Small Team, Big Model, AI Builds Itself

A startup is claiming a major breakthrough in artificial intelligence by using its own technology to build its products, a process that has long been debated by experts. Recursive, a new company, announced Flint, an open-weight AI model designed for coding tasks. The creation of Flint, according to the company, was largely handled by AI itself, a concept known as recursive self-improvement.

This approach, where AI systems are used to build other AI systems, is a contentious topic in the field. While it promises faster development, some researchers have expressed concern over the potential for machines to become uncontrollable. These fears were amplified in September when a researcher from Anthropic left the company, warning that recursive self-improvement could pose a threat to humanity, and several congressional bills have since been proposed to regulate the practice.

Recursive, however, views these concerns as premature. CEO and cofounder Richard Socher stated that the system used to create Flint was built under strict human oversight, ensuring it could not operate independently. The company's goal is to eventually achieve full recursive self-improvement, but Socher clarified that such a process would always be supervised by humans who could halt it 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's system, called the Eureka Machine, reportedly automated most elements of developing the Flint model, from designing new architectures and training methods to optimizing code for specific chips. According to Socher, this automation allowed a small team of researchers to produce a competitive model in just a few months. “Nearly every aspect of the entire model pipeline was actually done and improved upon by AI,” he told a reporter. Benchmark scores provided by the company suggest Flint outperforms other models of a similar size on coding tasks.

The company has secured $650 million in funding at a $4.65 billion valuation from investors including Google Ventures, 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 how to build AI systems capable of learning indefinitely.

Testing of Flint has shown promising results. The 40-billion-parameter model was run on a personal computer with 48 GB of RAM and was used to generate several programs, including a command-line to-do list, a program to search for research on recursive self-improvement, and a website explaining the concept. The model was described as both fast and highly capable, successfully completing every coding request without error.

Recursive is not the only organization pursuing this path. Researchers from Microsoft, Carnegie Mellon University, and the University of Virginia have published work detailing how an AI agent can modify its own code and model weights to improve itself. While the potential risks of creating a super-intelligent system remain a topic of debate, Socher believes the approach can make advanced 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 stated.