“Recursion” has emerged as the latest focal point in artificial intelligence discussions. Two nascent companies have adopted the name, and a growing number of organizations are integrating Recursive Self-Improvement (RSI) into their strategic roadmaps. Much like Artificial General Intelligence (AGI) before it, RSI has become a shorthand for a potentially transformative, even cataclysmic, AI advancement, despite some ongoing debate regarding its precise definition.
Fundamentally, RSI describes an AI system capable of continually enhancing itself. Once AI systems can manage their upgrade cycles more effectively than humans, this process could evolve into a self-contained loop, constrained solely by available computational power, rendering human involvement unnecessary or even counterproductive.
Regardless of its perceived threat, this vision is one that numerous AI laboratories are actively pursuing.
Earlier this month, renowned AI researcher Richard Socher unveiled Recursive Superintelligence, a venture explicitly targeting RSI. Socher informed TechCrunch at its launch, “Our main focus is to build truly recursive, self-improving superintelligence at scale, which means that the entire process of ideation, implementation, and validation of research ideas would be automatic.”
Several other distinguished researchers are independently working toward this identical objective, seeking the breakthrough that will enable recursive self-improvement.
Among the most prominent is Alex Karpathy, a notable figure from Tesla and OpenAI, who is leveraging agent swarms to train large language models (LLMs) on basic tasks for a project dubbed Auto-Research. Karpathy has maintained an unusual degree of transparency about this initiative, regularly sharing milestones via Twitter and making its foundational components accessible through a public GitHub repository. To date, the work has predominantly focused on minor enhancements to a GPT-2 scale model—as Karpathy observed in March, “It’s not novel, ground-breaking ‘research’ (yet)”—yet it has inspired many other researchers to pursue the RSI ambition. With Karpathy now engaged in pre-training efforts at Anthropic, he will have ample opportunities to apply this concept on a grander scale.
Adaption, founded by Cohere and Google alumna Sara Hooker, recently introduced a comparable tool named AutoScientist, designed to automate frontier training. Similar to Karpathy’s auto-researchers, this system trains agents to enact incremental improvements. However, Adaption’s primary goal is to streamline the training of full-scale frontier models. Should these systems empower researchers to advance the frontier, the process could rapidly escalate into a scenario closely resembling RSI.
Doris Xin, founder of Disarray, garnered significant RSI-related interest when her self-trained machine learning agent secured 28 medals in a recent Kaggle competition, surpassing numerous human-trained counterparts. Xin identifies reliability as the principal challenge in this domain.
“I would argue, given infinite compute and infinite time horizon, we are already there,” Xin stated. “I want to make an argument that this is not a creative endeavor, really. It’s just a lot of meat-and-potatoes engineering.”
Conversely, substantial evidence suggests that the AI industry is not yet meaningfully close to achieving recursive systems and continues to struggle with communicating its progress to a cautious public. Google CEO Sundar Pichai tacitly acknowledged this in a recent podcast interview.
“It’s a continuum, and we are all definitely making progress,” Pichai commented. “But in the way people describe R.S.I., that would represent a next level of acceleration and would have a lot of implications, but we aren’t quite there yet.”
Nevertheless, this continuum already encompasses a considerable number of self-improving AI systems. In January, a lead programmer for Anthropic’s Claude Code estimated that “close to 100%” of his team’s code was generated by the tool—a candid admission that Claude Code was, in essence, writing itself.
While engineers utilizing an AI tool doesn't automatically mean the tool can replace them, Anthropic appears to be approaching a point where AI could substitute for human engineers. A recent survey associated with the Mythos preview revealed that five out of 18 Anthropic engineers believed that, with further harness improvements, this iteration of Mythos could soon replace an L4 engineer—a mid-level programmer capable of undertaking complex projects autonomously.
However, some predictable weaknesses persisted.
The report detailed, “Some of Claude’s major reported weaknesses compared to an L4 include: self-managing week-long ambiguous tasks, understanding org priorities, taste, verification, instruction-following, and epistemics.”
In essence, its shortcomings lie in all aspects related to self-direction, which forms the fundamental basis of RSI. Yet, for most other tasks, Claude is deemed ready to contribute.
Just as with the term AGI, the AI industry remains unable to specify how far away it is from demonstrating a significant recursive system. When Georgetown’s Center for Security and Emerging Technology convened a panel of experts to examine RSI last year, the group’s assessments were sharply divided—some anticipated an imminent “superintelligence”-style explosion, while others predicted slower, more gradual progress leading to an eventual plateau. All participants, however, concurred that recursion rendered the future particularly unpredictable.
Helen Toner, director of CSET and a former OpenAI board member, informed TechCrunch that simply employing AI tools for AI research is insufficient to meet the criteria for RSI. “They’re just using AI for as much as they can,” Toner explained to TechCrunch. “And I think that is different from the classic definition of RSI, which is really that there are no humans needed.”
Toner referenced a recent post by METR’s Ayeja Cotra, which outlines distinct milestones on the path to AI research autonomy. One stage, termed “adequacy” by Cotra, would be reached when the system can conduct research without any human involvement, even if the output isn't as valuable or efficient. “Parity” occurs when an AI-only system performs research as effectively as a human-only system. “Supremacy,” the ultimate stage, is achieved when an AI-only system surpasses a collaborative human-AI research system.
Ultimately, Cotra concludes that AI is very close to the “adequacy” threshold, capable of generating some work independently—akin to the incremental advancements seen in Karpathy’s Auto-Research system. Cotra writes, “I wouldn’t be totally shocked if you told me this milestone had already passed, and I expect it to happen in the next couple years.”
She is less certain about the timeline for “parity,” but believes that once attained, it would “massively accelerate the pace of AI progress, leading to AI research supremacy within another year.”
Given that much of AI development is predicated on scaling laws, there's a strong inclination to assume RSI will follow a similar trajectory. Toner suggests that many pursuing AI research and development through RSI “think of it as a pretty smooth ladder, where you can just keep scaling up.”
However, even if AI researchers can achieve incremental improvements akin to Karpathy’s auto-researchers, greater challenges lie in fully automating the entire research process. Toner draws parallels to the history of computing, where humans have progressively delegated more processes while still maintaining overall direction.
“We went from machine languages to assembly language and compiled languages; you’re getting further and further from the guts of the computer,” Toner elaborated. “But the human is still, in some intuitive sense, running the show.”
Transcending this paradigm will necessitate overcoming substantial engineering and alignment obstacles. Despite significant investments, infinite computational power is not available, and the fundamental trade-off between human labor and machine intelligence will be difficult to resolve.
As for a complete recursive AI system leading to apocalyptic scenarios? The one point on which researchers largely concur is that, much like AGI, it has not yet arrived.
The Editorial Staff at AIChief is a team of professional content writers with extensive experience in AI and marketing. Founded in 2025, AIChief has quickly grown into the largest free AI resource hub in the industry.