Plus benchmarks are typically ran from the API, using a tagged version of 3.5-turbo and not the whatever extra tuning and prompt magic ChatGPT as a frontend has

Plus benchmarks are typically ran from the API, using a tagged version of 3.5-turbo and not the whatever extra tuning and prompt magic ChatGPT as a frontend has to actually get the better results that people see in practice. On the other hand, if the base instruct model is good enough to roughly match it then the fine tunes will be interesting for sure.