@francoisfleuret@giffmana@thesignalnow The cost is likely the answer. One of the major reasons people switched from LayerNorm to RMSNorm was "same performance, cheaper computation". Why fixing what is not broken? Although you could run some tests to see if it helps in any way
@sergeax@Serg_VSV_ Все ещё не объясняет, почему вторник... Пятница и суббота тоже подходят под эти критерии. Наверное, под конец недели не хочется уже никого выбирать
@C_3C_3 Funny, Russian propaganda says Soros invests into turning countries into American marionettes, Republicans say he is destroying USA. I guess some names just attract the image of pure evil
@platoff@Yetanothergrump Чем экспонента интереснее полинома со степенью 1000? На практике - ничем абсолютно. Это не "не просто", это для N=2 уже сильно больше атомов в наблюдаемой вселенной.
@_TarunKathuria@DanielleFong How much? Say p(P=NP) = 1%, p(ETH is false) = 10%. Then p(P=NP | ETH is false) = 10%. So this epsilon could be substantial ;)
@_TarunKathuria@DanielleFong Interestingly, it seems like it should be! Say H is P=NP, E is ETH is false. Then p(H|E) = p(E|H)p(H)/p(E)=p(H)/p(E), since p(E|H)=1. p(E)<=1, so p(H|E) >=p(H). So if your expectation p(E) was smaller than 1, then p(H|E)>p(H).