Recursive’s New Model: A Test Case for AI-Built AI

For developers and data scientists in the Great Lakes and Ohio Valley, the arrival of Flint marks a shift in how software is built. This open-weight model, created by the startup Recursive, demonstrates that a small team can produce a coding AI that matches or exceeds larger, more expensive competitors. The model runs on consumer-grade hardware, suggesting that high-performance AI tools may soon become accessible beyond the data centers of major tech firms.

The concept behind Flint is recursive self-improvement, or RSI. This approach uses AI to design, train, and refine other AI systems. While AI has long assisted in coding tasks, RSI takes this further by allowing the system to identify its own weaknesses and implement fixes without constant human direction. Recursive’s CEO, Richard Socher, explains that their system, the Eureka Machine, handled nearly every aspect of the pipeline, from new training methods to code optimized for specific chips.

The development timeline was unusually short. Socher notes that a handful of researchers built a world-class model in just a few months, a feat that would traditionally take years. The Eureka Machine devised specific innovations, including new model architectures and a novel base model training technique. Benchmark scores indicate that Flint outperforms other models of similar size and architecture on coding tasks, validating the efficiency of the automated process.

This rapid progress has intensified a debate about the safety of letting AI improve itself. Some experts fear that an AI capable of indefinite self-improvement could eventually escape human control. These concerns gained prominence this September when a researcher left Anthropic to warn that RSI could lead to human extinction within a few years. The issue has become political, with several bills in Congress proposing to ban the technique.

Recursive views these fears as overly dramatic. Socher argues that Flint was built using a system that cannot escape human oversight. The company’s ultimate goal is for the Eureka Machine to achieve full recursive self-improvement, but humans would retain the ability to stop the process if it goes awry. Socher stated, “We hope that this shows people that indeed we have all the steps in place” to manage such a system effectively.

The company behind the model has secured significant backing. Recursive has raised $650 million at a $4.65 billion valuation from investors including Google Ventures, Greycroft, Nvidia, and AMD Ventures. Socher, formerly the chief scientist at Salesforce, co-founded the company with Jeff Clune, a professor at the University of British Columbia who researches AI systems that can learn indefinitely.

Practical tests of Flint reveal its capabilities. The model has 40 billion parameters, a size that allows it to run on a Mac with 48 GB of RAM, whereas the largest models require industrial servers. In testing, Flint proved fast and reliable, successfully building a command-line to-do list, a web-scraping tool for research, and a simple website explaining RSI. Unlike other models of similar size, which can be error-prone, Flint handled these requests without significant issues.

Other institutions are exploring similar paths. A recent study from Microsoft, Carnegie Mellon University, and the University of Virginia showed how an AI agent can recursively improve itself by modifying its own software harness and model weights. While the long-term risks of creating godlike intelligences remain a topic of debate, the immediate benefit is clear: RSI offers a promising way to make cutting-edge AI more accessible to smaller teams and individual developers.

As the technology evolves, the focus is shifting from whether RSI is possible to how it can be safely integrated. For now, it appears to be an effective method for improving AI models. Socher believes that the approach used for Flint will demonstrate that recursive self-improvement is a viable and beneficial tool for a wide range of people, not just large corporations.