Jev 走红两周:29 篇论文与一堆克隆,账单从数百降到几美元
作者认为 Jev 暴露了 agent 循环里大量「点击/安全/继续」式决策本不该交给聊天模型,有人称账单从数百美元降到几美元。
JEV DIDN'T CREATE A NEW TREND. It exposed a layer everyone was paying a chat model to fake. Two weeks of limited access. 29 papers already on arxiv. A rush of clones. A company brain thread claiming the bill went from hundreds to a few dollars. That's how you know the idea is real. Not because every demo is good. Because people suddenly noticed how much of an agent loop is just: click or don't safe or unsafe this file or that file keep going or stop A chat model can answer those. It just answers them slowly, expensively, and in paragraphs you then have to parse. The interesting part isn't "replace Claude." It's stop waking Claude for work that was never language. Tomorrow the timeline will argue about OpenAI's next agent. That's fine. The quieter question is still the same one Jev forced into the open: how much of your agent is thinking and how much of it is just deciding
这条是引用着另一条写的:「Jev + Opus 5.5 搭「公司大脑」的 10 步架构」