Event Replay: Understanding Animals: AI Helps Scientists Interpret Language Across Species

[00:00:00] Speaker 1: Well, happy Science Week at OpenAI. We're so excited to see all of you. And I don't know if you guys heard, but OpenAI just launched Prism, the AI native workspace for scientific writing and collaboration. So thanks to Kevin Weill's team, VP for science, over there in the back. And is Victor here? Victor! Victor also is a part of that team and we owe a lot to him. So if you want to learn more about Prism, talk to Victor. He's in the audience today. So this is what our program is going to look like this afternoon, friends. We are going to start with Gasper Begas. He's a professor of linguistics and also part of the SETI program at UC Berkeley. He's going to present. And then our VP for science, Kevin Weill, is going to join him for a fireside chat. We'll do 20 minutes of Q&A afterwards, and then you all will have an hour to meet with each other. I want to call out a few awesome people that are here today and I'm going to continue to do this because I really want you guys to connect with each other. But we have Eva Silverstein here from Stanford. She's a physicist. We have Greg Schmidt, he's NASA Director of Solar System Exploration Emeritus. We have many more and I'm going to continue to introduce you to each other, but first please help me in welcoming Gasper.

[00:01:38] Speaker 2: Thank you so much. All right. So I want to talk about today how do we make discoveries in science. And up until today and these days, we had to do everything pretty much ourselves. We had some tools, but they're not as smart as they are today, so the human was in the loop and the only one. I think now with AI, we have this powerful tool and I think of AI as a tool for scientific discovery in analog of a metal detector. So if you need to find a ring on a beach, it's very good to have a metal detector. It's a tool. It's not yet an end product, but it's a tool that can help us push science forward. And I'm going to talk about three case studies for how we can do that.

[00:02:44] So what is science all about? Science is about finding causal explanations in the world. And I'm very passionate about AI interpretability because it pairs the extreme performance of these new models that we have with causal insight, causal explanation. So the three case studies for how AI interpretability can push forward science that I'm going to talk about today is LLMs as a medium for testing what are the limits of silicon intelligence? What are the limits of something that is intelligent but is not biological? Then I'm going to talk about AI and LLMs for simulating what is possible. These models are extremely creative, and they can help us challenge our assumptions. They can help us think outside a box by creating new synthetic data that does not exist, that might seem crazy to us but are actually super useful when we have very difficult questions at stake. Science usually deals with very difficult questions. If they weren't difficult, they probably wouldn't be science. Finally, I'm going to talk about how AI interpretability, so basically looking inside the networks, doing this kind of neurosurgery on the models, can help us discover something new. The case study will be whale communication and I'll talk about how exactly I was able to use AI interpretability to find something we didn't know before existed in whales, and now we know it does and how that can bring us closer actually to other species, to these wonderful intelligences that live in the ocean.

[00:04:41] I'm going to talk about whales. First thing, let's try to use LLMs to test the limits of silicon intelligence. For the longest, it was considered that you have to be a human to learn language. So either you have some special device in your brain or your language abilities are innate, but that was the assumption. For 60, 80 years, only humans can learn human language. By just looking at it, some intelligent creature cannot do that. And that assumption is now basically not true anymore. And the whole paradigm of language is changing. So we can test the most difficult aspects of language. So what is language? It can be many things. Language is very difficult to define it. And so instead of defining language, we can look at where language might be applied.

[00:05:40] Speaker 1: Instead of defining a language, we can look at animals and see what other species have that is similar or different to our language and what is uniquely human about language. When researchers were doing that, the one thing they came up with as the only human-defining aspect of language is recursion, or the infinity of language. I call it the infinity of language: this idea that you can take a limited set of words and express an infinite number of ideas. You do this by embedding one piece into a larger piece infinitely. For example, you can say something like, "John's grandmother's uncle." You can also say "John's grandmother's uncle's aunt" and infinitely make this sentence more and more complex, like "John's grandmother's uncle's aunt's brother." Now, I think your processing is probably not catching up, but with LLMs, you can actually test these sentences precisely.

[00:06:52] The phrase "the mouse and the cat that the dog painted thought sang" should be a good sentence in English, according to grammatical rules, but it’s probably very difficult for you to process who did what. Nonetheless, we can test models. For about 5-4 years, the LLMs have started to do language, and they are starting to do it extremely well. By being exposed to language, they can process these kinds of sentences that we thought only humans were able to process, and they do so with a high degree of accuracy. We came up with a lot of interesting sentences, like "the world view that the prose Nietzsche wrote expressed was unprecedented." Not only was the model able to parse this sentence properly, but it was also able to analyze it just like a linguist would, with syntactic trees.

[00:08:01] So, that's a case where we can use this kind of behavioral interpretability to understand how the model is doing language internally. Basically, the model can perform at the level of a PhD student in linguistics. We hire PhD students in linguistics to evaluate the model’s performance on creating syntactic trees of the most complex sentences. Not only can the models generate good English without grammatical mistakes, they can generate the most complex sentences and analyze them metalinguistically. We call the ability to analyze language itself metalinguistics. This is a higher degree of cognition; it’s metacognition. The model is also able to add another layer of complexity, such as in the sentence "the world view that the prose that the philosopher Nietzsche admired was unprecedented." This is pretty difficult for humans to parse.

[00:09:45] It shows us through these experiments that something we thought was uniquely human or that you had to be human to learn language is no longer true. This high level of metalinguistics and metacognition is something we’re starting to see in the models. That was three years ago with the chain of thought models like O1, which significantly changes our perspective on what language is and who can learn it. If you don’t have to be human to learn language, maybe other biological species have something complex and interesting as well. However, other species are very difficult to understand because we don’t share the exact same language.

[00:10:23] Speaker 2: So one case study where AI can help is to simulate data or to simulate what is possible, what the possible languages are. We have our own human language that we know and understand, but there are 6,000 languages in the world, and they all share a lot of commonalities. Other species might have something completely different. LLMs for synthetic data represent a really powerful approach in science because sometimes the problems are so hard that you need to exercise how you're going to tackle a problem. For example, the animal communication problem is quite challenging because sometimes you don’t even know what is meaningful in that communication system.

[00:10:46] We created a conlang crafter, like a genetic pipeline, where models are creating endless new languages. We have 6,000 to 7,000 languages in the world. So far, people have been making up languages, called conlangs. You may know Dothraki or Klingon, but it takes a lot of work to come up with Dothraki and Klingon, and it’s not easy.

[00:11:21] Speaker 1: And it's not easy. So now we have a pipeline where several models are making up languages, creating their complex grammar from phonology, morphology, syntax, and they're refining the grammar that can translate sentences and basically provide us with invaluable simulated data where all of a sudden you can generate thousands of new languages. Now, why would that be important? Maybe we have enough languages already. Why would you make additional, make up additional languages? Well, if you're approaching something new and complex and something you don't understand, it's really good to have an exercise in what's possible. So, you can take our conline crafter and tell it, okay, make up a language which is produced by an alien cephalopod species where phonemes are color values and gestures rather than consonants and vowels, and then put them in ASCII. And the model will do that. So, the model will just generate sentences in this language that is nothing like human language, but it gives us materials to practice on for when we need to decipher, say, something that animals have.

[00:12:59] Animals is the ultimate test ground. Trying to understand what they're saying, getting closer to them is something that I'm really passionate about. Before we do that, we need to have a better understanding of human language, in my opinion. So, for the longest we thought humans only can learn language. When that's not true anymore, can we use AI and modeling to better understand what our language actually is? And I think we need to approach this problem by combining all these three players now. And that's what I'm doing, what we do in our lab, combine the knowledge we have about linguistics and neuroscience with modeling and with what other species have. So, this is kind of mutual informative intelligences, humans, animals, and machines.

[00:13:56] But what language actually is? It's very difficult to define it, but we can try. AI helped us come up with a different framework that we're trying to use now to expand the definition of language to other species. When we learn language as children, we have no text, we have no Wikipedias, we need to learn from raw sounds by interacting with parents. We need to listen, we hear a lot of things, and we see a lot of things. And the child has this extraordinarily difficult task of pairing what it sees and what it hears and making that connection, right? So it sometimes hears a red apple and sometimes sees a red apple together. Sometimes it also just hears the red apple word and sometimes it also just sees a red apple and doesn't hear a word, but the child is able to, by interacting with the world, understand that red apple at the end of the day is red apple and that's how you call it and you know, it learns a word. It takes a lot of effort to learn a word.

[00:15:08] In a sense, this is, in machine learning, still an unsolved problem, how a child does it with so little data. So we try to approximate this process for how humans learn language as closely as possible with modeling. And we model language as informative imitation. So a child is born and it needs to imitate what it hears from the parents. It needs to learn to move its mouth such that the sound it produces approximates what parents are saying as closely as possible plus it carries some information. And this is the model that we've been able to devise. I won't go into details, but the model has a mouth and it needs to change the shape of the mouth. It is a GaN-based model that forces imitation and it forces informativity. So you think of it, we imitate and we encode information into the sound. Maybe whales do the same. So that was a useful conceptualization of language and it can spread to other species as well.

[00:16:18] We're able to show that these models are able to do a lot of things. They can learn words, they can put words together, they have the process of creating syntax, simple syntax on their own just by this idea of informative imitation. And also in the brain, artificial neural network comparison, the signals inside these models is very similar to the signal inside our brain when we listen to speech. So we have a model that does well on human language. This is actually the recording of the brain activity when you hear it.

[00:17:01] Speaker 1: Recording of the brain activity when you hear it in the speech, and it's very similar to our models. Can we use that approach to find something new about whales? Now, our whales are amazing species. They're sperm whales. This is off the coast of Dominica and what you see are two whales that are swimming toward a third whale and they exchange vocalizations that look nothing like human vowels, nothing like human speech, nothing like even human language. They sound more like Morse code, like alien Morse code. But they exchange these vocalizations and it's not for mating and it's not for some sort of anti-signal. It appears that something meaningful is being transmitted here. They have very complex social structures, they have dialects, and so they have this alien communication system that we have very little knowledge about. For the longest, this was considered a Morse code-based system where basically the number of clicks, they put clicks, the clicks you hear into groups of clicks that we call codas, and they exchange those codas in conversations. And for the longest, it was considered that this is a Morse code and that the number of clicks matters and the timing between them matters.

[00:18:34] You can hear click, click, click, click, click, click, click. There's one nice pattern, and you would say, okay, that really doesn't have anything to do with human language. We don't click. We don't do Morse code. We don't do stuff like that. Well, we use our model that learns human language really well and say, okay, if it can imitate humans well and do that informatively, why don't we try to train it on whales and their communications and see what would a model learn to be meaningful about this communication because we certainly don't know yet what is meaningful. So can a model learn something that might be useful to us? Then the crucial part is the internal interpretability. We have this technique called CDEF where we can take basically a single neuron inside the model and tell it what the neuron had encoded. What did the neuron learn? This is very precious because monosemanticity in deep learning does not always pan out, but in these models, we have pretty good evidence of monosemanticity where a single neuron represents something meaningful about the data.

[00:19:57] And we can causally manipulate that neuron to understand something about that data. When we apply this internal interpretability technique in our models that learn to imitate whales, the models suggested that yes, timing and the number of clicks are important, which is what we suspected for Morse code, right? Morse code is about timing and the presence of a signal or number of clicks. But also that the spectrum is important. So we were able to get a hint from internal interpretability that told us, okay, try to look into spectrum in a similar way as a metal detector will not tell you where exactly a ring on a beach is, but it tells you where to look for, where to dig for.

[00:20:43] And we have to do a lot of digging on our own as scientists, and then we realized the crucial final step in the discovery of these really rich structured spectral properties that you're going to see is that whales are just slow. We have our anthropocentric perception of timing. We are very fast, and they are slow. And if you have that, then you can start taking these slow click, click, click patterns and removing timing from them. And once you remove timing from them, once you basically speed them up so that they're more in line with the human perception of timing, you start noticing these spectral properties that people were not noticing before that are highly structured, highly controlled, and repetitive across whales. Across whales in Dominica, across whales elsewhere in the world.

[00:21:38] And that's where we discovered that, wow, these spectral properties are basically exactly analogous to our human vowels. Some quotas have one formant or one of these spectral bars, others have two. And this is very similar to how we humans distinguish vowels. You can even hear these differences with your naked ear. So those are two sped-up quotas. What sounded like a Morse code to us now starts sounding as pitch but with different spectral frequency qualities. And if you speed their entire conversations up, you start appreciating just how complex this might be. So they're going back and forth. What we consider this to be a simple repetition of a signal. We thought, oh, they're just repeating this.

[00:22:42] Speaker 1: I thought, oh they're just repeating this call constantly, this 1 plus 1 plus 3 code constantly. Actually now we know that those are two types. We call them the A vowel and the I vowel in analogy to human vowels. And then we realized, well actually, the way they produce these vowels is incredibly similar to human vowels as well. So we have vocal folds that we vibrate and then on top of vocal folds we have our mouth that we change the shape of to get different spectral properties. You see these red arrows show you these shadowed bars. Those are the spectral properties that distinguish say an ah from an E. And whales have that too, but instead of vocal cords, they have phonic lips. And instead of our mouth, they have an air sac inside their nasal complex. So, even production of these sounds is similar. And this is a case where AI basically with the discovery helped us realize that on so many levels, they're actually, were so much more similar than we thought. Okay, so here's a short demo of actually how vowels work in whales and humans.

[00:24:03] On the surface, sperm whale vocalizations look nothing like human speech. The whales click. But if we look more closely, we find near identical parallels between human vowels and whale clicks. So, how do we produce vowels? We say all vowels by vibrating our vocal cords. By adjusting the shape of our mouth, we can say different vowels such as A and I. When we say A, we lower our jaw. When we say I, we place the tongue much higher. These different mouth shapes result in different resonant frequencies, or these shadowed bars. For A, the bars are close together. For I, they are far apart. So let's look at the whales. Just as humans have vocal cords, whales have phonic lips. Just as we have mouths, they have an air sac which they change the shape of. But still, they click. On the surface, their clicks sound nothing like our vowels. But that's because their clicks are slow, and our vowels are fast. We needed to change our perception of timing. If we remove silences from their clicks and make their clicks faster, more adjusted to human perception of timing, patterns start appearing that look similar to our vowels. The red arrows point to shadowed bars, just like those we saw in humans.

[00:25:33] We found two vowels so far, the A and the I vowel, and we realize they exchange those vowels in conversation. You can hear what their conversation sounds like if we make their clicks faster, more in line with human perception of timing.

[00:26:03] And the nice thing is that these patterns are so clear, so discreet, that you can take a piece of paper and a pen and start transcribing them with our human letters. These clicks were recorded from a sperm whale named Pinchie, who was communicating with a nearby whale off the coast of Dominica, where SETI's research takes place. Pinchie is my favorite whale. She's an old lady, a grandma, and very chatty. It was the first time I saw the vowels in her.

[00:26:33] So what we're learning is that maybe we're not the only species that can have complex language. And maybe our language is just one form of communication, and it has some spectrum of complexity, and you can have it more or less complex. And my hope is that linguistics and these kind of discoveries can feed into this loop of science where science informs linguistics and linguistics informs science, and with other sciences is the same, right? So we learn from each other, and we make progress.

[00:27:11] For example, 150 years ago, Darwin proposed his revolutionary idea about evolutionary theory, and he was inspired by linguistics directly, because 60 years before him linguists realized that you can have an ancestor and evolution of language, and Darwin used that principle of evolution to apply to species. What we hope is that these discoveries that bring us closer to animals, that bring our language closer to animals, and also the research that shows that they have an incredible complexity in their communication system can help us make changes in other fields as well, not only in science but also in law. We have a legal paper where we argue and explore how finding language-like structure in animals can help us with the fight for their rights. We're able to show that their communication system is one of the most complex in the animal kingdom.

[00:28:22] Speaker 1: Most complex in the animal kingdom. And if you have a complex communication system, that's highly indicative of complex internal lives and internal worlds. We recorded the birth of a sperm whale. We showed that they can take responsibilities. It's a really exciting field where science can interact with the legal world and policy world. We also collaborate with artists a lot.

Speaker 1: So AI is bringing a new reality. Basically, it's introducing a new reality, which are these latent spaces. These internal hidden layers that we're starting to explore with techniques that I just proposed. We paired up with designers and artists in this project that we call Latent Spacecraft, where we use art to explore these new internal realities, these latent spacecraft, these latent spaces.

Speaker 1: And here's an example of a video where we trained the model on one of the most impressive works of art, The Finnegan's Lake, by James Jones. We are exploring through layers how the models themselves are creating language and what kind of insights we can get for philosophy, for architecture, for art history. So in the final layer, the models start saying coincidences that are very similar to The Finnegan's Lake and basically generating all anew. This will launch tomorrow, so you can have a preview at the QR code. It's called Latent Spacecraft, and it's launching in the Antikythera Journal tomorrow.

Speaker 1: So we have these powerful tools that can help us find new properties about the world, find new things, and learn about us, and learn that maybe we're not so exclusive as we used to think. So in a sense, it's very humbling, and it's exciting because for the first time, we're really studying in parallel these three mutually informative intelligences: the humans, the animals, and machines. Thank you.

[00:34:03] Speaker 1: I think it's a little bit more through the sperm whale communication work in particular. Do you have a sense of how sophisticated their language is? What kind of concepts are they expressing? Are they communicating socially or just communicating danger at a more tactical level? What have you learned? The cool thing is that there's a lot we don't know, but when you look at it, you feel that there's something really rich there. We do know, though, that they have extremely rich social meaning transmission in that language. They have dialects. They form complex societies, so they have families, then clans. Basically, you can hear—I can listen to a whale communication for about a minute, and I can know roughly where in the ocean they're from because their dialects are really specific. They express a lot of social meaning identity through their dialects. But the situation in which they speak makes you think that there's something really interesting going on there. They speak before they hunt, before they dive. They dive like two kilometers deep, which is mind-blowing. They go for 45 minutes in order to hunt, and then they socialize and hang out. They go to sleep. They have collaborative births. We recorded the first birth of a sperm whale; it was never recorded before. Eleven females came together to help deliver the baby and lift it up. And they were talking like crazy during the birth event, obviously. So, you would think that there's something interesting going on, also in terms of meaning, but we just haven't decoded it yet.

[00:35:54] Speaker 2: That's amazing. Did they talk?

[00:35:57] Speaker 1: This is sperm whale to sperm whale. Is there any sperm whale to other whale communication?

[00:36:02] Speaker 2: That's a really interesting question because when teams saw that their birth is taking place, all of a sudden, tens of pilot whales and hundreds of Fraser dolphins came around as well. So it was like a big Lion King party in the ocean. We're starting to study how the interaction works, and there's some studies by other teams showing that various species can interact and hunt together. So there might be more cross-species communication than we used to think, actually.

[00:36:37] Speaker 1: Fascinating. Speaking of cross-species communication, I told my 9-year-old daughter that I was interviewing you today, and I gave her a little bit of info about what you work on, and she said the question, which I think is on all of our minds, which is: when are we going to be able to talk to our dogs?

[00:36:54] Speaker 2: Well, we're already talking to our dogs big time, right? But it's a really interesting point because what we need with animals is just a channel, right? And people have come up with different channels to get into their inner worlds before. Like, you know, there was work on primates, there was work on dolphins; you can give them thumbs or various different interfaces. My favorite example is Alex the parrot, African gray parrot. I started actually in animal research with Irene Pepperberg, who is this amazing scholar. Basically, the nice thing about parrots is that they can mimic human language to perfection. The medium to their inner world is our language. And that's great because we're good at understanding our own language. We're pretty bad at understanding other languages. But it was only when Alex the parrot, this incredibly smart parrot, learned our language that we were able to appreciate just how smart parrots are because through our language it communicated to us that it could count, that it could compositionally distinguish shapes and colors and stuff like that—stuff that was considered impossible for an animal at the time. So, to your question, I think AI has the potential to build these interfaces. But I as a linguist, I'm primarily interested in how their systems work. So for now, I prefer to listen to whales. And with dogs, maybe AI can provide that kind of bridge that can help us get into their inner worlds.

[00:38:38] Speaker 1: You've definitely got a trillion-dollar company on your hands if you can decode dog for everybody.

[00:38:43] Speaker 2: Dog, yeah. Not whales, though. Maybe not whales, but you know. Yeah, I think it's just—you’re right. We can communicate with animals in various ways already, but it feels like something more fundamental in actually being able to understand what they're saying. And not just figure out sort of a pigeon between the two of us. What do you think changes about the world if we can really start to decode all kinds of animal language, even if imperfectly, and maybe communicate back at higher fidelity than we do today? It just feels like something changes about our sense of place in the world.

[00:39:22] Speaker 1: A big time. I think AI and animal communication together are basically, in a sense, humbling to humans.

[00:39:30] Speaker 2: Very. Because we're starting to notice that we're not as special as we used to think. I think in the 60s, in the age of symbolic AI, we were the species that can do symbols. Since Aristotle, basically, we are the species that can do symbols.

[00:39:43] Speaker 1: Basically, we are the species that have language and others don't. And I think that line is now blurred. And there's like a barrier. I mean, this kind of work shows you that these barriers are pretty much artificial and they're very difficult to defend. And, you know, I think we're, in a sense, hopeful because I think that will bring a lot of good things as well, like, you know, rights of animals. And I think we're going to start appreciating biological intelligence big time because, you know, we always create our identities in the relationship to the other. Now, AI is becoming the other. And then I think biological intelligence will become a value per se. And I think that's good.

[00:40:31] Speaker 2: Yeah. It's interesting to think if it's really more of a quantitative difference. We're all maybe at different places on a sliding scale relative to animals versus there's a binary classification where we can do these things and they can't. So maybe we'll switch gears to AI for a little bit. I'm curious how your use of AI and your experience of AI has changed over, say, the last 18 months. Like, have there been moments where thresholds have been crossed that have made it fundamentally different or more useful for you?

[00:41:03] Speaker 1: Scientifically, the threshold was... I tweeted about the ability of, I think it was GPT-4 at the time, of doing analysis of language on March 14th, the day it got out. Because up until then, it was kind of curious. The models were able to do language pretty well, but they were not yet able to analyze it. And I think that was a big thing for me that AI can do metalinguistics. Basically, it can be a graduate student of linguistics. And then the Conline Crafter is a direct consequence of that. I think also, we're all starting to learn how to use it as a tool, and I think that's important. But there were definitely these milestones, and one of them was that. The other was kind of humbling, especially because I had a quick mathematical problem to do, and it would take me about three or four days. I asked the model to do it, and the humbling part was like, okay, it did it well, I saw it did correctly. And the humbling part was like, okay, now go step by step and tell me exactly what you did. I was like, look at, I thought to myself, look at me, you know, asking these models to explain what they're doing. I mean, I've been there too. It's humbling for the first time it happens. But then, you know, I think you start appreciating where you are as a scientist, where you're still irreplaceable. And I think that's, you know, originality, boldness, finding, you know, something that might, you know, a vanilla model might think not, it might not work. I think that's where, you know, scientists might still have a role to play.

[00:43:13] Speaker 2: Yeah. Yeah. I'm really excited about the idea that every scientist has their own special, you have your area that you focused on where you are a world-leading expert. But no human is a world-leading expert in everything. And there's a certain friction in collaboration with other people. You're trying to do some math, in this case, you would have figured it out. But maybe there was harder stuff you were trying to do in some other adjacent field. And you didn't know anybody, or it would have been hard to find. So, you know, there are probably avenues where a lot of scientists say, you know, this is just not going to be worth it for me to go try and figure this out. And now there's, we live in a world where you have easy access to a model that's read every scientific paper that's infinitely patient. And I think that's going to be its own form of acceleration. And I'm curious, do you have a sense of like, is there a next threshold that you're waiting for?

[00:44:04] Speaker 1: Man, if only the models could do this for me. It is, I think, if only models can decipher something that isn't undecipherable. Okay, then decipher so far. Or, you know, in mathematics, obviously, there's a lot of thresholds, but like, I think the next big thing for me will be when they do something original that was not done before. And it's easily verifiable. And, you know, it's been difficult to bet against the models, but we'll see how this time it works.

[00:44:35] Speaker 2: Yeah, I mean, you talk about a breakthrough being GPT-4. I remember at the time how mind-blowing GPT-4 was. And now you think if I told all of you that you could only use GPT-4 for the rest, you'd be like, Oh, I can't make any progress with this. What do you expect me to, you know, it's, and that was not that long ago. It's incredible how fast this evolved. So if you controlled open AI research for a few months, what would you have us work on to accelerate science? Could be in your own field or broadly.

[00:45:06] Speaker 1: That's great. I have a suspicion that, because the models are post-trained to be pretty polite, that that interferes with their ability to do you know cutting-edge science. So I think I would give...

[00:45:24] Speaker 1: So I think I would give scientists access to slightly less post-trained, slightly less polite model that, you know, because often times you think of, you know, there's this idea of a crazy scientist, right? You have to be bold. You have to be against the flow often times. I mean, you know, when you're proposing something new, you're not a vanilla scientist, and you're exposed to a lot of dangers, right? You're really vulnerable when you're claiming something that hasn't been done before. And I think that boldness, like, you know, working together with scientists, increasing that boldness, I would also give scientists a lot of inference time.

[00:46:05] Speaker 1: I agree with you. And I think there are also really important avenues that we should explore in interpretability, internal, right? So if that was the case, the last case study, right? We needed to look inside the hood to understand what did the model actually learn? Because the model learned to produce real codas, right? So it learned to mimic. But to understand what is new and important, we needed to look inside, in the connections, in the neurons, manipulate the neurons. So the ability to... Obviously, not all models... As a user, you cannot get into the models' weights. But that ability, I would give scientists access to that.

[00:46:52] Speaker 1: Okay. So more direct and bold and a little less supportive by default. And interpretability. So on interpretability, it's a super hot topic right now. It's a very tough and active area of research. But you've gone deep and, as you said, used it to help understand some of the work you're doing. What did you learn in the process of digging into interpretability about AI models that you didn't know before? Sometimes going way into the weeds on something teaches you a bunch.

[00:47:25] Speaker 2: Yeah, for sure. One thing that was actually interesting is that we realized that the models, when they're trained on very difficult tasks, they start coming up with their own languages. They start encoding information into things that you don't expect as humans. And that's really chilling because they're starting to develop their own communication protocols. And that was mind-blowing because we were working with these models and there were these weird silences that we couldn't make sense of. And then we realized, when we looked inside, while the model is using that to transmit information, it's kind of internally communicating with its own protocols.

[00:48:18] Speaker 2: And I think as we move forward, that is going to be very important because we'll need to understand in the age of agentic AI when the model starts communicating with each other, the languages actually might not be the most efficient communication protocols. I'm almost sure.

[00:48:36] Speaker 1: Exactly. And so they'll probably invent new ones and understanding those will be crucial. And so I think that you can only get that by looking inside the models and the weights.

[00:48:51] Speaker 2: So let's talk about conlang for a bit. You gave a bunch of examples on the screen, right, and those examples looked very foreign. They were elements of it that kind of felt human, but they felt very different than at least any language that I've seen. So is there some sense... And you said, look, the space of languages is super broad, colors, gestures, etc. So is there some sense in which human languages exist on some lower dimensional submanifold of the space of all languages? And if that's true, what do you take from that?

[00:49:29] Speaker 1: Yeah, I mean, we just have one language, and it's probably evolved for efficiency given our biological intelligence, right? And if you take this idea that language is a continuum, we always thought we're at the end, right? I think that was always the case. We are at the end of the food chain. We are at the top of intelligence. But now we're seeing that maybe we're not. And maybe our language just falls somewhere on that continuum of complexity.

[00:50:03] Speaker 1: There's really interesting work on unsupervised translation where you can test how complex a language can be by creating new languages with AI as well. So I think now we're kind of rethinking whether we're really at the top, and maybe there's some complexity. And what a con-lang crafter can do for example is challenge that and push us towards more complexity because it's very difficult to imagine things that are more complex to your limits of complexity. By the way, there's also a unicorn. Imagine a con-lang crafter creating alien languages in video games.

[00:50:51] Speaker 1: You have all the sophisticated new languages there. Every time a user does something. But I think the important thing is that it challenges our imagination. And that's really, really important when you're approaching animals where we didn't know even where to start.

[Speaker 1] To start. Yeah. Do you have the Conlang decoder? Can you take a language that Conlang generates and then feed it back into a different model? Yeah. Doesn't have it in context and ask it to decipher and try and understand what it is. Well, actually, it's the necessary ingredient of Conlang crafter. There are two. It's a genetic pipeline where one model works with another model because they need to refine the grammar. So, yes, they're talking to each other and translating sentences. But then can you take something that is the output of Conlang and take it to a totally separate instance that's never seen any of it before and say, hey, help me understand what this means? That would be a cool exercise. No, we don't have that yet where the other model hadn't seen anything about the language. But it could be like a good simulation for like decoding some of these languages like Rongorongo, for example, is this beautiful writing system that have never been decoded. And that would be like a, you know, a big breakthrough for me.

[Speaker 2] Where is Rongorongo from?

[Speaker 1] Rongorongo is an island. It's a beautiful, beautiful script. You can see it on Wikipedia. It's really interesting.

[Speaker 2] Yeah. So when you're building something like Conlang, you're exploring the space of all possible languages. How do you determine what is allowed in a language? How do you draw the boundary and say this is on this side, if it's this, well, that's a language, but this is clearly not. That's not part of the language.

[Speaker 1] Well, you know, there are things in language that don't exist. For example, no human natural languages express negation by just changing the order of words. But, you know, in principle they could do that. So, you know, like we do questions. Sometimes some languages just change the order of some words and you get a question out of a declarative sentence. Negation doesn't work like that. Now why that is the case has been the cause of dispute in cognitive science and linguistics for decades, right? So, you know, the Chomsky Linguistics Circle would say, well, that's because humans come with a pre-native ability to do language, and then more usage-based approaches. People would say, well, no, language evolved for efficiency and that's just what makes it efficient, easy to learn, easy to communicate, easy to, you know, pass on to generations. And we think we all, a lot of the times, forget the cultural aspect of language. And there's another thing that AI can help with agentic AI, right? So now we can have agents that start talking to each other and we can simulate—that's something I didn't talk about—we can use these models that we're developing in our lab to simulate how language evolved in the first place. What were those, you know, initial mechanisms that allowed us to develop language?

[Speaker 2] Super cool. So I have one more question. We'll do a quick lightning round and then we'll go to Q&A, so start thinking about your questions. You take this research that you're working on, fast forward a few years, where do you think you'll be or where do you want to be with your research by 2030? And how does that compare to where you might have said in 2020 you'd be with your research in 2030 before, you know, the advent of modern AI?

[Speaker 1] Well, I think that 2020 to now, the one thing that happened is an LLM was able to learn language to the highest degree of complexity. That was a paradigm changer to me. From now on, I mean, it's difficult to predict the future. I hope we do a lot—I'm really driven by societal impact. You know, can we use these things that we learn to make changes in the legal world, to make changes, you know, in philosophy, this kind of cross-pollination of science? I think that's very exciting. I'm very hopeful about the neural stuff as well. So I show the similar signals in the brain and in the models. I think we can understand the human brain better with these simulations. We can—we're getting, you know, we're learning about the human brain and how it processes language really well. And that can have huge clinical implications as well. So I'm really hopeful about the brain-neural network kind of cross-pollination of ideas there.

[Speaker 2] That's awesome. All right, quick lightning round. So these are going to be fast questions, quick answers. You don't get to explain. You've just got to take a stance. Dogs or cats or sperm whales?

[Speaker 1] Sperm whales, of course, but probably dogs.

[Speaker 2] Okay. Smallest hill that you'd be willing to die on?

[Speaker 1] That's a pretty big hill, but okay.

[Speaker 2] What's the thing that working on whales has taught you?

[Speaker 1] That our lives are not—our environments are not taken for granted. They never see a tree. I see a tree every day. They never have air. I mean, they breathe air, but it's different. Gravity works differently. It's just different worlds, but the same intelligence. It's really fascinating.

[Speaker 2] Who's your scientific hero?

[Speaker 1] Ferdinand de Saussure, who was the first linguist that basically influenced all of the fields in humanities in the early 1900s.

[Speaker 2] One bet for 2026?

Speaker 1: [00:56:45] 2026, a lot of scientific discoveries, hopefully. [00:56:53] Just in general? [00:56:56] A lot of fields, from protein design. [00:56:59] I mean, the hypothesis, the ability to shrink the hypothesis space and have that metal detector to help you know where to dig is going to be huge. [00:57:09] I hope. [00:57:10] I like that analogy. [00:57:11] One bet for 2030? [00:57:14] That LLM's develop their own communication protocols that we don't understand. [00:57:21] Last one. [00:57:22] One piece of advice for the scientists in the audience. [00:57:26] I think it's an incredibly exciting time, and just use those tools to the best of your ability and learn how to use them. [00:57:37] And think about larger scale impact like societal impact, positive impact that can make. [00:57:45] It's awesome. [00:57:46] Thank you so much. [00:57:47] Okay, well, let's go to Q&A. [00:57:49] Okay, thank you so much, guys.

Speaker 2: [00:57:51] We're going to start with one online question, and it's from Daniel Green, one of our community leaders. [00:57:56] Do you think AI will ever let us even in a very limited way have two-way interaction with an animal? [00:58:02] Is it likely to be with a whale?

Speaker 1: [00:58:04] That's a great question. [00:58:06] You know, it's hard to make predictions, and imagine if I had the perfect translator between me and whales. [00:58:14] You know, I would say something, it would translate into whales. [00:58:17] I as a linguist would still not be satisfied. [00:58:20] I would still want to know how exactly syntax works and how exactly what are all the moving pieces, and is it really what they're saying that matters, or is the model picking up on context and other things? [00:58:33] And I think for that, we need to first listen and then maybe get to communication later. [00:58:45] Thank you so much. [00:58:46] Very exciting.

Speaker 3: [00:58:47] I'm not sure if I have a question or a comment, but I work with the animal labs, including the mouse in the lab, and one of the questions or comments is that is there any study or research on them, and how can we understand them better? [00:58:59] You know, for example, if I'm doing a glucose type of response test in an animal, then how can we study them, how can we understand them better?

Speaker 1: [00:59:01] You know, it's hard to understand that, but we're only at the beginning in many of these species. [00:59:19] I mean, there are like many cetacean species for which we don't know almost anything. [00:59:25] We don't know what they – you know, how they communicate. [00:59:27] We know sometimes they – you know, we only see them sometimes when they're strand. [00:59:33] Some species we're like not even listening in the right frequencies. [00:59:37] And, you know, sometimes we're like looking at the data, and we're not seeing the right thing. [00:59:41] So I'm very hopeful in animal research as well because there's so much new data that's coming in and so much to learn, and I mean, in a sense, we're like only at the beginning in many of these species.

Speaker 4: [00:59:53] I have a question over here. [00:59:57] Hi. I have a question about sort of the features of the sperm whale language. [01:00:01] And so as you were talking, I was thinking about what are your thoughts on the presence of like, say, swear words or like filler words in our language? [01:00:10] Do you think this is something that you'll see pretty soon or even exist within our language?

Speaker 1: [01:00:17] I think it's really hard to say, because these will be the last ones to decode probably. [01:00:24] Because, again, the challenge is, and it's fascinating to work on this because it's so challenging, but the challenge is, how do I even represent their worlds? [01:00:35] They don't see really well, because the light in the water travels pretty poorly. [01:00:40] I think the only thing that travels well in water is sound, so they probably pack a lot of information into sound. [01:00:47] But we have, like when you study elephants or if you study parrots, you have air, you have the concept of air, you have the concept of drinking, you have the concept of a tree, none of that exists in whales, for example. [01:01:00] And it's really hard to know how they even navigate their environment, because these clicks are used also for echolocation sometimes. [01:01:10] They're very different acoustically when they echolocate. [01:01:13] But they're basically flashing these sound lights like a disco. [01:01:17] Once in a while, you get a flash of light and then you see things. [01:01:21] They're probably flashing these beams of sounds and getting their environment from the feed, from the echo. [01:01:29] So it's really difficult to even know where the prey is, let alone. [01:01:34] But maybe stuff like relationship, akin terminology, or elephants have names or dolphins have signatures and whistles. [01:01:45] This might be easier than, you know, whether they can lie or whether they can swear or not.

Speaker 5: [01:01:51] Thank you.

Speaker 6: [01:01:56] Lex Leifheit. [01:01:57] So you said earlier, when you're approaching something you don't understand, it's really good to explore what is possible. [01:02:05] And there's so much creativity in how you, even how you described your work. [01:02:11] I'm curious, you mentioned training on the text of James Joyce. [01:02:16] But if you had unfettered access to living artists, musicians, hip hop, culinary artists, anything, what might you do?

[01:02:25] Speaker 1: What might you do first to continue to advance your work and the possibilities of AI? If I was able to work with artists? You just had unfettered resource. Someone said, which artists, what would you like to do to advance your science?

[01:02:45] Speaker 2: I also wanna know where the name Pinchy came from.

[01:02:47] Speaker 1: Oh, that is a very complex, biologists own the domain of naming whales and it's complex and not easy and I have nothing to do with the name of Pinchy. So I think these models are bringing up, bringing new realities, right? And new exploration techniques. I think art is so powerful because, in a sense, it was dealing with the novelty and the new explorations for a long time. Just having that artist's mind is powerful but also in this project, it really, for example, I mean, we kind of did have a collaboration with artists already and it's a great group at Meta Haven, they're in Netherlands, and they're designer groups that are designers and artists.

[01:03:52] And it was a wonderful collaboration because we took these new spaces, are called latent spaces and we tried to explore what kind of philosophical realities they're bringing. One thing that came directly out of that research is I realized, through my work on the brain and artificial neural networks, that a lot of what we think of as symbols or this kind of symbolic complex operations we do are very physical. Even if you look at our brain, the sound that I played you, you can hear sounds in our brain, sounds that you hear in the form of electric activity.

[01:04:40] So even our brain, we think of it as this amazing abstract computation device, it's very physical all the way up to the final layers. And AI is showing that as well, right? The models that I can trace how sounds get perceived in an AI model and in the brain, and they were physical all the way up. So one of the things that is really powerful for artists was basically there are no symbols. Everything is, I mean, symbols are epiphenomenal. What does that bring? What kind of philosophical implications does that bring?

[01:05:29] That is something that is really also to your question, that's something that I learned through this work. I realized, wow, we're physical all the way up to the final layers of abstraction and computation, and that also has profound philosophical implications. Sometimes because these models are so multidimensional, you need that artistic interface to really explore that. If you go down to our website, you'll be able to see both visual and acoustic stimuli for exploring internal layers and how language gets created through those layers. It's really quite powerful to experience it. And that's, I think, where art is really important in the process.

[01:06:08] Speaker 2: Yeah, I have a couple questions. One, do you think in this collaboration with AI, we ask different research questions? And I don't mean that in terms of productivity or efficiency, it's just the result is we ask different questions and what are the implications of that, one. And two, is there anybody studying the changes that are happening to human language already as a result of that collaboration?

[01:06:43] Speaker 1: Well, I hope that we continue asking the same questions, but I also hope that we change our questions, right? Because there's a new possibility. There are some things that we were not able to do before. But I also hope that we don't just do those questions. I hope that we keep asking the same questions. I think with more inference, actually the models might themselves start helping us frame our questions. But your point is really important. I think science starts with asking the right question, right? And maybe that's gonna be the most important thing in the next couple of years in science, where answering that question might not be as difficult.

[01:07:40] Right now, it was really difficult to answer, both to pose the question and to answer it. But I think the focus on asking the right question is gonna be even bigger.

[01:07:48] Speaker 2: Yeah, there is interesting work, but I'm not working on that. But there's linguists who are studying how we speak to AI devices. Speech is another really interesting thing.

[01:08:04] Speaker 1: For the longest, I started studying computer speech for a while. It was like not getting anywhere, not getting anywhere, not getting.

[01:08:06] Speaker 1: Anywhere, not again, anywhere, not getting until very recently. And basically now we're almost there, right? It's basically a done deal. How we react to computer-generated speech and how our language might change is interesting. In other ways as well, right? Not just like the way we speak, but like what we say to models, how we try to convince models. There's other research coming from Michigan; apparently you can flip political opinions. So they try to flip political opinion of models. And when humans try to convince models into something, it doesn't work. But the other models convincing other models, it works better. So that's something to chew on as well. I think there's a lot of these sociological problems, and I'm not a sociologist at all, but you can simulate societies and interactions and then pair it up with human behavior; that's going to be exciting. But that's a little bit out of my expertise.

[01:09:06] Speaker 2: Do you say please to Chet Chippitee?

[01:09:08] Speaker 1: Absolutely, absolutely. Who here says please to Chet Chippitee? I definitely do too. Well, actually, I have this fear that if I'm not polite enough, it will start giving me bad answers. Roko's basilisk will come.

[01:09:27] Speaker 3: Jasper, a quick thought experiment. If you were to try to construct a dictionary of mapping signifiers to the things they actually refer to, just wild speculation, what observations or interactions might you test to see if we could actually map that?

[01:09:45] Speaker 1: Yeah, like what they say to each other when they hang out, what they say to each other before they go to sleep. We have some evidence that in some situational spaces, they use some of these vowels more than others. So I think you have to start with everything in science. You have to start with the simple thing, so the thing that you actually observe. So how far they go, where they go, how long they go, for who they talk to, when they go to sleep, what they do before they go to sleep, what they do before they hunt, and stuff like that. And then you build on higher complexity, maybe names, individuals, how they address each other. So I would say start simple first and then see how far it can get.

[01:10:41] Speaker 4: Oh, hi, so I'm a student at UC Berkeley and—

[01:10:46] Speaker 2: Go Bears.

[01:10:46] Speaker 4: Yeah, so, and I used to work in the NeuroBat Lab, and so thank you for answering that. My question was about echolocation, but I guess my other question after that was, do you think that the way that you're collecting data about these sperm whales is impacting the data that you're collecting? Especially since I know that there's been studies about how animals will change communication behaviors and just change behaviors in general when they're around fishing boats. And so I'm curious about that and how you think your data and your interpretation might change?

[01:11:26] Speaker 1: Yeah, that's another way where AI can help in very pragmatic ways. So it was the case that we didn't even have good data on whales. And now we have drones that gently let the tag on a whale. We have CETI with teams that have studies that predict where whales will surface. The drone flies there and gently places the tag on the whales. We have underwater gliders or mini submarines that follow the whales with the precise reason of having them record minimally invasive, with a minimal invasion into their lives. But also, whales are pretty good at ignoring things. Like they can, it's not easy to trick them. So I think from that perspective, it's good, but like I was surprised at just how little we had data before now. And when you're making new discoveries and new claims, you better have the best kind of data possible. The only way I would be willing to claim anything about the whales is if I had, and I did, the data from the whale itself. Because for those of you who do ocean acoustics, it's horribly complex and can mess up the signal so badly. But if the mic is on the whale, then you have the cleanest data possible. And you know, some of the advances can also be in collecting data, like drones paired with AI that predict where animals would surface and have minimum invasive techniques for animals to collect data on animals in the wild that don't get disturbed. I think we're going to see a lot of that data coming in, and you know, we'll interpret it and probably find really cool new things.

[01:13:28] Speaker 5: Question, do you use AI for, for example, operating not in traditional systems, for example, instead of traditional numbers, use pedic analysis or even more non-traditional correlations in your results?

[01:13:46] Speaker 1: In your results. For example, remember the movie Arrival when we, like, generated this artificial language and they use Wolfram to generate this language? Do you use something like that with the help of AI? Thank you.

[01:14:02] Speaker 2: Yeah, I think the closest to what you're describing is, so the models that we've been developing in the lab don't really have tokenization. What you get is raw data, and then through layers, the model itself creates representations, and that's incredibly useful. One thing I've learned, you know, by working with these models, even the small ones, even the smaller custom-built models, is that they're incredibly smart clusters. You know, the world of statistics has tried to do clustering really well for the longest, and some clustering techniques are great, but these models are just amazingly smart clusters.

[01:14:49] So you can then follow the pathway of how they represent data through the layers by looking inside them, and then at the end of the day, you have this monosemanticity, where the model packs what is informative into these single neurons and kind of forgets or disregards stuff that is not really important. And that is huge for scientific discovery, right? Because it helps you shrink hypothesis space. It kind of beeps when the ring is closed and helps you basically... When we were doing this, there was a period where the model said, "Okay, spectrum is important, spectrum is meaningful," but we didn't know what to do with it. But if we didn't have that information, we probably would not have dug so long and tried so hard to find the patterns.

[01:15:43] At the end of the day, when we found the patterns, it was easy. In hindsight, everything is easy, right? But it took a lot of time, still a lot of time with data. Another piece of advice that I have for scientists is don't forget to spend time with data. Still, in this age of AI, spend time with data. Because humans are also very smart clusters. The only difference is that we are biological clusters, and they are silicon clusters. I think once you combine those two principles, biological and silicon clusters, we can get farther and make advances.

[01:16:16] Speaker 1: We're going to take one from the virtual audience now. This is from Andrej Holtz. If we let AI agents evolve and communicate freely, what would their language look like? Would they develop a hyper-efficient system completely unlike human languages?

[01:16:31] Speaker 2: That's why I'm so excited to be a linguist these days because I think that will happen, and people will need our tools to describe those languages. Linguistics has been a very interesting journey. We came from a science that ten years ago was not very popular, not very prominent, and now we're finally at the center. I think it's a similar problem to whales in some ways; they have some communication system. Now we have the tools to approach it. Agents will likely develop their own protocols, and we'll have an exercise on how to approach them and what kind of steps we can take. Fascinating.

[01:17:16] Speaker 1: On your side, Erica.

[01:17:19] Speaker 3: Hi, I'm Cameron. I'm a data scientist, and I'm also thinking about the quality of our data. You talk about two things, like AI is very useful for scientific discovery. You talked first about hypothesis generation, how these models are creating original yet verifiable hypotheses that you can dive deep into, and you have this real-world data, these audio files that you can see. Is this meaningful or is this random noise?

[01:17:46] But a second thing you touched on was synthetic data creation. We're kind of hit or miss with real-world data, but I feel synthetic data could be helpful for niche topics or rare events. So, two related questions. You can answer one or both. One is how are we as scientists thinking about evaluating the usefulness or truthfulness of these synthetic data sets? And secondly, how would this change the scientific discovery process, especially? How do you verify hypotheses trained on this mix of synthetic and real-world data, or just synthetic data, especially if there's a risk that the synthetic data might skew from real-world applicability?

[01:18:32] Speaker 2: I think what I meant by synthetic data, of course, you can just simulate and try to simulate within distribution so that you can do experiments that you would not be able to do on real data. In this case, I meant more like synthetic data that is way outside of distribution, that you're not using to simulate the actual data, but you're using to challenge your inside-the-box thinking in some ways.

[01:19:01] So it's not as much recreating data as it is letting the system take a stab at creativity, that silicon creativity that we might not have as biological creatures, and having that kind of out-of-distribution data challenge us to maybe see things that we otherwise wouldn't have. I meant more in that respect.

[01:19:27] Speaker 1: And for that, you actually don't want the data to look like real data, you want the data to look different. You want a language that has colors and gestures and whatnot, you know, the light beams, not words and phonemes and so on. Hi, Michel Bilecki, thank you. So I think arguably human communication is pretty inefficient. You can say the same sentence twice, depending on your cultural background, emotional state, it can mean a hundred different things. You said that agents are already inventing their own kind of communication protocols. Do you see a future where we might be interacting through agents to bolster or augment the efficiency of our own interhuman communication?

[01:20:13] Speaker 2: Well, actually, I wouldn't be so bold as to say that language is completely inefficient, because yes, we can say a sentence twice, depending on where you're from, but that's precisely the feature, not a bug, because you're expressing your social identity through language. We're building ourselves through language. So I think our language evolved to be efficient for our purposes. Now, what I think is true is that agents' purposes will be different from ours. So that's why our language is inefficient for agentic AI pipelines, whereas for us, it's good. I think I want to have diversity of speakers to get your social information, and so do whales, by the way.

[01:21:05] There’s a really interesting study by our colleague at SETI, Shane Gera and others. They showed that whales have dialects, but not only that. In areas of the ocean where they bump into each other, where groups of dialects, dialectal groups bump into each other more, they amplify, they increase their dialect features, which is very similar to this wonderful study that was done in the 60s by Bill Abov in Martha's Vineyard where you have islanders, Martha's Vineyard speakers speak a dialect, they are islanders. When they speak with each other they speak a dialect, but when they speak to an outsider they even speak more in that dialect. They increase their dialectal features because they want to show that they are from the island and increase their identity. And whales do that in a similar way.

[01:21:58] Speaker 1: And maybe agents won't want, or maybe they will, to express their social identity, but certainly language is beautiful precisely because it can also not only transmit meaning but a lot of social information between us. So I want to ask a question. So in chain of thought the model would say first I do this, then I do that, then I do that. And then when reasoning emerged, the model learned to speak a different language. I want to ask you if you would call that the language as a linguist. But basically if you see reasoning models, the way they reflect, they backtrack, they say let me try this, oh no, I'm wrong. So they speak in a very different way than the pre-reasoning models.

[01:22:44] Speaker 2: And as far as I understand, they speak to themselves, that helps them solve the math problems. And as far as I understand, this was discovered by reinforcement learning. They just, I mean, they had the vocabulary to reflect, but this behavior was amplified because they got the reward. I don't know if you call this a language that the model discovered or not, but that is the emergence of reasoning is perhaps fitting in your description already.

[01:23:13] Speaker 1: Yeah, and I love chain of thought, and there's so much research that needs to be done there. Also, one of the biggest debates, most consequential debates in the legal world that has implications for the legal world and philosophy is the relationship between language and thought. And in animals especially, right? For the longest, the way that the legal world worked is, first it was language, animals cannot feel pain, let's not admit them to our legal system. Then it was animals can't have complex thoughts, and then we've shown that they can have pretty complex internal thoughts. And now it's like they cannot have language.

[01:24:10] Speaker 2: But the notion of what language is relative to thought is also changing, so it was considered that language is complex thought. So Chomsky's model would be those are intrinsically related. And now we're seeing that language is maybe just this final externalization layer that communicates my complex thought to your complex thought. And from that, I approach animals from that perspective. We just need that window. Maybe it's their language. Maybe it's something that we come up with, an AI model or a language when we train a parrot. But chain of thought is so fascinating, and Professor Dimakis is also at Berkeley, so maybe we should pair up and do some work on chain of thought.

[01:25:03] Speaker 1: But it can answer some of the big questions because sometimes you can also have a model where chain of thought is gibberish, and you can post-train a model for the chain of thought to be explainable, but the performance doesn't improve dramatically. So that is very powerful and an interesting kind of simulation for this idea that maybe you can have complex thought without language and the language is just this final layer that communicates our internal worlds.