Researcher Proposes Giving Language Models 'Proprioception' to Make Introspection Verifiable
jonasgeiping · x · 2026-10-21
Researcher Jonas Geiping explains why he wants to give language models a sense of "proprioception" as a verifiable subset of broader introspection research.
- Introspection work often relies on self-reports that a neutral observer cannot verify; proprioception offers a verifiable internal signal.
- He asks three questions: (a) whether access to proprioception makes models safer (reasoning over failures, injections, spiraling states); (b) whether models become more legible to probes if their states are trained to be self-explainable; and (c) whether this validates Perez & Long (2023), who argue models calibrated on verifiable introspection questions will be more reliable on unverifiable ones.
- The quoted thread suggests a digital analogue: a model verbalizing which expert activated in a given layer.
More from AGI Musings
- Margaret Mitchell: click-driven AI misinformation is causing real harm — mmitchell_ai · 2026-10-21
- Mathematician Daniel Litt concedes his AI bet, but says AI won't solve math — ziv_ravid · 2026-10-21
- AI SaaS monetization is cooked: consumers small, enterprises pay $20/month — menhguin · 2026-10-21
- Tim Dettmers: Academia is about to have a renaissance — the best AI work of the next decade happens in university labs — StasBekman · 2026-10-21
- AI saves scientists 7 hours a week, but validation is the new bottleneck — AnneliesGamble · 2026-10-21
- UW's Amy X. Zhang brings CSCW theory to human-agent collaboration in Stanford talk — amyxzh · 2026-10-21