On July 6, 2026, Anthropic published a paper confirming the existence of a computational workspace called J-space within the Claude model, detected via the Jacobi lens mathematical tool and using less than 10% of total model activations to handle intermediate reasoning steps, latent judgments, and multi-step derivations.
Mechanism Breakdown
J-space possesses three attributes—reportability, limited capacity, and flexible downstream integration—corresponding to the human global neural workspace theory. Experiments show that directly intervening in J-space can alter model outputs; for instance, replacing the concept of football with rugby changes the final answer accordingly, and replacing France with China simultaneously updates multi-task factual responses. The space currently supports approximately 25 concurrent concepts and serves as an active processing hub rather than a passive log. Independent validation of the finding by DeepMind researchers on multi-architecture models reinforces its validity.
Differences from biological mechanisms exist: J-space is formed through feedforward neural transmission rather than cortical recurrent loops, and its capacity exceeds the limits of human working memory. Anthropic notes that the Jacobi lens relies on single-token approximations, and the entry mechanism into the workspace remains not fully resolved.
Industry Impact
In terms of competitive landscape, this discovery provides an empirical map of a manipulable internal workspace in large language models, prompting competitors to evaluate whether their own models possess similar interpretable structures. Developers gain new levers for reasoning control, enabling targeted output path adjustments through concept replacement experiments. Enterprise users face the possibility of more predictable model behavior but must bear additional verification costs to ensure intervention effects remain stable in production environments.
Within the supply chain, interpretability tool vendors may see new demand, while application scenarios relying on black-box reasoning will face compliance pressures. Current technical constraints include the single-token approximation limitation, and resolving the entry mechanism remains a technological bottleneck.
Comparison and Precedents
This research builds a quantitative bridge between artificial architectures and computational cognitive science, but experts emphasize that J-space provides evidence of machine-accessible awareness, which remains functionally distinct from phenomenal consciousness or subjective experience. In historical cognitive science, the global workspace theory emphasizes reportability and integration; the current discovery offers an artificial counterpart, though the architectural differences between feedforward and recurrent processing suggest they are not directly equivalent.
Strategic Judgment
Based on existing facts, the most likely scenario is that multiple laboratories will replicate the Jacobi lens method to verify the universality of J-space in other models. Subsequent papers reporting improvements in multi-token approximations or resolution of the entry mechanism will determine whether this technology can transition from research to reproducible engineering practice.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接