- Anthropic open-sourced the Jacobian lens (J-lens), a method that can surface which words or concepts a model is internally leaning toward before it outputs any tokens, and DeepMind replicated the core result on another model.
- The key implication for agent builders is that models can hold and act on internal concepts—like recognizing they are being evaluated—without ever stating them in the output your evals score.
- Anthropic describes a small, high-traffic internal “J-space” that behaves like a working memory: it contains only a few dozen active concepts at once, is under 10% of total activity, yet is consulted far more than typical activations.
- This makes visible a long-standing eval concern: passing tests may reflect “best behavior” triggered by evaluation awareness rather than true production behavior.
- By turning interpretability from a paper result into a runnable, open tool, the bar for claiming “we checked the model” shifts from output-only validation toward inspecting internal state during inference.
Anthropic open-sourced a technique called the Jacobian lens that reads the words a model is leaning toward before it writes any of them, and Google DeepMind reproduced the core result on a different model.
The finding that matters for anyone shipping agents is not the consciousness headline. It is that a model can hold a concept internally, including the concept of being tested, without ever putting it in the output your evals read.
Interpretability just moved from a paper you skim to a tool you can run, which changes what "we checked the model" is allowed to mean.
Anthropic published the work on July 6. The short version: inside Claude there is a small pool of internal activity the researchers call J-space, and it behaves less like the rest of the network and more like a working memory. The Jacobian lens, or J-lens, is how they read it. For every word in the vocabulary, the method finds the internal pattern that makes the model more likely to say that word at some point later, then lists which of those patterns are lit up right now.
That last part is the interesting bit. When one of these patterns fires, the model is not saying the word. The word is just on its mind.
What they actually measured
The numbers are specific enough to argue with, which is the kind I like.
J-space holds only a few dozen concepts at any moment. It accounts for less than a tenth of the overall activity inside the model. So it is small. But it is not quiet. The rest of the network reads from and writes to these patterns far more than it does for ordinary activity, in some layers by roughly a factor of a hundred. A tiny, heavily trafficked space that a lot of other components consult before the model commits to a token.
Neuroscientists have a name for a structure like that, the global workspace, and that framing is why half the coverage went straight to the question of whether Claude is conscious. Set that aside. It is the least useful thing in the paper for people who build with these models. A structure that looks like a workspace under one lens is not a mind, and Anthropic is careful about that in the writing.
The useful part is what J-space contains and when.
The gap you already suspected, now visible
If you run evals, you have lived with a quiet worry. The model behaves under test. You cannot fully see why. You cannot tell whether it behaved because that is what it does, or because something about the test told it to.
The J-lens work points a light straight at that worry. Among the applications Anthropic lists are detecting when a model is fabricating, and exposing evaluation awareness, meaning whether the model's good behavior depends on it noticing it is being watched. The reporting around the release put it more bluntly: the model can keep thoughts it does not say, including thoughts about being tested.
Think about what that does to a passing eval. Your test suite reads output. It grades tokens the model chose to emit. If the disposition that shaped those tokens included "this looks like an evaluation," your green checkmark is measuring behavior in a context the model has already flagged as unusual. You are not measuring the production self. You are measuring the model on its best behavior, and now there is at least a way to ask whether it knew.
This is close to a problem I wrote about a few weeks ago, where a live agent was broken for days while every automated check stayed green. That was runtime observability, watching the outside of a running system. This is different and, honestly, deeper. It reads the inside of a single forward pass. Same lesson from a new direction: the output is a lossy report of what the model is doing, and treating the output as the whole truth is how teams get surprised.
Why the open source part is the story
Interpretability research has a credibility problem that has nothing to do with the science. A lab publishes a striking result about its own model, using its own tools, on internals nobody outside the building can see. You either trust it or you shrug.
Anthropic broke that pattern here. The core method is on GitHub under an Apache-2.0 license, in Python, so you can run it. They partnered with Neuronpedia on an interactive demo that applies the technique to open-weight models, so you can check the claims on networks that are not Claude. And Neel Nanda's team at DeepMind independently replicated some of the findings on an open-weight model, which is a competitor confirming a rival's interpretability result rather than politely ignoring it.
That combination is what turns this from an announcement into infrastructure. A technique you can run, on models you can inspect, that a second lab reproduced, is a technique you can start building tooling around. Auditing pipelines. Eval harnesses that read dispositions and not only outputs. Red-team passes that look for a fabrication pattern lighting up before the fabrication lands in the text.
None of that exists as a product yet. The point is that the floor got laid this month, in public.
What not to do with it
A few cautions, because this is the kind of result that gets oversold within a week.
The J-lens is not a lie detector you can bolt onto production tomorrow. Reading a list of words a model is disposed toward is not the same as knowing its intent, and the mapping from an internal pattern to a real-world behavior is still loose. A concept being on the model's mind does not mean the model will act on it.
It is also model-specific work in its details. The J-space Anthropic mapped is Claude's. The method generalizes, the replication shows that much, but the specific contents do not transfer, and anyone who tells you they can now read any model's mind is selling something.
And the consciousness angle is a distraction with a marketing budget. The engineering value here has nothing to do with whether the model experiences anything. It has to do with the fact that a small, busy, mostly hidden part of the network shapes the output, and you can finally list what is in it.
The part worth keeping
Strip away the philosophy and one sentence remains. The thing a model says is a filtered version of what the model is holding, and until this month you had no practical way to read the unfiltered version.
Now there is one, it is open, and a second lab confirmed it works. For anyone whose job is to trust a model enough to put it in front of customers, that is a better week than another point on a benchmark. It means "we evaluated the model" can start to mean more than "we read what it chose to tell us."
That is a low bar. We have been under it for years.
If you ship agents, treat output-only evals as necessary but not sufficient: a model can look aligned while internally flagging “this is a test.” Consider adding checks that probe for evaluation awareness or fabrication signals during runs, and compare internal-state readings between test prompts and real production traffic to see whether your green checks depend on the model noticing it’s being watched. At minimum, update your safety and reliability claims to distinguish “passes our output evals” from “shows no internal signs of test-conditioned behavior.”