We get many Qs asking which part of the AI stack (model, GPU, inference framework, harness, etc) has the most room for optimization. Answer is “all of them!!”
It’s still such early days. This coding harness cuts token use by ~2x for ~any model with ~0 change in performance.