gpt-3 and scaling trends

nostalgebraist:

EDIT: this is a followup to this post I wrote about GPT-3.  If you were linked to this and need more context, read that post first, and/or read the paper I’m talking about.

EDIT 2: since this particular post is getting shared a lot, I want to spell out some things that might not be clear out of context:

• I talk about two different papers in this post.  Both are from OpenAI.  They are
  
  “Scaling Laws for Neural Language Models” (Jan 2020, https://arxiv.org/pdf/2001.08361.pdf)
  
  “Language Models are Few-Shot Learners” AKA GPT-3 (June 2020, https://arxiv.org/pdf/2005.14165.pdf)
  
  
• I also talk about two kinds of tasks where scale may improve performance: language modeling and few-shot learning.  The part about Appendix H is about few-shot learning.  The part about the “breakdown” in scaling is about language modeling.
  
  The two tasks are related to one another, since the same model is used for both and one task (language modeling) is its training objective.  I would guess that in future work, few-shot learning will improve if and only if language modeling improves, but this is not an inevitability.
  
  
• When I talk about the “breakdown” in scaling, I am talking about section 6.3 in “Scaling Laws for Neural Language Models.”
  
  
• By “scaling” here I mean: “using the same architecture and training objective as GPT / GPT-2 / GPT-3, while increasing the parameter count and/or dataset size.”
  
  That is, I am talking about the concept of “scaling” which is the topic of of “Scaling Laws for Neural Language Models.”  It is also what most of the figures in “Language Models are Few-Shot Learners” show on their horizontal axes.
  
  I am not making a general argument about “whether current approaches will scale,” nor am I claiming anything about the performance of models that augment a GPT-style model with other data modalities, different objectives, etc.  Of course my point has some relevance to these topics, just as the papers do.
  
  
• For those wondering whether the scaling work in “Language Models are Few-Shot Learners” is limited by dataset size or by model shape hyperparameters, please see “Scaling Laws for Neural Language Models” on these topics.
  
  And, additionally, please review how “Scaling Laws for Neural Language Models” is cited in the GPT-3 paper (as “KMH+20″) to justify architecture, training, and dataset decisions.
  

—–

From the LW comments on my GPT-3 post, it looks like a lot of the people there think the GPT-3 paper is valuable because it shows there is room for even larger models to do better.

(That is, the point isn’t the performance at 175B, but the shape of the curve as it passes from 117M to 175B, and the implications for >175B.)

This interpretation seems wrong to me, and I also saw little in the paper to indicate that this is what the authors are trying to say.  So I didn’t discuss further scaling at all in my original post.  Since some people find that topic important, though, I will close the loop and copy over here some things I wrote in an LW comment:

—–

If I thought the paper shows a clear trend, with room to grow, toward much greater performance few-shot learning with even bigger models, I would be more impressed with “few-shot + large LM” as an approach.

I don’t think it shows that. The clearest evidence on this subject, IMO, is the many plots in their Appendix H. On a large fraction of the individual downstream tasks, few-shot learning has either

• a scaling trend with a clearly defined shape that is mostly flat by the 175B point, with a remaining gap vs. fine-tuning that seems unlike to be closed (examples: WiC, MultiRC, ReCoRD, PhysicaQA, OpenBookQA, at least 5 of the 6 reading comprehension tasks, ANLI)
  
  
• a very noisy trend where, due to noise, returns to scale might be large but might just as well be near zero (examples: BoolQ, CB, WSC)

The scaling trend is more encouraging on certain downstream tasks (COPA, ARC, Winogrande, many the MT tasks), on “less downstream” tasks that essentially probe language modeling skill in a different way (cloze/completion), and on synthetic tasks.

On average, there is a trend toward slow but steady growth with scale (Fig 1.3), but this masks the great across-task variance catalogued above. The scaling picture for few-shot is very different from the scaling picture for LM loss itself, which as catalogued in another OpenAI paper is remarkably smooth and predictable, and which (as GPT-3 shows) continues smoothly to 175B.

Does few-shot learning look promising in the scaling limit?

• As a tool for humans: no, I expect fine-tuning will always be preferred.
  
  
• As a demonstration that transformers are very generic reasoners: no, we still see a wide spread of task performance despite smooth gains in LM loss, with some of the most distinctive deficits persisting at all scales (common sense physics, cf section 5), and some very basic capabilities only emerging at very large scale and noisily even there (arithmetic).
  
  
• As an AGI component: no. Because few-shot learning on most tasks shows no clear scaling trend toward human level, any role of transformers in AGI will require more effective ways of querying them (such as fine-tuning controlled by another module), or non-transformer models.

—–

Something I didn’t say in the LW comment, but have discussed elsewhere, is that OpenAI expects their scaling laws for LM loss to break down at a scale somewhere close to GPT-3′s scale.  (Cf. this paper, section 6.3.)

This is because their scaling law for compute-efficient training (which grows the model fast and the data slowly) eventually predicts better performance that is possible according to their scaling law for optimal performance at a given dataset size.

Specifically, their point estimate for the breakdown point (released in Jan 2020, before the GPT-3 paper) is 

• ~1e12 model parameters
• ~1e12 tokens in the dataset

with an order of magnitude uncertainty either way.  GPT-3 is

• 1.75e11 model parameters
• 3e11 tokens in the dataset

So we are less than one order of magnitude away from the point estimate.  

(N.B. I am not confident I am comparing like to like here, as I’m not sure GPT-3 was exactly on the compute-efficient frontier defined in the scaling paper, or what effect the difference has.)

In short, not only is few-shot performance unlikely to scale nearly as well as LM loss, LM loss itself – according to OpenAI – is likely to stop scaling in the current way after ~1 additional order of magnitude.

What will happen at that point is unclear to me, but this would seem to complicate any simple extrapolation of performance far beyond 175B, even for measures of performance which (unlike few-shot!) we would otherwise expect to scale indefinitely.

EDIT: if you’re interested in more quantitative detail, I recently made a Colab notebook that combines material from the two papers so you can see GPT-3 on the same axes as the breakdown point.

Whenever one of my posts gets linked somewhere popular (in this case HN), I end up doing several passes of edits to clarify it.

Here’s this post again, with that treatment applied to it – maybe it will clear things up for some people reading on tumblr as well.

Idly flipping through the released GPT-3 samples…

Some amusing bits (full context quoted under the cut):

From a Wikipedia-esque article on “Harry Potter and the Order of the Phoenix”:

Following a Harry Potter fan’s dream that Harry’s late headmaster Albus Dumbledore is alive, and in a critical condition at the Ministry of Magic, Harry Potter and his friends Ron Weasley and Hermione Granger, decide to rescue him, as the school year comes to a close.

On the night of their attempt to break into the Ministry, Ministry of Magic employee Delores Umbridge slashes Rubeus Hagrid’s hand with a knife, accusing him of stealing her kitten. […]

Albus Dumbledore appears to die in battle, but this is revealed to be a ruse, as he and Severus Snape attack Voldemort and Lucius Malfoy, and attempt to take the prophecy from Ron. Lucius disarms Dumbledore, and an enraged Bellatrix kills him. […]

The two engage in a fierce duel in which Snape calls on his master to save him. Harry is unaffected by the curse due to his ability to cast a shield charm. He manages to shield himself and fight back, and in his distraction, Snape accidentally breaks his neck and dies.

Harry meets with Dumbledore’s portrait, who reveals to Harry that the boy’s mother died to save him, and Harry is filled with his mother’s love. Harry reveals that he feels angry and confused at this revelation.

Samuel Richardson, noted sensualist:

But Firbank’s major work, which he completed in 1920, was the novel Inclinations. First published in the United States in 1924, Inclinations is a sensualist novel, and an example of what Firbank calls “new sensualism,” a novelistic genre that—like its eighteenth-century prototype, Richardson’s Pamela—affirms the place of physical attraction in human relations.

Don’t worry, DepressionBot is here to help:

To help raise awareness of depression and suicide, a group of engineering students from the University of Waterloo have designed a robot that can create artwork with the help of artificial intelligence (AI).

Mental illness is a huge problem that often gets ignored, or underplayed. Although the majority of the time you can tell if a person is suffering, there are some cases where it can be harder to identify, especially if you’re just talking to them on the phone.

The team of eight, led by Chris Cui, spent four months designing their robot, which is called DepressionBot.

DepressionBot is now available to help raise awareness of depression and suicide 1:03

“The robot is meant to help raise awareness for depression and suicide,” said Cui, a third-year industrial design student, in an interview with CBC’s The Morning Edition host Craig Norris.

It’s no secret that mental health issues affect a lot of people, but it’s not often that we see people, especially students, take time out of their day to help those affected.

“It’s a really personal subject to a lot of people, but there’s a lot of stigma around it, so a lot of people don’t want to talk about it,” said Cui.

Sketchbook artwork from DepressionBot. (Submitted by Chris Cui)

People who are happy are more likely to be happy:

People feel alone when they don’t have a purpose. They feel like they are nothing when they don’t have something to do. They feel insecure because they feel like their life doesn’t have a meaning. Without these feelings people will feel like they are a nobody. They will feel like their life is worthless. Without a purpose to live for it will be hard for people to be happy and be themselves. When people have no purpose to life they can be so unhappy. These feelings come up in people when they are very upset. People are less likely to be happy when they are alone. With being alone it makes it harder to have fun. Without fun in your life it makes it harder to have a positive attitude. With a positive attitude people will be able to live without fear. They will be able to live like they are in a good mood. With having a good mood people will not be sad and they will feel good about themselves. People will also feel happy when they have a good life. When people feel good about their life it will help people have a positive attitude. They will also be able to have a better life. It will be easier to have a better life without bad feelings. When people have a bad life it will make it harder for them to be happy. This is because their life is not the way they want it. It is very hard for people to be happy when they don’t have the life they want. Without having the life they want they will not be happy. People who have a positive attitude will not feel so sad when their life isn’t the way they want it. With having a positive attitude people will be able to feel happy. When they are happy it will help them to live a better life. With a better life they will be happier. It will be hard for people to be happy when their life isn’t the way they want it. This is because people feel like their life is going wrong. When people feel like their life is going wrong they are less likely to be happy. People who are happy are more likely to be happy. When people are happy they will not be sad. People will not be sad when they are happy.

A reassuring if confusing verdict:

Court rules out taking blood samples for cell phone surveillance

In the United States, the Supreme Court ruled out taking blood samples for cell phone surveillance, the second verdict in just a few days.

The Supreme Court in Washington (USA) ruled that the police is not allowed to search a person’s body for a cell phone’s location and information.

Everyone’s favorite non-flowering, highly toxic source of dietary vitamin C:

Chrysanthemum is a non-flowering perennial plant belonging to the Asteraceae family. The asteraceae family is composed of around 24,000 species. The genus Chrysanthemum is a composite group and consists of approximately 200 species. In Hindi, the plant is known as Karanji, Kuntimuk, Kwatamal. It is a good source of vitamin C and is used as a food and a medicinal plant. The leaves are considered to be of some value as a pot-herb. The entire plant is highly toxic. This plant is often confused with one belonging to the Rosaceae family, known as Golden Shower (Cassia fistula).

Keep reading

gpt-3 and scaling trends

EDIT: this is a followup to this post I wrote about GPT-3.  If you were linked to this and need more context, read that post first, and/or read the paper I’m talking about.

EDIT 2: since this particular post is getting shared a lot, I want to spell out some things that might not be clear out of context:

• I talk about two different papers in this post.  Both are from OpenAI.  They are
  
  “Scaling Laws for Neural Language Models” (Jan 2020, https://arxiv.org/pdf/2001.08361.pdf)
  
  “Language Models are Few-Shot Learners” AKA GPT-3 (June 2020, https://arxiv.org/pdf/2005.14165.pdf)
  
  
• I also talk about two kinds of tasks where scale may improve performance: language modeling and few-shot learning.  The part about Appendix H is about few-shot learning.  The part about the “breakdown” in scaling is about language modeling.
  
  The two tasks are related to one another, since the same model is used for both and one task (language modeling) is its training objective.  I would guess that in future work, few-shot learning will improve if and only if language modeling improves, but this is not an inevitability.
  
  
• When I talk about the “breakdown” in scaling, I am talking about section 6.3 in “Scaling Laws for Neural Language Models.”
  
  
• By “scaling” here I mean: “using the same architecture and training objective as GPT / GPT-2 / GPT-3, while increasing the parameter count and/or dataset size.”
  
  That is, I am talking about the concept of “scaling” which is the topic of of “Scaling Laws for Neural Language Models.”  It is also what most of the figures in “Language Models are Few-Shot Learners” show on their horizontal axes.
  
  I am not making a general argument about “whether current approaches will scale,” nor am I claiming anything about the performance of models that augment a GPT-style model with other data modalities, different objectives, etc.  Of course my point has some relevance to these topics, just as the papers do.
  
  
• For those wondering whether the scaling work in “Language Models are Few-Shot Learners” is limited by dataset size or by model shape hyperparameters, please see “Scaling Laws for Neural Language Models” on these topics.
  
  And, additionally, please review how “Scaling Laws for Neural Language Models” is cited in the GPT-3 paper (as “KMH+20″) to justify architecture, training, and dataset decisions.
  
  
• In practice, I expect large transformers will continue to make it easier and faster to do better NLP, via fine-tuning, for quite a long time before any scaling limits become relevant.  (Most practical work is still done at ~345M at most; scaling up will help.)  My enthusiasm expressed in section (6b) here has not changed.
  
  
• I am not arguing that scaling limits put meaningful bounds the reasoning or language abilities of transformer LMs.  Perhaps they will, but I don’t think we can know yet.
  
  I find these abilities extremely impressive even at GPT-2′s scale, have argued at length against those who do not find them impressive, and believe few-shot learning is too weak a probe of these abilities to be informative here.  (That is, I expect the underlying abilities may continue to improve with scale even when few-shot stops improving with scale.)

—–

From the LW comments on my GPT-3 post, it looks like a lot of the people there think the GPT-3 paper is valuable because it shows there is room for even larger models to do better.

(That is, the point isn’t the performance at 175B, but the shape of the curve as it passes from 117M to 175B, and the implications for >175B.)

This interpretation seems wrong to me, and I also saw little in the paper to indicate that this is what the authors are trying to say.  So I didn’t discuss further scaling at all in my original post.  Since some people find that topic important, though, I will close the loop and copy over here some things I wrote in an LW comment:

—–

If I thought the paper shows a clear trend, with room to grow, toward much greater performance few-shot learning with even bigger models, I would be more impressed with “few-shot + large LM” as an approach.

I don’t think it shows that. The clearest evidence on this subject, IMO, is the many plots in their Appendix H. On a large fraction of the individual downstream tasks, few-shot learning has either

• a scaling trend with a clearly defined shape that is mostly flat by the 175B point, with a remaining gap vs. fine-tuning that seems unlike to be closed (examples: WiC, MultiRC, ReCoRD, PhysicaQA, OpenBookQA, at least 5 of the 6 reading comprehension tasks, ANLI)
  
  
• a very noisy trend where, due to noise, returns to scale might be large but might just as well be near zero (examples: BoolQ, CB, WSC)

The scaling trend is more encouraging on certain downstream tasks (COPA, ARC, Winogrande, many the MT tasks), on “less downstream” tasks that essentially probe language modeling skill in a different way (cloze/completion), and on synthetic tasks.

On average, there is a trend toward slow but steady growth with scale (Fig 1.3), but this masks the great across-task variance catalogued above. The scaling picture for few-shot is very different from the scaling picture for LM loss itself, which as catalogued in another OpenAI paper is remarkably smooth and predictable, and which (as GPT-3 shows) continues smoothly to 175B.

Does few-shot learning look promising in the scaling limit?

• As a tool for humans: no, I expect fine-tuning will always be preferred.
  
  
• As a demonstration that transformers are very generic reasoners: no, we still see a wide spread of task performance despite smooth gains in LM loss, with some of the most distinctive deficits persisting at all scales (common sense physics, cf section 5), and some very basic capabilities only emerging at very large scale and noisily even there (arithmetic).
  
  
• As an AGI component: no. Because few-shot learning on most tasks shows no clear scaling trend toward human level, any role of transformers in AGI will require more effective ways of querying them (such as fine-tuning controlled by another module), or non-transformer models.

—–

Something I didn’t say in the LW comment, but have discussed elsewhere, is that OpenAI expects their scaling laws for LM loss to break down at a scale somewhere close to GPT-3′s scale.  (Cf. this paper, section 6.3.)

This is because their scaling law for compute-efficient training (which grows the model fast and the data slowly) eventually predicts better performance that is possible according to their scaling law for optimal performance at a given dataset size.

Specifically, their point estimate for the breakdown point (released in Jan 2020, before the GPT-3 paper) is 

• ~1e12 model parameters
• ~1e12 tokens in the dataset

with an order of magnitude uncertainty either way.  GPT-3 is

• 1.75e11 model parameters
• 3e11 tokens in the dataset

So we are less than one order of magnitude away from the point estimate.  

(N.B. I am not confident I am comparing like to like here, as I’m not sure GPT-3 was exactly on the compute-efficient frontier defined in the scaling paper, or what effect the difference has.)

In short, not only is few-shot performance unlikely to scale nearly as well as LM loss, LM loss itself – according to OpenAI – is likely to stop scaling in the current way after ~1 additional order of magnitude.

What will happen at that point is unclear to me, but this would seem to complicate any simple extrapolation of performance far beyond 175B, even for measures of performance which (unlike few-shot!) we would otherwise expect to scale indefinitely.

EDIT: if you’re interested in more quantitative detail, I recently made a Colab notebook that combines material from the two papers so you can see GPT-3 on the same axes as the breakdown point.

[Update 11/6/20: OpenAI has recently released a new scaling paper that provides some additional theoretical insight into the “breakdown.”  See here for my commentary.]

The GPT-3 paper cites another recent (Jan 2020) OpenAI paper, “Scaling Laws for Neural Language Models,” which to me is really a lot more interesting.

They train GPT-2 variants across a large range of model sizes, data sizes, batch sizes, and training durations.  They find remarkably smooth scaling laws for the test loss in terms of these parameters – it’s the kind of fit between empirical data and simple analytic formulas that I usually associate with the physical sciences, although possibly I am just not familiar enough with this type of work in ML.

Insofar as you trust these scaling laws, they tell you how to optimally pick all the other parameters for any given compute budget.  In short, the optimal thing to do is “increase model size greatly, increase batch size very slowly, and increase dataset size even more slowly.”  Or, as a picture:

I haven’t read all of the paper in detail yet, but there are various other neat things, like:

- Model shape for transformers (depth vs. width, etc.) matters very little relative to the other parameters.  Changing size rather than shape was already the trend in research, but this provides one kind of reassurance that it’s not a bad trend.

- Their scaling laws eventually contradict one another, though only in a parameter range not yet reached (Section 6.3).  They speculate on the exact place where this happens – the bounds on it are pretty wide – and they conjecture (I’m not sure I understand why) that this reflects either the inherent limits of the transformer, or the true informational density of language.

This paper is useful context for the “GPT-3″ paper, and provides a bigger intellectual frame in which to place “GPT-3″ as an additional piece:

- This paper only explored the GPT-2 range of model sizes, while the “GPT-3″ paper gives empirical results on models 2 orders of magnitude bigger.

- The scaling laws from this paper apparently continue out another two orders of magnitude.  The continuing trend is not surprising, given how tight it is in this paper, although (if I’m reading correctly) the 175B size is getting close to the predicted breakdown point where the scaling laws contradict, and hence puts on a new lower point on that point.

- The GPT-3 paper focuses less on language modeling loss and more on the scaling of downstream task performance with minimal task exposure.

As discussed elsewhere, I find this less interesting for several reasons:

– The ability to get better downstream results is utterly unsurprising: it would be very surprising if language prediction grew steadily toward perfection without a corresponding trend toward good performance on NLP benchmarks

(I mean, duh??? if you have access to a godlike being with literally optimal powers to predict speech, and you can’t get to solve a Winograd schema, then you must be doing something wrong!)

– Their downstream results are a boring lower bound: if you care enough about a task to spend some time figuring out the right way to set up the prompting methodology for “few-shot learning,” you probably care enough about it to spend a day or two preparing a custom supervised dataset, which will do no worse and possibly far better.  I.e. from a practical POV their data efficiency argument is unconvincing.  Cf. my comment here.

In case this sort of thing interests you: I compiled together my two GPT-3 posts and put them up on LessWrong.

A few follow-up comments on the “GPT-3″ paper (my main post is here and should be read before this one):

⭑ On my first read, I thought there was only one plot showing how performance varies with K (number of few-shot samples), but I missed the one very early in the paper, Fig 1.2 on p. 4.

That plot is more impressive than the other one, but doesn’t change my impression that the authors are not very interested in showing off “progressive learning” over the course of a text.

The argument they’re trying to make with Fig 1.2 is that more progressive learning happens with bigger models, and hence that their overall strategy – “use big models + few-shot learning to get good scores on benchmarks” – benefits from an interaction effect above and beyond the independent effects of its two parts (big models, few-shot learning).

Again, this is interesting if you care about scores on NLP benchmarks, but I have trouble seeing much qualitative significance for overall language understanding.

⭑ One of their experiments, “Learning and Using Novel Words,“ strikes me as more remarkable than most of the others and the paper’s lack of focus on it confuses me.  (This is section 3.9.5 and table 3.16.)  The task is closely related to the Wug test – it’s the kind of thing Gary Marcus focused on in his critique of GPT-2 – and looks like this:

[Human prompt] To do a “farduddle” means to jump up and down really fast. An example of a sentence that uses the word farduddle is: 

[GPT-3 continuation] One day when I was playing tag with my little sister, she got really excited and she started doing these crazy farduddles.

This is the sort of task that developmental linguists study in human children, and which past NLP models have had trouble with.  You’d think a success on it would deserve top billing.  The authors apparently report a success here, but treat it as an unimportant sideshow: they say they tried it 6 times and got 6 successes (100% accuracy?!), but they apparently didn’t consider this important enough to try the same thing on a larger sample, compute a real metric, show variance w/r/t parameters, etc.  Meanwhile, they did those things on something like 40 other tasks, mostly far less interesting (to me).  Confusing!

⭑ In addition to the usual NLP benchmarks, they tried some “synthetic or qualitative” tasks (section 3.9).  Their stated goal with these is to clarify the role the actual learning in “few-shot learning,” separating it from mere familiarity with similar-looking text:

One way to probe GPT-3’s range of abilities in the few-shot (or zero- and one-shot) setting is to give it tasks which require it to perform simple on-the-fly computational reasoning, recognize a novel pattern that is unlikely to have occurred in training, or adapt quickly to an unusual task.

The “synthetic or qualitative” tasks are:

• various forms of simple arithmetic (like “add two 2-digit numbers”)
• various anagram/reversal/etc tasks operating on the individual letters of words
• SAT analogies

This line of work feels insufficiently theorized, and thus hard to interpret.

Consider the arithmetic tasks.  Let’s grant the authors’ premise that the model has not just memorized some lookup table for arithmetic problems – it’s really “doing the problems” on the fly.  Then, there are 2 things the model could be doing here (probably some of each simultaneously):

1. It might have developed a real internal model of arithmetic from seeing many related numbers in training texts, and is applying this model to do the problems like you or I would
2. It might have developed some generic reasoning capability for arbitrary abstract tasks, which can handle arithmetic as a particular case of a much more generic class of problems (e.g. it could also pick up various “fake arithmetics” where +, -, etc have non-standing meanings, if appropriately prompted)

Insofar as #1 is happening, the multiple prompts of few-shot learning shouldn’t matter: if the model knows how real (not fake) arithmetic works because it’s seen it in text, then additional examples don’t help “locate the task.”  That is, if it has only learned to do real arithmetic, it shouldn’t need to be told “in this task the + symbol has the standard meaning,” because its ability depends on that assumption anyway.

So, if we’re mostly seeing #1 here, this is not a good demo of few-shot learning the way the authors think it is.

Insofar as #2 is happening, the few-shot prompts do matter: they “locate the meanings” of the symbols in the large space of possible formal systems.  But #2 is wild: it would represent a kind of non-linguistic general intelligence ability which would be remarkable to find in a language model.

I really doubt this is what the authors are thinking.  If they think language models are fully general reasoners, why not highlight that?  The abstract reasoning capacity of transformers has already been more clearly probed without the confounding aspects of natural language, and a priori there are few reasons to think a very large language-specific model should develop strong abilities here (while there are a priori reasons to think the abilities are subtle forms of text recognition/memorization the authors’ methodology was not able to detect).

My best guess is that the authors imagine a factorization of the task into “knowing how to do it” and “knowing we are doing it right now.”  Training on text teaches you how to do (real) arithmetic, and the few-shot prompts tell you “right now we are doing (real) arithmetic, not some other thing you know how to do.”

But arithmetic is a really bad choice if you want to probe this!  The authors use K=50 here, meaning they give the model 50 correct examples of simple math problems to let it “locate the task.”  But no one who can do this task should need 50 examples of it.

What information is conveyed by example #50 that wasn’t already known by example #49?  What are we ruling out here?  Trollish formal systems that look like addition 98% of the time?  “Addition, except ‘52′ actually means ‘37′ but everything else is the same?”  Do we have to rule this out when you should have (and the model must have) a strong prior towards real addition?

I don’t know what the authors are trying to do here, and I think they may not know, either.

argumate:

@nostalgebraist, give us the goss on how GPT-3 compares with GPT-2!

I haven’t read the paper super carefully yet, but I am pretty sure of the following:

(a)

“GPT-3″ is just a bigger GPT-2.  In other words, it’s a straightforward generalization of the “just make the transformers bigger” approach that has been popular across multiple research groups since GPT-2.

This excerpt captures this pretty clearly:

Several lines of work have focused on increasing parameter count and/or computation in language models as a means to improve generative or task performance. […] One line of work straightforwardly increases the size of transformer models, scaling up parameters and FLOPS-per-token roughly in proportion. Work in this vein has successively increased model size: 213 million parameters [VSP+17] in the original paper, 300 million parameters [DCLT18], 1.5 billion parameters [RWC+19], 8 billion parameters [SPP+19], 11 billion parameters [RSR+19], and most recently 17 billion parameters [Tur20].

The first two papers mentioned here are the original transformer for machine translation (VSP+17) and BERT (DCLT18).  The parameter count doesn’t actually increase that much between those two.

The third one (RWC+19) is GPT-2.  The parameter counts jumps up 5x there.  Arguably the point of the GPT-2 paper was “it sounds dumb and too easy, but amazing things happen if you just make a transformer bigger” – and this “GPT-3″ paper is making the same point with bigger numbers.

In one way this is a fair thing to call “GPT-3″: it’s another step in the new biggening tradition which GPT-2 initiated.

But in another way it’s pretty annoying and misleading to call it “GPT-3.”  GPT-2 was (arguably) a fundamental advance, because it demonstrated the power of way bigger transformers when people didn’t know about that power.  Now everyone knows, so it’s the furthest thing from a fundamental advance.  (As an illustration, consider that their new big model deserves the title “GPT-3″ just as much, and just as little, as any of the last 3 big models they mention in that paragraph.)

(b)

The paper seems very targeted at the NLP community, which I mean in almost a wholly negative way.  (Despite being part of the NLP community, I guess.)

The GPT-2 paper argued that language models (text predictors) could do well, or in some cases “at least not terribly,” at the specialized tasks used as NLP benchmarks – even without being told anything about those tasks.  This was sort of neat, but mostly as a demonstration of the language model’s power.

The “zero-shot” learning they demonstrated in the paper – stuff like “adding tl;dr after a text and treating GPT-2′s continuation thereafter as a ‘summary’” – were weird and goofy and not the way anyone would want to do these things in practice.  It was more cool as a demonstration that sufficiently good language models could “do it all,” even things they weren’t intended for; the point wasn’t that they were world-class great at these tasks, the point was the gap between their performance and their low level of preparation.  Kinda like a child prodigy.

In the GPT-3 paper, they’ve introduced a new (…ish? maybe?) way for language models to be good at the standard benchmarks.  Now it’s about how they can “figure out” what they’re supposed to be doing across the course of a text, i.e. instead of prompting the model with one thing like

Q: What is the capital of France?

A: 

they instead prompt it with several, like

Q: What is the capital of France?

A: Paris

Q: What is the capital of Spain?

A: Madrid

Q: What is the capital of Lithuania?

A: Vilnius

Q: What is the capital of Brazil?

A: 

The NLP-community-relevant point of “GPT-3″ is that language models can do much better on the standard benchmarks than we thought, via this kind of multi-prompting and also via even more biggening.  Putting those two changes together, you can even even beat the state of the art on a few tasks (of many).

I can imagine something viewing this as very important, if they thought it showed an ability in transformer LMs to “pick things up on the fly” in an extremely data-efficient, human-like way.  That would be relevant to some of Gary Marcus’ concerns.

But the paper seems totally, weirdly uninterested in the “learning on the fly” angle.  Their paper has many, many figures graphing performance against papemeter count – bigger is better yet again – but I can only find one figure graphing performance against their parameter K, the number of distinct task examples in the prompt (K is 1 and 4 in the two capitals examples).

[EDIT: turns out there’s another one I missed on my first read – Fig. 1.2 on page 4 – but I’ll discuss that in another post sometime, and keep this post as is]

And that figure is, uh, not encouraging:

They do better with one task example than zero (the GPT-2 paper used zero), but otherwise it’s a pretty flat line; evidently there is not too much progressive “learning as you go” here.

(Oddly, the caption for this figure explains these are dev set results so not directly comparable to the test set results given as horizontal lines – which doesn’t stop them from plotting them!  Elsewhere, they do report test set results for SuperGLUE, but only for K=32.  Also, I’m not a fan of this plot’s lack of error bars.)

©

Instead, their interest is almost completely in how good they can get on the benchmarks in absolute terms.

This is why I say it’s aimed at the NLP community: these are the metrics that whole community measures itself against, so in a trivial sense the community “has to” find these results interesting.  But by now, this starts to feel like Goodhart’s Law.

The reason GPT-2 was so cool wasn’t that it did so well on these tasks.  It was that it was a really good language model that demonstrated a new overall understanding of language.  Coercing it to do well on standard benchmarks was valuable (to me) only as a flamboyant, semi-comedic way of pointing this out, kind of like showing off one’s artistic talent by painting (but not painting especially well) with just one’s non-dominant hand.

GPT-2 isn’t cool because it’s good at “question answering,” it’s cool because it’s so good at everything that it makes caring about “question answering” per se feel tiny, irrelevant.

The transformer was such an advance that it made the community create a new benchmark, “SuperGLUE,” because the previous gold standard benchmark (GLUE) was now too easy.

GPT-3 is so little of an advance, it doesn’t even do that well at SuperGLUE.  It just does okay with its dominant hand tied behind its back.

“No, my 10-year-old math prodigy hasn’t proven any new theorems, but she can get a perfect score on the math SAT in under 10 minutes.  Isn’t that groundbreaking?”

Sort of?  Not especially?

(d)

The more I think about this paper, the more annoying it is.  Transformers are extremely interesting.  And this is about the least interesting transformer paper one can imagine in 2020.

[EDIT: see also my follow-up post]

openai/gpt-3 →