“A viable strategy for training models in specialized domains is to use a small set of high-quality data from a few experts to generate a thousand times more synthetic "lookalike" data.”