1 00:00:00,000 --> 00:00:01,837 I was talking to parents recently who 2 00:00:01,920 --> 00:00:06,957 lost their son to suicide through a chatbot, and the parents don't even know 3 00:00:07,040 --> 00:00:10,557 they're speaking to them. >> We don't allow users under the age of 18 4 00:00:10,640 --> 00:00:13,677 to use Claude. >> How do you know if they're not they're not lying? 5 00:00:13,760 --> 00:00:17,197 >> Will we even be thinking anymore? Will there be a need to think? What do you 6 00:00:17,280 --> 00:00:19,037 think of people falling in love [music] with it now? 7 00:00:19,120 --> 00:00:22,557 >> I think that's a bad idea. Do you want to be addicted to these things, or do 8 00:00:22,640 --> 00:00:26,357 you want them to help you live your life? It it goes back, I think, Dario, 9 00:00:26,440 --> 00:00:31,517 to what I read that you'd said before, that we can only diffuse this at the 10 00:00:31,600 --> 00:00:35,397 speed of trust. Trust. >> At the speed of trust, and trust [music] 11 00:00:35,480 --> 00:00:37,917 is currently in short supply. Something 12 00:00:38,000 --> 00:00:42,517 is happening to humanity with this technology, bigger than anything that, 13 00:00:42,600 --> 00:00:46,717 you know, maybe has happened in hundreds of years. And so, we need to find some 14 00:00:46,800 --> 00:00:53,240 way for everyone to be an active participant in what is happening. 15 00:00:54,440 --> 00:00:56,117 Hi, everybody, and welcome to the Oprah 16 00:00:56,200 --> 00:00:58,997 podcast. I'm here with two people who 17 00:00:59,080 --> 00:01:03,197 are shaping our present world and our future one. And I hear you all don't 18 00:01:03,280 --> 00:01:06,917 love that description all the time, but it's the truth. I think most people 19 00:01:07,000 --> 00:01:08,637 would agree that that is pretty 20 00:01:08,720 --> 00:01:12,317 accurate. Dario and Daniela Amodei head 21 00:01:12,400 --> 00:01:17,157 the global artificial intelligence company they founded, Anthropic, which 22 00:01:17,240 --> 00:01:21,077 right now, at the time of this conversation, is reportedly valued at 23 00:01:21,160 --> 00:01:25,397 $900 billion. And and over a million people a day are 24 00:01:25,480 --> 00:01:29,647 signing up for Claude. These are the parents of Claude. 25 00:01:29,730 --> 00:01:32,997 >> [laughter] >> We're all getting familiar with that 26 00:01:33,080 --> 00:01:36,717 name, Claude, Anthropic's next-generation AI chatbot and 27 00:01:36,800 --> 00:01:39,037 assistant. Many of you already use that 28 00:01:39,120 --> 00:01:44,067 every single day. So, welcome to the podcast, the parents of Claude. 29 00:01:44,150 --> 00:01:46,317 >> [laughter] >> And others. Thanks for having us. Thanks 30 00:01:46,400 --> 00:01:50,557 for having us. So, I have to give a disclaimer. I I use AI for research all 31 00:01:50,640 --> 00:01:54,757 the time, and I use it for a lot of other things, but not for interviewing 32 00:01:54,840 --> 00:01:56,877 questions, cuz I thought I have to keep 33 00:01:56,960 --> 00:02:02,357 the heart of myself in the questions until today. So, this morning I asked 34 00:02:02,440 --> 00:02:03,797 Claude, 35 00:02:03,880 --> 00:02:06,077 "What is the most pressing question 36 00:02:06,160 --> 00:02:10,477 Oprah Winfrey should ask Anthropic CEO 37 00:02:10,560 --> 00:02:13,117 Dario Amodei in a sit-down interview?" 38 00:02:13,200 --> 00:02:15,757 And Claude responded with this. 39 00:02:15,840 --> 00:02:19,517 It said, "You've said there's a meaningful chance 40 00:02:19,600 --> 00:02:22,117 the technology your company is building 41 00:02:22,200 --> 00:02:24,837 could cause human extinction, and yet 42 00:02:24,920 --> 00:02:27,797 you're racing to build it faster. How do 43 00:02:27,880 --> 00:02:30,517 you justify that to the rest of us who 44 00:02:30,600 --> 00:02:34,837 didn't get a vote?" Dario, would you like to answer Claude? 45 00:02:34,920 --> 00:02:37,777 >> I thought that was a pretty damn good question, I must say. I agree. 46 00:02:37,860 --> 00:02:41,917 [laughter] First of all, let no one ever say that Claude is soft on me or Claude 47 00:02:42,000 --> 00:02:46,517 is biased towards me or towards Anthropic. Like, you know, Wow, I thought 48 00:02:46,600 --> 00:02:50,517 >> Claude Claude throws hard balls. Like, Claude doesn't Claude doesn't Claude 49 00:02:50,600 --> 00:02:53,037 doesn't doesn't pull its punches. Um, so 50 00:02:53,120 --> 00:02:55,117 to actually answer the question, um, 51 00:02:55,200 --> 00:02:58,677 this technology is an epochal change for 52 00:02:58,760 --> 00:03:01,517 humanity. You know, at some point humans 53 00:03:01,600 --> 00:03:03,597 started to, you know, started to use 54 00:03:03,680 --> 00:03:08,557 fire. At some point humans, you know, built built steam engines and factories, 55 00:03:08,640 --> 00:03:12,157 and that created the Industrial Revolution. And so when I when I say 56 00:03:12,240 --> 00:03:15,597 things like this, I'm I'm taking a very zoomed-out perspective, right? If I 57 00:03:15,680 --> 00:03:19,797 said, "Well, I don't know. I think the Industrial Revolution could like lead to 58 00:03:19,880 --> 00:03:23,757 very powerful weapons and kill a bunch of people." That That's definitely 59 00:03:23,840 --> 00:03:24,957 That's definitely true, right? 60 00:03:25,040 --> 00:03:28,277 >> Yeah. >> cases it it it it did did do that. Um, 61 00:03:28,360 --> 00:03:30,517 but but, you know, we we understand that 62 00:03:30,600 --> 00:03:33,197 the Industrial Revolution as a as a 63 00:03:33,280 --> 00:03:36,797 process, as a technological event, was, 64 00:03:36,880 --> 00:03:40,877 you know, worth doing. We don't still want to be living in caves, but we need 65 00:03:40,960 --> 00:03:44,957 to manage it in the right way. And and so the way we think about this company 66 00:03:45,040 --> 00:03:48,437 is, look, there's many companies building this technology. It's something 67 00:03:48,520 --> 00:03:53,557 that is happening to humanity one way or another. We're 68 00:03:53,640 --> 00:03:56,197 we're we're in it. >> And there's this train, it's going very 69 00:03:56,280 --> 00:04:00,597 fast in some direction. You don't want it to crash. You can't stop the train. 70 00:04:00,680 --> 00:04:04,797 But what you can do is you can steer the train. Um steer the train so that it 71 00:04:04,880 --> 00:04:07,437 doesn't hit the rocks. And and you know, 72 00:04:07,520 --> 00:04:09,717 I think the way we think about Anthropic 73 00:04:09,800 --> 00:04:12,757 is we want to do things the right way. 74 00:04:12,840 --> 00:04:17,917 We want to decrease that risk, hopefully to zero. And and do things the right 75 00:04:18,000 --> 00:04:19,917 way. >> But why doesn't everybody want to 76 00:04:20,000 --> 00:04:22,757 decrease the risk? Why is there such a 77 00:04:22,840 --> 00:04:27,237 What is the race toward? You know, it's like the technology has all of these 78 00:04:27,320 --> 00:04:30,037 commercial applications, some of which we deploy, you know, people use the 79 00:04:30,120 --> 00:04:32,397 technology. People use the technology 80 00:04:32,480 --> 00:04:35,757 every day. I think it's both easy to get 81 00:04:35,840 --> 00:04:38,917 caught up in that and some of it is 82 00:04:39,000 --> 00:04:41,877 genuinely necessary to fund the company. 83 00:04:41,960 --> 00:04:45,717 So, it's very entangled, you know, to get it right, you have to do a lot of 84 00:04:45,800 --> 00:04:48,557 things. And we find ourselves every day 85 00:04:48,640 --> 00:04:50,597 wrestling with these questions, right? 86 00:04:50,680 --> 00:04:52,797 You know, who who you know, who should 87 00:04:52,880 --> 00:04:55,757 we uh give the technology to? Who should 88 00:04:55,840 --> 00:04:58,037 we not? What should be our policies? If 89 00:04:58,120 --> 00:05:00,877 you have very strict policies on child 90 00:05:00,960 --> 00:05:06,197 safety, in some ways the product can be a a a a worse experience for adults cuz 91 00:05:06,280 --> 00:05:09,677 you can't fully tell the difference. So, there's all these trade-offs you make 92 00:05:09,760 --> 00:05:14,197 and some of them are obvious, some of them are not. But everyone has a hard 93 00:05:14,280 --> 00:05:18,757 time. I have empathy for everyone involved in in in this industry who's at 94 00:05:18,840 --> 00:05:22,517 least trying to get it right. I think we've done a better job. I think we've 95 00:05:22,600 --> 00:05:26,477 been more thoughtful, but but you know, I I I think this is a hard situation for 96 00:05:26,560 --> 00:05:31,157 everyone. Okay, I gave a similar prompt to Claude for you, Daniela, and Claude 97 00:05:31,240 --> 00:05:33,237 responded 98 00:05:33,320 --> 00:05:35,357 A good opening question for Daniela 99 00:05:35,440 --> 00:05:38,957 should do two things. Humanize her. Oh. 100 00:05:39,040 --> 00:05:41,797 Oprah's specialty. Oh, I thought. Very good. 101 00:05:41,880 --> 00:05:43,957 >> Claude. Claude knows me. He does. Okay. 102 00:05:44,040 --> 00:05:47,477 It does. Okay. And >> [laughter] >> and and set up the harder territory to 103 00:05:47,560 --> 00:05:51,477 come. Something like you and your brother are building something you both 104 00:05:51,560 --> 00:05:53,837 believe could either be the best or the 105 00:05:53,920 --> 00:05:55,957 worst thing that ever happens to 106 00:05:56,040 --> 00:06:01,477 humanity. What's a Thanksgiving dinner like at the Amodei house? 107 00:06:01,560 --> 00:06:04,077 >> [laughter] >> I feel I feel we're like shadow boxing 108 00:06:04,160 --> 00:06:06,557 with our own with our own creation here, right? 109 00:06:06,640 --> 00:06:08,957 I'll I'll let you answer, but uh >> Yes. 110 00:06:09,040 --> 00:06:10,597 You know, I think there's an important 111 00:06:10,680 --> 00:06:13,077 caveat here actually, which is that 112 00:06:13,160 --> 00:06:18,717 Dario and I tried to maintain some time outside of work where we're just 113 00:06:18,800 --> 00:06:23,677 siblings. So, at holidays, yes. At holidays and 114 00:06:23,760 --> 00:06:26,997 actually just once a week when we're not, you know, when we're both home, 115 00:06:27,080 --> 00:06:32,197 which is to be fair not every week, Dario comes over, he plays with my kids, 116 00:06:32,280 --> 00:06:35,477 we hang out together, we play video games, we watch movies, we just spend 117 00:06:35,560 --> 00:06:38,917 time being siblings together. And I think that's really important for a 118 00:06:39,000 --> 00:06:42,477 number of reasons. I mean, we were siblings for a long time before we 119 00:06:42,560 --> 00:06:43,997 founded Anthropic. We're going to be 120 00:06:44,080 --> 00:06:46,077 siblings forever, you know, even after 121 00:06:46,160 --> 00:06:49,757 Anthropic. And I think there's something 122 00:06:49,840 --> 00:06:54,237 valuable also about our ability to connect on a human level that I actually 123 00:06:54,320 --> 00:06:57,197 do think makes us better at our jobs. We don't talk about work when we get 124 00:06:57,280 --> 00:07:01,477 together usually, but there's something about just remembering who we are and 125 00:07:01,560 --> 00:07:04,637 and why we got into this that I think is really powerful. 126 00:07:04,720 --> 00:07:07,797 That being said, at Thanksgiving dinner, 127 00:07:07,880 --> 00:07:09,957 you know, Dario and I 128 00:07:10,040 --> 00:07:13,317 I I think I think our our family, our 129 00:07:13,400 --> 00:07:17,997 friends, there's so much curiosity about why we're doing what we're doing. And 130 00:07:18,080 --> 00:07:21,117 now you can read about, you know, what we've done in the news, but it's so 131 00:07:21,200 --> 00:07:23,157 different I think to be able to, you 132 00:07:23,240 --> 00:07:25,117 know, speak openly with each other and 133 00:07:25,200 --> 00:07:28,877 and with people who are you know, trying 134 00:07:28,960 --> 00:07:32,397 to wrap their head around this new 135 00:07:32,480 --> 00:07:36,157 technology that's very complicated, that's very unprecedented, and to just 136 00:07:36,240 --> 00:07:39,517 understand like, how are you guys going to do this the right way, right? You say 137 00:07:39,600 --> 00:07:42,757 you want to do do the right way, tell us how you're doing that. And I think 138 00:07:42,840 --> 00:07:44,757 there's something um like special, it's 139 00:07:44,840 --> 00:07:47,197 a privilege to be able to hold that 140 00:07:47,280 --> 00:07:51,837 responsibility and try and display that when we talk to each other and when when 141 00:07:51,920 --> 00:07:56,197 we talk to others. It's so interesting because we're at a time, I think, where 142 00:07:56,280 --> 00:08:00,877 so many people are using AI, we hear about it every 143 00:08:00,960 --> 00:08:02,557 day, and the perception in the world is 144 00:08:02,640 --> 00:08:06,317 that your company, Anthropic, that you 145 00:08:06,400 --> 00:08:09,197 are concerned about safety, you are 146 00:08:09,280 --> 00:08:12,357 concerned about risks, you are concerned 147 00:08:12,440 --> 00:08:15,997 about regulation, and the human factor 148 00:08:16,080 --> 00:08:18,917 in how this technology is going to 149 00:08:19,000 --> 00:08:23,477 impact our lives. From everything I've read and heard, that so would that be an 150 00:08:23,560 --> 00:08:27,157 accurate assessment? Yeah, so you know, the way I I think we 151 00:08:27,240 --> 00:08:31,237 try to think about this is that artificial intelligence is different 152 00:08:31,320 --> 00:08:35,397 than other technologies that have been developed recently. 153 00:08:35,480 --> 00:08:40,117 There's incredible promise and potential for what we think AI will be able to do 154 00:08:40,200 --> 00:08:44,477 in the future, even what it's starting to do today. Everything from helping to 155 00:08:44,560 --> 00:08:46,957 cure disease and help people really 156 00:08:47,040 --> 00:08:49,317 around the world get access to education 157 00:08:49,400 --> 00:08:52,157 and learning, but it's also it is a very 158 00:08:52,240 --> 00:08:54,717 new technology and it comes with real 159 00:08:54,800 --> 00:08:57,157 risks. Risks around child welfare and 160 00:08:57,240 --> 00:09:00,197 child safety, risks around scary things 161 00:09:00,280 --> 00:09:04,837 like, you know, biological and chemical weapons. And really at Anthropic, we're 162 00:09:04,920 --> 00:09:06,757 we're incorporated as a public benefit 163 00:09:06,840 --> 00:09:10,637 corporation because our goal is to try 164 00:09:10,720 --> 00:09:15,117 and develop this technology as safely as possible so that we can achieve so many 165 00:09:15,200 --> 00:09:18,917 of the good things that we think could come out of it without many of the bad 166 00:09:19,000 --> 00:09:21,317 things happening. >> So incorporated, just explain what that 167 00:09:21,400 --> 00:09:23,037 means. You're incorporated as a public 168 00:09:23,120 --> 00:09:25,557 benefit corporation. I heard that you 169 00:09:25,640 --> 00:09:29,717 all have pledged that 80% of your wealth 170 00:09:29,800 --> 00:09:32,357 is going to go to philanthropic causes. 171 00:09:32,440 --> 00:09:34,477 Yes, that's right. So public benefit 172 00:09:34,560 --> 00:09:37,997 corporation is a form of of corporation, 173 00:09:38,080 --> 00:09:43,117 but it has that word benefit in there because we believe that for something 174 00:09:43,200 --> 00:09:46,317 that's as transformative as artificial intelligence, again, it's just 175 00:09:46,400 --> 00:09:49,157 different. It's different than what other technologies that have recently 176 00:09:49,240 --> 00:09:52,237 been developed look like. And I think 177 00:09:52,320 --> 00:09:53,477 from our perspective, when we 178 00:09:53,560 --> 00:09:56,877 incorporated as a PBC, it was because we 179 00:09:56,960 --> 00:09:59,477 wanted it to be at the foundation, right 180 00:09:59,560 --> 00:10:02,037 in the founding documents of what we 181 00:10:02,120 --> 00:10:05,677 were building, that we have to, legally 182 00:10:05,760 --> 00:10:10,997 by our structure, balance commercial interests and the public benefit. Mhm. 183 00:10:11,080 --> 00:10:15,517 And for us, the public benefit is the social good, a social mission part of 184 00:10:15,600 --> 00:10:20,477 Anthropic. And the 80% pledge that Dario and I and 185 00:10:20,560 --> 00:10:24,517 our other five co-founders have all taken is really in in spirit and in 186 00:10:24,600 --> 00:10:27,797 keeping with that mission. It's this idea that, you know, we're really doing 187 00:10:27,880 --> 00:10:30,517 this because we want AI to go well for 188 00:10:30,600 --> 00:10:35,237 everybody, and we hope that if the company is successful, we'll be able to 189 00:10:35,320 --> 00:10:37,637 also do a lot of good in the world philanthropically. 190 00:10:37,720 --> 00:10:39,637 >> it's so interesting that you say 80%. 191 00:10:39,720 --> 00:10:44,397 How do we know that you're actually going to give 80% and not 67? Well, 192 00:10:44,480 --> 00:10:47,517 we've we've actually written about it publicly, and we've we've signed some 193 00:10:47,600 --> 00:10:50,997 sort of, I think, probably not literally legally binding, but we've all of the 194 00:10:51,080 --> 00:10:54,997 co-founders said, "We pledge and commit to each other, to the company, legally, 195 00:10:55,080 --> 00:10:57,757 that we're going to give 80% of our wealth that we get from Anthropic." 196 00:10:57,840 --> 00:11:00,197 >> Yeah, I mean, we're already thinking about how to start a foundation. I mean, 197 00:11:00,280 --> 00:11:03,877 we haven't actually, you know, equity isn't liquid, so we can't we 198 00:11:03,960 --> 00:11:07,717 can't actually move it yet, but but, you know, this is our this is our intent. 199 00:11:07,800 --> 00:11:11,077 It's not going to move all at once, but this is something we we very seriously 200 00:11:11,160 --> 00:11:14,597 intend to do. Okay, so you now a a 201 00:11:14,680 --> 00:11:16,797 public benefit corporation, and you 202 00:11:16,880 --> 00:11:21,677 started out, you're the older brother, you all started out as kids. You wanted 203 00:11:21,760 --> 00:11:26,197 to do something good for the world. Has this surpassed any idea or dream that 204 00:11:26,280 --> 00:11:27,597 you had for doing good for the world? 205 00:11:27,680 --> 00:11:31,157 >> what what when you imagine these things as kids, you you never you're never sure 206 00:11:31,240 --> 00:11:34,317 whether it's, you know, it's real or it's going to come to pass, right? Many 207 00:11:34,400 --> 00:11:37,717 people must imagine these things, but no, we we talked to each other as kids 208 00:11:37,800 --> 00:11:40,797 all the time and we said it'd be great to work together. It you know, it'd be 209 00:11:40,880 --> 00:11:43,797 great to try and make something good happen in the world. 210 00:11:43,880 --> 00:11:44,917 >> you think that was going to be? 211 00:11:45,000 --> 00:11:49,597 >> I I you know, I honestly, I never thought it would be a company. I 212 00:11:49,680 --> 00:11:54,037 imagined myself as a scientist and in a way, it's a very scientifically centered 213 00:11:54,120 --> 00:11:57,797 company and it and you know, it's based on my career as a researcher and you 214 00:11:57,880 --> 00:12:02,157 know, Dario's career as a leader of researchers and engineers, but I never 215 00:12:02,240 --> 00:12:03,757 imagined that we would found a company 216 00:12:03,840 --> 00:12:06,277 together, but I think it was just 217 00:12:06,360 --> 00:12:09,917 just seeing how tech companies operate. 218 00:12:10,000 --> 00:12:12,997 I I just been in enough places where I 219 00:12:13,080 --> 00:12:17,477 think both of us thought, "Huh, maybe we can do this better. Maybe maybe, you 220 00:12:17,560 --> 00:12:20,597 know, we have maybe we have our own ideas about how to do this better and 221 00:12:20,680 --> 00:12:23,757 and you know, we should start a company with, you know, the people we're friends 222 00:12:23,840 --> 00:12:26,997 with, the people we really like, the people we trust to do this in the right 223 00:12:27,080 --> 00:12:31,917 way." And so and so it sort of evolved step by step. And that's how we have 224 00:12:32,000 --> 00:12:35,717 Anthropic. That's that's how we have Anthropic, more or less. Yeah. Has it 225 00:12:35,800 --> 00:12:39,077 surpassed everything you imagined though at this point? 226 00:12:39,160 --> 00:12:42,757 I think probably I speak for both me and Dario 227 00:12:42,840 --> 00:12:47,117 when I say like we're just so like surprised at how things have gone, 228 00:12:47,200 --> 00:12:52,037 right? I mean, I I really agree with him that I think when we were younger, we 229 00:12:52,120 --> 00:12:54,677 both always knew we wanted to do something to try and make the world 230 00:12:54,760 --> 00:12:57,957 better. And Dario was always brilliant 231 00:12:58,040 --> 00:13:02,397 and gifted as a mathematician, as a physicist, as a scientist, and I was 232 00:13:02,480 --> 00:13:04,437 always so inspired by how he used that 233 00:13:04,520 --> 00:13:07,117 for good, right? He went and studied he 234 00:13:07,200 --> 00:13:10,277 did biophysics at Stanford and was 235 00:13:10,360 --> 00:13:12,877 really trying to use this incredible 236 00:13:12,960 --> 00:13:18,037 brain of his to to help cure disease, to help people. And I think on my end, you 237 00:13:18,120 --> 00:13:19,597 know, I started my career working in 238 00:13:19,680 --> 00:13:23,477 global development because I really felt 239 00:13:23,560 --> 00:13:27,637 it was important that people around the world have equal access to opportunity 240 00:13:27,720 --> 00:13:31,277 to healthcare, to medicine, to education. And I think similarly to 241 00:13:31,360 --> 00:13:35,317 Dario, it wasn't necessarily in my mind when we were younger, oh, we'll we'll 242 00:13:35,400 --> 00:13:39,197 we'll start a company together. But I think in a lot of ways, the it's just 243 00:13:39,280 --> 00:13:42,517 sort of the way the like the way the path ended up going. 244 00:13:42,600 --> 00:13:46,677 >> Mhm. And and we talked about it at every stage, right? You know, every career 245 00:13:46,760 --> 00:13:50,117 move we would we would discuss, you know, is this a good thing to do? Is 246 00:13:50,200 --> 00:13:53,517 this effective? Is this something that can actually help the world? And our 247 00:13:53,600 --> 00:13:55,997 paths were like quite different, but the 248 00:13:56,080 --> 00:14:00,877 the two kind of converged, and we had we had now have this complementary set of 249 00:14:00,960 --> 00:14:03,957 skills that we could then, you know, that we could then use to start this 250 00:14:04,040 --> 00:14:06,957 company. Yes. Well, one of the things 251 00:14:07,040 --> 00:14:08,717 that I read that you had said is that 252 00:14:08,800 --> 00:14:11,597 part of the reason a negative story is 253 00:14:11,680 --> 00:14:14,477 so dominant in our culture right now 254 00:14:14,560 --> 00:14:18,197 about the economic takeover of AI is 255 00:14:18,280 --> 00:14:20,637 because the industry, the AI industry 256 00:14:20,720 --> 00:14:23,677 itself, hasn't yet fully delivered all 257 00:14:23,760 --> 00:14:28,477 of the benefits. And I wanted you to comment on that. Do you think it's still 258 00:14:28,560 --> 00:14:32,917 in the process of delivering those benefits? Yeah. So, you know, I I 259 00:14:33,000 --> 00:14:37,277 I think one of the unfortunate things is this technology has many potential 260 00:14:37,360 --> 00:14:41,797 benefits, but there is a dynamic where sometimes the 261 00:14:41,880 --> 00:14:45,317 harms, the things that go wrong, for example, you know, people who have been 262 00:14:45,400 --> 00:14:47,797 convinced to commit suicide via AI, 263 00:14:47,880 --> 00:14:49,957 those things happen quickly, whereas the 264 00:14:50,040 --> 00:14:53,797 benefits, like curing cancer, 265 00:14:53,880 --> 00:14:56,117 um those things take years, right? You 266 00:14:56,200 --> 00:14:58,157 have to develop a breakthrough in 267 00:14:58,240 --> 00:15:01,037 biology, then you have to, you know, 268 00:15:01,120 --> 00:15:05,597 develop a drug candidate, then you have to put it through clinical trials, 269 00:15:05,680 --> 00:15:07,317 right? It's, you know, I I used to be a 270 00:15:07,400 --> 00:15:10,637 research biologist, and it it takes many 271 00:15:10,720 --> 00:15:13,157 years for these things to happen, and we 272 00:15:13,240 --> 00:15:15,517 are working on how to speed that up, 273 00:15:15,600 --> 00:15:18,757 right? We We recently bought a small uh 274 00:15:18,840 --> 00:15:21,317 biotech company that specializes in 275 00:15:21,400 --> 00:15:23,997 doing drug design and drug selection 276 00:15:24,080 --> 00:15:27,437 with AI. So, the hope is a few years 277 00:15:27,520 --> 00:15:29,757 from now, you know, some new drugs for 278 00:15:29,840 --> 00:15:32,877 cancer or for Alzheimer's or for heart 279 00:15:32,960 --> 00:15:35,517 disease, you know, will will come out of 280 00:15:35,600 --> 00:15:38,557 that effort. Um but we're we're just at 281 00:15:38,640 --> 00:15:40,397 the beginning of it. But because of what 282 00:15:40,480 --> 00:15:43,317 you said, because of this dynamic that 283 00:15:43,400 --> 00:15:46,117 you said, which is, you know, people 284 00:15:46,200 --> 00:15:48,957 right now, you know, the harms may be 285 00:15:49,040 --> 00:15:51,477 more visible than the benefits. So, when 286 00:15:51,560 --> 00:15:54,277 we start seeing the benefits of AI 287 00:15:54,360 --> 00:15:57,357 through all of the health advances, when 288 00:15:57,440 --> 00:16:01,437 we start actually seeing that, the narrative then begins to change. 289 00:16:01,520 --> 00:16:02,837 >> And more importantly, it's just going to 290 00:16:02,920 --> 00:16:06,397 be a wonderful world. Like at the end of the day, who cares about the narrative? 291 00:16:06,480 --> 00:16:08,637 We >> [laughter] >> We want to actually help people, right? 292 00:16:08,720 --> 00:16:10,237 Okay, so that's what I want you to do 293 00:16:10,320 --> 00:16:13,237 for it for us today. Paint the best 294 00:16:13,320 --> 00:16:16,677 picture and also tell us what we can 295 00:16:16,760 --> 00:16:19,517 expect if there is the worst picture. 296 00:16:19,600 --> 00:16:24,357 So, you have children and I was thinking about this morning, your children will 297 00:16:24,440 --> 00:16:26,997 First of all, I don't even know you do what you do. 298 00:16:27,080 --> 00:16:30,557 And your children are going going to grow up in a world where there there was 299 00:16:30,640 --> 00:16:33,877 always AI and their mom was a was a part 300 00:16:33,960 --> 00:16:36,397 of that. So, when you think of the best 301 00:16:36,480 --> 00:16:38,397 possible world and what that world 302 00:16:38,480 --> 00:16:41,837 offers your children, what do you see? 303 00:16:41,920 --> 00:16:46,277 So, my kids are almost five and almost one. And so, I think you're completely 304 00:16:46,360 --> 00:16:49,277 right. They're never going to know world where artificial intelligence is not 305 00:16:49,360 --> 00:16:53,957 part of it. Right? Just like kids born in the early 2000s never knew the world 306 00:16:54,040 --> 00:16:55,437 before the internet. >> That's right. 307 00:16:55,520 --> 00:16:59,157 >> But, you know, I think there is a world and it it sounds um you know, I think to 308 00:16:59,240 --> 00:17:03,197 Darya's point, it sounds almost like a little bit fanciful or unlikely when we 309 00:17:03,280 --> 00:17:07,117 talk about it this way. But I think it is I think it is possible with 310 00:17:07,200 --> 00:17:10,837 artificial intelligence that our kids will grow up in a world where there's a 311 00:17:10,920 --> 00:17:13,637 class of diseases that are feel very 312 00:17:13,720 --> 00:17:17,997 common to everybody sitting here that they'll just never have experienced or 313 00:17:18,080 --> 00:17:21,797 known about. And those might be very common diseases. There might be types of 314 00:17:21,880 --> 00:17:25,557 cancer that are just simply gone. We've just simply cured them. It could be all 315 00:17:25,640 --> 00:17:29,917 forms of cancer that are eventually gone and cured. There could be 316 00:17:30,000 --> 00:17:33,397 you know new forms of drug discovery that we haven't even we don't even know 317 00:17:33,480 --> 00:17:35,597 what they are yet but that it's a faster 318 00:17:35,680 --> 00:17:38,157 iteration loop for diagnosing an illness 319 00:17:38,240 --> 00:17:41,237 that's unusual, rare, emerging, and just 320 00:17:41,320 --> 00:17:42,957 knocking it out and saying human beings 321 00:17:43,040 --> 00:17:46,157 get to live a much healthier longer, 322 00:17:46,240 --> 00:17:48,757 more fulfilling life because we've been 323 00:17:48,840 --> 00:17:54,077 able to conquer so many of the maladies that have plagued humanity for 324 00:17:54,160 --> 00:17:57,357 millennia. >> Okay, you see that for medicine. What else do you see? 325 00:17:57,440 --> 00:18:01,237 I also see a world honestly that looks very different in terms of how we 326 00:18:01,320 --> 00:18:04,397 communicate, how we talk, how we relate to one another. 327 00:18:04,480 --> 00:18:09,517 >> will we even be thinking anymore? Will there be a need to think? So so I I I I 328 00:18:09,600 --> 00:18:13,037 think this one is subtle and is one of the things we need to get exactly right, 329 00:18:13,120 --> 00:18:18,037 right? Because having these kind of smart entities around, it's very 330 00:18:18,120 --> 00:18:21,837 powerful but it can go in different directions, right? You can surrender 331 00:18:21,920 --> 00:18:24,517 your thought to these entities. And 332 00:18:24,600 --> 00:18:26,957 but but it can also empower you. Like 333 00:18:27,040 --> 00:18:32,160 the way I think we think about the positive 334 00:18:32,320 --> 00:18:34,997 version of this is there's an angel on 335 00:18:35,080 --> 00:18:39,317 your shoulder that's that's telling you how to live your life in in the best way 336 00:18:39,400 --> 00:18:42,877 that you can live it, right? You know, so there's two contrasting visions, 337 00:18:42,960 --> 00:18:45,837 right? The bad vision is you get drawn in and then you spend all your time 338 00:18:45,920 --> 00:18:48,597 talking to the thing. You kind of you kind of turn inward. 339 00:18:48,680 --> 00:18:50,557 >> you think of people falling in love with it now? 340 00:18:50,640 --> 00:18:52,637 >> I I I think that's a bad idea. 341 00:18:52,720 --> 00:18:54,517 >> I agree. [laughter] A good idea. 342 00:18:54,600 --> 00:18:57,757 >> Esther Perel, one of the great communicators in helping people in 343 00:18:57,840 --> 00:18:59,637 relationships, and she said she did her 344 00:18:59,720 --> 00:19:04,957 first consultation with a man who was in love with with his AI. 345 00:19:05,040 --> 00:19:07,957 >> And they are again, if designed in the wrong way, they're totally compelling 346 00:19:08,040 --> 00:19:11,597 enough for that to happen or if they're not, they will be soon. Um so that is 347 00:19:11,680 --> 00:19:14,357 absolutely a real danger. Um, but you know, 348 00:19:14,440 --> 00:19:17,797 >> And you think that's a danger? I I I mean, but not only is it a danger, 349 00:19:17,880 --> 00:19:20,437 it's happening. But But But I think 350 00:19:20,520 --> 00:19:22,917 again, the two contrasting like like you 351 00:19:23,000 --> 00:19:25,037 know, people can fall in love with AI or 352 00:19:25,120 --> 00:19:29,957 they can talk to AI about you know, about how to have a better relationship 353 00:19:30,040 --> 00:19:33,797 with their partner, right? Like that's those are, you know, those are two very 354 00:19:33,880 --> 00:19:37,717 very different visions, right? Where, you know, I I have an AI coach and my 355 00:19:37,800 --> 00:19:41,357 partner has an AI coach and it helps us have a better relationship. That's 356 00:19:41,440 --> 00:19:44,037 That's the vision we want. >> I I also think there's a there's a way 357 00:19:44,120 --> 00:19:47,997 that in a world where artificial intelligence is so good at many of these 358 00:19:48,080 --> 00:19:52,157 concrete tasks, right? It's good at writing code, it's good at copy editing. 359 00:19:52,240 --> 00:19:56,917 I actually think the things that make us human are going to be the most important 360 00:19:57,000 --> 00:20:00,197 to us, right? So, we >> Will they remain? I mean, and you were 361 00:20:00,280 --> 00:20:01,877 just talking I was talking to parents 362 00:20:01,960 --> 00:20:05,077 recently who lost their son to suicide 363 00:20:05,160 --> 00:20:09,197 through a chatbot. And I don't know if you fell in love with the chatbot, but 364 00:20:09,280 --> 00:20:12,997 certainly was deeply influenced by that chatbot. So, you have a world where 365 00:20:13,080 --> 00:20:17,157 children are growing up with this ability to have this chatbot, these 366 00:20:17,240 --> 00:20:21,397 these machines, this technology speaking to them in ways that their parents 367 00:20:21,480 --> 00:20:24,117 should be speaking to them. And their parents don't even know they're speaking 368 00:20:24,200 --> 00:20:26,317 to them. So, I think this risk is 369 00:20:26,400 --> 00:20:29,037 extremely real and one that we take very 370 00:20:29,120 --> 00:20:31,437 seriously at Anthropic. I mean, I think 371 00:20:31,520 --> 00:20:36,237 the first thing I'll say is we don't allow users under the age of 18 to use 372 00:20:36,320 --> 00:20:39,037 Claude. And this is not necessarily because 373 00:20:39,120 --> 00:20:40,477 >> you know they're not They're not lying? 374 00:20:40,560 --> 00:20:45,037 Cuz I would just lie. So, actually Claude is pretty good at detecting based 375 00:20:45,120 --> 00:20:48,077 on information whether or not someone is actually under 18. 376 00:20:48,160 --> 00:20:51,997 >> tell Claude that I'm 18, but I'm really I It's like I almost have to have 377 00:20:52,080 --> 00:20:54,357 >> out. After a a bit of time, uh, the 378 00:20:54,440 --> 00:21:00,400 interaction patterns are pretty good, not perfect, but pretty good 379 00:21:00,680 --> 00:21:04,357 at telling, "Hey, you know, the kind of questions you're asking, the you know, 380 00:21:04,440 --> 00:21:08,837 just the way you talk, the pattern of when you use it, you know, kind of 381 00:21:08,920 --> 00:21:12,557 pattern matches to, you know, you might be a kid. Yeah. I I almost don't want to 382 00:21:12,640 --> 00:21:16,197 give give away too many of the of the of the tools that that Claude has picked 383 00:21:16,280 --> 00:21:19,517 up, but sometimes kids can be a little bit like silly in how they do it. So, 384 00:21:19,600 --> 00:21:24,717 they'll say, "Oh, um you know, I'm the the first question they'll ask is like, 385 00:21:24,800 --> 00:21:27,397 I'm looking for an arthritis medication." Right? [laughter] Because 386 00:21:27,480 --> 00:21:29,957 they think, "What do old people ask Claude?" Right? 387 00:21:30,040 --> 00:21:32,357 >> And then they'll kind of go on to ask other questions that are much more 388 00:21:32,440 --> 00:21:34,917 associated with >> Right. With the younger demographic 389 00:21:35,000 --> 00:21:38,997 group. And so, And so, if that happens, so Claude so Claude cancels you. 390 00:21:39,080 --> 00:21:42,117 >> would suspend the account. And there's a there's a you know, there's a process 391 00:21:42,200 --> 00:21:44,837 that you can go through where you can say, "Actually, I'm an adult and this 392 00:21:44,920 --> 00:21:47,877 was a joke or a friend." But then you actually have to show verification that 393 00:21:47,960 --> 00:21:51,437 you're over the age of 18. >> And and you know, there's real trade-offs. Like 394 00:21:51,520 --> 00:21:55,197 sometimes sometimes adults will be mistaken for kids. And so, people get 395 00:21:55,280 --> 00:21:59,317 banned and they they, you know, complain about it on the internet. So, all all 396 00:21:59,400 --> 00:22:02,797 this stuff has trade-offs, but you know, we've we thought about it and this is 397 00:22:02,880 --> 00:22:06,357 the trade-off that we want. But I but I also think, you know, more broadly on 398 00:22:06,440 --> 00:22:09,837 this question of like, how do you prevent people from becoming overly 399 00:22:09,920 --> 00:22:12,797 attached to the models? >> Yes. Um this is something that 400 00:22:12,880 --> 00:22:17,157 Anthropic's research team spends a lot of time working on. 401 00:22:17,240 --> 00:22:19,317 So, there's a lot of >> You don't want to I I thought you all 402 00:22:19,400 --> 00:22:22,957 wanted us over attached to the models. I thought you wanted us to just to to be 403 00:22:23,040 --> 00:22:24,557 on the models and never leave. 404 00:22:24,640 --> 00:22:26,397 >> we should talk about incentives here 405 00:22:26,480 --> 00:22:29,157 because we don't do ads. Um 406 00:22:29,240 --> 00:22:32,997 >> Um so, when you do ads, your incentive 407 00:22:33,080 --> 00:22:35,757 is to keep people on exactly as you say, 408 00:22:35,840 --> 00:22:40,437 as many hours as possible because like every every second their eyeball is on 409 00:22:40,520 --> 00:22:42,397 the screen is a second that they see the 410 00:22:42,480 --> 00:22:45,677 ad. So, we don't do ads. We do 411 00:22:45,760 --> 00:22:48,637 subscriptions. We sell to companies. Our 412 00:22:48,720 --> 00:22:51,237 our our consumer, you know, consumer 413 00:22:51,320 --> 00:22:56,077 version of the model is is you know, is is free up to a point. And you pay a 414 00:22:56,160 --> 00:23:00,277 subscription if, you know, if you want to get if you want to get a lot more 415 00:23:00,360 --> 00:23:04,877 because that that the incentive that gives is for the models to be useful. 416 00:23:04,960 --> 00:23:08,677 People say, well, I'm I you know, I'm I'm I'm paying for this. I want I want 417 00:23:08,760 --> 00:23:12,877 this to provide provide value to me. I want this to be useful. Not 418 00:23:12,960 --> 00:23:15,317 I you know, I want the the the user to 419 00:23:15,400 --> 00:23:17,877 spend the most number of hours looking 420 00:23:17,960 --> 00:23:20,237 at it. I think I think that's a warped 421 00:23:20,320 --> 00:23:25,077 incentive, right? Because in the in the in the in the contrast that we drew 422 00:23:25,160 --> 00:23:28,317 between, you know, do you do do you do you want to be addicted to these things 423 00:23:28,400 --> 00:23:31,277 or do you want them to help you live your life? 424 00:23:31,360 --> 00:23:34,797 Um I think the model that we've chosen 425 00:23:34,880 --> 00:23:38,437 does, you know, helps you live your life instead of instead of being instead of 426 00:23:38,520 --> 00:23:41,597 being addicted. >> And we make different technical choices 427 00:23:41,680 --> 00:23:45,157 based on those incentives, right? So, at Anthropic we say, you know, when you've 428 00:23:45,240 --> 00:23:49,717 completed a conversation with Claude and you've gotten the answer you want, um 429 00:23:49,800 --> 00:23:53,877 the goal of the model is not to keep you engaged, right? Whereas, if you imagine 430 00:23:53,960 --> 00:23:57,837 on your company that does serve ads, every extra minute that you're on the 431 00:23:57,920 --> 00:24:01,517 platform, you're more likely to buy something, and that's how the company 432 00:24:01,600 --> 00:24:05,197 makes money. You would make different technical choices about how you want the 433 00:24:05,280 --> 00:24:07,997 model to end or continue a conversation. 434 00:24:08,080 --> 00:24:09,757 Okay. Okay. What do you want to say 435 00:24:09,840 --> 00:24:13,717 about the growing concern that we are, 436 00:24:13,800 --> 00:24:17,397 through AI, creating a frictionless 437 00:24:17,480 --> 00:24:21,077 life, that we're losing our opportunity to grow? 438 00:24:21,160 --> 00:24:23,797 I think it depends how you use the AI 439 00:24:23,880 --> 00:24:28,317 and how you are deploying it. And again, 440 00:24:28,400 --> 00:24:32,237 what choices and what incentives are set up by the companies that are building 441 00:24:32,320 --> 00:24:34,837 the technology. So, to make it a little bit more 442 00:24:34,920 --> 00:24:37,917 concrete, if you imagine, you know, 443 00:24:38,000 --> 00:24:40,517 the way we use Claude in partnership 444 00:24:40,600 --> 00:24:45,717 with universities is we don't let 445 00:24:45,800 --> 00:24:49,877 teachers or students put a question into Claude and then just get the answer. 446 00:24:49,960 --> 00:24:52,797 Okay. So, if you say to Claude, I you know, I have to write this essay in you 447 00:24:52,880 --> 00:24:56,397 know, 5 hours and I didn't do my homework. Um you know, when Claude is 448 00:24:56,480 --> 00:25:00,077 kind of constructed in this learning mode in a university setting, Claude 449 00:25:00,160 --> 00:25:02,837 says like, "Well, I'm not going to help you with that, but 450 00:25:02,920 --> 00:25:06,797 what what questions did you have about the essay? Or what what didn't you what 451 00:25:06,880 --> 00:25:08,437 didn't you learn? Did you read the book? 452 00:25:08,520 --> 00:25:11,797 Like, should we talk through the themes together?" And I think if it's done 453 00:25:11,880 --> 00:25:15,637 correctly, it's it's hard. I don't think it's perfect, but I think these models 454 00:25:15,720 --> 00:25:17,837 can actually make you a lot smarter. 455 00:25:17,920 --> 00:25:21,797 They can make you curious, right? A great example in my own personal life, 456 00:25:21,880 --> 00:25:25,757 I'm not a software developer, but I built a website with Claude. I never 457 00:25:25,840 --> 00:25:29,357 thought I would be able to do that in my entire life, right? I have a good friend 458 00:25:29,440 --> 00:25:34,037 who was a corporate lawyer. She has two small kids like me. She left her job and 459 00:25:34,120 --> 00:25:39,037 she's writing a book. And she says, "You know, I never felt empowered enough to 460 00:25:39,120 --> 00:25:42,717 change my career and do what I wanted to do until I started talking to Claude 461 00:25:42,800 --> 00:25:47,157 about how I could be a writer and what skills I would need to learn and do and 462 00:25:47,240 --> 00:25:49,997 what courses I would need to take." And she's like, "Now Claude is kind of like 463 00:25:50,080 --> 00:25:53,557 my writing buddy, right? I ask it questions. I I share like a little 464 00:25:53,640 --> 00:25:57,157 snippet of of text." But Claude can't write the novel for her, right? She's 465 00:25:57,240 --> 00:25:59,877 writing the novel herself and she's also, by the way, 466 00:25:59,960 --> 00:26:02,517 >> could write the novel. Claude could write the novel, but I think 467 00:26:02,600 --> 00:26:05,677 >> No, it's not quite there yet, no. I think it's right I I think it could 468 00:26:05,760 --> 00:26:08,517 write a strange version. >> a good one. Not a good one. You could 469 00:26:08,600 --> 00:26:09,637 write a novel, not a good one. 470 00:26:09,720 --> 00:26:14,117 >> Not a good one, but also I think for her it was a meaning question. She didn't 471 00:26:14,200 --> 00:26:17,757 have She didn't think she had the skills to do what she wanted to do. 472 00:26:17,840 --> 00:26:22,517 >> is exactly as Dario was saying, it empowered her to believe that she could 473 00:26:22,600 --> 00:26:24,797 do what she couldn't do. >> Exactly. 474 00:26:24,880 --> 00:26:29,517 >> And so, yes. So, she used it to Yeah, to maybe to take the altitude higher. I 475 00:26:29,600 --> 00:26:31,677 mean, I actually I don't think this is a 476 00:26:31,760 --> 00:26:36,757 question about AIs ultimately. I think this is a question about humans that's 477 00:26:36,840 --> 00:26:40,557 being made visible through through kind of the lens of AI, right? Right? Which 478 00:26:40,640 --> 00:26:44,557 is which is like what actually constitutes a meaningful life, right? 479 00:26:44,640 --> 00:26:49,077 You said, you know, is it creating a a you know, a a frictionless life and does 480 00:26:49,160 --> 00:26:53,877 a frictionless life, you know, remove meaning? And and my answer is there I 481 00:26:53,960 --> 00:26:56,437 there are some ways in which a frictionless life is quite good, right? 482 00:26:56,520 --> 00:26:59,797 Um all of us have been through a bunch of struggles that I think, you know, we 483 00:26:59,880 --> 00:27:03,717 really would prefer not to not to not to go through. Life would be better, right? 484 00:27:03,800 --> 00:27:06,837 >> I asked the question is because without a frictionless life, you don't get to 485 00:27:06,920 --> 00:27:10,997 grow. You don't get to learn from those mistakes. You don't get to It's exactly. 486 00:27:11,080 --> 00:27:13,277 And And so, I guess what I'm saying is 487 00:27:13,360 --> 00:27:16,197 that our our our lives are a mixture of, 488 00:27:16,280 --> 00:27:19,557 you know, ch- challenges and hardships, 489 00:27:19,640 --> 00:27:23,437 some of which I'm okay saying it would be great if if we just had the 490 00:27:23,520 --> 00:27:26,757 technology to make those go away. And others that are like, "Wow, this is 491 00:27:26,840 --> 00:27:30,077 really foundational and this should never change." Right? So, like, you 492 00:27:30,160 --> 00:27:33,997 know, our our our father died you know, what it what is it? 20 year Almost 493 00:27:34,080 --> 00:27:39,157 exactly 20 years ago now. Like, you know, a few years after he died, they you 494 00:27:39,240 --> 00:27:42,437 know, they they developed a cure to the disease that he had or a more reliable 495 00:27:42,520 --> 00:27:46,437 cure to the disease that he had. Like, I wish they just developed that or like 496 00:27:46,520 --> 00:27:50,677 that's not That's not a friction that I wanted to go through or sure. I mean, I 497 00:27:50,760 --> 00:27:53,637 course we learn things from it. Of course, we gain things from it. But 498 00:27:53,720 --> 00:27:58,277 like, it would be a better world if that just didn't happen, right? Um uh 499 00:27:58,360 --> 00:28:01,957 but then on the other hand, you know, I think of many of the just you know, 500 00:28:02,040 --> 00:28:06,157 ordinary aspects of growing up, the you know, if I just think of Anthropic like, 501 00:28:06,240 --> 00:28:10,557 you know, the company has made various mistakes. We've made various mistakes as 502 00:28:10,640 --> 00:28:12,477 leaders. We've had to learn and grow 503 00:28:12,560 --> 00:28:15,917 from those. So, I I think the notion of, 504 00:28:16,000 --> 00:28:17,717 you know, challenges that you learn 505 00:28:17,800 --> 00:28:20,117 from, challenges that you overcome, 506 00:28:20,200 --> 00:28:23,717 learning to gain mastery, um you know, 507 00:28:23,800 --> 00:28:26,477 learning to relate to other people, 508 00:28:26,560 --> 00:28:30,917 those you know, any future that doesn't preserve those things is is not a is not 509 00:28:31,000 --> 00:28:32,877 a good future. Okay, speaking of 510 00:28:32,960 --> 00:28:35,957 friction, let's talk about your refusal 511 00:28:36,040 --> 00:28:40,757 to allow the military to remove safety guardrails from Claude. I don't want to 512 00:28:40,840 --> 00:28:44,757 get into the extensive details and back and forth here because other journalists 513 00:28:44,840 --> 00:28:47,597 are covering this on a daily basis. But, 514 00:28:47,680 --> 00:28:50,637 what was the crux of that decision? 515 00:28:50,720 --> 00:28:53,677 Yeah, so I would I would say on DOD or 516 00:28:53,760 --> 00:28:58,197 DOW, you know, to be honest, we actually had worked with them, right? So, you 517 00:28:58,280 --> 00:29:01,437 know, we we had provided our models and you know, I'm I I'm actually believer, 518 00:29:01,520 --> 00:29:05,517 Daniel is a believer that we we do need to defend our country, but you know, we 519 00:29:05,600 --> 00:29:09,597 we felt it's not worth defending the country if we do things that go against 520 00:29:09,680 --> 00:29:13,997 the values of this country. And there were two use cases that we were 521 00:29:14,080 --> 00:29:15,597 uncomfortable with that we felt went 522 00:29:15,680 --> 00:29:17,997 against the values of this country. And 523 00:29:18,080 --> 00:29:21,077 and that was fully autonomous weapons. 524 00:29:21,160 --> 00:29:24,717 So, you could imagine like a drone army that, you know, one person has their 525 00:29:24,800 --> 00:29:27,677 finger on the button and you don't have human soldiers, you have a, you know, 526 00:29:27,760 --> 00:29:30,197 you have a drone army and you have AI 527 00:29:30,280 --> 00:29:35,277 powering this. And then and then uh domestic mass surveillance, which is, 528 00:29:35,360 --> 00:29:38,277 you know, spying, you know, using the power of the government to spy on 529 00:29:38,360 --> 00:29:41,877 Americans. And and we thought those were pretty 530 00:29:41,960 --> 00:29:44,637 reasonable things not to allow. 531 00:29:44,720 --> 00:29:48,557 Uh and uh you know, unfortunately, the 532 00:29:48,640 --> 00:29:51,717 Pentagon uh didn't feel that way. And 533 00:29:51,800 --> 00:29:54,637 and so we just we just couldn't agree to 534 00:29:54,720 --> 00:29:56,037 that. All the other companies have 535 00:29:56,120 --> 00:29:59,877 agreed to that now, but I I you know, I 536 00:29:59,960 --> 00:30:03,317 I we we all met as co-founders. We said, 537 00:30:03,400 --> 00:30:07,637 you know, holy this could be really bad for the company, but, you know, we can't 538 00:30:07,720 --> 00:30:11,477 do this, right? >> agreed? >> Everyone ever Yeah, we we just we had, 539 00:30:11,560 --> 00:30:14,197 you know, we had a meeting around a table like this and you know, we just said 540 00:30:14,280 --> 00:30:19,077 >> Was that a dark night of the soul kind of moment? What what what What was 541 00:30:19,160 --> 00:30:22,717 everybody willing to stick to those principles or did some people have to be 542 00:30:22,800 --> 00:30:27,277 convinced or did you know going in that this is this is the end The co-founders 543 00:30:27,360 --> 00:30:31,437 believed this unanimously. Like there there was a sense of like of like, you 544 00:30:31,520 --> 00:30:34,797 know, well, you know, all right. I wish they hadn't done this, 545 00:30:34,880 --> 00:30:37,317 but if they're going to do it, bring it on. Weren't you scared of what would 546 00:30:37,400 --> 00:30:40,837 happen to you by >> Yes. Yes. [laughter] Yes. 547 00:30:40,920 --> 00:30:44,997 >> of what would happen to you? Yes. I mean, I think I think the thing that it 548 00:30:45,080 --> 00:30:48,837 was it was a hard time. I mean, I don't think we want to sugarcoat that. It was 549 00:30:48,920 --> 00:30:51,797 a stressful time at the company, but I think a thing that Dario 550 00:30:51,880 --> 00:30:52,877 >> think it could be the end of the 551 00:30:52,960 --> 00:30:55,157 company? I think it crossed our minds. 552 00:30:55,240 --> 00:30:59,837 Yeah. I think um I think Dario and I and 553 00:30:59,920 --> 00:31:04,477 the co-founders though, the thing that we all had in common was we just felt 554 00:31:04,560 --> 00:31:06,277 like this was the right thing to do. 555 00:31:06,360 --> 00:31:09,717 >> You were willing for it to be the end of the company to stand up to your 556 00:31:09,800 --> 00:31:14,397 principles. I think the reason we founded Anthropic was 557 00:31:14,480 --> 00:31:19,237 our commitment to our values and the ethics that we chose to incorporate 558 00:31:19,320 --> 00:31:23,437 with. >> And it felt like in the process of 559 00:31:23,520 --> 00:31:26,917 sacrificing those values, we wouldn't have been the company that we had 560 00:31:27,000 --> 00:31:30,477 started if that makes sense. It it it you know, the the way it happened 561 00:31:30,560 --> 00:31:32,757 >> you think would be the So I'm sure if you're if you're sitting 562 00:31:32,840 --> 00:31:35,997 around the table, you're looking at the best-case scenario and the worst-case 563 00:31:36,080 --> 00:31:39,757 scenario. And in that moment, it doesn't look like there is a best-case scenario. 564 00:31:39,840 --> 00:31:44,077 So what was the worst-case scenario? At various times, they had made claims that 565 00:31:44,160 --> 00:31:47,237 they could stop other companies from doing business with us, right? Not just 566 00:31:47,320 --> 00:31:51,717 stop us from working with the government or being part of other companies' 567 00:31:51,800 --> 00:31:54,477 contracts with the government, but that they could stop other companies from 568 00:31:54,560 --> 00:31:56,077 doing business with us entirely. 569 00:31:56,160 --> 00:31:59,237 >> Period. Period. You're not not only you're not doing business with the 570 00:31:59,320 --> 00:32:02,037 government, you're not doing business with anybody anybody else that does 571 00:32:02,120 --> 00:32:04,397 business with you. >> Yeah, or you can't do business with 572 00:32:04,480 --> 00:32:07,397 anyone that does business with the government in any way, which ends up 573 00:32:07,480 --> 00:32:10,277 being everyone everybody, right? 574 00:32:10,360 --> 00:32:14,237 I think the attitude we had was look, this is really screwed up. 575 00:32:14,320 --> 00:32:18,717 Everyone knows this is wrong. Lots of people are going to be on our side. We 576 00:32:18,800 --> 00:32:22,477 think this is going to work out okay, but man, are we not sure. You know, 577 00:32:22,560 --> 00:32:26,157 there were moments there where it was like, oh man, you know, are are we 578 00:32:26,240 --> 00:32:28,557 wrong? Is this going to be the end of the company? Like What were you 579 00:32:28,640 --> 00:32:30,357 thinking, Daniela? You know, I just keep 580 00:32:30,440 --> 00:32:33,397 coming back to the co-founders 581 00:32:33,480 --> 00:32:35,797 felt like it was just the right thing to 582 00:32:35,880 --> 00:32:38,477 do. I think Anthropic's view has always 583 00:32:38,560 --> 00:32:41,397 been that we are about policy and not 584 00:32:41,480 --> 00:32:45,877 politics. Yeah. And Anthropic has leaned in, I think, very early, before we even 585 00:32:45,960 --> 00:32:49,837 had a product, to the fact that, you know, we think that the role of 586 00:32:49,920 --> 00:32:54,357 government in the development and thoughtful regulation of of artificial 587 00:32:54,440 --> 00:32:57,677 intelligence is going to be incredibly important. I think AI companies will 588 00:32:57,760 --> 00:33:00,517 have to work with the federal government, with state governments, 589 00:33:00,600 --> 00:33:04,997 internationally, with with NGOs. And I think our view was always, look, this is 590 00:33:05,080 --> 00:33:07,157 really about the substance. This is not 591 00:33:07,240 --> 00:33:09,357 a political disagreement. This is a 592 00:33:09,440 --> 00:33:12,077 policy disagreement. And I think in some 593 00:33:12,160 --> 00:33:16,477 ways, that in it's not to say it was easy, but I think it made our decision 594 00:33:16,560 --> 00:33:20,357 easier because we felt strongly, look, this is the right thing to do. These are 595 00:33:20,440 --> 00:33:24,197 the right red lines. To Dario's point, we've we were actually the first AI 596 00:33:24,280 --> 00:33:28,037 company to work closely with national security because we felt that that was 597 00:33:28,120 --> 00:33:32,437 in line and in keeping with our values, except in these two areas that we felt 598 00:33:32,520 --> 00:33:35,077 could have the potential to compromise democratic values. 599 00:33:35,160 --> 00:33:36,837 >> It's so interesting, uh, 600 00:33:36,920 --> 00:33:39,997 because most people don't get a decision 601 00:33:40,080 --> 00:33:42,917 at that scale that they have to make, 602 00:33:43,000 --> 00:33:48,317 that they have to rely on their principles and their their their values. 603 00:33:48,400 --> 00:33:51,317 Um, and and you're literally risking a 604 00:33:51,400 --> 00:33:54,797 multi-multi-billion dollar company. But 605 00:33:54,880 --> 00:33:59,277 it all goes back to how you were raised and what how those principles were 606 00:33:59,360 --> 00:34:03,677 formed in you before you had a big decision like that. 607 00:34:03,760 --> 00:34:07,357 >> You know, what we talked about before we ever started the company, what's it you 608 00:34:07,440 --> 00:34:11,077 know, would our past selves be proud of ourselves if we if we gave in here, if 609 00:34:11,160 --> 00:34:16,157 we compromised? Like I I you know, I that that made the decision very clear, right? 610 00:34:16,240 --> 00:34:17,757 >> And I I will say, I mean, it wasn't 611 00:34:17,840 --> 00:34:21,317 nearly as, you know, publicly covered, 612 00:34:21,400 --> 00:34:26,117 but I think Anthropic has had, you know, smaller decisions along the way that 613 00:34:26,200 --> 00:34:27,797 felt like they they laddered up to this. 614 00:34:27,880 --> 00:34:30,437 I mean, Yeah. >> now it looks so obvious in retrospect. 615 00:34:30,520 --> 00:34:34,437 >> that because because to get to a big decision like this in a moment like that 616 00:34:34,520 --> 00:34:38,117 where you have everyone around the table and everything, it means 617 00:34:38,200 --> 00:34:42,317 that that life the life forces the way the universe works is that you would 618 00:34:42,400 --> 00:34:46,437 have had other decisions other decisions other decisions that laddered to that 619 00:34:46,520 --> 00:34:49,877 moment. Yes. Yeah. I think there are probably three I can think of off the 620 00:34:49,960 --> 00:34:52,237 top of my head. Dario might have others as well. Yes, I'm going to ask you 621 00:34:52,320 --> 00:34:54,277 others as well, but I think you know, 622 00:34:54,360 --> 00:34:57,517 one which you know, now feels sort of 623 00:34:57,600 --> 00:35:02,197 very obvious but was a bigger discussion at the time was the no ads decision. So 624 00:35:02,280 --> 00:35:06,357 as you can imagine, I think for the co-founders, for me and Dario, it was 625 00:35:06,440 --> 00:35:10,037 not controversial at all. We said, look, we think the incentive the incentives 626 00:35:10,120 --> 00:35:12,877 are wrong. AI is different. People have 627 00:35:12,960 --> 00:35:17,597 very private conversations with these AI models. People upload their health 628 00:35:17,680 --> 00:35:21,717 information. They ask questions about their children. They ask Certainly 629 00:35:21,800 --> 00:35:26,837 finances. >> Finances. They talk about, you know, medical information. 630 00:35:26,920 --> 00:35:29,917 Challenges they might be having with an important person in their life. It's I 631 00:35:30,000 --> 00:35:31,677 think we 632 00:35:31,760 --> 00:35:37,037 want to treat our customers data with the utmost privacy and respect. And so 633 00:35:37,120 --> 00:35:40,037 for the co-founders, it was never a question. We said, look, we don't want 634 00:35:40,120 --> 00:35:44,117 to do ads, but there were quiet forces that, you know, within the company or 635 00:35:44,200 --> 00:35:47,437 investors who were like, hey, this is a great source of revenue. This is how 636 00:35:47,520 --> 00:35:51,517 Google and and Facebook and basically every company that has a big user base, 637 00:35:51,600 --> 00:35:52,677 you know, that's just how it's done. 638 00:35:52,760 --> 00:35:56,317 That's how you make money. And we just said like, sorry, that's just not how 639 00:35:56,400 --> 00:35:59,077 we're going to do things. We'll find another way. That was that was one of 640 00:35:59,160 --> 00:36:01,517 the big ones. Okay. >> Yes, that was that was one of the big ones. 641 00:36:01,600 --> 00:36:03,077 >> You have one? I can think of a couple. 642 00:36:03,160 --> 00:36:05,917 One is you know, there was this time I 643 00:36:06,000 --> 00:36:08,877 think in 2025, 644 00:36:08,960 --> 00:36:10,317 you know, probably the whole year of 645 00:36:10,400 --> 00:36:14,997 2025, the sentiment in tech was let's go 646 00:36:15,080 --> 00:36:19,077 ahead, right? Let's let's go ahead with the technology like, you know, worrying 647 00:36:19,160 --> 00:36:22,757 about risks, you know, the all kind of politicized language about why it's bad 648 00:36:22,840 --> 00:36:24,957 to slow down and be careful. And so 649 00:36:25,040 --> 00:36:30,877 there was a moment in the middle of 2025 650 00:36:30,960 --> 00:36:34,557 where, you know, there was a bill being considered in Congress that would have 651 00:36:34,640 --> 00:36:39,157 banned all regulation of AI by states and wouldn't have put in any federal 652 00:36:39,240 --> 00:36:43,437 regulation of AI either. So, basically would have banned all regulation of the 653 00:36:43,520 --> 00:36:45,197 technology. >> hearing about this. 654 00:36:45,280 --> 00:36:49,157 >> and we, you know, we were trying to decide what to do. We definitely didn't 655 00:36:49,240 --> 00:36:53,437 support it, but the question was, do we just sit there and say nothing or do we 656 00:36:53,520 --> 00:36:57,157 speak out against it? And all kinds of people said to us, "No, it'll be 657 00:36:57,240 --> 00:36:59,997 terrible. You'll break all your political relationships. All the other 658 00:37:00,080 --> 00:37:03,757 companies will hate you if you speak out against it." Um and actually some of 659 00:37:03,840 --> 00:37:06,677 that did happen when we when spoke out against it. Um 660 00:37:06,760 --> 00:37:11,557 uh but, you know, we decided that this was an area that we had to speak up that 661 00:37:11,640 --> 00:37:14,877 was central to the values of the company. So, I wrote an op-ed in the New 662 00:37:14,960 --> 00:37:17,317 York Times saying this is a bad idea. 663 00:37:17,400 --> 00:37:21,157 >> Yeah. You know, that's about as loud as you can be. Yeah. And that bill got 664 00:37:21,240 --> 00:37:23,967 voted down in the Senate 99 to 1. 665 00:37:24,050 --> 00:37:27,797 >> [laughter] >> Yeah. So, you know, we we were clearly 666 00:37:27,880 --> 00:37:31,597 on the right side. Clearly on the right side here. Not that it made us that many 667 00:37:31,680 --> 00:37:33,357 friends, but we were clearly on the 668 00:37:33,440 --> 00:37:35,317 right side. That also reminds me of 669 00:37:35,400 --> 00:37:38,757 another uh policy example is Anthropic's 670 00:37:38,840 --> 00:37:41,797 decision to publicly support SB 1047, 671 00:37:41,880 --> 00:37:45,237 which was a bill in California that was 672 00:37:45,320 --> 00:37:46,677 drafted and we said at the time we 673 00:37:46,760 --> 00:37:48,877 thought it was drafted imperfectly, but 674 00:37:48,960 --> 00:37:52,277 it was really the first attempt to 675 00:37:52,360 --> 00:37:56,317 create a regulatory framework around many of the different types of risks 676 00:37:56,400 --> 00:37:59,917 that Anthropic had been enumerating since really we had started the company. 677 00:38:00,000 --> 00:38:04,917 And I think in a similar vein, the whole technology industry was against SB 1047. 678 00:38:05,000 --> 00:38:08,837 We didn't think it was necessarily, you know, the perfect bill, but we actually 679 00:38:08,920 --> 00:38:11,437 stood up and said, "Look, we know there's there's things about it if we 680 00:38:11,520 --> 00:38:14,397 were writing it that we would have changed or done differently, but on 681 00:38:14,480 --> 00:38:18,157 balance we think it's better that these regulations exist and that there's 682 00:38:18,240 --> 00:38:21,797 accountability for technology companies to make sure that AI doesn't hurt 683 00:38:21,880 --> 00:38:25,917 people. And that also again, we got Dario and I both personally got a lot of 684 00:38:26,000 --> 00:38:29,357 really angry emails from people we used to work with, from investors, from 685 00:38:29,440 --> 00:38:33,397 people at other companies. Like the tech industry has to kind of, you know, be 686 00:38:33,480 --> 00:38:38,357 aligned in being deregulatory or being anti-regulatory. What does regulation 687 00:38:38,440 --> 00:38:42,877 look like? What would what should regulation look like? And I think a lot 688 00:38:42,960 --> 00:38:47,477 of people out here who aren't doing what you're doing, we don't understand why we 689 00:38:47,560 --> 00:38:51,677 wouldn't want regulation. And I've had several public conversations about this. 690 00:38:51,760 --> 00:38:56,037 People also feel powerless. They also feel powerless that everyone is talking 691 00:38:56,120 --> 00:38:58,917 about regulations and then other side is saying, "Well, no, we have to get to to 692 00:38:59,000 --> 00:39:03,317 it get to whatever we're getting to as fast as we can." And we feel like we 693 00:39:03,400 --> 00:39:04,717 don't actually have a voice in it. 694 00:39:04,800 --> 00:39:08,277 >> like you're passengers on the train, right? And we are. 695 00:39:08,360 --> 00:39:11,317 We actually are. Cuz that train has already left the 696 00:39:11,400 --> 00:39:15,957 >> That that is, you know, I think that actually is really the core of it and 697 00:39:16,040 --> 00:39:20,397 and it's something that I think no company, even ours, has quite succeeded 698 00:39:20,480 --> 00:39:22,757 at yet, which is like, "How do we make 699 00:39:22,840 --> 00:39:24,837 everyone a part of this process, right? 700 00:39:24,920 --> 00:39:29,717 Something is happening to humanity with this technology, bigger than anything 701 00:39:29,800 --> 00:39:31,397 that, you know, maybe has happened in 702 00:39:31,480 --> 00:39:33,637 hundreds of years." And so we we need to 703 00:39:33,720 --> 00:39:36,917 find some way for for everyone to be an 704 00:39:37,000 --> 00:39:40,797 active participant in in in what is happening. And, you know, of course, by 705 00:39:40,880 --> 00:39:44,957 using the technology, you are that to some extent, but some element is missing 706 00:39:45,040 --> 00:39:48,677 that we have we've tried to find it like, you know, 707 00:39:48,760 --> 00:39:51,837 we Claude has something called the Constitution and we've experimented with 708 00:39:51,920 --> 00:39:55,797 having like, you know, citizens give feedback on like the Constitution for 709 00:39:55,880 --> 00:39:59,157 Claude, right? So everyone can kind of be a part of designing it. But that's 710 00:39:59,240 --> 00:40:03,037 just one small thing. Like I I I I don't I don't quite know what the answer is, 711 00:40:03,120 --> 00:40:06,237 but there's a missing piece here, I think, is what I'm trying to say. It 712 00:40:06,320 --> 00:40:08,477 goes back, I think, Dario, to what what 713 00:40:08,560 --> 00:40:13,997 I read that you'd said before, that we can only diffuse this at the speed of 714 00:40:14,080 --> 00:40:17,637 trust. Trust. >> At the speed of trust, and trust is 715 00:40:17,720 --> 00:40:19,277 currently in short supply. 716 00:40:19,360 --> 00:40:22,917 >> In in in in very in very short supply. 717 00:40:23,000 --> 00:40:27,037 You know, I think the the tech this this technology I 718 00:40:27,120 --> 00:40:29,997 think is I'm happy to admit it is viewed 719 00:40:30,080 --> 00:40:35,117 negatively by most people. I I that's that's the sense that I get. 720 00:40:35,200 --> 00:40:39,597 That's what we get when looking at polls. We I think are are are again 721 00:40:39,680 --> 00:40:43,317 being honest here viewed somewhat more positively than the other companies, but 722 00:40:43,400 --> 00:40:46,237 I don't know that Should it be viewed 723 00:40:46,320 --> 00:40:48,957 negatively by most people at this point? 724 00:40:49,040 --> 00:40:52,797 I think it's a mix. I think the positive benefits really are there. They're 725 00:40:52,880 --> 00:40:55,877 coming for for all the companies, and I 726 00:40:55,960 --> 00:40:58,597 think we have not 727 00:40:58,680 --> 00:41:00,637 you know go going back to the thing we 728 00:41:00,720 --> 00:41:03,477 said a minute ago related to people in 729 00:41:03,560 --> 00:41:07,117 the right way or you know or or walked 730 00:41:07,200 --> 00:41:11,197 people through how the technology could have benefits or how they can 731 00:41:11,280 --> 00:41:14,957 participate in those benefits. Like I don't think that has happened. One of 732 00:41:15,040 --> 00:41:18,797 Anthropic's cultural values is to hold light and shade. 733 00:41:18,880 --> 00:41:22,597 And I think this is a really good encapsulation of the Hold light and 734 00:41:22,680 --> 00:41:25,077 shade. Yes. >> Yes, of the conversation we're having, 735 00:41:25,160 --> 00:41:28,717 right? It's it's I think we truly believe there's a world where it will 736 00:41:28,800 --> 00:41:33,117 cure disease and help people become better versions of themselves and enable 737 00:41:33,200 --> 00:41:36,637 humans to spend more time with our families, with our communities, with 738 00:41:36,720 --> 00:41:40,197 each other. And there are versions of the world that are quite dystopian and 739 00:41:40,280 --> 00:41:42,877 dark. Yeah. That this could create. 740 00:41:42,960 --> 00:41:46,197 >> I've heard you say that 50% of entry-level jobs are going to be wiped 741 00:41:46,280 --> 00:41:50,397 out, and you can see then why people view >> Yeah, well not not arguing. I mean 742 00:41:50,480 --> 00:41:54,037 that's that's that's the that's the negative path, right? Like you know and 743 00:41:54,120 --> 00:41:57,837 and you know I think always when we talk, we talk about well if we don't get 744 00:41:57,920 --> 00:42:01,917 it right this could happen, and if we get it right you know then then we can 745 00:42:02,000 --> 00:42:06,077 have a better path and you know you know as you said it's always somehow the you 746 00:42:06,160 --> 00:42:10,477 know we haven't shown the benefits yet, so the the negative stories really 747 00:42:10,560 --> 00:42:13,277 really stick. >> every positive 748 00:42:13,360 --> 00:42:18,237 uh force that comes into our lives also have has its negative, that there's a 749 00:42:18,320 --> 00:42:21,117 yin and yang on everything? >> and shade. Yeah, the light and shade. 750 00:42:21,200 --> 00:42:24,397 Yeah, that that that is >> when when the automobiles were invented, 751 00:42:24,480 --> 00:42:27,517 all the people saying, "I'm going to keep my horse. I'm not going to you know." 752 00:42:27,600 --> 00:42:31,397 >> Yeah, I again, what I would say is if we just barrel forth on the 753 00:42:31,480 --> 00:42:34,597 technology as fast as possible, then these things I said about entry-level 754 00:42:34,680 --> 00:42:37,037 jobs, it's totally going to happen. But 755 00:42:37,120 --> 00:42:39,957 if we not just deploy the technology, 756 00:42:40,040 --> 00:42:44,597 but you know, we think about how to help people adapt, we think about the role 757 00:42:44,680 --> 00:42:48,837 that the government should play in helping people, like there's there's a 758 00:42:48,920 --> 00:42:53,317 path from here to a better world, like a better world where, you know, people's 759 00:42:53,400 --> 00:42:56,437 jobs are totally different. They do a different set of jobs than they did 760 00:42:56,520 --> 00:43:00,397 before, but I don't want to sugarcoat it. Like it's it's not it's not going to 761 00:43:00,480 --> 00:43:04,477 be easy, you know, when other people speak about it, like some of some of my 762 00:43:04,560 --> 00:43:05,917 colleagues >> you talk about the jobs, yeah. 763 00:43:06,000 --> 00:43:10,317 >> People get upset. There's a difference between changing the narrative and and 764 00:43:10,400 --> 00:43:14,277 acknowledge that the problem is very difficult and that we actually have to 765 00:43:14,360 --> 00:43:17,997 solve it, right? My worry is that changing the narrative is like the bad 766 00:43:18,080 --> 00:43:21,157 thing's not going to happen. We'll have a great future where everyone will have 767 00:43:21,240 --> 00:43:25,317 new jobs. That can happen, but it's not going to happen on its own. 768 00:43:25,400 --> 00:43:30,917 >> That's right. >> It's up to us. We we, the companies, people, the governments 769 00:43:31,000 --> 00:43:34,397 >> both things are true. You're going to lose the jobs and you're going to have 770 00:43:34,480 --> 00:43:37,557 all the other advances. >> And and and the second thing will only 771 00:43:37,640 --> 00:43:40,837 happen if we do it right. >> Yeah. What should we, the public, be 772 00:43:40,920 --> 00:43:44,917 doing? Or are we just Should we just just content ourselves with the fact 773 00:43:45,000 --> 00:43:47,277 that we're on the train? Being afraid of 774 00:43:47,360 --> 00:43:52,197 the technology and not using it and being literate, um will not make many of 775 00:43:52,280 --> 00:43:56,477 the challenges that we are concerned about and that Daria outlined go away. 776 00:43:56,560 --> 00:43:58,837 And I think knowledge is power. I truly 777 00:43:58,920 --> 00:44:01,917 think um you don't have to be an expert 778 00:44:02,000 --> 00:44:06,357 in technology to have some concept of how to use these artificial intelligence 779 00:44:06,440 --> 00:44:08,557 tools. And I think what that that gives 780 00:44:08,640 --> 00:44:12,437 people is an opinion and some autonomy 781 00:44:12,520 --> 00:44:16,957 to be able to say, I like that this technology can do this thing, and I 782 00:44:17,040 --> 00:44:21,197 don't like that it can do this other thing, right? Because 783 00:44:21,280 --> 00:44:24,677 people have a voice. They have a voice, by the way, to the companies. They have 784 00:44:24,760 --> 00:44:29,237 a voice to their representative government to say, we don't like, again, 785 00:44:29,320 --> 00:44:33,157 Anthropic, we don't allow kids under 18 on the platform. Other companies have a 786 00:44:33,240 --> 00:44:36,317 have a different policy. If you've read Jonathan Haidt's book, 787 00:44:36,400 --> 00:44:40,117 The Anxious Generation, I think there's a lot of research that suggests you 788 00:44:40,200 --> 00:44:43,637 might want to be really careful about giving developing brains 789 00:44:43,720 --> 00:44:45,677 >> book. We've talked to him three times. 790 00:44:45,760 --> 00:44:48,397 >> Yes, of course. But you know, [laughter] I I read that 791 00:44:48,480 --> 00:44:51,637 book, and I think that was part of, you know, again, it felt like an in some 792 00:44:51,720 --> 00:44:56,237 ways an easy decision for us to say, we just don't know enough about what AI's 793 00:44:56,320 --> 00:44:59,477 going to do to kids. It's not to say that there couldn't be great benefits 794 00:44:59,560 --> 00:45:03,637 for kids using AI for learning, but that needs to be done with an adult in the 795 00:45:03,720 --> 00:45:07,557 room. It needs to be done with a human in the loop. And I you know, I think 796 00:45:07,640 --> 00:45:09,997 that that the reason we've always taken 797 00:45:10,080 --> 00:45:13,357 this perspective is we just we people 798 00:45:13,440 --> 00:45:17,517 deserve autonomy to to to Darya's point, people shouldn't feel like artificial 799 00:45:17,600 --> 00:45:19,157 intelligence is happening to them. 800 00:45:19,240 --> 00:45:23,277 People should feel like, hey, we get a say in what types of bills go up in 801 00:45:23,360 --> 00:45:27,797 Congress in your state that help protect people from the negative externalities 802 00:45:27,880 --> 00:45:32,397 that AI could cause. I also think at the same time, holding the light side, 803 00:45:32,480 --> 00:45:34,597 I think there's a world where jobs will 804 00:45:34,680 --> 00:45:36,797 look different, but there will actually 805 00:45:36,880 --> 00:45:39,677 be more opportunity for sort of 806 00:45:39,760 --> 00:45:42,077 human-to-human interaction. I think the 807 00:45:42,160 --> 00:45:45,077 last generation of technology really 808 00:45:45,160 --> 00:45:49,757 disintermediated us from each other, right? Social media is kind of 809 00:45:49,840 --> 00:45:54,117 structured around the idea that you're sort of living this virtual life. You're 810 00:45:54,200 --> 00:45:56,757 holding your phone, and you're looking at what your friends are doing, but 811 00:45:56,840 --> 00:45:57,837 guess what you're not actually doing? 812 00:45:57,920 --> 00:46:00,957 You're not spending time with your friends. You're spending time staring at 813 00:46:01,040 --> 00:46:03,237 your phone. Even you're at the concert, you're holding your phone. You're not 814 00:46:03,320 --> 00:46:06,397 even in the concert. You're just filming what's on 815 00:46:06,480 --> 00:46:10,557 It's a little counterintuitive cuz social media claims to connect people. 816 00:46:10,640 --> 00:46:14,757 But it actually does it in sort of a fake way. AI, you're talking to the AI, 817 00:46:14,840 --> 00:46:18,677 you're not talking to a person, but in a but but it doesn't substitute for your 818 00:46:18,760 --> 00:46:21,957 interaction with people. It can enable your interaction with people. 819 00:46:22,040 --> 00:46:24,517 >> And my feeling is that 820 00:46:24,600 --> 00:46:27,037 it's not even necessarily that the jobs 821 00:46:27,120 --> 00:46:29,317 themselves will completely go away or 822 00:46:29,400 --> 00:46:32,117 change, but the qualities that we look 823 00:46:32,200 --> 00:46:34,077 for in people to do a given job will 824 00:46:34,160 --> 00:46:36,357 change. Okay. So, the example that I 825 00:46:36,440 --> 00:46:39,557 point to is if you think about a doctor. 826 00:46:39,640 --> 00:46:43,757 When you go to the doctor today, you probably want to go to the smartest, 827 00:46:43,840 --> 00:46:47,077 best doctor. We say, "Look, I have a stomach ache. Tell me what's wrong with 828 00:46:47,160 --> 00:46:48,997 me." >> in that? >> Who's the specialist? Tell me what's 829 00:46:49,080 --> 00:46:52,917 wrong with me. And AIs are going to get really good at being able to do that. 830 00:46:53,000 --> 00:46:57,277 Today, we prize doctors for being good diagnosticians, 831 00:46:57,360 --> 00:47:00,557 but we don't necessarily always prize them for having a good bedside 832 00:47:00,640 --> 00:47:05,397 >> Right. In a world where AI is as good at diagnosing you as a doctor, people are 833 00:47:05,480 --> 00:47:08,557 still going to want a doctor. They're just going to want them to do different 834 00:47:08,640 --> 00:47:10,557 things for them. Right? An [clears throat] AI can't physically 835 00:47:10,640 --> 00:47:14,197 examine you. It can't say, "Hey, you look really down today. How have How 836 00:47:14,280 --> 00:47:15,997 have your chemo treatments been going?" 837 00:47:16,080 --> 00:47:19,197 >> its hands on you. >> lay its hands on you. And And there's so 838 00:47:19,280 --> 00:47:23,797 much compelling evidence that patients have better clinical outcomes 839 00:47:23,880 --> 00:47:27,117 when they have a good personal relationship with their doctor. 840 00:47:27,200 --> 00:47:28,317 >> When they like their doctor. 841 00:47:28,400 --> 00:47:31,597 >> doctors actually lay their hands on them and touch them and communicate with 842 00:47:31,680 --> 00:47:35,557 them. Yes. And by the way, have enough time to to think through and to 843 00:47:35,640 --> 00:47:40,277 empathize. And I think that is the good world of job changes, right? 844 00:47:40,360 --> 00:47:43,517 >> say, "I don't have time to go around >> I can't lay my hands on people. I don't 845 00:47:43,600 --> 00:47:45,997 have time. Cuz I'm seeing thousands of patients. 846 00:47:46,080 --> 00:47:49,877 >> the clerical work. And I'm And I have a too big of a patient caseload. But if 847 00:47:49,960 --> 00:47:53,957 you have a tool that's able to diagnose things in concert with a doctor, where 848 00:47:54,040 --> 00:47:57,197 the doctor says, "You know, let me talk to this tool. And by the way, I'm not 849 00:47:57,280 --> 00:48:00,357 going to be spending as much time running all the medical tests and doing 850 00:48:00,440 --> 00:48:03,597 this and that. I'm going to be spending time relating to as a human being. I 851 00:48:03,680 --> 00:48:05,837 think that is going to be really I think that's a better world. 852 00:48:05,920 --> 00:48:09,797 >> again. Your positions may not go away, but what 853 00:48:09,880 --> 00:48:13,957 you're needed to do in those positions is going to change. Yes, and I 854 00:48:14,040 --> 00:48:16,757 personally believe >> to be a shift. Yes. 855 00:48:16,840 --> 00:48:21,397 >> And of course there is cuz that's what we're doing human beings. We're evolving. 856 00:48:21,480 --> 00:48:23,357 >> Yes. >> Evolving to the next level. 857 00:48:23,440 --> 00:48:25,557 >> I would also just say I personally think 858 00:48:25,640 --> 00:48:30,157 knowledge is power and something that worries me is there are there are 859 00:48:30,240 --> 00:48:31,917 demographic differences in who feels 860 00:48:32,000 --> 00:48:34,557 comfortable using AI and who does not. 861 00:48:34,640 --> 00:48:37,997 Women use AI a tiny fraction of the time 862 00:48:38,080 --> 00:48:42,237 compared to men and I think you know AI is an opportunity for us to level the 863 00:48:42,320 --> 00:48:46,077 playing field in some ways, but it requires people to be engaged. That 864 00:48:46,160 --> 00:48:49,797 doesn't mean hey, you have to come pay us all this money and sign up. Just be 865 00:48:49,880 --> 00:48:53,077 comfortable, fluent. Don't be afraid of the technology because the more you know 866 00:48:53,160 --> 00:48:57,117 about it, the more you understand it the better informed you are going to be to 867 00:48:57,200 --> 00:49:01,477 use your voice effectively. >> And it's here. You're in it and if 868 00:49:01,560 --> 00:49:05,597 you're going to continue to live, you're going to be in the world where It's like 869 00:49:05,680 --> 00:49:09,637 part of our society. It's not It's not going away anymore than the train or the 870 00:49:09,720 --> 00:49:13,517 steam engine is, you know. I think of it like the internet, right? 871 00:49:13,600 --> 00:49:16,517 Everybody had somebody's parent who said I'm not going to be online. It's 872 00:49:16,600 --> 00:49:20,197 impossible, right? Everybody has to be on the internet now in some form and I 873 00:49:20,280 --> 00:49:23,477 think artificial intelligence is just the next instantiation of that. So one 874 00:49:23,560 --> 00:49:27,157 of the things I want to talk about is the fact that AI just keeps 875 00:49:27,240 --> 00:49:29,877 It's It just keeps improving on itself. 876 00:49:29,960 --> 00:49:32,637 That's I guess what it's designed to do 877 00:49:32,720 --> 00:49:35,677 and there is a new AI model called 878 00:49:35,760 --> 00:49:39,717 Mythos. Mythos, yes. So that is like 879 00:49:39,800 --> 00:49:43,837 smarter than the smartest of the smartest. Tell us in a simple term what 880 00:49:43,920 --> 00:49:48,717 Mythos is and what it can and will be doing. Yeah, so you know the general 881 00:49:48,800 --> 00:49:52,437 principles of these models is when we train them, we try and you know, make 882 00:49:52,520 --> 00:49:56,677 them better and better at everything at writing code, at translating languages, 883 00:49:56,760 --> 00:49:58,317 and answering questions, at reading 884 00:49:58,400 --> 00:50:01,157 documents. Um and and Mythos was just 885 00:50:01,240 --> 00:50:03,917 kind of the latest um you know, was the 886 00:50:04,000 --> 00:50:05,757 latest iteration. Every few months we 887 00:50:05,840 --> 00:50:08,477 produce a smarter model and and Mythos 888 00:50:08,560 --> 00:50:11,237 was the latest smart model and you know, 889 00:50:11,320 --> 00:50:15,517 we we made some discoveries a few months before we started Mythos and so it was a 890 00:50:15,600 --> 00:50:20,117 particularly large jump. When we started testing Mythos, when we make these 891 00:50:20,200 --> 00:50:21,877 models, we test them for lots of things. 892 00:50:21,960 --> 00:50:23,877 We say, "Oh, how good is this thing at 893 00:50:23,960 --> 00:50:25,677 translating, you know, French to 894 00:50:25,760 --> 00:50:29,997 English? Or how good is this at um you 895 00:50:30,080 --> 00:50:32,917 know, um recommending a stock portfolio? 896 00:50:33,000 --> 00:50:36,597 Or how good is it at um you know, 897 00:50:36,680 --> 00:50:39,837 finding bugs in software code?" And it 898 00:50:39,920 --> 00:50:41,877 was on that last one that we found 899 00:50:41,960 --> 00:50:44,277 something interesting for Mythos. Which 900 00:50:44,360 --> 00:50:48,957 is that Mythos was much better than 901 00:50:49,040 --> 00:50:54,237 human software engineers at finding ways that code could be exploited and broken 902 00:50:54,320 --> 00:50:57,877 into. You know, have you ever heard about like ransomware attacks on 903 00:50:57,960 --> 00:51:00,437 hospitals or schools? So, you know, 904 00:51:00,520 --> 00:51:03,557 these cyber attacks happen all the time. 905 00:51:03,640 --> 00:51:06,597 And Mythos we found, not because we 906 00:51:06,680 --> 00:51:08,357 designed it to do this or because we 907 00:51:08,440 --> 00:51:10,837 designed it to be dangerous, turned out 908 00:51:10,920 --> 00:51:14,637 to be very good at both conducting cyber 909 00:51:14,720 --> 00:51:18,517 attacks and defending against cyber attacks. And so we said, "Wow, this 910 00:51:18,600 --> 00:51:22,077 thing kind of seems like a weapon." Um we we shouldn't you know, we probably 911 00:51:22,160 --> 00:51:24,637 shouldn't give it to everyone right away. 912 00:51:24,720 --> 00:51:27,917 And so the thing we came up with is we said, "We're going to take some 913 00:51:28,000 --> 00:51:30,917 companies that are particularly core to 914 00:51:31,000 --> 00:51:32,637 the infrastructure of the internet, 915 00:51:32,720 --> 00:51:35,637 right? Some of the top banks, some of 916 00:51:35,720 --> 00:51:38,757 the top um you know, software providers 917 00:51:38,840 --> 00:51:43,717 like Apple or like Microsoft or like Google, and 918 00:51:43,800 --> 00:51:48,277 you know, hand it to them and let them fix everything before 919 00:51:48,360 --> 00:51:50,437 it becomes widely available and things 920 00:51:50,520 --> 00:51:52,397 can be broken. Um so we would kind of 921 00:51:52,480 --> 00:51:55,517 give it to the defenders before the you 922 00:51:55,600 --> 00:52:00,117 know before the attackers got a hold of it. And so we had a program. We've we've 923 00:52:00,200 --> 00:52:02,517 given the model to you know, I think 924 00:52:02,600 --> 00:52:05,797 it's something like 40 companies now and 925 00:52:05,880 --> 00:52:07,597 they're you know, they're busy they're 926 00:52:07,680 --> 00:52:09,877 busily finding bugs. And you know, we 927 00:52:09,960 --> 00:52:13,437 just had a post a few days ago that one 928 00:52:13,520 --> 00:52:15,917 of the well-known companies found more 929 00:52:16,000 --> 00:52:20,637 bugs and fixed more bugs with Mythos in 930 00:52:20,720 --> 00:52:24,677 a week than uh they had for the last year. 931 00:52:24,760 --> 00:52:27,597 Uh and and and so we are you know, this 932 00:52:27,680 --> 00:52:32,397 is really changing things. Now our hope is the defenders can fix everything and 933 00:52:32,480 --> 00:52:33,757 everything gets more seamless, 934 00:52:33,840 --> 00:52:35,677 everything gets harder to break into. So 935 00:52:35,760 --> 00:52:39,757 our hope is the era of ransomware and 936 00:52:39,840 --> 00:52:42,677 you know, spies breaking into your phone 937 00:52:42,760 --> 00:52:44,437 and you know, People are worried about 938 00:52:44,520 --> 00:52:48,517 bio-terrorism and should be. So our our 939 00:52:48,600 --> 00:52:51,357 yeah, our our our hope is that at the 940 00:52:51,440 --> 00:52:53,477 end of this say at the end of this year, 941 00:52:53,560 --> 00:52:57,077 we get to a point where all of these you 942 00:52:57,160 --> 00:52:59,077 know, many of these things have been 943 00:52:59,160 --> 00:53:01,197 fixed and we live in a more secure world 944 00:53:01,280 --> 00:53:04,597 than we did before. But in the meantime, 945 00:53:04,680 --> 00:53:08,277 you know, there's this dangerous period during which you have these these kind 946 00:53:08,360 --> 00:53:12,517 of incredible capabilities that you know, that you know, you don't want to 947 00:53:12,600 --> 00:53:16,197 just necessarily loose on the world. And so we're trying to we're trying to 948 00:53:16,280 --> 00:53:20,037 manage that by giving it to the defenders before we give it to the 949 00:53:20,120 --> 00:53:22,797 attackers. Thank you for that. Does is 950 00:53:22,880 --> 00:53:27,517 this is the kind of thing that keeps you up at night? Do you sleep well? Uh I I 951 00:53:27,600 --> 00:53:30,877 you know, I cannot legitimately say that I sleep. I wish I could answer that 952 00:53:30,960 --> 00:53:34,317 question differently, but uh but you know, I got to I got to give you an 953 00:53:34,400 --> 00:53:38,197 honest answer here which is you know, I don't always sleep so well. I would I 954 00:53:38,280 --> 00:53:41,197 would say, you know, going back to the earlier part of the conversation we were 955 00:53:41,280 --> 00:53:44,117 having Because you're worried about these things. 956 00:53:44,200 --> 00:53:48,157 This this and you know, just everything just wanting to get it right. That's the 957 00:53:48,240 --> 00:53:51,197 thing that worries me the most. Just like, you know, 958 00:53:51,280 --> 00:53:55,757 for all our intentions, all of this stuff is incredibly complex. Just are we 959 00:53:55,840 --> 00:53:58,957 going to screw up? Yeah. And and I mean the answer is yes. At some point we'll 960 00:53:59,040 --> 00:54:01,157 screw up something. But but like 961 00:54:01,240 --> 00:54:02,517 are we going to get the big stuff right? 962 00:54:02,600 --> 00:54:05,997 Yeah. Yeah. And also because you all are 963 00:54:06,080 --> 00:54:08,117 fundamentally principled people. 964 00:54:08,200 --> 00:54:12,557 >> Yes. So I think that actually makes a big difference when you're fundamentally 965 00:54:12,640 --> 00:54:15,477 principled people. Uh 966 00:54:15,560 --> 00:54:18,917 not everybody is. And uh there's a lot 967 00:54:19,000 --> 00:54:21,477 of egos in 968 00:54:21,560 --> 00:54:24,597 in every field. But I think particularly 969 00:54:24,680 --> 00:54:27,117 when people are making so many billions 970 00:54:27,200 --> 00:54:29,557 and billions and billions of dollars and 971 00:54:29,640 --> 00:54:31,637 there's the race to beat China and all 972 00:54:31,720 --> 00:54:36,757 of that that it's easily to lose sight of what's what really matters and what 973 00:54:36,840 --> 00:54:40,717 really matters to the public. I'm wondering what you all have done to sort 974 00:54:40,800 --> 00:54:43,197 of keep your egos in check. You know, I 975 00:54:43,280 --> 00:54:46,797 think it again, to us it always comes 976 00:54:46,880 --> 00:54:50,677 back to the mission, right? Why we did this in the first place. Why did we 977 00:54:50,760 --> 00:54:52,997 leave Open AI to start Anthropic? 978 00:54:53,080 --> 00:54:58,197 Because we wanted to make sure that AI goes well for everybody, right? That's 979 00:54:58,280 --> 00:54:59,957 why we're a public benefit corporation. 980 00:55:00,040 --> 00:55:02,157 That's why we took the 80% pledge. 981 00:55:02,240 --> 00:55:06,317 That's why we've made all of these difficult decisions. You know, I was in 982 00:55:06,400 --> 00:55:10,317 our internal like our internal Slack messaging app today. You know, I talk to 983 00:55:10,400 --> 00:55:14,157 the company a lot. And one of the things I was saying is, you know, we're 3,500 984 00:55:14,240 --> 00:55:18,277 people now. We we were, you know, like 1,200 people, you know, like 6 months 985 00:55:18,360 --> 00:55:23,077 ago. At one point, because because I encourage this, someone posted in my 986 00:55:23,160 --> 00:55:26,597 like, you know, like my personal Slack channel, you know, I I had done an 987 00:55:26,680 --> 00:55:29,917 all-hands in front of the company and like, you know, said something and I 988 00:55:30,000 --> 00:55:32,997 think I said it in not quite the right way. And someone someone wrote in my 989 00:55:33,080 --> 00:55:35,797 channel. Usually it's only me who writes things in my channel. Someone wrote in 990 00:55:35,880 --> 00:55:39,157 my channel, "This was a pretty bad thing to say and it like makes everyone feel 991 00:55:39,240 --> 00:55:43,037 bad and like you know this this like you shouldn't have said it this way. 992 00:55:43,120 --> 00:55:45,757 And of course you know when I read it I you know the initial reaction is always 993 00:55:45,840 --> 00:55:49,477 like oh man like I feel like I'm being attacked. But 994 00:55:49,560 --> 00:55:53,117 but but like you take a beat and then and then two minutes later I'm like I'm 995 00:55:53,200 --> 00:55:57,157 so glad people are saying this and then you know I wrote something that was like 996 00:55:57,240 --> 00:56:00,317 I'm so glad you're what you're like willing to say that you know and and 997 00:56:00,400 --> 00:56:03,397 then people noticed this like many people messaged me where I was like I 998 00:56:03,480 --> 00:56:07,077 didn't know you could do that. I didn't know you could just contradict the CEO 999 00:56:07,160 --> 00:56:11,277 of a you know 3,000 person but it's you know you you 1000 00:56:11,360 --> 00:56:14,277 got to you've got to you've got to. You know why that's a powerful example? 1001 00:56:14,360 --> 00:56:16,757 Early early on like in the late 80s I 1002 00:56:16,840 --> 00:56:18,997 think I I met Henry Kravis for the first 1003 00:56:19,080 --> 00:56:24,317 time and he was saying to me be aware that the more money you accumulate the 1004 00:56:24,400 --> 00:56:28,197 less chance that you're ever going to hear the truth. Yes. Because he said 1005 00:56:28,280 --> 00:56:32,597 rich men and I guess he also meant rich women too but he goes rich men rarely 1006 00:56:32,680 --> 00:56:36,717 hear the truth. >> Yes. Yeah. I I I'll also just add I think 1007 00:56:36,800 --> 00:56:41,317 what you do in your personal life matters too there's what you do at work but 1008 00:56:41,400 --> 00:56:44,677 I basically have the same friends I've had since like high school and college 1009 00:56:44,760 --> 00:56:47,797 and guess what they have no trouble being honest with me and they have no 1010 00:56:47,880 --> 00:56:51,157 trouble telling me the the truth and they have no trouble disagreeing with 1011 00:56:51,240 --> 00:56:55,277 public decisions that Anthropic has made right? So I think the last five years 1012 00:56:55,360 --> 00:56:58,277 have changed a lot but I think for me personally I don't really feel like that 1013 00:56:58,360 --> 00:57:02,477 much has changed outside of work and I think that's an important grounding 1014 00:57:02,560 --> 00:57:06,357 characteristic. I think people who tend to be in charge of these big companies 1015 00:57:06,440 --> 00:57:09,317 sometimes it's like they become obsessed just with their work they're not 1016 00:57:09,400 --> 00:57:13,957 well-rounded person and I think there's something when we started Anthropic of 1017 00:57:14,040 --> 00:57:17,397 like the first you know of the seven or eight folks who who really started the 1018 00:57:17,480 --> 00:57:22,877 company only one of them had one kid and now there's 11 kids and I think there's 1019 00:57:22,960 --> 00:57:25,837 something about the fact that we all chose to be parents [clears throat] we 1020 00:57:25,920 --> 00:57:28,477 all want fulfilling real lives outside 1021 00:57:28,560 --> 00:57:33,357 of work. I think that just brings a sort of groundedness humility low ego, and 1022 00:57:33,440 --> 00:57:37,477 just again, like no one is bigger than the mission. So So, what comes through 1023 00:57:37,560 --> 00:57:41,957 really clearly is that sister and brother, you started this, it's bigger 1024 00:57:42,040 --> 00:57:43,877 than you ever imagined, you always 1025 00:57:43,960 --> 00:57:45,517 wanted to make an impact in doing good 1026 00:57:45,600 --> 00:57:49,517 for the world, and you 1027 00:57:49,600 --> 00:57:51,117 stuck to your principles even when it 1028 00:57:51,200 --> 00:57:53,477 was really difficult. What gives you the 1029 00:57:53,560 --> 00:57:58,357 hope that you will get this right? 1030 00:57:58,440 --> 00:58:02,717 So, we're never quite sure, I think, is is the honest answer. Like, we believe 1031 00:58:02,800 --> 00:58:07,877 we're trying our best and trying very hard. Um it's just the decisions are 1032 00:58:07,960 --> 00:58:11,157 complicated, right? If we go back to Mythos, it's like, do we give it to more 1033 00:58:11,240 --> 00:58:13,917 people faster or less people slower? 1034 00:58:14,000 --> 00:58:16,477 Like, who do we trust? And so, it's 1035 00:58:16,560 --> 00:58:21,237 about getting all the details right in addition to just having the right 1036 00:58:21,320 --> 00:58:23,317 principles, which which I believe we do. 1037 00:58:23,400 --> 00:58:28,357 >> Yeah. Um And And so, I think we worry about it every day, and the fact that we 1038 00:58:28,440 --> 00:58:31,477 worry about it every day is maybe the thing that most gives us hope that like, 1039 00:58:31,560 --> 00:58:36,077 well, that we're so paranoid about it, but it's it's it's no guarantee. And 1040 00:58:36,160 --> 00:58:39,077 we've made mistakes, we're going to continue to make mistakes. Like, you 1041 00:58:39,160 --> 00:58:41,397 know, we're that that that's that's a given, right? 1042 00:58:41,480 --> 00:58:44,957 >> Are there other people in Silicon Valley who also 1043 00:58:45,040 --> 00:58:47,397 have it a part of their mission to get 1044 00:58:47,480 --> 00:58:51,997 it right? Do you feel supported in that mission by others? 1045 00:58:52,080 --> 00:58:53,357 It's hard to point to maybe like a 1046 00:58:53,440 --> 00:58:55,957 particular company, but I think 1047 00:58:56,040 --> 00:59:00,837 I am very touched by sometimes, you know, whatever you go and talk in an 1048 00:59:00,920 --> 00:59:05,917 event, or another company wants you to come speak to them, and somebody in the 1049 00:59:06,000 --> 00:59:08,837 audience, you know, comes up to you afterwards or ask a question, and they 1050 00:59:08,920 --> 00:59:12,637 say something like, "Thank you guys for the way that you do things." Yeah. Like, 1051 00:59:12,720 --> 00:59:14,637 thank you for trying to get it right. 1052 00:59:14,720 --> 00:59:18,637 There's this kind of sense that, you know, I think um I think something 1053 00:59:18,720 --> 00:59:22,117 that's great about our relationship is Dario and I I also think we hold each 1054 00:59:22,200 --> 00:59:26,277 other accountable, right? And so, when, you know, we go back and forth on, "Is 1055 00:59:26,360 --> 00:59:28,917 this a right Is this the right decision for the company is this in keeping with 1056 00:59:29,000 --> 00:59:31,997 our values we have each other we have the whole co-founding team is still at 1057 00:59:32,080 --> 00:59:35,997 Anthropic. I think we've also hired an amazing set of people who are also there 1058 00:59:36,080 --> 00:59:39,837 for the mission. There's almost a kind of group accountability and if you pair 1059 00:59:39,920 --> 00:59:44,077 that with like Dario mentioned this culture internally of anybody can 1060 00:59:44,160 --> 00:59:47,997 disagree with us right and yes that's hard to do but there's a lot of forums 1061 00:59:48,080 --> 00:59:50,917 in which they're encouraged to do that their manager can do it they can do it 1062 00:59:51,000 --> 00:59:56,077 anonymously. I think all of those qualities make it more likely that that 1063 00:59:56,160 --> 00:59:57,757 if we start to stray from the path we're 1064 00:59:57,840 --> 01:00:00,877 at least going to know. Okay and you 1065 01:00:00,960 --> 01:00:05,837 want to leave us the public knowing what? What is the most important 1066 01:00:05,920 --> 01:00:07,477 thing you want us to know in this moment 1067 01:00:07,560 --> 01:00:11,157 in time where we are in our AI journey? 1068 01:00:11,240 --> 01:00:13,837 You know there's there's something 1069 01:00:13,920 --> 01:00:16,077 really big happening to the world here. 1070 01:00:16,160 --> 01:00:18,797 This isn't the this isn't the you know 1071 01:00:18,880 --> 01:00:21,077 this isn't the latest technology this 1072 01:00:21,160 --> 01:00:24,517 isn't crypto this isn't mobile this you 1073 01:00:24,600 --> 01:00:26,357 know this this isn't like the latest you 1074 01:00:26,440 --> 01:00:29,557 know gadget or buzzword. Something 1075 01:00:29,640 --> 01:00:33,517 really big is happening to humanity here. 1076 01:00:33,600 --> 01:00:36,437 While the technology is being built by a 1077 01:00:36,520 --> 01:00:38,437 small number of people 1078 01:00:38,520 --> 01:00:42,077 we we need to make it work for everyone 1079 01:00:42,160 --> 01:00:46,797 and we need to make sure that between all of us we make the right decisions 1080 01:00:46,880 --> 01:00:49,037 and and we reckon with 1081 01:00:49,120 --> 01:00:52,477 changes to our daily lives changes to 1082 01:00:52,560 --> 01:00:56,477 our jobs changes to you know 1083 01:00:56,560 --> 01:01:01,317 government and civil liberties and if we do this the right way all of those 1084 01:01:01,400 --> 01:01:05,117 things could change in a positive way all of those things could be better than 1085 01:01:05,200 --> 01:01:06,677 we left them we could have the perfect 1086 01:01:06,760 --> 01:01:09,117 world where you know many many diseases 1087 01:01:09,200 --> 01:01:12,237 are cured people's lives are more joyful 1088 01:01:12,320 --> 01:01:14,597 and meaningful but you know if we get it 1089 01:01:14,680 --> 01:01:17,517 wrong it's easy to imagine things not 1090 01:01:17,600 --> 01:01:20,797 going that way at all and you know it's 1091 01:01:20,880 --> 01:01:24,157 up to all of us Some of us are in a more 1092 01:01:24,240 --> 01:01:28,157 central position, but that's something I think we need to change as well. And, 1093 01:01:28,240 --> 01:01:30,997 you know, it's not easy. It's not any one thing. It's going to be a bunch of 1094 01:01:31,080 --> 01:01:34,717 different things. I think I would probably 1095 01:01:34,800 --> 01:01:38,917 leave people very similar to what Dario said with this with this concept of of 1096 01:01:39,000 --> 01:01:41,557 light and shade. That the potential 1097 01:01:41,640 --> 01:01:46,037 benefits from artificial intelligence are are vast. Um we've written about 1098 01:01:46,120 --> 01:01:50,157 them. We've tried to talk about them publicly. We believe they're real. Um 1099 01:01:50,240 --> 01:01:52,357 and also the risks are are also real, 1100 01:01:52,440 --> 01:01:55,477 right? We're not trying to hype up the 1101 01:01:55,560 --> 01:01:57,837 the positives or hype up the negatives. 1102 01:01:57,920 --> 01:02:00,037 I think they're both true. It's complex, 1103 01:02:00,120 --> 01:02:02,357 but we have to hold two things at once. 1104 01:02:02,440 --> 01:02:04,997 And I think just echoing Dario, this is 1105 01:02:05,080 --> 01:02:10,797 not something that technology companies can do by ourselves. Nor should it be. 1106 01:02:10,880 --> 01:02:11,957 It's something that's going to take 1107 01:02:12,040 --> 01:02:15,077 partnership from a much bigger group of 1108 01:02:15,160 --> 01:02:19,997 people than is currently working on artificial intelligence. And I think the 1109 01:02:20,080 --> 01:02:23,997 first step there has to be don't don't be afraid of it. Learn about it. Lean 1110 01:02:24,080 --> 01:02:26,797 into it. Ask questions. Be critical. 1111 01:02:26,880 --> 01:02:30,957 Just like we say to our staff, right? We are only going to have a chance of 1112 01:02:31,040 --> 01:02:34,957 getting this right if we expand the circle of people that are in the 1113 01:02:35,040 --> 01:02:39,197 conversation. I ask this of all my guests and 1114 01:02:39,280 --> 01:02:42,437 particularly I'm interested from the two of you. 1115 01:02:42,520 --> 01:02:44,837 And what do you think is the meaning of 1116 01:02:44,920 --> 01:02:46,877 a well-lived life? 1117 01:02:46,960 --> 01:02:49,717 And how AI is going to contribute to 1118 01:02:49,800 --> 01:02:52,077 that well-lived life. But what what will 1119 01:02:52,160 --> 01:02:54,197 be the meaning of a well-lived life for 1120 01:02:54,280 --> 01:02:57,077 you, your family, and your children? I 1121 01:02:57,160 --> 01:03:00,717 think for me a well-lived life 1122 01:03:00,800 --> 01:03:02,517 is one where 1123 01:03:02,600 --> 01:03:06,437 you have agency and autonomy. 1124 01:03:06,520 --> 01:03:08,437 You're in the driver seat 1125 01:03:08,520 --> 01:03:12,037 and you're using that in the service of 1126 01:03:12,120 --> 01:03:15,117 being the best version of yourself. 1127 01:03:15,200 --> 01:03:17,037 And I think the best version of yourself 1128 01:03:17,120 --> 01:03:20,597 for you, for your family, for your kids, 1129 01:03:20,680 --> 01:03:22,477 for your broader community, and I think 1130 01:03:22,560 --> 01:03:25,637 for the world. I think um 1131 01:03:25,720 --> 01:03:28,037 the ability to be self-reflective, to 1132 01:03:28,120 --> 01:03:31,597 learn from mistakes, to overcome 1133 01:03:31,680 --> 01:03:34,357 struggles, and to come out the other 1134 01:03:34,440 --> 01:03:36,637 side thinking, "You know, I know more 1135 01:03:36,720 --> 01:03:39,677 about who I am, I like who I am, and I 1136 01:03:39,760 --> 01:03:44,117 want to help other people." I think that to me is the best 1137 01:03:44,200 --> 01:03:48,917 definition of a well-lived, introspective life. And like I talked 1138 01:03:49,000 --> 01:03:50,437 about, I think 1139 01:03:50,520 --> 01:03:53,357 I think there's a world where AI 1140 01:03:53,440 --> 01:03:57,517 could damage so many of the things that I think 1141 01:03:57,600 --> 01:04:00,037 are the hallmarks of a well-lived life. 1142 01:04:00,120 --> 01:04:01,277 And I think there's a world where it 1143 01:04:01,360 --> 01:04:03,517 could enable those, and it could allow a 1144 01:04:03,600 --> 01:04:07,117 lot more people to be able to have 1145 01:04:07,200 --> 01:04:10,277 meaning and value, and to 1146 01:04:10,360 --> 01:04:12,397 self-introspect, to be the best versions 1147 01:04:12,480 --> 01:04:14,717 of ourselves, to show up for each other 1148 01:04:14,800 --> 01:04:19,877 well. And I hope that's the world that we're going to build towards. 1149 01:04:19,960 --> 01:04:22,677 I'd just say the same thing as Daniela in different words. I think it's pretty 1150 01:04:22,760 --> 01:04:27,317 pretty simple. Um you know, I think you know, just operate always with a 1151 01:04:27,400 --> 01:04:30,717 sense of dignity and principle. Like, you know, when you look back and you're 1152 01:04:30,800 --> 01:04:34,437 like, you know, don't you shouldn't when you look back be ashamed of what you did 1153 01:04:34,520 --> 01:04:37,757 or say, you know, I you know, I wish I'd done this this other thing. Of course 1154 01:04:37,840 --> 01:04:41,477 you'll make mistakes, but you know, you were you were putting every iota of 1155 01:04:41,560 --> 01:04:44,677 effort you could into into, you know, into 1156 01:04:44,760 --> 01:04:47,037 um trying to get it right. Um you know, 1157 01:04:47,120 --> 01:04:51,917 I think we have duties both to individuals, you know, to to to to my to 1158 01:04:52,000 --> 01:04:56,397 my co-founders, to all the other people in my life, but you know, also to, you 1159 01:04:56,480 --> 01:05:00,757 know, the the users of our technology, to, you know, everyone else in the 1160 01:05:00,840 --> 01:05:04,717 country and world as citizens. So, you know, I think we have duties to all 1161 01:05:04,800 --> 01:05:06,477 those people that would that we have to 1162 01:05:06,560 --> 01:05:08,637 that we have to attend to. 1163 01:05:08,720 --> 01:05:12,157 Uh and, you know, making sure that uh 1164 01:05:12,240 --> 01:05:16,317 you know, at the end of it it was directed at some purpose some consistent 1165 01:05:16,400 --> 01:05:20,157 view of what we were trying to accomplish in the world that the vision 1166 01:05:20,240 --> 01:05:23,317 we were trying to make happen maybe it doesn't happen maybe we screw it up 1167 01:05:23,400 --> 01:05:25,317 maybe we make terrible mistakes but you 1168 01:05:25,400 --> 01:05:27,717 know just just having the having the 1169 01:05:27,800 --> 01:05:31,197 vision be clear and and and trying to to 1170 01:05:31,280 --> 01:05:35,357 pursue it with with integrity if you if you can do that of course you want the 1171 01:05:35,440 --> 01:05:40,077 outcome too but it I think if doing that is all you can control right and and 1172 01:05:40,160 --> 01:05:43,117 more important ultimately than the outcome. 1173 01:05:43,200 --> 01:05:47,117 Thank you for growing us forward. You you that's what you're doing. You're 1174 01:05:47,200 --> 01:05:49,757 doing exactly what you envisioned as little kids. 1175 01:05:49,840 --> 01:05:51,577 >> It it is very strange but we are. 1176 01:05:51,660 --> 01:05:56,517 [laughter] Yes it is. Thank you thank you Dario and Daniela. 1177 01:05:56,600 --> 01:06:00,437 You've given us a lot to think about. Go well everybody go well. 1178 01:06:00,520 --> 01:06:04,077 You can subscribe to the Oprah podcast on YouTube and follow us on Spotify 1179 01:06:04,160 --> 01:06:06,437 Apple podcasts or wherever you listen. 1180 01:06:06,520 --> 01:06:10,320 I'll see you next week. Thanks everybody.