Fair enough. Is it now safe to say that OpenAI could have done with a 8B model + $500 of fine tuning instead of running a (much) larger model on their GPU cluster?