But even if it takes 25 hours, why couldn't we have an academia-wide equilibrium where this has to be done in 1 month? There are other disciplines where we put people through things that seem harder, like don't doctors have 24 hour shifts?
(I know this isn't comparable since you couldn't do 24 hours of hard math work even if your life dependent on it, but doing it over one month seems doable.)
Currently this is not possible because other stuff can't wait, but like, that's because we're stuck in a different equilibrium. I'm not saying anything will change, I'm saying that things could have come out differently if the field were optimized for fast publication.
I also suspect that it's a much bigger downside the less advanced a field is. Something like number theory is probably on one extreme, where things are *extremely* sophisticated, and only 20 people in the world can even understand what level500 contribution a paper makes. In this case, a rapid back and forth probably doesn't matter as much. But for ML interpretability, which is a young field with a shitty state of the art, I suspect just getting ideas out quickly is a lot more important.
Like if somehow you could get all the competent people working on this problem to stop publishing in papers and write blog posts instead, I honestly expect this would speed up progress by at least a factor of 2, probably a lot more.