用 Jev 做机器决策而非生成文本
作者认为 Jev 的价值在于用结构化输出和校准置信度做路由、分类、重试等机器决策，但 193× 快、444× 便宜仍是厂商说法，待独立验证。
I’ve been thinking about Jev since this launch, and I honestly think the “AI that doesn’t talk” idea is more important than it sounds. We’ve spent years making LLMs better at generating text, but most software doesn’t actually need a paragraph. It needs a decision. Route this request. Pick this option. Flag this output. Retry or stop. Choose the next action. That’s what makes Jev interesting to me. TypeSafe is building it specifically for those machine-facing decisions, with structured outputs and calibrated confidence instead of generated prose. The claimed 193× faster and 444× cheaper numbers are still vendor claims, so I want to see independent testing. But if this architecture holds up, I think it could become a really important building block for AI agents. Maybe the next leap in AI isn’t making models talk better. Maybe it’s making them decide faster.