Agreed. Also think this is the most probable path for them to take.
How do you imagine these features getting exposed to end-users? (When LLMs are discussed people default to thinking in terms of chat interfaces but to your point that doesn't have to be a hard and fast rule.) Do you think there will some sort of primitives that'll get exposed as APIs to developers, like with CoreML[1] for example?
[1]: (https://developer.apple.com/machine-learning/core-ml/)