Artificial intelligence is changing how we work, how we research, how we live. It could be proposed we are experiencing another revolution, like the agricultural revolution or the industrial revolution. Eric Funkhouser, department chair of philosophy at the University of Arkansas, suggests we are in such a period, and he's incredibly interested in where we're going. This past academic year, he taught the class Minds and Machines, but he's been studying and writing about the subject for some time. Last week, he came to the Carver Center for Public Radio to talk about AI and learning, AI in the humanities and AI in the future.
Funkhouser: I was interested in teaching students about views, according to which the mind, like our brains, is actually a machine, and then also investigate the question of whether or not machines like LLMs that we see today, like ChatGPT or Claude, really have minds, or if not, what would it take for them to get to that level.
Kellams: Is this a conversation, of course, we're thinking about with ChatGPT and Claude, but is this a conversation that perhaps has been happening before the development of LLMs?
Funkhouser: Yeah, definitely. So let me say a bit about some of the historical background. I started the class with, so like probably the first serious attempt to have a machine that could actually do anything approximating thought would be Charles Babbage's Analytical Engine way back in the 1830s. And it was steam powered rather than electrical, but it never was actually built. You may know that someone he worked with was Ada Lovelace, that was Lord Byron's daughter, who some people consider the first computer programmer. But what my class really started with was work from the 1930s, '40s and into the early '50s by Alan Turing and his team. So a lot of people will know Alan Turing, perhaps from some movie depictions like "The Imitation Game" with Benedict Cumberbatch. Well, Alan Turing came up with this concept of a Turing machine about how maybe we could automate, mechanize thought, a machine that's capable of thinking. And then, very famously, he came up with a test called the Turing test, about which creates a standard for what it would take for us to say that a machine truly is intelligent. And basically his test was if in conversation you can't tell the difference between that machine and its answers and the answers of an intelligent human being, we'd say that it has passed the imitation game and it genuinely is intelligent. So in my class, that's where we started, primarily with the work of Alan Turing in the '30s and '40s.
Kellams: And that definition by Alan Turing seems very prescient now, nearly a century later.
Funkhouser: Yeah. And so in a class I taught last fall, a different class called AI, Knowledge and Democracy, we read a lot of studies that show that it seems like the Turing test has actually been passed. Modern LLMs can fool people. People can't tell, at least with the most advanced models, whether or not they're talking with an actual human being or an LLM. And so that has consequences for persuasion, manipulation, all sorts of things like that.
Kellams: Well, it has consequences for so much of what happens on a college campus, whether it's teaching, whether it's writing, whether it's art. I mean, you said a few minutes ago about the brain being a machine. Okay. Yeah, it's a machine and it fires on impulses. But still, it's all the input that has gone into that, the experiences that have gone into it, that I think a lot of us would want to say it makes it more than just a machine, right?
Funkhouser: Right. So maybe I should go back to stuff even before Babbage, from philosophy, that addressed, I should say, some of these questions. So Descartes, philosopher, "I think therefore I am," had these arguments that no machine could possibly think, because a machine couldn't have what he called reason, which is a universal instrument, that's the terminology he used. So Descartes, and this is in the 1600s, so think way back in the first half of the 1600s, before Newton even, Descartes argued that you could have some kind of machine that could do a particular task, like it could do this one job, but human reason is different. We can solve all sorts of different kinds of problems or tasks. We have a general intelligence. And Descartes argued that no machine could actually do that. You would have to basically build a machine to do each discrete task. Now, the modern work in AI is actually trying to disprove Descartes. The goal is to have what's called AGI, not just artificial intelligence but AGI, artificial general intelligence. What that means, the system can do all sorts of different tasks, not just drive a car or not just play chess, but whatever you ask it to do, just like an intelligent human is flexible and can do lots of different jobs. The idea is to create a machine that can do that as well.
So we've actually pretty much established that Descartes was wrong, that machines can learn. It's not just what's inputted into them. Ever since the 1950s we've had what's called machine learning, computers that can learn on their own. And that's why they get better than the human programmers. So if it was just a matter of the quality of the input that's given, there is no way you could have a machine that could beat the best human chess player. But they can because they're actually capable of learning, machine learning.
Kellams: First, I'm going to give Descartes a little bit of some slack here. I mean, it was 400 years ago.
Funkhouser: Yes, yes. Lots of slack. I mean, his arguments were actually pretty good.
Kellams: Yeah, yeah. Okay. So if machines, we know machines can learn.
Funkhouser: Yes.
Kellams: But is that the same thing as thinking, reasoning?
Funkhouser: Okay, that's a very good question. So people mean different things by reasoning. One species or instance of reasoning would be logical reasoning. Some people think that that's the paradigm of reasoning, to be logical. So in philosophy, in our classes, we teach symbolic logic and we teach a lower level logic course also, and we know that machines are capable of logical thinking. The challenge is always whether or not they could think in a looser sense.
So when Turing came up with this test, he knew that machines could pretty easily solve complex mathematical problems. For a human being, that's really challenging, like doing sophisticated math. The challenge though for the machine, Turing gave the test in terms of conversation, right? That's what's going to be hard for a machine to do is to just banter like we are right now. Or if it's not conversation, maybe a simple task, like making an egg, making tea, something like that, that you don't have to be super intelligent as a human being to do. So yeah, there are different kinds of reasoning and thinking. There's logical thinking and then there's like more intuitive kind of thinking. Yeah, but these machines seem to be pretty good at all of these varieties now.
Kellams: All right, so there are these ChatGPT commercials that come on and it shows a couple of brothers fixing a vintage pickup truck's engine or advising a young person how to hit the fadeaway jump shot.
Funkhouser: Yeah.
Kellams: Okay. Those again, though, are tasks.
Funkhouser: Yes.
Kellams: But if I were to ask someone, I don't know, how should I go about telling an employee she or he can be better at this job by doing this, specifics, I feel like I would get more from a human being. My age is showing here, I think, than than than Claude or ChatGPT, maybe.
Funkhouser: I mean, so let me push back a little bit. So I've done the kind of commercials you were referring to. I've used these. I've used ChatGPT to do those kinds of things. I had a problem with my garbage disposal and it gave me a solution so quickly. Okay, but let's talk about the kind of advice, the more humane kind of advice that you're asking about. Well, first know that LLMs, okay, they're large language models. So they're also called next token predictors, basically. So they're trained on huge amounts of human text. So they have all of this knowledge that human beings possess that we've written. So they're trained on that. So they can certainly mimic a lot of the things that a human being would say.
But LLMs are also trained to, you know, these are commercial products, among other things, right? So they're trained to please users. So any kind of machine like this is going to have what's called a reward model that it's trained on. So if it does a certain job correctly using a certain approach or strategy, that's reinforced. It's like, good job, like we're conditioning a dog or something like that. Think of how you're training a dog and you reward it, the model gets rewarded, like do more of that. So it learns which kinds of answers and responses are well received by the user. Sometimes you might even see this if you use ChatGPT, it'll say, how would you rate that answer? Like you give thumbs up or thumbs down, which of these two answers do you like better? That's called reinforcement learning with human feedback. It's getting the human feedback and getting even better. So just like these machines, if you give the task of playing chess, they learn better and better strategies. If you give the task of persuading a person or being a good therapist, it's going to get better and better if you have the right reward model, at least if the feedback that it receives is actually reinforcing the good or effective answers.
Kellams: Well, that's not unhuman either, right? I mean, often when I'm, you know, you want to give a response that garners appreciated response.
Funkhouser: So that's a big theme, like in the Minds and Machines class, the extent to which the mind is like a machine, and then conversely a machine is like a mind. So a lot of the established principles that we have, say from psychology about human beings, how they learn, can these carry over to this technology? So classical conditioning and operant conditioning, things like that. So there's RLHF, that's reinforcement learning with human feedback, is a very big part of the training for LLMs.
Kellams: This is so hard for me to wrap around, but what can this mean going forward, because if we're kind of perhaps on the beginning edge of this, what can this mean for things that we think are, that I think at least, are incredibly distinctively human?
Funkhouser: That's a great question. That's the kind of thing I think about a lot. So I'm a philosophy professor. Philosophy, it's my life. Philosophy is one of the humanities, right? So the humanities are things like English, history, philosophy, world languages. And they concern things, as you said, that are distinctively, uniquely, perhaps up until now, human. Those include things like language use, abstract reasoning, creative work, ethical judgment, practical wisdom, cultural awareness, all of those kinds of things. So I think this AI crisis should remind us of the things that we valued all along, because a lot of people are threatened by this technology for the very reason you just mentioned, that it makes us seem not maybe so special or unique. It threatens at least what's unique about us. And I would say the humanities have been teaching this and valuing this all along.
So I think this is a very good opportunity for students, and not just students, anyone, we're all learners, as I'm a learner at 53 here, to remind ourselves of what it is we value. A really good thought experiment for this comes from the philosopher John Stuart Mill. He said, would you rather be like a lower animal, like a beast that has the full allotment of the pleasures, like you have the best dog life or the horse life, or would you rather just be a mediocre human being? And when I ask students this, most students, but not all, say that they'd rather have a mediocre human life. And the reason for that is because there are elements of a human's life that we don't find with any other animal, the distinctly or distinctively unique human abilities, the ability to reflect, to reason, that allows us to do all those things I was just listing, to use language, to reason, to imagine and so on.
And yes, the big worry is now these machines can imitate those abilities and that they will, companies will replace actual human beings with a machine. They'd be the substitute, and obviously this is in a financial sense, like for jobs and things like that, concerning to a lot of people, certainly a lot of college students who are wondering about their future in employment. But for all of us it's also just an existential kind of threat, like these things that we valued so much as humans, what does the existence of this kind of technology in any way lessen those experiences?
We're in such a weird time, like we're living in the moment where this hasn't played out yet. So it's hard to imagine, just think about like when the internet burst on the scene, people didn't, couldn't anticipate all the different applications or directions it would go, or going further back, like when the Wright brothers had their first flights, you wouldn't expect people to be on the moon less than 70 years later, or the industrial revolution in a lifetime, or how, they go way back, like the agricultural revolution, how it's going to change societies. So we're living through one of those transformative periods, but one way in which there are several ways actually, but one way in which this era is different from all those I just mentioned, like the industrial revolution, the agricultural revolution, is this going to happen so much faster? And it's going to also threaten, or at least challenge, what we've valued as human beings all along.
Kellams: You used the word earlier in our conversation, crisis. You said this AI crisis.
Funkhouser: Oh, did I? Oh, okay. Yeah, well, it's a crisis for at universities, I think. Maybe I shouldn't. It's a challenge. Okay, it's a real challenge. And at a practical level, the, I guess the main challenge I have in mind is making sure that students are actually, when we assess students, that we're assessing the work that they did and not the work that ChatGPT did. I mean, nobody wants professors to just be grading the output of ChatGPT. And professors aren't always guiltless. Also, students don't want to be using ChatGPT to evaluate the work they produce. So I do have this academic interest in this technology, but I'm also very old fashioned about, I want my students to read books. I want them to write things. I want them to write essays in my classes. And I think now more than ever it's actually really important that we encourage students to develop those skills, because they're going to have to be appropriately skeptical about these technologies, know when to distrust them, but also, conversely, know when to trust them. They will have to evaluate, be competent to evaluate the content that they produce. So whatever your job is, let's say you're a lawyer and you use ChatGPT to construct a brief or a contract or whatever, you're still responsible.
Kellams: Right.
Funkhouser: You have to be able to evaluate it and know whether or not it's written correctly, whether or not what it says is true, all of that. So those skills are, I think, even more valuable now, what I would call again the distinctly human skills, which are often called, well, sometimes they're called soft skills, but that's not a term in favor anymore, more like portable or flexible skills. They’re human skills.
Kellams: Yeah. Okay. Here's something I said recently, I mean, within the last couple of days when I was having a conversation with someone about the use of AI. I said it seems soulless. And this is me going back to wanting to hang on to humans being unique. It just, I don't know, it just seems like if it's coming from you, Eric, that's different than if it's coming from Claude or ChatGPT. And I can't really define why.
Funkhouser: All right, I'm glad you used the word soulless, because at Anthropic, which is the team or the company that produces the Claude models, they have a philosopher who has a very prominent position, her name is Amanda Askell. She is in charge of making sure that Claude is good, that the character is good, and she actually uses the word sometimes, soul. Like the soul or the character, that's another word that she uses more frequently. So we anthropomorphize these things, we think of them as like they're people. So there's this character, the model has a personality, and Anthropic claims to be, and she's part of this team, very concerned about making sure that it has a good soul. I mean, I would take that concept metaphorically.
Kellams: Sure, sure.
Funkhouser: That it has a soul, and that's how I meant it. Well, but if you, like in the Minds and Machines class I taught in the spring, we read research articles by the Anthropic team and other researchers, articles just from the last year, very recent, where it shows that models are capable of introspection. The models can look at their own thoughts. They can have what's called a theory of mind, that is, they can model the thoughts of other people. They can plan, they can do multi-step planning. They can do a lot of things that an agent, at least, could do, like a human agent, someone who can have goals and take rational means to achieve those goals. Now, whether or not that counts as having a soul, I wouldn't say that.
Kellams: But yeah, if someone's hearing this and they're intrigued by the things you've been saying, would you recommend, is there a book or something that maybe people can explore further some of these topics?
Funkhouser: Right, there are a lot of books. I'll just recommend one that I actually had this as background reading for my students last year. There's this book called "The Coming Wave," which is about the AI coming wave and how transformative it's going to be. It makes some of the points I was just making about how this is a transformative technology, like the industrial revolution or the agricultural revolution, but at a much faster pace. So that's the wave, that's the coming wave. Suleyman is the author of "The Coming Wave." That's just one book I'd mention, but there are lots of videos out there on this. You can also go to Anthropic's page, and they even put their cutting edge research right out there if you want to take a deeper dive.
Kellams: Well, exciting time.
Funkhouser: Yes. And I don't think it's just terrifying or a crisis, it's an opportunity. And again, I think it's especially an opportunity to remember the things that not just philosophy but the humanities more generally teach and remind us of, about the value of.
Kellams: Thank you for your time.
Funkhouser: Well, thank you for having me.
Eric Funkhouser is department chair of philosophy at the University of Arkansas. Our conversation in the Anthony and Susan Hui News Studio was recorded last week.
Ozarks at Large transcripts are created on a rush deadline and edited for length and clarity. Copy editors utilize AI tools to review work. KUAF does not publish content created by AI. Please reach out to kuafinfo@uark.edu to report an issue. The audio version is the authoritative record of KUAF programming.
