A global workspace in language models - By wesg

In this post, wesg explains new evidence that Claude has developed something like a “global workspace”: a small internal J-space where reportable, controllable, reasoning-relevant thoughts appear before they are spoken. The work suggests that some LLM cognition is split between automatic processing and a more deliberate, verbally accessible workspace, with implications for interpretability, AI safety, and the harder question of what kinds of “conscious access” language models may possess. * 00:00 - Introduction * 07:28 - How we found the J-space * 10:21 - Claude reports what’s in its J-space * 12:24 - Claude can control its J-space on request * 14:37 - Claude thinks in its J-space * 18:45 - Claude’s automatic processing skips the J-space * 21:40 - Monitoring Claude’s thoughts for misbehavior * 27:01 - Other results * 30:16 - What about consciousness? * 37:12 - External commentary https://www.lesswrong.com/posts/3PaLrzxagpbnNtPLT/a-global-workspace-in-language-models Get full access to Askwho Casts AI at askwhocastsai.substack.com/subscribe