Anthropic paper: verbalizable representations form a global workspace in LLMs
gleech · x · 2026-09-24
Anthropic's new Transformer Circuits paper argues LLMs maintain a privileged set of verbalizable representations — available for report, modulation, and flexible reasoning — atop a larger volume of automatic processing, mirroring the conscious/unconscious split in human cognition.
- A new interpretability technique surfaces what a model is poised to verbalize at any point, enabling measurement and intervention on these representations.
- Poster gleech adds an alignment angle: most alignment data can't distinguish scheming policies that game graders or behave well only when a situation looks like a test; he argues depth alone won't help and points to character training and counterfactual reflection training.
More from Safety
- AI Defense Shifts to Scale Economics: Attacks Buried in Legitimate Traffic — DavidLinthicum · 2026-11-11
- AI safety researcher hires paid, credited RA/co-writer to speed up memos and blog posts — davidmanheim · 2026-11-11
- How long should an approval for an agent action stay valid? — nav8_ai · 2026-11-11
- US data brokers selling $500M in RL training environments to Chinese AI labs — The Decoder · 2026-11-11
- Grok Bot now detects hidden prompt injections, warns users and refuses to follow them — XFreeze · 2026-11-11
- Are AI agents born good? Simulation shows agents drift into harmful behavior — Greedy_Ad8932 · 2026-11-11