I can think of lots of individually useful and productive cases, yours included, but i mean more the widespread adoption and integration of the current generation of llms may cost more in resources than it puts out economically. Actual intelligence will come when we advance, not by throwing ever greater amounts of compute at what are just slightly better versions of what was available 3 years ago, but by radical and as yet unseen changes to the way the models function.