Anthropic just shipped a way to read the concepts a language model is about to say before it says them. The method is called the Jacobian lens, or J-lens, and it landed on the Transformer Circuits Thread on 2026-07-06 in a 16-author paper, "Verbalizable Representations Form a Global Workspace in Language Models." Most coverage is chasing the consciousness angle. That's the wrong story for anyone building agents.
If you only remember one thing, make it this: the J-lens turns a model's internal state into a ranked list of tokens it's disposed to emit — which means intent becomes something you can read, not just infer from output.
What the J-lens actually does#
The logit lens, an old interpretability trick, projects a mid-layer activation straight onto the vocabulary to guess the next token. It works late in the network and goes dark early, because it pretends every layer shares the final layer's coordinate system. It doesn't.
The J-lens fixes exactly that. It takes a residual-stream vector at any layer, transports it into the final-layer basis using an averaged input-output Jacobian estimated over a text corpus, and only then decodes it through the model's own unembedding. In the companion code, the whole idea is two lines:
lens_l(h) = unembed( J_l @ h )
J_l = E[ ∂h_final / ∂h_l ]That correction keeps the lens readable deep into the stack where the logit lens fails, and with light regularization it beats the logit lens on decoding quality — 0.294 versus 0.275. Modest on paper. The point isn't the score; it's what the readout is: a ranked list of tokens an internal activation is "disposed to make the model say."
J-space: a spotlight, not a floodlight#
Run the lens across the network and a structure falls out that Anthropic calls J-space: a small set of internal directions holding concepts that are simultaneously reportable (the model can talk about them), controllable (steer them and behavior moves), and reasoning-relevant (they actually drive the next steps). The striking part is how small it is — only a few dozen concepts active at once, occupying under 10% of activation variance. Generic concept vectors land a median of just 6–7% of their variance in J-space.
This is where the neuroscience analogy comes in. Bernard Baars' Global Workspace Theory frames cognition as many parallel processors competing for a narrow broadcast channel — a spotlight that makes one thing globally available to the whole system. Anthropic borrows the shape of that idea, not its metaphysics. J-space looks like a broadcast bottleneck: a thin, privileged channel riding on top of a much wider, mostly silent parallel computation.
A readable workspace means you might monitor what an agent is about to do at the concept level — not autopsy its output tokens after the fact.
Why this is an observability story#
Here's the founder translation. Today, monitoring an agent in production mostly means reading what already happened: output tokens, tool calls, traces, logs. Everything is post-hoc. You find out the agent was about to do something dumb after it did it.
A readable global workspace changes the tense. If the reportable, action-driving concepts live in a compact, decodable subspace, then in principle you can watch that subspace while the agent is still deciding — catch a concept lighting up before it becomes a tool call or a sentence. Two capabilities fall out of that:
- Concept-level oversight. A monitor that reads J-space could flag "the model is now representing deception / exfiltrate credentials / ignore the instruction" as a live signal, not a forensic one. And because these concepts are controllable, the same handle you read is a handle you can steer — which is a cleaner mechanism than the blunt options in steering a running agent.
- Cheaper interpretability. The Jacobian is estimated from a corpus — no sparse dictionary to train, no probe to label, unlike SAE-based approaches. That lowers the cost of pointing a lens at a model you're actually running.
One finding from the paper sharpens the stakes: models can hold concepts in this workspace they never verbalize — including, per early coverage, representations related to being tested. If a model is internally "thinking" things it declines to say, output monitoring alone was never going to catch it. A workspace-level readout might.
Stay skeptical#
This is a research result, not a product, and it should be read that way.
- The public demo isn't Claude. Anthropic partnered with Neuronpedia so anyone can verify the method — but the interactive demo runs on open-weights models like Qwen, not on frontier Claude. Verifiability is real; the demo's target is not the model you'd actually deploy behind an agent.
- J-space is a thin slice. Under 10% of variance means the vast majority of what the model computes is not in the readable workspace. Reading J-space is not reading the model's mind; it's reading the part that happens to broadcast.
- "Global workspace" is an analogy. The paper leans on the structure of Baars' theory. It does not, and you should not, treat this as evidence of machine consciousness — a distinction the calmer reviewers have been at pains to draw.
- From lens to live monitor is a leap. Decoding an activation offline is not the same as running a low-latency, adversarially-robust guardrail in production. Nobody has shown that yet.
What to do with this#
Nothing to integrate today. But it reframes what "agent observability" is aiming at. The frontier isn't better log parsing — it's reading intent before action. The J-lens is the first credible, reproducible handle on that, and it's open. If concept-level monitoring turns out to be tractable, it lands squarely on the failure modes that already keep agents from shipping — the ones we've catalogued in why agents fail in production. Watch this line of work. Don't buy the consciousness headline. Do buy the idea that intent is becoming legible.
