Why it matters

Superhuman AI is coming. Superalignment shapes long-term safety.

Advertisement

The architecture

Scalable oversight: humans can't check all outputs.

Weak-to-strong generalization.

Superalignment approachesScalable oversightAI helps oversee AIWeak-to-stronggeneralize supervisionInterpretabilityread AI thoughtsOpen research problem; multiple approaches being explored
Superalignment.
Advertisement

How it works end to end

Scalable oversight: use AI to help oversee more capable AI.

Weak-to-strong generalization: weaker supervisor generalize to stronger.

Interpretability: understand AI reasoning.