AI Research Circle: A Global Workspace Inside LLMs

​Do language models have anything like conscious access?

​Global workspace theory is a leading framework for how consciousness works in the human mind: many specialized processes compete for a shared stage, and whatever wins gets broadcast to the whole system. In a new interpretability paper, Anthropic researchers found evidence for a mechanism inside LLMs that behaves a lot like one.

​This session, Commons member David Valdman will walk us through the paper, "Verbalizable Representations Form a Global Workspace in Language Models":

• ​What global workspace theory says about conscious access, and why it's a useful lens for looking inside a model.

• ​The math and careful prompt engineering experiments the researchers used to support their claims.

• ​How far does the analogy go? We'll close with discussion and Q&A.

​This session is presented by David Valdman and will be filmed for the Rally SF YouTube channel.

​Pre-read (optional but encouraged):

• ​Anthropic's global workspace blog post: https://www.anthropic.com/research/global-workspace

• ​Interpreting GPT: the Logit Lens (a simpler technique this paper improves on): https://www.lesswrong.com/posts/AcKRB8wDpdaN6v6ru/interpreting-gpt-the-logit-lens

​Optional supplemental reading:

• ​The paper itself: https://transformer-circuits.pub/2026/workspace/index.html

• ​Global Workspace Theory: https://en.wikipedia.org/wiki/Global_workspace_theory

​About the AI Research Circle

​A community gathering at The Commons where we explore AI research together. No research background required. Just curiosity. Each session, we pick a paper or topic, break it down, and open it up for discussion. The goal: make cutting-edge ideas accessible and spark conversation across disciplines.

​Rally (rallysf.com) is a community for AI-first builders and operators in San Francisco.