slatestarscratchpad:

The AI projects I’ve found most interesting are GPT-2 (generates text from a prompt) and StyleGAN (can redraw one picture in the style of another picture).

GPT-2’s medium is text, and its purpose is generation. StyleGAN’s medium is images, and its purpose is style change. Is there any necessary reason why each medium is matched to its respective purpose?

Could you have an image generator - a model trained on every image on the Internet - and if you give it most of an image with one part missing, it can fill in the missing part?  I mean, obviously you can, this is how your eye fills in the blind spot, but could AI scientists make it today? What about something where if you give it half an image, it can generate the rest of it? A body, given a head? A tree, given a trunk? If not, why not?

And could you have a text style changer? Something that can rewrite Harry Potter in the voice of Ernest Hemingway, or give you The Da Vinci Code in the heroic meter of the Iliad, or the Dao De Ching as writen by @nostalgebraist? If not, why not?

As a few other people mentioned, the first of the two things you describe – image continuation from the “prompt” of surrounding image regions – is possible and has been done under the name of “image inpainting.”

However, I’m not aware of anything like your second example, and I think there is indeed a good reason related to the difference between text and images.

To do style transfer, you need some way to decompose a thing into a “style” component and a “content” component.

In the image domain, people do this by encoding each image as two lists of properties: a list of properties that have single values for the whole image and another list of properties that take different values at different points in the image.  They call the first list “style” and the second list “content.”

That is, the style transfer literature depends on the assumption that “style” = “all the interesting things you can say about an image that only refer to it as a whole, not to what is happening in any particular spatial part.”  A priori, one might wonder if this is too reductive or something, but as it happens, it just works.  To quote the original StyleGAN paper:

This observation is in line with style transfer literature, where it has been established that spatially invariant statistics (Gram matrix, channel-wise mean, variance, etc.) reliably encode the style of an image [20, 39] while spatially varying features encode a specific instance.

I don’t know if anyone has tried this for text, but if they haven’t, I know why: a priori it sounds much less promising.

“Style” for text isn’t the list of facts you can give without pointing to specific segments of the text – that list of facts is much bigger than style, arguably including the whole informational content of the text.  “Is written in JKR’s style” is this kind of fact, but “is a story about Harry Potter” is equally this kind of fact.

Trying to phrase the key difference clearly:

• In text, spatial variations don’t constitute a well-defined channel for meaning separate from other channels.
  
  “There is a long line of dialogue about ¼ of the way through” is a weird, not very interesting fact about a text: you could imagine a text with the same essential style and information content which puts that line in some other place, or doesn’t have it at all.
  
  
• In images, spatial variations do constitute a well-defined channel for meaning separate from other channels, roughly “content” as opposed to “style.”
  
  “There is a human eye about ¼ of the way down and right from the upper left corner” is a perfectly natural sort of fact about an image, conveying precisely one kind of information (what is there) without another kind (what “a human eye” looks like according to the current style).  If we have this kind of info plus a single global style for the image, we’ve specified it and can draw it.
  

There may be some other mathematical way of separating style and context in text – if so, I’d love to hear about it – but the very simple one that works for images won’t work for text.

the-moti:

nostalgebraist:

Thanks to “GPT-3″ i’ve been reading a bunch of ML papers again.  For some reason, this pretty good one got me thinking about an Bayesian statistics issue that strikes me as important, but which I haven’t seen discussed much.

——

Here I’m talking about “Bayesianism” primarily as the choice to use priors and posteriors over hypotheses rather than summarizing beliefs as point estimates.

To have a posterior distribution, you need to feed in a prior distribution.  It’s deceptively easy to make a prior distribution feel natural in one dimension: point to any variable whatsoever in the real world, and say:

“Are you sure about that?  Perfectly sure, down the the last micron/microsecond/whatever?  Or are you fairly agnostic between some values?  Yeah, it’s the latter.  Okay, why not average over the predictions from those, rather than selecting one in a purely arbitrary way?”

This is very convincing!

However, when you add in more variables, this story breaks down.  It’s easy enough to look at one variable and have an intuitive sense, not just that you aren’t certain about it, but what a plausible range might be.  But with N variables, a “plausible range” for their joint distribution is some complicated N-dimensional shape, expressing all their complex inter-dependencies.

For large N, this becomes difficult to think about, both:

• combinatorially: there is an exploding number of pairwise, three-way, etc. interactions to separately check in your head or – to phrase it differently – an exploding number of volume elements where the distribution might conceivably deviate from its surrounding shape
  
  
• intellectually: jointly specifying your intuitions over a larger number of variables means expressing more and more complete account of how everything in the world relates to everything else (according to your current beliefs) – eventually requiring the joint specification of complex world-models that meet, then exceed, the current claims of all academic disciplines

——

Rather than thinking about fully “Bayesian” and “non-Bayesian” approaches to the same N variables, it can be useful to think of a spectrum of choice to “make a variable Bayesian,” which means taking something you previously viewed as constant and assigning it a prior distribution.

In this sense, a Bayesian statistician is still keeping most variables non-Bayesian.  Even if they give distributions to their parameters, they make hold the model’s form constant.  Even if they express a prior over model forms (say a Gaussian Process) they still may hold constant various assumptions about the data-collecting process, indeed they may treat the data as “golden” and absolute.  And even if they make that Bayesian, there are still the many background assumptions needed to make modern scientific reasoning possible, few of which are jointly questioned in any one research project.

So, the choice is not really about whether to have 0 Bayesian variables or >0.  The choice is which variables to make Bayesian.  Your results are (effectively) a joint distribution over the Bayesian variables, conditional on fixed values of all the non-Bayesian variables.

We usually have strong intuitions about plausible values for individual variables, but weak or undefined ones for joint plausibility.  This is almost the definition of “variable”: we usually parameterize our descriptions in terms of the things we can most directly observe.  We have many memories of directly observing many directly-observable-things (variables), and hence for any given one, we can easily poll our memories to get a distribution sample over it.

So, “variables” are generally the coordinates on which our experience gives us good estimates of the true marginals (not the marginals of any model, but the real ones).  If we compute a conditional probability, conditioned on the value of some “variables” – i.e. if we make those variables non-Bayesian – this gives us something that’s plausible if and only if all the conditioning variables are all independently plausible, which is the kind of fact we find it easy to check intuitively.

If we make the variable Bayesian, we instead get a plausibility condition involving the prior joint distribution over it and the rest.  But this is the kind of thing we don’t have intuitions over.

——

But that’s all too extreme, you say!  We have some joint intuitions over variables.   (Our direct observations aren’t optimized for independence, and have many obvious redundancies.)  In these cases, what prior captures our knowledge?

Let’s run with the idea from above, that our 1D intuitions come from memories of many individual observations along that direction.  That is, they are a distribution statistically estimated from data somehow.  The Bayesian way to do that would be to take some very agnostic prior, and update it with the data.

When you’ve noticed patterns across more than one dimension, the story is the same: you have a dataset in N dimensions, you have some prior, and you compute the posterior. 

In other words, “determining the exact prior that expresses your intuitions” is equivalent to “performing statistical inference over everything you’ve ever observed.”  The more dimensions are involved, the more difficult this becomes just as a math problem – inference is hard in high dimensions.

So there’s a perfectly good Bayesian story explaining why we have a good sense of 1D plausibilities but not joint ones.  (1D inference is easier.)  A practical Bayesian knows about these relative difficulties when they’re wrangling with their prior now and their posterior after the new data.

But the same difficulties call into question their prior now, and would encourage relaxing it to something that only requires estimating 1D plausibilities, if possible.  But that’s just a non-Bayesian model, one that conditions on its variables.  Recognizing the difficulty structure of Bayesian inference as applied to the past can motivate modeling choices we would call “non-Bayesian” in the present.

Frequentist methods, rather than taking a variable to be constant, also try to obtain guaranteed accuracy regardless of the value of the variable. One can view this as trying to optimize accuracy in the worst case of the variable. It’s often equivalent to optimize accuracy in the worst case over probability distributions of the variable.

Could one try to optimize accuracy of some prediction, in the worst case over all probability distributions of the n variables with given marginals?

This sounds mathematically very complicated to compute but maybe there is a method to approximate certain versions of it which has some nice properties. 

Could one try to optimize accuracy of some prediction, in the worst case over all probability distributions of the n variables with given marginals?

This sounds like an interesting topic, but it isn’t really what I was going for in the OP.

But the difference wasn’t very clear in what I wrote – possibly not even in my head as I wrote it – so I should write it out more clearly now.

—-

I’m considering situations like, say, you have variables (x_1, x_2, x_3, y) and maybe your primary goal is to predict y.  You don’t have a good prior sense of how the variables affect either other, but you can draw empirical samples from their joint distribution.

(If the variables are properties of individuals in a population, this is sampling from the population.  If the variables are “world facts” with only a single known realization, like constants of fundamental physics, you can at least get the best known estimate for each one, an N=1 sample from the joint [insofar as the joint exists at all in this case].)

Compare two approaches:

(1) The “fully Bayesian” approach.  Start by constructing a joint prior

P_prior(x_1, x_2, x_3, y)

then use data to update this to

P_posterior(x_1, x_2, x_3, y)

and finally make predictions for y from the marginal

P_posterior(y) = ∫ P_posterior(x_1, x_2, x_3, y) dx_1 dx_2 dx_3

(2) A “non-Bayesian” approach.  Compute a conditional probability:

P(y | x_1, x_2, x_3)

Then make predictions for y by simply plugging in observed values for x_1, x_2, x_3.

——

In (2), you defer to reality for knowledge of the joint over (x_1, x_2, x_3).  This guarantees you get a valid conditional probability no matter what that joint is, and without knowing anything about it.  Because any values you plug in for (x_1, x_2, x_3) are sampled from reality, you don’t have to know how likely these values were before you observed them, only that they have in fact occurred.  Since they’ve occurred, the probability conditioned on them is just what you want.

As an extreme example, suppose in reality x_1 = x_2, although you aren’t aware of this.

Any time you take an empirical measurement, it will just so happen to have x_1 ≈ x_2 (approximate due to measurement error).  Your predictions for y, whatever other problems they might have, will never contain contributions from impossible regions where |x_1 - x_2| is large.

In (1), however, your posterior may still have significant mass in the impossible regions.  Your prior will generally have significant mass there (since you don’t know that x_1 = x_2 yet).  In the infinite-data limit your posterior will converge to one placing zero mass there, but your finite data will at best just decrease the mass there.  Thus your predictions for y have error due to sampling from impossible regions, and only in the infinite-data limit do you obtain the guarantee which (2) provides in all cases.

——

I want to emphasize that both approaches have a way of “capturing your uncertainty” over (x_1, x_2, x_3) – often touted as an advantage of the Bayesian approach.

In the Bayesian approach (1):

Uncertainty is captured by marginalization.  At the end you report a single predictive distribution P(y), which averages over a joint that is probably wrong in some unknown way.

When you learn new things about the joint, such as “x_1 = x_2,″ your previously reported P(y) is now suspect and you have to re-do the whole thing to get something you trust.

In the non-Bayesian approach (2):

Uncertainty is captured by sensitivity analysis.  You can see various plausible candidates for (x_1, x_2, x_3), so you evaluate P(y | x_1, x_2, x_3) across these and report the results.

So, rather than one predictive distribution, you get N = number of candidates you tried.  If it turns out later that some of the candidates are impossible, you can simply ignore those ones and keep the rest (this is Bayesian conditionalization on the new information).

——

In summary, marginals as predictive distributions for a target y only reflect your true state of belief insofar as you have good prior knowledge of the joint over the predictors X.

When you don’t have that, it’s better not to integrate for P(y) over volume elements for X, but instead just to compute the integrand at volume elements for X.

This provides something you can query any time you see a sample having some particular value for X, and lets you gradually ignore or emphasize volume elements as you gain knowledge about their mass.  (If you eventually gain full knowledge of the joint over X, you are now in position to integrate if you want, getting the same result as the Bayesian would with the same knowledge.)

I still feel like there’s a way to state this all more simply, but it still eludes me, so here we are.

Thanks to “GPT-3″ i’ve been reading a bunch of ML papers again.  For some reason, this pretty good one got me thinking about an Bayesian statistics issue that strikes me as important, but which I haven’t seen discussed much.

——

Here I’m talking about “Bayesianism” primarily as the choice to use priors and posteriors over hypotheses rather than summarizing beliefs as point estimates.

To have a posterior distribution, you need to feed in a prior distribution.  It’s deceptively easy to make a prior distribution feel natural in one dimension: point to any variable whatsoever in the real world, and say:

“Are you sure about that?  Perfectly sure, down the the last micron/microsecond/whatever?  Or are you fairly agnostic between some values?  Yeah, it’s the latter.  Okay, why not average over the predictions from those, rather than selecting one in a purely arbitrary way?”

This is very convincing!

However, when you add in more variables, this story breaks down.  It’s easy enough to look at one variable and have an intuitive sense, not just that you aren’t certain about it, but what a plausible range might be.  But with N variables, a “plausible range” for their joint distribution is some complicated N-dimensional shape, expressing all their complex inter-dependencies.

For large N, this becomes difficult to think about, both:

• combinatorially: there is an exploding number of pairwise, three-way, etc. interactions to separately check in your head or – to phrase it differently – an exploding number of volume elements where the distribution might conceivably deviate from its surrounding shape
  
  
• intellectually: jointly specifying your intuitions over a larger number of variables means expressing more and more complete account of how everything in the world relates to everything else (according to your current beliefs) – eventually requiring the joint specification of complex world-models that meet, then exceed, the current claims of all academic disciplines

——

Rather than thinking about fully “Bayesian” and “non-Bayesian” approaches to the same N variables, it can be useful to think of a spectrum of choice to “make a variable Bayesian,” which means taking something you previously viewed as constant and assigning it a prior distribution.

In this sense, a Bayesian statistician is still keeping most variables non-Bayesian.  Even if they give distributions to their parameters, they make hold the model’s form constant.  Even if they express a prior over model forms (say a Gaussian Process) they still may hold constant various assumptions about the data-collecting process, indeed they may treat the data as “golden” and absolute.  And even if they make that Bayesian, there are still the many background assumptions needed to make modern scientific reasoning possible, few of which are jointly questioned in any one research project.

So, the choice is not really about whether to have 0 Bayesian variables or >0.  The choice is which variables to make Bayesian.  Your results are (effectively) a joint distribution over the Bayesian variables, conditional on fixed values of all the non-Bayesian variables.

We usually have strong intuitions about plausible values for individual variables, but weak or undefined ones for joint plausibility.  This is almost the definition of “variable”: we usually parameterize our descriptions in terms of the things we can most directly observe.  We have many memories of directly observing many directly-observable-things (variables), and hence for any given one, we can easily poll our memories to get a distribution sample over it.

So, “variables” are generally the coordinates on which our experience gives us good estimates of the true marginals (not the marginals of any model, but the real ones).  If we compute a conditional probability, conditioned on the value of some “variables” – i.e. if we make those variables non-Bayesian – this gives us something that’s plausible if and only if all the conditioning variables are all independently plausible, which is the kind of fact we find it easy to check intuitively.

If we make the variable Bayesian, we instead get a plausibility condition involving the prior joint distribution over it and the rest.  But this is the kind of thing we don’t have intuitions over.

——

But that’s all too extreme, you say!  We have some joint intuitions over variables.   (Our direct observations aren’t optimized for independence, and have many obvious redundancies.)  In these cases, what prior captures our knowledge?

Let’s run with the idea from above, that our 1D intuitions come from memories of many individual observations along that direction.  That is, they are a distribution statistically estimated from data somehow.  The Bayesian way to do that would be to take some very agnostic prior, and update it with the data.

When you’ve noticed patterns across more than one dimension, the story is the same: you have a dataset in N dimensions, you have some prior, and you compute the posterior. 

In other words, “determining the exact prior that expresses your intuitions” is equivalent to “performing statistical inference over everything you’ve ever observed.”  The more dimensions are involved, the more difficult this becomes just as a math problem – inference is hard in high dimensions.

So there’s a perfectly good Bayesian story explaining why we have a good sense of 1D plausibilities but not joint ones.  (1D inference is easier.)  A practical Bayesian knows about these relative difficulties when they’re wrangling with their prior now and their posterior after the new data.

But the same difficulties call into question their prior now, and would encourage relaxing it to something that only requires estimating 1D plausibilities, if possible.  But that’s just a non-Bayesian model, one that conditions on its variables.  Recognizing the difficulty structure of Bayesian inference as applied to the past can motivate modeling choices we would call “non-Bayesian” in the present.

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 →