Someone on HN has educated me that gpt4 and 3 should be on a similar param count. This is based on inference times of gpt4 vs gpt3.5 pre-speedup (where distille

Someone on HN has educated me that gpt4 and 3 should be on a similar param count. This is based on inference times of gpt4 vs gpt3.5 pre-speedup (where distilled version was used only post-speedup in the turbo version).