Models have been trained on synthetic inputs for over a year now. Collapse only happens in very specific circumstances, it's not a limit in practice.

Model collapse as an insurmountable wall is a fiction dreamed up by Luddites, actual frontier labs understand the failure mode and avoid it.