Anthropic Research: A global workspace in language models

Anthropic published new interpretability research: a global workspace in language models.

The framing borrows one of neuroscience's most influential ideas. Of everything happening in your brain right now, only a tiny fraction is consciously accessible—the "global workspace" that broadcasts a handful of signals to the rest of the system. Anthropic reports finding a strikingly similar divide inside Claude: a huge amount of internal computation, and a narrow channel of it that actually reaches what the model says.

#The practical spin-off: the Jacobian lens

The paper didn't stay theoretical for long. Anthropic open-sourced the "Jacobian lens"—a readout of what a model is about to say, extracted from its internal layers before it says it. It's white-box (you need the activations), but that's exactly what makes it interesting for anyone running open-weight models or building monitoring tooling.

The community moved immediately: @EricBuess wired it into an agent hook within days of release—a preview of what "know what your agent is about to do" tooling could look like.

#Why it matters

This lands in the same window as Anthropic's turn-averaged sparse autoencoders work—both pushing interpretability from per-token curiosities toward practical, scale-ready monitoring of high-level model behavior. A readout of intent-before-output is precisely the kind of primitive that agent-security researchers (who spent this month arguing agents should be treated as untrusted) have been asking for.

Sources: Announcement, @EricBuess: Jacobian lens with Qwen