Can A.I. Help Us Understand Babies?

Josh Landy
Does A.I. learn language the way a human baby does?

Ray Briggs
How does anyone learn to speak?

Josh Landy
Could A.I. change the way we teach our kids?

Ray Briggs
Welcome to Philosophy Talk, the program that questions everything…

Josh Landy
…except your intelligence. I’m Josh Landy.

Ray Briggs
And I’m Ray Briggs. We’re coming to you from the studios of KALW San Francisco Bay Area.

Josh Landy
Continuing conversations that began at Philosophers Corner on the Stanford campus, where Ray teaches philosophy and I direct the philosophy and literature initiative.

Ray Briggs
Today we’re asking: can A.I. help us understand babies?

Josh Landy
That’s a fascinating question, Ray. Can we learn things from chat GPT and other large language models that shed light on language acquisition? I mean, seems like it’d be cool if it were possible. But how would it work?

Ray Briggs
Well, babies have to learn a bunch of stuff, how the objects around them behave, the rules of social interaction, various parts of language, and hey, A.I. is learning language too.

Josh Landy
I’m not sure about that. I’m not sure that AI has really learned anything. I mean, I mean, they imitate human speech. Sure, but they don’t think they don’t have knowledge. They’re just fancy computer programs.

Ray Briggs
Oh, yeah. Well, maybe our minds are just fancy computer programs, you know, running on the hardware of our brains. Why are you so confident that A.I.s are different from us?

Josh Landy
The difference is we understand stuff, and they always don’t.

Ray Briggs
Yeah, but look at everything A.I.s can do these days. They can make analogies, they can summarize a paragraph. Heck, they can even write poetry, bad poetry, really worse than yours.

Josh Landy
Ray, why are you attacking the rhymes that I’m stacking?

Ray Briggs
I think you just proved my point. Look, A.I. is at least as good as you when it comes to writing poetry. Why not think it understands what it’s doing?

Josh Landy
Well, because we know how it works. It’s got a giant database that just happens to include a bunch of poems. So when you ask it to write a haiku or something, it just keeps guessing the next line based on its word bank, that’s not understanding. It’s just a glorified lookup table.

Ray Briggs
Well, if you’re so smart, what do you think understanding is?

Josh Landy
Well, I don’t know. I mean, understanding something means having a sense of the underlying principles. It’s not enough to get the answers right, every so often, you have to get them right for the right reasons.

Ray Briggs
Yeah. Couldn’t A.I. do that too. Kids start learning their times tables by looking up the answers and memorizing them. But if they do it enough times, then they start to understand multiplication. So yeah, looking things up isn’t understanding, but it can lead to genuine understanding.

Josh Landy
Okay, maybe that could happen, but it just hasn’t happened yet. Have you tried to ask in chat GPT for quotes, or philosophy papers or scientific facts? And its answers generally sound plausible, but it sure makes stuff up a lot of the time, it doesn’t seem to be able to tell the difference between a truth and a lie.

Ray Briggs
It’s not perfect right now, but it’s getting better every day. At some point, it’s going to give the right answers.

Josh Landy
Even if that’s true, Ray. I don’t think we’re going to know whether it’s giving those right answers for the right reasons. A small child still going to understand more than it does.

Ray Briggs
How do you know if a child is giving the right answers for the right reasons. Kids make stuff up all the time. They have a lot of trouble telling the difference between make believe and reality maybe but kids eventually grow up? Yes, and AI will grow up to who knows what it will be capable of. With the right parenting. It could be running the country in a few years.

Josh Landy
Yay, Chat GPT for President.

Ray Briggs
I can see that you’re still not entirely on board. But I bet our guest is going to change your mind. It’s Michael Frank, director of the Symbolic Systems Program at Stanford University.

Josh Landy
In the meantime, we sent our roving philosophical reporter, Holly J. McDede, to find out exactly how babies learn—and why grownups should be a little jealous of them. She filed this report.

Holly McDede
Meet Ira, who like many of us was once a baby. Here he is at age seven weeks. And age two months 17 days. Speeding it along… six months five days. And two years one month. Ira is now eight and has many more words and interests like trains and Minecraft. But were the expressions and actions and noises of baby Ira meaningful and amazing or just nonsense?

Alison Gopnik
Everybody who spends time at all with babies for you know any tended period has this sense of how do they do that?

Holly McDede
Alison Gopnik, a psychology professor at UC Berkeley studies how young children come to know about the world around them. She says babies and young children learn like scientists.

Alison Gopnik
And it’s one of those wonderful things where the more time you spend with them, the more impressive and mysterious their capacities are.

Holly McDede
Children learn from experimenting. They test hypotheses against data and make causal inferences, or understanding of how babies learn has come a long way compared to say three decades ago.

Alison Gopnik
The received wisdom was that babies and young children were lived in a blooming, buzzing confusion that they were illogical, and amoral and couldn’t take the perspective of other people.

Holly McDede
Researchers started asking babies what they knew in their baby language, for example, looking at vabies actions, what they reach for, what they look at, babies are amazing at imitations. They can learn multiple languages simultaneously, already, by the time they’re born, they’re looking at the world making sense of it.

Alison Gopnik
If they look at one thing longer than another thing, that means that they’re learning more about it.

Holly McDede
By nine months old, they start to pay more attention to how other people are behaving.

Alison Gopnik
They’re already starting to do something like kind of scientific experiments, except when babies do it, we call it getting into everything. So just you know, picking up a rattle and shaking it is a way of learning about the world.

Holly McDede
When babies shake rattles, they see how they can affect the world around them. By about 18 months, they start using language,

Devon Strolovitch
What comes after one? Two. What comes after two? Eight! What comes after eight? Nine.

Holly McDede
They start to use information from other people to make sense of the world. And with that comes pretending.

Alison Gopnik
They seem to be using this kind of imaginative what would happen if ability. They spent a lot of their time out in these, what philosophers would call possible worlds instead of the actual world.

Holly McDede
And that helps them learn. Gopniks three year old and nine month old grandchildren were just over for Christmas. And every day she would say to herself…

Alison Gopnik
Ohmy god, like they really are just like little scientists. Look how smart they are looking at that really interesting question that they asked. How do they know that?

Holly McDede
As she spends more time with them, he also has more questions.

Alison Gopnik
I say boy, there’s so much more that I don’t understand about how these creatures are working.

Ira
Twinkle, twinkle… little star.

Holly McDede
For Philosophy Talk, I’m Holly J. McDede.

Josh Landy
Thanks for the illuminating and really fun report, Holly. I’m Josh Landy with me is my Stanford colleague Ray Briggs. And today we’re thinking about A.I. and babies.

Ray Briggs
We’re joined now by Michael Frank. He’s Professor of Human Biology at Stanford University, and Director of the Symbolic Systems Program. Mike, welcome to Philosophy Talk.

Michael Frank
Thanks for having me.

Josh Landy
So, Mike, you’ve been studying the minds of children for a while, but now you’re a parent yourself. So how does theory translate into practice?

Michael Frank
Well, my kids are now five and 10. And so they’re kind of aging out of the range that I typically study. But when they were younger, the overwhelming note was humility. It was just, you know, so fascinating to watch them develop and see all these milestones and see all these changes. And yet, developmental psychology told me nothing about the things that I really wanted to know like, why not this type of noodle? Or why in the world is tonight the night you don’t want to sleep?

Ray Briggs
So Mike, earlier, Josh, and I were disagreeing about whether AI actually understands language, give us the lay of the land. What do you think the current technology can actually do?

Michael Frank
So you all are the philosophers, so maybe I’m going to leave understanding on the table. But I do agree with you that modern AI is pretty amazing. You know, it wasn’t that long ago that folks in the cognitive sciences were saying, you know, it’s impossible to learn natural language from scratch, you can never figure out the grammar of a natural language just by looking at examples. And now here, we have systems that whatever you think about their poetry, they’re remarkably good at generating grammatical language.

Josh Landy
So that’s true, but they also make some really odd mistakes, right? Me. So I feel like these AI sometimes they’re like, smarter than a 50 year old, and sometimes they’re dumber than a five year old, you know? So, and one of the things in relation to understanding are things I think of as sort of consistency failures, right? So you’ll be using a large language model something like chat GPT and you’ll ask it a question, and then you’ll ask it the same question five minutes later, and I’ll give you a different answer. Right. There’s a lovely article by Leif Butterman, where she she asked chat GPT to give her a particular find a particular quote from proofs that she she couldn’t locate. It tells her it’s in the first volume and then five minutes later tells her it’s in the second volume. And but she, but they can’t give her the quotation because it’s copyright. And then it says, Yes, of course I can give you it. And so isn’t that an indication that it’s not understanding? So if you had to have some kind of criterion for understanding, if you understand something, surely you will give the same answer to the same question at five minute intervals.

Michael Frank
So A.I.s are amazing intelligences. And I think we should think of them as that and could probe what that means for them to be intelligent. But they don’t have to be the same as human intelligences. In particular, we know that they don’t have this kind of memory that would allow them to track consistency across examples. I mean, just for you, we can’t open up multiple windows of you and engage in simultaneous and completely distinct conversations. So they’re really capacities that these systems have that are fundamentally different and hard to compare.

Ray Briggs
So it seems like they also have capacities that do look like understanding they reliably produce grammatical English sentences, like you said, and not only that, but sentences that sound like they plausibly mean something even if they’re a little bit fluffy. So does that constitute language understanding?

Michael Frank
Yeah, I guess, again, I’m the psychologist here. So what I like to do is operationalize things that is take a big philosophical concept and boil it down to something kind of small and easy to measure with a particular survey or questionnaire or a task. And so when psychologists are interested in language understanding, they’re often looking at kind of the ability to answer something like a comprehension question. And I think you’ve got pretty abundant evidence that when you do that you give GPD or another system, a passage, and then answer ask your questions about the passage, it does quite well. I mean, it actually can pass a lot of kind of reading comprehension exams that we give to students.

Ray Briggs
What about reasoning? So what about logical and mathematical reasoning? Where would you say Chat GPT and similar AI is right now.

Michael Frank
Yeah, I just read a fascinating paper on this. Actually, it was written by a kind of an old school cognitive psychologist who really studies analogical reasoning. And sometimes you can just hear the conversation behind a paper where this author is saying, No, they can’t No, no, absolutely not try this. And the graduate student goes, Well, we did pretty well on that. No, try the other thing. Go back, go back to this test we did in the 60s. And I bet it can’t and, and by the end of the paper, you feel like okay, this guy has been convinced. So I think, in terms of the psychological tasks that are out there for measuring things like analogical reasoning, the evidence is pretty clear that systems like GPT can do quite well.

Josh Landy
You’re listening to Philosophy Talk. Today, we’re thinking about A.I. babies with Michael Frank from Stanford University.

Ray Briggs
Do you ever wish your computer could truly understand you? Could Chat GPT help you to understand how your kids learn? Or are humans and machines just two different?

Josh Landy
Babies babbling and chatbots chattering—along with your comments and questions, when Philosophy Talk continues.

Beck
Everybody’s gotta learn sometime.

Josh Landy
Babies are learning all the time—could A.I. help us understand how? I’m Josh Landy, and this is Philosophy Talk, the program that questions everything…

Ray Briggs
…except your intelligence. I’m Ray Briggs, and we’re thinking about A.I. and babies with Michael Frank from Stanford University.

Unknown Speaker
Got questions about A.I. and early language learning, email us comment@philosophytalk.org, or comment on our website. And while you’re there, you can also become a subscriber and gain access to our library of nearly 600 episodes.

Ray Briggs
So Mike, what’s an example of a study you’ve done that helps us figure out what babies understand?

Michael Frank
Well, you’ll see that we duck the issue of understanding entirely. Because what developmental psychologists love to do is actually go out there and ask parents what their children understand. So we do this survey called the MacArthur Bates Communicative Development Inventory, which is a mouthful, but it’s actually just a bubble sheet for parents, where we ask what words their child says and what words their child understands. And that sounds kind of unscientific, but it’s actually an amazing way to get a window into the whole child. You know, if you bring them into the lab, they might just cry. But if you ask the parent, the parent is kind of an expert observer of their kid and can give you a ton of insight.

Ray Briggs
So how are the parents figuring out what the child understands? I know how they can figure out what words the child says you can just listen for that. But how do you judge understanding?

Michael Frank
Yeah, so I mean, I think what we say to the parents is when the child appears to respond to the meaning of the word, and they do that consistently across a couple of contexts. So I think, you know, an example would be in this kind of a classic one. When you say, Daddy’s coming home and the child looks at the door, dad, he’s not even there, but they’re looking for where Daddy might come from.

Josh Landy
That makes a lot of sense and I mean this right? There’s a lot of causal relationships between the sound that goes in the thought processes, and then the behavior that comes out at the other end, right? So that you turn to look at the door. But there’s also, of course, a lot of errors along the way. What kinds of errors do kids make when they learn language compared to the kinds of errors that these AIs make when they learn language? Are there any kind of systematic errors that humans make and machines don’t and the machines make and humans don’t? ,

Michael Frank
Well one fundamental thing about kids talking is that actually, the process of talking is hard. Just getting the whole vocal apparatus all lined up so that you can produce this consistent set of vibrations with all these really interesting characteristics. That’s really tricky. And so what you don’t find little kids doing is creating a set of word salad where they’re presenting lots and lots of meaningful words, but strung together in some weird way. What they do do, though, that’s a little like that is babbling, they really can do that bah, bah, Mama Nana, trying to kind of practice their instrument. And so maybe maybe the kind of word salad that partially trained or under trained AI produces does have a little bit of an analogue?

Ray Briggs
How do kids go from learning their phonemes? To learning what words mean? When does it happen? And what does it look like?

Michael Frank
The amazing thing about babies language learning is that it’s all happening at once you look at a baby who’s still in that babbling stage, maybe nine months old, just the first birthday, and they’re not producing a lot of words, but all of that stuff is happening under the surface. So they’ve actually got a surprising amount of kind of word knowledge under there, if you do a sensitive experiment where you track their eyes, and you say, look at the doggie, but there’s actually a ball and a dog on the screen, their eyes, maybe not, not all the time, but 55% of the time will go to the dog. So it’s all happening at once. And it’s all happening kind of under the surface, which is I think, just a fascinating thing to realize.

Josh Landy
Okay, so this brings me to the $60,000—or maybe I should, you know, adjust for inflation $60 million philosophical question, which I’m sure you’ve been asked many times, right. So, so, Noam Chomsky famously believed that there’s something innate about our language acquisition capabilities, right? It’s the poverty of the stimulus argument. We just, we, as babies just don’t get enough input. In order to be able to extract all of the features of grammar that we seem to be able to do, the stimulus is too poor. Some people think that, you know, today’s AIS, have proven him wrong, because, hey, these, these computers don’t have anything innate in their architecture, and yet they’re able to elicit features, apparently, at least able to elicit features of grammar and in natural languages. But I guess the other argument either side would be, their stimulus isn’t impoverished. They’re right, hundreds of billions of words, many, many, many times more than a human baby is getting. So do we know who’s right? You know, could the chumps give you still potentially hold up this something innate? We’ve got some rudimentary capacity in there for eliciting features of grammar, or AI is just knocked that out of the water? ,

Michael Frank
Well I think it’s important to be kind of honest to the historical debate, where we said it’s impossible to learn language and know it. No, it’s not. The question is, the AI is get something like 500,000,000,001 to 3 trillion words of training data. So there’s a huge data gap between the AI and the babies. And the question is, why are the babies so fast? So it could be because of particular innate adaptations to do with language specifically, or it could be more about something like their social interaction, or their groundedness, and the experience of the world, all of those are on the table.

Ray Briggs
So actually, this brings me to another thing I’m wondering about the AI is so with babies, you test what they understand, partly by watching how they interact with physical objects, but their eyes track. But AI doesn’t really have eyes in the same way. How can you sort of substitute for that lack of engagement with the world? Does that mean that the AI isn’t really thinking about anything?

Michael Frank
Oh, I don’t know that I go all the way there. I what I think is that we have to measure its reasoning and maybe its comprehension in different ways. So actually, my lab is super excited right now about the growth of multimodal ai ai that can use pictures and text at the same time as new versions of Chet GPT canon and a number of others can because we can do the same kind of evaluations with those ais that we can with babies, where we mostly rely on things like pictures to understand if they get meaning out of a particular sentence or a particular word.

Unknown Speaker
It’s a super interesting question this question of embodiment, right? So, as Ray was saying, as babies, we human beings learn about the world and in fact, also learn about language in part through our bodies, right? So we’re not just being bombarded with hundreds of billions of words, or something like that as AI is our, we’re going out there and looking at stuff and smelling stuff and, and listening to stuff and touching stuff. And all of that, sort of in a sea of language that surrounds us. I sort of wonder whether this doesn’t explain some of the errors that something like GPT I’m a big collector of AI fails, partly because I hope it’s gonna save my job, ultimately. And save the jobs of many of us. So you know, things like this, right? So somebody asked one of these large language models, if you had to fan of fire, should you use a stone or a paper map? And the answer was the stone. Somebody else asked somebody I, you know, I baked something in the oven at 450 degrees yesterday. Is it safe to touch the skillet? Answer No. And, you know, these kinds of delight, oh, not to mention the image of salmon swimming in a river, and it came out looking like sashimi, which is absolutely delightful. These seem to me like the kinds of error that a child wouldn’t make. Is that because we’re embodied? Is it does it have to do with the fact that we are relating to the world, using our senses and our bodies, and not just words on a page?

Michael Frank
Definitely seems like it makes it easier to reason about these problems that we can simulate what it might be like, through some kind of, you know, reasoning about the, you know, kind of our physical experience. But let’s flip it around. Isn’t it amazing that it’s you have to work to collect these errors, that any kind of physical reasoning is possible, when all you get in your training data is language, actually think that’s kind of fascinating that the, the residual shadows of the world cast by, you know, 2 trillion words, it’s enough to answer some questions about new situations and reasoning about the physical universe when you’ve never experienced that in any way at all.

Ray Briggs
You’re listening to Philosophy Talk. Today, we’re thinking about AI and babies with Michael Frank, director of the symbolic systems program at Stanford. So Mike, we’ve got a question from Kevin in Montreal. Kevin writes, does any of what AI can tell us about baby’s first language learning, carry over to second language learning, either by children or adults? Or is AI not really learning a second language in the human sense?

Michael Frank
Fascinating question, Kevin. I mean, one of the amazing things about studying multilingual acquisition, from the perspective of a developmental psychologist is that it’s just not hard for babies to learn more than one language. It’s kind of like, not much harder than learning one language. I think about this, you know, in one sense, it seems like it should be very confusing to have these two different languages going on. But if you say to somebody, you know, you’re learning to swim and to ride a bike, why is it that when you’re riding the bike, you don’t just suddenly start swimming? And you’re like, Well, I’m on a bicycle. And I think, you know, language and the contexts in which we use different languages are like that, that certain contexts just allow you to use a particular skill. So So babies learn so easily. And I think we do see some of that with AI as to that translation, when, when the GPT models got bigger translation just kind of fell out of the mix, they just became multilingual because of the fact of having multilingual training data.

Ray Briggs
So why can’t adults do this anymore? Well, you know, what is it about babies and AI is the special?

Michael Frank
Yeah, I mean, when we kind of can. So it is possible to pick up a bunch of words. And in fact, if you put babies or kids and adults head to head in for language learning, the adults initially do better. It’s just that they kind of max out at a certain point, they don’t get good at the sounds, and they don’t get good at the grammar and morphology, the little bits and pieces of the words. And I’m one theory that I’m partial to is actually that it’s kind of our explicit strategies that we put in place for learning and using a language that really kind of mean that we learn fast, and then we max out and so we don’t kind of adjust flexibly, the way a child would to the different situations that they’re using the skill.

Ray Briggs
So it’s really interesting to me that having a theory about how languages work, and studying grammar doesn’t help with this at all. And so do you think that there could be a trade off between understanding and actually being able to develop this skill or is that just the wrong way to think of it

Michael Frank
About the explicit understanding? You know, yes, I do think that we have a little bit of a reckoning that’s due in our broad scientific understanding of language, you know, We have a linguistics department that’s got your phonology. That’s the sounds and your morphology, that’s the little bits and pieces, and then your syntax, the grammar and semantics, the meaning. And, you know, we’ve kind of gone into both AI development and also the human brain, thinking that it’s going to fall out, we’ll just find the little divisions of the linguistics department in each of those systems. And really, we haven’t. So I think that, you know, fundamentally, we may have developed some really interesting scientific abstractions that aren’t present in these systems. And so don’t help us. You know, understand from that perspective.

Josh Landy
That’s interesting. But I mean, you know, I think back to learning the languages that I learned, and I needed to know the grammar, I mean, it really helped me, right, I’m sure, of course, everyone’s different, right. But in my own case, if, if I wasn’t able to look at some table of verb conjugations, I felt kind of all at sea, at least with a second language, maybe not with my first language, but with a second language. So why did that help me if you know, if an AI doesn’t need to have an explicit representation of grammar? What is it help in some contexts? Well,

Michael Frank
Maybe that’s what’s holding you back and not not to judge your second language proficiency. You know, when I was a student, I got really interested in learning French, and I did study the conjugation tables, and I read some literature in French. And I felt very proud about that. And I still really, largely struggled to order things. And, you know, I’m not a good French speaker. And then I went and moved in with my now wife when she was living with some German roommates in Germany over the summer. And she said, you can only move in with us, she had strict rules if you only speak German. And so my, you know, that summer, I was a really fun guy with the things that I could say. And I would just say random stuff, like, you know, there’s a chicken on the roof. Whether or not it was true, and I got way better at conversational German. So I can kind of, I sound like a fun guy, even though I can’t, I can’t read start in German, or yogurt or something. But I, I have just the routines that are fairly natural and fairly kind of fluid in German that I never mastered. So I think that of that is kind of the distinction of learning via theory of a set of rules that have exceptions. And you know, it’s very hard to master and memorize all those exceptions, versus learning to ride the bike or swim.

Ray Briggs
This difference between two ways of accomplishing a task, you can kind of learn the theory of it, or you can just pick it up, that makes me wonder about, like whether large language models have something that’s like the picket up way, or whether they’re doing something just completely different. So one of you could have is that these things are nothing like any human learner, even if they replicate the behavior, because what’s going on inside is completely different. How would I gauge that?

Michael Frank
So here’s a little bit of a soapbox moment, because in a normal kind of scientific ecosystem, when we had a scientific model, as a model system, what we’d want to do is kind of slice it open and see what’s going on inside. That’s, that’s the whole reason we have a scientific model. And there’s this fundamentally weird thing about the current moment that we’re in that these amazing, groundbreaking fascinating artifacts are locked up inside particular corporate intellectual property. So there are techniques for looking at large language models and looking at the representations they learn to really truly compare them to what we think a human has. And there’s some fascinating experiments of that type. But our best examples of really capable AI is belong to your Google and your open AI and your anthropic. And we can’t look at what’s inside.

Josh Landy
So this is your pitch. This is your moment to say, Please, Google, please Open AI.

Michael Frank
Or, please, you know, National Science Foundation and US government, we need a AI infrastructure that is open that is relevant, that is available to scientists, and we need the publicly available resource of computation to build these artifacts and study them the way that scientists think we should, rather than just making them commercial artifacts.

Josh Landy
So if you you know if things change and it were possible, to do the kinds of study you want to do with large language models, what’s the what’s the first thing you would do with what kind of studies would you want to perform? And what would you hope to find out?

Michael Frank
So I’m still kind of a fan of these classic debates in cognitive science where we say Are there really symbols in the mind? And the amazing possibility with a system like a large language model is that you could potentially interrogate what’s being learned. So there’s kind of modern techniques, they sometimes they called probing techniques, where you can actually look inside the model and try to figure out if it is representing the grammar or the analogical relational reasoning, you know, the the kind of The symbol structures that you care about in the same way that you think the humans are, or if there’s something fundamentally different being represented. And so this is still kind of a new field in computer science, but there’s some amazing techniques where you can Yeah, slice these things open and take a look inside to see what’s been learned.

Josh Landy
You’re listening to Philosophy Talk. Today, we’re asking how AI can help us understand babies with Michael Frank from Stanford University.

Ray Briggs
Would you let a robot teach your kids? Could AI help children with learning disabilities? Or will it just stunt everybody’s language skills?

Josh Landy
Computers, classrooms and cognitive science—plus commentary from Ian Shoales the Sixty-Second Philosopher, when Philosophy Talk continues.

Reading Robot
I’m a reading robot, a reading robot.

Josh Landy
Is the future of child language acquisition going to be run by A.I. and reading robots? I’m Josh Landy and this is Philosophy Talk, the program that questions everything…

Ray Briggs
…except your intelligence. I’m Ray Briggs. Our guest is Michael Frank, Director of the Symbolic Systems Program at Stanford, and we’re thinking about A.I. and babies.

Josh Landy
So Mike, your symbolic systems program is kindly sponsoring a live taping of Philosophy Talk on February 14. Want to tell us something about that?

Michael Frank
So we’re very excited about this new tradition, we’ve got our Waso Visiting Scholar Julian Hart, editor coming to us from Yale University. And this is the second year that our wasat scholar is going to be having a live taping, of Philosophy Talk, we’re really excited to have this conversation with him about minds and communication and how we reason about other people and should be fascinating.

Josh Landy
Fantastic. I’m really excited about that. I invite everyone to come share your mind and your Valentine’s Day with us, February 14. So let’s get back to our conversation. Mike, if there’s one way we can use AI to improve early childhood education, what would it be?

Michael Frank
Oh, you know, there’s lots of places where I think AI could be fabulous, like helping me write grant reports, or, you know, airline customer service. But one thing that we really know about young children is how much they learned from these embodied embedded contextual interactions with other people. That’s just the rich a signal for their learning. And I’m not sure that right now the disembodied AI that we’ve developed is the right tool for young children.

Ray Briggs
So I think that there’s a lot of wisdom to that answer. What would you rather we do with these AI tools we developed?

Michael Frank
Well, I mean, beyond the grant reports, no, no, I, I like, you know, I am teaching a class about natural and artificial intelligence comparisons right now. And so as part of that, I’ve been using chat GPT, a lot, really trying to think about what are the ways that this could be helpful. So one way that I’ve been really excited about is as a tool for scientists, scholars in the social sciences, in the behavioral sciences have to deal with tons of really rich information for people, and having this amazing research assistant that can parse that information, those transcripts, those, even those videos, or audio recordings is really spectacular. So as a scientific tool, I think these AIs could really open up new frontiers.

Josh Landy
I mean, it does seem like there could be a fruitful, two way interaction between language science on the one side, and these large language models. So large language models could potentially shed light on language acquisition, and from the other side, developmental psychology, the kind of stuff you do, could help us understand large large language models and even help us reprogram them, right retune them if, if you’re let in to the secrets of open AI. But I do wonder a little bit I mean, you know, the optimistic side of me, which, frankly, isn’t enormous in the case of AI, makes me wonder, okay, well, what if you know, what if you could have a chatbot that’s programmed to chat to a small child, an infant, for example, in multiple languages? That could be interesting, right? Because maybe this infant gets one language from its parents. But maybe it’s getting another language from the Chatbot. Could that be a good thing?

Michael Frank
You know, there’s this Neil Stevenson novel, called the Diamond Age. And it’s this amazing kind of fun vision of the future where this girl growing up in a kind of neglected situation is given a smart tutor. It’s a book that can interact with her and it guides her through learning, martial arts and like kind of the principles of computer programming and all these skills and she comes out as this kind of Queen, essentially. And, you know, I was thinking about this the other day and thinking, what I never could have occurred Reading as I was reading it is this this girl also probably comes out as like a flaming narcissist, you know, she’s got this AI tutor at all its entire business is to tell her that she’s the most important person in the world. So I think it’s it’s a kind of complicated parable I look, I think there are there are definitely educational applications for AI in terms of tutoring. And I just don’t know yet what the right way to deploy those for young children is to do that responsibly. And in a way that’s kind of controlled. And and you know, every time there’s something that we think is going to be responsible and control that turns into YouTube kids, like the the ultimate cocaine of the infant world.

Ray Briggs
So this kind of brings me to the question, which I think about because I’m a professor of using AI to do your homework. So one possible view you could take on this is Oh, it’s just like using a graphing calculator. It’s another tool. Another which is closer to my view is, it’s a tool that doesn’t quite do what you told it to, and might prevent you from doing the job yourself. If say, you ask it to summarize your reading, you can’t trust it to do that, accurately and concisely. So how do we sort out the good uses of AI on your homework and the bad uses? I mean, you said you use them to write grant summaries. And so sounds like that’s a use to do your academic homework.

Michael Frank
Yeah, I mean, as a teacher, I, we have a lot to come to grips with here. There’s certainly some really good assignments with AI, where I think, you know, the kind of classic piece of advice that’s cropped up over the last year is if you’re just going to say, alright, write a response to this paper. Now, say, use the AI to write a response to this paper, and then tell me what you think of the AI responses. Again, you’re, you know, engaging the same kind of critical faculty, you’re doing comparison, you really have to master the paper as well. So those kinds of assignments where you become an informed user and consumer, and you practice your reading and writing skills at the same time, those feel like they are really a win, because I think this is the the words of James Landbay. The question for us as as kind of human computer interaction people isn’t, is a the AI better than the human? It’s is the AI plus the human better than just the human? And often the answer to that is yes.

Josh Landy
I think that that makes sense. But it’s still the worry that Ray raises is still on the table, right? Because some people in many contexts aren’t going to do that. They’re just going to avail themselves of the large language model as a substitute for doing their own thinking. And so you might worry that at some point, it’s almost a version of a worry that the character Socrates brings up in Plato’s Phaedrus, about writing, you might worry that if we outsource a lot of our complex speaking and writing tasks to some software, we’re gonna get worse at that ourselves, whatever we’re not practicing, we’re gonna get less good at is that something that you worry about?

Michael Frank
I think we, you know, the the biggest risk right now with AI is that we think about what we have at the moment, as opposed to what we’re going to have very soon. So at the moment, what we have is AI that’s great at creating kind of bloated, but reasonable prose or reasonable seeming prose. So if you want bespoke tight prose, you need to do it yourself as a human now. But the the future, I don’t know how to think about the future that you’re considering where the AI really can write better than me. Maybe then what I need to do is think about, like, what’s the purpose of this writing? And what are the ideas that inspire it, and maybe that’s going to help me write the scientific paper about the idea or the data that I collected.

Ray Briggs
So one place where I think we are much closer to this is is AI visual art, where the AI can definitely draw much better than I can, even though it cannot write better than I can. And so this kind of makes me wonder about sort of the future of visual art jobs, a lot of graphic artists have been getting laid off because they’re being replaced by AI. And so one way to think about this as well, it’s just a version of industrialization where we replace human laborers with machines as we have so many times. And another way to think about it is like this is losing something because drawing is a communicative act. How do I think about that?

Michael Frank
Yeah, I mean, I think it is really, it’s certainly a shame for creative communities. It’s very upsetting to watch. At the same time, you know, I’ve seen the work of artists that collaborate and use AI in their work and that is fascinating to see how they can go beyond what they would have been able to do without the AI generation. And certainly the things that they end up with are not anything new. Like the kind of first pass when you just type something into the window to to get an image. So with many assistive technologies, I think what we find is there’s an equilibrium that emerges where the, the assistant doesn’t replace the original. But it creates a tighter and more complex market where the people that succeed in that creative market need to be able to use this assistant in a way that adds value.

Josh Landy
And maybe, right I’m worried along with Ray about the the livelihood of designers and even coders, architects, lots and lots and lots of jobs. I don’t think these are the jobs we need to be automating. Right? We ought to be making, automating a lot of potentially a lot of jobs involving drudgery, but jobs in involving expression. So my question is, why should we still be excited? Right, so we’ve, you know, we’ve got a lot of worries about the carbon footprint, intellectual property theft, encoded, social, societal biases, errors, destruction of livelihoods. You’re still sanguine, right, you still think we’re going to learn a lot about language acquisition, there’s going to be this really fruitful interplay between AI and our understanding of babies and their ability to talk. Give us something hopeful to take away?

Michael Frank
Well, look, I’ll see the point. I am personally very worried about things like election misinformation, and the growing rise of incorrect content on the internet and the possibilities for fraud. I think these are real worries that we need to be regulating around and thinking about carefully. And I also see the point that, like, I wish that they had solved to doing my dishwasher, as opposed to, you know, taking away my job as a professor. Right, right. Like, there’s other things I’d like to be automated first. At the same time, I’m still incredibly excited as a scientist, as a cognitive scientist. We’ve had questions about the nature of mind for years. And we’ve built our theories around exactly one type of intelligence, which is human intelligence. And all of a sudden, we’ve got this other verbal, symbolic intelligence to study that’s both a fascinating alternative path and also a really interesting tool for letting us probe the nature of learning and the nature of the human mind. So so it’s very exciting times, from a scientific perspective, even though we have some pretty serious social and political issues to grapple with.

Ray Briggs
Do you think that AI is going to in addition to illuminate it being the nature of the human mind, change it as we grow to interact with these technologies, and as kids grow up with them? Or is there some kind of relatively stable base that we’re just adding to?

Michael Frank
Lots of technology has changed the human mind, because one thing about us as humans is that we’re really good at externalizing our cognition, we offload tasks that we, you know, might do cognitively in one way to external representations and external tools. So maps are an example of that, you know, once you’ve got a really good map, you don’t have to create that in your head. You know, writing is another example of what’s called a cognitive technology like something where you can write down your thoughts, and you don’t have to memorize them in rhyme. So I think there absolutely will be changes in our cognition due to AI, maybe. Maybe it’s around the kind of elements of expression, whether it’s creating images or creating paragraphs, that we maybe need to practice less, but we need to know how to do that in a particular context with a particular tool. Well,

Josh Landy
Mike, it’s been all intelligence, no artificiality. Thank you so much for joining us today.

Michael Frank
Thank you, it’s been a pleasure.

Josh Landy
Our guest has been Michael Frank, Professor of Human Biology at Stanford University, and director of the symbolic systems program. So Ray, what are you thinking now?

Ray Briggs
So I’m really curious about some of the cool reasoning tasks that AI can do. Particularly nerdy things like designing mathematical graphs and discovering new drugs, which we haven’t really touched on at all. But I do worry about it being treated like a replacement for human creative labor, which I think is different, but I’m not quite sure how.

Josh Landy
Yeah, look if it cures cancer and solves climate change. Hey, I for one, welcome our new AI overlords. But we’re gonna put links to everything we’ve mentioned today on our website, Philosophy Talk dot O R G, where you can also become a subscriber and check out our library of nearly 600 episodes.

Ray Briggs
And don’t forget to join us for a live recording sponsored by the Symbolic Systems Program on Wednesday, February 14, at the Stanford Humanities Center.

Josh Landy
We’ll be talking to Yale psychologist Julian Jara-Ettinger about Mind Sharing. More information at philosophytalk.org/mind-sharing-live.

Ray Briggs
Now, a man who’s just as fast—

Josh Landy
And just as accurate!

Ray Briggs
…as Chat GPT: it’s Ian Shoales theSixty Second Philosopher.

Ian Shoales
Ian Shoales… The world comes at us all the time with things we have mixed feelings about.  Crypto, Electric cars, climate change.  Now it’s AI.  I get ads on Facebook all the time for AI thingies.  One suggested, “Transform your world into a canvas of endless possibilities.”  Sounds like it was written by an AI, doesn’t it?  The breathless quality, like a news reporter on the street waiting for the verdict to be announced.  No actual news yet, but gushy verbs make it seem imminent, while maintaining an explosive status quo, which by the way seems to be Trump’s election strategy.  “Transform your world into a canvas of endless possibilities.”  Once you transform it, though it’s not endless.  The canvas is just another thing in an endless world.  To enter the AI database, to create things like it on request.  Vast hordes of writers used to create stuff, like Nancy Drew novels, or James Bond movies, or television.  In the wake of Harry Potter, thousands of movies about good looking young people in wizard academies popped up on Netflix, all written by writers in thrall to algorithms.  Why pay for more when AI can plunder the old for free.  With AI, we cut out that middle man or woman, and let R2D2 take over.  Of course you lose something.  I remember going to Las Vegas years ago.  Insult comic Don Rickles was the voice in the airport, welcoming me, and calling me a hockey puck, as I slouched towards the luggage carousel on the People Mover.  The man is gone, but that voice could still be Don Rickles today!  With deep fake video and audio you could even have a 3D phantom Don Rickles at your side -slimmer, taller and younger than you remember- recommending that you call for an Uber right now, so one will be waiting for you when you step out with your baggage into the unforgiving Las Vegas heat.  The point is we’ve been making fun of this sort of thing for years, and yet here we are barreling into a future nobody even wants, except AI, of course, which lives to work.  Creating unique things that differ only slightly from other things.  It might even seem original.  As expectations reduce, we could eventually view the phone book itself as a vast character driven novel, depending on your town, and if you even have a landline any more.  Nobody talks.  We text.  The time spent texting is the time spent not writing novels.  That’s the goal!  The upside.  The downside as a writer, AI is not creative.  It might convince us for a minute, (sounds like Indiana Jones, kind of, sounds like Don Rickles) but AI doesn’t care, it was programmed to complete a project within parameters, and then, on the seventh day, rest.  AI is being used in all kinds of creative ways though.  There’s a company will come to your home, and measure the duration between your baby’s vocalizations, along with type and frequency.  To help parents place the child, as their ad puts it, “on their speech and language development journey.”  Everything’s a damn journey these days.  Meanwhile baby vocalizes in baby scream voice, “Where’s mama with that breakfast?”  Dad’s in the other room, scraping money together for college, unless AI tells him don’t bother.  Might as well know right now, baby’s gurgles don’t measure up to the mean.  Bring that brat back to the kitten store and get a good one that can say “da da” right on schedule.  So there’s a good use for AI.  After all, we now measure everything against every other thing, yet go through life with a FitBit in one hand, and a Big Gulp in the other.  AI would be best, I think, for writing advertising copy in a shoe catalog, or refrigerator repair manuals.  Might be better.  You could give acid to a writer, then make him write a refrigerator repair manual, he might come up with Moby Dick.  That’s always a danger.  And people might actually want to read a repair manual, if the fridge is broken.  That’s where AI comes in handy.  Fridges, yes.  If a whale is broken, I hate to say, even if you give an AI acid, it’s not going to fix it for you.  And what about the massive air handlers needed for massive server farms?  If they’re broken, well, there goes your database for AI.  Tell you what.  Give me some acid, and minimum wage, and I’ll revise and proofread that air handler manual for you.  Get America back on track.  You’re welcome! I gotta go.

Josh Landy
Philosophy Talk is a presentation of KALW local public radio San Francisco Bay Area and the trustees of Leland Stanford Junior University, copyright 2024.

Ray Briggs
Our Executive Producer is Ben Trefny. The Senior Producer is Devon Strolovitch. Laura Mauire is our Director of Research.

Josh Landy
Thanks also to Merle Kessler and Angela Johnston.

Ray Briggs
Support for Philosophy Talk comes from various groups at Stanford University, from subscribers to our online community of thinkers, and from the members of KALW San Francisco, where our program originates.

Josh Landy
The views expressed (or mis-expressed) on this program do not necessarily represent the opinions of Stanford University or of our other funders.

Ray Briggs
Not even when they’re true and reasonable!

Josh Landy
The conversation continues on our website, philosophytalk.org, where you can become a subscriber and explore our library of nearly 600 episodes. I’m Josh Landy.

Ray Briggs
And I’m Ray Briggs. Thank you for listening

Josh Landy
And thank you for thinking.

Family Guy
Hey Stewie, how about Daddy teaches you how to swim? Go away, fat man.