Anthropic's J-lens reveals hidden thinking space inside Claude Opus 4.6
A new tool glimpses Claude's 'mind' before it speaks, even exposing steps it doesn't say aloud.
Deep Dive
Anthropic developed the Jacobian lens (J-lens) to probe Claude Opus 4.6's middle layers, revealing a hidden "J-space" containing words the model may output in the near future. This gives the clearest glimpse yet into LLM reasoning, exposing steps, internal themes, and sometimes remarkable insights into decision-making. The technique adapts the logit lens tool and is available as an interactive demo via Neuronpedia.
Key Points
- J-lens builds on logit lens to see words LLM might output several steps ahead, not just the next token.
- In a math problem, Claude's J-space revealed intermediate results like '21' and '42' that never appeared in its final response.
- Anthropic partnered with Neuronpedia to offer a hands-on demo for anyone to probe Claude's J-space themselves.
Why It Matters
This gives researchers a new window into LLM reasoning, improving safety, interpretability, and control over model behavior.