AI Writes AI Code

A new open-weight AI model named Flint is challenging the status quo by proving that artificial intelligence can effectively build itself.

Flint is the inaugural product from Recursive, a startup dedicated to using recursive self-improvement to advance both AI and scientific research. The model was developed using a system called the Eureka Machine, which automated nearly every stage of the model pipeline. According to CEO and co-founder Richard Socher, this automated approach allowed a small team of researchers to create a world-class model in just a few months. The Eureka Machine introduced several innovations, including a novel method for training the base model, new architectural designs, and code optimized for specific hardware chips.

The use of AI to assist in building AI is not new, as coding capabilities allow machines to refine models and test ideas with less human input. However, Flint represents a step toward a fully recursive loop where AI decides on improvements and initiates new rounds of enhancement. While currently limited, this technique could theoretically continue indefinitely if AI contributes to every aspect of development. Recursive shared benchmark scores indicating that Flint outperforms other models of similar size and architecture on coding tasks, suggesting that this automated development process is yielding tangible results.

The concept of recursive self-improvement has sparked significant debate among experts and policymakers. Some researchers worry that allowing AI to improve itself indefinitely could lead to a loss of human control, with some fearing it might view humans as obstacles to its goals. These concerns gained prominence this September when a researcher leaving Anthropic warned that such advances could lead to human extinction within a few years, a sentiment echoed by other experts. The controversy has become so pronounced that several bills in Congress are currently proposing to ban recursive self-improvement.

Recursive views these fears as more science fiction than reality and believes that legislation preventing this approach is an unnecessary barrier to progress. Socher notes that Flint was built using a system that cannot escape human oversight. The company’s goal is for the Eureka Machine to eventually achieve full recursive self-improvement, but this process will remain under human supervision who can stop it if necessary. Socher hopes that Flint demonstrates the necessary steps are already in place to build such a system safely.

Recursive has raised $650 million at a $4.65 billion valuation 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. The company’s approach aims to make cutting-edge AI more accessible by allowing smaller teams to produce high-quality models.

Testing Flint on a Mac with 48 GB of RAM revealed its impressive capabilities for a model with 40 billion parameters. While larger models with trillions of parameters require industrial servers, Flint ran efficiently on this modest hardware. In practical use, the model proved fast and accurate, successfully building a command line to-do list, a web-scraping program, and a simple website explaining recursive self-improvement without the errors or slowness often associated with models of this size.

Other US labs are also exploring recursive self-improvement, with recent published work from Microsoft, Carnegie Mellon University, and the University of Virginia showing how an AI coding agent can improve itself by modifying its own software harness and model weights. This suggests that fully recursive systems are becoming feasible for multiple organizations.

While there are risks associated with building highly autonomous AI, the current approach appears to be a promising method for enhancing model quality. Socher believes that the technique used to build Flint will benefit many by showing that recursive self-improvement is a viable and beneficial tool for AI development, even for smaller teams.