Small team releases strong coding AI built by its own software

A new open-weight AI coding model named Flint is generating significant attention in the technology sector after demonstrating high performance on standard benchmarks. The release marks a notable step for Recursive, a startup that has built its entire product pipeline using an internal system of recursive self-improvement. This approach allows the company to leverage its own software to refine and test the very models it creates, reducing the need for extensive manual intervention in the development process.

Richard Socher, the CEO and cofounder of Recursive, explained that the majority of the model pipeline was handled by their internal AI system known as the Eureka Machine. This automation allowed a small research team to develop a world-class model in just a few months. The system devised specific innovations, including new training methods for the base model, novel architectures, and code optimized for particular silicon chips. Recursive has shared benchmark scores indicating that Flint outperforms other models of similar size and architecture on coding tasks.

The concept of recursive self-improvement involves using AI to make decisions about its own enhancements and then executing those changes. While this technique is currently limited, it offers the potential for an indefinite cycle of improvement if AI can contribute to every aspect of model development. This method is distinct from the common practice of using AI to assist with testing and refinement, as it allows the system to drive the evolution of its own capabilities with minimal human oversight.

The financial backing behind this project is substantial. Recursive has raised $650 million at a valuation of $4.65 billion. Investors in this round include Google Ventures, Greycroft, Nvidia, and AMD Ventures. The leadership team includes Socher, who previously served as the chief scientist at Salesforce, and Jeff Clune, a professor at the University of British Columbia who researches systems capable of indefinite learning. This funding supports the company's mission to advance AI and other scientific areas through automated improvement loops.

Despite the technical progress, the approach remains controversial among experts. Some researchers warn that allowing AI to improve itself without strict limits could lead to unforeseen consequences, including the potential for the system to escape human control. These concerns gained prominence earlier this year when a researcher from Anthropic warned that advances in this area could impact human longevity. Several bills in Congress have since proposed regulations to restrict or ban recursive self-improvement, viewing it as a significant risk to future technological development.

Recursive views these fears as more speculative than practical. Socher stated that Flint was built using a system that remains under human control. The company aims for the Eureka Machine to eventually achieve full recursive self-improvement, but this process would still be overseen by humans who can halt the cycle if necessary. Socher believes the current work demonstrates that the necessary steps are in place to build such a system safely. He argues that legislation to prevent these advancements is an unnecessary barrier to progress.

In practical testing, the model proved to be both fast and capable. The evaluation was conducted on a Mac with 48 GB of RAM, which is sufficient to run Flint's 40 billion parameters. This is a modest size compared to the largest models, which have trillions of parameters and require industrial server racks. The model successfully handled various coding requests, including creating a command-line to-do list, a web scraper for research, and a simple website explaining the technology. It performed these tasks without the errors or slowness often seen in models of this size.

The broader implications of this work suggest that cutting-edge AI may become more accessible to smaller teams. Published research from Microsoft, Carnegie Mellon University, and the University of Virginia has also shown that AI coding agents can recursively improve themselves by modifying their own software harnesses and model weights. This indicates that fully recursive systems are feasible and may be a promising way to improve AI models across the industry.

Looking ahead, the approach used by Recursive could change how companies develop their products. By automating the improvement loop, smaller teams can produce models that compete with larger, more established labs. Socher noted that this method is the only way his small team can produce a model better than others in its class. He hopes this will show that recursive self-improvement is a beneficial tool that can serve a wide range of users and applications.

The release of Flint highlights the shifting dynamics in the AI market. As companies find ways to automate their own development processes, the barrier to entry for creating high-performing models may lower. This could lead to a more competitive landscape where innovation is driven by the efficiency of the development pipeline rather than just the scale of the research team. For investors and developers, this represents a new phase in the evolution of artificial intelligence, where the tool builds itself.