4 ms·
But surely the string "I don't understand the question" (and many other variants) is in the range of GPT-3's possible outputs - it just assigns it lower probabi
by Imnimo 4y ago
But surely the string "I don't understand the question" (and many other variants) is in the range of GPT-3's possible outputs - it just assigns it lower probability than it does the nonsensical answers.
- woojoo666 4y agoThe string is there but it hasn't been trained/taught how to use it and what it actually means
- Imnimo 4y agoIn what sense has it been trained how to use any other string that is not also true for the string "I don't know"? Surely that string is perfectly common in the corpus of internet text used to train these models. Obviously it hasn't been given some explicit model of "you have low confidence on this question therefore you should say you don't know", but neither has it been given any explicit model for any of its other capabilities.
- woojoo666 4y agoI don't have the training data so I can't say for sure, but I'm assuming here that the training data is a lot of "valid question" => "valid answer", without many examples of "nonsense question" => "i don't understand"/"that's nonsense" Edit: I want to add that expressing confidence is not the same as answering the prompt. If I ask somebody "give me a drawing depicting Obama's son", and they said "Obama doesn't have a son", I explicitly asked for a drawing and they are giving a speech response. I believe this kind of indirect response has to be taught, and can't be expected to come out naturally from a model that has only been trained on giving out direct responses.