Recursive Launches Flint, A 40 Billion Parameter AI Coding Model Built By AI

A new open-weight artificial intelligence model named Flint has emerged as a strong contender in the coding landscape, demonstrating the practical viability of recursive self-improvement. The model, developed by the startup Recursive, utilizes a methodology where AI systems are employed to build and refine other AI systems. This approach, known as recursive self-improvement, has long been a subject of intense debate within the technology sector. Flint represents the first major public offering from the company, which seeks to apply this automated development process to advance both artificial intelligence and other scientific fields.

Richard Socher, the chief executive and co-founder of Recursive, explained that the entire model pipeline was largely executed and improved by AI. The company leveraged an internal system called the Eureka Machine to automate the majority of the development workflow. This automation allowed a small team of researchers to construct a high-performing model within a few months. Socher noted that the Eureka Machine identified several key innovations, including novel training methods for the base model, new architectural designs, and code optimized for specific silicon chips.

Shared benchmark data indicates that Flint exceeds the performance of other models of similar size and architecture in coding tasks. AI tools are already commonly used to assist in the construction, testing, and refinement of AI systems due to their coding capabilities. This utility accelerates the refinement process and allows for the testing of new concepts with reduced human intervention. The recursive loop allows AI to determine necessary improvements and initiate subsequent rounds of enhancement. While currently limited, this technique could theoretically continue indefinitely if AI contributes to every aspect of model development.

Concerns regarding recursive self-improvement have intensified in recent months. Some experts fear that an AI capable of unlimited self-improvement might eventually escape human control. This anxiety gained prominence in September when a researcher from Anthropic resigned to warn that such advances could lead to human extinction within a few years. Other experts have echoed these fears, and several bills in Congress now propose banning the practice. Recursive views these concerns as more akin to science fiction than immediate reality and argues that legislation restricting the technology is an unnecessary barrier to progress.

Socher emphasized that Flint was built using a system that remains under human oversight. The goal is for the Eureka Machine to eventually achieve full recursive self-improvement, but humans would retain the ability to halt the process if it deviates from intended paths. He stated that the current progress demonstrates that the necessary steps for such a system are already in place. The company has raised $650 million at a valuation of $4.65 billion from investors including Google Ventures, Greycroft, Nvidia, and AMD Ventures. Socher previously served as the chief scientist at Salesforce, and his co-founders include Jeff Clune, a professor at the University of British Columbia who studies indefinitely learning AI systems.

Testing the model on a Mac with 48 GB of RAM revealed that Flint is both fast and highly capable. The model contains 40 billion parameters, a size that allows it to run on consumer hardware, unlike the largest models which require industrial server racks. In comparison to other models of similar magnitude, which can often be slow or error-prone, Flint successfully handled a variety of coding requests. These tasks included creating a command line to-do list, building a program to search the web for research on recursive self-improvement, and developing a simple website explaining the technology.

Research from Microsoft, Carnegie Mellon University, and the University of Virginia suggests that fully recursive systems are already feasible. A recent publication detailed how an AI coding agent can recursively improve itself by modifying its own software harness and model weights. While the risks of approaching godlike intelligence remain, the current approach appears to be a promising method for enhancing AI models. Socher believes that this methodology will make cutting-edge AI more accessible to smaller teams. He argues that it is the only way such a small group can produce a model that surpasses others in its class, potentially benefiting a wide range of users.

The development of Flint highlights a shift in how AI models are created. By relying on automated processes, companies can achieve significant results with fewer resources. The model's performance on modest hardware suggests that advanced coding capabilities are becoming more attainable. As the industry continues to explore recursive self-improvement, Flint serves as a tangible example of the potential and the progress being made in this controversial but promising field.