Jev 是什么：只输出概率、不聊天的决策模型

2xnmore @2xnmore原推发布于 2026.09.19

在 X 上打开原推

The most interesting AI model launched this week cannot write a single sentence. That is not a bug. It is the entire pitch, and it points at where the AI bill actually goes. Here is Jev, from TypeSafe AI, step by step in plain words. 1. What it actually is Jev does not chat. You do not get paragraphs back. You send it a chunk of your program's state as text, plus questions where you already defined the possible answers. It returns the answer and a probability. Pick one option from a list. Rate something on a scale. Or give a yes or no as a number between 0 and 1. That is the whole interface. Typed questions in, probabilities out. 2. Why anyone would want that Look at what AI agents actually spend money on. Route this email. Is this tool call safe. Score this lead. Is this query worth sending to the expensive model. Does this response look like a jailbreak. Those are yes or no decisions. Today most agents send them to a giant chat model, wait seconds, then parse a paragraph to find the answer. Paying frontier prices for an if statement is the most repetitive waste in production AI right now. Jev is built to eat exactly that layer. 3. Who built it TypeSafe AI, San Francisco, founded around 2024. The CEO is Diogo Almeida, a former OpenAI researcher who worked on InstructGPT, RLHF, and GPT-4. He left because he decided chat was the wrong interface for automation. Two years in stealth. A $40 million seed led by DCVC. First public model shipped September 15. 4. How it is trained, and why this part matters most Chat models are trained to produce text humans prefer. Math and code models are trained on answers that can be checked. Jev is trained on calibration. The goal is that when it says 70% confident, it is right about 70% of the time. That is the real product. A probability you can actually trust lets software branch on it. An overconfident model ruins that. Their data is made in house, mostly synthetic. The architecture is closed. 5. The numbers, and where they come from Their own launch figures: $0.042 per million input tokens. Output tokens free. 70 to 500 milliseconds per call. They claim up to 193x faster and 444x cheaper than a frontier chat model on their internal workflows. Credit to them for publishing the footnotes. Speed was tested on laptops sitting next to the cluster. They admit they cannot prove the price is not subsidized. The eval harness was built by their own team, and the answer key is an average of two frontier labs' models, which tilts the board toward those labs. So treat 444x as the ceiling on their best case, not your result. 6. What independent testers found A Vercel engineer compared Jev with GPT 5.6 Luna on a command safety classifier and reported 5 to 18 times faster at p95, and more accurate. They have not published the accuracy table. One team, Bryo, ran email classification against Gemini. Gemini was slightly more accurate. Jev was 10 to 20 times cheaper. The probability score was the feature they actually wanted. That 10 to 20 times range is the honest early number until more third parties publish. 7. What it cannot do No essays. No code generation. No multi step tool use. No open chat. Nothing that needs to invent a string. Text in only. No images, audio, or PDFs yet. Around 255 options per question. No customer fine tuning at launch. The first party API is still waitlisted, and it buckled under launch traffic, though gateways like Vercel already list it. And read this one carefully. "Cannot hallucinate" means it cannot return an answer outside your list. It can still be confidently wrong. The CEO admitted that himself. 8. Why the timeline is loud about it Two reasons. First, the money. If a cheap, fast decision model can absorb the yes or no layer of agents, the big labs lose token demand on their most repetitive slice of production. That is a margin story, and the labs selling output tokens do not need it to be true. Second, the category. If "System One model" sticks as a name, every classifier that is not a chatbot will get called that within months. The API shape will be copied fast. The moat, if there is one, is calibration, the synthetic data factory, and latency. 9. How to think about it in one line Jev is a bet that the decision layer of AI agents deserves its own model class, cheap, typed, and honest about its own confidence, instead of renting a genius to answer yes or no. The chat race taught everyone that intelligence means talking. This launch argues that most working intelligence is deciding, and deciding was never a conversation. So which part of your AI bill is actually thinking, and which part is just an expensive if statement?