Using Jev as a gatekeeper to cut AI pipeline costs
The author argues that running Jev first to filter most input and letting the expensive model see only what survives costs about $0.06 a day for screening papers.
whoever built this is f*cking genius you pay for a top AI model and then feed it everything. every email. every paper. every clip. every "is this even relevant?" it reads all of it at full price. Jev is the bouncer at the door. it's a model that can't write a word. it looks at what's coming in and answers one small question: -> worth reading or skip? -> urgent or can it wait? -> which of these 5 is the one? -> cut here or keep rolling? -> good enough to ship or send it back? every answer comes with a probability, in 70 to 500 milliseconds. then I looked at what people are already running on it and the pricing stopped looking real. every new arXiv paper screened each morning for about $0.06 a day a 97-minute podcast cut into 62 clips for about 3 cents of Jev per hour of video a simulated drone flown across a whole city for about $0.01 at $0.042 per million input tokens, 10,000 decisions of ~1K tokens come out around $0.42. Jev isn't smarter than Claude. it just never gets asked a hard question. the setup takes an afternoon: -> Jev looks at everything and throws most of it out -> the big model only sees what survived -> code does the clicking, cutting and sending and you start seeing the door everywhere. > the inbox. > the feed. > the reading list. > the support queue. > the raw footage. most of your AI bill is the expensive model reading things it should never have been shown. > let Jev guard the door. > let Claude think about what gets in. > let code do the actual work. the catch: text only, and still behind a waitlist. I broke down where to put this door in an agent today + the exact router setup in the article below
This post quotes another post: x.com/thegreatest_sv/status/2102802378378621355