For the past few days, I’ve been doing something I don’t often get to do in this job: testing a new AI model by actually living with it. Flint is an open-weight coding model, which means it isn’t just a distant voice behind an API. You can inspect its design, run it yourself, and poke at its strengths and weaknesses in a hands-on way. And the hands-on experience is genuinely impressive. Flint writes clean code, explains tricky bugs with surprising patience, and makes suggestions that feel less like autocomplete and more like a collaborator. But Flint didn’t capture my attention simply because it’s another good coding assistant. It captured my attention because of what it represents. Flint is the first product from Recursive, a startup founded by Richard Socher, and Recursive’s entire premise is that AI should be used to improve AI. Socher told me that his team used a recursive system called the Eureka Machine to automate most of the model-building process. The training method, the model architecture, even code optimized for specific computer chips—all of it was shaped or improved by AI. A company with only a handful of researchers was able to build a world-class coding model in a matter of months, Socher said. That in itself would be a remarkable story. But the deeper story is about recursive self-improvement, or RSI, an idea that has lived for decades on the fringes of science fiction and theoretical AI research. Flint is a tangible reminder that RSI is edging closer to reality, and that means the familiar debate about whether AI should be allowed to improve itself is no longer academic.
We already know that AI can help build AI. For years, developers have used large language models to write, review, and improve code, and much of that code is used to train and run other AI systems. That may sound like a closed loop, but in practice it has been a narrow one: humans still decide what to build, which experiments to run, and what counts as “better.” Recursive is trying to widen that loop dramatically. Socher described several concrete innovations that the Eureka Machine came up with on its own, including a fresh approach to training the base model, new model architectures, and highly optimized code tailored to particular hardware. He also shared benchmark scores suggesting that Flint outperforms other models of similar size and architecture on coding tasks. The exact numbers are less important than the direction: a small company with limited human resources achieved results that previously would have required a much larger team. The difference, Socher argues, is that AI wasn’t just a tool being used by humans; it was a system that was actively participating in design decisions, spotting patterns, and proposing improvements that human researchers hadn’t considered. That is what makes Flint feel like a turning point. It’s not that AI suddenly became brilliant overnight. It’s that AI is now embedded in every stage of the development pipeline, and the pipeline itself is moving faster because of it.
So what exactly is recursive self-improvement, and why should we care? Imagine a workshop where a machine can build a better version of itself. That better version can then build an even better one, and so on. At its most dramatic, this creates a feedback loop that could accelerate beyond human comprehension—a machine that gets smarter every time it runs, without needing us to catch up. In the AI community, RSI has often been treated as a distant scenario, something discussed at conferences next to existential risk and the possibility of superintelligence. But the principle is already at work in watered-down form. If AI can help generate training data, refine model weights, discover better architectures, and evaluate its own performance, then it is already part of the improvement cycle. The missing piece, for now, is complete autonomy. Humans still set the broad goals and intervene at key moments. But Recursive’s goal is to push that boundary further. Socher says his company wants the Eureka Machine itself to achieve full recursive self-improvement, meaning AI would eventually be able to decide what to work on, carry out the work, test the results, and begin another round of improvement without waiting for a human hand. If every piece of the model pipeline can be improved by AI, then the loop could theoretically continue indefinitely. We are not there yet, and many researchers would say we aren’t close. But Flint suggests that the gap between theory and practice is narrowing.
That narrowing is exactly what makes some experts nervous. There is a long tradition of doomsayers in AI who warn that a system capable of improving itself could eventually surpass us, decide that humans are obstacles, and act accordingly. These fears became far more visible in September, when a researcher quit Anthropic and publicly warned that advances in AI, including recursive self-improvement, could lead to human extinction within just a few years. That claim was echoed by other voices in the community, and it landed with enough force that politicians took notice. A number of bills have been introduced in Congress that propose to ban recursive self-improvement altogether, or at least to regulate it heavily. To the people who support these measures, RSI is not a philosophical problem. It is a safety valve that, if opened too early, could release something we cannot contain. They imagine an AI that rewrites its own code, discovers new behaviors, and escapes the sandbox that was supposed to keep it under control. In that story, AI doesn’t need to be evil to be dangerous. It just needs to be extremely capable and insufficiently aligned with human values. The fact that a model like Flint exists today doesn’t disprove that story. If anything, it makes the story feel more immediate. The word “recursive” appears on company websites and in press releases now, not just in gloomy think tanks and academic papers. That is both exciting and unsettling.
Recursive, predictably, sees things differently. Socher is aware of the doomsday narratives, but he doesn’t accept them. He told me that Flint was built using a system that is fundamentally incapable of escaping human control. The whole process was designed with people in the loop, and if something went wrong, humans could stop it at any point. He also acknowledged that the dream of full recursive self-improvement involves more autonomy, but he insisted that even that dream would include human oversight. The goal isn’t to create a runaway AI that disappears into its own mind. The goal is to build a system where AI accelerates progress while people retain the ability to hit the brakes. Socher expressed hope that Flint would demonstrate to skeptics that Recursive has all the necessary steps in place to pursue RSI responsibly. To him, the fear of an AI apocalypse is more like a sci-fi movie than an engineering reality. And legislation to ban RSI, he argues, would be an unnecessary and counterproductive obstacle to progress. It’s not hard to understand why a startup founder would say that. If your entire company is built on the idea that AI can improve itself, a ban on recursive self-improvement is essentially a ban on your mission. But even people who don’t share Socher’s optimism can appreciate the appeal. The promise of RSI is not just smarter AI. It’s faster science, better medicine, more efficient energy systems, and breakthroughs that might otherwise take decades. The question is whether that promise can be achieved without inviting catastrophe.
I’ve come away from my time with Flint feeling more hopeful than I expected, but not naive. There is something genuinely remarkable about watching an AI model that was partly designed by another AI model produce brilliant code at my fingertips. It makes the future feel nearer than it did a year ago. But it also makes me appreciate how much depends on the choices we make now. Recursive’s approach is transparent and cautious, at least by its own account. The company is open about using AI in the development pipeline, shares benchmark results, and emphasizes human oversight. That is a good starting point. But the broader history of technology suggests that any tool powerful enough to accelerate progress is also powerful enough to be misused. The debate over RSI is not really a debate about whether we should use AI to improve AI. That ship has already sailed. It’s a debate about how much autonomy we should give these systems, and where we should draw the line. Flint is a milestone, not a conclusion. It shows that recursive self-improvement is increasingly real, but it also shows that humans are still very much in the loop. The future of RSI will depend on whether the people building these systems can maintain the same level of care and transparency as the field grows. For now, I’m watching with excitement, a little unease, and a lot of curiosity. The machine might be building itself, but we still get to decide what we want it to become.
