Nice to see that there are not just radical positions in the AI safety crowd, and there is a drift away from alarmism and towards “let’s try various approaches, iterate and see what we can learn” instead of “we must figure out AI safety first, or else!” Also, Christiano’s approach “let’s at least ensure we can build something reasonably safe for the near term”, since one way or another, something will get built, has at least a chance of success.
My personal guess, as someone who knows nothing about ML and very little about AI safety, but a non-zero amount about research and development in general, is that the embedded agency problems are way too deep to be satisfactorily resolved before ML gets the AI to the level of an average programmer. But maybe MIRI, like the NSA, has a few tricks up its sleeve that are not visible to the general public. Though this does not seem likely, otherwise a lot of the recent discussions of embedded agency would be smoke and mirrors, not something MIRI is likely to engage in.
Nice to see that there are not just radical positions in the AI safety crowd, and there is a drift away from alarmism and towards “let’s try various approaches, iterate and see what we can learn” instead of “we must figure out AI safety first, or else!” Also, Christiano’s approach “let’s at least ensure we can build something reasonably safe for the near term”, since one way or another, something will get built, has at least a chance of success.
My personal guess, as someone who knows nothing about ML and very little about AI safety, but a non-zero amount about research and development in general, is that the embedded agency problems are way too deep to be satisfactorily resolved before ML gets the AI to the level of an average programmer. But maybe MIRI, like the NSA, has a few tricks up its sleeve that are not visible to the general public. Though this does not seem likely, otherwise a lot of the recent discussions of embedded agency would be smoke and mirrors, not something MIRI is likely to engage in.