1 00:00:35,840 --> 00:00:37,997 Please welcome Andrew Ross Sorkin and 2 00:00:38,080 --> 00:00:40,317 his guest, CEO and co-founder of 3 00:00:40,400 --> 00:00:44,197 Anthropic, Dario Amodei. 4 00:00:44,280 --> 00:00:47,757 Good afternoon, everybody. I hope you guys all had a great lunch. 5 00:00:47,840 --> 00:00:49,997 We have a huge afternoon starting with 6 00:00:50,080 --> 00:00:55,477 Dario Amodei here. He is one of the most consequential people here in the world 7 00:00:55,560 --> 00:00:59,357 of artificial intelligence. He's the co-founder and CEO of Anthropic, of 8 00:00:59,440 --> 00:01:02,397 course, known for its Claude model. Uh it's one of the fastest-growing 9 00:01:02,480 --> 00:01:03,797 technology companies in history and 10 00:01:03,880 --> 00:01:07,517 uniquely backed now by all three giants, 11 00:01:07,600 --> 00:01:12,277 uh tech giants, Amazon, Microsoft, and Google. That's new, at least one of them. 12 00:01:12,360 --> 00:01:16,797 Uh and by the way, he has been at this uh really longer than most. Uh he worked 13 00:01:16,880 --> 00:01:18,797 at Baidu, then Google, was an early 14 00:01:18,880 --> 00:01:20,757 employee at OpenAI where he led the 15 00:01:20,840 --> 00:01:25,037 development of uh ChatGPT-2 and 3. Uh 16 00:01:25,120 --> 00:01:28,037 and the reason that we wanted to speak with him this year more than anything 17 00:01:28,120 --> 00:01:32,197 else is because he has singularly been perhaps the most outspoken and candid 18 00:01:32,280 --> 00:01:36,557 person about AI when it comes to the way he's been thinking about jobs and job 19 00:01:36,640 --> 00:01:38,797 losses and selling chips to China and 20 00:01:38,880 --> 00:01:43,597 politics in our country and where all of this goes. So, welcome to you. Thank you 21 00:01:43,680 --> 00:01:46,237 for being here. >> Thank you for having me. Um we got a lot 22 00:01:46,320 --> 00:01:50,957 to talk about, um including, by the way, are we in an AI bubble? But I promise 23 00:01:51,040 --> 00:01:55,917 you we will get there. I I I'll start here, though, um which is 24 00:01:56,000 --> 00:01:56,997 I mentioned you were a research 25 00:01:57,080 --> 00:02:00,677 scientist uh back at Baidu uh 2014. 26 00:02:00,760 --> 00:02:03,237 And if I had sat with you then 27 00:02:03,320 --> 00:02:05,837 and said, "We're going to sit together 28 00:02:05,920 --> 00:02:08,597 in 2025 29 00:02:08,680 --> 00:02:13,517 talking about AI." You would have told me what? What would have been your 30 00:02:13,600 --> 00:02:17,597 expectation for what would have happened? So, I'll tell you what I am 31 00:02:17,680 --> 00:02:19,197 surprised by and what what and what I'm 32 00:02:19,280 --> 00:02:23,077 not. Um I'm not surprised by the 33 00:02:23,160 --> 00:02:26,357 economic impacts of the technology, the 34 00:02:26,440 --> 00:02:30,837 the value that it's creating, you know, the fact that, you know, I walk by any 35 00:02:30,920 --> 00:02:35,517 billboard in New York and, you know, it's it's it's kind of everything about AI. 36 00:02:35,600 --> 00:02:38,397 >> that this would this would be this would be real by now. 37 00:02:38,480 --> 00:02:41,957 In in some form that this would be real, that it would be central to the economy, 38 00:02:42,040 --> 00:02:45,037 that it would be central to national security. That it would be central to 39 00:02:45,120 --> 00:02:48,157 scientific research. I you know, I I I I 40 00:02:48,240 --> 00:02:52,117 don't think I imagined that I would be leading one of the companies in the 41 00:02:52,200 --> 00:02:55,357 space, right? I think that I think that would have been, 42 00:02:55,440 --> 00:02:58,637 you know, very surprising to me. I didn't think of that as as kind of as 43 00:02:58,720 --> 00:03:02,837 kind of my role at the time. And the exact way in which things happened, all 44 00:03:02,920 --> 00:03:07,317 of the, you know, all of the kind of strange lingo we've developed around 45 00:03:07,400 --> 00:03:12,277 language models, all of the kind of financialization of it, you know, if you 46 00:03:12,360 --> 00:03:16,317 think about it, if you think about the implications of the models becoming as 47 00:03:16,400 --> 00:03:20,477 smart and as powerful as they are and scaling in the way that they are, in the 48 00:03:20,560 --> 00:03:24,877 way that me and some of my colleagues had predicted, it all makes sense. But I 49 00:03:24,960 --> 00:03:28,277 I don't think I would have derived it from first principles. Okay, well, then 50 00:03:28,360 --> 00:03:31,797 let's go straight to the question I said I'd ask at the beginning. Um 51 00:03:31,880 --> 00:03:34,837 because maybe this is the place thinking about where this goes and the fact that 52 00:03:34,920 --> 00:03:37,917 I didn't think you would say, by the way, that you thought that this is where 53 00:03:38,000 --> 00:03:41,557 we would be in in 2025 because I think even people back then 54 00:03:41,640 --> 00:03:45,437 thought this would be a much longer road. But if you're right, 55 00:03:45,520 --> 00:03:49,077 do you look at the amount of economic uh muscle that's being put into this 56 00:03:49,160 --> 00:03:53,157 industry right now? I mean, it really does represent potentially almost all of 57 00:03:53,240 --> 00:03:55,997 the growth in the United States GDP right now, literally. 58 00:03:56,080 --> 00:03:59,637 >> Yes. That we are in some form of a bubble. 59 00:03:59,720 --> 00:04:02,277 Are we overspending? Does the math of 60 00:04:02,360 --> 00:04:05,677 this make sense? So, this is really 61 00:04:05,760 --> 00:04:10,317 complicated and I want to separate out the technological side of it from the 62 00:04:10,400 --> 00:04:14,997 economic side of it. On the technological side of it, I feel really 63 00:04:15,080 --> 00:04:18,477 solid. I think I'm one of the most bullish people around and I think it 64 00:04:18,560 --> 00:04:21,637 pencils out. On the economic side, you 65 00:04:21,720 --> 00:04:26,637 know, I I I have my concerns where even if the technology is is, you know, 66 00:04:26,720 --> 00:04:30,437 really powerful and fulfills all its promises, I think there may be players 67 00:04:30,520 --> 00:04:34,357 in the ecosystem who, if they just make a timing error, if they just get it off 68 00:04:34,440 --> 00:04:37,237 by a little bit, bad things could happen. So, let me go through both of 69 00:04:37,320 --> 00:04:40,477 them. On the technological side, the the 70 00:04:40,560 --> 00:04:45,877 reason that I'm not in honestly, by the pure technology, not that surprised is 71 00:04:45,960 --> 00:04:49,757 myself and some of the people who eventually became my my my co-founders, 72 00:04:49,840 --> 00:04:54,277 we were the first to document the scaling laws of AI, which is you put 73 00:04:54,360 --> 00:04:56,757 more compute, you put more data into AI. 74 00:04:56,840 --> 00:05:00,757 With small modifications, we've seen these things like reasoning models and 75 00:05:00,840 --> 00:05:02,797 test time compute. They're all tiny 76 00:05:02,880 --> 00:05:05,037 little tweaks. And I've been watching 77 00:05:05,120 --> 00:05:08,077 that trend for the last 12 years or so 78 00:05:08,160 --> 00:05:09,957 since I joined the field. And the thing 79 00:05:10,040 --> 00:05:13,117 that is most striking about all of it is 80 00:05:13,200 --> 00:05:17,877 as you train these models in this very simple way, you know, with a few simple 81 00:05:17,960 --> 00:05:21,517 modifications, they get better and better at every task under the sun. They 82 00:05:21,600 --> 00:05:25,277 get better at coding. They get better at doing science. They get better at 83 00:05:25,360 --> 00:05:27,117 biomedicine. They get better at the law. 84 00:05:27,200 --> 00:05:31,077 They get better at finance. They get better at at, you know, materials and 85 00:05:31,160 --> 00:05:35,077 manufacturing. And and that's just that's just a listing of all the sources 86 00:05:35,160 --> 00:05:37,557 of of of of value of value in our 87 00:05:37,640 --> 00:05:40,077 economy. If I just take Anthropic 88 00:05:40,160 --> 00:05:42,997 itself, which, because we work so much 89 00:05:43,080 --> 00:05:47,197 in the enterprise side, I think we're a good barometer, maybe a purer barometer 90 00:05:47,280 --> 00:05:50,717 than the others which kind of filter through consumers, which kind of have 91 00:05:50,800 --> 00:05:53,037 their habits and their use cases. We 92 00:05:53,120 --> 00:05:55,557 look at our revenue. It's grown 10x a 93 00:05:55,640 --> 00:05:58,597 year every year for the last 3 years. 0 94 00:05:58,680 --> 00:06:01,197 to 100 million 2023, 100 million to 95 00:06:01,280 --> 00:06:06,397 billion 2024, 1 billion to it's going to end somewhere between 8 and 10 at at the 96 00:06:06,480 --> 00:06:10,957 end of this year. Will it continue? I don't know, but the technology is 97 00:06:11,040 --> 00:06:15,277 driving there and the economic value is coming with it. I it will, you know, 98 00:06:15,360 --> 00:06:18,997 that that trend is going to slow down for sure, but it's still going to be 99 00:06:19,080 --> 00:06:21,317 really fast. And so, I have this 100 00:06:21,400 --> 00:06:25,357 confidence that the the ult- eventually 101 00:06:25,440 --> 00:06:28,037 the economic value is the eventually the 102 00:06:28,120 --> 00:06:32,877 economic value is going to be there. Um When you but let's just go to this 103 00:06:32,960 --> 00:06:36,557 because there are companies that are spending 100 billion dollars a year 104 00:06:36,640 --> 00:06:38,917 more. You're going to be spending 50. 105 00:06:39,000 --> 00:06:42,397 You look at what Sam Altman, who was here last year, plans to be spending. 106 00:06:42,480 --> 00:06:44,237 These are extraordinary numbers and this 107 00:06:44,320 --> 00:06:47,877 is all a bet, a big bet, that this is 108 00:06:47,960 --> 00:06:51,997 going to scale in this way. And and my question is, is there a real way to 109 00:06:52,080 --> 00:06:56,277 pencil this out or is this more of a 110 00:06:56,360 --> 00:06:58,837 a gut feeling at this point? So, let me 111 00:06:58,920 --> 00:07:00,517 let Yeah, so that really gets to the 112 00:07:00,600 --> 00:07:03,517 second part of it. And and I I I I will 113 00:07:03,600 --> 00:07:06,597 describe as transparently as I can, I 114 00:07:06,680 --> 00:07:08,677 think there's a real dilemma 115 00:07:08,760 --> 00:07:13,917 deriving from uncertainty in how quickly the economic value is going to grow and 116 00:07:14,000 --> 00:07:17,837 the lag times on building the data centers that that drives it. So, I think 117 00:07:17,920 --> 00:07:21,437 there's genuine uncertainty. There's genuine dilemma, which we as a company 118 00:07:21,520 --> 00:07:24,237 try to manage as responsibly as we can. 119 00:07:24,320 --> 00:07:28,437 And then I think there are some players who, you know, who who are YOLOing, who 120 00:07:28,520 --> 00:07:31,997 who pull pull the risk dial too far and I'm very concerned. 121 00:07:32,080 --> 00:07:35,197 >> Who is YOLOing? >> So, I That's a question I'm not going to 122 00:07:35,280 --> 00:07:36,237 answer. 123 00:07:36,320 --> 00:07:40,877 Um uh uh So, on the first one, 124 00:07:40,960 --> 00:07:43,357 put yourself in the position of of Anthropic. 125 00:07:43,440 --> 00:07:47,117 >> yourself >> Put put yourself in in the position of 126 00:07:47,200 --> 00:07:50,397 of Anthropic. Put yourself in my position. 127 00:07:50,480 --> 00:07:54,917 You you you've seen this revenue curve that goes up 10x a year for 3 years. 128 00:07:55,000 --> 00:07:57,957 You're like, "Okay, what's going to happen next year?" If I'm really dumb 129 00:07:58,040 --> 00:08:01,877 and I extrapolate the pattern 10 to 100 billion. I don't believe that, just to 130 00:08:01,960 --> 00:08:04,277 be clear. I don't I don't believe that at all. Even though it's happened in the 131 00:08:04,360 --> 00:08:08,237 last 3 years, just at this scale, I don't believe it. But that's one of the 132 00:08:08,320 --> 00:08:12,757 outer bounds of, you know, the outer limits of possibility. If I go in and 133 00:08:12,840 --> 00:08:16,077 I'm like, "This enterprise and that enterprise and this use case and this is 134 00:08:16,160 --> 00:08:20,157 our go-to-market motion." If I try and do it, you know, bottom-up, then, you 135 00:08:20,240 --> 00:08:24,957 know, maybe it's maybe it's 20 or 30 or something like that. So, there is well, 136 00:08:25,040 --> 00:08:26,877 what I've been calling internally this 137 00:08:26,960 --> 00:08:30,037 cone of uncertainty where I don't know 138 00:08:30,120 --> 00:08:34,237 if a year from now it's going to be 20 billion or, you know, it's it's going to 139 00:08:34,320 --> 00:08:36,517 be 50 or or like, you know, it's it's 140 00:08:36,600 --> 00:08:39,557 it's very uncertain. I try to plan in a 141 00:08:39,640 --> 00:08:41,357 conservative way, so I plan for the 142 00:08:41,440 --> 00:08:44,437 lower side of it. But but that is very 143 00:08:44,520 --> 00:08:47,037 disconcerting and you add to that the 144 00:08:47,120 --> 00:08:51,557 idea that building the data centers has a long lag 145 00:08:51,640 --> 00:08:53,677 time. It's like a year or two. So, I 146 00:08:53,760 --> 00:08:55,997 have to decide now, literally now, or in 147 00:08:56,080 --> 00:08:58,717 some cases a few months ago, how how how 148 00:08:58,800 --> 00:09:01,797 much compute I need to buy in you know, 149 00:09:01,880 --> 00:09:04,397 early 2024 to serve the models in early 150 00:09:04,480 --> 00:09:07,197 2027 when I get to that revenue amount. 151 00:09:07,280 --> 00:09:12,157 And there's two couple dangers. One is that if I if I don't buy enough compute, 152 00:09:12,240 --> 00:09:15,157 I won't be able to serve all the customers I want. I'll have to turn them 153 00:09:15,240 --> 00:09:19,237 away and send them to to to my competitors. If I buy too much compute, 154 00:09:19,320 --> 00:09:23,157 of course, I might not get enough revenue to pay for that compute. And and 155 00:09:23,240 --> 00:09:26,397 and, you know, and and, you know, in the extreme case, there's kind of the risk 156 00:09:26,480 --> 00:09:29,077 of going bankrupt. And how how much 157 00:09:29,160 --> 00:09:33,717 buffer there is in that in that cone, um is basically determined by my margins. 158 00:09:33,800 --> 00:09:36,957 If I have 80% margins, I can buy 20 billion dollars of compute and it could 159 00:09:37,040 --> 00:09:39,237 serve a 100 billion dollar a 100 billion 160 00:09:39,320 --> 00:09:41,437 dollars of revenue. But because the cone 161 00:09:41,520 --> 00:09:44,637 is so wide, it's hard to avoid, you 162 00:09:44,720 --> 00:09:46,797 know, making a mistake on on 163 00:09:46,880 --> 00:09:49,437 on on one side on one side or the other. 164 00:09:49,520 --> 00:09:52,877 Now, we've been a relatively responsible 165 00:09:52,960 --> 00:09:57,957 company, um and I think because we focus on enterprise, I think we have a better 166 00:09:58,040 --> 00:10:01,117 business model. I think we have better margins. I think we're being responsible 167 00:10:01,200 --> 00:10:05,157 about it. But again, let's say you have a different business model. Let's say 168 00:10:05,240 --> 00:10:09,477 you have a consumer business model. You know, your your your your your your kind 169 00:10:09,560 --> 00:10:14,677 of the source of your revenue isn't as good. Your margins aren't certain. And 170 00:10:14,760 --> 00:10:18,597 let's say you're a person who just kind of like constitutionally just wants to 171 00:10:18,680 --> 00:10:21,157 YOLO things or just likes big numbers. 172 00:10:21,240 --> 00:10:25,397 Um then then you may turn that dial pretty far. And so I think there's a 173 00:10:25,480 --> 00:10:29,237 real underlying risk. Whenever there's uncertainty, there's a risk of 174 00:10:29,320 --> 00:10:31,997 overextension. We all face it. I face 175 00:10:32,080 --> 00:10:34,037 it. All the other companies face it. 176 00:10:34,120 --> 00:10:36,397 There's there's an inherent risk when 177 00:10:36,480 --> 00:10:38,677 when the timing of the economic value is 178 00:10:38,760 --> 00:10:43,997 is uncertain. There's inherent risk of of under-reacting or overextension. And 179 00:10:44,080 --> 00:10:47,997 because the companies are competing with each other and frankly we genuinely need 180 00:10:48,080 --> 00:10:52,517 to compete with you know our authoritarian adversaries. Um 181 00:10:52,600 --> 00:10:56,197 there there's there's kind of a lot of pressure to push things. So I think 182 00:10:56,280 --> 00:10:58,117 there's some amount of irreducible risk 183 00:10:58,200 --> 00:11:01,237 here and I absolutely don't want to deny 184 00:11:01,320 --> 00:11:05,237 this. But at the same time as that I think there are some players who are not 185 00:11:05,320 --> 00:11:08,597 managing that risk well, who are taking unwise risks. 186 00:11:08,680 --> 00:11:11,917 >> All right, let me let me ask you about that with maybe you'll mention who that 187 00:11:12,000 --> 00:11:14,557 is. I think we all know who that is. 188 00:11:14,640 --> 00:11:18,037 Um You have said that you're going to break even. 189 00:11:18,120 --> 00:11:19,917 Uh this is privately at least to your 190 00:11:20,000 --> 00:11:22,877 investors by 2028 even with the spending 191 00:11:22,960 --> 00:11:25,277 plans, I think. Uh Sam Altman who you 192 00:11:25,360 --> 00:11:30,117 worked with and for says that he's going to do it by 2030. 193 00:11:30,200 --> 00:11:33,117 Uh I use his math, not yours. He would have 194 00:11:33,200 --> 00:11:38,877 to go from a $74 billion loss in the course of 2 years to be profitable 2 195 00:11:38,960 --> 00:11:43,677 years later. Does that make sense to you? So look, I don't I don't know the internal 196 00:11:43,760 --> 00:11:47,437 financials of any other company. I can't I can't say anything about what the 197 00:11:47,520 --> 00:11:50,397 economics of any other of of of of of 198 00:11:50,480 --> 00:11:52,557 any other company is. I will just go 199 00:11:52,640 --> 00:11:55,717 back to our own calculation and the cone 200 00:11:55,800 --> 00:11:59,837 of uncertainty where where where where where where we basically say we want to 201 00:11:59,920 --> 00:12:04,557 buy enough compute that we're confident, you know, even in the 10th percentile, 202 00:12:04,640 --> 00:12:07,637 you know, scenario like might be in a bad position, but like we think we can 203 00:12:07,720 --> 00:12:10,997 pay for it. There's there's some end of the curve where you know, things go so 204 00:12:11,080 --> 00:12:14,917 badly that you know, we can't pay that that that there's all there's always a 205 00:12:15,000 --> 00:12:19,917 tail risk. There's not zero, but we're trying to manage that risk well while 206 00:12:20,000 --> 00:12:23,557 also buying an amount an amount of compute that allows us to be competitive 207 00:12:23,640 --> 00:12:26,797 with the other players. We're very efficient training. We're very efficient 208 00:12:26,880 --> 00:12:30,717 in inference. We have good margins. I think I you know, I think I think we can 209 00:12:30,800 --> 00:12:35,157 manage it. I think the odds are on our side. What are we supposed to think of 210 00:12:35,240 --> 00:12:38,997 what I think people now describe as circular deals. Back in the day we 211 00:12:39,080 --> 00:12:43,277 called this vendor financing. But in this context, you have a situation where 212 00:12:43,360 --> 00:12:47,557 Nvidia in particular, but others as well, effectively have 213 00:12:47,640 --> 00:12:53,077 been taking stakes in companies and invariably those companies are using 214 00:12:53,160 --> 00:12:56,157 some of that money one way or the other considering money's fungible and going 215 00:12:56,240 --> 00:12:59,157 to buy Nvidia chips. Yes. So I mean, 216 00:12:59,240 --> 00:13:03,437 we've done some of these deals, not not at the same scale as as as some other 217 00:13:03,520 --> 00:13:07,277 players. But we've done some of these deals and I 218 00:13:07,360 --> 00:13:09,877 can just I can just I can just I can just I can just I can just kind of walk 219 00:13:09,960 --> 00:13:13,077 you through not a specific deal cuz I'm not going to go into the details, but 220 00:13:13,160 --> 00:13:17,077 but but kind of like a stylized what these deals often look like and kind of 221 00:13:17,160 --> 00:13:21,597 why they can make sense. So if you want to buy a gigawatt of compute, right? 222 00:13:21,680 --> 00:13:23,797 That buying the chips and building the 223 00:13:23,880 --> 00:13:26,397 chips, building everything, that costs, 224 00:13:26,480 --> 00:13:28,837 you know, roughly let's say $50 billion 225 00:13:28,920 --> 00:13:31,517 of of capital expense to to kind of to 226 00:13:31,600 --> 00:13:35,997 kind of fund that. And you can think of that as useful lifetime as people argue 227 00:13:36,080 --> 00:13:40,477 about it, but maybe it's 5 years. So that's 10 10 billion dollars for basic 228 00:13:40,560 --> 00:13:44,357 basically for 5 years. And and so if you're a company, you know, you're 229 00:13:44,440 --> 00:13:47,357 you're you're you're a company that's making, you know, 8 10 billion dollars 230 00:13:47,440 --> 00:13:48,877 of revenue. You think that's growing. 231 00:13:48,960 --> 00:13:50,437 You don't know how fast it's growing. 232 00:13:50,520 --> 00:13:53,517 You have to make the decision right now and you don't have $50 billion. You 233 00:13:53,600 --> 00:13:57,597 don't have $50 billion on you. So you know, a thing you can do, a deal you can 234 00:13:57,680 --> 00:14:01,237 make with a large player who has an incentive to to to to to do this because 235 00:14:01,320 --> 00:14:04,637 they're the one selling the chips or providing the cloud is they'll say, 236 00:14:04,720 --> 00:14:08,637 "Okay, you know, I'll give you I don't know, 20% of it. I'll invest $10 237 00:14:08,720 --> 00:14:12,237 billion." So that lets you pay for the first year and then for and then for the 238 00:14:12,320 --> 00:14:16,037 other years you can kind of pay as you go because I know you don't have $50 239 00:14:16,120 --> 00:14:20,157 billion now, but but you know, looking at how things are are are are are 240 00:14:20,240 --> 00:14:24,237 growing, that isn't a crazy bet. We're already almost we're already almost at 241 00:14:24,320 --> 00:14:28,797 at at we're already almost at the $10 billion of revenue. So it it kind of it 242 00:14:28,880 --> 00:14:31,917 you know, it takes a year to build a data center. It's financed for a year. 243 00:14:32,000 --> 00:14:34,357 So you're you're you know, you're basically saying I need to get $10 244 00:14:34,440 --> 00:14:37,357 billion of of revenue per year 2 years 245 00:14:37,440 --> 00:14:41,717 from now. Um so I don't think there's anything wrong with that. One player has 246 00:14:41,800 --> 00:14:45,197 capital and has an interest because they're selling the because they're 247 00:14:45,280 --> 00:14:49,477 selling the you know, they're selling the chips and the other player is pretty 248 00:14:49,560 --> 00:14:51,877 confident they'll have they'll have the revenue at the right time, but they 249 00:14:51,960 --> 00:14:55,197 don't have $50 billion at hand. So I don't think there's anything 250 00:14:55,280 --> 00:14:57,037 inappropriate about that in principle. 251 00:14:57,120 --> 00:15:00,957 Now if you start stacking these where they get to huge amounts of money and 252 00:15:01,040 --> 00:15:05,637 you're saying, you know, by by 2027 or 2028 I need to make $200 billion a year, 253 00:15:05,720 --> 00:15:09,477 then yeah, you can you you can you can overextend yourself, of course. It's all 254 00:15:09,560 --> 00:15:12,117 a matter of the size. >> But I think is one of the key questions 255 00:15:12,200 --> 00:15:15,917 behind the math of this entire industry is what's called the depreciation 256 00:15:16,000 --> 00:15:18,517 schedule, if you will, for these chips. 257 00:15:18,600 --> 00:15:21,437 And there seems to be a big debate about it. Meaning 258 00:15:21,520 --> 00:15:25,597 when you buy a new chip, is that chip going to work for you 259 00:15:25,680 --> 00:15:29,317 effectively for three or four years or is it going to work for you for six or 260 00:15:29,400 --> 00:15:31,597 seven or eight years or even 10 years? 261 00:15:31,680 --> 00:15:35,157 And depending on where you think the math really lands, 262 00:15:35,240 --> 00:15:38,277 all of this pencils out or really doesn't. 263 00:15:38,360 --> 00:15:39,957 What do you think the schedule is? 264 00:15:40,040 --> 00:15:43,397 >> from from from from our point of view, we make very conservative assumptions 265 00:15:43,480 --> 00:15:46,037 here. Um I don't think there's a particular 266 00:15:46,120 --> 00:15:49,837 depreciation schedule, right? When new chips come out, the issue isn't the 267 00:15:49,920 --> 00:15:53,437 lifetime of the chips. Chips chips keep working for a long time. The issue is 268 00:15:53,520 --> 00:15:57,317 new chips come out that are faster and cheaper and and and and and 269 00:15:57,400 --> 00:16:01,437 >> if your competitors have them. Yeah, yeah. And so and so kind of the value of 270 00:16:01,520 --> 00:16:05,237 old chips goes down someone. In fact, that that can happen you know, a year 271 00:16:05,320 --> 00:16:08,117 after you buy the chips because you know, now there are multiple companies, 272 00:16:08,200 --> 00:16:12,837 TPUs, GPUs coming out with coming out with new chips. So I think the way we 273 00:16:12,920 --> 00:16:15,997 think about it is we take into account that you know, the old chips are going 274 00:16:16,080 --> 00:16:18,997 to be less valuable as as as time goes 275 00:16:19,080 --> 00:16:23,317 on. And you know, we assume very aggressive, you know, 276 00:16:23,400 --> 00:16:27,157 kind of continuation of the of the chip efficiency curve. And again, I can only 277 00:16:27,240 --> 00:16:30,877 speak for Anthropic. Like we make conservative assumptions here and and we 278 00:16:30,960 --> 00:16:34,757 think we're going to be we think we're going to be okay in basically almost all 279 00:16:34,840 --> 00:16:37,677 worlds. Can't be literally all worlds, but we think we're going to be okay in 280 00:16:37,760 --> 00:16:41,877 almost all worlds. I can't speak for other companies. Again, I can imagine 281 00:16:41,960 --> 00:16:44,957 that there may be other players out there who are, you know, who are 282 00:16:45,040 --> 00:16:48,517 diluting themselves and and you know, making making assumptions that are very 283 00:16:48,600 --> 00:16:52,717 far inflected on the optimistic So we're clear, there's only two of you who are 284 00:16:52,800 --> 00:16:55,477 who are not >> I look, I I I just don't know who you're 285 00:16:55,560 --> 00:16:58,357 talking about, okay? I I just have no idea. 286 00:16:58,440 --> 00:17:02,957 Um Let me ask you this about the models themselves and how you 287 00:17:03,040 --> 00:17:05,477 see the competition. So one of the things that's happened literally in the 288 00:17:05,560 --> 00:17:10,557 last week and there has been a complete sort of meltdown in the valley over 289 00:17:10,640 --> 00:17:15,077 what's happening here. And Sundar Pichai was also here last year and it appears 290 00:17:15,160 --> 00:17:16,877 at least that his new model has gotten a 291 00:17:16,960 --> 00:17:22,157 lot of people excited about what he's doing. And that Google, which I used to 292 00:17:22,240 --> 00:17:25,877 think from the beginning given all of the data they have, should have been 293 00:17:25,960 --> 00:17:30,557 sort of the the winner by default. And you have a memo that went out from Sam Altman 294 00:17:30,640 --> 00:17:34,117 saying that code red, everyone's got to get back to their desks to figure out 295 00:17:34,200 --> 00:17:37,757 what the what the next thing is to break the to get to the next place. How do you 296 00:17:37,840 --> 00:17:42,277 stack rank right now where these models are and how important do you think it is 297 00:17:42,360 --> 00:17:46,997 in any given moment? Yeah. So this is one of the cases where I'm just very 298 00:17:47,080 --> 00:17:51,357 grateful that Anthropic is taking a different path. Right? On one hand 299 00:17:51,440 --> 00:17:54,917 >> being the enterprise doing The the path being the enterprise, right? Both of the 300 00:17:55,000 --> 00:17:58,517 players that you that you mentioned, right? Both of these other two players 301 00:17:58,600 --> 00:18:02,157 are, you know, primarily focused on the consumer. They try and do some 302 00:18:02,240 --> 00:18:05,197 enterprise work, but they're they're they're and and they're fighting in 303 00:18:05,280 --> 00:18:08,917 consumer. That is, you know, that is that is the reason for kind of the code 304 00:18:09,000 --> 00:18:13,077 red, the intense fighting, right? It's, you know, Google has a search monopoly 305 00:18:13,160 --> 00:18:16,557 that they're trying to defend. And the center of what OpenAI is doing is in is 306 00:18:16,640 --> 00:18:20,517 in consumer. So those those two are fighting it out. For both of them, 307 00:18:20,600 --> 00:18:24,997 serving businesses is secondary. And so what we found is over time we've 308 00:18:25,080 --> 00:18:29,157 optimized our models more and more for the needs of businesses. The one that's 309 00:18:29,240 --> 00:18:33,117 gone the fastest has been coding. You know, I think that has really, you know, 310 00:18:33,200 --> 00:18:36,557 moved forward the quickest, but we're starting to go beyond that to finance, 311 00:18:36,640 --> 00:18:38,997 biomedical, retail, you know, energy, 312 00:18:39,080 --> 00:18:41,957 manufac- all kind of all of that. 313 00:18:42,040 --> 00:18:46,597 And so and so and so what we find is that these model wars, as much as our 314 00:18:46,680 --> 00:18:48,437 models are like really good, like, you 315 00:18:48,520 --> 00:18:53,717 know, the one released last week, Opus 4.5, is hands down, I think almost 316 00:18:53,800 --> 00:18:56,477 everyone thinks, the best model for coding. So I think that's it's very 317 00:18:56,560 --> 00:19:00,397 important that we continue to have this this model superiority, but there's a 318 00:19:00,480 --> 00:19:03,677 way in which we're kind of going in a different direction or or on a different 319 00:19:03,760 --> 00:19:07,877 dimension. And so we we kind of have to worry less about this back and forth. We 320 00:19:07,960 --> 00:19:11,277 have a little bit of a privileged position where we can just keep growing 321 00:19:11,360 --> 00:19:14,117 and just keep developing our models and we don't have to do any code red. 322 00:19:14,200 --> 00:19:18,277 >> But what is the the moat around any of these businesses? 323 00:19:18,360 --> 00:19:22,917 And and when I say that, I assume if Google is success is as successful as it 324 00:19:23,000 --> 00:19:26,397 wants to be or OpenAI or any or or Meta, right? Anybody else who's involved in 325 00:19:26,480 --> 00:19:29,077 this, that they think that one day if we 326 00:19:29,160 --> 00:19:34,357 ever get to AGI, that all these models effectively will be able to do what 327 00:19:34,440 --> 00:19:37,757 either, you know, any of them do. And whether it's, you know, is there a moat? 328 00:19:37,840 --> 00:19:39,757 Is it the persistent memory? So I'm I 329 00:19:39,840 --> 00:19:42,797 use ChatGPT for certain things. It knows 330 00:19:42,880 --> 00:19:46,997 me now because it's I've been, you know, asking it different questions. 331 00:19:47,080 --> 00:19:50,837 Is Or do you think people just switch back and forth, whoever's got the the 332 00:19:50,920 --> 00:19:54,917 latest thing? Yeah. So I you look, I can only speak to the enterprise side. What 333 00:19:55,000 --> 00:19:57,677 I will say is is it is surprising how 334 00:19:57,760 --> 00:20:02,037 different the personality and capabilities of the models are if you're 335 00:20:02,120 --> 00:20:04,997 building for businesses versus if you're building for if you're building for 336 00:20:05,080 --> 00:20:08,557 consumers. Um, you just focus on different things. You focus less on 337 00:20:08,640 --> 00:20:11,717 engagement, you focus more on coding, 338 00:20:11,800 --> 00:20:16,797 high intellectual activities, scientific ability. Um, and you know, I don't think 339 00:20:16,880 --> 00:20:20,597 it's true that if we got AGI they would all converge to the same place. Is 340 00:20:20,680 --> 00:20:23,357 everyone in this audience converge to the same place? Is everyone in this 341 00:20:23,440 --> 00:20:27,397 audience a copy of everyone else because we're all agent? No, we're we're all 342 00:20:27,480 --> 00:20:33,037 specialized. Specialization exists, you know, in in in it exists alongside general 343 00:20:33,120 --> 00:20:35,597 intelligence. And then I think there's all the standard enterprise stuff as 344 00:20:35,680 --> 00:20:39,677 well, which is that um, you know, companies build relationships with you, 345 00:20:39,760 --> 00:20:42,237 they get used to using certain models. 346 00:20:42,320 --> 00:20:43,837 Uh, and and we're starting to see that. 347 00:20:43,920 --> 00:20:47,517 Like even our even our API business, which is basically just selling the raw 348 00:20:47,600 --> 00:20:51,197 model, you wouldn't think that would be very sticky. Companies have great 349 00:20:51,280 --> 00:20:54,997 difficulty switching from one model to another because they have downstream 350 00:20:55,080 --> 00:20:58,317 customers who use the model and they like the current model. And you prompt 351 00:20:58,400 --> 00:21:00,597 and interact with the models in different ways and they have different 352 00:21:00,680 --> 00:21:04,677 personalities. It's actually quite hard to switch. So so I think there really is 353 00:21:04,760 --> 00:21:08,237 a durable business here. Um, one quick AGI question and it's a science 354 00:21:08,320 --> 00:21:12,557 question, which is do you think just the way transformers work today and just 355 00:21:12,640 --> 00:21:17,317 compute power alone from a scalability sense that that is what would get to AGI 356 00:21:17,400 --> 00:21:19,917 or do you think there's some other ingredient and maybe there's a technical 357 00:21:20,000 --> 00:21:24,757 question, but I'm trying to keep it very very easy that has that be included in 358 00:21:24,840 --> 00:21:28,197 this that gets you to some place where this stuff is actually going to really 359 00:21:28,280 --> 00:21:30,837 think on its own? >> No, I think I think scaling is going to 360 00:21:30,920 --> 00:21:34,077 get us there. Again, with small every once in a while there'll be a small 361 00:21:34,160 --> 00:21:37,437 modification, you know, so small you may not even read about it. It's just 362 00:21:37,520 --> 00:21:41,157 something going on in the lab. I have been watching these scaling laws for for 363 00:21:41,240 --> 00:21:44,077 10 years. So what's your what's your what's your timeline now? 364 00:21:44,160 --> 00:21:46,077 There's no one particular point, right? 365 00:21:46,160 --> 00:21:49,877 That this is what I've said over I've never liked these terms AGI, artificial 366 00:21:49,960 --> 00:21:53,237 superintelligence. I don't know what it means. There's just an exponential. Just 367 00:21:53,320 --> 00:21:57,477 like we had an exponential with Moore's law, chips getting faster and faster 368 00:21:57,560 --> 00:22:01,317 until they could, you know, do any, you know, simple calculation, you know, 369 00:22:01,400 --> 00:22:04,997 faster than faster faster than any human. I think the models are just going 370 00:22:05,080 --> 00:22:07,717 to get more and more capable at everything. Every few months we release 371 00:22:07,800 --> 00:22:12,117 a new model, gets better at coding, it gets better at science, you know, now 372 00:22:12,200 --> 00:22:15,317 models are routinely winning, you know, math high school math olympiads, they're 373 00:22:15,400 --> 00:22:18,757 moving on to college math olympiads, they're starting to do new new new 374 00:22:18,840 --> 00:22:23,637 mathematics. For the first time I've had internal people at Anthropic say, I 375 00:22:23,720 --> 00:22:26,757 don't write any code anymore. I don't write I don't open up an editor and and 376 00:22:26,840 --> 00:22:28,837 and and write code. I just let Claude 377 00:22:28,920 --> 00:22:30,757 code write the first draft and and all I 378 00:22:30,840 --> 00:22:34,157 do is edit it. Um, we had never reached 379 00:22:34,240 --> 00:22:38,037 that point before and the drumbeat is just going to continue. And and I I I I 380 00:22:38,120 --> 00:22:41,477 I I don't think there's any privileged point around there's no point at which 381 00:22:41,560 --> 00:22:45,397 the models start to do something different. What what we're going to see 382 00:22:45,480 --> 00:22:48,717 in the future is just like we've going to we've seen in the past except more 383 00:22:48,800 --> 00:22:52,317 so. The models are just going to get more and more intellectually capable 384 00:22:52,400 --> 00:22:55,957 and, you know, the the revenue is going to keep adding zeros. 385 00:22:56,040 --> 00:23:00,477 Last couple policy questions. Um, you have been outspoken. We we spoke to uh 386 00:23:00,560 --> 00:23:04,037 the president of Taiwan Yes. >> Um, 387 00:23:04,120 --> 00:23:07,517 you have been outspoken about the idea that we should not be selling Nvidia 388 00:23:07,600 --> 00:23:11,317 chips, for example, the the most advanced chips to China. By the way, 389 00:23:11,400 --> 00:23:14,397 it's interesting that you now have a partnership with Nvidia. Uh, Jensen 390 00:23:14,480 --> 00:23:17,397 Huang, who's also been here, was not so happy with you when you made those 391 00:23:17,480 --> 00:23:20,837 comments. Do you have a new view on that? My my my 392 00:23:20,920 --> 00:23:24,277 view hasn't changed. So I I definitely will say, and this has this has always 393 00:23:24,360 --> 00:23:27,877 been the case, you know, I have an enormous amount of respect for Jensen 394 00:23:27,960 --> 00:23:31,597 and for Nvidia. You know, Jensen is an immigrant who came to the the US with 395 00:23:31,680 --> 00:23:34,837 nothing. He built the most, you know, the the most valuable company in the 396 00:23:34,920 --> 00:23:37,117 world. This this isn't personal. This is 397 00:23:37,200 --> 00:23:39,317 a policy question. This is a question of 398 00:23:39,400 --> 00:23:44,197 how to def how best to defend our national security. And there my view 399 00:23:44,280 --> 00:23:47,877 hasn't changed. It's it's it's the following, which is if we go back to my 400 00:23:47,960 --> 00:23:52,437 picture of the models getting smarter and smarter as we continue to improve 401 00:23:52,520 --> 00:23:57,237 them. A phrase I've used in an essay I wrote a year and a half ago was, 402 00:23:57,320 --> 00:24:00,197 eventually the models are going to get to the point where they look like a 403 00:24:00,280 --> 00:24:02,557 country of geniuses in a data center. 404 00:24:02,640 --> 00:24:05,837 And so once we get to that point, think 405 00:24:05,920 --> 00:24:07,557 about what that country of geniuses in a 406 00:24:07,640 --> 00:24:10,597 data center can do and which existing 407 00:24:10,680 --> 00:24:13,077 country on earth it's plopped down in. 408 00:24:13,160 --> 00:24:17,877 Um, if it's plopped down in an authoritarian country, I feel like they 409 00:24:17,960 --> 00:24:19,637 can outsmart us in every way. 410 00:24:19,720 --> 00:24:23,317 Intelligence, defense, economic value, 411 00:24:23,400 --> 00:24:27,317 R&D, you know, and I worry that they'll be able to 412 00:24:27,400 --> 00:24:30,837 oppress their own people, that they'll be able to have a perfect surveillance 413 00:24:30,920 --> 00:24:33,517 state. And so I have always felt that we 414 00:24:33,600 --> 00:24:37,957 need to have the advantage here. And this is a national security issue. 415 00:24:38,040 --> 00:24:40,477 Right? Some people are saying this is an 416 00:24:40,560 --> 00:24:45,237 they they they they they say this is an economic issue, that it's an analogy to 417 00:24:45,320 --> 00:24:49,357 like the internet or 5G. We need to diffuse our stack like we needed to beat 418 00:24:49,440 --> 00:24:53,037 Huawei in in, you know, in telecommunications. 419 00:24:53,120 --> 00:24:54,877 You know, I don't I don't see it that 420 00:24:54,960 --> 00:24:57,997 way. I think we're building a 421 00:24:58,080 --> 00:25:00,557 growing and singular capability that has 422 00:25:00,640 --> 00:25:03,357 singular national security implications 423 00:25:03,440 --> 00:25:06,237 and democracies need to get there first. 424 00:25:06,320 --> 00:25:09,237 It is absolutely it is absolutely an imperative. 425 00:25:09,320 --> 00:25:09,757 >> Do you think it could happen? 426 00:25:09,840 --> 00:25:12,357 >> we if we sell these chips to China, that 427 00:25:12,440 --> 00:25:16,037 just makes it more likely they will get there first. It's common sense. 428 00:25:16,120 --> 00:25:17,477 >> Okay, but do you think that could happen 429 00:25:17,560 --> 00:25:20,637 here? So we had Alex Karp here earlier 430 00:25:20,720 --> 00:25:23,717 and there has been lots of worries about 431 00:25:23,800 --> 00:25:25,397 surveillance here. Talk about it in a 432 00:25:25,480 --> 00:25:30,877 democracy, right? What what is your concern there? Um, I should say by the 433 00:25:30,960 --> 00:25:33,597 way, there was a period of time where you called the president, this was 434 00:25:33,680 --> 00:25:38,477 before he was the president, a feudal warlord at one point. Um, 435 00:25:38,560 --> 00:25:42,637 so how do you think about the president today, America today and this idea that 436 00:25:42,720 --> 00:25:44,957 AI and surveillance could come together? 437 00:25:45,040 --> 00:25:48,757 Yeah, so so look, I I want to I want to say over and over again, the you know, 438 00:25:48,840 --> 00:25:53,157 that I I you know, I think that the the tendency to drag this down into being 439 00:25:53,240 --> 00:25:57,397 about specific personalities and and, you know, and and and and specific 440 00:25:57,480 --> 00:26:00,637 fights I think is not helpful here. We should really think we should really 441 00:26:00,720 --> 00:26:05,037 think at a at a policy level here. And it's not about one administration, it's 442 00:26:05,120 --> 00:26:08,997 not about another administration. We should have principles here. And and I 443 00:26:09,080 --> 00:26:13,117 think that the principle I would give that that I think is is is very 444 00:26:13,200 --> 00:26:14,757 important is actually it can happen 445 00:26:14,840 --> 00:26:18,077 anywhere, right? You know, a a we should 446 00:26:18,160 --> 00:26:22,597 we should worry about concentration of power in democracies, not as much as we 447 00:26:22,680 --> 00:26:25,077 worry about it in authoritarian um, uh 448 00:26:25,160 --> 00:26:26,877 states, but you know, we we need to make 449 00:26:26,960 --> 00:26:30,037 sure that the technology is governed in 450 00:26:30,120 --> 00:26:34,637 in, you know, in a way that that allows people to participate, that gives people 451 00:26:34,720 --> 00:26:38,357 basic rights. And so the formulation I've always given when I think about how 452 00:26:38,440 --> 00:26:41,917 to apply these models for national security is I think we should 453 00:26:42,000 --> 00:26:44,517 aggressively use them in in every 454 00:26:44,600 --> 00:26:47,637 possible way except in the ways that 455 00:26:47,720 --> 00:26:51,597 would make us more like our authoritarian adversaries, right? We 456 00:26:51,680 --> 00:26:55,757 need to beat them, but we need to not do the things that would cause us to become 457 00:26:55,840 --> 00:26:58,837 them. Um, that is the one constraint we should observe. 458 00:26:58,920 --> 00:27:00,077 >> Okay, let me ask you a separate question 459 00:27:00,160 --> 00:27:03,397 and maybe you'll say it's a fight, but you can take it out of out of the the 460 00:27:03,480 --> 00:27:07,557 the person if you if you want, which is you have been very vocal about your 461 00:27:07,640 --> 00:27:11,197 concerns about the chip issue, but also what could 462 00:27:11,280 --> 00:27:15,877 happen to jobs, uh regulating reg- regulating this technology so it doesn't 463 00:27:15,960 --> 00:27:17,877 do bad things or other things like that. 464 00:27:17,960 --> 00:27:20,957 Uh, David Sacks, uh who works as the AI 465 00:27:21,040 --> 00:27:23,117 czar at the White House, uh said this 466 00:27:23,200 --> 00:27:26,157 about you. He says that Anthropic is 467 00:27:26,240 --> 00:27:31,397 running a sophisticated regulatory capture strategy based on fearmongering. 468 00:27:31,480 --> 00:27:32,997 It is principally responsible for the 469 00:27:33,080 --> 00:27:35,757 state regulatory frenzy that is damaging 470 00:27:35,840 --> 00:27:40,357 the startup ecosystem. Again, I don't think this should be about specific 471 00:27:40,440 --> 00:27:44,397 individuals, right? This isn't about any particular administration. This is about 472 00:27:44,480 --> 00:27:48,717 This is about policy questions. I mean, you know, going all the way back to 20 473 00:27:48,800 --> 00:27:52,957 2016 I've written papers about AI right before I even had a company, right? 474 00:27:53,040 --> 00:27:56,677 Before there there there even, you know, could have been a plan around anything 475 00:27:56,760 --> 00:27:59,357 like regulatory capture. And and by the 476 00:27:59,440 --> 00:28:01,517 way, almost all the AI regulation that 477 00:28:01,600 --> 00:28:03,957 we've supported has has exemptions for 478 00:28:04,040 --> 00:28:06,677 small players, right? The main um, AI 479 00:28:06,760 --> 00:28:10,357 bill we supported, SB 53, um, you know, 480 00:28:10,440 --> 00:28:14,517 doesn't even apply at all to startups with under $500 million in revenue. So 481 00:28:14,600 --> 00:28:18,717 we've been very careful about this. And you know, if there's one point again, I 482 00:28:18,800 --> 00:28:22,357 want to say again, I think people should focus on the policy. You can throw out 483 00:28:22,440 --> 00:28:26,117 these accusations and they don't match reality at all. They don't match the 484 00:28:26,200 --> 00:28:29,997 reality of the bills we've actually supported. They don't match the reality 485 00:28:30,080 --> 00:28:33,117 >> these two worlds right now, which is by the way, you know, Andreessen Horowitz 486 00:28:33,200 --> 00:28:36,997 and and others are are have a have one super PAC. You guys are building another 487 00:28:37,080 --> 00:28:40,557 super PAC to approach regulation of this industry 488 00:28:40,640 --> 00:28:44,557 completely differently. And the question is why? What do you see that they don't? 489 00:28:44,640 --> 00:28:47,317 >> So so I, you know, again, I I want to 490 00:28:47,400 --> 00:28:49,997 keep this at the policy level. What what 491 00:28:50,080 --> 00:28:55,477 how I see this technology, I I I am concerned and I can understand where 492 00:28:55,560 --> 00:28:58,477 folks are coming from, but I am concerned that there are some who see 493 00:28:58,560 --> 00:29:01,797 this technology as analogous to previous 494 00:29:01,880 --> 00:29:06,077 technological revolutions, as being like the internet, as being like 495 00:29:06,160 --> 00:29:09,837 telecommunications where yes, there are some issues, but, you know, the the 496 00:29:09,920 --> 00:29:13,677 market will figure it out, which I think was maybe a more reasonable view in in 497 00:29:13,760 --> 00:29:17,597 these previous technological revolutions. I think those who are 498 00:29:17,680 --> 00:29:20,557 closest to AI don't feel this way. If 499 00:29:20,640 --> 00:29:22,837 you pull the actual researchers who who 500 00:29:22,920 --> 00:29:26,397 work on AI, not investors who invest in 501 00:29:26,480 --> 00:29:31,717 some AI application companies, not general tech commentators who, you know, 502 00:29:31,800 --> 00:29:35,037 think they know something about AI, but the actual people who are building the 503 00:29:35,120 --> 00:29:38,717 technology, they're excited about the potential, but 504 00:29:38,800 --> 00:29:42,037 they're also worried. They're worried about the national security risks. 505 00:29:42,120 --> 00:29:45,477 They're worried about alignment of the models. They're worried about the 506 00:29:45,560 --> 00:29:47,797 economic impacts of the models. And for 507 00:29:47,880 --> 00:29:52,717 example, the idea that we would put a moratorium on all all all kind of 508 00:29:52,800 --> 00:29:56,837 regulation or all state regulation without a federal, you know, with with 509 00:29:56,920 --> 00:30:02,157 that without without a federal framework for 10 years, which, you know, was 510 00:30:02,240 --> 00:30:05,517 attempted in the summer and I think it was just attempted this last week again 511 00:30:05,600 --> 00:30:09,237 and it failed because it was very unpopular because even the average 512 00:30:09,320 --> 00:30:11,557 person understands that this is a new 513 00:30:11,640 --> 00:30:16,677 and and powerful technology. And so, you know, I think we I I am more than any 514 00:30:16,760 --> 00:30:19,797 you know, I I am maybe the most optimistic about the 515 00:30:19,880 --> 00:30:24,397 upsides, right? I wrote this whole essay Machines of Loving Grace where I said AI 516 00:30:24,480 --> 00:30:26,357 is going to, you know, extend might even 517 00:30:26,440 --> 00:30:31,477 extend the human lifespan to 150 10 years after we get the the country of 518 00:30:31,560 --> 00:30:35,117 geniuses in the data center because we'll have a virtual biologist that can 519 00:30:35,200 --> 00:30:38,717 make discoveries much faster than we can. That that, you know, it could 520 00:30:38,800 --> 00:30:43,117 drive, you know, economic growth to, you know, five or five or 10%. I'm 521 00:30:43,200 --> 00:30:46,477 incredibly optimistic about the technology. Frankly, much more 522 00:30:46,560 --> 00:30:50,037 optimistic than some of people who describe themselves as boosters of the 523 00:30:50,120 --> 00:30:53,317 technology. Um, but but nothing that 524 00:30:53,400 --> 00:30:55,997 powerful doesn't have a significant 525 00:30:56,080 --> 00:30:58,357 number of downsides. And and we as a 526 00:30:58,440 --> 00:31:01,397 society, as a polity, need to think 527 00:31:01,480 --> 00:31:05,837 ahead about those downsides. Saying that for 10 years we won't regulate that 528 00:31:05,920 --> 00:31:09,277 technology, it's like it's like saying I'm driving a car, I'm going to rip out 529 00:31:09,360 --> 00:31:12,837 the steering wheel because I don't need to steer for 10 years. Okay, so here's 530 00:31:12,920 --> 00:31:14,677 the the question about the downside then. 531 00:31:14,760 --> 00:31:17,557 >> One of the downsides, beyond hacking and everything else that I know you worry 532 00:31:17,640 --> 00:31:21,637 about, is jobs. You spoke about it on 60 Minutes recently. But I want to know not 533 00:31:21,720 --> 00:31:24,837 just that you you think that there's a good chance 534 00:31:24,920 --> 00:31:27,757 and I don't want to put words in your mouth that it could be, you know, half 535 00:31:27,840 --> 00:31:31,037 of all entry-level jobs get lost. I want to know what you think should be done 536 00:31:31,120 --> 00:31:35,197 about it. Absolutely. So, you know, I I I think at the end of the day I warn 537 00:31:35,280 --> 00:31:40,157 about these things not to be a prophet of doom, but because warning about them 538 00:31:40,240 --> 00:31:42,037 is the first step towards solving them. 539 00:31:42,120 --> 00:31:45,797 And if we don't warn about them, then we'll just uh, you know, we'll we'll 540 00:31:45,880 --> 00:31:49,877 we'll just kind of blindly walk into a landmine and blow us up. Um, if we warn 541 00:31:49,960 --> 00:31:53,517 about them, if we see the landmine, we can walk around it and we can avoid it. 542 00:31:53,600 --> 00:31:56,517 So, I have been thinking a lot about these ideas. I've been thinking about 543 00:31:56,600 --> 00:31:59,917 them inside Anthropic where we where, you know, Claude is starting to write a 544 00:32:00,000 --> 00:32:03,397 lot of our code and we're thinking about how the jobs change. So, I think there's 545 00:32:03,480 --> 00:32:07,757 several levels of it that maybe go from short-term to long-term or or just kind 546 00:32:07,840 --> 00:32:11,877 of requiring more and more of the resources of society to happen. I think 547 00:32:11,960 --> 00:32:15,677 some of it can happen in the private sector and even in our relationships 548 00:32:15,760 --> 00:32:17,997 with customers. Every customer we work 549 00:32:18,080 --> 00:32:19,877 with has the following trade-off and 550 00:32:19,960 --> 00:32:23,037 it's not either-or. They can, um, you 551 00:32:23,120 --> 00:32:28,117 know, they can increase efficiency by basically having AIs do what what humans 552 00:32:28,200 --> 00:32:29,437 used to do. And there's plenty of that. 553 00:32:29,520 --> 00:32:33,517 Things like, you know, insurance claims processing or know your customer, whole 554 00:32:33,600 --> 00:32:37,357 workflows that can just be done end-to-end via AI and I think we'll need 555 00:32:37,440 --> 00:32:41,317 a lot less humans for them. It it will increase efficiency, it'll save costs, 556 00:32:41,400 --> 00:32:45,717 do the same thing for lower cost and much less people needed. But you can 557 00:32:45,800 --> 00:32:49,437 also do things where you can create a lot of new value. And even in cases 558 00:32:49,520 --> 00:32:54,357 where AI does 90% of the job, not 100% but 90%, the humans can be 10 times more 559 00:32:54,440 --> 00:32:57,917 leverage and sometimes you need 10 times more of them to do 100 times what you 560 00:32:58,000 --> 00:33:02,317 did before because it's so efficient and valuable. And so encouraging companies 561 00:33:02,400 --> 00:33:05,797 to do as much of the second relative to the first. We know they're going to do 562 00:33:05,880 --> 00:33:08,037 the first. We're not trying to stop them from doing the first, but if they can do 563 00:33:08,120 --> 00:33:11,917 more of the second than the first, maybe more jobs can be created than than get 564 00:33:12,000 --> 00:33:13,997 lost. >> mean we need government incentives? Is that a policy 565 00:33:14,080 --> 00:33:18,117 >> So so uh again, that's that's that's level one. Level two is the involvement 566 00:33:18,200 --> 00:33:21,917 of the government. I don't see retraining programs as a panacea, but I 567 00:33:22,000 --> 00:33:25,117 think we're going to need to do some form of that. Companies are going to do 568 00:33:25,200 --> 00:33:28,917 it, companies are going to have to work with governments to do it. And I but I I 569 00:33:29,000 --> 00:33:33,157 I do think fiscally at some point the government is going to need to step in. 570 00:33:33,240 --> 00:33:35,637 I don't know if that's tax policy, but 571 00:33:35,720 --> 00:33:40,317 but this world of fast growth, right? We did this report where we said even 572 00:33:40,400 --> 00:33:45,557 current models it looks like they will increase productivity by 1.6% 573 00:33:45,640 --> 00:33:48,837 a year. That's almost, you know, almost a doubling of productivity. And the 574 00:33:48,920 --> 00:33:52,077 models are getting better and better. So I think we're going to get up to five 575 00:33:52,160 --> 00:33:54,717 five percent a year, maybe 10% a year. 576 00:33:54,800 --> 00:33:57,397 That's a big pie. That's a big pie that 577 00:33:57,480 --> 00:34:01,757 we can we can give out to the people who are who are not such fortunate 578 00:34:01,840 --> 00:34:05,597 beneficiaries, right? If the wealth concentrates, there there there there 579 00:34:05,680 --> 00:34:08,677 there really is a big pie here. So I think the government is going to need to 580 00:34:08,760 --> 00:34:11,237 have some role here. That's level two 581 00:34:11,320 --> 00:34:13,757 and I think level three is over the long 582 00:34:13,840 --> 00:34:16,637 run the structure of a society that has 583 00:34:16,720 --> 00:34:21,597 built powerful AI is just going to have to be different. If we if we go back to 584 00:34:21,680 --> 00:34:25,317 John Maynard Keynes, right? Economic possibilities for our grandchildren, 585 00:34:25,400 --> 00:34:28,997 right? He invented this idea of technological unemployment. He suggested 586 00:34:29,080 --> 00:34:33,397 that maybe his grandchildren would only have to work 15 or 20 hours hours a 587 00:34:33,480 --> 00:34:36,917 week, right? That's like a different way of structuring the society. Some people 588 00:34:37,000 --> 00:34:41,517 will always want to work, you know, as hard as possible. Um, you know, there 589 00:34:41,600 --> 00:34:45,517 there always be segments of society who who want to do that. But, you know, can 590 00:34:45,600 --> 00:34:50,117 can can we have a world where work doesn't for many people doesn't need to 591 00:34:50,200 --> 00:34:53,757 have the centrality that it does that that people find their their locus of 592 00:34:53,840 --> 00:34:56,797 meaning elsewhere or work is about is 593 00:34:56,880 --> 00:34:58,997 about different things. It's more about 594 00:34:59,080 --> 00:35:01,837 fulfillment than it is about economic 595 00:35:01,920 --> 00:35:04,717 economic survival. Um, there there's so 596 00:35:04,800 --> 00:35:06,997 many possibilities here. I think society 597 00:35:07,080 --> 00:35:10,317 is flexible and and society can I'm not 598 00:35:10,400 --> 00:35:14,837 suggesting anything top-down. I think society needs to restructure itself. We 599 00:35:14,920 --> 00:35:19,237 all need to figure out how to operate in the in in the post in the post-AGI age. 600 00:35:19,320 --> 00:35:23,597 So I think those three levels will go from fast and and easy for it for, you 601 00:35:23,680 --> 00:35:27,437 know, individual companies to do to requiring a lot of consensus and and and 602 00:35:27,520 --> 00:35:31,077 very slow to do. But over the years we're we're going to need to do all 603 00:35:31,160 --> 00:35:34,357 three of these things. Uh, Dario, I hope you come back so we can have a 604 00:35:34,440 --> 00:35:38,197 conversation as we have to do all of those things to figure out what comes 605 00:35:38,280 --> 00:35:41,557 next. Want to thank you for a fantastic conversation. Thank you, Andrew. Thank 606 00:35:41,640 --> 00:35:44,840 you. Thank you so much. Appreciate it.