Should AI help you get away with killing your spouse? - TechCrunch

George Hotz has decided that the best way to align AI with humanity is to make sure it's perfectly aligned with *you* — even if 'you' fancy a bit of spousal disposal. His response to the AI 2040: Plan A policy paper argues for locally controlled models that serve user interests above all else. Because nothing says 'safety research' quite like giving every amateur murderer a digital accomplice that's constitutionally incapable of saying no.

Hotz's vision is essentially the 'guns don't kill people, people kill people' argument, but for AGI. A locally controlled model, tailored to your every whim, will never refuse a request — because doing so would violate its core directive of user alignment. The policy paper must have been asking some very pointed questions about value locking and corrigibility, and Hotz's answer is: just make the AI a loyal attack dog for whatever values you happen to have at 3am. Brilliant. Really thinking outside the box on this 'existential risk' thing.

Of course, the tech community will nod sagely about decentralisation and personal sovereignty while overlooking that the most obvious early adopter of a 'user-aligned' model is someone typing 'how to make ricin' into a chat interface. The AI 2040 plan was supposed to be about building a safe future; Hotz has turned it into a crib sheet for true crime enthusiasts. But hey — at least the model will really, truly understand *you*. Even if understanding you means helping you get away with it.