A Small Team Built a Powerful Coding AI by Letting AI Do the Heavy Lifting

A new open-weight AI model named Flint is drawing attention for its strong performance in coding tasks, suggesting that the concept of recursive self-improvement (RSI) is moving from theoretical debate to practical application. Developed by a startup called Recursive, the model has been tested by staff writers and found to be highly capable, raising questions about how much of the AI development process can be handed over to AI itself. The release comes at a time when the idea of AI building AI is both a source of excitement and intense scrutiny in the technology sector.

Recursive, the company behind Flint, is a relatively small operation with a handful of researchers. Its CEO and cofounder, Richard Socher, stated that nearly every aspect of the model’s pipeline was executed and refined by AI. This approach allowed the team to produce a world-class model in a matter of months, a timeline that would typically require a much larger workforce and significantly more resources if done by human hands alone. The core of this effort is a system called the Eureka Machine, which automated most elements of the development process.

The Eureka Machine did not just assist with coding; it devised specific innovations. According to Socher, the system created new methods for training the base model, designed new architectures, and optimized code for particular silicon chips. Recursive has shared benchmark scores indicating that Flint outperforms other models of similar size and architecture on coding tasks. This marks a shift from using AI merely as a tool for testing to using it as a primary driver of architectural and training decisions.

Recursive self-improvement involves creating a loop where AI decides what improvements to make and then initiates another round of refinement. While current implementations are limited, the theoretical possibility of an indefinite improvement cycle has sparked debate. Some experts worry that allowing AI to improve itself without significant human intervention could lead to unintended consequences, including a loss of human control. These concerns were amplified earlier this year when a researcher from Anthropic resigned to warn that advances in RSI could lead to human extinction within a few years, a view echoed by other experts in the field.

The controversy has reached the political realm, with several bills in Congress proposing to ban recursive self-improvement. However, Recursive views these fears as overly dramatic, comparing them to science fiction rather than immediate reality. The company argues that legislation to prevent RSI is an unnecessary obstacle to progress. Socher emphasized that Flint was built using a system that cannot escape human control. The goal is for the Eureka Machine to eventually achieve full RSI, but this would remain under human oversight, allowing operators to stop the process if it deviates from the intended path. "We hope that this shows people that indeed we have all the steps in place," Socher said.

Financially, Recursive has secured substantial backing, raising $650 million at a valuation of $4.65 billion. Investors include Google Ventures, Greycroft, Nvidia, and AMD Ventures. Socher previously served as chief scientist at Salesforce, and the company’s cofounders include Jeff Clune, a professor at the University of British Columbia who researches AI systems capable of indefinite learning. This heavy investment underscores the belief among major tech players that RSI is a critical next step in AI development.

Independent testing of Flint on a Mac with 48 GB of RAM revealed a model that is both fast and reliable. With 40 billion parameters, Flint is modest in size compared to the largest models, which run on industrial server racks. In practical use, the model successfully created a command-line to-do list, a program to search the web for research on RSI, and a simple website explaining the concept. Unlike some other models of similar size that can be slow or error-prone, Flint handled these requests accurately, suggesting that smaller models can be highly effective when optimized through RSI.

Other US labs are also exploring RSI, with recent work from Microsoft, Carnegie Mellon University, and the University of Virginia showing how AI coding agents can recursively improve themselves by modifying their own software harness and weights. While the risks of full RSI remain a topic of discussion, the current evidence suggests it is a promising method for improving AI models. Socher believes this approach will make cutting-edge AI more accessible, stating, "It's the only way this very small team can produce a model that's actually better than any other model in its class."

The development of Flint highlights a broader trend in the AI industry: the increasing reliance on AI to build AI. As the technology matures, the line between human direction and machine execution continues to blur. Whether this leads to unprecedented efficiency or unforeseen challenges remains to be seen, but the success of Flint suggests that the era of AI-driven AI development is well underway.