News for hackers. Stuff that matters.
Re:On the Catastrophic Risk of AI by Vecr on Thursday June 01, 2023 @01:43PM (#36154669)
Well, it's quite likely not your objectives, it's probably a mistake in encoding what you want or something produced during training. If you pretended humans were evolution and AIs were humans, you could say evolution "wanted" humans to have sex for reproduction, but instead humans use condoms and birth control and stuff like that. It still satisfies the reward function (encoded as dopamine/etc in humans), but it's misaligned with what evolution "wants".

There are no comments on this comment yet.