You can get _better_ that SOTA AI performance by burning the weights directly to silicon, the problem is you're locked into that model