I finally got access to fine tune GPT-4o. GPT-4o mini became available a couple weeks ago, but have been aching for OpenAI to let me at 4o proper.
What you see here is a page that includes a reflection agent. The user first prompts one of our models, be it one of our fine-tuned models or a plain vanilla model. The output is a bio written for one music-based artist. Once that first artist bio is written it’s taken and integrated into a prompt that’s then fed to yet another model.
That new prompt includes additional instructions for the 2nd model used to refine the original bio. In that step we outline specific editorial qualities and other requirements, then turn GPT-4o loose to reflect on what’s already written, then to write a refined bio based on the original.
In this 2nd step, we’re not limited to OpenAI models. I can take essentially any model to use for this reflection step. For now, and for our testing purposes we’re using 4o. I’m curious to feed that first bio with the same instructions to Claude to see how its writing differs.
