Is that a good example, though? The reason I was investigating pun explanations there was precisely because that was something I expected GPT-3 to not be able to do because of how its data is preprocessed to strip away phonetics. (If you've read that far in my page, you must've also read all my material ranting about BPEs.)

The puns & rhyming are striking in that they are exceptions that prove the rule about GPT-3's general linguistic capabilities, and are clearcut examples of how BPEs sabotage GPT-3: you can see how its performance falls off a cliff as soon as you hit a BPE-related task (like rhyming or alliteration or anagrams or dad jokes). It's one of the first things I look for these days whenever someone says GPT-3 doesn't do something it seems like it ought to be able to do: "is this phonetics-related, or could in any way be damaged by BPE encoding?" (And despite my warnings, people still get tripped up by it sometimes, or don't believe me - if I had a buck for everyone in the OA Slack or forums who'd tried to make GPT-3 rhyme to prove me wrong...)

But I don't believe there is anything intrinsic to Transformers or self-supervised language modeling which requires this. BPEs are just a cheap performance hack that everyone does without thinking too much about. I expect that switching to character-level (or at least phonetics-aware) encodings and spending some more compute/data would fix all of the issues I diagnose.