Comment by znagengast

What if you split up the training down to the literal vector math, and treated every macbook like a thread in a gpu, with just one big computer acting as the orchestrator?