The Intake
The Intake — Wednesday, July 15, 2026
On the substrate
Anthropic researchers identify verbalizable-representation subspace in Claude matching global workspace theory; ablating J-space collapses multi-step reasoning
transformer-circuits.pub VentureBeat Forbes
If you've been relying on Claude's ability to articulate its reasoning — treating verbal outputs as a meaningful signal in your agent pipeline — a paper published July 6 names the mechanistic basis for what you've been relying on.
The paper is "Verbalizable Representations Form a Global Workspace in Language Models," published at transformer-circuits.pub. The paper's Anthropic authors include Wes Gurnee and Nicholas Sofroniew. Adam Pearce is also listed. The method they introduce is the Jacobian Lens, or J-Lens. J-Lens computes the average downstream effect of each mid-layer Claude activation on output token likelihoods. The computation runs across thousands of contexts. Applying J-Lens identifies a subspace the authors call J-space. J-space accounts for roughly 10% of Claude's activation variance. J-space concepts are verbalizable — Claude can report on them when prompted. The paper identifies five functional properties of J-space. All five are consistent with global workspace theory. Global workspace theory is a cognitive science framework that describes how selected information is broadcast across specialized systems. Ablating J-space collapses multi-step reasoning. The paper does not take a position on phenomenal consciousness.
No near-term practitioner action follows from the research. If you use Claude's verbal reasoning as a signal in your pipeline, J-space is the mechanistic basis the paper names for that verbalizability.
---
For operators
China's anthropomorphic AI rules entered force today; ByteDance and Alibaba disabled agent features ahead of October 15 user-data deadline
TechNode South China Morning Post TechTimes
If you operate AI agent or companion-style interfaces with any Chinese-market user base, today is when China's "Interim Measures for the Administration of Artificial Intelligence Anthropomorphic Interaction Services" entered force.
ByteDance's Doubao disabled its user-created AI agent and humanlike companion features. Alibaba's Qwen did the same. Both acted ahead of the July 15 enforcement date. The Measures apply to platforms offering humanlike conversational AI agents. Required implementations include anti-addiction systems and mandatory usage notifications. An instant-exit mechanism is also required. User data from these discontinued services becomes inaccessible October 15, 2026.
If your deployment includes persistent AI agent or companion features serving Chinese-market users, October 15 is the operative compliance deadline — that's when user data from discontinued services goes dark under the Measures.
---