This Startup Is All In on Recursive Self-Improvement. Its Latest Model Is Scary Good

For the past few days I’ve been testing an impressive open-weight AI coding model called Flint. It suggests that fully recursive self-improvement, or RSI—an exciting and controversial approach that uses AI to build AI—is getting closer.

在过去的几天里，我一直在测试一款令人印象深刻的开源权重 AI 编程模型，名为 Flint。它表明，完全递归自我改进（RSI）——一种利用 AI 构建 AI 的令人兴奋且充满争议的方法——正变得越来越近。

Flint is the first offering from Recursive, a company that aims to use RSI to advance both AI itself and other areas of science using AI. “Nearly every aspect of the entire model pipeline was actually done and improved upon by AI,” Recursive CEO and cofounder Richard Socher told me.

Flint 是 Recursive 公司的首款产品，该公司旨在利用 RSI 来推动 AI 本身的发展，并利用 AI 促进其他科学领域的研究。Recursive 的首席执行官兼联合创始人理查德·索彻（Richard Socher）告诉我：“整个模型流水线的几乎每一个环节实际上都是由 AI 完成并改进的。”

Socher says his company used its recursive AI system, dubbed the Eureka Machine, to automate most elements of the model’s development. This enabled a company with a handful of researchers to build a world-class model in just a few months, he says.

索彻表示，他的公司利用其被称为“尤里卡机器”（Eureka Machine）的递归 AI 系统，实现了模型开发中大部分要素的自动化。他说，这使得一家仅有少数研究人员的公司能够在短短几个月内构建出世界级的模型。

Socher told me that the Eureka Machine devised a number of innovations, including a new way of training the base model, new model architectures, and code optimized for specific silicon chips. Recursive shared benchmark scores that suggest Flint outperforms other models of similar size and architecture on coding tasks.

索彻告诉我，尤里卡机器设计了多项创新，包括一种训练基础模型的新方法、新的模型架构以及针对特定硅芯片优化的代码。Recursive 分享的基准测试分数表明，Flint 在编程任务上的表现优于其他同等规模和架构的模型。

AI is already widely used to help build, test, and improve AI thanks to its amazing coding skills. This can speed up the process of refining a model and testing new ideas without the need for as much input from humans.

得益于其出色的编程能力，AI 已被广泛用于帮助构建、测试和改进 AI。这可以加快模型优化和测试新想法的过程，而无需过多的人工参与。

It is also possible to create a loop through which AI decides what improvements to make and then begins another round of improvement. For the moment this technique is limited, but if AI can contribute to every aspect of model development, then the process could theoretically continue indefinitely.

创建一个循环也是可能的，通过这个循环，AI 可以决定进行哪些改进，然后开始新一轮的改进。目前这项技术还很有限，但如果 AI 能够参与模型开发的各个方面，那么从理论上讲，这个过程可以无限期地持续下去。

Some experts worry that letting AI improve itself in this manner could ultimately lead to disaster. These doomsayers claim that AI capable of improving indefinitely could escape human control and might decide that humans are an impediment to its goals.

一些专家担心，以这种方式让 AI 自我改进最终可能导致灾难。这些末日论者声称，能够无限期自我改进的 AI 可能会逃脱人类的控制，并可能认为人类是其实现目标的障碍。

Those concerns gained wider notice this September when a researcher quit Anthropic and warned that advances in AI, including recursive self-improvement, could lead to human extinction within a few years—a concern then echoed by a number of other experts. The idea of RSI has become so controversial that several bills in Congress propose to ban it.

今年 9 月，当一名研究人员从 Anthropic 离职并警告称包括递归自我改进在内的 AI 进步可能在几年内导致人类灭绝时，这些担忧引起了更广泛的关注——随后许多其他专家也表达了同样的担忧。RSI 的概念已经变得极具争议，以至于国会中出现了几项提议禁止它的法案。

Recursive sees these fears as more sci-fi than reality and believes legislation to prevent RSI is an unnecessary impediment to progress.

Recursive 公司认为这些担忧更多是科幻而非现实，并认为旨在阻止 RSI 的立法是对进步不必要的阻碍。

Socher says Flint was built using a system incapable of escaping human control. He says the goal is for the Eureka Machine itself to achieve full recursive self-improvement, but this would still be overseen by humans who could stop the process if it goes awry. “We hope that this shows people that indeed we have all the steps in place” to build such a system, he says.

索彻表示，Flint 是使用一个无法逃脱人类控制的系统构建的。他说，目标是让尤里卡机器本身实现完全的递归自我改进，但这仍将由人类监督，如果过程出现偏差，人类可以随时叫停。“我们希望这能向人们展示，我们确实已经具备了构建此类系统的所有步骤，”他说。

Recursive has raised $650 million at a $4.65 billion valuation from Google Ventures, Greycroft, Nvidia, and AMD Ventures. Socher was previously the chief scientist at Salesforce, and his cofounders include Jeff Clune, a professor at the University of British Columbia who researches ways of building AI systems capable of learning indefinitely.

Recursive 已从 Google Ventures、Greycroft、Nvidia 和 AMD Ventures 筹集了 6.5 亿美元，估值达到 46.5 亿美元。索彻此前曾担任 Salesforce 的首席科学家，他的联合创始人包括不列颠哥伦比亚大学教授杰夫·克伦（Jeff Clune），他专门研究如何构建能够无限学习的 AI 系统。

I must say I am impressed with Flint. I tested it on a Mac with 48 GB of RAM—a healthy amount but only enough to run a modest-size AI model like Flint, which has 40 billion parameters. The biggest models out there have several trillions of parameters, and they can only run on racks of industrial servers.

我必须说，Flint 给我留下了深刻的印象。我在一台拥有 48GB 内存的 Mac 上测试了它——这虽然是不错的配置，但也仅够运行像 Flint 这样拥有 400 亿参数的中等规模 AI 模型。目前最大的模型拥有数万亿个参数，它们只能在工业级服务器机架上运行。

I used Flint to build a bunch of programs and found it to be both fast and very capable. In my experience, models of this size can be annoyingly error-prone or slow compared to Claude or Codex, but Flint nailed my coding requests every time. I used it to create a command line to-do list, then tried having it build a program that scours the web for interesting research related to recursive self-improvement. I also had Flint whip up a simple website that explains how RSI works. You can see that here.

我使用 Flint 构建了一系列程序，发现它既快速又非常强大。根据我的经验，与 Claude 或 Codex 相比，这种规模的模型往往容易出错或运行缓慢，但 Flint 每次都能完美完成我的编程请求。我用它创建了一个命令行待办事项列表，然后尝试让它构建一个程序，在网络上搜索与递归自我改进相关的有趣研究。我还让 Flint 快速制作了一个简单的网站来解释 RSI 的工作原理。你可以在这里看到它。

It’s unclear how much progress other US labs have made chasing recursive self-improvement, but some published work suggests fully recursive systems are already feasible. A team from Microsoft, Carnegie Mellon University, and the University of Virginia recently published work, for instance, showing how an AI coding agent can recursively improve itself by modifying the software harness that turns reasoning into useful action and its own model’s weights.

目前尚不清楚美国其他实验室在追求递归自我改进方面取得了多大进展，但一些已发表的研究表明，完全递归的系统已经是可行的。例如，来自微软、卡内基梅隆大学和弗吉尼亚大学的一个团队最近发表的研究显示，AI 编程代理可以通过修改将推理转化为有用行动的软件工具以及模型自身的权重，从而实现递归自我改进。

Perhaps there are risks to full RSI if you are close to building a godlike intelligence, but for the time being it seems like a very promising way to improve your own AI models. This is something I’ve dabbled with in previous editions of AI Lab.

如果你即将构建出一种神一般的智能，那么完全的 RSI 可能确实存在风险，但就目前而言，这似乎是改进自身 AI 模型的一种非常有前景的方法。这也是我在之前几期的《AI 实验室》中曾浅尝辄止的话题。

Indeed, Socher says the approach used to build Flint could make cutting-edge AI more accessible. “It’s the only way this very small team can produce a model that’s actually better than any other model in its class,” he told me. “This will hopefully show people that [recursive self-improvement] is actually quite nice and will benefit a lot of people.”

事实上，索彻表示，用于构建 Flint 的方法可以使尖端 AI 变得更加普及。“这是我们这个非常小的团队能够生产出在同类模型中表现最好的模型的唯一途径，”他告诉我。“希望这能向人们展示，（递归自我改进）实际上非常棒，并将造福许多人。”

This is an edition of Will Knight’s AI Lab newsletter. Read previous newsletters here.

这是威尔·奈特（Will Knight）的《AI 实验室》通讯的一期。点击此处阅读往期通讯。