Gotta say in my opinion it didn't age well for a paper published in 2020 (but pre GPT-3). In practice, we have an O that is pretty good at describing how to build weapons against bears despite having absolutely no idea what "weapon" or "bear" refers to.
But also, I don't seem to understand how this relates to the discussion, so can you help me there?
It may not have an idea of a picture of a bear. But it certainly knows all kinds of words that would describe that bear. And it relates those to the bear and also to other things. So I would argue that there is some „idea“. It‘s just not a visual idea and there is not conscious reflection. As far as the first part is concerned, what about a blind person? I would argue they can have a very good idea, too.
In my understanding of the "Octopus Test" argument from the linked PDF, GPT-3 is exactly an octopus test, and has exactly the same information that the octopus would have about bears, so GPT-3 being able to perform this kind of task disproves the argument conclusively.
>Gotta say in my opinion it didn't age well for a paper published in 2020 (but pre GPT-3).
Agreed. The glaring issue with their argument falls out of their own introduction. What is the "linguistic intent" of the sentence "When was Malala Yousafzai born"? Well, it falls right out of a natural language training corpus which will, with high likelihood, be followed up with her actual birthdate! So there is some non-trivial signal within the training corpus of linguistic intent. Capturing this intent is then just a computational problem. But the objective function of predicting the subsequent word has within the space of solutions the linguistic intent of the writers of the training corpus.
