8 Jul 2026

Digest: Beijing weighs curbing overseas access to China's top AI models, experts respond to Anthropic's global workspace paper

Today's most significant AI capability, safety, and governance developments. (Beta — please verify against the linked sources.)

1. Beijing is looking at curbing overseas access to China's top AI models, sources say Recommended

   China's Ministry of Commerce has met with Alibaba, ByteDance and Z.ai about potentially restricting overseas access to the country's most capable AI models — including unreleased systems and widely used open-weight releases like Qwen and GLM — with a tiered access regime and national-security penalties for leaking proprietary model technology among the options discussed, per Reuters. Export controls have been the dominant governance story in recent weeks, driven almost entirely from Washington; serious movement toward a Chinese counterpart regime would make frontier-model access a two-sided control problem and could sharply curtail the global availability of Chinese open weights.

   reuters.com
2. Neuroscientists and AI-consciousness researchers publish invited commentary on Anthropic's global workspace paper

   Anthropic has published invited independent commentary on its recent 'global workspace' interpretability paper: Stanislas Dehaene and Lionel Naccache — originators of the global neuronal workspace theory of consciousness — assess how far the findings parallel the human workspace (asking directly whether Claude possesses a conscious one), while Eleos AI researchers weigh what the results mean for AI consciousness and moral status. Soliciting named outside experts to publicly stress-test a major interpretability result is a notable step toward independent scrutiny of lab research claims.

   www-cdn.anthropic.com
3. A Review of Anthropic's Global Workspace Paper

   In a detailed public review, Google DeepMind interpretability lead Neel Nanda endorses the paper's central scientific claim — that models contain a 'cognitive space' acting as working memory for intermediate variables during a forward pass — and reports independently replicating the core results on an open-weight model (Qwen 3.6 27B). He judges the J-Lens technique a useful if limited tool for alignment audits and model forensics, a meaningful independent validation given how often headline interpretability results fail to reproduce outside their original lab.

   Neel Nanda via LessWrong