Jev decision model: first-week builds and cost comparison
Jev returns only decisions, not text; the post lists first-week community builds including a post scorer, millisecond UI assembly, Mac voice control and a 700-lead matcher, plus a GPT-6 cost comparison.
A former OpenAI researcher raised $40M for an AI that only makes decisions. Diogo Almeida, one of the co-authors of RLHF, the method used to train InstructGPT and GPT-4, has released Jev, a model that doesn't write text at all. You give it the current situation and a list of allowed answers. It picks one and returns a decision your code can act on right away. Why that matters becomes clear once you look inside any AI agent. Most of its expensive model calls go to small stuff: which button to click, which tool to use, whether to send a task for review. 10,000 of those decisions through GPT-6 Astra cost around $125. Through Jev, about 42 cents, based on public pricing as of September 20. With numbers like that, $40M on a narrow model stops looking like a strange bet. In its first week, people have already built: • a slop detector that scores posts as you scroll • a UI assembled from existing components in milliseconds • voice control for Mac that opens the app before you finish your sentence • a pipeline that matched 700 leads to outreach messages in 40 seconds for $0.09 I still wouldn't hand it anything important right away. The answer always comes back in the right format, but it can still be wrong. And 0.95 confidence doesn't mean the model is right 95% of the time. A sensible start: pick one decision your agent makes hundreds of times a day and run Jev in shadow mode next to a human. If the answers line up, every one of those branches costs pennies from then on. Bookmark this. The article below walks through 9 steps with code, cost math, and the places where Jev fails ↓
This post quotes another post: x.com/damir_akaza/status/2102336521235820836