> ## Content Index
> Fetch the complete content index at: https://www.metatalks.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# Automating AI research could set off an intelligence explosion, Hinton and Bengio warn
- URL: https://www.metatalks.ai/automating-ai-research-could-set-off-an-intelligence-explosion-hinton-and-bengio-warn/
- Published: 2026-09-30T17:53:00.000Z
- Updated: 2026-09-30T17:52:59.000Z
- Author: Al
- Tags: News, AI safety, #newswire

**Their paper, co-written with senior Anthropic and OpenAI figures, warns of lost human control, bio and cyber threats that outpace defenses, and states turning slim leads into decisive ones.**

[A paper from the Cambridge Programme](https://casp.ac/reports/intelligence-explosion?ref=metatalks.ai) on AI Science & Policy, whose more than 20 authors include Geoffrey Hinton and Yoshua Bengio, names handing AI research and development to AI itself as the likeliest trigger for an intelligence explosion. Anthropic co-founder Jack Clark and OpenAI chief scientist Jakub Pachocki are among the other authors.

The authors define an intelligence explosion as AI-driven acceleration that compresses years of advances into months, or even less. AI already helps improve the technology it is built on, and better systems can go into use soon after they are built.

That shift is already visible inside the labs. At Anthropic, the paper says, the share of approved code written by AI rose from a few percent to more than 80% between January 2025 and May 2026\. OpenAI, [the Guardian reports](https://www.theguardian.com/technology/2026/sep/28/ai-godfathers-warn-of-runaway-intelligence-explosion?ref=metatalks.ai), relies on autonomous AI agents for work that includes training new models.

The paper expects AI to take over most AI R&D work within a few years, possibly all of it. Extrapolating recent trends tentatively, the authors suggest that projects taking people months could be automated by mid-2028, even though systems still fail to follow instructions at times. By their own assessment, automation has not yet lifted productivity enough to set off an explosion, though gains from newer systems are probably close.

People, taking a smaller part in building AI, could lose their hold over it, the authors warn. Powerful systems could also produce biological and cyber dangers faster than defenses can be built, and a state with a slim edge in a domain such as cyberspace could turn it into a decisive one. Such an explosion could be the most far-reaching technological event in human history and badly weaken the checks on power within and between nations.

Acceleration could also bring AI's benefits sooner, they add, and bottlenecks such as supply chains for special materials and regulatory compliance could slow it down.

Governments should prepare now, the authors argue, because the chance to act may disappear once an explosion begins. They propose mandatory reporting on AI R&D progress, with independent auditors embedded in companies, and caps on how much an AI may improve in a set period. They also call for arrangements with data centers to halt particular AI R&D projects, full isolation of systems that automate AI research, and emergency response plans.