points
Modern LLMs are trained on a large percentage of synthetic data.
This sentiment is largely legacy (even though just a couple of years old).