I think we'll see a lot more contextual engineering efforts soon. It is really inefficient to be uploading your entire codebase pretty much every request, which is what a lot of people are doing. When in reality, very few parts need the full context when programming. Although, big token doesn't seem to care, and often encourages this (including the editors).

I could see a front-end / back-end split in the future where a completely on-client LLM is used to trim down the request and context before shoving the request off to the back-end.