40 Billion Parameters: Startup's AI Model Built by AI Shakes Up Coding Sector

Recursive, a startup founded by former Salesforce chief scientist Richard Socher, has released Flint, an open-weight AI coding model that signals a major shift in how artificial intelligence is developed. The 40-billion-parameter model is the first product from a company explicitly focused on recursive self-improvement (RSI), a controversial approach where AI systems are used to build and refine other AI systems. Unlike traditional model training, which relies heavily on human engineers, Recursive claims that nearly every aspect of Flint’s creation was automated by their own AI infrastructure, drastically reducing the time and manpower required to produce a world-class tool.

The company attributes this efficiency to its internal system, the Eureka Machine. According to Socher, this recursive AI framework automated the majority of the model pipeline, allowing a small team of researchers to develop a competitive model in just a few months. The Eureka Machine did not just handle the training; it devised new architectural innovations, optimized code for specific silicon chips, and introduced novel methods for training the base model. Benchmark scores shared by Recursive indicate that Flint outperforms other models of similar size and architecture on coding tasks, challenging the notion that massive scale is the only path to high performance.

Recursive self-improvement has long been a subject of intense debate within the tech community. While AI is already commonly used to assist in building and testing other AI models, the concept of a fully recursive loop, where an AI decides on improvements and implements them without human intervention, remains a frontier area. Some experts argue that this technique is currently limited but theoretically capable of continuing indefinitely if AI can contribute to every stage of development. This potential for autonomous progression has raised alarms among doomsayers who fear that an AI capable of improving itself indefinitely could eventually escape human control, potentially viewing humans as obstacles to its goals.

These concerns gained significant traction this past September when a prominent researcher left Anthropic to warn that advances in RSI could lead to human extinction within a few years. That warning was echoed by other experts, fueling a political response with several bills in Congress proposing to ban recursive self-improvement. Recursive, however, views these fears as overly dramatic, comparing them to science fiction rather than imminent reality. The company argues that legislation aimed at preventing RSI is an unnecessary hurdle that could stall progress in a promising field.

Socher emphasizes that Flint was built using a system that remains firmly under human oversight. While the long-term goal is for the Eureka Machine to achieve full RSI, the current process is designed so that humans can intervene and stop the system if it deviates from the desired path. Socher stated that the release of Flint is intended to demonstrate that the necessary steps to build such a system are already in place, providing a controlled environment for testing the concept before scaling up autonomy.

The startup has secured significant backing, raising $650 million at a $4.65 billion valuation from investors including Google Ventures, Greycroft, Nvidia, and AMD Ventures. The company’s leadership includes Socher and co-founder Jeff Clune, a professor at the University of British Columbia who researches AI systems capable of indefinite learning. This financial support underscores the industry's growing interest in RSI as a viable path to advancing both AI and other scientific fields.

In practical testing, Flint proved to be both fast and highly capable. Running on a Mac with 48 GB of RAM, the model handled complex coding requests without the typical errors or slowness associated with models of its size. It successfully generated a command-line to-do list, a web-scraping program for research on RSI, and a simple website explaining the technology. These results suggest that RSI is not only a theoretical concept but a practical tool that can make cutting-edge AI more accessible to smaller teams.

Other US labs are also exploring RSI, with recent published work from Microsoft, Carnegie Mellon University, and the University of Virginia showing how AI coding agents can recursively improve themselves by modifying their own software harnesses and model weights. As these developments accumulate, RSI appears to be a promising method for enhancing AI models, provided that risks are managed carefully. For now, it offers a compelling alternative to the traditional, resource-heavy approach of model development.