# 前 OpenAI 研究员打造 Jev：不做文本生成、只做选项判断的 AI 模型

- 来源：The Decoder：AI News（RSS）
- 作者：Maximilian Schreiner
- 发布时间：2026-09-16 23:19
- AIHOT 分数：42
- AIHOT 链接：https://aihot.news/items/cmu48z4fr0ecnro4w8yflkec6
- 原文链接：https://the-decoder.com/former-openai-researcher-builds-an-ai-model-that-judges-options-instead-of-writing-text

## AI 摘要

初创公司 TypeSafe AI 推出模型 Jev，不生成文本，而是在软件内部输出窄域判断与概率，由开发者预设问题与候选答案、模型为选项打分。TypeSafe 称其响应时间为 70 至 500 毫秒，定价为每百万输入 token 0.042 美元、输出不收费，开发者需通过 waitlist 申请接入。公司宣称 Jev 不会产生幻觉，但该保证仅限输出结构，选项内的事实性错误仍可能出现。

## 正文

TypeSafe AI

Startup TypeSafe AI built a model called Jev that delivers fast judgments inside software instead of writing chat responses. The promised upside is speed and cost, but the limits are tight.

Startup TypeSafe AI has introduced an AI model called Jev that takes a different approach than typical chatbots. Instead of generating text, emails, or code, Jev is built to deliver narrow judgments and probabilities inside other programs. Co-founder and CEO Diogo Almeida previously worked at OpenAI and is among the authors of the InstructGPT research that laid the groundwork for ChatGPT.

The core idea behind Jev: Developers define questions and possible answers, and the model scores the options.

Sorting requests instead of answering customers

An example from the documentation shows what this could look like in practice. A customer writes to an online store saying they were charged twice for an order and asks for a refund on the second payment. Before anyone can handle the request, it has to be sorted. Is it about a payment, a delivery, or a return? Does the customer want money back, or just an explanation?

This is where developers connect the store software to Jev and define these questions along with the allowed answers. With each new message, the software hands the customer's text to Jev. What it gets back isn't a reply letter but, say, the label "payment issue" and a probability that the customer wants a refund.

The store software can then attach fixed rules. Payment issues go to accounting, refund requests get flagged. When the label is unclear, a staffer reviews the request. Jev provides the assessment, and the software decides what happens based on the programmed rules.

That doesn't mean the money is back yet. Whether there really was a double charge has to be sorted out later against the actual transactions. A separate chatbot could tell the customer the confirmed status. For the sorting alone, it isn't needed.

Sub-second response times could open up new uses

TypeSafe says Jev delivers answers in 70 to 500 milliseconds, many times faster than even the fastest current language models. To get there, the model skips step-by-step text generation and, the company says, computes several outputs in parallel. Adding more questions in the same call barely increases the response time.

视频 · 前往原文观看

That speed could allow a check before every response from an AI assistant. Does the draft contradict the conversation so far? Does it claim a refund that isn't recorded in the customer's account? TypeSafe describes checks like these in its example workflows. The company lists more possible uses in sales and customer service, like spotting buying intent, sorting requests by topic, or judging when a human should take over.

But regular language models can also output preset categories and data structures. OpenAI offers Structured Outputs for exactly this. A structured response format alone doesn't set Jev apart from the competition. To win, the model would have to handle these tasks not just faster or cheaper, but with good enough quality.

The published performance tests have limits. TypeSafe compares four workflows it built itself and uses responses from other AI models as its reference, not independently verified correct solutions. GPT-6 Astra is also missing from the evaluations. The model is very cheap, though. TypeSafe lists a price of $0.042 per million input tokens. It says it charges nothing for outputs.

No hallucinations doesn't mean no mistakes

TypeSafe markets Jev as a model that can't hallucinate. But that guarantee only covers the allowed output structure. The model won't produce an answer outside the preset options. A factually wrong choice within those options is still possible.

The model is meant to make many small judgments in the background more economical. Whether those judgments are reliable enough is something companies will have to test for their own tasks. Developers get access first through a waitlist.