For years, neural networks looked inscrutable, like tangled knots that needed to be unwound. We no longer think that's the case. AI models are far more interpretable than we thought. We think this is great news for interpretability.
Goodfire's @nickcammarata co-authored Zoom In, the paper that introduced polysemanticity and superposition. He spent months revisiting the vision model organisms that much of interpretability grew out of and studying modern language models up to trillions of parameters. His new essay explains what the early field, him included, got wrong, why there's reason for optimism today, and why there might be a bitter lesson for interpretability.
Must-read papers of the week
▪️ RoboJEPA
▪️ ARC: A Reasoning Recipe for Robot Foundation Models
▪️ Memento 3: Model-Based Recursive Self-Improvement through Reflective Rulebooks
▪️ Self-Retrospection Distillation
▪️ RoboQuest: Generalist Physical Agents that Search, Inspect and Test
▪️ Towards Looped Models Done Right, Part II: Rethinking at Fixed Points
▪️ Long-WAM: Scaling the Context of World-Action Models
▪️ Optimizing the Optimizer: Language Models Discover Faster Molecular Relaxation Algorithms
▪️ REMORY: Learning Residual Memory for Context Compaction
▪️ EVISKILL: Grounding Skill Evolution in Replayable Evidence
▪️ AgentGarten: Code Worlds for Evolving Agents
▪️ LoGRA: Scaling LLM Reinforcement Learning with Low-Rank Gradient Sketches
▪️ RealtimeWAM: One-Step Asynchronous World Action Models
Explore them to keep up with main AI trends. Here’s also the full list of stunning papers + links and our weekly AI news digest: https://t.co/h0nYIGoK0o
New models to know this week
▪️ RoboJEPA – latent world models trained across 12 robotic embodiments, with prediction accuracy linked to planning performance.
▪️ Long-WAM – world-action model that uses longer visual histories to improve robot control.
▪️ VioLA – generalist humanoid control model that learns body and hand movements mostly from human recordings.
▪️ UNITAS – 3D-native world-action model that represents observations, actions, and scene changes in a shared 3D space.
▪️ DreamTrue – robot world model trained to better predict when an action will go wrong.
▪️ Kandinsky 6.0 Video – open video generation models with synchronized audio, speech, and lip sync.
▪️ MiMo-V2.6 – omni-modal model family exploring how to scale reinforcement learning toward self-improvement.