Even when we extrapolate arbitrarily far into the future of very large, very capable ML models, as long as they are built on the current paradigm (Internet-scale self-supervision followed by fine-tuning for desired behavior), I think being Extremely Fucking Good at anything will still remain valuable, regardless of field.
At the upper echelon of any craft you're a tastemaker more than a skilled hand, and taste is like entropy — there is only so much you can distill from a limited dataset into a model before it starts losing fidelity.
Human preference data will always be a scarce resource.