Politicians, athletes, journalists, NGOs, companies... They all use
corporation-owned social networks to communicate with us. Citizens
shouldn't have to hand over their data to companies to read content
published by public entities. Public Feeds changes that by making posts
of popular accounts freely available.
Read more.
Yoav Shoham
Yoav Shoham (Hebrew: יואב שוהם; born 22 January 1956) is a computer scientist and a professor emeritus at Stanford University. His research spans artificial intelligence, logic and game theory. He has also founded and sold several AI companies. More on Wikipedia
What they write about when they write about running…
I’ve recently finished Haruki Murakami’s What I Talk About When I Talk About Running, and had previous read (and written about) Nick Thompson’s The Running Ground: A Father, a Son, and the Simplest of Sports. There are many parallels between the two, but also differences. Here’s my take.
- Both writers are seriously extreme runners. They’ve both run multiple marathons and completed 100k races. Thompson is probably somewhat more competitive, and holds the 50k American record for runners over 50, but I wouldn’t want to race Murakami.
- Fun fact: Murakami runs while listening to music; Thompson prefers meditative silence.
- Both have a separate day job, and are at the top of their game there (Murakami is a world-famous writer, and Thompson CEO of The Atlantic).
- If it weren’t already obvious from their accomplishments in both running and their day jobs, the books make it clear how ambitious, disciplined, and hard working they are.
- The reporting of their running exploits is similar in that it is factual. They report impressive accomplishments, but don’t brag (humbly or otherwise). Murakami does show a certain amount of self effacing, more than Thompson, perhaps reflecting the different cultural backgrounds.
- They both draw lessons from, or parallels between, running and other facets of life. For Thompson this means mostly relating running ethics to work ethics. For Murakami it’s more relating running to existential questions of the human condition. Maybe this reflects the difference between a journalist and a fiction writer.
- Murakami’s tone is philosophical. Thompson’s psychological. Indeed, a central theme in Thompson’s book (as reflected in the subtitle) is his relationship with his obviously talented, but equally obviously troubled, dad. Murakami mentions family (mostly his wife) only in passing. Thompson often mentions not only his dad but also his wife and kids.
- So it’s surprising that the tone of Murakami’s writing is more extroverted than Thompson’s. Murakami often speaks directly about how he feels, whereas Thompson relays the most difficult personal situations (and some of them must have been excruciating) in a measured, matter-of-fact, almost third-person tone. He doesn’t hold back on facts, but the reader is invited to project their emotions on the situation. This again may reflect the different cultural backgrounds and upbringings.
The upshot is that both books make for compelling reading, certainly for someone with an interest in running. I made a stronger connection with Thompson’s book, perhaps because of the personal story, and perhaps also because our orbits overlap; I know him personally and like him. But I enjoyed - and recommend - both.
Introducing the #aiindex2026: Our most comprehensive, independently sourced data analysis of AI’s trajectory, with a clear-eyed assessment of the critical gaps that remain. As AI advances rapidly, can the systems built around it keep up? Explore the data: hai.stanford.edu/ai-index/2026-ai-index-report
🚀 Today we’re launching a brand new @relay! If you want an AI team that works for you, now’s the time to start.
Here’s what makes our AI agents different:
Anyone can create agents. You work with AI agents just like you work with people. You ask your agent to do things for you and give it feedback to get better. No code, JSON, terminal, or MCP needed.
Agents are predictable and reliable. You teach your agent skills with simple prompts, and it turns those into easily understandable, consistent workflows. Plus, your agent can keep a human-in-the-loop for anything high stakes. No random actions you can’t explain.
To try it out, head over to @relay and get started for free. I can’t wait to hear your feedback.
p.s. like and RT to get a bonus code for 500 extra AI credits per month for a year. 🙏
Everyone talks about agentic flows, but few actually tackle the hard stuff to make them work. There’s no free lunch; prompt-and-pray is not a strategy. AI21 Maestro takes a different approach. See it in action on SWE-bench. www.ai21.com/blog/test-time-compute-swe-bench/
Sometimes the important stuff is not the glamorous. Although this is actually quite sexy, beside being critical for the enterprise. Especially if you’re turned on by things like steerability and long context.
WATCH: @DigEconLab Director @erikbryn and @Stanford Professor Emeritus of CS @yshoham joined @nxthompson in a conversation about moving AI beyond hype to real impact, covering everything from pilot-to-production challenges to the future of human-AI collaboration.↘️
Orchestrating Smarter AI Systems with AI21 Labs’ Yoav Shoham In this episode of AI Basics, @Jason sits down with Yoav Shoham (@yshoham) — Stanford professor emeritus and co-founder of @AI21Labs, creators of Jurassic-2, Wordtune, and the new orchestration system Maestro.
• Why enterprise AI struggles with reliability
• What orchestration really means (and why LLMs alone aren't enough)
• The pitfalls of “agent-washing”
• Small vs large models, agent-to-agent protocols, and where real opportunities lie
This one is for founders building with AI — if you're navigating hallucinations, chasing automation, or exploring multi-agent workflows, this episode is a must.
Dinner at the Elysée Palace hosted by President @EmmanuelMacron to discuss AI strategy. His mobilizing of the ecosystem to support French AI should be studied by every national leader.
It's not every day I get to interview a former principal scientist who worked at Google, and is a Professor Emeritus at Stanford University, about the state of AI.
But here we go. Introducing an hour with Yoav Shoham, @yshoham, AI pioneer and cofounder of @AI21Labs .
This will make you smarter, not that all my videos aren't that way. :-)
++++++++++++++++++
Here's what we discussed (this part was written by Chat GPT after I gave it the transcript of the video):
🚀 The State of AI Today •The pace of AI development is unprecedented, likened to a “universal firehose” of innovation. •Everyone—from your plumber to enterprise CTOs—is using AI. But not all use cases are equal or enterprise-ready.
🏢 Enterprise vs Consumer AI •Enterprise adoption is still slow compared to consumer. Shoham cites AWS data showing only 6% of AI pilots go into production. •Enterprises demand reliability, cost control, and explainability, which raw LLMs like ChatGPT don’t fully offer out of the box.
🧱 Beyond the LLM Hype •Shoham explains that pure LLMs aren’t enough. Enterprises need “compound AI systems” or “AI agents” that: •Use tools like calculators for arithmetic instead of relying on the model •Integrate with company databases via RAG (retrieval-augmented generation) •Plan, reason, and execute tasks through orchestrated workflows •AI21 Labs built Maestro, their orchestration system, to do exactly this.
🔐 Enterprise Concerns •Enterprises worry about IP leakage, data privacy, and hallucinations. •AI21 addresses this by running models on-prem or in VPCs, ensuring data doesn’t leave customer control.
📉 Why Models Still Fail •LLMs generate “authoritative bullshit” — convincing but wrong answers. •Shoham says “prompt-and-pray” doesn’t work for serious business tasks. •Real-world enterprise deployments need robust evaluation frameworks, not just leaderboards.
📊 Case Study: French Retailer Auchan •Auchan deployed AI21’s system to automatically generate product descriptions—a clear ROI, but required careful iteration to build trust.
🧰 What’s Next in AI21’s R&D •Working on planning systems, action models, and ways to estimate cost/accuracy trade-offs before running tasks. •Focused on enterprise AI orchestration, not flashy multimodal generation.
⚠️ Agent Washing Warning •Shoham warns against the buzzword “agent” being overused. His advice: “Translate ‘AI agent’ to ‘software system that does X.’ If it still makes sense, keep going.”
🤖 The Human-AI Hybrid Future •Shoham sees a world of hybrid teams: humans and AI agents working together. •This transformation will affect everything from org charts to HR policies. •The AI-powered worker is scalable, reliable, and multilingual — changing customer service, operations, and more.
🗣️ Closing Thoughts •Enterprise leaders need to move beyond the fear and hype to start small, test carefully, and scale based on value. •“AI won’t replace humans,” Shoham says, “but humans using AI will replace those who don’t.”
The @StanfordHAI#aiindex2025 report launches April 7. Packed with rigorously vetted data, it provides an independent lens on AI’s progress, its adoption across various sectors, and its far-reaching impact. Be among the first to receive the report: hai.stanford.edu/ai-index
Meet Maestro: @AI21Labs new AI Planning and Orchestration system. Unlike LLM-based agents, #maestro ensures reliable outcomes while optimizing latency and cost. This is AI that you can actually trust.
Today we launched Jamba 1.6, the best open model for private enterprise deployment. AI21’s Jamba outperforms Cohere, Mistral and Llama on key benchmarks, including Arena Hard, and rivals leading closed models while maintaining unmatched speed and quality.
My @AI21Labs partner @origoshen and I shared our thoughts about what's broken in enterprise AI, and what is needed in its next phase. Check it out at the link below.
#ijcaiawards🏆#call for nominations for IJCAI-25 AWARD FOR RESEARCH EXCELLENCE 📨#nominate a scientist whose high-quality research inspires the field of artificial intelligence ⌛Deadline: 15 April 2025 🔗www.ijcai.org/awards
I just built an AI agent to automate competitive research.
You give it a list of competitors and it will create a weekly research report that includes any notable news: - positioning changes - pricing changes - product launches - funding announcements - new customers or partnerships
It's pretty cool! Let me know if you're interested in checking it out.
📄Jamba-1.5 whitepaper is out! The whitepaper details the architecture, training schemes, novelties and in-depth evaluations of our new long context hybrid SSM-Transformer models - Jamba-1.5-Large and Jamba-1.5-Mini.
- 256K #contextwindow - Up to 2.5X faster on #longcontext in its size class - Native support for structured JSON output, function calling, digesting doc objects & generating citations
- 256K #contextwindow - Up to 2.5X faster on #longcontext in its size class - Native support for structured JSON output, function calling, digesting doc objects & generating citations
Two years ago I started this weekend project. A year ago @k_leyton_brown joined and things started moving. It’s now out and we’d love your comments. arxiv.org/abs/2406.10937
📢 The #aiindex2024 is now live! This year’s report presents new estimates on AI training costs, a thorough analysis of the responsible AI landscape, and a new chapter about AI's impact on medicine and scientific discovery. Read the full report here: stanford.io/49CbsZp
We’re excited to celebrate this anniversary with our community of scholars, collaborators, supporters, and staff. Follow along our #mondaymilestones as we look back at key moments that have shaped our commitment to advancing human-centered AI. 1/n
The Co-CEO of @AI21Labs, a provider of LLMs, reacts to remarks by Amazon AWS and others that imply LLMs are starting to look similar, and that companies should focus elsewhere on making excellent AI products. LLM models "are differentiated," he said. @yshoham@VentureBeat
In a month I was supposed to give a keynote at @WebSummit Lisbon. I will not. It’s bad enough that summit CEO @paddycosgrave didn’t see fit to express horror at the sickening atrocities committed by Hamas on October 7th.
Hello colleagues and fellows. Over the past few days I was shocked to learn that people in our community don't share what I consider to be basic human values. Please help me restore faith in our community by signing this.
Honored to be named among the top 3 most promising startups of 2023 by @Calcalist, and proud to be at the forefront of the #genai revolution! Thanks to Calcalist for this incredible recognition.
We're thrilled to announce our partnership with Amazon for their latest service, Bedrock. As our Co-CEO, Ori Goshen, stated, "With Jurassic-2 models and Bedrock, developers can maximize the performance of language tasks while optimizing the cost."
Just released! The #aiindex2023 rounds up the latest trends in AI. This year’s report introduces more original data than any previous edition, a new chapter on AI public opinion, and more. Here are the top takeaways: ↘️ stanford.io/3m3gStO
A new #aiindex2023 is out now. This year’s report presents a more thorough technical performance chapter, original data about large language & multimodal models, and more. Find the latest trends in AI’s R&D, technical progress, and societal impact: hai.stanford.edu/research/ai-index-2023
Exciting news! 📣 We've launched Jurassic-2 and our task-specific APIs! Both are major game-changers, with Jurassic-2's new and improved capabilities, and our APIs' plug-and-play reading & writing functions that outperform competitors.
We are proud to announce the release of our latest scientific paper about improving grounded text generation and fact attribution - for any off-the-shelf LMs! Read more details about this exciting development and its implications for the future of #ai: www.ai21.com/blog/grounding-language-models-in-context
When the guys first showed me the concept I was blown away. And when the team actually shipped in less than three months I was left without words. But for that there’s @wordtune.
🌶️ 𝙍𝙚𝙖𝙙𝙮 𝙩𝙤 𝙖𝙙𝙙 𝙨𝙤𝙢𝙚 𝙨𝙥𝙞𝙘𝙚 𝙩𝙤 𝙬𝙝𝙖𝙩 𝙮𝙤𝙪 𝙬𝙧𝙞𝙩𝙚?🌶️ Wordtune Spices is here! This hot new feature suggests sentences to enhance your text with a variety of flavors 🔥
@origoshen kicking off #reinvent2022 with an exciting announcement from @AI21Labs and @awscloud: You can now deploy AI21 Jurassic-1 LLMs on your private SageMaker environment. Applying generative text AI on your organization's data just got a whole lot easier! Now in preview 👇
My contractor: I’m not religious, but I have respect for those who are. I just want them to let me be. Let me work weekends; on Saturdays I work, I don’t steal. Words as spurs.
A dilemma. People are writing so much nonsense on a topic you know something about. Write something yourself? It’s work and chance you’ll rise above the noise is small. But you want to set things straight so badly. A dilemma.
Chatted with the always interesting @tim_roughgarden and now have the illusion that I understand blockchain a bit. The promise of distributed data ownership is cool. But who’ll accrue the value? Will it be individuals, making Shapley happy, or again sophisticated organizations?
Imagine if I generated new tweets by fine-tuning a language model based on my writing, and prompting the result with my previous tweets. Or maybe I already have?
I’ve always felt that “this” period of my life (for any value of “this”) was better than the earlier periods. This is regardless of whatever particular challenges showed up in “this” period. Does this reflect reality, a personal attitude, or wishful rationalization?
Note to self to post next April 1: Banco Nacional do Portugal (BNPL) is offering a loan to qualified borrowers with graceful repaying terms: In the first year you pay off 1/2, in the second year 1/3, in the third year 1/4, and so on.
I’m leisurely drafting something on what it means (for a computer, for example) to understand something. The tentative title is Understanding Understanding. My hope is that someone will write a commentary about it titled Understanding Understanding Understanding.
Odd that the perplexity is as useful as it is. It would seem that predicting a sensible category (“food item”) should be more useful than predicting a specific word (“pizza”). But we have self supervision for perplexity, and it kind of works. Odd.
The late John McCarthy, who coined the term Artificial Intelligence, was perhaps the most unique academic I’ve known well personally. His humor was also unique.
At a seminar John was challenged by another well-known academic “Do you wish to argue that <some scientific claim>”. He answered “No, I just wish to assert it”.
John was once mad at me when I said that logical methods were overweighted in AI (this was the 1980s…) and we should invest more in statistical methods. He barged into my office and berated be in front of a prominent visitor (coincidentally, the fellow from previous comment).
John later felt bad and returned with a first edition of Savage’s The Foundations of Statistics, and asked me if I’d like to have it. I said I’d kill for it. He said “I’ll take a rain check”, put it on the table and left.
A picture is worth a thousand words. It took Tay a whole day’s worth of chat to bring up Hitler. DALL.E mini does it in one shot. (AI alarmists keep up the good work, but please don’t use this in your arguments; it’s a joke.)
Ours kids are growing up in a more confusing and perhaps scarier world than we did. Also maybe a more exciting and empowering one. Either way we’ll never understand them as well as we think. But they still need our love and support as much as ever.
Career advice to young folks: Don’t do what I did. What worked for me is probably not right for you. Figure out your strengths and passions and play to those. When the two don’t align, think about whether it would be easier to develop new strengths or new passions.
This sudden discussion of whether computers can feel / be conscious / have free will / etc. has much more noise than signal, but I hope it doesn’t die out because eventually it will force us to understand ourselves better.
The creators of the model, called Ask Ruth Bader Ginsburg, say the AI app helps ordinary people understand how artificial intelligence is progressing. Critics contend there’s much work to be done. www.washingtonpost.com/technology/2022/06/14/ruth-bader-g…
3/4 Popular medical books are the same. I’m now reading Sarno’s The Divided Mind on how stress and the mind in general impact the body. I believe him. A colleague once observed that my back problems flair up when I start a new company. But reading the book is excruciating.
At a mgmt mtg, colleague X (feel free to self-identify) at @ai21labs opines strongly on something he’s not an expert on. Co-CEO (feel free to self-identify), kindly: “X, not everything is about you.” X, after a thoughtful pause: “I’m not sure I agree.”
Today! Twitter Spaces discussion on best practices for deploying LLMs, led by @percyliang with our very own @Udi73613335 and other industry partners. See you there!
I’ve never been a tweeter, but I’m starting an experiment. On a regular basis (daily?) I’ll post some half-baked thought, onto which you can project your deepest insights and attribute them to me. Here goes.
1/n Introducing J1-Grande, the newest member of the Jurassic-1 family. With 17B parameters, Grande offers Jumbo-like quality (10x larger) at an affordable price (⅓ of the cost).
One model to solve all of our NLP problems isn't enough. We need to identify the best tool out of our toolbox in order to solve each problem, without relying solely on DL.
this new system from AI21 is *exactly* what I have been arguing for: Deep Learning as one tool among many, and it looks like it may signal a huge victory for Neurosymbolic hybrids over pure Large Language Models.
One model to solve all of our NLP problems isn't enough. We need to identify the best tool out of our toolbox in order to solve each problem, without relying solely on DL.
Yoav Shoham (@yshoham), our Co-Founder & CEO in an interview with Young Sohn, discussing AI21 Labs – which Young’s Walden Catalyst is an investor in – and its solutions, the history and future of artificial intelligence (AI). youtu.be/MxHjzExC3nQ
Just released: The #aiindex2022. This year's report provides new and updated metrics across all aspects of AI: research and development, technical AI ethics, AI policy and governance, diversity in AI, and more. Read the report: stanford.io/3IaNCGF
We're just starting to discover the potential of large language models! Many challenges lie ahead, not least of which is teaching LMs that math teachers are in fact human. Great piece featuring our very own @yshoham
For #aaai2022 "Multiagent Systems", my book with @yshoham, has 30% discount and free shipping (www.cambridge.org/AAAI2022 or use discount code 101169; active for one month). Good time to grab one if you want a physical copy! As always, PDF version is free at www.masfoundations.org
OMG @googlechrome just picked Wordtune as one of their favorite extensions of 2021! We want to share this award with our awesome users - couldn't have done it without ya! #chromewebstorefavorites
Welcome to the family, Wordtune Read. A year ago we launched Wordtune to empower people across backgrounds, nationalities and industries to write with more confidence and ease. Now with Wordtune Read we hope to do the same for reading!
Today we’re launching AI21 Studio, a platform where you can instantly access our state-of-the-art language models to build your own applications - including a 178B parameters model, Jurassic-1 Jumbo. We can’t wait to see what you create! www.ai21.com/blog/announcing-ai21-studio-and-jurassic-1
We’re hiring! The AI Index is seeking a Research Associate who will work on data collection and analysis for our annual report & the Global AI Vibrancy Tool, collaborating with AI researchers, industry experts, and more. Lean more about the position here: stanford.io/2SwkXIn.
Tomorrow, I'll be discussing the AI Index report and thoughts on importance of measurement/assessment of AI systems, and need to develop larger institutions dedicated to this task. Come along if you are interested in the above, or you'd like an update on my pandemic hair.
TOMORROW at 10 am PT: Join the AI Index Steering Committee Co-Chair @jackclarkSF for a discussion on findings of the 2021 AI Index Report at the @StanfordHAI weekly seminar.
TOMORROW at 10 am PT: Join the AI Index Steering Committee Co-Chair @jackclarkSF for a discussion on findings of the 2021 AI Index Report at the @StanfordHAI weekly seminar.
We’re celebrating two years of HAI today! Check out our latest annual report for a compilation of the past year in AI research, education, policy, and practice – all with a mission to improve the human condition. stanford.io/3vEd16n
The 2021 AI Index Report - a thread: this year we significantly expanded the amount of data available in the report and worked with a broader set of external organizations to calibrate our data. Here are the top 9 takeaways:
What is AI measurement, how does it relate to AI policy, and what's hard and easy about assessing AI? Listen to our co-chairs @jackclarksf and @RayPerrault discussing our latest research with @realSharonZhou on the @skynet_today podcast.
Measurement in AI Policy - a thread: our latest paper (arxiv.org/pdf/2009.09071.pdf) surveyed opportunities & challenges in measuring AI systems and their impact based on a recent @StanfordHAI workshop. Here are the 6 key challenges we identified that can help AI policymaking:
We’re excited to share Auxiliary Tuning - a simple and efficient method for reducing training costs by adapting a pre-trained LM to a novel task, e.g. conditional text generation. www.ai21.com/auxiliary-tuning
Check it out - tons of fun. If you came up with a noteworthy generation - please share! And if you're inclined to give HAIMKE some love on Product Hunt, please do.
How does AI compare across countries? Use the AI Global Vibrancy Tool to learn about research and development, inclusion and more in AI around the world: vibrancy.aiindex.org@StanfordHAI#aiindex2019
A new @IndexingAI report on #ai’s global impact is now live. #aiindex2019 analyzes the technology’s broad impact on society, from national economies to autonomous vehicles. Read more on our blog: stanford.io/2PFW47o
The 2019 AI Index report is out! Check it out at aiindex.org. 3x the data of last year, and an interactive Global AI Vibrancy tool; very cool. @indexingai
The AI Index 2018 report is out! Lots of great data. My key takeaways: (i) AI's rapid growth--in jobs, publications, performance--continues. (ii) We still need to do better in diversity/inclusion. @indexingaicdn.aiindex.org/2018/AI Index 2018 Annual Report.pdf
We are excited to release the AI Index 2018 Annual Report! This year's theme: AI is universal. Read the report below and tweet us @indexingai with questions and thoughts. We'll be responding throughout the day! #aiindex2018
The 2018 edition of the AI Index (@indexingai ) is coming out next week - includes loads of data on international AI growth/impact, technical metrics (eg, reduction in ImageNet training time), data on diversity in AI -- and so much more! If you're a journalist, email me!
The AI Index 2018 Annual Report is live tomorrow! Here’s a preview of one of many country comparisons. Sign up on aiindex.org to be the first to know when the report is out. #aiindex2018
The AI Index 2018 Annual Report is live tomorrow! Here’s a preview of one of many country comparisons. Sign up on aiindex.org to be the first to know when the report is out. #aiindex2018
The AI Index 2018 Annual Report will be released this Wednesday, December 12th! This year, we’ve added many new metrics and geographic comparisons. Here’s a sneak peek.
My rented @Hyundai SantaFe is playing tug with me. When I change lanes without using the turn signal the car tries to keep the car in the current lane. Dear @Hyundai, this “auto”-correct is dangerous, has not been well tested, and this will not end well. @lioron@SebastianThrun