I would expect that the "classics" have all been thoroughly discussed on the Internet in all major languages by now. But if you could re-train a model from scratch and control its input, there are probably many theories you could test about the model's ability to connect bits of insight together.