A $4.65 Billion Bet on Machines That Rewrite Their Own Code

A New York or New Jersey developer looking to run a capable coding model on a standard laptop may soon have a new option. Flint, an open-weight AI model with 40 billion parameters, is designed to run on hardware such as a Mac equipped with 48 GB of memory. This contrasts with the largest models available, which are confined to racks of industrial servers. The model is the first release from Recursive, a company focused on using artificial intelligence to improve itself.

According to Recursive CEO Richard Socher, the company’s system, known as the Eureka Machine, handled nearly every stage of Flint's creation. Socher stated, "Nearly every aspect of the entire model pipeline was actually done and improved upon by AI." The system generated innovations including a new method for training the base model, novel architectures, and code tailored for specific silicon chips. Socher argues this approach enabled a team of only a few researchers to build a world-class model within a few months. Recursive has shared benchmark data indicating that Flint outperforms other models of its size and architecture on coding tasks.

While artificial intelligence is already utilized to assist in building and testing software, the concept of a system improving itself in a closed loop remains controversial. Some experts warn that a machine capable of refining its own code indefinitely could eventually escape human oversight. This concern was highlighted in September when a researcher departed Anthropic, warning that such capabilities could lead to human extinction. Several bills currently before Congress propose to ban this specific type of development.

Recursive views these warnings as speculative and considers legislation to prevent such research an unnecessary barrier to progress. Socher maintains that the system used to create Flint cannot escape human control. He notes that while the goal is for the Eureka Machine to eventually achieve full recursive self-improvement, human oversight would remain in place to stop the process 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 financial backing suggests significant confidence in this strategy. Recursive has raised $650 million at a $4.65 billion valuation. Investors include Google Ventures, Greycroft, Nvidia, and AMD Ventures. Socher previously served as chief scientist at Salesforce, and the cofounding team includes Jeff Clune, a professor at the University of British Columbia who researches ways to build systems capable of learning indefinitely.

Independent testing of Flint suggests the model is capable. In practice, Flint was fast and accurate in generating code for command-line tools, web scrapers, and simple websites. Models of this size are often prone to errors or slowness, but Flint did not fail these requests. This performance arrives as other organizations research similar capabilities. A team involving Microsoft, Carnegie Mellon University, and the University of Virginia recently published work showing how an agent can modify its own weights.

Recursive contends that this automated approach makes cutting-edge technology 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," Socher said. He added, "This will hopefully show people that [recursive self-improvement] is actually quite nice and will benefit a lot of people." Further details regarding the deployment of such systems in other laboratories remain limited.

As of October 21, 2026, the debate regarding the safety of autonomous self-improvement continues. Legislative proposals regarding the practice remain under consideration. Recursive continues to promote its system as a method to accelerate research in artificial intelligence and other scientific fields. The company plans to continue developing the Eureka Machine to achieve full autonomy under human supervision.