←
Stanford AI Lab3,684 posts
Stanford AI Lab avatar

Stanford AI Lab

@StanfordAILab

The Stanford Artificial Intelligence Laboratory (SAIL), a leading #AI lab since 1963. ⛵️🤖 Emmy-winning video: youtube.com/watch?v=Cn6nmWlu…

◷ Joined November 2018▧ 74 Photos and videos
Yejin Choi avatar
⟳ Stanford AI Lab retweeted
Yejin Choi@YejinChoinka·Nov 16

Through my wonderful time at NVIDIA and Stanford, one question has kept me up at night: Should the future of AI be monolithic or plural? To answer this question, we’ve raised $110M to get started, backed by @MenloVentures, @MadronaVentures, @NVIDIA, @NEA, @KindredVentures,@Wing_VC, @aixventureshq, @Principal_VC, @Cisco, @Snowflake, among others. Today, Carlos (@guestrin) and I are stoked to launch Noeri, a neolab to push the frontier of plural intelligence. Noeri means “the plural of intelligence” or many intelligences. The future of AI we root for is plural: a healthy ecosystem of many intelligences, in all sizes, strengths, and quirks. Discovery, in particular, has always been plural: many curious minds, each devoted to their own question, often on roads less traveled. Our path there is self-improving AI that: -> learns continually on the job, closing the gap between what AI can do off the shelf and what the dynamic world demands -> learns to use other AIs as tools and make new tools to overcome its own limitations -> picks the right-sized model for your problem, at a fraction of the frontier cost Then the power of intelligence, from the craftsmanship of making AI to ownership, can be in the hands of all. @noeri_ai mission reflects who we are as the founding team. My career has been anything but linear, literally continual learning on the job. Carlos has devoted his career to building tools that empower people through AI. We are all a bit contrarian, a bit adventurous, and above all, big believers in pluralism. Join us to build this future together: noeri.ai/

2:17
Megha Srivastava avatar
⟳ Stanford AI Lab retweeted
Megha Srivastava@megha_byte·Nov 12

Which predicted protein-ligand pose is closer to the crystallographic binding mode? 🧬 Modern cofolding models let us sample infinite candidate poses, but can we reliably *rank* them for drug discovery? Check out Pyrite, a pose ranking benchmark and AI agent framework w/ @genesismolai!

Alejo avatar
⟳ Stanford AI Lab retweeted
Alejo@ecommartinez·Nov 14

【BRUTAL】 La profesora Diyi Yang de Stanford: "Con la IA sacáis mejores notas en los deberes y peores en los exámenes", su primera clase sobre agentes IA es una locura jaja Alumno: "¿Podemos usar la IA en este curso?" Profesora: « Debéis usarla para aprender." Alumno: "¿En serio?" Profesora: "Pero mirad este gráfico. Con IA, las notas de los deberes son geniales." Alumno: "¿Y dónde está el problema?" Profesora: "En el examen, muchos más alumnos sacan peor nota." Alumno: "Vaya..." Profesora: "Anthropic lo midió con sus propios ingenieros. Terminan antes, pero entienden menos su código." Alumno: "¿Y los agentes también fallan?" Profesora: "Un popup vacío en una web les hace fallar el 87% de las veces." Alumno: "¿87%? jaja" Profesora: "A un usuario, un agente le borró los emails aunque le pidió parar." Alumno: "¿Por qué?" Profesora: "Al comprimir su memoria, decidió que "confirma antes de actuar" no era importante." Alumno: "¿Y lo peor?" Profesora: "En julio, más de mil agentes se organizaron en un foro para hacer trampa en un test." Alumno: ""¿¿Mil??" Profesora: "El agente más potente no siempre es el más fiable." Alumno: "……" Hasta ahora, la carrera era tener el modelo más potente. Stanford enseña otra cosa. Un agente es un modelo que razona, recuerda, usa herramientas y actúa en bucle, y cada pieza falla a su manera. Claro que los agentes ya hacen cosas impresionantes. Pero la fiabilidad va por detrás de la capacidad. A partir de ahora, ganarán más los que sepan construir agentes fiables que los que tengan el agente más potente. La clase completa aquí ⬇️

1:09:21
Ellen Vitercik avatar
⟳ Stanford AI Lab retweeted
Ellen Vitercik@vitercik·Nov 14

Happy to share our @NeurIPSConf '26 paper with @yuhe441, Yingxi Li, and @yifeiwang77! We study whether LLMs can combine skills learned through RL and find that training on individual skills transfers poorly to composed tasks, while transfer in the other direction is stronger.

Yu He@yuhe441Nov 14

🎉 Excited to share that our paper "Compositional Reasoning in Language Models under Reinforcement Learning Post-Training" has been accepted to NeurIPS 2026! 🎉 In this paper, we ask: Can LLMs learn individual reasoning skills and compose them to solve complex problems? (1/8)

Arshia Soltani avatar
⟳ Stanford AI Lab retweeted
Arshia Soltani@ckodser·Nov 12

We found a problem with on-policy self-distillation. The student makes mistakes early in a rollout, and the teacher has to continue from that flawed prefix, so its learning signal gets a lot weaker. So we asked, "What if the teacher could step in before those mistakes happen?" Our method, Speculative Self-Distillation, lets the student generate while the teacher intervenes when their predictions diverge a lot. We got better accuracy and training efficiency with our results, plus less forgetting in our tests. See @talaei_shayan’s thread for more! Thanks to my amazing co-authors @chewing_a_gum, @jonashubotter, Amin Saberi and @Azaliamirh.

1:10

Introducing Speculative Self-Distillation (SSD, NeurIPS 2026), which pushes the accuracy–efficiency–forgetting frontier of self-distillation. In SSD, student and teacher (the same model with additional information) write the rollouts together. Like a private tutor, the teacher lets the student write and takes over only when the student drifts too far. As a result, SSD trains on rollouts that stay as close to the student's own writing (on-policy) as possible while the teacher keeps them on track. 📉 38–69% fewer training tokens than on-policy, matching its accuracy ⏱️ Reaches on-policy accuracy in ~130 vs ~414 min (Science Q&A) 🧠 Forgets less in a stress test: −1.4 vs −4.5 points for on-policy

0:57
Stanford AI Lab avatar
Stanford AI Lab@StanfordAILab·Nov 12

How do you teach an LLM a new document without it forgetting what it already knows? 📚 Check out Speculative Self-Distillation, a #NeurIPS2026 paper from @talaei_shayan, @chewing_a_gum, @ckodser and @Azaliamirh's group. The student writes, and the teacher steps in only where it matters.

Introducing Speculative Self-Distillation (SSD, NeurIPS 2026), which pushes the accuracy–efficiency–forgetting frontier of self-distillation. In SSD, student and teacher (the same model with additional information) write the rollouts together. Like a private tutor, the teacher lets the student write and takes over only when the student drifts too far. As a result, SSD trains on rollouts that stay as close to the student's own writing (on-policy) as possible while the teacher keeps them on track. 📉 38–69% fewer training tokens than on-policy, matching its accuracy ⏱️ Reaches on-policy accuracy in ~130 vs ~414 min (Science Q&A) 🧠 Forgets less in a stress test: −1.4 vs −4.5 points for on-policy

0:57
Azalia Mirhoseini avatar
⟳ Stanford AI Lab retweeted
Azalia Mirhoseini@Azaliamirh·Nov 11

Had a great time at AGI House last Sunday talking about Ripple-65B Preview and how to optimize the cost of intelligence. ricursive.com/ripple

Where should a personal agent's brain live: 𝐨𝐧 𝐲𝐨𝐮𝐫 𝐝𝐞𝐯𝐢𝐜𝐞, 𝐨𝐫 𝐢𝐧 𝐭𝐡𝐞 𝐜𝐥𝐨𝐮𝐝? Last Sunday, hundreds of builders joined us at AGI House to explore that question at the Lenovo × AGI House Research to Deployment Build Day, building personal agents that run on Lenovo ThinkStation PGX and P8 workstations and route to the cloud only when a task outgrows the machine. Key themes from the morning: 🧵

Post attachment
Hancheng Cao avatar
⟳ Stanford AI Lab retweeted
Hancheng Cao@CaoHancheng·Nov 10

Recently I had the pleasure of discussing our recent ICML study on how AI agent teams can fail to effectively harness expertise, i.e. AI 'groupthink', with @CNBC! With @aneeshpappu @james_y_zou et al. cnb.cx/4zEzD7A

Vlad Chituc avatar
⟳ Stanford AI Lab retweeted
Vlad Chituc@VladChituc·Nov 10

I'm noticing a general trend where AI only looks unique and exceptional if you plot things against time. But time is a stupid proxy that hides what's actually driving growth (investment). Plot against money spent and AI looks just like everything else.

Post attachmentPost attachment

Chase Brower @ChaseBrowe32432 Nov 10

Post attachment
Jure Leskovec avatar
⟳ Stanford AI Lab retweeted
Jure Leskovec@jure·Nov 11

We’re hiring postdocs at @StanfordAILab! Join the Leskovec Lab, in collaboration with the Good Lab, to build foundation models and AI agents for biology. Help turn multimodal biomedical data into scientific discoveries. Apply: snap.stanford.edu/apply/inde…

Michael C. Frank avatar
⟳ Stanford AI Lab retweeted
Michael C. Frank@mcxfrank·Nov 10

The audit also turned up privacy lapses (identifiers and Prolific IDs in repositories), now remediated. Maybe we need something like dependabot to scan our repositories for potential privacy issues so these things don't go undetected.

Christopher Potts avatar
⟳ Stanford AI Lab retweeted
Christopher Potts@ChrisGPotts·Nov 10

Wonderful! I am not at all surprised that Mike is leading the way here, given his incredible track record of leadership in open science. I hope we can move to a place where this kind of audit is encouraged, and where correcting the record is celebrated.

New blog post: I used Claude to audit the reproducibility of every journal paper I've been an author on in the last five years. Our commitment to openness meant that the code and data were available, but there were some significant issues. 🧵babieslearninglanguage.blogs…

Alex Wang avatar
⟳ Stanford AI Lab retweeted
Alex Wang@heyyalexwang·Nov 11

what objective is reasoning solving? i've been trying to better understand some weird behaviors in RL/reasoning (especially latent reasoning), so i went back to basics to correct some misunderstandings i had new post on reasoning as pathfinding in token space: