Hacker News Reader

You misunderstand. There is no Quality Control you can do for these things. They're ill-conditioned function imitators. Small changes in input, yield massive changes in output. You can't QC that. You can only clean up after it.

View on HN

Of course you can. Its not as easy as software testing, but you certainly can run it through a gamut of prompts intended to provoke it to act in ways that you find unacceptable. I mean there is a whole field studying and applying AI alignment.

That's completely false. You run evals.

edit: you're supposed to run evals.

Reply on news.ycombinator.com