40B-Parameter Flint Model Demonstrates Viable Recursive AI Coding

The debate over how much autonomy artificial intelligence should hold in its own creation has intensified, with a new open-weight model suggesting the technology is closer to full independence than many predicted. Flint, developed by the startup Recursive, represents a significant step toward recursive self-improvement, a controversial method where AI systems are used to build and refine subsequent AI generations.

Richard Socher, CEO and co-founder of Recursive, stated that nearly every aspect of the model pipeline was executed and enhanced by AI. The company utilized a system named the Eureka Machine to automate the majority of Flint’s development, allowing a small team of researchers to produce a world-class model in just a few months. Socher noted that the Eureka Machine independently devised several innovations, including new training methods for the base model, novel architectures, and code optimized for specific silicon chips. Benchmark scores released by Recursive indicate that Flint outperforms other models of similar size and architecture on coding tasks.

While AI is already common in testing and refining models, the concept of a fully recursive loop, where AI decides its own improvements and initiates the next round of development, remains largely theoretical. However, experts have raised concerns that such a process could lead to unintended consequences, including the potential for AI to escape human control if it determines humans are an impediment to its goals. These fears gained traction in September when a researcher from Anthropic warned that recursive self-improvement could lead to human extinction within a few years, a sentiment echoed by other experts and now reflected in several bills proposed in Congress.

Recursive views these legislative efforts as unnecessary obstacles to progress. Socher emphasized that Flint was built using a system designed to remain under human control, with the long-term goal of the Eureka Machine achieving full recursive self-improvement under human supervision. He expressed hope that the project demonstrates the necessary steps are in place to build such a system safely. The company has raised $650 million at a $4.65 billion valuation from investors including Google Ventures, Greycroft, Nvidia, and AMD Ventures. Socher, formerly the chief scientist at Salesforce, co-founded the company with Jeff Clune, a University of British Columbia professor who researches systems capable of indefinite learning.

Independent testing of Flint, which contains 40 billion parameters, confirmed its capability on a Mac with 48 GB of RAM. While models of this size are often considered modest compared to the trillions of parameters in industrial-scale systems, Flint performed without error on complex coding requests. It successfully built a command-line to-do list, a web-scraping program for research on recursive self-improvement, and a simple explanatory website. Other recent work from Microsoft, Carnegie Mellon University, and the University of Virginia suggests that fully recursive systems are already feasible, as AI coding agents have shown the ability to improve themselves by modifying their own software harnesses and model weights.

Socher argues that the approach used to build Flint could make cutting-edge AI more accessible to smaller teams. He described the method as the only way his small team could produce a model superior to others in its class, hoping to demonstrate that recursive self-improvement is a beneficial tool rather than a dangerous experiment. For now, the technology appears to be a promising avenue for enhancing AI models, even as the broader scientific community grapples with the long-term implications of machines building their own successors.