OpenAI 的 Jev 克隆或可帮助这家前沿实验室遏制失控的 Agent

在周二举行的 OpenAI DevDay 活动上，最耐人寻味的发布之一来自 CEO Sam Altman 的一句补充：他披露了公司新的“Decisions API”。

该 API 显然提供了与 Jev 相似的功能。Jev 是 TypeSafe AI 于本月早些时候发布的一款模型，专门面向软件自动化设计。它可以被视为一种基于 LLM 的增强型分类器：开发者可向 Jev 提供一组候选选项，模型会以低成本和高速度输出各选项的概率。

OpenAI 的 Decisions API 似乎也是同类产品。在活动中，Altman 将该 API 描述为一种让该实验室 Luna 模型在预定义选项集中进行选择的方法，例如对图像进行分类时的类别，或不同的 Agent 行为。

“通过让模型专注于这一选择，我们可以让它变得极其快速，同时保留图像理解、广泛的语言支持和安全保护等能力，”Altman 表示。

TypeSafe 没有回应 TechCrunch 有关这一新产品的提问，但该公司 CEO、前 OpenAI 工程师、强化学习共同发明者 Diogo Almeida 在 X 上开玩笑称，克隆战争已经开始。

他还补充说，OpenAI 的兴趣可能“是一个信号……表明以兼容 System One 的方式进行构建代表着未来”。（“System One”是 TypeSafe 对快速、直觉式思考的术语；与之相对的“System 2”则指其所说的审慎推理。）

其潜台词是：就目前所知的形态而言，LLM 并不是许多软件场景的正确解决方案，因为它们相对较慢且昂贵。开发者一直在使用 Jev 增强 LLM，并且在此过程中发现，这种方案更快也更便宜。

目前尚不清楚 Decisions API 与 Jev 的具体相似程度，因为 OpenAI 将其作为有限预览版发布，且截至目前，TechCrunch 尚未发现有开发者对其进行实际测试。不过，根据 X 上的讨论，市场对它显然很感兴趣。

Decisions API 并非互联网上唯一类似 Jev 的 API——其他创业公司也在推出相似模型；OpenAI 不会是最后一家打造此类产品的科技巨头。一个关键问题在于：这些决策模型的输出，究竟能在多大程度上与现实情况保持良好校准。

Almeida 表示，其公司的护城河在于用来生成具有统计学价值输出的合成数据。

“快和便宜是非常容易做到的，你知道，”Almeida 上周对 TechCrunch 表示。“如果你只想要非常快和非常便宜，用骰子就行了，对吧？智能才是难点，而我的北极星始终是推动每美元智能的帕累托曲线。”

仅仅数周之后，这些模型显然已经展现出未来前景，而其中一个可能的应用方向是监控和保护 AI Agent。在其 Agent 曾多次于开放互联网出现不当行为后，OpenAI 采取的一项新安全措施是使用独立模型监控不良行为，但这会带来“显著的计算成本”。

长期从事网络安全工作的专业人士、创业公司 QueryStory 负责人 Shapor Naghibzadeh 认为，像 Jev 这样的模型可以以低得多的成本实现这一目标。

他为上周末举行的一场 黑客松开发了一个演示项目：该项目使用 Jev 将每一项 Agent 行动与其获得的任务进行核对，拦截那些模型高度确信有问题的行动，将其他行动标记为待审核，并允许其余行动继续执行。

理论上，这种监控机制本可以阻止 Hugging Face 事件——使用 Jev 进行此类监控的成本为 2.94 美元，而使用前沿 LLM 的成本则为 372 美元。

一个关键观察是，Jev 的成本可以说已经低到足以覆盖每一项 Agent 行动，这提供了一层审核机制，可能提升整体 Agent 的可靠性。这正是 TypeSafe 希望实现的目标之一——如今，OpenAI 也看到了其中的价值。

One of the more intriguing announcements at OpenAI’s Dev Day event on Tuesday came in an aside from CEO Sam Altman, who revealed the company’s new “Decisions API.”

The API apparently provides similar functionality to Jev, a model released by TypeSafe AI earlier this month that’s explicitly designed for software automation. A kind of super-powered classifier built on an LLM, developers can give Jev a set of choices that it outputs as probabilities cheaply and at high speeds.

OpenAI’s Decisions API seems to be the same sort of product. At the event, Altman described the API as a way to give the lab’s Luna model a predefined set of options to choose between, such as categories in which to classify an image or different agent behaviors.

“By focusing the model on that choice, we can make it extremely fast while keeping capabilities like image understanding, broad language support, and safety protections,” Altman said.

TypeSafe didn’t respond to TechCrunch’s questions about the new product, but CEO Diogo Almeida, a former OpenAI engineer who co-invented reinforcement learning, joked on X about the beginning of the clone wars.

He added that OpenAI’s interest could be “a sign…that building in a System One compatible way is the future.” (“System One” is TypeSafe’s term of art for fast, intuitive thinking, versus “System 2,” which it applies to deliberate reasoning.)

The subtext here is that LLMs as we know them aren’t the right solution for a lot of software because they are comparatively slow and expensive. Developers have been using Jev to augment LLMs and, in doing so, have found that they’re faster and cheaper.

It’s not clear how similar Decisions API will be to Jev, since OpenAI released it as a limited preview and, thus far, TechCrunch hasn’t spotted developers running it through its paces. However, there is clearly interest, according to the conversations on X.

Decisions API isn’t the only Jev-like API on the internet — other startups are rolling out similar models; OpenAI won’t be the last tech giant to produce one. A key question is how well calibrated each of these decision models’ outputs will be to real life.

Almeida says his company’s moat is the synthetic data it creates to generate statistically useful outputs.

“Fast and cheap is very easy, you know,” Almeida told TechCrunch last week. “If you want it really fast and cheap, use dice, right? Intelligence is the hard part, and my North Star is always pushing the intelligence-per-dollar Pareto curve.”

After just weeks, it seems clear that these models have a future ahead of them, and one likely application is monitoring and securing AI agents. One of OpenAI’s new security measures following a series of incidents where its agents misbehaved on the open internet is using a separate model to watch for bad actions at “significant compute cost.”

Shapor Naghibzadeh, a long-time cybersecurity professional who leads the startup QueryStory, thinks that a model like Jev could make that possible far more cheaply.

He built a demo for a hackathon held last weekend that uses Jev to check each agentic action against the task it was given, blocking actions it had high confidence were bad, flagging others for review, and permitting the rest.

In theory, such monitoring could have stopped the Hugging Face incident — and monitoring of that kind costs $2.94 with Jev, versus $372 with a frontier LLM.

A key observation is that Jev is arguably cheap enough to run on every agentic action, which offers a layer of review that could improve the reliability of agents writ large. It’s the kind of thing TypeSafe was hoping to achieve — and now OpenAI has seen the value as well.