Certain kinds of synthetic data may improve AI models' generalization to real data

A researcher argues that synthetic training data can sometimes benefit AI models. He cites work on training with original data plus an accumulating pool of synthetic data, research suggesting certain kinds of synthetic data can improve generalization to real data, and his own work on cases where model-written training data helps at least at later stages. He also points to STaR, which he describes as showing how models could improve by training on their own reasoning for verifiable tasks and filtering for correct answers.