RSI May Be the Next AGI Challenge: Defining Recursive Self-Improvement in AI

The term "recursion" has rapidly become one of the hottest topics in artificial intelligence. Several startups have embraced the concept directly, while many AI researchers and companies are increasingly discussing recursive self-improvement, or RSI, as a major long-term objective. Much like artificial general intelligence (AGI) before it, RSI has become a shorthand for the possibility of an AI-driven intelligence explosion, even though there is still debate over what the concept truly entails.

At its core, recursive self-improvement describes an AI system capable of continually enhancing itself. Once a system becomes better than humans at improving its own capabilities, the cycle could theoretically become self-sustaining. In such a scenario, AI would repeatedly refine its own performance, limited mainly by available computing resources, while human involvement becomes increasingly unnecessary.

Whether viewed as exciting or alarming, this vision is being pursued by many AI organisations.

Earlier this month, prominent AI researcher Richard Socher launched a company called Recursive Superintelligence, openly identifying RSI as its central mission. According to Socher, the goal is to build systems capable of automatically generating research ideas, implementing them, testing results, and continuously improving the entire process without human intervention.

He is far from alone. Several leading researchers are exploring pathways that could eventually lead to recursive self-improvement.

Among the most visible is former Tesla and OpenAI researcher Andrej Karpathy. Through a project known as Auto-Research, Karpathy is experimenting with networks of AI agents that collaborate to improve language models. He has regularly shared updates publicly and released parts of the project through an open-source GitHub repository.

So far, Auto-Research has mainly produced modest improvements on smaller-scale models. Karpathy himself acknowledged earlier this year that the work is not yet groundbreaking research. Nevertheless, the project has encouraged many researchers to pursue similar ideas. Now working on model pre-training at Anthropic, Karpathy may have opportunities to test these concepts on much larger systems.

Another example comes from Adaption, a company founded by former Cohere and Google researcher Sara Hooker. The startup recently introduced AutoScientist, a system designed to automate parts of frontier AI research. Like Auto-Research, the platform relies on agents making small, incremental improvements. However, Adaption's long-term ambition is to streamline the development of next-generation AI models, potentially moving closer to recursive improvement.

Interest in the field also increased after Disarray founder Doris Xin demonstrated a self-trained machine-learning agent that earned 28 medals in a major Kaggle competition, outperforming many human-built systems. Xin believes reliability remains the biggest challenge.

"I would argue, given infinite compute and infinite time horizon, we are already there," Xin said, suggesting that much of the challenge revolves around engineering execution rather than creativity.

Not There Yet

Despite growing enthusiasm, there is substantial evidence that truly recursive AI systems remain a distant goal.

Even Google CEO Sundar Pichai recently acknowledged the gap during a podcast interview.

"It's a continuum, and we are all definitely making progress," Pichai said. "But in the way people describe RSI, that would represent a next level of acceleration and would have a lot of implications, but we aren't quite there yet."

Still, examples of AI contributing to its own development are becoming increasingly common.

Earlier this year, one of Anthropic's lead developers of Claude Code estimated that nearly all of the team's code was generated by the tool itself. While that does not mean AI engineers have become obsolete, it illustrates how deeply AI systems are already integrated into software development workflows.

Anthropic has also explored how close its models are to replacing mid-level software engineers. In evaluations connected to its Mythos research preview, several engineers concluded that future versions could potentially perform many responsibilities typically handled by an L4-level engineer.

However, significant limitations remain.

Researchers noted weaknesses in areas such as managing long-term projects, understanding organisational priorities, exercising judgement, following complex instructions, and maintaining reliable reasoning. These capabilities are precisely the skills required for autonomous self-improvement.

In short, AI may be increasingly capable in execution, but it still struggles with self-direction.

Just as debates over AGI continue, researchers also disagree about how close the industry is to achieving meaningful RSI. A study conducted by Georgetown University's Centre for Security and Emerging Technology found a wide range of expert opinions. Some anticipated rapid progress toward superintelligence, while others expected slower advancement and eventual limitations.

Helen Toner, director of CSET and a former OpenAI board member, argues that simply using AI tools to assist with research does not constitute recursive self-improvement.

"They're just using AI for as much as they can," Toner explained. "And I think that is different from the classic definition of RSI, which is really that there are no humans needed."

Toner pointed to work by METR researcher Ajeya Cotra, who describes several milestones along the path toward AI-led research.

The first stage, known as "adequacy," occurs when an AI system can continue producing research without human involvement, even if its results are less effective. "Parity" is reached when AI conducts research as effectively as humans do. The final stage, "supremacy," occurs when AI-only research outperforms even collaborative human-AI efforts.

Cotra believes AI may be approaching the adequacy threshold within the next few years. However, predicting when parity or supremacy will arrive remains far more difficult.

Challenges Ahead

Many researchers assume recursive self-improvement will follow the same scaling patterns that have driven AI progress so far. Toner cautions that the reality may be far more complicated.

She compares the challenge to the history of computing itself. Over time, humans have moved further away from low-level machine operations through programming languages and software abstractions. Yet humans remain fundamentally in control.

"We went from machine languages to assembly language and compiled languages; you're getting further and further from the guts of the computer," Toner said. "But the human is still, in some intuitive sense, running the show."

Moving beyond that model presents major technical and safety challenges. Building systems that can fully direct, evaluate, and improve themselves requires far more than incremental automation. Limited computing resources, alignment concerns, and practical engineering hurdles all stand in the way.

For now, the vision of a fully autonomous recursive AI system remains largely theoretical. While researchers continue to make progress toward more capable self-improving tools, most agree on one thing: much like AGI, true recursive self-improvement has not arrived yet.