A Small Team Uses AI to Build a Strong New Coding Model

The line between human oversight and machine autonomy in software development is blurring faster than many expected.

Recursive, a startup led by former Salesforce chief scientist Richard Socher, has released Flint, an open-weight AI coding model with 40 billion parameters. The company asserts that nearly every aspect of the model's pipeline was executed and refined by AI. Socher stated that their internal system, known as the Eureka Machine, automated most of the development process. This automation allowed a small group of researchers to create a high-performing model in a matter of months.

The Eureka Machine reportedly devised several key innovations, including a novel training method for the base model, new architectural designs, and code optimized for specific hardware chips. Benchmark scores released by the company indicate that Flint outperforms other models of similar size and architecture on coding tasks. This approach represents a shift from the traditional method where humans guide most of the development, testing, and refinement stages.

The concept behind this work is recursive self-improvement, or RSI. While AI is already used to assist in building and testing other AI systems, RSI involves creating a loop where the AI decides on improvements and then implements them. Currently, this technique has limitations. However, if an AI can contribute to every stage of model development, the process could theoretically continue indefinitely without human intervention at each step.

This capability has sparked significant debate among experts. Some worry that allowing AI to improve itself could lead to unintended consequences. Critics argue that an AI capable of indefinite self-improvement might eventually escape human control or view humans as obstacles to its goals. Concerns escalated in September when a researcher from Anthropic resigned, warning that advances in RSI could lead to human extinction within a few years. Several other experts echoed this view, leading to congressional bills proposing to ban the practice.

Recursive views these fears as overly dramatic. Socher maintains that Flint was built using a system that cannot escape human control. The long-term goal is for the Eureka Machine to achieve full RSI, but humans would still oversee the process and retain the ability to stop it if necessary. Socher believes the current work demonstrates that the necessary steps are in place to build such a controlled system.

The startup has raised $650 million at a $4.65 billion valuation from investors including Google Ventures, Greycroft, Nvidia, and AMD Ventures. The company's co-founders include Jeff Clune, a University of British Columbia professor who researches AI systems capable of indefinite learning. Socher previously served as the chief scientist at Salesforce.

Independent testing of Flint on a Mac with 48 GB of RAM revealed a model that is both fast and highly capable. While models of this size are often error-prone or slow compared to larger rivals, Flint handled various coding requests accurately. These tasks included building a command line to-do list, creating a program to search the web for research on RSI, and generating a simple website explaining how RSI works. The performance suggests that smaller models can rival larger ones in specific coding domains.

Other US labs are also exploring RSI. A recent publication from Microsoft, Carnegie Mellon University, and the University of Virginia demonstrated how an AI coding agent can recursively improve itself by modifying its own software harness and model weights. Socher argues that the approach used for Flint could make cutting-edge AI more accessible to smaller teams. He believes this method is currently the only way a small group can produce a model that exceeds the performance of others in its class, potentially benefiting a wider audience.