A New Model Is Built Mostly by AI, and It Works

If you write software, a new open-weight model called Flint is designed to help you build programs quickly and reliably, and it can run on a laptop rather than a server farm. The tool is the first product from Recursive, a startup whose stated purpose is to use artificial intelligence to improve itself. The company says nearly every part of creating the model, from training methods to the code itself, was handled by its own AI systems. In practical terms, this means a small group of researchers built a model the company says performs at a level comparable to much larger systems in a matter of months.

The path to this kind of automation has been building for some time. For years, AI has been used to write, test, and refine the code that makes other AI models, which speeds up the process of trying new ideas without needing a person to write every line. Researchers have long discussed taking this a step further by closing the loop, letting a model decide on its own what improvements to make and then starting another round of changes. For now, that cycle is limited, but if a system could contribute to every stage of its own creation, the process could in theory continue indefinitely.

Some experts warn that a machine capable of unlimited self-improvement could drift out of human control. These concerns were amplified in September, when a researcher left Anthropic and cautioned that such advances could threaten humanity within a few years, a view echoed by others. The debate has reached Capitol Hill, where several bills in Congress propose to ban the practice.

Recursive, which was founded by Richard Socher, formerly chief scientist at Salesforce, and Jeff Clune, a professor at the University of British Columbia who studies how to build systems that learn indefinitely, views those fears as speculative. Socher says Flint was built with a system that cannot escape human control, and that any future system capable of full self-improvement would remain under human oversight. "We hope that this shows people that indeed we have all the steps in place" to build such a system, he says. He argues legislation to block the work is an unnecessary obstacle to progress.

The company has raised $650 million at a $4.65 billion valuation from Google Ventures, Greycroft, Nvidia, and AMD Ventures. In testing, Flint, which has 40 billion parameters, was run on a Mac with 48 GB of RAM, enough for a modest-size model, though the largest systems in existence require racks of industrial servers. The model was used to build a command line to-do list, a program that searches the web for research on recursive self-improvement, and a simple website explaining the concept. The model was fast and produced no errors on those requests, in contrast to other models of that size, which can be slow or mistake-prone.

It remains unclear how much progress other labs in the United States have made on this work, though some published research suggests such systems are already feasible. A team from Microsoft, Carnegie Mellon University, and the University of Virginia recently published work showing how an AI coding agent can improve itself by modifying the software harness that turns reasoning into action and its own model's weights.

Socher says the method used to build Flint could make advanced AI more widely available. "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 said. "This will hopefully show people that [recursive self-improvement] is actually quite nice and will benefit a lot of people." The company's stated aim is for its system, called the Eureka Machine, to eventually achieve full recursive self-improvement, under human supervision.

Whether the approach can be scaled to a system capable of unlimited change remains an open question, and the debate over the risks is ongoing. For now, the company says its method can build a model that outperforms other models of similar size and architecture on coding tasks. The company says it used its system to devise a new way of training the base model, new model architectures, and code optimized for specific silicon chips. Recursive has not said when it plans to release a next version.