Jev：200倍速决策AI，非文本生成的新架构
Jev Just Made AI 200x Faster…But There’s A Catch
Jev展示了非自回归的决策范式，对优化Agent推理延迟和成本有直接参考价值，建议关注其并行采样与置信度校准的工程细节。
A new AI called Jev is taking the world by storm. But why get excited for an AI system that can't even write text? It properly run over the internet. So let's try to separate the facts from the hype. Okay, so it does not write text. Instead, it makes decisions, a question, and some options go in. One answer comes out. Kind of like who wants to be a millionaire? But you don't wait [music] for the answer. It pops out instantly, up to about 200 times faster than our chatbots today.
一个名为 Jev 的新 AI 正在全球掀起热潮。但为什么我们要为一个甚至无法撰写文本的 AI 系统感到兴奋呢？它实际上是在互联网上运行的。所以让我们试着将事实与炒作区分开来。好吧，它确实不写文本。相反，它做的是决策：一个问题和一些选项输入，然后输出一个答案。有点像《谁想成为百万富翁》？但你不需要等待[音乐]答案。它会瞬间弹出，速度比今天的聊天机器人快大约 200 倍。
That is very useful. You can also play video games with it. Hooray. The question is the world state. And the potential answers are the moves you can make. And if it is extremely fast, it can play things in real time. But it gets better. LLMs are relatively slow. So when you code up something with a super smart AI, everything takes a long time. But along the way, you also have a few simple decisions, too. Which file do I open?
这非常有用。你甚至可以用它玩电子游戏。万岁。问题是世界状态，而潜在的答案是你可采取的行动。如果速度极快，它就能实时进行游戏。但这还不是最棒的。大型语言模型（LLMs）相对较慢。所以当你用超级智能 AI 编写代码时，一切都需要很长时间。但在过程中，你也会遇到一些简单的决策。比如：我该打开哪个文件？
Which failure matters. Retry or move on. And this is where Jev can help. Offload these decisions to Jev and it solves them nearly instantly. In some cases, it can meaningfully speed up a fablelike AI system and even cut down the token usage. It makes things cheaper and faster. That is amazing. Now, there is no research paper, but some of you fellow scholars probably say, "Okay, this is a classifier, an idea that dates back around 90 years now."
哪种失败是关键的。重试还是继续前进。这正是 Jev 能帮上忙的地方。将这些决策交给 Jev，它能近乎即时地解决它们。在某些情况下，它可以显著加速类似寓言般的 AI 系统，甚至减少 token 的使用量。这让事情变得更便宜、更快。这太棒了。现在，虽然没有研究论文，但你们这些同行学者可能会说：“好吧，这是一个分类器，这个想法距今已有约 90 年的历史了。”
So, what's new here? Well, first, there is no official research paper. There is only a highlevel description on the website, but it doesn't stop us fellow scholars from doing our research, does it? Now, hold on to your papers, fellow scholars. Come on, this time with more feel.
那么，这里有什么新东西呢？首先，没有官方的研究论文。网站上只有高层级的描述，但这并没有阻止我们同行学者进行研究，对吧？现在，握紧你们的论文，各位学者。来吧，这次要更有感觉一点。
Hold on to your paper. [music] Let that wisdom roll. Hold on to your paper. Science got your soul. [music] Hold on to your paper. The future just stars tonight. And here's the one who shows the spark that keeps our dreams alive. [music] Now we're talking actual song, by the way. So similar ideas have been explored for years and years now. There are open papers surprisingly close in spirit and we celebrate amazing science and research here.
握紧你的论文。[音乐] 让智慧流淌。握紧你的论文。科学拥有你的灵魂。[音乐] 握紧你的论文。未来今晚才刚刚开始。而这里有一位展示了让梦想保持活力的火花的人。[音乐] 顺便说一句，现在我们说的是真正的歌曲。因此，类似的思路已经被探索了很多年。有一些精神内核惊人相似的研究论文公开可用，而我们在这里庆祝卓越的科学与研究。
So I would like to say a big thank you to the scientists behind these earlier open works. So the basic idea in Jev is certainly not new but Jev also combined three ingredients. One, a new architecture that is built to make decisions instead of generating tokens. Two, it evaluates all possible answers all at once instead of one by one. They call it parallel sampling. and three, a new training method they call RLCD, where the AS confidence roughly matches how often it is right.
因此，我想向这些早期开源工作的背后科学家们致以热烈的感谢。Jev 的基本理念当然并非全新，但 Jev 还融合了三个要素。第一，一种旨在进行决策而非生成 token 的新架构。第二，它一次性评估所有可能的答案，而不是一次一个。他们称之为并行采样。第三，一种新的训练方法，他们称之为 RLCD，其中 AS 的置信度大致与其正确的频率相匹配。
So if [clears throat] it says 80% confident, it should be correct about 80 times out of 100, not always. Why is that good? Well, this way you can actually look at the confidence value and decide when to trust it or fall back to a heavier, smarter model. Once again, similar ideas have been explored for years now, but the way this one is put together, it is an incredible combination that gives us a really useful new tool.
所以如果[清嗓子]它表示有 80% 的置信度，那么它在 100 次中大约有 80 次是正确的，并非总是如此。为什么这很好？这样一来，你实际上可以查看置信度值，并决定何时信任它，或回退到更强大、更智能的模型。再次强调，类似的想法多年来已被探索过，但这种组合方式是一种令人惊叹的结合，为我们提供了一个真正有用的新工具。
Love it. And you fellow scholars are also pumping out open-source implementations of it as well. That is open science at its best. You are amazing. So, let's be level-headed and separate the facts from the hype. Is it awesome? Yes, one ingredient closer to fast local AI that we will be able to own forever. Now, is it also heavily amplified by hype and viral marketing? Yes. Is it new? Partly. Okay. I took a bit longer to make this for you.
太棒了。各位学者也在纷纷推出其开源实现。这是开源科学的最佳体现。你们太厉害了。那么，让我们保持冷静，将事实与炒作区分开来。它是否很棒？是的，它让我们离能够永久拥有的快速本地 AI 又近了一步。现在，它是否也被炒作和病毒式营销严重放大？是的。它是全新的吗？部分是的。好吧，我为制作这个内容花的时间稍微长了一点。
That's not great for views, but I do it to be more concise. Look up related papers and see how you fellow scholars use it in practice. That is a better video for you. Subscribe and hit the bell for more papers. I use Lambda to reproduce AI research papers often in minutes. It's also great to train your own models or fine-tune an existing one. Run inference or text to image or video. Easy peasy. Running a Deepseek chatbot or agent.
这对浏览量来说并不理想，但我这样做是为了更加简洁。查阅相关论文，看看各位学者在实践中如何使用它。那才是更适合你的视频。订阅并点击铃铛以获取更多论文解读。我经常使用 Lambda 在几分钟内复现 AI 研究论文。它也非常适合训练你自己的模型或对现有模型进行微调。运行推理，或文本转图像/视频。非常简单。运行 Deepseek 聊天机器人或智能体。
Super fast, super reliable. Lambda gives you powerful Nvidia GPUs to run your own experiments. I test ideas from the papers I cover and moments later, results. Love it. Seriously, try it out now at lambda.ai/papers. Ei peepers.
超级快速，超级可靠。Lambda 为你提供强大的 Nvidia GPU 来运行你自己的实验。我测试我所涵盖论文中的想法，片刻之后即可获得结果。太棒了。说真的，现在就去 lambda.ai/papers 试试看吧。爱你们哦。
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力