Hypothesis: ChatGPT trains on humans' neural networks, not just data

me.asu.red

1 pointjayhoon4 comments

This is known. The HF in RLHF upon which InstructGPT is fine-tuned is human feedback. With so many people offering their take on this new technology, perhaps it's important to at least read a little before writing about it.

This is "known" as "proven" or as in "suggested by others previously"? Obviously, answers are tuned by humans (as the OP agrees), but does the GPT model train against the human neural network?

The main question holds, still: will the human IQ limit the capabilities of GPT and other models?

> The main question holds, still: will the human IQ limit the capabilities of GPT and other models?

Unless the models of a way of assessing output quality independent of, superior to, and which trumps input provided by humans, yes.

This (inadvertently) can happen when the model size goes over what humans can oversee. Unless human input gets much higher weight.

Reply on news.ycombinator.com