交易 bot 用 Jev 做买/跳过/等待的快速判断

Polykinder @polykinder原推发布于 2026.09.25

在 X 上打开原推

Opus 5.5 writes the brief. Jev makes the calls. A thin code layer fires them. The one piece that makes every actual decision - Jev - can't write a single word. That's exactly why it wins. Most agent stacks wire one big language model into every step. Buy or skip. Now or wait. Thousands of yes-or-no calls a day, all sent through a model built to write essays. It works. It also crawls. You're paying a poet to flip a light switch. So the bot splits the job. Here's how one decision actually runs: new forecast comes in → structured into a clean market state → Jev reads it against the live price → Jev classifies: buy / skip / wait → only the rare edge case gets kicked up to the big model → order fires, or it doesn't → logged to a decision journal The big model never writes an essay about a routine buy. Jev just gets narrow questions: does anything here even change the current setup? is this routine, or one of the rare hard calls? does this need the expensive model, or can it close now? flag, skip, or wait? Everything easy dies at Jev in a quarter of a second, for a fraction of a cent. Only the genuinely hard calls reach the expensive model. Thinking happens rarely. Choosing happens constantly. Price them differently. We spent two years teaching AI to talk. The unlock was building one that shuts up and picks.

这条是引用着另一条写的：x.com/polykinder/status/2103131070002495608