Many of the previous pieces of text on the internet that are in “Human: …, AI: …” format were produced by less advanced AI’s. If GPT-3 had noticed the pattern that text produced by an AI generally makes less sense than human produced text, then it might be deliberately not making sense, in order to imitate less advanced AI’s.
Gwern said that if you give it a prompt with spelling mistakes in, it outputs text containing spelling mistakes, so this kind of deliberately producing low quality text is possible. Then again, I suspect the training dataset includes far more spelling mistake filled text than AI generated text.
This has been suggested a few times before, and I’ve wondered it as well, but I don’t really notice any difference in quality between prompts that explicitly mention or invoke an AI (like the chatbot dialogue one or the Transformer Poetry ones) and the ones which don’t. I suspect there is actually very little real AI/human dialogue text online (simply because most of it is way too bad to bother quoting or storing outside of large ML datasets), and it may well be outweighed by all the fictional dialogues (where of course the AI talks just as well as the human because anything else would be boring).
Many of the previous pieces of text on the internet that are in “Human: …, AI: …” format were produced by less advanced AI’s. If GPT-3 had noticed the pattern that text produced by an AI generally makes less sense than human produced text, then it might be deliberately not making sense, in order to imitate less advanced AI’s.
Gwern said that if you give it a prompt with spelling mistakes in, it outputs text containing spelling mistakes, so this kind of deliberately producing low quality text is possible. Then again, I suspect the training dataset includes far more spelling mistake filled text than AI generated text.
This has been suggested a few times before, and I’ve wondered it as well, but I don’t really notice any difference in quality between prompts that explicitly mention or invoke an AI (like the chatbot dialogue one or the Transformer Poetry ones) and the ones which don’t. I suspect there is actually very little real AI/human dialogue text online (simply because most of it is way too bad to bother quoting or storing outside of large ML datasets), and it may well be outweighed by all the fictional dialogues (where of course the AI talks just as well as the human because anything else would be boring).