Jev 输出带置信度的结构化决策示例
用「如何应对未知无人机」的选项列表演示 Jev 返回 decision+confidence，并说明它负责快判断、harness 保留控制权、难例再升级给 GPT/Claude。
On September 15, TypeSafe AI revealed Jev, its first “System One Model.” The interesting part is not whether it competes with GPT or Claude. Jev is built for a different job: turning state into a bounded decision with confidence. For example: { "question": "How should I respond to this unknown drone?", "options": [ "machine_gun", "laser", "missile", "spear", "call_911" ], "decision": "call_911", "confidence": 0.98 } Instead of generating a long answer, Jev gives the system a decision it can directly use. That makes it interesting for agent and harness engineering: Jev handles fast judgment. The harness keeps control. If the case needs deeper reasoning, escalate it to GPT, Claude, or another frontier LLM. Not every decision in an agent loop needs the biggest model.
这条是引用着另一条写的：「ChatGPT 联合发明人发布新模型 Jev」