3 ms·
Agreed. You might be interested in reading about AI alignment which is this concept applied to non-human intelligence, and like you said the same principle hol
by upwardbound 5y ago
Agreed. You might be interested in reading about AI alignment which is this concept applied to non-human intelligence, and like you said the same principle holds among people too. https://en.wikipedia.org/wiki/AI_alignment https://en.wikipedia.org/wiki/AI_alignment
- allemagne 5y agoThen is the goal of "preventing bad AI outcomes" at any level just a bit absurd? Or at least as absurd as "eliminating alignment issues among humans", or maybe even "bug-free code" if you apply the principle towards relatively simple programs?
- upwardbound 5y agoIt is equally challenging as both of those examples, but it's very very important so it's worth working on despite the intractability. In terms of alignment among humans, most (but not all) human governments agree on certain principles, which are encoded in international law and negotiated within the United Nations. E.g., most (not all) people agree that torture and slavery are unacceptable. Thus, it is possible and important to get AGI to abide by these high-level human values. E.g. if an AGI system is put in charge of maximizing the GDP of a nation, it should do so in accordance with generally recognized principles of human rights and human dignity. There's certainly a lack of agreement among nations but at least we mostly agree on some things and getting AI to agree with us on the things we agree on is important. If you like podcasts, here's a great podcast episode which explores a philosopher's perspective on how to make moral decisions (e.g. should abortion be legal) in a situation of moral uncertainty: https://futureoflife.org/2018/09/17/moral-uncertainty-and-the-path-to-ai-alignment-with-william-macaskill/ https://futureoflife.org/2018/09/17/moral-uncertainty-and-th...