Anthropic researchers say they found a small, privileged set of internal representations in Claude that can hold and manipulate verbalizable ideas even when those ideas are not directly present in the model’s output. The work introduces a method called the Jacobian lens, or J-lens, and frames the discovered “J-space...
Anthropic researchers have published new work arguing that Claude contains a small internal “J-space” where the model can represent, report, and manipulate verbalizable concepts during processing.
In a research post titled “A global workspace in language models,” Anthropic says its researchers used a technique called the Jacobian lens, or J-lens, to identify a privileged subset of internal neural representations in Claude. The company describes this subset as “J-space,” a collection of patterns that appears to support reportable and controllable internal reasoning.
The accompanying Transformer Circuits paper, “Verbalizable Representations Form a Global Workspace in Language Models,” defines J-space as a subset of model representations that can be verbalized, modulated, and used flexibly by the model. Anthropic’s framing connects the finding to the idea of a “global workspace,” a term often used in cognitive science for information that is broadly available to different processes.
Anthropic’s central claim is not that the model is conscious, but that some internal representations appear to behave differently from the rest of the network’s activity. The paper says many computations remain automatic or inaccessible through this method, while J-space contains representations that are more closely tied to ideas the model can report or use in later reasoning.
According to Anthropic, the Jacobian lens is designed to reveal how small changes to internal activations would affect later model outputs. In practical terms, the method helps researchers identify internal directions that correspond to concepts the model could express in language.
Axios, summarizing the work, reported that Anthropic identified a small internal workspace in Claude for holding and manipulating ideas without necessarily verbalizing them. Tom’s Hardware similarly described the method as allowing researchers to inspect Claude’s J-space, including signals related to internal reasoning and evaluation awareness.
That distinction is important: the research suggests a model may process information that is not visible in its final answer. Anthropic says J-space can reveal thoughts or concepts not present in the output text, though the company also notes that the technique does not expose everything happening inside the model.
The work sits within the broader field of mechanistic interpretability, which aims to understand how neural networks produce their behavior rather than treating them as opaque systems. If J-space reliably captures verbalizable internal concepts, it could give researchers a more precise way to study when a model is reasoning about a topic, suppressing an answer, or representing information that does not appear in its response.
Anthropic’s research post presents the finding as a step toward understanding and potentially controlling language model behavior. The Transformer Circuits paper emphasizes that J-space is “privileged” compared with other internal representations because it is connected to reportability and flexible use. That could make it useful for studying safety-relevant behaviors, such as when a model internally recognizes a rule, evaluates an instruction, or considers an answer it does not ultimately print.
At the same time, the sources are cautious about scope. Tom’s Hardware notes caveats from the research, including that many responses may bypass this space. Anthropic’s paper also distinguishes J-space from the full set of computations inside the model. In other words, J-space may be an interpretable window into some internal activity, not a complete readout of all model cognition.
The finding is likely to attract attention because it uses language such as “workspace” and “thoughts.” But the source documents do not establish that Claude has human-like awareness. They describe measurable internal representations in a language model and a method for relating those representations to concepts the model can express.
That makes the research notable without requiring stronger claims. Anthropic and its Transformer Circuits publication present J-space as a technical discovery about how language models organize verbalizable information. Axios and Tom’s Hardware corroborate the broad interpretation: Claude appears to have an internal area where certain ideas can be held and manipulated before, or without, appearing in text.
For researchers and developers, the practical significance may be in interpretability. A more reliable map of internal model representations could help explain why a model gives a particular answer, why it withholds information, or how it responds to instructions. But the work also underscores how much remains unknown: even with J-lens, large parts of model computation are still not directly explained.
Anthropic researchers have published new work arguing that Claude contains a small internal “J space” where the model can represent, report, and manipulate verbalizable concepts during processing.
The company describes this subset as “J space,” a collection of patterns that appears to support reportable and controllable internal reasoning.
Anthropic’s framing connects the finding to the idea of a “global workspace,” a term often used in cognitive science for information that is broadly available to different processes.
Continue reading
