WEBVTT - After The Builders: On Dario & Daniela Amodei and Mustafa Suleyman 0:00:15.860 --> 0:00:18.680 Welcome to Tech Stuff, Oswald Ossian here. It's another special 0:00:18.720 --> 0:00:22.020 edition of the podcast today, focused on Kaleidoscope's new series, 0:00:22.100 --> 0:00:25.270 The Builders, with Walter Isaacson. This time, we're going to 0:00:25.300 --> 0:00:28.820 discuss the third and fourth episodes, where Walter interviews Daniela 0:00:28.890 --> 0:00:33.070 and Dario Amodei, the two co-founders of Anthropic, and Mustafa Suleiman, 0:00:33.110 --> 0:00:37.229 the CEO of Microsoft AI. Dario and Daniela left OpenAI 0:00:37.310 --> 0:00:40.589 in 2020 after disagreements over AI safety with Sam Altman. 0:00:41.070 --> 0:00:44.030 Mustafa Suleiman was pushed out of DeepMind, which he co-founded 0:00:44.050 --> 0:00:48.430 in 2019, after two years of grinding negotiations with Google 0:00:48.530 --> 0:00:53.360 about AI safety and following complaints about his management style. Today, 0:00:53.380 --> 0:00:56.760 they represent different, perhaps even opposite sides of the debate 0:00:56.880 --> 0:01:01.280 on safety, with the Anthropic co-founders advocating for encoding ethics 0:01:01.420 --> 0:01:05.319 in AI, and Suleiman arguing for the primacy of human control. 0:01:05.980 --> 0:01:08.199 It's a fascinating pairing, and today I'm joined by my 0:01:08.240 --> 0:01:12.140 Kaleidoscope co-founder, Mangesh Hatikader, as well as two journalists who 0:01:12.200 --> 0:01:14.530 know more about these figures and the debate about the 0:01:14.560 --> 0:01:17.680 future of AI than anyone I know. We'd like to 0:01:17.700 --> 0:01:20.700 be joined by Sebastian Malaby, currently a senior fellow at 0:01:20.720 --> 0:01:23.220 the Council on Foreign Relations and the author of many books, 0:01:23.700 --> 0:01:26.950 most recently The Infinity Machine, Demis Hassabis' Deep Mind and 0:01:26.970 --> 0:01:30.850 the Quest for Superintelligence, and his forthcoming book recently announced 0:01:31.110 --> 0:01:33.149 and untitled is about Dari Amode. 0:01:33.430 --> 0:01:34.030 Great to be with you. 0:01:34.810 --> 0:01:37.610 And Natasha Tiku, Silicon Valley Culture and AI reporter for 0:01:37.630 --> 0:01:40.280 The Washington Post and regular Tech Stuff contributor. 0:01:40.319 --> 0:01:42.000 Welcome. Thanks for having me. 0:01:42.420 --> 0:01:43.860 Before we get into the builders, we'd love to hear 0:01:43.880 --> 0:01:45.979 a bit from both of you about what it's like 0:01:46.000 --> 0:01:49.260 to report on these people. Natasha, maybe starting with you, 0:01:49.280 --> 0:01:52.620 I'm curious about the kind of arc of reporting on 0:01:52.640 --> 0:01:55.940 these tech leaders, particularly in the age of the self-published 0:01:55.980 --> 0:01:56.600 viral essay. 0:01:56.620 --> 0:02:02.500 Yeah, it requires a lot of time spent reading thousands 0:02:02.540 --> 0:02:06.559 and thousands of words. I put actually recently all of 0:02:07.120 --> 0:02:11.410 Dario's essays in a Google Doc. Because I'm just highlighting 0:02:11.430 --> 0:02:13.290 and there's so many of them. And I had to, 0:02:13.490 --> 0:02:17.070 you know, even incorporate different tabs for his recent long 0:02:17.110 --> 0:02:21.770 form tweet responses. So yeah, it's kind of like, you know, 0:02:21.850 --> 0:02:24.940 Talmudic studies. Sometimes it seems like that. 0:02:26.060 --> 0:02:27.119 Do they write them all themselves? 0:02:27.820 --> 0:02:31.079 I've heard that it is a norm within Anthropic that 0:02:31.200 --> 0:02:33.780 it's actually very bad form if it seems like you 0:02:33.840 --> 0:02:36.500 have used Claude to do your own writing, which I 0:02:36.560 --> 0:02:38.340 think is like a beautiful irony. 0:02:38.360 --> 0:02:41.560 Yeah, I do believe he's written them by himself. 0:02:41.660 --> 0:02:44.989 But Jack Clark is also prolific, the other Anthropic co-founder, right? 0:02:45.010 --> 0:02:46.050 I mean, to keep up with all the things the 0:02:46.090 --> 0:02:50.970 Anthropic co-founders write, you'd have to be quite the reader. Sebastian, 0:02:50.990 --> 0:02:52.450 what about you? I mean, how do you get in 0:02:52.470 --> 0:02:56.110 the heads of these founders? I mean, Demis and now Dario. 0:02:57.100 --> 0:02:59.820 Well, with Demis, it was a question of sitting with 0:02:59.880 --> 0:03:01.820 him in a pub in North London for hours and 0:03:01.840 --> 0:03:04.910 hours and hours, more than 30 in total, and just 0:03:04.930 --> 0:03:08.090 getting him to go through his life story and reflect 0:03:08.190 --> 0:03:10.710 on different bits of it. He's a great narrator, so 0:03:10.730 --> 0:03:13.430 that was actually super fun. And he tells, you know, 0:03:13.470 --> 0:03:16.109 he self-dramatizes very well. And you have to maybe slightly 0:03:16.630 --> 0:03:18.690 filter a few things where you think it may be 0:03:18.730 --> 0:03:23.290 not quite true. But essentially, he is relatively trustworthy and 0:03:23.669 --> 0:03:27.010 extremely fluent. And the thing about him is he doesn't 0:03:27.070 --> 0:03:29.350 write that many essays. He did one this year in 0:03:29.410 --> 0:03:32.340 July on safety. But the way he's told his own 0:03:32.419 --> 0:03:34.900 story is through lectures and podcasts. There hadn't been a 0:03:35.260 --> 0:03:39.220 kind of long-form written book which had access to him, 0:03:39.360 --> 0:03:42.690 which kind of covered not only his childhood and all that, 0:03:42.720 --> 0:03:46.670 but the linkages between what he studied at Cambridge as 0:03:46.710 --> 0:03:50.370 an undergraduate, Claude Shannon and so forth, and then how 0:03:50.430 --> 0:03:53.560 he developed this crazy conviction that you could build powerful A.I., 0:03:54.430 --> 0:03:57.010 and founded a company to do that in 2010, two 0:03:57.030 --> 0:04:00.240 years before AI recognizes the photograph of a cat. I mean, 0:04:00.250 --> 0:04:01.240 the conviction was wild. 0:04:02.900 --> 0:04:05.240 So Mangesh, obviously you and I are quite proud of 0:04:05.320 --> 0:04:07.560 our series, The Builders, and there's a little bit of 0:04:07.720 --> 0:04:11.060 kind of fun, wild ride of how it all came about. 0:04:11.080 --> 0:04:12.840 Why don't you tell us about the Amadeus interview first? 0:04:13.860 --> 0:04:17.170 Yeah, so we're all obviously fascinated by the Amadeus, and 0:04:17.190 --> 0:04:21.210 we're trying to book them. We had reached out, both 0:04:21.270 --> 0:04:24.130 asking for it Dario and Daniela. We thought we'd be 0:04:24.170 --> 0:04:27.060 able to get them separately. And initially they kept saying, well, 0:04:27.100 --> 0:04:29.100 Dario is available this time. And we were pushing more 0:04:29.140 --> 0:04:32.880 for Daniela's voice as well. And then they said, you know, 0:04:33.279 --> 0:04:36.260 would you like to do them together? And we thought, perfect. 0:04:36.300 --> 0:04:39.240 You know, there's only one interview that they've done together before, 0:04:39.300 --> 0:04:47.890 and that's with Oprah. And so, you know, and so, 0:04:47.910 --> 0:04:50.770 you know, we managed to secure them. We're so excited. And, uh, 0:04:52.160 --> 0:04:54.100 And Dario got to the interview first and he was 0:04:54.140 --> 0:04:57.160 there about 10 minutes early and Walter was there as well. 0:04:57.260 --> 0:05:00.500 And so, you know, Walter is Southern and charming and 0:05:00.520 --> 0:05:02.630 he's trying to cajole him into chatting a little bit. 0:05:03.430 --> 0:05:05.570 And Dario kind of waves at him and then just 0:05:05.589 --> 0:05:08.490 looks down and he does not want to do any 0:05:08.550 --> 0:05:11.770 small talk. And Walter just kept saying like, you know, 0:05:11.790 --> 0:05:14.969 a little bit more, a little bit more. And Dario 0:05:14.990 --> 0:05:16.570 was like, you know, I'd really prefer to wait till 0:05:16.610 --> 0:05:20.469 Daniela's here. And so it was just silence until Daniela 0:05:20.510 --> 0:05:23.820 gets there. And then immediately, he turns the charm on. 0:05:23.860 --> 0:05:26.940 You know, like, he's so excited to talk. And you see, 0:05:27.440 --> 0:05:30.960 you know, how close they are. And you realize that, like, 0:05:31.060 --> 0:05:34.710 he's obviously a CEO who's a visionary but doesn't manage anyone. 0:05:35.089 --> 0:05:39.030 And she really both protects and translates his brilliance so 0:05:39.089 --> 0:05:41.390 effectively to a larger world. And it's kind of remarkable 0:05:41.410 --> 0:05:43.690 to watch them interact. And I really feel lucky that 0:05:43.710 --> 0:05:44.810 we managed to get them together. 0:05:45.190 --> 0:05:47.789 Natasha, Sebastian, maybe Sebastian, you first. You were both smiling 0:05:47.810 --> 0:05:49.270 at that account of the pregame. 0:05:49.300 --> 0:05:49.450 Yeah. 0:05:51.180 --> 0:05:53.940 Yeah, I guess the compartmentalizing is what people have to 0:05:54.000 --> 0:05:55.740 do when they're super busy. And, you know, if you've 0:05:55.760 --> 0:05:58.540 got that extra seven minutes before the real thing starts, 0:05:58.980 --> 0:06:00.720 you're going to use it to think of another big 0:06:00.779 --> 0:06:03.780 thought because the value of each incremental big thought is 0:06:03.820 --> 0:06:07.400 measured in the billions of dollars. You know, and the 0:06:07.440 --> 0:06:09.640 safety of humanity might depend on it. So this is 0:06:09.680 --> 0:06:13.620 serious stuff. So, look, I mean, I recognize that tendency 0:06:13.660 --> 0:06:16.500 of being super efficient with one time. And, you know, Demis, 0:06:17.140 --> 0:06:18.520 just to go back to him as a kind of 0:06:18.600 --> 0:06:23.010 baseline comparator, has this routine where he has two working days. 0:06:23.070 --> 0:06:24.750 He works a day in the office, he comes back, 0:06:24.790 --> 0:06:27.190 has dinner with the kids, goes up to his home office, 0:06:27.230 --> 0:06:29.630 works from 10 a.m. until 4 a.m. the next day. 0:06:29.710 --> 0:06:33.810 So he gets the sort of reading, thinking, cogitating time. 0:06:34.550 --> 0:06:37.890 And that's another example of being just aggressively efficient with 0:06:38.440 --> 0:06:41.080 how much you can eke out from each 24-hour unit. 0:06:41.400 --> 0:06:43.940 There's a scene in your book where Demis and Mustafa 0:06:43.980 --> 0:06:46.150 have dinner every Sunday night in the pub and they 0:06:46.190 --> 0:06:48.400 wait until last orders to order so they can eke 0:06:48.460 --> 0:06:51.350 out as much conversational time as possible. But Natasha, what 0:06:51.370 --> 0:06:51.589 about you? 0:06:52.110 --> 0:06:54.550 I mean, I watched the video, so it was just 0:06:54.610 --> 0:06:57.110 reminding me of times when Dario knew that he would 0:06:57.150 --> 0:07:00.570 get in trouble, so he just handed it over to Daniela. 0:07:01.010 --> 0:07:04.039 And I was watching Daniela's hands when Dario was getting 0:07:04.080 --> 0:07:05.880 a little bit too colorful. 0:07:06.400 --> 0:07:09.359 So it was a rich text, the video itself. 0:07:10.540 --> 0:07:13.420 Love a video podcast. Before we get into actual interviews, though, 0:07:13.460 --> 0:07:16.160 can you give us some context, Natasha, starting with you? 0:07:16.480 --> 0:07:18.790 What do we need to know about the Anthropic IPO 0:07:18.830 --> 0:07:19.150 this week? 0:07:19.850 --> 0:07:22.940 I think you need to know that It is one 0:07:22.990 --> 0:07:26.950 of the fastest growing companies in history, period. People are 0:07:27.030 --> 0:07:31.950 anticipating a valuation of potentially $ 2 trillion. 0:07:32.070 --> 0:07:34.670 There's also been reports of $ 1. 0:07:33.510 --> 0:07:35.790 5 trillion, but you have to keep in mind this 0:07:35.870 --> 0:07:39.350 is a company that is less than six years old. 0:07:40.010 --> 0:07:44.200 And the text of the SEC filing is exactly what 0:07:44.220 --> 0:07:46.420 you would expect from the Amadeus, right? Like they talk 0:07:46.460 --> 0:07:49.920 about bringing forth an age of miracles. If you buy 0:07:49.960 --> 0:07:51.780 a share in their company, don't think of it as 0:07:51.840 --> 0:07:54.340 a transaction. Think of it as being, you know, part 0:07:54.420 --> 0:07:58.010 of this momentous period in time. They talk about how 0:07:58.110 --> 0:08:02.470 each of their employees is worth $ 6. 6 million. Like that's 0:08:02.510 --> 0:08:07.190 how talent dense this organization is. And they also go 0:08:07.230 --> 0:08:10.800 through some of the existential issues risks, which they touch 0:08:10.920 --> 0:08:12.320 on in the interview as well. 0:08:13.900 --> 0:08:15.920 Sebastian, I mean, I know it's just been announced in 0:08:16.160 --> 0:08:19.160 Publishers Weekly, and it's sort of probably a quiet period, 0:08:19.500 --> 0:08:21.500 so to speak, on the book. But what's your biggest 0:08:21.560 --> 0:08:23.230 question about Dariam Ali going to be? 0:08:24.270 --> 0:08:26.390 Yeah, I mean, it is super early. I've only just 0:08:26.430 --> 0:08:29.450 begun interviewing, but I would say the questions going in are, 0:08:30.050 --> 0:08:32.150 you know, in a way, how did this extraordinary thing happen? 0:08:32.170 --> 0:08:36.930 You know, these guys left OpenAI, the seven founders of Anthropic. 0:08:36.980 --> 0:08:42.680 They left in 2020. They start the company in 2021. 0:08:42.679 --> 0:08:45.380 And at the time, you know, they're way behind. They 0:08:45.400 --> 0:08:49.080 don't have a big source of funding like Google DeepMind 0:08:49.120 --> 0:08:52.900 did because of Google's, you know, enormous balance sheet. They 0:08:52.920 --> 0:08:56.280 didn't have the head start that OpenAI had. OpenAI then 0:08:56.350 --> 0:09:00.990 goes on to do the chat GPT, goes viral in 2022, 0:09:00.990 --> 0:09:04.530 and Anthropic seems even more behind. They scrounge together some 0:09:04.590 --> 0:09:08.450 money miraculously from the effective altruist community with Sam Backman-Fried 0:09:08.490 --> 0:09:10.670 at the center of that. And then he goes to prison. 0:09:10.690 --> 0:09:13.370 I mean, it's like the setback after setback after setback. 0:09:13.429 --> 0:09:17.579 And yet it turns out that by 2026, these guys 0:09:17.620 --> 0:09:21.500 are at the forefront scientifically. They are the biggest revenue, 0:09:21.559 --> 0:09:23.900 fastest revenue ramp, I think, in the history of capitalism. 0:09:23.940 --> 0:09:26.559 I would go maybe one notch further than Natasha did. 0:09:26.640 --> 0:09:29.709 I think it is the biggest, fastest ever. And then 0:09:29.760 --> 0:09:32.569 also in policy terms, you know, they stick their necks out. 0:09:32.590 --> 0:09:35.569 They take principal stands. They take on the Trump administration. 0:09:36.090 --> 0:09:37.970 And they make the weather on the kind of policy 0:09:37.990 --> 0:09:40.109 and safety front. So I think there's a kind of 0:09:40.170 --> 0:09:45.240 triple here, right? The policy stuff, the business thing, and 0:09:45.250 --> 0:09:46.180 the scientific advance. 0:09:47.160 --> 0:09:50.260 It's so crazy how the tech news cycle is nowadays 0:09:50.320 --> 0:09:55.590 because Microsoft announced they've offered buyouts to 60,000 employees, more 0:09:55.610 --> 0:09:58.830 than 25% of their entire workforce. It's probably like the 0:09:58.890 --> 0:10:01.679 fifth biggest tech story of the week. But I mention 0:10:01.700 --> 0:10:03.840 it just because obviously the other person we're going to 0:10:03.860 --> 0:10:05.780 be talking about in this episode is Mustafa Suleiman of 0:10:05.800 --> 0:10:09.160 Microsoft AI. Have either of you followed these buyouts? I mean, 0:10:09.400 --> 0:10:13.390 the premise is unlocking more capital for compute, essentially, I think. 0:10:14.410 --> 0:10:16.470 Yeah, I mean, I haven't followed that specifically, but it's 0:10:16.510 --> 0:10:18.510 part of a big thing where, you know, when you're 0:10:18.530 --> 0:10:22.080 trying to onboard a lot of artificial intelligence. And I 0:10:22.340 --> 0:10:25.100 use that word deliberately. When you onboard a human, you 0:10:25.120 --> 0:10:26.900 have to kind of figure out how to use that person, 0:10:26.960 --> 0:10:28.500 train them and so forth. It's kind of the same 0:10:28.520 --> 0:10:30.319 with AI. You have to kind of figure out how 0:10:30.360 --> 0:10:34.429 to integrate AI into your business processes to make it productive. 0:10:34.970 --> 0:10:37.989 And that is an enormous suck on management time and 0:10:38.030 --> 0:10:41.510 management focus. And these guys just don't want to have 0:10:41.590 --> 0:10:44.520 too many humans knocking around when they're focused on this 0:10:44.559 --> 0:10:47.490 other thing. And it's already the case that, you know, in, in, 0:10:47.700 --> 0:10:51.320 in anthropic, for example, large amounts of the tasks that 0:10:51.340 --> 0:10:53.980 might have traditionally been done by humans, for example, thinking about, 0:10:54.580 --> 0:10:57.100 you know, working on the IPO. I'm told that there 0:10:57.140 --> 0:10:59.579 are large numbers of AI agents that are doing a 0:10:59.580 --> 0:11:00.140 lot of that work. 0:11:00.960 --> 0:11:03.800 Um, and so they can't write the essays, but they 0:11:03.820 --> 0:11:05.610 can write their perspectives. 0:11:06.830 --> 0:11:09.430 Well, they probably could write the essays. It's just that existentially, 0:11:09.490 --> 0:11:11.350 we don't want them to do that because writing an 0:11:11.390 --> 0:11:14.020 essay and expressing your view. If you're the leader of 0:11:14.059 --> 0:11:16.160 the glab, I mean, it's kind of like, that has 0:11:16.200 --> 0:11:16.800 to be you. 0:11:17.300 --> 0:11:19.500 And Natasha, what's the vibe in Silicon Valley? I mean, 0:11:19.520 --> 0:11:22.120 at the same time, there's any number of Anthropic employees 0:11:22.160 --> 0:11:25.420 who are now worth exactly $ 6. 6 million each versus all 0:11:25.440 --> 0:11:28.460 the Microsoft employees being bought out. Like, what's the mood? 0:11:29.220 --> 0:11:31.540 The mood is, you know, can I get a job 0:11:31.580 --> 0:11:35.020 at Anthropic? Otherwise, you'll be part of the permanent underclass. 0:11:35.100 --> 0:11:38.200 I mean, there's just so much money sloshing around. You 0:11:38.260 --> 0:11:43.059 see people expressing, you know, oh, here's me in a 0:11:43.080 --> 0:11:47.559 picture of Ilya six years ago, Ilya Sutskiver. You know, like, 0:11:47.620 --> 0:11:50.439 did I make the wrong decision? Why are some of 0:11:50.500 --> 0:11:53.450 the people who have the exact same background as me billionaires? 0:11:53.470 --> 0:11:57.220 You know, Matt Levy and his Bloomberg company. write up 0:11:57.320 --> 0:11:59.579 about the S one was great. It was basically like 0:11:59.700 --> 0:12:03.610 Claude is going to subsume all human endeavor. So why 0:12:03.640 --> 0:12:06.309 not buy a share of Claude? That's that was kind 0:12:06.330 --> 0:12:08.790 of the IPO pitch. And that is the mood right now, 0:12:08.890 --> 0:12:13.210 I would say, you know, kind of mouth agape envy. 0:12:13.429 --> 0:12:17.089 I mean, this is the feat of capitalism is is astounding. 0:12:17.510 --> 0:12:20.650 I keep thinking actually back to the Peter Thiel, Walter 0:12:20.710 --> 0:12:24.000 Isaacson interview where he said, can you think of a 0:12:24.040 --> 0:12:29.440 cult that has ever done this? It's phenomenal to think 0:12:29.480 --> 0:12:33.420 of somebody with deep ideological beliefs that is actually at 0:12:33.480 --> 0:12:34.740 the pinnacle of capitalism. 0:12:35.800 --> 0:12:38.550 Yeah, truly. Well, I mean, let's get into that. I 0:12:38.570 --> 0:12:40.950 mentioned at the top that kind of debates over safety 0:12:40.970 --> 0:12:45.790 in some sense inform both Mustafa's departure from Google. I 0:12:45.790 --> 0:12:47.730 want to hear more from both of you on that, 0:12:47.750 --> 0:12:49.949 but also, of course, Daria and Daniela from OpenAI. So 0:12:50.700 --> 0:12:54.199 This was one of Walter's more controversial lines of questioning 0:12:54.300 --> 0:12:56.860 in the interview. Probably the body language told its own story, 0:12:56.880 --> 0:12:57.580 but let's play it now. 0:12:58.400 --> 0:13:03.420 Dario, you say your departure from OpenAI was partly due 0:13:03.460 --> 0:13:07.080 to a breakdown of trust, an interesting word, you know, 0:13:07.120 --> 0:13:10.980 with Sam Altman. Explain how you felt there was a 0:13:11.040 --> 0:13:13.980 breakdown of trust and honesty there. 0:13:14.900 --> 0:13:18.990 What I will say is, look, leaving OpenAI and founding Anthropic, 0:13:19.090 --> 0:13:22.650 I mean, It ultimately comes down to the principles and 0:13:22.670 --> 0:13:24.610 how we're going to build the technology. That is what 0:13:24.830 --> 0:13:27.530 matters at the end of the day. But I will 0:13:27.630 --> 0:13:32.120 say that the moment when I realized that we needed 0:13:32.179 --> 0:13:35.180 to found our own company, that OpenAI was not the 0:13:35.220 --> 0:13:39.939 right place for us. Look, these safety issues are difficult. 0:13:40.020 --> 0:13:43.840 You can have lots of disagreements on them. I disagree 0:13:43.880 --> 0:13:46.660 with the co-founders and with lots of people at Anthropic, 0:13:47.059 --> 0:13:50.350 but the reason we've been able to resolve these disagreements 0:13:50.429 --> 0:13:53.209 is because there's always trust that everyone is trying to 0:13:53.250 --> 0:13:55.490 get to the right answer and everyone is being honest. 0:13:56.270 --> 0:13:58.850 My point in saying that there's a breakdown of trust 0:13:58.950 --> 0:14:01.790 is simply to make the abstract point that if you 0:14:01.910 --> 0:14:04.739 don't believe that about a place where you're working, if 0:14:04.780 --> 0:14:08.480 you don't believe that the disagreements are being held in 0:14:08.559 --> 0:14:11.679 good faith and that you're getting the real story from people, 0:14:12.170 --> 0:14:15.050 How can you possibly hope to resolve them? That's why 0:14:15.090 --> 0:14:17.890 we left. That's why we founded our own thing. And 0:14:18.050 --> 0:14:20.590 I think the last five and a half years are 0:14:21.210 --> 0:14:26.430 themselves just the best proof of how our philosophy differs 0:14:26.510 --> 0:14:29.490 from that of the other companies in the space. I 0:14:29.530 --> 0:14:32.670 think if you continue to work for someone who has 0:14:32.710 --> 0:14:36.390 a different vision than you, it becomes very unclear what 0:14:36.410 --> 0:14:37.530 you actually stand for. 0:14:38.070 --> 0:14:40.830 Wait a minute, though. Push back a bit here because 0:14:40.850 --> 0:14:43.930 you talk about differences in values and stuff, but that's 0:14:44.010 --> 0:14:49.380 fundamentally different than a difference in trust, that you feel 0:14:49.480 --> 0:14:52.220 that somebody says they have the same values, but they 0:14:52.260 --> 0:14:53.960 don't and you can't trust them. 0:14:54.920 --> 0:14:56.200 Honesty is one. 0:14:56.140 --> 0:14:56.900 Of our values. 0:14:58.160 --> 0:15:02.250 It's a very important value. It's hard to know what 0:15:02.290 --> 0:15:05.250 someone's values are if they don't have the value of honesty. 0:15:05.910 --> 0:15:07.030 And you felt that about Sam? 0:15:09.210 --> 0:15:11.130 I think the way, at least for me, I would 0:15:11.170 --> 0:15:15.900 put it is I think the co-founders are an unusually 0:15:15.920 --> 0:15:21.280 high integrity group of people. And all of us, I think, 0:15:21.320 --> 0:15:24.360 shared a very similar vision for how we wanted this 0:15:24.420 --> 0:15:26.620 technology to go and how we wanted to build it. 0:15:27.300 --> 0:15:29.500 But at a certain point, it just makes more sense 0:15:29.540 --> 0:15:31.730 to go do this as your own organization. 0:15:31.750 --> 0:15:32.350 How are you going to. 0:15:32.330 --> 0:15:37.670 Get someone else to implement or share your vision if 0:15:37.710 --> 0:15:42.330 they don't naturally do so. No matter how much they say, 0:15:43.230 --> 0:15:46.360 they share it. If that's not reflected in what the 0:15:46.410 --> 0:15:49.040 company actually does, what choice do you have? 0:15:49.900 --> 0:15:52.980 Natasha, I want your reaction, particularly, I loved your smile 0:15:53.020 --> 0:15:54.200 at Daniela's interjection. 0:15:54.980 --> 0:15:58.980 I think that was actually the part where he moved 0:15:59.000 --> 0:16:01.680 his hands and said, Daniela, you take over. I'm assuming 0:16:01.730 --> 0:16:05.250 that's because They said, Dario, you're way too charged up 0:16:05.350 --> 0:16:08.450 to actually make comments about Sam in any interview. Please 0:16:08.470 --> 0:16:11.729 defer to Daniela. And then he couldn't help himself. He like, 0:16:11.890 --> 0:16:15.320 you know, went back into it. That whole section was 0:16:15.380 --> 0:16:22.840 so fascinating because, you know, obviously this animosity and disapproval, 0:16:24.940 --> 0:16:27.450 hatred in some cases, I'm not sure, you know, I 0:16:27.470 --> 0:16:29.270 don't want to put words in anybody's mouth, but that 0:16:29.830 --> 0:16:33.770 animates so much of And you can see that throughout 0:16:33.850 --> 0:16:36.290 this whole interview, right? Like he comes back to it 0:16:36.350 --> 0:16:39.300 again and again, you know, wanting to make a company 0:16:39.340 --> 0:16:43.840 that is so different from the way that OpenAI operated. 0:16:44.140 --> 0:16:46.380 And a little bit of what he's alluding to, I think, 0:16:46.420 --> 0:16:51.320 in that clip is this emphasis on culture fit, right? 0:16:51.360 --> 0:16:56.190 Like which Anthropic famously has this interview process where, you know, 0:16:56.230 --> 0:16:58.170 you can sail through the technical interview process. 0:16:58.440 --> 0:16:59.420 You can do really well. 0:16:59.500 --> 0:17:02.140 And then they have this culture fit interview, which actually 0:17:02.200 --> 0:17:05.140 involves reading some of Dario's essays. 0:17:05.280 --> 0:17:07.490 I was talking to somebody who sailed. 0:17:07.290 --> 0:17:09.650 Through the technical part and then got cut off at the, 0:17:09.830 --> 0:17:12.330 at the culture part. And they asked like, do you 0:17:12.430 --> 0:17:15.850 share the same views on AI safety? How much of 0:17:16.010 --> 0:17:19.010 a issue is AI safety to you? And you're, you know, 0:17:19.070 --> 0:17:21.770 supposed to espouse some of the same values that, that 0:17:21.910 --> 0:17:26.570 the co-founders have. And I mean, certainly you can see 0:17:26.630 --> 0:17:30.030 that they do operate as one cohesive unit. That's why 0:17:30.090 --> 0:17:32.369 the company is so successful, but you can just tell 0:17:32.450 --> 0:17:36.170 that even five, six years later, Dario is so animated by, 0:17:36.800 --> 0:17:38.639 by the reason that he left open AI. 0:17:38.800 --> 0:17:41.659 You know, it's interesting. We had Elon Musk was the 0:17:41.700 --> 0:17:44.340 first episode of the builders and, you know, he was 0:17:44.380 --> 0:17:46.160 talking about how it would be good to get all 0:17:46.220 --> 0:17:49.060 the AI lab founders talking once a week and blah, blah. 0:17:49.540 --> 0:17:53.370 And Walter said, even Sam. And, uh, I mean, Elon 0:17:53.390 --> 0:17:55.369 was basically like, no, you can't really trust Sam. And 0:17:55.609 --> 0:17:58.450 it is funny how, you know, and there was obviously 0:17:58.470 --> 0:18:00.439 a kind of dust up between Elon and the Amadeus 0:18:00.460 --> 0:18:02.760 as well, which now seems to be resolved. But Sam 0:18:02.780 --> 0:18:07.220 seems to be uniquely disliked by his fellow AI titans, Sebastian. 0:18:07.600 --> 0:18:11.500 Yeah, that's true. I remember being in Silicon Valley talking 0:18:11.580 --> 0:18:15.940 about my book in April, and I was being hosted 0:18:16.040 --> 0:18:20.350 by the general partner of a prominent VC firm who 0:18:20.369 --> 0:18:24.390 had known Sam, you know, for 20 years. And since 0:18:24.430 --> 0:18:26.930 I knew he was, they were friends, I said to him, well, look, 0:18:26.950 --> 0:18:29.409 this New Yorker piece has just come out. We can 0:18:29.450 --> 0:18:32.730 summarize the New Yorker piece as follows. Sam Altman lies. 0:18:32.750 --> 0:18:35.669 And when he finished lying, he lies some more. Is 0:18:35.710 --> 0:18:39.470 that unfair? I said. And this VC who's been friends 0:18:39.490 --> 0:18:43.300 with him for 20 years says, nah, not really. I mean, 0:18:43.420 --> 0:18:44.379 this is what he's. 0:18:44.260 --> 0:18:44.900 Like, right? 0:18:45.020 --> 0:18:50.100 He, he absolutely tells people what they want to hear. And, 0:18:50.420 --> 0:18:53.160 you know, the sympathetic interpretation is this is what people 0:18:53.200 --> 0:18:56.790 do all the time in legislative affairs in Washington, D.C., right? 0:18:56.810 --> 0:18:58.530 You try to get a coalition for a bill in 0:18:58.570 --> 0:19:01.190 Congress by telling one member of Congress it's going to 0:19:01.230 --> 0:19:02.890 do this, which you want, and then the other member 0:19:03.250 --> 0:19:05.369 is going to do that, which you want. And you 0:19:05.390 --> 0:19:07.510 try to get people together and you're not, you know, 0:19:08.109 --> 0:19:10.590 maybe there's like a little bit of truth in both claims, 0:19:10.670 --> 0:19:12.690 but when you put them both together side by side, 0:19:12.710 --> 0:19:15.050 they can't both be quite true. But too bad, you 0:19:15.070 --> 0:19:17.369 put coalitions together. And you do this in business all 0:19:17.390 --> 0:19:19.770 the time. When you're raising money for some big real 0:19:19.810 --> 0:19:21.790 estate project or whatever it is, you've got to tell 0:19:22.090 --> 0:19:23.830 the local community one thing. You've got to tell the 0:19:23.850 --> 0:19:26.600 politicians another. You've got to tell your shareholders yet a 0:19:26.619 --> 0:19:30.080 third thing. The tenants get a fourth story. So people 0:19:30.160 --> 0:19:32.980 do this in the real world, but just Sam Altman 0:19:33.020 --> 0:19:35.700 does it to quite a pronounced degree. And in a 0:19:35.740 --> 0:19:39.910 repetitive game with very high stakes like AI, in the end, 0:19:39.950 --> 0:19:41.010 you burn your boats with people. 0:19:41.850 --> 0:19:43.780 Obviously, I'm sure most of us have watched the Elizabeth 0:19:43.960 --> 0:19:47.870 Holmes documentary, but one of the things which comes across 0:19:47.910 --> 0:19:49.950 from that and from that New York article you mentioned 0:19:50.010 --> 0:19:52.330 is that the default assumption is most of these people 0:19:52.390 --> 0:19:56.200 are lying through their teeth at all times. Natasha, and 0:19:56.220 --> 0:19:58.480 of course, Jensen Huang has implied that the Amodeus kind 0:19:58.520 --> 0:20:01.020 of safety discourse is kind of like a bad faith 0:20:01.600 --> 0:20:04.359 attempt at regulatory capture and marketing, etc. 0:20:04.380 --> 0:20:04.560 etc. 0:20:04.580 --> 0:20:07.100 But when you hear them speak, when you look at them, like, 0:20:07.359 --> 0:20:08.320 I don't know, I mean, this is the same thing 0:20:08.340 --> 0:20:10.780 I have with the Elizabeth Holmes documentary, but like, it doesn't, 0:20:10.800 --> 0:20:13.060 I mean, it looks like they really believe what they're saying. 0:20:13.980 --> 0:20:17.280 Yes, I think that trying to go down the road 0:20:17.359 --> 0:20:21.830 of gauging whether or not they're sincere about these beliefs 0:20:22.050 --> 0:20:22.710 is futile. 0:20:22.770 --> 0:20:25.050 Like, they obviously very strongly believe it. 0:20:25.130 --> 0:20:29.110 But actually, I was talking to Daniel Kokotajlo, who's this 0:20:29.250 --> 0:20:33.169 early OpenAI employee and whistleblower of sorts, you know, he 0:20:33.190 --> 0:20:36.250 wrote that AI 2027, because I was talking to him 0:20:36.310 --> 0:20:40.270 about like the trust that employees have in Dario versus Sam. 0:20:40.490 --> 0:20:45.520 And he was reminding me that actually OpenAI employees felt 0:20:45.680 --> 0:20:49.239 this way about Sam for a number of years. You know, 0:20:49.320 --> 0:20:53.419 now they feel very disillusioned maybe, but he was, he 0:20:53.460 --> 0:20:56.690 was cautioning me to see it more as like DeepMind 0:20:56.730 --> 0:21:00.330 has been around for like, that's a 20 to 15 0:21:00.330 --> 0:21:03.880 year old company and OpenAI is a 10-year-old company and 0:21:04.780 --> 0:21:07.540