In my experience, coding companions like copilot and code whisperer are really useful for getting to the actual problem solving and not fucking around with the best way to write a Kafka shard iterator. The programmer will still need to know how to actually make the business logic work in the code, they just won't need to have library or language specific knowledge to get there. If my only skill is being able to read and implement a coding manual, then yeah I'd be sweating, but if that were the case I probably already wasn't adding that much value to the business.
I am very eager to see what data and ML scientists are able to do with feature extraction/vectorization of non "language" information. The entire premise of LLM's is to learn the pattern of a prompt's vector to a response's vector and train a neural net to replicate that sort of response pattern but in natural language. If we could convert geological or stellar phenomena into a similar vectorized state, we would have all the ingredients to bake a "natural" understanding of higher order phenomena that humans are still studying but lack the man hours to explore it's depths. It would be a hell of a tool to test a hypothesis and rule things out as not possible or extremely unlikely without burning a few years of some poor grad students life.
We just need some sort of generational leap in computing along the lines of moving from vacuum tubes to transistors to have access to the computing power to train it. If silicon photonics ever gets worked out and commoditized like computers have, it'll change the world as we know it. We could build state machines that have a built in modeling of quantum phenomena that could compute the best probabilistic ML model weights and biases in O(1) time, rather than our current brute force and stochastic method of training an ML model