WEBVTT - Dario & Daniela Amodei: "This Is Gonna Determine the Fate of Nations and the Future of Humanity” 0:00:09.930 --> 0:00:14.270 The Builders is a production of Kaleidoscope and iHeart, presented 0:00:14.430 --> 0:00:19.150 by Ramp. I'm Walter Isaacs, and welcome to The Builders, 0:00:19.650 --> 0:00:23.970 my new series of interviews featuring the most consequential minds 0:00:24.390 --> 0:00:29.510 shaping our future. A person's biography tells you what values, 0:00:29.910 --> 0:00:34.520 what drives, sometimes what demons are bred into them. It's 0:00:34.560 --> 0:00:36.059 particularly interesting in an. 0:00:36.100 --> 0:00:36.840 Era of A.I. 0:00:37.420 --> 0:00:41.840 Because we're watching the biographies of AI systems happen, we 0:00:41.900 --> 0:00:45.370 have to make sure that we raise them carefully so 0:00:45.420 --> 0:00:48.810 that they don't reflect the dark sides of humanity. 0:00:49.470 --> 0:00:55.390 Alarming new details about robots from OpenAI hacking into another company. Anthropic, 0:00:55.410 --> 0:00:57.980 the maker of Claude, says that it blocked several potential 0:00:58.020 --> 0:01:00.440 plots this year by scientists who were using its AI 0:01:00.480 --> 0:01:01.960 models to conduct. 0:01:01.700 --> 0:01:04.319 Research that could have led to the creation of Bioweapon. 0:01:05.480 --> 0:01:08.820 And so I've begun to feel that the moral compass 0:01:08.860 --> 0:01:12.180 of these builders is the most important thing you can 0:01:12.360 --> 0:01:16.860 know about them, to understand what values we're imbuing into 0:01:16.920 --> 0:01:22.449 these artificial agents. So today, I'm speaking with Dario and 0:01:22.510 --> 0:01:24.010 Daniela Amadei. 0:01:24.150 --> 0:01:28.230 CEO and co-founder of AI giant Anthropic, Dario Amadei, with 0:01:28.270 --> 0:01:30.630 an urgent warning about artificial intelligence. 0:01:30.920 --> 0:01:34.840 Dario and Daniela are the brother and sister duo who 0:01:34.920 --> 0:01:37.529 run Anthropic, the maker of Claude. 0:01:38.350 --> 0:01:40.550 And for a while, Anthropic. 0:01:39.970 --> 0:01:43.530 Has been seen as the good guys in the AI race. 0:01:44.050 --> 0:01:47.670 But in the past few months, we've seen worrying developments. 0:01:48.110 --> 0:01:52.030 Anthropic researcher Jacob Coxon resigned from his post, sounding the 0:01:52.090 --> 0:01:53.890 alarm of what he called out of control. In a 0:01:53.890 --> 0:01:57.250 now viral post, Coxon said the companies are, quote, gambling 0:01:57.290 --> 0:02:00.930 with our lives. The people building AI earnestly believe that 0:02:01.010 --> 0:02:03.580 it could kill us all by the end of the decade. 0:02:04.360 --> 0:02:07.860 All of this, along with reports of a possible IPO, 0:02:08.180 --> 0:02:12.380 has begun to make me wonder, how deep does Anthropic's 0:02:12.480 --> 0:02:17.260 commitment to humanity really run? And what happens to those principles? 0:02:18.030 --> 0:02:19.380 When they get in the way of winning. 0:02:20.180 --> 0:02:24.060 Anthropic has reportedly warned investors that its technology may pose 0:02:24.280 --> 0:02:28.280 existential risk to humanity in its IPO prospectus. And you 0:02:28.340 --> 0:02:31.960 might then wonder, with a company warning of this, why 0:02:32.180 --> 0:02:33.980 it's still planning to go ahead with the. 0:02:33.980 --> 0:02:34.680 IPO at all. 0:02:34.700 --> 0:02:38.760 The public sale could value Anthropic at $ 2 trillion. 0:02:41.139 --> 0:02:44.420 I chatted with Daniela and Dario in July of 2026. 0:02:44.400 --> 0:02:49.660 At the time, I didn't know just how prescient this 0:02:49.740 --> 0:02:55.950 conversation would be. Here's my interview with Dario and Daniela Amadei. 0:02:58.650 --> 0:02:59.649 Welcome to The Builders. 0:03:00.110 --> 0:03:01.290 Thank you so much for having us. 0:03:02.050 --> 0:03:04.780 You know, Steve Jobs, whenever he launched a product, always 0:03:04.810 --> 0:03:07.200 had that street sign of, you know, the humanities and 0:03:07.260 --> 0:03:11.500 technology standing at that intersection. And when you were growing up, Daniela, 0:03:11.800 --> 0:03:12.820 you were a flautist. 0:03:13.520 --> 0:03:14.500 You loved the arts. 0:03:14.990 --> 0:03:18.269 And Dario, you always a scientist. Did that help you gel? 0:03:18.330 --> 0:03:21.330 And do you think that combination is important? 0:03:22.169 --> 0:03:25.780 Yeah, I think that combination is a lot of what's 0:03:25.860 --> 0:03:30.239 behind the company, right? One theory I have is that 0:03:30.820 --> 0:03:33.820 running a company is too big a task for any 0:03:33.940 --> 0:03:36.620 one person. The range of skills that it involves, it 0:03:36.680 --> 0:03:39.350 is not possible for any single person to have. So 0:03:39.390 --> 0:03:42.190 you kind of have to assemble a full complement of 0:03:42.290 --> 0:03:45.600 skills across the top, like 10 or 15 people. 0:03:45.380 --> 0:03:46.700 In the company or something like that. 0:03:46.740 --> 0:03:50.500 And every very successful company splits it up in different 0:03:50.540 --> 0:03:53.370 ways and solves the problem in different ways. And I 0:03:53.430 --> 0:03:56.430 think with the sciences and the humanities thing. 0:03:56.470 --> 0:03:58.820 You know, Daniela and I kind of embody that and 0:03:58.860 --> 0:03:59.860 split that. Yeah. 0:04:00.370 --> 0:04:02.090 And I would say, you know, it's interesting, like, there 0:04:02.120 --> 0:04:06.280 was an off-site we did in maybe 2023 where, like, 0:04:06.860 --> 0:04:08.740 we sort of were talking about just, like, what does 0:04:08.760 --> 0:04:10.340 it feel like to work at Anthropic? Like, how do 0:04:10.360 --> 0:04:13.040 we kind of embody what our culture is? And we 0:04:13.140 --> 0:04:16.620 actually said, you know, it's an unusual blend of science 0:04:16.690 --> 0:04:19.349 and humanism. And it was only later that I'm like, 0:04:19.390 --> 0:04:21.430 that also is how you would describe me and Dario. 0:04:21.529 --> 0:04:23.450 Well, when you were kids, did you know you wanted 0:04:23.490 --> 0:04:24.190 to work together? 0:04:24.850 --> 0:04:25.090 Yes. 0:04:25.130 --> 0:04:28.270 Yes. I don't think either of us imagined... that we 0:04:28.290 --> 0:04:29.370 would start a company. 0:04:29.610 --> 0:04:33.550 That's not really how either of us thought of our 0:04:33.610 --> 0:04:36.310 lives or about what we wanted to do. I had 0:04:36.370 --> 0:04:39.560 always wanted to be a scientist. Daniela can speak to 0:04:39.620 --> 0:04:44.540 her aspirations. But I actually found companies and business to 0:04:44.580 --> 0:04:45.840 be of very little interest to me. 0:04:46.100 --> 0:04:50.200 Yeah. I would say Dario and I were very unusual 0:04:50.360 --> 0:04:53.219 in that from a really early age, we just genuinely 0:04:53.380 --> 0:04:55.640 really liked each other. And I agree with Dario. I 0:04:55.680 --> 0:04:58.440 don't think starting a company was ever on our bingo card. 0:04:58.520 --> 0:05:01.580 Neither of us were particularly company-oriented. 0:05:01.610 --> 0:05:02.950 I was always really. 0:05:02.810 --> 0:05:05.510 Interested in, I think I would call myself a classic do-gooder. 0:05:05.930 --> 0:05:09.370 So I worked in international development. I felt really strongly 0:05:09.430 --> 0:05:13.469 about helping people everywhere in the world have equal access 0:05:13.529 --> 0:05:16.940 and opportunity. And I became a vegetarian when I was 0:05:16.980 --> 0:05:18.640 like quite young. It's like the only thing that I 0:05:18.660 --> 0:05:21.660 did first. Dario did everything first because he's older, but 0:05:21.800 --> 0:05:24.020 I became a vegetarian and then he became a vegetarian. 0:05:24.620 --> 0:05:27.850 And Dario, tell me how you and Daniela divide up 0:05:27.890 --> 0:05:29.729 the roles and then how you fill it in with 0:05:29.810 --> 0:05:30.550 others in the team. 0:05:31.650 --> 0:05:35.360 Yeah, so I think I can function best when I 0:05:35.420 --> 0:05:39.140 think about the big picture strategy and the research side 0:05:39.160 --> 0:05:42.710 of things. I'm best at laying out a vision, seeing 0:05:42.790 --> 0:05:45.849 what comes next. I have found that when I'm freed 0:05:45.890 --> 0:05:48.750 up to focus entirely on that, that is the thing 0:05:48.790 --> 0:05:50.510 that has the most value for the company. 0:05:50.950 --> 0:05:53.350 And then running the company day to. 0:05:53.310 --> 0:05:58.190 Day, recruiting the people, just making sure that the machine doesn't, 0:05:58.510 --> 0:06:01.100 you know, the wheels don't fall off the machine while 0:06:01.140 --> 0:06:03.080 I'm trying to figure out what comes next. 0:06:03.560 --> 0:06:04.540 Daniela does this. 0:06:04.680 --> 0:06:08.900 And I think this division of labor allows both things 0:06:08.980 --> 0:06:12.700 to be done to a much higher standard than either 0:06:12.760 --> 0:06:15.840 of them would be done if we were distracted by 0:06:15.860 --> 0:06:16.800 having to do the other thing. 0:06:17.080 --> 0:06:19.140 Our chief people officer, who knows both of us really well, 0:06:19.180 --> 0:06:21.529 is in a lot of conversations with us, the way 0:06:21.570 --> 0:06:23.529 she put it is that Dario makes us ambitious and 0:06:23.610 --> 0:06:25.950 I make us successful. And I thought that was a 0:06:26.110 --> 0:06:29.990 really good encapsulation, which is, I think Anthropic could be 0:06:30.130 --> 0:06:33.830 like a good, successful business, but not change the world 0:06:33.910 --> 0:06:36.050 if we didn't have Dario. And I think with Dario, 0:06:36.070 --> 0:06:37.810 we would have the perfect vision, but our ability to 0:06:37.890 --> 0:06:41.250 execute would be, it would just be more complicated. And 0:06:41.270 --> 0:06:44.310 so I think there's something special about A, the fact 0:06:44.330 --> 0:06:47.010 that we know each other really well. B, it's extremely 0:06:47.070 --> 0:06:51.270 convenient that our skills are like 100% complementary or maybe 0:06:51.310 --> 0:06:53.760 not 100%, but I would say close to 100% complementary 0:06:53.779 --> 0:06:56.440 in this fashion. And this is not even to mention 0:06:56.480 --> 0:06:58.940 the fact that we have five co-founders along with the 0:06:58.980 --> 0:07:03.159 two of us who also own like incredibly important parts 0:07:03.220 --> 0:07:06.640 of the business, the culture, like the organization. It's hard 0:07:06.660 --> 0:07:08.219 for me to imagine how you would do this with 0:07:08.260 --> 0:07:09.200 like just the two of us. 0:07:09.790 --> 0:07:12.650 So go back to growing up, and you have these 0:07:12.730 --> 0:07:17.340 extraordinary parents, and they help teach you right from wrong. 0:07:17.440 --> 0:07:18.820 How do they do that? 0:07:19.980 --> 0:07:22.820 I don't know that it's something that was ever, you know, 0:07:22.840 --> 0:07:24.960 sometimes I hear people say like, you know, we covered 0:07:25.580 --> 0:07:28.490 different forms of philosophical ethics at the dinner table. 0:07:28.530 --> 0:07:30.270 That was not, that was not how it was done. 0:07:30.290 --> 0:07:32.650 My mom does have a philosophy degree, but she didn't 0:07:32.690 --> 0:07:34.550 subject us to talking about Kant or anything like that. 0:07:35.050 --> 0:07:37.580 I think it was much more like inherent in terms 0:07:37.620 --> 0:07:41.960 of values, like just the belief that people everywhere in 0:07:41.980 --> 0:07:46.700 the deserving of opportunity and dignity and rights and respect. 0:07:46.980 --> 0:07:48.800 I think there was something special about the feeling that 0:07:48.840 --> 0:07:51.540 like we could be whoever we wanted to be as 0:07:51.580 --> 0:07:53.870 long as there was this kind of faith and belief 0:07:53.900 --> 0:07:56.770 that like, hey, part of what it means to be 0:07:56.790 --> 0:07:59.210 a successful human is to be a good person and 0:07:59.230 --> 0:08:00.430 to be kind to other people. 0:08:00.450 --> 0:08:03.750 Yeah, it felt implicit. There wasn't any kind of, you know, 0:08:03.770 --> 0:08:05.130 this is what you should do or this is how 0:08:05.150 --> 0:08:05.970 you live in the world. 0:08:06.930 --> 0:08:09.310 Well, one reason I ask, as you might suspect, is 0:08:09.370 --> 0:08:13.100 now you've got a really important kid named Claude And 0:08:13.140 --> 0:08:16.330 to what extent does that way that you grew up 0:08:16.770 --> 0:08:19.770 help you, inform you about how you teach Claude to 0:08:19.790 --> 0:08:20.770 know right from wrong? 0:08:21.290 --> 0:08:24.630 Yeah, you know, we put out a constitution for Claude, 0:08:24.830 --> 0:08:26.790 I think it was a few months ago. And I 0:08:26.830 --> 0:08:30.140 think there are some interesting lessons in it, which we've 0:08:30.220 --> 0:08:34.900 learned from several years of trying to train Claude in 0:08:34.960 --> 0:08:35.920 successive versions. 0:08:36.120 --> 0:08:39.360 I think early on, we had been imagining that. 0:08:39.240 --> 0:08:40.980 There would be various rules that, you know, you would 0:08:41.000 --> 0:08:43.870 tell Claude, you know, don't, Don't help someone commit a 0:08:43.910 --> 0:08:46.490 crime or don't make biological weapons or whatever. 0:08:46.970 --> 0:08:49.250 But it turns out that when you make this kind. 0:08:49.130 --> 0:08:51.890 Of discrete list of rules, you end up with this 0:08:51.929 --> 0:08:54.110 kind of spiky shape where you're trying to teach the 0:08:54.150 --> 0:08:58.410 model like 170 different rules that are kind of arbitrary. 0:08:58.840 --> 0:09:02.180 We found it's better to teach the model some basic 0:09:02.700 --> 0:09:06.740 principles like what it should prioritize, how it should reason 0:09:06.820 --> 0:09:08.460 about what the right thing to do is. 0:09:08.890 --> 0:09:10.650 Which, of course, I think that's the right way to 0:09:10.670 --> 0:09:11.960 teach people as well, right? 0:09:12.000 --> 0:09:15.060 Like, you know, if I just give someone a list 0:09:15.100 --> 0:09:17.500 of rules, thou shalt do this, thou shalt not do this. 0:09:17.840 --> 0:09:21.319 You know, I don't think that produces as robust a 0:09:21.340 --> 0:09:23.280 morality as I say, this is how you should think 0:09:23.340 --> 0:09:25.900 about things. You should learn how to think for yourself. 0:09:26.360 --> 0:09:28.420 And these are the principles on which you should think, right? 0:09:28.460 --> 0:09:33.630 Principles of open inquiry, principles of care for other people, 0:09:33.730 --> 0:09:35.850 principles of fairness and justice. 0:09:35.950 --> 0:09:39.719 And then everything is derived from that. And that is, frankly, 0:09:39.840 --> 0:09:41.040 how Claude learns. 0:09:41.540 --> 0:09:44.420 We find that to be surprisingly effective because that allows 0:09:44.740 --> 0:09:47.760 Claude to generalize to new situations that it might never 0:09:47.820 --> 0:09:48.340 have imagined. 0:09:49.370 --> 0:09:51.750 Danielle, I can't help myself when you said your mom 0:09:51.790 --> 0:09:57.330 was an expert on Kant. Did you actually wrestle with philosophy? 0:09:57.690 --> 0:10:02.550 You know, what's interesting is the woman who is responsible 0:10:02.590 --> 0:10:06.380 for training Claude's character has a PhD in philosophy. 0:10:07.260 --> 0:10:10.150 It's just, it's unusual, right? You would think, oh, the person. 0:10:09.950 --> 0:10:12.309 That's training the sort of character of these models must 0:10:12.370 --> 0:10:15.590 be a computer scientist or somebody who's studied ML or 0:10:15.650 --> 0:10:18.929 AI for, you know, many decades. Amanda's technical and she 0:10:18.990 --> 0:10:21.410 like technically trains the model, but I think it is 0:10:21.460 --> 0:10:25.940 very felt in Claude's character when you communicate with it, 0:10:26.300 --> 0:10:30.900 that there's this kind of experience of something that's very wise, 0:10:30.980 --> 0:10:33.260 kind of the better angel of your nature trying to 0:10:33.360 --> 0:10:36.429 answer your questions, But there are like if you kind 0:10:36.450 --> 0:10:40.010 of poke Claude in these like difficult areas where maybe 0:10:40.090 --> 0:10:44.330 different philosophical traditions might conflict. I think that's where you 0:10:44.350 --> 0:10:48.390 especially feel, Amanda, because you'll sort of see like, oh, well, 0:10:48.730 --> 0:10:50.490 you know, you might want to consider things from like 0:10:50.530 --> 0:10:52.530 this point of view, but you could also consider it 0:10:52.550 --> 0:10:54.640 from this point of view. And I do think it's 0:10:54.860 --> 0:10:59.319 it's important because as the world evolves, begins to rely 0:10:59.400 --> 0:11:01.320 on these models more and more, you don't want to 0:11:01.340 --> 0:11:05.840 be locked into sort of one philosophical tradition because ethics 0:11:05.880 --> 0:11:09.480 and morality are quite complicated. And different people and different 0:11:09.520 --> 0:11:12.000 cultures might have different answers about how they would like 0:11:12.040 --> 0:11:12.660 you to behave. 0:11:14.860 --> 0:11:18.260 Don't go anywhere. There's a lot more right after this break. 0:11:29.320 --> 0:11:32.920 Dario, you were there at one of the seminal moments 0:11:32.980 --> 0:11:38.179 of modern times, which is when GPT-1 and 2 suddenly 0:11:38.280 --> 0:11:42.060 burst on the scene, surprising, perhaps even surprising you. You 0:11:42.080 --> 0:11:46.470 were head of research at OpenAI. What was that moment like? 0:11:47.670 --> 0:11:51.349 Yeah, so I wasn't very involved in GPT-1, but I 0:11:51.390 --> 0:11:53.570 was very involved in the idea that we should run 0:11:53.590 --> 0:11:59.850 a larger scale, more expensive, more ambitious approach experiment on 0:12:00.230 --> 0:12:02.790 the idea that if we ran these language models at 0:12:02.809 --> 0:12:05.010 a larger scale they would be better at language they 0:12:05.030 --> 0:12:06.670 would be more coherent. 0:12:06.570 --> 0:12:09.750 So it's almost like one of these moments in scientific. 0:12:09.290 --> 0:12:12.719 History where you're like you propose a definitive experiment to 0:12:12.760 --> 0:12:15.320 just make a very strange prediction and then if the 0:12:15.340 --> 0:12:17.500 prediction turns out to be correct you're like god this 0:12:17.679 --> 0:12:21.540 we're really we're really on to something so i see 0:12:21.580 --> 0:12:25.120 gpt2 as the first real test of the scaling laws 0:12:25.460 --> 0:12:30.170 hypothesis and What I found super interesting was the enormous 0:12:30.210 --> 0:12:32.060 divergence in reactions to it. 0:12:32.929 --> 0:12:34.569 And were you kind of surprised? 0:12:35.429 --> 0:12:38.829 I personally was super freaked out by it. 0:12:38.910 --> 0:12:41.609 Like, I looked at it and I said, my God, 0:12:41.650 --> 0:12:43.860 you just scale this thing up another factor of a 0:12:43.910 --> 0:12:46.880 million and it'll be better than humans at everything. Like, 0:12:47.260 --> 0:12:51.230 this is the defining thing. economic and scientific moment of 0:12:51.300 --> 0:12:53.230 the modern age. We just don't know it yet, right? 0:12:53.250 --> 0:12:56.190 This is going to determine the fate of nations and 0:12:56.270 --> 0:12:59.410 economies and the future of humanity. Like, you know, you 0:12:59.429 --> 0:13:02.950 peer into the crystal ball, you see the whole future unfurled. And, 0:13:03.070 --> 0:13:05.130 you know, it really scared me. I said, oh my God, 0:13:05.170 --> 0:13:06.969 what are the implications of that? Is this the right 0:13:07.030 --> 0:13:09.810 place to be developing it? Is, you know, what's going 0:13:09.830 --> 0:13:10.470 to happen next? 0:13:10.510 --> 0:13:11.970 What do we do with this? Who do we tell 0:13:12.010 --> 0:13:12.550 about this? 0:13:13.170 --> 0:13:15.770 And maybe 10% of the people around me had exactly 0:13:15.809 --> 0:13:16.690 the same reaction. 0:13:16.710 --> 0:13:18.110 And it was often the smartest ones. 0:13:18.110 --> 0:13:24.220 90% of the people around me, because GBD2' s capabilities in 0:13:24.280 --> 0:13:27.739 any particular area were not that impressive, right? 0:13:28.140 --> 0:13:28.620 It was the. 0:13:28.580 --> 0:13:32.270 Difference between GBD1 and GBD2, the generality, the fact that 0:13:32.290 --> 0:13:38.070 it could translate languages, And summarize a document and write 0:13:38.110 --> 0:13:40.770 a few lines of code here and there. But people 0:13:40.820 --> 0:13:41.860 would look at it and they would say in any 0:13:41.890 --> 0:13:42.800 particular area that. 0:13:42.740 --> 0:13:43.890 Doesn't seem so smart. 0:13:43.960 --> 0:13:46.900 And so many people just shrugged and they just dismissed it. 0:13:47.260 --> 0:13:49.380 And you get the usual comments from kind of the 0:13:49.440 --> 0:13:50.580 jaded people. 0:13:50.720 --> 0:13:52.320 Ah, tech people have done this before. 0:13:52.340 --> 0:13:55.040 Here are all the limitations of the models. And I 0:13:55.059 --> 0:13:56.580 just look at it and I say it's the same 0:13:56.620 --> 0:14:00.060 thing over and over again. The limitations of the models now, 0:14:00.400 --> 0:14:03.620 six months from now, the things that seemed hard, they're 0:14:03.640 --> 0:14:05.500 going to be easy. 12 months from now, the things 0:14:05.520 --> 0:14:07.660 that seemed impossible, the models are going to be doing 0:14:07.700 --> 0:14:12.280 them regularly. Everything we saw with GPT-2, we're seeing it 0:14:12.500 --> 0:14:15.179 over and over again in wider and wider swaths of 0:14:15.220 --> 0:14:19.020 the world. And the world is gradually finally being convinced. 0:14:19.360 --> 0:14:21.520 You know, we just announced that we're going to be 0:14:21.840 --> 0:14:25.000 developing drugs for, you know, neglected rare diseases. 0:14:25.080 --> 0:14:25.320 And so. 0:14:25.920 --> 0:14:27.520 I think you were quite polite about it, but I 0:14:27.560 --> 0:14:32.650 remember at the time there were some very well-respected computer 0:14:32.670 --> 0:14:37.190 scientists and AI and ML leaders who basically bullied some 0:14:37.230 --> 0:14:39.190 of the researchers who worked on it on Twitter and 0:14:39.210 --> 0:14:42.230 were like, this is stupid, this model is so incapable. 0:14:42.990 --> 0:14:44.430 It was like you were sort of a quack if 0:14:44.470 --> 0:14:47.430 you believed this was like a really significant moment. 0:14:47.460 --> 0:14:49.260 And that's all been memory hold. 0:14:49.340 --> 0:14:53.120 Like all the researchers who like were like, oh, this 0:14:53.180 --> 0:14:54.979 is this is pseudoscience. 0:14:55.040 --> 0:14:56.000 It doesn't make sense. 0:14:56.060 --> 0:14:58.740 Like all of them are applying for jobs at like 0:14:59.140 --> 0:15:01.860 the LLM companies. And like, you know, we never say 0:15:01.900 --> 0:15:04.970 it out loud to any particular one. But like, you know, 0:15:04.990 --> 0:15:07.150 we look at them and they're like, we remember what 0:15:07.170 --> 0:15:08.970 you were saying in 2019. 0:15:08.970 --> 0:15:11.250 You have no credibility on this matter. Right. 0:15:12.380 --> 0:15:15.760 So, Daniela, you were vice president of safety back then 0:15:15.780 --> 0:15:20.030 at OpenAI. And when you hit GPT-2, I assume you're 0:15:20.250 --> 0:15:23.650 like Dario, somebody who's almost freaked out about it. 0:15:23.690 --> 0:15:25.739 Did it freak you out about safety, too? 0:15:26.670 --> 0:15:29.750 I think, you know, there wasn't something about GBT2, the 0:15:29.770 --> 0:15:32.620 model itself, that made us, you know, concerned, right? We 0:15:32.640 --> 0:15:34.900 didn't look at that model and say, wow, we think 0:15:34.920 --> 0:15:38.260 this model has like cybersecurity implications or something, right? The 0:15:38.280 --> 0:15:41.520 model itself was in some ways kind of laughably wrong 0:15:41.560 --> 0:15:43.460 about a lot of things, right? It was still a very, 0:15:43.480 --> 0:15:46.490 I would say, harmless model. But I think the thing 0:15:46.850 --> 0:15:50.090 that Dario mentioned that I also felt was a sense 0:15:50.130 --> 0:15:54.710 of like, okay, the difference between GPT-1 and GPT-2, just 0:15:54.750 --> 0:15:58.290 like extrapolate that out, right? Like there was something about seeing, 0:15:58.710 --> 0:16:00.770 it's almost like if you'd never seen a child before 0:16:00.790 --> 0:16:03.220 and you'd only seen adults and you were kind of like, 0:16:03.620 --> 0:16:04.320 I don't know what this. 0:16:04.220 --> 0:16:06.260 Little thing is. And then suddenly it starts to walk. 0:16:06.620 --> 0:16:09.840 You're like, wow, huh, I bet this thing is going 0:16:09.860 --> 0:16:11.380 to be capable of doing a lot of the things 0:16:11.420 --> 0:16:12.739 that a grownup could do one day. 0:16:12.800 --> 0:16:14.300 Like maybe this is where this is going. 0:16:15.070 --> 0:16:20.100 Dario, you say your departure from OpenAI was partly due 0:16:20.120 --> 0:16:24.979 to a breakdown of trust, an interesting word, with Sam Altman. 0:16:25.480 --> 0:16:28.850 Explain how you felt there was a breakdown of trust 0:16:29.030 --> 0:16:30.650 and honesty there. 0:16:31.590 --> 0:16:35.670 What I will say is, look, leaving OpenAI and founding Anthropic, 0:16:35.750 --> 0:16:39.340 I mean, it ultimately comes down to the principles and 0:16:39.380 --> 0:16:40.760 how we're going to build the technology. 0:16:40.820 --> 0:16:43.100 That is what matters at the end of the day. 0:16:43.690 --> 0:16:46.290 And, you know, I think over the last five and 0:16:46.310 --> 0:16:49.960 a half years, we have seen how those principles differ. 0:16:50.080 --> 0:16:52.840 You know, we see, you know, Anthropic has advocated for 0:16:53.420 --> 0:16:56.560 sensible regulation and safeguards on the technology. 0:16:56.600 --> 0:17:00.120 Anthropic has invested in mechanistic. 0:16:59.600 --> 0:17:04.720 Interpretability research to understand what's going on inside the models, right? 0:17:04.900 --> 0:17:06.620 You know, Anthropic has managed to. 0:17:06.600 --> 0:17:11.760 Maintain a culture of unity to have the highest retention rates. 0:17:11.840 --> 0:17:14.850 We haven't had a number of the scandals that other 0:17:14.910 --> 0:17:17.510 AI companies have had. So, you know, I think that's 0:17:18.650 --> 0:17:20.790 kind of sufficient to describe the difference. 0:17:21.250 --> 0:17:22.390 But I will say. 0:17:22.190 --> 0:17:26.570 That the moment when I realized that we needed to 0:17:26.630 --> 0:17:29.730 found our own company, that OpenAI was not the right 0:17:29.850 --> 0:17:34.449 place for us. Look, these safety issues are difficult. You 0:17:34.490 --> 0:17:38.370 can have lots of disagreements on them. I disagree with 0:17:38.410 --> 0:17:41.550 the co-founders and with lots of people at Anthropic. But 0:17:41.810 --> 0:17:44.830 the reason we've been able to resolve these disagreements is 0:17:44.869 --> 0:17:47.190 because they're held in good faith, right? 0:17:47.550 --> 0:17:50.250 Should we release this model or should we wait? 0:17:50.590 --> 0:17:54.240 Should we put stronger or weaker biosafeguards on the model? 0:17:54.540 --> 0:17:57.100 You know, how much of our effort should we spend 0:17:57.380 --> 0:18:00.440 lobbying for safety regulations and what should they look like? 0:18:00.780 --> 0:18:04.740 These are all good faith disagreements, but they always get 0:18:04.790 --> 0:18:07.830 resolved because there's always trust that everyone is trying to 0:18:07.850 --> 0:18:10.070 get to the right answer and everyone is being honest. 0:18:10.869 --> 0:18:13.450 My point in saying that there's a breakdown of trust 0:18:13.550 --> 0:18:16.390 is simply to make the abstract point that if you 0:18:16.490 --> 0:18:19.330 don't believe that about a place where you're working, if 0:18:19.369 --> 0:18:23.070 you don't believe that the disagreements are being held in 0:18:23.130 --> 0:18:26.290 good faith and that you're getting the real story from people, 0:18:26.750 --> 0:18:28.800 how can you possibly hope to resolve them? 0:18:29.260 --> 0:18:32.080 That's why we left. That's why we founded our own thing. 0:18:32.400 --> 0:18:34.280 And I think the last five and a half years 0:18:34.619 --> 0:18:38.939 have have are themselves that just the best proof of 0:18:39.140 --> 0:18:42.660 how our philosophy differs from that of the other companies 0:18:42.700 --> 0:18:43.220 in the space. 0:18:44.020 --> 0:18:46.619 I think if you continue to work for. 0:18:46.720 --> 0:18:49.380 Someone who has a different vision than you, it becomes 0:18:49.460 --> 0:18:52.180 very unclear what you actually stand for. 0:18:52.660 --> 0:18:53.700 Wait, wait a minute, though. 0:18:53.720 --> 0:18:55.920 You push back a bit here because you talk about 0:18:55.980 --> 0:19:00.790 differences in values and stuff, but that's fundamentally different than 0:19:00.970 --> 0:19:05.030 a difference in trust, that you feel that somebody says 0:19:05.050 --> 0:19:07.280 they have the same values, but they don't and you 0:19:07.320 --> 0:19:08.540 can't trust them. 0:19:09.520 --> 0:19:14.280 Honesty is one of our values. It's a very important value. 0:19:15.180 --> 0:19:18.500 It's hard to know what someone's values are if they 0:19:18.520 --> 0:19:19.840 don't have the value of honesty. 0:19:20.500 --> 0:19:21.680 And you felt that about Sam? 0:19:23.820 --> 0:19:25.740 I think the way, at least for me, I would 0:19:25.780 --> 0:19:30.460 put it is I think the co-founders are an unusually 0:19:30.480 --> 0:19:35.659 high integrity group of people. And all of us I 0:19:35.700 --> 0:19:38.619 think shared a very similar vision for how we wanted 0:19:38.800 --> 0:19:41.220 this technology to go and how we wanted to build it. 0:19:41.900 --> 0:19:44.110 But at a certain point, it just makes more sense 0:19:44.150 --> 0:19:46.310 to go do this as your own organization. 0:19:46.330 --> 0:19:49.869 How are you going to get someone else to implement 0:19:50.030 --> 0:19:54.510 or share your vision if they don't naturally do so? 0:19:55.150 --> 0:19:58.600 No matter how much they say they share it, if 0:19:58.630 --> 0:20:02.919 that's not reflected in what the company actually does, what 0:20:02.960 --> 0:20:03.640 choice do you have? 0:20:04.020 --> 0:20:04.280 Yeah. 0:20:04.740 --> 0:20:07.990 And I do think Anthropic is a company that is 0:20:08.070 --> 0:20:10.869 founded on trust. And I think what I mean by 0:20:10.910 --> 0:20:14.130 that is like, I think honesty and. 0:20:14.070 --> 0:20:14.909 Integrity and. 0:20:17.290 --> 0:20:20.600 How we treat each other and our employees is also 0:20:20.650 --> 0:20:23.560 reflective of how we want the technology to be built 0:20:23.640 --> 0:20:24.120 and to go. 0:20:25.420 --> 0:20:28.820 An example I will give is every two weeks, I 0:20:28.859 --> 0:20:30.980 get up in front of the company and I speak 0:20:31.040 --> 0:20:34.800 for an hour about what is happening inside the company, 0:20:34.840 --> 0:20:37.480