Got any links to explanations of why fine tuning open models isn’t a productive solution? Besides renting the GPU time, what other downsides exist on today’s SOTA open models for doing this?