The researchers building AlphaGo first trained a neural network to accurately reproduce the move an expert human would make for a given board state. The neural network didn't end up being a great player, and lost half it's games against the runner-up AIs. Why? Because its goal had been to accurately reproduce the movements it had learned, instead of the actual goal of winning. Have you heard the term cargo-culting? That's what it was doing: merely going through the motions it had seen, without understanding why and using that knowledge to pull its bag of great moves together into an actual strategy.

Just like how most students sit in class and do their homework because that's what they're expected to do, accurately reproducing their role, and never once thinking about what it would take for them to actually absorb the material and make it a part of themselves. Or how everyone is really impressed with a small set of famous people, and then just go right back to the usual motions of their daily lives, without once trying to work out the path through action-space that would lead to their own ascendency.

That neural net got to winning 85% of its games against the runner-up AI when they further trained it using Reinforcement Learning to just flipping win any way it could.

So don't merely do what others do. Sample and remix what others do to win.