Hinton, Bengio and lab researchers ask governments to prepare for AI that speeds up its own research

The only measured figure in the coverage comes from a company. Anthropic says AI generates 80% of its code ^[6]. A share of code tells you who wrote the code. The loop in the paper depends on something harder to count: how much of the research itself an improved system can take over, and how quickly ^[10]. Superpower Daily, in its report on the paper, gives the same caution: the 80% figure alone does not show the loop has reached the report's threshold ^[6].

The authors give two reasons for treating that loop as a likely route to rapid acceleration. AI already contributes to its own development, and improved systems can be deployed quickly once built ^[15]. The second reason governs timing. If a better research system goes into use soon after it is built, little time separates one round of improvement from the next. The authors define an intelligence explosion as a dramatic acceleration in AI progress driven by AI itself, and call it a possible outcome that has not begun ^[9].

Much of what they ask for is measurement, and I think it is the soundest part of the package: a change in the rate of progress can only be recognised against a baseline recorded beforehand. Developers would report progress on AI research and development, including through independent auditors embedded at the companies ^[4]. The stated purpose is to let governments follow how much research AI systems can do, so that a jump in capability does not arrive unannounced ^[4].

The proposals for slowing development are harder to specify. The report suggests limiting how much an AI system can improve over a given period, and working with data centers to pause certain AI research projects ^[14]. A limit on improvement needs an agreed unit of improvement, and the account of the report does not say what that unit would be. Auditors would need the same unit to tell whether a limit had been crossed.

On consequences, the authors hedge in both directions. The same acceleration could bring medical breakthroughs, and a discovery might still wait on special materials, supply chains or regulatory approval before it changes anything outside a lab ^[12]. AI might also help develop countermeasures faster ^[12]. They call the overall effects uncertain but argue that a short response window justifies preparing now ^[12]. "Once an intelligence explosion begins, the window for action may close," they wrote ^[13].

The 2028 date is narrower than the phrase "intelligence explosion" suggests. The forecast covers full automation of some research projects ^[3]. An embedded auditor could check directly whether a project had been fully automated. Meanwhile the authors say the productivity gains needed to set off the larger acceleration have not yet appeared ^[2].