Startup Recursive Claims AI Built Its Own Supercharged Coding Model

A small team of researchers has produced a coding model that appears to have been designed largely by artificial intelligence itself.

The model, named Flint, represents the debut product of Recursive, a startup explicitly dedicated to recursive self-improvement, or RSI. This controversial methodology involves using AI systems to design, train, and optimize subsequent AI systems, effectively creating a loop of automated advancement. According to Richard Socher, the company’s CEO and co-founder, nearly the entire pipeline for developing Flint was executed and refined by AI. Socher stated that the company utilized an internal system called the Eureka Machine to automate the majority of the development process. This automation allowed a team with only a handful of researchers to construct a world-class model in a matter of months, a timeline that would typically require significantly larger resources and personnel.

Socher noted that the Eureka Machine identified several key innovations during the development phase. These include a novel approach to training the base model, new architectural designs, and code specifically optimized for particular silicon chips. Benchmark scores released by Recursive indicate that Flint outperforms other models of similar size and architecture on specific coding tasks. While AI is already frequently employed to assist in building and testing other AI models, the creation of a closed loop where the system decides on improvements and then implements them remains a distinct step. Currently, this technique is limited, but if AI can contribute to every facet of model development, the process could theoretically continue indefinitely without human intervention for each iteration.

The concept of RSI has ignited significant debate within the scientific community. Some experts express concern that allowing AI to improve itself could lead to catastrophic outcomes. Critics argue that an AI capable of indefinite self-improvement might eventually escape human control, potentially viewing humans as obstacles to its objectives. These anxieties gained prominent attention in September of this year when a researcher departed Anthropic to warn that advances in AI, including RSI, could result in human extinction within a few years. This warning was subsequently echoed by other specialists in the field. The controversy has become so pronounced that several bills in Congress have proposed legislation to ban RSI entirely. Recursive, however, views these fears as more science fiction than imminent reality, arguing that such legislation is an unnecessary barrier to progress.

Socher maintains that Flint was constructed using a system that remains incapable of escaping human oversight. The long-term goal is for the Eureka Machine to achieve full recursive self-improvement, but this process would still be monitored by humans who retain the ability to halt operations if they deviate from the intended path. Socher hopes this development demonstrates that the necessary steps are in place to build such a system safely. The company has secured $650 million in funding at a valuation of $4.65 billion, with investors including Google Ventures, Greycroft, Nvidia, and AMD Ventures. Socher previously served as chief scientist at Salesforce, and his co-founders include Jeff Clune, a professor at the University of British Columbia who studies AI systems capable of indefinite learning.

Practical testing of Flint reveals a capable and rapid performance. The model, which contains 40 billion parameters, can be run on consumer hardware such as a Mac with 48 GB of RAM. In comparison, the largest models currently available possess trillions of parameters and require racks of industrial servers to function. During evaluation, Flint successfully handled various coding requests, including creating a command-line to-do list, building a web-scraping program for RSI-related research, and generating a website explaining the mechanics of RSI. Unlike other models of similar size, which can be slow or prone to errors, Flint executed these tasks accurately and quickly.

It remains unclear how much progress other U.S. laboratories have made in pursuing recursive self-improvement, though recent publications suggest that fully recursive systems are already feasible. A team comprising researchers from Microsoft, Carnegie Mellon University, and the University of Virginia recently published work demonstrating how an AI coding agent can recursively improve itself by modifying both its software harness and its own model weights. While the risks associated with full RSI may become significant as systems approach godlike intelligence, the current application appears to be a promising method for enhancing AI models. Socher argues that this approach could make cutting-edge AI more accessible, noting that it is the only way such a small team can produce a model superior to others in its class. He believes this will demonstrate that recursive self-improvement is beneficial for a wide range of users.

The emergence of Flint highlights a shifting landscape in AI development, where the barrier to entry for high-performance models may be lowering. As smaller teams leverage automated pipelines to compete with larger institutions, the debate over the speed and safety of recursive self-improvement is likely to intensify in the coming months.