Will fine-tuning LLMs be relevant?

Are we just going to use foundational models off the shelves, or will fine-tuning still be necessary? Seems like openai isn't into a rush to allow fine-tuning for gpt-4 to all users. Google has stated that their vision is "bigger models" (that's the whole update for gemini they came up with at the i/o event they held imo), believing that bigger context windows is the way to go. Llama does some interesting stuff with smaller models. What's your take on the whole thing? Will we need to fine-tune AI in a few years from now, or will we just dump tons of documents and that will do the job?