A small team used AI to build a coding model that rivals giants

A new AI model is proving that small teams can punch above their weight when they let algorithms build the algorithms. The system, known as Flint, has generated significant buzz in the tech community for its impressive coding abilities and its development process.

Flint is the debut product from a startup called Recursive, which is betting heavily on a concept known as recursive self-improvement. This approach involves using AI to design, train, and refine other AI systems. Richard Socher, the company's CEO and cofounder, explained that nearly every part of the model's creation pipeline was handled by their internal system, the Eureka Machine. This automated process allowed a small group of researchers to produce a world-class model in a matter of months. The Eureka Machine reportedly devised new training methods, novel architectures, and code optimized for specific hardware. Benchmark scores shared by the company indicate that Flint outperforms other models of similar size on coding tasks.

The idea of AI improving itself has long been a point of debate among experts. While AI is already used to assist in building and testing new models, creating a loop where the system decides its own improvements is a more radical step. Some worry that if an AI can improve indefinitely, it might eventually escape human control. These concerns gained traction earlier this month when a researcher from Anthropic warned that such advances could lead to human extinction within a few years. The controversy is now influencing policy, with several bills in Congress proposing to ban recursive self-improvement entirely. Recursive views these fears as overly dramatic and believes that legislative restrictions would unnecessarily hinder progress. Socher insists that the system used to build Flint cannot escape human oversight, and he hopes to eventually apply full recursive self-improvement to the Eureka Machine itself under human supervision.

The company has secured significant funding to pursue this vision, raising $650 million at a valuation of $4.65 billion. Investors include Google Ventures, Greycroft, Nvidia, and AMD Ventures. Socher, who previously served as chief scientist at Salesforce, is joined by cofounder Jeff Clune, a professor at the University of British Columbia who studies AI systems capable of learning indefinitely.

In practical testing, Flint performs well even on consumer hardware. The model contains 40 billion parameters, which is modest compared to the trillions found in the largest industrial models. It runs efficiently on a Mac with 48 GB of RAM. During evaluation, the model demonstrated speed and accuracy, successfully completing coding requests without the common errors seen in similar-sized competitors. Users have used it to create command line tools, web scrapers, and informational websites.

Other research institutions are also exploring similar techniques. A recent publication from Microsoft, Carnegie Mellon University, and the University of Virginia showed how an AI agent can recursively improve itself by modifying its own software harness and model weights. This suggests that fully recursive systems are becoming technically feasible. Socher believes that the methods used to create Flint will make cutting-edge AI more accessible to smaller organizations. He argues that this approach is the only way a small team can produce a model that beats larger competitors in its class, ultimately benefiting a wider range of users.

As the technology advances, the balance between speed and safety will remain a central focus. For now, the results suggest that recursive self-improvement is a promising pathway for developing efficient and capable AI models without requiring massive industrial infrastructure.