Recursive Debuts Flint, a 40 Billion Parameter Coding Model Built by AI

The race to build artificial intelligence that can effectively build itself is moving faster than many anticipated. This week, a startup called Recursive released Flint, an open-weight coding model that claims to be a significant step toward that goal. The release has caught the attention of tech observers who have been tracking the controversial concept of recursive self-improvement, often abbreviated as RSI.

Flint is the first product from Recursive, a company dedicated to using AI to advance both the field of AI and other scientific disciplines. Richard Socher, the CEO and co-founder of Recursive, explains that the development process was largely automated. He notes that nearly every part of the model pipeline was handled and improved by AI systems rather than human engineers. This approach allowed a small team to create a competitive model in a matter of months.

The company utilized a system called the Eureka Machine to drive this automation. According to Socher, this internal tool automated most elements of the development cycle. It identified new training methods for the base model, proposed novel architectures, and optimized code for specific hardware chips. Benchmark scores shared by Recursive indicate that Flint outperforms other models of similar size and architecture in coding tasks. This suggests that the automated process yielded tangible results in performance.

The idea of using AI to build AI is not new, but the level of autonomy in this case is notable. While AI is already used to help test and refine models, creating a loop where the AI decides on improvements and executes them is a more complex endeavor. Currently, this technique has limits, but if AI can contribute to every stage of development, the process could theoretically continue indefinitely. This potential for endless self-optimization is what excites some researchers and worries others.

Concerns about RSI have gained traction in recent months. In September, a researcher who left Anthropic warned that advances in this area could lead to human extinction within a few years. This warning was echoed by other experts, turning RSI into a hot topic in policy circles. Several bills in Congress have since proposed banning the practice. Recursive views these fears as overly dramatic and believes that legislation preventing RSI is an unnecessary barrier to progress. Socher argues that Flint was built with a system that cannot escape human control. He states that while the goal is for the Eureka Machine to achieve full RSI, humans would still oversee the process and could stop it if needed.

Recursive has raised $650 million at a valuation of $4.65 billion. Investors include Google Ventures, Greycroft, Nvidia, and AMD Ventures. The leadership team includes Jeff Clune, a professor at the University of British Columbia who studies how to build AI systems capable of learning indefinitely. Socher previously served as the chief scientist at Salesforce. The company believes its approach makes cutting-edge AI more accessible to smaller teams.

In my testing on a Mac with 48 GB of RAM, Flint performed impressively. The model has 40 billion parameters, which is modest compared to the largest models that have trillions of parameters. I used Flint to build a command line to-do list, a program to search the web for research on RSI, and a simple website explaining the concept. It handled these requests quickly and accurately. In my experience, models of this size are often slow or prone to errors, but Flint did not fail on any of my prompts.

Other institutions are also exploring similar paths. A team from Microsoft, Carnegie Mellon University, and the University of Virginia recently published work showing how an AI coding agent can recursively improve itself. They demonstrated how an agent can modify its own software harness and model weights. Socher says that his team's approach shows that RSI is a practical method for improving models. He believes this will benefit many people by making high-quality AI development less resource-intensive.

For now, RSI appears to be a promising tool for enhancing AI capabilities. While there are valid debates about the long term risks of creating systems that improve themselves indefinitely, the immediate results are clear. Flint suggests that small teams can compete with larger players by leveraging AI to automate the development process. As the technology matures, the line between human built AI and AI built AI may continue to blur.