What exactly did they do? Just transform data or generate completely new data for training? I've seen plenty of people that have used GPT-4 to help transform data, i.e. using GPT-4 to generate summaries of texts and then using that to train smaller models. Not sure why this wouldn't be allowed, as it's not technically data that is coming from GPT-4.

If they don't allow this it just seems like they are trying to prevent people from building smaller cheaper models that will perform better for a specific use case and gobbling up the market for as long as they can.