Gašper Beguš is an associate professor of linguistics at UC Berkeley, where he focuses on interpretable AI and combines linguistics, cognitive science, machine learning, neuroscience, and marine biology.
In this week’s conversation, Yascha Mounk and Gašper Beguš discuss what makes human language exceptional compared to animal communication, whether whales and other animals have true language capabilities, and how properties like cultural transmission and recursion distinguish human speech.
This transcript has been condensed and lightly edited for clarity.
Yascha Mounk: I have a very simple question for you to start off with. What is language and in what ways is the human capacity to communicate by the use of language exceptional in the animal kingdom, or in fact in the world beyond animals?
Gašper Beguš: Actually, it’s not a simple question at all. It’s an extremely complex question. Nobody has a really good definition of language. Maybe definitions are not even that useful. For a while, we’ve defined language as the thing that sets us apart from animals. So it’s the communication system that humans have. But if you look closely, I think the better question is what properties of human language exist and what are the uniquely human properties about our language.
Mounk: Before we get there, what’s an example of some form of language that animals might use, for example?
Beguš: Animals have alarm calls. There is some acoustic vocalization, and that denotes a danger from above or danger from the ground. Those alarm calls are oftentimes not learned vocalizations. With language, you want to have a learned culturally transmitted vocalization. Not something like a cry that you come preborn with, but something that you learn from your caregivers.
Mounk: A dog might yelp when they’re in pain, and a dog in the United States and in Germany and in South Africa is going to yelp more or less the same way, and dogs have an innate ability to understand that that is a fellow member of a species that’s in pain. But that’s not language because it’s not culturally transmitted. It doesn’t vary in this kind of way. It’s not specific enough. Is that what you’re saying?
Beguš: Yeah. For example, you take a vervet monkey or a Campbell’s monkey and have that monkey grow up in the other monkey’s world. That first species will still speak its own innate cries. Whereas if we want to have language, our language is obviously culturally learned. You can take any child, and any child can hear, at the beginning of their lives, any language contrast, any sounds of language, and then slowly they build their internal neural connections in such a way that they only hear and attend to the language that they actually hear. But in principle, a child can learn any language. So it’s learned communication. It changes a lot. We have dialects. The reason we have dialects is because at some point there’s a mother language, and then there are daughter languages, and the daughter languages arise because of mutations and changes in language. So those are all the things that you’re looking for.
Then there are more complex things you have to look for in animals. So is there a lexical principle? Can they talk about things that are not immediately present? I would say there are two broad ways to study language in non-humans. One is to take our language and train other animals in our language. Those would be called language-trained animals.
Mounk: I became obsessed a number of years ago with all of the very cute dogs on Instagram—and, I suppose, TikTok, though I don’t use TikTok—which use buttons with words. There’s a very long history of us fooling ourselves into thinking that animals are communicating with us. There’s a horse by the name of Clever Hans, and it turned out that, in some ways, I forget whether the owner knew that they were kind of cheating and giving hand signals or whether they were doing it unwittingly, I forget the details of the case, but it turned out that Hans wasn’t as clever as he seemed.
But there are some really remarkable clips in which it does seem like these dogs that have access to 100, 150 buttons construct these relatively long sentences, if you want to call them that. They don’t quite have grammar, which are situationally appropriate, and which don’t seem like they’re just pressing a random button like “love you” because they know that their owner seems to be happy when they press that button. There seems to be something more than that going on. Is there?
Beguš: Well, there are even more impressive studies done in the 90s. I started my research in animals with Irene Pepperberg, who was the pioneer of language-trained animals, who had one of the smartest parrots that ever lived, Alex. There’s work on bonobos. Kanzi the bonobo recently passed away, and they tried to bring some interface to the animals so that they could communicate with humans. There’s also work on dolphins, for example.
So Kanzi had this keyboard with symbols that had nothing to do with the object they described, and he was able to use them and even combine them. With parrots, it’s great because they don’t need any interface at all. They learned to mimic our human language to perfection. I have recordings of Alex, actually, and I’m studying them. It’s just so amazing how close they can mimic us and master mimicking our language to the degree that we are never able to do the other way around. So that was kind of the first insight into how smart animals are, because parrots probably were considered smart at that time, but not exceptionally smart.
Now you have Irene Pepperberg training Alex the parrot, and Alex could count up to three or four. Not only that, you could give it a plate of green balls and blue squares and ask him how many green balls, and it would count and distinguish compositionally between shape and objects and so on. That really was maybe the first time we really started to appreciate their intelligence. So that’s animal communication or language-trained animals, where you take human language and you try to instill it upon animals.
Mounk: That’s not a language that animals use to communicate with each other, but it shows that they have a capacity to process language in some kind of way, since we’re able to teach a version of our language to them and then they can use that to communicate with us. What about forms of language that animals have to communicate with each other? For those of you who are listening to this or reading the transcript, I see your face nicely framed by many whales in a beautiful poster hanging behind you.
We hope you’re enjoying the podcast! If you’re a paying subscriber, you can set up the premium feed on your favorite podcast app at writing.yaschamounk.com/listen. This will give you ad-free access to the full conversation, plus all full episodes and bonus episodes we have in the works! If you aren’t, you can set up the free, limited version of the feed—or, better still, support the podcast by becoming a subscriber today!
If you have any questions or issues setting up the full podcast feed on a third-party app, please email leonora.barclay@persuasion.community
Whales seem to be quite remarkable in the way they communicate, including the fact, as my producer Leo points out to me, that supposedly baby whales babble in the same way that human babies babble, seemingly to sort of learn how to communicate, to learn to make the sounds that they need to in order to then accurately communicate to other members of a species. Is that a good example of animal communication within one species of animals, and what kind of languages they have?
Beguš: Well, that’s where it becomes really difficult. Because you train an animal in your human language. You understand what the animal is saying because it’s your language. But if you’re observing animals in the wild, you don’t know that language. You don’t speak that language. You don’t even know what is meaningful in that language, if anything at all.
So you go from this terrain where you have your language and you understand most of the things. Kanzi was probably at the level of a three-year-old human baby. It’s always good to think of this animal research in developmental terms. So roughly, if you compare it to a human child, where are they, how many words, what kind of abilities?
Then you go into animals in the wild, and that’s where we’re really just starting in some ways. First of all, back in the day, there wasn’t as sophisticated recording equipment available. It just takes a lot of work. For example, whales live in the open ocean. They dive very deep. It’s not easy to record them and try to see what they do.
So a lot of animals communicate for mating purposes. They vocalize for mating purposes. So you have birdsong. Birdsongs that are learned, not innate. The father usually teaches male offspring to sing. The idea is, I’m very fit, because in addition to searching for food, I can also practice these beautiful notes.
Mounk: But I imagine that a skeptic would say that that doesn’t count as language because it’s a form of communication, obviously, but it’s really a form of signaling. It’s just showing that I have the extra time and resources to sit around with my dad and be taught this beautiful melody, and I can put protein, whatever else it takes, into my vocal cords so I have a nice, loud voice, and all of that is a signal of evolutionary fitness, in the same kind of way in which a peacock has a beautiful tail. The tail doesn’t mean anything, presumably. Perhaps it does, and we’ve missed it. But it’s not like the pattern of a tail has a semantic meaning. It’s a sign of evolutionary fitness.
So presumably, even though those melodies are passed on from bird to bird, there may be differences within a species in how they sing and so on. A skeptic would say that’s not language in the kind of way in which perhaps whales have language.
Beguš: That’s where we get to whales. The mating function is not that exciting from a perspective of human language or comparing human language to other species. But birds have alarm calls, and monkeys have alarm calls. Now the really interesting things happen with whales. The reason why I love studying whales is because they don’t vocalize for mating at all, at least the whales I study don’t. Humpback whales do vocalize for mating. Sperm whales form complex societies. Based on the dialects they speak, they form families and clans. Crucially, they communicate. They exchange vocalizations, learned vocalizations. We know they’re learned because the baby whales babble.
There’s a lot of variation and dialects, and so they exchange those communications while hunting, before hunting, before they dive two kilometers deep to hunt, and when socializing and during giving birth.
Mounk: Presumably one difference would be that if you take a bird from one locality and put it in the vicinity of other birds of that species, the birds of that species might go, wow, we haven’t heard that melody before. That might make the melody more attractive or less attractive to them. But they would still get the melody in the same way as they might get the melody of a bird in the vicinity. Whereas with whales, because it is culturally bound in that way, it appears to be the fact that the dialect they use changes depending on the part of the tribe or the group of whales you’re talking about. If you were able to take a whale from a completely different geography and put it in communication with other whales, they might struggle to understand each other in the way that I would struggle to understand somebody if I was dropped in the middle of Vietnam and people there didn’t happen to speak English.
Beguš: Some birds’ vocalizations are learned as well. It’s actually interesting. You cannot learn human language as a human unless you get linguistic input and you hear language from the beginning. So there are some really unfortunate cases of child neglect or abuse where children were not able to get language until they were around twelve. That’s the critical period. You’re never going to be able to fully master human language if you don’t get it. The same happens with birds. There are some experiments back in the day where, if a bird with learned communication is not exposed to the father’s song, it will never master it to the same level as if it’s exposed to it from birth. Now with whales, we cannot do such experiments. They would be ethically really problematic.
Mounk: I think with birds it even takes a number of generations to build back up the complexity of birdsong, though it was there previously and so on. I take the point that you made earlier that birdsong changes, it’s passed down from father to son and therefore changes. But I meant something slightly different. We may not be able to do the experiment, but there’s reason to believe that the function of whale language depends on mutual comprehension of the content of what they’re saying. Because different sets of whales have particular linguistic conventions that help them communicate, if you introduced a whale from a totally different group, they may not be able to communicate with each other what otherwise they would.
Whereas with a bird, where the purpose is just a mating call, the purpose is to show how beautiful the birdsong is. If you take a bird from the same species but a different geographic region, that may be a very different song from the local ones, but there wouldn’t be that barrier to comprehension. They would register that the bird sings an interesting song and that it seems to take a lot of evolutionary resources. So perhaps that’s a good signal of evolutionary fitness. It wouldn’t lose the communicative potential in the same way that it might in the case of whales.
Beguš: That’s certainly true for mating songs. The question is that some birds also have learned vocalizations for alarm calls, for example. That’s when it becomes interesting in birds. For birds, in a sense, they’re much easier to study, because there’s a lot of work in animal communication research on birds and elephants. We share so many things with birds and elephants. You can conceptualize a predator in birds or elephants. They have trees, they have air, they have ground, all these things that are familiar to us.
Then you come to whales and you don’t even know how to represent the presence of a predator because the ocean is so vast and the light doesn’t travel at all in the ocean. The light gets in water and gets lost so fast. We are so visual in some ways that we try to represent things visually, but whales probably don’t represent things as much visually as they do with sound. They navigate, orient themselves, hunt, and escape predators by emitting echolocation sounds and then hearing what comes back. So it’s really difficult to conceptualize their world and make a good model of their world, because underwater life is just so different.
Mounk: I think we have some different gradations. We have birds that have forms of communication, but it mostly seems to be things like mating calls. We have whales. It’s very hard to know exactly what’s going on, but it seems to share some additional features with human language, like the fact that there are different languages and dialects and so on. Is there something specific about human language? How is it that human language is different from what whales seem to be doing?
Beguš: Well, I can tell you roughly what the story was for sixty years and how we should change it. The idea was that human language is completely different, and for all sorts of reasons. One was that the main aspect of human language that was considered uniquely human was the so-called recursion. It’s a slightly complex property of language, but roughly it says that you can take one element and embed it into an element of the same type. Yascha said that Gasper said that whales have language. You can always make that sentence larger by saying Leo said that Yascha said that Gasper said that whales have language, and so on. It’s this kind of infinite loop of language.
That hasn’t been found. There are some attempts in birdsong, but in the animal kingdom in the wild that principle hasn’t really been found. It’s a very complex principle, and there’s so much we still don’t know about animals that I’m not really surprised that it hasn’t been found. There are other things, like the idea that animals cannot talk about things that are not immediately present. They don’t talk about things that are in the past or in the future. They have only immediate reference.
A lot of these things were debunked by research at various levels. With language-trained animals, bonobos were able to go to another room and talk about things. One favorite example is Alex the parrot saying “banari” for apple. He couldn’t say apple because it’s difficult for them to say P, so he made up a word, banarry. An apple tastes like a banana but looks like a cherry. He put words together, which shows productive compositional compounding.
Across the research, if you take these studies seriously, you realize that there’s very little that is uniquely human about language. Not every other species has all the properties of human language, but there are bits and pieces across animals, especially language-trained animals. We’re only starting to uncover the richness of animal communication in the wild. Slowly, the properties of human language that are exclusively human are being taken away.
The big step now is also artificial intelligence models. That’s a different story, but some of the things that we thought were uniquely human and never achievable are now achievable in these large language models.
Mounk: There are these incredibly cute Instagram videos of humans communicating with dogs in various ways, and that’s remarkable. But what’s even more astonishing is that I now communicate in natural language with some artificial entity that can easily pass itself off as human these days, that can pass the famous Turing test.
This is very different from what computers did when I was younger. For a long time, we had these symbolic languages that we built, programming languages, and they basically told the computer exactly what digits to manipulate in response to some kind of request. So that was this artificial interface that we needed to learn and that needed to be hard-baked into the computer for us to be able to communicate. That was very useful. It allowed me to run some statistical analyses in a language like R. It allowed much more sophisticated computer programmers to come up with things like Windows, so that we have a visual interface to still tell the computer what to do on the basis of me clicking, and that generates some kind of code, and that symbolic language tells the computer what to do. That’s just vastly different from what we’re able to do today.
Now I can say, hey Claude, tell me a poem, and it will. I can say, I don’t like this part of it. It speaks to me in some ways like a human, some ways not like a human, but using human language. Have these AI systems learned language? Are they cleverly pretending to have learned language? Would that even be a meaningful distinction for us, given that we can now have these in-depth, sophisticated communications in human language with this non-human entity, an entity that is not even an animal, unlike dogs and whales and birds and all these other things we’ve talked about?
Beguš: I think what these models have shown us is that you do not need language-specific or human-specific apparatus, either in your brain or anywhere else. They show us what properties of language are something a general cognitive device can learn. Because animals obviously are intelligent, and it was thought that they’re intelligent but don’t have language because you need something specific about language.
The reason for that was the idea that we get little data as humans. Within two or three years, a child learns language pretty well, and that is impressive. If you think about it, especially because I do a lot of modeling not with text but in a way that human babies actually learn from sound, a human baby needs to learn impressively. Just take a single word like apple. Imagine you’re learning the word apple. You see an apple and you hear “apple” sometimes, but sometimes you hear the word apple and you don’t see an apple. Sometimes you see an apple and you don’t hear anything.
Imagine all the things a child hears and all the things a child sees. It needs to make that match, that correspondence between the sound that denotes something and the actual object in the real world. That’s not an easy thing. We’re trying to model things with ecological data, meaning the same amounts of data that a child gets, and it’s not easy. Of course, artificial AI models are slightly different. They’re modeled according to our brains, but you still have to computer-program them. They have their own quirks, and they need more data, much more data. But at the end of the day, they learn things.
Mounk: So two points just to draw out here to make sure that people follow. First, we’ve gone over that in past conversations with Geoffrey Hinton and David Bau and so on. But if you can give people a very short primer on what you mean by AI systems being kind of like the human brain. Obviously, I assume you’re referring to the fact that they’re both forms of neural networks and that, in key ways, the particular form of artificial intelligence that we’ve gotten turned out to make much bigger advances than the alternative of symbolic AI that people had considered for a long time is in some ways deliberately modeled on the human brain.
The second point to draw out is this remarkable thing that humans actually go by much less data and use much less energy. This is a point that a good friend of mine, who’s a neuroscientist, made to me about his child. His child has encountered vastly less language than ChatGPT has in its training data and runs on cornflakes and the occasional hot dog. Whereas the amount of energy that is used to train the cutting-edge AI models is just vastly more. So just explain those two facets to us.
Beguš: Well, first of all, large language models that we mostly talk about these days don’t learn like human babies do at all. They learn from text, and human babies learn a lot about language before they can read. They can speak by the age of six when they start reading. So reading is kind of secondary. Text is just a secondary representation of language. That’s why they’re slightly different.
I think we also need to build models that are more like humans. But it’s still true. To finish my thought, what LLMs tell us is the limits of what can be learned with artificial intelligence. What are the limits? I don’t think they tell us how humans learn language. They just give us the limits, and that’s impressive. But they’re not really learning in the same way as humans do.
So to go back to your question about the relationship between the brain and AI, AI is, in a sense, a rough mathematical approximation of how neurons work. The idea is that you can learn concepts and really complex things through the connections between units that are called neurons.
It’s a really exciting time now because AI is advancing and neuroscience is advancing, and we’re starting to gain understanding. You can use an AI model to simulate what’s going on in our brain. You can use our brain to better understand what AI is doing. One of the big topics these days is how AI is learning and whether we can interpret it and look inside it. That’s something I’m really curious about as well. The idea is that we now have an artificial approximation of the human brain. There are a lot of interesting experiments we can do and potentially use those tools for all sorts of things, including scientific insight and trying to understand what animals are saying.
Mounk: Now, help me clear up a confusion that I have here. In my limited understanding of linguistics, I thought that the leading theory for how humans are able to acquire language so fast and are able to communicate was proposed by Noam Chomsky, and it’s called universal grammar. I will probably butcher the details here, but the broad idea is that there’s a set of grammatical possibilities that are hard-coded into the brain, into the architecture of our mind. Then there are a bunch of switches that flip one way or another depending on the kind of language that you encounter as a baby and as an infant. Rather than having to learn grammar from scratch, there are pre-existing modalities, and that makes it much more efficient. That helps to explain why, with cornflakes and hot dogs, dear Gabriel can now speak the English language, which is remarkable.
You can see that by the fact that there are grammatical rules that we as humans follow without ever being aware of them. I don’t think that more than 0.01 percent of the listeners to this podcast are going to be able to explain to me off the cuff the rule that governs the order adjectives go in the English language. Yet they will know that “the little blue house” is grammatical and “the blue little house” is not grammatical. The explanation for that seems to be rooted, in some ways, in universal grammar.
Presumably, we don’t really understand the architecture of the brain well enough to have replicated whatever structure this is supposed to be rooted in in these AI models. It’s not like people at Anthropic are studying Chomsky’s models of universal grammar and saying that they have to arrange layers of neurons in a very particular way to code for human language. All we did was throw a bunch of data at these very sophisticated machines, and they picked up on it. In fact, Claude today is 99.99 percent of the time going to say “the little blue house” and not “the blue little house” because it too has figured out the basic rules of language. Does that create a problem for Chomsky’s theory of universal grammar? Does that indicate that perhaps linguists were wrong about that?
Beguš: You can think of that historically. Linguistics of the 60s was influenced by the computer science of the 60s. Exactly what you said at the beginning. That’s how computers used to work. You had a preprogrammed program, and it performed rules according to what you preprogrammed. So the worldview was that there’s a code. There was a background code. There was probably a genetic code that had all those rules, all those switches, all those if-then statements. So language was considered an if-then statement.
Now the paradigm is changing. From logic and if-then statements, we went into neural nets. When you say ninety-nine percent of the time, exactly, those are not rules anymore. Those are tendencies based on trained neural networks. So from rules to neural networks in computer science meant that linguistics is now seeing things differently. People have been arguing for this approach before. The idea is that you learn all the rules from being exposed to data and making those connections inside your brain, those neural connections. At the end of the day, you’re speaking as you’ve been taught by your parents.
I have a pet hypothesis for how we can conceptualize language. We can say it’s informative imitation, and that’s it. Whatever comes after that is a matter of complexity. A child is born and needs to imitate what it hears from the parents. It makes neural connections such that, gradually, you’re approximating what parents are saying, and your vocalizations need to carry information that your caregivers are able to decode. It’s encoding information by imitation and decoding it. Once you have that, you loosen the limits and can admit more things into the idea of language. Then it becomes a matter of complexity. Maybe humans have more complex language than whales, but the basic principle is imitation and informative transmission of information through vocalizations.
I want to say one more thing about LLMs and what they contribute to linguistics. Because they are able to do language now, this idea that you need a prearranged, innate program, some genetic mutation that allowed language, is gone. All you need is a general-purpose pattern discoverer or recognizer. That is really important. That is really powerful.
Another big question in that arena is the relationship between language and thought. For the Chomskyan tradition, language and thought are intrinsically connected. You cannot have complex thought without complex language. What neuroscience is realizing now is that maybe language is just a final stage, an externalization algorithm. You have complex thought and complex processing, and language is taking that complex thought and linearizing it into sentences and words. There’s interesting work on aphasia patients showing that they can do mathematical operations but cannot express them. You can have a stroke in an area of the brain that damages language.
That is another paradigm shift. If language is just an externalization algorithm, then animals and other species might have complex inner lives. We just haven’t figured out the window into those lives yet. In some cases we have. With Alex the parrot, we used our own language. We taught Alex our language and were able to see that it could count and do complex operations. For other species, we just haven’t found that window yet. The hope is that we do.
Mounk: That last thought you had is fascinating to me, and that’s the first time I’m hearing this. To follow this carefully, the idea is that we might assume that what’s going on when you’re reflecting on something, should I choose this restaurant or should I choose that restaurant, should I take this job or should I take that job, should I become a monk or should I become an actor, is a linguistic process. It’s nearly like I’m writing out the arguments for doing A or the arguments for doing B and raising objections to it. Even if that’s not spoken aloud, and unless I’m a little bit nuts, even if I don’t fully understand those words, and even if I don’t have an internal monologue all of the time, what’s going on is basically in the form of language.
Now you’re saying that this is not how AI seems to reason. It’s not how AI seems to produce language. We force them to do that a little bit nowadays with reasoning models. We can see a quasi chain of thought, but that’s actually not what seems to be going on in the layers of neurons that are activating and all of that. Perhaps that’s also not what’s going on in the brain when I’m deliberating about something. What’s going on is some very different process. The last step of that is that, in my own brain, I translate that into language or have an internal voice. Or if I want to explain to you, Gasper, what it is that I’m thinking, I fall back on conventions in order to speak to you. I translate, unconsciously, this set of thoughts into language as a tool for communication. That’s quite a striking thought. Is this one hypothesis among others at the moment?
Beguš: Yeah, it’s pretty complex. Maybe all the concepts. Obviously, you have words to denote things. But when you do processing, are you thinking in this higher-level space that doesn’t necessarily tie each concept to a word? Then you go back and explain everything in words. That’s probably one model that I would subscribe to at this point. The jury is still out, but the idea is that complex thinking does not always involve discrete words.
Some people have an inner voice. Fedorenko says that other people don’t. But the idea is that I think both AI and the human brain operate in many ways quite similarly. There are interesting experiments you can perform in that area. The idea is that language does not penetrate all layers of complex thought. It’s probably easier to be intelligent if you have ways to label things, call objects, and represent them. But when you do complex math, you don’t need the entirety of language.
When you calculate eight times seven, maybe you remember that one from primary school. But if I ask you something more complex, you would do some sort of reasoning, but you would not narrate it in words. One of the most difficult exercises in elementary school is “explain your answer.” You have to externalize everything and explain it step by step.
One really interesting aspect about LLMs is the so-called chain of thought. LLMs become much better when they are not just answering immediately, that knee-jerk reaction to a question, but when they are able to internally subdivide complex questions into simpler ones and then give a final answer. When you asked me what language is, I first started dissecting things. When we allowed LLMs to do that, they became very good.
They also started doing language very well. When you mentioned that readers would not be able to explain why it’s “a big blue car” and not “a blue big car,” LLMs started explaining that. They became very good once they had chain of thought.
Going back to chain of thought, it’s interesting to look inside those LLMs and see what they’re doing. Sometimes they produce complete gibberish, yet the answer is still perfectly correct. The chain itself was something we trained them to produce, to explain how they got to the answer. That may indicate that what’s going on in thought is some kind of high-level combination of concepts that we are computing effectively, and then we go back to language and explain it. That would be the relationship between language and thought that I currently subscribe to. But there’s still a lot of work to be done for that question to be fully answered.
Mounk: How does that relate to this idea that I always hear bandied about in these discussions of artificial intelligence and LLMs, of these bots just being stochastic parrots? I think there are two things going on here. One is that I think this is a bit of a deepity. For those who don’t know the term, a deepity is the idea that love is but a word, and it’s also modern baby talk. The claim that love is but a word relies on a straightforward interpretation that is obviously true. Love is a word in the English language. But there’s a deeper sense that would be more interesting and actually turns out to be implausible, which is that there’s no such thing as romantic love in the world, that people don’t really feel it, that you can’t really trust anybody, or whatever else you might take to be the implication of that sentence. I think that interpretation is untrue, unrealistic, and shouldn’t be taken very seriously.
The seeming depth of the phrase trades on switching back and forth between those two meanings. I think when people talk about stochastic parrots, something similar is happening. There’s a sense in which it’s obviously true. The way these models are trained is next-token prediction. You give them a bunch of text and say, here are the first forty-seven words. What do you think the forty-eighth word, or more specifically token, is? Over a set of processes, they get better at doing that. So yes, they’re trained to predict the next token, and in that sense they’re like a stochastic parrot.
But there’s another sense people seem to imply, which is more like the famous example of monkeys spending an infinite amount of time typing randomly on a keyboard until they produce the works of Shakespeare, where there’s no goal-directed process going on. To me, in the first sense, yes, LLMs are stochastic parrots in a trivial and uninteresting way. In the second sense, it’s just not true. LLMs sometimes mess up, but it’s not the case that they just produce random gibberish and once in a blue moon it happens to be meaningful. Most of the time, they address my question in a clearly goal-directed way, which indicates some kind of understanding.
That’s my general thought listening to you. This is the first time I’m formulating it in this particular way about the stochastic parrot idea. The other thing is that the way you’ve been talking about how humans acquire language, and perhaps how other animals acquire language, sounds somewhat similar to stochastic parrots too. Once we have the right understanding of the sense in which other animals might be stochastic parrots, are we ourselves not also stochastic parrots?
Beguš: Those are some really deep questions. I think we are, in a sense, stochastic parrots. That’s a somewhat controversial opinion. I think we need to understand that LLMs do not learn language like humans. That’s both good and bad. I think people are conflating two things here.
We have LLMs. They can reason about incredibly complex things. It’s not just that they’re providing answers. With chain of thought, they can not only do language and be correct ninety-nine percent of the time with rules they were never taught, but they can also reason about language. They can have metalinguistic awareness. They can do higher-level reasoning and higher-level mathematics.
It’s hard to know exactly what it means to be a stochastic parrot, but the performance they show is impressive enough that I don’t think you can reduce everything they do to a very simple operation. I think what’s happening is that they learn concepts about the world in ways that are similar to how we learn concepts about the world. The mapping between concepts in our brain probably has some relationship to the mapping between connections in their artificial neural networks. They’re able to make those connections, process complex concepts, and deliver impressive performance. Five years ago, none of this conversation would have been happening, and that’s remarkable.
But that’s an engineering product. What becomes problematic is to ask whether they’re doing things in exactly the same way we are, given that we’ve built them to do something different. Humans learn language by listening, speaking, imitating, and getting messages across to the people around us. That distinction is important.
We can use the same architectures and principles from neural networks to build models that are more realistic. Models that need to speak, listen, and interact with the environment. I think we’re going to see very exciting results there, and even better improvements. That will help us understand whether we are similar or different and what the underlying mechanisms are. We need to distinguish between scientific simulations and an engineering product.
Mounk: That’s very interesting. What do you think the implications of this are for how much further progress is going to go on in LLMs? One very simple confusion I have is whether there’s something like an upper limit on intelligence in general, and then, secondarily, whether there’s something like an upper limit on the intelligence of a machine that is trained on the work product of a species that has an upper limit on its intelligence, even if that has biological reasons.
If the best we can feed LLMs in training is the best text we can come up with, and a bunch of other things, is there some way that they can bootstrap themselves off that and go off into a completely different realm of thought? Or is there a reason to think that inherently they’re going to stall out at the top level of human intelligence because that’s all the training that went into it? I don’t know if this question even makes sense, but help me puzzle through this.
Beguš: No, it makes perfect sense. It’s a million-dollar question. I think it’s one we need to look to other species to answer. We talked about the exclusivity of language. In previous worldviews, you needed a language-specific center, organ, instincts, whatever. The other worldview is that you just increase capacity and all these things that we cherish about intelligence and language emerge. Humans are very smart primates, and as you increase the complexity of the brain, things emerge on their own.
I’m much more in the second camp. In the modeling work I do, we build AI models like LLMs, but more realistic ones that learn from speech only and are more brain-inspired. A lot of the things that linguists cherish as symbolic computation emerge. I can show that something like a phoneme or a morpheme, which in the Chomskyan world was considered a rule or a symbol, emerges in neural computation. These neural networks are incredibly efficient, and symbolic-like computation emerges, as it does in our brain. If you increase capacity, the capacity for symbolic-like processing increases, and all of a sudden you have language.
In my opinion, language is a continuum and consciousness is a continuum. As you increase capacity, a lot of the things that we cherish about intelligence emerge. Under that worldview, if you increase the capacity of LLMs or other models, intelligence will increase or emerge. As for an upper limit, some people think LLMs are the wrong way to think about intelligence. But what we’re seeing is that from GPT-2 to 3 to 3.5 to 4, and so on, as you increase parameters, performance continues to increase. If you subscribe to this worldview, I don’t have a convincing argument that at some point they will just level off and stop improving.
One thing I observe in my work with AI models is that they’re not necessarily original. They’re very creative, but I define originality as proposing something that has never been proposed before. Modigliani is an amazing painter not just because he combined styles, but because he proposed a completely new one. That’s something I’m not seeing yet. That may be because companies don’t want to release very original models, since that could go in bad directions. In general, all else being equal, as you increase connections, intelligence increases.
In the rest of this conversation, Yascha and Gašper discuss using artificial intelligence to gain insights into Whalish, what creativity shows us about humanity, and what it means to be conscious. This part of the conversation is reserved for paying subscribers…
Listen to this episode with a 7-day free trial
Subscribe to Yascha Mounk to listen to this post and get 7 days of free access to the full post archives.
