What's this about?
Go champion Ke Jie is reported to have said he can now beat AI reliably — even giving it a nine-stone handicap — and that if he'd known this when he played AlphaGo, he could have won. The method as described: deliberately play badly in the opening, and the AI, having little training data for genuinely terrible positions, plays badly in response and gets picked off. Baoyu's caveat is important: it's reportedly a 2023 video, so it likely doesn't work against current systems.
What exactly is this exploiting?
A model trained largely on strong play has seen very few genuinely stupid positions, because strong players don't produce them. Its evaluations in those positions are unreliable — it's extrapolating far outside its training distribution. This is the same failure mode as adversarial examples: the system is excellent within the distribution it learned and can be brittle outside it. And crucially, a nine-stone handicap means giving the opponent an enormous head start, so the claim is that distribution shift can outweigh a very large skill gap.
What's the community saying?
- The best joke, and it's also the clearest explanation: "Understood. I play half the game, Ke Jie plays the other half, and together we beat the AI. We're both amazing."
- Baoyu's quotation of the proverb — never try to defeat a genuine idiot, because he'll drag your intelligence down to his level and then beat you with experience — which is a surprisingly precise description of the strategy.
- The dating caveat matters most. It's a 2023 video, and as the original poster noted, this is a data gap that can simply be filled — AlphaGo's imagined reply being "no problem, we just lack training data for that, it can be fixed."
- It's a genuine insight wearing a joke's clothing. Out-of-distribution brittleness is a real property of these systems, and a human exploiting it deliberately is a cleaner demonstration than most adversarial-example papers.