"Move 37" is the word-of-day – it's when an AI, trained via the trial-and-error process of reinforcement learning, discovers actions that are new, surprising, and secretly brilliant even to expert humans. It is a magical, just slightly unnerving, emergent phenomenon only
RESEARCH
-

AI Supercharges Human Agency, Creativity and Progress
By
–
“What could go right?” We need to be both pessimists and optimists about the future of AI, but often pessimism overshadows everything else. Reid’s latest book makes the case for how AI will superpower human agency, creativity, and progress. A brilliant and practical vision.
-

DeepRoles Release Date and Capabilities
By
–
When DeepRoles (agents from DeepSeek) will be released and what they will be able to do?
-
Core Technology Remains Fundamentally Matrix Multiplication Based
By
–
Not really — the core is still just big matmuls
-
RL vs SL: Demystifying Reinforcement Learning’s Core Concept
By
–
Yeah exactly. I get triggered when RL is dressed up in its full rigorous math formalism because it's gate-keeping an essentially trivial core idea. SL:
a token sequence comes from some 3rd party source (e.g. human demonstration), and you just train on it. RL:
you first sample a -
Consciousness vs Tokens: Understanding AI Model Awareness
By
–
It’s only profound to you, because you are conscious and self-aware and it therefore means something. To the model, it’s just a bunch of tokens..
-
Stanford AI Model Combines Clinical Notes and Images for Personalized Cancer Treatment
By
–
Multimodal approaches represent the future of AI in medicine. A new AI model developed by Stanford researchers considers clinical notes and images to help doctors personalize cancer treatments. @StanfordCancer
-
Plasma Simulation and Generative AI: Career Paths in Fusion Research
By
–
Ironically I shifted to fluid dynamics in 2022 (plasma simulation for fusion power) so we moved in opposite directions. Thanks for GigaGAN!
-
Sparse vs Dense Models: Pre-training and Inference Trade-offs
By
–
Its advantages, George, are best understood for pre-training. How it performs in inference remains a less well-understood area. And there may be trade-offs there where "dense" models do better.
-
DeepSeek Podcast: Exploring AI’s Current State and Future
By
–
I'm doing a podcast on DeepSeek and the state of AI soon. Let me know if you have questions / topic suggestions.