AI that helps build the next AI, which helps build the one after that. In three weeks of September 2026, OpenAI and Alibaba both said it out loud. Here is what it means, and why everyone mentions a brake in the same sentence.
A workshop where every new tool makes the next one faster. The first year you swap the hammer for a better one. The second year the hammer itself suggests how to make a better hammer. Somewhere around there you stop counting generations in years and start counting them in months.
Who said it, and how
On September 6, 2026 OpenAI's chief scientist Jakub Pachocki published the essay "An Alien Mind". He writes that internal results give him a strong expectation that the current pace could carry into recursive self-improvement. In the same text he writes that scaling has to be constrained by confidence in safety, and that frameworks like the Preparedness Framework and the Responsible Scaling Policy should become widely mandated safety bars.
On September 23 Sam Altman told the UN Security Council that systems able to improve themselves and their future versions call for extreme care. And in a post from September 24, Alibaba said its Qwen3.8-Max model had gone through 33 cycles of fully automated self-improvement in one month.
Where to be careful
The term is big, and the numbers behind it so far belong to the companies themselves. The 33 cycles are Alibaba's figure, with no independent check. The internal results Pachocki relies on aren't shown in the essay. That doesn't mean it isn't true. It means it can't yet be checked from outside, and that's exactly why Pachocki proposes that the safety bars be enforced by third-party auditors, government agencies or international bodies.