Recursive Launches Flint, an AI Model That Writes Its Own Code

Recursive, a startup founded by former Salesforce chief scientist Richard Socher, has released Flint, an open-weight artificial intelligence model designed to write code. The release marks a significant step toward recursive self-improvement, a method where AI systems are used to build and refine other AI systems. This approach allows smaller teams to develop high-performance models without the massive resources typically required by major tech firms.

Flint is the company's first product and features 40 billion parameters. It was developed over a few months by a small group of researchers using an internal system called the Eureka Machine. Socher stated that nearly every part of the model pipeline was handled and improved by AI. This included creating new training methods for the base model, designing new architectures, and writing code optimized for specific hardware chips.

The company shared benchmark scores showing that Flint performs better than other models of similar size and architecture on coding tasks. The Eureka Machine automated most of the development process, allowing the team to iterate quickly. This efficiency is a key advantage for startups that lack the vast compute resources of larger competitors.

Recursive self-improvement is a controversial topic in the AI community. Some experts worry that allowing AI to improve itself indefinitely could lead to systems that are difficult for humans to control. These concerns gained attention this September when a researcher from Anthropic resigned and warned that such advances could lead to human extinction within a few years. Several bills in Congress have proposed bans on this technology in response to these fears.

Socher views these concerns as more science fiction than immediate reality. He explained that Flint was built using a system that cannot escape human oversight. The goal is for the Eureka Machine to eventually achieve full recursive self-improvement, but humans will still monitor the process and can stop it if necessary. Socher said, “We hope that this shows people that indeed we have all the steps in place.”

Recursive has raised $650 million at a $4.65 billion valuation from investors including Google Ventures, Greycroft, Nvidia, and AMD Ventures. Socher's cofounders include Jeff Clune, a professor at the University of British Columbia who researches AI systems capable of learning indefinitely. The funding supports the company's mission to use AI to advance both artificial intelligence and other scientific fields.

Testing Flint on a Mac with 48 GB of RAM revealed that it is fast and capable. While 40 billion parameters is modest compared to the trillions found in the largest models, Flint performed well in practical tests. It successfully created a command line to-do list, a web scraper for research papers, and a simple website explaining recursive self-improvement. In these tests, it avoided the common errors and slowness associated with models of this size.

Other US labs are also exploring recursive systems. A recent publication from teams at Microsoft, Carnegie Mellon University, and the University of Virginia demonstrated how an AI coding agent can improve itself by modifying its own software and weights. This suggests that fully recursive systems are technically feasible and may become more common in the near future.

Socher believes that the approach used to build Flint will make cutting-edge AI more accessible to smaller organizations. He noted that this is the only way a small team can produce a model that outperforms its peers. The release of Flint offers a practical example of how AI can be used to build better AI, potentially benefiting a wider range of developers and researchers.

The model is available for public testing, allowing users to evaluate its capabilities firsthand. As the debate over recursive self-improvement continues, Flint stands as a tangible example of the technology's current potential. It demonstrates that AI-assisted development can yield high-quality results even with limited human intervention.