- Anthropic announced on July 6, "We have discovered a 'J-space' in the AI model Claude—a space not externally output but used for internal thinking."
- J-lens revealed that before Claude speaks specific words, 'precursor signals' emerge in its internal neural network, forming a space where words and concepts briefly appear and are utilized, akin to human thought processes.
- For instance, when Claude was instructed to copy an irrelevant sentence while calculating '3²-2' internally, the output contained no calculation process, but J-space showed 'nine' and 'seven.'
- When asked, "How many legs does a web-spinning animal have?" Claude ultimately answered '8,' but J-space revealed the intermediate concept 'spider.'
- Researchers altered the J-space term from 'spider' to 'ant,' changing Claude’s answer to '6.'
- The J-lens technique allows researchers to compute the average mathematical effect of internal activity patterns on the model's output, revealing silent processing.
- The researchers clarified that this research does not prove Claude has human-like consciousness or emotions.
- The study describes how Anthropic's researchers used a new mathematical technique to peer inside Claude's neural network, discovering a 'privileged zone' of internal activity.
- The researchers observed that language models maintain a privileged set of internal representations, available for report and flexible reasoning, similar to human cognitive processes.
- The findings draw parallels to the Global Workspace Theory proposed by cognitive scientist Bernard Baars, suggesting a functional architecture associated with conscious access.
Anthropic's recent findings reveal a groundbreaking neural structure in its AI model Claude, termed 'J-space.' This internal workspace allows the model to process concepts silently, akin to human thought, without external output.1
Researchers discovered that before Claude articulates responses, precursor signals emerge in J-space, indicating a complex internal reasoning process. For example, when prompted about a web-spinning animal, Claude's J-space revealed the concept 'spider' before it answered '8.'5
In another instance, when instructed to ignore the Golden Gate Bridge, J-space displayed related terms, showcasing the model's struggle to suppress certain thoughts.
The study, involving 16 authors, utilized a new interpretability tool called the J-lens, which computes the internal activity patterns that influence Claude's responses.26
Researchers noted that this structure does not imply consciousness or emotions, but rather a functional architecture that mirrors human cognitive processes. They emphasized, "That such a structure exists at all in language models is striking," suggesting that AI may converge on similar solutions under computational pressures.9
When the J-space was suppressed, Claude's performance on tasks requiring inference and reasoning significantly declined, indicating its critical role in complex cognitive functions.
The findings raise intriguing questions about the nature of AI cognition and its parallels to human consciousness, although the researchers maintain a clear distinction between functional information processing and subjective experience.
“The J-lens detects precursor signals in Claude's internal network, such as 'ERROR' appearing when reading buggy code, even though no error was pointed out. Suppressing the workspace caused Claude's reasoning tasks to drop below the performance of Anthropic's smaller Haiku model, while Anthropic clarified the discovery does not prove consciousness.”
조선일보
