As people note, you cannot substitute locally for the Azure GPU cloud that GPT-4 runs on. But I believe that will change, and maybe quickly. After years of expl

As people note, you cannot substitute locally for the Azure GPU cloud that GPT-4 runs on. But I believe that will change, and maybe quickly. After years of explosive exponential growth in model size, all of a sudden, small is beautiful. The precipitating factor is that running large models for research is very expensive, but pales in comparison to putting these things into production. Expenses rise exponentially with model size. Everyone is looking for ways to make the models smaller and run at the edge. I will note that PaLM 2 is smaller than PaLM, the first time I can remember something like that happening. The smallest version of PaLM 2 can run at the edge. Small is beautiful.