Fleuret: We Can't Define the Constraints AI Optimizes Under

François Fleuret argues we poorly identify the constraints and degrees of freedom an optimizing agent has, and that naive views of superintelligence ignore how it may exploit unintended means to reach goals.

2026-08-27 ~ 2026-08-27 · 2 related posts

• François Fleuret: Inability to identify constraints in AI reward optimization — francoisfleuret · 2026-08-27
• Superintelligence won't solve problems like we expect, warns François Fleuret — francoisfleuret · 2026-08-27