AI Progress Could Compress a Year into Five Weeks: Hinton, Bengio, and 22 Top Scholars Warn 'Intelligence Explosion' Is Closer Than Thought

The University of Cambridge's AI Science and Policy Programme (CASP) has released a paper co-authored by 22 frontier experts in AI, titled "What If Automated AI R&D Triggers an 'Intelligence Explosion'?" The author list includes Nobel laureate and Turing Award winner Geoffrey Hinton, Turing Award winners Joshua Bengio and Andrew Barto, as well as OpenAI chief scientist Jakub Pachocki, Anthropic co-founder Jack Clark, and Microsoft chief scientific officer Eric Horvitz, among others.

This roster itself constitutes a signal: those discussing the "intelligence explosion" are no longer just the longtermist circle, but include senior figures at the frontline institutions pushing frontier models. Yet the paper's conclusions are not radical. The authors collectively hold that although there is currently no evidence that an "intelligence explosion" process has occurred, and productivity gains from AI R&D automation have not crossed the corresponding threshold, the conditions supporting this feedback loop are maturing rapidly compared with a few years ago.

In other words, the paper's position is "not yet happened, but conditions are changing." This distinction matters: it neither treats the trend as an accomplished fact nor dismisses the trend as unworthy of monitoring simply because evidence is lacking.

The internal data from frontier labs cited in the paper is the most concrete part of the whole text and also the part that most requires caution. Internal data published by Anthropic this year shows that as of May 2026, over 80% of the code merged into the company's codebase was written by Claude, whereas when Claude Code entered research preview in early 2025, that share was still in the single digits.

The same source also claims that by August 2026, in AI R&D work evaluated internally at Anthropic, Claude was able to "lead" 26% of tasks while receiving only high-level human supervision, compared with about 1% a few months earlier; over 90% of R&D work had at least entered a deeper level of human-AI collaboration. On OpenAI's side, by mid-August, converted to an 8-hour workday, every one human workday invested by OpenAI's research department already corresponded to about 3.1 agent workdays.

These figures come from corporate self-reports and lack independent third-party verification, so they can only serve as corroboration of a trend, not as confirmed industry fact. On this basis, the paper notes that today's strongest AI systems can already complete some R&D tasks that previously required professional researchers hours or even days, whereas 2023 models could typically only reliably complete tasks measured in seconds—this comparison describes a change in the order of magnitude of task duration, not that capability has crossed the automation threshold.

The trend extrapolation cited in the paper suggests that if recent capability growth continues, by mid-2028 some AI R&D projects that originally required a human team months to advance could enter the scope of automation, but the authors explicitly describe this as a preliminary extrapolation.