September 29, 2026 · Research & Policy
A group of senior AI researchers and lab leaders published a warning that automating AI research and development could produce an “intelligence explosion”—compressing years of progress into months if software agents do more of the work required to build their successors. Coverage of the Cambridge CASP-linked report and related paper named contributors associated with OpenAI, Anthropic, Microsoft, and Meta, with additional attention on comments from pioneers including Geoffrey Hinton and Yoshua Bengio.
The argument is not that models are already self-improving without limit. It is that the share of research labor done by models is rising fast enough that governments should measure internal R&D automation and prepare options to steer, constrain, or adapt if the loop tightens. Anthropic has separately disclosed that Claude already performs a substantial share of the company’s own AI research, a data point that makes the paper’s scenario feel less hypothetical.
The Wall Street Journal and other outlets treated the document as a policy push: information-sharing among labs, early oversight, and metrics for how much of the research pipeline is automated. Critics will note that the same institutions racing to automate their own labs are asking governments to watch that race.
Still, a joint warning from people who run or advise the leading labs is different from outsider commentary. It lands in the same week as Astra’s cancellation, Nvidia’s hardware kill switch, and Florida’s courtroom language about extinction—three other signs that the industry’s own risk models are leaking into public institutions.
Key takeaway. The new policy ask is to measure how much of AI research is already done by AI. If that share keeps climbing, the report says progress could jump from years to months.
Photo: Unsplash (abstract science). Sources: The Neuron / WSJ coverage of the CASP report, Northeastern Global News summaries, Sept. 28–29, 2026.
