Using AI to solve most of alignment is a fundamentally doomed approach: an intuition
The author argues that using swarms of non-superhuman AIs to align superintelligence is fundamentally doomed because alignment is a systemic property of a whole architecture rather than an isolated component. The piece draws analogies to rocket engineering and gradient descent to demonstrate that optimizing solely for capabilities while hoping alignment is preserved or manageable through swarm oversight fails to account for how goal-directed systems generalize and optimize around constraints. Open article →