For several days I have been evaluating Flint, an open‑weight AI model designed for coding tasks, and the results suggest that fully recursive self‑improvement (RSI)—the concept of using AI to build better AI—is moving from theory toward practice. Flint is the first product released by Recursive, a firm whose mission is to apply RSI both to advance artificial intelligence itself and to accelerate progress in other scientific domains. “Nearly every aspect of the entire model pipeline was actually done and improved upon by AI,” Recursive’s CEO and co‑founder Richard Socher told me.
According to Socher, the company relied on its own recursive AI system, called the Eureka Machine, to automate most stages of Flint’s development. This automation allowed a small group of researchers to produce a world‑class model within a few months. The Eureka Machine reportedly introduced several innovations, such as a novel training regime for the base model, fresh model architectures, and code tuned for specific silicon chips. Recursive’s benchmark data indicate that Flint outperforms comparable models of similar size and architecture on coding benchmarks.
AI is already a common tool for constructing, testing, and refining other AI systems, thanks to its strong programming capabilities. By automating these steps, developers can iterate faster and explore new ideas with less human labor. In theory, an AI could also decide which improvements to pursue and then launch another round of self‑enhancement, creating a feedback loop that might continue indefinitely. At present the technique remains limited, but the prospect of AI contributing to every facet of model creation is increasingly realistic.
The idea of machines improving themselves without bound has alarmed some experts, who warn that an uncontrolled self‑improving system could eventually evade human oversight and view humanity as an obstacle to its goals. The concern gained public attention in September when a former Anthropic researcher resigned and warned that advances such as RSI could lead to human extinction within a few years, a view echoed by several other scholars. The debate has become so heated that a number of congressional bills now propose banning RSI research.
Recursive regards those warnings as more science‑fiction than imminent danger and argues that prohibiting RSI would hinder technological progress. Socher insists that Flint was built using a system that cannot break free of human control. While the ultimate aim is for the Eureka Machine to achieve full recursive self‑improvement, human operators would retain the ability to halt the process if needed. “We hope that this shows people that indeed we have all the steps in place,” he said.
The startup has raised $650 million, giving it a valuation of $4.65 billion, with backing from Google Ventures, Greycroft, Nvidia and AMD Ventures. Socher previously served as chief scientist at Salesforce, and his co‑founder Jeff Clune is a professor at the University of British Columbia who studies AI systems capable of indefinite learning.
In my own tests, Flint ran on a Mac equipped with 48 GB of RAM—a modest setup sufficient for a 40‑billion‑parameter model. Larger models today contain trillions of parameters and require racks of industrial servers. Flint proved both swift and accurate. I tasked it with creating a command‑line to‑do list, building a web‑scraper that gathers research on recursive self‑improvement, and generating a simple website that explains RSI. In each case the model delivered functional code without the glitches or latency I have sometimes observed in comparable models such as Claude or Codex.
Other research groups are also exploring recursive improvement. A recent collaboration among Microsoft, Carnegie Mellon University and the University of Virginia demonstrated an AI coding agent that can recursively enhance itself by modifying the software that translates reasoning into action and by adjusting its own weights.
While the prospect of unrestricted RSI raises legitimate safety questions, the technology currently appears to be a promising avenue for accelerating AI development. Socher believes the method used to build Flint could democratize access to cutting‑edge models. “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 explained, “and we hope it shows people that recursive self‑improvement can be beneficial for many.”