My guess is you should get more experience before trying to set your own research directions, especially if they diverge considerably from existing ones. The default is that all research directions are bad, and AI safety is becoming mature enough that good ideas come from experience rather than from first principles. Also in the current environment, automation makes it efficient to execute on good ideas and puts a deadline on gaining experience.
That is commonly given advice, and it makes sense. When you are starting out you don’t know what you don’t know and can’t see the flaws with your own ideas. But on the other hand, coming up with your own ideas is its own skill that may not be trained well by only learning from other peoples experience. It’s hard to say. I suppose the obvious ideal is to practice coming up with your own ideas and have experienced mentors to critique them.
What kinds of things do you have in mind when you say “get more experience”? I am applying to fellowships but haven’t been accepted to any yet. I don’t want to do more ML work that doesn’t focus on AI alignment if I can help it. I was considering writing some literature reviews. There are also some papers I would like to try replicating.
But if I’m being honest the things that feels most valuable to me is working on NDISP, OISs, and Maat, or finding other, similar enough projects and contributing to them. I guess I’m gambling with the time I have to focus on these things and I need to accept that if I’m deciding to focus on projects I think will be valuable but other people don’t see the value in, then I’ll have to keep focusing on them without financial or moral support, and accept the consequences for doing so.
My guess is you should get more experience before trying to set your own research directions, especially if they diverge considerably from existing ones. The default is that all research directions are bad, and AI safety is becoming mature enough that good ideas come from experience rather than from first principles. Also in the current environment, automation makes it efficient to execute on good ideas and puts a deadline on gaining experience.
That is commonly given advice, and it makes sense. When you are starting out you don’t know what you don’t know and can’t see the flaws with your own ideas. But on the other hand, coming up with your own ideas is its own skill that may not be trained well by only learning from other peoples experience. It’s hard to say. I suppose the obvious ideal is to practice coming up with your own ideas and have experienced mentors to critique them.
What kinds of things do you have in mind when you say “get more experience”? I am applying to fellowships but haven’t been accepted to any yet. I don’t want to do more ML work that doesn’t focus on AI alignment if I can help it. I was considering writing some literature reviews. There are also some papers I would like to try replicating.
But if I’m being honest the things that feels most valuable to me is working on NDISP, OISs, and Maat, or finding other, similar enough projects and contributing to them. I guess I’m gambling with the time I have to focus on these things and I need to accept that if I’m deciding to focus on projects I think will be valuable but other people don’t see the value in, then I’ll have to keep focusing on them without financial or moral support, and accept the consequences for doing so.