I think its when software gets 'better' at writing AI-software than humans, and when AI-software is capable of writing an AI-software that is better than itself.

The point being that the quality of AI software has some sort of 'exponential' (maybe not numerically but conceptually due to compounding effects) growth in its quality, and that during this growth it crosses the point where it is better at 'general intelligence' than humans.