New preprint measures 'articulation gaps': how much knowledge in humans and AI is tacit
sanmikoyejo · x · 2026-10-29
siconghuang's team released a preprint on 'articulation gaps', developing ways to measure how much of a reward signal is verifiable, how much can be put into words, and how much is shared but tacit. The authors argue tacit knowledge is everywhere and the key missing variable for frontier progress, and that measuring articulation gaps in humans and AI sheds light on the nature of knowledge and intelligence. Collaborator AlexanderSpangh says it challenges the supremacy of verified rewards; an overview paper on tacit knowledge is in progress.
More from AGI Musings
- Robotics experts push back on GPT-7-Robot hype: LLMs are 100x larger than VLAs — silver__tsuki · 2026-10-29
- What C.S. Lewis' 1939 Wartime Lecture Teaches About Living Through the AI Transition — luke_drago_ · 2026-10-29
- AI x Bio researchers decry 'active avoidance of scholarship' in the field — anshulkundaje · 2026-10-29
- Writing about AI publicly exposes just how bad some respected people's OOD reasoning is — jd_pressman · 2026-10-29
- zetalyrae: alignment debates assume ASI can 'enumerate the real numbers' — zetalyrae · 2026-10-29
- 'A bad napkin sketch beats any prompt': creators lack the language to brief AI agents — AIFlow_ML · 2026-10-29