The official openai-cookbook (https://github.com/openai/openai-cookbook) used to have an explicit, but buried, call out that instruction-following models like `text-davinci-003` were "Less diverse; less creative; sometimes harder to steer tone, style, etc." as opposed to base completion models like `davinci`.
It stood out to me because it seemed to be an internal admission that this training narrowed the potential of the models.
It stood out to me because it seemed to be an internal admission that this training narrowed the potential of the models.
Required a bit of digging but I found the old file in the history, the relevant text is in the comparison table at the bottom: https://github.com/openai/openai-cookbook/blob/c651bfdda64ac...