[datadrivenangel]

Local models are ~18-24 months behind the frontier on approximate intelligence, and then like 36-48 months behind the frontier on inference speed for nice hardware.