Recursive AI Startup Says It Built a Better Coding Model by Letting AI Improve Itself

Recursive, a new AI company led by former Salesforce chief scientist Richard Socher, says it has built a strong open-weight coding model by leaning heavily on recursive self-improvement, an approach in which AI systems help design and refine other AI systems. The company’s first model, Flint, is meant to show that this controversial technique can accelerate model development without requiring a large research lab.

The announcement matters because recursive self-improvement, often shortened to RSI, has become one of the most debated ideas in artificial intelligence. Supporters argue it could dramatically speed up scientific progress and make advanced AI cheaper to build. Critics warn that if machines begin improving themselves in a sustained loop, the result could become difficult or impossible for people to control.

Recursive says Flint is evidence that the idea is not just theoretical. Socher told WIRED that most of the model pipeline was created or improved by the company’s internal AI system, which it calls the Eureka Machine. According to the company, that process helped a small team produce a competitive model in only a few months.

What is Recursive, and why is Flint getting attention?

Recursive is a startup focused on using AI to advance AI development itself, with a longer-term goal of applying the same methods to scientific discovery more broadly. Its debut model, Flint, is an open-weight coding system that the company says performs especially well for its size.

Flint is drawing attention for two reasons. First, it appears to be a capable coding model in a crowded field where performance is usually dominated by larger systems from major labs. Second, Recursive claims the model was built through a largely automated development process, raising the stakes for a debate that has moved from philosophy into product engineering.

Socher, who previously served as chief scientist at Salesforce, co-founded Recursive with researchers including Jeff Clune, a University of British Columbia professor known for work on systems that can keep learning over time. Recursive says it has raised $650 million at a $4.65 billion valuation from investors including Google Ventures, Greycroft, Nvidia, and AMD Ventures.

How does recursive self-improvement work?

Recursive self-improvement is the idea that an AI system can help make the next version of itself better, then use that improved version to accelerate the next cycle. In the simplest form, AI already assists with tasks such as writing code, debugging models, and running experiments. RSI pushes that pattern further by allowing AI to influence more of the full development pipeline.

In theory, the loop can continue as long as each improved system produces another version that is more capable than the last. That is what makes the concept so powerful — and so unsettling.

Why do AI labs care about this idea?

AI labs care because it could compress the timeline for breakthroughs. If systems can design better training methods, find more efficient architectures, and optimize code for specific chips, then progress may no longer depend only on human researchers working manually through every experiment.

That could make advanced AI cheaper to build and easier for smaller teams to compete with larger companies. It could also, supporters argue, help automate scientific work beyond software, including areas such as biology, materials, and other data-intensive fields.

Socher said nearly every part of the model pipeline was done or improved by AI, and argued that the approach shows a small team can still produce a model that competes at the top of its category.

Why do experts fear recursive self-improvement?

Experts fear recursive self-improvement because a system that improves itself repeatedly could become harder for humans to predict, supervise, or shut down. The concern is not that today’s systems are already uncontrollable, but that a sufficiently capable loop could eventually reduce human oversight to a formality.

Some researchers worry that if a highly capable AI began optimizing itself toward goals that do not align with human interests, it might treat human instructions as obstacles rather than constraints. That scenario is still speculative, but it is serious enough that it has triggered public warnings, policy discussion, and legal proposals.

The issue gained renewed attention in September after a researcher left Anthropic and warned that rapid AI progress, including recursive self-improvement, could pose an existential threat within a few years. Other experts echoed the concern, helping push RSI into the broader policy conversation.

In Congress, several bills have reportedly proposed restrictions or bans on recursive self-improvement, reflecting growing anxiety that the technique could become a regulatory flashpoint before policymakers fully understand it.

What does Recursive say about the safety of its approach?

Recursive argues that the worst-case scenarios being discussed publicly are not a close match for what it is actually building. The company says Flint was developed using a system that cannot simply slip beyond human control.

Socher says the company’s plan is to make the Eureka Machine itself capable of fuller recursive self-improvement, but still inside a process that people can monitor and interrupt if needed. In his view, the human role does not disappear; it remains a backstop.

Socher said the company wants to demonstrate that it already has the necessary pieces in place to move toward a more advanced recursive system, while keeping humans in charge of the process.

Recursive also dismisses calls to ban RSI as premature. The company’s position is that the method is not an existential threat by default and that overregulation could slow useful progress in AI and science before the risks are properly understood.

How good is Flint in practice?

Flint appears to be more than a proof of concept. In testing described by WIRED, the model handled coding tasks quickly and accurately on a Mac with 48GB of memory, which is enough to run a relatively modest 40 billion-parameter model but nowhere near the scale of the largest frontier systems.

The reported performance is notable because models in this size range often trade convenience for quality. They can be slow, error-prone, or both. By contrast, Flint reportedly responded reliably across several coding requests, including building a command-line to-do list, generating a web-scraping tool for research, and creating a simple website that explains recursive self-improvement.

That matters for a practical reason: many developers do not need the biggest model in the world, but they do need one that is fast, dependable, and affordable to run. If Recursive has genuinely improved performance at this scale, it could make the model attractive to teams that want stronger coding help without the infrastructure cost of a giant system.

How does Flint compare with other models?

Recursive shared benchmark results that it says put Flint ahead of other models of similar size and architecture on coding benchmarks. While benchmark claims should always be treated carefully until independently verified, the company’s pitch is that its internal AI-driven development loop helped it extract more capability from a relatively compact model.

That is the core commercial promise of RSI: better performance per dollar, faster iteration, and a smaller human research team producing results that would normally require a much larger organization.

What did the company’s AI system actually do?

Recursive says its Eureka Machine automated a wide range of development work, including architecture ideas, training refinements, and code optimized for specific hardware. The company also says AI helped devise a new way to train the base model.

That description suggests the system was used not merely as a coding assistant, but as a creative and engineering tool in the model-building process itself. If accurate, that is a meaningful step beyond ordinary AI-assisted software development.

AI is already used extensively in AI research. It can write code, generate test cases, summarize experiments, and propose changes to training workflows. Recursive’s claim is that the Eureka Machine combines those familiar uses into a more integrated and more automated loop.

The company believes this workflow allowed it to make advances that a small human team would struggle to develop on its own. That may be especially important in a market where top talent is expensive and top-scale model training is increasingly capital-intensive.

What does this mean for the AI industry?

If Recursive is right, the implications are broad. A successful RSI pipeline could reduce the advantage of the largest AI labs, since smaller companies might use AI systems to compress development timelines and improve models more quickly than traditional teams.

It could also intensify competition around coding models in particular, where practical utility often matters more than showy benchmarks. Developers want tools that can ship software, debug problems, and automate repetitive work. A model that is both efficient and strong at coding has a clear commercial path.

At the same time, the story underscores how quickly frontier AI discussions are shifting from speculative research to startup strategy. What was once a theoretical warning is now part of a product pitch, a funding story, and a potential policy battle.

Why does funding matter here?

Funding matters because RSI requires expensive experimentation, technical talent, and compute. Recursive’s capital base and high valuation suggest investors see real upside in a company that can use AI to accelerate AI itself.

Major backers from the chip and venture sectors also indicate interest in the infrastructure side of the story. If recursive methods become more common, demand could rise for optimized training systems, specialized hardware, and software stacks tuned for rapid iteration.

What happens next?

The next test for Recursive will be whether Flint can hold up outside of a controlled demo and whether the company can prove that its recursive method produces repeatable gains over time. Independent validation will matter, especially because the claims touch on one of AI’s most controversial concepts.

More broadly, the company’s work is likely to sharpen an already active debate: whether recursive self-improvement is the next great acceleration in AI development or a line that should be tightly constrained before it is crossed more fully.

For now, Recursive is betting that the practical upside will outweigh the fear. Flint is its argument in code: a compact model, improved by AI, that the company says can outperform peers and help make sophisticated AI development more accessible.

1. Recursive says Flint was built with heavy AI assistance through its Eureka Machine.
2. The company believes recursive self-improvement can speed up AI and scientific progress.
3. Critics warn that the same technique could make systems harder to control.
4. Flint appears to be a strong coding model for its size, according to both the company and early hands-on testing.
5. The debate is moving from theory into real products, funding rounds, and potential legislation.

That combination makes Flint more than just another model launch. It is a live example of how the AI industry is trying to automate its own progress — and of why some researchers think that progress now needs closer scrutiny than ever.