We didn't even have LLMs 3 years ago... so lots of things can still happen. In fact, most researchers agree that the current way of building/using LLMs is not in its final form. In fact, many folks I know don't even think we will be using "LLMs" anymore. There will be a new way of building this. Just last month, MIT published an article capable of removing the context size limit completely by allowing the model itself to ask for the context instead of having to upload all instructions together with a huge context file. We don't even have truly distributed LLMs yet. There is so much to come.