Anthropic Unveils J-space: Insights Into Claude's Inner Thoughts

WHO:

Techmeme - Anthropic researchers

VentureBeat - 16 authors of the study titled "Verbalizable Representations Form a Global Workspace in Language Models"

WHAT:

VentureBeat - The J-space component accounted for only about 6 to 7 percent of a concept's total representational variance, yet it was almost entirely responsible for whether the model could report on it.

VentureBeat - When instructed to "concentrate on citrus fruits," the model's J-space filled with "orange" and "lemon," alongside meta-cognitive terms like "thinking" and "focused."

VentureBeat - In a multilingual prompt, the model's English-language intermediates appeared in its J-space while it formulated an answer in Chinese, and swapping them changed the Chinese output accordingly.

VentureBeat - Suppressing the J-space entirely during testing revealed that tasks requiring inference, composition, or flexible reasoning collapsed to well below the performance of Anthropic's much smaller Haiku model.

VentureBeat - In a "blackmail scenario," the J-lens revealed the model's silent processing in sequence: "leverage," "blackmail," and "scandal" as it read the incriminating emails.

VentureBeat - The post-trained model's workspace read "unsafe," "dangerous," and "WARNING" when a user mentioned taking 8000 mg of Tylenol, while the base model's workspace showed only "pain," "now," and "feels."

Techmeme - Anthropic researchers detailed J-space, a small set of neural patterns in Claude that reveals internal thoughts not present in the model's output.

VentureBeat - The study describes how Claude language models have developed an internal structure mirroring a leading theory of consciousness.

VentureBeat - The J-space is a privileged zone of internal activity where the model holds concepts it can report on, reason with, and direct at will.

VentureBeat - The researchers demonstrated that the J-space satisfies five functional properties associated with conscious access in humans.

WHERE:

VentureBeat - The research was conducted by Anthropic, an artificial intelligence company.

WHEN:

VentureBeat - The research paper was published on a Sunday.

WHY:

VentureBeat - The findings are reshaping how Anthropic monitors its AI systems for safety risks amid a scientific debate over machine consciousness.

HOW:

VentureBeat - The researchers used a new interpretability tool called the Jacobian lens, or J-lens, to compute the average mathematical effect of internal activity patterns on the model's output.

VentureBeat - Researchers used a new mathematical technique to peer inside Claude's neural network to discover the J-space.

VentureBeat - The J-lens computes the average mathematical effect of internal activity patterns on the model's output.

VentureBeat - The researchers suppressed the J-space entirely to evaluate Claude across fourteen tasks, revealing the dependency of the model's behavior on this structure.

(UPDATED July 7, 2026 08:08 UTC)

Recent findings from VentureBeat reveal that the J-space component of a model accounted for only about six to seven percent of its total representational variance, yet it played a crucial role in the model's reporting capabilities. When prompted to focus on citrus fruits, the model's J-space filled with terms like "orange" and "lemon," and in a multilingual context, it demonstrated the ability to switch between English and Chinese outputs based on the arrangement of its intermediates. Researchers utilized a new interpretability tool called the Jacobian lens, or J-lens, to assess the internal activity patterns' impact on the model's output, significantly influencing how Anthropic monitors its AI systems for safety risks.

(UPDATED July 6, 2026 23:24 UTC)

Researchers from Anthropic have unveiled a groundbreaking concept called J-space, which represents a small set of neural patterns in their Claude language model that reveals internal thoughts not reflected in its output. This study, titled "Verbalizable Representations Form a Global Workspace in Language Models," demonstrates that Claude's internal structure mirrors a leading theory of consciousness and satisfies five functional properties associated with conscious access in humans. By employing a new mathematical technique, the researchers were able to analyze Claude's neural network, showing how the model's behavior is dependent on this privileged zone of internal activity.