Throwing away determinism may well be a mistake though. I think it would be good if we treated in-determinism in AI in the same way that we do in normal software: as a defect.