@PosterInternet the USA have the best baseball team on paper. but, 1st, the delta in baseball isn't as large. 2nd, USA players are less interested in playing at WBC/not allowed to play at WBC by their MLB teams. and 3rd baseball is much more high variance than basketball (the sport we win most).
@DavidSacks and the solution is not to just say "ok then don't build ASI". ASI will arrive. this has been inevitable since the existence of the genus homo. you try to build it well (as they seem to be doing), but it's perfectly fine to talk about some of the potential downsides
@DavidSacks i think a lot of this is cope; it's much scarier for people who fear ASI if Dario/Sam are sincere in their belief in near-term superintelligence, so they cope and say they are insincere and this is some grand ploy for regulatory capture or to build hype.
@thomasahle the difficulty in answering this is much more about determining how hard the Millennium problems are than determining how capable models will be.
i.e. i'm confident i know roughly how good the models will be relative to humans in 5 years, but i have no clue how hard P = NP is
@regard_capital@punished_daniel dog what are you talking about. in what world is jacob jackson better. he has less impressive (although still very impressive) IOI results than ben/scott/andrew (and many other americans... e.g. rainboy, william), way lower CF and no hacker cup/code jam results... its not close
@FangYi11101@punished_daniel it makes a lot of sense. the non medalists are usually from small countries, don't grind olympiads, and are usually "just" smart people who go to a stem HS and are decent at math. pretty normal people. normal people don't pick their BF based on IOI dominance
@FangYi11101@punished_daniel i mean there are basically no candidates, at least from 2012-2014... like 10 people total and almost none were high scorers (a no medal doesn't strike me as the type to select their BF based on an IOI score). could just be talking about demi (obv overlapped at camp)
@bidiptas13 yea it's super sus. like they frame "content dependent selection" as the breakthrough even though standard attention does that as well (what do they think the QK matmul is for). but they don't say a single word about how their content dependent selection approach is subquadratic.
@sriramk And, I think right now China (despite US concerns) are not ASI-pilled. The only path for China to beat us at this point (given my short timelines) is a heavily nationalized effort; a full chip ban is one of the few things that might wake up the CCP and spark such an effort.
@sriramk I'm incredibly confident we have true ASI by around 2030 (so, fairly short timelines), but think it's probably optimal to sell smallish amounts of lower quality chips to China. You're not going to build ASI on 500,000 B30s, or even 2M B30s.