this article is so baffling to me man. i don't know what notion's evals are telling them, but i've tried to use the notion ai assistant, and unfortunately it's so deeply incapable i (and universally every coworker) instead interact with notion thru mcp via claude instead.
sure probably open source models are pareto-better at "highlight sentence make more formal" "summarize email" type tasks.... but no one cares about such tasks anymore! the bar is creeping ever higher. i don't want some dumb lightweight frontend additions. i expect to have something able to correctly determine *which* large chain of tasks to execute, handle all of them well, maybe use other weaker models as tools or subagents but always be able to reach into the internals as necessary. i don't want to waste my time with a long rabbithole because an open source model or haiku summarizes an email wrong!! just read the damn thing.
i "trust" notion as a platform to give me intelligence that's always behind the frontier, that's always less cohesive and worse than what i can expect if i simply exit it. and so no matter how cheap it is that means i just don't trust it at all! i wish they would get rid of every AI feature and quite honestly their block based database nonsense too and become a FUSE mount for markdown files with a multiplayer web view and editor. that's all i trust them to get right!
i'm picking on notion but it's not alone here. i think it's very telling that almost every "embedded ai" feature company out there is clearly and strictly worse than just fucking using claude code plus apis. your 500 person AI integration teams are *making things worse*! you are using bad models to save on costs *and we can all tell*! yea i know you can't get as good a deal on tokens as the labs do with their subs, i know you're struggling to figure out how to provide value here. i dont have an answer for ya. but using worse models and hoping ur mystical "optionality" and "product expertise" is gonna make up for it just ain't gonna work lmao.
gigafucked
@sarahmsachs
i think this is basically copeslop, and that the job of most "leaders" at most "ai companies" right now *is* basically just the production of copeslop. what else can they do?