I Built a Self-Improving AI, and So Can You

我构建了一个自我改进的 AI，你也可以

These days, the frontier AI labs are all racing to build self-improving models. Some believe it’s the surest route to superintelligence—as AI improves itself in a mind-melting loop, the thinking goes, it will eventually surpass human comprehension (and perhaps even control).

如今，前沿 AI 实验室都在竞相构建自我改进模型。一些人认为这是通往超级智能最稳妥的途径——人们普遍认为，随着 AI 在一种令人难以置信的循环中不断自我完善，它最终将超越人类的理解能力（甚至可能超越人类的控制）。

That’s all well and good, but I have a newsletter to produce. I wondered if recursive self-improvement might also be useful for me. Could I use AI to train and continually improve a model that automates some of this newsletter’s busywork?

这听起来固然不错，但我还得写时事通讯。我心想，递归式自我改进对我来说是否也有用？我能否利用 AI 来训练并持续改进一个模型，从而自动化处理这份通讯中的一些繁琐工作呢？

After a week or so of experimenting, the answer appears to be a resounding—and surprising—hell yes. What’s more, dabbling with self-improving models shows a different vision for how AI might unfold—one that doesn’t center on a handful of companies that control the whole industry.

经过一周左右的实验，答案似乎是一个响亮且令人惊讶的“绝对可以”。更重要的是，涉足自我改进模型展示了 AI 发展的另一种愿景——一种不以少数几家控制整个行业的公司为核心的愿景。

I started by trying out a simple self-improving loop

我从尝试一个简单的自我改进循环开始

To get my feet wet, I experimented with training a small language model from scratch—by which I mean I dumped all the hard work on Claude’s plate.

为了小试牛刀，我尝试从零开始训练一个小语言模型——我的意思是，我把所有艰苦的工作都甩给了 Claude。

I installed AutoResearch, which helps an off-the-shelf AI model build and improve a smaller model. AutoResearch is the brainchild of Andrej Karpathy, a superstar AI researcher who helped found OpenAI, led AI work at Tesla, and recently joined Anthropic.

我安装了 AutoResearch，它可以帮助现成的 AI 模型构建并改进一个较小的模型。AutoResearch 是 AI 巨星研究员 Andrej Karpathy 的心血结晶，他曾参与创立 OpenAI，领导过特斯拉的 AI 工作，最近加入了 Anthropic。

I fired up Claude and gave it the recommended instruction: “Hi, have a look at program.md and let’s kick off a new experiment!” While Claude did the hard stuff, I provided silicon (an Nvidia DGX, a desktop “supercomputer” designed for AI experimentation), the electricity (running hot for a few days straight), and a possibly ill-advised willingness to let the model skip all the usual permission checks in order to do its thing (let him cook!)

我启动了 Claude 并给它下达了推荐指令：“嗨，看看 program.md，让我们开始一个新的实验吧！”当 Claude 处理那些困难的工作时，我负责提供算力（一台专为 AI 实验设计的桌面级“超级计算机” Nvidia DGX）、电力（连续几天全负荷运转），以及一种可能不太明智的意愿：为了让模型发挥作用，我允许它跳过所有常规权限检查（让它放手去干吧！）

I checked in on the AutoResearch project every few hours and marveled as Claude adjusted parameters and training regimes, looked at how this changed the smaller model’s output, and went on refining it further.

我每隔几个小时就会查看一下 AutoResearch 项目，并惊叹于 Claude 如何调整参数和训练方案，观察这些调整如何改变小模型的输出，并继续对其进行进一步优化。

Here’s what an early version of that smaller language model produced when I prompted it to complete the phrase “In the beginning …”

以下是该小语言模型的早期版本在我提示它补全“In the beginning…”（起初……）这句话时产生的内容：

“In the beginning of the beginning of the end of the end of the end end of end end end end end end end end beginning end end end end…”

“在起初的起初的终结的终结的终结的终结的终结的终结……”

Not so brilliant. But later models, improved autonomously by Claude, got more coherent and less prone to insane, endless repetition. It’s hardly GPT-5, but it showed a promising path toward continual improvement.

表现并不出色。但后来由 Claude 自主改进的模型变得更加连贯，也不再那么容易陷入疯狂、无休止的重复。虽然它还远不是 GPT-5，但它展示了一条通往持续改进的有前途的路径。

My journey continued with something more complex—and useful

我的旅程继续深入，转向了更复杂且更有用的东西

I already use an agent that relies on Claude to help me find noteworthy research papers, so I decided to see whether it was possible to build something that went beyond that.

我已经在通过一个依赖 Claude 的智能体来帮我寻找值得关注的研究论文，所以我决定看看是否有可能构建出超越这一功能的东西。

I turned to a tool from a startup called Prime Intellect, which uses AI to train a custom model for a specific task. I collected 100 or so previous “Elsewhere on the frontier of AI” entries—the bits and bobs of research that follow the main essay in my newsletter. Then, I created a Prime Intellect training environment and asked Claude to help me build my own model, which it dubbed Frontier_Paper_Curator, to find and summarize interesting papers.

我转向了一家名为 Prime Intellect 的初创公司提供的工具，该工具利用 AI 为特定任务训练自定义模型。我收集了大约 100 条之前的“AI 前沿其他动态”条目——即我时事通讯中主文章之后的研究碎片。然后，我创建了一个 Prime Intellect 训练环境，并请 Claude 帮我构建我自己的模型（它将其命名为 Frontier_Paper_Curator），用于查找和总结有趣的论文。

Claude found more papers and generated a bunch of synthetic data to help with training. It then tapped yet another model to assess Frontier_Paper_Curator’s output, while the training environment also improved the model with reinforcement learning.

Claude 找到了更多的论文，并生成了一堆合成数据来辅助训练。随后，它调用了另一个模型来评估 Frontier_Paper_Curator 的输出，同时训练环境也通过强化学习对模型进行了改进。

Vincent Weisser, CEO of Prime Intellect, which recently received $15 million in funding, tells me that his company aims to make recursive self-improvement accessible to everyone—not just frontier labs. The models made by frontier labs might be brilliant, but democratizing this kind of AI training could produce just as capable specialized models, he says.

最近获得 1500 万美元融资的 Prime Intellect 公司首席执行官 Vincent Weisser 告诉我，他的公司旨在让递归式自我改进技术惠及所有人，而不仅仅是前沿实验室。他说，前沿实验室制造的模型可能很出色，但将这种 AI 训练民主化，同样可以产生能力相当的专业模型。

“Give every company access to frontier training infrastructure, and the collective creativity of the market unlocks far more than any handful of labs can,” Weisser says. “We don’t want one centralized, almost godlike intelligence, we want a billion intelligences that go into all the niches that create beautiful things.”

“让每家公司都能使用前沿的训练基础设施，市场的集体创造力所能释放的价值远超少数几个实验室，”Weisser 说。“我们不想要一个集中的、近乎神一般的智能，我们想要的是十亿个智能，它们深入到各个细分领域，创造出美好的事物。”

Prime Intellect isn’t the only company that sees the future this way. Adaption, another startup, offers a tool called AutoScientist, which automates AI model training. CEO Sara Hooker says Adaption is working with several large companies that are burning through tokens and don’t have in-house AI experts.

Prime Intellect 并不是唯一一家以这种方式看待未来的公司。另一家初创公司 Adaption 提供了一种名为 AutoScientist 的工具，可以实现 AI 模型训练的自动化。首席执行官 Sara Hooker 表示，Adaption 正在与几家大型公司合作，这些公司正消耗大量 Token，且内部缺乏 AI 专家。

When Anthropic decided to block certain requests to its latest model Fable 5, it exposed the risk of relying too heavily on one frontier model. And some executives, like Palantir’s Alex Karp, have warned that using frontier labs also means handing over your own data and control over your technology.

当 Anthropic 决定阻止对其最新模型 Fable 5 的某些请求时，这暴露了过度依赖单一前沿模型的风险。一些高管，如 Palantir 的 Alex Karp，也曾警告说，使用前沿实验室的服务也意味着交出自己的数据和对技术的控制权。

The ultimate goal for recursive self-improvement is for AI to apply novel ideas to a model and come up with its own insights. The tools available to the rest of us are more limited, but still impressive. After less than a day of cooking with Prime Intellect, I was able to create a surprisingly good model for finding and summarizing research.

递归式自我改进的终极目标是让 AI 将新颖的想法应用于模型，并提出自己的见解。我们其他人可用的工具虽然有限，但依然令人印象深刻。在使用 Prime Intellect 尝试不到一天后，我就成功创建了一个在查找和总结研究方面表现出色的模型。

Here’s one example entry it created for me: 以下是它为我创建的一个示例条目：

Researchers at iFLYTEK have developed iFLYTEK-Embodied-Omni, a unified multimodal AI model that integrates vision, language, and action generation into a single framework. Unlike prior embodied agents which treat visual understanding, future state prediction, and action generation separately, their model uses shared multimodal self-attention to enable close coordination—analogous to a brain-cerebellum collaboration—between a vision-language “high-level brain” and an action-generating “low-level cerebellum.” This approach reduces error compounding and interface bottlenecks common in cascaded pipelines. By training on a large diverse dataset including human and robot-annotated embodied videos and image-text data, and using a staged training strategy, they demonstrate a general-purpose embodied agent capable of joint reasoning, prediction, and control. This contributes a novel architectural and training paradigm toward more integrated, versatile robotic AI systems.

科大讯飞的研究人员开发了 iFLYTEK-Embodied-Omni，这是一个将视觉、语言和动作生成集成到单一框架中的统一多模态 AI 模型。与以往将视觉理解、未来状态预测和动作生成分开处理的具身智能体不同，他们的模型利用共享的多模态自注意力机制，实现了视觉-语言“高级大脑”与动作生成“低级小脑”之间的紧密协作——类似于大脑与小脑的协同工作。这种方法减少了级联流水线中常见的误差累积和接口瓶颈。通过在包含人类和机器人标注的具身视频及图文数据的大型多样化数据集上进行训练，并采用分阶段训练策略，他们展示了一个能够进行联合推理、预测和控制的通用具身智能体。这为更集成、更多功能的机器人 AI 系统贡献了一种新颖的架构和训练范式。

Not bad for a first try. 对于第一次尝试来说，这还不错。