In theory there’s an inflection point at which models become sophisticated enough that they can self-sustain with generating training data to recursively improve

That sounds surprising. Do you have a source?