@neuralreckoning
@alexh
I do not want to just harp on that problem and did try to look past it, but that is the part I am genuinely interested in - the argument he was trying to make, but there were some parts that were very clearly GPT taking the wheel (the bioRxiv/hypothes.is part was particularly jarring because one can verify it's not the case by going to any biorXiv paper, but there is a ton of press coverage and blog posts that say it should be true that would have been in the training corpus). That made me doubt the entirety of the rest of the text because I couldn't tell what part was actually intentional, with some actual experience and insight that he as a human being has, and which part was just infill.
It turns into a fundamental issue when you structure a piece around a concept and then you're missing big pieces of it that get filled in halfheartedly by a language model. Like however you feel about cathedral/bazaar framing, it has tangible meaning. it also commits you to a number of structuring assumptions about what works - like even a journal mediated PRC process like eLife is not exactly a "bazaar-like" process. That space between "still journals mostly" and "self-coordinated public review" is the whole topic, but it only gets contended with piecemeal and without much treatment of the problem as a landscape of interrelated problems. As someone who has read a shitload of longform LLM prose masquerading as human text at this point on purpose, that kind of paragraph-level coherence but macro-level incoherence is a distinct hallmark.
=> More informations about this toot | View the thread | More toots from jonny@neuromatch.social
=> View neuralreckoning@neuromatch.social profile | View alexh@fediscience.org profile
text/geminiThis content has been proxied by September (3851b).