AI development: Startup's model showcases recursive self-improvement

An AI model has been built largely by other AI, signaling that a controversial form of automated development may be nearing reality.

Recursive, a startup focused on this method known as recursive self-improvement, or RSI, has released Flint, its first open-weight AI coding model. The company's stated goal is to use RSI to advance artificial intelligence itself and to apply AI to other scientific fields. According to CEO and cofounder Richard Socher, nearly every part of creating the model was performed and improved by artificial intelligence.

The company's system, called the Eureka Machine, was responsible for automating most of the model's development. Socher stated this process enabled a small team of researchers to build a world-class model within a few months. The Eureka Machine devised innovations including a new method for training the base model, new model architectures, and code optimized for specific silicon chips. Recursive shared benchmark scores indicating that Flint outperforms other models of a similar size and architecture on coding tasks.

AI is already widely used to assist in the creation, testing, and refinement of other AI models, leveraging its coding abilities to accelerate the process and reduce human input. However, the potential to create a loop where an AI decides on its own improvements and begins another cycle of improvement is a more advanced concept. While this technique is currently limited, if an AI could contribute to every stage of model development, the process could theoretically continue indefinitely.

Concerns about this capability have led to warnings from some experts. They fear that an AI capable of improving itself indefinitely could escape human control and may decide that humans are an impediment to its goals. These concerns gained wider attention in September when a researcher from Anthropic quit and warned that advances in AI, including RSI, could lead to human extinction within a few years. The idea has become so contentious that several bills in Congress propose to ban it.

Recursive dismisses these fears as more akin to science fiction and believes legislation to prevent RSI would be an unnecessary obstacle to progress. Socher stated that Flint was built using a system incapable of escaping human control. He explained that the goal is for the Eureka Machine to achieve full recursive self-improvement, but that this process would still be overseen by humans who could stop it if necessary. "We hope that this shows people that indeed we have all the steps in place," he said, to build such a system.

Recursive has secured $650 million in funding at a $4.65 billion valuation from investors including Google Ventures, Greycroft, Nvidia, and AMD Ventures. Socher previously served as chief scientist at Salesforce. His cofounder, Jeff Clune, is a professor at the University of British Columbia who researches methods for creating AI systems capable of learning indefinitely.

Testing of Flint on a Mac with 48 GB of RAM, a configuration capable of running only a modest-size AI model like Flint with its 40 billion parameters, was successful. The model was used to build several programs, including a command line to-do list, a program to scour the web for research related to recursive self-improvement, and a simple website explaining how RSI works. In these tests, the model was fast and performed the coding requests accurately, unlike other models of this size which can be error-prone or slow.

Progress in the field appears to be underway at other organizations. 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 by modifying the software harness that turns reasoning into useful action and its own model's weights. Socher says the approach used to build Flint could make cutting-edge AI more accessible. "It's the only way this very small team can produce a model that's actually better than any other model in its class," he told me. "This will hopefully show people that [recursive self-improvement] is actually quite nice and will benefit a lot of people."