New Anthropic Research Can Read Claude's Internal Thoughts
AnthropicAI · x · 2026-07-07
Anthropic published a paper introducing an interpretability method called "J-space" (derived from the Jacobian matrix). Existing within the model's internal neural activations, J-space is distinct from Claude's output text or chain-of-thought, enabling the model to "think" about concepts without explicitly writing them out.
By observing J-space, researchers can read, audit, and shape what Claude is currently thinking. This provides tools to maintain trustworthiness as model capabilities improve, and also hints at striking similarities between language models and the human mind.
Related event: Anthropic Discovers Global Workspace Inside Claude(102 posts)→
More from Research
- Token Collapse: Study Finds Graph Soft Prompting Tokens in LLMs Fail to Encode Structure — Sapienza · 2026-10-27
- Terence Tao's SAIR Foundation revamps Mathathon with $1.5M compute per team — _sathvikr · 2026-10-27
- Mathathon 2.0 launches with Terence Tao's SAIR, $1.5M in compute for teams — _sathvikr · 2026-10-27
- The 'sweet lesson' is real: two kinds of ML problems need opposite strategies — mike64_t · 2026-10-27
- PixGR: Pixel-space generative decoding for fast feed-forward 3D reconstruction — kwangmoo_yi · 2026-10-27
- UniAction: unified action conditioning lets one world model handle mixed action types — ZeYanjie · 2026-10-27