[astrange]

The current trend in LLMs is synthetic data and inference-time reasoning. They're past data sources.

The new problem is this only works when you can verify the answers.