[Video] Jacobian Lens: Can we read the mind of an AI?

October 26, 2026

Can we tell if an AI is plotting against us? Or if it’s being sycophantic?

Anthropic released a new mechanistic interpretability tool called the Jacobian lens, or J-lens. And what are J-lenses?

References

• Why Models Know But Don’t Say: Chain-of-Thought Faithfulness Divergence Between Thinking Tokens and Answers in Open-Weight Reasoning Models (arXiv 2603.26410)
• Astra can do a concerning amount with no chain of thought (LessWrong)
• A global workspace in language models (Anthropic)
• Verbalizable Representations Form a Global Workspace in Language Models (Transformer Circuits)
• What language do language models speak? (Zoltan’s Blog)
• The most cited paper of the century is a brilliant hack (Welch Labs, YouTube)
• Eliciting Latent Predictions from Transformers with the Tuned Lens (arXiv 2303.08112)
• anthropics/jacobian-lens: jlens/fitting.py (GitHub)
• neuronpedia/jacobian-lens (Hugging Face)
• Baron-Cohen, Leslie & Frith: Does the autistic child have a “theory of mind”? (Cognition 21, 1985)
• A Review of Anthropic’s Global Workspace Paper (Neel Nanda, LessWrong)
• The J-lens offset is the model’s token frequency: z-scoring helps (Ameya Panchal, LessWrong)