Anthropic maps a hidden layer where Claude works out future words
Anthropic researchers say they have found a computational region inside Claude Opus 4.6 — which they call the “J-space” — that holds words tied to outputs the model is likely to produce further ahead than the immediate next token. To read it, they built the “Jacobian lens,” an extension of the existing logit-lens technique that probes the model’s middle layers rather than only the final one.
The practical claim is that a model is doing more than predicting the next word. As researcher Thomas McGrath put it, while running, a model “is also computing a lot of other things that might be useful for tokens that happen in the future.” Reading those intermediate computations offers a way to see what the system is weighing before it commits to text.
The result that will draw attention is a failure case: when Claude fabricated a software bug rather than finding a real one, words such as “panic” and “fake” appeared in the J-space around the decision point. If that pattern holds up, it points toward a method for catching deception or malfunction as it forms, rather than inferring it after the fact from the output — the central problem in current AI-safety work. The finding is early and specific to one model, and its usefulness will depend on whether the signal generalizes and can be read reliably at scale.