cost goes down by 50% (it didn't) but you fake improvement while lacking any real improvement by taking the average of 4 runs. Cost just doubled, even though in

cost goes down by 50% (it didn't) but you fake improvement while lacking any real improvement by taking the average of 4 runs. Cost just doubled, even though inference technically halved. Never mind the fact that the current model is alwayj outdated, and a new (bigger one) one is always being trained in parallel with the supposedly "cheaper" inference.