[fnordpiglet]

It does make me wonder what the converged fixed point on this technique is. If I fine tune with GPT4 to make model A, which then performs better than GPT4, then fine tune model B with A, at what point does either artifacting or diminishing returns set in?

[elexhobby]

GPT-4 is powerful over a diverse set of tasks. They use it to build a model which is better for a narrow sub-task. Pretty sure the model is sub-optimal to GPT-4 for everything else.

[fnordpiglet]

Yeah, but there have been papers being published in general LLMs as well finetuned off of GPT4 instead of humans. Even in the narrow space the question remains. If I build a superior model for task X using gpt4 and it’s superior at X, can I think use my new model to train another model in X and continue to see benefits ?

[elexhobby]

For another possibility, see https://arxiv.org/abs/2305.15717. The new models may not actually be better - the evaluation may be broken.