I've seen this objection raised a lot, but I think it betrays a misunderstanding of what GPT-3 is capable of doing.
The best, in fact the only way to generate truly convincing text output on most subjects is to understand, on some level, what you're writing about. In other words, to create a higher level abstraction than simply "statistically speaking, this word seems to follow that one". Once you start to encode that words map to concepts, you can use the resulting conceptual model to create output which is conceptually consistent, then map it backwards to words. There is what humans do with sensory data, and there is good evidence that GPT-3 is doing this too, to some degree.
Take simple arithmetic, such as adding two and three digit numbers. GPT-2 could not do this very successfully. It did indeed look like it was treating it as a "find the textual pattern" problem.
But GPT-3 is much more successful, including at giving correct answers to arithmetic problems that weren't in its training set.
So what changed? We aren't sure, but the speculation is that in the process of training, GPT-3 found that the best strategy to correctly predicting the continuation of arithmetic expressions was to figure out the rules of basic arithmetic and encode them in some portion of its neural network, then apply them whenever the prompt suggested to do so.
If this is the case, and it remains speculation at this point, would you still argue that GPT-3 doesn't "understand" arithmetic, on some level? I would argue that this abstraction, this mapping of words onto higher-level concepts, which can then be manipulated to solve more complex problems, is exactly what intelligence is, once you strip away biologically-biased assumptions.
Certainly, at this point GPT-3's conceptual understanding remains somewhat primitive and unstable, but the fact that it exhibits it at all, and sometimes in spookily impressive ways, is what has people excited and worried. We have produced AIs that can perhaps think conceptually about relatively narrow topics like playing Go, but we have never before created one that can do so one such a wide range of topics. And there is no suggestion that GPT-3's level of ability represents a maximum. GPT-4 and beyond will be more powerful, meaning that it can mine more and more powerful conceptual understanding from their training data.
> We aren't sure, but the speculation is that in the process of training, GPT-3 found that the best strategy to correctly predicting the continuation of arithmetic expressions was to figure out the rules of basic arithmetic and encode them in some portion of its neural network
I don't mean to attack you personally, but this is a perfect example of what I feel is wrong with so much neural network research. (And I understand that you are just commenting in a discussion, not conducting research.)
In a word, it's baloney. And it's a really common pattern in neural networks' recent history: "How did they perform reasonably well on this task? We aren't sure, but the speculation is that they magically solved artificial general intelligence under the hood." Usually this is followed up by "I don't know how it works, but let's see if a bigger network can make even prettier text." Meanwhile, "it's funny how our image classifiers grossly misperform if you rotate the images a little or add some noise."
A rigorous scientific approach would be aimed at actually figuring out what these models can do, why, and how they work. Rather than just assuming the most optimistic possible explanation for what's happening -- that's antithetical to science.
This is where you lost me. They included important caveats to indicate not being sure, which is important to me as an indication of healthy skepticism. And you substituted a specific example: making inferences about arithmetic, for an more expansive, uncharitable, easy-to-caricature claim of "gee we must have solved general AI!" which is much easier to attack. And, unlike your counterpart, who hedged, you just went ahead and categorically declared it to be baloney, making you the only person to take a definitive side on an unsettled question before the data is in. This is a perfect example of the anti-scientific attitude exhibited in Overconfident Pessimism [0].
I don't think it's known how GPT-3 got so much better at answering math questions it wasn't trained on, I do think the explanation that it made inferences about arithmetic is reasonable, I think the commenter added all the qualifiers you could reasonably ask them to make before suggesting the idea, and frankly I would disagree that there's some sort of obvious history of parallels that GPT-3 can be compared to.
There is an interesting conversation to be had here, and there probably is much more to learn about why GPT-3 probably isn't quite as advanced as it may immediately appear to be to those who want to believe in it. But I think a huge wrench is thrown in that whole conversation with the total lack of humility required to confidently declare it 'baloney', which is the thing that sticks out to me as antithetical to science.
Thanks for your reply. A couple responses to advance the conversation.
As a side note, it's worth mentioning that apparently, from other responses, it seems we have little idea how much arithmetic GPT-3 has learned, and it may not be much.
Anyway, I think the important distinction between my perspective and Overconfident Pessimism, which you attribute to me, is that I'm not talking about (im)possibility of achievement, I'm talking about scientific methodology or lack thereof.
In other words, I'm not saying (here) that some NLP achievements are impossible. I'm saying that we are not rigorously testing, measuring, and verifying what we are even achieving. Instead we throw out superficially impressive examples of results and invite, or provoke, speculation about how much achievement probably must have maybe happened somewhere in order to produce them.
We have seen several years of this pattern, so this is not a GPT-3 specific criticism; it's just that particular quote so neatly captured patterns of lack of scientific rigour that we have seen repeatedly at this point.
Probably the first example was image recognition. Everyone was amazed by how well neural nets could classify images. There was a ton of analogous speculation -- along the lines of 'we're not sure, but the speculation is the networks figured out what it really means to be a panda or a stop sign and encoded it in their weights.' The terms "near-human performance" and then "human-level performance" were thrown around a lot.
Then we found adversarial examples and realized that e.g. if you rotate the turtle image slightly, the model becomes extremely confident that it's a rifle. So, obviously it has no understand of what a turtle or a rifle is. And obviously, we as researchers don't understand what those neural nets were doing under the hood, and that speculation was extremely over-optimistic.
Engineering cool things can absolutely be a part of a scientific process. But we have seen countless repetitions of this pattern (especially since GANs): press releases and impressive-looking examples without rigorous evaluation of what the models are doing or how; invitations to speculate on the best-possible interpretation; and announcing that the next step is to make it bigger. I think this approach is both anti-science and misleading to readers.
> And it's a really common pattern in neural networks' recent history: "How did they perform reasonably well on this task? We aren't sure, but the speculation is that they magically solved artificial general intelligence under the hood." Usually this is followed up by "I don't know how it works, but let's see if a bigger network can make even prettier text."
Layperson here, but my impression is that "let's see if a bigger network can make even prettier text" has _worked_ far beyond the point most people expected it would stop working.
Also my layperson impression: most "researchers" that are on the cutting edge of cool things are more interested in seeing what cool things they can do than on doing rigorous science (which makes sense -- if you optimize for rigorous science, your stuff probably isn't as flashy as the stuff produced by people optimizing for flash).
> what cool things they can do than on doing rigorous science (which makes sense -- if you optimize for rigorous science, your stuff probably isn't as flashy as the stuff produced by people optimizing for flash.
Is this a new iteration on that zigzag quote?
> Zak phases of the bulk bands and the winding number associated with the bulk Hamiltonian, and verified it through four typical ribbon boundaries, i.e. zigzag, bearded zigzag, armchair, and bearded armchair.
From "The existence of topological edge states in honeycomb plasmonic lattices"
I don't know if this means something is truly wrong. AI is a mix of engineering and scientific research, just like most CS subfields. Recently, the emphasis has shifted towards engineering, as the applications of neural nets have skyrocketed after a few breakthroughs in performance.
It's similar to computer systems research. For example, a research paper on filesystems might tell us a simple trick which leads to better performance on NVMM. The paper may go into why the trick works, but it doesn't (and shouldn't need to) generalize and try to improve our general understanding of how to design filesystems on different hardware. We've been designing filesystems to this day and well, we are always still guessing about which approaches to use and hoping for the best. In the same vein, we don't even have a widely-accepted theory of how to use data structures yet.
So, I don't think that neural nets aren't scientific enough means that it's all BS. We have gaps in understanding, but the power of the models warrants a lot of continued work on finding useful applications.
Doesn't mean I don't think AI is over-hyped/overfunded though...
I agree with a lot of this, but I think there is a consistent pattern of AI announcements playing on humans' intuitions to create the impression that much more has been achieved than can actually be proven -- in fact, not even trying to prove anything. Part of this is that the researchers are humans too and may be misled themselves. But a rigorous research process would at least try to prevent that.
For example, people once thought playing chess was hard. So they thought that if a computer could beat the world champion, then computers would probably also be able to replace every job and so on. If you sent Deep Blue back in time to the 1960s, they wouldn't understand how it works so they'd probably assume that it since it could beat Petrosian in chess, it could probably drive cars and treat disease.
But then we built Deep Blue and realized that you don't need AGI to play chess; a very specialized algorithm will do it.
So we're like people in the 70s who've been handed Deep Blue. It's irresponsible, in my opinion, to over-hype it when we have no idea how it works.
> Meanwhile, "it's funny how our image classifiers grossly misperform if you rotate the images a little or add some noise."
Same thing arguably happens with humans with rotation. Our eyes even rotate in the roll axis to keep gravity aligned things upright. Most people can draw faces more accurately when copying from an upside down face than a right side up one.
> So what changed? We aren't sure, but the speculation is that in the process of training, GPT-3 found that the best strategy to correctly predicting the continuation of arithmetic expressions was to figure out the rules of basic arithmetic and encode them in some portion of its neural network, then apply them whenever the prompt suggested to do so.
I saw a lot of basic arithmetic in the thousands range where it failed. If we have to keep scaling it quadratically for it to learn log n scale arithmetic then we're doing it wrong.
I'm surprised you think it learned some basic rules around arithmetic. A lot of simple rules extrapolate very well, into all number ranges. To me it seems like it's just making things up as it goes along. I'll grant you this though, it can make for a convincing illusion at times.
The example of simple arithmetic is interesting. I think you might be right, that is evidence that GPT-3 might be generating what we might consider models of it's input data. Very simple, primitive and fragile models, but yes that's a start. Thank you.
A disembodied AI with a really good model might be able to do good theoretical science, but it would still need a way of acting in the physical world to do experimental science.
I would love to talk to someone who actually believes in the Chinese Room argument. To me it seems to be ignoring the existence of emergent behavior, and the same argument could prove that a human Chinese speaker doesn't understand Chinese either: his neurons are just reacting to produce answers depending on the input and their current state (e.g. neurotransmitters and action potentials).
The Chinese Room argument is fairly transparently circular; if you assume understanding involves something more than applying a sufficiently complex set of deterministic rules, then a pure system of deterministic rules cannot ever achieve understanding.
Of course if you accept the required premise of the argument, you must accept that either, one, we don't live in a universe that is a pure system of deterministic rules, or, two, nothing in the universe can have true understanding.
The Chinese Room argument, scientific materialism, or the existence of true understanding—you can have at most two of those in a consistent view of the universe.
John Searle came up with that argument to conclude that despite a hypothetical Chinese room being able to have a conversation with someone, it doesn't truly have understanding, so N seems to be at least 1.
To your point though, the more interesting case is people who would disavow the Chinese Room argument, but then end up using reflecting its views while argue against the intelligence of this or that system.
Practically everyone in my online bubble feels similarly, it seems, though I do think steelmanning it is a great way to explore the topic. Same with the Mary's Room argument.
Peter Watts explores this in the novel Blindsight. I don't want to give the plot away, but the main idea is really interesting, and relevant to this discussion.
I'm posting to second the recommendation for the novel. It is the most interesting exploration of the Mind's I (not a typo) that I've come across in modern sci-fi.
> We aren't sure, but the speculation is that in the process of training, GPT-3 found that the best strategy to correctly predicting the continuation of arithmetic expressions was to figure out the rules of basic arithmetic and encode them in some portion of its neural network, then apply them whenever the prompt suggested to do so.
I strongly disagree. GPT-3 has 100% accuracy on 2-digit addition, 80% on 3-digit addition, 25% on 4-digit addition and 9% on 5-digit addition. If it could indeed "understand arithmetic" the increase in number of digits should not affect its accuracy.
My perspective as an ML practitioner is that the cool part of GPT-3 is storing information effectively and it is able to decode queries easier than before to get the information that is required. Yet with things like arithmetic, the most efficient way would be to understand the rules of addition but the internal structure is too rigid to encode those rules atm.
I don't think training on language by itself is enough. Consider for example, if we found extraterrestrial transmissions from an alien civilization. We don't know what they look like, what they're made of or if they even have corporeal form. All we have is a large quantity of sequential tokens from their communications.
It's possible to train GPT3 to produce a facsimile of these transmissions, but doing so does not let us learn anything at all about these aliens, beyond statistical correlations like ⊑⏃⟒⍀ often occurring in close proximity to ⋏⟒⍙⌇ (what do they represent - who knows?). Just having the text is not enough, because we have no understanding of the underlying processes that produced the text.
That said, this is only a limitation of language models as they currently exist. I imagine it would be possible to train a ML model that encodes more of the human experience via video/audio/proprioception data.
I wouldn't be so sure we couldn't decode the meaning of an alien language given enough sample text. There have been some advances[1] towards learning a translation between two human languages in an unsupervised manner, meaning without any (language1, language2) sentence pairs to serve as the ground truth for building a translation. Essentially it independently learns abstract representations of the two languages from written text in each language, all the while nudging these abstract representations towards identical feature spaces. The result is a strong translation model trained without utilizing any upfront translations as training data.
The intuition behind this idea is that the structure inherent in a language is dependent upon features of the world being described by that language to some degree. If we can abstract out the details of the language and get at the underlying structure the language is describing, then this latent structure should be language-independent. But then translation turns out to simply be a matter of decoding and encoding a language to this latent structure. One limitation of this idea is that it depends on there being some shared structure that underlies the languages we're attempting to model and translate. It's easy to imagine this constraint holds in the real world as human contexts are very similar regardless of language spoken. The basic units and concepts that feature in our lives are more-or-less universally shared and so this shared structure provides a meaningful pathway to translation. We might even expect the world of intelligent aliens to share enough latent structure from which to build a translation given enough source text. The laws of physics and mathematics are universal after all.
This doesn't really make any sense to me. Shared structure is not enough to assume shared meaning. Even given the idea of Universal Grammar (which seems extremely likely, given the interchangeablility of human languages for babies), that tells us nothing about the actual words and their association with the human world.
Take the sentence 'I fooed a bar with a Baz' - can you infer what I did from this?
>Shared structure is not enough to assume shared meaning.
How do you define meaning? If we can find a mapping between the sequence of words in a language and the underlying structure of the world, then we by definition know what those words mean. The question then reduces to whether there will be multiple such plausible mappings once we have completely captured the regularity of a "very large" sequence of natural language text. I strongly suspect the answer is no, there will only be one or a roughly equivalent class of mappings such that we can be confident in the discovered associations between words and concepts.
The number of relationships (think graph edge) between things-in-the-world is "very large". The set of possible relationships between entities is exponential in the size of the number of entities. But the structure in a natural language isn't arbitrary, it maps to these real world relationships in natural ways. So once we capture all the statistical regularities, there should be some "innocent" mapping between these regularities and things-in-the-world. "Innocent" here meaning a relatively insignificant amount of computation went into finding the mapping (relative to the sample space of the input/output).
>Take the sentence 'I fooed a bar with a Baz' - can you infer what I did from this?
Write me a billion pages of text while using foo bar baz and all other words consistently throughout, and I could probably tell you.
You're relying on having texts covering all aspects of a language.
Here's a good example. Suppose I had a huge set of recipe books from a human culture - just recipes, no other record of a culture.
I might be able to get as far as XYZZY meaning "a food that can be sliced, mashed, fried and diced" but how would I really tell if XYZZY means carrot, potato, or tomato, or tuna?
> If we can find a mapping between the sequence of words in a language and the underlying structure of the world, then we by definition know what those words mean.
This seems a bit tautological to me: if being able to make certain mappings is understanding, then does this not amount to "once we understand something, understanding it is a solved problem?"
On the other hand, the apparently simplistic mappings used by these language models have achieved way more than I would have thought, so I am somewhat primed to accept that understanding turns out to be no more mysterious than qualia.
I doubt that just any mapping will do. One aspect of human understanding that still seems to be difficult for these models is reasoning about causality and motives.
I think it is a fairly common intuition that one cannot understand something just by rote-learning a bunch of facts.
I meant it in the sense of: given any reasonable definition of "understanding", finding a mapping between the sequences of words and the structure of the world must satisfy the definition.
>I think it is a fairly common intuition that one cannot understand something just by rote-learning a bunch of facts.
I agree, but its important to understand why this is. The issue is that learning an assignment between some words and some objects misses the underlying structure that is critical to understanding. For example, one can point out the names of birds but know nothing about them. It is once you can also point out details about their anatomy, how they interact with their environment, find food, etc and then do some basic reasoning using these bird facts that we might say you understand a lot about birds.
The assumption underlying the power of these language models is that a large enough text corpus will contain all these bird facts, perhaps indirectly through being deployed in conversation. If it can learn all these details, deploy them correctly within context, and even do rudimentary reasoning using such facts (there are examples of GPT-3 doing this), then it is reasonable to say that the language model captures understanding to some degree.
Ok, I think I'm getting some of your idea more clearly. Essentially, the observation is that we probably can't consistently replace the words in a novel with other words without preserving the meaning of the novel (consistently meaning each word is always replaced with the same other word).
I think the biggest problem with this argument is the assumption that '[the structure in a natural language] maps to these real world relationships in natural ways'. One thing we know for sure is that human language maps to internal concepts of the human mind, and that it doesn't map directly to the real world at all. This is not necessarily a barrier to translation between human languages, but I think it makes the applicability of this idea to translations between human and alien languages almost certainly null.
Perhaps the most obvious aspect of this is any word related directly to the internal world - emotions, perceptions (colors, tastes, textures etc) - there is no hope of translating these between organisms with different biologies.
However, essentially any human word, at least outside the sciences, falls in this category. At the most basic level, what you perceive as an object is a somewhat arbitrary modeling of the world specific to our biology and our size and time scale. To a being that perceived time much slower than us, many things that we see as static and solid may appear as more liquid and blurry. A significantly smaller or larger creature may see or miss many details of the human world and thus be unable to comprehend some of our concepts.
Another obstacle is that many objects are defined exclusively in terms of their uses in human culture and customs - there is no way to tell the difference between a sword, a scalpel, a knife, a machete etc unless you have an understanding of many particulars of some specific human society. Even the concept of 'cutting object' is dependent on some human-specific perceptions - for example, we perceive a knife as cutting bread, but we don't perceive a spoon as cutting the water when we take a spoonful from a soup, though it is also an object with a thin metal edge separating a mass of one substance into two separate masses (coincidentally, also doing so for consumption).
And finally, even the way we conceive mathematics may be strongly related to our biology (given that virtually all human beings are capable of learning at least arithmetic, and not a single animal is able to learn even counting), possibly also related to the structure of our language. Perhaps an alien mind has come up with a completely different approach to mathematics that we can't even fathom (though there would certainly be an isomorphism between their formulation of maths and ours, neither of our species may be capable of finding it).
And finally, there are simply so many words and concepts that are related to specific organisms in our natural environment, that you simply can't translate without some amount of firsthand experience. I could talk about the texture of silk for a long while, and you may be able to understand roughly what I'm describing, but you certainly won't be able to understand exactly what a silkworm is unless you've perceived one directly in some way that is specific to your species, even though you probably could understand I'm talking about some kind of other life form, it's rough size and some other details.
>human language maps to internal concepts of the human mind, and that it doesn't map directly to the real world at all.
I disagree. Mental concepts have a high degree of correlation with the real world, otherwise we could not explain how we are so capable of navigating and manipulating the world to the degree that we do. So something that correlates with mental concepts necessarily correlates with things-in-the-world. Even things like emotions have real world function. Fear, for example, correlates with states in the world such that some alien species would be expected to have a corresponding concept.
>There is no way to tell the difference between a sword, a scalpel, a knife, a machete
There is some ambiguity here, but not as much as you claim. Machetes, for example, are mostly used in the context of "hacking", either vegetation or people, rather than precision cuts of a knife or a scalpel. These subtle differences in contextual usage would be picked up by a strong language model and a sufficient text corpus.
> Mental concepts have a high degree of correlation with the real world, otherwise we could not explain how we are so capable of navigating and manipulating the world to the degree that we do.
This is obviously a strong association from human mental concepts to real world objects. The question is if the opposite mapping exists as well - there could well be infinitely many non-human concepts that could map onto the physical world. They could have some level of similarity, but nertheless remain significantly different.
For a trivial example, in all likelihood an alien race that has some kind of eye would perceive different colors than we do. With enough text and shared context, we may be able to understand that worble is some kind of shade of red or green, but never understand exactly how they perceive it (just as they may understand that red is some shade or worble or murble, but never exactly). Even worse, they could have some colors like purple, which only exists in the human mind/eye (it is the perception we get when we see both high-wavlength and low- wavelength light at the same time, but with different phases).
Similarly, alien beings may have a significantly different model of the real world, perhaps one not divided into objects, but, say, currents, where they perceive moving things not as an object that changes location, but as a sort of four-dimensional flow from chair-here to chair-there, just like we perceive a river as a single object, not as water particles moving on a particular path. Thus, it may be extremely difficult if not impossible to map between our concepts and the real world back to their concepts.
> Fear, for example, correlates with states in the world such that some alien species would be expected to have a corresponding concept.
Unlikely, given that most organisms on earth have no semblance of fear. Even for more universal mental states, there is no reason to imagine completely different organisms would have similar coping mechanisms as we have evolved.
> Machetes, for example, are mostly used in the context of "hacking", either vegetation or people, rather than precision cuts of a knife or a scalpel.
Well, I would say hacking VS cutting are not significantly different concepts, they are a matter of human-specific and even culturally-specific degrees, which would be unlikely to me to be uniquely identifiable, though some vague level of understanding could probably be reached.
That kind of gets into "what is it like to be a bat" territory.
The more imminent question is more of engineering than philosophy - what does it take for GPT-3 to not make the mistakes it does? This would require it to have some internal model for why humans generate text (persuasion, entertainment, etc.) as well as the social context in which that human generated the text. On a lower level it also needs to know about cognitive shortcuts that humans take for granted (object permanence, gravity)
Basically, some degree of human subjective experience must be encoded and fed to the model. That's a difficult problem, but not an intractable one.
I certainly agree that GPT-3 appears to be learning how to do mathematics. I suspect that if you gave it enough it might perhaps even learn the maths of physics.
I suspect that if it did that, it would be able to write a very convincing fake paper about how it designed and tested an Alcubierre drive, and that the main clue about the paper being fake being a sentence such as “we dismantled Jupiter for use as a radiation shield against the issue raised by McMonigal et al, 2012”.
Or, to put it another way, the hardest of hard SciFi, but still SciFi, not science.
Nothing you say convinces me that GPT-3 is exhibiting any conceptual understanding.
Imitating existing texts better is not conceptual understanding.
"Understanding" means you can explain why you made a decision. It means there exists a model with conceptual entities that you can access and make available to others.
What GPT-3 does is this: "I am given many answers to similar questions, and I build up a huge model that reflects these answers. If I'm given a new question, I come up with a response that's probably right, based on the previous answers, but there's no explanation possible."
Don't get me wrong - it's amazing! But it's not understanding anything yet.
Even humans have skills that we know but do not understand - like "walking" for most of us!
But on abstract question, we almost always have access to a complete set of reasons. "Why did you go back to the store?" "I left my bag there." "Why did you talk to that man?" "I know he's the manager, I'm a regular." "Why were you happy?" "I had my bag."
(Indeed, this is so common that people often "backdate" reasons for actions that didn't really have any reason at the time. But I digress.)
I wonder how well it would perform in accuracy if given a large number of simple but lengthy sums like 13453 + 53521. Increased set size would move it beyond simple input/output memorization. Although if it recurses properly and carries the digit it could be text parsing and have an accurate but probably very inefficently written math parser.
> Although if it recurses properly and carries the digit it could be text parsing and have an accurate but probably very inefficently written math parser.
I suspect this is how many humans do arithmetic (especially considering how many people conflate numbers with their representation as digits). So if GPT-3 is doing that, that's pretty impressive.
You don't have to wonder. In their paper: https://arxiv.org/abs/2005.14165
they state it has 0.7% accuracy on zero shot 5 digit addition problems and 9.3% accuracy on few shot 5 digit addition problems.
You do have to wonder, because as that section states, the BPEs may impede arithmetic, and as we've found using the API, if you use commas, the accuracy (zero and few-shot) goes way up.
Solmonoff induction would imply the algorithm that learns the rules of arithmetic will have the most concise model for the data. But, it is unclear these gpt-3 type algorithms are solomonoff learners.
>> But GPT-3 is much more successful, including at giving correct answers to arithmetic problems that weren't in its training set.
That's not exactly what the GPT-3 paper [1] claims. The paper claims that a search of the training dataset for instances of, very specifically, three-digit addition, returned no matches. That doesn't mean there weren't any instances, it only means the search didn't find any. It also doesn't say anything about the existence of instances of other arithmetic operations in GPT-3's training set (and the absence of "spot checks" for such instances of other operations suggests they were, actually, found- but not reported, in time-honoured fashion of not reporting negative results). So at best we can conclude that GPT-3 gave correct answers to three-digit addition problems that weren't in its training set and then again, only the 2000 or so problems that were specifically searched for.
In general, the paper tested GPT-3's arithmetic abilities with addition and subtraction between one to five digit numbers and multiplication between two-digit numbers. They also tested a composite task of one-digit expressions, e.g. "6+(4*8)" etc. No division was attempted at all (or no results were reported).
Of the attempted tasks, all than addition and subtraction between one to three digit numbers had accuracy below 20%.
In other words, the only tasks that were at all successful were exactly those tasks that were the most likely to be found in a corpus of text, rather than a corpus of arithmetic expressions. The results indicate that GPT-3 cannot "perform arithmetic" despite the paper's claims to the contrary. They are precisely the results one should expect to see if GPT-3 was simply memorising examples of arithmetic in its training corpus.
>> So what changed? We aren't sure, but the speculation is that in the process of training, GPT-3 found that the best strategy to correctly predicting the continuation of arithmetic expressions was to figure out the rules of basic arithmetic and encode them in some portion of its neural network, then apply them whenever the prompt suggested to do so.
There is no reason why a language model should be able to "figure out the rules of basic arithmetic" so this "speculation" is tantamount to invoking magick.
Additionally, language models and neural networks in general are not capable of representing the rules of arithmetic because they are incapable of representing recursion and universally quantified variables, both of which are necessary to express the rules of arithmetic.
In any case, if GPT-3 had "figure(d) out the rules of basic arithmetic", why stop at addition, subtraction and multiplication between one to five digit numbers? Why was it not able to use those learned rules to perform the same operations with more digits? Why was it not capable of performing division (i.e. the opposite of multiplication)? A very simple asnwer is: GPT-3 did not learn the rules of arithmetic.
I dunno to me it seems clear that there is nothing of what we call intelligence in these neural networks. And I think we could have a general AI that can problem solve in the world but have zero of what we know of as understanding and sel awareness
The best, in fact the only way to generate truly convincing text output on most subjects is to understand, on some level, what you're writing about. In other words, to create a higher level abstraction than simply "statistically speaking, this word seems to follow that one". Once you start to encode that words map to concepts, you can use the resulting conceptual model to create output which is conceptually consistent, then map it backwards to words. There is what humans do with sensory data, and there is good evidence that GPT-3 is doing this too, to some degree.
Take simple arithmetic, such as adding two and three digit numbers. GPT-2 could not do this very successfully. It did indeed look like it was treating it as a "find the textual pattern" problem.
But GPT-3 is much more successful, including at giving correct answers to arithmetic problems that weren't in its training set.
So what changed? We aren't sure, but the speculation is that in the process of training, GPT-3 found that the best strategy to correctly predicting the continuation of arithmetic expressions was to figure out the rules of basic arithmetic and encode them in some portion of its neural network, then apply them whenever the prompt suggested to do so.
If this is the case, and it remains speculation at this point, would you still argue that GPT-3 doesn't "understand" arithmetic, on some level? I would argue that this abstraction, this mapping of words onto higher-level concepts, which can then be manipulated to solve more complex problems, is exactly what intelligence is, once you strip away biologically-biased assumptions.
Certainly, at this point GPT-3's conceptual understanding remains somewhat primitive and unstable, but the fact that it exhibits it at all, and sometimes in spookily impressive ways, is what has people excited and worried. We have produced AIs that can perhaps think conceptually about relatively narrow topics like playing Go, but we have never before created one that can do so one such a wide range of topics. And there is no suggestion that GPT-3's level of ability represents a maximum. GPT-4 and beyond will be more powerful, meaning that it can mine more and more powerful conceptual understanding from their training data.