most training data is not of that nature, that's "pretraining" input, it's now much more comprised of agent traces in mid/late phases, which are expanded each generation as we push what we do today with the last generation into the model itself, expect more team training in the next round