← New Books Network
New Books Network
Yosef Grodzinsky, "How Deeply Human Is Language?: Chomsky, the Brain, and the AI Fantasy" (MIT Press, 2026)
00:48:21
Loading summary
Transcript58 lines
- [00:00]
Home Depot Advertiser
Make every get together chill this Memorial Day. Get up to an extra thousand dollars off select top brand appliances like lg. Plus get free delivery at the Home Depot. Tackle pool towels and camp laundry with a large capacity washer and host in style with the fridge serving craft ice, mini craft ice, cubed ice and crushed ice. Shop Appliance Savings now through June 3rd at the Home Depot offer valid May 14th through June 3rd US only. Free delivery on appliance purchases of $998 or more. See Store Online for details.
- [00:31]
Mariam Polukoti
So good, so good so good.
- [00:33]
Nordstrom Rack Advertiser
Everything you want for summer is at Nordstrom Rack stores now and up to 60% off. Stock up and save on the brands you love like Vince Sam, edelman frame and free people. Join the NordicLub to unlock exclusive discounts. Shop new arrivals first and more. Plus buy online and pick up at your favorite Rack store for free. Great brands, great prices. That's why you Rack,
- [00:58]
Microsoft Advertiser
study and play come together on a Windows 11 PC and for a limited time, college students get the best of both worlds. Get the unreal college deal everything you need to study and play with select Windows 11 PCs. Eligible students get a year of Microsoft 365 Premium and a year of Xbox Game Pass ultimate with a custom color Xbox wireless controller. Learn more@windows.com studentoffer law supplies last ends June 30 terms@ aka mscollegepc welcome to
- [01:29]
Professor Yousef Grudzinski
the New Books Network.
- [01:33]
Mariam Polukoti
Hello everyone, I'm Mariam Polukoti. I am with me here today Professor Yousef Grudzinski, who is the author of How Deeply Human Is Language. Professor Youssef is a psycho and neurologist who has long been working in in the interface between theoretical linguistics and clinical and cognitive neuroscience. He has studied the neural basis of language and has also been concerned with the diagnosis of language pathologies, their treatment, rehabilitation and he is the recipient of various prizes including Humbug Senior Nachel Award and a Senior Canada Research Chair in Neuro Linguistics, which he held when he was at McGill University, Montreal. Currently, Professor Youssef is a Professor Emeritus and Director of the Neuro Linguistic Lab at Edmund Lilly Safra center of Brain Science at the IBRE University of General Salem. He is also a Scientific Associate at the Institute for Brain Research, University Hospital, Dusselfdorf, Germany. And finally, he is a Senior Overseas Visiting Scholar at the center for Language and well Being. Shanghai Jiotong University is joining us from Shanghai, China this morning. You are welcome to the NBN professor yourself.
- [02:56]
Professor Yousef Grudzinski
Thank you, Mariam. It's a pleasure to be here.
- [02:59]
Mariam Polukoti
All right, I would like to ask you what dictates your choice of title for Bibu.
- [03:06]
Professor Yousef Grudzinski
Quite an interesting question, but let me start off with a confession of sorts. Fantasy or not? Like many millions, of course I use ChatGPT and Claude, of course I use AI. We'll get back to why and how. But I do think that AI generally and large language models, or LLMs in particular, are great in many, many ways. And now to your question how I chose this title to remind you the title is How Deeply Human Is Language? Colon Chomsky, the the Brain and the AI Fantasy. Well, the big issues the book discusses regard the nature of two human creations. One is a theory of what we humans know as speakers of our mother tongue, and the other is technologies to build machines that understand sentences and text and generate them when interacting with humans. And who knows, maybe with machines, with other machines too. So the book is about human language, quite obviously, and the question is whether it is exclusively human or can it be owned by machines as well. In the book I try to answer this question from a multidisciplinary perspective. I have a bit of multidisciplinary training myself, and so in writing this book, I tried to use it as best I could. But why Deeply? I called it deeply human. There's a scientific issue how can we build a theory of us? And an engineering problem how to design a useful machine that tries to emulate us. Both questions have preoccupied some of the best minds and brightest minds around. And in both domains, the word deep has played a central promotional role over the years. First of all, is language may be the hallmark of our humanity. Studying language gives us a chance to get a deep look into ourselves. But there are much more concrete reasons for this deeply that I use. So Noam Chomsky, the founder and forefather of modern linguistics, famously called a particular abstract level of linguistic analysis, which I explained in the book. He called it deep structure in machine learning, which changed its character about 15 years ago. They started calling a certain aspect of their models deep learning for reasons that I also explain in the book. So this is where the deeply part comes from. It's a gesture to both worlds that are explored in the book. Now, the book subtitle Chomsky the Brain and the AI Fantasy tries to allude to three issues on which the book focuses. Chomsky's generative revolution, neurological issues that concern language, otherwise known as neuro linguistics and the current AI revolution. To me, it's just natural. I started out in medical school, studied linguistics in parallel, and ended up working in neuro linguistics. So I tested most relevant clinical neuropsychology, experimental neuroscience, theoretical linguistics and later I worked for many years not as a physician, but as a researcher in university departments of neurology and neurosurgery as well as psychology, neuroscience and linguistics departments. So all three topics are explained from scratch in the book with many illustrations and examples. And yes, I use the word fantasy both as a teaser and because I believe that at president modeling the speaking brain with current AI is a kind of fantasy. I hope that readers, whether professional or just curious bystanders, will read and find out for themselves.
- [07:01]
Mariam Polukoti
Oh, that is quite impressive, at least from your response. You've not only answered my question telling me about what triggers the title of the book was it's quite an exposition about who you are, what you do, how you get here and what regards your interests. Exploring the intersect between linguistics and generative AI LLM specifically. Thank you for that. Brilliant response.
- [07:33]
Professor Yousef Grudzinski
Thank you.
- [07:34]
Mariam Polukoti
So the next question. In the prologue of the book, you tell some story that are bidirectional. Could you give the audience a glimpse of those stories?
- [07:45]
Professor Yousef Grudzinski
This is a great question. I'm really pleased with that. So let's begin with some basics. So what's a scientific story? That's the first question one might want to ask. It's a story about the structure of some aspect of the world that's told by a curious and inquisitive mind of a scientist or scientific community. So with any luck, if you tell a story about some aspect of the world, with any luck, the story can later be formalized by mathematical grammatical tools. So it can be state. I mean, you know, it helps to derive predictions that can be tested, whose veracity would lend empirical support to the theory and would add factual layers to the theory or the story or perhaps invalidate your story. That's also possible. And in the book I tell the story of how grammar, grammars and grammatical systems have developed over the centuries since the. You know, the first. You have to realize the first known grammar was written by an Indian Brahmin named Panini, who noticed regularities in Sanskrit some 2300 years ago, believe it or not. And he set out to formulate rules that account for these of quantitative approaches to language. Started some 120 years ago in Russia by a mathematician named Andrei Markov, who. Who developed a mathematical approach to text in order really, believe it or not, to debunk the notion free will as a way to argue against religion. God knows what that means. I tried to understand it and was unsuccessful. But his mathematical development was amazing. So his approach then developed into the computational world of computational linguistics that we all see. So there's a grammatical rule story, a quantitative story, and both are theoretically formulated and both have the same that they have common goals, which is to understand the nature of language through the use of theoretically respect, namely formal tools whose job is to explain language. And this is where agreement ends and the debate begins. So linguists see rule based grammatical well formedness of strings of words as a special human capacity, by virtue of which we were confident the sentences like who do you think Mary loves? Is a fine English question, but who do you think that Mary loves? Is not is odd. And also there are many, many examples like this, and I give many, many in the book. Computational linguists don't think that. And what's important for linguists here is that grammar is and its rules sit in the speaker's head and they do right from wrong. They're the ones to dictate to you how you should say something and what you cannot say, because grammatical rules bar you from saying that. Computational linguists focus on quantitative properties of words. Sentence how many times a word appeared in the context and who its neighbors were when it appeared. It becomes complex very quickly as intricate relations between individual words and contexts are uncovered by computational methods. Here the idea is, just to take an example, that the distinction of a word like boy from the word joy is recoverable solely by the fact that boy is more frequent than joy and that it appears in vastly different contexts. So these are two stories. The stories I tell, I should add, are also stories of individuals who are interspersed throughout the book, of courageous patients like Congresswoman Gabby Giffords, who was shot in the head in an assassination attempt some 15 years ago and survived. Of stubborn and tireless scientists like Frank Rosenblatt, who strove to build the first AI machine, and many, many others. So the scientific is intermixed with the personal, creating what I hope is a fabric that readers might find interesting. But going back to the theoretical versus computational linguistics, there are different perspectives, but they may also be complementary. Sadly, the points of convergence between the two worlds, the linguistic and the computational, over the past 30 years or so have been few and far between. I'll say something about it later. And it's told in the book, of course. But you have to remember complex quantitative relations between words. That's the computational side versus tight formal rules that tell you what you can and cannot say. This is the contrast. These are the two stories. Now, why has there been this sort of parallel developments? Why don't theoretical linguists and computational linguists talk to each other? Well, these are interesting sociological questions that may follow from cultural differences between science and engineering and the obvious relevant of engineering to industry and money making, as opposed to sadly abstract theoretical ideas with little cash value in the most concrete sense of the expression. I touch on these in the book. Of course, a good example of the difference can be given from a somewhat surprising direction, namely my personal experiences in neurosurgical operating room. But let's save it for later. We'll talk about it later. That's it.
- [13:14]
Mariam Polukoti
Oh my God. That was mind blowing. As a linguist, I really enjoyed this session because it's not just you telling the audience about your book, but you actually teaching me more about my discipline. Thank you very much.
- [13:31]
Professor Yousef Grudzinski
I'm very pleased and as I told you, I'm very pleased that I'm being interviewed by linguists. It's a pleasure.
- [13:38]
Mariam Polukoti
All right. How significant are these. Okay. The stories you have just told us, how significant are they to the message communicated in the book?
- [13:50]
Professor Yousef Grudzinski
So let me tell you something about storytelling. I was told that stories must embody without a conflict, there's no story. It's just uninteresting. And this book is a story, but what distinguishes it from fiction is that it's all grounded in facts. You don't need to travel very far to see a conflict. From what I've already said, a debate emerges between the rule based community and the quantitative community. And there's even been some mudslinging and name calling between celebs. Sadly, I would say so. Chomsky, the father of modern linguistics, who currently is pushing 98 and unfortunately can no longer respond, having suffered a stroke three years ago, he flatly dismissed AI, calling it a false promise. But even worse, the physics Nobel laureate Geoff Hinton, considered the father of current AI, has repeatedly called Chomsky crazy while distorting some of his claims. This is not helpful and it's most unfortunate and detrimental to the field, but it makes a basis for a good story. Now, I gotta say, I wrote the book not because it's a good story, but because I thought the state of affairs must change. It's detrimental and let me lower expectations. You get stories, but zero gossip. You don't get gossip. So it's difficult to wrap up a book in a short interview. But I can say that I want to tell the reader that a seemingly soft science, the art and science of words, verbal communication, is actually a very hard one that is best treated by rigorous scientific theories. That's the point I begin by telling how old age old observations were converted into grammars by thinkers and philosophers in India, Greece and later other parts of the world. I then move to current day linguistics and tell its story, namely unravel the intricacies and mysteries of modern theory to the uninitiated. I explain what's special about these theoretical proposals, why they're scientific, at what the leading ideas are, ideas and tools of current linguistics. Next I move to give the same treatment to the story of neural networks. I recount their story from both a computational and neurophysiological perspective and then move on to explain in plain language some of the leading ideas of current AI, including those of Geoff Hinton, like the famous backward propagation algorithm. Well, it's not exactly his idea, but he and his team were the first to make the best use of it for sure. Throughout, I make minimal use of technical terms and while I do assume readers attentive, reading it requires zero prior knowledge. This is important. Zero, really. It's all in the box, all explained. And now we come to the real beef. I scrutinize each of the stories and confront it with the sometimes grim reality of the reading world. I subject each framework to a conceptual and empirical critique and I try to do so with as much color, patient stories and experimental results as well as games. I play with ChatGPT and give the reader some ideas of how to continue playing as much as I can. To be more concrete, I take a peek at the speech and communication abilities of brain damaged individuals and at brain activation patterns during language comprehension in order to reflect on what story would fit the pattern would best fit the pattern they display. We'll talk about that later, but now I just want to note that the outcome of this discussion is very simple. It's really not about who's in the right and who's in the wrong. If the two camps, the computational and the rule based if these two camps work together, the world would be much, much better off. That's it.
- [17:56]
Red Bull Advertiser
Ready to soundtrack your summer with Red Bull Summer All Day Play. You choose a playlist that fits your summer vibe the best. Are you a festival fanatic, a deep end dj, a road dog, or a trail mixer? Just add a song to your chosen playlist and put your summer on track. Red Bull Summer All Day Play Red Bull gives you wings. Visit RedBull.com BrightSummerAhead to learn more. See you this summer.
- [18:21]
Canva Advertiser
When you finally find your thing, you want the whole world to know about that thing. So you use a thing called CANVA to make it an even bigger and better thing. Whether you want to create flyers for that thing, make presentations for that thing, or design merch for that thing, you can do anything so people can see your thing, feel your thing, love your thing. The next thing you know, it's a thing. Canva the thing that makes anything a thing.
- [18:53]
American Express Advertiser
Experience a membership that backs your business journey with American Express Business Platinum. When you pay with membership rewards points for all or part of an eligible flight booked with a qualifying airline through Amex Travel, you can get 35% of those points back up to 1 million points back per calendar year. American Express Business Platinum there's nothing like it. Terms apply. Learn more@americanexpress.com Business Platinum
- [19:22]
Mariam Polukoti
well, thank you very much. That is expository. I can tell from my interview with you and what I've read in the book that both together has exposed me to the fact that the book is not written for written sake. It is actually a confluence of so many things. You've taught us research approaches, you've taught us methodologies, how these and this could be added together, cooking so many things, put in different ingredients. It is beautiful. I motivate.
- [19:57]
Professor Yousef Grudzinski
Thank you. Thank you.
- [19:59]
Mariam Polukoti
All right, you draw a line between technology and science in the book. Could you shed light on the differences? Do they have a meeting point, if I may ask?
- [20:10]
Professor Yousef Grudzinski
Yes, of course I can. This is perhaps the single most important distinction I make in the book. So let's be clear. Science is about understanding the world, and that includes understanding ourselves, our physical and mental makeup, and everything else that deserves an explanation. Engineering, and that's an important contrast, is about building useful tools. Now, science and technology and engineering may go hand in hand, but not necessarily. We're fortunate to have an excellent historical perspective on the relationship between science and technology from the fascinating writings of Joel Mokir, an economic historian from Chicago who is the 2025 bank of Sweden Memorial Nobel Memorial Prize Laureate. This is really a mouthful. It's difficult to say, but he's basically an economics Nobel Prize winner last year. What he does he focuses on science and Technology in the 18th and 19th century England during the Industrial Revolutions, and he shows how as late as the 1820s, engineers had little if any interest in science. And this he learns by looking at engineering textbooks of the time. They were about doing, not understanding the nature of the world, about producing better steel, about finding new medicines, but not about understanding the principles behind these developments. And while today engineering is much more complex and demanding than it was, to an extent, even today some scientists make a similar point you know, I'm currently in China, I told you that, for a relatively long period of time, and I'm exposed to Chinese culture. So Confucius, one of the most important Chinese philosopher, who known here as Kong Tzu, said in one of his analects, he said, people may be made to follow a path of action, but they may not be made to understand it. I think that's a rather gloomy perspective on the humankind. But I think that some people are drawn to a path of action, like building useful machines, not caring much about understanding, whereas others choose to reflect on why things and language in particular, why things are the way they are. So Confucius distinction between acting or doing and understanding is quite valid. And let me give you an important example for doing without knowing. So computational linguistics is a form of engineering which nowadays tends to have no linguistics at all, pure engineering. If you peek into the curriculum of a course on this topic, like NLP courses, Natural Language Processing, you often see a quick allusion to Chomsky's historical role. But typically it's immediately followed by a quote from Fred Yellinek, a well known computer engineer who sometimes in the 1980s quipped that, and this is a quote, every time I fire a linguist, the performance of my system goes up. That's a funny saying, perhaps in a somewhat vicious way, but it's arguably Confucian acknowledging the hazards of knowing if you know too much, you slow down the system. So you might say with success stories such as ChatGPT, who cares about understanding if it only leads to slowing down our machines? But Joel Moker, the economic historian, also shows that once scientific principles are introduced, the process of discovery and creation of new things is improved and expedited. Take the story of the most famous painkiller, aspirin. This analgesic was used by medicine for centuries in different forms, mostly through preparations made of the bark of the willow tree. While its properties were not known, let alone understood, they didn't care, they just wanted to kill pain. But after humanity had enough chemistry in its belly, it could identify acetylsalicylic acid, which was later synthesized. Going to mass production was called aspirin and has had an extremely important impact on the well being of millions. And this story mocker shows us repeats itself time and again in medicine and in engineering. Use without knowledge or understanding may be good, but scientific understanding provides a leap in both quality and quantity. Now will this happen in language processing? Will linguists and AI engineers find a meeting point? Well, read and judge for yourself and we'll See?
- [24:53]
Mariam Polukoti
All right, all right. I must say debug is a must for everyone, especially linguists from those parts of the world, everywhere in the world. Because you know, as linguists who claim that we are in the humanities alone, don't let me say only linguists. People in the humanities really need to get this text because people in the humanities think AI is not human. In the humanities we do human, not gen AI and there's no way gen AI can meet up with human. But this book is actually telling us that not only preaching that we should embrace gen AI, it's also saying that we should try to bridge the gap between the engineering discipline and the science discipline which we represent as theoretical linguists cognitively.
- [25:46]
Professor Yousef Grudzinski
Absolutely.
- [25:47]
Mariam Polukoti
Computational linguists too. They are out there, but how many are they? Especially in this part of the world. So I think we are still. It is still linguistics. LLM is still linguistics. It still has a very light percentage of linguistics in it. And that is why we really need to work together with the engineering people. All right, the next question in the book you evaluated the language learning models from the neuro linguistic perspective. Why do you do this?
- [26:18]
Professor Yousef Grudzinski
Oh, this is really, this is an important question. So when you think about it, the theory of human ability, linguistic ability is actually a theory of the human brain's language ability. So take vision, let's consider vision. I think everyone would agree that if you want to study vision, whether in humans, monkeys, cats or rats, we should peek into their brains and maybe poke their brains to study vision. Research in fact is likely the most advanced area. Experimental and computational. The fairly long history of success. Some failures too. You never succeed without failing sometimes. But language is not always perceived as part of neuroscience. This is quite shocking. One obvious reason is that study animals and any ethical constraints barred from poking humans heads at will. But there's a critical reason is, and I regret to say to be an amateur sociologist once more is as we talked language, linguistics, grammar, they seem to belong only to the humanities for some reason. Something soft, poetic, emotional, soft stuff. That's true. We're emotional creatures with poetic abilities. But there's a critical aspect of our linguistic ability that's been shown time and again to be amenable to rigorous study with precise results. And that's the grammatical domain. And so the question we ask is this, whence does our grammar emanate if not our brains? And this is a general consideration which becomes clear once you compare neuro linguistics to other domains of cognitive neuroscience. But there are major empirical clues as well. The first couple of decades of my career were dedicated to the study of aphasia, a variety of language disorders resulting from some kind of brain damage. So for my doctoral dissertation, I worked with stroke victims in a pioneering aphasia ward in Boston at the Veterans Administration Hospital in Boston. By the way, my two doctor factors my two supervisors were neuropsychologist Edgar Zurif and linguist Noam Chomsky. What a combination. That was an unbelievable combination. I saw and tested patients with a variety of grammatical disorders that quite clearly show that the brain carves out language abilities along grammatical lines very clearly. And I and many colleagues have been amassing data of this kind for many years now. If you ignore this database, it means you live in fantasy. One recent case, which I discussed in some detail, is Gabby Gifford Story of Courage. Giffords was a young and promising Arizona congresswoman who survived an assassination attempt some 15 years ago. But much to her poor luck, the bullet she received passed through her skull and destroyed parts of her language area and she became grammatically impaired. Her story, which goes hand in hand with the story of her husband, ex space shuttle astronaut and current Arizona senator Mark Kelly, is really a story of resilience, dedication and courage. But it's also a story of how language is represented in the brain. To explain her pattern of language impairments, you need grammatical tools. No AI creation. I know and I believe I know quite a lot can account for these patterns, but grammatical theories can. Maybe we can listen for a few seconds to her speech from a documentary about her. She's recorded practicing her resignation from Congress speech, which she would give to Arizona constituents about her stepping down. You will see she called them people Arizona instead of the people of Arizona. And about a year after she was hit. So listen up.
- [30:23]
Mariam Polukoti
Action. Gotta step down people Arizona. Step down people Arizona right now. Work to do, work to do, work to do. Worked.
- [30:53]
State Farm Advertiser
This episode is brought to you by State Farm. You know those friends who support your preference for podcasts over music on road trips? That's the energy State Farm brings to insurance. With over 19,000 local agents, they help you find the coverage that fits your needs so you can spend less time worrying about insurance and more time enjoying the ride. Download the State Farm app or go online@statefarm.com like a good neighbor, State Farm
- [31:17]
Expedia and Visit Scotland Advertiser
is there Expedia and visit Scotland. Invite you to come experience the beauty that awaits in Scotland. The sweep of wild coastlines, quiet lochs and untamed landscapes. Fresh cuisine that feels rooted in the land. Come experience. Experience the kind of stillness that stays with you long after you leave Plan your Scottish escape today@expedia.com VisitScotland
- [31:49]
Firestone Tires Advertiser
Real value shows up in reliability. You don't have to second guess. Like a set of Firestone all season tires. They're designed to deliver confidence inspiring wet weather, traction and a quieter ride. No matter the road. Season after season, Firestone all season tires for durability you can count on, just like people count on you. Firestone always dependable since 1900.
- [32:18]
Professor Yousef Grudzinski
So you can hear the frustration, but you can also hear the determination. And you can hear that there's clearly a grammatical issue here. The sentences are bad. She's a bit non fluent. She has the words, she definitely has the intelligence and the memory. But most importantly is her sentences are deviant in grammatically systematic ways, which provides important hints about how the brain breaks down. It does so in a grammatically systematic manner. So taken together with much neuroimaging evidence of various kinds, we have a very important test of the biological validity of a theory of language. The brain distinguishes between the AI and the linguistic approaches because we see that the talking brain is carved out by grammatical principles. Now, does this diminish the force of large language models as good engineering feats? Hell no. Of course not. Engineers are not committed to the biology, but to making good working products useful tools. But even from an engineering perspective, looking at the biology may actually be useful. The human brain speaks and understands language, but it requires much less energy than large language models. Unlike AI machines, we don't walk around connected to some nuclear power station. And these guys do. Okay, so perhaps engineers should look at the way the brain is structured. There's a guideline for for building more efficient machines. It actually happened in robotics. Robotics people started looking at how animals walk in order to understand how robots move. So there's no reason why it shouldn't happen in language. Here's another hint from the brain. For the past seven years, I've been working in operating rooms in Germany, helping out in brain operations, where my talented neurosurgeon colleagues remove or ablate malignant tumors that are called gliomas from patients brains. Now who needs me there? You might ask the linguist. What do I know? Well, in these procedures, patients are awakened. At one point, you may have seen it on medical TV series, so that the surgeon can test for the involvement of the tissue surrounding the tumor in different functions. You want to remove as much tissue as you can, but you also want to steer clear from highly functional areas to save their functionality. You don't want your patient to be mute or not be able able to Speak. So you test the patient intraoperatively by gently blocking the functionality of a brain area for about five seconds and at the same time administering various tests. But how do you decide what tests to give? You need a perspective on language from which you design these tests. So it's important for the preservation of the patient's communicative abilities. And so it turns out that computational methods have really little to offer in terms of test designs. But we have, and the tests that we provide are actually quite valuable. So this success, limited as it is, is part of the linguistic story. You have to view language as a human ability, formulate a theory that articulates your view most generally and precisely, derive predictions that help you design tests and use them sometimes in ways that help someone. That's a reasonably good story. Clearly, computational approach has its good own success stories. We see them in the media constantly, but not here. That's it.
- [35:55]
Mariam Polukoti
Awesome. All right, so I really want to applaud this because your consideration of the LLMs from the neural linguistic perspective is actually to draw the lines between what is engineering and what is science, and how could these two be married to bring out a very good product. So it's a kind of bridging the gap between, I mean, the gap between the linguist and the engineers. This leads us to then the next question, which is so very similar to the previous question which you've just answered. Based on the discussion of the book, then, from the previous position you gave us, what are the prospects of the marriage between linguistics and engineering? Will the application of linguistic theories help LLM's design? Please shed more light on this. Thank you.
- [36:49]
Professor Yousef Grudzinski
Well, you're absolutely right. It's a key question and I really thank you for asking it. So to my mind, such a merit is a must for everyone. We all agree that our machines and theories must enable the creation or novel of novel, well formed behaviors. It may be navigation, it might be vision, it might be language, whatever. So the debate between proponents of AI and linguists is not about what needs to be explained. The debate is actually about what's the right approach to language. Maybe none of them. But between these two possibilities, so which is the best, the linguistic one or the engineering one? And my answer is no. We must work together. If we work together, there's a chance. So here are some problems. Linguistic theory is not easily implementable. You can't really turn it into a working program. So the CTO of Waymo Autonomous Vehicles, a big autonomous vehicle company in the United States, the most sophisticated and thoughtful computer scientist who I Happen to know once told me we tried this, we tried to implement grammars in machine. It simply didn't work. And it's true it didn't work. But that was long ago. I believe him. I actually know that he tried, but that was long ago. If grammatical and AI considerations go hand in hand, it likely will be a success. I really think so. In fact, AI critics like Gary Marcus repeatedly claim that large language models already contain concealed grammatical elements without which they would not function properly. So whether or not he's right, it's clear that grammatical constraints would make much more efficient machines. Here's an illustration. Okay. The backbone of an LLM computation lies in a process called matrix multiplication. Forget about details, just every task of LLM performs. Every task that an LLM performs depends on trillions or maybe more such multiplications. And these are power hungry electricity. So trimming them would save much energy. Much energy perhaps. Maybe we would need fewer chips. The application of grammatical constraints into these processes would likely result in such trimming, but I'm unaware of any attempt to do so. So that's a place for linguists to have a major impact on machines that learn. And of course, as I discussed already, we linguists and no one else know how to test these machines. And Geoff Hinton says that large language models are better than anything ever created in linguistics. He actually said it, believe it or not, half an hour after he was told he won the Nobel Prize. That was the most important thing for him to say. But ironically, he's really not in a position to opine about this because his examples don't indicate that he has any knowledge of the basic linguistic facts. He never studied linguistics. You know, I don't look for his certificate as a linguist, but his examples show that he never really tested. I talk about this in the book in great detail and show how when you test it seriously, these machines break down.
- [40:07]
Mariam Polukoti
That's close. Okay. Thus, the next question I'm about to ask is also related to previous ones. And this has to do with you commenting on the differences and the similarities between the LLM's behavior and the human language behavior. I'm asking this question because in the book you mentioned that you tried to compare. You prompted some AI to do something and you are able to get some results. So can you share this with us by briefly commenting on the differences and similarities between the behavior of language learning models and the human language behavior? Thank you.
- [40:48]
Professor Yousef Grudzinski
Yeah, I'll be very brief because I think we're running out of time. And I'll just say that on the behavioral level, the current machines appear to behave seamlessly. Really, I actually use them all the time. For instance, when I plan my outings in China, I don't read Chinese, and it's quite difficult. I Repeatedly consult with ChatGPT or with Claude, and there have been repeated claims, for instance by Geoff Hinton, that these machines pass the Turing test, the concept which I'm sure that many readers, listeners, have heard about. So the idea is this. A person is conversing with an unknown interlocutor, which could be a human or a machine. An English mathematician, Alan Turing, proposed. Well, actually, he really just considered that if a person cannot tell whether she's talking to a human or a machine, it's an indication that the machine successfully emulated human behavior in total completely. But let's try to understand what this tests. And for those of you who have read Turing's original article from 1950, it's a very vague proposal that he makes. What does this test actually test? It means that an average language speaker fails to distinguish between a human and this machine. But with all due respect, the average speaker is not a linguist. A linguist and no other professional knows how to test the abilities of these machines. So one important point I make in the book is that it takes a professional eye to really get to really test the machines thoroughly and uncover the limitation. There's a section that I called Getting under the Hood. I provide concrete examples to illustrate this point and also give guidance to readers who seek to play linguistic games with language bots. All this is from the behavioral level. We previously talked about the brain. Now, this is behavioral because this is what you ask. There's also a structural, namely biological level which we just discussed. What are the parts of the machine? So the discussion is not just about behavior, but also about the underlying brain mechanisms which I previously mentioned.
- [43:02]
Mariam Polukoti
All right, so I understand that you have made some point on how language behavior by humans is clearly different from how LLMs behave. And nobody could actually know how the level of competence of these LLMs except linguists who know how to manipulate language or twist the LLMs. That is interesting because currently I'm working on Genai chatbots and language transcription because my specialization is in phonetics. So we'll talk about that as study interview. Anyway, okay. An interesting thing that actually captivated my attention in the book is the fact that you went ahead not just to tell us, okay, these are the potentials of linguistics. You also identify some pitfalls of linguistic practices. And that leads me to the next question, as I like it. In the book, would you describe the pitfalls of linguistic practices that limit the development of the field? Thank you.
- [44:11]
Professor Yousef Grudzinski
Yeah. This is not a PR exercise for linguistics. Not at all. You know, this is just like a serious scrutiny of the problems that we have and what lies ahead and what we need to do. And there are many problems, of course. You know, I'm a linguist at heart, and I pride my community for its amazing accomplishments over the years. But still, there are fundamental problems, and we have to acknowledge them. So the first and foremost, which is an accusation that is leveled at us time and again, what is true is that linguistic theory has never made itself amenable to implementations. That is, it has long shied away from constructing theories in a manner that could turn them into working. In certain respects, this has been a mistake. Again, there are all kinds of debates that can be made here, but it's pretty clear that it's been a mistake, at least in terms of garnering public support. And public support is extremely important. The second thing is that people have always shied away from pr. So Chomsky did all that. He argued with philosophers, he promoted his fields. He debated computer scientists and other language experts. He was all over the place. In my generation and actually beyond, there's been very little interaction between linguists and adjacent fields. We've been too curled up within our cocoon, and now we're waking up to see it shrink. Glorious linguistic departments are being eliminated, whether they are merged into language science as parts of the university, in parts of the University of California, like at Irvine, for example, or into cognitive science departments, as in University College London, Brown University, or into foreign language schools, as in Michigan State. You see it in many places, budgets for linguistics research are shrinking. And while smart students still knock on the doors of elite schools like MIT or Oxford, the numbers do not begin to approach those of computer science department. So we live in dangerous times, not only for the world, but also for the science of linguistics. Another thing is that linguistic has not formed a sufficiently positive, sufficient, positive collaboration. Recent fields like neuroscience and education and most interdisciplinary, multidisciplinary creatures like myself are sadly few and far between. I'm very, very sad to say so. These are strategic criticisms, but they do have a strong impact on substance. And while it's easy to capture the public's imaginations with the new big Machine, I'm just thinking about the time where IBM's deep blue computer beat in the world champion Gary Kasparov in chess almost 30 years ago. You know, it's like, all over the place. So while Chomsky made a snide comment at the time that the competition between a human and a computer is like a weightlifting contest between a human and a bulldozer, the headlines were out there. It was an accomplishment and people recognized it. Mathematics, by the way, suffers from a similar PR problem. So if you prove some very difficult theorem, it's no news to anyone and it's one of the most difficult intellectual activities one can imagine. So linguistics is a little bit similar, but it doesn't have the social status of math. That's the problem. So you know you mostly impressed when you see something working.
- [47:39]
State Farm Advertiser
So these are problems Starting a business can seem like a daunting task unless you have a partner like Shopify. They have the tools you need to start and grow your business. From designing a website to marketing, to selling and beyond, Shopify can help with everything you need. There's a reason millions of companies like Mattel, Heinz and Allbirds continue to trust and use them. With Shopify on your side, turn your big business idea into Sign up for your $1 per month trial@shopify.com SpecialOffer Zootopia
- [48:11]
Canva Advertiser
2 has come home to Disney. Let's go get ready for a new case.
- [48:15]
Red Bull Advertiser
We're gonna crack this case and prove
- [48:16]
American Express Advertiser
we're victorious partners of all time. New friends, you are Gary Desnake and
- [48:21]
Mariam Polukoti
your last name Desnake Dream Team Habitats
- [48:26]
Professor Yousef Grudzinski
Zootopia has a secret reptile population. You can watch the record breaking phenomenon at home. You're clearly working at Zootopia 2, now available on Disney.
- [48:38]
Canva Advertiser
Rated PG.
- [48:39]
American Express Advertiser
I get so many headaches every month. It could be chronic migraine, 15 or more headache days a month, each lasting four hours or Botox.
- [48:48]
Botox Advertiser
Onobotulinum toxin A prevents headaches in adults with chronic migraine. It's not for Those who have 14 or fewer headache days a month. Prescription Botox is injected by your doctor. Effects of Botox may spread hours to weeks after injection, causing serious symptoms. Alert your doctor right away as difficulty swallowing, speaking, breathing, eye problems or muscle weakness can be signs of a life threatening condition. Patients with these conditions before injection are at highest risk. Side effects may include allergic reactions, neck and injection site pain, fatigue and headache. Allergic reactions can include rash, welts, asthma symptoms and dizziness. Don't receive Botox if there's a skin infection. Tell your doctor your medical history, muscle or nerve conditions including als, Lou Gehrig's disease, Myasthenia gravis or Lambert Eaton syndrome, and medications including Botulinum toxins as these may increase the risk of serious Side effects.
- [49:32]
American Express Advertiser
Why wait? Ask your doctor. Visit botoxchronicmigraine.com or call 1-800-4 for Botox to learn more.
- [49:40]
Mariam Polukoti
Thank you very much. Thank you very much. Because you've actually hit the nail on the head as far as linguistics and some of its pitfall, which makes it so highly relevant as concerned. All right. How central is grammar to the LLM's design? How achievable do you think it is? A total grammatical competence Is of the LLMs is achievable. You think that is achievable?
- [50:10]
Professor Yousef Grudzinski
Professor Youssef this is the shortest answer I will ever give you, which is. I just talked about it a little earlier. None. Linguistics is not a part of this computer science curriculum in any form of way today. It's just not. Is it achievable? I'm sure it's achievable. And I was talking about it a lot, about how we must work together. Otherwise, you know, it's really going nowhere.
- [50:36]
Mariam Polukoti
Okay, now to the last question, which is, I think will go a long way to say, to impact on our readers. What are your parting words to the audience and potential readers? Professor Youssef?
- [50:54]
Professor Yousef Grudzinski
So here's what I have to say. It's very, very simple and very short. Scientist or engineer? That's the question. So individuals who engage in any of these are mostly curious, creative, imaginative, critical. So what to choose? I hope some young listeners listen to this podcast and they're thinking, should I be a scientist? Should I be an engineer? So I want to do complex things with machines, which is very difficult, or sit in front of a black wall of a blank wall, sweat and just reflect on some hard problem and persist until a solution pops up. It's really, really a question of personality, intellectual taste, and also of one's financial aspirations. Either way, you must remember, just emulating language by a huge machine is very hard, but it doesn't come close to a theory about the nature of language. You know why? Because language is deeply human. It's a truly deep human capacity.
- [51:59]
Mariam Polukoti
Thank you very much. That's mind blowing. It is good to have you today, Professor Yosse.
- [52:05]
Professor Yousef Grudzinski
It was great to have you interview me. You're a great interviewer. You really are.
- [52:10]
Mariam Polukoti
Hopefully you'll join us again on the New Books Network sometime in the future.
- [52:16]
New Books Network Host
Thank you for listening to this episode of the New Books Network. We are an academic podcast network with the mission of public education. If you liked this episode, please share it with a friend and rate us on your preferred podcast platform. You can browse all of our episodes on our website newbooksnetwork.com Connect with us on Instagram and Bluesky with the handle EW Books Network, and subscribe to our weekly substack newsletter@newbooksnetwork.substack.com to get episode recommendations straight to your inbox.
- [52:59]
Bloomberg Advertiser
Some Follow the Noise Bloomberg follows the money because behind every headline is a bottom line. Whether it's the funds fueling AI or crypto's trillion dollar swing, there's a money side to every story. And when you see the money side, you understand what others miss. Get the money side of the story. Subscribe now@bloomberg.com.