Hinton and 22 researchers publish first RSI paper on AI R&D automation and intelligence explosion

The intelligence explosion is not a new topic. The paper revisits the idea proposed in 1965 by mathematician I.J. Good: if a machine is better than humans at designing machines, it can keep designing ever-smarter versions, forming a self-reinforcing feedback loop. This is regarded as the theoretical origin of the intelligence explosion.

The paper "What if automating AI R&D triggers an intelligence explosion?", co-authored by Hinton and others, brings this question back into an engineering context. The paper has 22 authors in total, including Geoffrey Hinton, Yoshua Bengio, Andrew Barto, and OpenAI chief scientist Jakub Pachocki. Reports describe it as Hinton's first recursive self-improvement (RSI) paper; this claim comes from media paraphrase, not from the paper itself.

The paper is no longer concerned with whether a single AI can modify its own code, but with whether AI can enter the entire process of developing the next generation of AI and make that process itself increasingly fast.

The paper cites internal data from Anthropic and other labs, saying the share of approved code generated by AI rose from low single digits in January 2025 to over 80% by May 2026. The same data also shows that in March 2026, Claude could autonomously complete about 1% of R&D work when receiving only high-level human supervision, and by August 2026 this share rose to 26%. These figures come from the labs' own accounts and lack independent verification.

The paper also mentions that as of September 2026, OpenAI's internal AI systems could frequently complete R&D tasks that originally took human employees several days.

But the paper has a clear definition of "complete R&D": proposing hypotheses, designing experiments, implementing experiments, analyzing results, identifying the causes of failure, and adjusting the direction. Only when these stages are linked into a long chain does AI approach the role of a true researcher. Automating any single stage can only count as AI-assisted research. Anthropic also emphasizes that Claude has not yet reached a fully unsupervised level in any category of tested AI R&D work.

Combining benchmark trends and internal lab progress, the paper argues that in the coming years AI may be able to automate most AI R&D work, or even achieve full automation. This is the paper's extrapolation, not a fact that has already occurred.

The paper calls the path in which AI participates in developing the next generation of AI, and the stronger AI then participates in the next round of R&D, a software-driven intelligence explosion, that is, an intelligence explosion driven by software improvement.

This path differs from building faster chips. Chip design, tape-out, and factory construction all have long cycles, whereas once algorithms, training methods, agent workflows, and synthetic data strategies are verified as effective, they can in theory be redeployed very quickly into the next round of R&D. AI R&D largely takes place in digital environments, code can be executed directly, experimental results can be returned quickly, and model capabilities can be immediately verified through metrics such as benchmarks, loss, and reward.

What the paper really cares about is whether the feedback cycle of AI progress will become shorter and shorter. As long as AI can take over the development of the next-generation model more and more completely, a self-improvement feedback loop has already appeared.

The paper introduces the concept of an effective R&D workforce. The authors estimate that if AI reaches the level of top AI researchers and inference costs are comparable to current frontier models, the existing compute of a leading AI company could theoretically support AI R&D labor equivalent to at least millions of top human researchers. Today, the research team of a frontier AI lab is still only on the order of a few thousand people.

The key difference between AI researchers and human researchers is replicability: training a top human researcher takes years of education, while AI researchers can simply add instances and can also work 24 hours a day; after the underlying model is updated, the entire AI R&D team can be upgraded simultaneously. This bypasses the hardest growth limit in human scientific research: the speed of talent cultivation.

But the paper cautions that 1 million AI researchers will not automatically bring 1,000 times the R&D speed. Scientific research has diminishing marginal returns; a large amount of repetitive work, competition for experimental resources, and the increasing difficulty of finding good ideas will all dilute the returns from added labor.

The paper cites research on historical data from three AI research subfields, saying the central estimate of the return on research investment r is about 1.2 to 1.9. r measures how much the pace of technological progress can rise after research investment increases. When r is low, added investment quickly diminishes; when r is high enough, newly added R&D capacity will further produce stronger AI researchers, and only then does the feedback loop begin to accelerate.

Under assumptions such as full automation of AI R&D, sustainable returns to research, and no short-term compute or data bottlenecks, the model concludes that the pace of AI technological progress could increase tenfold in roughly 1.5 years, equivalent to compressing one year of today's progress into about 5 weeks. The paper also estimates that if, after full automation of AI R&D, only the current trend of software efficiency improvement is maintained, the automated R&D workforce could expand a hundredfold within months to years.

These are model extrapolation results, not predictions. They depend on multiple undetermined conditions, including whether AI can reliably chain together dozens of R&D steps over long time spans and continuously advance the same research goal.

The overall tone of the paper is that current evidence is far from sufficient to prove an intelligence explosion has already occurred, and the productivity gains brought by AI R&D automation have not yet clearly crossed the threshold that triggers explosive acceleration. The authors caution that between AI R&D automation and an intelligence explosion lies a long string of conditions: first, AI must be reliable enough to work autonomously for long periods; second, it must actually improve R&D output; and only then must these new R&D results in turn significantly enhance AI's own R&D capabilities.

Even if millions of AI researchers appear in the future, compute constraints remain: AI researchers can be replicated, but GPUs cannot be replicated out of thin air, and many key experiments still have to queue for chips; training a truly frontier model may itself take months and cannot be compressed by parallel agents.

Data constraints also exist: natural internet data will not grow in sync with the number of AI researchers, and whether high-quality training signals can be supplemented through synthetic data, verifiable tasks, environmental interaction, and other means currently has no definitive answer.

Experiment-time constraints and diminishing marginal returns in scientific research form the other two walls. Training models, chip manufacturing, and the construction of new data centers all take time, and experiments cannot be parallelized indefinitely; easy algorithmic improvements may be found first, and the cost of subsequent breakthroughs becomes increasingly high. If the difficulty of scientific research rises faster than AI R&D capability improves, the intelligence explosion may also fail to get off the ground.