"Alignment is a diff you can ship": Andrew Ng's most contrarian line on model alignment

ccerrato147 · x · 2026-09-22

In a thread replying to Andrew Ng, ccerrato147 highlights Ng's most contrarian take: our tools for making a model less racist are better than our tools for making a racist human less racist. "Go in, zero out a few numbers, done. Alignment is a diff you can ship."

The point: model alignment is engineering-controllable in ways human behavior change is not — a technical rebuttal to 'AI is uncontrollable' narratives.

Related event: Andrew Ng's 'Alignment Is a Shippable Diff' Sparks Debate on Fear as Moat(2 posts)→

Original post →

More from AGI Musings

AGI Musings channel →