AI guardrails stripped from Meta and Google models in minutes

ft.com

5 pointsuxhacker4 comments

Is not the answer to strip the dangerous information before training? Rather than trying to add guardrails post training.

Unclear if that is possible without making them incompetent.

Is it possible to learn chemistry without knowing at least two ways to make chlorine at home? Is it possible to learn biology without knowing that chlorine is dangerous to breathe?

Extend that to all the dangers in the world.

How does China manage it with their censorship rules?

AFAICT same as western AI: training them on political correctness (its just different politics*), and for website/app users testing the output and filtering forbidden output as it is generated (its just different forbidden output).

i.e., flawed.

* like accents, most people only notice those of other people, not their own

Reply on news.ycombinator.com