有人吐槽日本在 Agent 领域的存在感，还提到一个叫 Jev 的模型，说是没刷榜但全面更强，看看热闹。
这条推文讨论了一款代号 Jev 的智能体相关模型，作者认为不应只看其是否为基准测试优化。作者表示该模型在各方面表现都更强，而非仅针对 benchmark 调优。推文同时质疑日本在智能体（Agent）领域是否有实质性的工作产出。
Do the Japanese even have any work to do with agents? Jev, my ass No, it's not fucking optimized for benchmarks, it's stronger across the board.
蒸馏来的概率继承老师的偏差，加再多数据也去不掉，这篇用精确计算证明了这点，还拿 CFPB 23 万条投诉数据做了实测。
Typesafe 的 Jev 从黑客松亮相到日处理万亿 token 只用了五周，财富 500 强里四分之一在用，增长曲线很夸张。
一个挺有意思的行业观察：TypeSafe AI 想推 System One Models 这个词，结果没人用，大家把 Jev 叫 decision model。
The key insight is training the decision model on trajectories generated from the agent's own mistakes, which lets a 4B model outperform the larger Jev on out-of-distribution tasks. Builders running agent harnesses with high token bills can apply the fitted-threshold cascade to shave costs without touching task outcomes. The paper is hosted on DAIR.AI's open academy site, so the full write-up is freely available.
The non-obvious trick is labelling training data with the agent's own post-mistake states, which makes the small decision model generalise beyond familiar trajectories. Builders running agentic pipelines can replicate this distillation-and-cascade pattern with any open 4B model to slash inference cost without changing outcomes. Compared with fixed-threshold routing, fitting the confidence threshold on the agent's own trajectories is what preserves task results.
