Post #3839713

@tef@mastodon.social @whitequark@social.treehouse.systems My intuition is that right now the models only "work" at a level of token spend & constant training that is economically unviable. It would be like if you had to burn a kilogram of solid diamond to produce a "cheap" digital watch, on the assumption that diamond would get cheaper. And yeah, maybe the truth is somewhere between the boosters "100x cheaper inference every year" and the skeptics "TCO isn't going down, ROI is zero" but I think it's closer to the latter than the former