WEBVTT 00:00:00.240 --> 00:00:04.877 from Vanity Fair. This is Fair Game. I'm Mark Aduchi. I sat down with Alman for 00:00:04.960 --> 00:00:06.717 two interviews in San Francisco. The 00:00:06.800 --> 00:00:08.797 first on September 11th and then again 00:00:08.880 --> 00:00:12.157 today, October 2nd. In those weeks, the 00:00:12.240 --> 00:00:16.637 story about AI seemed to turn upside down about half a dozen times. I should 00:00:16.720 --> 00:00:20.317 note before we begin that Condas, Vanity Fair's parent company, has a content 00:00:20.400 --> 00:00:22.557 licensing agreement with OpenAI. Both 00:00:22.640 --> 00:00:24.557 interviews with Altman are live now. 00:00:24.640 --> 00:00:28.200 This is [music] part two. 00:00:31.840 --> 00:00:33.597 Sam, it's good to see you again. You, too. 00:00:33.680 --> 00:00:39.117 >> Uh, we sat in this room three weeks ago and so much has happened since. This 00:00:39.200 --> 00:00:42.877 past week, you introduced the world to your new product, Dots. You described it 00:00:42.960 --> 00:00:45.757 as like an AI helper that always has 00:00:45.840 --> 00:00:49.917 your back, inspired by the cool versions of what we watched in movies growing up. 00:00:50.000 --> 00:00:53.517 Sam, what's a Dot? >> Before I get to literally what it is, 00:00:53.600 --> 00:00:55.677 this this is the product I have been waiting for. 00:00:55.760 --> 00:00:57.997 >> Um, I like this. I think more than any 00:00:58.080 --> 00:01:02.717 product we've ever launched it. The fact that it gets to know me as I work. The 00:01:02.800 --> 00:01:04.157 fact that it's just kind of like always 00:01:04.240 --> 00:01:07.277 working and taking on the stuff that I 00:01:07.360 --> 00:01:11.997 didn't want to do. I I noticed that and I'm sure many other people have this too 00:01:12.080 --> 00:01:16.077 that I like let the tasks I most don't want to deal with just pile up and pile 00:01:16.160 --> 00:01:20.477 up and it becomes this like gravity well of dread and it's just sitting there. 00:01:20.560 --> 00:01:22.077 And the fact that my dot is going 00:01:22.160 --> 00:01:24.797 through my email, going through my like 00:01:24.880 --> 00:01:29.597 to-do list of just stuff that is not pleasant for me to do and where I don't 00:01:29.680 --> 00:01:31.197 add much value and it's just like 00:01:31.280 --> 00:01:34.157 knocking that out with undefeable spirit 00:01:34.240 --> 00:01:36.957 and energy uh is amazing and I and I 00:01:37.040 --> 00:01:43.640 really feel like I have this thing working for me all 00:01:44.240 --> 00:01:48.477 >> it somehow feels assistant is the mental model that I had 00:01:48.560 --> 00:01:51.677 in mind when we started working on the And I think that is probably how people 00:01:51.760 --> 00:01:54.477 still are talking about it, but it feels 00:01:54.560 --> 00:01:58.717 much more like a sort of extension of me. 00:01:58.800 --> 00:02:01.677 >> Is it just an avatar of you? Like is that a good way of thinking about it 00:02:01.760 --> 00:02:05.037 that it it knows [clears throat] what you want to say to those emails and is 00:02:05.120 --> 00:02:09.757 getting getting back to them or >> No, I I like I I think that we've I have 00:02:09.840 --> 00:02:13.757 figured out this like good workflow where uh I called my dot yesterday while 00:02:13.840 --> 00:02:16.717 I had some time and I was just like here's a set of ideas. Went off and did 00:02:16.800 --> 00:02:19.837 a bunch of work. By the time I got back to my computer, there was a bunch of 00:02:19.920 --> 00:02:24.397 things where my doc had sent me a message in Slack being like, uh, I, you 00:02:24.480 --> 00:02:27.837 know, did all this work. These it's bad ideas for these reasons. I also found 00:02:27.920 --> 00:02:30.877 this other stuff on your computer. I proactively talked to three people on 00:02:30.960 --> 00:02:34.637 the team and I was like, all right, let's do these things and let's keep 00:02:34.720 --> 00:02:36.317 going. And it was just a completely new 00:02:36.400 --> 00:02:40.557 way of trying to get a new idea towards 00:02:40.640 --> 00:02:43.757 reality. I'm curious like for that to be 00:02:43.840 --> 00:02:48.317 effective it must have access to everything right because I remember 00:02:48.400 --> 00:02:52.077 we're almost the same age when I got to college I immediately got on Facebook 00:02:52.160 --> 00:02:55.117 and was like >> I think we were a little naive we just 00:02:55.200 --> 00:02:57.757 everyone was like free-for-all you know >> it was a freeall 00:02:57.840 --> 00:03:01.357 >> and no one even thought about data how 00:03:01.440 --> 00:03:04.877 does living through Facebook 00:03:04.960 --> 00:03:10.237 impact how you think about privacy with dots particularly 00:03:10.320 --> 00:03:14.957 we have to be extremely extremely um 00:03:15.040 --> 00:03:16.797 rigorous about privacy for this to work. 00:03:16.880 --> 00:03:18.317 But I I was actually thinking about this 00:03:18.400 --> 00:03:22.477 the other day. I I am 00:03:22.560 --> 00:03:25.677 kind of sad about where 00:03:25.760 --> 00:03:30.237 technology has evolved to in many ways because I feel like it is constantly 00:03:30.320 --> 00:03:34.077 taking my attention. You know, my phone I finally turned off 00:03:34.160 --> 00:03:35.837 almost all notifications on my phone. 00:03:35.920 --> 00:03:39.037 It's way better. But it's like a crazy way to live your life with your phone 00:03:39.120 --> 00:03:41.277 just buzzing at you non-stop. My hope 00:03:41.360 --> 00:03:43.117 for the world is that we're going to 00:03:43.200 --> 00:03:46.397 automate the bad parts of life and then 00:03:46.480 --> 00:03:51.757 give people more effort, ability, time to do the good parts and express 00:03:51.840 --> 00:03:55.277 their [snorts] creative spirit or, you know, push on 00:03:55.360 --> 00:03:59.517 something they care about or help their community. And in my own use, I found 00:03:59.600 --> 00:04:03.917 out that one of the things I value the most, no surprise to anyone, is these 00:04:04.000 --> 00:04:07.757 small moments of human connection. And if it's like my dot is messaging yours 00:04:07.840 --> 00:04:10.397 and like that feels really just bad like 00:04:10.480 --> 00:04:14.877 I I want to send you the text. So last time we spoke here you were really eager 00:04:14.960 --> 00:04:16.957 to talk about Astra the latest version 00:04:17.040 --> 00:04:20.317 of chat BT but last week you delayed the 00:04:20.400 --> 00:04:23.357 latest release of Astra 6.1 after the 00:04:23.440 --> 00:04:25.437 model failed to meet safety standards. 00:04:25.520 --> 00:04:27.997 How did you come to that decision and 00:04:28.080 --> 00:04:30.957 what was Astra doing that means it's not 00:04:31.040 --> 00:04:35.517 ready for the public? these these models are in the like 00:04:35.600 --> 00:04:39.837 serious capable seriously capable era now. And so the kind of decisions where 00:04:39.920 --> 00:04:44.717 I think a year ago we could have very reasonably said, you know, just put this 00:04:44.800 --> 00:04:47.997 out in the world. It's maybe a little unsafe, but nothing that bad is going to 00:04:48.080 --> 00:04:49.597 happen and people, you know, we want 00:04:49.680 --> 00:04:52.637 people to benefit from it. Is it's 00:04:52.720 --> 00:04:57.597 different now. And and I think that 00:04:57.680 --> 00:05:01.597 as the capabilities continue to increase, that the bar we will hold for 00:05:01.680 --> 00:05:05.997 ourselves on safety is going to go higher and higher. Now, of course, you 00:05:06.080 --> 00:05:10.157 could say, and I think some of our colleagues in the industry would say, 00:05:10.240 --> 00:05:13.117 that this stuff is just way too dangerous to release. Only a couple of 00:05:13.200 --> 00:05:17.677 companies should have it or one company or one person should have it. Uh, and 00:05:17.760 --> 00:05:21.837 that is also not right. So, there are kind of two major ways that this could 00:05:21.920 --> 00:05:24.557 go wrong. And I think the Astra 6.1 is 00:05:24.640 --> 00:05:29.117 an interesting thing to study here. You can have a big safety accident 00:05:29.200 --> 00:05:33.117 eventually. one that leads to like a real loss of control where AI, you know, 00:05:33.200 --> 00:05:36.877 escapes and it's out doing wrecking havoc on the internet and society can't 00:05:36.960 --> 00:05:41.597 really get control back. You know, AI is so much more capable than us and so 00:05:41.680 --> 00:05:44.397 smart that it's doing things that we don't understand and we don't know how 00:05:44.480 --> 00:05:47.837 to stop it. Like this is the thing from the sci-fi movies. 00:05:47.920 --> 00:05:51.677 On the other hand, you could have AI have way too much power concentration, 00:05:51.760 --> 00:05:53.277 lead to way too much power concentration 00:05:53.360 --> 00:05:55.597 in a small number of companies or people 00:05:55.680 --> 00:06:00.797 or maybe like one country and if you don't release at all, you can fall off 00:06:00.880 --> 00:06:03.037 that way. And again, I think there's a lot of people just like we just want the 00:06:03.120 --> 00:06:05.117 control here. We have to navigate this 00:06:05.200 --> 00:06:09.357 like pragmatic middle path down down the 00:06:09.440 --> 00:06:13.720 center of this and 00:06:14.560 --> 00:06:16.957 6.1 Astra the current model we had ready 00:06:17.040 --> 00:06:19.597 to go I think would have tipped too much 00:06:19.680 --> 00:06:21.757 towards the accident risk. So what we're 00:06:21.840 --> 00:06:25.437 doing instead is pushing towards a 00:06:25.520 --> 00:06:29.277 model that we will still be able to distribute very broadly and we're 00:06:29.360 --> 00:06:32.477 committed to do that even though it's going to take a little longer but much 00:06:32.560 --> 00:06:36.237 we can make much more confident claims about the alignment. So then this is a 00:06:36.320 --> 00:06:39.037 good example of self-regulation of saying we're not going to release this 00:06:39.120 --> 00:06:41.117 product because we've determined for XYZ 00:06:41.200 --> 00:06:44.557 reasons not yet it's not yet safe. Um, 00:06:44.640 --> 00:06:49.357 which brings us to this week. President Trump and AI 00:06:49.440 --> 00:06:52.637 executives, including Open AI's President Greg Brockman, signed an 00:06:52.720 --> 00:06:56.957 accord on super intelligence. It essentially is an agreement for the AI 00:06:57.040 --> 00:07:00.477 companies to self-regulate. As I read it, I mean, it sounds like Trump wants 00:07:00.560 --> 00:07:02.877 you guys to police yourselves. 00:07:02.960 --> 00:07:06.957 You told me 3 weeks ago, we need our governments to step up. And then went 00:07:07.040 --> 00:07:11.997 even further the next week when you went to the UN. You said if AI is to be 00:07:12.080 --> 00:07:15.757 democratic, the most important decisions cannot be made by labs in San Francisco 00:07:15.840 --> 00:07:21.517 alone. Does this new accord fall short in your view? I think it's a good start. 00:07:21.600 --> 00:07:25.837 I think it's like much better than nothing. Um, I remain of the belief that 00:07:25.920 --> 00:07:27.917 we are going to need 00:07:28.000 --> 00:07:32.077 more over time. Um, but for like the 00:07:32.160 --> 00:07:36.237 models of kind of useless this thing, it's a piece of paper that has a typo on 00:07:36.320 --> 00:07:40.317 it. [laughter] Um I don't think that's quite fair. I 00:07:40.400 --> 00:07:43.997 mean I mean yes in some sense it has no real power. On the other hand I think 00:07:44.080 --> 00:07:46.477 there is always like social force behind 00:07:46.560 --> 00:07:49.837 these things and um it's a gesture 00:07:49.920 --> 00:07:53.757 people will point to it and I think it'll have some power inside in between 00:07:53.840 --> 00:07:55.117 these companies. What do you think is 00:07:55.200 --> 00:07:57.597 the next step that you would like to 00:07:57.680 --> 00:08:01.197 see? I think with all of these questions 00:08:01.280 --> 00:08:04.397 about next steps, we are in 00:08:04.480 --> 00:08:06.877 the realm of powerful enough models that 00:08:06.960 --> 00:08:09.517 it is difficult to predict the specific 00:08:09.600 --> 00:08:11.917 things that will go wrong in what order. 00:08:12.000 --> 00:08:16.557 Um, and I think politics and policy is going to be very reactive to the 00:08:16.640 --> 00:08:18.717 specific things that go wrong. So 00:08:18.800 --> 00:08:24.157 >> if the next big thing that goes wrong is like someone makes a novel pathogen, I 00:08:24.240 --> 00:08:25.677 think that will have a very different 00:08:25.760 --> 00:08:28.077 response than if the next thing that 00:08:28.160 --> 00:08:32.397 goes wrong is more cyber security or 00:08:32.480 --> 00:08:37.197 people feeling like, you know, businesses can't succeed because the 00:08:37.280 --> 00:08:41.517 model companies stop giving access. So I I think 00:08:41.600 --> 00:08:43.917 policy makers still have not really felt 00:08:44.000 --> 00:08:47.117 the exponential that we're on. And I 00:08:47.200 --> 00:08:51.357 would guess the world in general will be caught off guard by the next thing that 00:08:51.440 --> 00:08:54.717 goes really wrong. But what's your stance on it? like what's your can you 00:08:54.800 --> 00:08:56.717 clarify you or your company's position 00:08:56.800 --> 00:08:59.517 on government regulation because you 00:08:59.600 --> 00:09:04.637 have said you've preached almost like regulate us 00:09:04.720 --> 00:09:06.317 >> sometimes the actions of the company 00:09:06.400 --> 00:09:09.277 whether it's lobbying in the EU or 00:09:09.360 --> 00:09:13.437 California or >> Greg Brockman's political donations 00:09:13.520 --> 00:09:18.157 don't always seem to be aligned with that idea. So can you clarify your 00:09:18.240 --> 00:09:20.397 stance on 00:09:20.480 --> 00:09:22.317 regulation? Yeah, I mean this is 00:09:22.400 --> 00:09:25.277 obviously not a binary thing. It's not 00:09:25.360 --> 00:09:30.317 like I I I think it's like really easy for people to say either you are for all 00:09:30.400 --> 00:09:34.237 regulation or you're against all regulation. And I don't agree with you 00:09:34.320 --> 00:09:38.077 in that framing. Um I think it's very reasonable for us to say like given what 00:09:38.160 --> 00:09:42.237 we see on the technology with the technology and given that we are trying 00:09:42.320 --> 00:09:45.437 to balance between these two really important things 00:09:45.520 --> 00:09:50.557 >> um we're going to be in the gray area in between here. there's going to like we 00:09:50.640 --> 00:09:53.117 really want people to have the technology. We really want the US to 00:09:53.200 --> 00:09:58.077 win. We really want this to be something that all people get to use and control 00:09:58.160 --> 00:10:01.437 and have a voice in the future. Not just like okay fine AI is giving you a cure 00:10:01.520 --> 00:10:03.037 for cancer, but really get to like use 00:10:03.120 --> 00:10:05.597 to shape where the future's going to go. 00:10:05.680 --> 00:10:09.437 And we are willing to accept some 00:10:09.520 --> 00:10:11.437 things going wrong to make sure that 00:10:11.520 --> 00:10:15.357 this is like a widely held and used and 00:10:15.440 --> 00:10:19.677 democratic technology. What do you mean by some things going wrong? 00:10:19.760 --> 00:10:22.237 Um, 00:10:22.320 --> 00:10:25.437 I think for example, I support open 00:10:25.520 --> 00:10:30.077 source models. Uh, there is a coming title wave of cyber security problems 00:10:30.160 --> 00:10:33.597 that are going to come from these open source models. Now, we're trying to help 00:10:33.680 --> 00:10:36.237 the world be ready for that by saying we've got a little bit of a lead. We 00:10:36.320 --> 00:10:40.957 want to help you defend. Um, but there are people in the industry who say we 00:10:41.040 --> 00:10:44.157 should ban open source because otherwise a bunch of systems are going to be 00:10:44.240 --> 00:10:47.357 hacked. And I don't think that price of 00:10:47.440 --> 00:10:51.517 liberty is worth it. And we're going to as a society, I think have to just 00:10:51.600 --> 00:10:55.117 accept some fairly severe cyber incidents from 00:10:55.200 --> 00:10:58.077 open models in exchange for the liberty that comes with that. 00:10:58.160 --> 00:11:01.197 >> What's the 00:11:01.280 --> 00:11:04.157 what's the limit of that that we accept? 00:11:04.240 --> 00:11:06.077 I mean that is the political process 00:11:06.160 --> 00:11:09.597 like how we determine what 00:11:09.680 --> 00:11:14.317 >> you know safety is always like negotiated between stakeholders like 00:11:14.400 --> 00:11:16.317 airplanes are generally thought of as 00:11:16.400 --> 00:11:21.357 very extremely safe things still sometimes crash so like the limits on 00:11:21.440 --> 00:11:25.437 how much we accept for the benefits like that is not something that I think 00:11:25.520 --> 00:11:29.357 highly regulated >> that's that's an industry with extreme regulation 00:11:29.440 --> 00:11:31.917 >> totally in the past week or so Jensen 00:11:32.000 --> 00:11:34.397 Hang and Bill Gates have come out as 00:11:34.480 --> 00:11:40.077 kind of very different uh with different approaches to this. Hang 00:11:40.160 --> 00:11:43.117 said recently in the New York Times, "The right answer is don't ship products 00:11:43.200 --> 00:11:46.157 until they're in control. It's really quite simple." 00:11:46.240 --> 00:11:50.237 Pacing being self-regulated. Whereas Bill Gates said this is the most 00:11:50.320 --> 00:11:54.557 dangerous thing that humans have ever gone near. And in other areas, do you 00:11:54.640 --> 00:11:56.477 just release drugs without the FDA? He 00:11:56.560 --> 00:12:01.357 brings up aviation. So if they're 00:12:01.440 --> 00:12:06.637 kind of opposites in those views and you're saying it's a gray area in the 00:12:06.720 --> 00:12:09.357 middle, can you define that a little bit more clearly? Like what are the 00:12:09.440 --> 00:12:13.757 principles for safety and regulation that you stick with? 00:12:13.840 --> 00:12:16.877 >> Here are some things that I think have gone well in other industries we should 00:12:16.960 --> 00:12:20.877 learn from. Um I think one principle for learning how to make a new technology 00:12:20.960 --> 00:12:26.077 safe is very clear [snorts] and transparent and kind of like extract 00:12:26.160 --> 00:12:30.157 maximum learning out of accident reporting. So, I think the NTSB, we 00:12:30.240 --> 00:12:33.357 talked about airplanes and cars. These have gotten remarkably safe and I think 00:12:33.440 --> 00:12:37.437 there there's been an amazing culture of accident reporting or incident 00:12:37.520 --> 00:12:40.397 reporting, I don't know what they call it, one of those two. Um, that has 00:12:40.480 --> 00:12:45.437 helped do that. So, I think that's a good thing. Um, another is yeah, I think 00:12:45.520 --> 00:12:47.117 like pre-release standards and a 00:12:47.200 --> 00:12:51.437 liability framework. Um, we earlier did 00:12:51.520 --> 00:12:54.397 uh testing before we released a model but not until after we finished 00:12:54.480 --> 00:12:58.477 training. We're now saying that actually even during training things can now go 00:12:58.560 --> 00:13:01.597 wrong. This is like a new stage where the models are so smart that they can 00:13:01.680 --> 00:13:05.197 get out of a sandbox, hack into a system, things like that. So, also in 00:13:05.280 --> 00:13:08.717 the news this week, it was reported that OpenAI President Greg Brockman backed 00:13:08.800 --> 00:13:10.877 out of a $25 million commitment to 00:13:10.960 --> 00:13:13.757 leading the future super PAC who that is 00:13:13.840 --> 00:13:17.997 generally speaking opposes strong AI AI regulation. uh he and his wife had 00:13:18.080 --> 00:13:20.557 previously already donated$2 million and 00:13:20.640 --> 00:13:23.997 they had donated $25 million to 00:13:24.080 --> 00:13:26.077 President Trump's re-election campaign. 00:13:26.160 --> 00:13:28.477 Did you personally ask Greg to withdraw 00:13:28.560 --> 00:13:33.197 his pledge and how much daylight is there between your views on regulation 00:13:33.280 --> 00:13:35.517 and Greg's? >> Can I finish my answer the other 00:13:35.600 --> 00:13:40.840 question because it actually touches on this one but then we'll answer it. Um 00:13:41.200 --> 00:13:43.437 what I was going to say in terms of like 00:13:43.520 --> 00:13:46.397 where we are on the answer for 00:13:46.480 --> 00:13:50.237 regulation it is uh you don't want to 00:13:50.320 --> 00:13:52.957 overregulate to a degree that people 00:13:53.040 --> 00:13:57.917 don't get the benefits and that new entrance and you know random kids with 00:13:58.000 --> 00:14:02.317 ideas cannot build all the stuff we're going to do in the future. And part of 00:14:02.400 --> 00:14:06.717 why I think the internet was so magical and why the US did so well is that we 00:14:06.800 --> 00:14:10.477 accepted some things going a little bit wrong to give people the freedom and 00:14:10.560 --> 00:14:14.957 liberty and you know that was that was good but 00:14:15.040 --> 00:14:18.637 unlike other kinds of technology not all like you know we've seen some 00:14:18.720 --> 00:14:23.037 extraordinarily powerful technologies in our history the potential downside here 00:14:23.120 --> 00:14:27.357 may not be the kind of thing normally goes wrong that we can correct we have a 00:14:27.440 --> 00:14:31.197 legal system for um but that something goes super wrong like in the extreme 00:14:31.280 --> 00:14:32.317 extreme case back to what we talked 00:14:32.400 --> 00:14:34.557 about of you know we we lose control of 00:14:34.640 --> 00:14:38.397 the future to AI which is horrific. So 00:14:38.480 --> 00:14:41.597 my stance on regulation is 00:14:41.680 --> 00:14:46.477 pro- innovation, let some things go normally wrong and 00:14:46.560 --> 00:14:51.037 trust that we have a normal process for doing this. But where this technology is 00:14:51.120 --> 00:14:52.637 extraordinarily powerful and 00:14:52.720 --> 00:14:55.357 fundamentally new, have guardrails to 00:14:55.440 --> 00:14:58.717 ensure that all of society is kept 00:14:58.800 --> 00:15:02.557 reasonably safe and we don't have we don't kind of fall off that path in 00:15:02.640 --> 00:15:04.397 either direction in what could be a very 00:15:04.480 --> 00:15:08.237 difficult or unreoverable way. So, so do 00:15:08.320 --> 00:15:11.757 you think Greg is on the same page with that? On to that. I was going to say I 00:15:11.840 --> 00:15:13.037 think Greg is on the same page as that. 00:15:13.120 --> 00:15:15.197 I don't want to speak for him. You'd have to ask him, but I think he would 00:15:15.280 --> 00:15:19.837 give you a very similar take. I did not ask him to donate or not, you know, stop 00:15:19.920 --> 00:15:22.077 donating. He's like can make his own person >> because when you look at it 00:15:22.160 --> 00:15:25.037 holistically, it's like you're saying something, the company's saying 00:15:25.120 --> 00:15:29.757 something, but then there's lobbying efforts or executive donations that seem 00:15:29.840 --> 00:15:34.317 to speak to a different thing. So, >> I don't think that's a fair cut. Like I 00:15:34.400 --> 00:15:38.557 I again I don't want to like you should ask Greg. I don't have to speak for him 00:15:38.640 --> 00:15:41.517 on this but I don't I don't think Greg would say he's like against all AI 00:15:41.600 --> 00:15:44.557 regulation or has any of his contributions have ever expressed that 00:15:44.640 --> 00:15:49.437 view. Okay. Let's talk about the Chinese state visit because when we were in this 00:15:49.520 --> 00:15:53.437 room you told me that you were looking forward to Xiinping meeting with 00:15:53.520 --> 00:15:56.397 President Trump at the White House. What you said was these two leaders could 00:15:56.480 --> 00:16:00.957 make tremendous progress and then you attended the state dinner. 00:16:01.040 --> 00:16:05.277 It didn't feel like there was tremendous progress from that visit, but you know, 00:16:05.360 --> 00:16:08.557 a lot happens behind closed doors. What was the biggest takeaway from that night 00:16:08.640 --> 00:16:11.517 that the public has not really heard about yet? 00:16:11.600 --> 00:16:15.917 >> I don't really know either. Um, I was I mean, it was like a nice dinner, but uh 00:16:16.000 --> 00:16:18.317 the substant I was not in the room for the substantive conversations. 00:16:18.400 --> 00:16:22.237 >> Okay. At that state dinner, Trump offered guests a tour of the ballroom. 00:16:22.320 --> 00:16:24.077 Yeah. >> Did you take him up on the offer? 00:16:24.160 --> 00:16:25.997 >> Very briefly. >> What was it like? 00:16:26.080 --> 00:16:28.157 >> It's a construction site. It's big. 00:16:28.240 --> 00:16:30.797 Anything more? It's it's still like a 00:16:30.880 --> 00:16:32.397 very raw construction site. 00:16:32.480 --> 00:16:35.437 >> Okay, because here's how that dinner 00:16:35.520 --> 00:16:40.317 looks, right? I mean, you have the leaders of this industry who are some of 00:16:40.400 --> 00:16:43.917 the richest people in the world with the two most powerful men in the world 00:16:44.000 --> 00:16:46.317 arguably, and Trump gets up there and 00:16:46.400 --> 00:16:50.557 doesn't say one thing about AI, which is 00:16:50.640 --> 00:16:53.997 what the whole world is hoping 00:16:54.080 --> 00:16:57.677 he will lead on or that someone will lead on. It was an opportunity for 00:16:57.760 --> 00:17:00.397 leadership that he seems to have just 00:17:00.480 --> 00:17:02.717 skirted to talk about this ballroom. 00:17:02.800 --> 00:17:05.007 It's the elephant in the ballroom. 00:17:05.090 --> 00:17:07.277 [snorts] >> I I heard they did talk about it 00:17:07.360 --> 00:17:10.557 earlier. I mean, you know, like I think 00:17:10.640 --> 00:17:15.517 as I have never hosted another president first day dinner obviously, but when 00:17:15.600 --> 00:17:17.197 I've done the business version of this, 00:17:17.280 --> 00:17:20.397 we will often do 00:17:20.480 --> 00:17:24.237 difficult work all day and then by the time dinner comes, people are like very 00:17:24.320 --> 00:17:28.317 tired or anything like ahead of time or after 00:17:28.400 --> 00:17:31.597 the CEOs. Um, but I know that the two presidents spent a lot of time. 00:17:31.680 --> 00:17:34.797 >> Okay. I worry about him representing an 00:17:34.880 --> 00:17:37.517 industry that he seems to not quite 00:17:37.600 --> 00:17:40.637 understand. You have thought about the 00:17:40.720 --> 00:17:43.117 subject of AI for longer than almost 00:17:43.200 --> 00:17:46.397 anybody. Is the extinction risk from AI 00:17:46.480 --> 00:17:50.797 overhyped? >> Well, the fact that there's any at all, 00:17:50.880 --> 00:17:53.357 the fact that it is like a nonzero zero, 00:17:53.440 --> 00:17:55.197 >> uh, I think is something that we all 00:17:55.280 --> 00:18:00.957 should take incredibly seriously. And part of why we formed OpenAI is the 00:18:01.040 --> 00:18:02.317 mission to make sure that doesn't happen. 00:18:02.400 --> 00:18:04.157 >> But you know that there's some people 00:18:04.240 --> 00:18:06.477 are saying that that it's very canny 00:18:06.560 --> 00:18:09.517 marketing or that it contributes to a 00:18:09.600 --> 00:18:11.197 bubble or that it's some collaborative 00:18:11.280 --> 00:18:13.837 effort to block newcomers to the AI 00:18:13.920 --> 00:18:16.557 market. Is that just cynical? Is that 00:18:16.640 --> 00:18:19.197 just um analysis on analysis? 00:18:19.280 --> 00:18:22.077 >> I assume some people are using it as a 00:18:22.160 --> 00:18:26.237 way to justify power seeking power concentration. Um, but that doesn't mean 00:18:26.320 --> 00:18:29.357 there's not a real risk here. The thing I don't like about the conversation is 00:18:29.440 --> 00:18:33.677 when people say like this is so risky, there's this, you know, whatever percent 00:18:33.760 --> 00:18:35.357 chance we better hurdle towards it as 00:18:35.440 --> 00:18:38.877 fast as we can. Like I think the 00:18:38.960 --> 00:18:42.717 constructive thing here is 00:18:42.800 --> 00:18:47.757 yes, there are risks. We are all here to mitigate those risks 00:18:47.840 --> 00:18:51.757 to let people get the benefits. But before we do that to ensure that there's 00:18:51.840 --> 00:18:54.077 not going to be a really bad thing and 00:18:54.160 --> 00:18:58.797 we need to talk with the world and kind of collaborate on how to make sure that 00:18:58.880 --> 00:19:02.717 that happens and people understand how that's happening. Um, 00:19:02.800 --> 00:19:07.917 we also need to work to avoid using this as a maybe not maybe marketing is too 00:19:08.000 --> 00:19:12.557 negative, but like a way to justify power concentration and society has got 00:19:12.640 --> 00:19:16.637 to like decide on the particulars of the 00:19:16.720 --> 00:19:19.677 trade-offs here. There was a report this week in the New York Times claiming that 00:19:19.760 --> 00:19:23.837 your company ignored warnings from employees to top executives months 00:19:23.920 --> 00:19:28.157 before the Hugging Pace incident. Do you refute that or is that 00:19:28.240 --> 00:19:31.837 >> I'm not familiar with that specific report? I mean, look, you can probably 00:19:31.920 --> 00:19:34.637 you can complain to me like, "Oh, you're delaying a model." Or you can complain 00:19:34.720 --> 00:19:36.797 to me like, "Oh, you're like, you know, 00:19:36.880 --> 00:19:38.797 ignoring warning signs or many other 00:19:38.880 --> 00:19:41.757 things and like the answer is 00:19:41.840 --> 00:19:44.557 >> we're going to make I'm going to make 00:19:44.640 --> 00:19:49.277 many individual decisions you clearly won't like and you'll say, "Oh, you're 00:19:49.360 --> 00:19:52.077 like marketing and you're scaring people or oh, you're delaying a release or oh, 00:19:52.160 --> 00:19:55.837 you're like trying to, you know, you're ignoring warning signs and you need 00:19:55.920 --> 00:19:59.677 something to write about every day." I get it. I really do. My goal is to 00:19:59.760 --> 00:20:04.557 navigate this path and I won't get every decision right. Um, but making sure that 00:20:04.640 --> 00:20:06.557 we 00:20:06.640 --> 00:20:11.997 let the whole world have agency and authority and 00:20:12.080 --> 00:20:14.637 the ability to self-determine their future and make their own lives and 00:20:14.720 --> 00:20:18.077 their communities lives better while avoiding 00:20:18.160 --> 00:20:21.917 a and you know not seeing the power concentrate in my company or any other 00:20:22.000 --> 00:20:24.077 company while avoiding safety things. 00:20:24.160 --> 00:20:27.677 And so I'm sure you can always find someone who's going to say like, "Ah, 00:20:27.760 --> 00:20:30.477 Sam shouldn't have made this decision or that decision." You can come try making 00:20:30.560 --> 00:20:34.157 them all. They're not so easy. Um, but if you look at the trajectory of what we 00:20:34.240 --> 00:20:36.877 have done over the last many years we've 00:20:36.960 --> 00:20:41.837 been offering this. Um, it's imperfect, but I think remarkably 00:20:41.920 --> 00:20:43.597 good relative to what I feared it could 00:20:43.680 --> 00:20:46.797 have been. I have two more questions. 00:20:46.880 --> 00:20:48.957 One is that when we were here last, I 00:20:49.040 --> 00:20:50.717 asked you if there were any politicians 00:20:50.800 --> 00:20:53.917 whose positions on AI you approved. 00:20:54.000 --> 00:20:56.157 If the Democrats do take the House, will 00:20:56.240 --> 00:20:58.957 that help or hinder OpenAI's goals 00:20:59.040 --> 00:21:01.277 towards regulation? 00:21:01.360 --> 00:21:04.797 >> The main thing I have learned about 00:21:04.880 --> 00:21:07.917 Washington in the last few years that 00:21:08.000 --> 00:21:09.997 I've, you know, kind of had a more 00:21:10.080 --> 00:21:12.477 active role is that I cannot predict how 00:21:12.560 --> 00:21:15.197 that place works at all. I remember when 00:21:15.280 --> 00:21:17.837 I first testified in front of the Senate 00:21:17.920 --> 00:21:20.477 in like was that [clears throat] 23 23. 00:21:20.560 --> 00:21:23.997 I was so convinced and the people seemed so genuine. We're going to get this done 00:21:24.080 --> 00:21:26.557 in 2 weeks. It's going to be non-controversial. This is just like 00:21:26.640 --> 00:21:31.277 basic common sense. It's going to be done. We won't need to see you again. 00:21:31.360 --> 00:21:32.637 I was like great. That's how it works. 00:21:32.720 --> 00:21:35.597 This is these is, you know, and then nothing has happened. And I was so 00:21:35.680 --> 00:21:37.437 convinced then it was going to happen. 00:21:37.520 --> 00:21:39.917 >> Sam, when the history of AI is written, 00:21:40.000 --> 00:21:43.277 who do you want to be? What role do you 00:21:43.360 --> 00:21:45.517 want to have played? 00:21:45.600 --> 00:21:47.837 Let me put it this way. 00:21:47.920 --> 00:21:52.637 Elon Musk is like the Mad King Nero 00:21:52.720 --> 00:21:56.077 character. I think Jensen Hong will be the guy who was selling picks and 00:21:56.160 --> 00:21:58.477 shovels during the gold rush. If you had 00:21:58.560 --> 00:22:02.077 to summarize who you would like to be 00:22:02.160 --> 00:22:04.237 when the history of AI is written, Sam, 00:22:04.320 --> 00:22:07.677 who do you specifically want to be? 00:22:07.760 --> 00:22:09.277 >> First of all, I never think this way. 00:22:09.360 --> 00:22:11.277 And I think it's probably like a super 00:22:11.360 --> 00:22:15.197 toxic way to live your life. Um like I I 00:22:15.280 --> 00:22:19.277 try to just stay really focused on my 00:22:19.360 --> 00:22:24.317 work and you know there's so much still to to do 00:22:24.400 --> 00:22:27.037 and I don't want to get like too much main character energy here. I think 00:22:27.120 --> 00:22:29.917 that's super bad. um that's probably 00:22:30.000 --> 00:22:32.957 healthy and yet I think that people 00:22:33.040 --> 00:22:38.877 would benefit from understanding your own personal 00:22:38.960 --> 00:22:39.917 mission. 00:22:40.000 --> 00:22:43.757 >> Yeah. I like 00:22:43.840 --> 00:22:46.957 the thing that has driven me through all 00:22:47.040 --> 00:22:50.077 of this even especially the hard days is 00:22:50.160 --> 00:22:54.317 uh I have such a belief that 00:22:54.400 --> 00:22:57.517 although there are challenges although there's reason for unease although there 00:22:57.600 --> 00:23:01.677 are things we have to get right as we navigate this path that putting this 00:23:01.760 --> 00:23:04.237 technology in the hands of everyone and 00:23:04.320 --> 00:23:08.877 giving people a legitimate new superpower relative to 00:23:08.960 --> 00:23:09.917 I think what any of us could imagine 00:23:10.000 --> 00:23:12.397 mentioned a few years ago will do 00:23:12.480 --> 00:23:16.157 tremendously more good than bad. And 00:23:16.240 --> 00:23:18.637 a big part of my own motivation to start 00:23:18.720 --> 00:23:23.437 and keep working in OpenAI is that I think there are many people many of the 00:23:23.520 --> 00:23:25.597 traditional parts of the community in 00:23:25.680 --> 00:23:27.597 the Bay Area that were thinking about AI 00:23:27.680 --> 00:23:30.717 that really believed only a small 00:23:30.800 --> 00:23:36.077 priesthood should get access to AI and that maybe that that 00:23:36.160 --> 00:23:38.557 group would like, you know, they'd be benevolent dictators. they would decide 00:23:38.640 --> 00:23:40.477 how the rest of us got to use it. I 00:23:40.560 --> 00:23:42.717 think there are other people pursuing AI 00:23:42.800 --> 00:23:47.917 that just want like more power for themselves or their company. Like maybe 00:23:48.000 --> 00:23:51.437 they're newer to this, but they're just like this is, you know, this is how 00:23:51.520 --> 00:23:55.677 we're going to win. And and and what I would really like if the history books 00:23:55.760 --> 00:23:58.797 ever write anything about me at all be great if they don't is just like the 00:23:58.880 --> 00:24:02.077 work that OpenAI did really made sure 00:24:02.160 --> 00:24:05.757 that AI was like a fundamental 00:24:05.840 --> 00:24:09.277 thing that every person got hugely empowered with. It was not something 00:24:09.360 --> 00:24:13.197 happened to them. It was something that people were like given was like a new 00:24:13.280 --> 00:24:17.517 fire or electricity or whatever. And it led to this incredibly great new society 00:24:17.600 --> 00:24:20.957 that people built for each other. Not that like a small group of people 00:24:21.040 --> 00:24:23.117 worshiping the machine god decided to 00:24:23.200 --> 00:24:26.877 like bequeath on all of us. Prometheus 00:24:26.960 --> 00:24:28.157 closer to that than other things you 00:24:28.240 --> 00:24:35.080 could come up with for sure. Okay, Sam, thank you very much. Thank you.