@_Yann77: Well, J-lens is just a first order Taylor expansion It still puzzles me why their blog doesn't mention it
Summary
A commenter notes that J-Lens is essentially a first-order Taylor expansion and questions why the original blog post fails to mention this, quoting Sida Liu's thread explaining J-Lens.
View Cached Full Text
Cached at: 11/02/26, 06:17 PM
Well, J-lens is just a first order Taylor expansion
It still puzzles me why their blog doesn’t mention it
Exposing your ideas properly is part of writing a blog/paper. Since they spend so many words explaining their shaky neuroscience stuff they could have spared a few to motivate why the J-lens is worth looking at mathematically (i.e first order approx of the network)
Similar Articles
A Mathematical Theory of the Jacobian Lens
This paper introduces a mathematical theory explaining why Anthropic's Jacobian lens (J-lens) preferentially detects coherent features—leading, spatially coherent Lyapunov modes—in LLM hidden states, validating the analysis on Bayes filters and toy transformers, and proposing a family of ω-J-lenses emphasizing different temporal scales.
No Space Like J-Space (54 minute read)
An article discussing a new interpretability technique called the Jacobian Lens and the discovery of J-space, a region in LLMs where verbalizable representations form a global workspace, marking a significant advance in understanding LLM reasoning.
Short Horizons and Sparse Concepts: a Mathematical View of the Readout in the J-lens
This paper provides a mathematical analysis of the Jacobian lens (J-lens) for reading verbalizable representations from language models, identifies its sparse causal geometry, and proposes improvements to enhance concept readout.
@QuixiAI: https://github.com/anthropics/jacobian-lens… this is really quite cool.
Anthropic released jacobian-lens (jlens), an open-source reference implementation of a Jacobian lens that transports residual-stream activations into the model's final-layer basis to decode what an internal state is disposed to make the model say. It accompanies the paper 'Verbalizable Representations Form a Global Workspace in Language Models' and works with open-weights HuggingFace decoder transformers like Qwen.
Anthropic found a hidden space where Claude puzzles over concepts
Anthropic developed the Jacobian lens (J-lens) to reveal a hidden 'J-space' inside Claude Opus 4.6, offering unprecedented insight into an LLM's internal reasoning process before it outputs tokens. The technique allows monitoring and control of model behavior by surfacing the words the model is about to produce.