Glyph on Nostr: thanks, this is good to have. I think it overstates its case and misreads the ...

thanks, this is good to have. I think it overstates its case and misreads the underlying papers (in particular, the whole "but models ARE trained with images now" thing is a complete red herring, because the cited paper specifically says "what’s interesting here is not that the tasks are impossible, but rather what makes them impossible"; the authors are aware that you COULD train on images), but it does have some useful points.