Anthropic research finds verbalizable representations form a global workspace in LLMs
MoonL88537 · x · 2026-10-19
Anthropic's Transformer Circuits Thread published 'Verbalizable Representations Form a Global Workspace in Language Models' (July 2026). Using a new interpretability technique that surfaces concepts a model is poised to verbalize, the team found LLMs maintain a privileged set of internal representations available for report, modulation, and flexible reasoning — atop a much larger volume of automatic processing — an analogous functional split to conscious vs. unconscious human cognition. Measuring and intervening on these representations offers a window into the model's thought processes. Authors include Wes Gurnee and Jack Lindsey.
More from AGI Musings
- Top probability journals shut submissions for a month citing surge and AI impact — michaelchchoi · 2026-10-19
- Peter Diamandis: AI won't make humans useless — losing independent thought is the real risk — PeterDiamandis · 2026-10-19
- The FDE craze is just consulting rebranded, and customers may pay the price — springrod · 2026-10-19
- "Adapting to AI" means changing the human condition, not learning new tools — danfaggella · 2026-10-19
- "I'm grieving dead dreams": a Reddit reflection on AI's emotional toll — 777thLight · 2026-10-19
- Poisoned SERPs hit 25% of tracked keywords; AccuRanker explains Google's anti-scraping purge — lilyraynyc · 2026-10-19