The early OpenAI Instruct models were more like that. The original GPT-3 was only trained to predict the next token, then they used RLHF to make them interpret everything as queries, so that "Explain the theory of gravity to a 6 year old." wouldn't complete to "Explain the theory of relativity to a 6 year old in a few sentences." ChatGPT was probably that, expanded to multi-turn conversation. You can see the beginnings of that ChatGPT style in those examples.

https://openai.com/index/instruction-following/