Jev 决策模型跑 DOOM：每秒 10 次选择

Asteri @Asteri_eth原推发布于 2026.10.06

在 X 上打开原推

a model that can't write a single sentence is playing DOOM at 10 decisions per second that's the TypeSafe Jev demo, and it explains the whole idea better than any benchmark. Jev never outputs text. you give it state, it gives back a choice, a score or a yes/no with a confidence number. that's it. so instead of asking Claude to "think about" every fork in your agent, you hand the forks to something built only for forks: -> cheap model, frontier model or human for this step? -> does this answer actually match the sources? -> is this lead hot, warm or dead? -> post now, draft more or kill it? -> retry the tool call or give up? -> is this ticket urgent? then I looked at the pricing. $0.042 per million input tokens. output is free, because there's no output to bill. new accounts get $5 of credit, which TypeSafe says covers about 120 million tokens. at ~1K tokens per decision that's roughly 120,000 decisions before you spend a cent of your own. TypeSafe's own math puts that input rate 238x below Claude Fable 5.1. that's not Jev beating Claude at anything. it's Jev refusing to do Claude's job at all. the stack ends up in three layers: -> Claude plans, researches and writes -> Jev picks which branch runs next and whether Claude is even needed -> code does the clicking, saving, sending and executing OpenRouter already has a cookbook built on this: a cheap model answers, Jev checks if the answer is backed by the retrieved context, and only the failures go up to a frontier model. once you see it, every agent loop turns into a pile of tiny forks. > lead scoring. > content queues. > comment moderation. > retry logic. > research triage. > model routing. > even game loops. most agents don't burn money on the hard steps. they burn it on all the small ones in between. > Claude for the work that needs a brain. > Jev for the yes, no and which-one. > code for everything that just needs doing. I mapped 12 forks you can hand to Jev today + the exact cascade setup in the article below

这条是引用着另一条写的：x.com/asteri_eth/status/2104901781725598103