40 Billion Parameters: New AI Model Demonstrates Rapid Self-Improvement

A new open-weight artificial intelligence model containing 40 billion parameters has demonstrated that fully recursive self-improvement (RSI) is becoming a practical reality for small development teams. The model, named Flint, is the first product from Recursive, a startup dedicated to using AI to advance both its own capabilities and broader scientific fields. The system was designed to minimize human intervention in the creation process, relying instead on an automated pipeline to generate a world-class coding tool in a matter of months.

Richard Socher, the CEO and cofounder of Recursive, stated that nearly every aspect of the model's pipeline was executed and refined by AI. The company utilized its own recursive system, known as the Eureka Machine, to automate the majority of the development workflow. This automation allowed a small group of researchers to produce a model that outperforms competitors of similar size and architecture in coding benchmarks, according to scores released by the firm.

The Eureka Machine did not merely assist with coding; it devised specific technical innovations. These include a new method for training the base model, novel architectural designs, and code optimized for specific silicon chips. While AI has long been used to help build and test other AI systems, the distinction here is the closed loop where the AI decides on improvements and executes them. This approach allows for iterative refinement without constant human oversight, a process that experts believe could theoretically continue indefinitely if the system achieves full recursion.

The concept of recursive self-improvement has sparked significant debate within the tech industry. Some experts warn that allowing AI to improve itself without limits could lead to a loss of human control, potentially resulting in an intelligence that views humans as obstacles to its goals. These concerns gained prominence in September when a researcher leaving Anthropic warned that such advances could lead to human extinction within a few years. The issue has become contentious enough that several bills in Congress propose banning RSI, although Recursive views these fears as science fiction and considers proposed legislation an unnecessary barrier to progress.

Recursive argues that its current system is incapable of escaping human control. Socher explained that while the goal is for the Eureka Machine to achieve full recursive self-improvement, the process remains overseen by humans who can halt it if necessary. The company aims to demonstrate that all necessary steps are in place to build such a system safely. To date, Recursive has raised $650 million at a $4.65 billion valuation from investors including Google Ventures, Greycroft, Nvidia, and AMD Ventures.

The team behind Flint includes Socher, who previously served as chief scientist at Salesforce, and Jeff Clune, a professor at the University of British Columbia who researches AI systems capable of indefinite learning. The model's performance was evaluated on a standard Mac with 48 GB of RAM, a setup sufficient for running its 40 billion parameters. By comparison, the largest models currently available feature several trillions of parameters and require industrial server racks to operate.

In practical testing, Flint proved to be fast and highly capable, avoiding the common errors or slowness often associated with models of this size. The system successfully generated a command-line to-do list, a web-scraping program for research on recursive self-improvement, and a simple website explaining the RSI process. Each request was completed accurately, suggesting that the model is a robust tool for developers.

Other research institutions are also exploring similar techniques. A recent study from Microsoft, Carnegie Mellon University, and the University of Virginia demonstrated how an AI coding agent can recursively improve itself by modifying both its software harness and model weights. This suggests that fully recursive systems are already feasible, even if the risks of creating a "godlike" intelligence remain a topic of discussion.

Socher believes that the approach used to build Flint will make cutting-edge AI more accessible to smaller teams. He noted that recursive self-improvement is the only way a small group can produce a model superior to others in its class. The company hopes this success will show that RSI is beneficial and manageable, rather than a dangerous leap into the unknown.