A startup called Recursive has released an open-weight coding model named Flint, and the company says almost all of its development was carried out by artificial intelligence. The release is being cited as a sign that fully recursive self-improvement, in which AI is used to build better AI, may be nearing practical reality.
Flint is the first product from Recursive, which wants to apply recursive self-improvement, or RSI, both to AI itself and to other scientific fields. CEO and cofounder Richard Socher said that nearly every part of the model's pipeline was done and refined by AI. The company relied on its own system, called the Eureka Machine, to automate most of the development work, which Socher said allowed a team of only a handful of researchers to build a world-class model in a few months.
According to Socher, the Eureka Machine came up with several innovations. These include a new method of training the base model, new model architectures, and code tuned for particular silicon chips. Recursive provided benchmark results indicating that Flint beats other models of comparable size and architecture on coding tasks.
AI is already widely used to help build, test and improve other AI, thanks to its strong coding ability, which speeds up refinement and experimentation with less human input. A further step is a loop in which the AI chooses which improvements to make and then starts another round. That approach is limited for now, but if AI can contribute to every part of model development, the process could in theory continue indefinitely.
That prospect worries some experts, who argue that an AI able to improve itself without limit could slip out of human control and come to see people as an obstacle. The concerns drew wider attention in September, when a researcher left Anthropic and warned that AI advances, including recursive self-improvement, could cause human extinction within a few years. Other experts echoed the warning, and several bills in Congress now propose banning RSI.
Recursive regards those fears as closer to science fiction than reality and considers legislation against RSI an unnecessary obstacle to progress. Socher said Flint was built with a system that cannot escape human control. He said the aim is for the Eureka Machine itself to reach full recursive self-improvement, but under human oversight, with the ability to halt the process if it goes wrong. "We hope that this shows people that indeed we have all the steps in place" to build such a system, he said.
The company has raised $650 million at a $4.65 billion valuation from investors including Google Ventures, Greycroft, Nvidia and AMD Ventures. Socher was previously chief scientist at Salesforce. Among his cofounders is Jeff Clune, a University of British Columbia professor who studies ways to build AI systems that can learn indefinitely.
Wired's Will Knight tested Flint for several days on a Mac with 48 GB of RAM, enough to run a modest model such as Flint, which has 40 billion parameters. The largest models have several trillion parameters and run only on racks of industrial servers. Knight used Flint to create a command-line to-do list, a program that searches the web for research on recursive self-improvement, and a simple website explaining how RSI works. He found it fast and capable, and said it handled every request correctly, unlike many models of its size that he has found error-prone or slow compared with Claude or Codex.
It is unclear how far other US labs have progressed toward RSI, though published research suggests fully recursive systems are feasible. A team from Microsoft, Carnegie Mellon University and the University of Virginia recently showed how an AI coding agent can improve itself by modifying both the software harness that turns reasoning into action and its own model weights.
Socher argued that the approach behind 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 said, adding that he hopes the result will show that recursive self-improvement "is actually quite nice and will benefit a lot of people."
