3 ms·

For GPT-3, I think an example makes it clearer. When GPT-3 was asked the question [1]: > Explain the pun in the following joke: “What does Adult Swim call the

For GPT-3, I think an example makes it clearer. When GPT-3 was asked the question [1]: > Explain the pun in the following joke: “What does Adult Swim call their physical retail stores? Brick and Morty.” It answered: > … The pun “Brick and Morty” alludes to the cable television network “Adult Swim”, which broadcasts a cartoon series called “Rick and Morty”, a humorous parody of “Back to the Future” and other science fiction films. “Brick and Morty” refers not to the characters in the cartoon, but to physical stores that sell DVDs and merchandise based on the cartoon. The pun conflates two meanings of “Brick and Mortar”, a brick-and-mortar store and a brick which is part of a building. It seems on track until you get to the last sentence, which demonstrates that it does not understand the pun at all. There's other examples on that site where it falls down, like adding two 5-digit numbers. Since it's able to answer well things that match the human-written text that it ingested, but falls down on simple things otherwise, that suggests that a lot what appears to be verbal understanding is actually from the (extremely large) amount of human-written source text it ingested, and not added by GPT-3 itself. [1] https://www.gwern.net/GPT-3#pun-explanations https://www.gwern.net/GPT-3#pun-explanations

gwern 5y ago
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.
- throwaway34241 5y ago
  You're right, it's not a good example. I either forgot the context or didn't read it thoroughly enough when it was first written. I should get more familiar with the current state of AI, I've never been a long-run skeptic but my projects are unrelated so I probably have a poor idea of how close we are.

gwern 5y ago

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.

throwaway34241 5y ago
You're right, it's not a good example. I either forgot the context or didn't read it thoroughly enough when it was first written. I should get more familiar with the current state of AI, I've never been a long-run skeptic but my projects are unrelated so I probably have a poor idea of how close we are.