Recursive's Flint: a 40 billion parameter coding model built by AI

Recursive is making a direct bet that recursive self-improvement, the idea of using AI to build AI, is a practical path to better models rather than a distant risk. CEO Richard Socher says the Eureka Machine automated most of Flint's development, from a new way of training the base model to new architectures and code optimized for specific silicon chips. That claim is the company's answer to critics who worry that letting AI improve itself could escape human control; Socher says Flint was built using a system incapable of escaping human control, and that humans could stop the process if it goes awry.

The debate has become pointed. A researcher quit Anthropic this September and warned that advances including recursive self-improvement could lead to human extinction within a few years, and several bills in Congress propose to ban RSI. Recursive frames those fears as more sci-fi than reality and argues legislation against RSI would be an unnecessary impediment.

Flint also serves as a proof point for the company's scale argument. Socher says the approach is the only way a very small team could produce a model better than any other in its class, and Recursive has raised $650 million at a $4.65 billion valuation from Google Ventures, Greycroft, Nvidia, and AMD Ventures. Outside Recursive, published work from Microsoft, Carnegie Mellon University, and the University of Virginia shows an AI coding agent can recursively improve itself by modifying its software harness and its own model weights, suggesting fully recursive systems may already be feasible.