Ok hear me out, I think this is not impressive. I would say this is well within the model training (it has very likely seen a ton of code for pong). If it can do other things that are less represented in the training then it would be more interesting to me