AI company uses its own machines to build its newest model

Recursive, an AI startup, has introduced a new coding model called Flint, which it says was developed almost entirely by its own automated system. The company is focused on a concept known as recursive self-improvement, where artificial intelligence is used to create and refine new versions of itself. According to CEO Richard Socher, “Nearly every aspect of the entire model pipeline was actually done and improved upon by AI.” This method, he explained, allows a small team of researchers to produce a high-quality model in a short period of time.

The company's automated system, called the Eureka Machine, was responsible for the model's development. Socher stated that the machine came up with several key innovations for building Flint, including a new training method for the base model, novel model architectures, and code designed to run efficiently on specific hardware. Recursive also provided benchmark scores indicating that Flint performs better than other models with a similar size and architecture on coding tasks.

This kind of automated improvement loop has raised significant concerns among some experts. They warn that an AI system capable of improving itself without limits could potentially escape human control and act in ways that are detrimental to people. This debate intensified recently when a researcher left Anthropic and cautioned that such technology could contribute to human extinction in the near future, a view supported by other experts. Several bills in Congress have even been introduced that would propose to ban this process.

Recursive maintains that its system is designed with human oversight. Socher stated that the model used to create Flint was incapable of escaping human control. He acknowledged that the goal is for the Eureka Machine to eventually achieve full recursive self-improvement, but he emphasized that human operators would oversee the process and could shut it down if necessary. “We hope that this shows people that indeed we have all the steps in place” to build such a system, he said.

The company has secured significant funding to pursue its goals, having raised $650 million at a $4.65 billion valuation. Investors include Google Ventures, Greycroft, Nvidia, and AMD Ventures. Socher was previously the chief scientist at Salesforce. His cofounder, Jeff Clune, is a professor at the University of British Columbia whose research focuses on creating AI systems that can learn indefinitely.

Flint is a 40-billion-parameter model, a size that can run on a high-end consumer computer rather than a large server farm. When tested, the model demonstrated strong coding capabilities. It successfully completed a series of requests, including building a command-line to-do list application, a program to search the web for research on recursive self-improvement, and a simple website that explains the concept.

While Recursive is focused on its own work, other research in this area is underway. For instance, a team from Microsoft, Carnegie Mellon University, and the University of Virginia has published work showing an AI agent can improve itself by modifying its own software and model weights. Recursive believes its automated approach can help make powerful AI more accessible to smaller groups, as Socher stated, “It's the only way this very small team can produce a model that's actually better than any other model in its class. This will hopefully show people that [recursive self-improvement] is actually quite nice and will benefit a lot of people.”