40 Billion Parameters: The New AI Model That Coded Its Own Way to Existence

The question hanging over the current wave of artificial intelligence is simple but terrifying: can we let the machines build the machines? While most labs rely on human engineers to tweak and refine their code, a new startup is betting that the AI should do the heavy lifting. This approach, known as recursive self-improvement, is no longer just a theoretical concept. It is now the engine behind a new open-weight model called Flint, which is generating strong results in coding tasks and reigniting a heated debate about how much control humans should retain over their own creations.

The numbers behind this development are striking for a company that describes itself as a small team. Recursive, the firm behind Flint, raised $650 million in funding at a valuation of $4.65 billion. Investors include Google Ventures, Greycroft, Nvidia, and AMD Ventures. Despite this capital, the company operates with only a handful of researchers. They claim that this small group managed to build a world-class model in just a few months by automating the entire development pipeline. The resulting model, Flint, contains 40 billion parameters, a size that allows it to run on standard consumer hardware rather than industrial server racks.

The people driving this effort are experienced players in the tech industry. Richard Socher, the CEO and cofounder of Recursive, previously served as the chief scientist at Salesforce. He is joined by Jeff Clune, a professor at the University of British Columbia who studies systems capable of indefinite learning. Socher explained that the company uses an internal system called the Eureka Machine to automate nearly every aspect of the model’s creation. This includes designing new architectures, optimizing code for specific silicon chips, and training the base model. The system reportedly devised several innovations that allowed Flint to outperform other models of similar size on benchmark tests.

The stakes in this debate have risen sharply this year. In September, a researcher from Anthropic quit to warn that advances in recursive self-improvement could lead to human extinction within a few years. This alarm was echoed by other experts, leading to a wave of concern that AI might eventually escape human control. The tension is so high that several bills in Congress now propose banning the practice entirely. Recursive, however, views these fears as overly dramatic. Socher argues that their system is designed with human oversight in place, ensuring that the process can be stopped if it goes wrong. He hopes to demonstrate that all the necessary steps for safe self-improvement are already in place.

For the average user, the practical implications are already visible. Testing on a Mac with 48 GB of RAM, Flint proved fast and capable. It successfully built a command line to-do list, a web scraper for research, and a simple website explaining how the technology works. While models of this size are often error-prone, Flint handled these requests without issue. This suggests that the technology could make high-performance AI more accessible to smaller teams and individuals who lack massive computational resources.

The broader scientific community is watching closely. Recent published work from teams at Microsoft, Carnegie Mellon University, and the University of Virginia suggests that fully recursive systems are feasible. These researchers showed how an AI coding agent can improve itself by modifying its own software harness and model weights. As these technologies converge, the line between human-directed development and autonomous evolution continues to blur. For now, the focus remains on whether this efficiency can be maintained without triggering the very risks that have lawmakers and experts on edge.