40 Billion Parameters: The New Benchmark for AI Self-Improvement

A new open-weight coding model named Flint has emerged, suggesting that the contentious concept of recursive self-improvement is moving from theory to practice. Released by a startup called Recursive, the model is described by its creators as a product of an automated system that used AI to build AI. This development lands at a moment of heightened scrutiny, following recent warnings from prominent researchers that such capabilities could accelerate technological risks.

The numbers behind Flint are specific and modest in scale relative to the industry giants. The model contains 40 billion parameters, a size that allows it to run on consumer hardware, such as a Mac with 48 GB of RAM. This stands in stark contrast to the largest models currently available, which possess several trillions of parameters and require industrial server racks to operate. Despite its smaller footprint, Recursive shared benchmark scores indicating that Flint outperforms other models of similar size and architecture on coding tasks.

The company behind Flint, Recursive, is a small team that raised $650 million at a $4.65 billion valuation. Investors include Google Ventures, Greycroft, Nvidia, and AMD Ventures. The startup is led by Richard Socher, a former chief scientist at Salesforce, and co-founded by Jeff Clune, a professor at the University of British Columbia who specializes in systems capable of indefinite learning. Socher asserts that nearly every aspect of the model's pipeline was executed and improved by AI, a process he claims allowed a handful of researchers to create a world-class model in just a few months.

The core mechanism driving this efficiency is a system dubbed the Eureka Machine. Socher explains that this recursive AI system automated most elements of development, devising innovations such as new training methods, novel model architectures, and code optimized for specific silicon chips. The goal is for the Eureka Machine to eventually achieve full recursive self-improvement, though Socher emphasizes that humans will retain the ability to oversee and stop the process if it goes awry. "We hope that this shows people that indeed we have all the steps in place," Socher said, addressing concerns about control.

The stakes of this technology are high and have become a point of political and scientific contention. Critics argue that AI capable of improving itself indefinitely could escape human control, potentially viewing humans as impediments to its goals. These fears gained traction this September when a researcher left Anthropic to warn that advances in AI could lead to human extinction within a few years. The controversy has reached the legislative level, with several bills in Congress proposing to ban recursive self-improvement, a move Recursive views as an unnecessary impediment to progress.

Independent testing of Flint reveals a tool that is both fast and capable. In recent days, the model successfully executed complex coding requests without error, a performance that contrasts with other models of similar size that can be slow or prone to mistakes. Tasks included building a command-line to-do list, creating a program to scour the web for research on recursive self-improvement, and generating a simple website explaining how the technology works. These results suggest that the model is reliable for practical applications, even if it is not yet a general-purpose intelligence.

The broader implications for the industry are significant. While it is unclear how much progress other U.S. labs have made, recent published work from teams at Microsoft, Carnegie Mellon University, and the University of Virginia indicates that fully recursive systems are feasible. These studies show AI coding agents can improve themselves by modifying their own software harnesses and model weights. Socher believes the approach used for Flint could make cutting-edge AI more accessible to smaller teams. "This will hopefully show people that [recursive self-improvement] is actually quite nice and will benefit a lot of people," he stated.

For now, recursive self-improvement remains a promising but debated method for enhancing AI models. While the risks of building a godlike intelligence are real, the immediate benefits of efficiency and capability appear tangible. As startups like Recursive push forward, the gap between theoretical potential and practical application continues to narrow, forcing the industry and regulators to confront the reality of AI that helps build its own successors.