Jev's first three days: flight lookup, 1,018 paper classifications, 500 emails
Developers used Jev for structured decisions: finding flights for $0.0039, classifying 1,018 papers for eight cents, and sorting 500 emails for three and a half cents.
A company spent two years building an AI model, then launched it able to do less than almost any model that came before it. It cannot write a single sentence, and that's the entire reason people can't stop talking about it TypeSafe AI's Jev refuses to generate text on purpose. The founder co-invented RLHF at OpenAI, then spent two years realizing something the rest of the industry still hasn't: most of the small decisions inside an agent were never writing tasks in the first place is this ticket urgent. which of five teams should own this. did the test actually pass. every one of those has exactly two or three valid answers, and for years, all of them got routed through a model built to generate language before landing on a word it could have returned instantly within three days of launch, developers were already proving exactly what that habit had been costing everyone. one found real flight results for $0.0039. another classified 1,018 research papers for eight cents total. a third sorted 500 real emails for three and a half cents. none of it used generation. all of it used a model returning nothing but probabilities against answers someone had already defined here's what should actually bother you about your own agent's bill. the expensive part was never the hard, creative work. it's dozens of tiny yes/no checks wearing a full paragraph as a disguise, billed at the price of language for an answer that only ever had two or three shapes to begin with you're still paying for a paragraph every time your agent could have just said yes full breakdown below
This post quotes another post: x.com/rvaniaaaa/status/2101709732314558639