Recursive Unveils Flint, an AI Model Developed by AI

The prospect of artificial intelligence systems that can improve themselves without human intervention has long been a subject of intense debate, and a new product from a startup named Recursive brings that debate into sharp focus. The company has released an open-weight AI coding model called Flint, which it claims was developed largely through the very process it is designed to advance: recursive self-improvement.

Recursive CEO and cofounder Richard Socher states that the company utilized its proprietary system, called the Eureka Machine, to automate the vast majority of the model's creation. "Nearly every aspect of the entire model pipeline was actually done and improved upon by AI," Socher said. He further explained that this approach enabled a company with a limited number of researchers to build what he describes as a world-class model in just a few months. The Eureka Machine is said to have devised several specific innovations, including new model architectures and a novel method for training the base model.

According to benchmark scores provided by Recursive, Flint performs better than other models of a similar size and architecture on coding tasks. This claim is consistent with firsthand testing, which found the model to be both fast and highly capable in generating functional code for various requests. The model's 40 billion parameters allow it to operate on consumer hardware, such as a laptop with sufficient memory, unlike larger models that require industrial server racks.

The development of Flint is situated within a wider and contentious conversation about the safety of recursive self-improvement, or RSI. Proponents see it as a powerful method for accelerating AI development, while some experts warn that allowing an AI to improve itself indefinitely could lead to unforeseen consequences. This concern gained prominence after a researcher at Anthropic resigned, stating that such advances could pose an existential risk to humanity. The issue has become controversial enough that several bills in Congress have proposed banning the technique.

Recursive has raised $650 million at a $4.65 billion valuation from investors including Google Ventures, Greycroft, Nvidia, and AMD Ventures. Socher, previously the chief scientist at Salesforce, asserts that the system used to build Flint was incapable of escaping human control. He stated that while the goal is for the Eureka Machine to achieve full recursive self-improvement, any such process would be overseen by humans who could halt it if necessary.

The potential of such technology is also being explored by other institutions. Researchers from Microsoft, Carnegie Mellon University, and the University of Virginia have published work demonstrating how an AI coding agent can recursively improve itself. For Recursive, the release of Flint is presented as a demonstration that its approach is practical and beneficial. "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.