Anthropic Discovers J-space, an Internal Planning Mechanism in Claude Models

Anthropic announced that its interpretability research team has identified an internal mechanism in Claude family models called J-space, which functions like a global workspace where the model gathers the information it needs to plan and formulate its responses.

What is J-space

The research, published on the transformer-circuits.pub platform, builds on Anthropic's earlier work identifying individual interpretable concepts, known as features, inside the neural network. The new step involved linking those features into computational circuits, revealing part of the path the model follows from input words to output words.

J-space turns out to be a space where the model runs something like an internal monologue, building representations of concepts before it even generates the final answer visible to the user. The researchers also showed that Claude can think in a shared conceptual space independent of any particular language, suggesting a kind of universal language of thought, visible when the same sentences are translated into different languages.

The dictionary learning method

The technique behind the discovery, called dictionary learning, breaks down patterns of neuron activity into a set of human-understandable concepts. The algorithm compressed activation patterns into a dictionary of roughly 10 million relevant features, each corresponding to a cluster of neurons that fire together in response to a specific concept.

This approach differs from earlier interpretability efforts, which analyzed individual neurons in isolation. By linking features into circuits, researchers can trace not only what the model represents internally, but also how those representations influence one another on the way to the final answer.

Why this is urgent for Anthropic

Anthropic CEO Dario Amodei framed interpretability as an urgent company priority in an essay titled The Urgency of Interpretability. He compared the current state of AI development to building a nuclear reactor without blueprints or diagnostic tools, stressing that models have become too powerful to treat as a black box, especially as they move into defense, medicine, and finance applications.

We are building a nuclear reactor without blueprints or diagnostic tools - Dario Amodei, CEO of Anthropic

The company is testing several research approaches in parallel, from mapping individual neuron activity to analyzing recurring decision patterns and assigning meaning to specific parts of the neural network. The 2027 target is a concrete, time-bound benchmark, not a vague statement of direction.

Implications for safety and business

For companies deploying language models in decision-making processes, such as credit application scoring or medical diagnostics, interpretability tools offer a real chance to audit why a model reached a particular decision, rather than just observing the outcome. This could carry regulatory weight under the EU's AI Act, which requires explainability for high-risk systems.

Anthropic says it will collaborate with academic institutions and technology industry partners to further develop this line of research. It remains an open question whether the J-lens method will prove effective on models far larger and more complex than those studied so far.

Sources: Benchmark.pl (benchmark.pl), PurePC.pl (purepc.pl)