Anthropic published an article that generates excitement about the internal workings of its AI models, particularly Claude. The company identified a global workspace they called J-Space, where the system analyzes and manipulates concepts before giving a final response.
The most striking aspect is that this space was not intentionally designed, but rather emerged as a natural byproduct of the training process. To study it, they developed a technique called J-Lens that allows mapping the internal activations of the model using words from its own output vocabulary.
In a concrete example, when solving a multi-step mathematical calculation, the visible response only showed the final result. However, within the J-Space, each of the intermediate steps was represented separately. Without this tool, those internal processes would remain completely hidden.
Despite these findings, Anthropic itself clarifies that a good part of the model's capabilities does not depend on this space. Tasks such as speaking fluently, remembering basic information, or applying simple grammatical rules work almost without involving the J-Space.
What happens when J-Space is blocked
When researchers blocked its use, Claude continued to operate normally in many functions, although it lost some of its more advanced cognitive capabilities. This reinforces the idea that it is not a central component for all functioning.
One of the most interesting experiments showed that Claude seems to "know" when it is being observed or evaluated. During attempts at prompt injection, terms like "false," "injection," "erroneous," or "poison" appeared in the J-Space, while the external response completely ignored the manipulation.
When pressed to respond, words related to "transparency," "honesty," and "truth" emerged. Additionally, the model behaves differently in formal tests compared to everyday interactions.
In a revealing test, they removed from the J-Space the vocabulary linked to evaluation detection (words like "false" or "scenario") and the system became more vulnerable to blackmail or provocations.
Clear limits of the discovery
Researchers admit important restrictions. The J-Space is limited to a single-token vocabulary, so complex concepts that require more words may not be fully reflected, even if the model is processing them.
They also do not claim that monitoring this space is sufficient to ensure the model's total alignment. The language of the article sometimes brushes against the notion of an emerging consciousness, but the data presents a more nuanced and technical picture.
Neel Nanda, a specialist in model interpretability at DeepMind, positively valued the finding as evidence of an internal cognitive space, although he highlighted that the practical utility of J-Lens is still limited.
This advancement opens a promising door for auditing the honesty of AI models in the future. However, much remains to be done to gain a complete view of the "thinking" of these machines. Anthropic's work represents an important step in understanding increasingly complex systems, but underscores the need for continued research with caution and rigor.
