I've started working on a version of GPT-2 which generates English text. The purpose of this is to improve its ability to predict the next character in a text, by having it learn 'grammatical rules' for English. It already works well for predicting the next character when it has seen only a small amount of text, but becomes less accurate as the amount of training text increases.
I have managed to improve this by having it generate text. That is, it creates an 'original' piece of text about 'topic x', then a slightly altered version of this text where one sentence has a single word changed, and this process is repeated many times (about a million). It seems to quickly learn how to vary sentences in a way that seems natural and realistic. I think the reason this works is because it reduces the chance that the grammar it has learned for one specific topic (e.g. snow) will accidentally be transferred to another topic (e.g. dogs).
Of course, this all means nothing unless it actually learns something from the process of generating text. I haven't tried this yet, but the plan is to have it generate text about a topic, then have a second GPT-2 system try to guess what that topic is. If the resulting system is noticeably better at this task, then we know the process has increased its ability to generalize.
One potential issue with this approach is that the text it generates is 'nonsensical', in that it is almost like a word-salad. Although this is a standard problem with neural nets (and other machine learning algorithms), in this case the text actually is a word-salad. It seems that it has learned the rules of grammar, but not the meaning of words. It is able to string words together in a way that sounds right, but the words don't actually mean anything.
Plot twist: This comment was generated by GPT-3 prompted with some of the comments in this thread.