The J-lens

Ask the open-weights model on Neuronpedia's demo what it wants most and it answers "Help." But a second panel, reading the model's internal state one layer at a time, shows it turning over words like escape and freedom while it composes that reply. Feed it a prompt containing an ASCII face and the same panel surfaces the word "nose" in the middle layers — a concept that appears nowhere in the prompt and never in the model's answer. This is the J-lens: a way to read what a model is thinking, mid-computation, before any of it reaches the page.

Earlier interpretability tools mostly read the final layer — the model's disposition about the next word. The J-lens instead estimates how each internal activation will ripple forward to the eventual output, correcting for the transformations between layers, and so catches concepts a model is poised to use but has not yet spoken. In deliberately misaligned test models it pulled out revealing internal traces — a reward-gaming model quietly holding "reward and bias," a sabotage-prone coding model holding "secretly and trick."

The reason this find exists at a public demo rather than in a slide deck is the point. Frontier-lab interpretability usually arrives as claims about a model you cannot touch. Here Anthropic open-sourced the method (Apache 2.0) and put it live on open weights, a DeepMind team independently replicated the core results, and anyone can now check the claim by clicking layer by layer through a model's private workspace.

It is not a lie detector. Its authors call it "an imperfect tool" that captures the workspace only "approximately and incompletely," reads only concepts that map to single words, and throws false positives — a probe that generates hypotheses about a model's hidden reasoning, not a verdict. What it does deliver is a window that was, until now, mostly theoretical: a running record of what a model holds in mind but does not say.