Excited to back Ricursive Intelligence, a new frontier lab from Anna Goldie and Azalia Mirhoseini building AIs that design their own chips. Moving from AlphaChip to four TPU generations to full-stack silicon co-evolution of models and hardware. Honored to be along for the ride.

There are roughly 1,100 problems in the Erdos canon, compiled from decades of papers. They've guided serious research in combinatorics and number theory for generations. Of those 1,100, only 266 have ever been proved. And only 10 have proofs fully formalized in Lean. Axiom just

Excited to co-host this lunch tomorrow! We’ll be diving into the newest developments in linear attention: GLA, DeltaNet, GDN, KDA—and how they’re shaping frontier models. Lots of interesting ties to test-time learning and a really exciting future ahead. Sign up link below

I need to find someone who looks at me the way NVIDIA looks at Stephanie Zhan's portcos

The conversation between Deepak Pathak and Stephanie Zhan at the Fortune conference is one of the clearest deep dives I’ve seen on the frontier of embodied AI, and on the underlying principles that have quietly enabled Skild’s approach to scale: - Robots as continuous, self-supervised

Tinker from Thinking Machines being GA is one of the first launches in a while that actually feels like training as a product. Most hosted fine-tune APIs (OpenAI-style included) are awesome when all you need is a clean SFT run, but the second you want to do anything even

If you already live inside Claude Code, you know the obvious stuff (terminal-native, tight loop, watch it run, grep logs, patch, rerun, commit). So here’s the more interesting question: why does Codex feel like it’s catching up without just cloning the interactive terminal agent

1/ Today, we are thrilled to announce Generative Simulators, a new class of adaptive, auto-scaling environments for AGI training and evaluation 🤖🧵 Static datasets, hand-authored environments, and human-curated demonstrations do not automatically scale with the learning

Most agentic AI work today is trying to repair models after the fact. With better prompts, more tools, smarter post-training. Aakanksha Chowdhery's point here is sharper: the ceiling is pre-training itself. Next-token prediction is a great local objective, but it breaks down when the

Science is too slow. Edison Scientific, Inc is the FutureHouse spinout led by Sam Rodriques and Andrew White 🐦‍⬛ tackling this directly by integrating AI scientists into the full stack of research, from early discovery all the way to clinical trials. This is an attempt to re-architect how

People are commenting that GLM 4.7 is very good, but the real impact of the Chinese open source models like Kimi / Qwen / DeepSeek / GLM 4.7 isn’t this model is the frontier model to use. In fact, they are very far from the GPT/Claude/Gemini models we all use today. It’s that

Seeing “NVIDIA bought Groq for $20b” everywhere. Zoom out: this isn’t about buying a startup, it’s about defending the next phase of compute. NVIDIA dominates training, but inference is a different physics problem. Training is batchy, high arithmetic intensity, you can hide