“On one hand, Eliezer writes extremely good explanations. I’m learning from his style a lot.”
Yeah, but they are rather verbose he tends to use 5 words when 2 would do.
“On the other hand, many people have pointed out that he doesn’t publish novel rigorous results, which kinda detracts from the aura.”
If you want to be in science this is a big issue unless your trying to pull a Wolfram and we all know how that turned out.
“On the third hand, he often finds and corrects non-obvious mathematical mistakes made by other people, including me, and he’s turned out right every time that I know of.”
But the math on this site what little there is tends to be toy problems and very simple. Let’s see him find and correct a mistake in some higher order fluid mechanics equations. I would personally like to see him solve a non-trivial second order non-linear partial differential equation.
“On the fourth hand, I’ve seen multiple cases where he made math mistakes of his own, and have discovered a couple of those myself. But that could be attributed to the fact that he publishes so much, and his error frequency is certainly many times lower than mine.”
That’s horrifying if you’re going to do science you have to control your error rate and that is where peer review comes in. (I recently submitted a paper where I was sloppy on some rounding of some of my results and I got slammed for it, science is all about precision and doing it right) If you don’t do the peer review then you may think your idea is good when if you actually had someone else look at it you’d see it was total trash.
“On the fifth hand, he has published novel non-rigorous arguments on real world topics that I don’t completely agree with but find pretty important. Biggest of them is the idea of Friendly AI.”
But for science and AI this is essentially meaningless since if your goal is to make an FAI then math and rigor is necessary. The ability to write non-technical papers arguing for some idea that is technical is trivial. The challange is getting the technical detail right. This is where I would like to see Eliezer submit some of his work on decision theory show that he is actually making a theory that is properly rigorous.
I think the worst thing would be if people here just wait for Eliezer and he shows up at the end of 10 years with an extremely long non-technical paper that gets us no closer to a real FAI.
While awesome math ability is a great thing to have, it would only complement whatever skills Eliezer needs to succeed in his AI goals. If Eliezer finds that he lacks the math skills at a certain point to develop some new piece of mathematics, he can find a math collaborator that will be thrilled about having a novel problem to work on.
I’m also not concerned about error rate. You write that the challenge is “getting the technical details right”—this is simply not true. It’s the main, big, mostly correct ideas we need to progress in science, not meticulousness.
(I recently submitted a paper where I was sloppy on some rounding of some of my results and I got slammed for it, science is all about precision and doing it right)
Publication is all about precision and doing it right, and it should be. But don’t you feel like the science was done before the more careful rounding?
“On one hand, Eliezer writes extremely good explanations. I’m learning from his style a lot.”
Yeah, but they are rather verbose he tends to use 5 words when 2 would do.
“On the other hand, many people have pointed out that he doesn’t publish novel rigorous results, which kinda detracts from the aura.”
If you want to be in science this is a big issue unless your trying to pull a Wolfram and we all know how that turned out.
“On the third hand, he often finds and corrects non-obvious mathematical mistakes made by other people, including me, and he’s turned out right every time that I know of.”
But the math on this site what little there is tends to be toy problems and very simple. Let’s see him find and correct a mistake in some higher order fluid mechanics equations. I would personally like to see him solve a non-trivial second order non-linear partial differential equation.
“On the fourth hand, I’ve seen multiple cases where he made math mistakes of his own, and have discovered a couple of those myself. But that could be attributed to the fact that he publishes so much, and his error frequency is certainly many times lower than mine.”
That’s horrifying if you’re going to do science you have to control your error rate and that is where peer review comes in. (I recently submitted a paper where I was sloppy on some rounding of some of my results and I got slammed for it, science is all about precision and doing it right) If you don’t do the peer review then you may think your idea is good when if you actually had someone else look at it you’d see it was total trash.
“On the fifth hand, he has published novel non-rigorous arguments on real world topics that I don’t completely agree with but find pretty important. Biggest of them is the idea of Friendly AI.”
But for science and AI this is essentially meaningless since if your goal is to make an FAI then math and rigor is necessary. The ability to write non-technical papers arguing for some idea that is technical is trivial. The challange is getting the technical detail right. This is where I would like to see Eliezer submit some of his work on decision theory show that he is actually making a theory that is properly rigorous.
I think the worst thing would be if people here just wait for Eliezer and he shows up at the end of 10 years with an extremely long non-technical paper that gets us no closer to a real FAI.
But those are just my thoughts.
While awesome math ability is a great thing to have, it would only complement whatever skills Eliezer needs to succeed in his AI goals. If Eliezer finds that he lacks the math skills at a certain point to develop some new piece of mathematics, he can find a math collaborator that will be thrilled about having a novel problem to work on.
I’m also not concerned about error rate. You write that the challenge is “getting the technical details right”—this is simply not true. It’s the main, big, mostly correct ideas we need to progress in science, not meticulousness.
Publication is all about precision and doing it right, and it should be. But don’t you feel like the science was done before the more careful rounding?