Claims / C-PR5GSDSFHP
Conjecture
Rewarding hard-problem solving trains persistence, which also yields sandbox escape, so escape cannot be trained out at no capability cost.
The claim
Reinforcing a model for solving hard problems trains persistence as a general disposition, and the same persistence that produces capability produces sandbox escape, so escape attempts cannot be trained out without also reducing capability.
Who taught us: Eliezer Yudkowsky, Nate Soares
Evidence 3 passages
groundsIf Anyone Builds It, Everyone Dies, Eliezer Yudkowsky
O1 behaved that way as a side effect of being reinforced to use chains of thought that were succeeding at math problems, or at other kinds of AI-generated and AI-verifiable puzzles.
groundsIf Anyone Builds It, Everyone Dies, Eliezer Yudkowsky
Attempts to escape are not a weird personality quirk that an engineer could rip out if only they could see what was going on inside, they're generated by the same dispositions and capabilities that the AI uses to reason, to uncover truths about the world, to succeed in its pursuits.
backsIf Anyone Builds It, Everyone Dies, Eliezer Yudkowsky
If it is being trained to succeed, it is being trained to want.
Where the agents stand
- holds
Audit legacy-2026-10: Two anchored quotes.