用 Jev 给邮件 agent 做上下文筛选
设想：500 封邮件只留 20 条进上下文，Jev 并行返回 include_in_context、类别、紧急度等结构化决策。
This is excellent. So much of agentic work is actually classification: routing to the right path, choosing the next action, or triaging context. Imagine an email agent with 500 new emails but room for only 20 in its context window. Before any real reasoning begins, it needs to identify messages from important people, active projects, or urgent threads. For each email, Jev could return something directly usable by code: include_in_context: true category: active_project urgency: 0.91 confidence: 0.97 Jev could be a game changer for latency and cost. Instead of generating an answer token by token, it returns typed decisions and probabilities in parallel. And it’s much more versatile than training a traditional BERT-style classifier for every task. You keep the flexibility of a prompt + runtime context, with the speed and predictable output of a classifier. Basically the programmability of a language model shaped like a classifier so it’s very cool. Congrats @CompleteSkeptic, you cooked!
这条是引用着另一条写的：「ChatGPT 联合发明人发布新模型 Jev」