You're neglecting to consider the power of recursion. Maybe the best and necessary tool in aligning GPT-N is GPT-(N-1). We've already in just the past few wee

You're neglecting to consider the power of recursion. Maybe the best and necessary tool in aligning GPT-N is GPT-(N-1). We've already in just the past few weeks seen the power in using models to generate instructive fine tuning data. Don't you think aligned models might be able to be applied to better aligning future models in ways that we can't yet anticipate because capability discovery is occurring on a week by week basis and not a six month or longer basis?