Over 20 AI Researchers Warn of an 'Intelligence Explosion' — A Forecaster's Own Numbers Say Not Yet

A paper released September 28 by more than 20 AI researchers — including Turing Award winner Geoffrey Hinton, Turing Award winner and Mila founder Yoshua Bengio, OpenAI chief scientist Jakub Pachocki, Anthropic co-founder Jack Clark, Microsoft chief scientific officer Eric Horvitz, and UC Berkeley's Dawn Song — argues that once AI systems can automate their own research and development, progress could accelerate sharply enough to compress a year of advances into months or weeks. The paper, produced through Cambridge's Programme on AI Science & Policy, frames this as a distinct risk category from today's chatbot harms: not a model behaving badly, but a feedback loop in which AI-accelerated AI research outpaces society's ability to govern what comes out the other end.

The authors back the claim with internal figures rather than speculation. At Anthropic, the share of newly approved code written by AI reportedly rose from the low single digits in January 2025 to more than 80% by mid-2026, and the fraction of R&D work done under AI supervision climbed from roughly 1% to 26% between March and August 2026 alone. Some of the paper's own projections suggest months-long research tasks could be substantially automatable by mid-2028 — not a prediction of collapse, but evidence that the trendline the authors are worried about is already visible inside the labs they help run.

Once an intelligence explosion begins, the window for action may close.

The policy ask is notably procedural rather than alarmist: government visibility into how much AI R&D is being automated inside labs, mandatory incident reporting, independent evaluators, defined capability thresholds that trigger review, and the groundwork for an emergency brake if growth outpaces oversight. That lines up with what Bengio told the UN Security Council five days earlier, on September 23, when he called AI's dangers "real and imminent" and urged binding international coordination on loss-of-control risk specifically — suggesting the paper is less a standalone alarm than a technical companion to a governance push already underway at the UN level.

The clearest counterweight came a day before the paper's release, not after it. In a September 27 analysis, technologist and forecaster Ramez Naam used OpenAI's own published research-automation data to estimate how strong the self-improvement feedback loop would need to be to become self-sustaining — and found current performance well short. By his math, each additional point of AI research capability is currently buying only a 2-3% productivity gain on real research tasks, against a roughly 15-19% gain per point that he calculates would be needed for the loop to run away on its own — a five-to-tenfold gap. He points to AI agents managing only about 15 minutes of autonomous work at 80% reliability versus forecasts of 4-11 hours, and to a 124-fold increase in token usage over the same period producing only a 1.6-fold increase in experimental throughput, as evidence that diminishing returns, not runaway acceleration, is still the dominant pattern in the data both sides are reading.

Notably, the paper's own authors concede the scenario is "far from certain," citing compute limits, data scarcity, and hard-to-automate research tasks as plausible brakes — which is close to Naam's objection, if not his confidence. A separate and distinct critique surfaced October 1, when AI researcher Timnit Gebru, accepting the 2026 Right Livelihood Award, argued that framing AI risk around hypothetical future catastrophe distracts from present-day concentration of power among the small number of companies — several of them represented among the paper's own signatories — now seeking exemption from independent oversight. Taken together, the week's record isn't a story of consensus breaking down so much as three groups measuring different things: internal automation metrics that are rising fast, a feedback-loop threshold that by one outside estimate isn't close to being met, and a corporate-accountability argument that treats the whole technical debate as somewhat beside the point.