Researchers at Anthropic have identified a previously unknown internal structure within their Claude AI model, which they have termed “J-space.” This area appears to function as a workspace for the AI’s unspoken thoughts and active concepts, emerging organically during the model’s training process. The discovery has drawn attention from neuroscientists, as the structure exhibits similarities to leading theories on how the human brain processes conscious thought.
An Internal Notepad for AI
The “J-space” acts as an internal notepad for Claude, holding concepts that the model actively uses and steers while generating responses. Unlike the text-based “chain-of-thought” that is often visible in AI outputs, these internal hints remain off-screen, even when they guide the AI’s next answer. Researchers demonstrated the significance of this space by manipulating it: when an internal “spider” pattern was swapped for an “ant” pattern during a query about legs, the AI’s response shifted from mentioning eight legs to six.
Further experiments showed that while Claude could still engage in conversation and recall facts after the J-space was deleted, its ability to complete multi-step problems significantly degraded. This suggests that the J-space plays a crucial role in the AI’s more complex reasoning processes.
Implications for AI Consciousness Debate
Anthropic has faced criticism, notably from Microsoft AI head Mustafa Suleyman, regarding its discussions about AI consciousness. While the researchers emphasize that this discovery does not confirm whether Claude is conscious or capable of feeling, the existence of an undesigned, brain-like workspace within the model is precisely what fuels the lab’s ongoing inquiries. The finding adds a new layer to the complex debate surrounding the nature of artificial intelligence and its potential parallels with biological cognition.
In other AI developments, Chinese tech giant Tencent has open-sourced its Hy3 model. The model is designed for efficiency, using a small subset of its parameters per request, which allows it to run on less hardware compared to larger models. While Hy3’s performance in coding tests is surpassed by some other models, it reportedly excels in web research and tool use according to Tencent’s own evaluations. The release features a permissive Apache 2.0 license, making it broadly accessible.
Additionally, Replit is enabling users to build mobile app prototypes quickly. The platform’s AI tools can assist in defining user flows and generating initial app versions, which can then be tested on mobile devices via the Expo Go app. This process allows developers to iterate on app ideas by focusing on specific improvements identified during testing.
Steve Lopez is the Editorial Page Editor for News Raise. He covers Health. He has won more than a dozen national journalism awards for his reporting and column writing at seven newspapers and four news magazines.
