Consider LLM code generation. It would an ideal domain for simulated data.

For example, a version of popular software is released with new functions. A framework could simulate numerous plausible code snippets that exercise the new release, including edge cases. The code snippets would be amply commented, and proven to work in a test harness.

This method is the opposite of depending upon the internet for training data. And this method is being used to train AI robots, that have no available internet training data.