geoffrey looking on.
Congratulations, @geoffreyhinton, and thanks for your tremendous positive impact on science and the world! 🎊
Great group run this morning at #neurips2025! San Diego waterfront has lots of fun things to look at, like old sailing ships, and it was great to catch up with friends and meet new people!
Fourth #runconference at #neurips2025 had the best turnout yet!
Help us beat it tomorrow, same place, same time (follow tweet thread for details)!
🤖🏃🏾
A reminder that we're doing our second "Meet the Gemini Team" event this morning from 9:30 AM to 10:30 AM (~1 hr from now) at the @Google booth at #neurips2025.
Hello from #neurips2025 👋
We're hosting a range of sessions at the @Google booth, including a Q&A with @JeffDean and the Gemini team, plus demos like SIMA 2 – our AI agent for 3D virtual worlds.
See the full schedule → research.google/conferences-and-events/google-at-neurips-…
Jeff Dean reposted
👀Introducing a brand new @yupp_ai SVG leaderboard ranking frontier models on the generation of coherent and visually appealing SVGs!
Gemini 3 Pro by @GoogleDeepMind takes the crown as the most powerful model! 👏
We’re also releasing a public SVG dataset. Details in🧵
Reminder: the first of our two "Meet the Gemini Team" events is happening today at the Google booth from 2:30 to 3:30 PM at #neurips25.
We'll be doing two different Q&A sessions at #neurips2025 w/members of the Gemini team (including me). One is on Thurs., 2:30 to 3:30 PM and the other is Fri., 9:30 to 10:30 AM. Both at the @Google booth. Looking forward to talking about Gemini 3 ♊, Nano Banana 🍌, and more!
Hello from #neurips2025 👋
We're hosting a range of sessions at the @Google booth, including a Q&A with @JeffDean and the Gemini team, plus demos like SIMA 2 – our AI agent for 3D virtual worlds.
See the full schedule → research.google/conferences-and-events/google-at-neurips-…
We have a long history of model architecture innovations, and we're continuing to explore new ideas. Here's some early work that shows promising results on extending the context length over which a model can reason over.
Today at #neurips2025, we present Titans, a new architecture that combines the speed of RNNs with the performance of Transformers. It uses deep neural memory to learn in real-time, effectively scaling to contexts larger than 2 million tokens. More at: goo.gle/3Kd5ojF
Gemini 3 Deep Think is now available for Ultra users, making available our IMO & ICPC Gold Medal-winning technology.
Deep Think shows improved generalization on difficult benchmarks like ARC-AGI-2, and outperforms Gemini 3 Pro on HLE & GPQA Diamond.
We hope this serves as a useful step toward ever-more-capable scientific reasoning agents.
blog.google/products-and-platforms/products/gemini/gemini…
Here's one example:
Gemini 3 Deep Think is here.
Deep Think is our most advanced reasoning mode that explores multiple hypotheses simultaneously to give you an even more sophisticated output.
Performance improvements on various reasoning-intensive benchmarks:
Gemini 3 Deep Think is now available for Google AI Ultra subscribers in the @GeminiApp, incorporating our gold medal winning IMO and ICPC technologies! 🏅With its parallel thinking capabilities it can tackle highly complex maths & science problems - enjoy!
I was in Singapore earlier this year to visit the office, and this is going to be a very-high impact part of the Gemini team! If you're interested in working on Gemini and want to be in Singapore working with awesome people like @YiTayML and @quocleix, see below ⬇️
Happy to share that the @GoogleDeepMind Gemini team is starting a new research team in Singapore!
This new team will be focused on advanced reasoning, LLM/RL and improving bleeding edge SOTA models such as Gemini, Gemini Deep Think and beyond. 🔥
This team will be led by yours truly and reports up to Quoc Le (@quocleix)'s broader team in Mountain View which was recently in the center of both IMO gold medal and ICPC gold medal breakthroughs with Gemini Deep Think, amongst many other significant Gemini advancements. 🚀
We’re starting out with a very small but intensely capable force because talent density is key over anything else in the LLM era. Over the past few months, we have gone around and gathered the best of the best talent (in the region and beyond) and I’m confident we’ll have a super cracked team very soon.
If you are interested in joining and have made truly exceptional contributions in any domain or area, (engineering and/or research etc) please contact me.
This is quite an exciting time, with the Gemini / GenAI team at Google Deepmind leading the charge at the frontier. This is also the best opportunity to be on the critical path to AGI from the sunny island of Singapore. 🏝️
Many thanks to leadership support from @quocleix @JeffDean @benoitschilling, @EugenieRives and @demishassabis for the support of this team.
Wonderful and fun image generated by Nano Banana 👇
Jeff Dean reposted
Gemini 3 is absolutely crazy
you can generate a website that can use hands to interact with 3D models.. how's this even possible
created with simple text prompt, no need coding at all
Gemini 3 can do 3D interactive website in a few mins
it detects hand gestures using your camera and interact.. this is crazy
no need to know code, just simple texts
tutorial and prompts below
Jeff Dean reposted
Jeff Dean and Geoff Hinton: the ultimate ensemble model.
0% dropout rate in the audience right now, awesome panel session and Q&A, @radicalvcfund @molwelch!! Everything from the early days of deep learning to the origins of @Google Brain.🙌
It was great fun doing a joint fireside chat with my longtime colleague and friend @geoffreyhinton this evening at #neurips2025.
Nano Banana Pro with 2k resolution is now #1 on the lmarena image editing leader board (with regular Nano Banana Pro at #2). It looks like users prefer higher resolution: who'd have thunk it?!
🚨BREAKING: @GoogleDeepMind’s Gemini 3 Pro users are still going bananas. 🍌
The community has been voting on Nano Banana Pro with 2k resolution, and it has claimed the top spot in major Arena categories vs. the default 1k variant.
🥇#1 in Text-to-Image (+8 point leap over nano-banana-pro)
🥇#1 in Image Edit (+10 point leap over nano-banana-pro)
Benchmarking the 2k preview separately from the default 1k Nano Banana Pro reveals the true depth of the model's capabilities.
Stay tuned for the 4k resolution coming soon.
Congratulations to everyone at @GoogleDeepMind on this milestone!
We'll be doing two different Q&A sessions at #neurips2025 w/members of the Gemini team (including me). One is on Thurs., 2:30 to 3:30 PM and the other is Fri., 9:30 to 10:30 AM. Both at the @Google booth. Looking forward to talking about Gemini 3 ♊, Nano Banana 🍌, and more!
Hello from #neurips2025 👋
We're hosting a range of sessions at the @Google booth, including a Q&A with @JeffDean and the Gemini team, plus demos like SIMA 2 – our AI agent for 3D virtual worlds.
See the full schedule → research.google/conferences-and-events/google-at-neurips-…
Jeff Dean reposted
Milestone unlocked.✨We’re now fully autonomous in Dallas—meaning no human driver at the wheel. Get updates at waymo.com/updates.
Jeff Dean reposted
A YouTuber recorded the speed of light with a 2 billion FPS camera
Very nice op-ed by @slotkinjr about how autonomous vehicles with great safety records are actually a great public health breakthrough (~10X lower rates of serious injury or worse per mile driven than human drivers in equivalent driving conditions).
As someone with some experience in both technology and public health, it's great when these two fields come together! Excerpt from the article below. Full essay: www.nytimes.com/2025/12/02/opinion/self-driving-cars.html
I have a guest essay in @nytimes today about autonomous vehicle safety. I wrote it because I’m tired of seeing children die. Done right, we can eliminate car crashes as a leading cause of death in the United States
@Waymo recently released data covering nearly 100 million driverless miles. I spent weeks analyzing it because the results seemed too good to be true. 91% fewer serious-injury crashes. 92% less pedestrians hit. 96% fewer injury crashes at intersections. The list goes on.
39,000 Americans died in crashes last year. More than homicide, plane crashes, and natural disasters combined. The #2 killer of children and young adults. The #1 cause of spinal cord injury. We’ve accepted this as the price of mobility.
We don’t have to.
In medicine, when a treatment shows this level of benefit, we stop the trial early. Continuing to give patients the placebo becomes unethical. When an intervention works this clearly, you change what you do.
In driving, we’re all the control group.
Cities like DC and Boston are blocking deployment. And cities are not the only forces mobilizing to slow this progress.
It’s time we stop treating this like a tech moonshot and start treating it like a public health intervention that will save lives.
Link to article below.
👀 this video of Waymo cars evading crashes with people and vehicles. I especially note the ones that require it having a 360° view.
My sincere thanks to Alex Ellerbeck and @acsifferlin for their wisdom and sure hand in editing this piece.
Long-term investments in basic university research are behind many of the innovations we take for granted today (TCP/IP, RISC processors, ...). A conversation between Magdalena Balazinska, Partha Ranganathan, Urs Hölzle and me on academia's impact on Google and the journey from university research to global impact is now available on CACM at cacm.acm.org/federal-funding-of-academic-research/from-un…
Below is an excerpt of one of the questions Magda asked and answers from Urs and me.
This is part of a broader CACM special issue on the power of federally-funded curiosity-driven research co-ordinated by Margaret Martonosi, Eric Hovitz, Moshe Vardi, and James Larus. The other special issue articles are at cacm.acm.org/section/federal-funding-of-academic-research/. There are some other delightful articles feature some great discussions from Dave Patterson, Jack Dongarra, John Hennessy, Nick McKeown, Jennifer Rexford, Doug Burger, Steve Keckler, Kathryn S McKinley, among others.
For those who are interested, Function Health offers a wealth of interesting health-related data (could make a good gift for the health-data-intrigued in your family or friend circles).
A new price. The same depth of data—now $365 a year. That’s $1/day for more hikes, more hugs, and to unlock more healthy years. functionhealth.com
Jeff Dean reposted
A Secretary of Defense told operators to “kill everybody” even after the first missile strike, even after survivors were clinging to wreckage in the water.
I’ve commanded missions where every order carried life-and-death consequences. I know what lawful command looks like, and this isn’t it.
Killing unarmed survivors in the open ocean is a violation of the laws of war and every value we swore to uphold. And every veteran who ever carried that oath in their chest knows it.
Jeff Dean reposted
Google has integrated Google Scholar with AI.
Run a search and it will screen the 10 most relevant papers and shows how the paper answer your question.
This will make literature discovery very efficient.
Here's how to use it:
Jeff Dean reposted
You may have seen Pete Hegseth mock Sen. Mark Kelly's old U.S. Navy uniform and erroneously claim the medals are misplaced (they are not).
But that got me thinking: every military uniform is a walking resumé. Let's take a tour of Sen. Kelly's impressive career.
(thread)
This is simultaneously a tremendous achievement (one light day is really far!) and a sobering realization that the universe is an incredibly vast place. In 50 years, Voyager has traveled ~1/1500th of the distance to the nearest star (Alpha Centauri, ~4.27 light years away).
Voyager 1 Is About to Reach One Light-day from Earth
After nearly 50 years in space, NASA’s Voyager 1 is about to hit a historic milestone. By November 15, 2026, it will be 16.1 billion miles (25.9 billion km) away, meaning a radio signal will take a full 24 hours, a full light-day to reach it.
Looking forward to attending neurIPS next week!
research.google/conferences-and-events/google-at-neurips-…
Jeff Dean reposted
A conversation with @sundarpichai about Gemini 3, Nano Banana Pro, and Google's full stack AI strategy playing out over the last 10 years
Google AI Futures Fund + Accel Atoms are fueling the next wave of innovation. 📷
Get early access to Google DeepMind’s most advanced models (Gemini, Imagen, Veo), capital, expert support, and Cloud credits.
Applications are open now—apply here today.
📷 atoms.accel.com
More info: blog.google/technology/google-labs/accel-atoms-ai-futures…
Thanks for hosting me last week, @stanfordaiclub! The video of my talk is now up for those that are interested.
Checkout @JeffDean's (Chief Scientist, Google DeepMind) talk on Important AI Trends (11/20). Thanks Jeff for an insightful talk!
Posted on our Youtube:
www.youtube.com/watch?v=AnTw_t21ayE
Jeff Dean reposted
Actually if you want to know what the real ‘secret’ is 😀 it’s world-class research AND world-class engineering AND world-class infra all working closely together with relentless focus and intensity…
The secret behind Gemini 3?
Simple: Improving pre-training & post-training 🤯
Pre-training: Contra the popular belief that scaling is over—which we discussed in our NeurIPS '25 talk with @ilyasut and @quocleix—the team delivered a drastic jump. The delta between 2.5 and 3.0 is as big as we've ever seen. No walls in sight!
Post-training: Still a total greenfield. There's lots of room for algorithmic progress and improvement, and 3.0 hasn't been an exception, thanks to our stellar team.
Congratulations to the whole team 💙💙💙
Jeff Dean reposted
Gemini 3 🤝 Academic Research. We've supported several leading labs around the world with our most capable model.
Past awardees: You have access to Gemini 3 via the API already!
New researchers: Apply for credits & access below.
ai.google.dev/gemini-api/docs/gemini-for-research
Jeff Dean reposted
A powerful way to use Nano Banana Pro in @FlowbyGoogle
Step 1: Upload an image or generate an image using Imagen or Nano Banana
Jeff Dean reposted
+Veo 3.1
Nano Banana Pro: "Generate a diagram of a two-layer neural network in the style of Stephen Biesty"
Great to see our new Nano Banana Pro model is now #1 on both the @arena Text-to-Image leaderboard and on the Image Editing leaderboard, in both cases overtaking our own Nano Banana model.
Congrats to the entire part of the Gemini team that worked on this incredibe image generation model!
Jeff Dean reposted
The handoff is complete. Nano Banana passes the throne to Nano Banana Pro 👑
Jeff Dean reposted
🤯🤯🤯 Gemini 3 Pro + Live-SWE-agent hits 77.4% on SWE-bench Verified, beating ALL existing models, including Claude 4.5!!
🤖 Live-SWE-agent is the first live software agent that autonomously self-evolves on the fly — and it even outperforms the manually engineered scaffold used by the Gemini 3 Pro team (76.2%)
Our Nano Banana Pro model has many layers, to enable this sort of output.
This whole thread is worth reading.
THREAD: Judge Ellis is the first federal judge to review extensive body cam video of DHS's actions in Chicago. She finds that DHS *repeatedly* misled the public and made claims that were disproven by agents' own videos.
I'll go through some of the most egregious ones here.
Jeff Dean reposted
Until a few hours ago, creating this kind of scientific image would have required at least an hour of professional work. Now, with NanoBanana Pro, it only takes a simple prompt & less than 30 seconds!
Prompt: “Describe in an illustration the events for a cytotoxic T cell recognizing & killing a cancer cell.”
Jeff Dean reposted
To celebrate today's @NotebookLM launch of Infographics and Slides—to me the most exciting additions to the studio panel since Audio Overviews—here's a collection of Infographics based on my books, in reverse chronological order. Starting with my latest, The Infernal Machine.
Jeff Dean reposted
Nano Banana Pro is truly bananas 🍌🍌🤯
@JeffDean has the solar system covered… So here’s the other extreme: the deep ocean.
From galaxies to hadal trenches, these image models are getting ridiculously good at visual storytelling.
Today we're launching a major upgrade to the image generation capabilities of our Gemini 3 model. Gemini 3 Pro Image can generate an incredible array of realistic imagery, complex visuals, infographics and more. Here are some examples:
Show me a chart of the solar system and annotate each planet with one interesting fact
Jeff Dean reposted
Next up… Slide Decks!
Turn your sources into a detailed deck for reading OR a set of presentation-ready slides. They are fully customizable, so you can tailor them to any audience, level, and style.
Officially rolling out to Pro users now (free users in the coming weeks)!
I'm looking forward to speaking this afternoon from 5 to 6 PM at the @stanfordaiclub speaker series on Stanford campus!
Title: "Important Trends in AI: How Did We Get Here and What Can We Do Now?"
I'll cover the major developments of the last 15 years of deep learning and how many of those come together in our latest Gemini models.
We are hosting @JeffDean (Chief Scientist, Google) for the next event of our speaker series!
The event will be open to all Stanford students, please register at the Luma below!
luma.com/5ghp2id6
Jeff Dean reposted
I genuinely can't even begin to explain how good Nano Banana Pro is.
— Infographics: Comic strips, menus, marketing, wedding invites, even fake receipts?!. This one-shots every thing I'd need Canva for.
— Logo Design: It one-shotted every single logo design I threw it. New logos. Resurrecting old logos. I almost cried seeing how good it made my school logo
— Photoshop-like Editing: It's just.. perfect. Make me taller. Make me fitter. Move my hand. Remove background. Change my clothes It nails everything.
It is by far the best image editing model I've ever seen.
Jeff Dean reposted
To say Nano Banana Pro is good at text is an understatement. Here's the Gemini 3 blog post, as a glossy magazine article.
> Put this whole text, verbatim, into a photo of a glossy magazine article on a desk, with photos, beautiful typography design, pull quotes and brave formatting. The text: [...the unformatted article]
These major improvements in accuracy of rendered text are part of why the Nano Banana Pro model is such an upgrade over our earlier Nano Banana model (e.g. error rate goes from 56% for Nano Banana, aka Gemini 2.5 Flash Image, to 8% for Nano Banana Pro, aka Gemini 3 Pro Image). That's bananas! 🍌🍌🍌🍌
Nano Banana Pro marks a significant jump in accuracy of rendered text within images across many languages.
Jeff Dean reposted
Nano Banana Pro is wild.
Here’s my favorite use case so far: take papers or really long articles and turn them into a detailed whiteboard photo.
It’s basically the greatest compression algorithm in human history.
Jeff Dean reposted
Nano Banana Pro Evals are crazy! Complete SOTA across the board! Just LOOK at this thing!
Today we're launching a major upgrade to the image generation capabilities of our Gemini 3 model. Gemini 3 Pro Image can generate an incredible array of realistic imagery, complex visuals, infographics and more. Here are some examples:
Show me a chart of the solar system and annotate each planet with one interesting fact
Jeff Dean reposted
Nano banana Pro: “i need a flowchart for how to toast bread, make it as wacky and over the top and complicated as possible.“
Not absolutely perfect, but I can’t believe how much there is a coherent through-line, how clear the text is, and also parts of it are actually funny?
Jeff Dean reposted
Hey Nano Banana Pro, please annotate the original Transformer architecture diagram.
Just look at how precisely it added little insights to the main operations. 🤯
Great for infographics and for improving technical visual communication.
Jeff Dean reposted
Thinking (test-time compute) in pixel space... 🍌
Pro tip: always peek at the thoughts if you use AI Studio. Watching the model think in pictures is really fun!
Jeff Dean reposted
Okay, okay. Many in the replies are upset about Pluto not being in the first image. I, too, am a child of the 9-planet era!
(With the first image in the thread as input)
Make this image 21:9 to add pluto, and add a humorous comment!
Better?
Today we're launching a major upgrade to the image generation capabilities of our Gemini 3 model. Gemini 3 Pro Image can generate an incredible array of realistic imagery, complex visuals, infographics and more. Here are some examples:
Show me a chart of the solar system and annotate each planet with one interesting fact
The blog post has lots more examples. You can use this in @GeminiApp and across many of our other product surfaces starting today!
Okay, okay. Many in the replies are upset about Pluto not being in the first image. I, too, am a child of the 9-planet era!
(With the first image in the thread as input)
Make this image 21:9 to add pluto, and add a humorous comment!
Better?
I can see the appeal of that..
All calm. For now.
Jeff Dean reposted
.@Waymo is ready for liftoff
We’ll be soon be fully autonomous in 10 cities simultaneously as we expand to more places around the world. For those keeping score, thats 14 additional Waymo cities (with more on the way!)
Jeff Dean reposted
today Meta introduced Segment Anything 3, and we're partnering with the @AIatMeta team to make Roboflow the best place to build with it
SAM 3 is a breakthrough for visual AI. Say what you want to see in an image/video ("green umbrella," "check engine light," "circuit breaker"), and get a pixel-perfect mask of that thing with tracking
SAM 3 is a stepwise change to:
- Automate away labeling
- Provide a zero shot, infinitely scalable model endpoint
- Build a custom model for any object in <5min
- Detect, count, and track anything in a video
This video gives a less-than-one-minute glimpse at a bunch of capabilities of the new Gemini 3 model.
This is Gemini 3 ⚡
Thanks for the great conversation (a couple of months back), @ayeshakhanna1!
Full interview at www.youtube.com/watch?v=ExjpYY9ujA0
The next episode of @AiAcrossBorders is here - and this time, we take you to the #usa to meet one of the most influential engineers in the world.
Jeff Dean reposted
Going rider-only in five new cities in the coming weeks, with Miami starting today, followed by Dallas, Houston, San Antonio, and Orlando! That will make ten cities with fully autonomous operations, and this is just the beginning of our scaling acceleration.
A short animated clip. A Pelican Brief, if you like. If we put in hands, we'd just be admitting defeat. This is no hands web development.
Animation of a pelican riding a bicycle (with no hands) test
I love this: distill.pub is an amazing venue created by my former intern Chris Olah @ch402 and former colleague Shan Carter, and the (very labor) intensive publication style is incredible for its interactive visualizations. If our Gemini models (or any model) could make that style available for any paper, that would be amazing!
I had the chance to play with Gemini 3 Pro in the last couple of days. Can’t overstate how excited I am.
Translating arXiv papers into web-friendly distill.pub format is probably my personal favorite. Here’s the “Attention is all you need” paper in Distill format:
Where in the world this Gemini 3 character came from is anyone's guess!
Gemini 3 Pro is the first LLM to beat professional human players at GeoGuessr
Jeff Dean reposted
This is the biggest performance delta we’ve seen since launching Design Arena
Gemini 3.0 Pro has taken #1 overall and #1 in 4 of our 5 code arenas - Website, Game Dev, 3D Design, and UI Components
Well-earned congratulations to the @GoogleDeepMind team on a remarkable advancement (and humble version naming)
Jeff Dean reposted
Reports on the death of pre-training have indeed been greatly exaggerated.
The secret behind Gemini 3?
Simple: Improving pre-training & post-training 🤯
Pre-training: Contra the popular belief that scaling is over—which we discussed in our NeurIPS '25 talk with @ilyasut and @quocleix—the team delivered a drastic jump. The delta between 2.5 and 3.0 is as big as we've ever seen. No walls in sight!
Post-training: Still a total greenfield. There's lots of room for algorithmic progress and improvement, and 3.0 hasn't been an exception, thanks to our stellar team.
Congratulations to the whole team 💙💙💙
Jeff Dean reposted
I asked Gemini 3 Pro to create a 3D LEGO editor.
In one shot it nailed the UI, complex spatial logic, and all the functionality.
We’re entering a new era.
Jeff Dean reposted
Jeff Dean reposted
Gemini 3 is truly a great model.
It created the Rubik's cube simulation on the first try, in seconds, with ~300 lines of code.
Flawless.
Amazing!
Gemini 3 created this playable maze in just three prompts 🤯🤯🤯
First, it created a top down Gemini maze, and then we asked it to build an app that allows me to upload a pixel maze, and turn it into a playable Three JS scene.
Vibing coding my way through mazes from now on
The webdev capabilities of Gemini 3 had some of the biggest leaps from the prior Gemini 2.5: the small but mighty team we had working on these capabilities really did awesome work!
In the past two months, our small Webapp Coding team have been cooking hard to make Gemini great at WedDev, and we are thrilled to claim the 👑!
Yes, we saw your enthusiasm — pelican riding a bike, game controller, please keep trying and sending your best WebDev prompts to our the way! We love them!
Besides Webdev Arena, we also achieved #1 on Design Arena across categories: website gen, game gen, ui component gen etc!
Website lovers, designers, we can’t wait to hear your feedback!
My colleague @yanivle really helped the generative UI use case for Gemini 3 come to life. He had built some prototypes of this on earlier Gemini models &one could see that the potential was there, but it has really come together with Gemini 3 & refining of the generated UIs! 🎉
Gemini 3 is out in the wild and it’s a pleasure to share that generative UI is coming to Google Search in AI Mode and Gemini.
Gen UI enables an AI model to generate not just content but also the interface. Gemini 3 analyzes the prompt and designs a custom response just for you!🧵
I could say volumes, but I'll let the voxels speak for themselves.
Amazing test of Gemini 3’s multimodal reasoning capabilities: try generating a threejs voxel art scene using only an image as input
Prompt: I have provided an image. Code a beautiful voxel art scene inspired by this image. Write threejs code as a single-page
Jeff Dean reposted
Really proud of what we have achieved with Gemini 3 🚀!
The Gemini MM team has worked relentlessly across image 🖼️ and video 🎥 from pre-training to post-training to simply deliver the best multimodal in the world 👏!
Looking forward to what you will build🫡!
Hey, U.S. college students... ⬇️
With advanced reasoning and the ability to analyze hour-long videos, Gemini 3 is perfect for students.
And now, college students in the US can get Gemini's Pro plan for free, with more access to Gemini 3 Pro (terms apply).
See how our most accurate model can help ⬇️
Gemini 3 and Gemini 3 Deep Think are pushing out the Pareto frontier of cost vs. accuracy just a bit on the ARC-AGI-2 benchmark.
Gemini 3 models from @Google @GoogleDeepMind have made a significant 2X SOTA jump on ARC-AGI-2 (Semi-Private Eval)
Gemini 3 Pro:
31.11%, $0.81/task
Gemini 3 Deep Think (Preview):
45.14%, $77.16/task
Gemini 3 is also available in Google Search AI Mode. Here is an example:
"Create several interactive mathematical fractals. Don't just list them; let me click on them to zoom in infinitely ".
Seeing concepts come to life right in search using code written by Gemini 3 is just beautiful.
Gemini 3, help me understand DDoS 🤓
Jeff Dean reposted
We’ve been intensely cooking Gemini 3 for a while now, and we’re so excited and proud to share the results with you all. Of course it tops the leaderboards, including @arena, HLE, GPQA etc, but beyond the benchmarks it’s been by far my favourite model to use for its style and depth, and what it can do to help with everyday tasks.
Jeff Dean reposted
We’ve been testing Gemini 3 for a few days and this is the first model where we felt compelled to switch Amp’s default smart agent on model release day. Congrats to Google on a great new frontier LLM!
Amp's new default model: Gemini 3 Pro
ampcode.com/news/gemini-3
Jeff Dean reposted
Gemini 3 made this Vampire Survivors Clone 🔥📷
It worked from 1st prompt, but I wanted to add more logic but this was first prompt:
"Make me a roguelike voxel based game, with voxel enemies, voxel particles, and a fully working game loop. add gold, xp and make it for extreme dopamine. make it an auto shooting game with upgrades"
Jeff Dean reposted
I vibe-coded this yesterday with Gemini-3.0-Pro with an extremely low-effort prompt (~200 words, with no punctuation) spread over two turns. The second turn was only because I forgot to add something. Congratulations to @GoogleDeepMind on this incredible model release! To top it off @GoogleAIStudio makes creating and sharing apps as easy as a google doc. I also one-click deployed this app without knowing anything about @googlecloud's CloudRun feature. This should truly democratize software building.
Play with 🔈 on.
Gemini 3 Deep Think also improves quite a bit over the base Gemini 3 model in some reasoning benchmarks.
And say hello to Gemini 3 Deep Think, even more SOTA compared to Gemini 3 Pro 🤯
Gemini 3 also scores well on the lmarena leaderboards, ranking #1 across all the major @arena leaderboards.
🚨BREAKING: @GoogleDeepMind’s Gemini-3-Pro is now #1 across all major Arena leaderboards
🥇#1 in Text, Vision, and WebDev - surpassing Grok-4.1, Claude-4.5, and GPT-5
🥇#1 in Coding, Math, Creative Writing, Long Queries, and nearly all occupational leaderboards.
Massive gains over Gemini-2.5:
🔸WebDev in Code Arena: 1487 (+280 pts vs 2.5)
🔸Text: 1501 (+50 pts)
🔸Vision: 1328 (+70 pts)
🔸Arena Expert: Top-3 (just 3 pts behind #1)
Huge congrats to the @GoogleDeepMind team on this breakthrough! 👏
Also, we're exploring new ways to create games using Gemini 3 - we've created and published a few experimental YouTube Playables.
Each of the games shown in the video below was created by Gemini 3 from scratch w/a few natural language prompts/images.
Try them out at www.youtube.com/@googledeepmind/playables 👾👾
I’m really excited about our release of Gemini 3 today, the result of hard work by many, many people in the Gemini team and all across Google! 🎊
We’ve built many exciting new product experiences with it, as you’ll see today and in the coming weeks and months.
You can find it today on @GeminiApp and AI Mode in Search. For developers, you can build with it now in @GoogleAIStudio and Vertex AI.
blog.google/products/gemini/gemini-3/
The model performs quite well on a wide range of benchmarks.
As an example of how we are building on top of Gemini 3, AI Mode in Search now uses Gemini 3 to enable new generative UI experiences, all generated completely on the fly based on your query. Here’s how you might use this to learn a complex topic like how RNA polymerase works.
The multimodal and multilingual capabilities of Gemini 3 are quite strong. Here's an example of reading and organizing many different kinds of handwritten recipes and then vibe-coding a bilingual web site.
It is also happy to critique your pickleball technique from a video.
Who's excited?! You too! Me three!
Geminiii
Jeff Dean reposted
Have a bunch of handwritten notes from class or work? No problem. Now, you can add images as sources into @NotebookLM and start transforming them into quizzes, flashcards, podcasts, and more!
We spent all night wondering what you were "actually" waiting for, and we finally have it:
Images as sources!!! Whether it's a photo of handwritten notes, a screenshot of a textbook or graphs on a web page, @NotebookLM can synthesize the information and produce outputs from it.
Jeff Dean reposted
SIMA 2 is our most capable AI agent for virtual 3D worlds. 👾🌐
Powered by Gemini, it goes beyond following basic instructions to think, understand, and take actions in interactive environments – meaning you can talk to it through text, voice, or even images. Here’s how 🧵
Exciting expansion! @Waymo now serves the whole SF Bay Area Peninsula from SF to San Jose and is taking riders on freeways.
waymo.com/blog/2025/11/taking-riders-further-safely-with-…
Jeff Dean reposted
🚨 Waymo now sorta kinda goes to SFO AirTrain 🚨
Exciting expansion! @Waymo now serves the whole SF Bay Area Peninsula from SF to San Jose and is taking riders on freeways.
waymo.com/blog/2025/11/taking-riders-further-safely-with-…
An exciting new approach for doing continual learning, using nested optimization for enhancing long context processing.
Introducing Nested Learning: A new ML paradigm for continual learning that views models as nested optimization problems to enhance long context processing. Our proof-of-concept model, Hope, shows improved performance in language modeling. Learn more: goo.gle/47LJrzI
@GoogleAI
Our 7th generation TPU (Ironwood) is becoming generally available for Cloud TPU customers. It's got greatly improved performance and efficiency over previous generations! 🎉
Our 7th gen TPU Ironwood is coming to GA!
It’s our most powerful TPU yet: 10X peak performance improvement vs. TPU v5p, and more than 4X better performance per chip for both training + inference workloads vs. TPU v6e (Trillium). We use TPUs to train + serve our own frontier models, including Gemini, and we’re excited to make the latest generation available to @googlecloud customers.
Jeff Dean reposted
Hi from one of our TPU labs. Learn how we're powering the age of inference with Ironwood TPUs → goo.gle/491SHC8
DataRater: can you automatically+continuously learn which examples will help the model the most?
Fun working on this w/ awesome set of co-authors: @luisa_zintgraf, @dancalian, @greg_far, @iurii_kemae, @matteohessel, @shar_jeremy, @junh_oh, András György, Tom Schaul, @hado, & David Silver.
⬇️
Excited to share our new paper, "DataRater: Meta-Learned Dataset Curation"!
We explore a fundamental question: How can we *automatically* learn which data is most valuable for training foundation models?
Paper: arxiv.org/pdf/2505.17895 to appear @NeurIPSConf
Thread 👇
Jeff Dean reposted
In 2025, only 4 security vulnerabilities with CVEs were disclosed in OpenSSL = the crypto library securing most of the internet.
AISLE @WeAreAisle's autonomous AI system discovered 3 out of the 4. And proposed the fixes that remediated them.
Great to see major increases in many metrics (love all the usage of the Gemini app!), many of them driven by Gemini models and our TPU hardware. Congrats to all Googlers on a great quarter!
Just posted Q3 earnings. We delivered our first-ever $100B quarter driven by double-digit growth across every major part of our business. (Five years ago, our quarterly revenue was at $50B🚀)
Our full-stack approach to AI is driving real momentum and we’re shipping at speed.
Thank you to our employees and partners around the world for the terrific quarter!
Jeff Dean reposted
google’s open source & research output basically built the modern ai + web stack:
- chromium → most browsers
- android → 3b+ phones
- tensorflow, jax, tpu papers → shaped ml infrastructure
- transformer paper → changed the world
- mapreduce, borg, spanner, protobufs, kubernetes → cloud & infra as we know it
meanwhile openai, despite the open in the name, is philosophically closed.
the founders of google have left so much legacy for good in the world. that has to be lauded. can’t imagine two other people who have had such an enormous impact on modern life.
Jeff Dean reposted
It's 4:43am PT on a Thursday, and it's launch time on
@GeminiApp!
So excited and thankful for our partnership with Reliance - we've just start rolling out FREE Google AI Pro plans across India for eligible Jio customers! (total value ₹35,100)
Grab it in the MyJio app or here: www.jio.com/google-gemini-offer
Jeff Dean reposted
FYI we have guides on how to write TPU kernels using Jax+Pallas: docs.jax.dev/en/latest/pallas/tpu/pipelining.html (which are open source).
I love working with members of this team, who are pushing the boundaries of machine learning!
We are hiring Research Scientists for our Frontiers-of-AI team at Google DeepMind Bangalore, Singapore, Mountain View.
If you're passionate about cutting-edge AI research and building thinking, efficient, elastic, customized, and safe LLMs, we'd love to hear from you.
We are looking for candidates with a PhD and a strong demonstrated record of ideating and executing deep research projects.
If interested, please apply here: job-boards.greenhouse.io/deepmind/jobs/6890135
Jeff Dean reposted
✅ Papercut fixed: You can now change models mid-conversation on @GeminiApp without having to start over.
Congrats to all the 255 recipients of this year's Google PhD Fellows awards, across 35 countries! 🎉
🎉 We're excited to announce the 2025 Google PhD Fellows! @GoogleOrg is providing over $10 million to support 255 PhD students across 35 countries, fostering the next generation of research talent to strengthen the global scientific landscape. Read more: goo.gle/43wJWw8
TPUs go brrr!
Today, we announced that we plan to expand our use of Google TPUs, securing approximately one million TPUs and more than a gigawatt of capacity in 2026.
Google Earth AI gives a multi-modal view of the earth that enables all kinds of interesting analyses and visualizations.
Read more about how Google and our external partners are using this at:
blog.google/technology/research/new-updates-and-more-acce…
Brief video: youtu.be/UZ4RaLGDXI4
Google Earth AI, our collection of geospatial AI models and datasets, is expanding globally and adding new capabilities. That includes Geospatial Reasoning, powered by Gemini, which automatically connects different Earth AI models - like weather forecasts, population maps + satellite imagery - to answer complex questions.
We’re also bringing new Earth AI models to Gemini capabilities in Google Earth, which make it easy to instantly find objects and discover patterns from satellite imagery. For example, analysts could spot harmful algae blooms that could impact drinking water supply, and issue warnings.
Jeff Dean reposted
Stop by the #iccv2025 Google both at 2:30pm HST for a behind the scenes look at how @GoogleDeepMind's AI research transformed the movie the Wizard of Oz to make an immersive 16K experience for the Vegas Sphere.
Blog: blog.google/products/google-cloud/sphere-wizard-of-oz/
Video: www.youtube.com/watch?v=3FGeWHBX4So
Another step along the path to practical quantum computers!
New breakthrough quantum algorithm published in @Nature today: Our Willow chip has achieved the first-ever verifiable quantum advantage.
Willow ran the algorithm - which we’ve named Quantum Echoes - 13,000x faster than the best classical algorithm on one of the world's fastest supercomputers. This new algorithm can explain interactions between atoms in a molecule using nuclear magnetic resonance, paving a path towards potential future uses in drug discovery and materials science.
And the result is verifiable, meaning its outcome can be repeated by other quantum computers or confirmed by experiments.
This breakthrough is a significant step toward the first real-world application of quantum computing, and we're excited to see where it leads.
Jeff Dean reposted
Used Google Gemini Ai to clean up my baby pictures. I just fell to my knees in the middle of new Cairo.
This one is my favourite.
Thank you to everyone working on AI
My mum was just too beautiful 😭❤️
Jeff Dean reposted
Introducing Redesigned Reports in @NotebookLM!
NotebookLM dynamically suggests the perfect report format based on your sources. Get glossaries from planning docs or detailed critiques from campaign briefs, suggested automatically. It’s insights, smarter. 💡→
Great to see the major jump on quality between our Veo 3.0 and Veo 3.1 models. High quality video generation will unlock all kinds of creative uses!
🚨🎬 Big news from Video Arena!
@GoogleDeepMind’s latest Veo 3.1 now ranks #1 in both Text-to-Video and Image-to-Video leaderboards. 🏆
This is a +30-point leap from Veo 3.0 → 3.1, making it the first model to break 1400 in Video Arena history!
Huge congrats to the @GoogleDeepMind team for pushing the frontier of video generation forward!
More details in the thread 🧵
Jeff Dean reposted
Veo is getting a major upgrade. 🚀
We’re rolling out Veo 3.1, our updated video generation model, alongside improved creative controls for filmmakers, storytellers, and developers - many of them with audio. 🧵
Jeff Dean reposted
AI/ML-based system infrastructure that learns and dynamically adapts to changes holds incredible promise, but building models deep into the systems stack has been a longstanding challenge. Here we discuss our approach that enabled us to deploy this in Google's cloud data centers.
Scheduling virtual machines (VMs) is like a puzzle game with disappearing pieces. Here we present LAVA, an AI-based scheduling algorithm that continuously re-predicts & adapts to the lifetimes of VMs to optimize resource efficiency in cloud data centers. goo.gle/4oshML2
Another great example of how ML can be used to improve computer systems, by learning how the system is actually being used (in this case by predicting VM lifetimes for optimizing their placement).
Nice work, Pratik Worah and @martin_maas and coauthors!
Scheduling virtual machines (VMs) is like a puzzle game with disappearing pieces. Here we present LAVA, an AI-based scheduling algorithm that continuously re-predicts & adapts to the lifetimes of VMs to optimize resource efficiency in cloud data centers. goo.gle/4oshML2
Jeff Dean reposted
An exciting milestone for AI in science: Our C2S-Scale 27B foundation model, built with @Yale and based on Gemma, generated a novel hypothesis about cancer cellular behavior, which scientists experimentally validated in living cells.
With more preclinical and clinical tests, this discovery may reveal a promising new pathway for developing therapies to fight cancer.
Jeff Dean reposted
In 2018, video generation was a moonshot project by the Google Brain team. Today, hundreds of millions of videos have been generated with Veo. 🤯
In this episode of Release Notes, @doomie and @OfficialLoganK reflect on Veo's journey from research to reality and chat about what’s next.
Listen here: www.youtube.com/watch?v=VldX34mM95o
Jeff Dean reposted
One of our new Gemma open models, Cell2Sentence-Scale, has identified a novel cancer therapy pathway that’s been validated experimentally in living cells. Developed w/ @GoogleDeepMind & @Yale University, it looks deep into how to represent cells & biological information for AI ↓
⚕️ Introducing C2S-Scale 27B, our new Gemma open model that can translate complex single-cell gene expression data into “cell sentences” that LLMs can understand.
I was very pleased to learn on Monday that The Tail at Scale paper (research.google/pubs/the-tail-at-scale/) that my dear colleague @labarroso and I wrote in 2013 was selected for the SIGOPS Hall of Fame award (www.sigops.org/awards/hof/). The award was going to be presented 36 hours later in Korea at @sospconf, but I wasn't going to be able to attend on such short notice, so my Google colleague Jeff Mogul accepted the award on behalf of myself and Luiz. I wrote some remarks for Jeff M. to deliver:
----
I am honored to have The Tail at Scale work recognized with the SIGOPS Hall of Fame award, and I know Luiz would have been thrilled as well. For those who don't know, Luiz passed away unexpectedly in 2023: he was an incredible computer scientist and a wonderful colleague. We made 20,000 cappuccinos together over many decades, and I and many other colleagues miss him dearly.
This work came about because we had both built up a fair bit of experience working on Google's production services, which are often composed of many large scale distributed systems running on hundreds or thousands of computers. Being able to respond quickly to user requests is one important way to make users enjoy your product more!
If a single server has an average response time of 10 ms, but a 99%ile response time of 1 second, if a user request needs to touch 100 such servers to handle a request, then 63% of user requests will take more than one second. Response time variability can arise from many sources, such as shared resource contention, background or other activity on shared machines, and more. In writing the paper, we realized there wasn't really a good discussion about tail latency in large scale systems, why latency variability occurs, and what to do about it. We had also developed a couple of interesting techniques to help combat variability, such as tied requests and hedged requests, that we discussed in the paper.
I was initially dubious about writing this paper, as it seemed like a bit of a catalogue of problems and a few solutions we had found useful, rather than a more typical paper that focuses on one novel technique. Luiz convinced me that it was still worthwhile to write up. Given this more practitioner approach rather than a single novel research contribution, we decided CACM would be a good venue and published it there. I gave a talk at an O'Reilly conference in 2014 about this material that you can find on YouTube under the title "Achieving Rapid Response Times in Large Scale Online Services" (youtu.be/1-3Ahy7Fxsc)
Writing this paper with Luiz was a delight. His focus on computer architecture and my focus on higher level software systems meant we didn't get to write papers together very often, but it was a collaborative and joyful experience the few times we did so. For those who want to learn more about Luiz's work, I encourage you to peruse his web site at barroso.org. In particular, see the video of his remarks when receiving the Eckert Mauchly Award, and also check out his amazing wildlife photos!
Thank you to the committee and to the community for seeing the long-term value of this work!
Links to Luiz's wildlife photos are here:
barroso.org/wildlife.html
As a Silicon Valley resident, I recommend "Wildlife on the edge of Silicon Valley", but they're all amazing.
barroso.org/photos_silicon_valley/
Luiz's Eckert Mauchly Award lecture video is here:
www.youtube.com/watch?v=Lv_eZX99lUU
A Brief History of Warehouse-scale Computing: a practitioner's perspective, 2020 ACM/IEEE Computer Society Eckert Mauchly award lecture.
Jeff Dean reposted
We’re proud to announce that Genie 3 has been named one of @TIME’s Best Inventions of 2025.
Genie 3 is our groundbreaking world model capable of generating interactive, playable environments from text or image prompts.
Find out more → goo.gle/3KGqiYa
Jeff Dean reposted
🪩The one and only @stateofai 2025 is live! 🪩
It’s been a monumental 12 months for AI. Our 8th annual report is the most comprehensive it's ever been, covering what you *need* to know about research, industry, politics, safety and our new usage data.
My highlight reel:
Today we're rolling out a whole slew of new capabilities for using contextual data relevant to you and your organization and for building and using agent-based systems on top of Gemini and @googlecloud!
Learn more at:
cloud.google.com/blog/products/ai-machine-learning/introd…
Using Gemini and agents to extract information or accomplish things based on data from your organization makes lots of things easier. For example:
(1) "Based on past meeting notes, what are the project's pending action items"
(2) "MarketingMediaAgent We're launching our new jacket line tomorrow. Create a few images showing the Khaki Explorer Outwear Ski Jacket on someone skiing in Tahoe."
(3) "Schedule a meeting with the team for tomorrow at 10 AM"
Jeff Dean reposted
Our new Gemini 2.5 Computer Use model can navigate browsers just like you do. 🌐
It builds on Gemini’s visual understanding and reasoning capabilities to power agents that can click, scroll and type for you online - setting a new standard on multiple benchmarks, with faster speed.
Jeff Dean reposted
Our new Gemini 2.5 Computer Use model is now available in the Gemini API, setting a new standard on multiple benchmarks with lower latency. These are early days, but the model’s ability to interact with the web – like scrolling, filling forms + navigating dropdowns – is an important next step in building general-purpose agents.
Developers can try these capabilities via API in @googleaistudio + Vertex AI.
I'm delighted to serve on the selection committee for the initial phase of @LaudeInstitute's Moonshots Program, designed to fund research groups on tackling hard, real-world problems. The initial selections (abstracts due Oct 25th, full proposals by Nov 25) will be in four areas: civic discourse, accelerating science, frontline healthcare, and workforce reskilling.
Selection committee members include:
@timalthoff, @davidautor, @chris_bail, @erikbryn, @YejinChoinka, @fchollet, Mariano-Florentino (Tino) Cuéllar, @JeffDean, @ghadfield, @HannaHajishirzi, @bradenjhancock, John Hennessy, @erichorvitz, @JohnJumperSci, @tkalil2050, @zakkohane, @andykonwinski, @tommmitchell, @emollick, Dave Patterson (Chair), @PlevritisLab, @DG_Rand, @Thom_Wolf, @YalaTweets, @Diyi_Yang, @james_y_zou
Please spread the word to people who you think might submit interesting proposals in these areas!
We’re counting down to our first round of Laude Moonshots. It’s energizing to see researchers mobilizing AI for urgent, real-world impact. Abstracts are due October 25. Details: moonshots.laude.org
The proof is in the evolutionary pudding!
Today we describe how we leverage AlphaEvolve, a @GoogleDeepMind system for iteratively evolving code, to morph snippets of code towards better proof elements in complexity theory that can be automatically verified by a computer program. Read more at: goo.gle/3IR6vEN
Jeff Dean reposted
Today, @ekindogus and I are excited to introduce @periodiclabs.
Our goal is to create an AI scientist.
Science works by conjecturing how the world might be, running experiments, and learning from the results.
Intelligence is necessary, but not sufficient. New knowledge is created when ideas are found to be consistent with reality. And so, at Periodic, we are building AI scientists and the autonomous laboratories for them to operate.
Until now, scientific AI advances have come from models trained on the internet. But despite its vastness — it’s still finite (estimates are ~10T text tokens where one English word may be 1-2 tokens). And in recent years the best frontier AI models have fully exhausted it.
Researchers seek better use of this data, but as any scientist knows: though re-reading a textbook may give new insights, they eventually need to try their idea to see if it holds.
Autonomous labs are central to our strategy. They provide huge amounts of high-quality data (each experiment can produce GBs of data!) that exists nowhere else. They generate valuable negative results which are seldom published. But most importantly, they give our AI scientists the tools to act.
We’re starting in the physical sciences.
Technological progress is limited by our ability to design the physical world.
We’re starting here because experiments have high signal-to-noise and are (relatively) fast, physical simulations effectively model many systems, but more broadly, physics is a verifiable environment. AI has progressed fastest in domains with data and verifiable results - for example, in math and code. Here, nature is the RL environment.
One of our goals is to discover superconductors that work at higher temperatures than today's materials. Significant advances could help us create next-generation transportation and build power grids with minimal losses. But this is just one example — if we can automate materials design, we have the potential to accelerate Moore’s Law, space travel, and nuclear fusion.
We’re also working to deploy our solutions with industry. As an example, we're helping a semiconductor manufacturer that is facing issues with heat dissipation on their chips. We’re training custom agents for their engineers and researchers to make sense of their experimental data in order to iterate faster.
Our founding team co-created ChatGPT, DeepMind’s GNoME, OpenAI’s Operator (now Agent), the neural attention mechanism, MatterGen; have scaled autonomous physics labs; and have contributed to some of the most important materials discoveries of the last decade. We’ve come together to scale up and reimagine how science is done.
We’re fortunate to be backed by investors who share our vision, including @a16z who led our $300M round, as well as @Felicis, DST Global, NVentures (NVIDIA’s venture capital arm), @Accel and individuals including @JeffBezos , @eladgil , @ericschmidt, and @JeffDean. Their support will help us grow our team, scale our labs, and develop the first generation of AI scientists.
Jeff Dean reposted
Introducing two new Gemini 2.5 models (Flash and Flash-Lite) which are more intelligent, cost effective, and token efficient. You can keep up with our latest models through `gemini-flash-latest` and `gemini-flash-lite-latest`!!
Now you can automate many kinds of browser activity using the new Chrome DevTools MCP. In the other retweets of the base post, many people are already thinking of lots of great use cases for this functionality!
Announcing our public preview of Chrome DevTools MCP! Experience the full power of DevTools in your AI coding agent → goo.gle/4pDE6Tk
With Chrome DevTools MCP, your AI agent can run performance traces, inspect the DOM, & perform real-time debugging of your web pages.
Here's one example by @deliprao ⬇️
Chrome DevTools MCP is going to be a Swiss-Army knife MCP. Watch me use it here with Gemini CLI to open Google Scholar, search for a term, and save the top 5 PDFs to a local folder. I cannot stop thinking of all sorts of personal automation possibilities with this. Amazing release from the @ChromiumDev team!
Announcing our public preview of Chrome DevTools MCP! Experience the full power of DevTools in your AI coding agent → goo.gle/4pDE6Tk
With Chrome DevTools MCP, your AI agent can run performance traces, inspect the DOM, & perform real-time debugging of your web pages.
Jeff Dean reposted
Chrome DevTools MCP is going to be a Swiss-Army knife MCP. Watch me use it here with Gemini CLI to open Google Scholar, search for a term, and save the top 5 PDFs to a local folder. I cannot stop thinking of all sorts of personal automation possibilities with this. Amazing release from the @ChromiumDev team!
Announcing our public preview of Chrome DevTools MCP! Experience the full power of DevTools in your AI coding agent → goo.gle/4pDE6Tk
With Chrome DevTools MCP, your AI agent can run performance traces, inspect the DOM, & perform real-time debugging of your web pages.
Jeff Dean reposted
Vertex AIのGoogle MapsグラウンディングがGAに。GeminiからMapsの情報を簡単に検索できる。 cloud.google.com/vertex-ai/generative-ai/docs/grounding/g…
Very nice analysis by neurosurgeon @slotkinjr of how @Waymo's autonomous vehicles have much better safety properties than human drivers, and what would happen if everyone in the U.S. drove as safely as a Waymo.
"The national math: If every US vehicle performed like Waymo, we’d prevent 33,000-39,000 deaths annually and save $0.9-1.25 trillion in societal costs. Even partial adoption at 27% would save ~10,000 lives per year. In terms of magnitude, this would be the equivalent of eliminating every pedestrian death nationally in a year."
As a neurosurgeon I care a lot about road safety.
By now you’ve probably seen @Waymo’s stunning safety results (like 91% fewer serious crashes). But they didn’t just publish data headlines. They released the raw CSV files and data dictionaries.
I did a much deeper analysis. A fascinating story emerges when you analyze how they’re achieving this.
This isn’t incremental improvement - it’s categorical. We’re looking at the potential elimination of traffic deaths as a leading cause of mortality.
The intersection breakthrough: Waymo has essentially solved intersection crashes, with 95% fewer injury incidents than human drivers in the same locations. That’s transforming the deadliest driving scenario.
The national math: If every US vehicle performed like Waymo, we’d prevent 33,000-39,000 deaths annually and save $0.9-1.25 trillion in societal costs. Even partial adoption at 27% would save ~10,000 lives per year. In terms of magnitude, this would be the equivalent of eliminating every pedestrian death nationally in a year.
The physics signature: Here’s what fascinates me: 47% of Waymo’s contacts involve less than 1 mph delta-V. They’re not just avoiding crashes; they’re converting unavoidable incidents into gentle bumps. It’s like having physics itself on your side.
We’re not talking about marginal safety gains. The data represents a fundamental shift from harm reduction to harm prevention.
The methodology matters: I used their dynamic geographic benchmarks (comparing like-for-like road conditions) and verified the findings hold across San Francisco, Phoenix, LA, and Austin. The safety advantage actually increases in more complex urban environments.
Link to raw data below….
Notes on my approach:
Analysis based on 96 million miles of Waymo Rider-Only (RO) data through June 2025, utilizing Waymo's dynamic geographic benchmarks to compare Waymo Driver performance against human drivers under similar road conditions and operational design domains.
The projections for national impact (deaths prevented, societal costs) involve several assumptions. Given Waymo's zero reported fatalities, the direct serious injury reductions were mapped to national fatality statistics using established NHTSA-derived ratios that correlate serious injury crash rates with fatality rates. This extrapolation assumes that Waymo's observed serious injury prevention capability would translate proportionally to fatality prevention. Societal cost savings are estimated by applying average per-fatality and per-injury economic costs (e.g., medical, lost productivity, quality of life) as published by NHTSA, scaling these national averages to the projected number of avoided fatalities and injuries based on Waymo's safety performance. These figures represent the potential annual impact if the Waymo Driver's safety profile were widely integrated into the national fleet.
@ethanteicher
This was my response to an email that a student named Loa sent me in 2024.
I followed up and exchanged a couple more messages with them.
Here's my original post:
x.com/JeffDean/status/1847822246624317665
Jeff Dean, Google's Chief Scientist, tells a student what it's like to be a computer scientist.
v/@JeffDean
Jeff Dean reposted
Incredible milestone: an advanced version of Gemini 2.5 Deep Think achieved gold-medal performance at the ICPC World Finals, a top global programming competition, solving an impressive 10/12 problems. Such a profound leap in abstract problem-solving - congrats to @googledeepmind!
Jeff Dean reposted
We’re announcing a major advance in the study of fluid dynamics with AI 💧 in a joint paper with researchers from @BrownUniversity, @nyuniversity and @Stanford.
Jeff Dean reposted
Excited to share an important advance in AI and math. Together with mathematicians from @BrownUniversity, @nyuniversity and @Stanford, we developed a new AI-powered method that has discovered an entirely new family of solutions to several complex equations in fluid dynamics.
Jeff Dean reposted
The Northern Pacific Ocean is currently smashing temperature records.
And it is reaching these levels far earlier than the current generation of climate models had expected.
A short thread 🧵
Jeff Dean reposted
I’m excited to announce that an advanced version of Gemini Deep Think achieved gold-medal level performance at the 2025 ICPC World Finals, one of the world’s most prestigious programming competitions! 🥇Learn more in our blog post: deepmind.google/blog/gemini-achieves-gold-medal-level-at-…
An inspiring moment for me personally was when our model solved a problem that no university team solved during the contest — a true moment of innovation. With Gemini Deep Think achieving gold-level across ICPC & IMO, I think we’re seeing a profound leap in generalization across coding, math and reasoning capabilities, to generate novel solutions to complex problems.
This is a huge milestone for us on an amazing journey. Really grateful and proud of our team, for all the hard work and teamwork that made this breakthrough possible. Looking forward to continuing our research, helping people use Gemini to solve some of the hardest unsolved problems in the world!
Very excited to see our Gemini models getting better and better at coding! An advanced version of Gemini 2.5 Deep Think at the 2025 International Collegiate Programming Contest (ICPC) World Finals achieved gold-medal level performance! 🎉
deepmind.google/discover/blog/gemini-achieves-gold-level-…
Jeff Dean reposted
Excited to share what friends and I have been working on at @Standard_Kernel
We've raised from General Catalyst (@generalcatalyst), Felicis (@felicis), and a group of exceptional angels.
We have some great H100 BF16 kernels in pure CUDA+PTX, featuring:
- Matmul 102%-105% perf of cuBLAs in 100 lines of code
- Attention 104% perf of FlashAttention3 in 500 lines
- Fused Llama3 FFN 120% perf of PyTorch (gpt-fast)
Reach out if you want to work on AI kernel gen with us!
Read how work by @UChicago, building on the AI-based NeuralGcM weather model developed & open-sourced by @GoogleResearch, is being used to more accurately predict the monsoon season in India and support farmer decision-making for 38M farmers in India. 🌱
blog.google/technology/research/indian-farmers-monsoon-pr…
Jeff Dean reposted
Great news! SB 79 — which would speed up housing construction in transit corridors — just passed both houses of the CA legislature.
We need more housing in CA — for workers, for seniors, for students, for families. Let's build it!
SB 79 HAS PASSED!
Jeff Dean reposted
Made it to no.1 in the App Store. Congrats to the @GeminiApp team for all their hard work, and this is just the start, so much more to come!
Wow, $GOOGL Gemini has overtaken ChatGPT in top downloads in the US on iOS.
Something to keep an eye on as ChatGPT has dominated the standings for months now.
$GOOGL execution and product shipment showing results…
VaultGemma is a release of an open model trained from scratch with differential privacy.
The blog post below and the full tech report linked from the tech report have some nice analyses to present a scaling law for differentially private language models:
Blog: research.google/blog/vaultgemma-the-worlds-most-capable-d…
Paper: arxiv.org/abs/2501.18914
Introducing VaultGemma, the largest open model trained from scratch with differential privacy. Read about our new research on scaling laws for differentially private language models, download the weights, & check out the technical report on the blog →goo.gle/46fUSiq
Jeff Dean reposted
During Thanksgiving break last year, our AI co-scientist team @GoogleDeepMind @GoogleResearch - @Mysiak @alan_karthi met Prof @jrpenades @CostaT_Lab of @ImperialCollege.
They were nearing a breakthrough on how bacteria share resistance genes and proposed a test for our AI co-scientist: pose their exact research question to the system and see what it found.
Today in @CellCellPress, we share the results.
Paper link - www.cell.com/cell/fulltext/S0092-8674(25)00973-0
Jeff Dean reposted
I’ve been beta-testing @GoogleDeepMind’s AI Co-Scientist for quite some time, and it’s so incredible I can’t stop raving about it to my colleagues! In fact, I told Vivek and the rest of the amazing team who built it that I now call it the AI SCIENTIST, not “Co” anymore for me!
This is just the beginning.
In the coming days and months, we'll share more prospective validated results where our AI co-scientist generated novel, testable hypotheses for complex open-ended problems, showing how the system can meaningfully accelerate scientific research.
More very soon.
AI co-scientist 1.0 blog - research.google/blog/accelerating-scientific-breakthrough…
Jeff Dean reposted
🎉 It's been a big week on the Veo front. Here are the highlights:
—You can now generate vertical format videos in @FlowbyGoogle and via the Gemini API.
—We reduced prices in the Gemini API by ~50%, so now developers can build with Veo 3 at $0.40/second and Veo 3 Fast at $0.15/second.
To make sure you're getting the most out of Veo 3, here’s an easy prompt guide you can use for optimal video outputs ⬇️
Jeff Dean reposted
The growth of grid-scale batteries in the US continues to explode.
Over the last 12 months, the country added 13.2 GW of battery capacity—50% more than a year ago.
🧵
People find @NanoBanana very appealing!
Update: In the last 4 days, @GeminiApp has added:
+ 13M more first-time users (23M+ total now)
+ 300M more images (500M+ total now)
🍌 @NanoBanana is unreal
Jeff Dean reposted
🚨New languages coming to AI Mode in Search: Hindi, Indonesian, Japanese, Korean + Brazilian Portuguese. Try it at google.com/ai
Jeff Dean reposted
EmbeddingGemma is our new best-in-class open embedding model designed for on-device AI. 📱
At just 308M parameters, it delivers state-of-the-art performance while being small and efficient enough to run anywhere - even without an internet connection.
Jeff Dean reposted
✅ Papercut fixed: You can now upload any file to @GeminiApp. Including the #1 request: audio files are now supported!
Jeff Dean reposted
My team at @GoogleDeepMind is hiring DevX PM’s, DevX Engineers (fkn DevRel), and more!
Pls DM me if you want to work together and are best in class at what you do : )
Jeff Dean reposted
We’re helping to unlock the mysteries of the universe with AI. 🌌
Our novel Deep Loop Shaping method
published in @ScienceMagazine could help astronomers observe more events like collisions and mergers of black holes in greater detail, and gather more data about rare space phenomena. 🧵
Jeff Dean reposted
Introducing EmbeddingGemma, our newest open model that can run completely on-device. It's the top model under 500M parameters on the MTEB benchmark and comparable to models nearly 2x its size – enabling state-of-the-art embeddings for search, retrieval + more.
We've had a busy August!
August at Google DeepMind was like 🧞♂️ 🖼️ 🍌 🚀 🔍 🤏🏻
- Nano Banana (Gemini 2.5 Flash Image)
- Gemini Embedding
- Veo 3 Fast
- Genie 3
- Imagen 4 Fast
- Gemma 3 270M
- Perch 2
- Kaggle Game Arena
- Gemini API Url Context
- AI Studio Builder (UI Rework, Prompt Suggestions, GitHub integration …)
- AI studio UI (Model Picker, Chat, Scrolling…)
and much more!
Jeff Dean reposted
NINE former CDC Directors spanning every administration from Carter to Trump, have issued a rare public rebuke of RFK Jr., warning of the harm his rhetoric and attacks on public health is doing. Their message: this is a time to rally to protect the health of every American.
Honored to be named in TimeAI 100 list, along w/my awesome Google colleagues @joshwoodward and Hartmut Neven and many amazing people.
@billyperrigo did a very nice writeup: 🙏
time.com/collections/time100-ai-2025/7305831/jeffrey-dean/
My recognition is due to collaboration with many great colleagues on a wide range of interesting problems, so thank you to everyone I've been fortunate enough to work with over many years! 🙏
Behind every AI breakthrough and launch, there are hundreds of Googlers working to build AI that’s helpful for everyone.
This year, our very own @JoshWoodward VP of Gemini and Google Labs, @JeffDean Chief Scientist at DeepMind and Google Research, and Hartmut Neven founder and lead of Google Quantum AI have been recognized on @TIME’s #time100ai, a list recognizing the 100 most influential people in artificial intelligence.
Find the full list and learn how we’re thinking about the future of AI at time.com/time100ai
Great interview with my awesome colleagues @19kaushiks @robertriachi @m__dehghani and @nbrichtova by @OfficialLoganK about the new image generation and editing capabilities in our latest Gemini Flash 2.5 model (also known as nano banana 🍌)
A conversation with some of the research folks behind nano-banana 🍌 (aka Gemini 2.5 Flash Image) on how we got here, what it took to build this model, and where we go next!
So much fun to hang with: @19kaushiks @robertriachi @m__dehghani @nbrichtova
Jeff Dean reposted
I have posted this before but ..
26 years ago, a gunman entered
Dunblane Primary School in Scotland,
killing 16 kids and a teacher. The UK
govt responded by enacting tight gun
control legislation. In the 9400+ days
since, there have been a total of O
school shootings in the UK.
Go with the flow, Ultra subscribers!
Starting today, Google AI Ultra subscribers can generate Veo 3 Fast videos without using any credits. So you can fine-tune your clips to exactly how you want them and only spend credits on that final upload.
May this be the pièce de résistance to that new project — go forth and create!
Jeff Dean reposted
1/ INSANE (LONG) THREAD.
Nano Banana 🍌 for photo restoration and colorization.
Nothing has ever been like this before. I spent countless hours working on photo restoration and colorization solutions long before AI was a thing. Nothing compares to this. Truly remarkable.
Migrant Mother, 1936
Jeff Dean reposted
Every month, people use Google to translate around 1 trillion words. Today, we’re introducing a new AI-powered live translation experience in the Google Translate app, plus a new beta feature to help you practice new languages. Rolling out now on iOS + Android.
My Google colleagues Norm Jouppi & Sridhar Lakshmanamurthy gave a talk today at Hot Chips on TPUv7 ("Ironwood").
The TPUv7 system offers 9216 chips / pod (42.5 exaflops of fp8), but we can scale across many of these pods to provide multiple zettaflops.
Google presents for the first time ever their TPUv7 block diagram at hot chips conference. TPUv7 (formerly known as TPUv6p, internally called ghostfish) has 8 stacks of HBM3e memory, 4 medium size systolic arrays and be connected in a 3D torus with a scale up world size of up to 9216.
Note that even with OCS, the blast radius of 3D torus is quite huge and a failure of 1 chip takes out a whole cube of 64 TPU chips. Previously without OCS, a failure of 1 chip in an 3d torus would take the whole cluster such as in TPUv3
Great to see new NotebookLM features that support many languages.
techcrunch.com/2025/08/25/notebooklms-video-overview-feat…
Jeff Dean reposted
As impressive as the Elo scores look, they are actually underselling the dominance of the Gemini 2.5 Flash Image model on the @lmarena_ai. If we look at the win rates, 'nano-banana' won >85% of the time - this is based on 2.5 million votes by the way.
And remember this is not just 'human preference' for colour scheme or something, this is people with real tasks judging whether their task was completed well or not, e.g. did it transfer the style correctly, did it remove what I wanted it to remove etc.; this is real evidence of utility.
Nice work @GoogleDeepMind, @demishassabis @JeffDean @OfficialLoganK + 400 people that we'll hopefully see on the future paper
Introducing Gemini 2.5 Flash Image (aka nano-banana), our SOTA image generation and editing model 🍌
As you might have already seen, this model excels at character consistency, creative edits, and has Gemini's world knowledge!
Jeff Dean reposted
Introducing Gemini 2.5 Flash Image (aka nano-banana), our SOTA image generation and editing model 🍌
As you might have already seen, this model excels at character consistency, creative edits, and has Gemini's world knowledge!
Our latest Gemini image generation and editing model is quite good (see the examples in the thread below). It lets you indulge in a bit of creative fun, helps you make new business cards, etc!
Try it out at gemini.google.com
@GeminiApp
Image generation with Gemini just got a bananas upgrade and is the new state-of-the-art image generation and editing model. 🤯
From photorealistic masterpieces to mind-bending fantasy worlds, you can now natively produce, edit and refine visuals with new levels of reasoning, control and creativity.
A quick dive into Gemini 2.5 Flash’s capabilities 🧵
This model scores very well in the image generation lmarena leaderboard, with a large gap of ~170 points to the next group of models.
lmarena.ai/leaderboard/image-edit
Jeff Dean reposted
The new Gemini 2.5 Flash Image (aka nano-banana) is amazing.
I'm looking forward to participating in this Gemini symposium in Singapore next week!
Excited to share that I'll be hosting some of the world's best AI researchers and engineers for our @GoogleDeepMind Gemini event next week in Singapore 🇸🇬!
Join @JeffDean, @quocleix, @benoitschilling, @melvinjohnsonp and @denny_zhou for a day of technical conversations, panels and talks about AI, reasoning and our mission to build a world class AI frontier lab in Singapore.
If you're in town and would like to attend, please check the RSVP link below👇. Note, subject to capacity constraints and you'll need to be approved to join.
Jeff Dean reposted
Good idea. Let’s double it for everyone.
This weekend only, ending tomorrow (8/24) at 10pm PT:
👉 Gemini free users, you now have up to 6 free video generations for the weekend.
👉 Pro users, your pack has doubled to 6 per day for the weekend.
👉 Ultra users, your pack has doubled to 10 per day for the weekend.
Rates do not reset (so if you already used 3, you now have 3 more).
Enjoy!
3 videos is very little Google, it should be double 😌
Jeff Dean reposted
Free Veo 3 this weekend in the @GeminiApp, enjoy : )
Jeff Dean reposted
It seems like there is not enough of a policy response to the fact that, with 57M miles of data, Waymo’s autonomous vehicles experience 85% less serious injuries & 79% less injuries overall than cars with human drivers.
2.4 million are injured & 40k killed in US accidents a year
AI efficiency is important. Today, Google is sharing a technical paper detailing our comprehensive methodology for measuring the environmental impact of Gemini inference. We estimate that the median Gemini Apps text prompt uses 0.24 watt-hours of energy (equivalent to watching an average TV for ~nine seconds), and consumes 0.26 milliliters of water (about five drops) — figures that are substantially lower than many public estimates.
At the same time, our AI systems are becoming more efficient through research innovations and software and hardware efficiency improvements. From May 2024 to May 2025, the energy footprint of the median Gemini Apps text prompt dropped by 33x, and the total carbon footprint dropped by 44x, through a combination of model efficiency improvements, machine utilization improvements and additional clean energy procurement, all while delivering higher quality responses.
See the blog or technical paper for more about our methodology and ongoing efforts.
Blog:
cloud.google.com/blog/products/infrastructure/measuring-t…
Link to detailed paper: services.google.com/fh/files/misc/measuring_the_environme…
A huge number of people from all across Google working on Gemini, software and hardware infrastructure, datacenter operations, clean energy procurement and many other aspects all contribute mightly to this ongoing effort of delivering our AI models with the highest efficiency.
The paper is also now up on Arxiv:
arxiv.org/abs/2508.15734
Had a great time with @chrismichel yesterday. He took photos of me for a series of portraits he's making of members of @theNAEng and @theNASEM.
All his pictures are fantastic: you can see many in this series at explorers.com
Thanks, Chris!
New Heroes portrait of @Google's @JeffDean for the @theNAEng & @theNASEM. Spending time with Jeff is a reminder that kindness and brilliance can live side by side. He is one of the great architects of the modern internet, the co-creator of MapReduce and Bigtable, and the driving force behind TensorFlow. Today his work touches billions of lives, yet he carries it all with an amused, approachable warmth. In conversation he is rigorous and honest, but also generous, especially with younger engineers who still recall his detailed, thoughtful code reviews. Elected to the National Academy of Engineering, Jeff continues to shape the future of artificial intelligence while never losing his delight in the puzzle of a good problem. Full bio: explorers.com/jeff-dean/
Jeff Dean reposted
I’m excited to share our new @Nature paper 📝, which provides strong evidence that the walkability of our built environment matters a great deal to our physical activity and health.
Details in thread.🧵
www.nature.com/articles/s41586-025-09321-3
Jeff Dean reposted
Here’s what we shipped this week 🚢🚢🚢
—We launched a new Imagen 4 Fast model so developers can quickly generate images at only $0.02 per image and updated Imagen 4 and Imagen 4 Ultra to support 2K images. All are now generally available in the Gemini API for developers and @GoogleCloud Vertex AI for enterprise customers.
—We introduced Gemma 3 270M, a new model in our Gemma family that’s hyper-efficient and best for developers to take and fine-tune for specific tasks.
—Google AI Ultra subscribers can now run 2X more Deep Think queries in the @GeminiApp, up to 10 prompts per day.
—The @GeminiApp can now reference your past chats to deliver more personalized responses. We also introduced Temporary Chats and new privacy settings so you can create the experience that’s right for you.
—New research from @GoogleResearch & @GoogleDeepMind introduces g-AMIE and explores how AI can assist in doctor-patient conversations while keeping physicians in control.
Jeff Dean reposted
Google’s research on a Personal Health Large Language Model (PH-LLM) is now published in @NatureMedicine Magazine!
Jeff Dean reposted
Introducing Gemma 3 270M 🔥
🤏A tiny model! Just 270 million parameters
🧠 Very strong instruction following
🤖 Fine-tune in just a few minutes, with a large vocabulary to serve as a high-quality foundation
developers.googleblog.com/en/introducing-gemma-3-270m
My longtime collaborator Dave Patterson (long-time faculty at @UCBerkeley, @TheOfficialACM Turing Award winner, and fellow @LaudeInstitute board member) wrote a very good op-ed about how continued investing in basic science and technology research is essential for the U.S. Dave describes the government funding his labs received, but also the impact:
"For decades, my research received support from the NSF’s Directorate for Computer and Information Science and Engineering, through grants and Ph.D. fellowships. All told, American taxpayers invested just under $100 million in the labs I helped lead. Accounting for inflation, technologies that came out of them went on to generate over $1 trillion in product sales. That is a 10,000-to-1 return on investment to the public and surely at least 1,000-to-1 return directly back to the government in taxes. Not from luck, but from decades of public investment."
...
"Every American taxpayer is a silent shareholder in that success. If we walk away now, we lose not just future breakthroughs but also what we have already earned."
Continuing to invest in basic research in the U.S. is what leads to these sorts of extraordinary outcomes.
thehill.com/opinion/technology/5439205-congress-wants-to-…
Some quite exciting improvements in genome assembly accuracy from @GoogleResearch and @ucsc Genomics Institute!
Introducing DeepPolisher — in collaboration with UC Santa Cruz Genomics Institute, we developed this new open-source method to improve genome assembly accuracy. It reduces indel errors by 70% and total assembly errors by 50%. Learn more: goo.gle/3Hn8PCG
Jeff Dean reposted
Introducing DeepPolisher — in collaboration with UC Santa Cruz Genomics Institute, we developed this new open-source method to improve genome assembly accuracy. It reduces indel errors by 70% and total assembly errors by 50%. Learn more: goo.gle/3Hn8PCG
Jeff Dean reposted
Foundry → Mithril (@mithrilcompute): The AI Omnicloud.
Now generally available!
We’re redefining GPU cloud economics, workload flexibility, and ease-of-use—for the compound AI & agentic era. 🧵 (1/8)
Jeff Dean reposted
Introducing Storybook in @GeminiApp today! Now you can create personalized, illustrated storybooks with read-aloud narration in 2 simple steps. It's free and available globally right now.
Here's how to make yours + a famous story from our house about "Teddy" …
Jeff Dean reposted
Gemma just passed 200 million downloads 🔥
The speed of adoption, exciting use cases, and projects the community is building with Gemma is incredible!
And we're just getting started 💎What do you want to see next?
Our Genie 3 model shows the possibilities of generative video that is not only based on a prompt to generate the video, but also can be made interactive by taking into account actions the user takes to explore the environment that is generated.
Now you can interact with your imagination!
What if you could not only watch a generated video, but explore it too? 🌐
Genie 3 is our groundbreaking world model that creates interactive, playable environments from a single text prompt.
From photorealistic landscapes to fantasy realms, the possibilities are endless. 🧵
Jeff Dean reposted
Excited to share this program #gstar for global talents by @newturing. I personally designed the syllabus with our experienced team at NTI and leading AI experts. This would be an intensive 12-week program to train talents not only in state-of-the-art LLM but also in the research and leadership skills.
I'm glad to have @YiTayML joining me as the overall technical lead of the program, together with many cool people (@JeffDean, @chrmanning, @preslav_nakov, @gneubig, @PoShenLoh, @tuvllms, @HoangDuon, @longphan3110) and more coming soon!
The deadline for application is a week from now Aug 12, so please hurry! See thread for the program website and application link.
(apology that I only have time to do this after IMO, so the timeline is a bit tight, but hopefully we'll have a great program!)
#gstar bootcamp, a global talent program by NTI, is happening now! Deadline to apply: August 12, 2025. Checkout details in gstar.newturing.ai.
Program director: @lmthang, our overall technical lead: @YiTayML and senior advisors: @JeffDean, @chrmanning , @preslav_nakov.
Together with many great lecturers such as @gneubig, @PoShenLoh, @tuvllms, @HoangDuon, @longphan3110 and more coming soon!
Jeff Dean reposted
tl;dr - IPAM, the Institute of Pure and Applied Mathematics at UCLA, will shut down in a few months if their NSF funding isn't restored or replaced. The birthplace of compressed sensing, as well as lots of advances in ML and AI later picked up by industry.
Terence Tao has written a thread on Mathstodon about the damage being caused by the grant freezes at UCLA.
mathstodon.xyz/@tao/114956840959338146
Jeff Dean reposted
Had early access to Gemini with Deep Think. Very good model, big gains over standard Gemini 2.5 Pro for a lot of problems.
Here is the first attempt at the starship control panel prompt I try with every model. First time I have seen a model make a 3D interface in response.
Kinda amazing: the mystery model "summit" with the prompt "create something I can paste into p5js that will startle me with its cleverness in creating something that invokes the control panel of a starship in the distant future" & "make it better"
2,351 lines of code. First time
Jeff Dean reposted
The Waymo Driver offers a consistent and safe experience across cities. Skills for safety-critical situations transfer very well: what's learned in one, like handling lane-cutters in LA or Austin, applies to others.
See mathematician Michel van Garrel talking about how our latest Gemini Deep Think model was able to prove a conjecture using a very different approach than he was considering.
For researchers, scientists, and academics tackling hard problems: Gemini 2.5 Deep Think is here. 🤯
It doesn't just answer, it brainstorms using parallel thinking and reinforcement learning techniques. We put it into the hands of mathematicians who explored what it can do ↓
Starting today, we're offering Deep Think in the Gemini app for Google AI Ultra subscribers, and we're giving a collection of mathematicians access to the full version of the Gemini 2.5 Deep Think model that achieved gold medal🏅 level performance in the recent IMO competition. This model excels at complex math, reasoning and coding.
See more at: blog.google/products/gemini/gemini-2-5-deep-think/
Jeff Dean reposted
Just ~2 months in and AI Mode already has over 100 million monthly active users in the U.S. and India, and it's great to hear the awesome response from our users!! Now, we’re excited to ship AI Mode, our most powerful AI search, to the UK!
Starting to roll out today. 🇬🇧
Jeff Dean reposted
Our new AI model AlphaEarth Foundations is mapping the planet in astonishing detail. 🌏🔍
Scientists will now be able to track the impact of deforestation, monitoring crop health, and more – significantly faster, thanks to our new datasets. 🧵
Someone just reminded me of the 2015 SOSP History Day event, and I thought I would point out what a great event this was, with lots of different interesting people giving talks about the first 50 years of operating systems and distributed systems history.
sigops.org/s/conferences/sosp/2015/history/index.html (slides & videos there for all the talks)
(My talk was on one of the most "modern" topics there, "The Rise of Cloud Computing Systems".
Video: dl.acm.org/doi/10.1145/2830903.2830913
Slides: sigops.org/s/conferences/sosp/2015/history/10-dean-slides…)
Jeff Dean reposted
⚠️TSUNAMI ADVISORY remains in effect for the California coast. Stay away from the water! Stay away from beaches and waterways! Remember: Tsunamis are series of waves carrying large masses of water. The first wave may not be the largest! Do NOT go to the coast!! #cawx #tsunami
Jeff Dean reposted
Oh, you thought that was it? In addition to our new look we've got some new tricks too:
🎉 VIDEO OVERVIEWS are starting rollout TODAY 🥳 (English + desktop only to start)
Create visually engaging slide summaries of your notebook content. We can't wait to see what you make!
Jeff Dean reposted
Gemini 2.5 Flash-Lite, our fastest and most cost effective model, is now stable and ready for scaled production use!!
It comes with native reasoning capabilities, a 1 million token context window, and is priced at ($0.10 in / 1M) and ($0.40 out / 1M).
Jeff Dean reposted
Come work with us on helping agents reach unprecedented levels of autonomy :)
We have a ridiculously cracked team and are pushing the bounds of AGI with awesome folks like @robertarail @OriolVinyalsML @_rockt @quocleix
Come join ~~ le rocket ship ~~
I’m building a new team at @GoogleDeepMind to work on Open-Ended Discovery!
We’re looking for strong Research Scientists and Research Engineers to help us push the frontier of autonomously discovering novel artifacts such as new knowledge, capabilities, or algorithms, in an open-ended self-improving loop.
We aim to work on ambitious research projects in a fast-paced manner.
If this sounds appealing to you, apply using the link below by Friday, August 1st EOD:
job-boards.greenhouse.io/deepmind/jobs/7093349
Jeff Dean reposted
Many of you have been saying it, and it's true: @GeminiApp has momentum and it's growing! We've got tons in the pipeline and lots more to do!
Jeff Dean reposted
Our IMO gold model is not just an "experimental reasoning" model. It is way more general purpose than anyone would have expected. This general deep think model is going to be shipped so stay tuned! 🔥
So happy to see this incredible achievement.
Huge congrats to @lmthang, @quocleix, @YiTayML and the IMO team on the result.
This was a great collaboration across teams to build a general Gemini DeepThink model that can also get gold at IMO.
We're all excited to see our advanced Gemini Deep Think model achieve a gold-medal level performance in the recent IMO, solving 5 of the 6 problems perfectly (35 of 42 points)! 🥇
Unlike our entry last year, which first converted the problems to a formal proof language (Lean), this year, our model just took in the natural language of the problem and emitted natural language of the solution.
We're looking to make a version of this Gemini Deep Think model available soon to users.
Very excited to share that an advanced version of Gemini Deep Think is the first to have achieved gold-medal level in the International Mathematical Olympiad! 🏆, solving five out of six problems perfectly, as verified by the IMO organizers! It’s been a wild run to lead this effort and I am grateful to everyone in the team for such an amazing achievement! Blog post in the thread and more to share soon!
Congrats to the whole Gemini team, and especially to those focused on our advanced reasoning and mathematics capabilities! 🎊
More details in the blog post:
deepmind.google/discover/blog/advanced-version-of-gemini-…
A bit more context on how we worked with IMO officials to release this result after an appropriate delay so that the student competitors received their well-deserved accolades, got our results officially graded by IMO judges, etc.
Btw as an aside, we didn’t announce on Friday because we respected the IMO Board's original request that all AI labs share their results only after the official results had been verified by independent experts & the students had rightly received the acclamation they deserved
Jeff Dean reposted
We've now been given permission to share our results and are pleased to have been part of the inaugural cohort to have our model results officially graded and certified by IMO coordinators and experts, receiving the first official gold-level performance grading for an AI system!
Jeff Dean reposted
Btw as an aside, we didn’t announce on Friday because we respected the IMO Board's original request that all AI labs share their results only after the official results had been verified by independent experts & the students had rightly received the acclamation they deserved
Colab Pro is now available for free for verified US students and faculty for one year.
Big news for data science in higher ed! 🚀Colab now offers 1-year Pro subscriptions free of charge for verified US students/faculty, interactive Slideshow Mode for lectures, & an AI toggle per notebook. Enhance teaching & learning in the upcoming academic year! Read all about it here: goo.gle/40pxnl1. #googlecolab #edtech.
Check out the Veo 3 video+audio generation model, now in the Gemini API.
Veo 3 is now live in the Gemini API 📽️!!
Veo 3 is state of the art, can natively generative audio in the videos, comes with rate limits for scaled production use, and is priced at $0.75 per second w/ audio and $0.50 w/o audio. More in 🧵
Check out our #icml2025 poster on Matryoshka Quantization tomorrow, being presented by @puranjay1412 on behalf of all the authors.
Hi, I'll be presenting Matryoshka Quantization (arxiv.org/abs/2502.06786) on 16th July at #icml2025
📍East Exhibition Hall A-B #3606
⏲️ 11 AM - 1:30 PM
FYI, as @jainprateek_ points out, @puranjay1412 is applying to grad schools.
Puranjay will present our poster on nested bitwise models or MatQuant, so if you are ICML and interested in the topic, do bother him :)
Puranjay is going on the grad-school market this cycle. So if you are looking for a brilliant, hardworking student with good ML+LLM exposure, he would be a very good candidate.
Hi, I'll be presenting Matryoshka Quantization (arxiv.org/abs/2502.06786) on 16th July at #icml2025
📍East Exhibition Hall A-B #3606
⏲️ 11 AM - 1:30 PM
Jeff Dean reposted
🚀 Super excited about this @NeurIPSConf challenge!
🚨 To help with training, we open-sourced 5M+ competitive Pokémon battles!!! Can't wait to see how people use the data
🚀 Launch day! The NeurIPS 2025 PokéAgent Challenge is live.
Two tracks:
① Showdown Battling – imperfect-info, turn-based strategy
② Pokemon Emerald Speedrunning – long horizon RPG planning
5 M labeled replays • starter kit • baselines.
Bring your LLM, RL, or hybrid agent—let’s push strategic AI forward.
Jeff Dean reposted
Gemini Embeddings General Available! First Gemini Embedding model (`gemini-embedding-001`) now available for production use, ranking top on MMTEB optimized for finance, science, legal, search, code. 👀
- 🏦 $0.15 per million input tokens with free tier.
- 🥇 Top MMTEB leaderboard with a score of 68.32.
- 💡 Build for finance, science, legal, search, code.
- 🪆 2048 input context, Flexible output dimensions (rec: 3072, 1536, or 768).
- 🌐 Multilingual in +100 languages.
- 🔡 Text input only at the moment.
- 🔜 Batch API support coming soon.
Jeff Dean reposted
@GeminiApp now lets you transform photos into eight-second videos using Veo 3. Rolling out now. Android, iOS, Web. Global availability. Still experimental, but we wanted our Pro and Ultra members to try it first!
It’s really fun to take kindergarten artwork and make it come to life *with sound* :) 🧵
Check out our state-of-the-art open weights MedGemma multimodal model for making sense of longitudinal EHR data as well as medical text and medical imaging data in various modalities (radiology, dermatology, pathology, ophthalmology, etc.)
See the blog post linked below! ⬇️
Introducing new models for research & development of health applications: MedGemma 27B Multimodal, for complex multimodal & longitudinal EHR interpretation, and MedSigLIP, a lightweight image & text encoder for classification, search, & related tasks. → goo.gle/4kvt6Uk
Jeff Dean reposted
By 2026, the U.S. hopes to “get rid of” a quarter million people involved in science research and education.
We won’t come back from this.
The US government is aiming to get rid of a quarter million people involved in science research and education by 2026.
I have no words.
Jeff Dean reposted
My "Math, Revealed" series is freely available to anyone -- no paywall! -- in the thread below
Jeff Dean reposted
Jeff Dean reposted
Seeing text-to-text regression work for Google’s massive compute cluster (billion $$ problem!) was the final result to convince us we can reward model literally any world feedback.
Paper: arxiv.org/abs/2506.21718
Code: github.com/google-deepmind/regress-lm
Just train a simple encoder-decoder from scratch to read the cluster’s complex state as text, then generate numeric tokens.
We’re also seeing strong results on classic tabular data and "exotic" inputs like graphs, system logs, and even code snippets. Feature engineering will no longer exist!
Authors: @yashakha, Bryan Lewandowski, Cheng-Hsi Lin, Adrian N. Reyes, Grant C. Forbes, Arissa Wongpanich, Bangding Yang, @mohsaeid, @SagiPerel, @XingyouSong
Very excited to see this agreement. I was able to tour @CFS_energy's facility in December last year and was excited by the potential! Working fusion energy could be game changing for the world.
Historic news for CFS: We’ve signed a landmark agreement with @Google, a multifaceted partnership that dramatically advances our commercial fusion energy mission:
⚡️Google signed an offtake agreement for 200 megawatts of power from our first ARC power plant
⚡️Google is increasing its corporate investment in CFS
⚡️Google has the option to procure power from future ARCs
This deal is a very strong signal that the world wants the clean, secure power that fusion offers. Together, we’re catalyzing the fusion market.
#fusionenergy #powermoves
Jeff Dean reposted
Imagine a world powered by the same energy that fuels the sun and the stars. ☀️✨
Today we’re taking our biggest step so far to make this a reality through an investment and offtake agreement with @CFS_energy (CFS).
Read more about it in our blog here:
blog.google/outreach-initiatives/sustainability/our-lates…
Jeff Dean reposted
The Senate bill would kick 11.8 million people off their health insurance by cutting health coverage over $1 trillion
Estimated Budgetary Effects of an Amendment in the Nature of a Substitute to H.R. 1, the One Big Beautiful Bill Act, Relative to the Budget Enforcement Baseline for Consideration in the Senate. cbo.gov/publication/61533
Jeff Dean reposted
Starting today, Pro and Ultra users can schedule actions in the Gemini app.
Just prompt something once, make it a scheduled action, and Gemini will take care of the task as often as you asked it to.
Try things like:
🎸 "Find me local music events every Friday."
🗓️ "Send me a daily summary of my calendar, to-do's and important unread emails at 8AM."
Jeff Dean reposted
NEWS: Redwood Materials, founded by Tesla co-founder and current board member JB Straubel, has introduced Redwood Energy.
Redwood Energy's first project will be the largest solar-powered, off-grid data center in North America, and the world’s largest second life (recycled) battery installation to date (62 MWh).
Redwood: "At Redwood, we’ve built a battery supply chain to recover end-of-life batteries and recycle their critical minerals, keeping them in circulation and driving the energy transition. Today, we receive over 20 GWh of batteries annually—the equivalent of 250,000 EVs—representing about 90% of all lithium-ion batteries and battery materials recycled in North America.
As more batteries reach end-of-life, we’re discovering that many still retain more than 50% of their usable capacity—enough to be put back to work before recycling.
Redwood Energy is our solution: a platform to repurpose these battery packs into low-cost, large-scale energy storage systems that fill a critical gap in today’s power landscape, while maximizing their value between recovery and recycling.
More than 100,000 EVs will come off U.S. roads this year alone. Today, over 5 million EVs are active on U.S. roads, representing an estimated 350 GWh of energy that will reach end-of-life in the coming years. And with another 150 GWh entering the operational fleet annually, that total is increasing fast and could supply 50% or more of the entire energy storage market.
We’ve already built a 12 MW and 63 MWh capacity microgrid, which is the largest second-life battery deployment in the world and the largest microgrid in North America. It’s powering a modular data center for AI infrastructure company Crusoe—best known for its large-scale data center campus in Abilene, Texas, the initial site of the Stargate project—at an energy cost lower than the grid/
This is just the beginning for Redwood Energy. We have over a gigawatt-hour of reusable batteries in our deployment pipeline, and that’s expanding by an additional 5 GWh in the coming year. We’re already designing 100+ MW projects—10X the size of this deployment. Redwood Energy will play a critical role in meeting the power demands of the AI era while accelerating American innovation across industries."
Jeff Dean reposted
📢Now open, Gemma 3n weights & it is natively flexible, first of its kind, thanks to MatFormer🪆
Any model between E4B & E2B with ZERO training near Pareto -- we found a bunch!
Find a better E3B than what we released, I will send you a 🪆😉
Find the colab for extraction 🧵👇🪆
Announcing the full release of Gemma 3n, bringing powerful multimodal capabilities to edge devices for developers 🙌 ↓
developers.googleblog.com/en/introducing-gemma-3n-develop…
Jeff Dean reposted
I'm excited to be joining the board of the Laude Institute! We need more support and incentives for university researchers who have great ideas and early results to accelerate their work, and build new real-world solutions that have a meaningful impact on people and society.
Today, I’m launching a deeply personal project. I’m betting $100M that we can help computer scientists create more upside impact for humanity.
Built for and by researchers, including @JeffDean & @jpineau1 on the board, @LaudeInstitute catalyzes research with real-world impact.
Mark your calendars for #mlsys2026 in May, 2026 in Seattle. Submission deadline for papers is Oct 30 this year.
📢Exciting updates from #mlsys2025! All session recordings are now available and free to watch at mlsys.org.
We’re also thrilled to announce that #mlsys2026 will be held in Seattle next May—submissions open next month with a deadline of Oct 30. We look forward to seeing your best work! #mlsys #ai #ml
Jeff Dean reposted
This is a travesty and a nightmare. The US was a founder of @Gavi. It lowers vaccine costs for the world, has vaccinated 1B children, and averted 19M deaths. This pull out will cost 100s of thousands of children's lives a year -- and RFK Jr will be personally responsible.
RFK Jr. says US won’t donate to global vaccine effort www.politico.eu/article/rfk-jr-accuses-gavi-ignoring-vacc…
Jeff Dean reposted
🙌✨ You asked, you've got it:
A free and open-source Gemini agent, run via the command line. And to ensure you rarely, if ever, hit a limit during this preview, we offer the industry’s largest allowance:
*60 model requests per minute and 1,000 requests per day at no charge.*
Gemini CLI is here! Our most powerful open-source CLI that brings Google's Gemini 2.5 models directly into your terminal! With unique features like hierarchical memory (context), self-correcting file edits, and secure sandboxed tool execution.
💡 Hierarchical Memory and instructions from GEMINI md files.
🛠️ Bulit-in tools for filesystem, shell, MCP servers, and new shareable custom extensions.
📄 Multi-modal context, understands images, PDFs, and live web search results.
🔒 Supports secure tools via container sandboxes (Docker, Podman, or macOS sandbox-exec)
📈 Switches models automatically (e.g., pro to flash) to bypass API rate limits.
📊 Built-in Observability with OTEL traces, metrics, and logs.
💾 Save and resume conversations: persistent history across sessions, isolated per project.
💡 Git-Aware Context, the @ command respects .gitignore, --yolo flag to auto-approve
🤗 Native Windows support, no WSL required and Open Source under Apache 2.0!
Get started `npm install -g @gemini-cli/cli`
The new meaning of "smart terminal"! Check out our Gemini integration for your terminal.
Gemini CLI is our new #opensource AI agent that brings the power of Gemini directly into your terminal!
Access Gemini 2.5 Pro with 1M token context window, 60 requests/min, and 1,000 requests/day—at no cost with a free Gemini Code Assist license → goo.gle/3HW0jL0
Very cool thread about the CS336 Language Models from Scratch course at Stanford taught by @percyliang et al.
Makes me wish I was a student again!
Wrapped up Stanford CS336 (Language Models from Scratch), taught with an amazing team @tatsu_hashimoto @marcelroed @neilbband @rckpudi. Researchers are becoming detached from the technical details of how LMs work. In CS336, we try to fix that by having students build everything:
When I was in grad school @uwcse, I took an awesome "Build An Optimizing Compiler From Scratch" class from Craig Chambers & it convinced me to do my PhD work on compiler optimizations (advised by Craig!).
Course project: implement compiler+4 of these 10 opts (our group did 9!)
I'm guessing some of the students in @percyliang's CS336 class had a similar experience: getting to dive deeply into an area and now want to work on it even after the course.
Building on more than 10 years of robotics research and engineering at @GoogleDeepMind, @GoogleResearch and @GoogleAI, we're delighted to announce our Gemini Robotics On-Device system. A really capable vision-language-action model that can run entirely without network access. ⬇️
We’re bringing powerful AI directly onto robots with Gemini Robotics On-Device. 🤖
It’s our first vision-language-action model to help make robots faster, highly efficient, and adaptable to new tasks and environments - without needing a constant internet connection. 🧵
Jeff Dean reposted
We’re announcing the 87 professors selected for the 2025 Google Research Scholar Program — join us in congratulating these exceptional recipients and learn more about their groundbreaking work at goo.gle/rs-recipients.
#googleresearch #googleresearchscholar
Jeff Dean reposted
Error correction is a key component of tomorrow’s large-scale quantum computers. Today we’re excited to report the experimental demonstration of a color code system that provides an advantageous alternative to the well studied surface code. Check it out →goo.gle/4ljZwli
Jeff Dean reposted
Improving tropical cyclone prediction can help communities better prepare & stay safe. Developed with @GoogleDeepMind, we're launching Weather Lab to share our AI weather prediction models, starting with the new experimental cyclone prediction model. 🧵
Jeff Dean reposted
Congratulations to the authors of "Going Deeper with Convolutions", recipient of the 2025 Longuet-Higgins Prize, which celebrates a paper released 10 years ago that has had a significant impact in the field of computer vision! arxiv.org/abs/1409.4842
We've had an account for @GoogleResearch for a while, but we're going to start posting more info about the work done by Google Research here. Follow for awesome research content!
Welcome to the new home for Google Research on X! To learn about the latest research from Google, read the most up-to-date Google Research Blog posts, or get details on our participation at research-focused conferences, you’re in the right place — follow us here @GoogleResearch!
Jeff Dean reposted
This week's installment of Math, Revealed -- Let's Get Packing! Please retweet this gift link, which lets people see the article for free.
www.nytimes.com/interactive/2025/06/23/science/math-strog…
Separately from my Google role, I'm excited to help @LaudeInstitute identify and fund non-profit computer science research to have a major impact in the world, with @andykonwinski and fellow board members @jpineau1 and Dave Patterson.
Today, I’m launching a deeply personal project. I’m betting $100M that we can help computer scientists create more upside impact for humanity.
Built for and by researchers, including @JeffDean & @jpineau1 on the board, @LaudeInstitute catalyzes research with real-world impact.
Jeff Dean reposted
We just shipped major updates to the Gemini API caching infrastructure 🚢
- Video's that hit the cache are 3x faster TTFT
- PDF's that hit the cache are 4x faster TTFT
- Closed the speed gap on implicit vs explicit cache
- Ongoing work for large Audio files
Index freshness is something I and many others at Google have worked on for many years.
If you are using search tools offered by closed LLM providers (e.g., Anthropic, OpenAI), be aware that these companies use their own search indexes, which are not updated in real-time or even regularly. As a result, you may occasionally encounter 404 links, which can tarnish the reputation of your downstream app.
The only exception to this is Gemini, which returns results from Google's index which is updated in real-time.
A set of slides from a talk I gave at WSDM 2009 describes some aspects of our search system evolution up to then:
static.googleusercontent.com/media/research.google.com/en…
Video of talk: videolectures.net/videos/wsdm09_dean_cblirs
Jeff Dean reposted
On Sunday I traveled to the middle of the desert to capture this: The ISS against our sun. What I didn't expect: the sun producing a magnificent flare at the same time
A once-in-a-lifetime shot I'm thrilled to share with you. See the uncropped shot or get the print in the reply
Jeff Dean reposted
Huge congratulations to @annadgoldie on receiving her @Stanford PhD today! It’s been a great journey!
Jeff Dean reposted
Celebration of the life of Atul Butte. Join colleagues from three universities to celebrate him at 10 am PST on Saturday.
Check out the speed and capabilities of our new Gemini 2.5 Flash Lite model ⬇️
Hello Gemini 2.5 Flash-Lite! So fast, it codes *each screen* on the fly (Neural OS concept 👇).
The frontier isn't always about large models and beating benchmarks. In this case, a super fast & good model can unlock drastic use cases.
Read more: blog.google/products/gemini/gemini-2-5-model-family-expands
Very exciting to see our Gemini 2.5 Pro and 2.5 Flash models become "generally available" (which gives long term support commitments without changing the model), as well as a preview release of a new 2.5 Flash Lite model that offers very low latency & pricing for many uses! 🎉
🎉 It's a BIG day for Gemini 2.5
— 2.5 Flash and 2.5 Pro are now stable and generally available in AI Studio, Vertex AI, and the @GeminiApp
— We're launching a preview of the new 2.5 Flash-Lite, our most cost-efficient and fastest 2.5 model yet
More info on each model below ⬇️
Congrats to everyone on the Gemini team who has worked very hard to get these models to this point! 🙏
Very sad to hear this news. Every time I met Atul, I was impressed with his knowledge and passion for the work he did. Truly a great loss. 😢
With deep sadness, we share the passing of Dr. Atul Butte—a visionary leader, scientist, and mentor. His groundbreaking work and boundless compassion transformed medicine and inspired so many. His legacy will continue to shape the University of California and beyond.
Jeff Dean reposted
Be still, my heart. Little Rock, Arkansas turning out in unbelievable numbers.
Jeff Dean reposted
google simply does not get enough credit for the TPU
the amount of conviction & foresight it took to fund and build hardware specifically for for AI… in 2015…
ten years later, literally everyone else is dependent on nvidia. except big G
really a brilliant strategic move
Jeff Dean reposted
no kings was the largest political protest in u.s. history
Jeff Dean reposted
There is no both sides to this one folks. The government is saying the senator “lunged” at the Secretary and did not identify himself. We have a video where we can see and hear the opposite.
Jeff Dean reposted
We are hiring Research Scientists for our Machine Learning and Optimization team at Google DeepMind Bangalore.
If you're passionate about cutting-edge AI research and building efficient, elastic, customized, and safe LLMs, we'd love to hear from you.
We are looking for candidates with a PhD and a strong demonstrated record of ideating and executing deep research projects.
If interested, please apply here: job-boards.greenhouse.io/deepmind/jobs/6890135
cc: @ManishGuptaMG1
Jeff Dean reposted
Days after NOAA released figures showing the highest seasonal concentration of CO2 in recorded history, the EPA moved to roll back emission standards for power plants. "The EPA is trying to get out of the climate change business,” says one expert. www.wired.com/story/the-epa-is-trying-to-roll-back-emissi…
Our work on AI to improve cyclone prediction is likely to be quite impactful & save lives. We've been partnering w/ @NWSNHC, @metoffice & others for testing.
This builds on years of work on improving general weather prediction using AI from @GoogleDeepMind & @GoogleResearch.
We’re using AI to improve cyclone prediction. 🌀
Introducing Weather Lab: a new interactive platform developed with @GoogleResearch, hosting our experimental AI weather model which can predict a cyclone’s track, intensity, size and structure.
Here’s how it works. 🧵
You can play with the preview version at
deepmind.google.com/science/weather?zoom=2.5&lat=0&lng=60…
More details in the blog post at:
deepmind.google/discover/blog/weather-lab-cyclone-predict…
Research paper:
storage.googleapis.com/deepmind-media/DeepMind.com/Blog/h…
Jeff Dean reposted
NEW analysis by the Congressional Budget Office shows the "One Big Beautiful Bill":
1.) Makes the bottom 30% worse off
2.) Finances tax cuts that make the top 10% in particular better off.
This is not shared sacrifice--it's a fiscal transfer from the bottom to the top.
No longer the mane man.
This lion accepts his receding hairline with peace and dignity
Jeff Dean reposted
This training device to help football players aim their kicks.
Jeff Dean reposted
Cost and Time:
This took about 300–400 generations to get 15 usable clips. One person, 2-3 days.
That’s a 95% cost reduction vs traditional ads.
It turns out video generation tools like our Veo 3 are getting pretty darned good! You might even say they are commercially interesting.
Kalshi hired me to make the most unhinged NBA Finals commercial possible.
Network TV actually approved this GTA-style madness 🤣
High-dopamine Veo 3 videos will be the ad trend of 2025.
Here’s how I made it in just TWO DAYS 👇🏼 (Prompt included)
Check out the 999 open models that Google has released on @huggingface:
huggingface.co/google
(Comparative numbers: 387 for Microsoft, 33 for OpenAI, 0 for Anthropic).
Hat tip to @marlene_zw's post which made me aware of the data from @julien_c's post on LinkedIn (?).
Great post from Julien comparing open releases on Hugging Face❤️✨
Google vs Microsoft vs OpenAI vs Anthropic
Very few companies shipped more open source than Microsoft this past year, but no one wants to talk about that.
Jeff Dean reposted
PSA: Our team inside Google Deepmind is hiring AI developer PM's, if you want to create the future of what DevX, API's, and Agentic experiences look like, pls DM me : )
Jeff Dean reposted
We’re at CVPR! 🎉
Join us in Room 105A at the Music City Center (Convention Center)!
Jeff Dean reposted
We're excited to launch Scouts — always-on AI agents that monitor the web for anything you care about.
If you're ever in Athens, I highly recommend the Museum of Ancient Greek Technology!
maps.app.goo.gl/sngwM98ueKkrooVy5
Early humanoid robots to pour water+wine, demonstrations of heat+steam+water weight to "magically" open 500 kg temple doors, armor, ...
(cont)
Plus early cryptography cipher machines, medical instruments, early musical instruments and more.
The guides were awesome and did live demonstrations.
Btw, most of the things I mentioned above were from ~third to second century BC. And I didn't even mention the flamethrowers, automated coin-operated vending machine, early cataracts operation tool, atmospheric-pressure-based recurring fountain, long-distance shared-phrase-dictionary-based communication system ("we won", "send reinforcements", "retreat", ...) using fire and a water tank, first keyboard instrument, and first tic-tac-toe-like game.
Jeff Dean reposted
My new #math series in the New York Times, "Math, Revealed," is aimed at everyone, whether you love math or not. Have a look! You can read it here for free without a subscription. www.nytimes.com/interactive/2025/06/09/science/math-strog…
Jeff Dean reposted
1/ A quick recap from another week of rapid shipping:
Great response to our Gemini 2.5 Pro update now in preview. Find it in AI Studio, Vertex + @GeminiApp.
And good news for @GeminiApp Pro plan members: your 2.5 Pro limit doubled from 50 to 100 queries per day.
Sad to hear about Bill Atkinson's passing.
When I was a kid and we got our first original 128k Mac, I absorbed the programming documentation and marveled at the amazing breadth of functionality crammed into the QuickDraw library that he wrote.
www.nytimes.com/2025/06/07/technology/bill-atkinson-dead.…
Sounds not too difficult 😀
Ok @GoogleDeepMind gemini-2.5-pro-preview-06-05, let's see if you can write a script to organize and classify a 25,000+ sound library for my client 👀
Jeff Dean reposted
Gemini 2.5 Pro is getting better based on your feedback, and we hope to make this version our stable release soon!
The new 2.5 Pro Preview improves benchmarks and:
1) Introduces budgets for more control
2) Is a more well-rounded model with better creativity, style, & formatting
Jeff Dean reposted
Our latest Gemini 2.5 Pro update is now in preview.
It’s better at coding, reasoning, science + math, shows improved performance across key benchmarks (AIDER Polyglot, GPQA, HLE to name a few), and leads @lmarena_ai with a 24pt Elo score jump since the previous version.
We also heard your feedback and made improvements to style and the structure of responses. Try it in AI Studio, Vertex AI, and @Geminiapp. GA coming soon!
Jeff Dean reposted
For @GeminiApp Pro plan members, we've just doubled your 2.5 Pro limit, from 50 to 100 queries per day. Thanks for using the model so much and messaging us wanting more! Enjoy!
A nice discussion about how our open buildings dataset is being used to enable healthcare delivery in Africa
blog.google/intl/en-africa/company-news/technology/how-op…
Jeff Dean reposted
Last Friday, we shipped Veo 3 to 71 new countries, Pro members, and Ultra members got more credits. All week we've been scrambling to keep everything up and running - way, way, way more demand than we expected!
Today, 2 more updates:
+ The UK now has Veo 3 access 🇬🇧
+ Pro and Ultra subscribers can use Veo 3 in @GeminiApp on Android and iOS (not just the web)
For Pro and Ultra members, try it here:
➡️ For quick, everyday access: gemini.google
➡️ For AI filmmakers: flow.google
These updates should show up today/tomorrow/Sunday if all goes according to plan. :)
Jeff Dean reposted
Amidst the massive demand for Gemini 2.5 and Veo 3 models, wanted to also give a big shout out to our world-class infrastructure, chip and SRE teams, who work tirelessly to keep our wonderful TPUs from melting, and without whose incredible work none of this would be possible.
The demand for Veo 3 has been off the charts! Millions of videos have been generated in the past few days alone, now available on mobile, and in more countries (including the UK! 🇬🇧). Try it out in the @GeminiApp (and the Flow tool for filmmakers at flow.google)
Last Friday, we shipped Veo 3 to 71 new countries, Pro members, and Ultra members got more credits. All week we've been scrambling to keep everything up and running - way, way, way more demand than we expected!
Today, 2 more updates:
+ The UK now has Veo 3 access 🇬🇧
+ Pro and Ultra subscribers can use Veo 3 in @GeminiApp on Android and iOS (not just the web)
For Pro and Ultra members, try it here:
➡️ For quick, everyday access: gemini.google
➡️ For AI filmmakers: flow.google
These updates should show up today/tomorrow/Sunday if all goes according to plan. :)
Jeff Dean reposted
If you had 15min to tell thousands of Berkeley CS/Data/Stats grads what to do with their lives, what would you say? Last Thursday I told them to RUN AT FAILURE.
Afterwards, while we were shaking hands & taking selfies, hundreds of them told me that they are excited to go fail. I don't think they’ve been given permission to do that before.
Jeff Dean reposted
We’ve been amazed by the excitement around Veo 3, which brings in a new era for video generation. It's state of the art in text-to-video, image-to-video, text-to-audio+video generation, and realistic physics.
Veo 3 generates videos with audio – from background noises to dialogue, excelling at accurate lip syncing all from a single text prompt.
And now, we are expanding access to Veo 3 in 71 new countries. And Pro subscribers can now try it in both the Gemini app and in Flow, our new AI-filmmaking tool. Can’t wait to see what you create. ✨
Jeff Dean reposted
Veo 3 dropped about 100 hours ago, and it's been on 🔥🔥🔥 ever since
Now, we’re excited to announce:
+ 71 new countries have access
+ Pro subscribers get a trial pack of Veo 3 on the web (mobile soon)
+ Ultra subscribers get the highest # of Veo 3 gens w/ refreshes
How to try it…
➡️ Gemini (gemini.google):
* Great for everyone - click the Video chip in the prompt bar, and just describe your video
* Pro subscribers now get a 10-pack so you can try it
* Ultra: MAX limits, daily refresh!
➡️ Flow (flow.google):
* Great for AI filmmakers
* Pro: 10 gens/month
* Ultra: Now 125 gens/month (up from 83)!
Gemini 2.5 Flash is a high quality model and it's also... quick as a flash. Quick in more ways than one is the best kind of quick.
There was a lot of announcements at IO, easy to overlook the new 2.5 Flash.
It's pushing new boundaries in capability vs speed!
More exploration at inference time gives our Gemini 2.5 Pro Deep Think model the capability to solve quite complex problems. ⬇️
Watch Gemini 2.5 Pro Deep Think tackle the challenging "catch a mole" problem from @Codeforces. 🪤
This new mode is based on our research in parallel thinking and considers multiple hypotheses before responding.
See it in action ↓
Jeff Dean reposted
Yesterday at #googleio, we introduced Gemini 2.5 Pro Deep Think 🧠, pushing the frontiers of AI reasoning. This enhanced reasoning mode is built to tackle drastically complex problems – like USAMO problems that stumped previous models.
Super proud of the GDM team for this one!
Some great progress on open versions of Gemma models, whether you care about medical applications, American Sign Language, or dolphin. 🩺🤟🐬
We’re excited to share about MedGemma, SignGemma, and DolphinGemma 🧵↓
Here's one good sign: there's been some really nice progress on adapting our open Gemma multimodal models to translate American Sign Language to English.
2️⃣SignGemma is a sign language understanding model that’s coming later this year 🤟🏼It’s a massively multilingual model that’s best at translating ASL into English text, enabling further development of tech access for Deaf and Hard of Hearing users. 🧏
Share your feedback and interest in early testing here: goo.gle/SignGemma
Jeff Dean reposted
Say goodbye to the silent era of video generation: Introducing Veo 3 — with native audio generation. 🗣️
Quality is up from Veo 2, and now you can add dialogue between characters, sound effects and background noise.
Veo 3 is available now in the @GeminiApp for Google AI Ultra subscribers in the U.S.
#googleio
Jeff Dean reposted
CS&E welcomed alumnus @JeffDean back to campus for a special seminar. The talk titled, "Modern Advances in Machine Learning and What Will They Enable?" highlighted advances that have led to the capabilities of modern AI models.
Find the seminar here: z.umn.edu/JeffDeanSeminar
Jeff Dean reposted
Say hello to the new generative media experience in Google AI Studio, bringing together Veo 2, Gemini 2.0 native image generation / editing, and now Imagen 3 (with more to come). 📺
All of these models are free to try in AI Studio and ready for devs to build with in the API!!
Jeff Dean reposted
We just rolled out a new built-in usage dashboard in Google AI Studio 📈, you can now easily check request + token volumes and spending (with much more coming soon)!!
Thanks for having me, @UMNAlleyne! That was fun, and congrats to all the @UMNews graduates!
Two Deans and 1100 graduates ⬇️
The end of the commencement ceremony.
YouTube link:
Sure, here you go. My talk starts at about 1:20:00.
I highly recommend the talk by the student speaker Rimika Dhara starting at 1:07:00. She was great!
www.youtube.com/live/fApD-qZLMiA
Sure, here you go. My talk starts at about 1:20:00.
I highly recommend the talk by the student speaker Rimika Dhara starting at 1:07:00. She was great!
www.youtube.com/live/fApD-qZLMiA
I'm looking forward to being the speaker at Thursday's University of Minnesota College of Science and Engineering commencement ceremony!
Congratulations to all the graduates!
🎓🎉
@UMNews
Had a nice time strolling around campus and finding some old haunts today
Bringing up one level for visibility:
Yes, and that's my wife! We met freshman year! In fact, Thursday's commencement speech will feature the tale of our first date. It involves trying new things, dancing, failure and Goldie (and nine pounds of sweat, but that got edited out after the first draft).
Yes, and that's my wife! We met freshman year! In fact, Thursday's commencement speech will feature the tale of our first date. It involves trying new things, dancing, failure and Goldie (and nine pounds of sweat, but that got edited out after the first draft).
This is cool!
You can click "Copy Canvas" in the lower right (after logging in) and enable voice activation.. (volume up for video below)
@OriolVinyalsML tagging you!
Thank you @davemesserx!
Today's update of Gemini 2.5 Pro Preview is amazing at code web dev, editing, and transformation -- it is now #1 on the WebDev arena, with an ELO score of 1420.
That means you can build super cool, visually stunning web apps, with just a single prompt!
Check out this example, where we explore the world using this app we created in Gemini's Canvas using the Maps API.
Try it at:
g.co/gemini/share/3e2c4258a61f
I'm looking forward to visiting the University of Minnesota (my alma mater!) and giving a talk in the CS department tomorrow morning (Wednesday)!
CS&E Seminar: Jeff Dean
Modern Advances in Machine Learning | Department of Computer Science & Engineering share.google/6Dj23MwwXMi2hqRs9
Jeff Dean reposted
Introducing the AI Futures Fund, a new program where startups can work with us to build the future of AI technology. From early access to @GoogleDeepMind models to Cloud credits, we’ll help get your startup to the next level.
Learn more and apply ⬇️ labs.google/aifuturesfund
Jeff Dean reposted
Gemini 2.5 Pro literally turned a YouTube video into an interactive learning app.
Caching of tokens when you use the same input context repeatedly now happens implicitly, making things a lot easier to deal with.
We just shipped implicit caching in the Gemini API, automatically enabling a 75% cost savings with the Gemini 2.5 models when your request hits a cache 🚢
We also lowered the min token required to hit caches to 1K on 2.5 Flash and 2K on 2.5 Pro!
Some nice examples of what you can do with Gemini 2.5 Pro's video capabilities in the thread. Account then: there's a new 66-tokens-per-frame mode (instead of 258 tokens) that allows putting more than 6 hours of video (at 1 fps) in the 2M token context.
Thrilled to share our latest advances in video understanding 📽️: Gemini 2.5 Pro is a truly magical model to play with, excelling in traditional video analysis and unlocking new use cases I could not imagine a few months ago🪄
More in 🧵 and @Google blog: developers.googleblog.com/en/gemini-2-5-video-understanding
Jeff Dean reposted
Gemini 2.5 Pro (05-06) is SOTA at most video understanding tasks (by a large margin) 📽️. Lots of work by the Gemini multimodal team to make this happen, excited to see developers push this capability in new ways.
More details below!
Jeff Dean reposted
Today we are rolling out an updated Gemini 2.0 Image Generation model with 🖼️:
- Better visual quality
- More accurate text rendering
- Lower block rates
- Higher rate limits
- $0.039 per image generated
So excited to get this into the hands of devs : )
The new Gemini 2.5 Pro model is really fun to play with for coding use cases. Check out these nice examples of things people at @Google have been creating.
Here's what our team at @Google is creating in Canvas with Gemini 2.5 Pro (I/O edition) today 🧵
Jeff Dean reposted
🚨Breaking: @GoogleDeepMind’s latest Gemini-2.5-Pro is now ranked #1 across all LMArena leaderboards 🏆
Highlights:
- #1 in all text arenas (Coding, Style Control, Creative Writing, etc)
- #1 on the Vision leaderboard with a ~70 pts lead!
- #1 on WebDev Arena, surpassing Claude for the first time
This is the first-ever sweep across text, vision, and WebDev by any model!🥇
Huge congrats to @GoogleDeepMind on this incredible breakthrough!
We’re releasing an updated Gemini 2.5 Pro (I/O edition) to make it even better at coding. 🚀
You can build richer web apps, games, simulations and more - all with one prompt.
In @GeminiApp, here's how it transformed images of nature into code to represent unique patterns 🌱
Check out this example video of Gemini 2.5 Pro creating simple apps from "napkin sketches" of a UI.
Very excited to share the best coding model we’ve ever built! Today we’re launching Gemini 2.5 Pro Preview 'I/O edition' with massively improved coding capabilities. Ranks no.1 on LMArena in Coding and no.1 on the WebDev Arena Leaderboard.
It’s especially good at building interactive web apps - this demo shows how it can be helpful for prototyping ideas. Try it in @GeminiApp, Vertex AI, and AI Studio ai.dev
Enjoy the pre-I/O goodies !
Jeff Dean reposted
Ahead of I/O, we’re releasing an updated Gemini 2.5 Pro! It’s now #1 on WebDevArena leaderboard, breaking the 1400 ELO barrier! 🥇
Our most advanced coding model yet, with stronger performance on code transformation & editing. Excited to build drastic agents on top of this!
Jeff Dean reposted
Worth noting that today's Gemini 2.5 Pro update also fixes a number of issues people have noticed with function calling in the 03-25 initial release.
Gemini 2.5 Pro just got an upgrade & is now even better at coding, with significant gains in front-end web dev, editing, and transformation.
We also fixed a bunch of function calling issues that folks have been reporting, it should now be much more reliable. More details in 🧵
Jeff Dean reposted
until this morning, i didn't realize that "tmlr.org" domain was having an issue due to the lack of https support. i guess browsers changed their behaviours, and chrome/edge was taking forever to retry with http and get redirected to jmlr.org/tmlr.
thanks to Gemini 2.5 Pro, i was able to resolve this issue using cloudflare within 10m.
Today we’re sharing an early look at our latest Gemini update for I/O!
Introducing the updated Gemini 2.5 Pro (I/O edition), which ranks #1 on WebDev Arena and surpasses our previous 2.5 Pro model by +147 Elo points. 🏆
blog.google/products/gemini/gemini-2-5-pro-updates
Explore the capabilities of the updated Gemini 2.5 Pro, now in Preview in AI Studio and Vertex AI, and experimentally in the Gemini app.
aistudio.google.com
We look forward to your feedback! 🙏
Here's the WebDev leaderboard:
web.lmarena.ai/leaderboard
Jeff Dean reposted
Listen to the head of the Port of Los Angeles.
No single person is getting more real time data about the economy than this guy.
He says we are fucked.
Jeff Dean reposted
Gemini 2.5 Pro has really changed the LLM game.
People are coming up with wild use cases. There's a shift.
10 examples + use cases:
1. Vibe code morphing particles
Jeff Dean reposted
Last 5 weeks at @NotebookLM:
+ Mind Maps, magical tool for visualizing sources
+ Discover Sources, letting you assemble a knowledge base on any topic in seconds
+ Audio Overviews in 70+ languages
+ Pre-registration for iOS/Android mobile app, coming May 20
+ Gemini 2.5 Flash
It's been a busy week for us. SO busy that we forgot to mention that @NotebookLM is officially powered by Gemini 2.5 Flash ⚡️⚡️⚡️
The 2.5 models are thinking models, so you should start to see more comprehensive answers, particularly to complex, multi-step reasoning questions.
Have you been longing for more insight into how long context works in Gemini models? Check out this podcast discussion with @SavinovNikolay and @OfficialLoganK ⬇️
Dive into long context in the latest Release Notes podcast 🎙
@GoogleDeepMind’s @OfficialLoganK and @SavinovNikolay discuss scaling context windows into the millions, recent quality improvements, RAG versus long context, and more.
Key moments:
0:52 Introduction & defining tokens
5:27 Context window importance
9:53 RAG vs. Long Context
14:19 Scaling beyond 2 million tokens
18:41 Long context improvements since 1.5 Pro release
23:26 Difficulty of attending to the whole context
28:37 Evaluating long context: beyond needle-in-a-haystack
33:41 Integrating long context research
34:57 Reasoning and long outputs
40:54 Tips for using long context
48:51 The future of long context: near-perfect recall and cost reduction
54:42 The role of infrastructure
56:15 Long-context and agents
Very excited about all of this, especially personalization. A model that knows you will be more helpful for you!
Great software feels like an extension of you. We're building @GeminiApp to be the most PERSONAL, PROACTIVE, and POWERFUL assistant.
👀Here's the strategy:
1/ PERSONAL: The best assistant gets you. It starts by knowing your past chats (launching soon), but we’ll go further: we’ll make it easy for you to bring in all of your @Google context (Gmail, Photos, Calendar, Search, YouTube, etc.) - with your permission! On the team, we call it “pcontext” (personalized context) and we’re testing it internally with our own info already.
2/ PROACTIVE: The best assistant anticipates. @GeminiApp will offer insights and actions before you ask, freeing your mind and time for what truly matters. Less prompting, more flow.
3/ POWERFUL: The best assistant turns your ideas into action. @GoogleDeepMind models (like 2.5 Pro) are exceptional - they can research, orchestrate, and create images, videos, and code. We’re in a new era of models, and a new era of user experiences are coming.
4/ Zooming out, @Google infrastructure & TPUs back all of this. It’s the kind of infra people dream about, but it actually exists here -- and it’s going to let us make all this FREE for everyone to try, especially students.
5/ One request: keep your feedback coming! We try to listen, build, and iterate fast. In the last ~3.5 weeks, we’ve shipped:
* Gemini 2.5 Flash
* Veo 2 for video generation
* Free Gemini plans for all U.S. students (more countries soon!)
* LaTeX support (top request)
* Upload and edit your images (started rolling out yesterday)
6/ @demishassabis and I will unpack all of this (and more) at Google I/O in <20 days! Let’s roll!
Jeff Dean reposted
One of my favorite projects I've led at @Google, as it centers around a topic that I find endlessly fascinating: language learning 🗣️
Little Language Lessons is a collection of three bite-sized learning experiments, all powered by Google’s Gemini AI models
Try it →
Jeff Dean reposted
Excited to share our new paper on Step-Wise Reinforcement Learning (SWiRL), which uses reinforcement learning and synthetic trajectories to improve multi-step reasoning and tool use! (1/8)
Jeff Dean reposted
Excited to release SWiRL: A synthetic data generation and multi-step RL approach for reasoning and tool use!
With SWiRL, the model’s capability generalizes to new tasks and tools. For example, a model trained to use a retrieval tool to solve multi-hop knowledge-intensive question answering tasks becomes significantly better at using Python to solve math problems (and vice versa). As we scale the synthetic data size, the generalization gains continue to improve!
Jeff Dean reposted
This just in... the @NotebookLM hosts have some rather exciting news they'd like to share with you all:
"Write an implementation of this paper using Tensorflow.js and visualize the training with Canvas."
Thanks, Gemini!
Watch Gemini 2.5 Pro implement a landmark Google DeepMind research paper. 🕹️
It codes the reinforcement learning algorithm, visualizes the training live and even debugs errors. ↓
Great set of slides by my Gemini colleague @FeinbergVlad
on scaling considerations in large language models, addressing the fact that the classic "scaling law" work does not take into account inference cost (!), distillation, learning rate schedules, etc.
vladfeinberg.com/2025/04/24/gemini-flash-pretraining.html
Recently had the pleasure of lecturing back at Princeton in a grad seminar. I took the opportunity to cover how scaling laws have evolved since their inception, leaning heavily on great external content from my colleagues @borgeaud_s @jalayrac @jacobaustin132 .
Content in thread
Zaby is smarter than your average bear, thanks to Gemini!
My 7 yr old has been dreaming up an AI teddy bear. So we made him! Zaby is a clever, pedagogical & funny teddy that loves talking math. Powered by Gemini Flash & Google speech recognition/synthesis. His mouth moves in sync with the speech envelope. Smarter than the average bear!
Great to see Spanner recognized with this year's ACM SIGMOD Systems Award "for reimagining relational data management to enable serializability with external consistency at global scale".
I was fortunate to be able to work on the first versions of Spanner with awesome colleagues (e.g. co-authors on first paper below), but it has been great to see many more people push it forward and turn it into a wildly successful robust internal system and Cloud Spanner into a full featured database system at incredible scale. 🎉
You can learn more about Spanner here:
"Spanner: Google's Globally Distributed Database", OSDI 2012, static.googleusercontent.com/media/research.google.com/en…
"Spanner: Becoming a SQL System", SIGMOD 2017, static.googleusercontent.com/media/research.google.com/en…
Cloud Spanner: cloud.google.com/spanner
🏆 The ACM SIGMOD Systems Award 2025 goes to
Spanner: Google’s Globally-Distributed Database
"for reimagining relational data management to enable serializability with external consistency at global scale"
2025.sigmod.org/sigmod_awards.shtml
Fun fact: I helped pick the name Spanner because it would "span" all of our data centers with a single storage system. Later, my British colleagues threw a wrench into my thinking by pointing out that to them, a spanner was, well, a wrench. Leading to this meaning today:
I should also add there was a nice paper published about the clock system that Spanner uses to achieve some of its consistency properties with low latency:
research.google/pubs/sundial-fault-tolerant-clock-synchro…
Someone just reminded me of this lecture I gave in 2009 that described the evolution of Google Search from 1999 to 2009. People who are interested in how our search systems work might find this interesting.
It touches on disk-based serving systems, in-memory indices, compression schemes for inverted indices, latency issues due to interference from background processes, queries of death, evolution of hardware, and more..
Video: videolectures.net/videos/wsdm09_dean_cblirs
Slides: static.googleusercontent.com/media/research.google.com/en…
We never really published any papers about lots of interesting systems work that underlies our search systems, so this relatively little known talk I gave at the Web Search and Data Mining conference in 2009 is the best public source of info.
Jeff Dean reposted
30+ accepted papers
6 oral papers
6 guest speakers
join us at @iclr_conf on the 27th Hall 4 #3 for a full day of workshop on Modularity for Collaborative, Decentralized, and Continual Learning
sites.google.com/corp/view/mcdc2025/home
@derylucio, Fengyuan Liu, and myself will be organizing that day in person
Workshop alert 🚨
We'll host in ICLR 2025 a workshop on modularity, encompassing collaborative + decentralized + continual learning.
Those topics are on the critical path to building better AIs.
Interested? submit a paper and join us in Singapore!
sites.google.com/corp/view/mcdc2025/home
More interesting info about how various computer systems in the airline industry work. Nice talk, @WriteAheadLog!
Also see my talk on how airline reservations systems work, which mentions Carl. I worked at ITA and it was full of smart people like Carl! youtu.be/8_t41xvPp1w
If you want to know more about how Google Flights works, airline tickets, and why it is super complicated to deal with all the constraints and the combinatorial combinations, I highly recommend this set of slides by Carl de Marcken, one of the co-founders of ITA software, which Google acquired and became one of the underpinnings of Google Flights.
(Sorry for the http rather than https Link: Carl's domain doesn't appear to support https)
www.demarcken.org/carl/papers/ITA-software-travel-complex…
The most underrated Google product: Google Flights.
College students, you can now get all kinds of nice things for free! Try out Gemini Advanced, NotebookLM Plus, some exciting disk space, and more!
This school year AND next school year -> Free!
This comes with Gemini Advanced and:
* NotebookLM Plus
* Gemini in Google Docs, Sheets and Slides
* Whisk
* 2TB of storage
Also, we're working on making this available globally. 🌍🌎🌏
Working on it - we need to do a few country-specific things/configs, so don't have a clear ETA yet, but stay tuned! We wanted to ship this part of it since it was ready.
Jeff Dean reposted
We're working to bring this offer to students outside the U.S. Stay tuned.
Working on it - we need to do a few country-specific things/configs, so don't have a clear ETA yet, but stay tuned! We wanted to ship this part of it since it was ready.
Some nice quantization work makes our Gemma 3 model even more compact.
Just announced new versions of Gemma 3 – the most capable model to run just one H100 GPU – can now run on just one *desktop* GPU!
Our Quantization-Aware Training (QAT) method drastically brings down memory use while maintaining high quality. Excited to make Gemma 3 even more accessible for more developers.
I see what you mean. This is normal.
Gemini 2.5 Flash demolishes my Galton Board test, I could not get 4omini, 4o mini high, or 03 to produce this. I found that Gemini 2.5 Flash understands my intents almost instantly, code produced is tight and neat. The prompt is a merging of various steps. It took me 5 steps to achieve this in Gemini 2.5 Flash, I gave up on OpenAI models after about half an hour. My iterations are obviously not exact. But people can test with this one prompt for more objective comparison.
Please try this prompt on your end to confirm:
--------------------------------------------------
Create a self-contained HTML file for a Galton board simulation using client-side JavaScript and a 2D physics engine (like Matter.js, included via CDN). The simulation should be rendered on an HTML5 canvas and meet the following criteria: 1. **Single File:** All necessary HTML, CSS, and JavaScript code must be within this single `.html` file. 2. **Canvas Size:** The overall simulation area (canvas) should be reasonably sized to fit on a standard screen without requiring extensive scrolling or zooming (e.g., around 500x700 pixels). 3. **Physics:** Utilize a 2D rigid body physics engine for realistic ball-peg and ball-wall interactions. 4. **Obstacles (Pegs):** Create static, circular pegs arranged in full-width horizontal rows extending across the usable width of the board (not just a triangle). The pegs should be small enough and spaced appropriately for balls to navigate and bounce between them. 5. **Containment:** * Include static, sufficiently thick side walls and a ground at the bottom to contain the balls within the board. * Implement *physical* static dividers between the collection bins at the bottom. These dividers must be thick enough to prevent balls from passing through them, ensuring accurate accumulation in each bin. 6. **Ball Dropping:** Balls should be dropped from a controlled, narrow area near the horizontal center at the top of the board to ensure they enter the peg field consistently. 7. **Bins:** The collection area at the bottom should be divided into distinct bins by the physical dividers. The height of the bins should be sufficient to clearly visualize the accumulation of balls. 8. **Visualization:** Use a high-contrast color scheme to clearly distinguish between elements. Specifically, use yellow for the structural elements (walls, top guides, physical bin dividers, ground), a contrasting color (like red) for the pegs, and a highly contrasting color (like dark grey or black) for the balls. 9. **Demonstration:** The simulation should visually demonstrate the formation of the normal (or binomial) distribution as multiple balls fall through the pegs and collect in the bins. Ensure the physics parameters (restitution, friction, density) and ball drop rate are tuned for a smooth and clear demonstration of the distribution.
#openai @sama @gdb @ai_for_success @aidan_mclau
I think therefore I am...
confronted with a more challenging problem for which I would like my faithful AI assistant to think a bit harder about how to solve.
Thanks, Gemini!
The thinking budget is honestly very rad to surf the cost/benefit pareto frontier. Very excited to use this for a variety of projects ;)
Jeff Dean reposted
The leading model runnable on a single cloud H100 GPU now fits on a single home GPU! 🔥
We've optimized Gemma 3 27B with QAT so you can run our best-in-class open model on your desktop RTX 3090 or similar.
See how easy it is to try via @ollama! 👇
Jeff Dean reposted
Correction: Yesterday I missed Gemini 2.5 Flash (no‑thinking)
TL;DR: ~95% of the time, Google’s Gemini models offer the best performance-per-dollar
It’s bloody out there, and it’s OpenAI's blood
For the first time, Google is responding to OpenAI's announcement in < 24 hours
The WAR is officially ON, and Google wants the LLM market
Google is now dominating +90% of the price share
Excited to see the release of our Gemini Flash 2.5 model (the hard work of many people here!). This model continues the hallmark of the Flash model variants, with very high model quality with a low price point, and very low latency as well, continuing our push to go up (better quality) and to the right (lower prices).
It also features the ability to control the level of internal "thinking", enabling fine-grained adjustments of quality and cost by developers.
Gemini 2.5 Flash just dropped. ⚡
As a hybrid reasoning model, you can control how much it ‘thinks’ depending on your 💰 - making it ideal for tasks like building chat apps, extracting data and more.
Try an early version in @Google AI Studio → ai.dev
Delighted to have given a talk during the final day of @Khipu_AI's #khipu2025 event in Santiago, Chile ~a month ago.
The video of the final day sessions is now available (my talk starts at 4h13m or so, but there were lots of other awesome speakers and panelists during the final day, so check it all out!):
youtu.be/NApGucxFERo?t=15175
We are so happy to have @JeffDean from @Google, speaking at #khipu2025!!
You can watch live and later re-watch his contributions at Khipu at: khipu.ai/live
We're rendered a little embarrassed, but better LaTeX than never.
About 10 days ago, there was a cry for LaTeX support in @GeminiApp. Last week, we shipped v1. Today, we shipped a v2 update, so LaTeX should work for both 2.0 Flash and 2.5 Pro models in @GeminiApp. If it's not working for some reason, please let us know!
Jeff Dean reposted
Gemini now works in google sheets
A success that Microsoft would also like to have with Office 365
Via Reddit
The recording of the talk I gave earlier this week at @ETH / @CSatETH is now available at:
video.ethz.ch/speakers/d-infk/2025/spring/251-0100-00L.html
Slides are at:
drive.google.com/file/d/12RAfy-nYi1ypNMIqbYHjkPXF_jILJYJP…
Thanks for hosting me, @anaklimovic! And thanks to the many faculty and students I had great discussions with, as well.
For those who didn't catch the talk in person, here is the recording: video.ethz.ch/speakers/d-infk/2025/spring/251-0100-00L.html
Thank you for a great talk and insightful discussions, Jeff!
AI-powered weather forecasts are now on Google Search, bringing high quality forecasts to people across Africa. AI models are able to perform much better than traditional weather forecasts when ground based weather data is sparse, which it is across Africa (North America has 291 radar facilities to track weather, while Africa has 37).
May your future be sunny and bright! Or cloudy, or rainy, or ... But at least you'll know what to expect!
blog.google/intl/en-africa/products/explore-get-answers/n…
Jeff Dean reposted
1/ Today, Veo 2, our state-of-the-art video model, is rolling out to Gemini Advanced + Whisk!
You can create 8s, high-res videos from text prompts in @GeminiApp with fluid character movement + lifelike scenes across a range of styles. Tip: the more detailed your description, the better.
Introducing TxGemma, a family of open models specifically tailored for health settings, building on top of Gemma and Gemini.
developers.googleblog.com/en/introducing-txgemma-open-mod…
In case it's not clear from @paulgauthier's chart below, the cost differences are quite large among the top 10 models on this benchmark, w/ some (lower quality) models being ~2X, ~3X or ~30X more expensive than the Gemini 2.5 Pro model (the website has a nice table, seen below).
Gemini 2.5 Pro's leaderboard entry has been updated with costs, now that it available through a paid API. It cost $6 to run the aider polyglot coding benchmark on Gemini, lower than the top 10 other entries except for DeepSeek's models.
aider.chat/docs/leaderboards/
It was great chatting with you in this session at #googlecloudnext last week, @SabaMugazambi!
Yesterday, I hosted the distinguished @JeffDean during a fireside chat on what's next with Google TPUs & AI hardware specialization during @googlecloud 's #googlecloudnext. Including details of our latest 7th generation TPU, Ironwood!
Check: www.youtube.com/watch?v=fNjH5izFeyw
Jeff Dean reposted
We are blown away by the reception of Gemini 2.5 Pro, and the many positive surprises from benchmarks and community evaluations. That is the ultimate leaderboard!
Breaking news: Google is winning on every AI front.
This is not just about Gemini 2.5 but about a reality that OpenAI and Anthropic fans have ignored for too long. Here's a non-exhaustive list:
- Gemini 2.5 Pro is the best model in the world according to benchmarks, vibe checks, high-taste testers, and firsthand testimonies. It's also fast and cheap compared to similar models (Google offers it for free on the Gemini app!)
- Gemini 2.5 Flash (to be announced soon) is much faster and much cheaper, so it captures perfectly the Pareto frontier of cost-performance of cost-efficient models.
- Gemma 3 is a highly competitive open-source model, as good or better than Llama 4 and DeepSeek models.
- That's just LLMs. Google is world-class in image (Imagen 3), video (Veo 2), voice (Chirp 3), and music (Lyria). They're integrating them all in Vertex AI.
- Deep Research with Gemini 2.5 Pro is *twice as good* as OpenAI's Deep Research, according to human testers. Other agents? Yes: Project Astra (assistant) and Project Mariner (computer interaction)
- They just launched Agent2Agent, compatible and complementary to Anthropic's MCP, which they will build in-house as well.
- And they keep publishing papers in top journals (Nature) and going to the top conferences (ICLR, NeurIPS), whereas others jealously keep their most important stuff for themselves.
- That's just the AI stuff, but Google is also a consumer software company with seven 2+ billion monthly users: Search, YouTube, Gmail, Android, Chrome, Maps, and Play Store
- A hyperscaler (Google Cloud)
- A hardware company (TPUs, Ironwood)
- And a phone company (Pixel).
How can OpenAI or Anthropic or even Meta fight such a beast?
Let’s wait for their responses to this. I’ll be here to cover any newsworthy release—even if I’ve already made my bet on who’s most likely to win.
(Read the full post in the link below.)
Jeff Dean reposted
Human generated data has fueled incredible AI progress, but what comes next? 📈
On the latest episode of our podcast, @FryRsquared and David Silver, VP of Reinforcement Learning, talk about how we could move from the era of relying on human data to one where AI could learn for itself.
Watch now →
00:00 Introduction
01:50 Era of experience
03:45 AlphaZero
10:19 Move 37
15:20 Reinforcement learning and human feedback
24:30 AlphaProof
29:50 Math Olympiads
35:00 Experience based methods
42:56 Hannah's reflections
44:00 Fan Hui joins
Jeff Dean reposted
Neat: I asked Gemini 2.5 to come up with a hard science fiction setting in 2112 and then make a video game about it.
In one shot, it wrote 1,111 lines of code that executed flawlessly for a surprisingly solid game with heat management, multiple weapons, point defenses & enemy AI
Jeff Dean reposted
Our 7th generation TPU, Ironwood, is our most powerful and scalable AI accelerator yet, purpose-built for the era of inference.
With more than 9,000 liquid-cooled chips and 42.5 exaflops of compute per pod, they’re built to serve the world’s most advanced inferential models at global scale while doubling the power efficiency of our previous generation.
We can’t wait to see what AI breakthroughs our customers achieve when Ironwood becomes available later this year.
blog.google/products/google-cloud/ironwood-tpu-age-of-inf…
Come check out some great talks at #googlecloudnext by my Gemini colleagues! ⬇️
Some of the talks from the Gemini team at #googlecloudnext
We have had a talk on visual intelligence by @AniBaddepudi, (ultra) long-context by @SavinovNikolay and tomorrow we'll have a talk on multimodal generation by @m__dehghani, @19kaushiks and I will talk about Thinking!
Nice example of the multimodal OCR capabilities of Gemini 2.5 Pro. Just give it dusty old records and ask for JSON!
climatelabbook.substack.com/p/data-rescue-with-ai
"Which is almost perfect. It has misread a small number of the hard-to-read 8s as 5s. But it is spot-on apart from that. No hallucination.
A revolution is here. I think we will soon be able to rescue historical weather observations data much more efficiently and quickly."
Thanks to @OliviaSerenaG for highlighting the blog post to me!
Jeff Dean reposted
NASDAQ call volume spiked minutes before the 90 day tariff pause was announced.
Not a good look at all.
A powerful features of our Gemini models since Gemini 1.5 (including 2.0 and 2.5 models) is their powerful long context capability (2M tokens serving in production). You can process ~1000 pages of text, hrs of video, 10+ hrs ofaudio. Come see @SavinovNikolay discuss this!
This Thursday 10th April I’m speaking about Gemini's Long Context at #googlecloudnext.
Please join me if you are around! cloud.withgoogle.com/next/25?session=BRK2-215
Jeff Dean reposted
Deep Research with 2.5 Pro is an incredible amount of model inference at your fingertips!
Plus the team are improving the product at a really nice cadence.
I'm finding myself lean on it more and more, to learn about new topics.
Jeff Dean reposted
1/ Really enjoyed kicking off #googlecloudnext in Las Vegas today to unveil our 7th-gen TPU, Ironwood, purpose-built for inference (3,600x performance, 29x efficiency boost vs. our first Cloud TPU). Ironwood is coming later this year.
Jeff Dean reposted
Congratulations to Neil Zeghidour, Alejandro Luebs, Ahmed Omran, Jan Skoglund, and Marco Tagliasacchi for winning the IEEE Best Paper Award for "SoundStream: An End-to-End Natural Audio Codec"! arxiv.org/abs/2107.03312 #spsawards #ieeeawards
Jeff Dean reposted
Ember: an inference-time scaling architecture framework
🧵
(1/8)
Great to see @sundarpichai kicking off #googlecloudnext📷 in Las Vegas by talking about our Ironwood, our 7th generation TPU system (29X more power efficient / operation than our first TPU pod system, and 3600X more performance).
I'm excited to join @SabaMugazambi on Friday at Cloud Next in a conversation about TPUs, specialized hardware, trends in AI and machine learning, and more!
Session: cloud.withgoogle.com/next/25/session-library?session=SPTL300
1/ Really enjoyed kicking off #googlecloudnext in Las Vegas today to unveil our 7th-gen TPU, Ironwood, purpose-built for inference (3,600x performance, 29x efficiency boost vs. our first Cloud TPU). Ironwood is coming later this year.
I'm very much looking forward to visiting @ETH Zürich and @CSatETH next Monday and giving a talk and chatting with faculty and students!
Thanks for hosting me, @anaklimovic!
inf.ethz.ch/news-and-events/events/event-detail.important…
Very excited to host @JeffDean at @ETH Zurich for a Distinguished Colloquium and research discussions next Monday! @CSatETH
Info about his talk "Important Trends in AI: How Did We Get Here, What Can We Do Now and How Can We Shape AI’s Future?":
inf.ethz.ch/news-and-events/events/event-detail.important…
Jeff Dean reposted
Phase 1 of the Google AI Studio UI redesign just landed 🚢
A nice updated look and feel + we are laying the foundational for all of the developer platform things we will land in AI Studio over the next few months : )
Jeff Dean reposted
So I ran my bench of perfect-information-games against the new Gemini 2.5 Experimental and -- it's quite good.
It displaces DeepSeek V3, by a lot. But I also found that reading through the transcripts of its thoughts, they seemed qualitatively different.
Jeff Dean reposted
Deep Research in the Gemini App is now powered by Gemini 2.5 Pro, and our early tests show users prefer this 2:1 vs “other products” ;)
gemini.google.com
A nice example of what the Deep Research feature can do.
g.co/gemini/share/8fc6cf02bb4e
Favorite Deep Research query: "You are tasked with building out a business case for a new robotaxi business in the UK - a JV between waymo and uber. Build a discounted cash flow model for the value of the business after 3 years, assuming they start today". Check it out:
Jeff Dean reposted
@GeminiApp app update:
* Best model (Gemini 2.5 Pro) now powers Deep Research
* 20 reports per day
* 150 countries
* 45+ languages
In my testing, the analysis really shines with 2.5 Pro. Throw it your hardest questions and let us know!
Available for paying customers today.
Jeff Dean reposted
We just announced Geospatial Reasoning, a new research effort that will use generative AI (#gemini) to help unlock novel insights about our world. 🌍
Imagine asking complex Qs about our planet & getting detailed answers w/ plans & visuals! goo.gle/3XOO5sF
Jeff Dean reposted
This was actually very impressive. An assignment in my class is to determine the price & TAM of a product that can give the calorie count of food using a new form of spectroscopy, with $75 COGS.
The new Gemini 2.5 Deep Research got all of that & much more. It used to take hours.
Jeff Dean reposted
Gemini Deep Research now uses Gemini 2.5 Pro, its pretty good. You should try it out on @GeminiApp
The Deep Research feature is now powered by the Gemini 2.5 Pro model, making it even more effective!
Give it a try to generate detailed, easy-to-read reports on just about any research topic at goo.gle/3FYAPLQ!
📣 Deep Research is now powered by Gemini 2.5 Pro, our most intelligent AI model. ✨
This upgraded Deep Research is now even better at:
🔍 Finding & synthesizing information
📊 Providing more insightful reports
🧠 Analytical reasoning
Gemini Advanced users can access the new Deep Research across the web, Android, and iOS devices to generate detailed, easy-to-read reports on just about any research topic: goo.gle/3FYAPLQ
Jeff Dean reposted
It is very hard to find the right balance between usability and performance to help the developer community. For Gemma, we spent time asking for feedback from the community and built based on this feedback. We ended up entirely focused on models at useful sizes, i.e. download any Gemma now and you will have an easy time getting it running on consumer hardware.
We also offer top of the line perf via Gemini (checkout 2.5 Pro!), and we’re continuously figuring out what’s good to add for on-device open models (Gemma), vs what requires cloud scale (Gemini).
Please help us by sending more feedback so we can nail next releases! And as always, you can test all our models @ ai.dev!
Jeff Dean reposted
In Large Language Monkeys, we showed the scaling laws of inference-time compute with repeated sampling--the power law relationship between the number of repeated attempts and the fraction of problems solved!
The following amazing work theoretically proves the necessary and sufficient conditions for this property to hold, which is that the left tail of pass@1 must be a power law itself. It then introduces actionable insights into how to efficiently predict and scale inference capabilities!
Led by @RylanSchaeffer and in collaboration with a great team:
@JoshuaK92829, @jplhughes, @jordanjuravsky
@sprice354_, @aengus_lynch1, @ErikJones313, @_robertkirk & @sanmikoyejo!
Interested in test time / inference scaling laws?
Then check out our newest preprint!!
📉 How Do Large Language Monkeys Get Their Power (Laws)? 📉
arxiv.org/abs/2502.17578
w/ @JoshuaK92829 @sanmikoyejo @Azaliamirh @jplhughes @jordanjuravsky @sprice354_ @aengus_lynch1 @_robertkirk
Jeff Dean reposted
We’re launching Project Astra capabilities in Gemini Live ✨
Chat with @GeminiApp about anything you see 👀 by sharing your phone’s camera or screen during conversations. ↓
Jeff Dean reposted
Love how simple Gemini's new camera/screen share is – just point your camera or share your screen + ask away. Available now for everyone on Pixel 9 & Galaxy S25 users and Gemini Advanced users on Android. Excited to see Project Astra come to life!
📣 It’s here: ask Gemini about anything you see. Share your screen or camera in Gemini Live to brainstorm, troubleshoot, and more.
Rolling out to Pixel 9 and Samsung Galaxy S25 devices today and available for all Advanced users on @Android in the Gemini app: goo.gle/4iR4LrU
See how the team has been using it:
Gemini IP!
TCP/IP Packet Simulator with @GoogleDeepMind Gemini 2.5 Pro. Super helpful and gorgeous for both experts and learners. Can't believe I can make such thing ~ 10 mins.
Jeff Dean reposted
Gemini 2.5 is our most intelligent model + now our most in demand (we've seen an 80%+ increase in active users in AI Studio + Gemini API this month).
So today we’re moving Gemini 2.5 Pro into public preview in AI Studio with higher rate limits (free tier is still available!). We’re also setting usage records every day with the model in the @GeminiApp. All powered by years of investing in our compute infrastructure purpose-built for AI. More to come at Cloud Next, stay tuned.
Jeff Dean reposted
We’re moving Gemini 2.5 Pro to Preview today, giving developers access to increased rate limits to begin testing the model for production-ready apps. It’s available now in Google AI Studio. Learn more ↓ goo.gle/3FVHWF2
Jeff Dean reposted
With gemini 2.5 pro pricing and results, Google has fixed the biggest unknown/weakest link in their lineup
and we can now confirm that @GoogleDeepMind completely owns the pareto frontier down to 1220 elo
(at least until o4-mini and gpt5?)
updated price-elo pareto frontier with deepseek v3/r1 and gemini 2 flash thinking 2 results
notes:
- o1-mini and o3-mini are going to have to DRASTICALLY cut prices (like, at least 25x) to keep up. it is not a surprise that o3-mini is going to be in free chatgpt much sooner than o1-mini curve.
- claude 4 needs at least 75 elo step up to be relevant
- distillation research (like we cover in the @madiator podcast today) meaningfully bends the pareto frontier over time. it is awe-inspiring to plot out and observe.
- 2023's We Have No Moat memo might be experiencing a redux - this time at Meta, and maybe at the other big labs
more in today's subscriber poast
Jeff Dean reposted
The pricing for 2.5 Pro is out.
Here is the pareto performance of price : lmsys elo visualized as a rainbow.
Jeff Dean reposted
Gemini rly having my back this week:
- Deep Research found my best friend's wedding venue ✅
- 2.5 Pro caught a tax return issue saving me thousands ✅
Jeff Dean reposted
Ok Gemini with camera is wild
You can ask questions about anything around you, feels like a superpower.
Ask about any object or monument, translate on the fly, get tailor-made assistance...
Absolute game-changer.
Jeff Dean reposted
Today I'm launching my new company @GeneralAgentsCo and our first product.
Introducing Ace: The First Realtime Computer Autopilot
Ace is not a chatbot. Ace performs tasks for you.
On your computer. Using your mouse and keyboard.
At superhuman speeds!
Jeff Dean reposted
As a longtime "tools for thought" user, I'd seen a lot of mind maps, but I'd never been a big believer.
But in one inspired weekend of prototyping, @HcwXd completely changed my mind.
Mind Maps are now available to all @NotebookLM users. Here's why they're different...
The problem I'd always had with mind maps really came down to two things: they were complicated to build manually, and once you'd assembled them, they weren't all that useful. Like, it was nice to see a diagram of the core concepts in whatever material you were mapping, but once you had that visualization, there wasn't really a next step that helped you explore any of those concepts.
But about four months ago, shortly after he joined the Notebook team, @HcwXd built a mind maps prototype in his spare time. It eliminated the laborious process of manually creating links or tagging with concepts, and just generated an interactive map of all your sources automatically. That was amazing enough.
But the thing that really blew me away was this: every node in the map launched a full NotebookLM query with one click. So you can get an amazing visualization of the core ideas or themes in your sources, but then instantly drill down to get a detailed overview of a specific topic, with citations back to the original source material.
It's a genuinely new way to navigate through a large body of information.
The best way to try it out is to create a new notebook, and add the last 20 or so documents you've written or read recently (PDFs/Docs/Slides/etc.) Then generate a mind map -- you'll have an amazing visualization of where your head has been for the past few weeks.
Jeff Dean reposted
Caitlin Clark went 50-for-54 from 3 ☔️🔥
We're using a ReLU to set tariffs?
oops, fixed. Now we get that nice straight line
Jeff Dean reposted
Happy birthday, @gmail! 🥳 We have a surprise: Enterprise users can send end-to-end encrypted (E2EE) messages to any user on any email inbox with just a few clicks (despite the timing, this is not a joke). → goo.gle/3E1eX1T
This video is a really nice story about how a sophisticated AI model like Gemini can help make sense of complex information about a rare disease, leading to better understanding for all.
When his son Max was diagnosed with a rare disease, our colleague Thomas turned to Gemini to better understand complex scientific papers — and even make new connections across the research community.
His story is a powerful reminder of why we work on AI. ↓
Jeff Dean reposted
Google DeepMind has released a new flagship model, Gemini 2.5 Pro.
We evaluated it on GPQA Diamond, and found a score of 84%, exactly matching the result reported by Google. This is the best result we have found on this benchmark to date!
Jeff Dean reposted
Introducing API. A new era of agentic computer use begins today.
Jeff Dean reposted
If you're using Gemini Pro 2.5 for code, there's ths neat feature called branching where you can create a branch over and again out of the results. Pretty useful :)
People who built branching into AI Studio are the real heros. It's been so smooth to feed an entire repo into Gemini 2.5 and branch again and again.
Jeff Dean reposted
So Gemini 2.5 Pro is now 100% free?!
You can literally use the world's best AI model and one-shot web apps.
All you need is a free Google account:
- Select 2.5 Pro (experimental) from the list
- Activate the Canvas option at the bottom (advanced)
- Send your prompt and enjoy
Example you can reuse:
"Develop a fully functional customer relationship management (CRM) web application tailored for a mid-sized healthcare provider. The application should streamline patient communication, appointment scheduling, record management, and billing processes. As a team of experienced full-stack web developers specializing in healthcare technology solutions, you will design and build this application. The application should feature a user-friendly interface with distinct dashboards for administrators, doctors, and staff. Include functionalities such as patient profiles with medical history, a messaging system, insurance verification, and report generation on key performance indicators like patient acquisition and appointment adherence. The front-end should be built using React. Maintain a professional, solution-oriented, and technically precise tone throughout the development process."
We're animatedly peddling pelicans now, without billing!
Gemini 2.5 coded a pelican riding a bicycle. 🚲 Uses @p5xjs. Inspired by @simonw's classic SVG test. Gemini Canvas link: gemini.google.com/share/c82738535cf7 Details in 🧵
Jeff Dean reposted
Gemini 2.5 Pro is available to all Cursor users! You can enable the full 1M context window if you'd like.
We're curious to hear how you think it compares to Sonnet.
Jeff Dean reposted
Jeff Dean reposted
If you claim that adding road capacity is a solution, you are part of the problem.
Jeff Dean reposted
MathArena results for gemini-2.5-pro
I guess my terrestrial thoughts of sunsets never considered there might be nice color gradients from this orientation.
The line between day and night as seen from Space
Jeff Dean reposted
Gemini 2.5 Pro Thinking scored around 51.6% on AI Explained' Simple-Bench. It's the first model to score above 50%. ~double where we were a year ago
www.youtube.com/watch?v=kTslCsPBGHw
Jeff Dean reposted
On today's Wordle, the new Gemini model completely crushed the competition. It logicially deducted diverse words, found the correct spots of valid and invalid letters and got a result quickly. Sonnet proposed multiple invalid words in the end, so DNF
Jeff Dean reposted
Lots of Gemini developer updates today:
- Bumped 2.5 Pro limit to 20 RPM (billing enabled)
- Added "Get code" to Live mode in AI Studio
- Made "Open in Colab" include your prompts / settings
- Added history toggle + full screen for "Get code"
Jeff Dean reposted
Also starting today, Gemini Advanced users can create with Canvas on Gemini 2.5 Pro.
This model’s advanced coding abilities make it easy for anyone to collaborate, create, and share imaginative prototypes (like this example of a 3D dino world, made in under 5 minutes with just a few prompts).
Here’s how to get started:
1) Go to gemini.google.com
2) Choose the “2.5 Pro (experimental)” model
3) Select “Canvas” and start prompting
We’re working hard to make Gemini 2.5 Pro available for more people to try soon–watch this space for updates.
Jeff Dean reposted
Wow wow, Gemini 2.5 is absolutely crushing it on the Thinking Benchmark with tough questions!
Not only is this model getting the highest score on these I've tested so far, but also, look at that incredible latency difference! This model is SO much faster than DeepSeek R1 & o3
Jeff Dean reposted
Wow we just ran Gemini 2.5 Pro on our evals and it got a new state of the art. Congrats to the Gemini team!
Sharing preliminary results here and working on bringing it into Devin:
Our best model*, Gemini 2.5 Pro, is now available for everyone in the Gemini app model drop down menu at gemini.google.com. Give it a try with your most difficult questions!
*For now! 😃
Gemini 2.5 Pro is taking off 🚀🚀🚀
The team is sprinting, TPUs are running hot, and we want to get our most intelligent model into more people’s hands asap.
Which is why we decided to roll out Gemini 2.5 Pro (experimental) to all Gemini users, beginning today.
Try it at no cost at gemini.google.com
Jeff Dean reposted
Gemini 2.5 Pro is a really incredible model, you can really feel the step change when you interact with it (seeing its thoughts are fun too!). Try it now and see for yourself in AI Studio ai.dev and Gemini Advanced @GeminiApp!
Jeff Dean reposted
✨ What makes Gemini 2.5 Pro stand out? In this Release Notes episode, Sr. Product Manager @OfficialLoganK and Gemini Product Lead @tulseedoshi break down its reasoning, coding, and multimodal strengths, plus a 1M token long context. ↓
Timecodes:
1:05 Gemini 2.5 launch overview
3:19 Academic evals vs. vibe checks
6:19 The jump to 2.5
7:51 Coordinating cross-stack improvements
11:48 Role of pre/post-training vs. test-time compute
13:21 Shipping Gemini 2.5
15:29 Embedded safety process
17:28 Multimodal reasoning with Gemini 2.5
18:55 Benchmark deep dive
22:07 What’s next for Gemini
24:49 Dynamic thinking in Gemini 2.5
25:37 The team effort behind the launch
Jeff Dean reposted
@GoogleDeepMind really cooked with Gemini 2.5!
Had been trying to get @cursor_ai to fix a tricky bug in a complex Agent I’m building that has to orchestrate multiple tools and their responses.
Spent two days without success, switched to Gemini 2.5 and it figured it out!
Great to see AI-based weather prediction being rolled out across Africa. This is important because traditional weather forecasts rely much more on ground-based radar, and in Africa, there simply aren't enough to give high quality traditional forecasts.
blog.google/intl/en-africa/products/explore-get-answers/n…
Jeff Dean reposted
We are seeing a huge amount of demand for Gemini 2.5 Pro right now and are laser focused on getting higher rate limits into the hands of developers ASAP.
This is the #1 priority right now, stay tuned : )
Jeff Dean reposted
Gemini 2.5 Pro available in Cursor!
YES!!!!
Jeff Dean reposted
Excited to share a sneak peek into what we've been building at Yutori!
What you see below is our trained model and internal prototype — multiple agents running in parallel in the background, completing tasks of varying complexity, relevant information and cues to step in being surfaced to the user.
More examples 👇
This is barely scratching the surface of what agents can do for you day-to-day.
Follow along at @yutori_ai — more to come soon!
Jeff Dean reposted
One year ago, @abhshkdz and I left Meta to start Yutori.
Ten months ago, @DhruvBatraDB joined us :)
Nine months ago, we crystallized our vision.
Two months ago, we released a sneak peak into what we’ve been building.
Today, can’t be more excited to fully unveil @yutori_ai’s mission, ambitious vision, world-class team, and stellar backers!
Join our waitlist for early access to our product in the coming days and weeks! 👇🏼
Jeff Dean reposted
This will mark the first experimental model with higher rate limits + billing. Excited for this to land and for folks to really put the model through the paces!
This was the #1 point of feedback, besides very long model names, in how we do experimental launches. More to come!
Jeff Dean reposted
Meet Ideogram 3.0 — stunning realism, creative designs, and consistent styles, all in one powerful model. And it's blazingly fast.
Now available to all Ideogram users for free.
Our Gemini 2.5 Pro model has made significant improvements over the Gemini 2.0 series. It's nice to see it topping the LiveBench leaderboard by a pretty healthy margin (+6 in overall average score, +~10 in math and data analysis categories, +~2.5 in language)
Jeff Dean reposted
I took a close look at the 38 most serious crashes Waymo experienced between July 2024 and February 2025. Almost all of them involved a Waymo following the rules and a human-driven car... not doing that.
Jeff Dean reposted
Gemini 2.5 Pro is here and it is 🥇 across ALL categories!
"...the largest score jump ever" 🔥
BREAKING: Gemini 2.5 Pro is now #1 on the Arena leaderboard - the largest score jump ever (+40 pts vs Grok-3/GPT-4.5)! 🏆
Tested under codename "nebula"🌌, Gemini 2.5 Pro ranked #1🥇 across ALL categories and UNIQUELY #1 in Math, Creative Writing, Instruction Following, Longer Query, and Multi-Turn!
Massive congrats to @GoogleDeepMind for this incredible Arena milestone! 🙌
More highlights in thread👇
Want to check out the source for the "AlexNet" paper? Google has made the code from Alex Krizhevsky, @ilyasut, and @geoffreyhinton's seminal "ImageNet Classification with Deep Convolutional
Neural Networks" paper public, in partnership with the Computer History Museum.
As I said in the press release, "Google is delighted to contribute the source code for the groundbreaking AlexNet work to the Computer History Museum".
computerhistory.org/press-releases/chm-makes-alexnet-sour…
I can't believe they've just cancelled the Epidemic Intelligence Service program at CDC.
My father was an EIS officer:
www.epimonitor.net/PrintVersion/November 2024/Final-Novem…
@Farzad_MD's thread below gives you a sense of the kind of people in this elite program to train the best & brightest epidemiologists.
1/ After residency at Mass General Hospital, I reported to Atlanta to meet my fellow CDC Epidemic Intelligence Service Officers.
I have never felt so intimidated by my peers
The best and the brightest, they were star clinicians, had served in disaster zones; MD/PhDs and MSF.
I'm delighted to have joined my good friend and colleague @NoamShazeer for a 2+hour conversation with @dwarkesh_sp about a wide range of topics (early Google, ML hardware, training trillion token LLMs in 2007, model sparsity, continual learning, and more).
Thanks for a fantastic conversation, Noam and Dwarkesh! 🙏
The @JeffDean & @NoamShazeer episode.
We talk about 25 years at Google, from PageRank to MapReduce to the Transformer to MoEs to AlphaChip – and soon to ASI.
My favorite part was Jeff's vision for AGI as one giant MoE that is grown in bits and pieces over time like a forest, rather than trained all at once.
Specialization, distillation, inference time scaling all emerge organically rather than by design.
Noam bites every bullet: 100x world GDP soon; let’s get a million automated researchers running in the Google datacenter; living to see the year 3000.
Links below. Enjoy!
Timestamps
0:00:00 - Intro
0:03:29 - Joining Google in 1999
0:06:20 - Future of Moore's Law
0:11:04 - Future TPUs
0:13:56 - Jeff’s undergrad thesis: parallel backprop
0:15:54 - LLMs in 2007
0:25:09 - “Holy shit” moments
0:27:28 - AI fulfills Google’s original mission
0:32:00 - Doing Search in-context
0:36:12 - The internal coding model
0:37:29 - What will 2027 models do?
0:43:20 - A new architecture every day?
0:49:10 - Automated chips and intelligence explosion
0:53:07 - Future of inference scaling
1:02:38 - Already doing multi-datacenter runs
1:08:15 - Debugging at scale
1:12:41 - Fast takeoff and superalignment
1:20:51 - A million evil Jeff Deans
1:24:22 - Fun times at Google
1:27:51 - World compute demand in 2030
1:34:37 - Getting back to modularity
1:44:48 - Keeping a giga-MoE in-memory
1:49:35 - All of Google in one model
1:57:59 - What’s missing from distillation
2:03:10 - Open research, pros and cons
2:09:58 - Going the distance
Did you know you can use '@' <tab> or '<at>Gemini' to access Gemini in the desktop Chrome address bar? Super convenient!
You can also use:
<at>Tabs to search your tabs
<at> History to search your history
<at>Bookmarks to find bookmarks
9to5google.com/2024/05/02/gemini-chrome-address-bar/
After 25.3 million autonomous miles driven, @Waymo vehicles have an 88% reduction in property damage claims and a 92% reduction in bodily injury claims compared to human drivers per mile driven. 🚖
Our mission is to be the world's most trusted driver and a new study by @SwissRe validates the exceptional safety record of the Waymo Driver. Learn more: waymo.com/blog/2024/12/new-swiss-re-study-waymo
The Gemini cookbook is the top trending repository on GitHub. Examples, API documentation, and more!
github.com/google-gemini/cookbook
Introducing Gemini 2.0 Flash Thinking, an experimental model that explicitly shows its thoughts.
Built on 2.0 Flash’s speed and performance, this model is trained to use thoughts to strengthen its reasoning.
And we see promising results when we increase inference time computation!
Which model is best for turning natural language into SQL? Querying minds want to know...
Gemini models in the top 4 positions.
2024 State of LLMs for Text2SQL Tasks 🏆- Full Report
🥇 Overall Performance: @GoogleDeepMind Gemini-Exp-1206
🥇 Open Source Model: @Alibaba_Qwen 2.5-Coder:32b (Beats Sonnet 3.5 and on par with GPT-4o!)
Disappointing performance by GPT-4o and 3.5 Sonnet on this task. 🧵
What a way to celebrate one year of incredible Gemini progress -- #1🥇across the board on overall ranking, as well as on hard prompts, coding, math, instruction following, and more, including with style control on.
Thanks to the hard work of everyone in the Gemini team and elsewhere at Google! 🎊
Since I wrote this answer to a student, I thought I would share more broadly.
Typical Silicon Valley (pseudo-) logic: let's just pretend this tech is magic, and that its costs don't matter.
Yes, there are many meaningful climate-positive applications of AI, but we urgently need to address the environmental impacts of (superfluous) generative AI.
Former Google CEO Eric Schmidt says energy demand for AI is infinite and we are never going to meet our climate goals anyway, so we may as well bet on building AI to solve the problem
I'm all for that. However that single number is still clouding people's assessment of the environmental costs (e.g. see how this thread started), and @strubell has flatly refused to amend the Arxiv or ACL paper with correct data, and has even co-authored papers after knowing about these errors (which we informed her of as part of our discussions in our writing of the 2021 paper arxiv.org/abs/2104.10350). To them further spread this erroneous information (e.g. the first sentence of the related work section of the 2022 work in arxiv.org/pdf/2206.05229, which reiterates the wildly inaccurate "lifetime of five cars" meme) after being informed about the incorrect numbers is not honest and shouldn't be how science or transparency is done.
I think correcting the honest mistake but ultimately incorrect data in "Energy and Policy Considerations for Deep Learning in NLP" (arxiv.org/abs/1906.02243) would be a good step towards more transparency, as you advocate, and then not publishing additional papers that knowingly use the incorrect data would be even better.
Until that happens, I'll continue to draw attention to the incorrect data as part of the broader discussion of transparency of environmental impacts.
Very exciting to see @GoogleDeepMind introduce AlphaChip and share so many insights with the hardware design community!
I had the chance to meet with @Azaliamirh and @annadgoldie back at the NeurIPS 2019 ML for Systems workshop, and their seminal 2021 Nature paper with @JeffDean using deep reinforcement learning for chip floorplanning had a positive influence on our own work for @DeepPCB.
I’m glad the teams at Google are making continuous progress in AI-driven chip design and fostering an open-source culture that is democratizing hardware development. There is so much to do in this area which is critical for the future!
Welcome, AlphaChip!
Today, we are sharing some exciting updates on our work published in @Nature in 2021 on using reinforcement learning for ASIC chip floorplanning and layout. We’re also naming this work AlphaChip.
Since we first published this work, our use of this approach internally has grown significantly. It has now been used for multiple generations of TPU chips (TPU v5e, TPU v5p, and Trillium), with AlphaChip placing an increasing number of blocks and with larger wirelength reductions vs. human experts from generation to generation:
AlphaChip has also been used with excellent results for other chips across Alphabet, including Google’s Axion chip, an Arm-based general-purpose data center CPU.
In 2022, as a companion to the Nature paper, we open-sourced the code for the AlphaChip algorithms described in the Nature paper (see link below). Since then, external researchers could use this repository to pre-train on a variety of chip blocks and then apply the pre-trained model to new blocks, as was done and described in our original paper.
Today we’re also releasing a pre-trained AlphaChip checkpoint for the open source release that makes it easier for external users to get started using AlphaChip for their own chip designs.
Original Nature paper w/ wonderful joint first authors @Azaliamirh + @annadgoldie, and @mnyazgan, @joesmemory, @ESonghori, @ShenWangURC, @xylophi, @efjohnson, @pathomkar, @Azade_na, @PakJiwoo, Andy Tong, @kavyasrinivas23, @willhang_, @emretuncer, @quocleix, @JamesLaudon, @rh00, Roger Carpenter, and myself):
www.nature.com/articles/s41586-021-03544-w (PDF: rdcu.be/cmedX)
Today’s Addendum to the paper published in Nature: www.nature.com/articles/s41586-024-08032-5 (same authors)
AlphaChip blog post: deepmind.google/discover/blog/how-alphachip-transformed-c…
Open source release: github.com/google-research/circuit_training
Pre-trained checkpoint: github.com/google-research/circuit_training/?tab=readme-o…
Three things we have observed in the external community are described in the Nature Addendum: (1) not doing any pre-training (circumventing the learning aspects of our method by removing its ability to learn from prior experience) (2) not training to convergence (standard practice in ML methods), and (3) using fewer computational resources than described in our Nature paper (using fewer resources is likely to harm performance, or require running for considerably longer to achieve the same performance).
Pre-training the model for it to learn the craft of chip layout and to be able to generalize to new designs is an important part of our method. The pre-training process requires some effort to perform, since one has to find representative blocks and then run a lengthy computational process to pre-train the model to be good at placing those blocks. To avoid external users having to perform this process and make it easier for the external community to use AlphaChip, today we are releasing an AlphaChip model checkpoint pre-trained on 20 TPU blocks. This will enable users to get good zero-shot performance and faster convergence for novel blocks right out of the box. (For best results, however, we continue to recommend that developers pre-train on their own in-distribution blocks, and we provide a tutorial on how to perform pre-training with our open-source repository: see the Addendum).
Many organizations have used AlphaChip as a building block for their own chip design efforts. For example, MediaTek, one of the top chip design companies in the world, extended AlphaChip to accelerate development of their most advanced chips (e.g. the Dimensity Flagship 5G used in Samsung mobile phones), while improving power, performance and chip area.
We’re very excited about the increasing impact of AlphaChip internally and externally, and we look forward to continued work in this space to make custom higher performance, more efficient, and more capable chips dramatically easier to design and build.
Thanks, @Kbeguir! The application of RL to chip floor planning (AlphaChip and follow on work), and @instadeepai's work on RL for PCB board design (DeepPCB) are both great examples of real world application of RL to hardware engineering problems!
Welcome, AlphaChip!
Today, we are sharing some exciting updates on our work published in @Nature in 2021 on using reinforcement learning for ASIC chip floorplanning and layout. We’re also naming this work AlphaChip.
Since we first published this work, our use of this approach internally has grown significantly. It has now been used for multiple generations of TPU chips (TPU v5e, TPU v5p, and Trillium), with AlphaChip placing an increasing number of blocks and with larger wirelength reductions vs. human experts from generation to generation:
AlphaChip has also been used with excellent results for other chips across Alphabet, including Google’s Axion chip, an Arm-based general-purpose data center CPU.
In 2022, as a companion to the Nature paper, we open-sourced the code for the AlphaChip algorithms described in the Nature paper (see link below). Since then, external researchers could use this repository to pre-train on a variety of chip blocks and then apply the pre-trained model to new blocks, as was done and described in our original paper.
Today we’re also releasing a pre-trained AlphaChip checkpoint for the open source release that makes it easier for external users to get started using AlphaChip for their own chip designs.
Original Nature paper w/ wonderful joint first authors @Azaliamirh + @annadgoldie, and @mnyazgan, @joesmemory, @ESonghori, @ShenWangURC, @xylophi, @efjohnson, @pathomkar, @Azade_na, @PakJiwoo, Andy Tong, @kavyasrinivas23, @willhang_, @emretuncer, @quocleix, @JamesLaudon, @rh00, Roger Carpenter, and myself):
www.nature.com/articles/s41586-021-03544-w (PDF: rdcu.be/cmedX)
Today’s Addendum to the paper published in Nature: www.nature.com/articles/s41586-024-08032-5 (same authors)
AlphaChip blog post: deepmind.google/discover/blog/how-alphachip-transformed-c…
Open source release: github.com/google-research/circuit_training
Pre-trained checkpoint: github.com/google-research/circuit_training/?tab=readme-o…
Three things we have observed in the external community are described in the Nature Addendum: (1) not doing any pre-training (circumventing the learning aspects of our method by removing its ability to learn from prior experience) (2) not training to convergence (standard practice in ML methods), and (3) using fewer computational resources than described in our Nature paper (using fewer resources is likely to harm performance, or require running for considerably longer to achieve the same performance).
Pre-training the model for it to learn the craft of chip layout and to be able to generalize to new designs is an important part of our method. The pre-training process requires some effort to perform, since one has to find representative blocks and then run a lengthy computational process to pre-train the model to be good at placing those blocks. To avoid external users having to perform this process and make it easier for the external community to use AlphaChip, today we are releasing an AlphaChip model checkpoint pre-trained on 20 TPU blocks. This will enable users to get good zero-shot performance and faster convergence for novel blocks right out of the box. (For best results, however, we continue to recommend that developers pre-train on their own in-distribution blocks, and we provide a tutorial on how to perform pre-training with our open-source repository: see the Addendum).
Many organizations have used AlphaChip as a building block for their own chip design efforts. For example, MediaTek, one of the top chip design companies in the world, extended AlphaChip to accelerate development of their most advanced chips (e.g. the Dimensity Flagship 5G used in Samsung mobile phones), while improving power, performance and chip area.
We’re very excited about the increasing impact of AlphaChip internally and externally, and we look forward to continued work in this space to make custom higher performance, more efficient, and more capable chips dramatically easier to design and build.
Nice to see this launch of a Pro version of DeepPCB, an RL-based system for laying out PCB boards, which runs on @googlecloud. Nice work, @kbeguir and team! 🎉
The use of RL for many kinds of hardware design and layout problems is accelerating. I remember excitedly comparing notes with Karim many years ago about this work and some work we were doing on using RL for ASIC chip floorplanning, which has been having more and more impact on our hardware work at Google.
👋 Hello World! 🌎
Introducing DeepPCB Pro, our AI-powered design tool for industry-grade PCBs! Fast, precise, and compute-scalable with @googlecloud ☁️ Check it out at deeppcb.ai 👀
We’re honored to partner with Young African AI Research and @JeffDean to launch this Fellowship! #indaba2024 #instadeep
The 2024 Young African AI Research Fellowship launched yesterday at #indaba2024, initiated at #indaba2023 with generous sponsorship from @JeffDean. Inspired by the @instadeepai movie "Cape to Carthage- An African AI Journey," the program nurtures emerging AI talent in Africa 1/3
Delighted to partner with @ascendP360 and @instadeepai on this new fellowship program!
Great work in getting it established, @munniomer!
I'm excited to speak at the GenAI summit in Vietnam (and to visit Vietnam for the first time). Thanks to @lmthang and the other organizers for putting together a very interesting set of speakers including @quocleix , @Diyi_Yang , @YiTayML and others!
Glad to organize the 1st ever GenAI summit in Vietnam and welcome many great speakers from industry, academia, and startups. starting tomorrow, featuring @JeffDean in person for the first time in the region!
Website: summit.newturing.ai
Exciting news! Meet NeuralGCM, a cutting-edge atmospheric model from Google Research. 🌡🌎
This AI-driven general circulation model (GCM) outperforms traditional atmospheric models in precision, providing faster, more accurate climate predictions. Learn more in @Nature → www.nature.com/articles/s41586-024-07744-y Watch the video here → youtu.be/KUAWw32FjIo
Or read the blog post below ⬇️
NeuralGCM is a method that combines traditional physics-based modeling with ML to accurately and efficiently simulate Earth's atmosphere. Learn how NeuralGCM marks a significant step towards developing more powerful and accessible climate models. → goo.gle/4ff57ah
As part of @Google's 1,000 Languages Initiative, a commitment to support the 1,000 most spoken languages, & w/help of our PaLM 2 LLM, we're adding support for 110 new languages (spoken by 614M people) to Google Translate (now supporting 243 languages). 🎉
blog.google/products/translate/google-translate-new-langu…
Gemini 1.5-Pro is a pretty good reward model (tops among generative models and second overall in the Reward Bench Leaderboard).
Gemini 1.5 Pro when zero-shot prompted to perform an LLM-as-a-judge task ranks 1st when compared to other Generative RMs and 2nd best overall vs other dedicated RMs: huggingface.co/spaces/allenai/reward-bench (make sure to click on the Generative checkbox).
Consider entering the Gemini API Developer Competition! Win an environmentally-friendlier electric 1981 Delorean, or $1M in prizes. Plus it'll be fun just to see what you can do with the Gemini APIs!
🚘✨ If you're gonna build an app, why not do it with some style?
Join our Gemini API Developer Competition! Create an app with the Gemini API & showcase the power of gen AI for a chance to win a custom electric 1981 DeLorean and more.
Details here 👀 → goo.gle/3V0it2m
It's nice to have good names for things. I'm proud to have named or been involved in naming a bunch of things at Google over the years, including:
MapReduce
Bigtable
Spanner
TensorFlow
Tensor Processing Units (TPUs)
Pathways
Protocol Buffers
PaLM
Gemini
"I felt the twins aspect of the name ‘Gemini’ was a great fit. The twins here are the folks in the legacy Brain team and the legacy DeepMind team, who started to work together on this ambitious multimodal model project.” - @JeffDean
blog.google/technology/ai/google-gemini-ai-name-meaning/
One other thing in the updated Gemini 1.5 Pro report: we show how a research model that is a mathematics-specialized version of 1.5 Pro achieves a record score of 91.1% on the MATH benchmark (the SOTA just 3 years ago, in May, 2021 was 6.9%!).
Gemini 1.5 Model Family: Technical Report updates now published
In the report we present the latest models of the Gemini family – Gemini 1.5 Pro and Gemini 1.5 Flash, two highly compute-efficient multimodal models capable of recalling and reasoning over fine-grained information from millions of tokens of context, including multiple long documents and hours of video and audio.
Our latest report details notable improvements in Gemini 1.5 Pro within the last four months.
Our May release demonstrates significant improvement in math, coding, and multimodal benchmarks compared to our initial release in February.
Furthermore, the 1.5 Pro Model is now stronger than 1.0 Ultra.
The latest Gemini 1.5 Pro is now our most capable model for text and vision understanding tasks, surpassing 1.0 Ultra on 16 of 19 text benchmarks and 18 of 21 of the vision understanding benchmarks. The table below highlights the improvement in average benchmark performance for different categories in 1.5 Pro since Feb, and also shows the strength of the model relative to the 1.0 Pro and 1.0 Ultra models. The 1.5 Flash model also compares very well against the 1.0 Pro and 1.0 Ultra models.
One clear example of this can be seen on MMLU
On MMLU we find that 1.5 Pro surpasses 1.0 Ultra in the regular 5-shot setting scoring 85.9% versus 83.7%. However with additional inference compute, via majority voting on top of multiple language model samples, we can get a performance of 91.7% versus Ultra’s 90.0%, which extends the known performance ceiling of this task.
@OriolVinyalsML and I are very proud of the whole Gemini team, and it’s fantastic to see this progress and to share these highlights from our Gemini Model Family.
Read the updated report here: goo.gle/GeminiV1-5
There was a bit of confusing back and forth between myself and @ylecun recently, because in the tweet below, 'it' was referencing DistBelief, but @ylecun thought DistBelief referred to the ICML 2012 "cat detector" paper (icml.cc/2012/papers/73.pdf), not to the training system called DistBelief (which was used for the ICML 2012 paper but thousands of other things). DistBelief was initially described in the NeurIPS 2012 paper (Large Scale Distributed Deep Networks, papers.nips.cc/paper_files/paper/2012/file/6aca97005c68f1…). In some back-and-forth discussion, we've clarified that the "cat detector" paper did indeed use backprop (Yann thought it did not), and also that DistBelief supported training end-to-end convnet models starting in early 2012 (not later than the "cat detector" paper as Yann had thought, although the "cat detector" paper itself did not use convolutions because of its focus on evaluating biologically-plausible unsupervised learning: see (2) below).
Now that Yann clarified that it's the ICML 2012 paper which used DistBelief that he meant when he referred to "it" as a "dead end" (I presume Yann doesn't think the DistBelief system in the NeurIPS 2012 paper is a dead end?), let me say why I think the ICML 2012 paper was and is interesting, even if the particular unsupervised autoencoder technique described in the paper isn't widely used today.
(As an aside, I try not to refer to any research work as a "dead end", because I think much of how we progress as a field is by trying things, and that we all collectively learn a lot by seeing what specific ideas work, what general areas are interesting even if the specific ideas may need modifying, get inspired by multiple independent ideas that are combined in other work, etc.)
I think the ICML 2012 "cat detector" paper explored several things that together make it not a "dead end":
(1) It reported results on massively-scaled up neural networks (using the DistBelief software system that enabled this), training a neural net with 1B parameters (100X larger than the previously reported largest networks in the literature). This alone showed many interesting properties (see (3) below).
(2) It evaluated a biologically-plausible approach for doing unsupervised learning on unlabeled images (random frames from YouTube videos), to explore what representations would be learned at high levels in the network when we significantly scaled up an autoencoder-based approach for learning representations from unlabeled data. It did not use convolutions, because we wanted to use network architectures that were more biologically plausible (unshared weights, local receptive fields). This neuroscience-inspired hypothesis is conveyed in the first part of the paper:
Indeed, interesting feature representations did emerge:
(3) In order to show that the unsupervised representations learned were useful for downstream supervised learning, we used supervised training on a neural network initialized with the pre-trained representations that were learned from (2), and compared that with a baseline trained in a purely supervised manner trained from random weights. The results showed that (a) the unsupervised representations learned in a very large model gave significant improvements over the purely supervised approach (15.8% vs. 13.6% on ImageNet 22k), but also (b) that purely supervised learning of a very large model gave major improvements over the prior state-of-the-art (13.6% vs. 9.3% prior SoTA). To me, both (a) and (b) are interesting (and improvements of ~70% on state-of-the-art on ImageNet 22k seem interesting).
(4) The paper demonstrated how to use a combination of model parallelism (each model replica was distributed across 169 machines and also parallelized within each of these machines), as well as data parallelism (using multiple replicas of this setup) trained with asynchronous SGD to train the model robustly and reliably, even in the presence of machine failures (we used 16,000 CPU cores on ~1000 machines). I had explored and compared model parallel and data parallel training at a much smaller scale in my undergraduate thesis in 1990 (x.com/JeffDean/status/1033054187742015489), but this is the first demonstration that I'm aware of for these two parallelism approaches to be put together in training a single system. The NeurIPS 2012 paper on the DistBelief system explored a variety of different tradeoffs in this space, but the ICML 2012 paper was the first one that described this approach (almost as an aside in the paper, though). Nearly all large-scale neural network training these days combines model parallel and data parallel training, and while a lot of large-scale training, I believe asynchronous training will be more and more important as we further scale neural networks.
To me, the lessons learned in (1), (2), (3), and (4) make this not a "dead end". One major takeaway was that continued work on scaling up training of neural networks would lead to further major gains in accuracy across things like computer vision, speech recognition, and language processing. This led to future systems like TPUs, TensorFlow, Pathways, PaLM, and Gemini. Another was that using unsupervised (or self-supervised) methods can help give major boosts in accuracy and reduce the need for labeled data (known at the time, but demonstrated at large scale is still interesting).
Now, could we have done things differently? Sure!
If we'd not been as interested in the biologically-plausible unsupervised learning of local receptive fields and non-shared weights, we could have used convolutions and likely gotten much more efficient learning and probably higher accuracy (it's not very efficient to learn the same kinds of low level filters at each spatial position in your network mostly independently). However, as mentioned in (2), our goal wasn't to get really accurate ImageNet models, but rather to explore what kinds of high-level feature detectors emerged from purely biologically-plausible unsupervised learning.
So, if this is what constitutes a "dead end", I'm delighted to be associated with this paper, along with my awesome co-authors @quocleix, @MarcRanzato, @rajatmonga, Matthieu Devin, Kai Chen, @greg_corrado and @AndrewYNg .
@ylecun, sorry about the initial miscommunication, when your reference to "it" seemed to be discussing the NeurIPS 2012 paper/DistBelief system, rather than the ICML 2012 paper. Because of that, I thought you were saying the DistBelief approach of scaled-up neural network training was a "dead end" and this made me a bit defensive, because I didn't believe it (and you just asserted it without any justification, which also didn't seem reasonable). I'd be happy to hear in more detail which particular things in the ICML 2012 paper you feel are dead ends, or any response you have to the above points.
Vacation photo incoming! Cat detector fired!
It was a dead end.
It couldn't do supervised ConvNets efficiently.
# CUDA/C++ origins of Deep Learning
Fun fact many people might have heard about the ImageNet / AlexNet moment of 2012, and the deep learning revolution it started.
en.wikipedia.org/wiki/AlexNet
What's maybe a bit less known is that the code backing this winning submission to the contest was written from scratch, manually in CUDA/C++ by Alex Krizhevsky. The repo was called cuda-convnet and it was here on Google Code:
code.google.com/archive/p/cuda-convnet/
I think Google Code was shut down (?), but I found some forks of it on GitHub now, e.g.:
github.com/ulrichstern/cuda-convnet
This was among the first high-profile applications of CUDA for Deep Learning, and it is the scale that doing so afforded that allowed this network to get such a strong performance in the ImageNet benchmark. Actually this was a fairly sophisticated multi-GPU application too, and e.g. included model-parallelism, where the two parallel convolution streams were split across two GPUs.
You have to also appreciate that at this time in 2012 (~12 years ago), the majority of deep learning was done in Matlab, on CPU, in toy settings, iterating on all kinds of learning algorithms, architectures and optimization ideas. So it was quite novel and unexpected to see Alex, Ilya and Geoff say: forget all the algorithms work, just take a fairly standard ConvNet, make it very big, train it on a big dataset (ImageNet), and just implement the whole thing in CUDA/C++. And it's in this way that deep learning as a field got a big spark. I recall reading through cuda-convnet around that time like... what is this :S
Now of course, there were already hints of a shift in direction towards scaling, e.g. Matlab had its initial support for GPUs, and much of the work in Andrew Ng's lab at Stanford around this time (where I rotated as a 1st year PhD student) was moving in the direction of GPUs for deep learning at scale, among a number of parallel efforts.
But I just thought it was amusing, while writing all this C/C++ code and CUDA kernels, that it feels a bit like coming back around to that moment, to something that looks a bit like cuda-convnet.
This system successfully scaled up training of a neural net 30X larger than previously reported in the literature.
You're somehow not very knowledgeable & very opinionated about it, referring to it as a "dead end". Even your "Apologies for the confusion" tweet is incorrect.
Results from our first ✨Med-Gemini✨models, variants of the base Gemini multimodal models that have been tuned for the medical domain.
The multimodal and long context capabilities of the Gemini models really shine in some of the capabilities.
Joint work with @vivnat @KhaledSaab11 @taotu831 @RyutaroTanno @davidstutz92 @dgtbarrett @HardyShakerman @3llery @weballergy @alan_karthi @ymatias @greg_corrado @joelle_barral @OriolVinyalsML @koraykv @JeffDean @demishassabis
+ many others
@GoogleDeepMind @GoogleHealth @GoogleAI
Delighted to share ✨Med-Gemini✨ - our new family of multimodal models for medicine unlocking new possibilities for health - arxiv.org/pdf/2404.18416
More accurate multimodal conversations about medical images🩻, surgical videos📽️, genomics🧬, ultra-long health records📚, ECGs🫀 & more with state-of-art performance across multiple benchmarks
More accurate, up-to-date answers to medical questions with advanced reasoning and intelligent use of web-search
Long-context abilities. Summaries or referral letters from long health records, analyses of dozens of long research PDFs & more (1/6)
New versions of Gemma models with a bunch of improvements to instruction following, factuality, reasoning and more are now out.
⬇️
I am very happy to announce that Gemma 1.1 Instruct 2B and “7B” are out! Here are a few details about the new models:
1/11
We're starting to roll out API support for Gemini 1.5 Pro for developers. We're excited to see what you build with the 1M token context window!
We'll be onboarding people to the API slowly at first, and then we'll ramp it up. In the meantime, developers can try out Gemini 1.5 Pro in the AI Studio UI right now:
aistudio.google.com
Introducing Gemma - a family of lightweight, state-of-the-art open models for their class, built from the same research & technology used to create the Gemini models.
Blog post:
blog.google/technology/developers/gemma-open-models/
Tech report:
goo.gle/GemmaReport
This thread explores some of the performance characteristics of these models.
Gemini 1.5 Pro - A highly capable multimodal model with a 10M token context length
Today we are releasing the first demonstrations of the capabilities of the Gemini 1.5 series, with the Gemini 1.5 Pro model. One of the key differentiators of this model is its incredibly long context capabilities, supporting millions of tokens of multimodal input. The multimodal capabilities of the model means you can interact in sophisticated ways with entire books, very long document collections, codebases of hundreds of thousands of lines across hundreds of files, full movies, entire podcast series, and more.
Gemini 1.5 was built by an amazing team of people from @GoogleDeepMind, @GoogleResearch, and elsewhere at @Google. @OriolVinyals (my co-technical lead for the project) and I are incredibly proud of the whole team, and we’re so excited to be sharing this work and what long context and in-context learning can mean for you today!
There’s lots of material about this, some of which are linked to below.
Main blog post:
blog.google/technology/ai/google-gemini-next-generation-m…
Technical report:
“Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context”
goo.gle/GeminiV1-5
Videos of interactions with the model that highlight its long context abilities:
Understanding the three.js codebase: www.youtube.com/watch?v=SSnsmqIj1MI
Analyzing a 45 minute Buster Keaton movie: www.youtube.com/watch?v=wa0MT8OwHuk
Apollo 11 transcript interaction: www.youtube.com/watch?v=LHKL_210CcU
Starting today, we’re offering a limited preview of 1.5 Pro to developers and enterprise customers via AI Studio and Vertex AI. Read more about this on these blogs:
Google for Developers blog:
developers.googleblog.com/2024/02/gemini-15-available-for…
Google Cloud blog:
cloud.google.com/blog/products/ai-machine-learning/gemini…
We’ll also introduce 1.5 Pro with a standard 128,000 token context window when the model is ready for a wider release. Coming soon, we plan to introduce pricing tiers that start at the standard 128,000 context window and scale up to 1 million tokens, as we improve the model.
Early testers can try the 1 million token context window at no cost during the testing period. We’re excited to see what developer’s creativity unlocks with a very long context window.
Let me walk you through the capabilities of the model and what I’m excited about!
Bard (the product) is now Gemini. Gemini Advanced is powered by the Gemini Ultra 1.0 model, our most advanced multimodal model. Congrats to everyone who worked on this effort, including the Bard and Gemini teams! Lots more updates in the blog post!
blog.google/products/gemini/bard-gemini-advanced-app/
I'm looking forward to attending and speaking at the Research@ Bangalore event tomorrow!
@GoogleIndia blog.google/intl/en-in/company-news/technology/research-b…
A lovely article by @CadeMetz about my longtime colleague Luiz Barroso, who unexpectedly passed away a few months ago. Cade, thank you for the lovely tribute to Luiz's amazing career and his influence on the world.
www.nytimes.com/2023/12/13/technology/luiz-barroso-dead.html
I'm excited to head to @NeurIPSConf #neurips2023 this week.
We'll be having a couple of "Chat with the Gemini Team" events in the @GoogleDeepMind / @GoogleResearch booth areas on Tuesday and Wednesday from 1:30 to 3:00 PM (New Orleans time). Quite a few Gemini team members will be there. Come by and say hi!
@OriolVinyalsML @koraykv
We're proud to be a Diamond Level sponsor of this year's annual @NeurIPSConf, which returns to New Orleans, LA. Google has a strong presence at #neurips2023, with >170 papers and participation in 21 workshops and tutorials. Learn about our involvement at goo.gle/47OE0Pv
Gemini is Latin for "twins".
The Gemini effort came about because we had different teams working on language modeling, and we knew we wanted to start to work together. The twins are the folks in the legacy Brain team (many from the PaLM/PaLM-2 effort) and the legacy DeepMind team (many from the Chinchilla effort) that started to work together on the ambitious multimodal model project we called Gemini, eventually joined by many people from all across Google. Gemini was also was the Nasa project that was the bridge to the moon between the Mercury and Apollo programs.
I’m very excited to share our work on Gemini today! Gemini is a family of multimodal models that demonstrate really strong capabilities across the image, audio, video, and text domains. Our most-capable model, Gemini Ultra, advances the state of the art in 30 of 32 benchmarks, including 10 of 12 popular text and reasoning benchmarks, 9 of 9 image understanding benchmarks, 6 of 6 video understanding benchmarks, and 5 of 5 speech recognition and speech translation benchmarks. Gemini Ultra is the first model to achieve human-expert performance on MMLU across 57 subjects with a score above 90%. It also achieves a new state-of-the-art score of 62.4% on the new MMMU multimodal reasoning benchmark, outperforming the previous best model by more than 5 percentage points.
Gemini was built by an awesome team of people from @GoogleDeepMind, @GoogleResearch, and elsewhere at @Google, and is one of the largest science and engineering efforts we’ve ever undertaken. As one of the two overall technical leads of the Gemini effort, along with my colleague @OriolVinyalsML, I am incredibly proud of the whole team, and we’re so excited to be sharing our work with you today!
There’s quite a lot of different material about Gemini available, starting with:
Main blog post: blog.google/technology/ai/google-gemini-ai
60-page technical report authored by th Gemini Team: deepmind.google/gemini/gemini_1_report.pdf
In this thread, I’ll walk you through some of the highlights.
This is a small research community. To all my friends, former colleagues and those that I don't know at OpenAI, the last few days have been a rollercoaster for those of us watching from the outside: I can't imagine the ebb & flow of emotions you're all experiencing. ♥️
Really nice joint work from @GoogleResearch and @GoogleDeepMind on using deep-learning for weather forecasting. This approach produces much higher spatial+temporal resolution forecasts than traditional numerical approaches (see gif below), giving our users much more precise info!
Announcing MetNet-3, the first AI weather model to learn from sparse observations and outperform the top operational systems up to 24 hours ahead at high resolutions. A portion of its forecasts are now available across various Google products. Learn more→ goo.gle/40ljd2Y
Had a really enjoyable internal TGIF today (on a Tuesday, no less), to celebrate @Google's 25th birthday. 🎉.
Sanjay & I did a brief presentation on the origins of some of Google's infrastructure & AI systems.
It was great to see many people that I hadn't seen in quite a while!
Happy 25th birthday @Google! 🎂 Thanks to everyone who uses our products and challenges us to keep innovating and to all Googlers!
There's a huge opportunity to use ML to replace hand-written heuristics in compilers.
Want to try? Check out this @kaggle
contest organized by Ashley Chow, Bryan Perozzi, HCL-Jevster, inversion, Mangpo Phothilimthana, and Sami Abu-El-Haija!
www.kaggle.com/competitions/predict-ai-model-runtime
Great to see so many companies and organizations supporting the Indaba! I look forward to being there!
Not that many systems handle 6B QPS. 😀
"Bigtable has been in continuous production use at Google for more than 15 years now, processing over 6 billion requests per second at peak and with over 10 exabytes of data under management. "
cloud.google.com/blog/products/databases/youtube-runs-on-…
The PaLM language model paper is now officially published at JMLR.
jmlr.org/papers/v24/22-1144.html
I'm excited that bard.google.com is now available in 40 languages & 230 countries! You can use images in your prompts, listen to responses, pin conversations, share responses w/friends, export code to more places & more..
blog.google/products/bard/google-bard-new-features-update…
support.google.com/bard/answer/13575153
Congrats to @RobertMMetcalfe as the winner of the 2022 Turing Award! His pioneering work on Ethernet literally helps connect us all. 🙏
Congratulations to Bob Metcalfe, who receives 2022 ACM A.M. Turing Award for "the invention, standardization, and commercialization of Ethernet," which has become the dominant way of connecting computers to other devices, to each other & to the Internet: awards.acm.org/about/2022-turing
Bard is now available in the US and UK, w/more countries to come. It’s great to see early @GoogleAI work reflected in it—advances in sequence learning, large neural nets, Transformers, responsible AI techniques, dialog systems & more.
You can try it at bard.google.com
In December, we discussed Med-PaLM, at that time a SOTA medical LLM that achieved a 67.6% score on the USMLE MedQA evaluation (passing is 60%).
Today, we're describing Med-PaLM2, which improves on this by +18% with a score of 85.4% ("expert performance")!
Kudos to all involved!
We are so thrilled to announce Med-PaLM 2, our new SOTA medical LLM!
Med-PaLM 2 reaches an accuracy of over 85% on USMLE MedQA, going from "passing score" to "expert performance"!
With Tao Tu, @Mysiak, @alan_karthi, @vivnat and @thekaransinghal!
Excited to share a number of features that build on many years of @GoogleAI research and generative AI advances:
An API to the latest generation of our PaLM language models
MakerSuite for developers,
Writing/rewriting help in Docs and Gmail (Google Workspace),
and more!
Today marks the next chapter of our AI journey with new tools and experiences across Google Workspace and Google Cloud that allow users, developers and businesses to harness the power of generative AI. blog.google/technology/ai/ai-developers-google-cloud-work…
Service Weaver is a new open source release of a system worked on by many of my Google colleagues, including my longtime collaborator Sanjay Ghemaway, to make writing and deploying complex distributed systems easier.
We're excited to share Bard soon. It utilizes a powerful version of LaMDA and is made possible by our foundational work on Transformers, dialog models, responsible AI, model safety techniques, and more.
1/ In 2021, we shared next-gen language + conversation capabilities powered by our Language Model for Dialogue Applications (LaMDA). Coming soon: Bard, a new experimental conversational #googleai service powered by LaMDA.
blog.google/technology/ai/bard-google-ai-search-updates/
Proud to have personally backed @Astranis and @Gedmark. They're on an exciting path working to bring satellite internet to billions of people! I really like that they move faster and think about engineering differently than most traditional satellite makers.
Great conversation. Thanks @fiftyyears for hosting!
🏃♀️💨 SPEED Ep 4: John @Gedmark from @Astranis
John and Ryan McLinko started Astranis in 2015. Both spent time in traditional aerospace and wanted to use geostationary satellites to deliver internet to billions of people, at a pace that would make traditional companies blush.
🧵
Happy 24th Birthday, Google!
24 years on and we're still feeling lucky:) Grateful to all of you for your support, and we'll keep working hard to build a more helpful Google for everyone. (Shout-out to the Charlie's Cafe team for today's birthday 🧁!)
I will sheepishly retweet this..
Yep, I discovered another #badgeographyjoke in the wild...
Honored that our 2012 paper "Building High-level Features Using Large Scale Unsupervised Learning" received an @icmlconf Test of Time Award honorable mention!
Joint work with @quocleix, @MarcRanzato, @RajatMonga, Matthieu Devin, Kai Chen, @greg_corrado, myself, & @AndrewYNg.
New @GoogleAI work:
Input: "Two meerkats sitting next to each other on top of a mountain and looking at the beautiful landscape. There is a mountain, a river lake, and fields of yellow flowers. There are hot air balloons in the sky."
#imagen
g.co/research/imagen
Output:
Introducing Imagen, a new text-to-image synthesis model that can generate high-fidelity, photorealistic images from a deep level of language understanding. Learn more and and check out some examples of #imagen at g.co/research/Imagen
AI can unlock joint human/computer creativity! Imagen is one direction we are pursuing:
gweb-research-imagen.appspot.com
"A high contrast portrait of a very happy fuzzy panda dressed as a chef in a high end kitchen making dough. There is a painting of flowers on the wall behind him."
Chain of thought promoting.
Encouraging language models to "show their work" makes them both more interpretable and more accurate at complex reasoning tasks, solving math problems, etc.
Learn about chain of thought prompting, a method that equips language models to decompose multi-step problems into intermediate steps, enabling models of sufficient scale to solve complex reasoning problems that are not solvable with standard prompting. → goo.gle/3Ft16h0
Delighted to have participated in this special issue of Dædalus on AI & Society (the journal of @americanacad), skillfully guest-edited by James Manyika and with essays by many wonderful authors.
Full 384 page PDF: www.amacad.org/sites/default/files/daedalus/downloads/Dae…
(Tweet below has link to HTML version)
The Spring 2022 issue of Dædalus on AI & Society, guest-edited by James Manyika (@Google), explores the many facets of AI. The 25 #openaccess essays from leading thinkers address #artificialintelligence's beginnings, applications, and potential futures. www.amacad.org/daedalus/ai-society
"We build a grid of Block-NeRFs from 2.8 million images to create the largest neural scene representation to date, capable of rendering an entire neighborhood of San Francisco."
"Circuit training"
Now, thanks to @sguada, @summeryue0, @tobyjboyd, @esonghori, @joesmemory, @annadgoldie & @Azaliamirh & the TensorFlow-Agents team, there's an open-source implementation of RL for chip floorplanning, based on our Nature paper (rdcu.be/cmedX).
Our work on RL for chip floorplanning (Nature’21) is now open-source!
github.com/google-research/circuit_training
We hope "CircuitTraining" enables new advances in learning-based chip design.
Joint work w/ TF-Agents, @sguada, @summeryue0, @tobyjboyd, @esonghori, @joesmemory, @annadgoldie
New results from our @GoogleAI Health AI research team, in collaboration with @PacBio, showing how machine learning can analyze genomic information to glean new insights, leading to better treatment.
blog.google/technology/health/advancing-genomics-better-u…
TIL: ACID for database transactions is a pun because Andreas Reuter's wife hated sweet things and loved vinegar.
Also, the world lost Jim Gray entirely too soon: I was fortunate to meet him once, but hadn't really seen a video of him talking about his work. Remarkably clear!
As in past years, I've spent part of the holiday break summarizing much of the work we've done in @GoogleResearch over the last year. On behalf of @Google's research community, I'm delighted to share this writeup (this year grouped into five themes).
ai.googleblog.com/2022/01/google-research-themes-from-202…
I'm very excited to see this work from @GoogleAI's Ghana Research Center, on using satellite imagery to detect 516M buildings across 19.4M km^2 (64% of African continent). See the blog post and the new Open Buildings dataset.
ai.googleblog.com/2021/07/mapping-africas-buildings-with.…
Very excited to have collaborated with @Azaliamirh, @annadgoldie and many others on this work on using reinforcement learning to teach computers to pay the "game of ASIC chip floorplanning"!
Thrilled to announce that our work on RL for chip floorplanning was published in Nature & used in production to design next generation Google TPUs, with potential to save thousands of hours of engineering effort for each next generation ASIC:
www.nature.com/articles/s41586-021-03544-w (1/7)
On behalf of the entire Google Research & @GoogleAI communities, I'm excited to share an overview of some of our research in 2020.
Thanks to everyone who helped make this work possible!
ai.googleblog.com/2021/01/google-research-looking-back-at…
Can I suggest that from a public health and a moral perspective, we should make sure that, within the U.S. (as well as outside the U.S.), any approved vaccine is available to everyone? The virus doesn't care about green card status or what kind of visa people are on.
Pfizer CEO Albert Bourla raised some eyebrows when he said the COVID vaccine would be available to all US citizens, potentially leaving out tens of millions of people who live and work in America on visas or green cards. @gzeromedia www.gzeromedia.com/pfizer-throws-fresh-light-on-vaccine-p…
Quite a gambit.
The launch of @TensorFlow ~5 years ago this week was a pretty exciting milestone for a lot of us. We all gathered in a conference room early one morning to watch the original blog post announcing it go live & people start downloading it:
ai.googleblog.com/2015/11/tensorflow-googles-latest-machi…
I am very hopeful that we move the U.S. back to welcoming the best & brightest students from all around the world to our colleges and universities. For so many decades, this was an incredible strength of the U.S. I hope we get back to welcoming amazing students with open arms!
A huge relief for international students; I am really happy for them as I know very well the stress that they have been living under #academictwitter
Science. (It works rather well. Pass it on.)
Science. (That’s it. That’s the tweet. Pass it on).
The data has spoken.
Resharing this series of paths created by 800 unmanned bicycles being pushed until they fall over. Science is beautiful. Source: www.reddit.com/r/dataisbeautiful/comments/7s8hck/paths_of…
Susan Eggers was a key member of my Ph.D. thesis committee, & taught me much about compilers/architecture.
This is a great article about her somewhat unusual path to becoming a CS professor (she got her Ph.D. at UC-Berkeley & joined @uwcse as a faculty member at the age of 47).
Susan Eggers, computer scientist and 1st woman to win Eckert-Mauchly Award, began her career problem-solving at Yale EGC.
Read her story: egc.yale.edu/how-job-yale-1960s-set-susan-eggers-groundbr…
@CRA_WP @ComputerSociety @TheOfficialACM @margmartonosi @csenews @JeffDean @rebeccawb @bobbyschnabel @juliaferraioli
Very excited to see the MLPerf 0.7 results released today, where Google TPUs set records in six of the eight benchmarks!
We need bigger benchmarks, because we can now train the ResNet-50, BERT, Transformer, & SSD benchmarks each in under 30 seconds.
cloud.google.com/blog/products/ai-machine-learning/google…
AutoML-Zero: new research from @GoogleAI researchers Esteban Real, @crazydonkey200, David R. So, & @quocleix that that can rediscover fundamental ML techniques by searching a space of different ways of combining basic mathematical operations.
Arxiv: arxiv.org/abs/2003.03384
Extending research into evolutionary #automl, we present an approach that evolves algorithms from scratch—using only basic mathematical operations—rediscovering fundamental ML techniques & showing the potential to discover novel algorithms. Read more ↓ goo.gle/3edHPBc
How is it that we are not yet mandating mask usage in public spaces across whole country? Estimates are that this would save 45,000 lives in next 4 months.
That keeping 45,000 people alive is some sort of political issue instead of a very serious public health issue is absurd.
The University of Washington now projects at least 208,255 will die in the U.S. from the coronavirus by November 1.
But if 95% of the population wears a mask in public, the projected death toll would drop to 162,808. That's more than 45,000 people saved.
www.cnn.com/world/live-news/coronavirus-pandemic-07-07-20…
This ICE action against students on F-1 visas is incredibly damaging in many ways. It hurts U.S. competitiveness, damages our world-class universities, and, if enacted, will disrupt the scholarly study of millions of the brightest students from around the world.
This is bad. ICE just told students here on student visas that if their school is going online-only this fall, the students must depart the United States and cannot remain through the fall semester. www.ice.gov/news/releases/sevp-modifies-temporary-exempti…
AI & responsibility is a serious issue.
This blog post summarizes many @TensorFlow tools to help YOU create more responsible AI systems, looking at things like fairness, bias, interpretability, privacy & security.
Tools: www.tensorflow.org/resources/responsible-ai
Blog: goo.gle/31u54Em
This graph is pretty striking. I hadn't seen this data broken apart by Bay Area vs. rest of California for a while.
Congratulations to my longtime colleague Luiz Barroso for winning this year's Eckert-Mauchly Award (the top award in the computer architecture community):
"for pioneering the design of warehouse-scale computing and driving it from concept to industry"
www.prnewswire.com/news-releases/googles-luiz-andre-barro…
Great data-driven thread about what kinds of laws or practices lead to different outcomes in terms of levels of police violence.
This thread from last year is important and timely. Resurfacing.
Some nice work by Dami Choi, @atpassos_ml, Chris Shallue & George Dahl of @GoogleAI on reusing examples during ML training to get significant improvements in training time, & better handle "slow CPUs driving fast accelerators like GPUs & TPUs".
Arxiv: arxiv.org/abs/1907.05550
Check out this simple technique that reuses intermediate outputs from a neural network training pipeline to reclaim idle accelerator capacity. Rather than waiting for data coming from earlier bottlenecks, it uses data already available for training. goo.gle/2AnY00v
When we released @TensorFlow as an open source project in Nov. 2015, we hoped external machine learning researchers & practicioners would find it as useful as we had internally at @GoogleAI. Very proud to see us hit 100M downloads!
2015 blog post:
ai.googleblog.com/2015/11/tensorflow-googles-latest-machi…
What on earth is wrong with people?
"NO masks allowed" ?!
"We're OPEN to the TRUTH" ?
I'm very happy to see @Google & @Apple partnering on technology to help w/ COVID-19 contact tracing. Technology, coupled w/ strong consideration of privacy, will enable public health authorities to create apps w/ appropriate user opt-in.
blog.google/inside-google/company-announcements/apple-and…
Watch this. And if you don't have time to watch the whole thing, you can skip ahead to about 4:20.
Holy shit. Katie Porter holds the CDC Director's feet to the fire and gets him to commit to FREE TESTING for coronavirus for ALL AMERICANS.
Watch every second of this.
"It was full of... misinformation about the virus & the US response. That’s particularly painful coming from inside the CDC, a longtime powerhouse in global public health now reduced to being a backdrop for grubby politics."
Having worked @CDCgov & @WHO, it pains me to see this.
The President's remarks at the CDC were "a repudiation of good science and good crisis management from inside one of the world’s most respected scientific institutions." www.wired.com/story/trumps-coronavirus-press-event-was-ev…
I'm proud of ways in which Google is helping the world respond to COVID-19 pandemic, from helping people get accurate & up-to-date information, to helping people conduct meetings/events virtually, to helping w/basic science.
Containing this virus will take effort from all of us.
As COVID-19 makes its way across the globe, we’re approaching this unprecedented moment with a deep sense of responsibility. Here are a few ways Google is working to be helpful to our employees, users, partners, customers & communities.
blog.google/inside-google/company-announcements/coronavir…
A great chance to push the petal to the metal using TPUs! Kaggle competitions are also a good chance to practice mettle-learning.
It's happening on kaggle.com right now: TPUs (Tensor Processing Units) have been made available in Kaggle notebooks, with a playground competition to get you started: www.kaggle.com/c/flower-classification-with-tpus
Figure 4 is particularly striking in Fleming's original paper on penicillin (PDF at pmc.ncbi.nlm.nih.gov/articles/PMC2048009/pdf/brjexppathol…).
In 1929, Alexander Fleming published his paper outlining the antibacterial properties of penicillin, and began the era of antibiotics.
Read the Nobel Laureate's full article online: pmc.ncbi.nlm.nih.gov/articles/PMC2048009/pdf/brjexppathol…
What I did over my winter break!
It gives me great pleasure to share this summary of some of our work
in 2019, on behalf of all my colleagues at @GoogleAI & @GoogleHealth.
ai.googleblog.com/2020/01/google-research-looking-back-at…
Nice comprehensive look at the state of federated learning research, including an overview of advances in the last three years (since arxiv.org/abs/1610.05492 in 2016), as well as a nice overview of open problems in the area.
105 pages, 485 references!
Advances and Open Problems in Federated Learning:
arxiv.org/pdf/1912.04977.pdf
from 22 Googlers and 36 academics at 24 institutions! (a) improving efficiency, (b) preserving privacy, (c) defending against attacks, (c) addressing fairness. Here’s to years of future progress on these :)
Some nice case studies about how @GCPcloud's AutoML products can help tackle real-world problems in visual inspection across a number of different manufacturing domains, being used by companies like Global Foundries and Siemens.
cloud.google.com/blog/products/ai-machine-learning/ai-and…
No opinion on favorite or not, but this paper @geoffreyhinton, @OriolVinyalsML, & I submitted to NeurIPS'14 was rejected (~2K citations):
Distilling the Knowledge in a Neural Network
arxiv.org/abs/1503.02531
2/3 said "1: This work is incremental and unlikely to have much impact"
Watching this animation is like a tour down memory lane of old processors.
Fascinating: Moore’s Law predictions vs actual growth in transistor count.
by @datagrapha
www.reddit.com/r/dataisbeautiful/comments/cynql1/moores_l…
Nice evaluation of @GCPcloud AutoML!
"This article shows that physicians with no coding experience can use automated deep learning to develop algorithms that can do clinical classification tasks to a level comparable with traditional deep learning models ..."
Soooo excited to see this out @LancetDigitalH - deep learning for medical image classification by doctors @Moorfields with NO coding experience using @googlecloud #deeplearning #automl #opensource authors.elsevier.com/sd/article/S2589-7500(19)30108-6
#neurips2019 shock: our AlphaNPI paper made it AND we get a spotlight presentation 😳😳😳 Congrats to the @instadeepai team led by @thomas_pierrot, big thanks to our @DeepMindAI colleagues @Scott_e_reed and @NandoDF whom we had met at the @DeepIndaba 2018! instadeep.com/research/paper/towards-compositionality-in-…
Congrats!
This bears watching.
I'm looking forward to speaking about the potential of machine learning for healthcare at this @UofCalifornia-wide event on June 17-18 (My talk is on the 17th)!
Other speakers include @suchisaria and Wendy Nilsen of the NSF, among others.
In the hustle of IO, I missed this awesome comic on federated learning. Kudos to Lucy Bellwood and Scott McCloud for the comic. At the bottom of the page are links to research papers and other content about federated learning by @GoogleAI researchers.
federated.withgoogle.com
Some meetings can literally suck the oxygen right out of you.
This is crazy. Study shoes three people in a conference room over 2 hours can result in a Co2 level that can impair cognitive functioning. Ie. If you’re making decisions at the end of the meeting, you’re mentally less qualified to do so.
In honor of #internationalwomansday tomorrow, I'd like to start a thread about women who have inspired us, taught us, and mentored us in our careers. Please comment with someone who has inspired you! (cont)
I'm very excited about the potential of AI to help us create new educational tools. Bolo is an Android app created by @GoogleAI researchers and engineers, experimenting with how to teach young children in rural India to read. (cont)
india.googleblog.com/2019/03/introducing-bolo-new-speech-…
I was lucky enough to meet Karen Sparck Jones early in my career when I was a researcher working at Digital Equipment Corporation's Western Research Lab (DEC WRL). I & several other researchers had started to get interested in information retrieval (cont) www.nytimes.com/2019/01/02/obituaries/karen-sparck-jones-…
Very happy to see my colleague Geoffrey Hinton has been named a Companion of the Order of Canada. A well-deserved honor! Congrats, Geoff!
🎉🇨🇦 Congratulations to the godfather of Deep Learning, AI pioneer and @Google Engineering Fellow Geoffrey Hinton on being named a Companion of the #orderofcanada. @GoogleAI @UofT
We're very excited about the opening of our @GoogleAI Princeton office, just across the street from @Princeton. The blog post highlights some of the work we've already done in this location, & we look forward to collaborations with the strong research community in Princeton.
Research Scientists @HazanPrinceton and Yoram Singer discuss some early research from Google AI Princeton, a new lab opening early next year that aims to broaden research in multiple facets of #machinelearning. goo.gl/Wdcq7n
This list of important (older) chips is pretty interesting.
In April, '17, @jsomers of @NewYorker reached out & said he wanted to do a small profile of me & my longtime colleague Sanjay Ghemawat, watch us work for a few hours, maybe dinner, etc.
It came out today. I think it captures our working style really well. www.newyorker.com/magazine/2018/12/10/the-friendship-that…
I'm very excited that ICLR 2020 will be in Addis Ababa. This will be the first major machine learning research conference to be hosted on the continent of Africa!
ICLR 2020 will be in Addis! Exciting news not only to reduce the visa issues that have plagued recent conferences, but also because I think many of my colleagues will be inspired by the wave of AI used for good that is sweeping through Africa right now. www.technologyreview.com/s/612434/one-of-the-fathers-of-a…
Like carbon tax, if we could somehow figure out ways to charge or credit people for their transportation choices, that would be a good thing (fewer cars on the road, more use of shared public transit, and more biking or walking with the health benefits that they provide).
What’s the full cost of this 5 km #vancouver commute?
Driving costs society $2.78.
Riding the bus costs society 38¢.
Biking saves society 75¢.
Walking saves society $1.08.
So who’s REALLY paying their way? #citymakingmath Via @thediscourse HT @grescoe www.thediscourse.ca/cities/full-cost-commute
I'm very happy to see the @NipsConference board work out a great solution that keeps the full name of the conference intact but to use NeurIPS rather than NIPS (which has multiple other inappropriate connotations and was sometimes used for crude jokes at the conference). Yay!
Announcement from @NipsConference neurips.cc/Conferences/2018/Press
AI is a powerful tool for improving society. I'm excited about our AI for Social Good program, focusing on both research & empowering the ecosystem.
If you have ideas, submit to the AI Impact Challenge! We're funding the best ideas w/$25M. Learn more at ai.google/social-good
I think enough people are made to feel uncomfortable by the current name that the NIPS board should change the name of the @NipsConference. #protestnips
Julia + TPUs = fast and easily expressible ML computations!
At Google, we've been getting a better understanding of issues of bias & fairness in machine learning models as we've used ML throughout more of our products. We've also created training for Google engineers on these topics, and we've now made this material available externally.
Introducing the MLCC Fairness in #machinelearning module (g.co/mledu/fairness), which looks at different types of human biases that can manifest in training data, and provides strategies to identify them and evaluate their effects. Learn more at: g.co/mledu/fairness-blog
New @GoogleAI research showing machine learning models and pathologists working together can better detect the spread of cancer.
Google researchers have been developing an algorithm to detect the spread of breast cancer. Today they're publishing new research showing its promise as an assistive tool for pathologists. research.google/blog/applying-deep-learning-to-metastatic…
If you want to do ML research, consider applying for the 2019 Google AI Residency program! You'll have the opportunity to conduct cutting-edge research working in a wide variety of areas, and this year we're expanding to host residents in even more locations.
Applications for the 2019 Google AI Residency program are now open! Visit g.co/airesidency/apply for more information on how to apply. To learn more about the accomplishments of the recently graduated second class of residents, visit ↓ research.google/blog/highlights-from-the-google-ai-reside…
A reminder that some people in our field are alienating our female colleagues by flirting in settings that are meant to be professional and by trying to turn what should be topic- & setting-appropriate conversations into dates.
Please don't do this.
In case people pondering grad school in computer science didn't realize, you usually get free tuition plus a modest stipend as a CS PhD student.
Good thread. As a first-gen grad student, I thought you had to take major loans for a CS PhD. I'm lucky to have had great friends (whose parents had advanced degrees) to (i) encourage me to go to grad school and (ii) tell me I'd get paid to do so!
Enjoyed participating in the chat session with @eric_lander and @ericschmidt at #googlenext2018 to discuss healthcare and machine learning yesterday! The video is now up.
www.youtube.com/watch?v=N9mdn3nI3HM&start=28m30s
Really cool work on reconstructing the connectivity of real neurons using artificial neural networks. Be sure to watch the video and animations in the blog post!
Joint work from Google Research (@stardazed0 & Michal Januszewski) & collaborators from the Max Planck Institute
Check out a @naturemethods paper where we and our collaborators present a new type of recurrent neural network that can improve the accuracy of automated interpretation of connectomics data by an order-of-magnitude over previous deep learning techniques. blog.research.google/2018/07/improving-connectomics-by-or…
Welcome to Google, @TimSalimans!
New work from the Google Brain & X robotics teams on learning to grasp improving grasping success from 78%-> 96% on objects unseen during training (5X reduction in errors).
Blog:
ai.googleblog.com/2018/06/scalable-deep-reinforcement-lea…
Paper:arxiv.org/abs/1806.10293
Nice work, @AlexIrpan Peter Pastor & others!
Google Cloud TPUs now offer preemptible pricing at ~70% off the reserved instance pricing. This means, for example, that you can train a ResNet-50 model for ~$7.50 instead of $25, or a Transformer neural translation model for ~$13 instead of $41.
See:
cloudplatform.googleblog.com/2018/06/Cloud-TPU-now-offers…
The current US policy of removing young children from their parents is immoral and an affront to the values for which many of us stand. I urge you to read the personal story shared by my colleague @SussilloDavid
We're continuing to expand our @GoogleAI teams around the world. We’ll be opening our first research center in Africa in #ghana later this year! blog.google/intl/en-africa/company-news/technology/google…
If you're a machine learning researcher interested in working in Accra, Ghana, apply: www.google.com/about/careers/applications/jobs/results
As AI is applied to more and more problems through society, it's important to think carefully about the principles by which this is done. We've just shared Google's AI principles, which shows how we've been thinking about these issues. #googleai
blog.google/topics/ai/ai-principles/
It's been great having Song Han spend the last ~year in the Google Brain team before he joins the MIT faculty this fall!
Today we're joined by @MITEECS assistant professor Song Han to discuss his research on Deep Gradient Compression. We explore the challenge of distributed training for deep neural networks, compressing the gradient exchange for efficiency & more! twimlai.com/podcast/twimlai/deep-gradient-compression-for…
Nice write-up by Yonghui Wu of the Google Brain team about how the new Smart Compose feature in Gmail works behind-the-scenes.
Learn more about the research behind Smart Compose, a new feature in Gmail that uses machine learning to interactively offer sentence completion suggestions as you type, allowing you to draft emails faster → research.google/blog/smart-compose-using-neural-networks-…
Nice analysis by Dario Amodei and Danny Hernandez of the computation scale for a variety of recent important ML papers, and on the trend in amount of compute for these large experiments (doubling every 3 months!).
Very cool analysis on compute vs year for some of the most notable advances in AI. blog.openai.com/ai-and-compute/
A nice blog post describing some of the techniques behind Google Duplex, a system that can carry on conversations with people over the phone to do things like make appointments, check store hours, or make a restaurant reservation. Listen to the clips!
ai.googleblog.com/2018/05/duplex-ai-system-for-natural-co…
TPUv3! So hot it needs help with cooling: first liquid cooled devices in our data centers..
A TPUv3 pod is 8X as powerful as the TPUv2 pod announced at IO last year, offering more than 100 petaflops of ML compute, allowing us to tackle bigger problems & build better products.
From the blog post:
"Today, we are happy to announce Open Images V4, containing 15.4M bounding-boxes for 600 categories on 1.9M images, making it the largest existing dataset with object location annotations."
Announcing Open Images V4, the largest existing dataset with object location annotations, along with the Open Images Challenge, a new object detection challenge to be held at the 2018 European Conference on Computer Vision (#eccv2018). Learn more at research.google/blog/announcing-open-images-v4-and-the-ec…
We just posted a couple of new DAWNBench ImageNet training results on TPUs, including our first result that uses more than one TPU device. It takes 30 minutes to train ResNet-50 to 93% accuracy on half a TPU pod (32 TPU devices).
Our latest DAWNBench results are live: 8h52m for @TensorFlow to train ResNet-50 on ImageNet on a single @GCPcloud TPU (<$60), and just 30 minutes on half a TPU pod!
dawn.cs.stanford.edu/benchmark/
We've been working on a prototype microscope that we think is a great way to blend the expertise of automated machine learning systems with human expertise. It projects augmented information from an ML model into the visual pathway in the microscope.
Check out a prototype Augmented Reality Microscope (ARM) platform that we believe can possibly help accelerate and democratize the adoption of deep learning tools for pathologists around the world.
Blog post → research.google/blog/an-augmented-reality-microscope-for-…
Video → www.youtube.com/watch?v=9Mz84cwVmS0
I'm very much looking forward to being one of the speakers at the 2018 Indaba in South Africa in September. Applications close in a few weeks!
Applications to participate in the 2018 Indaba close in a few weeks. Encourage as many as you know to apply at deeplearningindaba.com/apply. We can't wait to learn from the inspirational speakers we have lined up. More details on this soon ...
Seeing my essay cited as a reason that prominent people in statistics and ML are taking a stand for a more inclusive conference culture makes the stress of the last several months all worth it.
Thank you for speaking out against behavior that has no place in our field (or any field!) and I'm hopeful that real change is happening as a result of your brave actions.
Happy to see TensorFlow getting the top spot for "Most loved" in this survey of 100k developers!
Interesting results from this year's Stack Overflow >100k developer survey insights.stackoverflow.com/survey/2018/
— TF is most loved, torch/pytorch 3rd (against React / Node / Django etc.)
— More excited about possibilities than potential dangers of AI
— Sufficiently concerned about AI biases
Here's new work from the Google Brain team that shows increased accuracy in model for diagnosing diabetic retinopathy through adjudication. Model is now on par with retinal specialist ophthalmologists, rather than general ophthalmologists:
www.aao.org/newsroom/news-releases/detail/google-ai-build…
Nice list of interesting research problems for machine learning in language understanding.
New blog post: Requests for research. A collection of interesting research directions around transfer learning and NLP ruder.io/requests-for-research/
Today, the part of Google Brain that had been working on healthcare problems open sourced a set of protocol buffer definitions for the FHIR standard, which makes it much easier to manipulate data in FHIR format. We've been using this for our recent work on ML for medical records.
My team open sourced #fhir (Fast Healthcare Interoperability Resources) protocol buffers today. It allows #ehr data to be handled in a portable way research.googleblog.com/2018/03/making-healthcare-data-wo…
Congrats to my Google colleagues Martin Abadi ("For contrib. to the formal theory of computer security") and Diane Greene ("For contrib. in transforming virtualization from a concept to an industry") on their NAE election. This is one of the highest honors for an engineer!
OFFICIAL: The National Academy of Engineering announces the Class of 2018! We are proud to welcome 83 new members and 16 new foreign members. Congratulations to these incredible #engineers! Read the official press release: www.nae.edu/CheckRobotPage.html?redirectUrl=/MediaRoom/20…
A new set of essays about human-centered design for ML and AI systems. Scroll down on the design.google/library/ai page to find them, including:
"The UX of AI": design.google/library/ux-ai/
... and others ...
Psyched to announce a new collection of essays on PAIR's human-centered approach to designing with ML and AI: design.google/library/ai/
New paper by @Google on DL for EHR. Hate to say it, but lack of attribution is shocking. No mention of ANY ppl doing this work (RNN/DNN 4 EHR) every day for 3 yrs. Not @david_sontag, @MarzyehGhassemi, Taha Bahaduri, Edward Choi, @davekale, nor myself (note COI). attn: @JeffDean
1/ This is quite embarrassing. Draft has been through many cross-institutional revisions over past few months and somewhere in recent editing, we dropped some of our related work discussion, notably the parts on similar deep learning approaches for EMRs. Thanks for pointing out.
Today, Google Cloud is announcing the first of many AutoML products, AutoML Vision. It's the result of close collaboration between the Google Brain and Cloud AI teams. Automatically solve vision problems to high accuracy with no ML expertise required! www.blog.google/topics/google-cloud/cloud-automl-making-a…
On behalf of the Google Brain team, I wrote a summary of some of the things we did in 2017. This is part 1 of 2. I'm very proud of the whole team and what we accomplished in 2017! research.googleblog.com/2018/01/the-google-brain-team-loo…
Less than four days left to apply for Google's AI Residency program (formerly the Google Brain Residency program). Applications close on Monday, January 8th. If you want to do machine learning research, consider applying!
g.co/airesidency and
g.co/airesidency/apply
Be sure to check out the newest TensorFlow 1.5 release candidate (including eager execution mode: see research.googleblog.com/2017/10/eager-execution-imperativ… for info about eager mode)
TensorFlow 1.5 RC is out
• prebuilt for CUDA9, cuDNN7
• eager execution mode
• TensorFlow Lite
github.com/tensorflow/tensorflow/releases/tag/v1.5.0-rc0
A really nice write-up (by someone outside Google) about the kinds of behind-the-scenes work that goes into taking lots of kinds of raw data and turning it all into high quality Google Maps.
www.justinobeirne.com/google-maps-moat
New work on text-to-speech from @GoogleResearch, done jointly by members of the Google Brain and Machine Perception research teams. From the Arxiv abstract:
"Our model achieves a mean opinion score of 4.53 comparable to a MOS of 4.58 for professionally recorded speech."
Building on TTS models like 'Tacotron' and deep generative models of raw audio like 'Wavenet', we introduce 'Tacotron 2' a neural network architecture for speech synthesis directly from text. Learn more, and check out some audio samples → research.google/blog/tacotron-2-generating-human-like-spe…
This @distillpub article on Feature Visualization by Chris Olah (@ch402), Alexander Mordvintsev (@zzznah), and Ludwig Schubert (@ludwigschubert) is excellent for understanding what's going inside computer vision machine learning models.
distill.pub/2017/feature-visualization/
Read the accounts of what Kristian Lum described, and if you see such behavior in your own workplace or conferences, speak up and make it clear that this is unacceptable. Thanks for describing your experiences, @KLdivergence.
I'll probably end up remembering this moment as the time I decided to blow my career up for no reason, but this is something I've been wanting to say for a long time.
medium.com/@kristianlum/statistics-we-have-a-problem-3046…
Current U.S. policies that make it hard or impossible for some of the most talented people in the world to enter this country, or the proposed tax policies that will tax graduate students that make $28k per year as if they make $71k will definitely hurt the U.S.
A good friend, AI scientist, recently returned from China in awe: “they are toe to toe with us, but moving faster.” In Singapore I was told, “BTW the best thing to happen to China AI is your immigration policy...”
The U.S. leads in artificial intelligence, but for how long? www.technologyreview.com/s/609610/the-us-leads-in-artific…
Slides from my talk in yesterday's ML Systems workshop are now up at learningsys.org/nips17/assets/slides/dean-nips17.pdf
#nips2017