Early results with language, convoy operations, and information technology troubleshooting suggest that, at their best, lightweight digital simulations can train superbly. However, the notion of training delivered on personal computers also gives rise, apparently by spontaneous generation, to a horde of ‘games for training’ fanatics whose products ensure that, at the median, such simulations are dreadful. Even the best of breed suffer because training individuals to perform as part of a larger military unit often forces the rest of the team and other echelons, above and below, to become expensive and difficult-to-schedule training devices. Moreover, current military operations require soldiers to spend most of their time doing exactly what they did not join the military to do: deal with very sticky situations involving human and cultural interactions. A successful second training revolution must deal with these new missions. To do so simulations will need ‘people engines’ to control characters in virtual cultural landscapes in the same way ‘physics engines’ control physical simulation objects. A handful of plausibility arguments suggest that this might be doable, but there are few existence proofs. Finally, and always, whatever we create will fail if training is not embedded from the beginning.