Founder Mode activated from my new office ♾
Oriol Vinyals
Oriol Vinyals (born 1983) is a Spanish machine learning researcher. He has developed most of his career at DeepMind. He is a co-founder of Discovery Loop, and was technical lead on Gemini until August 2026, along with Jeff Dean. More on Wikipedia
-
Oriol Vinyals
@OriolVinyalsML
-
Post not available here, view on X
Oriol Vinyals
@OriolVinyalsML
We'll gladly take model credits Tibo : ) Let's catch up.
-
Oriol Vinyals
@OriolVinyalsML
Thirteen years ago, a quick email exchange with @JeffDean brought me to @Google. After an unforgettable journey from Brain, to DeepMind, and to @GoogleDeepMind, working on seq2seq or distillation, and co-leading huge projects such as AlphaStar, and Gemini, I'm saying goodbye to GDM.
The best part? The same person who brought me in is challenging me to take the next leap.
Excited to build Discovery Loop (@DiscoLoopAI) alongside @JeffDean, @quocleix, and @Sanjay_Ghemawat
Deeply grateful to everyone who made GDM feel like home. Time to try something new and to keep it drastic! ✨♾ -
Oriol Vinyals
@OriolVinyalsML
Fun fact: @woj_zaremba wanted to completely remove that paper from existence a year or so after. I forgot why though -- Woj?
Ofir Press
@OfirPress
I started doing language modeling research in Jan 2016, using the LSTM from the "RNN Regularization" model by @woj_zaremba @OriolVinyalsML and @ilyasut, trying to improve its perplexity on Penn Treebank (1M toks). This led to the weight tying method, sometimes still used today.
-
Oriol Vinyals
@OriolVinyalsML
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
-
Oriol Vinyals
@OriolVinyalsML
1/ Today at Google I/O, we’re launching Gemini 3.5 Flash ⚡️⚡️⚡️!
Our mission was clear: bring frontier-level intelligence with unprecedented speed.
3.5 Flash delivers drastic intelligence (beating 3.1 Pro on almost every benchmark), at Flash speeds. 🧵Post not available here, view on X
Oriol Vinyals
@OriolVinyalsML
3/ If you push Gemini 3.5 Flash with complex agents and tasks, it will shine brighter. Here, it generates fully tactile, interactive HTML/SVG hardware simulations, complete with bump-mapped metals, spring physics, and procedural audio, in a single shot.
This is what happens when you connect reasoning across the full stack simultaneously at extremely low latency. 💡🔉
Oriol Vinyals
@OriolVinyalsML
4/ Gemini 3.5 Flash is available today globally - we can't wait to see the incredible agents and apps you all build with it!
Check out our blog for more. blog.google/innovation-and-ai/models-and-research/gemini-… -
Oriol Vinyals
@OriolVinyalsML
3.1 Flash-Lite 🔦 > 2.5 Flash ⚡️
It happened again. Our newest Flash-Lite model outperforms the previous generation's Flash tier.
Smarter. Faster. Cheaper.
blog.google/innovation-and-ai/models-and-research/gemini-… -
Oriol Vinyals
@OriolVinyalsML
Banana^2
Arena.ai
@arena
🚨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! -
Oriol Vinyals
@OriolVinyalsML
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-…Post not available here, view on X
Oriol Vinyals
@OriolVinyalsML
"Generate an SVG of a rollercoaster"
Oriol Vinyals
@OriolVinyalsML
"make it first person view (i want to see the rollercoaster in front of me)"
-
Oriol Vinyals
@OriolVinyalsML
Personal update: After an amazing 10 years in London, it's time for a major change. One-way ticket back to California 🌞! I'm incredibly excited to return to the Bay Area to continue building Gemini and pushing us toward the age of AGI 🚀
-
Oriol Vinyals
@OriolVinyalsML
An updated & faster Gemini 3 Deep Think is taking off! 🚀 Our smartest mode to date!™️
PhD-level reasoning to the most rigorous STEM challenges (models' gotta think harder). Gold medal-level results on Physics & Chemistry Olympiads. 🧪💻
Full details: blog.google/innovation-and-ai/models-and-research/gemini-…
Oriol Vinyals
@OriolVinyalsML
Fun to see Deep Think’s real-world impact. Check out how it’s helping researchers catch errors in high-level mathematics research papers. As "just" a math undergrad, I couldn't even dream to do any of this myself!
This media couldn't be displayed. See it on the original post. -
Oriol Vinyals reposted
Adam Brown
@A_G_I_Joe
New paper out today, proving a novel theorem in algebraic geometry with an internal math-specialized version of Gemini. This was a collaboration between @GoogleDeepMind (Professor Freddie Manners and @GSalafatinos, hosted by the Blueshift team) and Professors Jim Bryan, Balazs Elek, and Ravi Vakil.
arxiv.org/abs/2601.07222 -
Oriol Vinyals reposted
Logan Kilpatrick
@OfficialLoganK
Recapping an incredibly year of Gemini progress with @OriolVinyalsML @JeffDean and @NoamShazeer live, join us : )
x.com/i/spaces/1eaJbjvBOooJX -
Oriol Vinyals
@OriolVinyalsML
Classic LLM dilemma: smart vs fast. The new reality with Gemini 3 Flash: Why not both?
That trade-off has never felt smaller. Congrats to the team! ⚡️ -
Oriol Vinyals
@OriolVinyalsML
See you at 2:30 pm with @JeffDean and the Gemini team @NeurIPSConf!
-
Oriol Vinyals
@OriolVinyalsML
On my way to Barcelona to receive a Doctor Honoris Causa from my alma mater, @la_UPC. Truly honored! 🎓
Join Thursday for my Master Class, "From AI to AGI: The Quest for True Intelligence." Hope to see you there! telecos.upc.edu/ca/esdeveniments/master-class-del-dr-orio…
"Create an image at 41.4036° N, 2.1744° E, January 1st, 1983, 15:00 hours." -
Oriol Vinyals reposted
Demis Hassabis
@demishassabis
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…
Oriol Vinyals
@OriolVinyalsML
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 💙💙💙 -
Oriol Vinyals reposted
Oriol Vinyals
@OriolVinyalsML
Confirmed
-
Oriol Vinyals
@OriolVinyalsML
Visual Retweeting powered by Nano Banana Pro 🍌
Prompt used: Create a clever infographic to illustrate <copy_paste_tweet>
Oriol Vinyals
@OriolVinyalsML
(Original tweet x.com/OriolVinyalsML/status/1990854455802343680)
Oriol Vinyals
@OriolVinyalsML
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 💙💙💙 -
Post not available here, view on X
Oriol Vinyals
@OriolVinyalsML
Confirmed
-
Oriol Vinyals
@OriolVinyalsML
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 💙💙💙 -
Oriol Vinyals
@OriolVinyalsML
🤔🤔🤔
Sundar Pichai
@sundarpichai
🤔🤔
Alphabetting
@wintermoat
Prediction markets are betting on Gemini 3 release next week $goog
-
Oriol Vinyals reposted
Google DeepMind
@GoogleDeepMind
If you want to learn AI from the experts, keep reading. 💡
Together with @UCL, we made a free AI Research Foundations curriculum – available now on Google Skills.
With lessons from a Gemini Lead like @OriolVinyalsML, you'll explore how to code better, fine-tune an AI model and more. → www.skills.google/collections/deepmind -
Oriol Vinyals
@OriolVinyalsML
We've just upgraded Gemini 2.5 Flash image generation & editing! 🍌🍌🍌
Besides topping leaderboards, it topped my model usage this month. It keeps subjects consistent, you can make precise edits & combine creative elements. Have fun with it @GeminiApp @GoogleAIStudio
Oriol Vinyals
@OriolVinyalsML
Also it went kind of viral on LMArena (as nano-banana), 2.5M votes before it released 🤯
-
Oriol Vinyals
@OriolVinyalsML
Incredible evolution of "Neural Video Games": from GQN (2018) to Genie3 (2025). The future is exciting!
deepmind.google/discover/blog/genie-3-a-new-frontier-for-… -
Oriol Vinyals
@OriolVinyalsML
Gemini 2.5 Deep Think is rolling out to Google AI Ultra subscribers!
It uses advances in parallel reasoning & RL to tackle tough math & science problems. Awesome feeling to have IMO medal capabilities at your fingertips! Such a good model!🥇
blog.google/products/gemini/gemini-2-5-deep-think/ -
Oriol Vinyals
@OriolVinyalsML
We already are. Just through a fairly narrow communication bottleneck.
Lisan al Gaib
@scaling01
what if I told you that OpenAI, Google, Anthropic and xAI will all be working together in a few years
-
Oriol Vinyals
@OriolVinyalsML
Drastic progress on maths with Gemini 2.5! As a math undergrad, I am impressed 🤯
🥈 -> 🥇 ✅
Formal -> Informal ✅
Specialized model -> General model ✅
Available soon ✅
Huge thanks to IMO and congrats to all participants!
Blog: deepmind.google/discover/blog/advanced-version-of-gemini-… -
Oriol Vinyals
@OriolVinyalsML
good trajectory++ ⭐️🍿
Demis Hassabis
@demishassabis
good trajectory...
Similarweb
@Similarweb
Google's AI domains are thriving.
Month-over-month traffic growth in May:
@GoogleDeepMind: 4,258,300 → 11,189,832 (+162.78%)
@GoogleLabs: 9,796,459 → 19,792,943 (+102.04%)
@GoogleAI: 1,349,027 → 2,042,165 (+51.38%)
@NotebookLM: 48,340,200 → 64,906,295 (+34.27%)
@GeminiApp: 409,432,013 → 527,738,672 (+28.90%) -
Oriol Vinyals
@OriolVinyalsML
"A Neural Conversational Model" is 10 years old, w/ @quocleix . TL;DR you can train a chatbot with a large neural network (~500M params!). Samples 👇
This paper was received with mixed reviews, but I'm glad all the critics are now riding the LLM wave 🌊 arxiv.org/abs/1506.05869 -
Oriol Vinyals
@OriolVinyalsML
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 -
Oriol Vinyals
@OriolVinyalsML
Introducing the new Gemini 2.5 Pro preview, with a +24 LMArena Elo score over its predecessor! 📈 It leads on tough coding (AIME, AIDER), science (GPQA), and reasoning (HLE).
Style/structure have improved thanks to your feedback. Learn more:
goo.gle/4kKynYo -
Oriol Vinyals
@OriolVinyalsML
Cool benchmark! Good to see Gemini 2.5 topping another leaderboard, though we are quite far from the summit 😅🪜🏔️
alex zhang
@a1zhang
Can GPT, Claude, and Gemini play video games like Zelda, Civ, and Doom II?
𝗩𝗶𝗱𝗲𝗼𝗚𝗮𝗺𝗲𝗕𝗲𝗻𝗰𝗵 evaluates VLMs on Game Boy & MS-DOS games given only raw screen input, just like how a human would play. The best model (Gemini) completes just 0.48% of the benchmark!
🧵👇 -
Oriol Vinyals reposted
Hashem Al-Ghaili
@HashemGhaili
Prompt Theory (Made with Veo 3)
What if AI-generated characters refused to believe they were AI-generated?This media couldn't be displayed. See it on the original post. -
Oriol Vinyals reposted
Thang Luong
@lmthang
As highlighted yesterday at #googleio, Gemini 2.5 Pro with #deepthink mode achieved a new state-of-the-art performance of 49.4% on USAMO 2025, an extremely challenging Math benchmark for AI that requires detailed proof writing.
I was analyzing the model outputs with our math experts for Problem #2, a challenging algebra problem. We were really surprised that it can nail a complex solution, something we haven’t seen before in previous models! This video shows how Gemini 2.5 Pro with DeepThink solves this problem:
- First the model decided that it will prove by contradiction.
- The model then came up with different ideas such as using Rolle's theorem or Newton's inequalities.
- The model later decided to proceed with Rolle's theorem and conducted a series of advanced algebraic manipulations before arriving at a final correct proof.
The results are quite fascinating, thanks to our latest research in parallel thinking techniques!
Thang Luong
@lmthang
It's been a wild run over the last 2 weeks to land Deep Thinking just in time for I/O! Heroic efforts across many teams @GoogleDeepMind from research to production that felt almost impossible last month! Proud of our research team for the relentless effort over the past 6 months towards advanced deep reasoning with parallel thinking, with state-of-the-art performances across math (USAMO), coding (LiveCodeBench), and multimodality (MMMU)! #googleio
-
Oriol Vinyals
@OriolVinyalsML
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!
Oriol Vinyals
@OriolVinyalsML
Deep Think uses new techniques, allowing the model to consider multiple hypotheses before responding. It's not just about math; it also gets an impressive score on LiveCodeBench 💻, MMMU & more!
Learn more: blog.google/technology/google-deepmind/google-gemini-upda…
Oriol Vinyals
@OriolVinyalsML
Because we're defining the frontier with Deep Think, we're taking extra care. This includes first getting input from safety experts and trusted testers🛡️ via the Gemini API to gather crucial feedback before wider release.
-
Oriol Vinyals
@OriolVinyalsML
Veo2 was already SOTA, but Veo3 is something else! Been having a blast playing with it, and loving all the creativity now that it's out (more in 🧵).
Me walking to I/O thinking about Gemini3, Veo4, ...
Oriol Vinyals
@OriolVinyalsML
Spaghetti benchmark.
Google Gemini
@GeminiApp
Bon appétit 🍝
Oriol Vinyals
@OriolVinyalsML
Ultimate AGI benchmark: humor!
fofr
@fofrAI
NO WAY. It did it. And, was that, actually funny?
Prompt:
> a man doing stand up comedy in a small venue tells a joke (include the joke in the dialogue) -
Oriol Vinyals
@OriolVinyalsML
Today we introduced Gemini Diffusion⚡️ (& DeepThink, Veo3, Imagen4, 2.5 updates...).
It's been a dream of mine to remove the need for "left to right" text generation. It's so fast, that we had to *slow down* the video during the presentation.
deepmind.google/models/gemini-diffusion/ -
Oriol Vinyals
@OriolVinyalsML
We may be delaying AGI if we keep playing this game, but OK! I used Copy Canvas and here is my addition to the Cartographer. Coding this in two sentences... drastic & wild times! @NoamShazeer I nominate you : )
g.co/gemini/share/c6ad686dbd1c
Jeff Dean
@JeffDean
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!
Tulsee Doshi
@tulseedoshi
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.This media couldn't be displayed. See it on the original post. -
Oriol Vinyals
@OriolVinyalsML
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!
Oriol Vinyals
@OriolVinyalsML
With this update, you can create even more complex web apps from a single prompt. See a side by side (new vs old 2.5 Pro) on building interactive learning apps from YouTube videos.
Try the new Gemini 2.5 Pro in AI Studio, Vertex AI, Gemini app & Canvas. aistudio.google.com -
Oriol Vinyals
@OriolVinyalsML
It's not only about how long your context is, but how well you use it. Great to see Gemini 2.5 models dominating MRCR and other benchmarks on long context!
See 2.5 Pro tackle a complex coding task by reasoning over an entire repo (>500k tokens). Performance and effective use of the (loooong) context windows are what really matter!This media couldn't be displayed. See it on the original post.
Oriol Vinyals
@OriolVinyalsML
For some benchmarks & analysis:
Dillon Uzar
@DillonUzar
Another update - Ran Gemini 2.5 Flash (Auto Thinking and Non-Thinking). See the comparison below to other thinking models.
Interesting curve for Gemini 2.5 Flash Non-Thinking! Meanwhile Gemini 2.5 Flash Thinking (Auto) matches Gemini 2.5 Pro!
I'm still working on o3 access and will update when I have access. Some error on OpenAI's side when generating the Persona link. DM me if you are willing to grant me access in the meantime. -
Oriol Vinyals
@OriolVinyalsML
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!
Alberto Romero
@Alber_RomGar
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.) -
Oriol Vinyals
@OriolVinyalsML
MCP here we come! ♊️🚀
We’re collaborating w/ the MCP team, and want to hear directly from you again! What do you need from Gemini models to build more powerful applications with MCP? Lmk in comments
Sundar Pichai
@sundarpichai
To MCP or not to MCP, that's the question. Lmk in comments
-
Oriol Vinyals
@OriolVinyalsML
I somehow missed this, wow! Any analysis from the solutions on USAMO of the improvements vs other models? ♊️
Mislav Balunović
@mbalunovic
Big update to our MathArena USAMO evaluation: Gemini 2.5 Pro, which was released *the same day* as our benchmark, is the first model to achieve non-trivial amount of points (24.4%). The speed of progress is really mind-blowing.
-
Oriol Vinyals
@OriolVinyalsML
Cool leaderboard by @alexandr_wang and team @scale_AI! Good to see so much purple 🟪
-
Oriol Vinyals
@OriolVinyalsML
Glad that our frontier models are literally defining the (Pareto) frontier across the board!
Jack Rae
@jack_w_rae
The pricing for 2.5 Pro is out.
Here is the pareto performance of price : lmsys elo visualized as a rainbow. -
Oriol Vinyals
@OriolVinyalsML
A mere ~16 point jump on livebench.ai. Such a good model! Gemini 2.5 Pro ♊️
-
Oriol Vinyals reposted
Logan Kilpatrick
@OfficialLoganK
: )
-
Oriol Vinyals
@OriolVinyalsML
Introducing Gemini 2.5 Pro Experimental! 🎉
Our newest Gemini model has stellar performance across math and science benchmarks. It’s an incredible model for coding and complex reasoning, and it’s #1 on the @lmarena_ai leaderboard by a drastic 40 ELO margin. Only a handful of model releases have leaped ahead so strongly in ELO. 📈
ELO score differences map directly to win rate: e.g. a 400 ELO difference yields a ~91% win rate. Incredible that since 1.5, just a year ago, we jumped 200 ELO (300 since 1.0).
Here’s a fun example where Gemini 2.5 Pro writes code to create an animated swarm of colorful boids swimming in a rotating hexagon. 💫
Try the model for free today in AI Studio. It’s also available to Gemini Advanced users in @geminiapp. aistudio.google.com/app/prompts/gemini-2.5-pro-exp-03-25
Blog: goo.gle/4c3NitOThis media couldn't be displayed. See it on the original post. -
Oriol Vinyals
@OriolVinyalsML
🤔
AshutoshShrivastava
@ai_for_success
More news is coming that Nebula on LMSYS Arena is actually a Google model, probably Google Gemini 2.0 Pro Thinking Model.
It is too good at coding too, Looks like Google is making sure to nail coding.
Can’t wait 🔥🔥
Thanks @AnalogPvt for the info. -
Oriol Vinyals
@OriolVinyalsML
Mind blown 🤯 Gemini 2.0 Flash interleaved image generation is a game changer. I explained AlphaFold to my 2yo daughter's class for science week on Friday 🧑🔬. Prompt: "Write an illustrated story about AlphaFold for toddlers." Instant magic! ✨
aistudio.google.com -
Oriol Vinyals
@OriolVinyalsML
Big day for the robots too! 🚢🤖🚢🤖🚢🤖🚢🤖🚢🤖
Sundar Pichai
@sundarpichai
We’ve always thought of robotics as a helpful testing ground for translating AI advances into the physical world. Today we’re taking our next step in this journey with our newest Gemini 2.0 robotics models. They show state of the art performance on two important benchmarks - generalization and embodied reasoning - which enable robots to draw from Gemini’s multimodal understanding of the world to make changes on the fly + adapt to their surroundings. This milestone lays the foundation for the next generation of robotics that can be helpful across a range of applications.
-
Oriol Vinyals
@OriolVinyalsML
Loving our open models family! Left as an exercise for the open source community: add reasoning on top of Gemma3 and become the best open model 🚀💙
Sundar Pichai
@sundarpichai
Gemma 3 is here! Our new open models are incredibly efficient - the largest 27B model runs on just one H100 GPU. You'd need at least 10x the compute to get similar performance from other models ⬇️
-
Oriol Vinyals
@OriolVinyalsML
Gemini 2.0 Flash debuts native image gen! Create contextually relevant images, edit conversationally, and generate long text in images. All totally optimized for chat iteration.
Try it in AI Studio or Gemini API. Blog: developers.googleblog.com/en/experiment-with-gemini-20-fl… -
Oriol Vinyals
@OriolVinyalsML
Still trying to figure out who reivewer#2 was! 💙
-
Oriol Vinyals
@OriolVinyalsML
Gemini 🤝 Search. Check out AI Mode, our new Labs experiment with the potential to make Google Search drastically more helpful for your hardest questions. AI Mode offers more advanced reasoning, thinking and multimodal capabilities, bringing the best of Gemini 2.0 to Search.
Sundar Pichai
@sundarpichai
AI Overviews are one of our most popular Search features ever. They keep getting better as models improve, and now, Gemini 2.0 is powering responses for coding, advanced math, and multimodal queries in the US, with more to come.
Today, we’re also introducing our newest Labs experiment for Search: AI Mode. You’ll get AI responses using Gemini 2.0’s advanced reasoning, thinking, multimodal capabilities + new ways to explore even more of the web. -
Oriol Vinyals
@OriolVinyalsML
We are hiring ♊️ deepmind.google/about/careers/
-
Oriol Vinyals
@OriolVinyalsML
Distillation has been on the news (!) due to @deepseek_ai. The paper arxiv.org/abs/1503.02531 was actually rejected from NeurIPS 2014 due to lack of novelty 🧐 (true-ish), and lack of impact 🙃.
Thanks reviewer#2 (literally), and thanks for @arxiv!
@geoffreyhinton @JeffDean
Oriol Vinyals
@OriolVinyalsML
PS @SchmidhuberAI in 1991 published something related to compressing history length in RNNs, which may be very relevant for long context, an important topic of research today.
-
Oriol Vinyals
@OriolVinyalsML
The Gemini 2.0 family grows! I love astronomy, so what better way to celebrate than to code up an animated 3D scene of the solar system with Gemini 2.0 Pro Exp 🪐🌞 Let’s keep the balls rolling!
Google DeepMind
@GoogleDeepMind
Gemini 2.0 is now available to everyone. ✨
⚡ Start using an updated 2.0 Flash in @Google AI Studio, @GoogleCloud’s #vertexai and in @GeminiApp.
We’re also introducing:
🔵 2.0 Pro Experimental, which excels at coding.
🔵 2.0 Flash-Lite, our most cost-efficient model yet.
🔵 2.0 Flash Thinking Experimental, in @GeminiApp.
🧵 goo.gle/40OsRfj -
Oriol Vinyals
@OriolVinyalsML
I uploaded my PhD thesis (98K tokens) and asked our latest Flash thinking model to find flaws in it. "Potential over-reliance on empirical results without sufficient theoretical grounding." A Deep Learning thesis, indeed.
Kara 🦇 🔊
@0xkarasy
I just tested one of my dissertation (over 40,000 words) with Google AI Studio. I left ONE wrong correlation coefficient value and guess what it picked up first.
"For hypotheses H4 and H5, you state the correlation coefficient is 0.501 for both Cultural Alignment -> PU and Quality of Chatbot Responses -> PU. Double-check if these are actually identical in your SPSS output. If so, it's just a coincidence, but worth verifying."
Absolutely amazing catch.
Oriol Vinyals
@OriolVinyalsML
Quite funny to ask it to predict my future career (pretty accurately I must say!), or to roast the thesis (Q_Q).
Oriol Vinyals
@OriolVinyalsML
Great & super fast progress from our first Thinking release in December, kudos to @jack_w_rae @NoamShazeer and the whole Gemini team!
-
Oriol Vinyals
@OriolVinyalsML
🤔
Logan Kilpatrick
@OfficialLoganK
🤔
-
Oriol Vinyals reposted
Google DeepMind
@GoogleDeepMind
Join our VP (Drastic) Research, Gemini co-Tech Lead @OriolVinyalsML and our podcast host @FryRsquared as they discuss the evolution of our AI models, from AlphaGo to Gemini.
They also cover agentic capabilities and why giving AI access to tools could lead to a new era of problem-solving.
Listen now ↓
Timecodes:
00:00 Intro
02:30 Games and early AI agents
04:28 Weights
09:27 Architectures and a digital brain
10:24 Agentic behaviour
13:31 Digital body
14:09 Scaling
19:02 Data
20:59 Complex understanding and knowledge
25:14 Post training challenges
30:43 Reasoning
33:11 Planning
34:19 Systems 2
37:00 Memory
40:54 Gemini and agentic capabilitiesThis media couldn't be displayed. See it on the original post. -
Oriol Vinyals
@OriolVinyalsML
seq2seq 10 years later. 💙@ilyasut @quocleix. Talk 👇
Vincent Weisser
@vincentweisser
.@ilyasut full talk at neurips 2024 "pre-training as we know it will end" and what comes next is superintelligence: agentic, reasons, understands and is self aware
This media couldn't be displayed. See it on the original post. -
Oriol Vinyals
@OriolVinyalsML
Besides model quality, it has been incredible to see the advances in infrastructure that happened in the last year (efficiency, speed, easier to use & ship).
I am sure things are still far from perfect, so please give (constructive) feedback to @clmt and @OfficialLoganK 😅
Clement Farabet
@clmt
Very excited to share all the new developer features we shipped for Gemini 2 –
Live mode aistudio.google.com/live audio, video and screen sharing!
Spatial understanding aistudio.google.com/starter-apps/spatial
And some more coming soon e.g. native audio/image out.
developers.googleblog.com/en/the-next-chapter-of-the-gemi… -
Oriol Vinyals
@OriolVinyalsML
Check out this example which showcases one of the most exciting research directions: self-improvement.
In it, you see this behavior emerging (!) when the model realizes (with a "Oops!") that it did a mistake, and fixes it to create the cute image.
Wild times.
Kaushik Shivakumar
@19kaushiks
Full conversation that was used to generate this image.
-
Oriol Vinyals
@OriolVinyalsML
Gemini 2.0 Flash ⚡️ has arrived!
2.0 Flash > 1.5 Pro (again!) 📈
Interacts with a browser 🤖
Native image generation 🖼️
and much more!
Try it out aistudio.google.com/prompts/new_chat
As a preview of what is possible, wishing you all a Drastic Holiday powered by 2.0! -
Oriol Vinyals
@OriolVinyalsML
A year ago we launched Gemini 🚀 Back then I shared with the team that it would be great to be in a position to beat ourselves. Just a year later this is starting to happen. But we have formidable competitors, so I expect this to age poorly 😅. GL HF!
aistudio.google.com/app/prompts/new_chat?model=gemini-exp… -
Oriol Vinyals reposted
Oriol Vinyals
@OriolVinyalsML
Scaling up a simple / boring recipe is not what many researchers wanted to hear at the time.
-
Oriol Vinyals
@OriolVinyalsML
Such a great honor, thanks a lot @NeurIPSConf and congrats to my esteemed co-authors @ilyasut & @quocleix! The 2014 talk also stood the test of time IMO. Here is a slide from it (powerful models of today == large transformers). Believe it or not this talk was controversial at the time (I got a lot of negative feedback at NeurIPS).
NeurIPS Conference
@NeurIPSConf
Announcing the NeurIPS 2024 Test of Time Paper Awards:
blog.neurips.cc/2024/11/27/announcing-the-neurips-2024-te…Post not available here, view on X
Oriol Vinyals
@OriolVinyalsML
Scaling up a simple / boring recipe is not what many researchers wanted to hear at the time.
-
Oriol Vinyals
@OriolVinyalsML
Ensemble submission anytime soon?
Sam Altman
@sama
good new model out!
-
Oriol Vinyals
@OriolVinyalsML
🍿🍿🍿
Arena.ai
@arena
Woah, huge news again from Chatbot Arena🔥
@GoogleDeepMind’s just released Gemini (Exp 1121) is back stronger (+20 points), tied #1🏅Overall with the latest GPT-4o-1120 in Arena!
Ranking gains since Gemini-Exp-1114:
- Overall #3 → #1
- Overall (StyleCtrl): #5 -> #2
- Hard Prompts (StyleCtrl): #3 → #1
- Coding: #3 → #1
- Vision: #1
- Math: #2 → #1
- Creative Writing #2 → #1
Congrats again @GoogleDeepMind! The LLM race is on fire — progress is now measured in days!
See more analysis below👇 -
Oriol Vinyals
@OriolVinyalsML
What wall?
Arena.ai
@arena
Massive News from Chatbot Arena🔥
@GoogleDeepMind's latest Gemini (Exp 1114), tested with 6K+ community votes over the past week, now ranks joint #1 overall with an impressive 40+ score leap — matching 4o-latest in and surpassing o1-preview! It also claims #1 on Vision leaderboard.
Gemini-Exp-1114 excels across technical and creative domains:
- Overall #3 -> #1
- Math: #3 -> #1
- Hard Prompts: #4 -> #1
- Creative Writing #2 -> #1
- Vision: #2 -> #1
- Coding: #5 -> #3
- Overall (StyleCtrl): #4 -> #4
Huge congrats to @GoogleDeepMind on this remarkable milestone!
Come try the new Gemini and share your feedback! -
Oriol Vinyals
@OriolVinyalsML
Raptor 1: Traditional ML: SIFT + SPM + SVM
Raptor 2: Deep Learning: min_θ loss(NeuralNet(data, θ))
Raptor 3: LLMs: "Please solve this coding puzzle"
Raptor 4: ??? -
Oriol Vinyals
@OriolVinyalsML
Richly deserved @NobelPrize prize for @geoffreyhinton, @demishassabis, and John Jumper. I've had the privilege to work alongside these brilliant minds. Here are some personal insights into their remarkable qualities:
-When I joined Google, I was blown away by how much @geoffreyhinton codes and runs experiments. It was surreal to have him implement some ideas we were discussing around distillation in the early days, defying expectations of seniority.
-I've always been impressed by @demishassabis incredible strategic mind, and how accurate his predictions tend to be. In late summer, our AlphaStar agent could barely play the game and was beaten by the worst player in the team. Demis boldly decided to schedule a match against pro player @TLOgg in early December. Had that match been literally a week earlier, our agent wound't have been able to compete. Since then, I don't second guess Demis' calls anymore : )
-When the pandemic hit, I decided to go individual contributor mode and joined AlphaFold a few months before the submission to CASP14. Huge respect for John Jumper who led the team in a style which I strive to match: leading by example, knowing absolutely every single detail of the code / project, and commanding a massive amount of technical knowledge.
Congrats to them and the rest of awardees!! 🏅 -
Oriol Vinyals
@OriolVinyalsML
NaN papers accepted at #neurips2024
-
Oriol Vinyals
@OriolVinyalsML
It was always great fun to discuss ideas and results with @NoamShazeer and @dan_defr Welcome back & let's go! ♊️♊️🚀🚀
techcrunch.com/2024/08/02/character-ai-ceo-noam-shazeer-r… -
Oriol Vinyals reposted
sarah guo
@saranormous
New 🔥 @NoPriorsPod drop: @OriolVinyalsML
VP Deep Learning @GoogleDeepMind. Topics:
*Gemini
*infinite context windows
*where compute goes: pre-training/post-training/test-time search
*reward modeling beyond games
*how we should prep for AGI -
Oriol Vinyals
@OriolVinyalsML
Our experimental version of Gemini 1.5 Pro in AI Studio debuted at #1 on the LMSys leaderboard. 1300 ELO, nice round number 🔥
With the Gemini team, we are focused on both improving our models by iterating and finessing our recipe (deep learning FTW!), and also on massive bets that could yield big quality jumps. If you have what it takes, join us!
boards.greenhouse.io/deepmind/jobs/6143663
Arena.ai
@arena
Exciting News from Chatbot Arena!
@GoogleDeepMind's new Gemini 1.5 Pro (Experimental 0801) has been tested in Arena for the past week, gathering over 12K community votes.
For the first time, Google Gemini has claimed the #1 spot, surpassing GPT-4o/Claude-3.5 with an impressive score of 1300 (!), and also achieving #1 on our Vision Leaderboard.
Gemini 1.5 Pro (0801) excels in multi-lingual tasks and delivers robust performance in technical areas like Math, Hard Prompts, and Coding.
Huge congrats to @GoogleDeepMind on this remarkable milestone!
Gemini (0801) Category Rankings:
- Overall: #1
- Math: #1-3
- Instruction-Following: #1-2
- Coding: #3-5
- Hard Prompts (English): #2-5
Come try the model and let us know your feedback!
More analysis below👇 -
Oriol Vinyals
@OriolVinyalsML
AI is beating me at the things I love, one step at a time (coding, StarCraft, mathematics, ...). AlphaProof solved the most difficult 2024 IMO problem (P6). Answer doesn't fit in this tweet 🥈
Google DeepMind
@GoogleDeepMind
We’re presenting the first AI to solve International Mathematical Olympiad problems at a silver medalist level.🥈
It combines AlphaProof, a new breakthrough model for formal reasoning, and AlphaGeometry 2, an improved version of our previous system. 🧵 dpmd.ai/imo-silver -
Oriol Vinyals
@OriolVinyalsML
Extremely proud of my partner in crime @FortunatoMeire and all of #graphcast team for this award! The thing is heavy 🪙🪙🪙
Demis Hassabis
@demishassabis
Huge congrats to Pete Battaglia & our #graphcast team who won the #macrobertaward, the UK’s top award for engineering innovation! GraphCast provides super fast and highly accurate weather predictions up to 10 days in advance, including early warnings for extreme weather events.
-
Oriol Vinyals
@OriolVinyalsML
Gemma 2 has arrived! The arena for open models is also heating up. Extra exciting as 27B is ~3x smaller than Llama3 70B and ~10x smaller than NVIDIA's Nemotron 340B 🔥♊️💙
Google DeepMind
@GoogleDeepMind
We're excited to unveil Gemma 2. 🛠️
Available in both 9B and 27B parameters, it delivers the best performance for its size - unlocking more possibilities for developers to build and deploy with AI. → dpmd.ai/45Q6yba -
Oriol Vinyals
@OriolVinyalsML
Scale In Peace 💙
Ilya Sutskever
@ilyasut
I am starting a new company:
SSI Inc.
@ssi
Superintelligence is within reach.
Building safe superintelligence (SSI) is the most important technical problem of our time.
We've started the world’s first straight-shot SSI lab, with one goal and one product: a safe superintelligence.
It’s called Safe Superintelligence Inc.
SSI is our mission, our name, and our entire product roadmap, because it is our sole focus. Our team, investors, and business model are all aligned to achieve SSI.
We approach safety and capabilities in tandem, as technical problems to be solved through revolutionary engineering and scientific breakthroughs. We plan to advance capabilities as fast as possible while making sure our safety always remains ahead.
This way, we can scale in peace.
Our singular focus means no distraction by management overhead or product cycles, and our business model means safety, security, and progress are all insulated from short-term commercial pressures.
We are an American company with offices in Palo Alto and Tel Aviv, where we have deep roots and the ability to recruit top technical talent.
We are assembling a lean, cracked team of the world’s best engineers and researchers dedicated to focusing on SSI and nothing else.
If that’s you, we offer an opportunity to do your life’s work and help solve the most important technical challenge of our age.
Now is the time. Join us.
Ilya Sutskever, Daniel Gross, Daniel Levy
June 19, 2024 -
Oriol Vinyals reposted
Arena.ai
@arena
Big news – Gemini 1.5 Flash, Pro and Advanced results are out!🔥
- Gemini 1.5 Pro/Advanced at #2, closing in on GPT-4o
- Gemini 1.5 Flash at #9, outperforming Llama-3-70b and nearly reaching GPT-4-0125 (!)
Pro is significantly stronger than its April version. Flash’s cost, capabilities, and unmatched context length make it a market game-changer!
Huge congrats to @GoogleDeepMind on the incredible
Gemini launches! Can't wait to see what new applications Gemini unlocks!
More breakdown analysis below👇 -
Oriol Vinyals
@OriolVinyalsML
For Reinforcement Learning to be successful on top of LLMs, it is critical to have a very powerful and accurate reward model. Reward in most language tasks isn't as clearly defined as, say, in chess, where the winning condition is a simple computation.
Generative reward models are the most powerful class as they are themselves general. It's great to see how the best LLMs are starting to top this important benchmark!
huggingface.co/spaces/allenai/reward-bench -
Oriol Vinyals
@OriolVinyalsML
Today we have published our updated Gemini 1.5 Model Technical Report. As @JeffDean highlights, we have made significant progress in Gemini 1.5 Pro across all key benchmarks; TL;DR: 1.5 Pro > 1.0 Ultra, 1.5 Flash (our fastest model) ~= 1.0 Ultra.
As a math undergrad, our drastic results in mathematics are particularly exciting to me!
In section 7 of the tech report, we present new results on a math-specialised variant of Gemini 1.5 Pro which performs strongly on competition-level math problems, including a breakthrough performance of 91.1% on Hendryck’s MATH benchmark without tool-use (examples below 🧵).
Gemini 1.5 is widely available, try it out for free here aistudio.google.com & read the full tech report here: goo.gle/GeminiV1-5
Oriol Vinyals
@OriolVinyalsML
Here are some examples of the model solving problems from the Asian Pacific Mathematical Olympiad (APMO) that has stumped prior models. The top example is cool because it is a proof (rather than a calculation). The solutions are to the point and "beautiful".
This clearly shows the opportunity to build amazing systems on top of Gemini, like the recently announced Project Astra, or e.g. AlphaCode2, which was built on top of Gemini 1.0, and sits at an estimated 85th percentile on average on Code Contests, performing better than 85% of entrants. -
Oriol Vinyals reposted
Jeff Dean
@JeffDean
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 -
Oriol Vinyals
@OriolVinyalsML
We must go deeper.
Google DeepMind
@GoogleDeepMind
We watched #googleio with Project Astra. 👀
This media couldn't be displayed. See it on the original post. -
Oriol Vinyals
@OriolVinyalsML
Live from #googleio, we’re announcing significant updates to our Gemini family of models. More multimodal, real time, faster, better, longer context...
We can’t wait to see the amazing things people will do with this tech, as we continue making it more helpful and useful for billions of people.
Huge congrats & massive thanks to the Gemini team. So exciting & humbling to be working with a team with such high talent density. Join us if you want to help us build the future ♊️
dpmd.ai/gemini-flash -
Oriol Vinyals
@OriolVinyalsML
Gemini 1.5 Pro has entered the (LMSys) Arena! Some highlights:
-The only "mid" tier model at the highest level alongside "top" tier models from OpenAI and Anthropic ♊️
-The model excels at multimodal, and long context (not measured here) 🐍
-This model is also state-of-the-art in audio 🔊
-It is free 🤑
Try it out: aistudio.google.com
Arena.ai
@arena
More exciting news today -- Gemini 1.5 Pro result is out!
Gemini 1.5 Pro API-0409-preview now achieves #2 on the leaderboard, surpassing #3 GPT4-0125-preview to almost top-1!
Gemini shows even stronger performance on longer prompts, in which it ranks joint #1 with the latest GPT-4-Turbo🔥
Big congrats to @GoogleDeepMind on shipping this powerful Gemini API to developers & community. Very excited to see what app can be built on top! -
Oriol Vinyals reposted
Logan Kilpatrick
@OfficialLoganK
New @Google developer launch today:
- Gemini 1.5 Pro is now available in 180+ countries via the Gemini API in public preview
- Supports audio (speech) understanding capability, and a new File API to make it easy to handle files
- New embedding model!
developers.googleblog.com/2024/04/gemini-15-pro-in-public… -
Oriol Vinyals
@OriolVinyalsML
We are rolling out Gemini 1.5 Pro API so that you can keep building amazing stuff on top of the model like we've seen in the past few weeks.
Also, if you just want to play with Gemini 1.5, we removed the waitlist: aistudio.google.com
Last, but not least, we pushed the model to see if it can do 10M on all modalities. Seems to be working 🚀♊️
Jeff Dean
@JeffDean
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 -
Oriol Vinyals
@OriolVinyalsML
A new Large Language Model a day keeps the GPUs away.
-
Oriol Vinyals reposted
Mckay Wrigley
@mckaywrigley
The future of fixing bugs?
Just record them.
I filmed 3 separate bugs in an app and gave the videos to Gemini 1.5 Pro with my entire codebase.
It correctly identified & fixed each one.
AI is improving insanely fast.This media couldn't be displayed. See it on the original post. -
Oriol Vinyals reposted
rohan anil
@_arohan_
The team is working hard to bring audio inputs to the AI Studio interface for Gemini 1.5 Pro. We have an internal version that handles audio and video and can sample the video less frequently to increase the length of content that can be handled. @karpathy, thanks for the awesome tutorial. 🙏 Here's the prompts we used and the output for something in the direction you suggested.
@TeplyashinDenis @joe_stant @SavinovNikolay @machelreid @ankurbpn -
Oriol Vinyals
@OriolVinyalsML
It certainly has been a fun year @Google: enjoy playing with our open source models Gemma, built from the same research and technology used to create the Gemini models. 💙♊️🚀
Blog: blog.google/technology/developers/gemma-open-models/
Tech report: storage.googleapis.com/deepmind-media/gemma/gemma-report.pdf -
Oriol Vinyals
@OriolVinyalsML
Gemini 1.5 Pro launched last week and already we're seeing the community produce some amazing interactions with long context. Below 👇 are some highlights and cool posts from folks who have gotten early access.
A few thoughts from discussions / reactions from the community so far:
-RAG ("retrieval-augmented generation") isn’t done for, even though we can do 1M or more tokens in context now. In fact, RAG has some nice properties that can enhance (and be enhanced by) long context. RAG allows you to find relevant information, but the way the model accesses it may be too restrictive due to compression. Long context may help bridge that gap, similar to how L1/L2 cache & main memory work together in modern CPUs.
-Regarding questions about in-context learning abilities, and whether the documents users may put in the context have also been trained on, I find the distinction irrelevant. By putting a document (which the model may have been trained on) in memory, you help it in many different ways: you "refresh" its knowledge about it; you allow the model to reason about complex queries related to the document or do complex computation and reasoning over the relevant knowledge directly. For example, in the Kalamang example, it’s more powerful to put the Kalamang grammar book directly in the prompt, allowing the model to reason about the language more powerfully and "in context", than it would be to have that book fully embedded in the network weights.
-We are also looking to land many inference optimizations that will make the models substantially faster. Stay tuned!
Thread with cool interactions with Gemini 1.5 Pro. Keep them coming! 🧵Post not available here, view on X
Oriol Vinyals
@OriolVinyalsML
Taelin
@VictorTaelin
Gemini 1.5 destroyed GPT-4-Turbo. I guess long contexts overwhelm it, while Gemini 1.5 doesn't care at all. Complete breakdown:
www.reddit.com/r/singularity/comments/1atjz9v/ive_put_a_c…?
Oriol Vinyals
@OriolVinyalsML
Sully
@SullyOmarr
Ok, gemini 1.5 is really really good.
seriously the most impressive jump i've seen in models in a while when it comes to long context
just tried with a paper and asked it "what does Figure 5 show?"
it contextualized the whole thing and answered based on that tiny section + figure.Post not available here, view on X
Oriol Vinyals
@OriolVinyalsML
Rob Lennon 🗯 | AI Whisperer
@thatroblennon
Spent the last few days building a megaprompt for a client.
This one takes in a 1-hour meeting transcript and outputs a report that becomes the basis for a paid consulting deliverable.
Ran it in 4 LLMS:
- ChatGPT (GPT-4 Turbo)
- GPT-4 Turbo in OpenAI Playground
- Claude-2 via Poe
- Gemini 1.5 Pro*
(Note, this is the closed beta that most people don't have access to. This model is smaller than Gemini Ultra but has a huge context window. Gemini Ultra couldn't handle a prompt of this size.)
GPT-4 Turbo was adequate and accurate.
Claude-2 had a more nuanced analysis but wasn't quite as thorough as GPT-4.
Gemini 1.5 Pro provided the best of both worlds. Kind of like the thoroughness of GPT-4 with the nuance of Claude-2.
Oriol Vinyals
@OriolVinyalsML
elvis
@omarsar0
The Gemini 1.5 Pro model guide is live!
With support of up to 1 million tokens context length, you may be wondering what's possible with Gemini 1.5 Pro.
My overall impression after our first round of testing is that Gemini 1.5 Pro is among the most powerful long context LLMs available today.
I've published a summary of Gemini 1.5 Pro's capabilities along with concrete examples in the prompting guide. These are just preliminary tests. I will continue to analyze and document the model's capabilities and limitations. Stay tuned!
From preliminary experiments, Gemini 1.5 Pro shows impressive capabilities around multimodal reasoning, video understanding, long document question answering, code reasoning on entire codebases, and in-context learning.
One insight from testing this model is that we will have different kinds of LLMs that support different types of use cases. Gemini 1.5 Pro is not meant to be a model to reign among all. The long context LLMs are not meant to cover every use case imaginable, they are meant to unlock complex use cases that were unimaginable before with LLMs.
Link to guide below ↓ -
Oriol Vinyals
@OriolVinyalsML
Amazing job from my friends @OpenAI on Sora! Also, sorry if I spoil this video for you, but I can't unsee the cats' 3 front legs 😆
-
Oriol Vinyals reposted
Jeff Dean
@JeffDean
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! -
Oriol Vinyals
@OriolVinyalsML
To show what’s possible with the drastically huge context window in Gemini 1.5 Pro, we prompted it with the three.js examples code - over 100,000 lines of code/800k+ tokens!
(That’s not even the max, it can handle millions of tokens 😀)
Gemini was able to process all the code in one go and correctly answer questions about it:
-When we asked Gemini to find the top 3 examples within the codebase to help us learn a specific skill, it looked across 100s of possible examples and came back with super-relevant options. Understand ✅
-Gemini was able to locate a specific animation from within the set, correctly reason and explain what code is used to control it, and show us exactly how to customize that code to make a specific adjustment to the animation. Code ✅
-Given only a screenshot from one of the demos in the codebase, Gemini was able to find the right demo - and then clearly explain how to modify the code to achieve a specific change to the image. Edit through images ✅
-When we asked Gemini about changing the text and style in an example, it identified the exact lines of code to change, showed us how to change them, and gave an explanation about what had been done and why. ✅✅✅
Just a few examples of what’s possible! Even more info and examples in the blog. Obviously, like with any machine learning model, it sometimes doesn’t get it right. That’s why Humans + AI is the real deal!
Blog: blog.google/technology/ai/google-gemini-next-generation-m…
Congrats to the teams for this drastic model! 1.5 🚀 ♊💙This media couldn't be displayed. See it on the original post. -
Oriol Vinyals
@OriolVinyalsML
Gemini 1.5 has arrived. Pro 1.5 with 1M tokens available as an experimental feature via AI Studio and Vertex AI in private preview.
Then there’s this: In our research, we tested Gemini 1.5 on up to 2M tokens for audio, 2.8M tokens for video, and 🤯10M 🤯 tokens for text. From Shannon’s 1950s bi-gram models (2 tokens), and after being mesmerized by LSTMs many years ago able to model 200 tokens, it feels almost impossible that I would be talking about hundreds of thousands of tokens in context length, let alone millions. ♊️💙
Tech report: goo.gle/GeminiV1-5
Demis Hassabis
@demishassabis
In December we began the Gemini Era, and we’ve continued to make relentless progress since. Today we’re thrilled to introduce the next generation: Gemini 1.5 - hugely enhanced performance, highly efficient architecture & long-context length breakthrough
blog.google/technology/ai/google-gemini-next-generation-m… -
Oriol Vinyals reposted
Jascha Sohl-Dickstein
@jaschasd
Have you ever done a dense grid search over neural network hyperparameters? Like a *really dense* grid search? It looks like this (!!). Blueish colors correspond to hyperparameters for which training converges, redish colors to hyperparameters for which training diverges.
This media couldn't be displayed. See it on the original post. -
Oriol Vinyals
@OriolVinyalsML
Just over a year ago, a few of us got together to brainstorm a brand new project called Gemini ♊️. So grateful for the impact across @Google. Huge thanks to everyone who's supported this journey – and we ain't slowing down! 🚀🌊 #gemini #ai
Google
@Google
Today, we’re entering the next chapter of our Gemini era by bringing our #geminiai models to more of our products, starting with Bard — which will now be called Gemini. goo.gle/4bxymmT
This media couldn't be displayed. See it on the original post. -
Oriol Vinyals
@OriolVinyalsML
Fun game. Clocking 17973 citations: "Distilling the knowledge in a neural network" @geoffreyhinton, @OriolVinyalsML, @JeffDean
Reviewer 38 (NeurIPS 2014): "This work is incremental and unlikely to have much impact even though it may be technically correct and well executed."
Yann LeCun
@ylecun
My most-cited, never-accepted, ArXiv-only paper has over 1880 citations.
"Deep Convolutional Networks on Graph-Structured Data"
Mikael Henaff, Joan Bruna, Yann LeCun
arxiv.org/abs/1506.05163 -
Oriol Vinyals
@OriolVinyalsML
Evaluation of LLMs is very hard and nuanced (especially academic evals which are leaked massively). Evals that rely on human judgement are far superior, so it feels good that Bard Gemini Pro (free tier) climbed pretty high on lmsys🪜 Looking forward to Gemini Ultra release! #ultraiscoming 🌊♊️
Try it at bard.google.com
Arena.ai
@arena
🔥Breaking News from Arena
Google's Bard has just made a stunning leap, surpassing GPT-4 to the SECOND SPOT on the leaderboard! Big congrats to @Google for the remarkable achievement!
The race is heating up like never before! Super excited to see what's next for Bard + Gemini Ultra release. -
Oriol Vinyals
@OriolVinyalsML
Hi 2024! 🎉
Summary of 2023 for many of the people I love working with: Gemini ♊️
May 2024 bring us more capable Neural Networks, loss functions converge quickly, and may we never see an Inf or a NaN -
Oriol Vinyals
@OriolVinyalsML
Incredibly proud of the team behind Imagen 2. What a week!
My attempt at generating a logo for Gemini ♊️. Help me generate a better one please, I'm not great at prompting 🙃
Google DeepMind
@GoogleDeepMind
Meet Imagen 2: our most advanced text-to-image diffusion technology. ✨
It features high-quality, photorealistic outputs and stronger consistency with your prompts. 🖼
Now available to use via @GoogleCloud’s #vertexai platform. → dpmd.ai/x-imagen2 -
Oriol Vinyals
@OriolVinyalsML
Time to build! 🔧♊️
Sundar Pichai
@sundarpichai
Today developers can start building with our first version of Gemini Pro through Google AI Studio at ai.google.dev.
Developers have a free quota and access to a full range of features including function calling, embeddings, semantic retrieval, custom knowledge grounding, chat functionality and more. It supports 38 languages across 180+ countries. Gemini Ultra is coming early next year. We’re excited to see what you build! blog.google/technology/ai/gemini-api-developers-cloud/ -
Oriol Vinyals
@OriolVinyalsML
On my way to @NeurIPSConf #neurips2023 where I'm presenting 0 papers 🙃
Come chat with the Gemini ♊️ team at @GoogleDeepMind / @GoogleResearch booths on Tue/Wed from 1:30 to 3:00 PM (NOLA time). Quite a few Gemini team members will be there!
Oriol Vinyals
@OriolVinyalsML
Great to meet so many awesome people today @NeurIPSConf!
-
Oriol Vinyals
@OriolVinyalsML
Why "Gemini"? ♊
-I liked the name because the monumental effort of training LLMs resonated with the spirit of launching rockets 🚀. Gemini is @NASA's program which preceded Apollo.
-Merging two teams together led @JeffDean to consider "Gemini" a fitting name.
Thrilled for what is yet to come! 🚀 ♊
Jeff Dean
@JeffDean
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. -
Oriol Vinyals
@OriolVinyalsML
Really happy to see the interest around our “Hands-on with Gemini” video. In our developer blog yesterday, we broke down how Gemini was used to create it. developers.googleblog.com/2023/12/how-its-made-gemini-mul…
We gave Gemini sequences of different modalities — image and text in this case — and had it respond by predicting what might come next. Devs can try similar things when access to Pro opens on 12/13 🚀. The knitting demo used Ultra⚡
All the user prompts and outputs in the video are real, shortened for brevity. The video illustrates what the multimodal user experiences built with Gemini could look like. We made it to inspire developers.
When you’re building an app, you can get similar results (there’s always some variability with LLMs) by prompting Gemini with an instruction that allows the user to "configure" the behavior of the model, like inputting “you are an expert in science …” before a user can engage in the same kind of back and forth dialogue. Here’s a clip of what this looks like in AI Studio with Gemini Pro. We’ve come a long way since Flamingo 🦩 & PALI, looking forward to seeing what people build with it! -
Oriol Vinyals
@OriolVinyalsML
seq💜💙❤️seq
@ilyasut @OriolVinyalsML @quocleix
Sasha Rush
@srush_nlp
2014->2024, 100M->1T, same people, pretty similar model. It's such a crazy story.
-
Oriol Vinyals
@OriolVinyalsML
Exciting times, welcome Gemini (and MMLU>90)! State-of-the-art on 30 out of 32 benchmarks across text, coding, audio, images, and video, with a single model 🤯
Co-leading Gemini has been my most exciting endeavor, fueled by a very ambitious goal. And that is just the beginning! A long 🐍 post about our Gemini journey & state of the field.
The biggest challenges in LLMs are far from trivial or obvious. Evaluation and data stand out to me. We've moved beyond the simpler "Have we won in Go/Chess/StarCraft?" to “Is this answer accurate and fair? Is this conversation good? Does this complex piece of text prove the theorem?” Exciting potential coupled with monumental challenges.
The field is less ripe further down the model pipeline. Pretraining is relatively well understood. Instruction tuning and RLHF, less so. In AlphaGo and AlphaStar we spent 5% of compute in pre-training and the rest in the very important RL phase, where the model learns from its successes or failures. In LLMs, we spend most of our time on pretraining. I believe there’s huge potential to be untapped. Cakes with lots of cherries, please 🎂
@Google has demonstrated its ability to move fast. It has been an absolute blast to see the energy from my colleagues and the support received. A “random” highlight is coauthoring our tech report with a co-founder. Another is coleading with @JeffDean. But beyond individuals, Gemini is about teamwork: it is important to recognize the collective effort behind such achievements. Picture a room full of brilliant people, and avoid attributing success solely to one person.
On a personal note, recently I celebrated my 10 year anniversary at Google, and it’s been 8 years since @quocleix and I co-authored “A Neural Conversational Model”, which gave us a glimpse of what was, has, and is yet to come. Back then, that line of work received a lot of skepticism. Lessons learned: whatever your passion is, push for it!
Zooming back out, there’s lots of change in our field, and the stakes couldn’t be higher. Excited for what’s to come from Gemini, but humbled by the responsibility to “get it right”. 2024 will be drastic. Welcome Gemini!
blog.google/technology/ai/google-gemini-ai/ -
Oriol Vinyals reposted
Jeff Dean
@JeffDean
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. -
Oriol Vinyals
@OriolVinyalsML
Curiously, AlphaGo was using imitation learning (cake, yummy 🎂) + RL refinement (cherry, yes plz 🍒). A recipe that should sound familiar to many today.
Aravind Srinivas
@AravSrinivas
This also happened in 2016. Whole world was behind end to end RL as the solution to general intelligence after AlphaGo. And @ylecun gave a visionary and controversial (then, obvious now) talk on how we should instead work on generative models and self-supervised learning.
-
Oriol Vinyals
@OriolVinyalsML
Welcome (back) ((hopefully 🎢))!
-
Oriol Vinyals
@OriolVinyalsML
💙 ❤️ 💛💙💚♥️
Sam Altman
@sama
i love the openai team so much
-
Oriol Vinyals
@OriolVinyalsML
Music + AI == ❤️ I've been playing with this system for a while and it's incredible. Very proud of the teams who worked so hard on #lyria, and looking forward to see what creative minds can achieve with this! Music will never be the same 🎶🎶🎶
Blog: deepmind.google/discover/blog/transforming-the-future-of-…
Sample: www.youtube.com/watch?v=1gjuHUy0IMM&t=7s -
Oriol Vinyals
@OriolVinyalsML
Tweet2Paper service: strong approve.
-
Oriol Vinyals
@OriolVinyalsML
It took a lot of Deep RL finesse (& some Transformers ❤️) to get the first change in over a decade accepted into the LLVM standard C++ library. AlphaDev, published @Nature by my brilliant colleagues!
www.deepmind.com/blog/alphadev-discovers-faster-sorting-a… -
Oriol Vinyals
@OriolVinyalsML
"Context length" is the new "model size" 🤔
-
Oriol Vinyals
@OriolVinyalsML
Best feeling in the (Neural Network) world: loss is going down 📉 Happy Sunday!🌞
-
Oriol Vinyals
@OriolVinyalsML
Thanks for all you've done for many of us, @geoffreyhinton Keep distilling your knowledge back to the world, and may your gradients never vanish!
Geoffrey Hinton
@geoffreyhinton
In the NYT today, Cade Metz implies that I left Google so that I could criticize Google. Actually, I left so that I could talk about the dangers of AI without considering how this impacts Google. Google has acted very responsibly.
-
Oriol Vinyals reposted
Tivadar Danka
@TivadarDanka
I described some of the most beautiful and famous mathematical theorems to Midjourney.
Here is how it imagined them:
1. "The set of real numbers is uncountably infinite." -
Oriol Vinyals
@OriolVinyalsML
𝗚𝗼𝗼𝗴𝗹𝗲 𝗗𝗲𝗲𝗽𝗠𝗶𝗻𝗱
Oriol Vinyals
@OriolVinyalsML
-
Oriol Vinyals
@OriolVinyalsML
Introducing Fictitious Optimization & Obfuscation Learning. FOOL is a cutting-edge method based on the latest in artificial stupidity & irrational reasoning. Its key advance is to randomly change the direction of gradients during optimization. It may also slow down AGI research.
-
Oriol Vinyals
@OriolVinyalsML
To whoever invented USB-C: THANK YOU!!
-
Oriol Vinyals reposted
Sundar Pichai
@sundarpichai
We're expanding access to Bard in US + UK with more countries ahead, it's an early experiment that lets you collaborate with generative AI. Hope Bard sparks more creativity and curiosity, and will get better with feedback. Sign up: bard.google.com
blog.google/technology/ai/try-bard/ -
Oriol Vinyals
@OriolVinyalsML
Criticism is easy, and gets a lot of attention these days. And, as we all know, attention is all (some people) need. To those who build: you rock!
Christopher Manning
@chrmanning
This is truly an opinion piece. Not even a cursory attempt is made to check easily refutable claims (“they may well predict, incorrectly”). Melodramatic claims of inadequacy are made not of specific current models but any possible machine learning approach
www.nytimes.com/2023/03/08/opinion/noam-chomsky-chatgpt-a… -
Oriol Vinyals
@OriolVinyalsML
Making predictions <<< Making them happen
-
Oriol Vinyals
@OriolVinyalsML
Welcome to @DeepMind! Looking forward to working together on drastic research! Also, an additional excuse to visit New York and the Bay Area 🌉🗽🇺🇸
Behnam Neyshabur
@bneyshabur
Excited to announce that the entire Blueshift team has joined @DeepMind! We will be working with @OriolVinyalsML and others to advance capabilities of LLMs developed by DM / Alphabet! We hope to continue to grow DM's presence in Bay Area and New York in the coming months :-)
-
Oriol Vinyals
@OriolVinyalsML
Language modeling is compression gone wrong.
-
Oriol Vinyals
@OriolVinyalsML
Chain ⛓️ Rule(s) rules! Appreciation thread of one of the most interesting coincidences in machine learning. Two rules, both named "Chain Rule", happen to be absolutely critical to recent advances in ML & AI. A 🧵 on the Chain Rule of Probability & the Chain Rule of Calculus👇
Post not available here, view on X
Oriol Vinyals
@OriolVinyalsML
This is very important in Deep Learning: it allows calculating derivatives of highly complex neural networks, composed of stacked layers of simple functions. Gradient descent is used to update the neural net weights, allowing it to learn from data and perform complex tasks.
Oriol Vinyals
@OriolVinyalsML
The thread above was, of course, written in collaboration with one of our powerful Large Language Models. Long live Chain ⛓️ Rules!
-
Oriol Vinyals
@OriolVinyalsML
Zero accepted papers to #iclr2023 😢 / Zero submitted 🙃. Acceptance rate = NaN #achievementunlocked
-
Oriol Vinyals
@OriolVinyalsML
Many people mistakenly think that large language models that generate one word at a time is the end game.
Connecting LLMs with tools, e.g.: search engines, python interpreters, etc., is super exciting, and leverages tools' power, robustness & more!
Oriol Vinyals
@OriolVinyalsML
E.g. our own Sparrow connects a chatbot to @Google search. A cool demo by @perplexity_ai shows something similar, and also allows you to query a @Twitter database for e.g. "histogram of top words from my tweets" (somehow I seem to use @ylecun a lot???).
Oriol Vinyals
@OriolVinyalsML
Forgot the actual histogram 🙃
-
Oriol Vinyals
@OriolVinyalsML
Protein folding competition #casp15 has followed the same pattern observed in computer vision back in 2012, revolutionized by Deep Learning. Following dominance of #alphafold2 in #casp14, all top methods this year build on top of AF2.
Oriol Vinyals
@OriolVinyalsML
#alphafold2 indeed was the ImageNet moment in protein folding. What field will be next? Gotta ❤️ neural nets!
Oriol Vinyals
@OriolVinyalsML
Biology's ImageNet moment has arrived: #alphafold2 solves Protein Folding at the #casp14 competition!
Blog: deepmind.com/blog/article/alphafold-a-solution-to-a-50-ye… -
Oriol Vinyals
@OriolVinyalsML
The "Parents AI test". My dad just sent me a picture and told me: "there's no way AI can guess what this is". Challenge accepted! #flamingo
-
Oriol Vinyals
@OriolVinyalsML
Congrats Argentina 🇦🇷 and Messi! Also pictured: Harry Potter and the Goblet of Fire 🪄🏆
FIFA World Cup
@FIFAWorldCup
The moment he's dreamed about 🏆
-
Oriol Vinyals
@OriolVinyalsML
So proud of the team behind #alphacode, published today @ScienceMagazine!
Paper: scim.ag/Nn
Oriol Vinyals
@OriolVinyalsML
Beautiful cover page too!
-
Oriol Vinyals reposted
Google DeepMind
@GoogleDeepMind
In @ScienceMagazine, we present #alphacode - the first AI system to write computer programs at a human level in competitions.
It placed in the top 54% of participants in coding contests by solving new and complex problems.
How does it work? 🧵 dpmd.ai/alphacode-science -
Oriol Vinyals
@OriolVinyalsML
2001: A Space Odyssey appreciation 🧵 I just rewatched 2001 for the n-th time (n>20). This movie was hugely ahead of its time, so many cool topics:
-Origin of (human) intelligence 🧠
-Pushing frontiers of space exploration 🚀
-And, of course, AI 🤖
Oriol Vinyals
@OriolVinyalsML
A few notes on 2001's AI:
-HAL is a wink to IBM (H+1 A+1 L+1=IBM). IBM's DeepBlue was "predicted" in the movie when HAL beat Frank at chess : )
-Conversational AI is at the center of current research.
-As AI advances, topics such as alignment, safety, etc., become critical.
Oriol Vinyals
@OriolVinyalsML
If you do watch it, pretend to be in 1968, when the film was originally released. I'd also highly recommend Christopher Nolan's 2021 4K "unrestored" version. 2001 is not only my all time favorite movie, but the reason why I chose science and AI.
-
Oriol Vinyals
@OriolVinyalsML
Time to test your Spanish skills! : )
EL PAÍS
@el_pais
"Creo que conseguir crear una inteligencia artificial general será uno de los avances científicos más profundos que podrá alcanzar la humanidad, porque ni siquiera entendemos nuestra propia inteligencia", considera Vinyals. elpais.com/tecnologia/2022-12-03/oriol-vinyals-nuestra-ge…
-
Oriol Vinyals
@OriolVinyalsML
Playing video games at a "pro" level can help productivity at work thanks to increased keyboard / mouse speed & precision. I certainly "micro" code and emails better thanks to @StarCraft. Careful balance is needed though 🤣
-
Oriol Vinyals
@OriolVinyalsML
Beautiful exhibit from @fpa paying rendition to AlphaFold. I tried to get diffusion models to generate something as nice, but couldn't. Humans > AI : )
-
Oriol Vinyals reposted
khipu.ai
@Khipu_AI
There is 1 week left to apply to join us at Khipu 2023! ⏳
It's simple to apply to attend Khipu and there is no requirement to submit a paper, so head over to our website khipu.ai now to get your application started! 🚀
#khipu2023 -
Oriol Vinyals
@OriolVinyalsML
This neural network architecture that was showcased at the @Tesla AI day is a perfect example of Deep Learning at its finest. Mix and match all the greatest innovations to do something drastic and super ambitious. Congrats!
Post not available here, view on X
Oriol Vinyals
@OriolVinyalsML
This isn't ML-new, but who cares? Applied at the level of ambition of full-scale real world impact, with the right team, execution, (and compute/data!), you can do things that felt impossible before. Both the architecture and cool use of language heavily reminded me of AlphaStar.
Oriol Vinyals
@OriolVinyalsML
With all the debates on who-claimed-what-when, ego wars, research-prophecies, and other non-sense that is filling my Twitter feed these days, it is sobering to see what good execution can achieve. Congratulations to all those involved in this and similar projects!
-
Oriol Vinyals
@OriolVinyalsML
As a "failed" scientist => converted deep learner, the dream is to be able to influence & advance major scientific fields. If you are interested in working with us @DeepMind on weather & climate modeling, please apply!
boards.greenhouse.io/deepmind/jobs/4580940
Scott Duncan
@ScottDuncanWX
When a major hurricane meets the jet stream...
-
Oriol Vinyals
@OriolVinyalsML
Interested in exploring the limits of large-scale models? Consider joining our brand new Scalable Deep Learning team in California as a Research Engineer or Research Scientist, see job descriptions below!
boards.greenhouse.io/deepmind/jobs/4452868?t=bbda0eea1us
boards.greenhouse.io/deepmind/jobs/4445808?t=bbda0eea1us -
Oriol Vinyals
@OriolVinyalsML
Sundar Pichai
@sundarpichai
One year after open-sourcing its AlphaFold Protein Structure Database, @DeepMind and @emblebi are expanding it from nearly 1M to over 200M protein structures, covering almost every organism that has had its genome sequenced, a huge milestone.
-
Oriol Vinyals
@OriolVinyalsML
It's out! Thanks @lexfridman for the interesting conversation! I will make sure to acknowledge you in some future papers as our chat already sparked interesting research ideas!
Lex Fridman
@lexfridman
Here's my conversation with @oriolvinyalsml, Research Director and Deep Learning Lead at @DeepMind. He is one of the most brilliant researchers in AI, and so it's always fun to sit down with him to chat about both technical & philosophical topics. www.youtube.com/watch?v=aGBLRlLe7X8
-
Oriol Vinyals reposted
Lex Fridman
@lexfridman
Here's my conversation with @oriolvinyalsml, Research Director and Deep Learning Lead at @DeepMind. He is one of the most brilliant researchers in AI, and so it's always fun to sit down with him to chat about both technical & philosophical topics. www.youtube.com/watch?v=aGBLRlLe7X8
-
Oriol Vinyals
@OriolVinyalsML
Tomorrow I'll be talking about pre-training at 9:30AM EST in the massive Hall F #icml2022 room. Come find out what chinchillas🐹, flamingos🦩 and Gatos😺 have in common!
pretraining.github.io
icml.cc/virtual/2022/workshop/13457 -
Oriol Vinyals
@OriolVinyalsML
Great to be at #icml2022! I missed poster sessions in person. Deep Learning Is All I Need ♥️
-
Oriol Vinyals
@OriolVinyalsML
Amazing progress since I visited and tried to "teach" the robots how to play. 🏓🏓🏓
Laura Graesser
@lgraesser3
🏓So excited to share i-S2R in which we train robots to play table tennis cooperatively with humans for up to 340 hit rallies. That’s 4+ mins of play without dropping the ball. The robot makes the human work hard to keep up! 🏓
sites.google.com/view/is2r
arxiv.org/abs/2207.06572
1/nThis media couldn't be displayed. See it on the original post. -
Oriol Vinyals reposted
Ian Goodfellow
@goodfellow_ian
I'm excited to announce that I've joined DeepMind! I'll be a research scientist in @OriolVinyalsML 's Deep Learning team.
-
Oriol Vinyals
@OriolVinyalsML
It is great to welcome @goodfellow_ian to the Deep Learning team and @DeepMind!
Ian and I intersected briefly whilst he was finishing THE Deep Learning book, but still were able to work together on a cool paper arxiv.org/abs/1412.6544 Looking forward to much more!
Ian Goodfellow
@goodfellow_ian
I'm excited to announce that I've joined DeepMind! I'll be a research scientist in @OriolVinyalsML 's Deep Learning team.
-
Oriol Vinyals
@OriolVinyalsML
Had a great time chatting with @lexfridman earlier today, looking forward to the episode!!
Lex Fridman
@lexfridman
I'm visiting DeepMind soon to talk about AI. Should be fun.
-
Oriol Vinyals reposted
Liam Fedus
@LiamFedus
Presenting our survey on emergent abilities in LLMs!
arxiv.org/abs/2206.07682
What's it about? Certain downstream language tasks exhibit an interesting behavior: eval curves are flat/random up to a certain model scale, until -- poof -- things start to work.
1/7 -
Oriol Vinyals reposted
Mai Giménez 🖤🔻
@maidotgimenez
Out of context this is probably the best tweet #academictwitter will ever have. Meow 🐈⬛
Oriol Vinyals
@OriolVinyalsML
Gato is not the end. Gato is the beginning. Meow 🐈
-
Post not available here, view on XPost not available here, view on X
Oriol Vinyals
@OriolVinyalsML
Gato is not the end. Gato is the beginning. Meow 🐈
-
Oriol Vinyals
@OriolVinyalsML
Summary of the last month (!) of projects released in Deep Learning. Huge congrats to the teams involved! #imagen #flamingo #dalle2
Post not available here, view on X
Oriol Vinyals
@OriolVinyalsML
🤯
Oriol Vinyals
@OriolVinyalsML
Original pictures from #imagen in full res.
-
Oriol Vinyals
@OriolVinyalsML
The most exciting aspect of our recent LM 🐹, VLM 🦩 and Gato 🐈 works is not that these models are ~SOTA on the tasks they've trained on, but that they are reasonably good at ANY task**. A 🧵 (1/4)
Post not available here, view on X
Oriol Vinyals
@OriolVinyalsML
We've come a long way, but there's more to do! **These models aren't really good at ANY task, but MANY. Good means they sometimes succeed, but are below human capabilities. Also, our datasets reflect real world biases. If you are interested in any of these, join us and help! : )
Oriol Vinyals
@OriolVinyalsML
PS Metal Learning🤘
-
Oriol Vinyals
@OriolVinyalsML
Happy that my friend @ilyasut has been elected a Fellow of the Royal Society! As he has also been gifting us some interesting tweets lately, I have challenged a Language Model 🐹 to see if it can compete. Results in thread🧵(1/8)
The Royal Society
@royalsociety
Dr Ilya Sutskever FRS is elected a Fellow of the Royal Society. He has been a major contributor to several of the most disruptive advances in artificial intelligence, from object recognition to natural language processing. #rsfellows royalsociety.org/people/ilya-sutskever-35834/
Post not available here, view on X
Oriol Vinyals
@OriolVinyalsML
"I think that the main contribution of deep learning is not that it's deep, but that it's learning. Learning is at the core of intelligence."
"The world is deep learning. And you're just a perceptron."
"Deep learning is a beautiful lie" (4/8)
Oriol Vinyals
@OriolVinyalsML
"The beauty of deep learning is that it is a bit like alchemy. You combine data, networks, and parameters, and out comes this almost magical thing called "intelligence"."
"AI is like the universe: most of the work is already done"
"there is no free lunch" (5/8)Post not available here, view on X
Oriol Vinyals
@OriolVinyalsML
"if it feels wrong, it's probably wrong"
"when you're sure you're right, you're probably wrong"
"the best way to win a debate is to not be in it"
"The most efficient way to travel is to not travel at all"
"the best time to change is always" (7/8)
Oriol Vinyals
@OriolVinyalsML
"the best way to predict the future is to make it happen"
Congrats again @ilyasut!! Would you tweet any of the above? 🤣 (8/8) -
Oriol Vinyals reposted
Jordi A.
@jordiae
AlphaStar (2019) vs. Gato (2022) architectures:
-
Oriol Vinyals reposted
Google DeepMind
@GoogleDeepMind
Gato🐈a scalable generalist agent that uses a single transformer with exactly the same weights to play Atari, follow text instructions, caption images, chat with people, control a real robot arm, and more: dpmd.ai/Gato
Paper: dpmd.ai/Gato-paper 1/ -
Oriol Vinyals reposted
JB Alayrac
@jalayrac
It is great to see the excitement about 🦩! As shown in different examples during the last few days, interacting with 🦩 has been quite fun, unique and sometimes mind blowing. However,🦩 has clear limitations as detailed in this 🧵! 1/11
-
Oriol Vinyals reposted
Roman Ring
@Inoryy
10 yrs ago @karpathy wrote a blog post on the outlook of AI: karpathy.github.io/2012/10/22/state-of-computer-vision/ in which he describes how difficult it would be for an AI to understand a given photo, concluding "we are very, very far and this depresses me."
Today, our Flamingo steps up to the challenge. -
Oriol Vinyals reposted
Drew Jaegle
@drew_jaegle
I’m hiring a Research Scientist to work on anymodal architectures and representation learning! If you’re interested in building systems for all the world’s data 📷⚗️🎙️📚🔬, please apply to @DeepMind's RS London role and mention you’d like to work with me boards.greenhouse.io/deepmind/jobs/469515
-
Oriol Vinyals reposted
Katie Millican
@millikatie
Data quality has a massive effect on the performance of multimodal models like Flamingo and DALL-E 2. If you'd like to do high-impact data research, get in touch!
boards.greenhouse.io/deepmind/jobs/4089743 -
Oriol Vinyals
@OriolVinyalsML
Flamingo, a drastic VLM that achieves SOTA on many vision + language tasks! Fun fact: (real) flamingos change colors depending on what they eat. Our Flamingo model behaves differently depending on the inputs it's fed 🍕🦩
www.deepmind.com/blog/tackling-multiple-tasks-with-a-sing…
Oriol Vinyals
@OriolVinyalsML
We should have totally incorporated the fun fact in the manuscript, to compete with this other 💎 from the distillation paper.
-
Oriol Vinyals
@OriolVinyalsML
Love this work! In very "deep learning" fashion, you can train a neural net to generate causal structure directly from observational data, end to end 🤯. Deep Learning❤️ Causal Reasoning
Nan Rosemary Ke
@rosemary_ke
Supervised causal induction: In this work, we learn to induce causal structure by
treating the inference process as a black box and design a neural network architecture
that learns the mapping from data to graph
structures via supervised training.arxiv.org/abs/2204.04875 -
Oriol Vinyals
@OriolVinyalsML
Chinchilla: A Neural Network Language Model, smaller yet more powerful, brought to you by updated scaling laws.
Google DeepMind
@GoogleDeepMind
Chinchilla: A 70 billion parameter language model that outperforms much larger models, including Gopher. By revisiting how to trade-off compute between model & dataset size, users can train a better and smaller model. Read more: dpmd.ai/chinchilla 1/3
-
Oriol Vinyals
@OriolVinyalsML
This is great work by my colleagues @KelseyRAllen, @neuro_kim and @ZePoLiTaT (et al). It shows why we should all be excited about Neural Simulators!
Kelsey Allen
@KelseyRAllen
Could machines design new tools for us? With my co-first authors @neuro_kim and @ZePoLiTaT, we are excited to announce "Physical Design using Differentiable Learned Simulators" where we solve complex, physical design problems by differentiating through GNN-based simulators. 1/9
-
Oriol Vinyals reposted
Katie Millican
@millikatie
We're building a brand new team at @DeepMind dedicated to data, and we're looking for Data Research Engineers to join us.
Feel free to reach out, or apply online here:
boards.greenhouse.io/deepmind/jobs/4089743 -
Oriol Vinyals reposted
Andrej Karpathy
@karpathy
New (small!) language model Chinchilla (70B) outperforms much larger Gopher (280B), GPT-3 (175B), Jurrasic-1 (178B), MT-NLG (530B) arxiv.org/abs/2203.15556 Important new LM scaling laws paper from DeepMind. Go smaller, train longer. Many misconfigurations likely continue to lurk.
-
Oriol Vinyals reposted
Yilun Du
@du_yilun
Monday (March 28th) at 4PM EST @OriolVinyalsML (@DeepMind) will give a in person talk at the @MIT_CSAIL Embodied Intelligence seminar about The Deep Learning Toolbox: From AlphaFold to AlphaCode! Check it out in person in 34-101 and the livestream at:
www.youtube.com/channel/UCnXGbvgu9071i3koFooncAw -
Oriol Vinyals
@OriolVinyalsML
... I'm a deep learning researcher and we've annoyed almost ALL statisticians : )
Daniela Witten
@daniela_witten
That’s nothing. I’m a statistician who does research and I’ve annoyed HUNDREDS of statisticians
Mark Zobeck
@MarkZobeck
I'm a doctor who does research.
I've annoyed DOZENS of statisticians.
Here are 5 proven ways to exasperate them.
#medtwitter #epitwitter #statstwitter -
Oriol Vinyals reposted
Karsten Roth
@confusezius
Visual Similarity 🤝 Language in
"Integrating Language Guidance into Vision-based Deep Metric Learning"
#cvpr22 @OriolVinyalsML @zeynepakata @ELLISforEurope
Equip visual similarity models w/ lang. semantics at little cost!
arxiv.org/abs/2203.08543
github.com/ExplainableML/LanguageGuidance_for_DML
1/3 -
Oriol Vinyals reposted
Karsten Roth
@confusezius
🚀Supercharge Metric Learning w/
"Non-isotropy Regularization for Proxy-based Deep Metric Learning"
#cvpr22 @OriolVinyalsML @zeynepakata @ELLISforEurope
Norm. Flows + proxies = strong generalization & super fast training.
arxiv.org/abs/2203.08547
github.com/ExplainableML/NonIsotropicProxyDML
1/4 -
Oriol Vinyals
@OriolVinyalsML
#alphacode paper FINALLY up on @arxiv! (it somehow took 6 weeks to go live): arxiv.org/abs/2203.07814
Also, wonderful commentary on some #alphacode solutions by Petr Mitrichev, a legend in competitive programming, up at alphacode.deepmind.com -
Oriol Vinyals
@OriolVinyalsML
Sepp Hochreiter
@HochreiterSepp
ArXiv arxiv.org/abs/2203.00555: Transformers up to 1,000 layers using a new normalization technique. Outperforms baselines on machine translation IWSLT-14 (De-En) and WMT-17 (En-De). New SOTA on multilingual benchmark with 7,482 translation directions with a 200-layer model.
-
Oriol Vinyals
@OriolVinyalsML
There is so much wind in London that our building is moving / swinging 😨😨😨 #stormeunice
-
Oriol Vinyals
@OriolVinyalsML
Google Maps 2.0 aka Neural Google Maps aka NERF Google Maps. Amazing!
AK
@_akhaliq
Block-NeRF: Scalable Large Scene Neural View Synthesis
abs: arxiv.org/abs/2202.05263
project page: waymo.com/research/block-nerf/This media couldn't be displayed. See it on the original post. -
Oriol Vinyals reposted
Danielle Belgrave
@DaniCMBelg
I'm hiring Research Scientists in DeepMind to work on the intersection of ML, science and healthcare. Apply for the @DeepMind RS role and express interest to work with me here: deepmind.com/careers/jobs/469515. My DMs are open, so feel free to reach out to me if you have questions.
-
Oriol Vinyals
@OriolVinyalsML
#alphacode: a drastic version of seq2seq & search. The Bitter Lesson strikes again, exposing the power of general methods that scale w/ increased compute! Like AlphaStar, an agent outperforms me at something I love, competitive programming 🤣 @codeforces
dpmd.ai/Alpha-Code -
Oriol Vinyals reposted
Google DeepMind
@GoogleDeepMind
Introducing #alphacode: a system that can compete at average human level in competitive coding competitions like @codeforces. An exciting leap in AI problem-solving capabilities, combining many advances in machine learning!
Read more: dpmd.ai/Alpha-Code 1/ -
Oriol Vinyals reposted
AK
@_akhaliq
Improving language models by retrieving from trillions of tokens
abs: arxiv.org/abs/2112.04426
With a 2 trillion token database, Retrieval-Enhanced Transformer (Retro) obtains comparable performance to GPT-3 and Jurassic-1 on the Pile, despite using 25× fewer parameters -
Oriol Vinyals
@OriolVinyalsML
Another great example of "The Deep Learning Devil is in the Details".
Yann LeCun
@ylecun
Some of the helpful tricks make complete sense: larger kernels, layer norm, fat layer inside residual blocks, one stage of non-linearity per residual block, separate downsampling layers....
2/N -
Oriol Vinyals
@OriolVinyalsML
2021 personal highlights, a🧵. Despite being a challenging year globally due to the pandemic 😷🦠, but thanks to many incredible collaborators, it's been an exciting year research-wise 🤖 Some highlights below.👇
Oriol Vinyals
@OriolVinyalsML
Diversity and inclusion. I kept engaged through our efforts @DeepMind, mentorship and as a member of @Khipu_AI community, supporting AI in Latin America, where we had a fireside chat w/ @geoffreyhinton (youtu.be/9ahlP4u_Ras). I was also a mentor for: docs.google.com/spreadsheets/d/1OR7xEIvJICLbQ9FA2eKNwaWFS…
Post not available here, view on X
Oriol Vinyals
@OriolVinyalsML
Linking many of the exciting topics of the year, I gave a talk @stanfordnlp about how many advances in NLP created some of the breakthroughs in agents, architectures, vision, and biology.
Talk: docs.google.com/presentation/d/e/2PACX-1vQgDdaVQYe_hDDy6-…
Oriol Vinyals
@OriolVinyalsML
This could not have been possible without great collaborators from @DeepMind & beyond, and the amazing research community, who keep pushing the boundaries of what's possible with machine learning whilst trying to make our field better & more inclusive. Looking forward to 2021++!
-
Oriol Vinyals reposted
The TWIML AI Podcast
@twimlai
“All the data is just a sequence of bytes. You can represent any data structure, input, or output as a sequence of bytes. And that is a very romantic way to think about machine learning,” - @OriolVinyalsML of @DeepMind
-
Oriol Vinyals
@OriolVinyalsML
This is a great thread, +1e100. @karpathy didn't mention my (biased!) favorite example of 2021. Models designed to generate words (transformers) & model language (BERT) were reused in #alphafold to solve the protein folding problem, mapping a bunch of letters, to 3D coordinates🤯
Andrej Karpathy
@karpathy
The ongoing consolidation in AI is incredible. Thread: ➡️ When I started ~decade ago vision, speech, natural language, reinforcement learning, etc. were completely separate; You couldn't read papers across areas - the approaches were completely different, often not even ML based.
-
Oriol Vinyals
@OriolVinyalsML
"Extrapolating from a Single Image to a Thousand Classes using Distillation". Good title! @geoffreyhinton actually tried something like this in the original distillation paper. He called it "mythical digits" 🦄😆🧵👇
arxiv.org/abs/2112.00725
cc @y_m_asano @aaqib_saeed
Oriol Vinyals
@OriolVinyalsML
From arxiv.org/abs/1503.02531, what we did on #mnist is to entirely remove a single class from the training set, and showed that distillation was able to still classify that "mythical digit" successfully. In the new paper, they show you can ignore all classes / supervision. 👍👍👍
-
Oriol Vinyals
@OriolVinyalsML
This looks like a great survey on a great topic! (going to my "to read" stack : )). Clearly lots of work and ❤️ went into it. TRAIN=TEST 😆 Congrats to all the coauthors! @_robertkirk @yayitsamyzhang @egrefen @_rockt
arxiv.org/abs/2111.09794 -
Oriol Vinyals
@OriolVinyalsML
I will be a mentor at this wonderful event -- sign up soon before it's too late : ) Mentoring is one of my favorite things to do, as at least for me it is a great learning experience!
khipu.ai
@Khipu_AI
Mentoring opportunity to close the #khipu2021 Event Series. Tue Nov 23rd at 4PM Brasilia (7PM GMT) on GatherTown. It's a space to discuss your projects, ideas or questions with an expert on the topic of your choice. Free registration (DEADLINE NOV 16) khipu.ai/signup
-
Oriol Vinyals
@OriolVinyalsML
The intersection of AI + Art is fascinating! Our take on this is presented in Arnheim 2. As an example of what you could do with it, a drawing of @ylecun using only MNIST digits. #yamnist
Source code: github.com/deepmind/arnheim/blob/main/arnheim_2.ipynb
Upcoming AI+Art NeurIPS Workshop: @ML4CDworkshop
Oriol Vinyals
@OriolVinyalsML
Image created by @chrisantha_f & @MirowskiPiotr
-
Oriol Vinyals
@OriolVinyalsML
The Deep Learning Devil is in the Details. I love this work from @IrwanBello and collaborators in which they show how training "tricks" improve ~3% absolute accuracy on ImageNet, progress equivalent to years of developments and research!
Paper: arxiv.org/abs/2103.07579 -
Oriol Vinyals
@OriolVinyalsML
We released a few self-play games of AlphaStar vs AlphaStar as part of their training, brought to you by @LaughNgamez. Some unusual @StarCraft never seen before : )
www.youtube.com/watch?v=z2izAP-AI5M -
Oriol Vinyals reposted
Google DeepMind
@GoogleDeepMind
We’ve acquired the MuJoCo physics simulator (mujoco.org) and are making it free for all, to support research everywhere. MuJoCo is a fast, powerful, easy-to-use, and soon to be open-source simulation tool, designed for robotics research: dpmd.ai/mujoco-blog
-
Oriol Vinyals reposted
Skanda
@skandakoppula
We've released code to accompany our work on contrastive detection (dpmd.ai/3jsOV9l), now available at github.com/deepmind/detcon!
-
Oriol Vinyals
@OriolVinyalsML
The Deep Learning team is looking for interns, consider applying to work with us! Some projects our past interns have worked on: discovering the VAE "posterior collapse", VQ MuZero, AlphaStar, & more!
Google DeepMind
@GoogleDeepMind
✨ Applications for internships are now open✨
Interning at DeepMind offers a unique opportunity to closely collaborate with our team & work on various projects that aim to advance science and benefit humanity.
Interested? Learn more: dpmd.ai/internships -
Oriol Vinyals reposted
Google DeepMind
@GoogleDeepMind
From packing an umbrella to preparing for extreme conditions, predicting short term weather patterns is crucial for daily life.
New research with the @metoffice and SOTA model advances the science of Precipitation Nowcasting - the prediction of rain: dpmd.ai/nowcasting 1/4 -
Oriol Vinyals
@OriolVinyalsML
If you love Graph Neural Nets, language & Wikipedia, this is for you! We open sourced data & baselines for the challenging task of generating text from knowledge graphs (& vice versa!) #graph2seq #seq2graph #wikigraphs
Source github.com/deepmind/deepmind-research/tree/master/wikigraphs
Paper arxiv.org/abs/2107.09556
Oriol Vinyals
@OriolVinyalsML
-
Oriol Vinyals
@OriolVinyalsML
Four years ago, we open sourced pysc2. Deep Learning + RL then went from solving some "simple" tasks using @StarCraft, to becoming GrandMaster a couple years later with #alphastar. Good memories & fun / humble beginnings!
Oriol Vinyals
@OriolVinyalsML
Launching StarCraft2 Learning Environment (Linux)! A3C baseline on mini & full game + policies trained on 1M replays.deepmind.com/blog/deepmind-and-blizzard-open-starcraft-ii…
-
Oriol Vinyals reposted
Andrej Karpathy
@karpathy
Perceiver IO is good reading/pointers for neural net architectures arxiv.org/abs/2107.14795 esp w.r.t. encoding/decoding schemes of various modalities to normalize them to & from Transformer-amenable latent space (a not-too-large set of vectors), where the bulk of compute happens.
-
Oriol Vinyals
@OriolVinyalsML
Perceiver is an architecture based on Transformers and derived ideas (e.g., Set Transformers) that allow you to create a learned "bottleneck" so that you can process "any" modality. Simple, effective, and just "wants to work".
Google DeepMind
@GoogleDeepMind
To tackle all the challenges we meet while solving intelligence, we need tools that are as adaptable as possible. Announcing the paper & code for Perceiver IO, an architecture that handles a wide range of data and tasks, all while scaling gracefully: dpmd.ai/perceiver-IO 1/4
-
Oriol Vinyals reposted
Danielle Belgrave
@DaniCMBelg
Today is my first day @DeepMind! I'm really excited about the journey and collaborations ahead.
-
Oriol Vinyals reposted
Max Jaderberg
@maxjaderberg
Very excited to release our new work: Open-Ended Learning Leads to Generally Capable Agents. tldr; algorithm that dynamically shapes task distributions to train agents on huge task space, resulting in surprisingly general behaviour dpmd.ai/open-ended-blog
Thread: (1/n) -
Oriol Vinyals
@OriolVinyalsML
Good thread / summary of the fireside chat today with @geoffreyhinton!
Dickson
@DicksonWuML
Thread of learnings from the Fireside chat between @geoffreyhinton & @OriolVinyalsML
-
Oriol Vinyals
@OriolVinyalsML
Combinatorial Optimization is the "anti" Deep Learning domain: no structure and extremely non-smooth. It is nonetheless exciting to see what DL can do on these. Here, (and with a little help from classical solvers), we show some cool results at scale!
Google DeepMind
@GoogleDeepMind
Mixed Integer Programming is an NP-hard optimisation problem arising in planning, logistics, resource allocation, etc.
Presenting a solver with neural heuristics that learns to adapt to the problem domain, outperforming SCIP on Google-scale MIPs: dpmd.ai/13349 (1/) -
Oriol Vinyals
@OriolVinyalsML
I was introduced to DL by this Matlab code to classify MNIST from www.cs.toronto.edu/~hinton/MatlabForSciencePaper.html. This changed my career & that of many others. Very excited & nervous to host a fireside chat w/ @geoffreyhinton in the wonderful @Khipu_AI AI series. Don't miss it!
www.crowdcast.io/e/fireside-chat-with-7/register
Oriol Vinyals
@OriolVinyalsML
More details:
-Event time: Tuesday, July 27h, 17hs Brasilia (UTC-3)
-Slack for the @Khipu_AI 2021 Event Series: join.slack.com/t/khipu-event-series/shared_invite/zt-qlx4…
-There will be a post-event social gathering to have a more informal chat with the LATAM & AI community! -
Oriol Vinyals
@OriolVinyalsML
#alphafold is Deep Learning at its finest. Take a (small!) dataset of aminoacid sequences -> 3D structure. Use some of the best ideas from ML, NLP, CV (transformers, GNNs, BERT, distillation), sprinkle loads of innovation for proteins (Evoformer, MSA representation), and voila.
Oriol Vinyals
@OriolVinyalsML
Also, I'll never get tired to point this out: datasets enabled breakthroughs as much as the techniques that made it possible. A few favorites: WMT (Machine Translation), ImageNet (Computer Vision), PDB (Protein Folding). THANK YOU! @drfeifei @PDBeurope @buildmodels
-
Oriol Vinyals reposted
Google DeepMind
@GoogleDeepMind
Today with @emblebi, we're launching the #alphafold Protein Structure Database, which offers the most complete and accurate picture of the human proteome, doubling humanity’s accumulated knowledge of high-accuracy human protein structures - for free: dpmd.ai/alphafolddb 1/
-
Oriol Vinyals
@OriolVinyalsML
A tribute to AlphaFold by VQGAN + CLIP. Can you guess the caption that generated this image from scratch?
@Nature paper: www.nature.com/articles/s41586-021-03819-2
Code: github.com/deepmind/alphafold -
Oriol Vinyals reposted
Demis Hassabis
@demishassabis
Last year we presented #alphafold v2 which predicts 3D structures of proteins down to atomic accuracy. Today we’re proud to share the methods in @Nature w/open source code. Excited to see the research this enables. More very soon!
idp.nature.com/transit?redirect_uri=https%3A%2F%2Fwww.nat…
github.com/google-deepmind/alphafold -
Oriol Vinyals
@OriolVinyalsML
An Image is Worth 50K Pixels: Perceivers propose convolution free image (& other modalities) processing (cc @ylecun for checking lack of convs 😉).
Paper (appearing @icmlconf): arxiv.org/abs/2103.03206
Google DeepMind
@GoogleDeepMind
Many models bake in domain knowledge to control how input data is processed. This means models must be redesigned to handle new types of data.
Introducing the Perceiver, an architecture that works on many kinds of data - in some cases all at once: dpmd.ai/perceiver (1/) -
Oriol Vinyals reposted
khipu.ai
@Khipu_AI
Join us in this exciting #khipu2021 fireside chat with @OriolVinyalsML hosting @geoffreyhinton on July 27 9pm UK, 5pm Brasília time. It's FREE, you can register at www.crowdcast.io/e/fireside-chat-with-7/register. After the chat we will have a social on Gathertown and keep the conversation going! #latam
-
Oriol Vinyals
@OriolVinyalsML
❄️Multimodal Few-Shot Learning with Frozen Language Models❄️
-Freeze an LM & train vision net to do captioning (simple)
-Model can also do VQA, miniImageNet, ... (effective)
Paper: arxiv.org/abs/2106.13884
w/@mtsimpoukelli @jacobmenick @serkancabi @arkitus @FelixHill84 -
Oriol Vinyals reposted
Demis Hassabis
@demishassabis
Brief update on some exciting progress on #alphafold! We’ve been heads down working flat out on our full methods paper (currently under review) with accompanying open source code and on providing broad free access to AlphaFold for the scientific community. More very soon!
-
Oriol Vinyals
@OriolVinyalsML
We (still) require more MineRLs! Awesome competition to test your machine learning skills (whilst enjoying MineCraft), happening @NeurIPSConf competitions. See you there, and GL HF!!
MineRL Project
@minerl_official
We are so excited to announce that the first round of the 2021 @NeurIPSConf MineRL Diamond Competition has officially started! 💎 Submit your agents here: www.aicrowd.com/challenges/neurips-2021-minerl-competition
We have some exciting new changes this year. Keep reading to learn more!
1/4This media couldn't be displayed. See it on the original post. -
Oriol Vinyals
@OriolVinyalsML
MuZero removed simulators in MBRL vs AlphaGo. VQ Models for Planning generalize to partial observable & stochastic environments. How?
1. Discretize states w/ VQVAE
2. Train a LM over states
3. Plan w/ MCTS using the LM
Led by @yazhe_li & @sherjilozair
arxiv.org/abs/2106.04615 -
Oriol Vinyals
@OriolVinyalsML
Few Shot Learning Is All You Need #vaccinated #neurips2021
-
Oriol Vinyals
@OriolVinyalsML
Turns out this is what I was optimizing for all along : )
🧢🧢🧢 cap-measure.org -
Oriol Vinyals reposted
Yaroslav Ganin
@yaroslav_ganin
New paper day.
Our take on generation of structured objects:
Protocol Buffers + Transformers + Pointer Nets
We showcase the method on 2D CAD sketches (geometric primitives & relations between them). Should work for other domains too.
arxiv.org/abs/2105.02769
Mandatory samples: -
Oriol Vinyals reposted
yobibyte
@y0b1byte
Today's notes on "Learning Deep Generative Models of Graphs" (arxiv.org/abs/1803.03324) by @liyuajia @OriolVinyalsML @redpony @rpascanu & @PeterWBattaglia. Read to learn about how to generate graphs with graph nets!
www.notion.so/Learning-Deep-Generative-Models-of-Graphs-d… -
Oriol Vinyals
@OriolVinyalsML
Linear Algebra and Calculus is really all you need.
Michael Bronstein
@mmbronstein
What is in common between CNNs, GNNs, LSTMs, Transformers, DeepSets, mesh CNNs? In a new post with @joanbruna @TacoCohen @PetarV_93 we show this zoo of neural nets can be seen through the lens of symmetry. #geometricdeeplearning is all you need!
towardsdatascience.com/geometric-foundations-of-deep-lear… -
Oriol Vinyals
@OriolVinyalsML
Honoured to be giving a keynote @ucsd_cse in a few hours! It's always good to look back at your origins and where you come from. @UCSanDiego for me was the beginning of a love story with sequences, which changed my (research) life.
Sign up here: cseopenhouse.ucsd.edu
UCSD CSE
@ucsd_cse
There's still time to register! Tomorrow from 11:00AM to 5:30PM PST, come hear from our keynote speaker @OriolVinyalsML , discover current research, and network with past and present members of @ucsd_cse.
Sign up here: cseopenhouse.ucsd.edu
Oriol Vinyals
@OriolVinyalsML
-
Oriol Vinyals
@OriolVinyalsML
Is machine translation harder than Go? Enjoy this study of decoding methods for MT lead by @RemiLeblond
Paper: arxiv.org/abs/2104.05336
Rémi Leblond
@RemiLeblond
Check out our latest paper arxiv.org/abs/2104.05336! We investigate using MCTS with a value network for machine translation decoding. Joint work with the amazing @jalayrac, @laurentsifre, @Miruna_Pislar, JB Lespiau, Ioannis Antonoglou, Karen Simonyan and @OriolVinyalsML 1/4
-
Oriol Vinyals
@OriolVinyalsML
Thanks for posting @CaroPetitjean!
Here are the slides: docs.google.com/presentation/d/e/2PACX-1vQT9UTWraCYr2H2qG…
Caroline Petitjean
@CaroPetitjean
Interested by self-supervised learning?
I highly recommend @OriolVinyalsML's video at @NeurIPSConf workshop on SSL, which is crystal clear, with many interesting insights on motivations and history behind SSL.
slideslive.com/38938406/the-infonce-loss-in-selfsupervise… -
Oriol Vinyals reposted
khipu.ai
@Khipu_AI
Excited to announce the *Khipu 2021 Event Series*: monthly online meetings to support the advancement of AI talent in Latinamerica. Our inaugural event on April 27 will feature Doina Precup and Rich Sutton & other outstanding speakers.Register at crowdcast.io/khipu/latest [1/4]
-
Oriol Vinyals
@OriolVinyalsML
I am proud to have been the very first report of Samy Bengio. His kindness, values, and leadership, is what I pursue as a manager. Thank you, Samy!
www.reuters.com/article/us-alphabet-google-research-bengi… -
Oriol Vinyals
@OriolVinyalsML
Congrats! When will we get a (short) neural animated movie? : )
Google DeepMind
@GoogleDeepMind
The paper “Learning Mesh-Based Simulation with Graph Networks” shows how to effectively learn physics simulation on irregular meshes. The model can simulate vastly different systems with significant speedups over the ground truth solver: dpmd.ai/2PubkYS #iclr2021 (3/3)
This media couldn't be displayed. See it on the original post. -
Oriol Vinyals
@OriolVinyalsML
Love the condensed / "TLDR" style format 😍. Thanks @y0b1byte for featuring our work on program induction! 💻
yobibyte
@y0b1byte
Today's notes on "Strong Generalization and Efficiency in Neural Programs" (arxiv.org/abs/2007.03629) by @liyuajia, Felix Gimeno, @pushmeet @OriolVinyalsML. Read to learn abt neural program induction &designing state/action space for strong generalisation. www.notion.so/Strong-Generalization-and-Efficiency-in-Neu…
-
Oriol Vinyals
@OriolVinyalsML
2016 -> 2021 Understanding Deep Learning (Still) Requires Rethinking Generalization
A "refreshed" version of our paper from 2016, with a cool perspective from @prfsanjeevarora @TheOfficialACM
Perspective: cacm.acm.org/magazines/2021/3/250716-technical-perspectiv…
Paper: cacm.acm.org/magazines/2021/3/250713-understanding-deep-l… -
Oriol Vinyals reposted
Xavier Bresson
@xbresson
.@ylecun @wellingmax @OriolVinyalsML @69alodi @PeterWBattaglia @StefanieJegelka @WoodyOsher and I are organizing next week the @ipam_ucla workshop on
"Deep Learning and Combinatorial Optimization"
Great line-up of speakers :)
www.ipam.ucla.edu/programs/workshops/deep-learning-and-co… -
Oriol Vinyals reposted
Meire Fortunato
@FortunatoMeire
Very glad to share this work (arxiv.org/abs/2010.03409)! Rescuing my math/simulation background :-) My first paper in ML was Pointer Networks (arxiv.org/abs/1506.03134), where, amongst other things, we were looking into mesh generation. Awesome to see how loops form :-D
-
Oriol Vinyals reposted
Via lliure RAC1
@vialliure
📌 El sabadellenc @OriolVinyalsML lidera un equip d'investigació en intel·ligència artificial a Google.
🌍 Coneix quina és la seva feina i què ens espera en un futur pròxim a l'entrevista que li hem fet al @vialliure!
www.rac1.cat/programes/via-lliure/20210123/491692610817/o… -
Oriol Vinyals
@OriolVinyalsML
A privilege being able to talk about my passion in my native tongue, Catalan : ) Thanks @vialliure @rac1!
Via lliure RAC1
@vialliure
📌 El sabadellenc @OriolVinyalsML lidera un equip d'investigació en intel·ligència artificial a Google.
🌍 Coneix quina és la seva feina i què ens espera en un futur pròxim a l'entrevista que li hem fet al @vialliure!
www.rac1.cat/programes/via-lliure/20210123/491692610817/o… -
Oriol Vinyals
@OriolVinyalsML
I am seriously considering a more drastic version of this: tweet IS paper.
Kyle Cranmer
@KyleCranmer
Love it... not papers with code... more like paper IS code.
@fperez_org you'll love it... Jupyter notebook + Lattice Field Theory! -
Oriol Vinyals
@OriolVinyalsML
Neural networks still surprise me as rich information retrieval systems. The generalization and "creativity" exhibited by @OpenAI's DALL·E is amazing! Good luck finding these images in the training distribution (see thread).
Oriol Vinyals
@OriolVinyalsML
Here I show the most similar image in the web on the right, with the generated one on the left. Bonus points if you can correctly caption the image on the left : )
-
Oriol Vinyals
@OriolVinyalsML
2020++
-
Oriol Vinyals
@OriolVinyalsML
Happy online #neurips! The (lack of) networking is challenging 🤔
Yesterday, I was "hanging out" in gathertown. Weds, I'll be mentoring #wiml with the brilliant @orussakovsky on long term research planning.
I also plan to go to socials. Any other recommendations for networking?
Oriol Vinyals
@OriolVinyalsML
Also yesterday I gave a different / personal kind of talk @_LXAI , "A Road towards the Chain Rule" (www.latinxinai.org/neurips-2020)
As for socials, these sound interesting: "Lapsed Physicists Wine-and-Cheese", "Starting/transitioning your career in ML during the pandemic".
CU around! -
Oriol Vinyals reposted
Petar Veličković
@PetarV_93
On Thursday, I will be presenting our latest advance in neural algorithmic reasoning: the Pointer Graph Network.
Spotlight talk + Q&A: Thursday 3:10--3:20pm GMT
Poster Session: Thursday 5--7pm GMT
with Lars Buesing, Matt Overlan, @rpascanu, @OriolVinyalsML and @BlundellCharles. -
Oriol Vinyals reposted
Marta Garnelo
@GarneloMarta
What giving a @NeurIPSConf tutorial in 2020 really looks like...
🎊Join us for the 'Designing Learning Dynamics' tutorial this Monday at 10:30 AM (GMT)! 🎊 sites.google.com/view/dld-tutorial/home -
Oriol Vinyals reposted
Oriol Vinyals
@OriolVinyalsML
In science, everyone wins.
-
Post not available here, view on X
Oriol Vinyals
@OriolVinyalsML
In science, everyone wins.
-
Oriol Vinyals
@OriolVinyalsML
Biology's ImageNet moment has arrived: #alphafold2 solves Protein Folding at the #casp14 competition!
Blog: deepmind.com/blog/article/alphafold-a-solution-to-a-50-ye…
Oriol Vinyals
@OriolVinyalsML
Lots of cool details in the blog!
-
Oriol Vinyals reposted
WiML
@WiMLworkshop
1/5 #wiml is excited to present the recent work of Meire Fortunato (@FortunatoMeire) on “Learning Mesh-Based Simulation with Graph Networks”. Read more here: arxiv.org/abs/2010.03409
#latinxwiml @Khipu_AI #womeninstem #latinxinai #graphneuralnetworks #gnns #simulation -
Oriol Vinyals reposted
LatinX in AI (LXAI)
@_LXAI
We are honored to have Oriol Vinyals as one of our keynote speakers in the #lxai Workshop at #neurips2020. Do not miss his intervention: "A Road towards the Chain Rule".
-
Oriol Vinyals
@OriolVinyalsML
Reminder that this deadline is coming up!
Application deadline: December 1st, 2020, 23:59 CEST
Oriol Vinyals
@OriolVinyalsML
Interested in a PhD in ML? Applications are open to *anyone* wishing to pursue a PhD in ML within ELLIS. Excited to be working with many talented students!
Portal opens: Oct 1st
Deadline: Dec 1st
Ping me for advice along the process!
ellis.eu/de/news/ellis-phd-program-call-for-applications -
Oriol Vinyals
@OriolVinyalsML
It's that time of the year: PhD applications. I just re-read my statement of purpose / research statement from 2008. It was epically bad, as it more or less read like a written CV. @nlpnoah offers great advice which I wished I had access to!
PS Thanks @berkeley_ai : )
Oriol Vinyals
@OriolVinyalsML
"A great statement is distinctive and shows awareness, curiosity, and forward thinking."
-
Oriol Vinyals reposted
khipu.ai
@Khipu_AI
Hola Khipu community, oi! It's been a full year since #khipu2019. We’re excited to share a documentary with highlights of our first Khipu edition. What were your favorite memories? Share them with #khipu_ai”. (short version youtu.be/uDWkdVl1NOQ, full youtu.be/Vq_PKNagLR4and)
-
Oriol Vinyals reposted
Sasha Rush
@srush_nlp
These tutorial slides on "High Perf NLP" are really impressive. Every slide is current to the minute. Amazing set of diagrams.
gabrielilharco.com/publications/EMNLP_2020_Tutorial__High…
(@gabriel_ilharco @Tim_Dettmers @IuliaTurc @kentonctlee Felipe Ferreira Cesar Ilharco) -
Oriol Vinyals
@OriolVinyalsML
One of the first things I did after joining @Google was to implement an LSTM using TF's predecessor "DistBelief". The model definition had 16245 lines! Once we moved to @TensorFlow, the same model could be defined in 44 lines. Happy Bday @TensorFlow! 🎂🎂🎂🎂🎂
Jeff Dean
@JeffDean
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… -
Oriol Vinyals
@OriolVinyalsML
May our gradients never vanish. #sciencewins
-
Oriol Vinyals
@OriolVinyalsML
21 papers asked questions in their titles in NeurIPS 2019. I am not sure the answer to those is always No : ) Some favourites:
Are Sixteen Heads Really Better than One?
Can Unconditional Language Models Recover Arbitrary Sentences?
When does label smoothing help? -
Oriol Vinyals
@OriolVinyalsML
Fixed : )
-
Oriol Vinyals
@OriolVinyalsML
I enjoy visualizations of trends in #ml. I added some prehistoric data points to give some further perspective on the "Language models: Welcome to the Billion Parameter club" slide from the excellent @stateofai : )
-
Oriol Vinyals reposted
Ilya Sutskever
@ilyasut
It can be hard to tell where memorization ends and generalization begins
-
Oriol Vinyals
@OriolVinyalsML
bugpropagation: a bug that goes undetected due to gradient descent "taking care" of it. Bonus: fixing the bug makes results worse.
Never happened to me, of course! ; )
Kory Mathewson
@korymath
Alrighty #machinelearning, I'm looking for your number 1 favourite ML/#ai term.
Mine is "loopy belief propagation" - approximate inference algorithm on graphical models with loops
Faves @danijarh @elasri_layla @FelixHill84 @hardmaru @MirowskiPiotr @ladognome @transformsai? -
Oriol Vinyals
@OriolVinyalsML
Internships are a great learning opportunity. I learned a lot during my 5 (!) internships, and kept learning as well as an intern host. Things I learned as a host? VAEs, GANs, GNNs, open ended learning, ...
Consider interning in our Deep Learning team! deepmind.com/careers/jobs/2357199
Oriol Vinyals
@OriolVinyalsML
Since many of you asked: if you are willing to relocate to the countries listed in the website, DeepMind will help sponsoring a working permit there.
-
Oriol Vinyals
@OriolVinyalsML
Recent conversation with a friend:
@ilyasut: what's your take on openreview.net/pdf?id=YicbFdNTTy?
@OriolVinyalsML: my take is: farewell convolutions : ) -
Oriol Vinyals
@OriolVinyalsML
I feel so lucky & happy to see my alma mater and current employee empowering the next generation of ML researchers! @UCSDJacobs @DeepMind
UCSD CSE
@ucsd_cse
We are grateful to CSE alum @OriolVinyalsML and @DeepMind for the generous gift supporting #artificialintelligence research and enhanced #diversity
@UCSDJacobs cse.ucsd.edu/about/news/deepmind-gift-will-give-boost-mac… -
Oriol Vinyals
@OriolVinyalsML
Very nice to see this done! Also, this is clearly a busy research area, so try to think "outside of the box" -- or "outside of the circles" : ) #xformers
Yi Tay
@YiTayML
Inspired by the dizzying number of efficient Transformers ("x-formers") models that are coming out lately, we wrote a survey paper to organize all this information. Check it out at arxiv.org/abs/2009.06732.
Joint work with @m__dehghani @dara_bahri and @metzlerd. @GoogleAI 😀😃 -
Oriol Vinyals reposted
Xavi Giró
@DocXavi
Today we start a new teaching adventure for our @UPCTelecos @dseupc undergrads, with an ambitious course on reinforcement and unsupervised learning.
Built by an awesome team @la_UPC Barcelona, and supported by a first-class line up of guest speakers.
telecombcn-dl.github.io/drl-2020/ -
Oriol Vinyals
@OriolVinyalsML
Interested in a PhD in ML? Applications are open to *anyone* wishing to pursue a PhD in ML within ELLIS. Excited to be working with many talented students!
Portal opens: Oct 1st
Deadline: Dec 1st
Ping me for advice along the process!
ellis.eu/de/news/ellis-phd-program-call-for-applications
Oriol Vinyals
@OriolVinyalsML
For those interested, email is better than DM, and should be easy to find : )
-
Oriol Vinyals
@OriolVinyalsML
GNNs are so cool! But we still need more advances to predict London traffic🚦🤦
Google DeepMind
@GoogleDeepMind
Exciting results from our partnership with @googlemaps! Using advanced Graph Neural Networks we were able to improve the accuracy of ETAs in major cities across the world by up to 50% deepmind.google/blog/traffic-prediction-with-advanced-gra…
-
Oriol Vinyals reposted
khipu.ai
@Khipu_AI
We have 14 complimentary tickets to @Aistats2020 to reassign to the Khipu community. Please fill out the form forms.gle/1kJY2kE8YJ5JS5du6 if you are interested.
-
Oriol Vinyals
@OriolVinyalsML
Mesh generation will keep pushing the boundaries of the Deep Learning Toolbox. Very happy to see code released from @charlietcnash et al, as it will help anyone interested in pushing the field further!
Source: github.com/deepmind/deepmind-research/tree/master/polygen
Paper: arxiv.org/abs/2002.10880 -
Oriol Vinyals
@OriolVinyalsML
I have a joke about a joke about meta learning.
-
Oriol Vinyals
@OriolVinyalsML
I will host an informal mentoring session / discussion as part of the wonderful #icml2020 mentoring initiative in 10 minutes. Feel free to drop by! (details are posted in the RocketChat🚀 #mentoring channel)
-
Oriol Vinyals
@OriolVinyalsML
slide_level++
Feryal
@FeryalMP
Here is Meire's drawing of her career path! Love the creativity, both in her career and how she illustrated it!
-
Oriol Vinyals reposted
Yujia Li
@liyuajia
Excited to share our new work on learning to discover neural programs that strongly generalize and outperform hand-coded algorithms for sorting (quick sort) and other tasks, in number of execution steps.
arxiv.org/abs/2007.03629
w/ @FelixAxelGimeno @pushmeet @OriolVinyalsML -
Oriol Vinyals
@OriolVinyalsML
Interested in RL, data efficiency & imitation learning? We require more MineRLs! @NeurIPSConf
MineRL Project
@minerl_official
We are pleased to announce that the 2020 @NeurIPSConf MineRL Competition has officially launched! This is an exciting collaboration between @mldcmu, @OpenAI, @PreferredNet, @AIcrowdHQ, @MSFTResearch, and @DeepMind. Check it out and sign up here: minerl.io/competition
-
Oriol Vinyals reposted
Ferran Alet
@FerranAlet
This Wednesday (2pm EST) @OriolVinyalsML will present at the @MIT Embodied Intelligence Seminar an overview of recent progress in Deep Learning: architectures, self-supervised learning, meta-learning, and more! We will live-stream the talk on this channel: www.youtube.com/channel/UCnXGbvgu9071i3koFooncAw
-
Oriol Vinyals reposted
Cristian Canton
@cristiancanton
DeepVision Workshop (@DeepVision2020) at @CVPR
Today!
sites.google.com/view/deepvision2020
A grat set of speakers: Yann LeCun (@ylecun), Joaquin Quiñonero (@jquinonero), Oriol Vinyals (@OriolVinyalsML) and Vladlen Kaltun plus 13 papers on fairness and learning with little data. Enjoy! -
Oriol Vinyals reposted
Petar Veličković
@PetarV_93
Excited to share Pointer Graph Networks! 🧑💻
Combining classical pointer-based data structures with GNNs, we recover strong results on algorithm execution + useful latent graphs.
w/ Lars Buesing, Matt Overlan, @rpascanu, @OriolVinyalsML, @BlundellCharles
arxiv.org/abs/2006.06380 -
Oriol Vinyals
@OriolVinyalsML
Thanks for doing this, @le_roux_nicolas! Nice to see lots of us being able to help. I'm in position 378 ^_^
-
Oriol Vinyals
@OriolVinyalsML
Scale *still* delivers! Congrats @OpenAI on showing very nice zero/few-shot language capabilities of GPT-3. #timelesstweet
Paper: arxiv.org/abs/2005.14165
Endless Samples: github.com/openai/gpt-3
Oriol Vinyals
@OriolVinyalsML
Also amazing how pushing performance on next word prediction just works. Elegant and simple.
-
Oriol Vinyals reposted
Ilya Sutskever
@ilyasut
A Moore’s law but for DL algorithms:
OpenAI
@OpenAI
Since 2012, the amount of compute for training to AlexNet-level performance on ImageNet has been decreasing exponentially — halving every 16 months, in total a 44x improvement.
By contrast, Moore's Law would only have yielded an 11x cost improvement: openai.com/blog/ai-and-efficiency/ -
Oriol Vinyals
@OriolVinyalsML
Very happy to see such a cool application of VQ-VAEs by our friends @OpenAI! Congrats!
OpenAI
@OpenAI
Introducing Jukebox, a neural net that generates music, including rudimentary singing, as raw audio in a variety of genres and artist styles. We're releasing a tool for everyone to explore the generated samples, as well as the model and code: openai.com/blog/jukebox/
-
Oriol Vinyals
@OriolVinyalsML
See you later today for some generative fun! sites.google.com/corp/view/gforg #iclr2020 @phillip_isola @kilianq @cephaloponderer @puneetdokania @hugo_larochelle
Prabhu Pradhan
@PrabhuPradhan
@iclr_conf Calling all 'generators' (+ friends) for our Generative ML & Generalization Social (sites.google.com/view/gforg/), 28th April- 4pm GMT🗓️.
Speakers incl. @OriolVinyalsML @phillip_isola @kilianq @puneetdokania.
Please share! 🧠👩🎨👨💻
Post questions here app.sli.do/event/mopvannt -
Oriol Vinyals reposted
Maithra Raghu
@maithra_raghu
Presenting this at @iclr_conf *today*!
Talk and Slides: iclr.cc/virtual/poster_rkgMkCEtPB.html
Poster Sessions: (i) 10am - 12 Pacific Time, (ii) 1pm - 3pm Pacific Time
Thanks to the organizers for a *fantastic* virtual conference, hope to see you there!
#iclr #iclr2020
Oriol Vinyals
@OriolVinyalsML
Rapid Learning or Feature Reuse? Meta-learning algorithms on standard benchmarks have much more feature reuse than rapid learning! This also gives us a way to simplify MAML -- (Almost) No Inner Loop (A)NIL. arxiv.org/abs/1909.09157 With Aniruddh Raghu @maithra_raghu Samy Bengio.
-
Oriol Vinyals
@OriolVinyalsML
Datasets over Algorithms. Creating good benchmarks and datasets is critical for the advancement of #ai. Thanks ImageNet, WSJ, ALE, PTB, GLUE, WMT, etc.!
(Old) Source: www.spacemachine.net/views/2016/3/datasets-over-algorithms -
Oriol Vinyals
@OriolVinyalsML
Single author (!) #metaresearch paper from NeurIPS2019 on reproducibility in ML. These are the top factors that affect reproducibility (w/ many caveats, discussed in the paper).
arxiv.org/abs/1909.06674
Oriol Vinyals
@OriolVinyalsML
Note that the author did NOT use / look at code if it was available. Thus why pseudo code > code available.
-
Oriol Vinyals reposted
Sundar Pichai
@sundarpichai
We've published an early release of our #covid19 Community Mobility Reports in 130+ countries which provide insights that could help public health officials understand the impact of response strategies such as social distancing: google.com/covid19/mobility
blog.google/technology/health/covid-19-community-mobility… -
Oriol Vinyals
@OriolVinyalsML
Nice visualization of progression of methods that went into Agent57, achieving best (so far...) results in ATARI.
Brought to you from colleagues @DeepMind:
Paper: arxiv.org/abs/2003.13350
Blog: deepmind.com/blog/article/Agent57-Outperforming-the-human… -
Oriol Vinyals reposted
Maithra Raghu
@maithra_raghu
A Survey of Deep Learning for Scientific Discovery
To help facilitate using DL in science, we survey a broad range of deep learning methods, new research results, implementation tips & many links to code/tutorials
Paper arxiv.org/abs/2003.11755
Work with @ericschmidt
Thread⬇️ -
Oriol Vinyals
@OriolVinyalsML
The city of London supporting our #nhs heroes. #clapfornhs
-
Oriol Vinyals reposted
Sander Dieleman
@sedielem
New blog post: 'Generating music in the waveform domain'
benanne.github.io/2020/03/24/audio-generation.html
A comprehensive overview of the field and some personal thoughts, based on a tutorial I gave at @ismir2019 with @jordiponsdotme and Jongpil Lee back in November. Comments / feedback welcome! -
Oriol Vinyals
@OriolVinyalsML
Graphics is definitely among my favorite machine learning applications!
See more in: www.matthewtancik.com/nerf -
Oriol Vinyals
@OriolVinyalsML
Like many others, I am in social distancing mode. At the same time, we have so many tools at our disposal to not feel isolated. Any ideas or ongoing efforts to do (online) unconferences, meetups, AMAs, etc., please let everyone know!
www.washingtonpost.com/graphics/2020/world/corona-simulator/ -
Oriol Vinyals reposted
Peter Battaglia
@PeterWBattaglia
Excited to present “Learning to Simulate Complex Physics with Graph Networks”.
arxiv.org/abs/2002.09405
Our model can generate realistic simulations, and generalizes to much larger systems and longer trajectories than its training.
w/ @spectralhippo @RexYing0923 @jureThis media couldn't be displayed. See it on the original post. -
Oriol Vinyals
@OriolVinyalsML
Transformers are a special case of Graph Neural Networks. This may be obvious to some, but the following blog post does a good job at explaining these important concepts.
Chaitanya K. Joshi
@chaitjo
Excited to share a blog post on the connection between #transformers for NLP and #graphneuralnetworks (GNNs or GCNs).
graphdeeplearning.github.io/post/transformers-are-gnns/ -
Oriol Vinyals
@OriolVinyalsML
Amazing how far mesh generation has been pushed! I post a video from Pointer Networks (2015), in which results like the one below by @charlietcnash et al seemed impossible.
Charlie Nash
@charlietcnash
Excited to share PolyGen, a generative model of 3D meshes. PolyGen sequentially generate meshes vertex-by-vertex, and face-by-face using Transformers.
With @yaroslav_ganin @arkitus @PeterWBattaglia @DeepMind
arxiv.org/abs/2002.10880 -
Oriol Vinyals reposted
Chelsea Finn
@chelseabfinn
Want to learn about meta-learning? Lecture videos for CS330 are now online!
www.youtube.com/playlist?list=PLoROMvodv4rMC6zfYmnD7UG3LV…
Topics incl. MTL, few-shot learning, Bayesian meta-learning, lifelong learning, meta-RL & more:
cs330.stanford.edu
+ 3 guest lectures from Kate Rakelly, @svlevine, @jeffclune -
Oriol Vinyals
@OriolVinyalsML
Excited to visit @Stanford tomorrow at 3pm for a talk in the StatsML Seminar on "Large Scale Imitation Learning: From AlphaGo to AlphaStar", where I'll explain how pure imitation yields an agent stronger than 85% of active @StarCraft players.
More info: statsml.stanford.edu/seminar.html -
Oriol Vinyals
@OriolVinyalsML
Rapid unsupervised learning progress thanks to contrastive losses, approaching supervised learning!
-40% Multitask SSL arxiv.org/abs/1708.07860 (2017)
-50% CPC arxiv.org/abs/1807.03748 (2018)
-70% AMDIM/MOCO/CPCv2/etc (2019)
-76.5% SimCLR arxiv.org/abs/2002.05709 (2020, so far)
Oriol Vinyals
@OriolVinyalsML
Though I'm not a big fan of the evaluation protocol: linear classification on top of unsupervised features learned on ImageNet.
-
Oriol Vinyals reposted
Quoc Le
@quocleix
Happy to announce that we've released a number of models trained with Noisy Student (a semi-supervised learning method). The best model achieves 88.4% top-1 accuracy on ImageNet (SOTA).
Enjoy finetuning!
Link: github.com/tensorflow/tpu/tree/master/models/official/eff…
Paper: arxiv.org/abs/1911.04252 -
Oriol Vinyals
@OriolVinyalsML
Less is more.
Zachary Lipton
@zacharylipton
The field is already self-correcting. Good departments/labs are clearing their eyes, caring less about paper count, seeing through the noise. Don't worry so much about the ICML deadline. Slow down, relax, try to do work you're proud of, submit when it's ready.
-
Oriol Vinyals
@OriolVinyalsML
seq2seq still delivers : ) Incredible how far we've gotten in ~5 years of progress in neural conversational models, with relatively small changes. More exciting is that there's still LOTS to be done!
Paper: arxiv.org/abs/2001.09977
Blog: ai.googleblog.com/2020/01/towards-conversational-agent-th… -
Oriol Vinyals
@OriolVinyalsML
Although neural networks usually require massive datasets to do impressive things, for me the highlight of #alphafold is the fact that it achieved state-of-the-art using only 30K training examples.
Code: github.com/deepmind/deepmind-research/tree/master/alphafo…
Paper: www.nature.com/articles/s41586-019-1923-7 -
Oriol Vinyals
@OriolVinyalsML
The "Deep Learning Toolbox" has greatly expanded in the last decade thanks to our wonderful research community. Also, important progress has been made to make our community more inclusive and less toxic. Still, there's LOTS to do, and I plan to keep focusing on advancing both.
Oriol Vinyals
@OriolVinyalsML
The figure above is an (old) slide I usually show in lectures and talks. Mix and matching those components creates cool architectures such as the one used in #openaifive or #alphastar.
-
Oriol Vinyals reposted
Robert Lange
@RobertTLange
Many gems in @OriolVinyalsML Deep RL workshop talk at #neurips2019 on AlphaStar. Including scatter connections, imitation-based regularization, the league & the unique problem decomposition.
-
Oriol Vinyals
@OriolVinyalsML
I have loved relativity theory and black holes my whole life. Super exciting to have been at #katiebouman talk at #neurips2019. Great example of the research diversity in our field!
-
Oriol Vinyals
@OriolVinyalsML
Exciting day tomorrow for those who love Deep RL! #neurips2019 Deep RL Workshop (West Exhibition Hall C):
-We will start by presenting #alphastar at 9:00 (also as poster)
-@OpenAI Five will follow at 9:30
Full schedule: sites.google.com/corp/view/deep-rl-workshop-neurips-2019/…
Oriol Vinyals
@OriolVinyalsML
Also check out the @minerl_official competition! (in a parallel session)
-
Oriol Vinyals
@OriolVinyalsML
This was one of the highlights of #neurips2019 for me! Met so many brilliant UG/MS students and discussed anecdotes, issues we all go through, advice, paper titles, etc. Given the size & growth, @NeurIPSConf meetups are a great way for communities to get together.
Prabhu Pradhan
@PrabhuPradhan
We had an amazing event Buds@Neurips owing to our wonderful mentors @MilindTambe_AI @AnimaAnandkumar
@pulkitology
@OriolVinyalsML
@white_martha
@akapoor_av8r
@stefmlee
@hayasick
#katiebouman
#billfreeman
@raskarmit
Thankyou so much for this!
@neu_rips @skoularidou @NeurIPSConf -
Oriol Vinyals
@OriolVinyalsML
Pointer Networks -------> @FortunatoMeire
Welcome to Twitter!😃 -
Oriol Vinyals
@OriolVinyalsML
I enjoyed Yoshua's #neurips2019 talk! We still have a lot of work to do to define WHAT the benchmarks should be towards System 2 to research HOW to make progress on them. In ML, setting the right problem is often more important than its solution.
Talk: slideslive.com/38922304/from-system-1-deep-learning-to-sy… -
Oriol Vinyals
@OriolVinyalsML
#neurips2019 Wednesday: Fun day ahead!
-9:20 to 10:05 Meet the #alphastar team at the @DeepMindAI stand
-CAS Score for evaluating gen. models, 10:45 Poster #111
-VQVAE2, 17:00 Poster #140 -
Oriol Vinyals
@OriolVinyalsML
#neurips2019 Daily Highlights
-@celestekidd's great talk included precious advice to men
"You deserve to benefit from the innovations and ideas of the women at your institutions. We deserve to be interacted with as equal colleagues. We need each other to innovate and thrive."
Oriol Vinyals
@OriolVinyalsML
-Awesome work by Sanaa Mansoor et al on protein structure generation w/ VAEs
(@celestekidd full advice: www.docdroid.net/u5r0j3S/kidd-neurips.pdf) -
Oriol Vinyals reposted
Aäron van den Oord
@avdnoord
Unsupervised pre-training now outperforms supervised learning on ImageNet for any data regime (see figure) and also for transfer learning to Pascal VOC object detection
arxiv.org/abs/1905.09272v2 -
Oriol Vinyals
@OriolVinyalsML
#neurips2019 tutorial on efficient deep learning by Vivienne Sze (@eems_mit). The tutorial will focus on evaluation and metrics, rather than techniques, which is a great idea!
Oriol Vinyals
@OriolVinyalsML
Slides available at eems.mit.edu/publications/tutorials/
-
Oriol Vinyals
@OriolVinyalsML
I am looking forward to #neurips2019! Let's remind ourselves to use scientific and empirical evidence to back our claims, and to be kind to each other and set a good example to current and future generations : )
-
Oriol Vinyals
@OriolVinyalsML
A nice 8 minute "Two Minute Paper" summary of #alphastar : )
Two Minute Papers
@twominutepapers
DeepMind’s AlphaStar: A Grandmaster Level StarCraft 2 AI
▶️Full video (ours): www.youtube.com/watch?v=jtlrWblOyP4
📜Source paper: deepmind.com/blog/article/AlphaStar-Grandmaster-level-in-…
#ai #deeplearning #science #twominutepapers #alphastar #deepmind -
Oriol Vinyals
@OriolVinyalsML
This is so cool! #alphastar learned a lot from players, and I'm hoping there is something players can learn from #alphastar : ) Congrats @ENCE_Serral on winning #hscxx!
Joona Sotala
@Serral_SC2
BTW game 2 on ephemeron was a alphastar build, almost won even though biles hit hard!
-
Oriol Vinyals reposted
Andrés Miguel Torrubia Sáez
@antor
PODCAST: SOFTWARE 2.0 es mio nuevo podcast sobre IA.
Entrevistamos a @OriolVinyalsML uno de los investigadores que ha creado AlphaStar en @DeepMindAI
soundcloud.com/andres-torrubia/oriol-vinyals-parte-1 -
Oriol Vinyals
@OriolVinyalsML
Thanks so much everyone who stopped by and played AlphaStar at BlizzCon - including @ENCE_Serral! Definitely one of the highlights of the weekend : ) For anyone who missed it, @Artosis has put together a cool summary and also casted the games. Take a look!
Dan Stemkoski
@Artosis
AlphaStar vs Serral
Thanks to @DeepMindAI for making it possible for me to analyze these games.
Thanks to @ENCE_Serral Serral for playing these as well!
I tried to do it in a way that non-StarCraft people could understand and appreciate
Here they are:
www.youtube.com/watch?v=OxseexGkv_Q&list=PLojXIrB9Xau29fR… -
Oriol Vinyals
@OriolVinyalsML
Standing ovation to close up @Khipu_AI 2019. Huge congratulations to the organizers and participants! This was a brilliant week! #ai #latam #khipu2019
Talk Videos: tv.vera.com.uy/buscar?search=Khipu
Oriol Vinyals
@OriolVinyalsML
(all videos will be made available in the next few days)
-
Oriol Vinyals
@OriolVinyalsML
So exciting to be at @Khipu_AI, where today I'll present our work on #alphastar at 11 AM (GMT-3). Super grateful to be part of this magical moment for LATAM & AI!
Program: khipu.ai/program/
Live Stream: tv.vera.com.uy/evento/53626
Online Questions: app.sli.do/event/qkxzvzhf/live/questions -
Oriol Vinyals
@OriolVinyalsML
So incredible to see #alphastar and the amazing game of @StarCraft featured on the cover of @nature! GG!
Nature Portfolio
@NaturePortfolio
This week on the Nature cover: Game plan. AI program learns to play StarCraft II to Grandmaster level. Browse the issue here: go.nature.com/2QgMo5p
-
Oriol Vinyals
@OriolVinyalsML
There are (roughly) two kinds of reviewers: those who prefer novelty over SOTA, and those who prefer SOTA over novelty.
-
Oriol Vinyals
@OriolVinyalsML
If you are in the London area, stop by if you want to chat about Deep Learning / RL / #alphastar!
Imperial College AI Network
@ImperialAI
"AlphaStar, Reinforcement Learning and AI", this should be an exciting talk by @OriolVinyalsML - come along to @imperialcollege @ImperialDSI tomorrow!
www.eventbrite.co.uk/e/alphastarreinforcement-learning-an… -
Oriol Vinyals
@OriolVinyalsML
Amazing two days at the #blizzcon #alphastar demo. We had great games, chats & discussions with the community. Glad that @ENCE_Serral, @Artosis, @LiquidTLO, and many others could stop by. We felt the @StarCraft magic and will never forget this week! GGWP!
-
Oriol Vinyals
@OriolVinyalsML
Exciting to be mentioned in the #blizzcon opening ceremony! We prepared a fun mix of #alphastar agents available for anyone to play. Come try them out and meet the team in the Arcade!
Oriol Vinyals
@OriolVinyalsML
Our demo stations are open! So cool to see people playing #alphastar : ) We’ve made a range of agents available, at different skill levels. Can’t wait to see you all. GL HF!
-
Oriol Vinyals
@OriolVinyalsML
The #alphastar team will be at #blizzcon! Looking forward to meeting fellow @StarCraft fans : ) Catch us on @ThePylonShow with @Artosis & @LiquidTLO on Saturday 11:30am PDT, and keep an eye out for more surprises!
-
Oriol Vinyals reposted
nature
@Nature
DeepMind’s AlphaStar beat all but the very best humans at the fast-paced, strategy sci-fi game StarCraft II. go.nature.com/2q20Lzt
-
Oriol Vinyals
@OriolVinyalsML
#alphastar @nature: Grandmaster level as all 3 @StarCraft races on battle.net, w/ a pro approved interface (camera & APM limits). 2 years ago I thought this was impossible!
How? Imitation learning (Diamond) -> multiagent League (Grandmaster)
deepmind.com/blog/article/AlphaStar-Grandmaster-level-in-…
Oriol Vinyals
@OriolVinyalsML
Special thanks to the @StarCraft community, and @Blizzard for their support! : )
@nature Paper: storage.googleapis.com/deepmind-media/research/alphastar/…
@nature Video: www.youtube.com/watch?v=6eiErYh_FeY
Replays: storage.googleapis.com/deepmind-media/research/alphastar/…
Blog: deepmind.com/blog/article/AlphaStar-Grandmaster-level-in-… -
Oriol Vinyals
@OriolVinyalsML
Very cool research, congrats! Next: generate full video games : )
Julian Togelius
@togelius
Why is machine learning rarely used for generating game levels? Because game levels are harder to generate than, say, faces. And because there aren't enough levels to train on. HOWEVER! In our new paper, we learn to generate pretty good levels after training on only 5 levels!
-
Oriol Vinyals reposted
Chris Olah
@ch402
Do you formally know Monte-Carlo and TD learning, but don't intuitively understand the difference? This is for you.
distill.pub/2019/paths-perspective-on-value-learning/ (with @samgreydanus) -
Oriol Vinyals reposted
Jeff Dean
@JeffDean
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" -
Oriol Vinyals
@OriolVinyalsML
Rapid Learning or Feature Reuse? Meta-learning algorithms on standard benchmarks have much more feature reuse than rapid learning! This also gives us a way to simplify MAML -- (Almost) No Inner Loop (A)NIL. arxiv.org/abs/1909.09157 With Aniruddh Raghu @maithra_raghu Samy Bengio.
-
Oriol Vinyals
@OriolVinyalsML
I totally second @NandoDF on this! Focus on long term impact, rather than on number of papers. Hopefully the incentive systems are there to protect us from short term optimization, though sometimes they are part of the problem :(
Nando de Freitas
@NandoDF
I believe I have written more papers than Alan Turing + John Nash! Numbers of papers alone is a wrong misleading metric. Please focus instead on writing good papers that advance the field, help the world, and that you’ll be proud of when you look back in 20 or 50 years.
-
Oriol Vinyals
@OriolVinyalsML
The paper "Understanding deep learning requires rethinking generalization" mostly asked questions. Glad to see some answers / new theories since then!
Ian Osband
@IanOsband
Looking back over the year, the one paper that gave me the best "aha" moment was...
Reconciling Modern Machine Learning and the Bias-Variance Tradeoff:
arxiv.org/abs/1812.11118
The "bias-variance" you knew was just the first piece of the story! -
Oriol Vinyals
@OriolVinyalsML
A good example of our field moving (too) fast: the all attention layer was actually released in the original transformer paper from 2017, with similar findings.
Code: github.com/tensorflow/tensor2tensor/blob/master/tensor2te…
Paper: arxiv.org/pdf/1706.03762v3.pdf (only in v3)
AI at Meta
@AIatMeta
Facebook AI researchers are sharing an all-attention layer to simplify the Transformer model and an adaptive attention span method to make it more efficient. Even with a much simpler architecture, these methods match or improve state-of-the-art results. ai.facebook.com/blog/making-transformer-networks-simpler-…
-
Oriol Vinyals
@OriolVinyalsML
One of the coolest aspects of SPIRAL is to use computer programs (as opposed to games) as environments for RL agents. Another is to use a discriminator that provides rewards to the agent.
TensorFlow
@TensorFlow
SPIRAL is an RL agent that generates images by interacting with off-the-shelf graphics software. We've released 9 agents trained to paint portraits of imaginary people in just 19 brush strokes on #tfhub!
For more, see:
→ goo.gle/2Mky1f9
→ goo.gle/2TAAvXk -
Oriol Vinyals
@OriolVinyalsML
Really great to see PBT working on a challenging supervised learning task, helping @Waymo to train neural nets faster, more efficiently & reducing false positives by 24% absolute!
Google DeepMind
@GoogleDeepMind
We’re excited to announce the results of our research collaboration with @Waymo, applying Population Based Training (PBT) to help make the process of training neural nets in their self-driving cars more effective and efficient. deepmind.com/blog/how-evolutionary-selection-can-train-mo…
-
Oriol Vinyals
@OriolVinyalsML
Excited that AlphaStar will be joining battle.net for scientific evaluation & for a limited time! The updated agent:
-Plays as and versus all races
-Plays multiple ladder maps
-Uses a camera interface and more restricted APMs
FAQ: starcraft2.com/en-us/news/22933138
GL & HF!
STARCRAFT
@StarCraft
StarCraft update! As part of ongoing scientific research into artificial intelligence, players can now opt-in for the chance to be matched against experimental versions of DeepMind's StarCraft II agent, AlphaStar, on the competitive ladder.
Learn more: blizz.ly/DeepmindLadder -
Oriol Vinyals
@OriolVinyalsML
Great post by Prof. David McAllester on why discrete representations matter, based on our findings in VQ-VAE2. "Vector quantization seems to be a minimal-bias way for symbols to enter into deep models."
arxiv.org/abs/1906.00446
machinethoughts.wordpress.com/2019/06/25/the-inevitabilit… -
Oriol Vinyals
@OriolVinyalsML
Reminder that applications to @Khipu_AI close this Friday. EVERYONE can / should apply! (speakers @chelseabfinn @kchonyc @JeffDean @soumithchintala @lspecia @NandoDF @DaniCMBelg Sapiro Bengio Ferrer Lopez-Paz Zadrozny Vidal Ferrante Abadi)
khipu.ai
@Khipu_AI
Hello world! We are happy to announce the first edition of Khipu: Latin American Meeting in Artificial Intelligence in November, in Montevideo, Uruguay.
Outstanding set of speakers. Applications now open for ALL!
Apply and see details in www.khipu.ai #khipu2019 -
Oriol Vinyals
@OriolVinyalsML
Good benchmarks may be more important than "novel" algorithms to advance the machine learning field. Congratulations to the ImageNet authors for receiving the @cvpr2019 test of time award! @jiadeng Wei Dong @RichardSocher @lijiali_vision Kai Li @drfeifei
www.image-net.org/papers/imagenet_cvpr09.pdf -
Oriol Vinyals reposted
Dario TLO Wünsch
@TLOgg
Back in England for DeepMind’s Alphastar StarCraft artificial intelligence project. No new information I can divulge but I will say I’m enjoying every minute of my time with this amazing team. @ London, United Kingdom www.instagram.com/p/BynAWPeBv2I/
-
Oriol Vinyals
@OriolVinyalsML
Surprising how simple ideas can yield such a good generative model!
-Mean Squared Error loss on pixels
-Non-autoregressive image decoder
-Discrete latents w/ straight through estimator
w/ @catamorphist & @avdnoord
VQ-VAE-2: arxiv.org/abs/1906.00446
Code: github.com/deepmind/sonnet/blob/master/sonnet/examples/vq… -
Oriol Vinyals reposted
Aäron van den Oord
@avdnoord
VQVAE-2 finally out!
Powerful autoregressive models in a hierarchical compressed latent space. No modes were collapsed in the creation of these samples ;)
Arixv: arxiv.org/abs/1906.00446
With @catamorphist and @vinyals
More samples and details 👇 [thread] -
Oriol Vinyals
@OriolVinyalsML
Evaluating generative models is hard! We propose Classification Accuracy Score from classifiers trained on generated data:
-Accuracy of 43% when trained purely on BigGAN samples (vs 73%)
-Naive data augmentation doesn't work (yet!)
Paper: arxiv.org/abs/1905.10887
cc @SumanRavuri -
Oriol Vinyals
@OriolVinyalsML
The distillation paper was actually rejected from #neurips2014 : ) As I'm sure many people are working towards #neurips2019, optimize for long term impact, not probability of acceptance.
Dmytrii S
@ml_review
Distilling the Knowledge in a Neural Network #nips2014
By Geoffrey Hinton @OriolVinyalsML @jeffdean
On transferring knowledge from an ensemble or
from a large highly regularized model into a smaller, distilled model
arxiv.org/abs/1503.02531 -
Oriol Vinyals
@OriolVinyalsML
Come meet the team and talk to us about #alphastar tomorrow, Thursday, at 10:30am at the @DeepMindAI #iclr2019 booth!
-
Oriol Vinyals reposted
Mostafa Dehghani
@m__dehghani
If you are at #iclr2019, make sure that you learn about our work, "Universal Transformers", w/ @sgouws, @OriolVinyalsML, @kyosu, and @lukaszkaiser (Thursday 11am-1pm, poster session at Great Hall BC, #62). You can also check out this blog post about UT:
mostafadehghani.com/2019/05/05/universal-transformers -
Oriol Vinyals reposted
Google DeepMind
@GoogleDeepMind
Our latest work on neural network models for reasoning about similarity between graph structured objects, with implications for a broad spectrum of applications: arxiv.org/abs/1904.12787
By @liyuajia, @calbeargu, Thomas Dullien, @OriolVinyalsML and @pushmeet -
Oriol Vinyals reposted
Lex Fridman
@lexfridman
Here's my conversation with Oriol Vinyals (@OriolVinyalsML) on the Artificial Intelligence podcast. He and his colleagues are behind many recent breakthrough ideas in AI. At @DeepMindAI he co-lead the development of AlphaStar that beat humans in StarCraft. www.youtube.com/watch?v=Kedt2or9xlo
-
Oriol Vinyals
@OriolVinyalsML
GL HF!!
Greg Brockman
@gdb
Sneak peek of the venue being set up for tomorrow’s Dota event. Get excited :)
-
Oriol Vinyals
@OriolVinyalsML
Giving away my free registration to @icmlconf (non-transferable after I transfer it to you). Reply/email me if you are interested! #icml2019
-
Oriol Vinyals
@OriolVinyalsML
Proud to take part in strengthening AI in Latin America! @Khipu_AI #khipu2019
khipu.ai
@Khipu_AI
Hello world! We are happy to announce the first edition of Khipu: Latin American Meeting in Artificial Intelligence in November, in Montevideo, Uruguay.
Outstanding set of speakers. Applications now open for ALL!
Apply and see details in www.khipu.ai #khipu2019 -
Oriol Vinyals
@OriolVinyalsML
Very proud of my alma mater, and excited about future collaborations with brilliant students and faculty! Go Bears!
Berkeley AI Research
@berkeley_ai
Announcing the BAIR Open Research Commons @Berkeley_AI @UCBerkeley to accelerate cutting-edge collaborative AI research with founding sponsors @Amazon, @Facebook, @Google, @Samsung, and @WaveComputing!
bair.berkeley.edu/blog/2019/03/24/bair-commons/ -
Oriol Vinyals
@OriolVinyalsML
Excited to visit Boston (Monday) & NYC (Tuesday) next week to talk about #alphastar! @MIT @BU_Tweets @nyuniversity
Boston: calendar.mit.edu/event/alphastar_mastering_the_real-time_… & www.bu.edu/hic/2019/03/01/alphastarmastering-the-real-tim…
NYC: wp.nyu.edu/cilvr/2019/03/04/march-12th-2019-oriol-vinyals… -
Oriol Vinyals
@OriolVinyalsML
Thanks a lot, Laure! It was my pleasure to give the lecture, and hopefully the video can be widely shared soon!
Laure Ho
@LaureHohoho
Excellent talk by @OriolVinyalsML from @DeepMindAI on RNNs and LSTMs @ucl! Even briefly addressed the newly released GPT-2 by @OpenAI. I am highly tempted to play with its ability to interpret TL;DR and summarize text 🤓
-
Oriol Vinyals
@OriolVinyalsML
Progress of samples from statistical language models.
web.stanford.edu/~jurafsky/slp3/3.pdf
www.cs.utoronto.ca/~ilya/pubs/2011/LANG-RNN.pdf
arxiv.org/abs/1602.02410
arxiv.org/abs/1801.10198
blog.openai.com/better-language-models/
(this is just a sample! 😃)
Oriol Vinyals
@OriolVinyalsML
Similar, but from faces (h/t @goodfellow_ian):
Ian Goodfellow
@goodfellow_ian
4.5 years of GAN progress on face generation. arxiv.org/abs/1406.2661 arxiv.org/abs/1511.06434 arxiv.org/abs/1606.07536 arxiv.org/abs/1710.10196 arxiv.org/abs/1812.04948
-
Oriol Vinyals
@OriolVinyalsML
Scale + compute: the deep learning formula that does not disappoint. Congrats @AlecRad @ilyasut et al!!
OpenAI
@OpenAI
We've trained an unsupervised language model that can generate coherent paragraphs and perform rudimentary reading comprehension, machine translation, question answering, and summarization — all without task-specific training: blog.openai.com/better-language-models/
-
Oriol Vinyals
@OriolVinyalsML
Most surprising facts about #alphastar:
1. A Neural Network can play full StarCraft -- without hard coding
2. Pure imitation learning achieves Gold/Platinum
3. Naive Self-Play "gets stuck", but AlphaStar league yields a GM agent
Full highlights: www.youtube.com/watch?v=6EQAsrfUIyoThis media couldn't be displayed. See it on the original post. -
Oriol Vinyals reposted
Liv Boeree
@Liv_Boeree
Seems like a good time to retweet this chart...
Incredible feat by @DeepMindAI - #alphastar just thoroughly defeated the world’s best humans in
#starcraft2 . Wow. -
Oriol Vinyals
@OriolVinyalsML
Happy that we could share #alphastar progress with you all! Good Games @LiquidTLO and @Liquid_MaNa, and @Artosis and @RotterdaM08 for a great show! You can see all the details in the blog.
deepmind.com/blog/alphastar-mastering-real-time-strategy-…
Oriol Vinyals
@OriolVinyalsML
I hope you enjoy these visualisations, which are great! Huge kudos to @maxjaderberg and @wojczarnecki
-
Oriol Vinyals
@OriolVinyalsML
Great to have @LiquidTLO at @DeepMindAI! Good Luck & Have Fun : )
Dario TLO Wünsch
@TLOgg
A lot of excitement in the room. Livestream @ 18:00 GMT: youtube.com/c/deepmind twitch.tv/StarCraft
-
Oriol Vinyals
@OriolVinyalsML
We’ve spent the past several months hard at work on #starcraft2 - tomorrow we’ll be hosting a livestream demonstration! Excited to hear what you all make of it : )
Tune in @ 6:00pm GMT: www.youtube.com/c/deepmind / www.twitch.tv/starcraft -
Oriol Vinyals reposted
Google DeepMind
@GoogleDeepMind
Join us and @Blizzard_Ent this Thursday at 6:00pm GMT for an exciting #starcraft demonstration, hosted by @Artosis and @RotterdaM08!
Livestream on YouTube: www.youtube.com/c/deepmind
Read more about #starcraft2 as an environment for AI research: deepmind.com/blog/deepmind-and-blizzard-open-starcraft-ii… -
Oriol Vinyals
@OriolVinyalsML
"Failing" is too often portrayed as something negative. Failures, and not successes, have been a higher drive for me both personally and professionally (and there have been plenty!). If you never fail, you are doing it wrong : )
-
Oriol Vinyals
@OriolVinyalsML
True story :) @JeffDean converts bugs to better algorithms.
Rajat Monga
@rajatmonga
In our early days at Google Brain, @JeffDean wrote an implementation of k-means. @OriolVinyalsML realized it wasn't "correct" but turned out it worked better than the standard algorithm. We had to call it j-means :)
-
Oriol Vinyals
@OriolVinyalsML
Very interesting visualization of 80 years of #ai research represented as connectionist (~neural nets) vs symbolic (~rule based) . I wonder what this will look like in another 80 years!
Source: neurovenge.antonomase.fr -
Oriol Vinyals
@OriolVinyalsML
Pleasure to be back at #blizzcon! Great discussion with @day9tv about why @StarCraft II is such an incredible AI research platform and a natural progression from DeepMind’s work on Atari and Go. @Blizzard_Ent
-
Oriol Vinyals
@OriolVinyalsML
In celebration to open sourcing Graph Networks, I assembled some slides (adapted from Yujia Li) which should serve as a nice tutorial -- feedback welcome!
Tutorial: docs.google.com/presentation/d/e/2PACX-1vSA16UYIp9wf8F9T7…
Open Source: github.com/deepmind/graph_nets -
Oriol Vinyals
@OriolVinyalsML
Between the no name change decision and not doing more to ensure a more fair distribution of the waitlist tickets (other than a FIFO queue), I am quite disappointed @NipsConference today! #protestnips
-
Oriol Vinyals
@OriolVinyalsML
Even I can play the piano now! Congrats!
Chris Donahue
@chrisdonahuey
Excited to announce Piano Genie, an intelligent controller that allows anyone to improvise on the piano! This was my internship project at @GoogleMagenta with @iansimon and @sedielem. Blog post with more information: magenta.tensorflow.org/pianogenie
-
Oriol Vinyals
@OriolVinyalsML
Best GAN samples ever yet? Very impressive ICLR submission! BigGAN improves Inception Scores by >100.
Paper: openreview.net/pdf?id=B1xsqj09Fm
Lots more samples: drive.google.com/drive/folders/1lWC6XEPD0LT5KUnPXeve_kWeY… -
Oriol Vinyals
@OriolVinyalsML
It has been 250 -> 500 -> 1000 -> 1600. Disappointed at sub-exponential grow :) #winteriscoming #linearorexponential
Yann LeCun
@ylecun
1600 submissions to ICLR.
Up from about 1000 last year. -
Oriol Vinyals
@OriolVinyalsML
Interesting papers using github.com/deepmind/pysc2 showing the difficulty of StarCraft2 1vs1 for Deep RL (even w/ handcrafting & simplifications): arxiv.org/abs/1809.09095, arxiv.org/abs/1809.07193
Note that Blizzard AIs are easy to beat as human. Looking forward to more progress! -
Oriol Vinyals reposted
Ali Eslami
@arkitus
Conditional Neural Process implementation by Marta Garnelo! Neural networks meet stochastic processes. Useful for few-shot regression, classification and meta-learning:
1. Repo: github.com/deepmind/conditional-neural-process
2. Notebook: github.com/deepmind/conditional-neural-process/blob/maste…
3. Paper: arxiv.org/abs/1807.01613 -
Oriol Vinyals
@OriolVinyalsML
For evaluating generative models I prefer downstream applications or metrics based on test data. But if you are into sample quality, here is a pretty powerful non-parametric model.
Demo: images.google.com -
Oriol Vinyals
@OriolVinyalsML
Given that NIPS sold out in 10 minutes, have you reserved spots for underrepresented groups so that we see more diversity this year? @NipsConference @hannawallach @hugo_larochelle
NeurIPS Conference
@NeurIPSConf
#nips2018 The main conference sold out in 11 minutes 38 seconds
-
Oriol Vinyals
@OriolVinyalsML
Highly recommended, hopefully it will be recorded for others to enjoy! Assuming independence, the more seeds you plant, the more chances one will sprout. 🌱🎲 #mlss2018
Shakir Mohamed
@shakir_za
First time I'll be visiting Madrid ✈️! Looking forward to speaking at the Machine Learning Summer where I'll give 3 lectures with the aim of 'planting the seeds of probabilistic thinking'. 🌱🎲 #mlss2018
-
Oriol Vinyals
@OriolVinyalsML
Regardless of the result, #openai5 reaffirms that with current techniques, learning >>> hand crafted rules for complicated video games. Congrats and good luck on next matches! openai.com/five/
-
Oriol Vinyals reposted
Google AI
@GoogleAI
Check out Universal Transformers, new research from the Google Brain team & @DeepMindAI that extends last year's Transformer (a neural network architecture based on a self-attention mechanism) to be computationally universal. blog.research.google/2018/08/moving-beyond-translation-wi…
-
Oriol Vinyals
@OriolVinyalsML
See you tomorrow at 12:00 at the Program Induction #nampi Workshop where I'll talk about #spiral from @yaroslav_ganin and also some unpublished work with @j_gauthier and @ilyasut. #icml2018 uclmr.github.io/nampi/
-
Oriol Vinyals
@OriolVinyalsML
Universal Transformers propose to augment Transformers with Recurrence in depth and Adaptive Computation Time. This model outperforms Vanilla Transformers in MT / bAbI / LA / LTE.
Paper: arxiv.org/abs/1807.03819
Code: Soon in github.com/tensorflow/tensor2tensor -
Oriol Vinyals reposted
Aäron van den Oord
@avdnoord
Our latest work is out!
Representation Learning with Contrastive Predictive Coding (CPC).
Autoregressive modeling meets contrastive losses in the latent space.
Learn useful representations in an unsupervised way.
-> On Audio, Vision, NLP and RL.
Arxiv: arxiv.org/abs/1807.03748 -
Oriol Vinyals
@OriolVinyalsML
Welcome back, gradients! This method is orders of magnitude faster than state-of-the-art non-differentiable techniques.
DARTS: Differentiable Architecture Search by Hanxiao Liu, Karen Simonyan, and Yiming Yang.
Paper: arxiv.org/abs/1806.09055
Code: github.com/quark0/darts -
Oriol Vinyals
@OriolVinyalsML
Neural Video Games: distilling video games with Neural Networks so accurately that you can play as if you were playing the real thing in real time. Congrats @arkitus @DeepSpiker & al! #neuralvideogames science.sciencemag.org/content/360/6394/1204
-
Oriol Vinyals
@OriolVinyalsML
Graph Neural Networks / Relational Networks are models worth studying. We wrote a pretty comprehensive review about them which I hope you will find helpful (code forthcoming!). arxiv.org/abs/1806.01261
-
Oriol Vinyals reposted
Yaroslav Ganin
@yaroslav_ganin
Just learned that our paper "Synthesizing Programs for Images using Reinforced Adversarial Learning" (aka SPIRAL) got a long talk at @icmlconf. If you haven't already, go check it out: arxiv.org/abs/1804.01118
-
Oriol Vinyals
@OriolVinyalsML
A nice update on StarCraft2 minigames thanks to Relational Architectures + IMPALA: 196 vs 177 (GrandMaster) vs 104 (previous best agent). More results in arxiv.org/abs/1806.01830
This media couldn't be displayed. See it on the original post. -
Oriol Vinyals
@OriolVinyalsML
Very nice results on imitation learning from (few) YouTube videos. Unsupervised learning of Embeddings >> Pixel Embeddings. arxiv.org/abs/1805.11592
-
Oriol Vinyals reposted
Aäron van den Oord
@avdnoord
VQ-VAE (arxiv.org/abs/1711.00937 and avdnoord.github.io/homepage/vqvae/) is now open source in DM-Sonnet! Here's an example iPython notebook on how to use it for images: github.com/deepmind/sonnet/blob/master/sonnet/examples/vq…
-
Oriol Vinyals
@OriolVinyalsML
One of the most exciting aspects of AI is how it can empower us. Great example of this applied to art with the "texturer" from @nonoesp nono.ma/suggestive-drawing
-
Oriol Vinyals
@OriolVinyalsML
Very cool analysis on compute vs year for some of the most notable advances in AI. blog.openai.com/ai-and-compute/
-
Oriol Vinyals
@OriolVinyalsML
Very impressive application of deep learning from Chen Chen & @intel for "Seeing in the Dark" with regular cameras. Code available, too! github.com/cchen156/Learning-to-See-in-the-Dark
-
Oriol Vinyals
@OriolVinyalsML
I just signed this. I hope that @NatMachIntell will try to find a good compromise here. See you at #iclr2018, where we use OpenReview, the drastic version of open access!
Thomas G. Dietterich
@tdietterich
Several machine learning researchers have signed a statement regarding the upcoming launch of Nature Machine Intelligence. If you agree, I encourage you to sign this as well. openaccess.engineering.oregonstate.edu
-
Oriol Vinyals
@OriolVinyalsML
A great talk by my friend @ilyasut on two important topics -- meta learning, and self play. I'd add program synthesis as a third exciting area which I'm looking forward to learning more at #iclr2018. See you there! ✈️🇨🇦
Ilya Sutskever
@ilyasut
a talk on some of our recent work: www.youtube.com/watch?v=9EN_HoEk3KY
-
Oriol Vinyals
@OriolVinyalsML
Great SIGGRAPH paper which uses RL/ML for motion generation. I predict many more applications of machine learning to graphics & other domains -- looking forward to that! arxiv.org/abs/1804.02717
-
Oriol Vinyals
@OriolVinyalsML
RL + GANs: Program synthesis with an agent that uses a paint program to fool a discriminator. Paper+Blog: deepmind.com/blog/learning-to-generate-images/
-
Oriol Vinyals
@OriolVinyalsML
Fun interview from @Blizzard_Ent on our views on the role of @StarCraft in #agi.
Justin Groot
@JustinGroot3
Also interviewed ex-Brood War god and current AI researcher @OriolVinyalsML about DeepMind's efforts to beat humans at StarCraft II:
starcraft2.com/en-us/news/21509421 -
Oriol Vinyals
@OriolVinyalsML
GANgratulations! Looking forward to more progress on generative models and also their evaluation (arxiv.org/abs/1511.01844 for a thought provoking note on the subject).
Ian Goodfellow
@goodfellow_ian
MIT Technology Review listed GANs as 1 of 10 breakthrough technologies this year: www.technologyreview.com/lists/technologies/2018/
-
Oriol Vinyals
@OriolVinyalsML
Very cool dataset / work from FAIR! Exciting direction for dialogue research, where you condition on "personality". Seq2Seq is certainly more entertaining! arxiv.org/abs/1801.07243
-
Oriol Vinyals
@OriolVinyalsML
Evolution > RL (for now...) for architecture search. New SOTA on CIFAR10 (2.13% top 1) and ImageNet (3.8% top 5). 🔥 450 GPU / 7 days & 900 TPU / 5 days 🔥 arxiv.org/abs/1802.01548
-
Oriol Vinyals
@OriolVinyalsML
Two years ago I left @GoogleBrain to join @DeepMindAI (two amazing teams!). My farewell speech was generated by an LSTM trained on all my emails.
-
Oriol Vinyals
@OriolVinyalsML
"Optimal" running, brought to you by Deep RL. arxiv.org/abs/1801.00690
-
Oriol Vinyals
@OriolVinyalsML
Excellent summary of AI developments -- so good I don't have to write one on my own! =]
Denny Britz
@dennybritz
I wrote up a (not so brief) summary of AI developments that stood out to me in 2017 www.wildml.com/2017/12/ai-and-deep-learning-in-2017-a-yea…
-
Oriol Vinyals
@OriolVinyalsML
May your gradients flow and your learning rates be large in 2018! For an exciting topic of research, here are my Meta Learning NIPS Symposium slides. drive.google.com/file/d/15fvtfWNw2T4DcU7Kw1cfT9ZjJTwAFGOg…
-
Oriol Vinyals
@OriolVinyalsML
An agent playing StarCraft2 from raw pixels (+ rendering StarCraft2 on Linux). Top left shows the 64x64 screen aka "what the agent sees".
-
Oriol Vinyals
@OriolVinyalsML
As someone who is very close to the issue (my wife is a ML researcher), I'm very happy to (finally) see this! I have reported inappropriate behavior before and will continue doing so. #diversity #ml #wiml
Thomas G. Dietterich
@tdietterich
The NIPS Foundation has posted a "Statement on inappropriate behavior" at #nips2017. See nips.cc
-
Oriol Vinyals
@OriolVinyalsML
Deep Learning: Practice and Trends. Tutorial tomorrow at 8AM @ Hall A. @scott_e_reed @NandoDF #nips2017 #coffee #jetlag
Oriol Vinyals
@OriolVinyalsML
Slides here: docs.google.com/presentation/d/e/2PACX-1vQMZsWfjjLLz_wi8i…
-
Oriol Vinyals
@OriolVinyalsML
Slides from our #nips2017 Tutorial on Deep Learning with @scott_e_reed & @NandoDF posted here: docs.google.com/presentation/d/e/2PACX-1vQMZsWfjjLLz_wi8i…
-
Oriol Vinyals reposted
Maithra Raghu
@maithra_raghu
Deep Learning: Bridging Theory and Practice happening tomorrow at #nips2017! Final Schedule: ludwigschmidt.github.io/nips17-dl-workshop-website/ @lschmidt3 @OriolVinyalsML @rsalakhu We have an exciting program with talks by Yoshua Bengio @goodfellow_ian Peter Bartlett Doina Precup Percy Liang Sham Kakade!
-
Oriol Vinyals
@OriolVinyalsML
Tomorrow I'll be talking about StarCraft2 at the NIPS Deep Reinforcement Learning Symposium, at 7:50PM. Full schedule: sites.google.com/corp/view/deeprl-symposium-nips2017/
-
Oriol Vinyals
@OriolVinyalsML
Great stuff from @maxjaderberg et al on hyperparameter search based on evolving populations that actually outperforms random search! deepmind.com/blog/population-based-training-neural-networks/
-
Oriol Vinyals
@OriolVinyalsML
Almost 3000x speedup with Parallel WaveNet (Google Speech production model). Brought to you by @avdnoord et al. Drastic! More details & link to paper: deepmind.com/blog/high-fidelity-speech-synthesis-wavenet/
-
Oriol Vinyals
@OriolVinyalsML
Before solving intelligence, shall we solve automatically finding reviewers' conflicts of interests? #iclr2018 #programchairing
-
Oriol Vinyals
@OriolVinyalsML
Happy to be back at @la_UPC and give a talk about RL, StarCraft2, and AlphaGo. See you Monday November 20 at 11:30am in #barcelona! You must register to attend: www.eventbrite.com/e/upc-bsc-deep-learning-open-lecture-o…
-
Oriol Vinyals
@OriolVinyalsML
[Sutskever, Vinyals & Le, 2017] @OpenAI @DeepMindAI @GoogleBrain 🙂 #friendship #deeplearning #seq2seq
-
Oriol Vinyals
@OriolVinyalsML
Great work by @avdnoord who managed to train a generative model w/ discrete variables. Audio samples here: avdnoord.github.io/homepage/vqvae/ #nips2017
-
Oriol Vinyals
@OriolVinyalsML
#blizzcon2017 mode on! Together with @Blizzard_Ent & @DeepMindAI we are organizing a SC2 AI Workshop with many talented participants!
-
Oriol Vinyals
@OriolVinyalsML
#iclr18 submissions grow to 1003 from 490! Last 24 hours before submissions deadline shown below for both years. #lastminute #deeplearning
-
Oriol Vinyals reposted
hardmaru
@hardmaru
I'm blown away by the method and results in this paper. Progressive growing neural nets may be a trend we will see in 2018.
This media couldn't be displayed. See it on the original post. -
Oriol Vinyals
@OriolVinyalsML
Quite addicting to witness self-play improvements over time. Go #alphagozero! deepmind.com/blog/alphago-zero-learning-scratch/
-
Oriol Vinyals
@OriolVinyalsML
See you at Blizzcon! We are organizing (w/ @Blizzard_Ent) a StarCraft II AI workshop. Apply soon (limited spots)! us.battle.net/sc2/en/blog/21048078/announcing-the-starcra…
-
Oriol Vinyals
@OriolVinyalsML
WaveNet becomes the voice of Google Assistant! 1000x faster & better, brought to you by a neural net : ) deepmind.com/blog/wavenet-launches-google-assistant/
-
Oriol Vinyals
@OriolVinyalsML
Machine Learning applied to Computer Graphics is quite exciting IMO. PS No deep learning here! physicsforests.com
-
Oriol Vinyals
@OriolVinyalsML
Amsterdam is such a beautiful city! Come say hi tomorrow! @quvalab www.meetup.com/Deep-Learning-AI/events/243203062/
-
Oriol Vinyals
@OriolVinyalsML
A cool viz on progress with Deep Learning for Video Games. Need bigger circles for next years : ) Source: arxiv.org/abs/1708.07902
-
Oriol Vinyals
@OriolVinyalsML
Pretty nice video summary describing our @StarCraft paper (colab between @Blizzard & @DeepMindAI). www.youtube.com/watch?v=St5lxIxYGkI
-
Oriol Vinyals
@OriolVinyalsML
If you are at #icml today, come check out the Video Games and Machine Learning #vgml workshop! Great speakers! @syhw syhw.github.io/vgml_workshop_icml2017/
-
Oriol Vinyals
@OriolVinyalsML
Launching StarCraft2 Learning Environment (Linux)! A3C baseline on mini & full game + policies trained on 1M replays.deepmind.com/blog/deepmind-and-blizzard-open-starcraft-ii…
-
Oriol Vinyals
@OriolVinyalsML
Thanks all for coming and great questions during the tutorial! Slides at sites.google.com/view/seq2seq-icml17/ #icml2017 #seq2seq
Lisha
@lishali88
Tutorial on #seq2seq with @OriolVinyalsML and Navdeep Jaitly. #icml2017
-
Oriol Vinyals
@OriolVinyalsML
Very well timed switch to Neural MT by Facebook given our #icml2017 tutorial on seq2seq with Navdeep! #28hourflight code.facebook.com/posts/289921871474277/transitioning-ent…
-
Oriol Vinyals
@OriolVinyalsML
This is a very cool / useful viz of VAEs from @dpkingma!
Durk Kingma
@dpkingma
A figure I made for explaining variational autoencoders (VAEs) as part of a larger work-in-progress.
-
Oriol Vinyals
@OriolVinyalsML
Retweeting to support my optimism about using AI for good.
Andrew Ng
@AndrewYNg
Kudos Mark Zuckerberg for speaking out against AI fear mongering. fortune.com/2017/07/25/mark-zuckerberg-elon-musk-artifici…
-
Oriol Vinyals
@OriolVinyalsML
When you fix a bug in your code, and results get worse. #bugpropagation #deeplearning
-
Oriol Vinyals
@OriolVinyalsML
Hoping to see more work on model based RL. This is our small contribution, with loads left to do / improve upon! deepmind.com/blog/agents-imagine-and-plan/
-
Oriol Vinyals
@OriolVinyalsML
The LSTM Strikes Back: sub 60 perplexities on LM PTB (single model). All you need is a (well tuned) LSTM :) arxiv.org/abs/1707.05589
-
Oriol Vinyals
@OriolVinyalsML
Nice progress on few-shot learning in the past year! 10% absolute improvement on miniImagenet: arxiv.org/abs/1707.03141
-
Oriol Vinyals
@OriolVinyalsML
Agents having fun in Parkour! Cool paper from colleagues at DeepMind arxiv.org/abs/1707.02286
-
Oriol Vinyals
@OriolVinyalsML
NoisyNets: great to see more work on exploration in RL! Also very simple: add noise to parameters. Fortunato et al, arxiv.org/abs/1706.10295
-
Oriol Vinyals
@OriolVinyalsML
A few weeks after FB new Neural MT achieved 25.1 BLEU in 18 days x 1 GPU, 28.4 is attained in 3 days x 8 GPUs. Nice progress!
Google AI
@GoogleAI
Announcing the Tensor2Tensor open-source system for training a wide variety of deep learning models in #tensorflow research.googleblog.com/2017/06/accelerating-deep-learnin…
-
Oriol Vinyals
@OriolVinyalsML
Navdeep and I are giving a tutorial at #icml (August) on seq2seq. What do you prefer to see (most) in a tutorial? 2017.icml.cc/Conferences/2017/Tutorials
-
Oriol Vinyals
@OriolVinyalsML
Score overflow at 1M points in PacMan. Nicely done : ) www.maluuba.com/hra
-
Oriol Vinyals
@OriolVinyalsML
It's never clear how to measure overfitting in GANs, but generating up to 4kx4k res faces is pretty drastic! mtyka.github.io/machine/learning/2017/06/06/highres-gan-f…
-
Oriol Vinyals
@OriolVinyalsML
Cool architecture for relational reasoning by @santoroAI @dnraposo et al! arxiv.org/abs/1706.01427
-
Oriol Vinyals
@OriolVinyalsML
CfP is out! GL HF!
Gabriel Synnaeve
@syhw
The CfP for our Video Games and Machine Learning workshop at ICML2017 has arrived syhw.github.io/vgml_workshop_icml2017
(cc @togelius @OriolVinyalsML) -
Oriol Vinyals
@OriolVinyalsML
US leaving Paris Climate Accord... Time to train a DeepRL adviser to the president? www.deeprlbootcamp.berkeley.edu #covefefe
-
Oriol Vinyals reposted
Demis Hassabis
@demishassabis
#alphago wins game 1! Ke Jie fought bravely and some wonderful moves were played.
-
Oriol Vinyals
@OriolVinyalsML
A typical training curve in Montezuma's Revenge (note: there are several random seeds which overlap) 😭 #nips #rl #exploration
-
Oriol Vinyals
@OriolVinyalsML
Quote from the future: "Remember when we could fit #iclr in a single stadium?". #iclr17 #deeplearning
-
Oriol Vinyals
@OriolVinyalsML
Nice work by our friends at FAIR! Research trend LSTM->CNN. I'm certainly hoping for an LSTM comeback. #lstmsvscnns code.facebook.com/posts/1978007565818999/a-novel-approach…
-
Oriol Vinyals
@OriolVinyalsML
This is excellent, congrats! I heard the team has since moved to Las Vegas 😃
DeepStack
@DeepStackAI
The DeepStack paper was just published in the print edition of Science, and we made the front cover! science.sciencemag.org/content/356/6337/508
-
Oriol Vinyals
@OriolVinyalsML
It is rare these days to see single author papers. Congrats Daniel (who amazingly is an undergrad student)! #iclr2017
Oriol Vinyals
@OriolVinyalsML
Cc @hexahedria
Oriol Vinyals
@OriolVinyalsML
Paper is: openreview.net/forum?id=HJ0NvFzxl¬eId=HJ0NvFzxl
-
Oriol Vinyals
@OriolVinyalsML
Some vision convnets as seen by Chiyuan at #iclr2017
-
Oriol Vinyals
@OriolVinyalsML
Cool work!! We've come a long way since image captioning. And the good news is, there's still lots to improve, too!
Devi Parikh
@deviparikh
Visual Dialog: code, demo (i.e., a chatbot that can see), AND code for demo: all now available at www.visualdialog.org!
-
Oriol Vinyals
@OriolVinyalsML
"The key idea is to cache hidden states to avoid redundant computation" a.k.a. rediscovering the backpropagation algorithm.
Miles Brundage
@Miles_Brundage
"Fast Generation for Convolutional Autoregressive Models," Ramachandran et al: arxiv.org/abs/1704.06001 21x+183x speedup 4 WaveNet+PixelCNN++
-
Oriol Vinyals
@OriolVinyalsML
This is an exciting opportunity, with travel grants and such for those traveling from around Africa & World! #mldiversity
Shakir Mohamed
@shakir_za
Announcing the Deep Learning Indaba; to stimulate participation of fellow Africans in Machine Learning. Help spread the word! .@DeepIndaba
This post is unavailable. See it on X. -
Oriol Vinyals
@OriolVinyalsML
Pretty sketchy research: arxiv.org/abs/1704.03477
-
Oriol Vinyals
@OriolVinyalsML
Next chapter for #alphago in May! Thankfully after the #nips deadline :) deepmind.com/blog/exploring-mysteries-alphago/
This media couldn't be displayed. See it on the original post. -
Oriol Vinyals
@OriolVinyalsML
Another Wrapper for TensorFlow (WTF++)! As a user of #sonnet, I enjoy its variable sharing semantics. github.com/deepmind/sonnet
-
Oriol Vinyals
@OriolVinyalsML
Great work from my colleagues @OpenAI. Would be fun to collaborate! blog.openai.com/unsupervised-sentiment-neuron/ blog.openai.com/learning-to-communicate/ blog.openai.com/evolution-strategies/
-
Oriol Vinyals
@OriolVinyalsML
Pretending to do real science (as opposed to ML) is fun! Neural Message Passing for Quantum Chemistry by Justin&Sam arxiv.org/abs/1704.01212
-
Oriol Vinyals
@OriolVinyalsML
A typical overfitting curve.
-
Post not available here, view on X
Oriol Vinyals
@OriolVinyalsML
In a few years, once Global Warming fully kicks in!
-
Oriol Vinyals
@OriolVinyalsML
Our lip reading work will be presented as an oral at CVPR! You can read more at arxiv.org/abs/1611.05358
Oriol Vinyals
@OriolVinyalsML
send me an email : )
-
Oriol Vinyals
@OriolVinyalsML
Looks like machine learning conferences selling out is the norm now!
-
Oriol Vinyals reposted
Xavi Giró
@DocXavi
Today at 4pm an open lecture by @OriolVinyalsML from @DeepMindAI in our Master in Computer Vision.
CVC_UAB
@CVC_UAB
Remember everybody! Open Lecture tomorrow at the @Enginyeria_UAB @UABBarcelona @PRUAB
-
Oriol Vinyals
@OriolVinyalsML
London to Scotland! 🚂🚂 Looking forward to Edinburgh Deep Learning Workshop.
-
Oriol Vinyals
@OriolVinyalsML
Happy Women's Day! Hoping to see more diversity in our field -- this pic is from a talk Meire gave today. #internationalwomensday
-
Oriol Vinyals
@OriolVinyalsML
ML & health: awesome news! (89% ML vs 73% human) Something my parents can appreciate more than detecting cats! research.googleblog.com/2017/03/assisting-pathologists-in…
-
Oriol Vinyals
@OriolVinyalsML
Networks computing their own synthetic gradients. Now with some theory brought to you by Woj and Grzegorz. arxiv.org/abs/1703.00522
-
Oriol Vinyals
@OriolVinyalsML
Working on ICML vs playing with affinelayer.com/pixsrv/. #choices
-
Oriol Vinyals
@OriolVinyalsML
A preview from my lecture @ucl tomorrow.
-
Oriol Vinyals
@OriolVinyalsML
@iclr2017 decisions are out! 15 orals, and 181 posters out of 507 submissions. I attach some stats for those curious.
-
Oriol Vinyals
@OriolVinyalsML
I typically don't post political statements on Twitter. But #muslimban goes beyond politics into basic human rights.
Andrew Ng
@AndrewYNg
Speak up on twitter if you haven't yet about #muslimban. #deleteuber influenced Travis Kalanick. Every voice, including yours, matters.
-
Oriol Vinyals
@OriolVinyalsML
This seems pretty entertaining especially if you have a PhD : )
x.com/i/moments/822706973531631616 -
Oriol Vinyals
@OriolVinyalsML
Excited to be teaching a lecture @UCL! Hopefully we can keep the Academia <-> Industry relationship as bidirectional as possible : )
Google DeepMind
@GoogleDeepMind
#deepmind teams up with @UCL to teach Masters training module. Read more on our collaborations with academia here: deepmind.google/blog/our-collaborations-with-academia-to-…
-
Oriol Vinyals
@OriolVinyalsML
A year ago I left Brain to join DeepMind. My farewell speech was of course generated by an LSTM trained on all my emails.
-
Oriol Vinyals
@OriolVinyalsML
Nice summary of Brain's 2016! Despite officially moving a year ago, I'm still involved in many collaborations! research.googleblog.com/2017/01/the-google-brain-team-loo…
-
Oriol Vinyals
@OriolVinyalsML
Great to see #alphago tested "in the wild"! Congrats to the team!
Demis Hassabis
@demishassabis
Excited to share an update on #alphago!
-
Oriol Vinyals
@OriolVinyalsML
Generative models from kids drawings -- sadly, not a neural net (yet?) www.instagram.com/thingsihavedrawn/
-
Oriol Vinyals
@OriolVinyalsML
I was hoping for arxiv to be quiet during holidays, so I could decrease my "to read" stack. Oh well... #maybenextyear
-
Oriol Vinyals
@OriolVinyalsML
Planning to do a fair bit of reading / discussing over the holiday break for #iclr #greatpapers #openreview!
-
Oriol Vinyals
@OriolVinyalsML
The first rule of #rocketai is you don't talk about #rocketai.
-
Oriol Vinyals
@OriolVinyalsML
I'll be giving a talk today at the RNN symposium, featuring #cakes and #starcraft at 3:40PM.
-
Oriol Vinyals
@OriolVinyalsML
Very nice image captioning results!
Richard Socher
@RichardSocher
Next level for deep vision-language models: knowing when to look
arxiv.org/abs/1612.01887
By intern Jiasen Lu + researcher Caiming Xiong etal -
Oriol Vinyals
@OriolVinyalsML
One Shot Learning with seq2seq poster tonight at #nips2016. I attach the poster here for those who cannot make it. arxiv.org/abs/1606.04080
-
Oriol Vinyals
@OriolVinyalsML
Going to #nips2016 AND going to my hometown : ) See you there! #barcelona
-
Oriol Vinyals
@OriolVinyalsML
Happy Birthday @StarCraft!
ESPN Esports
@ESPN_Esports
On this day in esports, 18 years ago, @StarCraft: Brood War -- a game that launched a thousand careers -- was released.
-
Oriol Vinyals
@OriolVinyalsML
Pix2Pix -- Great work from Berkeley! And with source code already! arxiv.org/abs/1611.07004 github.com/phillipi/pix2pix
-
Oriol Vinyals
@OriolVinyalsML
Lip Reading Sentences in the Wild! New Dataset + nice results on BBC News. seq2seq+att : ) arxiv.org/abs/1611.05358 www.youtube.com/watch?v=5aogzAUPilE
-
Post not available here, view on X
Oriol Vinyals
@OriolVinyalsML
Efectivamente soy -=TheSaint=- : )
Oriol Vinyals
@OriolVinyalsML
Sip! Sabeis algo de Alfa?
-
Oriol Vinyals reposted
Dan Stemkoski
@Artosis
It was just such a pleasure to speak with @OriolVinyalsML @TimoEwalds Paul Keet and Kevin Calderone about AI in @StarCraft . So awesome guys
-
Oriol Vinyals
@OriolVinyalsML
StarCraft keeps coming back in my life : ) Together with Blizzard, we are opening StarCraft2 for AI research! @Blizzard_Ent @DeepMindAI
-
Oriol Vinyals reposted
Chris Olah
@ch402
Comparison of different GAN variants. I couldn't fit the full version in our paper.
-
Oriol Vinyals
@OriolVinyalsML
Cool work by Nal et al!!
Nal
@nalkalc
New neural net for Language and Machine Translation! Fast and simple way of capturing very long range dependencies arxiv.org/abs/1610.10099
-
Oriol Vinyals
@OriolVinyalsML
Nice building! About to present #emtechmit
-
Oriol Vinyals reposted
Shakir Mohamed
@shakir_za
Thought-provoking connection between different areas of machine learning. From my colleagues .@pfau .@OriolVinyalsML
Stat.ML Papers
@StatMLPapers
Connecting Generative Adversarial Networks and Actor-Critic Methods. (arXiv:1610.01945v1 [cs.LG]) arxiv.org/abs/1610.01945
-
Post not available here, view on X
Oriol Vinyals
@OriolVinyalsML
ST vectors used many times, I would spell since it is also a main contribution :)
-
Oriol Vinyals
@OriolVinyalsML
I've been obsessed with sequences for a while now. Video generation is perhaps the more drastic kind of sequence. Nice to see progress! : )
Google DeepMind
@GoogleDeepMind
Excited to share our new generative model! #videopixelnetworks predict detailed future video frames arxiv.org/abs/1610.00527
-
Oriol Vinyals
@OriolVinyalsML
Traveling to SF on Wednesday. Looking forward meeting old friends & colleagues! Ping me if you are around!
-
Oriol Vinyals
@OriolVinyalsML
Nice work from my colleagues at Brain. Congrats!
Google AI
@GoogleAI
Learn about the GNMT system, which utilizes state-of-the-art training techniques to improve machine translation - goo.gl/v6bvD5
-
Oriol Vinyals
@OriolVinyalsML
ImageNet results out!
Andrej Karpathy
@karpathy
ImageNet ILSVRC 2016 results are out image-net.org/challenges/LSVRC/2016/results congrats to Trimps-Soushen (0.02991 error), & FAIR team with 0.03031.
-
Oriol Vinyals
@OriolVinyalsML
Another open source announcement! It took a while due to original research being done pre-TensorFlow, but our im2txt model is finally here!
Google AI
@GoogleAI
We’ve made the latest version of our image captioning system available as an open source model in #tensorflow - research.googleblog.com/2016/09/show-and-tell-image-capti…
-
Oriol Vinyals
@OriolVinyalsML
We just open sourced the SOTA single model (~30 PPL) trained on lm_1b. Paper: arxiv.org/abs/1602.02410 Code: github.com/tensorflow/models/tree/master/lm_1b
-
Oriol Vinyals
@OriolVinyalsML
Generating audio at 16000 samples / second, one sample at a time. Crazy times : )
Google DeepMind
@GoogleDeepMind
Excited to share our latest research on speech and music generation! Hear our #wavenets speak for themselves: deepmind.google/blog/wavenet-a-generative-model-for-raw-a…
-
Oriol Vinyals
@OriolVinyalsML
Excited to have won the MIT TR35 innovator award! Very grateful to do research alongside brilliant colleagues : ) www.technologyreview.com/lists/innovators-under-35/2016/p…
-
Oriol Vinyals
@OriolVinyalsML
I was very surprised about these results regarding SGD / Backprop. TL;DR: learn all the things : )
Google DeepMind
@GoogleDeepMind
Just out: our new paper unlocks neural nets by removing the dependency on backpropagation using Synthetic Gradients! arxiv.org/pdf/1608.05343
-
Oriol Vinyals
@OriolVinyalsML
Nice summary of our paper from @karpathy !
Andrej Karpathy
@karpathy
Another "finally got to read this paper" & "might as well post my notes" on Matching Networks for One Shot Learning github.com/karpathy/paper-notes/blob/master/matching_netw…
-
Oriol Vinyals reposted
OReillyAI
@OReillyAI
.@OriolVinyalsML discusses sequence-to-sequence #machinelearning, neural networks & the Traveling Salesman Problem oreil.ly/2a3LkfR
-
Oriol Vinyals
@OriolVinyalsML
Brexit indeed.
Edward Grefenstette
@egrefen
Well done Iceland :)
-
Oriol Vinyals
@OriolVinyalsML
Awesome to be at ICML this week : ) Congrats to the team, and looking forward catching up with many colleagues.
Google DeepMind
@GoogleDeepMind
#deepmind Team win 2 of 3 best paper awards at #icml2016! Congrats to all! arxiv.org/pdf/1601.06759 goo.gl/UIpzZt
-
Oriol Vinyals reposted
Google DeepMind
@GoogleDeepMind
PixelCNN 2.0: New state of the art generative model for conditional natural image synthesis arxiv.org/abs/1606.05328
-
Oriol Vinyals
@OriolVinyalsML
Excited to share my first DeepMind paper! Seq2Seq + Learning2Learn for One Shot Learning on ImageNet, Omniglot, & LM arxiv.org/abs/1606.04080
-
Oriol Vinyals reposted
Google
@Google
Our hearts are with the Orlando victims, their families and the LGBT community. #loveislove
-
Oriol Vinyals reposted
Andrej Karpathy
@karpathy
New blog post: "Deep Reinforcement Learning: Pong from Pixels" karpathy.github.io/2016/05/31/rl/ on policy gradients
-
Oriol Vinyals reposted
Hugo Larochelle
@hugo_larochelle
The #iclr2016 talks are now on videolectures.net: videolectures.net/iclr2016_san_juan/
Enjoy! :-) -
Oriol Vinyals reposted
Carlos Ciller
@carlos_ciller
"The 1st step in #ml is to start doing something better than random, then you overfit, then you regularize!" @OriolVinyalsML #mlss2016 #dl
-
Oriol Vinyals
@OriolVinyalsML
Very exciting to be back @UPC after so many years! x.com/JordiTorresBCN/status/731563855034355712
-
Oriol Vinyals
@OriolVinyalsML
Couldn't agree more : )
Isabelle Augenstein
@IAugenstein
Brainstorming at @uclmr : how to recommend "what method works best for NLP task X"? Our baseline is ready #nlproc
-
Oriol Vinyals
@OriolVinyalsML
GG!
Demis Hassabis
@demishassabis
#alphago wins game 5! One of the most incredible games ever. To comeback from the initial big mistake against Lee Sedol was mind-blowing!!!
-
Oriol Vinyals reposted
Demis Hassabis
@demishassabis
#alphago won game 3 and the match! Historic moment. In complete awe of Lee Sedol’s incredible genius, and proud of the amazing AlphaGo team!
-
Oriol Vinyals
@OriolVinyalsML
Waking up at 3AM London time for the last two days was totally worth it. Lots of excitement & respect for #leesedol and #alphago.
-
Oriol Vinyals
@OriolVinyalsML
Our new Language Modeling paper: arxiv.org/abs/1602.02410 With a single RNNLM (no ensembling), we reduce perplexity from ~50 down to 30.0!
-
Oriol Vinyals
@OriolVinyalsML
Fun times ahead!
Demis Hassabis
@demishassabis
Thrilled to officially announce the 5-game challenge match between #alphago and Lee Sedol in Seoul from March 9th-15th for a $1M prize!
-
Oriol Vinyals reposted
Demis Hassabis
@demishassabis
A 20-year dream comes to fruition - our program AlphaGo masters the game of Go, achieving a grand challenge of AI googleblog.blogspot.com/2016/01/alphago-machine-learning-…
-
Oriol Vinyals
@OriolVinyalsML
My next step: DeepMind. I'll certainly miss all the friends left behind, Google Brain, and the sun, but I'm thrilled about this new chapter.







































































































































































































































































































