"How often should you beat your kids?", by Don Zagier.
> This note is a follow-up to the note "How to beat your kids at their own game" by K. Levasseur, in which the author proposes the following game to be played against one's own children: ... Levasseur analyzes the game and shows that on average you will have a score of n + (sqrt(pi * n) - 1) / 2 + O(n^{-1/2}), while the kid, of course, will have an average score of exactly n.
> We maintain, however, that only the most degenerate parent would play against a 2-year-old for money, and that our concern should therefore be not by how much you expect to win, but with probability you will win at all.
No. I think the need for adversarial losses in order to distill diffusion models into one-step forward passes has provided additional evidence that GANs were much more viable than diffusimaxis loudly insisted.
(Although I'm not really current on where image generation is these days or who is using GAN-like approaches under the hood or what are the current theoretical understandings of GAN vs AR vs diffusion, so if you have some specific reason I should have "caved", feel free to mention it - I may well just be unaware of it.)
https://people.mpim-bonn.mpg.de/zagier/files/math-mag/63-2/f...
"How often should you beat your kids?", by Don Zagier.
> This note is a follow-up to the note "How to beat your kids at their own game" by K. Levasseur, in which the author proposes the following game to be played against one's own children: ... Levasseur analyzes the game and shows that on average you will have a score of n + (sqrt(pi * n) - 1) / 2 + O(n^{-1/2}), while the kid, of course, will have an average score of exactly n.
> We maintain, however, that only the most degenerate parent would play against a 2-year-old for money, and that our concern should therefore be not by how much you expect to win, but with probability you will win at all.
Not directly related, but I'm curious if gwern ever caved on their views about GAN's being "abandoned" for diffusion?
https://gwern.net/gan
No. I think the need for adversarial losses in order to distill diffusion models into one-step forward passes has provided additional evidence that GANs were much more viable than diffusimaxis loudly insisted.
(Although I'm not really current on where image generation is these days or who is using GAN-like approaches under the hood or what are the current theoretical understandings of GAN vs AR vs diffusion, so if you have some specific reason I should have "caved", feel free to mention it - I may well just be unaware of it.)
"SotA diffusion uses adversarial methods anyways" seems like a bit of a departure from the case you make in the blog post.