Because training is largely running inference, and then correcting any mistakes. Over and over again.
(I say largely, because inference also does stuff like sample top-k completions etc.)
This is why that rap looks so weird: so high-quality in general, yet so totally lacking a critical element we expect from rap. GPT-3 has a pretty good idea how to write rap, but it just doesn't know what words sound like. It is blind to that. It can't see that difference between rap and poems. To it, rap is just a sort of poetry, and some words are chosen for reasons it can't understand, no matter how many rap lyrics or poems it looks at.
The better the models get at writing, the more this difference throws us for a loop: "how can it be so good at ABC, and yet fail so totally at D?"
At least that's what I am getting from Gwern's write-up. I might have misunderstood Gwern, or Gwern might be wrong, of course.
And eg it would probably perform worse at translating from English to French than a naive system. (Unless you preprocess your corpus extensively to figure out when you get English and when you get a snippet of French.) GPT-3 is surprisingly good at translating English to French.
Another problem is that English has not just one text-to-phonetic transcription: different accents pronounce words differently. For just one example:
> n most non-rhotic accents, if a word ending in written "r" is followed immediately by a word beginning with a vowel, the /r/ is pronounced, as in water ice. That phenomenon is referred to as "linking R". Many non-rhotic speakers also insert an epenthetic /r/ between vowels when the first vowel is one that can occur before syllable-final r (drawring for drawing). The so-called "intrusive R" has been stigmatized, but many speakers of Received Pronunciation (RP) now frequently "intrude" an epenthetic /r/ at word boundaries, especially if one or both vowels is schwa. For example, the idea of it becomes the idea-r-of it, Australia and New Zealand becomes Australia-r-and New Zealand, the formerly well-known India-r-Office and "Laura Norder" (Law and Order). The typical alternative used by RP speakers (and some rhotic speakers as well) is to insert a glottal stop wherever an intrusive R would otherwise have been placed.
https://en.wikipedia.org/wiki/Rhoticity_in_English
Btw, even without any explicit pronunciation data, I would expect the system to get a pretty good idea of how pronunciation works by essentially using the same technique human linguists use to reconstruct old pronunciations:
They observe what kinds of mistakes people make.
Eg when you see people mixing up "there" and "their" and "they're" in the corpus, that tells you that in modern English these three are pronounced almost the same.
From spelling mistakes in words like carburetor you can figure out that unstressed vowels in English are pretty much all pronounced the same: as a schwa. https://en.wikipedia.org/wiki/Schwa
You can also learn a lot from observing rhymes.
Your example of their/they're/there has some data because the whole words can be mistaken for each other, but even if you take a billion pages of prose, you won't get data to deduce that 'there' rhymes with 'pear' but not 'here', that 'great' rhymes with 'straight' but 'cough', 'dough' and 'through' do not rhyme. A model can't learn something that's simply not represented in the training data.
So you either have to bring in external information (i.e. pronunciation models, or dictionary data with pronunciation guides, or audio recordings) or you have to have sufficient rhyming text inside your training data - i.e. train it on corpora of poetry instead of prose.
Also, I'm not sure if the variation of pronunciation is critical - any accent variations that affect every instance of some sound similarly would preserve rhyming. There are certain differences e.g. I recall a discussion about some parts of Shakespeare which rhyme perfectly in pronunciation of that period but do not in modern English, but I think that the variations of modern English accents should be mostly OK from that perspective.
Yes, the latter. Just mix some rhyming text into your corpus. If you take eg Wikipedia, there's already plenty of rhyming stuff in there. Nursery rhymes, song lyrics, etc. Similar for other corpora.
Doesn't matter how many parameters it has; and as a matter of fact, the GPT models have not improved their rhyming noticeably even as they go from a hundred million parameters to over a hundred billion parameters.
(And I have extensive evidence consistent with my interpretation at the link, ranging from several years of failure by many people to prove me wrong by eliciting rhymes even a fraction as good as its non-rhyme poetry to fixing GPT-3 performance on tasks by working around BPEs to different models by different groups but the same data encoding also showing the same utter failure to rhyme.)
Now if you ran an experiment comparing MLM (or any LM) on rhyming tasks with different encodings, then you could certainly make a statement like "BPE is worse at rhyming than other encodings" and it would be scientifically supportable. And that very well might be true. But your extreme conclusion is not supportable.
A rhyme dictionary would still not replicate human rhyming capabilities. Think about neologisms or misspellings. A model can memorize every single entry in the rhyming dictionary (and let's say this somehow cashes out as apparent rhyming proficiency in being able to for every word recall an entry in the rhyme dictionary of valid other words), but it would not be able to write something like "Jabberwocky" inventing a bunch of new words or phrases or names which rhyme. (How would it know to rhyme "wabe" and "outgrabe" when they appear in no dictionaries - because they were just invented?) A model which has "actually" learned to rhyme would be able to take new words (not necessarily invented by it, but possibly invented by humans after it was trained, or invented on the spot for a prompt, or part of a new fictional work like worldbuilding) and rhyme them appropriately. A model which has memorized a rhyming dictionary would not.
It's beyond me that Facebook squandered its data value by half in not including even a privately tracked 'dislike' button.
The value in what people both like and dislike is going to be massive in the next few years, as it can help train the evaluation of content in both general and personalized scales.
I'd wager with where the tech is today we Tinder could generate synthetic matches that would be far preferred over any natural matches, even at equal or lower physical attractiveness scores in generated photos.
Pandora might be relevant again for music soon as their like/dislike data could allow for the generation of new music in line with preferred tastes and not just selection.
Transistor models like GPT-3 are only part of the equation, and as we continue to improve not only generation of new data, but evaluation of data too -- the application of the technology is likely going to move faster than anything we've seen before in the next few years.
Codex generated 35% of the code written by people with access to it, and Microsoft and OpenAI have the data on what was and wasn't selected. That data alone could produce an AI linter that could have significant value in catching sophisticated natural programmer errors - but in combination with Codex too?
(Keep in mind it was only summer 2020 when the tweet showcasing normal GPT-3 writing HTML was blowing minds, and only last summer Copilot was announced.)
We really haven't seen anything yet.
Things are moving so fast that anyone building products relying on 3rd party AI at a foundational level right now should absolutely be concerned with how they are planning around obsolescence within ~3 years for whatever they are building on top of in their product design and implementation.
(Not 'Transistor' - what I get for typing it out on my phone and not proofreading)