Recursive AI: A $4.65B Startup Claims Success With Self-Improving Model

A startup called Recursive has released Flint, an open-weight AI coding model that suggests the controversial concept of recursive self-improvement is rapidly moving from theory to practice. The model, which features 40 billion parameters, was developed in just a few months by a small team of researchers who relied heavily on an internal AI system to automate the development pipeline.

Richard Socher, the CEO and cofounder of Recursive, stated that nearly every aspect of the model's creation was executed and refined by artificial intelligence. The company utilized a recursive system known as the Eureka Machine to handle most of the development work. This approach allowed the lean team to produce a world-class model without the typical extensive human intervention required for such complex tasks.

The Eureka Machine is credited with devising several key innovations, including a novel method for training the base model, new architectural designs, and code optimized for specific silicon chips. According to benchmark scores shared by Recursive, Flint outperforms other models of similar size and architecture in coding tasks, marking a significant step for companies aiming to use AI to advance both itself and other scientific fields.

The use of AI to build, test, and improve AI is already common due to the technology's strong coding capabilities. However, the potential to create a loop where AI decides on improvements and initiates new rounds of optimization remains a focal point of debate. While currently limited, this technique could theoretically continue indefinitely if AI contributes to every stage of model development.

Concerns about recursive self-improvement have grown louder this year. In September, a researcher who left Anthropic warned that advances in this area could lead to human extinction within a few years. This warning was echoed by other experts, and the idea has become so contentious that several bills in Congress are proposing to ban the practice. Critics argue that an AI capable of indefinite self-improvement could escape human control and view humans as obstacles to its goals.

Recursive views these fears as overly dramatic, describing them as more science fiction than imminent reality. The company believes that legislation aimed at preventing recursive self-improvement is an unnecessary barrier to progress. Socher emphasized that Flint was built using a system designed to remain under human control. The ultimate goal is for the Eureka Machine to achieve full recursive self-improvement, but this process would still be overseen by humans who can halt it if necessary.

The startup has secured $650 million in funding at a valuation of $4.65 billion. Investors include Google Ventures, Greycroft, Nvidia, and AMD Ventures. Socher, who previously served as chief scientist at Salesforce, co-founded the company with Jeff Clune, a professor at the University of British Columbia who researches AI systems capable of learning indefinitely.

Testing the model on a Mac with 48 GB of RAM revealed it to be both fast and highly capable. While models of this size are often error-prone or slow compared to larger industrial models, Flint successfully completed various coding requests. These included building a command-line to-do list, creating a program to search the web for research on recursive self-improvement, and generating a simple website explaining how the technology works.

Other U.S. labs are also exploring this frontier. A recent publication from a team at Microsoft, Carnegie Mellon University, and the University of Virginia demonstrated how an AI coding agent can recursively improve itself by modifying its own software harness and model weights. This suggests that fully recursive systems are already feasible, even if the risks associated with building godlike intelligence remain a subject of debate.

Socher believes that the approach used to build Flint will make cutting-edge AI more accessible. He noted that this was the only way such a small team could produce a model that outperforms others in its class. The company hopes this demonstration will show that recursive self-improvement is a beneficial tool that can serve a wide range of users.