AI labs across the globe are sharpening their focus on recursive self‑improvement (RSI), a concept that envisions machines that can continuously upgrade themselves without human input. The push mirrors the hype that once surrounded artificial general intelligence, but the technical hurdles remain formidable.

Earlier this month, former Meta and Salesforce researcher Richard Socher announced the launch of Recursive Superintelligence, a venture explicitly built around RSI. Socher told TechCrunch that the goal is to automate the entire research pipeline—ideation, implementation, and validation—so that the system can improve at scale without human oversight.

Parallel to Socher’s effort, former Tesla and OpenAI lead Andrej Karpathy is running a project he calls Auto‑Research. By deploying swarms of agents to train large language models on modest tasks, Karpathy hopes to achieve incremental gains that could eventually cascade into a self‑improving loop. He has been unusually transparent, posting milestones on Twitter and releasing code on GitHub. So far, the improvements sit at a GPT‑2‑scale level, and Karpathy admits they are “not novel, ground‑breaking ‘research’ (yet).”

Adaption, a startup founded by Cohere and Google alum Sara Hooker, entered the arena with AutoScientist, a tool that trains agents to make incremental improvements on frontier‑scale models. The aim is similar to Karpathy’s: simplify the training of massive models by letting AI agents handle the fine‑tuning steps that traditionally require human engineers.

Concrete success stories are emerging. Disarray founder Doris Xin saw her self‑trained machine‑learning agent capture 28 medals in a recent Kaggle competition, outpacing many human‑trained entries. Xin points to reliability as the primary obstacle, arguing that with “infinite compute and infinite time horizon, we are already there.”

Industry leaders remain cautious. Google CEO Sundar Pichai, in a recent podcast interview, acknowledged progress but warned that the kind of acceleration described by RSI is still beyond reach. “It’s a continuum, and we are all definitely making progress,” he said, adding that the next level would have “a lot of implications, but we aren’t quite there yet.”

Anthropic’s own internal tools hint at how quickly the boundary is shifting. Its Claude Code system now writes close to 100 % of the code its engineers use, effectively authoring itself. A survey of Anthropic staff revealed that five out of 18 engineers believe the upcoming Mythos model could replace a mid‑level (L4) programmer, though they note weaknesses in self‑management, priority setting, and verification—skills essential for true RSI.

Experts stress that using AI to assist research does not equal RSI. Helen Toner, director of Georgetown’s Center for Security and Emerging Technology, told TechCrunch that “they’re just using AI for as much as they can,” which falls short of the classic definition where no humans are needed. Ajeya Cotra of the Machine Intelligence Research Institute outlines a roadmap: first “adequacy,” where an AI can conduct research alone; then “parity,” matching human performance; and finally “supremacy,” surpassing collaborative human‑AI teams. Cotra believes the adequacy threshold may be within a few years, but parity and supremacy remain speculative.

The community remains split on timelines. A CSET‑organized expert panel last year found opinions ranging from imminent superintelligence explosions to slower, plateau‑bound progress. All agree, however, that recursion makes forecasting especially difficult.

As funding pours in and compute resources expand, the engineering and alignment challenges loom large. Infinite compute is not a reality, and balancing human labor against machine intelligence will be a persistent tension. For now, recursive self‑improvement stays a tantalizing promise—one that the AI field is racing toward but has yet to fully capture.

This article was written with the assistance of AI.
News Factory SEO helps you automate news content for your site.