Yeah, I actually was succesful generating small weights for a small NN, see the following https://www.reddit.com/r/OpenAI/comments/zghtvu/hallucinating_optimized_model_weights_with_chatgpt/?utm_source=share&utm_medium=android_app&utm_name=androidcss&utm_term=1&utm_content=share_button.

If you frame it in the right way, for small enough datasets you can generate a small network. Have yet to confirm this for slightly more complicated data though.