factual
Self-improvement less necessary than once thought for danger
Recursive self-improvement is now less central to the danger than it seemed a decade ago, because modern AI like AlphaZero became superhuman in its domains without deeply understanding and redesigning itself, suggesting AIs could reach dangerous capability without a complete self-rewrite—though a sufficiently powerful unaligned system you don't control will still self-improve because that helps almost any goal.
factualpending
Speaker
Eliezer YudkowskyEvidence Quote
“alpha zero got to be superhuman in the domains it was working on without doing a bunch of without understanding itself and redesigning itself in a deep way”
Created: 6/14/2026, 2:07:58 AM
Your notes
Loading notes...