Physicist Argues AI Training Should Incentivize 'Hard-to-Vary' Explanations
shyamalanadkat · x · 2026-08-09
The author argues that the model of discovery in AI matters as much as the model of physics. Knowledge grows through conjecture and refutation, not merely by induction over data.
Therefore, AI training should incentivize explanations that are "maximally hard to vary" (Deutsch's criterion, roughly equivalent to extreme compression plus counterfactual robustness). Subsequently, the system should select experiments that maximize expected refutation rather than expected confirmation.
More from AGI Musings
- AI Could Create Unprecedented Wealth But Make Millions Feel Economically Worthless — VraserX · 2026-08-09
- AI Safety Policy Should Be Physics: Controlling Reachable States Over Actions — pstryder · 2026-08-09
- Gwern's Essay: Building Personalized 'Guardian Angel' LLMs for Productivity and Cognitive Security — morgymcg · 2026-08-09
- Breaking the Supervised Learning Curse: Models Must Learn 'In the Wild' — tensorqt · 2026-08-09
- AI Billionaires Pledge to Give Away Fortunes: A Multi-Billion-Dollar Pinky Promise? — Wired AI · 2026-08-09
- China Builds Billion-Agent Society Simulator with Terrifying Predictive Power — pbaylies · 2026-08-09