Anthropic Ships Jacobian Lens as Unmaintained Research Code

Anthropic released Jacobian Lens on GitHub July 2, code that decodes what a model's activations are nudging it to say, per the repo.

Interpretability researchers use tools like this to check whether a model's internal state matches what it later claims to be thinking. That's a live question for anyone building safety tools on top of a model's self-reports.

The method: average the Jacobian, the derivative of the model's final hidden state with respect to an earlier layer, across a corpus. That matrix becomes a linear map: feed it any layer's activation and it projects into the model's output vocabulary, skipping every layer in between.

It's built for open-weights decoder transformers. The formula is lens_l(h) = unembed(J_l @ h). The examples run on Qwen, not Claude, and the repo renders interactive layer-by-position visualizations of what an activation would say at each depth.

The release sits alongside a new paper, "Verbalizable Representations Form a Global Workspace in Language Models," on transformer-circuits.pub. Its claim: models route information through a shared internal format specific enough to decode this way. Jacobian Lens is the demonstration.

Anthropic marks the project explicitly as unmaintained, a reference implementation not accepting contributions, per the repo. That's an unusual thing to flag on a repo that already has 1,394 stars.

I read that label as the real news. Interpretability code built to prove one paper's point isn't the same thing as a library meant to run against your own model in production. Anthropic is telling you which one this is, before someone files an issue and waits for a merge that won't come.

Worth watching whether anyone reproduces the global workspace result on a model besides Qwen, since the repo doesn't include Claude itself.