GLOSSARY
Synthetic Data
Artificially generated training data — model outputs or simulations standing in for real records when real data is scarce, private or expensive.
Synthetic data ranges from fully simulated sensor logs to — increasingly common — outputs of one LLM used to train another (“distillation at dataset scale”). It sidesteps privacy constraints, fills long-tail scenarios real collections miss, and costs a fraction of human labeling. Frontier labs openly use synthetic reasoning traces to train their newest models; the practice is now standard, not exotic.
The known failure mode is model collapse: successive generations training on their own outputs gradually lose diversity and drift. The working rule — blend synthetic data with fresh human data, verify with evals, and never close the loop entirely on a model's own emissions.