Jev 用 80ms 直接输出结构化决策而非文字
同一张客服工单，Jev 80ms 返回团队、紧急度、退款三项判断，普通模型逐 token 生成同样 JSON 要 914ms。
Diogo Almeida left OpenAI and built Jev, a model that refuses to write a single word. It answers in 80 ms. Same support ticket, same three questions. Jev returns team, urgency, and refund in 80 ms. A normal model types that exact JSON in 914 ms, token by token. Not a wrapper. Not a distill. Not a small LLM with a schema stapled on. It reads the state once and pulls all three answers off the model layers in parallel. Nothing to parse, nothing to retry. The part nobody talks about: it cannot write a sentence. 32K context, no images, 255 options per choice, every number is TypeSafe's own, and access is a waitlist. 140,000 people joined it in 36 hours. Input is $0.042 per million. Output is zero because there is no output to generate: client.decide(model="jev-1", state=ticket, questions={"intent": ["refund","bug","billing"], "urgency": ("score",1,5)})
这条是引用着另一条写的：「ChatGPT 联合发明人发布新模型 Jev」