That's an INCREDIBLY good point about synthetic training data. During model training, AI agents could pretty much start their own coding projects, based on AI-generated wish-lists of features, and verify progress on their own. This leads to limitless training data in the area of coding.
Coding might be cooked.