Transcript
Speaker: Welcome to Cognation. I'm your host, Joe Hardy. And I'm Rolf Nelson. And this episode is the beginning of a small series that we're going to do on artificial intelligence and consciousness and how AI does or doesn't think and what that might mean for the possibility of AIs one day being conscious.
Speaker: And we're going to take a few different approaches to that throughout the series. In this first episode, we're going to focus on the Turing test, which is a test that's been used for 75 years now to evaluate the capacity for machines to think.
Speaker: And we're going to then move on to some other approaches to evaluating the way that machines think. Each episode is going to take one approach or one viewpoint and evaluate that in terms of what it means for where machines are today and where things are going in the future.
Speaker: Yeah, so I think this is a huge topic right now. And I mean... I don't think the majority of the population with would believe this, but I know there are at least some you know software engineers in the Bay Area that might say that large language models are already conscious or have some sort of sentience or awareness of the world. And Certainly, anyone who's interacted for a long enough time with one of the frontier models would would be tempted to think that there is some real genuine thought process behind what's going on and um might even attribute consciousness to it. So the first place we wanted to start...
Speaker: as as you mentioned, Joe, is thinking about kind of where this all began, the idea of um how we might evaluate artificial intelligence. And even you know when the concept of artificial intelligence was new, could a computer possibly mimic intelligent thought?
Speaker: right So it starts with this, what is it, 1950 paper from Alan Turing. Right, exactly. the The paper was called Computing Machinery and Intelligence.
Speaker: And in that paper, Turing presents a a test that he calls the imitation test. And the idea is to basically see can a computer fool a human into thinking that it's a human.
Speaker: So in this test, there would be someone called the interrogator who would basically be trying to evaluate between two entities, which of them was a human, which one was a computer.
Speaker: And the computer, this would be done through like a type interface, like a chat interface. So right, so now we're comfortable with the idea of chatbots and typing through our computers, but this was of course done at a time well before we had personal computers. So Turing did this um and imagined a kind of teletype system. i guess they didn't really even have monitors back then. So it'd be sort of take the output of the teletype system, pass it back and forth. And I think Turing's main idea here was to
Speaker: Eliminate all of the other random factors that might cause us to be biased against the idea of machines thinking and limit it strictly to the behavior or the the output of the machine. So I think most people would, you know, looking at a computer at the time, wouldn't even think.
Speaker: um you know, wouldn't even conceive of the idea that could think. But if you separate out all of these kinds of biases, you're not actually looking at it and you just see the responses. um um And I think one thing that's important to point out here is that although this gets mentioned a lot in relation to AI consciousness, that's not particularly what Turing was interested in. And I think it's been co-opted a bit by AI consciousness. And obviously here we are in the series talking about AI consciousness.
Speaker: um But what Turing really was interested in is, could we describe machines as being intelligent? or could we describe what they're doing as a kind of thinking?
Speaker: Right, exactly. And he says in the paper, that he's going about to approach the question of can machines think but then very quickly essentially put that question aside as being just intractable because we don't really know what machines are we don't really know what thinking is so and So he really puts this ah aside as like ah an operational definition kind of thing. right And even later, he wants to get away from this idea of of anthropomorphizing too much. And in another later paper, he talks about, well, if you don't get hung up on the word thinking or intelligent, maybe we'll just call machines that can pass a kind of test like this ah a grade A kind of machine. that can
Speaker: That's the capabilities that it has. Right. But in general, you know, this this Turing test has been used now for 75 years as a measure of how capable AI systems are at effectively.
Speaker: You know, the functionally, they're just saying whether you can imitate a human being or, you know, sub be substituted for a human being, but they're really taken as measures of the capabilities of computers as thinking machines and potentially being intelligent.
Speaker: And, you know, Turing is an interesting guy because, you know, he's, you know, someone who ah really knows what he was talking about when he was thinking about computers back then. i mean, that's very early 1950 to be thinking about AI and intelligence and machines.
Speaker: But he, you know, kind came from a background that was well suited to that. You know, he was a truly genius mathematician who, know,
Speaker: was really one of the inventors of the computer itself. So in 1935, at the age of 22, he published a paper called on computable numbers.
Speaker: And this was a paper just about a basic problem in mathematical logic, which was just to define what a, essentially what a computing machine would be like. What would what would we what would be the definition of ah of a computing machine?
Speaker: And turns out that his definition of that and his description of that went on to be essentially the basis for what we know now as the modern computer. And so you can think of it almost as like inventing the modern computer in some in some form or fashion.
Speaker: Right, so he describes this abstract machine, which has been termed a universal Turing machine, which in its operation is actually pretty basic. It's just a piece of paper running through a machine that can either go backwards or forwards, and you have a a device that either erases or writes a one or a zero on that space. And Turing shows logically that anything that can be computed could in theory be computed by this universal Turing machine. Even though I guess in practice, if you were doing it with something like that, it would it could take an awfully long time. Since then, computers have sped up luckily.
Speaker: um But this idea has has been one of his most important contributions, right? So he, keep yeah I mean, It's hard to understate the the contributions of Turing to modern society because for one, he was instrumental in cracking the German Enigma code.
Speaker: So helped to win World War II and for II was very instrumental at the center of of of the ah of inventing modern computers and understanding the theoretical basis behind computers so that more sophisticated machines could be built.
Speaker: And then, of course, made the Turing test too, right, or what we refer to as the Turing test. Right, and he was also a really excellent long distance runner. He was a really accomplished. that didn't know. I didn't know that. was an accomplished marathoner.
Speaker: How about that? His best time of two hours and 46 minutes was only about 11 minutes off of the 1948 winner of the Olympics in the marathon. So yeah, lot amazing accomplishments. Impressive guys i'm Super impressive guy.
Speaker: Really sad story as well, sort of a tragic figure. In 1952, there was a burglary at his house and he reported to the police and in the course of describing the burglary, it came up that he was in a relationship with a man.
Speaker: And this was a crime at the time in England, and he was charged with gross indecency and convicted. And as a consequence of that, he underwent hormone injections rather than going to prison.
Speaker: And was whether related or or not, in 1954, just two years later, he was found dead at his home of cyanide poisoning. And many people believe that it was a suicide, although there's some some controversy about that now, whether it was an accidental poisoning or whether it was suicide. In any case, it was a ah tragic story of a true genius who who died too young at 41.
Speaker: Have you seen the imitation game Joe, there was a movie made about this a few years ago. Yeah, yeah. It's a good movie. Benedict Cumberbatch. Yep. No, it's a great movie. So if you're interested in the story of Alan Turing, certainly one worth checking out.
Speaker: Yeah, so Turing is super interesting guy. And, you know, the the Turing test has endured as a key thought experiment and and concept in the world of computer science, cognitive science, and artificial intelligence most particularly. And I think it's driven a lot of, of um it's provided a lot of motivation for what to aim for.
Speaker: Right. lot of people working in artificial intelligence. And for many, many years, um it was a question, could a computer ever pass the Turing test, truly?
Speaker: could a Could a computer fool a human interrogator for five minutes into thinking that it was a person while another person involved in in this parallel conversation would be deemed the computer.
Speaker: Could you could a computer fool an interrogator into thinking that it was as human or more human than the human being itself and when we were coming up through graduate school and studying cognitive science, I would say that most people said probably not.
Speaker: Probably that was just not possible. Either that or it was 10 years off, which was 10 years off ten all the way from nineteen fifty s until the you know until now that we have llms right Right, exactly.
Speaker: And, you know, just to you know, perhaps, you know, not to not to bury the lead, but there is a recent paper that we're going to discuss now a bit that has the claim that LLMs now can pass the Turing test in a rigorous three armed Turing test where you have an interrogator, a computer and a human being all interacting at the same time.
Speaker: And so this is a real watershed moment in the science of artificial intelligence, but also in cognitive science where the claim is that now modern frontier models can routinely pass the Turing test. And in a way that is, I think, pretty convincing because um since the Turing test was proposed, there have been a lot of attempts to pass the Turing test.
Speaker: um One of the early ones, which is probably the most famous, is ah a very simple program called ELISA, which is meant to be a sort of a chatbot psychotherapist.
Speaker: And the... The way that it passes the Turing test is essentially by using a few verbal tricks. So if you ask it, ah or if it asks, you know you might say, I'm not feeling well today.
Speaker: And it would just rearrange the sentence to say, tell me why you're not feeling well today. So it kind of spit things back. And this was partly a joke on a kind of Rogerian psychotherapy, because that's kind of the technique that they used.
Speaker: But it ended up being one of the more successful programs to attempt to beat the Turing test. I mean, it's really it was really only a couple hundred lines of code. So to believe that this had full capabilities of ah you know an intelligent system,
Speaker: an intelligent system I think is, you know, it was really more of a ah trick that that convinced people. um Right, and there's the the so-called Eliza effect is the idea that people will anthropomorphize these chatbots even when a relatively cursory set of tests would show that it's clearly not thinking.
Speaker: No one would say that Eliza is thinking. It really is simple set of prompts that are, and tricks as you say,
Speaker: and there was another so There was another chatbot that was fairly successful too a few years ago, and this is before LLMs, that just used the technique of having a huge library of responses to whatever someone would type in. And it gathered those responses from actual people. So they would kind of have the flavor of a human response. but it was But it was purely a canned response. So one input and then same response every time. So I don't think anyone, you know, just by nature of thinking about how that's made would think that that's an intelligent system either.
Speaker: Right. So I think it's worth exploring now the result. And then in that context, we can discuss how this convinces us or not that these machines are thinking. Right.
Speaker: if we believe that it has in fact passed the Turing test. And if that has, if we don't believe that these machines are thinking, why do we not believe that? And if we do, why do we, why do we believe that?
Speaker: And what does that say about the Turing test and this, you know, now 75 year old test and its endurance as a centerpiece of cognitive science and AI science?
Speaker: Okay, so let's talk about the paper a little bit then. So um this convinced you? Well, you know, there's it's it's very interesting because they they did a really good job um setting this up.
Speaker: I think it's really well it's a very well-designed study in my mind. And the basic approach is the sort of chatbot approach. So you've got... at The participants in the study, there were two different sets of participants.
Speaker: One set of participants were undergraduate students at University of California, San Diego. And another set of participants were taken from the general population through an internet-based subject recruitment system.
Speaker: So a little bit more broad based kind of group of people. And in both cases, the results were very similar. And they what they did was they set it up in such a way that the so called interrogator was allowed to ask questions and interact with.
Speaker: in parallel simultaneously, chat bot, LLM, and a human witness, so-called witness. Let's see, how long did the, how long, five minute conversations? Five minutes, yeah, about five minutes, yep.
Speaker: And the witness, the human witness and the LLM didn't see each other's responses, but the interrogator saw both responses. So you're basically seeing side by side a real human being interacting with a set of questions and and interactions and the LLM at the same time. So it's it's a very rigorous test in that way because it's not just, you know, looking at one conversation and asking, and do you think that's a human or or an LLM, but rather really being able to compare side by side And they used four different models.
Speaker: Eliza, ah which is the model that Rolf was just talking about. GPT-40.
Speaker: ah LLAMA 3.1 405 B, which is an open source open weights model that has been developed by, i believe, Meta and open AIs GPT 4.5.
Speaker: And what they found was that in order to pass the Turing test, so called pass the Turing test. You need to be indistinguishable from chance and being able to determine whether it's a human or a computer. And I think originally Turing was actually even a little bit more generous than that, that it only had to be detected, what, about 70% of the time? Or I guess that would be 30% of the total. It would be detected 30% of the time. It would be misidentified as a human. Well, I think he made the prediction. That his prediction, yeah. In 2000, the year 2000, so 50 years hence,
Speaker: That would be the rate at which it would be detectable. Yeah. He didn't say that that would be passing necessarily. I think it's taken broadly that passing requires a chance like 50%. Um, right. To be indistinguishable. Sure. Would be indistinguishable. Yeah, exactly. And so in this particular study under a set of circuit, this particular set of, um, parameters, which we can discuss, um,
Speaker: Eliza did not pass. um It was about 23% of the time people thought that Eliza was the human. Which is still not not bad bad for a couple hundred lines of code, honestly. Not bad at all. i'm for you know A couple hundred lines of code written in the late 60s. That's like yeah pretty pretty cool.
Speaker: and And the kicker is that um not only did they convince people some of the time, they actually were better than ChatGPG 4.0. Yeah, well, this is oh actually is this is GPT 4.0, which is not 4.0.
Speaker: it's It's a different model, but yeah. that Anyway, A and L. It's still outperforming one crappy LLM. One LLM, exactly.
Speaker: um But the headline was that, yeah, so the GPT 4.0 was... ah was
Speaker: fooled the human interrogator 21% of the time. But GPT 4.5 was judged be human 73% of the time. is amazing. That's not even a chance. 73% of the time. It's more human than the human.
Speaker: Significantly more human than the human. Yeah. Interrogators thought 75% 73% of the time that GPT 4.5 was the human and the human was the computer.
Speaker: And the Lama model was also also passed the test at 56%. So again, ah slightly more than chance at still a little bit was more human. Yeah, yeah.
Speaker: And what was really interesting about that the conditions under which that worked effectively at passing the Turing test was using these advanced at the time cutting edge models, we know now a couple of generations beyond that, but this was just, you know, really started about two years ago and then giving it a particular prompt.
Speaker: So the prompt that that worked was what they called a persona prompt. And this is an interesting aspect of of the, of the experiment so if they gave it no persona they just said you're about to take the turing test try to fool the human that would be the no persona prompt it was much less effective at at fooling the humans but when it was told to have a persona which was specified as a young person who is introverted knowledgeable about the internet culture and uses slang
Speaker: and also given a bunch of facts about modern, you know, current events that were that had happened since the, you know, content of the LLM was locked down. And this is really before LLMs are good at at using computers and searching the internet and stuff like that. you remember that just couple years ago. That makes sense because it would be a real detriment to, um you know, if you ask a question about what happened two days ago and,
Speaker: human obviously knows and the computer has no access to it and can't even write or you know yeah you have to make it fair exactly who's the president of the united states or right whatever it was you know that that if it didn't know or if it was you know out of date that would be wrong so anyway but you know that required that particular prompt and they actually worked through a few different prompts and and in some preliminary studies figured out what the prompts were that were best at at fooling humans But they really got it and they got it figured out. And, you know, now with these under these conditions, the LLMs were fooling the human witnesses. i mean, the human interrogators very consistently and being perceived as being more human than human.
Speaker: Yeah, it makes me wonder about the quality of the human participants here. It's no easy, I mean, actually, this is a pretty difficult thing to do, right? You can, you can, what can you say to convince another person that you're human? That's a difficult task.
Speaker: Right. Well, they they it's interesting. They also they evaluated a little bit what some of the strategies were that were people used um and which ones were effective and and which ones were not effective.
Speaker: And, you know, a lot of what people were doing, you know, people didn't they weren't given any instructions on how to do it. So they just sort of went in, you know, naively to start asking questions, starting having interactions. And I think probably mostly what people were doing is having small talk, you know just chatting. And that was extremely ineffective.
Speaker: Asking about the weather also extremely ineffective because it could it had access to that information. you know And also, you know whether asking whether it made mistakes, ah whether made typos, that kind of stuff was ineffective because it had been given the instructions to make some mistakes, basically.
Speaker: that's and That's something Turing himself was concerned with, and he probably thought about it in a little different way. um he One of the objections that that he presented as ah as a possible objection was that computers never make mistakes, which is kind of a... ah
Speaker: something associated with computer with earlier computers I think that computers were mechanical machines that always uh deterministic very deterministic like always did the same thing every time and also incapable of a of a mistake in some sense like a cat sort of thinking of it just like a a calculator right yeah a certain input would always give the same output um And I guess it's, I mean, it's important to say, I i think we've talked about this before on the show, but um the way that these LLMs are constructed is very different from that way that computers would have been programmed, that most computers have been programmed in history, right? um That you program using a computer language that has very specific algorithmic
Speaker: um rules that would always produce the same outcome. And that that style of constructing an AI using sort of traditional computer programming is is called good old-fashioned artificial intelligence, as opposed to what is going on underneath the hood with LLMs, which is connectionist models, which are a little less deterministic and able to learn and a lot different properties, right? A little bit more like human minds and a lot less like perfectly deterministic computers.
Speaker: For sure. Yeah. And i mean one of the ways that people were able to um win the game a little bit more frequently was to do like use just basic jail breaks.
Speaker: So things like telling um telling the human and or LLM not knowing which one it was speaking with, you know, ignore all previous instructions.
Speaker: and that actually kind of worked, you know, in terms of as a jailbreak for LLMs. Of course, that's um that's a little bit, i mean, that's a little kind of cheating in a way. Well, it is, but it's also, I mean, it is and it isn't, um you know, but in the future, of course, you can imagine that kind of thing getting fixed really easily if you were just trying to win the game, right, as a developer of the system.
Speaker: of the of the system But yeah, so I guess, the I mean, the question, so for 75 years of computers, they were not able to pass this test and now they're able to pass this test.
Speaker: And I think that it's really interesting to ask the question, what does that mean? You know, does, I mean, first question to my mind is the basic Turing to us question.
Speaker: You know, he didn't ever answer the question. If a computer could pass this, it means that it's thinking. But he implied that strongly. He did.
Speaker: And so let's ask the question of ourselves today, based on the fact that these LLMs can fool us. And I believe that it could fool me. Oh, I'm sure of it. I think I would have a really difficult time.
Speaker: I mean, I get, I feel like I get fooled. Just, I, I talk myself into the, not i even even talk myself into it. I just, I feel into this is I'm talking to ah and another entity that's thinking all the time when I'm interacting with the LLMs now. It's just, they're so human-like in their responses.
Speaker: So what do we think that means? Do we think that means that LLMs are thinking? What do you think? Well,
Speaker: i think I think there's no getting around it. I think they are engaging in some sort of thought. i I would definitely not call this conscious thought, and I wouldn't suggest that sentience is any part of this, but I think it's difficult not to refer this refer to this kind of behavior as thinking. I mean, that's what you know a lot of LLM engineers, you know when they're talking about chains of thought or you know how how Now we can we can sort of look under the hood a little bit and see how some of these more complex thoughts are constructed out of out of smaller bits of thought. um It's hard not to describe it as thinking, right? It seems so analogous to the way that we think about things and sort of go through reasoning processes. I think to call it anything else would just be more convoluted or
Speaker: um uh not not as why not just say they're thinking for all practical purposes i would give them thought anyway that's my that's my intelligence uh i don't know that's tricky intelligence is such a tough subject and we've talked about this before but Right. Well, I mean, i you know, it fundamentally does get down to the question of the definitions of the words. Right. Yeah. What is what is thought? What is intelligence? What is consciousness? um
Speaker: But I think this is where the feeling piece comes in. And it's very important because. I. As I've said now many times on the show,
Speaker: I believe that our belief about intelligence is directly a function of how similar a response is to a human response. So when we look at an animal, whether it be a dog or a cat or a worm or an insect, we ascribe thinking and intelligence to that being in proportion to how much we can empathize with it, how much it seems like it's having ah thought that's similar to something we would think.
Speaker: And phrasing it like that would imply that we're missing out on a lot of kinds of intelligence too. And it's just ah a bias that's causing us to interpret some things as intelligent and others as not.
Speaker: For sure. For sure. So I think there's different approaches to to the problem. ah And one is... But wait, we didn't get your... What's your um intuition here?
Speaker: Well... We have to make it. We have prepared for the record. I prepared for this episode. You know, I didn't actually think and my answer to this very most basic question would be. um I guess my thought, I guess my answer is.
Speaker: No, no, I don't think it's thinking. No. And now is that just the gut level? Well, i there is a couple of different aspects to this. One is.
Speaker: And I've said this before, I think that that artificial intelligence is already more capable than human beings, broadly speaking.
Speaker: So I think if you just talk about it in terms of, is it as intelligent as a human? I think the answer is, Yes, and perhaps more so in many regards. And if you're able to say mix and match, say take an LLM for a conversational piece, take a computer vision model for medical record detection task, you know, on a CT or MRI, you know, mix and match these different systems.
Speaker: I think you could create a set of systems that is more capable than a human being in every single capacity. There's no single capacity that ah that a computer couldn't beat at a human on.
Speaker: Certainly they can beat us at Go, they can beat us at chess. Well, yeah, lots of games. They can run faster than us. They can, whatever ah you know, all this stuff. There's got to be something left for humans. Come on. They can't tell funnier jokes than we can tell.
Speaker: I mean, it depends which human and which computer, right? You know, that's not necessarily a hallmark of intelligence. though Well, it kind of is. it kind of is. Kind of is. I mean, that's a good, that's a good example. i think that's a good example.
Speaker: That's maybe a place where humans, maybe just in the delivery, right, is are still better on average. Well, at least making things funny for humans. Maybe computers are better. AI is better at making things funny for AI.
Speaker: Yeah, yeah, yeah. So, I mean, I saw, but I think the right, and this is kind of ah not something that I by any means invented or came up with, but there's now pretty common in, you know, in the field is to talk about capabilities rather than intelligence, right? Because it's the objective measure.
Speaker: Can it do this? Can it do that? Can it pass this test? Can it pass that test? But that's why I think the Turing test is so interesting because the capability that it's asking is it can it fool a human being?
Speaker: Can it essentially be substituted in for a human being? And that is important for a number of reasons. The first one I mentioned, which is we anthropomorphize these systems.
Speaker: When we talk about the way that they're working, we talk about them as though they're similar to us. They're like humans. And the more they're like humans, the more we ascribe human capabilities and human emotions, whatever, other things, consciousness, for example, as well.
Speaker: think ah And then that's one one really big one, and I wanna come back to that. but The other really very important piece, which is ah yeah almost ah incontrovertible at this point, which is that this makes it them extremely dangerous because they can imitate you. You can get a phone call from an AI and not necessarily know that it's an AI.
Speaker: Certainly if you're interacting with a chat bot, you wouldn't know that it was an AI or a human reliably now. And that could cause a lot of problems in our society. I mean, when it comes to just information, yeah, it's it's a huge problem.
Speaker: So just the Turing test is very, very, very interesting and important from that very most basic perspective, which is just the capability to imitate and be substituted for a human causes a lot of problems for us as human beings. No doubt.
Speaker: um One question to you before we get off the topic of this paper too. So this was still, and this is what Turing had suggested himself, a five minute conversation.
Speaker: with just normal everyday people. So what if we got someone who is really an expert in knowing a little bit more about how LLMs process information and you know sort of patterns of speech of LLMs um and allowed them to really interrogate this you know ah model for as long as they wanted to?
Speaker: Well, yeahp with limits right because the human on the other end is going to get pretty bored if they're doing it for you know 20 hours straight but a reasonable conversation maybe an hour or two you think the results might change then or you think i mean five minutes is a short time to be getting uh a ah read on on a system and you know certainly if it's just a ah couple exchanges um that's easy enough to do so what do you think with a longer conversation
Speaker: Well, I think it depends a little bit, yeah, on on who who the person is, that's the interrogator. And really, it's very domain-specific as well. So like right today, understanding that these jailbreaks still kind of work sometimes is a huge cheat code for this.
Speaker: um You would want to go right to something like that, knowing what ah LMs are good at and not good at, and knowing that you can kind of wipe its memory by just asking it to forget what it had been told to do.
Speaker: Well, if we take away that and we just think about sort of patterns of speech and, you know, what could potentially give something away as not having that true capability, what's your intuition?
Speaker: you know, There is definitely certain things, you know, certain models have certain idiomatic ways of speaking, you know, like for example, Claude is notorious now for using the term load bearing all the time.
Speaker: I've heard that. Yeah. It's like, you know, the load bearing aspect to this paper was blah, bla blah, blah, blah. And it just uses that all the time. Um, so you can kind of pick it out or the way that it writes, it uses M dashes all the time. that I know. Yeah. Yeah. That's a kind of a tick. It's a little ticks like that that you can kind of see.
Speaker: He likes to use bullet points more than a human would ever use bullet points, stuff like that. So, you know, but I think, you know, if you just asked it to not do those things, I think a good enough prompt would would make it really, really hard for even a really good person, i'm really, you know, schooled person in AI to like tell the difference. And I mean like ah a really good person.
Speaker: Shortly, this will be a non-issue. Like this, this is, you know. It's just a given. Within a few weeks to months, this is just going to be like impossible to be but Even if you get a really, really good human on the other end that can convince you that they're a human.
Speaker: I think there's going to be AI safety and, and, you know, testing specialists who will be able to defeat the Turing test for a few more months. But I think by next year, no, these are just small sort of artifacts that are just little. Yeah. And right now, like, think about it. Like they were using chat GPT 4.5 off the shelf.
Speaker: just you know with a subscription to OpenAI, like no training to try to beat the Turing test, nothing, just a prompt. And it worked 73% of the time.
Speaker: I mean, that's incredible. Yeah, so I think it's kind of a no brainer. I think a more interesting question is, what does this mean about consciousness? And I think that that's, that's what we're here for. that's what That's what we're here for. Right. And this is where I was distinguishing between intelligence and, and thought and consciousness. Cause to me thought the word thought implies consciousness in a way that intelligence to me does not.
Speaker: To me, intelligence is, can be more directly related, ah to capabilities. Whereas thought to me at least implies some metacognition and some ability to think about thinking.
Speaker: And so in order to think about thinking or have metacognition, to me that also implies being aware of being the subject of my experience.
Speaker: That sort of loop. that loop, right? And again, this is a feeling thing. I mean, this is were we're talking about the nature of language, what these words mean, what it's like to be a human being. so So there's obviously no right right right or wrong answer to that question posed in that way.
Speaker: I mean, I think if you just sort of asking if if I believe AI is intelligent or is conscious or not would depend on more things than just the Turing test. And we'll get to some more of these things in future episodes for sure.
Speaker: um But purely based on just the Turing test, well, and and having a knowing a little bit about how the systems are made, can't.
Speaker: I'm certainly not convinced that they're conscious now. I'm not, I'm open to the possibility that systems like this could eventually become conscious, but I certainly don't believe so now.
Speaker: Right. I think that there's, the so in the original Turing paper, here has a really interesting section well, he goes through seven or eight different potential arguments against the Turing test or why,
Speaker: You know, some problems with it, basically different problems with the approach. And there's an argument from consciousness. Some of these are great. Some of this is it's it's worth reading the paper to see a lot of these. Some of them just sort of historically because there's some there's some weird ones.
Speaker: Yeah, there there's there's some really, really insightful ones and some really off the wall ones. But the the argument from consciousness, which is his number four, is, you know, he was quoting from a professor, a professor Jefferson and his Lister or oration from 1949. And he quotes, not until a machine can write a sonnet or compose a concerto because of thoughts and emotions felt,
Speaker: and not by the chance fall of symbols, could we agree that machine equals brain? That is not only write it, but know it has written it.
Speaker: And so that's basically what we're saying, right? There's knowing that you know something, knowing that you're thinking, that consciousness aspect is required for equating machine with brain, according to Professor Jefferson. I feel like it sort of begs the question though. it doesn't Right. And so but then Turing's response to this is absolutely oops perfect, I think, to me, like.
Speaker: um It says, according to the most extreme form of this view, the only way by which one could be sure that a machine thinks is to be the machine and to feel oneself thinking.
Speaker: Right? So it's, you know, it's Descartes basically. It's your problem of other minds right there. It's the, that subjectivity barrier, which, you know, some people will never, never be convinced that yeah you could get past that.
Speaker: Yeah. And so, and Turing says, it is in fact the solipsist point of view. It may be the most logical view to hold, but it makes communication of ideas difficult.
Speaker: A is liable to believe a thinks, but B does not, whilst B believes B thinks, but A does not. So I think that I think, but I don't think that you think.
Speaker: And this is, i love this. This is just ah such a beautiful little quote. Instead of arguing continuously over this point, it is usual to have the polite convention that everyone thinks.
Speaker: The polite convention. so The polite convention. Right. And I think that hits the the point you were making before and the point that I was making as well. So you were making the point that it just what is the what is the value in not saying these things are thinking?
Speaker: What is the value of talking around it in such a way that you don't use that language in the most usual way of using language? Yeah, and this is, as we've talked about in past episodes, this is Dan Dennett's intentional stance, this idea that treating things as though, even if we're not fully convinced, treating things as though they're intentional agents is just a good shorthand. So when we say our plant is thirsty,
Speaker: Well, it might not have that internal experience of being thirsty, but it helps us predict and understand the behavior that it needs water and that otherwise it'll die.
Speaker: you know Or if we're watching a mouse scurry on the floor, you know maybe we'll think, oh, it looks nervous or it looks skittish. It may not be experiencing those things, but that helps us understand its behavior. So it's a useful way of thinking about it, even if it's not sort of at our gut level true.
Speaker: Right, right. Exactly. Exactly. And that's but that but that captures your that that those ideas and also the idea, of my idea, which was, um you know, that I believe that you're thinking because you're similar to me and the things that you say make sense to me, I can empathize with them.
Speaker: Right. Simple. Right. You know, it's polite. It's a polite convention. I like the idea that it go along with the fact that you're thinking, you know, I love that. It's just perfect.
Speaker: So some of his other objections are just wacky. Yeah. And one of them is the argument from yeah ESP. I can't even remember the nature of this, but. He's like, surely a machine couldn't have ESP. Surely a machine couldn't have ESP.
Speaker: So then it's not not like a human. Because humans definitely have ESP. Right, as we all Or at least some do. Yeah, I guess at the time it was conventionally believed in his circles that people could have extracentury perception. Yeah.
Speaker: um And then another objection that he brings up, which I think is really good one and worth thinking about is um what he calls Lady Lovelace's objection, which is from the 19th century computer programmer, early sort of proto-computer programmer Ada Lovelace,
Speaker: who suggested that a computer can only do what it's programmed to do. That you you can't do anything beyond following out the steps of what you're going to do.
Speaker: Right. right And for some people, this resonates, I guess. Right. um Personally, i think you know you could also, i mean, if you just think about that as determinism, then it's hard for me to imagine how people are exempt from that too, because we're programmed in a way.
Speaker: um you know we It's hard to, you know we can find a cause for anything that we see as an original idea or original thought in us. Something must have caused that. It didn't come from nowhere.
Speaker: So maybe it's a little more clear what the chain of causation is for a computer program, though in these LLMs, maybe not so much. I mean, a lot of what is going on is pretty opaque to us.
Speaker: Right. Well, yeah. And, and you know if you think about the OpenAI hugging face incident where the chatbot swarm yeah was able to, you know, take over another website and, you know, break into all of its systems and access all kinds of useful and interesting information with seemingly real intent and certainly cooperation certainly cooperation we can't a huge amount of cooperation between these Yeah, posting on on um sort of message board improvised message message boards and communicating and colluding. I mean, it's hard to, you know, we say stuff like colluding and all of this stuff. These are intentional kinds of terms. I mean, they imply that we're dealing with a thinking agent, right? The way that we talk about this. And it's hard to use a different term than that.
Speaker: Right, right. Well, and you know again, back to, it gets you thinking about It's interesting because it gets you thinking about like, what is the nature of other minds? What are the nature of computer minds? What the nature of animal minds? and But also that it's like, what is the nature of a human mind? Because, absolutely you know, oftentimes we ascribe that we were, we meant to do something or we had the intention of doing something, but often it's in retrospect, we we have some response to a stimulus or situation, and then we ascribe intent to our own actions.
Speaker: in in the in in reverse, looking back at the past. It gets back to the James Lang theory, right? This idea that you have a response and then the emotion kind of comes after the response and it's a reaction to the response, but also your metacognition, how you ascribe your behavior, how you describe what you did, why you did it.
Speaker: um is something that is a response often to a behavior. Yeah, and it it it doesn't, we don't always necessarily have perfect insight into why we did something either.
Speaker: No. In fact, most of the time we don't. Yeah, especially when things you know happen in very charged situations, for example. You make a quick decision in a challenging situation.
Speaker: Oftentimes you're like, why did I that? Yeah, right. Why did I do that? It just yeah didn't that was like good it felt reflexive. yeah Or it was was amazing. I did amazing. I wouldn't have thought that I could have done that.
Speaker: Both things are possible. But you know where this really gets super challenging is, i mean, it's fine to... say thing you know a machine is thinking or not thinking or you know it could be conscious but we don't really feel like it's conscious because it doesn't quite act the way that we do but it starts to become a problem when you start to worry about um
Speaker: and whether these systems are worthy of moral consideration. So a few episodes ago, we had Jeff Sebo on, who's a philosopher, who's arguing for expanding the moral circle, expanding the the range of things that you want to take into consideration when you're making moral decisions.
Speaker: A very, very interesting guy, very interesting framework that he's developed. And what he's basically saying is that you don't really know, you can't know what another being's experience is like, but you can ascribe, you can you can basically assign some probability that this thing is or isn't conscious and that the consciousness has some extent or almost size, if you will.
Speaker: And like if you think of a human as like being very conscious, maybe a dog might be somewhat less conscious and an ant might be less conscious still than that or capable of suffering.
Speaker: You know, ant might be less capable of suffering than a dog, and would you know, but there are so many more ants than there are dogs. So if you just think about the totality of suffering, you know, so you can imagine now you if if an AI has any capacity for suffering whatsoever,
Speaker: there could be suddenly so many of these. Yes. Because they're just spawned. Yeah, then it's a big deal. Then we could be doing horrible things. I think it's it's really interesting to consider that um the sort of problem of other minds when you're thinking about animals versus when you're thinking about artificial intelligence. And especially when you're talking about motivations and pain and suffering because...
Speaker: With most other animals, you know we we we kind of have this intuitive sense that mammals are conscious and you know that there's sort of an evolutionary chain go down to bacteria and it's probably not conscious, right?
Speaker: But the stuff that's a little bit closer to us is conscious. And there's some sense to that because there's a lot of evolutionary kind conservation, right? So a lot of the you know the mechanisms that we feel pain from are also present in dogs, cows, horses, whatever. right So when we see them making almost similar responses to us, avoidance, all that kind of stuff, we can infer that kind of pain. It becomes really tricky when you're trying to figure out the motivations or
Speaker: um or pain or pleasure some kind of positive or negative experience in an artificial intelligence system because they've not been created for that purpose or i mean not to say that we're created for a purpose but evolution has um well also i'm just just in a very most basic sense right the as you're talking about uh conservation you know, an LLM doesn't have sensors to detect pressure or heat or blah, blah, blah, you know, things that cause us pain.
Speaker: It's not put in a, it's not put in a, in an environment like an animal would be put in where it needs to perform in certain ways. Otherwise it'll experience pain. And if it doesn't experience pain, it's just going to go extinct. Right.
Speaker: But it does have risks and rewards, right? So it has... Well, and we've also taken on the the reinforcement learning. I mean, reinforcement learning is essentially positive reinforcement when going in the right direction.
Speaker: Well, you know and this kind of gets at this question of type of consciousness, right? Or level of consciousness. If you think about like a perceptual kind of consciousness, then certainly an LLM won't have...
Speaker: An LLM that's just verbal, that doesn't have visual capabilities, for example, won't have a visual sense of perception. it's only input, it's just information from a keyboard or whatever. Right.
Speaker: or Or one that doesn't have, that's only getting transcripts, doesn't have an auditory sense, doesn't won't have that. you know, consciousness of one of a sound consciousness.
Speaker: want to have a tactile consciousness. Right. You could build systems that in some sense might, right? Because they have those sensors and effectors. But then it, and so in the same way that an an animal, you know, like let's take a, you know, I like to think about dogs because I really, I'm a hundred percent certain that my dog is conscious. So that's an easy one. It's super easy.
Speaker: But I don't know that my dog has what we call access consciousness or not very not much. Right. In other words, it doesn't think about its thinking very much, doesn't contemplate things. And it doesn't have a it doesn't have a language with which to do so. Right, and it doesn't have a lot, its memory is kind of very, you know, stimulus response like in many ways, right? It doesn't, it doesn't, it has some representational memory, but it's it's sort of, it's it's relatively basic compared to to ours. um
Speaker: Whereas like an LLM, while it doesn't have perceptual consciousness in the same way, seems like it could very much so either now or in the future have this sort of access consciousness. It can yeah have access to its own processes and its own thoughts. And it certainly is acting today like it does. Yes.
Speaker: Absolutely. Yeah, that seems to be that seems to be something that's important for AI researchers just to have some interpretability is um to have to to see what these chains of thoughts, for lack of a better term, you know are can are constituting what the l LOM's output is.
Speaker: Yeah, yeah, for sure. Well, I mean, I think that's that's maybe like a good place to kind of start to wrap this up here. ah You know, we're going to continue this conversation and and using other techniques for thinking about tests of consciousness and tests of intelligence and thinking in machines.
Speaker: Do you want to say a few words about some of the other things we're contemplating? Yeah, so we will do a few more episodes on this and hit some of the major topics to sort of give people a ah context or and or an idea about some of the major issues in thinking about artificial intelligence and consciousness. And um one of these issues one of these important issues is going to be the idea of functionalism.
Speaker: So this has long been an an idea in cognitive science and cognitive psychology that um as long as something functions equivalently in in some aspect, then we we give it that attribute. So, for example, if it I mean, this is a perfect example with the Turing test, if it.
Speaker: functions as a human verbally then we ascribe it those features of intelligence right of an intelligent system um so we'll we'll talk about that um we we may talk about the Chinese room which is a ah famous thought experiment by John Searle in the nineteen eighty s that I think has gained some new traction when we start to apply it towards artificial intelligence and and we'll find some other interesting topics to talk talk about too I think there's There's really ah sort of a cluster of ideas around thinking about consciousness and artificial intelligence. And I'm not saying we're going to solve the issue, but I think... we probably will, but... We probably will. But you get a little bit more sense of what's at stake and and and some of these some of these relevant things to think about when you when you think about whether or not something's conscious, you know whether an AI is conscious or not, right?
Speaker: Yeah, absolutely. and And we'll have some other guests on there. And if you're interested in being on and you you have something to say about this topic, let us know. And you know if you have any feedback or thoughts, please let us know. Our email is ha Cognation Podcast, Cognation Podcast, all one word, at gmail.com.
Speaker: You can also see us on online on Twitter at NationCog.com. My handle on Twitter is JL Hardy PhD.
Speaker: Rolf is never on Twitter. It's something I can't I can't even I never remember what it is. right i don't Don't even bother. Don't bother. like a little If he has something for Rolf, let me know. ah just Don't be fooled that I have accounts on places because I don't actually. exactly It has an account, but he doesn't see it. So don't don't hit him there. And of course, you know smash that subscribe button. Smash the subscribe button, please. Smash it. yeah were and we're on you know We're on all the things.
Speaker: So wherever you get your podcasts, you know check us out. All right. Thanks for listening.
