Hinton Co-Authors First RSI Paper, Asking Whether Automated AI R&D Could Trigger an Intelligence Explosion

The paper's central concern is not whether a single AI can rewrite code or tune parameters, but whether AI can enter the full pipeline of developing next-generation AI and make that pipeline itself faster. According to QbitAI, the authors argue that once AI can fully participate in building its successors, the RSI loop could close. If AI R&D automation crosses a threshold, progress that now takes years could be compressed into months or less, and a future AI lab's R&D workforce could theoretically grow from a few thousand people to millions of researchers. In an extreme case, a year's AI progress could take about five weeks. The paper also says an intelligence explosion may not require superintelligence first; high automation efficiency could start the feedback loop.

The paper traces the idea to mathematician I.J. Good, who in 1965 imagined a machine better than humans at designing machines, leading to a self-reinforcing cycle. Today's RSI debate has shifted from self-modifying code to the more practical question of whether AI can take over the entire R&D process.

QbitAI reported that the paper cites internal data from Anthropic and other leading labs. In January 2025, AI-generated approved code was only a low single-digit share. By May 2026, that share exceeded 80 percent. In March 2026, Claude could autonomously complete about 1 percent of R&D work with only high-level human supervision; by August 2026, that figure had risen to 26 percent. Anthropic stressed that the number does not mean Claude completed a quarter of all its research, but measures internal AI R&D work that AI can handle under high-level oversight. Anthropic also said Claude has not reached fully unsupervised performance in any tested AI R&D category. As of September 2026, OpenAI's internal AI systems could often complete R&D tasks that human employees needed days to finish, the paper said.

The key change is the length of tasks AI can handle independently. The paper argues that benchmark scores alone do not determine whether AI can work like a researcher; more important is whether it can chain dozens of steps and pursue a research goal over hours, days or longer. Full AI R&D includes proposing hypotheses, designing experiments, implementing them, analyzing results, diagnosing failures, adjusting direction and starting again. Automating any single step is AI-assisted research. The sensitive point is when those steps become a long chain. AI R&D is especially suited to automation because it happens in digital environments, code can be executed directly, results return quickly and model capabilities can be verified through benchmarks, loss and reward. The paper calls this path a software-driven intelligence explosion, distinct from faster chips because algorithms, training methods, agent workflows and synthetic data strategies can be redeployed quickly.

The paper's most striking section concerns copying AI researchers. If a lab has 1,000 top researchers and develops an agent at the level of a top human AI researcher, it can expand by creating more instances rather than waiting for years of education. AI researchers can work around the clock, run many tasks at once and update together when the underlying model improves. The paper introduces the concept of an effective R&D workforce. The authors estimate that if AI reaches top-researcher level and inference costs remain similar to today's frontier models, a leading AI company's existing compute could theoretically support at least millions of top-human-researcher equivalents. That is several orders of magnitude above today's frontier lab research teams of a few thousand people.

But the paper warns that a million AI researchers would not automatically produce a thousandfold speedup. Research has diminishing returns, and large numbers of researchers can duplicate work, compete for experimental resources and find fewer good ideas. The key variable is returns to research effort, or r, which measures how much faster technology advances when research input increases. The paper cites a study of historical data from three AI research subfields, with a central estimate of r around 1.2 to 1.9. Assuming full automation of AI R&D, similar returns and no sudden compute or data bottlenecks, the model finds that AI progress could accelerate tenfold in about 1.5 years: a year of progress in roughly five weeks. If only current software efficiency trends continue, the automated AI R&D workforce could expand 100-fold in months to years, through both more researchers and stronger individual researchers. The dangerous feature of an intelligence explosion, the paper says, is that the source of acceleration is itself accelerated.

Still, the paper's tone is cautious. QbitAI reported that the authors say current evidence is far from proving an intelligence explosion has occurred, because productivity gains from AI R&D automation have not clearly crossed the threshold needed for explosive acceleration. AI must first be reliable enough to work autonomously for long periods, actually increase R&D output and then produce results that significantly improve AI's own R&D capacity. Even with millions of AI researchers, several barriers remain. Compute cannot be copied like software; GPUs cannot be created from nothing, and frontier model training may take months. Data is another limit: high-quality natural data does not grow with the number of AI researchers, and it remains unclear whether synthetic data, verifiable tasks or environmental interaction can provide enough training signal. Experiment time is a third constraint: model training, chip manufacturing and data center construction take time, and a million agents would still share limited GPUs, clusters and training windows. Finally, research itself may become harder. Easy algorithmic improvements may be found first, and each further gain may require more effort. If research difficulty rises faster than AI R&D capability, an intelligence explosion will not start.

The paper does not conclude that an intelligence explosion is certain. But QbitAI reported that the evidence is enough to move the question from science fiction to a practical issue that needs preparation. RSI is moving away from its sci-fi origins. The report said that in retrospect, 2026 may be remembered not for another model scoring higher on a benchmark, but as the year humans began to realize that AI's biggest scaling might happen in the act of researching AI itself.

Editor's Summary

Geoffrey Hinton and 21 co-authors have published a paper asking whether automating AI R&D could trigger a recursive self-improvement loop and an intelligence explosion. The paper cites steep increases in AI-generated code and autonomous R&D work at leading labs, while stressing that compute, data, experiment time and diminishing research returns still block a proven explosion. It argues the question has shifted from science fiction to a real issue requiring preparation.