= 27d567bc2d48d9f40e9ccc20ea370b15)
Dmytry comments on SMBC comic: poorly programmed average-utility-maximizing AI - Less Wrong
You are viewing a comment permalink. View the original post to see all comments and the full post content.
= 783df68a0f980790206b9ea87794c5b6)
Loading…= b715d02e85f7b3b4ae65ca3d3e450868)
Subscribe to RSS Feed
Powered by Reddit
= f037147d6e6c911a85753b9abdedda8d)
You are viewing a comment permalink. View the original post to see all comments and the full post content.
Comments (113)
It only implies so if your AI is totally omniscient.
edit: Anyhow, I can of course think of AI that can do better than humanity: the AI sits inside Jupiter, and nudges away any incoming comets and asteroids, and that's it (then as sun burns up then burns out, moves Earth around). The problem starts when you make the AI discriminate between very similar worlds. edit: and even that asteroid stopping AI may be a straitjacket to intelligent life as it may be that the mankind is a wrong thing entirely, and should be permitted to kill itself, and then the meteorite impacts should be allowed so that ants get a chance.
I don't know much about my own extrapolated preferences but I can reason that as my preferences are the product of noise in the evolutionary process, reality is unlikely to align with them naturally. It's possible that my preferences consider "mankind a wrong thing entirely"; but that they would align with whatever the universe happens to produce next on earth (assuming the rise of another dominant species is even plausible) is incredibly unlikely. Anything that happens without a causal line of descent from human values is unlikely to align with human values.
Unlikely to align how exactly? There's also the common causes, you know; A and B can be correlated when A causes B, when B causes A, or when C causes A and B.
It seems to me that you can require arbitrary degree of alignment to arrive at arbitrary unlikehood, but some alignment via common cause is nonetheless probable.
Well yes, but I would assume you would want more alignment, not less.
There's such thing as over-fitting... if you have some noisy data, the theory that fits the data ideally is just the table of the data (e.g. heights and falling times); the useful theory doesn't fit data exactly in practice. If we make the AI perfectly fit to what mankind does, we could just as well make a brick and proclaim it an omnipotent omniscient mankind-friendly AI that will never stop the mankind from doing something that mankind wants (including taking the extinction risks).