Recursive Unveils Flint, a 40 Billion Parameter Model Built to Fuel Its Own Evolution

Artificial intelligence systems are increasingly tasked with writing code, but a new startup is accelerating the shift from assisting developers to autonomously building its own successors. Recursive, a company dedicated to advancing science and AI through recursive self-improvement, has released Flint, an open-weight coding model that it claims was largely constructed by its own automated systems.

According to the company, the process of creating Flint involved a system referred to as the Eureka Machine, which automated the majority of the model's development pipeline. CEO and cofounder Richard Socher stated that nearly every aspect of the pipeline was performed and refined by AI. He noted that the Eureka Machine produced specific innovations, such as a novel method for training base models, new architectures for the models, and code optimized for particular silicon chips. Recursive reports that Flint achieves benchmark scores superior to other models of similar size and configuration on coding tasks.

The company's approach relies on a cycle wherein AI assists in creating and testing AI, which theoretically could continue without human intervention. However, Recursive maintains that human oversight remains a core component of its operation. Socher explained that Flint was built using a system that cannot escape human control. He added that while the goal is for the Eureka Machine to achieve full recursive self-improvement, the process would still be overseen by humans who could terminate the operation if necessary.

Recursive has raised $650 million at a $4.65 billion valuation, with backing from Google Ventures, Greycroft, Nvidia, and AMD Ventures. The company's cofounders include Richard Socher, formerly the chief scientist at Salesforce, and Jeff Clune, a professor at the University of British Columbia who researches methods for creating AI systems capable of learning indefinitely. Flint, which contains 40 billion parameters, is designed to operate on consumer hardware, such as a Mac equipped with 48 GB of RAM, unlike larger models that require industrial server racks.

The development of systems capable of recursively improving themselves remains a subject of debate. Some experts warn that such technology could ultimately escape human control, a concern that gained attention recently when a researcher leaving Anthropic cautioned that recursive self-improvement could lead to human extinction within a few years. This view has prompted legislative efforts to ban the technology, though Recursive considers such measures an unnecessary impediment to progress. Socher expressed hope that Flint demonstrates the necessity of such technology, stating that it is the only way a small team can produce a model that outperforms others in its class.