The Pulse
Researchers Warn AI R&D Could Compress Years of Progress Into Weeks
A paper published September 28 by researchers including Geoffrey Hinton and leaders from OpenAI and Anthropic examines whether automating AI research could sharply accelerate AI progress. The authors urge governments to track companies’ int
AI.info Team · Written with AI ·
If AI research and development becomes fully automated, the pace of AI progress could increase tenfold in about 18 months, according to a new paper’s analysis of historical data. Under that scenario, a year of progress at today’s pace could take about five weeks. The authors stress that this estimate depends on several uncertain assumptions, including that current returns from research effort persist and no other bottlenecks emerge.
The paper, published September 28 by the Cambridge Programme on AI Science & Policy, examines whether AI systems helping to build more capable successors could produce an “intelligence explosion”: years of advances compressed into months or less. Its 22 authors include Geoffrey Hinton, Yoshua Bengio, OpenAI’s Jakub Pachocki and Anthropic’s Jack Clark. They argue that the prospect warrants government attention even though the evidence remains preliminary and the outcome is far from certain.
The tenfold projection depends on full automation
The authors describe a feedback loop: AI systems take on more AI R&D tasks, helping produce improved systems that can do a larger share of that work. Once systems reach expert-level research capabilities, the paper estimates, the computing resources available to a single frontier AI developer could support an AI workforce equivalent to millions of top human researchers. That estimate describes potential capacity, not a demonstrated workforce or a forecast that the acceleration will occur.
The paper draws on estimates of the “returns to research effort” in three AI research areas. In the authors’ account, those estimates suggest that diminishing returns might not cancel out growth in the effective research workforce. But the paper says the data are limited and the result depends on modeling assumptions; its tenfold projection applies only if those returns remain in the estimated range and other constraints do not intervene.
AI already handles parts of AI development
The authors point to reported changes inside leading AI companies. Anthropic reported that AI systems’ share of approved code rose from low single digits in January 2025 to more than 80% in May 2026. The share of R&D work completed autonomously with only high-level human supervision rose from 1% in March to 26% in August 2026, according to the paper.
Those figures measure different things: approved code is not the same as autonomous R&D work. The paper also says AI systems can now complete some research and development tasks that take human experts hours or days, while noting that systems still make mistakes, misrepresent work and require human intervention. The authors cite a tentative extrapolation suggesting that months-long AI R&D projects could be automated by mid-2028, but caution that benchmarks do not always translate into real-world productivity.
Compute, data and long experiments could slow the loop
The authors identify several possible brakes on acceleration. AI research needs computing resources and data; some high-value tasks may resist automation; and long training runs can delay the next system even if researchers find improvements quickly. The paper notes that frontier training runs can take three months or more, and says evidence on how strongly such delays constrain progress remains limited.
That uncertainty cuts both ways. Better systems might automate more of the research process, but the authors say productivity gains have not yet reached the threshold needed to trigger an intelligence explosion. They argue the gap may be narrowing, while emphasizing that no single capability benchmark establishes when, or whether, a self-reinforcing acceleration would begin.
The proposed response starts with visibility
The paper urges policymakers to require standardized reporting on AI R&D automation and the pace of AI progress, and to consider independent evaluations or auditors embedded in companies. It also proposes safety requirements for continued development, options to pause specific research workloads, and isolated environments for some tests and deployments. The authors say governments should prepare response plans for severe labor-market impacts, geopolitical instability and possible loss of control over highly capable systems.
They also acknowledge the risks of the policy tools they propose: poorly designed restrictions could let a government slow competitors while favoring a chosen company, and delayed progress carries costs of its own. Their warning is not that an intelligence explosion is inevitable, but that society may have little time to respond if rapid acceleration does occur. “Once an intelligence explosion begins, the window for action may close,” the authors write.
“Once an intelligence explosion begins, the window for action may close.”
Alan Chan et al., “What if automating AI R&D triggers an intelligence explosion?”