>understanding the results This will be a long one, so bear with me, if you wish, it's likely to be sloppy but this preoccupied my thoughts for weeks. 1) While 700+ OAI math papers is massive, I like ICLR paper submissions more, because big numbers are pretty for punchlines. Let's say that 50000 papers are accepted for ICLR somehow (plausible given pervious acceptance rates). That is half a million pages (+ addendums and references) of increasingly terse neuralese. a) how many human ML/AI researchers are needed to digest this volume of data, and how much time would it take them? b) will everybody sit tight and twiddle their thumbs while Team Understanding works in a coal mine, or will they continue to throw slop grenades at increasing caliber and quantity? Who will take a bet on former? 2) LMs are perfectly alright with producing bullshit papers and bullshit research, everybody has seen examples. Humans are also perfectly alright with producing bullshit papers and bullshit research, so much so that one could say that science was in crisis before AI became a thing. So task is not just understanding half a million pages, but validating that data. Math could be blessed here (if one would use such phrase), because, well Lean (ugh). What to do with ML/DL papers? Run everything on your own to confirm? LMs are better (== faster) at validating, but offense is favored: it's cheaper to generate than to validate. 3) We are in regime of intelligence explosion, as in right now. And it's down to throughput: it's possible to generate knowledge artifacts at unprecedent scale. Generation is much cheaper (in time and watts) than (machine) validation, and vastly cheaper than human understanding. Therefore, we might be heading into epistemic crisis. There is dumb little game I play with Chat, which goes like "bro, take in consideration only papers with high number of citations" (my logic being, it was validated by peers, which does not mean it valid, but what else can I do?) But I have no idea what to do moving on. It's not just that understanding will go the way of dodo and actually reading AI generated code, but I have no idea how to rely on AI to understand if we're washed in waves of slop. (Again and again, few big AIcos will be in advantage, because they can hope to maintain some epistemic hygiene internally. Sigh. And few is down to two right now, pretty much. Sigh, sigh.)