The thought of an AI trained by youtube comments is scary as hell!
6 years ago | 4
androkguz I was going to say that we'd probably assume that the AGI would rewrite itself to remove the human requirement, but I'm not actually sure that's the case. If the machine's utility function literally gives no value without a human double check, you could maybe lean on the phenomenon where it's against your current function to accept a new one, and hope that the math keeps the idea of it rewriting its function out of the realm of possibility. Maybe. But the machine would operate veeeeeery slowly, even if you got the utility function going, I'm afraid.
5 years ago | 0
Hi Rob. I like your series on AI safety. I was wondering though why you've never mentioned anything about the following safety method: have AIs make plans that always require a final revision from a human before execution. I recognize this has its own problems but I think it's worth considering. The stamp collector satisfycer with the estimated espected value function would prefer strongly plans that don't destroy the world and are simple to understand because they are more likely to get approved. The tea maker won't break the house because that's unlikely to be approved.
6 years ago | 0
Robert Miles AI Safety
I'm trying out this 'posts' thing to let you all know that:
1) There's a new Computerphile video with me in it: https://www.youtube.com/watch?v=rURRY...
and 2) I made a follow up 2nd channel video, making fake AI-generated YouTube comments:
https://www.youtube.com/watch?v=M6EXm...
6 years ago | [YT] | 60