Isn't there more and more research coming out that at a certain point (200B~), parameters have significantly decreasing returns and it's better to just then do some supervised learning ontop of the base model?