LLMs are injective and invertible. In our new paper, we show that different prompts always map to different embeddings, and this property can be used to recover input tokens from individual embeddings in latent space. (1/6)

Sharing our work at NeurIPS Conference on reasoning with EBMs! We learn an EBM over simple subproblems and combine EBMs at test-time to solve complex reasoning problems (3-SAT, graph coloring, crosswords). Generalizes well to complex 3-SAT / graph coloring/ N-queens problems.

Can AI automate jobs? We created the Remote Labor Index to test AI’s ability to automate hundreds of long, real-world, economically valuable projects from remote work platforms. While AIs are smart, they are not yet that useful: the current automation rate is less than 3%.

Year 2 of the Amazon Nova AI Challenge is here, focused on trusted software agents. 10 university teams will advance agentic AI for software eng, balancing capability & safety - some will build defenses, others will probe for weaknesses. Apps open Nov 10! amazon.science/nova-ai-challe…

Last year, I led an unlearning effort at Amazon with the Nova Responsible AI and Pretraining teams, focusing on controlling model knowledge and behavior. We hosted an LLM unlearning challenge at Semeval and developed LUME unlearning bechmark a multitask optimization method, and a

Cool new work on progressively growing a dynamic vocabulary that merges tokens using LZW compression. arxiv.org/abs/2506.01084

Introducing GEN-0, our latest 10B+ foundation model for robots ⏱️ built on Harmonic Reasoning, new architecture that can think & act seamlessly 📈 strong scaling laws: more pretraining & model size = better 🌍 unprecedented corpus of 270,000+ hrs of dexterous data Read more 👇

Tomorrow, we are excited to welcome Weiyan Shi to the Stanford NLP Seminar! Date and Time: Thursday, November 6, 11:00AM — 12:00 PM Pacific Time. Zoom Link: stanford.zoom.us/j/93941842999?… Title: Beyond the Surface: How Post-Training Artifacts Shape LLM Diversity and Safety

How do you make LLMs safer and more aligned with human values? A challenge is understanding the hidden reward signals they learn.Our new paper introduces Failure-Aware Inverse RL, a method to uncover these signals by focusing on what LLMs get wrong. Paper: arxiv.org/abs/2510.06092

Today, we are launching the most rigorous ongoing source of expert forecasts on the future of AI: the Longitudinal Expert AI Panel (LEAP). We’ve assembled a panel of 339 top experts across computer science, AI industry, economics, and AI policy. Roughly every month—for the next

I'm teaching a new "Intro to Modern AI" course at CMU this Spring: modernaicourse.org. It's an early-undergrad course on how to build a chatbot from scratch (well, from PyTorch). The course name has bothered some people – "AI" usually means something much broader in academic

RL is bounded by finite data😣? Introducing RLVE: RL with Adaptive Verifiable Environments We scale RL with data procedurally generated from 400 envs dynamically adapting to the trained model 💡find supervision signals right at the LM capability frontier + scale them 🔗in🧵

Announcing a private AI bug bounty program to strengthen the Amazon Nova foundation models. Building on 30+ findings and $55,000+ in rewards from the public program, the new track partners with security researchers and academics to strengthen AI security. amazon.science/news/amazon-la…

The Anthropic perspective on interpretability is prominent and significant, but not inevitable. My own take is quite different. (Clip from a talk I gave; YouTube link in the thread):

The value of fast iteration in AI is overrated. The best results are obtained by knowing the right things to do and doing each thing with neurotic precision and attention to detail.

Expanding hidden states without increasing block dimensions. Classical line of approaches from the era when everyone tried to make variants of skip connections, but it could be worth trying.

cool paper introducing better steering method tl;dr instead of using a fixed steering coefficient, optimize s.t. we get max steering while staying within distribution arxiv.org/abs/2510.13285

✨ New course materials: Interpretability of LLMs✨ This semester I'm teaching an active-learning grad course at Tel Aviv University on LLM interpretability, co-developed with my student Daniela Gottesman. We're releasing the materials as we go, so they can serve as a resource for anyone

Super excited about this cyber-sec challenge in collaboration with amazing teams!