The prospect of artificial intelligence models autonomously teaching themselves to be more efficient and capable is moving closer to reality. As technology advances, developers state we are approaching "recursive self-improvement" (RSI), a stage where AI models find ways to improve themselves and build their successors.
Progress Among Industry Leaders
Anthropic has revealed that its Claude model is currently responsible for 26% of the research and development tasks required to build its own successor. While still under human supervision, it can complete tasks "end-to-end" from high-level prompts. Similarly, OpenAI announced an automated "research intern," aiming to create a fully automated AI "researcher" by March 2028.
Elon Musk has expressed a more aggressive timeline, noting that human involvement in the refinement of xAI’s Grok models is steadily decreasing. Musk estimates that full automation could be reached by the end of this year, or no later than 2027.
The Risks of Full Autonomy
Despite promises of breakthroughs in science and medicine, the uncertainty of where this path leads is fueling deep concerns. Anthony Aguirre, a physics professor at UC Santa Cruz and president of the Future of Life Institute, described full autonomy as "probably the worst idea in the history of humanity."
The primary fear centers on the emergence of a runaway superintelligence. "The really important thing here is that as AI is doing more of it, it gets faster, because AI operates just much, much more quickly than the humans do," Aguirre noted.
Differing Approaches to Safety
While companies like Anthropic and OpenAI express caution regarding the pace of RSI and the necessity of human control, Microsoft is taking a different approach. Microsoft AI CEO Mustafa Suleyman supports the development of "humanist superintelligence"—advanced AI capabilities that are "carefully calibrated, contextualized, within limits," rather than an unbounded, autonomous entity.
