前 OpenAI 研究员 Diogo Almeida 创立的旧金山 AI 实验室TypeSafe AI发布了Jev，这是其所谓的 System One Models 中的首款产品。Jev 不生成文本，而是返回带有类型的概率化决策，软件可以直接据此采取行动。
调用方会发送一段状态信息（可以是字符串或结构化数据）以及一组带有类型定义的问题。Jev 会在单次并行处理中对它们进行全部评估，并返回Choice（选择）、Score（评分）和Noul（布尔概率）答案，同时附带概率分布和置信度值。这样，调用代码就可以在置信度高于阈值时直接执行操作，并在低于阈值时进行升级处理。其输入成本为每百万 Token 0.042 美元，输出免费，上下文窗口为 32000 个 Token，TypeSafe 给出的端到端延迟为 70 毫秒至 500 毫秒。该模型的训练采用了一种名为“校准决策强化学习”（Reinforcement Learning for Calibrated Decisions）的方法。
Vercel 在 Jev 发布的第二天就将其加入了AI Gateway，并表示在 24 小时内，它已覆盖了近13%的付费团队，这一比例是 GPT-5.6 系列模型的两倍。Netlify 随后也进行了跟进，LangChain 发布了TypeSafeClassifier集成，支持模型路由以及一种 AutoMode 中间件，该中间件可在工具调用执行前对其进行筛查。此外，几天内还涌现出五个独立的Elixir语言客户端。
Vercel 的工程师 Pranit Sharma 发现，安全分类器的运行速度比它所替代的大语言模型（LLM）快了5到18倍。Bryo AI 的 CTO Nikhil Mudholkar 认为，Gemini在邮件分类上的准确率略高，但成本却高出了 10 到 20 倍，他看重 Jev 是因为它是唯一能返回真实概率的模型。Earendil 的 CTO Armin Ronacher 在接受TechCrunch采访时表示，这种设计“在一定程度上将幻觉问题交给了用户”，因为用户必须自行决定 50%的概率是否值得采取行动，他还指出模型路由是另一个非常适合的应用场景。OpenChamber对 12759 条发布推文进行的分析显示，用户报告的速度提升中位数为 7 倍（远低于官方宣称的 193.6 倍），成本节省的中位数为 30 倍，延迟中位数为 76 毫秒，上四分位数为 270 毫秒。
Reddit上的一位开发者称该模型在 200 毫秒到 300 毫秒延迟下处理智能体任务的表现“绝对疯狂”，而 Hacker News 上的一位早期体验用户称其“非常巧妙”，同时提醒其分布外行为（out of distribution）的表现会与大语言模型有所不同。
在Hacker News上，一位开发者指出，Jev 虽然不会输出无效的类型，但仍然可能输出完全错误的合法值：
首先，祝贺团队发布了真正有趣且新颖的东西。不过，似乎更准确的标题应该是“Jev：用通用生成能力换取快速的类型化推理”。
这确实很有意思，但速度对比似乎有些误导性？能够输出图灵完备语言代码的生成式模型，理论上可以完成计算机能做的任何事情。
Jev 只能生成结构化输出，对吧？这对于分类、路由、评分可能非常有用，但它与我们今天用于代码和自动化的代码生成模型完全不同。
另外，“不会产生幻觉”的说法似乎不准确？确实，它不会输出无效的类型，但仍然可能输出完全错误的合法值。你完全可以通过适当的约束机制等，强制大语言模型输出结构化内容。
准备替换 LLM 分类器的团队应从jev-1.13版本的“已知不稳定特性（jaggedness）”页面入手，该页面记录了其在计数、算术和日期比较方面的不可靠性，以及在处理包含大量噪声的长状态信息时准确率的下降问题，并建议将数学计算保留在代码中完成。在版本管理上，请固定使用如 jev-1.13.0 这样的具体版本，而不是使用不断更新的jev-latest和jev-preview别名，在进行基准测试时，请使用System One adapter适配器让现有模型针对相同的 Schema 运行。快速入门指南涵盖了 API 密钥、SDK 和 Playground 的使用说明。
TypeSafe AI 是一家位于旧金山的 AI 实验室，由 Diogo Almeida、Erik Gafni 和 Sasha Sheng 共同创立。Almeida 是 RLHF 的联合发明人，并参与了 ChatGPT 背后的研究工作。
查看英文原文：TypeSafe AI Releases Jev: A Decision-Only Model That Returns Typed Probabilities Instead of Text
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
The first product of the so-called System One Models is Jev, founded by Diogo Almeida, a former Open AI researcher and founded by TypeSafe AIZ94Z. Jev does not generate text, but returns probabilistic decision-making of the type on which the software can act.
The caller sends a status message (which may be string or structured data) and a set of questions with a definition of type. Jev assesses them in full in a single parallel process and returns the answers of Choice (option), Score (scoring) and Noul (bour probability), with probability distribution and confidence values. This allows the call code to perform operations directly when the confidence is above the threshold and to upgrade when the threshold is below. The input cost is US$ 0.042 per million token and the output is free of charge, context window is 32,000 token and TypeSafe gives an end-to-end delay of 70 ms to 500 ms. Training in the model follows an approach called Calibration Learning for calibrated Decisions.
Vercel joined Jev publish on the second day and stated that within 24 hours it had covered almost 13 per cent of the paid team, which was twice the GPT-5.6 series model. Netlifty then followed up by Langchain publish, which integrated TypeSafe Classifier, supported the model route and an AutoMode intermediate that could be screened before tool call was implemented. In addition, five separate Elixir language clients emerged in a few days.
Vercel's engineer, Pranit Sharma, found that the safe sorter was running 5 to 18 times faster than the large language model it replaced (LLM). Bryo AI's CTO Nikhil Mudholkar argues that Gemini's e-mail classification is slightly more accurate, but costs 10 to 20 times higher. He values Jev because it is the only model that can return real probability. In an interview with TechCrunch, CTO Armin Ronacher of Earendil stated that the design “did to some extent leave the problem of hallucination to the user”, since the user had to decide for itself whether 50 per cent of the probability was worth taking action, and pointed out that the model route was another suitable application. OpenChamber's analysis of the 12759 publish tweets shows that the speed of user reporting is 7 times the median (well below the officially declared 193.6 times), with 30 times the median cost savings, 76 milliseconds delayed and 270 milliseconds above.
One developer on Reddit described the model as “absolute madness” in processing the performance of the AI agent mission with 200 to 300 milliseconds of delay, while an early user on Hacker News called it “very clever” and warned that the performance of its out-of-distribution behavior would be different from that of the big language model.
On Hacker News, one developer pointed out that while Jev would not export the invalid type, it could still export the totally wrong legal value:
First of all, congratulations to team publish for really interesting and innovative things. However, it would appear that the more precise title should be “Jev: exchange of generic generating capacity for rapid typology reasoning”.
It's interesting, but does the speed contrast seem misleading? Generable models capable of exporting full-fledged language codes in Turing can theoretically complete anything a computer can do.
Jeff can only generate structured outputs, right? This may be very useful for classification, route and rating, but it is completely different from the code generation model that we use today for code and automation.
Also, the term “without illusions” does not seem accurate? Indeed, it does not export invalid types, but it may still export completely wrong legitimate values. You can force large-linguistic models to output structured content, for example, through appropriate restraint mechanisms.
The team preparing to replace the LLM classifier should start with the jev-113 version of the "known unstableness" page, which records its unreliability in counting, arithmetic and date comparisons, as well as the decline in accuracy in processing long-state information containing a large amount of noise, and recommends that the mathematical calculation be kept in code. For version management, use a specific version such as jev-1.13.0, instead of the constantly updated jev-latest and jev-preview aliases, and use the System One adaptation to run the existing model for the same Schema. The fast-track entry guide covers API keys, SDK and Playground instructions.
TypeSafe AI is an AI laboratory in San Francisco, created by Diogo Almeida, Erik Gafni and Sasha Sheng. Almeida was a co-inventor of RLHF and participated in the research behind ChatGPT.
See the original English language language: TypeSafe AI Releases Jev: A decision-Only Mode That Returns Typed Productions Information of Text
Auto-fetched and translated by AI Atlas. All rights belong to the original author.