You won't feel much difference in performance for the next 4 years but my guess is local LLM inference going to be much better post 3-4 generation.