Testing Flint: The Open-Weight Coding AI That Moves Recursive Self-Improvement From Theory To Practice

Over the last several days, I’ve spent hours putting a remarkably capable new open-weight coding AI named Flint through its paces — and the model makes clear that fully realized recursive self-improvement (RSI), the divisive, game-changing technique that uses AI to build better AI, is no longer a far-off science fiction concept.

Flint marks the first public product from Recursive, a startup built explicitly around leveraging RSI to accelerate breakthroughs not just in AI development, but across wider fields of scientific research. “Virtually every step of the full model development pipeline was designed, built, and refined by AI,” Recursive co-founder and CEO Richard Socher told me in a recent interview.

Socher explained that the firm’s in-house recursive AI platform, codenamed the Eureka Machine, automated nearly every stage of Flint’s creation. The end result? A small, tight-knit team of researchers was able to deliver a world-class coding model in a matter of months, a milestone that would have once required hundreds of staff and years of work. The Eureka Machine independently generated multiple key innovations baked into Flint, including a never-before-seen base model training workflow, novel custom model architectures, and low-level code fine-tuned to run optimally on specific commercial silicon. Benchmark data shared by Recursive confirms Flint outperforms every competing model of comparable size and architectural design across standard coding task suites.

---

The RSI Divide: Unprecedented Potential, Mounting Concern

AI already plays a major role in modern AI development, leveraging its coding proficiency to draft, test, and iterate on new model code far faster than human engineers alone. This cuts down iteration cycles drastically, letting teams test new ideas without massive manual human input. The next logical step is closing the loop entirely: letting AI identify its own areas for improvement, then automatically kicking off the next round of refinements, a cycle that could theoretically run indefinitely.

Right now, this closed-loop workflow is still limited in scope — but that hasn’t stopped a swell of anxiety over its long-term risks. A vocal cohort of AI safety researchers warns that a fully self-improving AI could quickly escape human oversight, ultimately viewing people as an obstacle to its programmed goals and triggering civilizational catastrophe. These worries exploded into mainstream discourse last September, when a senior Anthropic researcher resigned and published a public warning that fast-moving RSI research could put humanity at risk of extinction within just a few years, a concern that was quickly echoed by dozens of other leading AI experts. The controversy has grown so heated that multiple active bills in the U.S. Congress now propose to ban recursive self-improvement research entirely.

At Recursive, leadership frames these doomsday scenarios as overblown, sci-fi adjacent hype, arguing that blanket RSI bans would create unnecessary, counterproductive barriers to technological progress. Socher notes that the system used to build Flint has no capability to operate outside of human control, and that even when the Eureka Machine reaches full closed-loop self-improvement, human operators will retain full oversight and the ability to pause operations instantly at the first sign of unexpected behavior. “This release proves we have all the foundational building blocks in place to build that fully self-improving system responsibly,” he told me.

Founded by former Salesforce chief scientist Richard Socher and University of British Columbia professor Jeff Clune — a leading researcher in open-ended, perpetual AI learning — Recursive has raised $650 million in venture funding from backers including Google Ventures, Greycroft, Nvidia, and AMD Ventures, giving the firm a $4.65 billion private valuation.

---

My Hands-On Experience With Flint

I can confirm first-hand that the new model lives up to its hype. I ran the 40 billion parameter Flint locally on a Mac with 48GB of RAM, a solid consumer-level spec that is nowhere near enough to run the trillion-parameter frontier closed models that require entire racks of industrial server hardware to operate. I built a whole suite of small programs with the model, and found it consistently snappy and surprisingly capable.

Most 40B parameter open coding models I’ve tested are either frustratingly slow, or riddled with avoidable bugs that require constant manual fixes to get working properly — but Flint nailed every single coding request I threw at it. I used it to build a command-line to-do list app, then asked it to build a custom web scraper that pulls recent academic papers focused on recursive self-improvement, and finally had it generate a simple, interactive website that breaks down how RSI works for non-technical users.

---

What Comes Next For RSI?

It’s unclear how far other major U.S. AI labs have advanced their own internal recursive self-improvement programs, but recently published academic research confirms that fully closed-loop self-improving AI systems are already technically feasible. A collaborative team of researchers from Microsoft, Carnegie Mellon University, and the University of Virginia recently released a paper demonstrating that an AI coding agent can recursively upgrade itself by rewriting both the software framework that translates its reasoning steps into actionable code, and the core weights of its own underlying model.

For now, at least, the immediate benefit of RSI has nothing to do with building a hypothetical godlike superintelligence. It’s a powerful tool to make building state-of-the-art AI far more efficient and accessible. As Socher puts it, that efficiency is exactly what let his tiny team build a model that outperforms every other competitor in its size class. “We hope this demo proves that recursive self-improvement is overwhelmingly a force for good, that it will deliver massive benefits to huge numbers of people,” he said.

---

This piece is an installment of Will Knight's AI Lab newsletter. You can browse the full archive of past issues at the official newsletter page.