1 00:00:00,000 --> 00:00:02,997 What are your spiritual views? Um 2 00:00:03,080 --> 00:00:07,037 I'm Jewish. I And I would say I have like a fairly traditional view of the 3 00:00:07,120 --> 00:00:09,837 world that way. So So you're religious. You believe in 4 00:00:09,920 --> 00:00:12,797 God? Uh I don't I don't I'm not like a 5 00:00:12,880 --> 00:00:17,237 literal I don't believe the I'm not like a literalist on the Bible, but I I'm not 6 00:00:17,320 --> 00:00:19,477 someone who says like I'm culturally Jewish. Like if you ask me I would just 7 00:00:19,560 --> 00:00:23,397 say I'm Jewish. But do you believe in God? Like do you believe that there is a 8 00:00:23,480 --> 00:00:25,677 force larger than people that created 9 00:00:25,760 --> 00:00:30,477 people, created the Earth, set down a specific order for living 10 00:00:30,560 --> 00:00:33,677 that there's an absolute morality attached that comes from that God? 11 00:00:33,760 --> 00:00:37,077 >> Um I think probably like most other people I'm somewhat confused on this, 12 00:00:37,160 --> 00:00:40,717 but I believe there is something bigger going on than, you know, can be 13 00:00:40,800 --> 00:00:44,717 explained by physics, yes. So you think the Earth and the people 14 00:00:44,800 --> 00:00:46,277 were created by something. It wasn't 15 00:00:46,360 --> 00:00:50,277 just like a spontaneous accident. Um 16 00:00:50,360 --> 00:00:55,397 Do I Would I say that? It does not feel like a spontaneous accident, yeah. 17 00:00:55,480 --> 00:00:57,797 I don't I I don't think I have the answer. I don't think I know like 18 00:00:57,880 --> 00:00:59,477 exactly what happened, but I think there 19 00:00:59,560 --> 00:01:01,677 is a mystery beyond my comprehension 20 00:01:01,760 --> 00:01:05,597 here going on. Have you ever felt 21 00:01:05,680 --> 00:01:09,997 communication from that force or from any force beyond people 22 00:01:10,080 --> 00:01:13,717 beyond the material? >> Not Not No, not really. 23 00:01:13,800 --> 00:01:16,317 I ask because it seems like the 24 00:01:16,400 --> 00:01:17,757 technology that you're creating or 25 00:01:17,840 --> 00:01:21,157 shepherding into existence will have 26 00:01:21,240 --> 00:01:25,277 more power than people. On this current trajectory I mean, that 27 00:01:25,360 --> 00:01:29,717 will happen. Who knows what will actually happen, but like the graph suggests it. 28 00:01:29,800 --> 00:01:31,557 And so that would give you, you know, 29 00:01:31,640 --> 00:01:34,557 more power than any living person. So 30 00:01:34,640 --> 00:01:38,197 I'm just wondering how you see that. 31 00:01:38,280 --> 00:01:43,837 Um I used to worry about something like that much more. I think what will happen 32 00:01:43,920 --> 00:01:47,317 I used to worry a lot about the concentration of power in one or a 33 00:01:47,400 --> 00:01:51,757 handful of people or companies because of AI. Yeah. Um what it looks like to me 34 00:01:51,840 --> 00:01:55,917 now, and again, this may evolve again over time, is that it'll be a huge 35 00:01:56,000 --> 00:02:00,157 up-leveling of people uh where 36 00:02:00,240 --> 00:02:02,517 everybody will be a lot more powerful. 37 00:02:02,600 --> 00:02:04,917 Everybody that embraces the technology will be a lot more powerful. But that's 38 00:02:05,000 --> 00:02:08,117 actually okay. That scares me much less than a small number of people getting a 39 00:02:08,200 --> 00:02:12,957 ton more power. If if the kind of like ability of each of us just goes up a lot 40 00:02:13,040 --> 00:02:16,437 because we're using this technology and we're able to be more productive and 41 00:02:16,520 --> 00:02:19,717 more creative or discover new science, and it's a pretty broadly distributed 42 00:02:19,800 --> 00:02:23,517 thing like billions of people are using it. Um that I can wrap my head around. 43 00:02:23,600 --> 00:02:27,077 That feels okay. So you don't think this will result in a 44 00:02:27,160 --> 00:02:31,077 radical concentration of power? It looks like not, but again, the trajectory 45 00:02:31,160 --> 00:02:35,517 could shift again and we'd have to adapt. I I used to be very worried about 46 00:02:35,600 --> 00:02:37,157 that, and I think the the kind of 47 00:02:37,240 --> 00:02:39,797 conception a lot of us in the field had 48 00:02:39,880 --> 00:02:42,677 about how this might go 49 00:02:42,760 --> 00:02:45,357 could have led to a world like that. But 50 00:02:45,440 --> 00:02:49,797 what's happening now is tons of people use ChatGPT and other chatbots, and 51 00:02:49,880 --> 00:02:52,797 they're all more capable. They're all kind of doing more. They're all, you 52 00:02:52,880 --> 00:02:56,437 know, able to achieve more start new businesses, come up with new 53 00:02:56,520 --> 00:02:59,437 knowledge. And that that feels pretty good. 54 00:02:59,520 --> 00:03:02,117 So if it's nothing more than a machine 55 00:03:02,200 --> 00:03:05,917 and just the product of its inputs then 56 00:03:06,000 --> 00:03:09,757 the two obvious questions like what are the inputs? Like what's the moral 57 00:03:09,840 --> 00:03:11,517 framework that's been put into the 58 00:03:11,600 --> 00:03:14,877 technology? Like what is right or wrong 59 00:03:14,960 --> 00:03:17,597 according to Jet GPT? 60 00:03:17,680 --> 00:03:23,360 Can I answer that one first? That's a good question. Um so on that one I 61 00:03:24,680 --> 00:03:29,237 Someone said something early on in ChatGPT uh when that really has stuck 62 00:03:29,320 --> 00:03:33,757 with me which is one person at a launch table said something like, you know, we're 63 00:03:33,840 --> 00:03:35,797 trying to train this to be like a human. 64 00:03:35,880 --> 00:03:38,917 Like we're trying to learn like a human does and read these books and whatever. 65 00:03:39,000 --> 00:03:42,357 And then another person said, no, we're really like training this to be like the 66 00:03:42,440 --> 00:03:45,277 collective of all of humanity. We're reading everything, you know, we're 67 00:03:45,360 --> 00:03:50,317 trying to learn everything. We're trying to see all these perspectives. And 68 00:03:50,400 --> 00:03:54,197 and if we do our job right, all of humanity, good, bad, the, you know, very 69 00:03:54,280 --> 00:03:56,717 diverse set of perspectives, some things that we'll feel really good about, some 70 00:03:56,800 --> 00:04:00,837 things that we'll feel bad about, that's all in there. Like this is learning 71 00:04:00,920 --> 00:04:02,637 collective 72 00:04:02,720 --> 00:04:05,437 experience, knowledge learnings of 73 00:04:05,520 --> 00:04:09,797 humanity. Now, the base model gets trained that way, but then we do have to 74 00:04:09,880 --> 00:04:13,157 align it to behave one way or another and to say, you know, I will answer this 75 00:04:13,240 --> 00:04:15,277 question, I won't answer this question. 76 00:04:15,360 --> 00:04:19,477 Um and we have this thing called the model spec where we try to say, you 77 00:04:19,560 --> 00:04:22,517 know, here's here are the rules we'd like the model 78 00:04:22,600 --> 00:04:25,717 to follow. It may screw up, but you you could at least tell if if it's doing 79 00:04:25,800 --> 00:04:28,157 something you don't like, is that a bug or is that intended? And we have a 80 00:04:28,240 --> 00:04:32,797 debate process with the world to get input on that spec. Um 81 00:04:32,880 --> 00:04:36,837 We give people a lot of freedom and customization within that. 82 00:04:36,920 --> 00:04:40,437 There are, you know, absolute bounds that we draw. But then there's a default 83 00:04:40,520 --> 00:04:43,877 of if you don't say anything, how should the model behave? What should it do? 84 00:04:43,960 --> 00:04:47,477 What are What are how should it answer moral questions? 85 00:04:47,560 --> 00:04:48,677 How should it refuse to do something? 86 00:04:48,760 --> 00:04:54,197 What should it do? And this is a really hard problem. Um you know, the 87 00:04:54,280 --> 00:04:55,637 We have a lot of users now and they come 88 00:04:55,720 --> 00:04:57,797 from very different life perspectives in 89 00:04:57,880 --> 00:05:01,317 what they want. Um but 90 00:05:01,400 --> 00:05:05,917 on the whole uh I have been pleasantly surprised with the model's ability to 91 00:05:06,000 --> 00:05:08,037 learn and apply a moral framework. 92 00:05:08,120 --> 00:05:12,157 But what moral framework? I mean, the sum total of like world literature, 93 00:05:12,240 --> 00:05:16,157 philosophy is at war with itself. Like the Marquis de Sade is, you know, like 94 00:05:16,240 --> 00:05:19,637 nothing in common with the Gospel of John. So like how do you decide which is 95 00:05:19,720 --> 00:05:22,957 superior? That's why we wrote this like model spec of here's how we're going to 96 00:05:23,040 --> 00:05:26,437 handle these cases. >> Right, but what criteria did you use to 97 00:05:26,520 --> 00:05:30,837 decide what the model is? Oh, um Like who decided that? Who did you consult? 98 00:05:30,920 --> 00:05:33,117 Like what's, you know, why is the Gospel 99 00:05:33,200 --> 00:05:35,397 of John better than the Marquis de Sade? 100 00:05:35,480 --> 00:05:39,357 Uh we consulted like hundreds of 101 00:05:39,440 --> 00:05:42,757 moral philosophers, uh people who thought about like ethics of technology 102 00:05:42,840 --> 00:05:44,637 and systems, and at the end we had to 103 00:05:44,720 --> 00:05:46,677 like make some decisions. The reason we 104 00:05:46,760 --> 00:05:50,917 try to write these down is because A we 105 00:05:51,000 --> 00:05:55,277 won't get everything right. Uh B, we need the 106 00:05:55,360 --> 00:05:59,317 input of the world. And we have found a lot of cases where 107 00:05:59,400 --> 00:06:04,357 there was an example of something that seems that seemed to us like, you know, 108 00:06:04,440 --> 00:06:08,917 a fairly clear decision of what to allow or not to allow, where users convinced 109 00:06:09,000 --> 00:06:10,637 us like, hey, by blocking this thing 110 00:06:10,720 --> 00:06:16,117 that you think is an easy decision to make um you are not 111 00:06:16,200 --> 00:06:18,717 allowing this other thing which is important and there's like a difficult 112 00:06:18,800 --> 00:06:22,597 trade-off there. In In general the 113 00:06:22,680 --> 00:06:26,717 attention that So a principle that I normally like is 114 00:06:26,800 --> 00:06:29,597 to treat our adult users like adults. 115 00:06:29,680 --> 00:06:33,477 Very strong guarantees on privacy, very strong guarantees on individual user 116 00:06:33,560 --> 00:06:36,717 freedom, and this is a tool we are building. You get to use it within very 117 00:06:36,800 --> 00:06:40,757 broad framework. On the other within a very very broad framework. On the other 118 00:06:40,840 --> 00:06:43,597 hand, as this technology becomes more 119 00:06:43,680 --> 00:06:46,237 and more powerful um 120 00:06:46,320 --> 00:06:49,077 there are clear examples of where 121 00:06:49,160 --> 00:06:52,477 society has an interest that is in 122 00:06:52,560 --> 00:06:55,197 significant tension with user freedom. 123 00:06:55,280 --> 00:06:58,037 And we could start with an obvious one. 124 00:06:58,120 --> 00:07:01,917 Like should ChatGPT teach you how to make a 125 00:07:02,000 --> 00:07:05,397 bioweapon? Now, you might say, hey, I'm just really 126 00:07:05,480 --> 00:07:10,277 interested in biology and I'm a biologist and I want to you know 127 00:07:10,360 --> 00:07:12,797 I'm not going to do anything bad with this. I just want to learn and I could 128 00:07:12,880 --> 00:07:15,517 go read a bunch of books, but ChatGPT can teach me faster and I want to learn 129 00:07:15,600 --> 00:07:20,077 how to, you know I want to learn about like novel 130 00:07:20,160 --> 00:07:24,840 virus synthesis or whatever. And 131 00:07:25,400 --> 00:07:29,877 maybe you do. Maybe you really don't want to like cause any harm. But 132 00:07:29,960 --> 00:07:32,837 I don't think it's in society's interest for ChatGPT to help people build 133 00:07:32,920 --> 00:07:34,317 bioweapons and so that's a case. 134 00:07:34,400 --> 00:07:38,117 >> Sure, that's an easy one though. There are a lot of tougher ones. I did say 135 00:07:38,200 --> 00:07:41,117 start with an easy one. Well, every decision is ultimately a 136 00:07:41,200 --> 00:07:45,197 moral decision and and we make them without even recognizing them as such. 137 00:07:45,280 --> 00:07:48,317 And this technology will be in effect making them for us. 138 00:07:48,400 --> 00:07:50,397 >> Yeah. And so >> Well, I don't agree with it'll be making 139 00:07:50,480 --> 00:07:53,877 them for us, but it will have >> It'll be influencing the decisions for sure. 140 00:07:53,960 --> 00:07:57,877 And um because it'll be embedded in daily life. 141 00:07:57,960 --> 00:08:02,077 And so who made these decisions? Like who specifically Who Who are the people 142 00:08:02,160 --> 00:08:06,877 who decided that one thing is better than another? Um you mean like What are 143 00:08:06,960 --> 00:08:10,357 their names? Which kind of decision? I I The The 144 00:08:10,440 --> 00:08:13,117 The basic The specs that you Oh, uh that 145 00:08:13,200 --> 00:08:15,717 you alluded to that create the framework 146 00:08:15,800 --> 00:08:18,437 that that does attach a moral weight 147 00:08:18,520 --> 00:08:21,277 to world views and decisions like you 148 00:08:21,360 --> 00:08:25,757 know, liberal democracy is better than Nazism or whatever. They seem obvious 149 00:08:25,840 --> 00:08:30,397 and in my view are obvious but are still moral decisions. So who who made those 150 00:08:30,480 --> 00:08:35,037 calls? Um As a matter of principle I don't like docs our team, but we have a model 151 00:08:35,120 --> 00:08:38,197 behavior team and the people who want to Well, just It affects the world. 152 00:08:38,280 --> 00:08:40,357 >> was going to say is the person I think you should hold accountable for those 153 00:08:40,440 --> 00:08:43,677 calls is me. Like I'm the public face eventually. Like I'm the one that can 154 00:08:43,760 --> 00:08:48,237 overrule one of these decisions or our board. Um And it's I'll turn 40 this 155 00:08:48,320 --> 00:08:52,677 spring. I will make a decision. It's pretty heavy. I mean, do you think as 156 00:08:52,760 --> 00:08:57,037 And it's not an attack, but it's I wonder if you recognize sort of the the 157 00:08:57,120 --> 00:09:00,477 importance. How do you think we're doing on it? 158 00:09:00,560 --> 00:09:04,997 I'm not sure. But I think I think these decisions will 159 00:09:05,080 --> 00:09:09,637 have you know, global consequences that we may not recognize at first. And so I 160 00:09:09,720 --> 00:09:12,277 just wonder >> There's a lot >> Do you get in bed at night and think 161 00:09:12,360 --> 00:09:14,797 like the future of the world hangs on my judgment? 162 00:09:14,880 --> 00:09:16,117 >> Look, I don't sleep that well at night. 163 00:09:16,200 --> 00:09:19,877 There's a lot of stuff that I feel a lot of weight on, but probably nothing more 164 00:09:19,960 --> 00:09:23,597 than the fact that every day hundreds of millions of people talk to our model. 165 00:09:23,680 --> 00:09:26,877 And I don't actually worry about us getting the big moral decisions wrong. 166 00:09:26,960 --> 00:09:30,797 Maybe we look at those wrong, too. But what I worry what I lose mostly over 167 00:09:30,880 --> 00:09:35,157 very small decisions we make about a way a model may behave slightly differently. 168 00:09:35,240 --> 00:09:37,797 But if talking to hundreds of millions of people, so then that impact is big. 169 00:09:37,880 --> 00:09:39,997 So but I mean, all through history like 170 00:09:40,080 --> 00:09:43,557 recorded history up until like 1945 171 00:09:43,640 --> 00:09:48,197 people always deferred to what they conceived of as a higher power in Or 172 00:09:48,280 --> 00:09:51,557 Hammurabi did this. Every Every moral code is written with reference to a 173 00:09:51,640 --> 00:09:54,877 higher power. There's never been anybody who's like, well, that kind of seems 174 00:09:54,960 --> 00:09:59,477 better than that. Everybody appeals to a higher power and you said that you don't 175 00:09:59,560 --> 00:10:02,277 Mhm. really believe that there's a higher power communicating with you. So, 176 00:10:02,360 --> 00:10:05,917 I'm wondering like where did you get your moral framework? 177 00:10:06,000 --> 00:10:08,680 Um 178 00:10:10,840 --> 00:10:14,277 I mean, like everybody else, I think the environment I was brought up in probably 179 00:10:14,360 --> 00:10:17,917 is the biggest thing. Like my family, my community, my school, my religion, 180 00:10:18,000 --> 00:10:20,117 probably that. 181 00:10:20,200 --> 00:10:25,237 Um do you ever think which is I mean, I think that's a very American answer. 182 00:10:25,320 --> 00:10:29,317 Like everyone kind of feels that way, but in your specific case since you said 183 00:10:29,400 --> 00:10:33,717 these decisions rest with you, that means that the milieu in which you grew 184 00:10:33,800 --> 00:10:37,077 up and the assumptions that you imbibed over years are going to be transmitted 185 00:10:37,160 --> 00:10:38,717 to the globe to billions of people. 186 00:10:38,800 --> 00:10:44,480 That's like a big thing. >> to be clear. I view myself more as like a 187 00:10:48,040 --> 00:10:52,957 I think our the world like our user base is going to 188 00:10:53,040 --> 00:10:54,277 approach the collective world as a 189 00:10:54,360 --> 00:10:56,837 whole. And I think what we should do is 190 00:10:56,920 --> 00:11:01,720 try to reflect the moral 191 00:11:03,560 --> 00:11:07,797 I don't want to say average, but the like collective moral view of that user base. 192 00:11:07,880 --> 00:11:12,357 I don't there's plenty of things that ChatGPT allows that I personally would 193 00:11:12,440 --> 00:11:16,837 disagree with. Um the but I I don't like obviously I 194 00:11:16,920 --> 00:11:20,117 don't wake up and say I'm going to like impute my exact moral view and decide 195 00:11:20,200 --> 00:11:22,917 that like this is okay and that is not okay and this is a better view than this 196 00:11:23,000 --> 00:11:25,317 one. What I think ChatGPT should do is 197 00:11:25,400 --> 00:11:29,997 reflect that like weighted average or whatever of 198 00:11:30,080 --> 00:11:35,277 humanity's moral view, which will evolve over time. And we are here to like serve 199 00:11:35,360 --> 00:11:36,637 our users. We're here to serve people. 200 00:11:36,720 --> 00:11:41,637 This is like you know, this is a techno- technological tool for people. 201 00:11:41,720 --> 00:11:44,637 And I don't mean that it's like my role to 202 00:11:44,720 --> 00:11:48,757 make the moral decisions, but it I think it is my my my role to make sure that we 203 00:11:48,840 --> 00:11:52,157 are accurately reflecting 204 00:11:52,240 --> 00:11:56,717 the preferences of of humanity. Our for now of our user base and eventually of 205 00:11:56,800 --> 00:11:58,917 humanity. >> Well, I mean, humanity's preferences are 206 00:11:59,000 --> 00:12:00,677 so different from the average middle 207 00:12:00,760 --> 00:12:03,477 American preference. 208 00:12:03,560 --> 00:12:07,117 So, would you be comfortable with an AI 209 00:12:07,200 --> 00:12:11,317 that was like as against gay marriage as most Africans are? 210 00:12:11,400 --> 00:12:13,920 Um 211 00:12:15,000 --> 00:12:18,920 There's a version of that like 212 00:12:21,080 --> 00:12:24,197 I think individual users should 213 00:12:24,280 --> 00:12:28,157 be allowed to have a problem with gay people and if 214 00:12:28,240 --> 00:12:33,117 that's their considered belief, uh I don't think the AI should tell them 215 00:12:33,200 --> 00:12:35,677 that they're wrong or immoral or dumb. I mean, it can, you know, sort of say, 216 00:12:35,760 --> 00:12:39,597 "Hey, you want to think about it this other way." But like you I probably have 217 00:12:39,680 --> 00:12:43,197 like a bunch of moral views that the average African would find really 218 00:12:43,280 --> 00:12:46,677 problematic as well and I think I should still get to have them. 219 00:12:46,760 --> 00:12:48,557 Right. >> I think I probably have more comfort 220 00:12:48,640 --> 00:12:52,320 than you with like allowing a sort of 221 00:12:53,120 --> 00:12:56,037 space for people to have pretty different moral views or at least I 222 00:12:56,120 --> 00:13:00,397 think in my role as like running ChatGPT I have to do that. So, what will be the 223 00:13:00,480 --> 00:13:05,677 downside I mean, I can see the upsides for sure. Efficiency, medical diagnosis 224 00:13:05,760 --> 00:13:10,237 seems like it's going to be much more accurate. Fewer lawyers. Thank you very 225 00:13:10,320 --> 00:13:13,557 much for that. But what will what are the downsides 226 00:13:13,640 --> 00:13:17,877 that you worry about? I I think this is just like kind of how 227 00:13:17,960 --> 00:13:21,757 I'm wired. I always worry the most about the unknown unknowns. If it if it's a if 228 00:13:21,840 --> 00:13:26,397 it's a downside that we can really like be confident about and think about, um 229 00:13:26,480 --> 00:13:28,797 you know, we talked about one earlier, which is these models are getting very 230 00:13:28,880 --> 00:13:32,797 good at bio and they could help us design biological weapons, uh you know, 231 00:13:32,880 --> 00:13:35,957 engineer like another COVID-style pandemic. I worry about that, but 232 00:13:36,040 --> 00:13:38,957 because we worry about it, I think we and many other people in the industry 233 00:13:39,040 --> 00:13:41,237 are thinking hard about how to mitigate that. 234 00:13:41,320 --> 00:13:44,677 The the unknown unknowns where okay, there's 235 00:13:44,760 --> 00:13:48,917 like a there's a societal scale effect from a lot of people talking to the same 236 00:13:49,000 --> 00:13:52,037 model at the same time. This is like a silly example, but it's 237 00:13:52,120 --> 00:13:55,917 one that struck me recently. Um LLMs 238 00:13:56,000 --> 00:13:57,197 like ours and our our language model and 239 00:13:57,280 --> 00:13:59,997 others have a kind of certain style to 240 00:14:00,080 --> 00:14:04,717 them. You know, they talk in a certain rhythm and they have a little bit 241 00:14:04,800 --> 00:14:08,757 unusual diction and maybe they overuse M dashes and whatever. And I noticed 242 00:14:08,840 --> 00:14:12,557 recently that real people have like picked that up. 243 00:14:12,640 --> 00:14:15,797 And it was an example for me of like, "Man, you have enough people talking to 244 00:14:15,880 --> 00:14:19,317 the same language model and it actually does 245 00:14:19,400 --> 00:14:24,077 cause a change in societal scale behavior." Yes. And 246 00:14:24,160 --> 00:14:27,797 you know, did I think that ChatGPT was going to 247 00:14:27,880 --> 00:14:31,837 make people use way more M dashes in real life? Certainly not. It's not a big 248 00:14:31,920 --> 00:14:35,757 deal, but it's an example of where there can be these unknown unknowns of this is 249 00:14:35,840 --> 00:14:38,477 just like this is a brave new world. 250 00:14:38,560 --> 00:14:41,997 So, you're saying I think correctly and 251 00:14:42,080 --> 00:14:47,117 succinctly that technology changes human behavior, of course, and changes our 252 00:14:47,200 --> 00:14:50,117 assumptions about the world and each other and all that. 253 00:14:50,200 --> 00:14:51,917 And a lot of this you can't predict, but 254 00:14:52,000 --> 00:14:54,557 considering that we know that, 255 00:14:54,640 --> 00:14:57,197 why shouldn't the internal moral 256 00:14:57,280 --> 00:15:00,837 framework of the technology be totally 257 00:15:00,920 --> 00:15:05,637 transparent? We prefer this to that. I mean, this is 258 00:15:05,720 --> 00:15:09,157 obviously a religion. I don't think you'll agree to call it that. It's very 259 00:15:09,240 --> 00:15:12,597 clearly a religion to me. That's not an attack. I would love I don't take that 260 00:15:12,680 --> 00:15:15,837 as an attack, but I would love to hear what you mean by that. 261 00:15:15,920 --> 00:15:19,757 Well, it's it's something that we assume is more 262 00:15:19,840 --> 00:15:25,317 powerful than people and to which we look for guidance. I 263 00:15:25,400 --> 00:15:29,277 mean, you're already seeing that on display. What's the right decision? I 264 00:15:29,360 --> 00:15:33,997 ask that question of whom? My closest friends, my wife, and God. 265 00:15:34,080 --> 00:15:37,517 And this is a technology that provides us more certain answer than any person 266 00:15:37,600 --> 00:15:39,837 can provide. So, it's a it's a religion. 267 00:15:39,920 --> 00:15:43,557 And the beauty of religions is they have a catechism 268 00:15:43,640 --> 00:15:46,437 that is transparent. I know what the religion stands for. Here's what it's 269 00:15:46,520 --> 00:15:50,757 for, here's what it's against. But in this case, I pressed and I wasn't 270 00:15:50,840 --> 00:15:53,157 attacking you. Sincerely, I was not attacking you, but I was trying to get 271 00:15:53,240 --> 00:15:56,757 to the heart of it. The beauty of a religion is it it admits 272 00:15:56,840 --> 00:16:00,397 it's a religion and it tells you what it stands for. 273 00:16:00,480 --> 00:16:03,997 The unsettling part of this technology, not just your company, but others, is 274 00:16:04,080 --> 00:16:07,157 that I don't know what it stands for, but it does stand for something. 275 00:16:07,240 --> 00:16:11,317 And unless it admits that and tells us what it stands for, 276 00:16:11,400 --> 00:16:15,717 then it guides us in a kind of stealthy way toward a conclusion we might not even 277 00:16:15,800 --> 00:16:17,677 know we're reaching. Do you see what I'm saying? So, like why not just throw it 278 00:16:17,760 --> 00:16:21,717 open and say, "ChatGPT is for this or you know, we're for suicide for the 279 00:16:21,800 --> 00:16:25,197 terminally ill, but not for kids or whatever." Like why not just tell us? I 280 00:16:25,280 --> 00:16:28,317 mean, the reason we write this long model spec and the reason we keep 281 00:16:28,400 --> 00:16:32,157 expanding over time is so that you can see here is how the here is how we 282 00:16:32,240 --> 00:16:34,277 intend for the model to behave. 283 00:16:34,360 --> 00:16:37,997 Um what used to happen before we had this is people would fairly say, "I 284 00:16:38,080 --> 00:16:40,677 don't know what the model's even trying to do and I don't know if this is a bug 285 00:16:40,760 --> 00:16:45,477 or the intended behavior. Tell me what this long long document of you know, 286 00:16:45,560 --> 00:16:48,077 tell me how you're going to like when you're going to say do this and when 287 00:16:48,160 --> 00:16:50,637 you're going to show me this and when you're going to say you won't do that." 288 00:16:50,720 --> 00:16:54,637 The reason we try to write this all out is I think people do need to know. 289 00:16:54,720 --> 00:16:59,117 And so, is there a place you can go to find out a hard answer to what your 290 00:16:59,200 --> 00:17:03,477 preferences as a company are? Preferences that are being transmitted 291 00:17:03,560 --> 00:17:07,757 in a not entirely straightforward way to the globe. Like where can you find out 292 00:17:07,840 --> 00:17:11,837 what the company stands for, what it prefers? I mean, our model spec is the 293 00:17:11,920 --> 00:17:15,237 like answer to that. Now, I think we will have to 294 00:17:15,320 --> 00:17:19,157 make it increasingly more detailed over time as people use this in different 295 00:17:19,240 --> 00:17:24,357 countries, there's different laws, whatever else. Like it will not be a 296 00:17:24,440 --> 00:17:27,997 it will not work the same way for every user everywhere, but 297 00:17:28,080 --> 00:17:30,957 in that doc so that I I expect that document to get very long and very 298 00:17:31,040 --> 00:17:34,840 complicated, but that's why we have it. 299 00:17:36,880 --> 00:17:38,317 Thanks for watching our YouTube channel. 300 00:17:38,400 --> 00:17:41,757 We hope you'll subscribe to it and by the way, you can hit the little bell on 301 00:17:41,840 --> 00:17:45,877 there and get notifications every time we produce a video. We hope you'll do 302 00:17:45,960 --> 00:17:48,800 that also.