Recursive Unveils Flint, a 40B-Parameter Model Built by AI to Prove RSI Viability

Recursive, a startup valued at $4.65 billion, has introduced Flint, an open-weight AI coding model designed to demonstrate the practical viability of recursive self-improvement (RSI). The company aims to use this methodology to advance both artificial intelligence and scientific research. According to CEO and co-founder Richard Socher, nearly every aspect of the model’s development pipeline was executed and refined by AI itself. This approach allows a team of just a handful of researchers to produce a world-class model within months, significantly reducing the reliance on human labor in the development cycle.

The core of this strategy is the Eureka Machine, a recursive AI system that automated most elements of Flint’s creation. Socher stated that this system devised several key innovations, including new training methods for the base model, novel architectures, and code optimized for specific silicon chips. Recursive shared benchmark scores indicating that Flint outperforms other models of similar size and architecture on coding tasks. The model contains 40 billion parameters, a scale that runs on consumer-grade hardware such as a Mac with 48 GB of RAM, unlike the largest frontier models that require racks of industrial servers.

The concept of RSI, which involves using AI to build and improve AI, has become increasingly controversial. While AI is already widely used to assist in testing and refining models, a fully recursive loop where AI decides its own improvements is a more radical step. Some experts worry that such systems could eventually escape human control, with concerns gaining prominence this September when a former Anthropic researcher warned that RSI could lead to human extinction within years. These fears have been strong enough that several bills in Congress propose banning the technology. However, Recursive views these concerns as more science fiction than immediate reality, arguing that legislation against RSI is an unnecessary barrier to progress.

Socher emphasized that Flint was built using a system that cannot escape human oversight. The goal is for the Eureka Machine to eventually achieve full recursive self-improvement, but this process would remain under human supervision to allow for intervention if necessary. He noted that the current steps are in place to build such a system safely. The company has raised $650 million 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 researches AI systems capable of indefinite learning.

In practical testing, Flint demonstrated high capability and speed. It successfully generated a command-line to-do list, a program to search for RSI-related research, and a simple website explaining the technology. These results suggest that the approach could make cutting-edge AI more accessible to smaller organizations. Socher stated that this method is the only way a small team can produce a model that surpasses others in its class. He believes this will show that recursive self-improvement is beneficial and not just a theoretical risk. While other labs like Microsoft and Carnegie Mellon have published work on recursive agents, Recursive’s Flint marks a significant step in making this technology tangible for the broader market.