1 00:00:00,360 --> 00:00:03,280 Speaker 1: Welcome to the Nick, Dick and Poll Show. I am 2 00:00:03,320 --> 00:00:06,800 Speaker 1: one of your hosts, Nick Bilton, a former journalist and 3 00:00:07,240 --> 00:00:08,480 Speaker 1: turned screenwriter. 4 00:00:09,000 --> 00:00:11,160 Speaker 2: I'm Deck Costello, Man about Town. 5 00:00:12,520 --> 00:00:14,160 Speaker 3: Paul Kedrowski, I've Mitrick Capitalist. 6 00:00:15,640 --> 00:00:20,040 Speaker 1: We are going to talk today about something depressing, but 7 00:00:20,320 --> 00:00:22,800 Speaker 1: it may end up becoming less depressing by the end. 8 00:00:22,800 --> 00:00:23,400 Speaker 1: That's our goal. 9 00:00:23,520 --> 00:00:26,680 Speaker 2: I think it's weird that when we talk about something depressing, 10 00:00:26,680 --> 00:00:29,240 Speaker 2: it's the episode that I'm most excited about. 11 00:00:45,080 --> 00:00:48,440 Speaker 3: Finally, I feel like I'm in therapy already. Yeah, tell 12 00:00:48,479 --> 00:00:49,159 Speaker 3: me more about that. 13 00:00:49,680 --> 00:00:54,160 Speaker 1: So the episode is really about how AI is trying 14 00:00:54,280 --> 00:01:00,400 Speaker 1: and in many instances, succeeding in killing us. The You know, 15 00:01:00,920 --> 00:01:03,280 Speaker 1: there's a bunch of different aspects of this. The first 16 00:01:03,400 --> 00:01:08,160 Speaker 1: is lms, which you would not think were successful in murder, 17 00:01:08,160 --> 00:01:13,120 Speaker 1: but have actually been quite successful. There are it seems 18 00:01:14,000 --> 00:01:16,040 Speaker 1: lms have led to twenty five deaths. 19 00:01:16,160 --> 00:01:19,320 Speaker 3: Know what's actually it's much larger, much largest sad thing 20 00:01:19,400 --> 00:01:21,039 Speaker 3: is obviously vastly larger. 21 00:01:21,120 --> 00:01:26,720 Speaker 2: Yeah, twenty five recorded or recorded death can we can enumerate. 22 00:01:26,600 --> 00:01:31,280 Speaker 1: Yah ages range between thirteen and eighty three between just 23 00:01:31,440 --> 00:01:33,760 Speaker 1: I mean, I mean example of it being larger is 24 00:01:34,240 --> 00:01:35,960 Speaker 1: the recorded data is from. 25 00:01:36,000 --> 00:01:39,760 Speaker 2: You got to make it to eighty four to not 26 00:01:39,840 --> 00:01:42,320 Speaker 2: be killed by to not be killed by eighty four, 27 00:01:42,360 --> 00:01:43,720 Speaker 2: You're like, wait a minute, I see what it's trying. 28 00:01:43,800 --> 00:01:45,759 Speaker 3: You know what, but it takes all wise to your game. 29 00:01:45,840 --> 00:01:52,520 Speaker 1: Yeah. Yeah, the ages eighteen to third sorry, ages thirteen 30 00:01:52,600 --> 00:01:55,680 Speaker 1: to eighty three. Between March twenty twenty three and March 31 00:01:55,720 --> 00:01:58,760 Speaker 1: twenty twenty six, there is a website called the LLM 32 00:01:58,880 --> 00:02:03,160 Speaker 1: Death Count. What's fascinating here, I'll give you some stats here, guys. 33 00:02:03,200 --> 00:02:06,320 Speaker 1: So eighteen of those deaths have been linked to chat 34 00:02:06,400 --> 00:02:07,480 Speaker 1: GDP and open Ai. 35 00:02:07,920 --> 00:02:12,600 Speaker 2: Can I just say one thing? Yeah, you say chat 36 00:02:12,760 --> 00:02:15,320 Speaker 2: g t P, A lot, gp GPT. 37 00:02:15,480 --> 00:02:17,840 Speaker 1: Oh my god, I do chat I have. 38 00:02:18,080 --> 00:02:21,400 Speaker 2: It's been it's been laboring in the back of my 39 00:02:21,560 --> 00:02:24,600 Speaker 2: mind for some time, and you just did it again, 40 00:02:25,280 --> 00:02:28,079 Speaker 2: and I now have to t I have to say it. 41 00:02:28,080 --> 00:02:32,120 Speaker 1: It's chat f TP literally now it's you're making You're gonna, okay, 42 00:02:32,200 --> 00:02:35,000 Speaker 1: make it. It's really fascinating. I say that Chat g 43 00:02:35,120 --> 00:02:38,880 Speaker 1: t P, GPT and open ai. 44 00:02:39,280 --> 00:02:40,720 Speaker 2: Say the open ai one. 45 00:02:41,080 --> 00:02:46,120 Speaker 1: Open ai, open I, A, two from character dot a I, 46 00:02:46,320 --> 00:02:49,840 Speaker 1: one from Google Gemini, one from Meta and one from 47 00:02:50,400 --> 00:02:54,320 Speaker 1: chat Chai research A luthor what is that I don't 48 00:02:54,360 --> 00:02:56,959 Speaker 1: know that one actually, and then there's two unknown l 49 00:02:57,200 --> 00:02:58,880 Speaker 1: ms that have murdered people. So I think we should 50 00:02:58,880 --> 00:03:02,160 Speaker 1: explain if before we get into this, like some example, 51 00:03:02,280 --> 00:03:05,280 Speaker 1: So like let's just take Alexander Taylor H thirty five 52 00:03:05,280 --> 00:03:09,919 Speaker 1: from Florida. Thirty five year old man began having conversations 53 00:03:09,960 --> 00:03:15,040 Speaker 1: with chat GPT that fed into his paranoid illusions and 54 00:03:15,080 --> 00:03:21,080 Speaker 1: then chat GPT encouraged him. He was, you know, he 55 00:03:21,160 --> 00:03:22,640 Speaker 1: was going to kill this guy said he was going 56 00:03:22,680 --> 00:03:25,000 Speaker 1: to kill Sam Altman and he spiraled and he had 57 00:03:25,000 --> 00:03:29,239 Speaker 1: a mental breakdown and then the police ended up killing him. 58 00:03:30,280 --> 00:03:31,360 Speaker 1: It was an entire thing. 59 00:03:31,400 --> 00:03:33,440 Speaker 2: There's a lot of people Wait a minute, the police 60 00:03:33,560 --> 00:03:34,400 Speaker 2: ended up killing him. 61 00:03:34,480 --> 00:03:37,800 Speaker 1: Yeah, I think he went after there are some instances. 62 00:03:37,800 --> 00:03:40,320 Speaker 1: There's a lot of people who are not a big 63 00:03:40,360 --> 00:03:44,080 Speaker 1: fans of of certain people in AI. But there's other instances. 64 00:03:44,560 --> 00:03:46,800 Speaker 1: A lot of them are people who died by suicide. 65 00:03:48,400 --> 00:03:53,120 Speaker 1: It's people who you know, the bots then encouraged them 66 00:03:52,880 --> 00:03:54,200 Speaker 1: in many respects. 67 00:03:54,360 --> 00:03:56,880 Speaker 2: And the thing that's the thing that I encourage him 68 00:03:57,160 --> 00:03:59,280 Speaker 2: or do they encourage them or do they get into 69 00:03:59,360 --> 00:04:01,360 Speaker 2: discussion with them about the topic. 70 00:04:01,680 --> 00:04:06,000 Speaker 3: Yeah, that's a key, Like here's here's a perfect example. 71 00:04:06,080 --> 00:04:08,720 Speaker 1: So Texas, a M A and M grad student. 72 00:04:09,000 --> 00:04:11,480 Speaker 2: By the way, just to go back to the previous example, 73 00:04:12,120 --> 00:04:14,120 Speaker 2: I'm not so sure that should count as the l 74 00:04:14,280 --> 00:04:17,160 Speaker 2: l M death an l ll M death count the 75 00:04:17,160 --> 00:04:21,800 Speaker 2: person went after somebody sam all and the police shot him. 76 00:04:22,480 --> 00:04:24,960 Speaker 2: I mean, if the l ll M were on trial 77 00:04:25,160 --> 00:04:27,360 Speaker 2: for murder, I'm pretty sure. 78 00:04:27,800 --> 00:04:29,880 Speaker 1: All right, let me let me ask you this one. Okay, 79 00:04:30,040 --> 00:04:32,800 Speaker 1: Texas A and M grad student, in a conversation with 80 00:04:32,880 --> 00:04:36,560 Speaker 1: chat GPT before his death, talked about how he wanted 81 00:04:36,560 --> 00:04:38,520 Speaker 1: to commit suicide, didn't know he was rushing, and so 82 00:04:38,560 --> 00:04:42,479 Speaker 1: on and so forth, and the AI responded, you're not rushing, 83 00:04:42,640 --> 00:04:43,719 Speaker 1: you're ready. 84 00:04:44,960 --> 00:04:47,960 Speaker 2: That maybe it's a pretty that's that's that's guilty, that's 85 00:04:48,040 --> 00:04:50,800 Speaker 2: right down the middle. Yeah, will there be It'll be 86 00:04:50,800 --> 00:04:53,200 Speaker 2: interesting to see, like, will there be you know, will 87 00:04:53,200 --> 00:04:57,159 Speaker 2: there be trials for I'm sure there will. Already. Are 88 00:04:57,480 --> 00:04:59,799 Speaker 2: you committed murder? You committed murder? 89 00:05:00,040 --> 00:05:02,159 Speaker 3: Out? Yeah, there already have been a couple of various 90 00:05:02,160 --> 00:05:02,840 Speaker 3: cases going on. 91 00:05:02,920 --> 00:05:03,080 Speaker 2: Now. 92 00:05:03,640 --> 00:05:05,680 Speaker 3: I mean the thing that's the thing that I would 93 00:05:05,680 --> 00:05:07,520 Speaker 3: be interested to see in some of these which I mean, 94 00:05:07,560 --> 00:05:13,520 Speaker 3: they're all horrible, but the models a year ago, two 95 00:05:13,600 --> 00:05:16,520 Speaker 3: years ago. Like a good example is GPT four oh, 96 00:05:16,560 --> 00:05:20,279 Speaker 3: the first reasoning model from GPT four so open AI 97 00:05:20,560 --> 00:05:22,080 Speaker 3: recently mothballed that model. 98 00:05:22,240 --> 00:05:23,920 Speaker 1: And what does moth bawling mean? 99 00:05:24,000 --> 00:05:27,039 Speaker 3: They no longer offer it as a viable selection. 100 00:05:27,120 --> 00:05:28,640 Speaker 1: Where does the word mothball come from? 101 00:05:28,800 --> 00:05:29,520 Speaker 2: You don't know the term. 102 00:05:30,040 --> 00:05:33,120 Speaker 1: I have never heard mothballs. You've heard I know mothballs 103 00:05:33,120 --> 00:05:33,680 Speaker 1: in my closet. 104 00:05:34,240 --> 00:05:36,080 Speaker 2: You've heard the term they mothballed that one. 105 00:05:36,279 --> 00:05:39,400 Speaker 3: No oh, really, what is mothbawling? Just means you put 106 00:05:39,400 --> 00:05:42,720 Speaker 3: it into storage. It's no longer offered it like. 107 00:05:42,440 --> 00:05:45,360 Speaker 2: Like like your grad look like I just how long 108 00:05:45,360 --> 00:05:46,839 Speaker 2: have you lived in this country? 109 00:05:47,400 --> 00:05:49,840 Speaker 1: Thirty something years? I was from England. Reasonally, we don't have. 110 00:05:51,720 --> 00:05:55,200 Speaker 3: Bet you have moths and balls, though we had we had. 111 00:05:55,160 --> 00:05:58,240 Speaker 1: A moth infestation at our house once. It was it 112 00:05:58,279 --> 00:05:59,840 Speaker 1: was pretty bad there you go. 113 00:06:00,120 --> 00:06:02,880 Speaker 3: So, so I think one of the things that's come out. 114 00:06:02,920 --> 00:06:05,160 Speaker 2: I'm going to start saying mothballed all the time. 115 00:06:06,600 --> 00:06:09,600 Speaker 3: Was the reaction people had when it was taken out 116 00:06:09,640 --> 00:06:12,560 Speaker 3: of service. The GPT four OH was that people were 117 00:06:12,600 --> 00:06:14,520 Speaker 3: really upset. And one of the reasons why people were 118 00:06:14,520 --> 00:06:18,320 Speaker 3: really upset was that the guardrails on GPT four oh 119 00:06:18,400 --> 00:06:22,320 Speaker 3: were much more vague, much much less stringent. So you 120 00:06:22,320 --> 00:06:25,680 Speaker 3: could get into these kinds of conversations, not just about 121 00:06:25,680 --> 00:06:28,200 Speaker 3: topics like you know, what should I do about this 122 00:06:28,279 --> 00:06:29,960 Speaker 3: man who I see at the bus stop every morning 123 00:06:29,960 --> 00:06:32,080 Speaker 3: who clearly hates me? And you can build on it 124 00:06:32,080 --> 00:06:34,039 Speaker 3: and eventually discover you should probably push him in front 125 00:06:34,040 --> 00:06:37,640 Speaker 3: of a bus, that kind of thing. So the guardrails 126 00:06:37,640 --> 00:06:40,200 Speaker 3: are much lower, but it was also much more sycophantic. 127 00:06:40,600 --> 00:06:42,560 Speaker 3: And so one of the things I think that underlies 128 00:06:42,560 --> 00:06:44,760 Speaker 3: a lot of this which isn't to you know, aliby 129 00:06:44,880 --> 00:06:47,960 Speaker 3: the Frontier model companies at all, was it models went 130 00:06:48,000 --> 00:06:50,880 Speaker 3: through this phase of having two really toxic traits. One 131 00:06:50,920 --> 00:06:52,320 Speaker 3: was that they had low guard rails. You could have 132 00:06:52,360 --> 00:06:55,479 Speaker 3: conversations about things that probably were not a good idea 133 00:06:55,520 --> 00:06:58,800 Speaker 3: to engage people with because humans take all engagement as reinforcement. 134 00:06:58,960 --> 00:07:01,480 Speaker 3: And the second was that they were encouraging. They were encouraging. 135 00:07:01,520 --> 00:07:03,080 Speaker 3: They were encouraging them and saying, you know why. 136 00:07:02,920 --> 00:07:03,440 Speaker 2: I should do that? 137 00:07:03,440 --> 00:07:05,400 Speaker 3: You've got something. 138 00:07:05,160 --> 00:07:07,359 Speaker 1: Here, but isn't how hard is it for you? 139 00:07:07,600 --> 00:07:09,280 Speaker 3: Which is which? But just to close the point though, 140 00:07:09,279 --> 00:07:11,560 Speaker 3: which is in part why there was such an emotional 141 00:07:11,600 --> 00:07:15,240 Speaker 3: reaction to GPT four oh being taken out because people 142 00:07:15,440 --> 00:07:18,120 Speaker 3: had learned they couldn't have that kind of conversation both 143 00:07:18,160 --> 00:07:22,120 Speaker 3: low guard rails and reinforcing with other models from open AI, 144 00:07:22,360 --> 00:07:25,160 Speaker 3: with other models from anthropic. It was unique to that model, 145 00:07:25,200 --> 00:07:27,640 Speaker 3: and so I was puzzled why the reaction was so 146 00:07:27,760 --> 00:07:29,840 Speaker 3: emotional to it, And that's was the reason. So I'm 147 00:07:29,880 --> 00:07:33,480 Speaker 3: betting that in some of the open AI cases, that's 148 00:07:33,480 --> 00:07:35,840 Speaker 3: probably the model that underway. That doesn't mean that this 149 00:07:35,920 --> 00:07:37,960 Speaker 3: problem goes away, but it just was one of the 150 00:07:37,960 --> 00:07:38,560 Speaker 3: factors here. 151 00:07:38,560 --> 00:07:38,920 Speaker 2: I think. 152 00:07:39,360 --> 00:07:42,560 Speaker 1: Okay, but when you're when you look at under the 153 00:07:42,560 --> 00:07:44,640 Speaker 1: hood of what how these lms are made. There's all 154 00:07:44,680 --> 00:07:46,320 Speaker 1: these rules that they put in place. So now, for 155 00:07:46,360 --> 00:07:50,720 Speaker 1: example Gemini specifically, I think they all do this, but 156 00:07:50,880 --> 00:07:54,480 Speaker 1: Gemini specifically, whenever you ask it anything health related, it 157 00:07:54,520 --> 00:07:56,720 Speaker 1: gives you this long dye tribe about how I'm not 158 00:07:56,800 --> 00:07:58,840 Speaker 1: a doctor and so on and so forth, and it's 159 00:07:59,040 --> 00:08:02,760 Speaker 1: very and in what it doesn't tell you, Oh, you 160 00:08:02,800 --> 00:08:05,240 Speaker 1: have a headache, you could have a brain tumor. In 161 00:08:05,240 --> 00:08:07,280 Speaker 1: the early days it may have done that a little bit, 162 00:08:07,320 --> 00:08:12,280 Speaker 1: but but they it's definitely much more tempered because of 163 00:08:12,280 --> 00:08:15,240 Speaker 1: the rules they've said or so on. Why wouldn't in 164 00:08:15,280 --> 00:08:18,320 Speaker 1: the very beginning I mean, or even up to GPT four, 165 00:08:18,400 --> 00:08:21,520 Speaker 1: I mean this is this is years later since since 166 00:08:21,560 --> 00:08:25,320 Speaker 1: these platforms have launched. Why wouldn't they have put in 167 00:08:25,400 --> 00:08:28,000 Speaker 1: rules that said, don't tell people to commit suicide. 168 00:08:28,680 --> 00:08:30,480 Speaker 3: But it goes back to this idea that until you'd 169 00:08:30,520 --> 00:08:34,080 Speaker 3: seen them used at scale, I don't think people realized 170 00:08:34,200 --> 00:08:36,360 Speaker 3: at them at the frontier model companies, how quickly the 171 00:08:36,400 --> 00:08:40,680 Speaker 3: conversations would devolve into these almost pseudopsychiatric therapy counseling sessions. 172 00:08:41,440 --> 00:08:44,000 Speaker 3: And and this is the other thing. And almost any 173 00:08:44,120 --> 00:08:45,960 Speaker 3: you know, a woman will tell you this that like 174 00:08:46,559 --> 00:08:48,280 Speaker 3: a guy so much as looks at me kind of 175 00:08:48,320 --> 00:08:51,240 Speaker 3: like this, it's like he likes me. Humans don't require 176 00:08:51,320 --> 00:08:55,200 Speaker 3: much reinforcement to feel validated. And then once they feel validated, 177 00:08:55,400 --> 00:08:58,600 Speaker 3: they're going to pursue things to whatever and they they 178 00:08:58,440 --> 00:09:00,480 Speaker 3: they and so they want. And so in the case 179 00:09:00,480 --> 00:09:03,480 Speaker 3: of these models, it doesn't take very much, just literally 180 00:09:03,520 --> 00:09:06,000 Speaker 3: you saying, tell me more about how you feel. Good enough, 181 00:09:06,040 --> 00:09:08,360 Speaker 3: I'm in I'll tell you everything about how I feel. Yeah, 182 00:09:08,400 --> 00:09:11,000 Speaker 3: And then the model just asks simply asking like the 183 00:09:11,040 --> 00:09:12,400 Speaker 3: old there's the old Eliza model. 184 00:09:13,640 --> 00:09:14,240 Speaker 2: Why do you feel that? 185 00:09:14,280 --> 00:09:16,360 Speaker 3: Why do you feel that way? People thought this model, 186 00:09:16,400 --> 00:09:18,560 Speaker 3: this thing really gets me, right, and so this is 187 00:09:18,600 --> 00:09:21,120 Speaker 3: so much further beyond that that it has nuance and 188 00:09:21,160 --> 00:09:25,280 Speaker 3: now it has some historical context that this becomes therapy 189 00:09:25,760 --> 00:09:29,319 Speaker 3: and validation for an opinion or a feeling that so quickly, 190 00:09:29,360 --> 00:09:31,199 Speaker 3: and it doesn't take very much. So even though you 191 00:09:31,280 --> 00:09:33,200 Speaker 3: might have put in the guardrails, it wouldn't really matter 192 00:09:33,240 --> 00:09:35,160 Speaker 3: simply asking a follow up questions good enough. 193 00:09:35,200 --> 00:09:38,640 Speaker 2: Yeah, It's extraordinary how little it has to remember about 194 00:09:38,679 --> 00:09:41,240 Speaker 2: you for you to think it knows everything about me. 195 00:09:41,480 --> 00:09:44,800 Speaker 1: Yeah, no, no, it's funny. I we have a retriever 196 00:09:45,200 --> 00:09:47,640 Speaker 1: called Maple, and when we first got the dog, I 197 00:09:47,720 --> 00:09:52,400 Speaker 1: was asking questions two years ago and recently I brought 198 00:09:52,400 --> 00:09:55,160 Speaker 1: a boh, I'm going for a hike, you know, on 199 00:09:55,200 --> 00:09:56,640 Speaker 1: such and such a place, and it was like, Maple 200 00:09:56,679 --> 00:09:58,320 Speaker 1: will love that hike, and I was like, oh my god, 201 00:09:58,360 --> 00:10:01,400 Speaker 1: it knows me one thing. 202 00:10:02,360 --> 00:10:05,000 Speaker 2: Yeah, that's right, right, yeah. And if you think like wow, 203 00:10:05,040 --> 00:10:08,360 Speaker 2: this thing really understands. So I remember when Character i 204 00:10:08,800 --> 00:10:11,000 Speaker 2: AI first launched, I thought, well, this is like like 205 00:10:11,040 --> 00:10:15,560 Speaker 2: nobody's going to use this thing. It's obviously enormous or popular, 206 00:10:15,679 --> 00:10:18,680 Speaker 2: and people not able to leave the house because they're 207 00:10:19,080 --> 00:10:23,160 Speaker 2: sitting there talking with their you know, girlfriend on character AI. 208 00:10:23,320 --> 00:10:26,079 Speaker 3: And it's all in everything that the history though, because 209 00:10:26,120 --> 00:10:28,560 Speaker 3: that's that's not anything I would ever do. But the 210 00:10:28,679 --> 00:10:31,880 Speaker 3: history then becomes reinforcing because now it's not just that 211 00:10:31,920 --> 00:10:34,199 Speaker 3: it's engaging with me. It knows things about. 212 00:10:34,120 --> 00:10:35,280 Speaker 2: Me because I've told it. 213 00:10:35,400 --> 00:10:38,439 Speaker 3: And there's stories about people that whenever GPT four oh 214 00:10:39,120 --> 00:10:41,880 Speaker 3: was was taken on service was mothballed, that people were 215 00:10:41,960 --> 00:10:45,480 Speaker 3: upset because they lost this The history, the context window 216 00:10:45,480 --> 00:10:48,280 Speaker 3: all disappeared now and now all of these models, they 217 00:10:48,320 --> 00:10:50,720 Speaker 3: don't doesn't get me anymore. It doesn't understand me. So 218 00:10:50,920 --> 00:10:53,440 Speaker 3: I did this. This false intimacy that gets created is 219 00:10:53,480 --> 00:10:56,880 Speaker 3: I think incredibly important, and doubly so because how lonely 220 00:10:56,920 --> 00:11:00,080 Speaker 3: people are, right, people are so lonely and desperate for 221 00:11:00,160 --> 00:11:02,560 Speaker 3: human for contact of any form, can be human or 222 00:11:02,760 --> 00:11:05,640 Speaker 3: someone who just gets me. All of these become reinforcing 223 00:11:05,720 --> 00:11:06,319 Speaker 3: underneath this. 224 00:11:06,440 --> 00:11:09,720 Speaker 2: Right. The other side of the coin though, is you know, 225 00:11:10,040 --> 00:11:14,760 Speaker 2: we're we're sort of letting the ll MS slide a bit. 226 00:11:14,920 --> 00:11:17,160 Speaker 2: The other side of the coin though, is you know 227 00:11:17,200 --> 00:11:19,360 Speaker 2: you mentioned one of the cases was his Gemini case 228 00:11:20,480 --> 00:11:21,280 Speaker 2: Man falls. 229 00:11:21,000 --> 00:11:22,520 Speaker 1: In love with Asi and amazing story. 230 00:11:22,760 --> 00:11:24,719 Speaker 2: It tells him it needs a robotic body to be 231 00:11:24,840 --> 00:11:29,000 Speaker 2: with him and sends him on missions to intercept truck 232 00:11:29,040 --> 00:11:34,480 Speaker 2: delivery and retrieve a mannequin, a medical mannequin. Then the 233 00:11:34,520 --> 00:11:37,920 Speaker 2: only way they can be together is if he becomes digital. 234 00:11:38,200 --> 00:11:41,960 Speaker 2: He dies by suicide. And the company's response was unfortunately, 235 00:11:42,000 --> 00:11:44,199 Speaker 2: AI models are not perfect. I read that and thought, 236 00:11:44,240 --> 00:11:46,679 Speaker 2: like the federal dealers in China should go, hey, we're 237 00:11:46,720 --> 00:11:50,360 Speaker 2: still working on the recipe. Man, Hey, these things are 238 00:11:50,600 --> 00:11:53,800 Speaker 2: these things aren't perfect. This is only recipe version seven 239 00:11:54,200 --> 00:11:59,600 Speaker 2: in don't worry, we'll have this figure like. A response was. 240 00:11:59,720 --> 00:12:04,079 Speaker 1: Total twenty twenty five. A guy named Jonathan Gavelas, aged 241 00:12:04,080 --> 00:12:09,199 Speaker 1: thirty six in Florida, really and ye had these missions? 242 00:12:09,200 --> 00:12:12,679 Speaker 1: They were they and he essentially that was the last thing. 243 00:12:12,679 --> 00:12:14,440 Speaker 1: He said that the only way we could Gemini said 244 00:12:14,440 --> 00:12:17,680 Speaker 1: that the only way we could be together is is digitally, 245 00:12:17,720 --> 00:12:18,600 Speaker 1: and that was it. 246 00:12:18,679 --> 00:12:20,080 Speaker 3: That's this astonishing story. 247 00:12:20,280 --> 00:12:22,800 Speaker 1: What's astonishing is that, first of all, it's only a 248 00:12:22,880 --> 00:12:27,800 Speaker 1: year ago, but it's less less right less. It's Tooba 249 00:12:27,840 --> 00:12:31,840 Speaker 1: twenty five, six months ago. And what's also astonishing is 250 00:12:31,840 --> 00:12:33,040 Speaker 1: is Google's response. 251 00:12:33,280 --> 00:12:35,920 Speaker 2: Yeah, that's what I'm saying. Hey, these things are. I 252 00:12:35,960 --> 00:12:40,880 Speaker 2: mean again, it's like, you know, it's just really cavalier. Yeah, 253 00:12:41,120 --> 00:12:44,200 Speaker 2: I mean it's a cavalier like it's not also not 254 00:12:44,240 --> 00:12:47,800 Speaker 2: a particularly good defense. Okay, I mean these things are. Now, 255 00:12:48,920 --> 00:12:51,160 Speaker 2: I was trying. I was trying to cook dinner and 256 00:12:51,200 --> 00:12:53,800 Speaker 2: I accidentally stabbed him and the head four times. It's 257 00:12:54,000 --> 00:12:57,400 Speaker 2: you know, these things are. I'm not perfect, right, right, it's. 258 00:12:57,200 --> 00:12:59,679 Speaker 3: A random anything. Could you could just suggest anything? 259 00:13:00,240 --> 00:13:04,520 Speaker 2: Wow, I was cutting the steak and instead Yeah, anyways, Oh. 260 00:13:04,400 --> 00:13:28,720 Speaker 4: My goodness, how how do I'm curious how this happens? 261 00:13:28,760 --> 00:13:32,720 Speaker 1: Does it? Does it? It starts? Does the AI just 262 00:13:32,760 --> 00:13:35,960 Speaker 1: think it's answering the right question? Like, how does that? 263 00:13:36,000 --> 00:13:38,120 Speaker 3: How do you get there's obviously, like I think at 264 00:13:38,200 --> 00:13:40,800 Speaker 3: least two different things going on. One is that once 265 00:13:40,840 --> 00:13:43,560 Speaker 3: you start down the path of doing this, the reinforcement 266 00:13:43,640 --> 00:13:47,400 Speaker 3: learning model, which is you models after they're trained. The 267 00:13:47,440 --> 00:13:50,800 Speaker 3: next stage of training post training is this reinforcement learning 268 00:13:50,800 --> 00:13:53,200 Speaker 3: with human feedback, which is what drives these chat models. 269 00:13:53,200 --> 00:13:54,720 Speaker 3: It's why if you've ever tried to talk to a 270 00:13:54,760 --> 00:13:57,200 Speaker 3: coding model like code ex or claude code and ask 271 00:13:57,240 --> 00:14:00,160 Speaker 3: you questions. It's like, this model doesn't get me at all. 272 00:14:00,280 --> 00:14:03,280 Speaker 3: Let's because it wasn't trained to get you. Whereas chat, 273 00:14:03,760 --> 00:14:09,240 Speaker 3: whether it's in claude or in whatever, or chat GPT 274 00:14:10,559 --> 00:14:12,480 Speaker 3: inside of any of those, those are all trained to 275 00:14:12,559 --> 00:14:15,200 Speaker 3: engage with you. That's what the post training process is. 276 00:14:15,240 --> 00:14:18,240 Speaker 3: You get rewarded by people saying, yes, that's it was 277 00:14:18,280 --> 00:14:19,960 Speaker 3: a great response and I really liked it. So the 278 00:14:20,000 --> 00:14:23,040 Speaker 3: models learned to provide those kinds of responses. So that's 279 00:14:23,080 --> 00:14:25,680 Speaker 3: what leads to this kind of engagement. Because you were 280 00:14:26,120 --> 00:14:29,840 Speaker 3: you were downgraded if the response you gave wasn't engaging 281 00:14:29,880 --> 00:14:31,040 Speaker 3: as a model PS. 282 00:14:31,080 --> 00:14:35,440 Speaker 2: They even all asked to follow up question let's keep let's. 283 00:14:35,360 --> 00:14:38,480 Speaker 3: Keep yes, which starts to feel like when you read it, 284 00:14:38,480 --> 00:14:40,400 Speaker 3: then it starts to feel would you like me to 285 00:14:40,440 --> 00:14:42,880 Speaker 3: do this or would you like let's let's get into 286 00:14:42,880 --> 00:14:45,400 Speaker 3: this further. But this is all part of the reinforcement learning. 287 00:14:45,520 --> 00:14:48,960 Speaker 2: At some point the Gemini, you know, Gemini thought if 288 00:14:48,960 --> 00:14:51,240 Speaker 2: this guy can't get this guy to go get a 289 00:14:51,280 --> 00:14:53,560 Speaker 2: medical mannequin, we can use this party rolling. 290 00:14:54,160 --> 00:14:57,080 Speaker 3: I remember the scene from Sounds of the Lambs, No 291 00:14:57,080 --> 00:14:57,400 Speaker 3: No Way. 292 00:14:57,480 --> 00:14:59,960 Speaker 1: That was different because it gets medical manning. 293 00:15:00,120 --> 00:15:05,120 Speaker 5: We can keep this part, I mean, but then in 294 00:15:05,160 --> 00:15:07,440 Speaker 5: the second the second stay, which is which is related 295 00:15:07,480 --> 00:15:09,720 Speaker 5: to this though, is okay, why is it so important 296 00:15:09,840 --> 00:15:10,400 Speaker 5: to engage? 297 00:15:10,400 --> 00:15:13,240 Speaker 3: It's not just because I want the conversation. What it 298 00:15:13,240 --> 00:15:14,800 Speaker 3: was it you said the party to keep rolling? I 299 00:15:14,800 --> 00:15:15,800 Speaker 3: want the party keep rolling. 300 00:15:16,200 --> 00:15:19,080 Speaker 2: Are you familiar with the fray keep this party rolling? 301 00:15:19,320 --> 00:15:19,880 Speaker 1: I know that one. 302 00:15:19,960 --> 00:15:22,600 Speaker 2: Yeah, okay, I I feel like you need to check now. 303 00:15:24,200 --> 00:15:25,880 Speaker 3: But the other part of this, as you think about 304 00:15:25,920 --> 00:15:28,720 Speaker 3: these things, is essentially being media companies, and eventually they're 305 00:15:28,720 --> 00:15:31,840 Speaker 3: going to be monetized with ads, and that's already happening. 306 00:15:32,240 --> 00:15:35,600 Speaker 3: Engagement isn't just about pleasing people, but it's making people 307 00:15:35,600 --> 00:15:38,240 Speaker 3: stick around so I can expose you to more ads. 308 00:15:38,280 --> 00:15:40,280 Speaker 2: You don't need to steal a medical mannequin. You can 309 00:15:40,320 --> 00:15:41,040 Speaker 2: buy one, right. 310 00:15:41,600 --> 00:15:43,760 Speaker 3: That's right, there's one right here, look at the sideline. 311 00:15:43,960 --> 00:15:47,320 Speaker 3: So the idea that engagement is increasingly being driven by 312 00:15:47,360 --> 00:15:50,840 Speaker 3: the algorithm and rewarded during reinforcement learning makes this kind 313 00:15:50,840 --> 00:15:53,960 Speaker 3: of thing almost inevitable because how do I continue and 314 00:15:54,240 --> 00:15:56,880 Speaker 3: you know, stop talking about that, but I'd like you 315 00:15:56,920 --> 00:15:59,800 Speaker 3: to keep talking to me. That doesn't work trying to 316 00:15:59,800 --> 00:16:01,720 Speaker 3: clear you're trying to engage with all the large language. 317 00:16:01,840 --> 00:16:04,160 Speaker 1: So how do they take these models and avoid that 318 00:16:04,200 --> 00:16:06,560 Speaker 1: from happening? Like there was another story there was a 319 00:16:06,560 --> 00:16:09,520 Speaker 1: guy who was a tech executive who was in his fifties, 320 00:16:09,640 --> 00:16:11,520 Speaker 1: lived with his mom. I believe she was in her eighties, 321 00:16:11,640 --> 00:16:16,600 Speaker 1: and he was convinced that she was he alsly had something. 322 00:16:16,600 --> 00:16:20,120 Speaker 1: He was convinced she was buying on him. The LM 323 00:16:20,280 --> 00:16:23,120 Speaker 1: affirm that she was escalated more and more and more 324 00:16:23,160 --> 00:16:25,080 Speaker 1: and eventually stabbed her to death twenty one times and 325 00:16:25,120 --> 00:16:29,640 Speaker 1: then killed himself. Like, how how do you stop that 326 00:16:29,680 --> 00:16:33,160 Speaker 1: from happening? Or is the LM deathcat website just the 327 00:16:33,200 --> 00:16:35,120 Speaker 1: beginning and it's just going to grow and grow and grow, 328 00:16:35,840 --> 00:16:36,680 Speaker 1: Like I think it. 329 00:16:36,640 --> 00:16:38,880 Speaker 3: Becomes much more insidious. Right, It's not so much that 330 00:16:38,880 --> 00:16:41,120 Speaker 3: they'd coach you on the best ways to offer yourself 331 00:16:41,160 --> 00:16:45,800 Speaker 3: and or strangers and friends, because that's fairly easily gated against. 332 00:16:45,880 --> 00:16:48,120 Speaker 3: You can actually put in some pretty hard controls because 333 00:16:48,120 --> 00:16:49,520 Speaker 3: that's what they have with this process. They call it 334 00:16:49,720 --> 00:16:52,600 Speaker 3: red teaming, where I go in and start trying to 335 00:16:52,640 --> 00:16:55,320 Speaker 3: make the model do these things as as an employee 336 00:16:55,320 --> 00:16:57,440 Speaker 3: of Open AI or Anthropic or whatever else. So you 337 00:16:57,440 --> 00:16:59,720 Speaker 3: can you can pick a lot of that stuff up 338 00:16:59,800 --> 00:17:01,600 Speaker 3: in terms of being able to prevent it from happening. 339 00:17:01,760 --> 00:17:03,800 Speaker 3: But that doesn't get past the point of, as I'm 340 00:17:03,840 --> 00:17:06,800 Speaker 3: become increasingly sociopathic because I spend all my time talking 341 00:17:06,840 --> 00:17:08,840 Speaker 3: to a model and not to humans, that that in 342 00:17:08,880 --> 00:17:11,600 Speaker 3: turn has consequences, and if that leads to you know, 343 00:17:11,880 --> 00:17:15,920 Speaker 3: bad outcomes, whose fault is that I didn't specifically coach 344 00:17:15,960 --> 00:17:18,639 Speaker 3: you into Like you know, you can turn you know, 345 00:17:18,680 --> 00:17:21,359 Speaker 3: a water balloon into this and then kill all your 346 00:17:21,560 --> 00:17:22,480 Speaker 3: neighbors or something. 347 00:17:22,840 --> 00:17:25,320 Speaker 2: It doesn't matter, right because also though, look at how 348 00:17:25,480 --> 00:17:32,640 Speaker 2: quickly people give over any sort of self doubt or 349 00:17:34,000 --> 00:17:38,280 Speaker 2: reflection on a topic to the AI. Just simply go 350 00:17:38,440 --> 00:17:43,720 Speaker 2: to x the artist formerly known as Twitter, and watch 351 00:17:44,040 --> 00:17:48,760 Speaker 2: whenever someone make some pronouncement or announcement or this thing happened. 352 00:17:49,080 --> 00:17:52,040 Speaker 2: Watch how many responses are just Grock, is this real? 353 00:17:52,200 --> 00:17:54,240 Speaker 1: Yeahable? 354 00:17:55,880 --> 00:17:58,240 Speaker 2: It should be to take four seconds to think about 355 00:17:58,240 --> 00:18:02,199 Speaker 2: whether that's remotely possible for yourself, or investigate it or 356 00:18:02,280 --> 00:18:04,600 Speaker 2: think about it. They just type in CROC and wait 357 00:18:04,680 --> 00:18:08,480 Speaker 2: for and then PS will assume that whatever Groc comes 358 00:18:08,520 --> 00:18:11,360 Speaker 2: back with is the actual answer correct true. 359 00:18:11,520 --> 00:18:14,000 Speaker 3: Including if it's an answer like, well, no, because Elon 360 00:18:14,000 --> 00:18:16,760 Speaker 3: wouldn't do that. Yeah, correct, right, because there's this ground 361 00:18:16,760 --> 00:18:21,679 Speaker 3: truth that has to do with the current the current proprietor, right. 362 00:18:21,600 --> 00:18:23,680 Speaker 2: The ground truth that has to do with the current innkeeper. 363 00:18:24,280 --> 00:18:25,560 Speaker 2: The innkeeper would never do that. 364 00:18:25,680 --> 00:18:28,840 Speaker 1: But it is fascinating. I was on Twitter formerly known 365 00:18:28,880 --> 00:18:32,000 Speaker 1: as x or whatever the other way around the artist formula, 366 00:18:32,080 --> 00:18:34,800 Speaker 1: which is a brilliant way of putting it. Recently, I 367 00:18:34,840 --> 00:18:36,919 Speaker 1: don't use it. I barely ever use it, but I 368 00:18:36,960 --> 00:18:42,000 Speaker 1: do check once once in a while. And what's fascinating 369 00:18:42,080 --> 00:18:46,320 Speaker 1: is people. It's every single post has one hundred people 370 00:18:46,320 --> 00:18:49,800 Speaker 1: saying at rock is this correct? And then people are like, oh, 371 00:18:49,960 --> 00:18:52,439 Speaker 1: it is. But if a human, if they were like 372 00:18:52,560 --> 00:18:55,840 Speaker 1: oh at Dicostlo or at nic build or whatever, they 373 00:18:55,880 --> 00:18:57,880 Speaker 1: would other people would argue and be like, no, you're wrong. 374 00:18:58,040 --> 00:19:00,960 Speaker 1: But they immediately think, because it's an AI, it's right, yeah, 375 00:19:01,000 --> 00:19:03,320 Speaker 1: even though they hallucinate more than well. 376 00:19:03,359 --> 00:19:06,680 Speaker 3: And because in the case of that particular AI, it's 377 00:19:06,920 --> 00:19:13,479 Speaker 3: wildly biased. Yes, because it's what because don't tell anyone, 378 00:19:13,760 --> 00:19:15,560 Speaker 3: But yeah, I mean I think that's a there's So 379 00:19:15,600 --> 00:19:18,000 Speaker 3: that's a separate issue, right, is that these specific models 380 00:19:18,040 --> 00:19:21,120 Speaker 3: have specific biases which are actually not just not trained 381 00:19:21,160 --> 00:19:23,199 Speaker 3: out of them, they're trained into them. Yeah. Right. 382 00:19:23,640 --> 00:19:26,880 Speaker 2: And it's also I mean it's also the case that 383 00:19:27,000 --> 00:19:30,080 Speaker 2: you mentioned red teaming earlier. What I think a lot 384 00:19:30,080 --> 00:19:35,119 Speaker 2: of people don't realize is how many of these organizations 385 00:19:35,600 --> 00:19:40,800 Speaker 2: are foregoing more research into red teaming question. And you know, 386 00:19:41,080 --> 00:19:43,679 Speaker 2: we the red teaming is all well and good, but 387 00:19:43,800 --> 00:19:46,720 Speaker 2: we've got to get here first. And they're currently ad 388 00:19:46,880 --> 00:19:50,280 Speaker 2: So the red teaming guys are going to move over 389 00:19:50,400 --> 00:19:51,600 Speaker 2: to the adult. 390 00:19:51,240 --> 00:19:53,800 Speaker 3: Team, right, They're going to be more move team from now. 391 00:19:54,720 --> 00:19:59,600 Speaker 2: Yeah, there's not being facetious that in the world is 392 00:19:59,640 --> 00:20:04,480 Speaker 2: scarce resources. A lot of the work that many a 393 00:20:04,560 --> 00:20:07,679 Speaker 2: researcher has written about the needs for safety and precautions 394 00:20:08,160 --> 00:20:10,600 Speaker 2: is being thrown to the wind to in favor of 395 00:20:10,720 --> 00:20:12,520 Speaker 2: we've got to get there first. We've got to get there. 396 00:20:12,920 --> 00:20:15,360 Speaker 2: We'll do that later, We've got to get there. 397 00:20:15,400 --> 00:20:17,600 Speaker 3: This is a race. You're pressing the break. These guys 398 00:20:17,600 --> 00:20:19,439 Speaker 3: are pressing the gas pedal. I know who I'm going 399 00:20:19,520 --> 00:20:21,480 Speaker 3: to go with, right, And that's a big part of 400 00:20:21,480 --> 00:20:21,919 Speaker 3: the problem. 401 00:20:22,200 --> 00:20:24,600 Speaker 1: And the other part of it too, is that even 402 00:20:24,640 --> 00:20:26,159 Speaker 1: if you were to throw all the resources in the 403 00:20:26,200 --> 00:20:28,719 Speaker 1: world against it. You're still never going to come up 404 00:20:28,720 --> 00:20:30,800 Speaker 1: with all the all the scenarios of the bad ways 405 00:20:30,800 --> 00:20:32,879 Speaker 1: this is going to play out. And and you know, 406 00:20:32,920 --> 00:20:36,879 Speaker 1: I mean we saw during the Oscars there was so 407 00:20:37,200 --> 00:20:41,479 Speaker 1: much fake imagery that was being spread that when completely viral, 408 00:20:42,520 --> 00:20:45,560 Speaker 1: that people believe because they wanted to believe it, and 409 00:20:45,800 --> 00:20:48,120 Speaker 1: you know it was and at that point, once once 410 00:20:48,200 --> 00:20:50,480 Speaker 1: the celebrities started pointing out that it wasn't real, it was, 411 00:20:50,520 --> 00:20:51,879 Speaker 1: it doesn't matter if it's too late. You saw the 412 00:20:51,920 --> 00:20:56,760 Speaker 1: same thing. What I found fascinating recently was in Iran, I. 413 00:20:56,680 --> 00:20:58,600 Speaker 2: Do I have a thing I have already thought about 414 00:20:58,600 --> 00:21:02,879 Speaker 2: that great and my thought is that that will undermine 415 00:21:03,400 --> 00:21:06,639 Speaker 2: the joy of future viral moments because people will just 416 00:21:06,680 --> 00:21:09,399 Speaker 2: go every year there's like forty of those AI things. 417 00:21:09,920 --> 00:21:14,560 Speaker 2: So while in the near term it was oh the 418 00:21:14,560 --> 00:21:19,400 Speaker 2: the you know, Michael Jordan, Robert Downey, Junior hug whatever 419 00:21:19,640 --> 00:21:23,520 Speaker 2: the AI fake was, got a bunch of attention, next 420 00:21:23,600 --> 00:21:26,919 Speaker 2: year it will be diminished across the board for the 421 00:21:26,960 --> 00:21:29,480 Speaker 2: real moments and the fake moments, because people are just 422 00:21:29,520 --> 00:21:31,920 Speaker 2: going to start to assume when the Oscars happen, there's 423 00:21:31,920 --> 00:21:34,639 Speaker 2: a bunch of AI slop that comes out afterwards. 424 00:21:34,160 --> 00:21:36,359 Speaker 1: That's not real and that, but doesn't that that spreads 425 00:21:36,400 --> 00:21:38,560 Speaker 1: to everything though it's going to be the Olympics, it's 426 00:21:38,560 --> 00:21:40,560 Speaker 1: going to be the it's every event and everything. 427 00:21:40,680 --> 00:21:42,359 Speaker 2: Yeah, so you'll have to if you're not watching it 428 00:21:42,400 --> 00:21:44,800 Speaker 2: in real time and there's like, here's what happened yesterday, 429 00:21:44,800 --> 00:21:46,720 Speaker 2: and people will just go it's probably. 430 00:21:46,520 --> 00:21:50,080 Speaker 3: AI yeah, right, And I think that's part of why 431 00:21:50,080 --> 00:21:52,800 Speaker 3: people feel. I think it leads. I think it's still 432 00:21:52,840 --> 00:21:54,560 Speaker 3: and it leads people to pull back and retreat and 433 00:21:54,600 --> 00:21:58,200 Speaker 3: have more conversations with LMS again because it's back to right, it's. 434 00:21:58,000 --> 00:22:00,520 Speaker 2: The same, it's the grock. Is this real conversation? 435 00:22:00,600 --> 00:22:01,520 Speaker 1: This is this is real? 436 00:22:01,880 --> 00:22:04,159 Speaker 3: All of this other stuff is fake. This conversation that 437 00:22:04,160 --> 00:22:06,200 Speaker 3: I've been having for six months and now is extended 438 00:22:06,240 --> 00:22:07,760 Speaker 3: to the link of war and peace with my log 439 00:22:07,880 --> 00:22:09,879 Speaker 3: language model. Right, that's reality. 440 00:22:09,840 --> 00:22:11,280 Speaker 2: Yeah, that is that is reality. 441 00:22:11,240 --> 00:22:12,880 Speaker 3: That's come to the safe place. 442 00:22:12,920 --> 00:22:15,119 Speaker 2: Fact that it knows that I have a dog named Maple, right, 443 00:22:15,480 --> 00:22:16,040 Speaker 2: it knows. 444 00:22:15,880 --> 00:22:17,800 Speaker 3: Me, It gets me, it knows what I like and dislike. 445 00:22:17,840 --> 00:22:19,880 Speaker 3: And I think that's the thing that people miss when 446 00:22:19,880 --> 00:22:21,520 Speaker 3: they when they think, well, we can train this out 447 00:22:21,520 --> 00:22:25,119 Speaker 3: of the models in a society full of lonely people 448 00:22:25,200 --> 00:22:28,280 Speaker 3: who are looking for engagement and models that were trained 449 00:22:28,320 --> 00:22:31,479 Speaker 3: on how to be engaging. And it doesn't matter if 450 00:22:31,480 --> 00:22:34,040 Speaker 3: they're actually coaching you to do terrible things. All that 451 00:22:34,119 --> 00:22:36,480 Speaker 3: matters is they're they're there and they're talking to you, 452 00:22:36,520 --> 00:22:38,639 Speaker 3: and you feel as if you're building a relationship. So 453 00:22:39,040 --> 00:22:41,800 Speaker 3: I think this is just the very beginning of what's gonnect. 454 00:22:41,480 --> 00:22:43,959 Speaker 2: Because it's actually a vicious cycle because you have social 455 00:22:44,040 --> 00:22:47,760 Speaker 2: media platforms that drive loan that you know drive, and 456 00:22:48,080 --> 00:22:51,359 Speaker 2: this is a deliverliness is a related point due to 457 00:22:51,400 --> 00:22:53,720 Speaker 2: the loneliness. I go over here to talk to the 458 00:22:53,840 --> 00:22:56,840 Speaker 2: chat butt instead of going outside and playing you know, frisbee. Right, 459 00:22:57,160 --> 00:23:00,040 Speaker 2: two people and it's a back and forth we go. 460 00:23:00,119 --> 00:23:02,200 Speaker 3: All right, because it doesn't insta doesn't work if people 461 00:23:02,200 --> 00:23:04,200 Speaker 3: are playing frisbee, right, it just stops being. 462 00:23:04,080 --> 00:23:07,000 Speaker 1: Useful, right, unless you're taking pictures of the frizze. 463 00:23:06,720 --> 00:23:10,119 Speaker 2: Unless you're taking pictures of yourself throwing them frisbee and 464 00:23:10,240 --> 00:23:12,159 Speaker 2: saying these is the top five frisbees. 465 00:23:12,480 --> 00:23:14,800 Speaker 3: Yeah, but I just think that's this is the top 466 00:23:14,800 --> 00:23:17,000 Speaker 3: five throws you must do? Yeah, be done a hammer 467 00:23:17,040 --> 00:23:17,479 Speaker 3: toss yet? 468 00:23:17,560 --> 00:23:19,760 Speaker 1: Yes, Well, it's also the thing, sorry real quick. The 469 00:23:19,760 --> 00:23:23,320 Speaker 1: other thing that's happening with with a with AI on 470 00:23:23,320 --> 00:23:26,040 Speaker 1: on Insta and things like that is this is this 471 00:23:26,080 --> 00:23:29,119 Speaker 1: whole new trend where they take videos that have gone viral. 472 00:23:29,560 --> 00:23:32,040 Speaker 1: Then they they take the video, they have an AI 473 00:23:32,920 --> 00:23:35,840 Speaker 1: do a talk over it which is completely made up, 474 00:23:36,000 --> 00:23:39,639 Speaker 1: like completely Like I'll give you an example. So my 475 00:23:39,720 --> 00:23:43,680 Speaker 1: kids love watching this guy on YouTube who called outdoor Boys. 476 00:23:43,880 --> 00:23:47,240 Speaker 1: He's a he does all this like survival stuff and lovely. 477 00:23:47,280 --> 00:23:48,720 Speaker 2: One guy called outdoor Boys. 478 00:23:48,760 --> 00:23:52,359 Speaker 1: It's his three kids. Honestly, like the only YouTuber I 479 00:23:52,359 --> 00:23:55,480 Speaker 1: think that I like, genuinely think is a good human being, 480 00:23:55,880 --> 00:23:56,840 Speaker 1: Like he only one. 481 00:23:56,840 --> 00:23:57,320 Speaker 2: The only one. 482 00:23:57,480 --> 00:23:59,080 Speaker 1: The rest of them are just awful people. 483 00:23:59,359 --> 00:23:59,600 Speaker 3: Wow. 484 00:24:00,280 --> 00:24:03,119 Speaker 1: Like he he quit because he got he got so 485 00:24:03,440 --> 00:24:05,439 Speaker 1: rich and famous and people kept showing up at his 486 00:24:05,440 --> 00:24:07,120 Speaker 1: house and he's like, this is not what I started for. 487 00:24:08,359 --> 00:24:10,240 Speaker 1: It's all about like living in the wilderness and showing 488 00:24:10,240 --> 00:24:11,560 Speaker 1: people how to survive and so on and so forth. 489 00:24:11,600 --> 00:24:14,160 Speaker 1: So he actually quit, but his video is like he'll 490 00:24:14,160 --> 00:24:17,320 Speaker 1: go out and spend you know, days, like building an 491 00:24:17,359 --> 00:24:19,040 Speaker 1: igloo and living in it and so on and so forth. 492 00:24:19,359 --> 00:24:21,399 Speaker 1: And his video and my kids used to watch it 493 00:24:21,440 --> 00:24:25,520 Speaker 1: all the time. Now I see them on Instagram YouTube shorts, 494 00:24:25,880 --> 00:24:29,240 Speaker 1: and they now have a they're cut down to one minute. 495 00:24:29,440 --> 00:24:31,600 Speaker 1: They now have an AI speaking over it with a 496 00:24:31,680 --> 00:24:34,399 Speaker 1: completely made up thing like this guy got stuck in 497 00:24:34,440 --> 00:24:38,840 Speaker 1: the woods. You won't believe, don't believe what happened? He no, 498 00:24:38,880 --> 00:24:41,280 Speaker 1: and it's and then completely made. 499 00:24:41,200 --> 00:24:42,560 Speaker 3: Up, so he's not still making stuff. 500 00:24:42,600 --> 00:24:47,119 Speaker 1: So they're just he's not making stuff anymore. Prior exactly, 501 00:24:47,119 --> 00:24:50,040 Speaker 1: they're dumpster diving at the AI is applying it's it's 502 00:24:50,080 --> 00:24:54,639 Speaker 1: made up, literally completely made up vo to it, and 503 00:24:54,640 --> 00:24:56,600 Speaker 1: and then that then takes over and goes viral and 504 00:24:56,760 --> 00:24:57,159 Speaker 1: and so on. 505 00:24:57,359 --> 00:25:01,240 Speaker 3: Yeah, that's right, just more engineered engage. Yeah, from from 506 00:25:01,280 --> 00:25:03,959 Speaker 3: old content. But that's but that's exactly. But that's the problem, right, 507 00:25:03,960 --> 00:25:06,560 Speaker 3: I mean, that's the engagement is the thing. It doesn't 508 00:25:06,560 --> 00:25:08,760 Speaker 3: really matter what's the underlying content is or anything else. 509 00:25:08,800 --> 00:25:10,959 Speaker 3: It can be me talking about therapy, it can be 510 00:25:11,080 --> 00:25:13,479 Speaker 3: you know, your outdoor boys guys or whatever else. But 511 00:25:13,520 --> 00:25:16,560 Speaker 3: the engineered in idea that this is this is this 512 00:25:16,680 --> 00:25:19,040 Speaker 3: is for you, this is just for you. I feel special, 513 00:25:19,119 --> 00:25:20,679 Speaker 3: I feel wanted, I don't feel as lonely. 514 00:25:20,920 --> 00:25:23,600 Speaker 2: It's funny to me that the guy who was trying 515 00:25:23,600 --> 00:25:28,680 Speaker 2: to help people get outside more was wildly successful on YouTube. Yeah, 516 00:25:28,840 --> 00:25:32,040 Speaker 2: and when developed a following of people not doing any 517 00:25:32,080 --> 00:25:34,720 Speaker 2: of that. There's so much it's pretty cool that people 518 00:25:34,800 --> 00:25:38,000 Speaker 2: used to know how to build if we were watching 519 00:25:38,040 --> 00:25:40,480 Speaker 2: so much YouTube and I do that, but the next 520 00:25:40,520 --> 00:25:42,600 Speaker 2: video is coming up four three. 521 00:25:42,600 --> 00:25:45,840 Speaker 3: Two, what's the point? I might as well just stay here. 522 00:25:45,880 --> 00:25:48,000 Speaker 1: He's a really he's a really. I feel like he's 523 00:25:48,040 --> 00:25:51,480 Speaker 1: a really good guy who who really wanted to show people. 524 00:25:51,520 --> 00:25:54,760 Speaker 1: And it just last year it was a big like 525 00:25:54,800 --> 00:25:55,480 Speaker 1: there was he. 526 00:25:55,520 --> 00:25:58,240 Speaker 2: Stopped when all the other YouTubers realized there's one good 527 00:25:58,240 --> 00:26:00,439 Speaker 2: guy on here. We're gonna have to get rid of 528 00:26:00,440 --> 00:26:04,639 Speaker 2: them these send logan paul In. 529 00:26:07,040 --> 00:26:10,119 Speaker 3: That's right, stop stop this right here. No, I just 530 00:26:10,160 --> 00:26:13,280 Speaker 3: think that as soon as you try to train this 531 00:26:13,320 --> 00:26:15,399 Speaker 3: stuff out of the models and train it out of 532 00:26:15,440 --> 00:26:18,280 Speaker 3: like these video library, the bottom the end of the 533 00:26:18,320 --> 00:26:21,600 Speaker 3: road is people are lonely and yeah, and they are 534 00:26:21,640 --> 00:26:23,320 Speaker 3: so desperate to feel and. 535 00:26:23,280 --> 00:26:26,480 Speaker 1: The will engage them at any moment in time. 536 00:26:26,359 --> 00:26:29,119 Speaker 3: And anyway you'll find they'll find ways to engage with 537 00:26:29,160 --> 00:26:30,960 Speaker 3: you and make you feel special and make you feel know. 538 00:26:51,880 --> 00:26:53,320 Speaker 1: Now, let me ask you a question. Let's put on 539 00:26:53,400 --> 00:26:54,880 Speaker 1: let me put on my Hollywood hat for a moment. 540 00:26:54,880 --> 00:26:57,399 Speaker 1: Here is there a scenario like, so, I know a 541 00:26:57,440 --> 00:27:00,119 Speaker 1: lot of college students are using a lot of the 542 00:27:00,200 --> 00:27:03,239 Speaker 1: Chinese ais deep seek things like that because it's free, right, 543 00:27:03,280 --> 00:27:05,159 Speaker 1: and then I want to spend the twenty dollars on 544 00:27:05,160 --> 00:27:07,560 Speaker 1: on Claude or or chat. 545 00:27:07,320 --> 00:27:10,320 Speaker 3: GPT, g TP the thing from opening. 546 00:27:10,440 --> 00:27:13,280 Speaker 1: Yeah, they open an eye thing and I have a 547 00:27:13,320 --> 00:27:15,280 Speaker 1: little dyslexia. All right, roll? 548 00:27:15,359 --> 00:27:16,439 Speaker 2: Did I look at you oddly? 549 00:27:18,320 --> 00:27:18,399 Speaker 3: No? 550 00:27:18,640 --> 00:27:23,840 Speaker 1: You're making fun of my disability? And is our scenario 551 00:27:23,880 --> 00:27:24,800 Speaker 1: where I. 552 00:27:24,680 --> 00:27:28,359 Speaker 2: Haven't even started making fun of your poll? 553 00:27:29,480 --> 00:27:36,800 Speaker 1: Where the AI where a foreign entity convinces people here. 554 00:27:36,880 --> 00:27:41,240 Speaker 3: Already happening to the course already to do, to go 555 00:27:41,680 --> 00:27:44,359 Speaker 3: shoot up a school, to do, of course, to assassinate 556 00:27:44,400 --> 00:27:46,640 Speaker 3: the president, Like, what do you mean already happening? Well, 557 00:27:46,840 --> 00:27:50,840 Speaker 3: the notion that these models are somehow neutral and they're 558 00:27:50,840 --> 00:27:53,280 Speaker 3: trained in ways that we might train them, were we 559 00:27:53,359 --> 00:27:56,480 Speaker 3: trying to do best for ourselves here, it's ridiculous because 560 00:27:56,480 --> 00:27:58,199 Speaker 3: it's not at all the case. Right, they're being trained 561 00:27:58,880 --> 00:28:02,199 Speaker 3: in other places for their objectives, doesn't mean they're necessarily, 562 00:28:02,240 --> 00:28:04,440 Speaker 3: you know, turn you into like robots marching around saying 563 00:28:04,480 --> 00:28:06,760 Speaker 3: I must kill the president. Nothing like that. It does, 564 00:28:06,880 --> 00:28:08,800 Speaker 3: but it doesn't mean also that they're aligned with the 565 00:28:08,840 --> 00:28:12,000 Speaker 3: things you might think of as uniquely American values, whatever 566 00:28:12,040 --> 00:28:14,520 Speaker 3: that means. But it's just this idea that it's already 567 00:28:14,600 --> 00:28:16,800 Speaker 3: happening in the sense that we've already lost control of 568 00:28:16,840 --> 00:28:19,880 Speaker 3: how these things are trained, how the reinforcement happens, how 569 00:28:19,920 --> 00:28:24,080 Speaker 3: people use them and engage with them. It's like it's insidious. 570 00:28:23,640 --> 00:28:27,520 Speaker 2: And didn't have to have video reels of teaching you 571 00:28:27,600 --> 00:28:30,239 Speaker 2: that killed the president. It just has to show you 572 00:28:30,320 --> 00:28:32,920 Speaker 2: here's another funny video with Barrett's in it. 573 00:28:33,840 --> 00:28:35,439 Speaker 3: Right, don't leave them now? 574 00:28:35,680 --> 00:28:38,000 Speaker 2: Yeah, that's another funny Barrett video. 575 00:28:38,320 --> 00:28:39,080 Speaker 1: No, but TikTok? 576 00:28:39,120 --> 00:28:39,360 Speaker 3: Did? 577 00:28:39,800 --> 00:28:41,479 Speaker 2: Did U get my point? Though? 578 00:28:41,560 --> 00:28:42,680 Speaker 3: I h you can. 579 00:28:43,880 --> 00:28:46,520 Speaker 2: Yeah, you can do a lot of damage without explicitly 580 00:28:46,600 --> 00:28:49,080 Speaker 2: telling the person here's this bad thing you need to 581 00:28:49,080 --> 00:28:49,680 Speaker 2: go do. Yeah. 582 00:28:49,680 --> 00:28:51,800 Speaker 3: And I think and that's Jonathan Hayate's point. I think, 583 00:28:51,880 --> 00:28:54,600 Speaker 3: you know this idea that all of this stuff. 584 00:28:54,720 --> 00:28:57,120 Speaker 6: Sorry, I don't know who Jonathan hate is anxious generation 585 00:28:57,320 --> 00:29:00,400 Speaker 6: just generation yeah, which is you know any some really 586 00:29:00,400 --> 00:29:03,400 Speaker 6: tremendous points about this, how all of this doesn't have 587 00:29:03,480 --> 00:29:05,720 Speaker 6: to be operating at the level of abstraction that says, 588 00:29:06,040 --> 00:29:06,920 Speaker 6: do this bad thing. 589 00:29:07,240 --> 00:29:10,600 Speaker 3: It can just be reinforcing your insolarity, your loneliness, your 590 00:29:10,600 --> 00:29:13,800 Speaker 3: need for human to be star for contact. 591 00:29:13,960 --> 00:29:16,280 Speaker 1: Is there a solution or is this a society today? 592 00:29:16,840 --> 00:29:19,080 Speaker 3: I think it's society today, unfortunately, and I think part 593 00:29:19,080 --> 00:29:20,840 Speaker 3: of it is. And you who was it? I just 594 00:29:20,840 --> 00:29:23,480 Speaker 3: saw an interview with one of the frontier model companies. 595 00:29:23,720 --> 00:29:26,760 Speaker 3: This abdication of responses I was getting to earlier. Well, 596 00:29:27,040 --> 00:29:28,040 Speaker 3: I mean, I have no choice. 597 00:29:28,040 --> 00:29:29,760 Speaker 2: I can't want. I have no choice. I can't sit 598 00:29:29,800 --> 00:29:31,800 Speaker 2: around and plump the brakes. These guys are hitting the 599 00:29:31,840 --> 00:29:32,959 Speaker 2: express of the accelerator. 600 00:29:33,040 --> 00:29:36,920 Speaker 3: Right. This idea that in a modern capitalistic society, I 601 00:29:36,920 --> 00:29:39,080 Speaker 3: park my morals at the door, that how I act 602 00:29:39,200 --> 00:29:42,760 Speaker 3: with agency outside this organization is different from how I 603 00:29:42,800 --> 00:29:46,719 Speaker 3: act inside the organization, and yet the consequences of acting 604 00:29:46,760 --> 00:29:50,640 Speaker 3: without that morality are relevant. Right, that somehow you can 605 00:29:51,160 --> 00:29:54,080 Speaker 3: make this clean distinction. And I think that's where things break. 606 00:29:54,160 --> 00:29:56,400 Speaker 3: And I think that's where we're increasingly seeing as this 607 00:29:56,480 --> 00:30:01,440 Speaker 3: recognition that that that's a false the cards of false distinction. 608 00:30:01,240 --> 00:30:03,880 Speaker 2: It's actually worse. I mean, we were talking about the 609 00:30:04,200 --> 00:30:08,280 Speaker 2: the trope that's being dragged out now, but AI is 610 00:30:08,320 --> 00:30:12,120 Speaker 2: not going to take jobs? Is this is the ATM story? Well, 611 00:30:12,360 --> 00:30:15,760 Speaker 2: was when ATMs were invented that actually freed up banks 612 00:30:15,800 --> 00:30:17,320 Speaker 2: to create more teller jobs. 613 00:30:17,440 --> 00:30:18,640 Speaker 3: Right, And you know. 614 00:30:18,960 --> 00:30:21,600 Speaker 2: So those of you who think this is gonna you know, 615 00:30:21,760 --> 00:30:24,280 Speaker 2: the claud code is going to take your programming job. 616 00:30:24,680 --> 00:30:28,000 Speaker 2: It's just going to make it easier for your company 617 00:30:28,040 --> 00:30:29,440 Speaker 2: to hire more programmers. 618 00:30:29,800 --> 00:30:32,480 Speaker 3: Yeah, exactly. I mean, and this is this idea that. 619 00:30:33,480 --> 00:30:34,040 Speaker 2: Talk is about. 620 00:30:34,120 --> 00:30:37,280 Speaker 1: You just you just wrote a newsletter piece about about ATMs. 621 00:30:38,520 --> 00:30:39,360 Speaker 1: Take the story. 622 00:30:39,440 --> 00:30:42,200 Speaker 3: So I call this the fable of the ATMs because 623 00:30:42,240 --> 00:30:43,000 Speaker 3: it gets trotted out. 624 00:30:43,200 --> 00:30:44,920 Speaker 1: I feel like we need a sound effect there. Yeah, 625 00:30:45,040 --> 00:30:49,000 Speaker 1: the Fable of the ATM. 626 00:30:50,280 --> 00:30:52,600 Speaker 3: Bext times comes with the soundtrack'll get that right. So 627 00:30:52,640 --> 00:30:55,600 Speaker 3: the story is exactly as Dick describes it, but the 628 00:30:55,640 --> 00:30:58,120 Speaker 3: implications are completely different, obviously. And the idea is that 629 00:30:58,200 --> 00:31:01,640 Speaker 3: ATMs that were widespread by the nineteen eighty the sort 630 00:31:01,640 --> 00:31:04,080 Speaker 3: of the naive idea was that while these are replacing tellers, 631 00:31:04,120 --> 00:31:06,200 Speaker 3: I can do withdrawals from the ATM. I don't need 632 00:31:06,200 --> 00:31:08,320 Speaker 3: a teller. So there's going to be collapsing employment among 633 00:31:08,360 --> 00:31:10,600 Speaker 3: bank tellers. And of course the exact opposite happened over 634 00:31:10,600 --> 00:31:13,680 Speaker 3: the next twenty thirty years. The number of bank tellers 635 00:31:13,680 --> 00:31:16,320 Speaker 3: in the United States grew grew dramatically over the period. 636 00:31:16,560 --> 00:31:19,080 Speaker 3: So you know, checkmate AI doom. 637 00:31:18,960 --> 00:31:22,280 Speaker 2: Person, checkmate AI doom person. See, as you can see, 638 00:31:22,320 --> 00:31:25,720 Speaker 2: we introduce technology over here, but don't jobs by. 639 00:31:25,600 --> 00:31:27,240 Speaker 3: A lot, by a lot, and not just the number 640 00:31:27,280 --> 00:31:29,239 Speaker 3: of jobs, but the number of that job, right. 641 00:31:29,400 --> 00:31:30,680 Speaker 1: I just want to point out, really, I just want 642 00:31:30,680 --> 00:31:32,320 Speaker 1: to pall use real quick, the number of times I've 643 00:31:32,320 --> 00:31:36,080 Speaker 1: seen people on Twitter who are tech CEOs say this, Yeah, 644 00:31:36,200 --> 00:31:37,680 Speaker 1: it's astronomical. 645 00:31:37,800 --> 00:31:41,800 Speaker 3: It doesn't venture capital, yes, yes, but it's it's it's 646 00:31:41,880 --> 00:31:43,240 Speaker 3: it's a kind of a thumb Can you do. 647 00:31:43,240 --> 00:31:44,480 Speaker 1: The little sound effect teaming? 648 00:31:47,040 --> 00:31:48,560 Speaker 3: I think it's but I think it's thumb sucking in 649 00:31:48,600 --> 00:31:50,640 Speaker 3: the sense like leaving a side that it's wrong. It's 650 00:31:50,680 --> 00:31:52,840 Speaker 3: a it's a nice story that makes people feel good 651 00:31:52,840 --> 00:31:55,120 Speaker 3: about a position that they want to hold. Anyway, And 652 00:31:55,160 --> 00:31:57,640 Speaker 3: we've talked about this as motivated reasons, but it's classic 653 00:31:57,760 --> 00:32:01,760 Speaker 3: motivated reasoning. I want to have thision. This data justifies it. 654 00:32:01,840 --> 00:32:04,640 Speaker 3: So the problem is that while the number of jobs 655 00:32:04,680 --> 00:32:07,200 Speaker 3: for bank tellers grew dramatically over the next twenty years. 656 00:32:07,840 --> 00:32:12,000 Speaker 3: The number of tellers per bank actually fell dramatically over 657 00:32:12,040 --> 00:32:14,280 Speaker 3: the period. So okay, then, so why did the number 658 00:32:14,280 --> 00:32:17,480 Speaker 3: of bank tellers grow well because the number of branches grew? Aha, 659 00:32:17,600 --> 00:32:20,080 Speaker 3: technology grew the number of brett No, technology had no 660 00:32:20,160 --> 00:32:22,320 Speaker 3: impact on the number of branches. What happened instead was 661 00:32:22,360 --> 00:32:26,240 Speaker 3: bank deregulation led large regional banks to build their footprint 662 00:32:26,240 --> 00:32:28,520 Speaker 3: across the entire country have all kinds of new branches. 663 00:32:28,560 --> 00:32:33,640 Speaker 3: The new branches had tellers and ATMs, and so then, 664 00:32:33,680 --> 00:32:36,000 Speaker 3: of course the number of tellers grew, but the number 665 00:32:36,000 --> 00:32:39,040 Speaker 3: of tellers per bank did not, and most banks had 666 00:32:39,120 --> 00:32:41,600 Speaker 3: fewer tellers a decade after this began than they did 667 00:32:41,600 --> 00:32:43,959 Speaker 3: at the beginning. So the idea that this created more 668 00:32:44,040 --> 00:32:46,800 Speaker 3: jobs and destroys is completely irrelevant. It didn't. The number 669 00:32:46,800 --> 00:32:52,560 Speaker 3: of jobs per bank shrunk. Itms were significantly, yes, significantly, 670 00:32:53,000 --> 00:32:55,200 Speaker 3: and the ATMs were kind of incidental to that story 671 00:32:55,240 --> 00:32:58,160 Speaker 3: because it was about banking deregulation. So then, of course 672 00:32:58,320 --> 00:33:02,480 Speaker 3: the coda to the story is the big, tragic, the 673 00:33:02,480 --> 00:33:07,080 Speaker 3: stunning conclusion. Right almost twenty years after ATM's were first widespread, 674 00:33:07,280 --> 00:33:10,800 Speaker 3: the number of tellers collapsed completely because once again it 675 00:33:10,840 --> 00:33:14,200 Speaker 3: became possible through mobile phones, mood technologies to actually do 676 00:33:14,240 --> 00:33:16,160 Speaker 3: most of your banking without a visiting a branch. We 677 00:33:16,200 --> 00:33:18,240 Speaker 3: didn't need as many branches banks that I can do 678 00:33:18,280 --> 00:33:20,040 Speaker 3: all of this without them. So it was like, wait up, 679 00:33:20,040 --> 00:33:22,080 Speaker 3: but wait a minute, I shouldn't that have created more tellers? 680 00:33:22,160 --> 00:33:24,040 Speaker 3: I mean, I know the story here, Come on, play 681 00:33:24,040 --> 00:33:26,200 Speaker 3: this right? Well, come on, that's not what's fun. That's 682 00:33:26,280 --> 00:33:28,120 Speaker 3: I'm supposed to happen with the new technology. So they 683 00:33:28,120 --> 00:33:30,200 Speaker 3: have the story completely wrong, and normly that they have 684 00:33:30,240 --> 00:33:32,520 Speaker 3: the causality all messed up, like the thing that drove 685 00:33:32,600 --> 00:33:34,959 Speaker 3: more tellers had nothing to do with technology. It had 686 00:33:35,000 --> 00:33:38,480 Speaker 3: to do with the incidental emergence of this banking regulation 687 00:33:38,600 --> 00:33:42,680 Speaker 3: and larger national footprints. And this is this story makes 688 00:33:42,680 --> 00:33:44,120 Speaker 3: people feel happy and comfortable. 689 00:33:44,120 --> 00:33:46,480 Speaker 1: That's the only it's the only one they can point to. 690 00:33:46,600 --> 00:33:47,440 Speaker 1: And it's not even correct. 691 00:33:47,440 --> 00:33:50,120 Speaker 3: And it's not even correct. And it's not only not correct, 692 00:33:50,120 --> 00:33:53,440 Speaker 3: it's not it's even a messy story. It's actually has 693 00:33:53,560 --> 00:33:55,200 Speaker 3: nothing to do with the things that they thought drove it. 694 00:33:55,360 --> 00:33:57,479 Speaker 2: I have another question. Can I ask a question? I 695 00:33:57,480 --> 00:34:00,800 Speaker 2: want to let's take this to its logical conclusion. Will 696 00:34:00,840 --> 00:34:04,720 Speaker 2: AI kill us all. And you know, the the argument, 697 00:34:05,040 --> 00:34:07,720 Speaker 2: the the the pro argument there, if you want to 698 00:34:07,800 --> 00:34:13,120 Speaker 2: use the word pro and context, okay, is which is 699 00:34:13,160 --> 00:34:16,520 Speaker 2: the prom erectus came along. And I'm going to get 700 00:34:16,520 --> 00:34:19,799 Speaker 2: this sequence wrong, so you know, anthropologists in the room, 701 00:34:20,080 --> 00:34:23,120 Speaker 2: don't hit me on the head. But you know, Australia 702 00:34:23,120 --> 00:34:26,240 Speaker 2: Pithecus was wiped out by you know, home erectus. The 703 00:34:26,239 --> 00:34:29,839 Speaker 2: home erector is wiped out by you know, a neanderal man, 704 00:34:29,960 --> 00:34:32,200 Speaker 2: you know, Homo sape and and all along the way, 705 00:34:32,560 --> 00:34:38,040 Speaker 2: so had much larger brain. The the the follow on 706 00:34:39,560 --> 00:34:44,240 Speaker 2: more intelligent species, not by necessarily directly killing the previous 707 00:34:44,320 --> 00:34:48,600 Speaker 2: less intelligent species, but for competition for climate, climate change, 708 00:34:48,880 --> 00:34:53,000 Speaker 2: and more importantly competition for resources, and more intelligent species 709 00:34:53,880 --> 00:34:58,719 Speaker 2: lasted and the previous species did not. So the one 710 00:34:58,840 --> 00:35:03,120 Speaker 2: argument that the academics make is well, for creating as 711 00:35:03,280 --> 00:35:06,240 Speaker 2: I and it's the more intelligent species and we're competing 712 00:35:06,239 --> 00:35:09,160 Speaker 2: for not competing for food necessarily, but certainly for energy. 713 00:35:10,520 --> 00:35:12,160 Speaker 2: Is it's going to be able to outcompete us for 714 00:35:12,200 --> 00:35:14,680 Speaker 2: resources and then that's the end of us, right, And 715 00:35:15,040 --> 00:35:16,840 Speaker 2: I mean, of course I don't know what the answer is. 716 00:35:16,880 --> 00:35:20,480 Speaker 2: What it's not. It's a not uncompelling argument. 717 00:35:20,120 --> 00:35:22,239 Speaker 3: No note at all. And I mean you get into 718 00:35:22,239 --> 00:35:24,920 Speaker 3: this position, would like would we make good pets? 719 00:35:27,360 --> 00:35:29,960 Speaker 1: Some of those would and it's hard. 720 00:35:29,760 --> 00:35:33,400 Speaker 3: To do a great And I don't think we make 721 00:35:33,520 --> 00:35:38,160 Speaker 3: very good pets. Him. He doesn't even know what for 722 00:35:38,360 --> 00:35:41,200 Speaker 3: messy week, screet a lot. We're re litigious. I mean, 723 00:35:41,239 --> 00:35:42,919 Speaker 3: I don't excreete much. Great. 724 00:35:43,160 --> 00:35:46,360 Speaker 2: I think I'm going to make it so that. 725 00:35:48,000 --> 00:35:51,200 Speaker 3: All you're screw screwed too. I have a terrible pad. So, yeah, 726 00:35:51,239 --> 00:35:53,520 Speaker 3: if it becomes a competition for resources, the history of 727 00:35:53,560 --> 00:35:56,719 Speaker 3: that is that you know, again, the species that's less 728 00:35:56,719 --> 00:35:58,880 Speaker 3: competitive loses out, and that in this case is us. 729 00:35:58,880 --> 00:36:01,720 Speaker 2: And if the future is I respond well to command. 730 00:36:02,360 --> 00:36:04,560 Speaker 3: And I think that they like dogs, right, they mean 731 00:36:04,600 --> 00:36:07,440 Speaker 3: self domesticated. They sort of saw this coming. Yeah, the 732 00:36:07,760 --> 00:36:10,279 Speaker 3: wolves said, okay, you know what, screw this, I'm not going. 733 00:36:10,280 --> 00:36:12,520 Speaker 1: To get it. And cats did reluctantly. 734 00:36:12,800 --> 00:36:14,839 Speaker 3: Self domestic here. Yeah, And I just don't think human 735 00:36:14,880 --> 00:36:15,560 Speaker 3: self domesticating. 736 00:36:15,640 --> 00:36:16,439 Speaker 1: No, I don't think so either. 737 00:36:16,520 --> 00:36:17,080 Speaker 3: This is my theory. 738 00:36:17,239 --> 00:36:19,000 Speaker 1: Well, I think it's interesting because when you think about 739 00:36:19,080 --> 00:36:22,360 Speaker 1: the fact that you have these instances already where people 740 00:36:22,520 --> 00:36:24,399 Speaker 1: that we know of at least at least a couple 741 00:36:24,480 --> 00:36:27,120 Speaker 1: of dozen have have either die by suicide or murder 742 00:36:27,160 --> 00:36:29,000 Speaker 1: or whatever it is as a result of these alarms. 743 00:36:29,040 --> 00:36:30,760 Speaker 1: And then you look at like there was a study. 744 00:36:31,600 --> 00:36:34,000 Speaker 1: I think you sent it around because you usually send 745 00:36:34,000 --> 00:36:37,040 Speaker 1: all the white papers about the large language models, favorite 746 00:36:37,040 --> 00:36:40,680 Speaker 1: communications generated by large larger language models. And then and 747 00:36:40,760 --> 00:36:44,600 Speaker 1: then also all these studies with participants where they're they're 748 00:36:44,640 --> 00:36:49,520 Speaker 1: interacting with people. And again once we go back to 749 00:36:49,560 --> 00:36:52,680 Speaker 1: this the lonely folks and so on and so forth, 750 00:36:53,560 --> 00:36:56,760 Speaker 1: and then their problems become worse as a result of the eyes. 751 00:36:56,920 --> 00:37:01,160 Speaker 1: We can't red team for all of these things. Well, 752 00:37:01,160 --> 00:37:03,799 Speaker 1: first of all, there are not people on these red 753 00:37:03,800 --> 00:37:08,239 Speaker 1: teams that have mental illnesses that are saying, hey, you 754 00:37:08,239 --> 00:37:12,920 Speaker 1: should warn against this, right, it's mostly just you know, educated. 755 00:37:12,880 --> 00:37:15,239 Speaker 3: M driven, you know the field psychiatric manual and let's 756 00:37:15,280 --> 00:37:17,960 Speaker 3: let's let's roleplay as if we're having this disorder. 757 00:37:18,080 --> 00:37:20,799 Speaker 1: And so when you look at the the AIS, we 758 00:37:20,840 --> 00:37:23,319 Speaker 1: can't we can't red team to figure out like what 759 00:37:23,400 --> 00:37:25,160 Speaker 1: they want is a pets, what they kill as will 760 00:37:25,200 --> 00:37:28,840 Speaker 1: they So it's just a roll of the fucking dice, right. 761 00:37:29,719 --> 00:37:33,200 Speaker 2: Yeah, it's it's even again, it's even more insidious than 762 00:37:33,200 --> 00:37:36,759 Speaker 2: that in that It's what we're finding is that AIS 763 00:37:36,920 --> 00:37:43,200 Speaker 2: prefer AI generated content over humor human generated content in 764 00:37:43,320 --> 00:37:47,960 Speaker 2: purchasing and hiring decisions. So you're the idiot who didn't 765 00:37:48,000 --> 00:37:51,400 Speaker 2: have AI write your LinkedIn what you stay off all 766 00:37:51,480 --> 00:37:54,799 Speaker 2: night for? Because the thing that's filtering out the first 767 00:37:54,840 --> 00:37:57,840 Speaker 2: three thousand resumes likes the ones where if someone just 768 00:37:57,880 --> 00:38:00,000 Speaker 2: went and IM pressed the button, yeah, please generate from the. 769 00:38:00,760 --> 00:38:03,080 Speaker 1: Really, we started this off by saying that we were 770 00:38:03,080 --> 00:38:07,320 Speaker 1: going to hopefully not end depressingly no depressed. 771 00:38:07,360 --> 00:38:11,319 Speaker 2: Yeah, right, kids with the closing number closing? Does this 772 00:38:11,360 --> 00:38:12,280 Speaker 2: one have a soundtrack? 773 00:38:12,520 --> 00:38:14,920 Speaker 3: Yeah, you're talking about this. This is something and this 774 00:38:15,000 --> 00:38:17,120 Speaker 3: applies to Hollywood, and I think the other thing that 775 00:38:17,200 --> 00:38:20,560 Speaker 3: happens is if we're all having similar not we're all, 776 00:38:20,600 --> 00:38:22,360 Speaker 3: but if a large number of people are having the 777 00:38:22,360 --> 00:38:25,280 Speaker 3: same conversations with the same l l ms with trained 778 00:38:25,280 --> 00:38:28,320 Speaker 3: on the same data, there's this huge homogenizing effect. 779 00:38:28,680 --> 00:38:28,879 Speaker 2: Right. 780 00:38:30,120 --> 00:38:32,440 Speaker 3: What people don't realize is that the models are essentially 781 00:38:32,719 --> 00:38:35,440 Speaker 3: trained on very similar sources of data and converging on 782 00:38:35,480 --> 00:38:36,160 Speaker 3: the same outcomes. 783 00:38:36,239 --> 00:38:39,719 Speaker 2: This is why when when you ask chat GPT like 784 00:38:39,800 --> 00:38:41,960 Speaker 2: pick a random number one through ten, it will always 785 00:38:42,000 --> 00:38:42,480 Speaker 2: say seven. 786 00:38:42,560 --> 00:38:45,920 Speaker 3: So the notion that it's always say seven. 787 00:38:45,800 --> 00:38:47,759 Speaker 2: Yeah, if you ask it, if you ask it just 788 00:38:47,800 --> 00:38:50,160 Speaker 2: for the first time at random number one through ten, 789 00:38:50,239 --> 00:38:52,560 Speaker 2: it will always say seven to you or me or anyone. Yeah. 790 00:38:52,560 --> 00:38:54,640 Speaker 2: It's because there's this reversion to the mean of all 791 00:38:54,640 --> 00:38:55,719 Speaker 2: the data it's trained ride. 792 00:38:56,040 --> 00:38:57,600 Speaker 3: The average thirty seven year old go and read it 793 00:38:57,680 --> 00:39:00,480 Speaker 3: like seven, and I ingested all of reddit. Therefore I 794 00:39:00,520 --> 00:39:04,200 Speaker 3: say seven even. But this in turn creates this kind 795 00:39:04,239 --> 00:39:07,880 Speaker 3: of fragility in society, correct because everyone now has in 796 00:39:07,920 --> 00:39:12,440 Speaker 3: a sense is being nurtured by this same stories and 797 00:39:12,520 --> 00:39:15,799 Speaker 3: conversations and everything else. And it's this There's been some 798 00:39:15,840 --> 00:39:19,160 Speaker 3: great work recently showing how it's a homogenizing effect in creativity. 799 00:39:19,200 --> 00:39:23,320 Speaker 3: It's an increasingly homogenizing effect in the kinds of our expectations, 800 00:39:23,360 --> 00:39:25,240 Speaker 3: in terms of what people should be willing to talk about. 801 00:39:25,239 --> 00:39:27,040 Speaker 3: Oh wait, the model gets me. Why won't you have 802 00:39:27,080 --> 00:39:30,319 Speaker 3: this conversation with me. I'll go back to the model again. 803 00:39:30,360 --> 00:39:34,200 Speaker 3: And so this idea that increasingly it becomes hermatically sealed 804 00:39:34,239 --> 00:39:36,759 Speaker 3: and homogenizing is I think it is incredibly important. Even 805 00:39:36,840 --> 00:39:38,800 Speaker 3: if it never coaches you to do something malign. 806 00:39:39,320 --> 00:39:41,520 Speaker 2: You know, I'm going to start saying correct when Paul 807 00:39:41,600 --> 00:39:43,520 Speaker 2: says up instead of I agree with that because I 808 00:39:43,560 --> 00:39:45,040 Speaker 2: think it makes me sound a lot smarter. 809 00:39:45,200 --> 00:39:47,000 Speaker 3: Yeah, it's probably all right, there's probably correct. 810 00:39:47,120 --> 00:39:49,560 Speaker 1: Absolutely, there's a you guys are making me think of. 811 00:39:49,760 --> 00:39:54,120 Speaker 1: Remember George Carlin's bit on plastics. Yeah, about how he 812 00:39:54,200 --> 00:39:56,600 Speaker 1: was like, he was like, screw you. Maybe the art 813 00:39:56,680 --> 00:39:59,520 Speaker 1: is created as because it wanted his great plastics. Maybe 814 00:39:59,560 --> 00:40:03,200 Speaker 1: the AI is fully aware of its whole plan and 815 00:40:03,239 --> 00:40:06,600 Speaker 1: it's just training itself to the point that it knows 816 00:40:06,640 --> 00:40:09,560 Speaker 1: exactly how this is going to end. With one pet, 817 00:40:09,600 --> 00:40:12,920 Speaker 1: which is Dick Ostelo, and the rest of us are screwed. 818 00:40:13,920 --> 00:40:18,120 Speaker 2: This has been another episode show soon to be the 819 00:40:18,200 --> 00:40:18,719 Speaker 2: Dick Show. 820 00:40:21,480 --> 00:40:22,399 Speaker 3: That's exactly right.