I would love to know how this adversarial training doesn't end up overfitting. Or, put another way, I'd love to see another piece of software (or even a human b

I would love to know how this adversarial training doesn't end up overfitting. Or, put another way, I'd love to see another piece of software (or even a human being?) exploit Alpha Go's overfitted strategy.