And it also suffers from the tired assumption that GPT-3 (or any language models) should, or are designed to in any way, give reasonable answers[1]. All GPT-3 does is give
likely continuations, given the training corpus.
The prompts here are too short, and it could likely just be writing mediocre fiction continuations. Fiction tends to not be reasonable much of the time (to create story conflict).
> "To understand why, it helps to think about what systems like GPT-3 do. They don’t learn about the world—they learn about text and how people use words in relation to other words. What it does is something like a massive act of cutting and pasting, stitching variations on text that it has seen, rather than digging deeply for the concepts that underlie those texts."
This is another pet peeve of mine. It has long been shown experimentally[1] that neural networks such as image recognition and text prediction networks such as GPT-3 do understand deep concepts that underlie texts (not perfectly yet, of course), from emergent abstractions and similar cognitive tools employed by human brains.
[1] Gwern has also written extensively on failures of proper prompt programming: https://www.gwern.net/GPT-3#prompts-as-programming
[2] For example, using feature map and kernel visualization. In object classification or detection CNNs, specialized filters arise for detecting common observed object classes, like faces. Moreover, there is a hierarchical assembly of objects from elementary components (e.g. from lines, to limbs, to humans).
Deep visualization toolbox: https://www.youtube.com/watch?v=AgkfIQ4IGaM
See this comment: https://news.ycombinator.com/item?id=24195009 for an extended discussion.