part of the pro-openai case is that, in general, and including in a research context where people are doing things that are quite difficult and quantifying them fairly well, "the biggest possible model, applied bluntly to this specific task" almost always beats specialized/cleverer/smaller

if this weren't the case things would be a lot less centralized; the fact that "a giant budget for a huge gpu cluster" REALLY IS a definitive advantage most of the time has seriously narrowed how many parties can do cutting-edge research and caused a lot of it to be held privately for profit