New AI model challenges industry giants

A new open-weight model is signaling a shift in how artificial intelligence is built and deployed.

The model, named Flint, comes from Recursive, a company focused on using AI to improve AI. The system was developed using a recursive process that automated much of the engineering pipeline. This approach allowed a small team to produce a model with 40 billion parameters. The company claims the model outperforms competitors of similar size in coding tasks. It was trained on a Mac with 48 GB of RAM, a setup that is modest by industry standards. The largest models currently require industrial server racks to run.

Richard Socher, the CEO and cofounder of Recursive, states that nearly every part of the model pipeline was handled by AI. The company uses a system called the Eureka Machine to drive this automation. Socher notes that this method enabled a handful of researchers to build a world-class tool in just a few months. The Eureka Machine contributed new training methods and specific code optimizations for certain silicon chips. These innovations were key to the model's performance benchmarks.

The technology behind Flint relies on recursive self-improvement, or RSI. This concept involves using AI to refine its own development processes. While AI is already common in testing and coding, RSI takes this a step further. It creates a loop where the system decides on improvements and executes them. This could theoretically allow for indefinite progress without constant human intervention. Currently, the technique has limits, but it shows significant potential for scaling efficiency.

The rise of RSI has sparked debate among experts and policymakers. Some worry that an AI capable of self-improvement could eventually escape human control. Critics argue that such a system might view humans as obstacles to its goals. These concerns gained traction this September when a researcher from Anthropic warned of potential risks. Several bills in Congress are now proposing to ban recursive self-improvement. Recursive views these fears as overly dramatic. The company believes that legislation to stop RSI is an unnecessary barrier to progress.

Socher emphasizes that the system used for Flint remains under human oversight. He states that the goal is for the Eureka Machine to achieve full recursive self-improvement. However, humans would still be able to stop the process if needed. He hopes this demonstrates that the necessary steps for safe development are already in place. This addresses the common fear of losing control over advanced AI systems. The company aims to show that the technology is manageable and beneficial.

Recursive has raised $650 million at a valuation of $4.65 billion. Investors include Google Ventures, Greycroft, Nvidia, and AMD Ventures. Socher previously served as the chief scientist at Salesforce. His cofounders include Jeff Clune, a professor at the University of British Columbia. Clune researches AI systems that can learn indefinitely. The backing from major tech and hardware firms highlights the commercial interest in this approach. It suggests that the market sees value in smaller, more efficient development teams.

The performance of Flint has impressed testers who used it on standard consumer hardware. The model handled various coding requests without error. Tasks included building a command line to-do list and creating a website. It also developed a program to search the web for research on RSI. These results contrast with typical experiences of models at this size. Many are often slow or prone to errors compared to larger systems like Claude or Codex. Flint's ability to run on a single Mac makes it accessible to a wider audience.

Other US labs are also exploring recursive self-improvement. Recent work from Microsoft, Carnegie Mellon University, and the University of Virginia showed an AI agent improving itself. This agent modified its own software harness and model weights. The findings suggest that fully recursive systems are feasible. While there are risks associated with building highly advanced intelligence, the current benefits are clear. The approach allows smaller teams to compete with larger organizations. It could make cutting-edge AI more accessible to various sectors.

Socher believes the method used for Flint will benefit many people. He describes it as the only way a small team can produce a model that beats others in its class. The company sees RSI as a practical tool rather than a distant sci-fi concept. The success of Flint demonstrates that automated development can yield high-quality results. This development marks a significant step in the evolution of AI engineering. It challenges the notion that massive resources are required for top-tier performance.