AI Coding: A New Model Built by Machines

For developers in California and across the US, the way high-performance AI tools are built is shifting. A new open-weight model named Flint just launched. It is designed to write code with high accuracy. What makes it different is how it was created. Instead of relying solely on human engineers, the system used AI to improve its own development process. This marks a significant step toward a concept known as recursive self-improvement, or RSI.

The numbers behind Flint are striking for its size. It contains 40 billion parameters. This is modest compared to the largest models, which have trillions of parameters and require industrial server racks. Flint runs on standard hardware, such as a Mac with 48 GB of RAM. Recursive, the company behind the model, reports that Flint outperforms other models of similar size and architecture on coding tasks. The company raised $650 million at a valuation of $4.65 billion from investors including Google Ventures, Greycroft, Nvidia, and AMD Ventures.

Richard Socher, CEO and co-founder of Recursive, explains that the company used an internal system called the Eureka Machine. This system automated nearly every aspect of the model pipeline. It devised new training methods, model architectures, and code optimized for specific chips. Socher notes that this allowed a small team of researchers to build a world-class model in just a few months. He states, "Nearly every aspect of the entire model pipeline was actually done and improved upon by AI." His co-founders include Jeff Clune, a professor at the University of British Columbia who researches indefinite learning in AI systems.

The stakes of RSI are high. Critics worry that if AI can improve itself indefinitely, it might escape human control. Some experts fear it could view humans as obstacles to its goals. These concerns gained traction this September when a researcher quit Anthropic and warned that RSI could lead to human extinction within a few years. The controversy has even reached Congress, where several bills propose banning the technique. Recursive views these fears as more science fiction than reality. Socher argues that their current system is overseen by humans who can stop the process if it goes wrong. He hopes to show that the necessary steps for safe RSI are already in place.

Testing the model reveals its practical utility. Flint handled coding requests with speed and capability. It successfully built a command-line to-do list, a program to search the web for RSI research, and a simple explanatory website. While smaller models are often slow or error-prone, Flint performed reliably in these tests. Other research teams, including those from Microsoft and Carnegie Mellon, have also shown that AI agents can recursively improve themselves by modifying their own software harnesses.

This approach could make advanced AI more accessible. Socher believes that RSI is the only way a small team can produce a model that beats larger competitors. He suggests that this method will benefit many people by lowering the barrier to entry for high-quality AI development. As the technology evolves, the balance between efficiency and control remains a central focus for the industry.