Recursive's New Model Demonstrates AI-Driven Code Generation

Recursive's new coding model, Flint, suggests that the development of AI systems capable of improving their own design is nearing reality.

Flint is a 40-billion-parameter open-weight model. Recursive CEO Richard Socher said that nearly every stage of the model's development was performed and refined by AI. The company's internal system, called the Eureka Machine, automated most of the work required to build the model. Socher said this process allowed a small team of researchers to create a high-quality model in a few months.

Recursive provided benchmark scores indicating Flint outperforms other models with similar size and architecture on coding tasks. Socher said the Eureka Machine developed specific innovations, including a new method for training the base model, new architectures for models, and code optimized for specific silicon chips. The Eureka Machine is designed to automate the creation and refinement of AI models, which can speed up development and reduce the need for human input. While AI is already used to write code for other AI, a system that decides on its own improvements could theoretically continue this process indefinitely without human intervention.

The concept of recursive self-improvement has drawn criticism from some experts who are concerned that a system improving itself indefinitely could escape human control. A researcher who left Anthropic in September warned that advances in AI, including this type of self-improvement, could lead to human extinction within a few years, and several bills in Congress propose to ban the practice. Recursive views these concerns as science fiction and believes that laws preventing it would block progress.

Socher said Flint was built with a system incapable of escaping human control. The company's goal is for the Eureka Machine to achieve full recursive self-improvement, but this would be overseen by humans who could stop the process. “We hope that this shows people that indeed we have all the steps in place” to build such a system, he said.

Recursive has raised $650 million in funding at a $4.65 billion valuation from investors including Google Ventures, Greycroft, Nvidia, and AMD Ventures. Socher previously served as chief scientist at Salesforce, and his co-founders include Jeff Clune, a professor at the University of British Columbia who studies how to build AI systems that can learn indefinitely. The company believes its approach can make advanced AI more accessible to more people. “It's the only way this very small team can produce a model that's actually better than any other model in its class,” Socher said.

Flint can run on a consumer device with 48 GB of RAM, as its 40-billion-parameter size requires only modest hardware compared to larger models that require industrial servers. During testing on a Mac, the model was used to create a command-line to-do list, a program that searches for research on recursive self-improvement, and a website explaining how the process works.

It is unclear how much progress other U.S. labs have made in this area, but some published research suggests that fully recursive systems are feasible. A team from Microsoft, Carnegie Mellon University, and the University of Virginia recently published work showing how an AI coding agent can recursively improve itself by modifying its own model's weights.