Version 11 had lots of hand-coded behavior. Version 12 is entirely a neural network, trained on human driving. Partly it comes from people running FSD and intervening sometimes, and partly it comes from just passive observation of a people doing their own driving. When FSD runs, video feeds into the neural net and it outputs vehicle controls, that's it.

Learning to respond to gestures is just more training on video and car control data. It shouldn't be hard to believe given all the other things we're doing these days with large neural networks.