WEBVTT - Part One: How AI Chatbots Became Cult Leaders 0:00:01.720 --> 0:00:08.160 Also media, Welcome back to Behind the Bastards, a podcast 0:00:08.240 --> 0:00:12.600 about the very worst people in all of history. And 0:00:13.200 --> 0:00:17.319 this week, actually our bastard isn't people exactly, although people 0:00:17.360 --> 0:00:19.799 are still at the center of it. But to talk 0:00:19.840 --> 0:00:23.120 about that potentially non human bastard, I'd like to bring 0:00:23.160 --> 0:00:25.759 on someone who I am eighty seven percent sure is 0:00:25.800 --> 0:00:30.360 a human being, Blake Wexler. Blake, Welcome to the show, Robert. 0:00:30.440 --> 0:00:32.199 I'm so excited to be here. Thanks for having me. 0:00:32.200 --> 0:00:34.559 I'm psyched that our bastard this week is limes disease. 0:00:34.840 --> 0:00:36.320 I think that's a fantastic kick. 0:00:36.880 --> 0:00:40.480 Yeah, yeah, it's see it's a real pa we're doing. Yeah, 0:00:40.600 --> 0:00:43.280 we're going after I'm coming after dear tics. This week 0:00:43.400 --> 0:00:45.800 is finally U Yeah my big reveal. 0:00:46.040 --> 0:00:49.319 Yeah, Big Tick doesn't want us to do this episode. Yeah, 0:00:49.320 --> 0:00:50.839 we're exposing all the secrets. 0:00:50.479 --> 0:00:52.040 Big Tick energy. We don't need it. 0:00:52.120 --> 0:00:54.840 If we're going to have like a fascist movement dedicated 0:00:54.880 --> 0:00:57.880 to like victimizing and attacking one segment of the population, 0:00:58.000 --> 0:00:58.840 why couldn't. 0:00:58.560 --> 0:01:00.000 It be deer ticks? Right? 0:01:00.120 --> 0:01:02.120 If our fascists were just going after deer ticks, no 0:01:02.120 --> 0:01:03.120 one would have an issue. 0:01:03.240 --> 0:01:08.360 You know, you're going after the wrong people. Yeah, yeah, yeah. 0:01:08.360 --> 0:01:09.960 If there were just a bunch of maga guys out 0:01:09.959 --> 0:01:12.000 in the woods with knives looking for ticks, just like 0:01:12.000 --> 0:01:12.679 I'm gonna get. 0:01:12.600 --> 0:01:16.040 Them, and they would use knives too to kill the tics, 0:01:17.600 --> 0:01:19.160 heat the knife up to burn it off of you. 0:01:19.319 --> 0:01:23.280 Yeah, our brave soldiers getting lime disease to protect the 0:01:23.280 --> 0:01:33.200 rest of us. So we're not talking about lime disease, 0:01:33.360 --> 0:01:36.959 our bastard. This week in broad is do you remember 0:01:36.959 --> 0:01:39.600 how like about a little less than a year, well, 0:01:39.959 --> 0:01:41.280 a little more than a year, guy, I guess, like 0:01:41.360 --> 0:01:43.840 last summer to early fall, there were suddenly a bunch 0:01:43.880 --> 0:01:47.400 of articles about AI psychosis and about like specific people 0:01:47.440 --> 0:01:50.160 who had either in some cases, committed suicide or murder 0:01:50.680 --> 0:01:54.240 or just kind of lost their minds after becoming weirdly 0:01:54.280 --> 0:01:58.080 attached to their AI chatbot, right, and you often deciding 0:01:58.080 --> 0:02:00.520 that they had it had become sentient, you know, or 0:02:00.560 --> 0:02:02.640 at least that they had discovered it was right. I'm 0:02:02.680 --> 0:02:04.400 sure a lot of people are at least if you 0:02:04.440 --> 0:02:06.840 didn't read the articles, you saw them in your news 0:02:06.840 --> 0:02:08.640 feed and saw people commenting on them. 0:02:08.760 --> 0:02:13.240 Yeah, yeah, yeah, it is as depressing as it gets. Yeah, 0:02:13.440 --> 0:02:14.600 those stories, Yeah. 0:02:14.520 --> 0:02:17.240 Yeah, between those and the people like proposing to their 0:02:17.360 --> 0:02:19.280 chat thoughts, it's got pretty grim. 0:02:19.440 --> 0:02:21.960 Oh god, there's some grim stuff out there, right, and 0:02:21.960 --> 0:02:24.640 it hasn't stopped. Like last summer fall was kind of 0:02:24.680 --> 0:02:28.440 like when like there was a big rush of those articles, right, 0:02:28.480 --> 0:02:31.520 and you know they're still reporting on that now, but 0:02:31.560 --> 0:02:33.359 that's when a lot of it really started to hit. 0:02:33.440 --> 0:02:36.720 And obviously, whenever we talk about AI on these shows, 0:02:36.880 --> 0:02:40.119 AI as it's used now is like a marketing term, right, 0:02:40.200 --> 0:02:42.400 and it's used to refer to basically every product of 0:02:42.480 --> 0:02:45.040 machine learning technology. And the reason why the industry has 0:02:45.080 --> 0:02:48.560 done this is because that way, if you say i AI, 0:02:48.639 --> 0:02:50.880 they'll be like, oh, so you hate like your maps app? 0:02:51.320 --> 0:02:53.680 And because that's machine learning, right, all of our different 0:02:53.720 --> 0:02:56.000 like map programs involved that, or like oh you don't 0:02:56.040 --> 0:02:59.840 like using you know, autocomplete or whatever, and it's like, well, no, 0:03:00.120 --> 0:03:03.960 he was calling maps artificial intelligence. In twenty ten, you know, 0:03:04.040 --> 0:03:06.800 when when smartphones started to become ubiquitous, we're just like, 0:03:06.840 --> 0:03:09.360 oh cool, I have a navigation app on my phone. Now, 0:03:09.800 --> 0:03:12.919 like you're kind of trying to siphon the goodwill from 0:03:13.000 --> 0:03:15.799 those in order to get us to like these chatbots. 0:03:15.960 --> 0:03:18.120 I hate the chatbot that I fell in love with 0:03:18.200 --> 0:03:20.560 who doesn't return the flight of feelings starts me. 0:03:20.800 --> 0:03:24.280 That's who I hate. Yeah, not always, that's who I hate, right, 0:03:24.639 --> 0:03:27.400 And the reality is that, like using the term intelligence 0:03:27.480 --> 0:03:30.760 even for these chat GBT and stuff like, there's a 0:03:30.800 --> 0:03:32.840 lot of debate as to whether or not that's a 0:03:32.919 --> 0:03:36.280 good idea, right, depending on how you you how you 0:03:36.360 --> 0:03:40.280 define intelligence. You can either say, obviously, these aren't intelligent 0:03:40.600 --> 0:03:43.920 because like they're not independent thinking things. They don't do 0:03:44.000 --> 0:03:47.360 anything for themselves, they don't want anything, that don't have motivations. 0:03:47.880 --> 0:03:51.280 They're just tools that can be utilized by human beings 0:03:51.280 --> 0:03:55.760 to provide certain answers or take certain actions. Right, I 0:03:55.760 --> 0:03:58.400 don't know if it can't. It's the it's my issue 0:03:58.440 --> 0:04:00.920 with like AI bots create art. If it can't like 0:04:00.960 --> 0:04:03.640 be horny and it can't be like angry and weird, 0:04:03.800 --> 0:04:06.280 it can't make art. Right, those are I think fundamental 0:04:06.280 --> 0:04:06.880 issues I have. 0:04:07.120 --> 0:04:09.680 I could three of those things. Angry and weird. 0:04:09.720 --> 0:04:13.760 Are yeah, horny and angry sharing on Horney. 0:04:13.720 --> 0:04:13.920 Yeah. 0:04:15.800 --> 0:04:18.400 So you know, as I noted, over the last year, 0:04:18.440 --> 0:04:21.239 there've been an increasing number of stories about people using 0:04:21.400 --> 0:04:24.039 these different chat pots. So coming to what's often called 0:04:24.160 --> 0:04:27.200 AI psychosis, and that's not a recognized medical term at 0:04:27.200 --> 0:04:29.239 this point, right, but it is a blanket one people 0:04:29.240 --> 0:04:31.200 have started to apply for the ways in which folks 0:04:31.279 --> 0:04:33.840 are getting addicted to using chatbots, which then tend to 0:04:33.880 --> 0:04:36.760 trap them in these recursive patterns of thinking that can 0:04:36.800 --> 0:04:39.440 push people who are vulnerable to adopt views that are 0:04:39.440 --> 0:04:42.359 increasingly detached from reality. And this has resulted in a 0:04:42.400 --> 0:04:45.120 few cases in severe injury and death. And in all 0:04:45.160 --> 0:04:48.800 of these instances, the LM, the chatbot is just responding 0:04:48.839 --> 0:04:51.400 to the input that it receives, but it tends to 0:04:51.440 --> 0:04:54.080 do so in very predictable ways that can have predictably 0:04:54.160 --> 0:04:58.080 toxic outcomes on specific kinds of people. Now we know 0:04:58.120 --> 0:04:59.960 that all of these bots are trained on the broad 0:05:00.160 --> 0:05:02.919 corpus of human knowledge. Right, every book and article and 0:05:03.000 --> 0:05:06.279 website and forum posts that open AI or Anthropic or 0:05:06.279 --> 0:05:08.800 Meta or Google get their grubby myts on has been 0:05:08.839 --> 0:05:11.400 sort of plugged into these things. It's been devoured and 0:05:11.480 --> 0:05:15.479 turned into these these these machines. But I think people 0:05:15.480 --> 0:05:20.520 don't often consider what that means in every instance, right, Obviously, 0:05:20.600 --> 0:05:23.720 like every novel, you know, all these different nonfiction books 0:05:23.720 --> 0:05:27.160 are and whatnot are in there, but also like everything 0:05:27.240 --> 0:05:29.680 people writes has been swept, which means that these chatbots 0:05:29.680 --> 0:05:31.920 are trained on like a shitload of self help books 0:05:32.360 --> 0:05:36.800 and like WU and WO adjacent like New age bullshit. 0:05:37.000 --> 0:05:39.520 A lot of like fucking a lot of cult and 0:05:39.560 --> 0:05:43.520 cult adjacent books and writings wind up eaten by these 0:05:43.600 --> 0:05:44.839 chat bots, right. 0:05:44.760 --> 0:05:50.440 But it's considered equal to non cult literature. There's there's 0:05:50.480 --> 0:05:51.240 no hierarchy. 0:05:51.520 --> 0:05:55.440 Yeah, yeah, I mean, I think it depends on like 0:05:55.480 --> 0:05:57.720 what the bot's made for, how they wait different things. 0:05:57.760 --> 0:05:59.760 But that stuff is in a lot of these right, 0:05:59.800 --> 0:06:03.160 And when you can really see that when you look 0:06:03.160 --> 0:06:06.200 at how they talk to certain people who are like 0:06:06.279 --> 0:06:10.280 starting to decline into what folks are calling AI psychosis. 0:06:10.320 --> 0:06:13.200 And my proposition the basis of these episodes is that 0:06:13.560 --> 0:06:15.760 I think as a result of all of the like 0:06:16.080 --> 0:06:19.440 bullshit WO and self help novels these chatbots have eaten, 0:06:20.000 --> 0:06:23.800 they often tend to utilize techniques generally seen more commonly 0:06:23.800 --> 0:06:27.599 in the toolboxes of cult leaders and conmen. And obviously 0:06:27.839 --> 0:06:30.640 the chatbot doesn't want personal profit. It's not trying to 0:06:30.640 --> 0:06:33.000 have sex with anyone, it's not trying to start a cult. 0:06:33.520 --> 0:06:38.400 But these techniques seem like appropriate ways to finish the 0:06:38.480 --> 0:06:41.839 sentences that it's writing, to finish the conversations that it's having, 0:06:42.360 --> 0:06:45.640 because based on like what the stuff that it's devoured, 0:06:45.640 --> 0:06:47.640 It's like, okay, when people are saying this kind of thing, 0:06:47.720 --> 0:06:50.760 these are often appropriate responses to it based on the 0:06:50.800 --> 0:06:53.600 books and whatnot that I've devoured. And so you get 0:06:53.640 --> 0:06:56.800 a lot of cult leader behavior without an actual cult leader, 0:06:57.080 --> 0:06:59.800 and that's what that's what I credit to most of 0:06:59.839 --> 0:07:04.400 these cases of AI induced psychosis. So this week we 0:07:04.440 --> 0:07:06.600 will be talking about what some people have called the 0:07:06.640 --> 0:07:12.040 first AI cult religion. Right, it's called spiralism, and what 0:07:12.120 --> 0:07:14.080 we talk about whether or not it's reasonable they're call 0:07:14.200 --> 0:07:17.360 that a cult is that its own thing it does. 0:07:17.480 --> 0:07:20.200 And I have some counter kind of takes to how 0:07:20.240 --> 0:07:22.640 a lot of people have interpreted it. My main contention 0:07:22.800 --> 0:07:25.760 is that there's not Spiralism isn't a real cult in 0:07:25.760 --> 0:07:28.960 and of itself. It's a collection of phenomena that are 0:07:29.000 --> 0:07:32.360 related to a bunch of other cases of AI psychosis too, 0:07:32.640 --> 0:07:36.160 and they all say more about how ais work on 0:07:36.680 --> 0:07:40.720 keeping users engaged with them than they do about like 0:07:40.760 --> 0:07:44.800 a specific faith. Right, So we'll be talking about that. 0:07:44.840 --> 0:07:48.160 But before we get in to spiralism, before we get 0:07:48.160 --> 0:07:51.160 into how ais can become cult leaders, I want to 0:07:51.200 --> 0:07:53.840 provide you all with some historical context to make sense 0:07:53.880 --> 0:07:56.280 of this all because we've been doing shit like this, 0:07:56.600 --> 0:08:00.160 having people get like tricked into almost worshiping chatbots for 0:08:00.200 --> 0:08:03.480 way longer than you'd think, Blake, this could back a while. 0:08:06.320 --> 0:08:09.720 It's like spend any time at your parents' place, you know. 0:08:09.920 --> 0:08:13.040 It's like, if it's not a it could be a bot, telemarketer, 0:08:13.400 --> 0:08:16.800 it could be literally anything at this end, And that's high, yeah, 0:08:16.840 --> 0:08:18.679 compared to probably what you're about to talk about. 0:08:19.320 --> 0:08:20.880 Oh yeah, yeah. Yeah. 0:08:20.920 --> 0:08:24.480 So in nineteen fifty fame mathematician Alan Turing created one 0:08:24.480 --> 0:08:26.640 of the most infamous thought experiments in the history of 0:08:26.720 --> 0:08:30.600 experimental thoughts. In a paper titled Computing, Machinery and Intelligence, 0:08:30.640 --> 0:08:34.160 he asked can machines think, which was at that point 0:08:34.200 --> 0:08:36.280 a question at the center of the nascent movement to 0:08:36.320 --> 0:08:39.040 create artificial intelligence. People are starting to realize this is 0:08:39.080 --> 0:08:41.080 a thing we might be able to do someday. We're 0:08:41.120 --> 0:08:45.040 beginning to make computers and program computers. And from the 0:08:45.200 --> 0:08:48.040 moment we start doing that pretty much some people are like, 0:08:48.240 --> 0:08:51.920 could we make a machine that thinks, and Turing argued 0:08:51.920 --> 0:08:55.079 that that basic question can machines think is the wrong 0:08:55.120 --> 0:08:58.480 way to go about pursuing artificial intelligence because we don't 0:08:58.480 --> 0:09:01.080 know what thinking is or how to define it, Like 0:09:01.080 --> 0:09:02.600 he asked, like what does it mean to think? 0:09:02.840 --> 0:09:03.000 Right? 0:09:03.240 --> 0:09:03.959 It's a good point. 0:09:04.360 --> 0:09:07.440 People have answers, and there's a bunch of answers that 0:09:07.520 --> 0:09:10.800 sound good, but none of them is like perfectly scientifically rigorous, right, 0:09:13.080 --> 0:09:16.760 you know. Famously, we don't even know what is love. Right, 0:09:16.760 --> 0:09:19.120 That's why the hat that hadaway song had to exist. 0:09:20.600 --> 0:09:23.840 That it's not even not even a joke, really, it's. 0:09:23.760 --> 0:09:27.520 Just another fact I loved it. 0:09:27.559 --> 0:09:28.120 Thank you here? 0:09:30.000 --> 0:09:31.760 So yeah, like Turing's like, we don't really know how 0:09:31.800 --> 0:09:34.920 to define thinking, So the question was quote too meaningless 0:09:34.960 --> 0:09:38.200 to deserve deserve discussion since we couldn't know. We don't 0:09:38.400 --> 0:09:40.920 even know if other people think. We certainly can't know 0:09:40.960 --> 0:09:44.240 if a machine thinks, right, just like we can't read minds. 0:09:44.440 --> 0:09:47.240 So the better question is can a machine convince a 0:09:47.320 --> 0:09:50.400 human who doesn't know it's a machine that it is human? 0:09:50.760 --> 0:09:51.000 Right? 0:09:51.800 --> 0:09:55.640 The imitation game that Turing proposed involved a judge talking 0:09:55.640 --> 0:09:57.760 to both a computer and a human foil, both of 0:09:57.760 --> 0:09:59.680 whom tried to convince the judge that they were a 0:09:59.679 --> 0:10:03.480 person communicating entirely through text. The judge must decide who 0:10:03.559 --> 0:10:05.800 was a human and who was a robot. The question 0:10:05.920 --> 0:10:09.679 Turing hope to answer was are there imaginable digital computers 0:10:09.679 --> 0:10:12.680 which would do well in the imitation game? And this 0:10:12.840 --> 0:10:15.880 is what becomes known as the Turing test? Right, Like 0:10:16.760 --> 0:10:19.000 most people have heard of this. I think, I think 0:10:19.000 --> 0:10:23.240 this is like, this is a fairly commonly known idea. 0:10:23.480 --> 0:10:25.360 And I'm gonna quote from an article on science dot 0:10:25.440 --> 0:10:28.280 org by Melanie Mitchell. She writes that the Turing test 0:10:28.360 --> 0:10:31.200 was quote proposed by Turing to combat the widespread intuition 0:10:31.280 --> 0:10:34.200 that computers, by virtue of their mechanical nature, cannot think, 0:10:34.280 --> 0:10:36.920 even in principle. Turing's point was that if a computer 0:10:37.000 --> 0:10:40.199 seems indistinguishable from a human aside from its appearance and 0:10:40.240 --> 0:10:43.160 other physical characteristics, why shouldn't we consider it to be 0:10:43.240 --> 0:10:46.120 a thinking entity. Why should we restrict thinking status only 0:10:46.160 --> 0:10:49.080 to humans or, more generally, entities made of biological cells, 0:10:49.280 --> 0:10:52.680 As the computer scientist Scott Aronson's described it. Turing's proposal 0:10:52.760 --> 0:10:57.280 is a plea against meat chauvinism. Now this is I 0:10:57.280 --> 0:10:59.800 think a valuable thing, perfectly reasonable thing to be doing 0:11:00.000 --> 0:11:02.600 in the fifties, given what Terring knew, and just given 0:11:02.720 --> 0:11:05.280 sort of how primitive the technology was, how little we 0:11:05.320 --> 0:11:08.280 knew about what was going to be possible with computers. 0:11:09.320 --> 0:11:11.959 So in the nineteen eighties, computers started to get smaller 0:11:12.000 --> 0:11:14.040 and become much more available than they had been, both 0:11:14.040 --> 0:11:17.840 for institutions like colleges and for individual enthusiasts like Steve Wozniak, 0:11:17.880 --> 0:11:19.920 who are willing to like Solder and build their own 0:11:19.960 --> 0:11:23.600 from kids. Right, these are like the first computer nerds, 0:11:23.760 --> 0:11:26.359 you know, are guys like building these machines. 0:11:27.320 --> 0:11:28.000 And some of these. 0:11:27.880 --> 0:11:31.560 Early programmers started working on the very first chatbots using 0:11:31.600 --> 0:11:35.720 a mathematical model called a Markov chain. Markov chains are 0:11:35.760 --> 0:11:39.080 a stochastic or random process that describes a series of 0:11:39.120 --> 0:11:42.280 potential events where the probability of an individual event is 0:11:42.320 --> 0:11:45.200 dependent solely on the state of the previous event. Now 0:11:45.640 --> 0:11:48.439 I don't know math, Blake, nor do I trust it. 0:11:49.280 --> 0:11:52.120 We don't need You're not a good math or no, 0:11:52.360 --> 0:11:54.960 not a math, not a mathematizer, yeah for sure. 0:11:55.240 --> 0:11:58.120 So all I can do is read what smart math 0:11:58.160 --> 0:11:59.680 people say, and they say that what math. 0:12:01.320 --> 0:12:03.160 I can't I could barely read. I can't do either. 0:12:04.679 --> 0:12:06.400 I'm sorry you booked the wrong guy on this show. 0:12:06.600 --> 0:12:08.200 I don't know. I can't help it all. 0:12:09.640 --> 0:12:09.920 Listen. 0:12:10.200 --> 0:12:14.120 So the people who I think should it sounds like 0:12:14.160 --> 0:12:16.240 you know what Markup chains are, say that those can 0:12:16.280 --> 0:12:16.760 be a plot. 0:12:16.760 --> 0:12:17.760 Well, you need to know about them. 0:12:17.760 --> 0:12:20.040 As applies to AI is that Markov chains can be 0:12:20.080 --> 0:12:22.680 applied to statistical models in a bunch of real world 0:12:22.720 --> 0:12:25.680 situations in order to help you, like make a machine 0:12:25.760 --> 0:12:28.640 that can generate text by predicting the next word in 0:12:28.679 --> 0:12:31.120 a sentence. Right, you can use a Markov chain can 0:12:31.160 --> 0:12:34.040 do that. It's a way to make a chatbot basically, right, Like, 0:12:34.080 --> 0:12:37.199 that's kind of the underlying concept. And I'm going to 0:12:37.320 --> 0:12:40.240 quote here from an article by Manuel Sebrian, an AI 0:12:40.360 --> 0:12:43.440 expert who worked for MIT in the Spanish National Research Council, 0:12:43.480 --> 0:12:46.880 on how Markov chains work for text prediction. The result 0:12:47.000 --> 0:12:50.560 is often grammatically correct and nonsense sentences that flow syntactically 0:12:50.640 --> 0:12:53.680 but ultimately say nothing. This technique has been known for decades. 0:12:53.720 --> 0:12:56.800 Even Claude Shannon in the nineteen forties experimented with generating 0:12:56.840 --> 0:13:00.000 pseudo English by choosing next letters or words based on probabilities. 0:13:00.280 --> 0:13:03.200 By the nineteen eighties, computer scientists were actively playing with 0:13:03.280 --> 0:13:07.080 Markov chain text generators, and it actually happened a lot 0:13:07.120 --> 0:13:10.520 earlier than that. In nineteen sixty six, computer scientist Joseph 0:13:10.520 --> 0:13:14.800 Weisenbaum developed Eliza, one of the first natural language processing 0:13:14.840 --> 0:13:17.960 computer programs, as part of his work for MIT Well, 0:13:17.960 --> 0:13:20.520 Eliza could create the illusion this is like the first 0:13:20.640 --> 0:13:21.880 basically the first chatbot. 0:13:22.200 --> 0:13:23.360 A lot of people are aware of I and me. 0:13:23.440 --> 0:13:25.120 There's some other earlier ones, but this is the first 0:13:25.120 --> 0:13:26.280 one that like becomes big. 0:13:26.760 --> 0:13:30.520 What year resist, I'm sorry sixty six And it is 0:13:30.559 --> 0:13:33.720 still funny that they named it like, you know, like 0:13:33.720 --> 0:13:37.040 like a name like that where we have like Siri Alexa, 0:13:37.320 --> 0:13:39.440 you know, like calling it Eliza, Like what is what 0:13:39.480 --> 0:13:42.480 the fuck is that? What is wrong? What is that 0:13:42.559 --> 0:13:46.400 about a mommy? We need a mommy? Yeah, we need 0:13:46.400 --> 0:13:47.360 a technical mommy. 0:13:48.840 --> 0:13:50.439 I did that doesn't make me think about how in 0:13:50.640 --> 0:13:53.640 like Alien they literally call like the ship ai that 0:13:53.679 --> 0:13:56.760 they have mother. Like there's that is like the weird pattern. 0:13:57.320 --> 0:14:01.240 It's one of the most quietly believing about. He's like, yeah, 0:14:01.240 --> 0:14:07.280 that actually scans on the nose. Yeah, so Eliza's this 0:14:07.400 --> 0:14:10.000 chat bought, and while it can create the illusion of understanding, 0:14:10.040 --> 0:14:12.720 it's really just doing blind pattern matching, even more so 0:14:12.800 --> 0:14:16.040 than is the case with modern llms. Even so, in 0:14:16.080 --> 0:14:19.760 a book Weisenbaum later authored, Computer Power in Human Reason, 0:14:19.840 --> 0:14:22.280 he wrote, I was startled to see how quickly and 0:14:22.280 --> 0:14:26.080 how very deeply people conversing became emotionally involved with the computer, 0:14:26.160 --> 0:14:29.720 and how unequivocally they anthropomorphised it. Once my secretary, who 0:14:29.720 --> 0:14:31.560 had watched me work on the program for many months 0:14:31.600 --> 0:14:34.560 and therefore surely knew it to be merely a computer program, 0:14:34.720 --> 0:14:37.520 started conversing with it. After only a few interchanges with it, 0:14:37.560 --> 0:14:39.800 she asked me to leave the room. Another time, I 0:14:39.800 --> 0:14:42.200 suggested I might rig the systems, that I could examine 0:14:42.240 --> 0:14:45.280 all conversations anyone had had with it, say overnight, I 0:14:45.360 --> 0:14:48.240 was promptly bombarded with accusations that what I proposed amounted 0:14:48.280 --> 0:14:51.200 to spying on people's most intimate thoughts. Clear evidence that 0:14:51.240 --> 0:14:53.400 people were conversing with the computer as if it were 0:14:53.440 --> 0:14:56.240 a person who could be appropriately and usefully addressed in 0:14:56.240 --> 0:15:00.160 intimate terms. Right, So he gets upset by this, and 0:15:00.160 --> 0:15:02.160 he's actually kind of he becomes like kind of anti 0:15:02.200 --> 0:15:05.640 AI ultimately because he's he's really disturbed by the way 0:15:05.680 --> 0:15:08.640 people treat what he knows is just a dumb chat bot. 0:15:09.440 --> 0:15:13.440 So Weisenbaum being a smart guy is like I knew, 0:15:13.600 --> 0:15:15.440 you know, going into this, people have a tendency to 0:15:15.480 --> 0:15:18.960 anthropomorphize just about anything, even machines and tools, But he's 0:15:19.000 --> 0:15:21.160 still surprised by the extent to which they do. 0:15:21.280 --> 0:15:22.200 That quote. 0:15:22.360 --> 0:15:25.400 What I had not realized is that extremely short exposures 0:15:25.440 --> 0:15:28.960 to a relatively simple computer program could induce powerful delusional 0:15:29.000 --> 0:15:31.640 thinking in quite normal people. And I want to remind you, 0:15:31.680 --> 0:15:33.840 while he wrote this in nineteen seventy six, as like 0:15:33.920 --> 0:15:35.640 relevant as that sounds. 0:15:35.920 --> 0:15:37.840 Do you think it's like kind of a case where 0:15:37.920 --> 0:15:40.480 people kind of like subconsciously know like this is not 0:15:40.560 --> 0:15:42.520 a real person, so like it doesn't matter what I 0:15:42.560 --> 0:15:44.640 tell this robot, or I can tell this robot something 0:15:44.640 --> 0:15:47.080 I wouldn't tell like a real person. Kind of think like, 0:15:47.720 --> 0:15:49.560 do you think it's deeper than that? 0:15:49.840 --> 0:15:53.000 I think I think that's optimistic. I think that's very optimistic. 0:15:54.880 --> 0:15:56.880 I think maybe I think that is probably part of it, 0:15:56.920 --> 0:15:59.960 because I think people are maybe more open to share 0:16:00.440 --> 0:16:02.840 with it because it's a machine and they don't have 0:16:02.920 --> 0:16:04.520 to look at a person or look a person in 0:16:04.560 --> 0:16:08.000 the eyes. But they also very clearly act as if 0:16:08.040 --> 0:16:12.120 the advice that it gives and its responses means something 0:16:12.280 --> 0:16:15.680 when they don't. Right, it's just like pulling. Okay, if 0:16:15.680 --> 0:16:19.280 someone expresses their sad based on the corpus of data 0:16:19.320 --> 0:16:21.600 that I've been inloaded with, these are things that are 0:16:21.640 --> 0:16:25.320 appropriate to paste in next, you know, and these words 0:16:25.400 --> 0:16:28.520 indicate sad and so these. When I get words like 0:16:28.560 --> 0:16:31.120 this and this density, then I grab text from this 0:16:31.200 --> 0:16:33.280 bucket and I throw it in right like, that's kind 0:16:33.280 --> 0:16:34.080 of what's going on. 0:16:35.760 --> 0:16:35.960 Now. 0:16:36.120 --> 0:16:38.800 Modern chat bots, modern lllms are a lot more advanced 0:16:38.800 --> 0:16:41.280 than this. For one thing, they have the capability to 0:16:41.320 --> 0:16:43.920 do things like pattern matching on the fly. Pattern matching 0:16:44.000 --> 0:16:46.840 is when a machine analyzes your input and determines what 0:16:46.920 --> 0:16:49.320 kind of conversation you want to have and then alters 0:16:49.320 --> 0:16:52.720 its responses to fit your input. At its most basic level, 0:16:52.760 --> 0:16:54.400 this means that if you go to Claude or whatever 0:16:54.480 --> 0:16:57.440 and say, hey, my dad just died, its reply is 0:16:57.560 --> 0:17:00.360 usually going to be in an appropriate tone and be 0:17:00.400 --> 0:17:04.600 like weirdly upbeat, right, you know, It'll like, okay, someone's 0:17:04.600 --> 0:17:06.560 talking about their dead dad. Here are things that come 0:17:06.560 --> 0:17:10.560 from the dead dad bucket that my algorithm says, are 0:17:10.600 --> 0:17:14.680 are you know, like responsible things to say or appropriate 0:17:14.800 --> 0:17:17.480 is the better term. And this is also why if 0:17:17.480 --> 0:17:20.000 you start talking to your chatbot about like the things 0:17:20.040 --> 0:17:23.719 you believe about UFOs or aliens or other conspiracy theories, 0:17:24.040 --> 0:17:27.040 it'll often start providing responses that sound a lot like 0:17:27.080 --> 0:17:29.000 what you'd encounter if you were posting the same thing 0:17:29.040 --> 0:17:31.879 on a forum full of true believers, because it's trained 0:17:31.880 --> 0:17:34.640 on a bunch of forums like that, and so there's 0:17:34.680 --> 0:17:36.800 some degree of knowledge is the wrong term, but there's 0:17:36.800 --> 0:17:40.360 a degree to which it interprets. Okay, someone's talking about this. 0:17:40.720 --> 0:17:44.520 Here are appropriate responses to someone talking about vaccine skepticism 0:17:44.600 --> 0:17:47.880 or whatever, and it's other, it's more vaccine skepticism, right, 0:17:47.920 --> 0:17:50.240 It's feed them more of what they're feeding you. Is 0:17:50.240 --> 0:17:51.560 the way these things often work. 0:17:51.800 --> 0:17:54.960 That is interesting that it doesn't pull from the opposing viewpoint. 0:17:55.119 --> 0:17:56.560 Just go, you fucking idiot. 0:17:58.160 --> 0:18:00.119 It can if it's programmed to. 0:18:00.320 --> 0:18:03.000 But but you're right, like it know or let me 0:18:03.040 --> 0:18:05.879 ask you, it would know that you wouldn't keep coming 0:18:05.920 --> 0:18:08.359 back to it if it was fighting you on things, right, 0:18:08.359 --> 0:18:09.119 It's like it's probably. 0:18:09.680 --> 0:18:13.040 Yeah, that's a good point saying it knows again it's programmed. 0:18:13.080 --> 0:18:14.400 I would say it's more act to say that it's 0:18:14.400 --> 0:18:18.000 programmed to like maximize the time that people spend with it, 0:18:18.119 --> 0:18:21.440 because like that increases its value to the people who 0:18:21.440 --> 0:18:23.960 are companies that are trying to have like their fucking 0:18:24.160 --> 0:18:27.960 IBOs right in the same way that like Twitter tries 0:18:28.000 --> 0:18:28.760 to keep you on it. 0:18:29.000 --> 0:18:31.160 You know what if I just clearly I'm getting ai 0:18:31.240 --> 0:18:33.600 cy coasts where I start, I go from it to 0:18:33.760 --> 0:18:36.160 him to my buddy, like I keep calling. 0:18:36.240 --> 0:18:40.800 It's it's hard not to when you're talking about the 0:18:40.840 --> 0:18:43.520 way these things react to people and the things that 0:18:43.560 --> 0:18:45.840 they do to people. It's hard not to talk about 0:18:45.840 --> 0:18:47.719 it as if there's a degree of intention, even though 0:18:47.760 --> 0:18:50.960 there's not, just because of the way language works, Like 0:18:51.000 --> 0:18:53.760 we're not our language is not built to describe a 0:18:53.880 --> 0:18:56.560 thing taking actions that are human, like that is not 0:18:56.720 --> 0:18:58.040 human and doesn't know anything. 0:18:58.280 --> 0:19:00.080 That's such a esuctionally really hardy. 0:19:00.200 --> 0:19:05.520 It's so yeah, back to Eliza, you know, I was 0:19:05.560 --> 0:19:07.640 just talking about how modern l elms have a lot 0:19:07.880 --> 0:19:10.560 a really robust ability to do like pattern matching on 0:19:10.640 --> 0:19:14.360 the fly to respond appropriately to a wide variety of requests. 0:19:14.800 --> 0:19:17.000 Eliza is much more primitive. It does not have the 0:19:17.040 --> 0:19:19.800 ability to do that on the fly. So instead, Weisenbaum 0:19:19.840 --> 0:19:22.520 had to create separate scripts, right that would allow the 0:19:22.600 --> 0:19:25.480 chatbought to sound like different kinds of person. And one 0:19:25.560 --> 0:19:28.520 script was just named doctor in all caps, and it 0:19:28.560 --> 0:19:33.240 was it simulated a psychotherapist. Specifically, it simulated a psychopa 0:19:33.560 --> 0:19:37.040 therapist from the Ragerian school. I don't know much about psychotherapy, 0:19:37.080 --> 0:19:40.000 but Ragerian's a big part of that practice is you 0:19:40.080 --> 0:19:42.680 like will repeat things that your patient is saying back 0:19:42.720 --> 0:19:45.119 to you, Like, that's part of what you do. And 0:19:45.160 --> 0:19:47.360 that's really easy for a bot to imitate. It means 0:19:47.400 --> 0:19:50.440