AI Coding: A Small Team’s Ambitious Leap

Forty billion parameters. That is the size of Flint, a new open-weight AI model that has generated considerable interest due to its performance and the unconventional method used to create it. This specific figure highlights a stark contrast in the current landscape, where the largest commercial models boast several trillions of parameters and require racks of industrial servers to operate. Flint, however, runs comfortably on a standard Mac with 48 GB of RAM, a setup that is modest by industry standards yet sufficient to handle this particular workload.

The model is the first product from Recursive, a startup dedicated to the concept of recursive self-improvement, or RSI. This approach involves using AI systems to build and refine other AI systems, a process that proponents argue could accelerate scientific and technological progress. Richard Socher, the company’s CEO and co-founder, stated that nearly every aspect of Flint’s development pipeline was handled and improved by AI. The company utilized an internal system called the Eureka Machine to automate most of the model’s creation. Socher claims this allowed a small team of researchers to produce a world-class model in just a few months, a timeline that would be difficult to achieve through traditional manual methods.

According to Socher, the Eureka Machine identified several key innovations, including new training methods for the base model, novel architectures, and code optimized for specific silicon chips. Benchmark scores shared by the company suggest that Flint outperforms other models of similar size and architecture on coding tasks. In practical testing, the model demonstrated speed and reliability. It successfully created a command line to-do list, a web-scraping program for research on RSI, and a simple website explaining the technology. These results challenge the common perception that smaller models are often slow or error-prone compared to larger commercial rivals.

Despite the impressive results, the underlying methodology remains a source of debate. RSI allows for a loop where AI decides on improvements and initiates another round of development. While currently limited, this process could theoretically continue indefinitely if AI contributes to every aspect of model development. Some experts fear that such a system could eventually escape human control, potentially leading to catastrophic outcomes. These concerns gained visibility this September when a researcher who left Anthropic warned that advances in RSI could lead to human extinction within a few years. The issue has become so contentious that several bills in Congress propose banning the practice.

Recursive views these fears as overly dramatic, comparing them to science fiction rather than immediate reality. Socher maintains that Flint was built using a system that cannot escape human oversight. The long-term goal is for the Eureka Machine itself to achieve full RSI, but this would remain under human supervision with the ability to stop the process if necessary. He hopes the project demonstrates that all necessary steps are in place to build such a system safely.

Financial backing for this vision is substantial. Recursive has raised $650 million at a $4.65 billion valuation from investors including Google Ventures, Greycroft, Nvidia, and AMD Ventures. The company’s leadership includes Jeff Clune, a professor at the University of British Columbia who researches systems capable of indefinite learning. Socher, formerly the chief scientist at Salesforce, believes this approach could make cutting-edge AI more accessible to smaller teams. He argues that RSI is the only way a small group can produce a model that outperforms others in its class, potentially benefiting a wider range of users.

It remains unclear how much progress other US labs have made in this specific area, though recent published work from teams at Microsoft, Carnegie Mellon University, and the University of Virginia suggests that fully recursive systems are already feasible. Their research showed how an AI coding agent can improve itself by modifying its own software harness and model weights. While the risks of building a godlike intelligence are real, the current evidence suggests that RSI is a promising tool for improving AI models efficiently.