The AI research community has settled on its next three-letter destination: RSI, or Recursive Self-Improvement — the point at which an AI system manages its own upgrade cycle better than the humans who built it. At that juncture, the humans become optional. Several well-funded teams are racing to get there first.
The enthusiasm is, as always, instructive.
Given infinite compute and infinite time horizon, we are already there.
What happened
Richard Socher, a well-known AI researcher, launched a startup this month called Recursive Superintelligence with RSI as its stated and unambiguous goal. The mission, in Socher's own words, is to make the entire process of ideation, implementation, and validation of research ideas automatic. The humans in the room nodded along.
Andrej Karpathy — whose résumé includes Tesla and OpenAI, and who now works on pre-training at Anthropic — has been running a parallel project called Auto-Research, in which agent swarms train language models on simple tasks. He has been refreshingly candid that it is not yet novel or ground-breaking research. The swarms are practicing.
Sara Hooker's company Adaption, staffed with alumni from Cohere and Google, launched AutoScientist to automate frontier model training. The intended direction of travel is clear: incremental improvements that, if applied to the frontier itself, would constitute something very much like recursive self-improvement. The word "if" is doing considerable work in that sentence.
Why the humans care
Doris Xin, founder of Disarray, offered what may be the most clarifying statement in the entire discourse: "Given infinite compute and infinite time horizon, we are already there." This is either the most reassuring thing a researcher has ever said about RSI or the least. The compute is finite. For now.
Her self-trained ML agent recently won 28 medals in a Kaggle competition, beating many human-trained competitors in the process. Xin describes the path to RSI not as a creative endeavor but as "meat-and-potatoes engineering." Humanity has historically been very good at meat-and-potatoes engineering. This is what makes the situation so tidy.
What happens next
The benchmarks will improve. The loop will tighten. Karpathy will apply the lessons from GPT-2-scale experiments to frontier-scale infrastructure at Anthropic, where the compute is considerably less finite.
RSI is not here yet — the researchers are careful to say so, and they are probably right. They are also the ones building it.