I personally don't see a future where common users will ever have to know the phrase "fine-tuning" or worry about it. The most I can see is "Do you consent to share your information with Apple/Meta/X/Microsoft/OpenAI's knowledge engine?" and if you agree, everything they have on you will power an extremely powerful all-encompassing knowledge engine. Probably with some daily recommendations to integrate a new domain into it, like, "We noticed you're into Lord of the Rings, so we went ahead and made your knowledge engine familiar with the collected works of Tolkein, all historical academic and modern interpretations and criticisms, transcripts of the movies, and generative AI fan fiction capabilities."
I don't think the major barrier to the idea would be consumer awareness. For the near-term the major barrier will be cost. Just as one example, together.ai offers fine-tuning service at an advertised cost of $0.001 per 1k tokens used [1]. That will get pricey for even small datasets. No doubt this will come down, but I don't see consumers paying $1000 for a customized AI model that they then have to pay inference costs to run. Maybe once we get consumer devices that have sufficiently capable AI accelerators (e.g. Apple Neural Engine) to run sufficiently capable llm models, then customers would be willing to customize and run local.
The second point is, we don't know if fine-tuned models, vector search or more-massive general purpose llm models is the right way to go.
But for business-to-business, I think this might be a viable business. If you had a whole bunch of ready-to-go open-source fine-tune datasets for commercial applications you might find a market of businesses that want to run their own models for a variety of reasons.
