Since GPT-3 transforms sentences into high-dimensional vectors with connections between all the words and of course its internal weights trained on both the English language and all languages in general, I find that it has a basically flawless ability to interpret normal language.

I tested a prompt where the input was English and the output was Swedish, and it did a really good (but not always perfect) job at translating Swedish. I would assume the amount of Swedish in its training set is relatively tiny, so I find its ability to do that kind of translation in addition to all this other stuff really impressive. This also shows in its ability to translate really well between English and bash shell commands, or JSON data structures, or XML.

If this kind of thing interests you, look into the field of zero-shot learning.

https://en.wikipedia.org/wiki/Zero-shot_learning

Zero-Shot Translation with Google’s Multilingual Neural Machine Translation System

https://ai.googleblog.com/2016/11/zero-shot-translation-with...

Zero-Shot Learning—A Comprehensive Evaluation of the Good, the Bad and the Ugly

https://ieeexplore.ieee.org/document/8413121

also available at:

https://arxiv.org/abs/1703.04394

mbsariyildiz / zsl_eval

> Python implementation of zero-shot/generalized zero-shot evaluation framework proposed by Xian et al in Zero-Shot Learning - The Good, The Bad, The Ugly.

https://github.com/mbsariyildiz/zsl_eval

This is a wild guess- When GPT-3 got confused by "no legs", I don't think it's because it would have any trouble correctly interpreting "no legs" in other contexts. In this context it was searching in a semantic space set up by "a head, a tail, is brown", and the general form of a question looking for an animal. GPT-3 didn't find anything in the semantic space matching "no legs" so it assumed that was an error and corrected it. It's actually fairly close to how a human would solve the riddle, except when the human noticed nothing matched the explicit and implicit prompts, they would correct the implicit request for an animal, and then reinterpret the rest of the riddle.

GPT-3 generates output word by word. And it doesn't make distinctions between prompt words and words it has generated. Generated words are selected somewhat randomly (if temperature >0). So "the train of GPT-3 thought" can be easily derailed by unfortunate random selection of a single word.

When it generated "I don't know what you mean by", it had to continue with something relevant. In this case "no legs" was selected.

In other words GPT-3 doesn't have machinery to stop and think "What I'm rambling about?".

> a head, a tail, is brown

I wonder if brown would have been an issue, even without the animal context, as a search for object with these properties wouldn't have turned up the cent which is copper color

it'd be interesting to ask gp3 what color is a cent and then redo the riddle with that

A head, tail, is brown, no legs. For an animal that’s got me thinking of fish rather than typos :)