Large language models are increasingly used as substitutes for human participants. I distinguish two routes to such silicon samples: foundation models fine-tuned on behavioral data, and off-the-shelf models steered by persona prompts. On the first route, I discuss Centaur, a language model fine-tuned on human choices from many behavioral experiments, which predicts decisions better than established cognitive models and generalizes to tasks beyond its training set. On the second route, I show that persona prompting works for an unexpected reason: demographic personas predict published experimental effects no better than personas describing pets and breakfasts. Prompt ensembles benefit from disagreement between prompts, not from information about people, a pattern well explained by classical test theory and the wisdom of crowds. Together, the two routes show that silicon samples are only as good as the behavioral information available to the model, whether learned in training or supplied in context. I close with an outlook on the use of silicon samples in behavioral research.