If OpenAI develops GPT-4, with 1T parameters, I wouldn't be surprised to see a performance gain larger than the jump between GPT-2 and GPT-3.
What GPT-3 shows us is that we're going to have ML systems that can write at the level of an average human pretty soon now.
If you give me a huge corpus of chinese texts and a very long time, I might be able to figure out what character goes with what other, find the various structures in the text and then be able to generate a somewhat convincing made up chinese text while still not understanding a word of it.
These GPT-3 demos are impressive because they look like real text with proper syntax and grammar, but they still express absolutely nothing. It reads like a long series of rambling that goes nowhere. There's no intent behind it.
It reminds me of these videos of apes imitating humans by using their tools, banging hammers ineffectively. They are able to copy the appearance of the behavior, but not the reasoning behind it. They don't get why we bang hammers or what it achieves.
My point is, GPT-3 is operating at human levels for certain contexts. I think it would get passing grades on essays in a lot of schools in the US, for instance, just based on syntax and grammar.
Given what I've seen so far with GPT-3, that simple idea would have to have already been discussed at length on forums on the internet and in the corpus.
Usually books have facts and studies that they use as supporting points. Many of the connections they make between the subject material and their thesis are unique, and this forms their supporting argument. GPT-3 is rearranging words and sentences to resemble structures it's seen before, but it does not create novel facts.
The interesting part is that GPT-3's leap in performance can be attributed to scaling. That's easier to do than inventing completely new approaches. Scale data, scale compute, scale money, then you have something you couldn't have invented directly.
Can I ask- who says this? Is it your personal opinion? Is it the conclusion of the article above, as you understand it? Is it a commonly held opinion of some of the experts in language modelling?
I am asking because everytime there is a claim like "X is important because Y" and someone points out that "Y" is not that interesting, if someone else then says "X is important because Z" and Z is not Y, it's very difficult to have a productive conversation, because it's very difficult to know what we are talking about. Of course, this is the internets and not scientific debate (typically carried out in peer reviewed publications) but if the goalposts keep moving all the time, it's pointless to even try to have a conversation about the merits and flaws of such a complex system. That, with all due respect.
Now, regarding whether GPT-3 is slowing down, it isn't, but it's not going very fast either. Like I say, the curve in the middle of the article that shows acccuracy as a function of parameters is quite flat. Depending on how you want to define diminishing returns, the image painted by the accuracy plot is not that far from it and in any case average accuracy is pretty disappointing.
>> What GPT-3 shows us is that we're going to have ML systems that can write at the level of an average human pretty soon now.
Like I say, there's no good metrics for this kind of task. We have no way to determine what is writing "at the level of an average humen" (let alone what is an "average human"), except eyballing output and expressing a subjective opinion. Anyone might claim that GPT-3 is already capable of writing "at the level of an average human". Anyone might claim that GPT-2 is. Or a Hidden Markove Model, or an n-gram model. Such claims really don't mean anything at all.
It is important to note that this is exactly the task that OpenAI has publicised the most, with GPT-3: a poorly defined task with no good metrics. This insistence in promoting an ability that cannot be objectively evaluated as being a strong point of the model is strong evidence that the model is not nearly as good as advertised.
DL is getting increasingly specialized hardware, there's plenty of growth there. Plus what we're seeing here is GPT scaling up without algorithmic changes. Algorithms are advancing too.
I might try to generate soft-science essays with GPT-3 at one of my universities to see if it passes through TA filters.
>> GPT-3 can also finally do arithmetic, something GPT-2 was unable to do well.
Is a preposterous claim that is very poorly supported by the data in the GPT-3 paper [1]. Figure 3.10 in the paper summarises the results. The authors tested addition between two to five digits, subtraction between two to five digits and multiplication between two digits drawn uniformly at random from [0,100]. There was also a composite task of addition, subtraction and multiplication with single-digit numbers (e.g. 6+(4*8), etc).
On all tasks, other than two and three digit addition and subtraction, accuracy was uniformly under 20%. The other four tasks achieved high accuracy with more parameters.
Of course, this doesn't show that the larger models "learned arithmetic". Two- and three-digit addition and subtraction are likely to be much better represented in a natural language dataset than other operations (and note of course the conspicuous absence of division). So it's safe to assume that the model has seen all the operations it's asked to repeat and knows their results by heart. Remember that for two and three digit addition and subtraction one only needs a dataset with the numbers up to 999, which is really tiny and easy to memorise.
Edit: the authors note that they "spot checked" whether the model is simply memorising results, by searching for three-digit addition examples in their dataset. Out of 2000 three-digit addition problems they failed to find more than 17% in their dataset which "suggests" that the model had not ever seen the problems before. Or, it "suggests" the search was not capable of finding many more existing matches. In any case, why only "spot-check" three-digit addition? Who knows. The paper doesn't say. Certainly, one- and two-digit addition and subtraction should be much more common in a natural language dataset. The authors also say that the model often makes mistakes such as not carrying a one, so it must be actually performing arithmetic! Or, it's simply reproducing common arithmetic mistakes in its dataset. Overall, this sort of "testing" of arithmetic prowess simply doesn't cut the mustard.
Edit 2: Also, no information about how many arithmetic problems of each type were tried. One? Ten? One hundred? Where all arithmetic tasks tested with the same number of problems? Unknown.
_____________
I found these series of tweets almost unbelievable in how well GPT-3 seems to reason about function composition f(f(x)).
"I wonder if the AI would be better at math if you told it to show it's work":
As I say in previous comments, no. It sounds much more like a system that can reproduce resutls it has seen in trainig, but has no general concept of arithmetic.
This would also explain the square root example example easily. Also, the examples in the OP's linked tweet are very simple examples of square roots and function composition that are very likely to have been lifted verbatim from some textbook, or who knows what ...and that's the problem, because who knows what the model has flat memorised and what it's composing from smaller components.
>> It clearly can perform arithmetic, and we don't need to trust OpenAI now that users can interact with the model.
The paper I link above performed a systematic evaluation of GPT-3's arithmetic ability. Playing around with the OpenAI API and eyballing a few results is not going to give a clearer understaning of its abilities.
In general, hitting a language model with a few queries is never going to give any clear understanding of its capabilities. Systematic evaluation is always necessary and the average user (or the non-average user) is not going to be able to do that.
In which GPT-3 answers the question "what is one hundred and five divided by three?" with "35.7". It also gave several other close-but-not-correct answers. It seems pretty unlikely these are all present in the training set, and surely can't all have been lifted verbatim.
I agree systematic testing is probably more useful, but find it really hard to believe this is all happening without any sort of model of arithmetic.
If there is enough arithmetical structure in the training corpus, eventually the best way to predict the training corpus is just to learn arithmetic rather than memorize every instance of arithmetical structure. Transformers have been shown to be equivalent to graph neural networks, so in some sense they have the power to self-discover novel architectures in service to learning a data set. So it is quite reasonable that it could have learned generic rules of arithmetic.
In any case, I think you're applying an overly permissive criterion for learning "generic rules of arithmetic". It's clear from the paper linked above that GPT-3 is extremely limited in its ability to return correct results given arithmetic operations as input. The only task that it performs with 100% accuracy is two-digit addition. It cannot even perform two-digit subtraction with perfect accuracy and it's all downhill from there.
Furthermore, like I say division is conspicuously absent from the set of tested tasks reported in the paper, as are any operations with five digits or more [EDIT: sorry, that's "operations with more than five digits"- my bad.]. Going again by my heuristic from an earlier comment, that researchers publish positive results and avoid publishing negative results, this tells us that GPT-3 can't perform any division at all with any accuracy and it can't perform any arithmetic operations with five digits with any accuracy [EDIT: again, that's "more than five digits". Apologies].
That is hardly the hallmark of a system that has "learned generic rules of arithmetic". It is far more likely that GPT-3 has learned to reproduce results that it has seen during training. Even more so since, as I say in my comment above, it is much better at operations that are likely to be found more often in a natural language corpus.
The cliff for GPT-3's arithmetic ability is likely due to the fact that it can't do recursive/recurrent calculations. That is, it can't reprocess and refine a tentative answer to improve it. You can't do arbitrary arithmetic with a finite amount of substrate without this sort of recursion or recurrency. The fact that it can only do two digits with 100% accuracy could be a hardware or architecture limitation.
Because otherwise, how do you know that your system has learned the "rules of arithmetic", as per your comment, and not something completely different? And like I say in my other comments, there's a very obvious alternative about what that something completely different could be: a representation of already seen results.
Besides, GPT-3 is a piece of software, it's not a child or a grown up human, who can make mistakes because their memory fails or because they get overwhelmed by the complexity of executing a complex set of rules. If a piece of software implements a set of rules, it's usually able to execute them right every time, without failure, certainly so for relatively simple rules like arithmetic. Pocket calculators with tiny resources can do that and they can do it with very long sequences of numbers, so why would a huge language model, running on very expensive hardware, fail?
>> The cliff for GPT-3's arithmetic ability is likely due to the fact that it can't do recursive/recurrent calculations.
Well, yes, exactly that. If a system can't represent recursion then it can't represent arithmetic between arbitrary numbers. Hell, without recursion, a system can't even count to arbitrary numbers. So in what sense can GPT-3 be said to have "learned the rules of arithmetic"? Learned them, how, if it can't represent them?
Actually, your observation about recursion is the first thing I'd have normally said, but it doesn't seem to be commonly understood that neural networks (and propositional, attribute-value, learners in general) can not represent recursion. Similarly, such systems can't represent non-ground values, that is they can't represent the concept of a variable. But that's a big part of why they can't build general theories. In terms of arithmetic, it means they can't represent the relation x + y = z because they can't represent x, y and z as universally quantified variables. The only remaining alternative is to represent every ground expression, like 1 + 1 = 2, 1 + 2 = 3, etc. But that's not the rules of arithmetic! That's only some instances of specific operations. That is why GPT-3 hasn't learned arithmetic and can't learn arithmetic, no matter how much data it is fed. It's just not possible to represent the rules of arithmetic in a propositional language. A first-oder language and the ability to define relations recursively are necessary.
Edit: OK, sorry, my claim about a first order language being necessary is maybe hard to substantiate outside of Peano arithmetic. But, recursion and the ability to represent variables are absolutely necessary. See primitive recursive functions: https://en.wikipedia.org/wiki/Primitive_recursive_function.
Presumably because it answers correctly for examples it hasn't explicitly seen in training. While its plausible that it has seen all two-digit sums in during the course of training, its not a given.
>Besides, GPT-3 is a piece of software, it's not a child or a grown up human, who can make mistakes because their memory fails or because they get overwhelmed by the complexity of executing a complex set of rules.
GPT-3 can become "overwhelmed" by the complexity of the problem extending beyond its feed-forward computation window.
>If a piece of software implements a set of rules, it's usually able to execute them right every time, without failure, certainly so for relatively simple rules like arithmetic.
But a computer system that "computes" through manipulations of language representations is fundamentally different than computer systems that came before. Carrying over the intuition from computers as bit-manipulators to manipulators of language representations is a mistake.
> so why would a huge language model, running on very expensive hardware, fail?
Impedance mismatch? It turns out performing tasks on a computational substrate not suited to those tasks comes with severe drawbacks. But we already knew that.
>So in what sense can GPT-3 be said to have "learned the rules of arithmetic"? Learned them, how, if it can't represent them?
It could know how to sum individual digits through memorization and learn the carry rule. It may be incapable of recursion and thus incapable of summing arbitrarily long digits. But learning the carry rule is most of the way there.
>Similarly, such systems can't represent non-ground values, that is they can't represent the concept of a variable.
I see no reason to accept this. Multi-layer networks seem to be well-suited for abstract representations and manipulations on non-ground values. Ground-values are the input into the network, but higher layers represent on the abstract properties of the ground-values within its receptive field, rather than the particulars of the ground-values. For example, the location and direction of an edge rather than the particular in the form of an edge.
Yes, I'm aware it's very difficult to get people to believe this outside of AI research. Of course, it is entirely uncontroversial and very well understood by researchers. For example, I was in a presentation by a gentleman who works at DeepMind last year and who works on neuro-symbolic integration and he was asked a question along the lines of "how can you model first order logic without variables?" and he pointed out that he had a footnote on one of his slides where he was noting this limitation and that work was underway to address it.
Regarding arithmetic, none of the points made in your comment made in the GPT-3 paper. In fact, the paper makes no attempt to explain what makes GPT-3 capable of performing arithmetic, other than to say that the mistakes in carrying a one suggest that it's actually trying to perform computation and failing. So I have to ask, where do these points come from?
What I mean is, you seem to have a theory about how GPT-3 works. Where does it come from? I apologise if this comes across as personal or unfair, but many commenters in this thread and similar conversations express strong opinions and give detailed explanations about how GPT-3 and similar models work. I am always left wondering where all this information comes from, given that usually it can't be found in the sources I'd expect to find it, namely the work that is being discussed (namely, the GPT-3 paper, in this case).
Sure, neural networks don't operate on proper variables and so in the context of neuro-symbolic processing I'm sure this is a significant hurdle. But in general, abstract representations is part-and-parcel of what makes deep learning powerful. And such an abstract representation is all that's needed for a neural arithmetic unit.
Here[1] is a study on GPT-2 that demonstrates its middle layers develop a representation of syntax and part-of-speech, the sorts of abstract representations that would be needed to develop a mechanism to do abstract arithmetic.
>What I mean is, you seem to have a theory about how GPT-3 works. Where does it come from?
Studies like the one mentioned, and reasonable extrapolation from knowledge of DL and other transformer architectures. We are not totally ignorant on how GPT-3 works.
[1] https://aletheap.github.io/posts/2020/07/looking-for-grammar...
I do not quite understand the relation between "abstract representations that would be needed to develop a mechanism to do abstract arithmetic" and variables. I'm also not sure what you mean by "abstract arithmetic", or what mechanisms you mean. Can you please explain?
Also, I had thought we shared an understanding that the ability to represent primitive recursive functions (which presupose the ability to represent variables and recursion) is necessary to represent arithmetic. Your above comment now makes me doubt this, also. Can you clarify?
Finally, the link above is a blog post. I wouldn't call it a study. But, can you say where in that post I can find the theory about GPT-3's function that you express above?
This means we could potentially dig out thousands of uses out of it by carefully crafting triggers. It's a more general kind of tool than what we're accustomed to work with. It opens a new direction that might become a large field in five years, if they manage to make it run on a regular computer.
<rant>I imagine a GPT-3 like model coupled with search (having Google in an internal loop), trained with multimedia - images, videos, audio, papers, code so it has grounded concepts, and being able to generate text, images, video and code as output. Then I imagine having curated thousands of tasks from the community and added in the training set so it becomes much more efficient, and having all these capabilities exposed as a general AI library. It will be able to work with any modality and you will be able to describe your task in natural language. All of these are possible today. GPT-3 has shown the power of learning to predict on 500B tokens.</>
Can you elaborate on what kind of training data was used here? I'm curious.
It's like the proverbial witch's pot where they put everything in and out comes the magic.
But it also shows it's not very good on any of those tasks. In any case, machine translation, despite its great popularity as a natural language processing task, is another AI task for which we do not have good metrics.
>> Just by a simple example of translation, it understands the task and does translation. By another example, it can do math, and by another one it can do reasoning, or write react apps.
I think in general, there's a tendency to overestimate the capabilities of GPT-3 for various reasons. Speaking of "understanding" and "reasoning" is really not justified.
On the one hand, it's a language model and most of the tasks it's applied to are tasks for which we don't have very good metrics or benchmarks. Like I say in my earliest comment above, natural language understanding metrics are very bad at measuring "understanding" and we don't even have a commonly agreed definition of what that means. Basically, many benchmarks are defined as classification tasks, e.g. with multiple choice questions supposedly testing a model's understanding, but without any way to ensure that a system is not overfitting to statistical regularities in the dataset - and, indeed, language models have often been shown to do exactly that (e.g. see [1]).
On the other hand, OpenAI very aggressively promotes its systems (not just GPT-3) to users outside of AI research and those users have no way to perform a systematic evaluation of such claims, so they are left with good old eyballing [2] of stuff like language generation or translation, etc. It's all too easy for such users to be impressed by a few hand-picked examples provided by OpenAI itself, or by other users who also don't have the capacity for systematic evaluation (and who hand-pick their results out of undue excitement, rather than for any other reason).
The result is that there is a public perception that OpenAI's language models are much better than they really are. If memory serves, OpenAI made a very big todo about how GPT-2 was so good it was dangerous, etc. Well, now we have GPT-3 which is reportedly even more better- but it's served as an API. Doesn't sound that dangerous and it all sounds a lot more as hype than actual progress.
____________
[1] Right for the Wrong Reasons: Diagnosing Syntactic Heuristics in Natural Language Inference
https://arxiv.org/abs/1902.01007
[2] I keep saying that word, but it's actually semi-formal terminology. There was a paper about evaluating the results of grammar induction algorithms that used it. I'll see if I can find it.
I believe such a situation would generate much less debate in software engineering: "my program passes all my unit tests, but it still crashes". Well, yes. Your program passes all your unit tests, because your unit tests are missing the point, not because your code works.
It is very good in all those tasks. In the paper all the numbers comparing GPT-3 with other models on those benchmark are comparing GPT-3 in a zero/one/few shot setting (so no gradient step), against the previous state of the art finetuned for hours or days on the specific task with millions of gradients steps. If you had time and money to finetue GPT-3 on the specific task there is every reason to believe the gap would be huge.
This is the big thing about GPT-3, the promise of not needing to finetune anymore. This is huge in term of productivity but also because it allows you to use the model in settings for which there are basically no dataset available to finetune on.
On the contrary, there is every reason to assume that GPT-3 cannot significantly improve its results on those tasks with extensive fine tuning. Because, if GPT-3 could significantly improve its performance on those tasks with extensive fine-tuning, the OpenAI paper would be reporting those results (edit: OpenAI sure has the time and money to finetune their model).
As we all know by now, there is strong bias in never reporting negative results in machine learning as in other research fields, so we can be reasonably certain that if there is an obvious experiment to perform and that experiment is missing from a paper, it was attemtped and the results were poor.
> "OpenAI sure has the time and money to finetune their model"
This model is absolutly gigantic. In term of training it's a nightmare. I am pretty sure they don't have the capacity to train 10 of them in parallel. So finetuning on all the downstream task needs to be done basically in serial and takes forever. They might have a lot of money, but time is as important for them than for anyone else, their lab isn't at the edge of black hole.
If what you are saying is true I am pretty sure they would have reported it because it would be an extremely interesting and important result in my opinion. If finetuning gave you no advantage over a few shot settings, it would basically mean that the model already know everything there is to know about the task just from it's pretraining and any additional training is useless as the model is not learning anything.
Finally, given the pre-training curves with various model sizes, we clearly haven't reached saturation there, there is no indication anywhere that we have reached saturation on downstream task.
So, for me the far more likely explanation is that fine-tuning on downstream tasks is indeed very costly (time and/or money) even by their standard and isn't even on topic for this paper.
Outside of academia I see this thing very often. "Oh, I'm sure if it was easy to do that, they'd have done it". No. This is not how a piece of research work is evaluated, not even a piece of work in deep learning, a field that has abandoned all pretensions to science in recent years.
My personal advice (speaking as someone who has been attacked by the hyenas and paid my pound of flesh) is that one should always demand the highest standard of proof for any claim in a research paper. That, if one really wishes to know what's going on. Intellectual curiosity and scientific wonderment should not result in gullibility.
Do you not see the cognitive dissonance? You were precisely claiming that because it should be easy to do they most likely have done it, and if they didn't report results it's because it failed.
You are making this unsupported claim with 0 evidence to back you up.
In any case, it's an obvious thing to do and there's no obvious explanation of why they didn't, given that they could.
Further down there's a plot titled "Aggregate Performance Across Benchmarks" where we can see that performance is on average about 50%. I don't know what the baseline for this plot should be (what is the expected average if the tasks are solved by a random classifier?) but comparing this plot with the plot at the top of the article, it doesn't really look like there's a huge improvement in accuracy with a huge increase in the number of parameters. In fact, it's quite the contrary: there's a small increase and a very smooth, almost linear curve. So that's an exponential increase in the use of resources for an almost linear increase in performance? That's not that impressive.
So it appears that the big thing about GPT-3 is that it's big.
It should also be noted that the public interest about GPT-3 is mostly focused on its ability to generate text, for which there is no good metric. So basically, GPT-3 is big, but it's not that good in tasks for which there are formal benchmarks (such as they are, because Natural Language Understanding benchmarks are often very poorly made and don't really measure what they say they measure) and we can't really tell how good it is in the one task that interests most people the most.