That's not how technology typically advances. Technology advances in waves. There's an initial breakthrough that leads to rapid progress in certain areas. But after some time there are diminishing returns as the the applications of that breakthrough are fully explored, and progress slows or stops until the next breakthrough. Predicting that the trajectory of progress for the next 20 years will match the progress that has been made in the first 2 years after the LLM breakthrough seems naive at best.

I would also disagree with the idea that LLM tech is "exponentially improving" now. There was the initial release of ChatGPT which was an enormous step forward, but since then its been small iterations on the same fundamental technology. We are already seeing significantly diminishing returns in our ability to improve these models. Most available training data has already has already been fed in, 10xing the compute results in only marginal improvements in performance, etc.