Oh, I 100% agree we are not there. Every time a new model comes out I cook up a few undergrad math homework questions to see if they’ve gotten any better, and s

Oh, I 100% agree we are not there. Every time a new model comes out I cook up a few undergrad math homework questions to see if they’ve gotten any better, and so far they haven’t. But, the thing I like about “reducing” AGI to solving math is that it kind of sidesteps these thorny questions about interacting with the real world to conduct experiments, etc. It’s something that, in principle, can be done with the current text interfaces, and it’s pretty easy to tell whether or not we’ve gotten there.