Abstract
Learning means extracting information from old situations that aids predictions or decisions in new situations. In traditional parametric statistics, this information takes the form of real numbers. Recent advances in NLP have made it possible to extract parameters in the form of natural-language strings. Estimating and applying these nonparametric models involves constructing and interpreting the strings using pretrained LLMs, so those processes are unusually opaque, but the strings themselves are interpretable and might generalize well. I'll review some examples, and will contrast them with earlier work that also recovered string-valued parameters.