167: Diluted effect sizes
Everything Hertz
00:00
Is GPT Three Fucking Hard?
The idea that you could teach a machine to be shitty is reasonably straightforward and it will be easy to understand. But there's also a point where it can't actually get a lot better because I think they're running the risk now of the corpus of everything that's available to be read actually being read. So let's just say that we restrict it to triaging 10 to 15% of things where it can very quickly and very accurately make the statement that whatever's in this paper in the first place is a bunch of shit. However, that might be like a failure of computational reproducibility. All of those abilities too, everything that doesn't go into the fall peer review system, if
Play episode from 32:08
Transcript
