The suggestion that sufficiently large datasets make theory unnecessary — that correlation suffices once the sample approaches the population — enjoyed a decade of currency and has not survived contact with practice. Its error is not that correlation is uninformative but that a dataset, however large, is a record of what was measured under the conditions that obtained, and using it to anticipate behaviour under different conditions requires exactly the causal understanding the claim proposed to dispense with. Systems that performed impressively on historical data and failed when the underlying process shifted are now numerous enough to constitute a literature. The interesting residue of the episode is that prediction and understanding, which had been treated as the same achievement, turned out to come apart in a way that is instructive about both.