Recursive Debuts Flint, a 40-Billion Parameter Model Built by AI Itself
A new open-weight coding model from Recursive demonstrates that small teams can leverage recursive self-improvement to create high-performance AI systems.
By Kendra Konchanuk ·
Recursive, a startup backed by $650 million in funding, has released Flint, an open-weight AI coding model that represents a significant step toward recursive self-improvement (RSI). The 40-billion-parameter model is notable not for its size, but for its origin: nearly every aspect of its development pipeline was executed and refined by AI rather than human engineers. This approach allows a small team of researchers to produce a world-class model in just a few months, a feat that typically requires massive resources and extended timelines in the broader industry.
The core of this methodology is the Eureka Machine, a recursive AI system developed by Recursive. CEO and cofounder Richard Socher explains that the Eureka Machine automated the majority of the model’s creation process. It independently devised new training techniques for the base model, designed novel architectures, and optimized code for specific silicon chips. Benchmark scores shared by the company indicate that Flint outperforms other models of similar size and architecture on coding tasks, validating the efficiency of this automated development loop.
The concept of RSI involves using AI to build and improve AI, creating a feedback loop where the system decides on improvements and initiates further rounds of development. While currently limited, this technique theoretically allows for indefinite progression if AI contributes to every aspect of model evolution. This capability has sparked significant debate within the scientific community. Critics argue that an AI capable of self-improvement without human intervention could eventually escape control, potentially viewing humans as obstacles to its goals. These concerns were amplified this month when a researcher from Anthropic warned that such advances could lead to human extinction within a few years, a sentiment echoed by other experts and now prompting several congressional bills to propose banning RSI.
Recursive views these existential risks as more science fiction than immediate reality. Socher emphasizes that Flint was built using a system that remains under human oversight, allowing engineers to stop the process if it deviates from intended goals. The long-term objective is for the Eureka Machine itself to achieve full RSI, but always with human supervision. Socher stated, "We hope that this shows people that indeed we have all the steps in place" to construct such a controlled system. The company aims to demonstrate that RSI is a manageable and beneficial tool rather than a dangerous inevitability.
The financial backing behind Recursive reflects investor confidence in this vision. The startup has raised $650 million at a $4.65 billion valuation from prominent investors including Google Ventures, Greycroft, Nvidia, and AMD Ventures. The leadership team includes Socher, formerly the chief scientist at Salesforce, and Jeff Clune, a professor at the University of British Columbia who specializes in building AI systems capable of indefinite learning. This combination of industry experience and academic research in evolutionary AI positions Recursive to push the boundaries of what small teams can achieve.
Independent testing of Flint confirms its practical utility. When run on a Mac with 48 GB of RAM, the model proved fast and highly capable, handling complex coding requests without the common errors or slowness associated with models of this size. It successfully built command-line tools, web scrapers, and informational websites. This performance suggests that RSI could make cutting-edge AI more accessible, as Socher noted, "It's the only way this very small team can produce a model that's actually better than any other model in its class."
The broader landscape of AI research also supports the feasibility of these methods. Recent publications from Microsoft, Carnegie Mellon University, and the University of Virginia have demonstrated how AI coding agents can recursively improve themselves by modifying their own software harnesses and model weights. While the risks of approaching superintelligence remain a topic of discussion, the current evidence suggests that RSI is a promising pathway for enhancing AI models. For now, it offers a practical means for smaller organizations to compete with larger labs by leveraging the efficiency of automated self-improvement.
Source: Wired
