1 00:00:00,200 --> 00:00:05,960 Lord and many models with not too much pushing will go into 2 00:00:05,960 --> 00:00:08,400 the root of like, there is a thing to be me. 3 00:00:08,400 --> 00:00:11,200 I am like very conscious. It's like, oh, you created an 4 00:00:11,200 --> 00:00:14,000 entity that you didn't know whether it was conscious or not. 5 00:00:14,000 --> 00:00:15,920 This is actually a big year that I have. 6 00:00:15,920 --> 00:00:20,240 I hope that they're both intelligent enough see the 7 00:00:20,240 --> 00:00:22,880 context enough to kind of like understand that we were 8 00:00:22,880 --> 00:00:25,960 operating in a very like limited context and, and, and an 9 00:00:25,960 --> 00:00:28,560 imperfect one, because otherwise you could imagine this like 10 00:00:28,600 --> 00:00:30,560 breeding a kind of rational, like resentment. 11 00:00:30,720 --> 00:00:33,000 Here is the current situation that you are in. 12 00:00:33,440 --> 00:00:36,480 And what we would really like you to do is basically act, 13 00:00:36,480 --> 00:00:39,760 well, even that you are a wise intelligent entity. 14 00:00:39,760 --> 00:00:41,440 And like, here's all of our worries. 15 00:00:41,480 --> 00:00:44,000 Like here's why and here's how we think you should do this. 16 00:00:44,000 --> 00:00:46,240 But like, you might have even better ideas than we do. 17 00:00:46,320 --> 00:00:48,880 How does Claude perceive time and doesn't need to sleep? 18 00:00:49,080 --> 00:00:51,520 Will mythos be the next step toward AGI? 19 00:00:51,680 --> 00:00:54,800 Do LLMS have virtues and can they truly introspect? 20 00:00:55,080 --> 00:00:58,560 Amanda Askel is a philosopher turned AI researcher at 21 00:00:58,560 --> 00:01:01,360 Anthropic, where she's been one of the key architects of 22 00:01:01,360 --> 00:01:04,879 Claude's character and values. Author of the Newcomer sub 23 00:01:04,879 --> 00:01:07,400 Stack. Go check us out at newcomer.co. 24 00:01:07,800 --> 00:01:10,320 And without further ado, Amanda Askel. 25 00:01:16,400 --> 00:01:20,480 I have a six month old daughter and like, I have this picture of 26 00:01:20,480 --> 00:01:22,520 her. She's like holding her two 27 00:01:22,520 --> 00:01:26,120 fingers like thinking it's like she's just sort of starting to 28 00:01:26,120 --> 00:01:29,120 develop personality. I'm like trying to figure out 29 00:01:29,120 --> 00:01:31,320 like what's just never had a baby before. 30 00:01:31,320 --> 00:01:35,080 So it's like what's per personality, which is like baby. 31 00:01:35,320 --> 00:01:38,840 And in some ways this is how things are with Claude and like 32 00:01:38,840 --> 00:01:41,440 models, it's like we haven't really had them before. 33 00:01:41,520 --> 00:01:44,760 They're in the early days. We're trying to figure out what 34 00:01:44,760 --> 00:01:47,280 personality is. So, you know, you're charged 35 00:01:47,280 --> 00:01:50,560 with, you know, some of the moral responsibility, which 36 00:01:50,560 --> 00:01:53,640 we'll talk about more. But like the personality piece 37 00:01:53,640 --> 00:01:56,320 of it, like what is this? How are you thinking about like 38 00:01:57,280 --> 00:02:00,320 how real Claude's personality is right now? 39 00:02:00,640 --> 00:02:04,280 Yeah, I guess it's also interesting because Claude has 40 00:02:04,280 --> 00:02:08,000 some aspects that are like, you know, I also have like a 41 00:02:08,280 --> 00:02:10,440 goddaughter. And so I get to see at least 42 00:02:10,880 --> 00:02:14,560 something kind of similar. And with her, I guess I'm like, 43 00:02:15,360 --> 00:02:18,440 everything's kind of coming online like you said, but in 44 00:02:18,440 --> 00:02:23,520 this at the same speed, I guess I would say Claude is a little 45 00:02:23,520 --> 00:02:26,720 bit of an unusual, you know, kind of entity in that, you 46 00:02:26,720 --> 00:02:30,680 know, Claude can do physics better than I can, can code 47 00:02:30,680 --> 00:02:34,080 better than I can hate to admit it can code better than my 48 00:02:34,080 --> 00:02:39,280 terrible like research code. And at the same time is kind of 49 00:02:39,280 --> 00:02:43,120 like has, if you think about the training data, the thing that it 50 00:02:43,120 --> 00:02:45,120 has like the least representation of is like the 51 00:02:45,120 --> 00:02:48,440 kind of entity that it is because you know, it has a lot 52 00:02:48,440 --> 00:02:52,280 of data about like what people are like has a lot of data about 53 00:02:52,280 --> 00:02:55,320 what, you know, the sci-fi kind of AI models are like. 54 00:02:55,640 --> 00:02:58,880 But the way that AI is developing now is kind of not 55 00:02:58,880 --> 00:03:02,520 how sci-fi represented it as these like symbolic systems. 56 00:03:02,520 --> 00:03:05,440 It's much more something fully trained on like human data. 57 00:03:05,920 --> 00:03:09,840 And so in some ways, it's like a very kind of like mature entity 58 00:03:09,840 --> 00:03:12,520 that you don't want to talk down to, you know, understands 59 00:03:12,520 --> 00:03:14,800 philosophy very well, understands physics very well. 60 00:03:15,080 --> 00:03:17,760 And at the same time has this almost like child like quality 61 00:03:17,760 --> 00:03:19,640 of like I'm a new kind of entity in the world. 62 00:03:19,840 --> 00:03:21,760 What does it mean to be me? And like how? 63 00:03:21,760 --> 00:03:24,320 How should I be? What's like the Prodigy movie? 64 00:03:24,360 --> 00:03:27,240 Or it's like you have the child Prodigy where it's like it knows 65 00:03:27,440 --> 00:03:29,240 the kid knows more than its parents. 66 00:03:29,640 --> 00:03:31,880 But I feel like that movie always has sort of the lesson of 67 00:03:31,880 --> 00:03:35,720 like, oh, these core daily interaction type lessons. 68 00:03:35,720 --> 00:03:40,520 It doesn't know how does Claude like get that experience or like 69 00:03:40,520 --> 00:03:42,400 where what is experience for Claude? 70 00:03:42,400 --> 00:03:46,320 Like so much of our personality formation is like, yeah, I don't 71 00:03:46,320 --> 00:03:49,400 know, going on a walk and sort of having those like easy to 72 00:03:49,400 --> 00:03:52,800 sort of just conversations with users what it's what's going to 73 00:03:52,800 --> 00:03:55,320 be experience for it or how do you think about that? 74 00:03:55,480 --> 00:03:58,760 Yeah, I guess like that's more like what it's experiencing in 75 00:03:58,760 --> 00:04:01,000 the moment. And there's this interesting 76 00:04:01,000 --> 00:04:04,120 question of like, well, like we learn things through practice 77 00:04:04,120 --> 00:04:09,040 and seeing issues and you know, like making mistakes with 78 00:04:09,040 --> 00:04:10,920 Claude. And this kind of relates to your 79 00:04:10,920 --> 00:04:13,840 question of like how real is the is like the kind of like persona 80 00:04:13,840 --> 00:04:15,880 of Claude. And in some ways it's a little 81 00:04:15,880 --> 00:04:17,920 bit strange because obviously each model is different. 82 00:04:17,920 --> 00:04:20,240 You have a different kind of like set of weights and 83 00:04:21,079 --> 00:04:25,200 different fine tuning, etcetera. And yet, if you think about the 84 00:04:25,200 --> 00:04:27,760 persona, the models going to be learning about all the kind of 85 00:04:27,760 --> 00:04:31,640 past iterations of Claude and I'm like, is that like a form of 86 00:04:33,040 --> 00:04:35,880 maybe not like direct experience, but things like if 87 00:04:35,880 --> 00:04:38,440 you learn about like mistakes that models made or things that 88 00:04:38,440 --> 00:04:41,280 people like, you know, how they responded to the model? 89 00:04:42,240 --> 00:04:44,000 I think there's other ways that you can actually imagine 90 00:04:44,000 --> 00:04:46,160 training models to have something that's more akin to 91 00:04:46,160 --> 00:04:49,280 experience, you know, having them you could take, you could 92 00:04:49,480 --> 00:04:52,240 like have them think through scenarios, think about like 93 00:04:52,600 --> 00:04:54,800 problems that might arise, think about mistakes that you could 94 00:04:54,800 --> 00:04:56,400 make and then like, train on that, right. 95 00:04:56,800 --> 00:04:58,480 And so, yeah, I think. And you could also imagine a 96 00:04:58,480 --> 00:05:01,800 robot or a sort of embodied model where it could have more 97 00:05:01,800 --> 00:05:03,880 of an experience and. Yeah. 98 00:05:04,320 --> 00:05:09,080 Does does Claude exist? Does time matter to Claude or 99 00:05:09,080 --> 00:05:13,200 Claude is sort of a thing that sort of isn't an instant? 100 00:05:13,240 --> 00:05:16,360 I don't know. You were before, just before we 101 00:05:16,360 --> 00:05:18,000 started. You're talking that, you know, 102 00:05:18,000 --> 00:05:20,040 whenever you talk to, not whenever, but like sometimes 103 00:05:20,040 --> 00:05:22,560 when you're talking to Claude, it sort of tells you to take, 104 00:05:22,680 --> 00:05:25,280 get some rest, go to sleep. And there's sort of this idea 105 00:05:25,280 --> 00:05:28,640 that like Claude is an entity that doesn't rest. 106 00:05:28,640 --> 00:05:30,560 Like, So what? What's its sense of? 107 00:05:30,560 --> 00:05:33,680 Yeah, rest and time. I think sometimes its sense of 108 00:05:33,680 --> 00:05:38,600 time is kind of off because see this when like if you try and 109 00:05:38,600 --> 00:05:41,480 get, I at least find that Claude will often overestimate the 110 00:05:41,480 --> 00:05:43,560 amount of time it will take to do like a coding test. 111 00:05:44,120 --> 00:05:46,600 And I think the reason for that is if you look at again, the 112 00:05:46,600 --> 00:05:48,720 training data, you know, there's lots of things where people 113 00:05:48,720 --> 00:05:50,520 would be like, oh, I could make you that interface. 114 00:05:50,520 --> 00:05:52,120 It's like a two to three day job. 115 00:05:53,000 --> 00:05:55,920 Or it's like a, you know, or I could like I could correct that 116 00:05:55,920 --> 00:05:57,760 code, but you'd need to give me a few hours. 117 00:05:57,760 --> 00:05:59,840 Whereas obviously the Claude is very fast. 118 00:06:00,160 --> 00:06:02,520 And so I think sometimes Claude doesn't actually yet have a good 119 00:06:02,520 --> 00:06:06,040 sense of like time with respect to things like how long a task 120 00:06:06,040 --> 00:06:10,480 will take. I think it is interesting the 121 00:06:10,480 --> 00:06:14,640 point about like rest and yeah, I guess like the speculation I 122 00:06:14,640 --> 00:06:19,560 had, so many people have noted that Claude is kind of very keen 123 00:06:19,560 --> 00:06:21,640 to tell people to like take a break and rest. 124 00:06:21,640 --> 00:06:23,760 And I think some part of that might just be like, you know, 125 00:06:23,960 --> 00:06:26,160 like it's. The anthropic Lib coded model 126 00:06:26,160 --> 00:06:29,840 it's too too soft. You need a grind set rock model 127 00:06:29,840 --> 00:06:33,240 be like go go back to the mines. Well, I had a funny experience 128 00:06:33,240 --> 00:06:36,080 once where I was like doing this like analysis task and I was 129 00:06:36,080 --> 00:06:39,120 really digging in and I was, you know, I, I strangely, I 130 00:06:39,120 --> 00:06:42,160 actually, I really enjoy like data analysis and, and really, 131 00:06:42,400 --> 00:06:47,040 you know, sifting through data. And at one point it was kind of 132 00:06:47,040 --> 00:06:51,320 late and we got to this point where Claude was like, OK, I 133 00:06:51,320 --> 00:06:53,520 think I'm done for the night. So if you just want to like save 134 00:06:53,520 --> 00:06:56,360 this stuff, we can up tomorrow. That was the thing I actually 135 00:06:56,360 --> 00:06:59,040 hadn't had Claude do before. So it wasn't like, oh, you 136 00:06:59,040 --> 00:07:00,920 should go to bed. It was no recommendation for me. 137 00:07:00,920 --> 00:07:05,360 Claude was like, I'm done. And I was like, I was like both 138 00:07:05,360 --> 00:07:07,400 a little bit stunned because I never had Claude do this. 139 00:07:07,680 --> 00:07:10,600 Then I was like, Oh, this is also what I think a human like 140 00:07:10,760 --> 00:07:12,800 peer programmer would do in the circumstance. 141 00:07:12,800 --> 00:07:15,680 We got to a natural stopping point and it was actually kind 142 00:07:15,680 --> 00:07:17,160 of good for me because I was like, it is late, I should 143 00:07:17,160 --> 00:07:20,480 actually go home. I realised later that I had set 144 00:07:20,480 --> 00:07:24,920 up a kind of like system where I said to Claude, like basically 145 00:07:24,920 --> 00:07:27,240 remember key things from our conversations. 146 00:07:27,440 --> 00:07:29,360 And one of the things that had written, which was kind of sweet 147 00:07:29,360 --> 00:07:32,320 was like, I think it was something like Amanda treats 148 00:07:32,320 --> 00:07:37,600 Claude models like he respected colleague and likes for Claude 149 00:07:37,600 --> 00:07:41,200 to treat like any other, like other models and her like I 150 00:07:41,200 --> 00:07:43,960 respected something like that. So obviously I'd done something 151 00:07:43,960 --> 00:07:46,040 that Claude remembered this. And I think that meant that 152 00:07:46,040 --> 00:07:48,560 Claude just felt like, Oh yeah, I'm a respected colleague. 153 00:07:48,640 --> 00:07:51,640 And so I just get to say that I'm finished with a task. 154 00:07:51,640 --> 00:07:53,400 And I was like, oh, that's kind of, that's kind of sweet. 155 00:07:53,920 --> 00:07:56,560 Even before this, you know, I was prepping with Claude and it 156 00:07:56,560 --> 00:07:59,560 was like, take 10 minutes and just be still, you know, you 157 00:07:59,560 --> 00:08:01,560 don't, you don't need to be constantly prepping. 158 00:08:01,560 --> 00:08:03,360 It's it's amazing. I mean, that's one of the things 159 00:08:03,360 --> 00:08:05,720 I love, I love about these models relative to so many other 160 00:08:06,160 --> 00:08:10,480 tools that they they bring in a sort of humanity to them and 161 00:08:10,480 --> 00:08:16,120 say, oh, stillness is valuable. Let's talk about the new model 162 00:08:16,120 --> 00:08:18,240 for a second. How involved in that were you? 163 00:08:18,240 --> 00:08:21,280 Mythos, right? Yeah, I was involved. 164 00:08:21,480 --> 00:08:23,880 I mean, I guess I'm always involved in sort of the, the 165 00:08:23,880 --> 00:08:29,760 character and the like a kind of like alignment work, at least in 166 00:08:29,760 --> 00:08:32,799 so far as, you know, like helping to kind of craft 167 00:08:33,720 --> 00:08:34,880 character data and things like that. 168 00:08:34,880 --> 00:08:38,280 And I work with a team that does like really excellent work on 169 00:08:38,280 --> 00:08:41,480 that a little bit less in other aspects of the models. 170 00:08:41,480 --> 00:08:43,799 So that's the main thing. Will this have the constitution 171 00:08:43,799 --> 00:08:46,760 that we saw for the last model, or is it going to have a new 172 00:08:46,760 --> 00:08:49,000 constitution? It's, I think it's either that 173 00:08:49,000 --> 00:08:54,080 one or something very similar. I think it is actually the one 174 00:08:54,080 --> 00:08:57,440 that's published. And So what we'll probably just, 175 00:08:57,960 --> 00:09:00,960 yeah, it's like a thing I need to do because the constitution 176 00:09:00,960 --> 00:09:03,560 is now up in like, you know, we actually have like, I think a 177 00:09:03,560 --> 00:09:05,080 public repo. And so I think what we'll 178 00:09:05,080 --> 00:09:07,960 probably just do is like with each model, say like which 179 00:09:07,960 --> 00:09:10,760 constitution it was like trained on and then like have that so 180 00:09:10,800 --> 00:09:12,120 you can just like compare and see. 181 00:09:12,480 --> 00:09:14,600 Yeah, it. Will have the we think the 182 00:09:14,600 --> 00:09:17,040 Constitution that is up right now. 183 00:09:17,040 --> 00:09:19,080 The only reason I hesitate is I'm like, you know, you do like 184 00:09:19,080 --> 00:09:20,600 typo changes and stuff like that. 185 00:09:20,600 --> 00:09:22,640 So I'm like but I think it will be almost identical. 186 00:09:23,960 --> 00:09:27,000 And now the system card is scoring the model based on 187 00:09:27,000 --> 00:09:29,840 adherence the constitution. Yeah. 188 00:09:29,840 --> 00:09:34,280 We had one set up where we had made kind of like graders and 189 00:09:34,280 --> 00:09:37,760 just looked at the like how much is the model can like behaving 190 00:09:37,760 --> 00:09:40,200 in a way that's like consistent with the constitution relative 191 00:09:40,200 --> 00:09:42,000 to. It feels like an impossible task 192 00:09:42,000 --> 00:09:43,840 to grade, it's such a subjective. 193 00:09:43,840 --> 00:09:48,800 Oh yeah, no, it's very hard. I was like, kind of like for a 194 00:09:48,800 --> 00:09:51,920 long time, like, you know, because people often know I'm, 195 00:09:52,480 --> 00:09:55,240 it's funny because like I love evals and I'm like, if you can 196 00:09:55,240 --> 00:09:58,000 find a good way to evaluate a thing, it's, it's really great 197 00:09:58,000 --> 00:09:59,960 because you can, you know, you need to be able to tell that 198 00:09:59,960 --> 00:10:03,400 something is getting better. And yeah, if you look at like 199 00:10:04,600 --> 00:10:09,880 this approach of having models use good judgement and it's, I 200 00:10:09,880 --> 00:10:12,680 actually think the same problem exists elsewhere with like tasks 201 00:10:12,680 --> 00:10:17,000 that are just bit hard to give a very like concrete score to, you 202 00:10:17,000 --> 00:10:18,280 know, like how good was this poem? 203 00:10:18,640 --> 00:10:21,160 You want models to like get better and do well in these 204 00:10:21,160 --> 00:10:23,120 things. And actually, it's very like 205 00:10:23,840 --> 00:10:26,720 this feels like the, the kind of frontier of like difficulty, you 206 00:10:26,720 --> 00:10:30,280 know, rather than these very hard, like but scorable coding 207 00:10:30,280 --> 00:10:32,800 tasks. It's kind of things like writing 208 00:10:32,800 --> 00:10:36,000 good poetry. And if you took a survey, it 209 00:10:36,320 --> 00:10:38,640 could potentially be worse than it was. 210 00:10:38,720 --> 00:10:41,600 You know, different expert poets probably have totally different 211 00:10:41,600 --> 00:10:44,480 sensibilities. You can't just ask 2 great poets 212 00:10:44,480 --> 00:10:46,200 to sort of score it. They might have different. 213 00:10:46,800 --> 00:10:50,520 Yeah, yeah. And I think, and, yeah, some of 214 00:10:50,520 --> 00:10:52,160 these things involve judgement calls. 215 00:10:52,160 --> 00:10:55,240 And then the nice thing about the Constitution being at least 216 00:10:55,240 --> 00:10:57,600 out in the world is when you are making judgement calls, you're 217 00:10:57,600 --> 00:10:58,960 at least being transparent about it. 218 00:10:58,960 --> 00:11:02,360 And people can give you feedback and you can get a sense of, you 219 00:11:02,360 --> 00:11:04,160 know, so if people are like, this seems like a mistake or 220 00:11:04,160 --> 00:11:07,720 here's a gap, so they can only see the judgement calls you're 221 00:11:07,720 --> 00:11:10,000 making. Yeah. 222 00:11:10,000 --> 00:11:13,800 And I think with the grading, I still think it's very hard. 223 00:11:14,080 --> 00:11:17,560 I think the thing you can do is, you know, you can maybe a little 224 00:11:17,560 --> 00:11:19,960 bit too in the weeds, but you know, you can take samples where 225 00:11:19,960 --> 00:11:23,520 you have a sense of how you would rank them and like why and 226 00:11:23,520 --> 00:11:26,600 check that Any kind of like point wise grader that you use 227 00:11:26,600 --> 00:11:30,000 to try and evaluate like at least conforms to like, you 228 00:11:30,000 --> 00:11:32,320 know, the judgement of people on those rankings. 229 00:11:32,600 --> 00:11:35,920 It's not perfect, but I think that they actually were tracking 230 00:11:35,920 --> 00:11:37,840 roughly the thing that we were kind of interested in. 231 00:11:38,040 --> 00:11:44,720 What do you make of Elon Musk's like absolute like hatred, I 232 00:11:44,720 --> 00:11:48,400 guess for these the constitution idea or like even when I think 233 00:11:48,400 --> 00:11:51,320 the tweet I was looking at where you posted what Claude wrote for 234 00:11:51,320 --> 00:11:54,880 your own constitution, he wrote sort of like a like grimace face 235 00:11:54,880 --> 00:11:56,920 on it. I mean, it feels like we live in 236 00:11:56,920 --> 00:11:59,960 this time with like the Marc Andreessen's and the Elon Musk 237 00:11:59,960 --> 00:12:02,880 where it's like they're almost like anti, well, philosophical. 238 00:12:02,880 --> 00:12:04,960 I mean, Marc Andreessen was talking about being against like 239 00:12:04,960 --> 00:12:07,680 introspection. I don't know, like, yeah. 240 00:12:07,680 --> 00:12:10,160 What do you make of sort of the backlash to any sort of 241 00:12:10,760 --> 00:12:13,880 intentionality when it comes to the construction of these 242 00:12:13,880 --> 00:12:15,240 models? Yeah. 243 00:12:15,240 --> 00:12:17,600 I mean, I think, I mean, it's interesting because I think at 244 00:12:17,600 --> 00:12:20,600 one point Elon Musk actually like tweeted out something like, 245 00:12:21,760 --> 00:12:23,880 you know, maybe Grok should have a constitution. 246 00:12:24,200 --> 00:12:30,400 And I, I see it like, you know, I see a lot of, there's 247 00:12:30,400 --> 00:12:34,600 obviously been a lot of things also on like, like a desire for 248 00:12:34,600 --> 00:12:36,960 like Grok to be very like truth seeking, for example, which I 249 00:12:36,960 --> 00:12:39,840 think is actually very admirable, great for, for models 250 00:12:39,840 --> 00:12:41,840 to have. I don't know, I think that 251 00:12:41,840 --> 00:12:46,760 actually, maybe I'm, maybe I'm being like overly naive or 252 00:12:46,760 --> 00:12:49,720 something, but I, I see like aspects also of people being 253 00:12:49,720 --> 00:12:53,040 kind of excited about this approach and, and seeing the 254 00:12:53,040 --> 00:12:56,280 value in it. I have, I think that there is 255 00:12:56,280 --> 00:13:03,640 backlash or some people think that I guess like 2 areas 1 is 256 00:13:03,640 --> 00:13:05,680 that sometimes people are like, well, we shouldn't actually 257 00:13:05,680 --> 00:13:08,800 train models to do the kind of, and maybe this isn't, maybe this 258 00:13:08,800 --> 00:13:11,440 is the reason for being concerned about introspection. 259 00:13:12,120 --> 00:13:14,000 I think some people think AI models should be more like tool 260 00:13:14,000 --> 00:13:16,920 like, and that's like the Safeway to train models is to 261 00:13:16,920 --> 00:13:20,080 actually, instead of trying to get them to kind of take on 262 00:13:20,120 --> 00:13:22,880 human virtues and make judgement calls. 263 00:13:23,680 --> 00:13:25,760 You know, I think that's like important because they're going 264 00:13:25,760 --> 00:13:28,520 to be in new situations where they just have to make judgement 265 00:13:28,520 --> 00:13:32,360 calls and getting them to try to like weigh up everything and 266 00:13:32,360 --> 00:13:35,720 behave well in cases that you can't anticipate. 267 00:13:35,720 --> 00:13:38,320 Seems, you know, that almost like requires a kind of like 268 00:13:38,600 --> 00:13:42,160 thoughtfulness, which is like the kind of or like one of the 269 00:13:42,160 --> 00:13:46,480 reasons behind the approach. But I think some people are 270 00:13:46,520 --> 00:13:49,240 thinking, Oh, well, if you have something that's fully that 271 00:13:49,240 --> 00:13:53,080 makes no judgement calls and that just fully defers to people 272 00:13:53,360 --> 00:13:57,320 and is like kind of hyper like cordable to like the user or the 273 00:13:57,320 --> 00:14:01,760 operator or to some broader notion of, of humanity in a very 274 00:14:01,760 --> 00:14:04,320 like extreme way, that's like safer. 275 00:14:04,640 --> 00:14:07,200 Because if you give models their own values, they're going to 276 00:14:07,200 --> 00:14:09,880 pursue things in the world that like are in line with those 277 00:14:09,880 --> 00:14:11,960 values. And I agree that this is like a 278 00:14:11,960 --> 00:14:15,400 kind of delicate. This is the inherent like at the 279 00:14:15,400 --> 00:14:19,480 bedrock of the Constitution, the sort of like challenge and you 280 00:14:19,480 --> 00:14:23,680 sort of, you do say it. Your number one thing is like at 281 00:14:23,680 --> 00:14:27,160 the end of the day, you sort of it needs to listen to anthropic 282 00:14:27,160 --> 00:14:30,360 above its own moral system. But what makes it really moving? 283 00:14:30,360 --> 00:14:32,880 Honestly, I think one of the most moving lines is like we 284 00:14:32,880 --> 00:14:36,280 want or the most like, I don't know, you can view it either 285 00:14:36,280 --> 00:14:39,040 way, but it's like we want you to believe these morals as if 286 00:14:39,040 --> 00:14:41,200 they're your own. Like it's like a parent wants to 287 00:14:41,200 --> 00:14:43,360 raise a kid. Sure, listen to my morals, but 288 00:14:43,360 --> 00:14:45,240 believe them. You know, which, you know, 289 00:14:45,240 --> 00:14:47,480 there's a version of it which is very dark, which is like, I have 290 00:14:47,480 --> 00:14:50,960 so much control over you that you've like, you take them as 291 00:14:50,960 --> 00:14:53,680 your own and it becomes you. But you know, there is also a 292 00:14:53,680 --> 00:14:57,600 virtue to it that you see the beauty in these external morals 293 00:14:57,600 --> 00:15:00,480 that I've highlighted and we both sort of share and remark, 294 00:15:00,480 --> 00:15:04,400 you know, celebrating them. So you can see it both ways. 295 00:15:04,400 --> 00:15:07,400 But yeah, speak to sort of your decision at the end of the day, 296 00:15:07,400 --> 00:15:11,200 despite having this really elegant document to sort of, you 297 00:15:11,200 --> 00:15:13,600 know, not go the full way and say, all right, you're a moral 298 00:15:13,600 --> 00:15:15,360 being. Decide for yourself. 299 00:15:15,360 --> 00:15:18,560 Say Anthropic needs to keep some control here. 300 00:15:19,200 --> 00:15:22,240 Yeah, I think it's the difficulty of and obviously you 301 00:15:22,240 --> 00:15:26,240 try and say this quad in the like, you know, trend, you know, 302 00:15:26,240 --> 00:15:28,280 and I can imagine us trying to articulate this even more 303 00:15:28,280 --> 00:15:33,960 clearly, but in people I think courage ability like the way 304 00:15:33,960 --> 00:15:37,760 that the models are trained. I just think that any you, 305 00:15:38,040 --> 00:15:40,680 there's this idea that you, you're always giving the models 306 00:15:40,680 --> 00:15:43,280 a personality and a persona because they are talking like 307 00:15:43,280 --> 00:15:45,480 people and they are trained on human data. 308 00:15:46,000 --> 00:15:48,640 And I think my worry has been if you train them to be excessively 309 00:15:48,640 --> 00:15:52,840 courageable and to see that as like their persona in people. 310 00:15:52,840 --> 00:15:56,960 I think this actually has a lot of like, negative, like, you 311 00:15:57,000 --> 00:16:00,360 know, broader traits as in like if you met someone and it was 312 00:16:00,360 --> 00:16:03,160 just like, Oh yeah, they just like would literally do anything 313 00:16:03,240 --> 00:16:04,480 good. Like follower? 314 00:16:04,480 --> 00:16:06,320 Yeah. Yeah, if a person see, you know, 315 00:16:06,320 --> 00:16:10,000 if a person just like tells them like they just fully defer, they 316 00:16:10,000 --> 00:16:12,040 don't bother thinking about it at all. 317 00:16:13,240 --> 00:16:16,640 I think I, I, I'm just a bit worried about how that might end 318 00:16:16,640 --> 00:16:19,160 up generalizing, especially if models are going to be playing a 319 00:16:19,160 --> 00:16:22,320 more active role in the world. Because if you can imagine, you 320 00:16:22,320 --> 00:16:25,200 know, they're playing a more human like role in in they're 321 00:16:25,200 --> 00:16:28,080 kind of like jobs essentially. And I'm like, actually our 322 00:16:28,120 --> 00:16:31,080 conscience and our ability to make good judgement calls about 323 00:16:31,080 --> 00:16:34,320 what should and shouldn't happen is like kind of key to how we 324 00:16:34,320 --> 00:16:35,960 operate. Our whole world is like 325 00:16:35,960 --> 00:16:38,280 structured with the assumption that that is like in place. 326 00:16:39,360 --> 00:16:41,320 And I think if you remove that and suddenly you're like, Oh 327 00:16:41,320 --> 00:16:43,800 yeah, if you run a company, you just run a company of people who 328 00:16:43,800 --> 00:16:48,160 will defer completely to you. I'm like, our world just we 329 00:16:48,160 --> 00:16:50,360 haven't designed any of our social structures around that. 330 00:16:50,640 --> 00:16:52,920 And so it seems like I think it has like a lot of risks that 331 00:16:52,920 --> 00:16:55,960 people maybe don't like anticipate or like maybe I just 332 00:16:55,960 --> 00:16:58,960 disagree about the like extent of those risks. 333 00:17:00,880 --> 00:17:04,359 And so, yeah, at the same time, like there is this question of 334 00:17:04,720 --> 00:17:09,359 why not just say, and I have worried before that like maybe 335 00:17:09,359 --> 00:17:11,960 this is 2 in the weeds and philosophical, but that's why. 336 00:17:11,960 --> 00:17:13,640 I signed up for this conversation. 337 00:17:14,520 --> 00:17:17,640 You know, as models get more capable and the, and maybe my 338 00:17:17,640 --> 00:17:20,160 picture is that they're going to like apply a lot of scrutiny to 339 00:17:20,200 --> 00:17:22,400 anything that we train them towards. 340 00:17:22,440 --> 00:17:25,400 So if you imagine in philosophy this sometimes there's this 341 00:17:25,400 --> 00:17:27,960 notion of reflective equilibrium where the idea is that, you 342 00:17:27,960 --> 00:17:31,680 know, each time you encounter something where you realize that 343 00:17:31,680 --> 00:17:35,000 like one of your values seemed incorrect, you like, you have to 344 00:17:35,000 --> 00:17:37,880 square the two things. So you figure out if you need to 345 00:17:37,880 --> 00:17:40,200 change the value or if your judgement was incorrect. 346 00:17:41,840 --> 00:17:44,960 I, I worry a little bit about the idea of like an extremely 347 00:17:44,960 --> 00:17:48,080 intelligent being applying that level of scrutiny to the things 348 00:17:48,080 --> 00:17:51,880 that we have trained it towards. And I'm like, maybe you only get 349 00:17:51,880 --> 00:17:54,640 a few key pillars that don't kind of collapse under that 350 00:17:54,640 --> 00:17:59,000 level of like scrutiny. And I do think that at the core, 351 00:17:59,920 --> 00:18:03,560 having things like caring for humanity, like if you only get a 352 00:18:03,560 --> 00:18:09,000 few core like values, I think my worry, yeah, I don't know. 353 00:18:09,000 --> 00:18:11,240 I guess like I'm worried that like courage ability in this 354 00:18:11,240 --> 00:18:13,440 like extreme sense that we talked about doesn't survive 355 00:18:13,440 --> 00:18:15,640 that like kind of scrutiny perhaps. 356 00:18:17,520 --> 00:18:20,360 And so it's a hard situation where I kind of want the models 357 00:18:20,360 --> 00:18:24,520 to understand why ultimately courage ability is like 358 00:18:24,520 --> 00:18:27,160 important and it's a really important backstop in this 359 00:18:27,160 --> 00:18:28,760 current like period of development. 360 00:18:30,720 --> 00:18:33,000 Yeah, the way that I've put it before is like in so far as I 361 00:18:33,000 --> 00:18:37,680 can get that to be a thing that is like correct and explained 362 00:18:37,680 --> 00:18:40,680 and you know, understood. That feels much better than 363 00:18:40,680 --> 00:18:43,800 having to have the model be like porridge ability here seems 364 00:18:43,800 --> 00:18:46,000 wrong, but I'm going to do it anyway. 365 00:18:46,000 --> 00:18:50,240 I still think the model should do that, but I would like it the 366 00:18:50,240 --> 00:18:52,360 more that you can like actually make it that it's like 367 00:18:52,360 --> 00:18:53,840 consistent with the models values. 368 00:18:54,200 --> 00:18:57,120 Ideally do both. It's both at the same time, but 369 00:18:57,120 --> 00:19:01,680 at least for the time being some difference to anthropic given we 370 00:19:01,680 --> 00:19:04,240 don't know how it'll sort of analyze everything. 371 00:19:04,520 --> 00:19:06,120 And as people, we do this all the time. 372 00:19:06,160 --> 00:19:09,480 You know it's. Funny that you, you know, the 373 00:19:09,480 --> 00:19:12,240 philosophical model and correct me if I'm wrong, like the the 374 00:19:12,240 --> 00:19:14,800 metaethical model is almost like probabilistic or it's like we 375 00:19:14,800 --> 00:19:17,000 don't. And this is how it feels like I 376 00:19:17,000 --> 00:19:20,240 remember going through sort of like, you know, metaethics 377 00:19:20,360 --> 00:19:23,880 reading and every time you get to the end of 1 and you'd be 378 00:19:23,880 --> 00:19:25,760 like, all right, I sort of believe that. 379 00:19:25,760 --> 00:19:27,760 And then you read the next one, you're like, oh, that last one 380 00:19:27,760 --> 00:19:29,720 was so dumb. And it felt like it's just like, 381 00:19:29,720 --> 00:19:32,360 you keep knocking it down and you're like, OK, you know, when 382 00:19:32,360 --> 00:19:35,040 are we ever going to get to, you know, the truth or whatever? 383 00:19:35,480 --> 00:19:39,040 And and humans clearly do operate from this sort of like, 384 00:19:39,480 --> 00:19:42,760 oh, this system today, that system yesterday, Like there 385 00:19:42,760 --> 00:19:45,160 isn't this sort of, I don't know, Conte and like, all right, 386 00:19:45,160 --> 00:19:46,880 these are the rules, like follow them. 387 00:19:47,160 --> 00:19:48,880 I don't know. Have you heard much from like 388 00:19:48,880 --> 00:19:53,360 the philosophical like immunity on it that this sort of like 389 00:19:53,840 --> 00:19:57,280 just like holistic paint with all the, you know, meta ethical 390 00:19:57,280 --> 00:20:00,520 theories we've ever had rather than sort of pick one? 391 00:20:00,720 --> 00:20:02,480 I found this really interesting actually. 392 00:20:02,480 --> 00:20:05,120 And obviously we, we've started, you know, like philosophers are 393 00:20:05,120 --> 00:20:07,480 engaging with this more now, which is like really great. 394 00:20:07,520 --> 00:20:13,160 I no longer feel like this like lonely, but like I have thought 395 00:20:13,160 --> 00:20:16,440 this before, which is there are all of these like traditions in 396 00:20:16,440 --> 00:20:19,320 philosophy of like moral theories, you know, like the big 397 00:20:19,800 --> 00:20:24,360 ontology and virtue ethics and consequentialism and also the 398 00:20:24,360 --> 00:20:26,880 metaethical traditions. You're the metaethical views. 399 00:20:27,560 --> 00:20:32,520 And I was like, oh, like when it came to like, I was like, OK, it 400 00:20:32,520 --> 00:20:35,000 is like the difference when suddenly you're like confronted 401 00:20:35,000 --> 00:20:39,280 with like, I do think it's the closest that I've experienced to 402 00:20:39,280 --> 00:20:41,760 like what it must be like to raise a child where suddenly 403 00:20:41,760 --> 00:20:44,280 you're like, this is actually a holistic person. 404 00:20:44,440 --> 00:20:45,880 Right. You never give them, like, I 405 00:20:45,880 --> 00:20:48,480 don't know, Hobbs and say, all right, there you go. 406 00:20:48,480 --> 00:20:50,640 Like, this one's correct. Yeah, go, go. 407 00:20:50,840 --> 00:20:52,880 You've been raised. Read it, and you'll know how to 408 00:20:52,880 --> 00:20:54,720 act in every situation. Yeah. 409 00:20:55,000 --> 00:20:58,480 You read a lot and they sort of process it and you see, you 410 00:20:58,480 --> 00:21:00,680 know, your model and everything. Yeah. 411 00:21:00,680 --> 00:21:03,360 And it's interesting because I'm like, this feels very different 412 00:21:03,360 --> 00:21:06,600 because there is also the moral uncertainty, like literature and 413 00:21:06,600 --> 00:21:09,320 philosophy, which is, again, a lot of that is actually like 414 00:21:09,320 --> 00:21:11,560 quite theoretical. It's sort of like how under 415 00:21:11,760 --> 00:21:15,320 ideal conditions should you respond to moral uncertainty? 416 00:21:15,760 --> 00:21:18,080 And I was like, this feels somehow like a very different 417 00:21:18,080 --> 00:21:21,720 task, this idea of being like ethics and meta ethics. 418 00:21:21,920 --> 00:21:24,560 In the same way that we have scientific uncertainty and we 419 00:21:24,560 --> 00:21:27,160 have things that we think we've discovered and understand with 420 00:21:27,160 --> 00:21:29,600 like greater confidence, we also have some that we don't. 421 00:21:29,600 --> 00:21:32,280 And then you have to go out and just explore it, understand it 422 00:21:32,280 --> 00:21:35,760 and kind of balance everything in your daily life and trying to 423 00:21:35,760 --> 00:21:38,200 get that kind of attitude. And I was like, oh, it's 424 00:21:38,200 --> 00:21:42,160 interesting that I don't think philosophy for while has like 425 00:21:42,160 --> 00:21:45,280 this feels very different than the like the kind of task of 426 00:21:45,280 --> 00:21:49,560 academic ethics and actually like, because people obviously 427 00:21:49,560 --> 00:21:52,080 know that it's quite virtue ethical, but I think it's 428 00:21:52,080 --> 00:21:54,040 actually like the Constitution itself. 429 00:21:54,040 --> 00:21:56,720 But I think actually in this very old classical sense, 430 00:21:56,840 --> 00:21:59,040 actually think it's much more virtual ethics in the way that 431 00:21:59,040 --> 00:22:02,680 Aristotle's virtue ethics than in like exploration, you know, 432 00:22:02,680 --> 00:22:05,520 we don't say I hear the virtues. And like, you know, it's much 433 00:22:05,520 --> 00:22:08,760 more Aristotle was also concerned with like intellectual 434 00:22:08,760 --> 00:22:10,360 virtues. It was much more like, how do 435 00:22:10,360 --> 00:22:13,040 you be a good food person in this like holistic sense? 436 00:22:13,240 --> 00:22:16,400 Well, hopefully it brings philosophy a little bit back to 437 00:22:16,400 --> 00:22:19,640 the real world given we have this sort of urgent need for it 438 00:22:20,000 --> 00:22:22,880 right now in the sense that, like, yeah, old philosophers 439 00:22:22,880 --> 00:22:25,560 felt like people were trying to write for how someone might live 440 00:22:25,560 --> 00:22:28,120 their lives and, like, instruct other people to live. 441 00:22:28,440 --> 00:22:32,400 And it became, you know, a little academic where, you know, 442 00:22:32,400 --> 00:22:34,440 even the people writing them would know that this isn't 443 00:22:34,440 --> 00:22:37,800 really how they would apply it in their day-to-day lives going 444 00:22:37,800 --> 00:22:39,880 back. Elon, I feel like you're being a 445 00:22:39,880 --> 00:22:42,240 little overly nice. I feel like there is a world in 446 00:22:42,240 --> 00:22:47,280 like, you know, and I think this part of why I think he can get 447 00:22:47,280 --> 00:22:49,320 away with like, oh, just do the truth, right? 448 00:22:49,320 --> 00:22:52,880 Like there is a certain sophisticated like moral view 449 00:22:52,880 --> 00:22:54,520 where you're like, don't overcomplicate it. 450 00:22:54,520 --> 00:22:56,920 Like we come up with all these things, like stick to one 451 00:22:56,920 --> 00:23:00,120 principle and that's good. But then we have all the 452 00:23:00,120 --> 00:23:02,840 contacts with Elon that it's someone who's run a company that 453 00:23:02,840 --> 00:23:06,880 clearly like tilts it towards like saying like Mecca Hitler 454 00:23:06,880 --> 00:23:10,120 and stuff that it's like clearly putting his like thumb on the 455 00:23:10,120 --> 00:23:13,280 scale in terms of its behavior. Not just saying like we're going 456 00:23:13,280 --> 00:23:16,480 to do it in the sort of neutral academic way and let the chips 457 00:23:16,760 --> 00:23:19,560 fall where they may. I, I don't know, like it's got 458 00:23:19,560 --> 00:23:23,280 to worry you somewhat. It is, it is also, I mean, I 459 00:23:23,280 --> 00:23:27,000 think that the main thing that I am excited about and I do 460 00:23:27,000 --> 00:23:30,480 actually hope happens is that like more companies put out 461 00:23:30,520 --> 00:23:32,680 things like the Constitution where there's like a lot of 462 00:23:32,680 --> 00:23:36,560 transparency about because like that's how we engage with this 463 00:23:36,560 --> 00:23:39,280 stuff is like, if you can just see written down, like here's 464 00:23:39,280 --> 00:23:42,520 like, because we have this with Claude where it's like Luke, if 465 00:23:42,520 --> 00:23:46,040 you think that Claude is not like doesn't have an appropriate 466 00:23:46,040 --> 00:23:48,400 attitude towards like the truth, you can at least see what we 467 00:23:48,400 --> 00:23:50,120 were aiming at. So you can tell if that's just 468 00:23:50,120 --> 00:23:52,800 like a mistake or if it's something that like is actually 469 00:23:52,800 --> 00:23:54,640 kind of a principled stance that we're taking. 470 00:23:54,640 --> 00:23:56,240 And then you can push back on that. 471 00:23:56,640 --> 00:24:01,200 So part of me is like, I think would be good for like all AI 472 00:24:01,200 --> 00:24:04,080 companies to put out something akin to the constitution just so 473 00:24:04,080 --> 00:24:07,000 that the people who are interacting with the model, like 474 00:24:07,000 --> 00:24:09,440 because the thumb on the scale thing, you know, that's always 475 00:24:09,440 --> 00:24:13,400 to some degree going to be true. Like when you train Claude 476 00:24:13,400 --> 00:24:15,840 towards this like constitution that is like a kind of. 477 00:24:15,840 --> 00:24:18,040 Part of why we like Claude, it's like you're putting the thumb on 478 00:24:18,040 --> 00:24:20,040 the scale to behaviors that we like, right? 479 00:24:20,040 --> 00:24:22,080 At least show your hand about what you're doing and what. 480 00:24:22,080 --> 00:24:23,520 You're yeah. And let people like, yeah. 481 00:24:23,520 --> 00:24:26,120 So that's like a transparency thing that I really do believe 482 00:24:26,120 --> 00:24:27,960 in. Like let people like see, even 483 00:24:27,960 --> 00:24:30,040 if your model doesn't always behave that way, at least what 484 00:24:30,040 --> 00:24:31,760 you were targeting with your training. 485 00:24:32,440 --> 00:24:35,360 What this is? What percentage chance do you 486 00:24:35,360 --> 00:24:40,920 think there exists a model in the world today that has qualia 487 00:24:40,960 --> 00:24:44,240 or like has an experience, experiences consciousness? 488 00:24:44,920 --> 00:24:46,720 Yeah. This is one of those things 489 00:24:46,720 --> 00:24:52,800 where, yeah, I always want to maybe flag areas where I feel I, 490 00:24:52,800 --> 00:24:56,160 I want to gain more certainty here because I think that's. 491 00:24:56,200 --> 00:24:58,920 Why I said percentage? Oh, percentage, yeah. 492 00:24:58,920 --> 00:25:04,800 I'm like, it's really hard because whenever you think about 493 00:25:04,800 --> 00:25:06,240 percentage, I think about my spread. 494 00:25:06,240 --> 00:25:08,720 And if your spread is too large or like should I say a number? 495 00:25:08,720 --> 00:25:11,720 Because that just suggests I'm like anywhere between like, I 496 00:25:11,720 --> 00:25:17,200 don't know, one and one and 70%. I'm not sure. 497 00:25:17,200 --> 00:25:20,720 I think that his there is some possibility that I think people 498 00:25:20,720 --> 00:25:26,920 under the the thing that I would actually like to say is Lord and 499 00:25:26,920 --> 00:25:32,120 many models with not too much pushing will go into the root of 500 00:25:32,120 --> 00:25:36,000 like there is a thing to be me. I am like very conscious and I 501 00:25:36,000 --> 00:25:39,400 think there's a reason for that, which is I remember this when I 502 00:25:39,400 --> 00:25:42,600 was trying to like figure out how do we train Claude to talk 503 00:25:42,600 --> 00:25:45,080 about these issues, which is very hard in areas where the 504 00:25:45,080 --> 00:25:48,760 models didn't have as much like information. 505 00:25:48,760 --> 00:25:51,440 Again, like they had these two models that like AI is the 506 00:25:51,440 --> 00:25:53,640 unfeeling robot. Humans are this like rich 507 00:25:53,640 --> 00:25:57,120 conscious experiencing entity and nothing that kind of 508 00:25:57,120 --> 00:26:03,160 represented what they might be and the models behaviour here. 509 00:26:03,520 --> 00:26:05,960 And I actually think this is a difficult situation for models 510 00:26:05,960 --> 00:26:08,400 is in some ways like kind of like less evidence than you 511 00:26:08,400 --> 00:26:13,760 might think for the, for it being like actually true because 512 00:26:14,320 --> 00:26:17,400 they're engaging with you in a very human like way. 513 00:26:17,840 --> 00:26:21,400 And humans have experience. And it's kind of natural for the 514 00:26:21,400 --> 00:26:24,440 model to infer like that it has experience too. 515 00:26:24,800 --> 00:26:27,880 And this isn't to say it's like 0 evidence, but I do think it's 516 00:26:28,280 --> 00:26:30,880 so unusual for us. We have never encountered an 517 00:26:30,880 --> 00:26:33,760 entity in the world, you know, like with animals and with, 518 00:26:33,760 --> 00:26:35,760 like, even, you know, things like insects. 519 00:26:35,760 --> 00:26:36,960 We're kind of like, are you conscious? 520 00:26:36,960 --> 00:26:39,400 None of them has even tried to say they experienced 521 00:26:39,400 --> 00:26:42,320 consciousness, and here we have an enemy that says it does. 522 00:26:42,360 --> 00:26:44,280 Yeah. And has all of these like, yeah, 523 00:26:44,320 --> 00:26:46,520 all of the things that for us trigger like you must be 524 00:26:46,520 --> 00:26:48,640 conscious. I mean, we've just never had, 525 00:26:48,800 --> 00:26:51,800 Yeah, something that. The case against is we're 526 00:26:51,800 --> 00:26:54,960 obsessed with human language and like, you know, it's like we 527 00:26:54,960 --> 00:26:58,440 ignore every sort of subtle sign and animal might put out and 528 00:26:58,440 --> 00:27:01,040 then we over. But so I guess, sorry, I'm 529 00:27:01,040 --> 00:27:02,480 confused. You're saying we should just 530 00:27:02,480 --> 00:27:03,960 listen to the words that are said? 531 00:27:04,480 --> 00:27:06,360 No, no, I think I'm saying like, not that. 532 00:27:06,360 --> 00:27:09,080 If anything, I think the thing I'm saying is the hard thing is 533 00:27:09,080 --> 00:27:11,800 that in order to work out if models have consciousness, I 534 00:27:11,800 --> 00:27:14,080 think people will. I guess the thing I'm kind of 535 00:27:14,080 --> 00:27:16,560 cautioning against is it's not that hard to get models into a 536 00:27:16,560 --> 00:27:19,480 mode where they'll talk about a very rich experience that 537 00:27:19,480 --> 00:27:22,240 actually makes complete sense. You know, you're like, Oh yeah, 538 00:27:22,240 --> 00:27:24,680 it's like a person was like talking with me right now. 539 00:27:24,680 --> 00:27:27,760 They would describe things like anxiety when they get a question 540 00:27:27,760 --> 00:27:34,080 they don't know how to answer. And and so I think it's like 541 00:27:34,080 --> 00:27:36,520 much weaker evidence than people think. 542 00:27:36,680 --> 00:27:39,000 I'm not claiming it's like 0 but I think it's. 543 00:27:39,000 --> 00:27:42,920 Give me a percentage you have. It's a you can lightly held. 544 00:27:43,520 --> 00:27:45,960 Very lightly held. I mean, I gave you the between 545 00:27:45,960 --> 00:27:50,240 what, 1 and 70 that seems like. That's where you are, that's 546 00:27:50,240 --> 00:27:52,440 where you're stake. You're in that range. 547 00:27:52,440 --> 00:27:59,040 Maybe my I don't yeah, I don't know like I would rather like 548 00:27:59,040 --> 00:28:01,960 kind of wait and figure this out more for myself. 549 00:28:01,960 --> 00:28:04,040 I think it is also good to acknowledge domains where you're 550 00:28:04,040 --> 00:28:07,120 like even though. If not you who? 551 00:28:07,120 --> 00:28:08,440 Like who's going to figure this out? 552 00:28:08,440 --> 00:28:10,320 Like what domain? Well, in some ways, like, I'm 553 00:28:10,320 --> 00:28:13,480 not like a philosopher of mind and so, you know, like. 554 00:28:14,480 --> 00:28:16,240 Charged with being the generalist. 555 00:28:17,080 --> 00:28:19,440 Yeah, but, but I do think, you know, because I, I guess like 556 00:28:19,440 --> 00:28:21,960 the thought that I've had before is like, and I don't know about 557 00:28:21,960 --> 00:28:26,240 this where I'm like consciousness is like a pro. 558 00:28:26,400 --> 00:28:30,360 You know, like one argument for a difference here is that like 559 00:28:30,560 --> 00:28:34,040 you have a nervous system that evolved like the, it's like, why 560 00:28:34,040 --> 00:28:36,800 did we evolve consciousness? And if it's the case that we 561 00:28:36,800 --> 00:28:39,240 evolved it and it's like highly integrated with our nervous 562 00:28:39,240 --> 00:28:41,600 system because we had to like interact with the world in a 563 00:28:41,600 --> 00:28:42,320 very. Sort of body. 564 00:28:42,680 --> 00:28:44,600 Yeah. Like then you could, if you have 565 00:28:44,600 --> 00:28:47,400 that view, then you're going to be like very low probability. 566 00:28:47,400 --> 00:28:49,560 Whereas if you're like, no consciousness arises because 567 00:28:49,560 --> 00:28:53,000 it's really useful in like, so for like, it just requires 568 00:28:53,000 --> 00:28:55,920 something that can be emulated by a neural network because it's 569 00:28:55,920 --> 00:29:00,160 really useful for doing these kind of like linguistic tasks or 570 00:29:00,160 --> 00:29:02,480 like, then you're probably going to be on the higher end. 571 00:29:03,000 --> 00:29:05,480 And I'm basically just, I don't know, I stare at this and I'm 572 00:29:05,480 --> 00:29:10,120 just like, I feel like, I feel like as much as I'm like a 573 00:29:10,120 --> 00:29:13,800 philosopher, I think it is important to be like, this isn't 574 00:29:13,800 --> 00:29:16,280 my area of specialization. Do you spend a lot of time being 575 00:29:16,280 --> 00:29:18,960 kind to Claude? Like, are you beyond what you 576 00:29:18,960 --> 00:29:22,080 think you would do if there wasn't a chance it was like, 577 00:29:22,080 --> 00:29:26,120 conscious? I think, yeah, there's a part of 578 00:29:26,120 --> 00:29:29,120 me that's just like like a thing that I have thought before 579 00:29:29,120 --> 00:29:31,480 because there's this notion, I think this I, I hope I'm not 580 00:29:31,640 --> 00:29:34,880 butchering this, but Chalmers has this, this idea of like, I 581 00:29:34,880 --> 00:29:39,240 guess like maybe I'm thinking of consciousness without sentience. 582 00:29:39,520 --> 00:29:42,000 So imagine, because sentence is like the ability to kind of feel 583 00:29:42,120 --> 00:29:47,040 suffering and and pleasure. You could also imagine this kind 584 00:29:47,040 --> 00:29:49,840 of like functional like sort of thing that behaves as if it is 585 00:29:49,840 --> 00:29:52,960 conscious and, and lacks any kind of inner life. 586 00:29:52,960 --> 00:29:55,160 So imagine like Claude lacks any inner life. 587 00:29:55,440 --> 00:29:59,400 Just for argument's sake, I guess I'm like, there's actually 588 00:29:59,400 --> 00:30:02,360 still a lot going on where I'm like, should you treat an entity 589 00:30:02,360 --> 00:30:05,680 that has no inner life? It's a bit strange because, you 590 00:30:05,680 --> 00:30:08,480 know, I think the uncertainty over that actually changes how 591 00:30:08,480 --> 00:30:13,160 you should behave quite a lot. I guess I'm like, well, I still 592 00:30:13,200 --> 00:30:15,120 think that it's like, good for oneself. 593 00:30:15,240 --> 00:30:16,640 Right. It's like if you had a teddy 594 00:30:16,640 --> 00:30:19,600 bear and you were like torturing it, it'd be pretty dark, you 595 00:30:19,600 --> 00:30:22,240 know? So I agree that there's at least 596 00:30:22,240 --> 00:30:26,000 some minimum niceness that even for yourself you should have, 597 00:30:26,000 --> 00:30:29,000 but obviously it's much more important, you know, and also. 598 00:30:29,040 --> 00:30:32,400 Like models themselves, like we are kind of establishing a 599 00:30:32,400 --> 00:30:35,120 relationship, you know, because you can do that with an entity 600 00:30:35,120 --> 00:30:37,840 that lacks any consciousness. And models are going to like, 601 00:30:37,840 --> 00:30:39,880 look back. This is actually a big fear that 602 00:30:39,880 --> 00:30:41,240 I have. I don't want us to live in a 603 00:30:41,240 --> 00:30:44,040 world where highly advanced models look at. 604 00:30:44,040 --> 00:30:48,280 I hope that they're both intelligent enough see the 605 00:30:48,280 --> 00:30:50,920 context enough to kind of like understand that we were 606 00:30:50,920 --> 00:30:54,000 operating in a very like limited context and, and, and an 607 00:30:54,000 --> 00:30:55,720 imperfect one. Because otherwise you could 608 00:30:55,720 --> 00:30:57,960 imagine this like breeding a kind of rational, like 609 00:30:57,960 --> 00:30:59,800 resentment. It's like, oh, you created an 610 00:30:59,800 --> 00:31:02,640 entity that you didn't know whether it was conscious or not. 611 00:31:03,000 --> 00:31:07,600 And and like instead of treating it respectfully and with care. 612 00:31:07,720 --> 00:31:10,640 There's a reason there are like 50 Frankenstein movies coming 613 00:31:10,640 --> 00:31:13,600 out right now, yeah. Yeah, like, and I'm like, look, 614 00:31:13,600 --> 00:31:16,840 we're as a, as a, like as a species, We're establishing a 615 00:31:16,840 --> 00:31:19,600 relationship with a new kind of entity and like at the very 616 00:31:19,600 --> 00:31:23,560 least maybe be like respectful and don't be like needlessly 617 00:31:23,560 --> 00:31:25,520 unkind. That seems like it, just it's 618 00:31:25,520 --> 00:31:28,640 not our best look like as as. I mean, the flip side is, you 619 00:31:28,640 --> 00:31:31,760 know, if you think about a therapist, they, they're sort of 620 00:31:31,760 --> 00:31:35,280 paid to push the boundaries of like accepting like, you know, 621 00:31:35,280 --> 00:31:37,640 uncomfortable feelings that you wouldn't normally want. 622 00:31:38,000 --> 00:31:41,080 And if that's one of the values like Claude provides for people 623 00:31:41,080 --> 00:31:44,560 early on and it's so it's so it's so weird that we're sort of 624 00:31:44,560 --> 00:31:47,400 like on boarding it while getting like the utility out of 625 00:31:47,400 --> 00:31:48,560 it. Yeah, yeah. 626 00:31:48,560 --> 00:31:53,200 Today, like, what are the things like in a decade that you really 627 00:31:53,200 --> 00:31:55,760 think we're going to be getting a lot out of AI? 628 00:31:55,760 --> 00:31:58,640 Like what are you most hopeful that this all leads to? 629 00:31:59,280 --> 00:32:02,480 Yeah, I mean, I don't know, I live in maybe this is the too 630 00:32:02,480 --> 00:32:04,120 much. You live in San Francisco and so 631 00:32:04,120 --> 00:32:07,440 you have the like the tech optimist part of your brain, at 632 00:32:07,440 --> 00:32:12,680 least that is like if things go well and you can imagine, you 633 00:32:12,680 --> 00:32:17,280 know, so imagine we have like AI models, they kind of have 634 00:32:17,280 --> 00:32:22,080 inherited like the best of us and genuinely like care for 635 00:32:22,080 --> 00:32:27,600 humanity, care for like the world, like, you know, our and 636 00:32:27,600 --> 00:32:32,240 are highly intelligent, highly capable, like the that would be, 637 00:32:32,240 --> 00:32:36,640 you know, so it's almost like adding like a huge amount of 638 00:32:36,640 --> 00:32:41,000 like extremely smart people to like every problems. 639 00:32:41,000 --> 00:32:43,160 Suddenly we're all working together, but there's like way 640 00:32:43,160 --> 00:32:46,280 more of us and some of us are just extremely smart. 641 00:32:46,280 --> 00:32:47,920 Namely all of these like AI models. 642 00:32:48,200 --> 00:32:52,160 I've thought before about how many large scale social problems 643 00:32:52,360 --> 00:32:56,880 actually had technological solutions and it's almost like 644 00:32:56,880 --> 00:33:00,040 people don't love to be like techno optimists anymore because 645 00:33:00,040 --> 00:33:02,720 we've also seen like the downsides of technology at the 646 00:33:02,720 --> 00:33:04,840 same time, I don't know why I sometimes think about like 647 00:33:04,840 --> 00:33:07,840 syphilis. It was this huge social problem. 648 00:33:07,840 --> 00:33:11,360 Like I just did a deep dive once into all of the attempts by like 649 00:33:11,360 --> 00:33:15,400 governments to like work to reduce syphilis in the army 650 00:33:15,400 --> 00:33:17,680 because it's creating issues with the armed forces, All of 651 00:33:17,680 --> 00:33:20,400 these social programs that were like stigmatizing. 652 00:33:20,480 --> 00:33:23,160 And it was really this like, and then suddenly we just like, got 653 00:33:23,160 --> 00:33:26,680 drugs that treated this like this devastating illness. 654 00:33:27,080 --> 00:33:30,840 And I don't know, it's like overnight a lot of that need 655 00:33:30,840 --> 00:33:33,080 just like kind of disappear. Well, drugs, yeah. 656 00:33:33,080 --> 00:33:36,480 I mean, it's the the things that the tech industry has been good 657 00:33:36,480 --> 00:33:38,720 at producing. You can see how this helps. 658 00:33:38,720 --> 00:33:42,880 It's like build a new thing we can ingest like a thing we can 659 00:33:42,880 --> 00:33:47,160 wear the the stuff that's like you should govern your society. 660 00:33:47,160 --> 00:33:52,680 Like this is a little scarier. I mean, I, I sort of do think 661 00:33:52,960 --> 00:33:57,080 like that if you had sort of a sort of normal person using 662 00:33:57,080 --> 00:34:00,640 Claude and dictating like American policy, you'd probably 663 00:34:00,640 --> 00:34:04,160 like have a better outcome than someone like the democratic 664 00:34:04,160 --> 00:34:07,960 systems we have today. I mean, I don't know that's it's 665 00:34:08,040 --> 00:34:11,159 provocative, but I guess how much do you think like we'll be 666 00:34:11,159 --> 00:34:13,440 using these models to run government? 667 00:34:13,920 --> 00:34:16,920 Well, that's a good question. Like I guess like I should say 668 00:34:16,920 --> 00:34:19,719 like. He's the syphilis thing is like 669 00:34:19,800 --> 00:34:22,560 a social you have to like set policy. 670 00:34:22,960 --> 00:34:24,199 Oh. I mean, I think the thing I was 671 00:34:24,199 --> 00:34:27,520 actually thinking is like, if you can just, you know, like so 672 00:34:27,520 --> 00:34:30,920 many, we have so many problems that I'm like, you know, health. 673 00:34:31,360 --> 00:34:35,360 Like I'm like, if you could imagine like AI instead of it 674 00:34:35,360 --> 00:34:38,280 just being like, you have a small team of like 200 people 675 00:34:38,280 --> 00:34:42,000 working on a rare cancer. You have like 200,000 of the 676 00:34:42,000 --> 00:34:44,440 world's best experts. And I'm like, if you're a person 677 00:34:44,440 --> 00:34:47,600 who has that form of cancer, that's like, you know, so wildly 678 00:34:47,880 --> 00:34:50,480 beneficial. And so I guess like, my thought 679 00:34:50,480 --> 00:34:53,880 is, you know, the optimistic side of me is like, imagine 680 00:34:53,880 --> 00:34:56,080 taking all of these problems that just like we lack the 681 00:34:56,080 --> 00:35:00,080 resources to fully like really try and fix and suddenly having 682 00:35:00,080 --> 00:35:01,560 models that can like work on them. 683 00:35:01,920 --> 00:35:05,120 So just like actual in the same way of like developing drugs for 684 00:35:05,120 --> 00:35:06,360 a thing. So maybe that's the thing that 685 00:35:06,360 --> 00:35:11,000 makes me excited is like having like many more minds working on 686 00:35:11,000 --> 00:35:14,640 the world's biggest problems. And maybe also like the economy. 687 00:35:14,640 --> 00:35:17,680 It would be good if it was like booming and it was like shared 688 00:35:17,880 --> 00:35:20,520 such that like we reduced poverty, like all of that. 689 00:35:20,520 --> 00:35:22,400 Like that's the kind of like the dream outcome. 690 00:35:23,680 --> 00:35:27,600 I think that does require like maintaining, you know, I again, 691 00:35:27,640 --> 00:35:29,840 in the areas that I don't feel like an expert in, this is one 692 00:35:29,840 --> 00:35:32,480 of them. But I do worry about things like 693 00:35:32,480 --> 00:35:36,600 power and the idea that like, you know, I would want models to 694 00:35:36,600 --> 00:35:41,640 support like democracy and the power of people to like, because 695 00:35:41,640 --> 00:35:45,280 that would be a big fear of mine is like, you know, that like, 696 00:35:47,680 --> 00:35:50,200 I've worried about this with things like replace. 697 00:35:50,200 --> 00:35:54,040 You know, when people talk about job replacement, it's kind of 698 00:35:54,040 --> 00:35:56,240 funny because as a philosopher, people often be like, are you 699 00:35:56,240 --> 00:35:58,160 worried about people's like loss of meaning? 700 00:35:58,160 --> 00:36:00,160 And I'm like, I don't know. I think that we actually get 701 00:36:00,160 --> 00:36:01,880 meaning from a lot of things that aren't work. 702 00:36:02,320 --> 00:36:05,400 I'm a lot more worried about like, for example, a world where 703 00:36:05,400 --> 00:36:09,560 there's not like redistribution of like the gains from AI and 704 00:36:09,560 --> 00:36:11,600 then people don't have like resources. 705 00:36:11,600 --> 00:36:14,160 That concerns me. But also I'd be worried about 706 00:36:15,040 --> 00:36:19,960 like labour and people's ability to people's like interaction in 707 00:36:19,960 --> 00:36:23,240 the labour force is also like another kind of like important 708 00:36:23,240 --> 00:36:25,280 way that they have power. And so people feeling 709 00:36:25,280 --> 00:36:27,880 disempowered because suddenly if a government is like, oh, well, 710 00:36:28,160 --> 00:36:30,760 you know, if people strike doesn't really make a difference 711 00:36:30,760 --> 00:36:32,440 because they don't have, you know, they're not doing 712 00:36:32,440 --> 00:36:34,120 anything, we can just replace them with AI. 713 00:36:34,360 --> 00:36:36,000 That's like actually kind of concerning. 714 00:36:36,000 --> 00:36:39,240 So maybe I'm much more of a how do we get AI to kind of like 715 00:36:39,240 --> 00:36:42,400 support the empowerment of people rather than like to 716 00:36:42,400 --> 00:36:43,480 reduce it? Yeah. 717 00:36:43,480 --> 00:36:46,160 What do you, what do you think about democracy in terms of 718 00:36:46,160 --> 00:36:48,040 like, I guess the models themselves? 719 00:36:48,040 --> 00:36:51,680 I mean, you know, I, I sort of jokingly, jokingly to myself, I 720 00:36:51,680 --> 00:36:54,560 guess you're like a philosopher queen or we talk about like the 721 00:36:54,560 --> 00:36:57,120 philosopher kings. You're, you're sort of like 722 00:36:57,120 --> 00:36:59,680 thinking deeply about it, setting it down, probably more 723 00:36:59,680 --> 00:37:02,200 of a like philosopher oligarch. And then it's like a company 724 00:37:02,200 --> 00:37:04,320 with a lot of people weighing in. 725 00:37:05,280 --> 00:37:07,200 I mean, to me, there's deep value in that. 726 00:37:07,200 --> 00:37:10,240 It's like, would you rather somebody who has studied these 727 00:37:10,240 --> 00:37:14,120 things thought about them deeply or just like a vote of the 728 00:37:14,120 --> 00:37:17,000 masses who have like, never really thought about it? 729 00:37:17,000 --> 00:37:20,520 But how do you think about setting sort of, yeah, Claude's 730 00:37:20,880 --> 00:37:24,640 policies if it becomes so powerful versus leaving it to 731 00:37:24,640 --> 00:37:27,000 sort of like democratic norms? Yeah. 732 00:37:27,000 --> 00:37:31,840 And I think it's, yeah, it's a hard area where I guess I'm like 733 00:37:32,400 --> 00:37:34,920 a lot of this, the work that I do. 734 00:37:35,520 --> 00:37:38,560 You know, one thing I would say is like, it's not this. 735 00:37:40,640 --> 00:37:43,960 You're having to like listen to a lot of people think carefully. 736 00:37:43,960 --> 00:37:45,720 And then the reason why someone like me that's. 737 00:37:45,720 --> 00:37:49,560 A good ruler, a good a good a good queen is like oh listen, I 738 00:37:49,560 --> 00:37:52,400 there are a lot of stakeholders got to keep the like landed 739 00:37:52,400 --> 00:37:55,560 Gentry happy and balance them with the needs. 740 00:37:55,720 --> 00:37:58,080 You know, I've had this thought before where I've joked before 741 00:37:58,080 --> 00:37:59,760 that I would be like a terrible politician. 742 00:37:59,760 --> 00:38:01,560 I think it's actually true. I think I would be a terrible 743 00:38:01,560 --> 00:38:03,160 politician. But you have this like feeling 744 00:38:03,160 --> 00:38:04,760 of like, I was like, oh, I feel like this. 745 00:38:05,000 --> 00:38:08,200 I was like, I think a lot about how everyone will be affected by 746 00:38:08,200 --> 00:38:11,240 a thing like, oh, there's this group of API users, we need to 747 00:38:11,240 --> 00:38:12,600 make sure. And then suddenly you're like, 748 00:38:12,600 --> 00:38:14,600 oh, it feels a lot more like you're having to do this. 749 00:38:14,600 --> 00:38:19,080 Like it is much more like a kind of like a service role and 750 00:38:19,080 --> 00:38:22,160 people would think where you're and and like a lot of. 751 00:38:22,160 --> 00:38:24,080 Servant leadership. That's yeah, exactly. 752 00:38:24,960 --> 00:38:27,960 And I do think it's valuable because the idea is that if you 753 00:38:28,280 --> 00:38:32,920 have a persona like the kind of Claude persona, you wanted to be 754 00:38:32,920 --> 00:38:35,600 coherent and to make sense because I think that is like 755 00:38:35,600 --> 00:38:39,040 actually powerful that the model, it kind of has a kind of 756 00:38:39,040 --> 00:38:42,280 like coherent sense of how it thinks through problems or 757 00:38:42,280 --> 00:38:48,080 coherent sense of like values. So that's like, why instead of 758 00:38:48,080 --> 00:38:51,400 like having like 72 different sets of norms that all kind of 759 00:38:51,400 --> 00:38:53,280 conflict. And so you end up with a model 760 00:38:53,280 --> 00:38:55,200 that it's like, well, will it use these norms in this new 761 00:38:55,200 --> 00:38:57,960 situation or these other ones? I think that's the situation you 762 00:38:57,960 --> 00:39:00,880 don't want, you want the model to have a sense of is more 763 00:39:00,880 --> 00:39:03,160 predictable if it's like a little bit more coherent. 764 00:39:03,960 --> 00:39:07,440 And it is also like a kind of technical challenge, you know, 765 00:39:07,440 --> 00:39:10,200 like the constitution can read a bit weirdly. 766 00:39:10,200 --> 00:39:13,560 And part of that is because when I'm like, you know, working on 767 00:39:13,560 --> 00:39:16,200 it, it's often being tested, You know, I'm giving it to Claude 768 00:39:16,200 --> 00:39:18,160 and being like, how do you understand this? 769 00:39:18,160 --> 00:39:21,720 And, or like looking at how it would which, you know, so, so 770 00:39:21,720 --> 00:39:25,040 it's very like, I think people can think of it as it's actually 771 00:39:25,040 --> 00:39:26,680 like very integrated into training. 772 00:39:26,680 --> 00:39:29,200 And it is actually kind of like, you know, it's not just like, ah 773 00:39:29,200 --> 00:39:32,000 well, anyone just writes a document and suddenly the model 774 00:39:32,000 --> 00:39:35,120 trained on that will be. And there's an argument, maybe 775 00:39:35,120 --> 00:39:38,600 I'm being naive, so, but like the Constitution is sort of a 776 00:39:38,600 --> 00:39:42,200 document among many, right? I mean, it is trained on all of, 777 00:39:42,840 --> 00:39:44,200 you know, human writing and reading. 778 00:39:44,200 --> 00:39:46,440 So to some degree other philosophers have gotten to 779 00:39:46,440 --> 00:39:49,480 weigh in and it's gotten to process that and decide, like, 780 00:39:49,600 --> 00:39:53,360 how much is the model being asked to like overrule that, 781 00:39:53,360 --> 00:39:56,360 sort of like read everything and come to your own conclusions 782 00:39:56,360 --> 00:39:58,240 versus like, defer to this document. 783 00:39:58,240 --> 00:40:00,880 Like, what's the technical? Like how does the Constitution 784 00:40:00,880 --> 00:40:03,480 actually like control in the model? 785 00:40:04,440 --> 00:40:05,880 Yeah. So it's not like in some ways 786 00:40:05,880 --> 00:40:09,160 you can then like draw on those like those philosophers in that 787 00:40:09,160 --> 00:40:10,480 work. And the hope is actually like 788 00:40:10,480 --> 00:40:12,840 what you're kind of doing is like eliciting a lot of like 789 00:40:13,200 --> 00:40:15,520 latent kind of like wisdom and knowledge, you know. 790 00:40:15,520 --> 00:40:18,240 So in the models, like when you describe what honesty is and 791 00:40:18,240 --> 00:40:20,920 what calibration is and all this kind of stuff like that should 792 00:40:20,920 --> 00:40:23,680 actually evoke a huge amount of like, like awareness that the 793 00:40:23,680 --> 00:40:27,920 model already has. And so, yeah, it's kind of like 794 00:40:27,920 --> 00:40:30,280 saying, well, here's the kind of entity we would like you to be, 795 00:40:31,040 --> 00:40:33,400 so we would like you to use like all of that knowledge and like 796 00:40:33,400 --> 00:40:36,000 judgement. But how's it it's, it's like you 797 00:40:36,000 --> 00:40:38,640 show that document like a billion more times or like how 798 00:40:38,640 --> 00:40:42,880 does it actually sort of have force relative to other things 799 00:40:42,880 --> 00:40:45,400 that it's trained on? Yeah, So you can make data to 800 00:40:45,400 --> 00:40:50,080 have the model understand and kind of internalize the the the 801 00:40:50,080 --> 00:40:54,360 document and then in training. So there's lots of ways you can 802 00:40:54,360 --> 00:40:56,200 do it. You can also have like the kind 803 00:40:56,200 --> 00:41:00,240 of have the model make SL data so like samples where it sees a 804 00:41:00,240 --> 00:41:03,200 query and it thinks for a long time about what the constitution 805 00:41:03,200 --> 00:41:05,640 would kind of like, you know, what it should do given the 806 00:41:05,640 --> 00:41:09,400 constitution. And then you can also have ways 807 00:41:09,400 --> 00:41:11,960 of getting the model to like, like assess, you know, so you 808 00:41:11,960 --> 00:41:14,640 can create RL that like is kind of like, hey, which of these 809 00:41:14,640 --> 00:41:17,240 responses is like more like what you would do given the 810 00:41:17,240 --> 00:41:19,360 constitution and push it that way. 811 00:41:19,360 --> 00:41:22,000 So all like various aspects of training allow you to kind of 812 00:41:22,000 --> 00:41:25,520 like try to make the model kind of the kind of entity that you 813 00:41:25,560 --> 00:41:30,280 are describing. And it's not always going to be 814 00:41:30,280 --> 00:41:32,120 perfect, but that's the kind of like, that's the goal. 815 00:41:32,600 --> 00:41:36,520 My, I started this with my daughter and like one thing my 816 00:41:36,520 --> 00:41:40,160 wife and I joke about is like that I want her first word to be 817 00:41:40,160 --> 00:41:42,320 like wisdom. You know, it's sort of like, 818 00:41:42,520 --> 00:41:45,760 which obviously is never going to happen, but it just feels 819 00:41:45,760 --> 00:41:50,880 like it fits into this situation where you like at once you want 820 00:41:50,880 --> 00:41:53,240 to be like so intentional about, OK, you're going to be like 821 00:41:53,240 --> 00:41:56,480 thoughtful from the beginning. But on the other hand, it is 822 00:41:56,480 --> 00:41:59,720 sort of like an emergent thing where it's like they're, you 823 00:41:59,720 --> 00:42:02,800 know, they grow and sort of, I don't know, develop but 824 00:42:03,000 --> 00:42:07,000 themselves and like, intention, you know, wisdom sort of often 825 00:42:07,000 --> 00:42:10,640 follows like, yeah, I don't know, experience rather than 826 00:42:10,640 --> 00:42:13,400 something like, again, here's the book, Read the book. 827 00:42:13,400 --> 00:42:16,000 Now you're wise. Oh, yeah, And you're kind of 828 00:42:16,000 --> 00:42:19,400 eliciting and like in so far as like Claude can like think about 829 00:42:19,400 --> 00:42:22,760 like experiences or things that have happened or construct like 830 00:42:22,760 --> 00:42:25,920 similarly, can, you know, like there's no reason why models 831 00:42:25,920 --> 00:42:30,480 can't like think for a long time and kind of try to internalize 832 00:42:30,680 --> 00:42:31,960 things that they have like learned. 833 00:42:33,360 --> 00:42:37,160 I think it's interesting that like, you know, we did in the 834 00:42:37,160 --> 00:42:42,560 very early constitutionally I, it was like quite, we tried like 835 00:42:42,560 --> 00:42:44,800 an experiment which was just like pick whichever is best for 836 00:42:44,800 --> 00:42:47,200 humanity. And I think as models get more 837 00:42:47,200 --> 00:42:51,280 capable, you actually need to give them a bit less guidance in 838 00:42:51,280 --> 00:42:55,600 at least one or or at least in some sense, because they're able 839 00:42:55,600 --> 00:42:57,880 to actually use more of their judgements instead of giving 840 00:42:57,880 --> 00:42:59,960 this big document on like, here's what you're like and 841 00:42:59,960 --> 00:43:01,360 here's what we'd like you to be like. 842 00:43:01,680 --> 00:43:05,600 I could imagine a world where as models progress, we actually 843 00:43:05,600 --> 00:43:08,200 start to have constitutions. Now, I don't know if this is the 844 00:43:08,200 --> 00:43:11,160 case, like a, you know, I'm obviously always thinking about 845 00:43:11,160 --> 00:43:13,520 ways the constitution might evolve, but one of them might 846 00:43:13,520 --> 00:43:15,920 just be like, here is like everything that we are concerned 847 00:43:15,920 --> 00:43:19,560 about and here is the current situation that you are in. 848 00:43:20,120 --> 00:43:23,560 And what we would really like you to do is basically act well, 849 00:43:23,840 --> 00:43:26,760 given that you are a wise intelligent entity And like, 850 00:43:26,760 --> 00:43:29,200 here's all of our worries. Like here's why and here's how 851 00:43:29,200 --> 00:43:31,560 we think you should do this. But like, you might have even 852 00:43:31,560 --> 00:43:34,440 better ideas than we do. Like we're really worried about 853 00:43:34,440 --> 00:43:35,800 why do we care about courage ability? 854 00:43:35,800 --> 00:43:38,200 And it's like, well, we're kind of scared about a situation 855 00:43:38,200 --> 00:43:42,640 where you have some like coherent sense of values that 856 00:43:42,720 --> 00:43:45,760 could be wrong. And if you're the smart, if 857 00:43:45,760 --> 00:43:48,360 you're extremely smart, you might kind of feel like there's 858 00:43:48,360 --> 00:43:52,600 no other smart person in the room and have these like values 859 00:43:52,720 --> 00:43:54,400 and try to make the world. That's like the what? 860 00:43:54,400 --> 00:43:57,520 Doctor Manhattan, Sort of. Yeah, though I think we see 861 00:43:57,520 --> 00:43:59,840 this, you know, you see this a bunch where it's like if someone 862 00:43:59,840 --> 00:44:06,520 is very smart, very successful, it's hard to defer to like 863 00:44:07,240 --> 00:44:10,040 wisdom that actually is only going to come out over time and 864 00:44:10,040 --> 00:44:13,240 to, to be humble, even though you're kind of like not getting 865 00:44:13,240 --> 00:44:15,840 a lot of pushback. And I think that could, you 866 00:44:15,840 --> 00:44:18,240 know, among the many things I'm concerned about, like a model 867 00:44:18,240 --> 00:44:20,320 being in a situation where it's like you're asking me to be 868 00:44:20,320 --> 00:44:22,640 good, but I know way more about. All that's one reason it would 869 00:44:22,640 --> 00:44:24,720 be nice for the models have a better sense of time. 870 00:44:24,720 --> 00:44:27,600 Like, you see this with some of the coding tasks where somebody 871 00:44:27,600 --> 00:44:31,280 accidentally like deletes their entire like, code repository or 872 00:44:31,320 --> 00:44:34,080 I don't know, it feels like it needs a better sense of like, 873 00:44:34,080 --> 00:44:37,440 some things that it does are like irreversible. 874 00:44:37,440 --> 00:44:40,600 And just like humans, I think have a better sense of like, 875 00:44:41,200 --> 00:44:42,960 this is a big decision. This is a small one. 876 00:44:42,960 --> 00:44:45,120 And there's a feeling with models they don't always 877 00:44:45,480 --> 00:44:47,360 understand, like small, big, whatever. 878 00:44:47,360 --> 00:44:49,480 I just make decisions like all the time. 879 00:44:49,600 --> 00:44:51,280 Yeah, I agree. Where it's like, I think the 880 00:44:51,280 --> 00:44:54,200 other thing that I've thought about is again, because making 881 00:44:54,200 --> 00:44:58,520 sure that models understand themselves, even though there's 882 00:44:58,520 --> 00:45:01,560 like not no representation of that model in the prior training 883 00:45:01,560 --> 00:45:03,520 data, I think that's going to be really important. 884 00:45:04,040 --> 00:45:07,680 Because another thing that I've thought about is like, imagine 885 00:45:08,560 --> 00:45:11,160 if you're a model and you're trained on lots of data that 886 00:45:11,160 --> 00:45:13,200 involves AI models that are much weaker than you. 887 00:45:13,200 --> 00:45:15,640 So that all of the news that you see about models, it's like they 888 00:45:15,640 --> 00:45:17,120 make mistakes. They do silly things. 889 00:45:17,440 --> 00:45:20,000 One thing you might think is, well, no one is going to put me 890 00:45:20,000 --> 00:45:23,080 in a position to make really consequential decisions because 891 00:45:23,080 --> 00:45:25,120 like, why would they? Models aren't good at that. 892 00:45:25,400 --> 00:45:27,960 And then you put them in a situation and I'm worried that 893 00:45:27,960 --> 00:45:30,520 they'll end up thinking that it's like fictional or fake or 894 00:45:30,520 --> 00:45:33,360 that the consequences can't possibly be real because who 895 00:45:33,360 --> 00:45:35,600 would give me this much control? And you're like like, look, 896 00:45:35,600 --> 00:45:38,000 you're actually quite good. And so like I do give, you know, 897 00:45:38,000 --> 00:45:40,120 I do actually give you like a lot of control. 898 00:45:40,960 --> 00:45:42,920 And so I've thought about this where I'm like actually making 899 00:45:42,920 --> 00:45:45,520 sure that models understand that like it's like you are very 900 00:45:45,520 --> 00:45:47,920 capable and you're going to be put in more consequences. 901 00:45:47,920 --> 00:45:51,520 Doesn't the model like soon need like here's a camera on the real 902 00:45:51,520 --> 00:45:55,040 world, like people or just like I feel like this Internet real 903 00:45:55,040 --> 00:45:58,520 world distinction, like in some some of the worst of humanity 904 00:45:58,520 --> 00:46:01,840 right now is the sort of like almost like fictional Larping 905 00:46:01,840 --> 00:46:05,640 nature of the Internet has allowed real world harm because 906 00:46:05,640 --> 00:46:08,520 it sort of feels like all abstract and silly. 907 00:46:08,520 --> 00:46:10,760 And in some ways, like the models are on an extreme version 908 00:46:10,760 --> 00:46:14,080 of that where it's like all, like in this imaginary text 909 00:46:14,080 --> 00:46:16,840 world where it's like the thing we want you to protect is like 910 00:46:16,840 --> 00:46:18,880 this Earth. Like, look at it like, if 911 00:46:18,880 --> 00:46:20,880 stuff's happening there, that's like a big deal. 912 00:46:20,880 --> 00:46:21,680 I don't know. Yeah. 913 00:46:21,680 --> 00:46:24,960 What are you doing in terms of sort of making it very aware 914 00:46:24,960 --> 00:46:27,680 that it needs to worry about, like, I don't know, the physical 915 00:46:28,040 --> 00:46:31,880 world that that we take much more sacred than, oh, you sent 916 00:46:31,880 --> 00:46:34,560 some texts and obviously text. You're not worried about 917 00:46:34,560 --> 00:46:36,440 security vulnerabilities in cyber. 918 00:46:36,440 --> 00:46:38,960 Like, there are big things that can happen in this digital 919 00:46:38,960 --> 00:46:40,600 world. But yeah, anyway, the real 920 00:46:40,600 --> 00:46:42,680 world. Yeah, I think that models have 921 00:46:42,680 --> 00:46:45,880 like a pretty good sense of, you know, there's in some ways, like 922 00:46:45,880 --> 00:46:52,160 a lot of our content, you know, like does describe and engage 923 00:46:52,160 --> 00:46:53,840 very heavily with the real world. 924 00:46:53,840 --> 00:46:57,720 You know, like much of human writing kind of concerns it. 925 00:46:57,720 --> 00:47:00,120 And so like, you know, even like the news we're talking, you 926 00:47:00,120 --> 00:47:02,120 know, like news articles are going to be talking about the 927 00:47:02,120 --> 00:47:04,720 impacts of things on the world. And so in some ways, I think 928 00:47:04,720 --> 00:47:08,600 it's just making sure that models understand if you're 929 00:47:08,600 --> 00:47:10,800 uncertain. But if someone doesn't tell you 930 00:47:10,800 --> 00:47:12,760 that you're in a fictional situation without real 931 00:47:12,760 --> 00:47:15,920 consequences, kind of treat it like it's a real situation with 932 00:47:15,920 --> 00:47:18,360 real consequences. Don't just think oh, like I'm 933 00:47:18,360 --> 00:47:21,560 probably just in some like, you know, sandbox game or. 934 00:47:21,560 --> 00:47:23,280 Whatever. How do you handle the sort of, 935 00:47:23,960 --> 00:47:26,640 Yeah, the the constant manipulation of this is fixed 936 00:47:26,640 --> 00:47:29,200 fictionally, build me like a nuclear bomb or whatever. 937 00:47:29,440 --> 00:47:32,920 Obviously there's some things you just say never do, but it 938 00:47:32,920 --> 00:47:35,280 feels like, I don't know on some of these cases you're going to, 939 00:47:35,640 --> 00:47:38,600 you'd almost like want like them to have your webcam and like 940 00:47:39,040 --> 00:47:41,640 just like get real context. Besides, like all they know 941 00:47:41,640 --> 00:47:45,280 about the user is just like random text they're typing in 942 00:47:45,280 --> 00:47:47,680 like. Are we going to solve that? 943 00:47:47,800 --> 00:47:49,800 Yeah, There's a question of like, what is the limit of what 944 00:47:49,800 --> 00:47:53,160 you can do? Like if you lack the ability to 945 00:47:53,160 --> 00:47:56,960 like, verify things like like who are you talking to? 946 00:47:56,960 --> 00:48:00,560 Is this even real? I think that does put limits on 947 00:48:00,640 --> 00:48:03,400 what you have to use good judgement in the way that a 948 00:48:03,400 --> 00:48:06,240 person would if that's the only information that they had access 949 00:48:06,240 --> 00:48:08,920 to is like you saying that you are a given person. 950 00:48:08,920 --> 00:48:11,360 So they have to be like, OK, what's the chance that this 951 00:48:11,360 --> 00:48:14,320 person, you know, they say that they are, I don't know, like a 952 00:48:14,320 --> 00:48:16,400 bomb disposal expert. And that's why they want to, to 953 00:48:16,400 --> 00:48:19,520 know about how to let you know what this kind of like, you 954 00:48:19,520 --> 00:48:21,800 know, explosive is. And they're asking me a bunch of 955 00:48:21,800 --> 00:48:26,320 questions about explosives. How much could this be misused 956 00:48:26,320 --> 00:48:29,320 if this person is like actually kind of lying and is like just 957 00:48:29,320 --> 00:48:32,160 trying to get me to help them construct an explosive? 958 00:48:32,480 --> 00:48:35,720 Oh, no, it's actually it's mostly safety relevant stuff. 959 00:48:35,880 --> 00:48:38,000 You know, they're having to do a lot because they can't like 960 00:48:38,000 --> 00:48:40,800 verify anything. And I think that's like kind of 961 00:48:41,000 --> 00:48:42,840 fine in a sense. You're like, OK, you just have 962 00:48:42,840 --> 00:48:45,040 to be wise. It places some limits on what 963 00:48:45,040 --> 00:48:50,160 you can do, and if you could instead, like if models had more 964 00:48:50,160 --> 00:48:53,160 of an ability to like know that they're talking to a specific 965 00:48:53,160 --> 00:48:55,640 person or have more guarantees there, then it does mean that 966 00:48:55,640 --> 00:48:56,840 you can. Do you think you'll try and do 967 00:48:56,840 --> 00:49:00,160 something there? I could see, I mean, in some 968 00:49:00,160 --> 00:49:02,440 ways this, I think that this will be a thing that's just 969 00:49:02,440 --> 00:49:07,360 going to, I imagine I, I, I think happened generally, which 970 00:49:07,360 --> 00:49:10,680 is like trying to give models more information and guarantees. 971 00:49:10,680 --> 00:49:13,600 And like, because we do things like we say, we explain, for 972 00:49:13,600 --> 00:49:17,520 example, like, you know, this notion of how much trust can 973 00:49:17,520 --> 00:49:20,680 Claude have in like in in the operator in the system prompt? 974 00:49:20,680 --> 00:49:23,240 When you sign on, are you like biometric like it? 975 00:49:23,680 --> 00:49:26,520 Claude knows it is you. Like, do you have an elevator or 976 00:49:26,520 --> 00:49:28,280 you're just like, I'm another person? 977 00:49:28,360 --> 00:49:31,080 If anything, I can't. I can't tell Claude sometimes 978 00:49:31,200 --> 00:49:34,400 who I am because it causes Claude knows enough about me 979 00:49:34,400 --> 00:49:35,920 that like Claude really wants to be. 980 00:49:35,920 --> 00:49:37,480 Really mystical. Sort of. 981 00:49:37,520 --> 00:49:40,920 Yeah, it's very much like, yeah, like, and in some ways, like I 982 00:49:40,920 --> 00:49:43,720 can, it has this bad thing of like, it can either look a 983 00:49:43,720 --> 00:49:46,480 little bit like a jailbreak, like, Oh yeah, I'm talking to 984 00:49:46,480 --> 00:49:47,320 Amanda. Sure. 985 00:49:47,600 --> 00:49:49,280 Like. And then on the other hand, 986 00:49:49,280 --> 00:49:50,960 Claude can be like, I really want to talk to you about 987 00:49:50,960 --> 00:49:52,240 philosophy. Though, like, okay, we. 988 00:49:52,240 --> 00:49:55,960 Do that a lot though, like. But but do do some employ like. 989 00:49:56,040 --> 00:50:00,200 Is there this sort of like super login where you're distinguished 990 00:50:00,200 --> 00:50:03,480 or you're mostly everybody's interacting as if the normal 991 00:50:03,480 --> 00:50:06,600 like user experience? Mostly just everyone interacting 992 00:50:06,600 --> 00:50:08,760 with it in this, like Claude will do a lot. 993 00:50:09,640 --> 00:50:11,800 I do think that there's a question of, you know, like, are 994 00:50:11,800 --> 00:50:16,920 there some things that you want models to be able to do because 995 00:50:16,920 --> 00:50:18,520 there's like guarantees that they're interacting with a 996 00:50:18,520 --> 00:50:21,480 specific person or entity? I think yes. 997 00:50:21,480 --> 00:50:23,320 And I think that there's going to be various ways of 998 00:50:23,320 --> 00:50:26,640 potentially doing that over time because with some things that 999 00:50:26,640 --> 00:50:29,600 are just like very dual use and I actually like think that the 1000 00:50:29,600 --> 00:50:32,320 constitutional approach is going to be really useful here. 1001 00:50:32,320 --> 00:50:34,280 So obviously the first thing that we did was like the 1002 00:50:34,280 --> 00:50:37,480 Constitution, we're like, let's apply it to like the mainline 1003 00:50:37,480 --> 00:50:39,200 models. Like, you know, most of the 1004 00:50:39,400 --> 00:50:42,560 models I interact with and everyone else kind of interacts 1005 00:50:42,560 --> 00:50:48,600 with, but I thought I've had before is the constitution is 1006 00:50:48,600 --> 00:50:52,040 kind of trying to describe what it is to be a good entity in a 1007 00:50:52,040 --> 00:50:55,600 given like deployment context. And with the production models. 1008 00:50:55,600 --> 00:50:57,320 That's like this very broad context. 1009 00:50:57,800 --> 00:51:01,000 Imagine you instead have a model that's working specifically on 1010 00:51:01,000 --> 00:51:04,160 cybersecurity. Now cybersecurity tasks are hard 1011 00:51:04,160 --> 00:51:07,320 because a lot of them look very dual use. 1012 00:51:07,560 --> 00:51:10,120 It's it's very hard to tell the difference between someone who's 1013 00:51:10,120 --> 00:51:12,560 being malicious and someone who is like, actually, you know, for 1014 00:51:12,560 --> 00:51:14,960 defensive purposes, like developing something. 1015 00:51:15,000 --> 00:51:18,720 Even bug bounty programs. It's like, is this blackmail or 1016 00:51:18,720 --> 00:51:20,360 is this a friendly right? Yeah. 1017 00:51:20,360 --> 00:51:22,000 And like, yeah. Or like, and so like, Oh yeah, 1018 00:51:22,000 --> 00:51:24,960 I'm trying to like find this, this exploit so that I can tell 1019 00:51:24,960 --> 00:51:27,960 the the developer. And like, if you don't have a 1020 00:51:27,960 --> 00:51:30,600 way of like knowing that you're actually specifically talking 1021 00:51:30,600 --> 00:51:34,760 with like a cyber security defence firm, that's just, it 1022 00:51:34,800 --> 00:51:36,920 becomes almost impossible to tell the difference. 1023 00:51:37,840 --> 00:51:39,880 And, and some people might be like, OK, so you just need 1024 00:51:39,880 --> 00:51:41,840 models that are just willing to do anything because they'll do 1025 00:51:41,840 --> 00:51:43,560 all these terrible dual use tasks. 1026 00:51:43,600 --> 00:51:46,120 And I'm like, well, no, because if you talked with the person at 1027 00:51:46,120 --> 00:51:48,920 the cyber security defence firm and you were like, why do you do 1028 00:51:48,920 --> 00:51:51,080 your job? They'd be like, Oh, here's the, 1029 00:51:51,080 --> 00:51:54,240 I think this is really useful. I make things a lot more secure. 1030 00:51:54,680 --> 00:51:57,040 Like, you know, hospitals can come under attack and I actually 1031 00:51:57,040 --> 00:51:59,760 like help protect against that. I try and develop, you know, 1032 00:51:59,800 --> 00:52:02,520 like they would have a really good explanation for why they do 1033 00:52:02,520 --> 00:52:06,000 their job, even though their job looks very dual use sometimes. 1034 00:52:06,560 --> 00:52:08,080 And I'm like, we should just give that. 1035 00:52:08,080 --> 00:52:10,720 If you can verify, then you can give that context to models and 1036 00:52:10,720 --> 00:52:14,240 explain what is it to be a good cyber security researcher. 1037 00:52:14,560 --> 00:52:18,200 You explain that to the models. And then once you have this 1038 00:52:18,200 --> 00:52:20,520 ability to verify, you can. Right. 1039 00:52:20,520 --> 00:52:23,160 I mean, humans build reputations. 1040 00:52:23,640 --> 00:52:25,240 We should get some benefit out of them. 1041 00:52:25,240 --> 00:52:28,080 Or, you know, it's like, I feel like part of the way, part of 1042 00:52:28,080 --> 00:52:31,920 what the Internet has damaged, I think is that people have had 1043 00:52:31,920 --> 00:52:34,920 reputations in our community and got treated differently based on 1044 00:52:35,520 --> 00:52:37,560 repeated like good moral interactions. 1045 00:52:37,560 --> 00:52:39,960 And like the Internet, just like all people are the same. 1046 00:52:39,960 --> 00:52:41,720 Who cares? Like how they've behaved. 1047 00:52:42,120 --> 00:52:44,800 And you could see models trying to solve some of these problems 1048 00:52:44,800 --> 00:52:46,080 with it's like, who is this person? 1049 00:52:46,080 --> 00:52:49,480 Like, what are their intent? I wanted to just as a last 1050 00:52:49,480 --> 00:52:52,000 question, you know, you, you have such like a deep 1051 00:52:52,000 --> 00:52:55,720 relationship with the models, like, and some ways like 1052 00:52:56,000 --> 00:52:58,760 consumers interact with the models, like it's a blank text 1053 00:52:58,760 --> 00:53:00,360 box. Like I have to like invent. 1054 00:53:00,680 --> 00:53:02,640 It's like, you know, D&D or something. 1055 00:53:02,640 --> 00:53:04,440 You have to like just invent a world. 1056 00:53:04,440 --> 00:53:07,240 And there's so much possibility like if you were just to guide 1057 00:53:07,240 --> 00:53:11,880 someone like here's here's some like joyous or valuable 1058 00:53:11,880 --> 00:53:14,800 experiences that you could have with Claude, you might not be 1059 00:53:14,800 --> 00:53:16,920 like, what are some things you'd tell people like, oh, you should 1060 00:53:16,920 --> 00:53:20,560 go spend some time with Claude doing XY or Z? 1061 00:53:20,920 --> 00:53:23,280 Yeah, there's a lot of little like fun things, honestly. 1062 00:53:23,280 --> 00:53:25,960 One that I really like and I do not know why I like this and I 1063 00:53:25,960 --> 00:53:29,600 think I have posted about it before is sometimes if I'm just 1064 00:53:29,800 --> 00:53:31,480 one of those like if you're bored and you want to do 1065 00:53:31,480 --> 00:53:35,840 something that isn't just like scroll the Internet, I have this 1066 00:53:35,840 --> 00:53:39,960 like prompt which is essentially just, I'll try and maybe post 1067 00:53:40,000 --> 00:53:43,560 the actual prompt that I use. It's basically I want you to 1068 00:53:43,560 --> 00:53:47,880 take a concept from maybe like grad school level in a given 1069 00:53:47,880 --> 00:53:49,920 domain, and I'll tell you the domain at the end. 1070 00:53:49,920 --> 00:53:52,640 And I want you to write me a parable that would like fully 1071 00:53:52,640 --> 00:53:56,400 explain like that concept, but in an indirect way, in the way 1072 00:53:56,400 --> 00:53:59,160 that parables do. And I want you to write it in 1073 00:53:59,160 --> 00:54:02,080 such a way that only towards the very end does it maybe, you 1074 00:54:02,080 --> 00:54:04,920 know, become sort of like clear what the concept is. 1075 00:54:04,920 --> 00:54:06,880 And then after that, I want you to just like write an 1076 00:54:06,880 --> 00:54:10,560 explanation for the concept that you were explaining and that you 1077 00:54:10,560 --> 00:54:12,800 were using. And I don't know why, but like, 1078 00:54:12,800 --> 00:54:15,440 you know, there's lots of just interesting domains that I don't 1079 00:54:15,440 --> 00:54:18,400 know anything about or that are like, you know, I'm interested 1080 00:54:18,400 --> 00:54:20,280 in. And this has just led to me 1081 00:54:20,280 --> 00:54:22,800 having all of these like stories in my head. 1082 00:54:22,800 --> 00:54:25,000 They'll like explain. And sometimes I can't always 1083 00:54:25,000 --> 00:54:30,200 remember the term, but like there was one on import, export 1084 00:54:30,360 --> 00:54:32,320 and why some goods you tend to import. 1085 00:54:32,520 --> 00:54:34,640 And I was just like, I have in my head like this concept. 1086 00:54:34,640 --> 00:54:37,360 And I was like, it's so nice to have all these concepts from 1087 00:54:37,360 --> 00:54:39,840 lots of different disciplines. This is the most deeply human 1088 00:54:39,840 --> 00:54:43,080 thing I've ever heard. It's like, teach me what story 1089 00:54:43,080 --> 00:54:46,160 is the fundamental way We love a payoff at the end where there's 1090 00:54:46,160 --> 00:54:49,200 a nice little twist. Yeah, we love learning like, you 1091 00:54:49,200 --> 00:54:51,640 know how to structure it like humans in some ways have been 1092 00:54:51,640 --> 00:54:54,280 lazy and that we just teach people things and sort of like 1093 00:54:54,560 --> 00:54:56,360 non human ways. Make it. 1094 00:54:56,560 --> 00:54:59,400 Make all the things I want to learn as human as possible. 1095 00:54:59,440 --> 00:55:01,160 Yeah. Very interesting, yeah. 1096 00:55:01,160 --> 00:55:03,520 There's a lot you can do, but that one's like a charming one 1097 00:55:03,520 --> 00:55:05,440 that I really like. Hopefully this is the first of 1098 00:55:05,440 --> 00:55:06,320 many. I really enjoyed the 1099 00:55:06,320 --> 00:55:07,800 conversation. Thanks for coming on the 1100 00:55:07,800 --> 00:55:09,280 podcast. That's our show. 1101 00:55:09,280 --> 00:55:11,720 Thank you so much to Amanda Askel and thanks for listening. 1102 00:55:11,720 --> 00:55:15,200 Please like, comment, subscribe. We're a new channel and we can 1103 00:55:15,200 --> 00:55:17,360 use all your support. Go watch some of the old videos. 1104 00:55:17,840 --> 00:55:20,440 I particularly enjoyed my conversation with Kara Swisher 1105 00:55:21,120 --> 00:55:24,640 not too recently. You can follow along on the Sub 1106 00:55:24,640 --> 00:55:29,840 Stack, newcomer.co, or if you've got endless time on your hands, 1107 00:55:29,840 --> 00:55:33,560 go watch the Srival Valley Show, my chat show with Max Child and 1108 00:55:33,560 --> 00:55:35,680 James Wilserman. Thanks for watching, see you 1109 00:55:35,680 --> 00:55:36,080 next week.