I avoided trying to do anything about AI alignment for a long time because, knowing the kind of work Eliezer (and now MIRI) was doing and what they think is relevant to solving AI alignment, I thought I didn’t have anything meaningful to contribute unless I went back and became more skilled at things I didn’t find fun to work on. I’ve talked to other people who have expressed similar intimidation out of doing work in the space for one reason or another. That changed for me when I realized I had a lot of knowledge and ways of thinking that are different from those of folks currently doing most of the work in AI alignment, and thus given my interest in solving AI alignment I might have some non-negligible-even-if-low probability of making a contribution by bringing that different perspective to bear on the problem. It remains to be seen if this works out, but on the margin I think more people working on AI alignment is better and writing something like this may help convince some of those who could work on AI alignment but feel like they can’t because they’re not interested in machine learning or the depths of decision theory to work on it.
And if it helps any, this is basically the same advice Eliezer has given elsewhere in slightly different form, where he said something along the lines of “if most people had the same disregard for high chance of failure as I do and they were only successful 10% of the time they’d still turn out pretty successful”. This formulation of it in terms of a particular metaphor is just something that struck me yesterday and figured I’d share since it might click for some people when his expression of the idea didn’t.
I had exactly the same line of thoughts until last year, when I decided to write several articles about AI safety, on the topics which seem to be underexplored.
For what it’s worth, I thought your posts trying to cover underexplored topics weren’t very strong, but your recent post about the Fermi paradox was unexpectedly good. So it seems like doing stuff that interests you is more fruitful than doing stuff because it’s important/underexplored/etc. At least that’s always been true for me.
An authorial note on why I wrote this:
I avoided trying to do anything about AI alignment for a long time because, knowing the kind of work Eliezer (and now MIRI) was doing and what they think is relevant to solving AI alignment, I thought I didn’t have anything meaningful to contribute unless I went back and became more skilled at things I didn’t find fun to work on. I’ve talked to other people who have expressed similar intimidation out of doing work in the space for one reason or another. That changed for me when I realized I had a lot of knowledge and ways of thinking that are different from those of folks currently doing most of the work in AI alignment, and thus given my interest in solving AI alignment I might have some non-negligible-even-if-low probability of making a contribution by bringing that different perspective to bear on the problem. It remains to be seen if this works out, but on the margin I think more people working on AI alignment is better and writing something like this may help convince some of those who could work on AI alignment but feel like they can’t because they’re not interested in machine learning or the depths of decision theory to work on it.
And if it helps any, this is basically the same advice Eliezer has given elsewhere in slightly different form, where he said something along the lines of “if most people had the same disregard for high chance of failure as I do and they were only successful 10% of the time they’d still turn out pretty successful”. This formulation of it in terms of a particular metaphor is just something that struck me yesterday and figured I’d share since it might click for some people when his expression of the idea didn’t.
I had exactly the same line of thoughts until last year, when I decided to write several articles about AI safety, on the topics which seem to be underexplored.
For what it’s worth, I thought your posts trying to cover underexplored topics weren’t very strong, but your recent post about the Fermi paradox was unexpectedly good. So it seems like doing stuff that interests you is more fruitful than doing stuff because it’s important/underexplored/etc. At least that’s always been true for me.