WEBVTT 00:00:08.320 --> 00:00:11.720 Hello everyone. 00:00:13.120 --> 00:00:14.877 >> Well, welcome to lunch. Not for you and 00:00:14.960 --> 00:00:19.517 me, but welcome to lunch. Um, 00:00:19.600 --> 00:00:21.677 uh, I think we've had a, uh, a great day 00:00:21.760 --> 00:00:25.117 so far. and uh and I'm honored to have 00:00:25.200 --> 00:00:29.517 with me uh Sam Alman uh joining us for 00:00:29.600 --> 00:00:31.997 our discussion about 00:00:32.080 --> 00:00:34.957 uh AI and and where we're going from 00:00:35.040 --> 00:00:37.597 here. So um 00:00:37.680 --> 00:00:39.677 so thanks for thanks for joining us. You 00:00:39.760 --> 00:00:42.397 have really the greatest countries in 00:00:42.480 --> 00:00:44.717 the world all around you. So would you 00:00:44.800 --> 00:00:47.757 mind giving us a little history like how 00:00:47.840 --> 00:00:50.317 do how do you go from being Sam Alman to 00:00:50.400 --> 00:00:54.957 being the head of AI and really at the frontier like sort of walk us through 00:00:55.040 --> 00:00:57.037 how you got from there to here? 00:00:57.120 --> 00:01:00.077 >> First of all thank you very much for having me. Um as you said this is 00:01:00.160 --> 00:01:03.357 greatest countries in the world and it's it's a great honor to get to address all 00:01:03.440 --> 00:01:06.717 of you at the same time. Um 00:01:06.800 --> 00:01:11.597 I I was always an AI nerd. I was like a real sci-fi nerd as a kid. I thought AI 00:01:11.680 --> 00:01:15.517 was the coolest thing. I always wanted to work on it. Um, I never like really 00:01:15.600 --> 00:01:19.437 thought I'd actually get to do it. I kind of can't believe this happened to 00:01:19.520 --> 00:01:23.277 me, but I I always I always thought it was awesome. I went 00:01:23.360 --> 00:01:25.037 to school to study it at the time. 00:01:25.120 --> 00:01:28.797 Nothing was working and I kind of gave up on it and I went on this side quest 00:01:28.880 --> 00:01:30.397 of entrepreneurship, which was also 00:01:30.480 --> 00:01:35.437 great. Um, and then in 2012, 00:01:35.520 --> 00:01:38.397 um, everything changed. None of us paid 00:01:38.480 --> 00:01:42.717 enough attention, but there was a a real watershed moment where all of a sudden 00:01:42.800 --> 00:01:47.037 deep learning started to work and it got better with more compute. And I paid a 00:01:47.120 --> 00:01:50.877 lot of attention to that for the next couple of years. And by the time 2015 00:01:50.960 --> 00:01:55.117 rolled around, I was like, hey, this is really this is going to happen. Like we 00:01:55.200 --> 00:01:58.557 got to take we got to try this could be like the most important thing ever. Uh 00:01:58.640 --> 00:01:59.997 and so we thought for a while about what 00:02:00.080 --> 00:02:03.597 to do and started OpenAI. Uh and 00:02:03.680 --> 00:02:09.117 >> so what year did you start OpenAI? uh the very end of 2015 um we announced it 00:02:09.200 --> 00:02:10.557 and then the first day of work was like 00:02:10.640 --> 00:02:13.757 January of 2016. 00:02:13.840 --> 00:02:17.037 >> So all right so you start in uh in in 00:02:17.120 --> 00:02:19.757 2016 and and what are your like what's 00:02:19.840 --> 00:02:24.477 the big goal that you start with and then how does that evolve? 00:02:24.560 --> 00:02:26.077 So we knew that there was this magic 00:02:26.160 --> 00:02:28.157 piece of technology. We knew that we 00:02:28.240 --> 00:02:31.997 could apply compute 00:02:32.080 --> 00:02:34.397 to a problem and to learn it. And we 00:02:34.480 --> 00:02:39.277 knew that the more compute we used, the smarter and more capable the system got. 00:02:39.360 --> 00:02:45.037 We didn't really know anything else. We didn't know how we were going to um 00:02:45.120 --> 00:02:47.437 what kind of system we were going to build. This is way before the idea of 00:02:47.520 --> 00:02:49.277 language models. We were trying to train 00:02:49.360 --> 00:02:54.637 a robot hand to do a Rubik's cube. Um we were trying to, you know, show that we 00:02:54.720 --> 00:02:57.997 could play video games. And we were really developing the algorithms and 00:02:58.080 --> 00:03:00.157 trying to understand how this amazing 00:03:00.240 --> 00:03:03.197 property of the universe worked. And and 00:03:03.280 --> 00:03:06.317 then not until about 2018 00:03:06.400 --> 00:03:08.957 did we start making progress on 00:03:09.040 --> 00:03:10.797 unsupervised learning. And then the 00:03:10.880 --> 00:03:15.437 first GPT1 model and not until 2020 did 00:03:15.520 --> 00:03:17.997 we train GPT3 which was the first model 00:03:18.080 --> 00:03:22.637 that I think actually was over some threshold of of usability. Even then, 00:03:22.720 --> 00:03:26.477 most people in the world looked at this and said, you know, okay, the model can 00:03:26.560 --> 00:03:29.357 like generate a co coherent sentence. 00:03:29.440 --> 00:03:31.037 That was really the limit at the time. 00:03:31.120 --> 00:03:33.837 Kind of impressive, but like what is this going to what is this going to 00:03:33.920 --> 00:03:36.557 mean? And why is why is this going to 00:03:36.640 --> 00:03:41.997 like matter to the economy or society or anything else? Um, but we had discovered 00:03:42.080 --> 00:03:45.597 by this point these scaling laws which said, well, if we could, you know, apply 00:03:45.680 --> 00:03:47.757 like a trillion dollars of compute to 00:03:47.840 --> 00:03:52.797 this idea, something quite amazing would happen. And people said, "Well, you 00:03:52.880 --> 00:03:55.277 can't apply a trillion dollars of pizza, so it doesn't matter." And we sort of 00:03:55.360 --> 00:03:59.277 said, "Why not?" Like, let's try. 00:03:59.360 --> 00:04:01.117 So, 00:04:01.200 --> 00:04:04.237 you've got the idea and and it's moving 00:04:04.320 --> 00:04:07.917 forward, but when when does it come 00:04:08.000 --> 00:04:10.957 together that it's really bottling 00:04:11.040 --> 00:04:15.677 knowledge? like how how does it 00:04:15.760 --> 00:04:19.757 >> different people at OpenAI would give you different answers to this but I 00:04:19.840 --> 00:04:22.557 would say mo the kind of consensus to 00:04:22.640 --> 00:04:28.920 the degree there is one was the time of GPT4 this was like 00:04:29.680 --> 00:04:32.157 I believe this was like March of 2023 00:04:32.240 --> 00:04:37.037 but I may be wrong on the timeline there um we had that model for maybe eight 00:04:37.120 --> 00:04:40.877 months before it launched so we had had time internally to really get to know it 00:04:40.960 --> 00:04:44.957 and that was the time where people even some of the people internally had been 00:04:45.040 --> 00:04:47.757 skeptics of this approach would looked 00:04:47.840 --> 00:04:51.037 at it and said this is going to go very 00:04:51.120 --> 00:04:54.637 far. Um and it hasn't been that long 00:04:54.720 --> 00:04:56.157 since that moment you know three and a 00:04:56.240 --> 00:04:59.597 half years now. Uh and 00:04:59.680 --> 00:05:04.717 and the fact that we now have a system 00:05:04.800 --> 00:05:07.517 uh and we're launching our new model soon which is another step forward and 00:05:07.600 --> 00:05:12.157 there be many more step forwards coming soon after that. um but that it is so 00:05:12.240 --> 00:05:13.837 capable that it is discovering new 00:05:13.920 --> 00:05:16.557 knowledge, doing science, able to add 00:05:16.640 --> 00:05:19.357 huge economic value, create entire comp 00:05:19.440 --> 00:05:24.397 pieces of complex software. Um this is 00:05:24.480 --> 00:05:28.557 this was like a very crazy moment. I think it hit us. 00:05:28.640 --> 00:05:30.557 Yeah. Over a period of time but centered 00:05:30.640 --> 00:05:34.637 around this GPT4 period. 00:05:34.720 --> 00:05:38.157 >> So you understand that 00:05:38.240 --> 00:05:41.037 you can sort of capture education right? 00:05:41.120 --> 00:05:45.677 You capture knowledge. Um, how how 00:05:45.760 --> 00:05:52.440 should the the G20 economies think about this? 00:05:54.080 --> 00:05:57.837 I I don't want to overstate things, but I'm like very excited about our new 00:05:57.920 --> 00:05:59.917 model, so forgive me if I get a little 00:06:00.000 --> 00:06:05.480 bit too enthusiastic here. Um I 00:06:06.240 --> 00:06:10.557 I think this for example will be the greatest boom in entrepreneurship and 00:06:10.640 --> 00:06:11.997 the creation and development of small 00:06:12.080 --> 00:06:15.117 businesses um the world has ever seen. 00:06:15.200 --> 00:06:17.837 If you have a great idea but you didn't 00:06:17.920 --> 00:06:20.797 have the resources or the ability to get 00:06:20.880 --> 00:06:25.357 enough capital together or hire enough experts to work on it or to figure out 00:06:25.440 --> 00:06:29.037 what it's going to take to invent a new piece of science or manage a complex 00:06:29.120 --> 00:06:32.077 supply chain. You can do that with AI 00:06:32.160 --> 00:06:36.077 now. Um, and I remember after ChateBT 00:06:36.160 --> 00:06:40.477 launched, one of the stories that has stuck with me the most from that moment 00:06:40.560 --> 00:06:42.797 was a guy that managed a laundromat. 00:06:42.880 --> 00:06:45.757 Apparently, laundromats are usually good businesses. Wouldn't have been obvious 00:06:45.840 --> 00:06:49.357 to me. Apparently, they are. Um, but this one was not going well. And he was 00:06:49.440 --> 00:06:52.397 using Chat GBT for everything. um to 00:06:52.480 --> 00:06:54.637 make his marketing materials to you know 00:06:54.720 --> 00:06:56.717 manage give him advice on legal 00:06:56.800 --> 00:07:01.757 contracts to manage like relationships with vendors this long list of stuff he 00:07:01.840 --> 00:07:03.437 he was really the first person I met 00:07:03.520 --> 00:07:07.037 that had figured out how to use AI to 00:07:07.120 --> 00:07:11.517 effectively run run a business and this was amazing to me at the time now I know 00:07:11.600 --> 00:07:14.957 you all have heard stories about this a bunch and it's no longer surprising but 00:07:15.040 --> 00:07:17.837 in 2023 the fact that someone was using 00:07:17.920 --> 00:07:22.477 chat GBT broken and embarrassing though it was at the time to let them manage a 00:07:22.560 --> 00:07:23.757 business that otherwise they couldn't 00:07:23.840 --> 00:07:27.037 have done was was like quite remarkable. 00:07:27.120 --> 00:07:31.037 Um in 2025 we saw a real acceleration of 00:07:31.120 --> 00:07:36.477 this with coding agents and and white collar work agents. Um I think we'll see 00:07:36.560 --> 00:07:39.437 a third thing soon where we'll have these long-term persistent agents that 00:07:39.520 --> 00:07:40.877 can work like a real virtual 00:07:40.960 --> 00:07:46.157 collaborator constantly helping. But um 00:07:46.240 --> 00:07:49.197 it it's it's it's amazing to me what is 00:07:49.280 --> 00:07:52.797 happening with the boom in 00:07:52.880 --> 00:07:54.637 entrepreneurship and business creation 00:07:54.720 --> 00:07:58.557 because of AI. If I were a country um 00:07:58.640 --> 00:08:02.077 you know many of you are going to take very different approaches to how you 00:08:02.160 --> 00:08:06.237 regulate AI, how you welcome it into your country, how you use it. But for 00:08:06.320 --> 00:08:09.837 sure I think it is non-negotiable that you have to use it. the economic growth 00:08:09.920 --> 00:08:15.117 and benefit to people that can come from this um the value to a country is too 00:08:15.200 --> 00:08:17.997 high to ignore. Uh there will be different ways people try to do this but 00:08:18.080 --> 00:08:20.557 this is like you know the smartest 00:08:20.640 --> 00:08:24.077 people in the world uh decided to move 00:08:24.160 --> 00:08:27.437 to your country to learn your language 00:08:27.520 --> 00:08:32.317 um to work very hard on whatever problem you'd like um to go discover new things 00:08:32.400 --> 00:08:36.317 that you need and you can put as much of this to work as you want. Some of you 00:08:36.400 --> 00:08:39.117 will choose to build data centers. Some of you will choose to rent data centers 00:08:39.200 --> 00:08:41.517 from other people because you don't want them in your country. I think all that's 00:08:41.600 --> 00:08:44.317 fine. Um, but we want to help. We want 00:08:44.400 --> 00:08:47.037 to support not just our company, our 00:08:47.120 --> 00:08:49.597 whole industry. Like we want to bring 00:08:49.680 --> 00:08:53.597 this incredible magic of, 00:08:53.680 --> 00:08:57.997 you know, intelligence in a bottle, education in a bottle um to all of you. 00:08:58.080 --> 00:09:00.237 I think this is like this is an 00:09:00.320 --> 00:09:02.957 incredible moment for the world. what 00:09:03.040 --> 00:09:09.197 people can do with this and the economic 00:09:09.280 --> 00:09:11.597 boom that will come with it. The sort of 00:09:11.680 --> 00:09:14.157 individual increased quality of life. 00:09:14.240 --> 00:09:18.717 People talk about, you know, medical care and what that means, education, 00:09:18.800 --> 00:09:20.797 what that means, services that only very 00:09:20.880 --> 00:09:25.997 rich people used to have access to, now anybody can have with AI. Um, and the 00:09:26.080 --> 00:09:29.757 ability for people to just come up with ideas and get them to fruition so 00:09:29.840 --> 00:09:32.637 quickly, uh, is going to be incredible. 00:09:32.720 --> 00:09:37.117 Do you do you think that um 00:09:37.200 --> 00:09:41.677 your company can create the context 00:09:41.760 --> 00:09:45.277 um that people can use to create uh 00:09:45.360 --> 00:09:47.117 their outcomes or do you think there'll 00:09:47.200 --> 00:09:50.557 be an industry sort of surrounding you 00:09:50.640 --> 00:09:54.237 that will contextualize it? Because if you think about it, you say the the 00:09:54.320 --> 00:09:57.437 smartest people in the world move in to my neighborhood and they're willing to 00:09:57.520 --> 00:10:01.037 work for me. But if I don't know you, that's not really a fun dinner party 00:10:01.120 --> 00:10:03.917 having the smartest people in the world sitting with you and they can't really 00:10:04.000 --> 00:10:08.237 have a conversation, right? So, how do we how do we think about context that 00:10:08.320 --> 00:10:11.757 that's not going to come from us and I don't think that should. Um, we are we 00:10:11.840 --> 00:10:14.477 will provide this this engine, you know, 00:10:14.560 --> 00:10:19.357 we will provide this incredible 00:10:19.440 --> 00:10:24.477 commodity of AI that can you can use for whatever you want. But the people of the 00:10:24.560 --> 00:10:27.517 world, and I have complete trust in the people of the world's creativity and 00:10:27.600 --> 00:10:29.997 ambition, um will have to figure out 00:10:30.080 --> 00:10:32.477 what to ask it to do, what ideas to to 00:10:32.560 --> 00:10:34.957 to throw this at, what what problems and 00:10:35.040 --> 00:10:37.517 what context in a particular society or 00:10:37.600 --> 00:10:42.077 company is important to understand what to do. Um so that won't be our role. I 00:10:42.160 --> 00:10:45.917 don't think we'd be good at it even if it were. Um but we will be able to 00:10:46.000 --> 00:10:48.717 enable everyone around the world to bring that context and those questions 00:10:48.800 --> 00:10:52.477 and those ideas. So 00:10:52.560 --> 00:10:55.837 should should the governments of the world 00:10:55.920 --> 00:10:59.437 think of AI 00:10:59.520 --> 00:11:03.357 and companies your company differently 00:11:03.440 --> 00:11:06.237 than they thought about prior technology 00:11:06.320 --> 00:11:09.517 meaning the if if what you're bringing 00:11:09.600 --> 00:11:12.397 them is you know the smartest people in 00:11:12.480 --> 00:11:14.877 the world an educated 00:11:14.960 --> 00:11:17.597 capacity then if they look at that as 00:11:17.680 --> 00:11:22.477 technology and they uh keep it out of their country or they 00:11:22.560 --> 00:11:24.557 constrain it. Are they are they 00:11:24.640 --> 00:11:30.840 constraining their capacity to think and grow and build? 00:11:31.600 --> 00:11:35.677 I am always tempted to say this time is different. This is really special. You 00:11:35.760 --> 00:11:40.637 know this is and I suspect 00:11:40.720 --> 00:11:42.477 well to zoom out even further than that 00:11:42.560 --> 00:11:46.957 as a kid my understanding of 00:11:47.040 --> 00:11:50.877 the world was that there were these few punctuated moments in history each of 00:11:50.960 --> 00:11:54.957 which was a very different thing. You had um the agricultural revolution and 00:11:55.040 --> 00:11:58.477 then you had the industrial revolution and then you had the computer revolution 00:11:58.560 --> 00:12:02.637 and these were all completely different unrelated things and there were not that 00:12:02.720 --> 00:12:06.797 many lessons to learn from them and they were best understood in isolation and 00:12:06.880 --> 00:12:08.557 even if you zoomed into one pick the 00:12:08.640 --> 00:12:10.877 industrial revolution um the right way 00:12:10.960 --> 00:12:15.037 to understand that was um 20 or 30 00:12:15.120 --> 00:12:18.317 important pieces of technology that were all developed over a period of a couple 00:12:18.400 --> 00:12:19.837 hundred years that came together in this 00:12:19.920 --> 00:12:23.117 important way. uh 00:12:23.200 --> 00:12:26.237 as an adult I look at it as you know it 00:12:26.320 --> 00:12:30.237 has been a single revolution a single exponential of technology. Each 00:12:30.320 --> 00:12:34.317 generation has built on the scaffolding of the generations before put a new 00:12:34.400 --> 00:12:37.917 brick in place. People have been able to see or layer of scaffolding to stick 00:12:38.000 --> 00:12:40.877 with the analogy. People have been able to climb and see a little higher and do 00:12:40.960 --> 00:12:45.277 the new thing. Um, if you think about what it took to get us to this point 00:12:45.360 --> 00:12:49.517 where we can build AI, um, it was all of those previous revolutions uh, stacked 00:12:49.600 --> 00:12:54.397 together. And it was the collective discovery of science throughout society 00:12:54.480 --> 00:12:56.077 and the global economy that we all 00:12:56.160 --> 00:12:59.117 contribute to that made this incredible 00:12:59.200 --> 00:13:04.237 thing possible and now happen. And it is tempting whenever you're in the 00:13:04.320 --> 00:13:06.637 moment to say this is the last revolution. This is the biggest 00:13:06.720 --> 00:13:10.157 revolution. There will be more after this. There will be uh bigger things 00:13:10.240 --> 00:13:13.517 after this. people will use AI to invent whatever thing that comes next. It's 00:13:13.600 --> 00:13:15.037 probably very difficult for us to 00:13:15.120 --> 00:13:19.517 imagine right now. Um and 00:13:19.600 --> 00:13:24.157 and I think the right answer is to look at the history of technology as one 00:13:24.240 --> 00:13:27.437 continuous exponential. The history of the economy and society as one 00:13:27.520 --> 00:13:31.677 continuous exponential. Um and I really mean that. It's not just the technology. 00:13:31.760 --> 00:13:35.597 It's this social fabric and the kind of complex structure 00:13:35.680 --> 00:13:38.477 of the world that allows this technology to flourish and people to figure out 00:13:38.560 --> 00:13:43.037 what to do with it. And you have to embrace it continuously all the way 00:13:43.120 --> 00:13:46.237 through. If you miss out on any part of it, um you're not ready for the next 00:13:46.320 --> 00:13:49.517 thing that will come. If you if you 00:13:49.600 --> 00:13:53.597 don't enable your people and your businesses with AI, uh they will not 00:13:53.680 --> 00:13:57.037 invent the next level of the scaffolding. But there will be more 00:13:57.120 --> 00:14:01.117 things and I think it is tempting to say this one is different and it clearly is 00:14:01.200 --> 00:14:05.597 really important. Um but I would view it as a continuum. 00:14:05.680 --> 00:14:07.517 So 00:14:07.600 --> 00:14:10.557 we all read about the the various 00:14:10.640 --> 00:14:13.677 companies at the frontier. Can you can 00:14:13.760 --> 00:14:15.597 you sort of differentiate them for us? 00:14:15.680 --> 00:14:18.157 What what do you what do you view open 00:14:18.240 --> 00:14:21.517 AI as best at? What what's the you know 00:14:21.600 --> 00:14:24.477 your DNA if you will that separates you 00:14:24.560 --> 00:14:28.877 from the others and what are you great at? And just give us that sort of from 00:14:28.960 --> 00:14:34.237 your perspective. Um 00:14:34.320 --> 00:14:36.957 I think from a from what matters to this 00:14:37.040 --> 00:14:39.917 room um we are the pragmatic centrists. 00:14:40.000 --> 00:14:45.757 We we realize like we want to be we want to be accurate in how we talk 00:14:45.840 --> 00:14:47.917 about what we see in the future. We don't want to fall into the trap of 00:14:48.000 --> 00:14:51.357 blind optimism. We don't want to fall into the trap of dumerism. This is 00:14:51.440 --> 00:14:55.117 clearly complex. There's clearly a combination of both excitement and 00:14:55.200 --> 00:14:57.757 anxiety of transformation and risks that 00:14:57.840 --> 00:15:02.717 the world has to navigate. Um but 00:15:02.800 --> 00:15:05.437 by getting pulled into either extreme we 00:15:05.520 --> 00:15:06.957 are not going to successfully figure out 00:15:07.040 --> 00:15:09.757 the challenges in front of us. Um we 00:15:09.840 --> 00:15:14.717 want to be a reliable partner. Uh we want like we really do view what we are 00:15:14.800 --> 00:15:18.397 doing as a platform as a service to the world. We realize that different people 00:15:18.480 --> 00:15:23.037 around the world are going to use it very differently and we want to put 00:15:23.120 --> 00:15:27.917 sufficient constraints on our system to avoid catastrophic risks. But we really 00:15:28.000 --> 00:15:30.557 do want to respect all of your sovereignity and that you're going to 00:15:30.640 --> 00:15:35.037 use it in very different ways. And we personally may not always like it. Um, 00:15:35.120 --> 00:15:39.357 but I don't like everything that people do with electricity either. And I still 00:15:39.440 --> 00:15:40.877 think everybody should get to have it. 00:15:40.960 --> 00:15:44.920 Um, so, 00:15:45.280 --> 00:15:49.677 oh, and then also I would say like we want the whole global economy to 00:15:49.760 --> 00:15:53.517 succeed. We do not want to be the only company. Um, we do not want to suck up 00:15:53.600 --> 00:15:56.877 all the value. We do not think that is good for global stability. nor do we 00:15:56.960 --> 00:16:00.077 think that's possible for the way that people need to flourish and people need 00:16:00.160 --> 00:16:03.197 to stay at the center of this. Oh, I'll get to that one more thing after this. 00:16:03.280 --> 00:16:07.437 Um, and we want to like enable every 00:16:07.520 --> 00:16:11.037 entrepreneur, every government, every big company in the world to thrive with 00:16:11.120 --> 00:16:15.117 this technology. Uh, the last thing I'll say that I think differentiates us from 00:16:15.200 --> 00:16:18.717 some of the other companies is we are very firmly and very proudly on team 00:16:18.800 --> 00:16:21.277 humanity. Um, we would like to build 00:16:21.360 --> 00:16:23.277 tools for people. We would like people's 00:16:23.360 --> 00:16:26.317 lives to be better. Um, we think that 00:16:26.400 --> 00:16:29.757 people need to be at the center of the 00:16:29.840 --> 00:16:31.917 economy and society and global 00:16:32.000 --> 00:16:34.877 decision-m um, and that people should 00:16:34.960 --> 00:16:37.277 run the future in AI. 00:16:37.360 --> 00:16:42.237 >> So, it it seems that entrepreneurship 00:16:42.320 --> 00:16:44.637 should be really the fire starter of all 00:16:44.720 --> 00:16:47.517 this, right? You're are you the you know 00:16:47.600 --> 00:16:49.997 the the kindling of it? I mean maybe 00:16:50.080 --> 00:16:53.117 sort of tell me how entrepreneurship and 00:16:53.200 --> 00:16:57.677 and uh open AAI sort of work together. 00:16:57.760 --> 00:16:59.917 >> Before OpenAI I used to run a startup 00:17:00.000 --> 00:17:05.597 accelerator and I had always been interested not always after I came to 00:17:05.680 --> 00:17:09.037 college I'd always been interested in entrepreneurship and then having had 00:17:09.120 --> 00:17:11.197 this incredible front row seat I really 00:17:11.280 --> 00:17:14.957 became convinced that entrepreneurship 00:17:15.040 --> 00:17:19.037 was maybe the most important 00:17:19.120 --> 00:17:21.837 novel force in all of capitalism. Um, 00:17:21.920 --> 00:17:25.757 and I fell in love with it and 00:17:25.840 --> 00:17:28.477 watching what new companies were able to 00:17:28.560 --> 00:17:32.637 do and watching how 00:17:32.720 --> 00:17:37.037 large of a percentage of the new ideas that mattered in the tech industry over 00:17:37.120 --> 00:17:40.557 the last, you know, 20 years or whatever came from startups and not big 00:17:40.640 --> 00:17:43.597 companies. Um, now having seen us in 00:17:43.680 --> 00:17:45.437 other industries, I am a huge believer 00:17:45.520 --> 00:17:50.840 in entrepreneurship. I also think that, 00:17:51.040 --> 00:17:53.677 you know, like I talked to a lot of people around the world and they would 00:17:53.760 --> 00:17:56.717 much rather run a great lifestyle business than go work as like a 00:17:56.800 --> 00:18:00.237 corporate cog. And I think that's great, too. Like I think that's just we should 00:18:00.320 --> 00:18:03.357 enable that. Um, one of the things I got 00:18:03.440 --> 00:18:07.277 excited about early with AI, of course AI is going to transform big companies 00:18:07.360 --> 00:18:08.717 and of course AI is going to transform 00:18:08.800 --> 00:18:12.717 governments, but what AI can do 00:18:12.800 --> 00:18:15.757 for entrepreneurs and for people that 00:18:15.840 --> 00:18:18.237 want to take an idea and bring it into 00:18:18.320 --> 00:18:23.037 the world quickly um is really quite amazing. When I was running the startup 00:18:23.120 --> 00:18:26.317 accelerator, it lasted for three months and the work that 00:18:26.400 --> 00:18:30.237 >> it lasted for three months. You were in it for three months and the work that we 00:18:30.320 --> 00:18:32.157 expected a startup to accomplish in 00:18:32.240 --> 00:18:34.637 three months is now probably doable in 00:18:34.720 --> 00:18:38.557 like 17 minutes with codeex. This is 00:18:38.640 --> 00:18:42.637 just amazing. The world is not caught up with what this is going to mean. Not 00:18:42.720 --> 00:18:47.197 just for the economic value that gets created from that, but how quickly 00:18:47.280 --> 00:18:50.637 people can test ideas, build things, get feedback from real customers. We're all 00:18:50.720 --> 00:18:53.597 just going to get much better stuff. 00:18:53.680 --> 00:18:57.197 So, you use the term uh I want to sort 00:18:57.280 --> 00:19:01.357 of flush it out a bit a bit non-negotiable. 00:19:01.440 --> 00:19:04.077 Can you maybe uh help everybody 00:19:04.160 --> 00:19:09.037 understand what you mean about that and and what that means for each of uh our 00:19:09.120 --> 00:19:14.120 partner ministers who are here today with them? 00:19:14.800 --> 00:19:17.357 >> Uh again, I'm biased. I'm trying to talk 00:19:17.440 --> 00:19:19.277 my book here. Forgive me for this 00:19:19.360 --> 00:19:22.077 analogy. Um, I think it would be 00:19:22.160 --> 00:19:26.797 approximately as bad of an idea to say we're not going to have AI in our 00:19:26.880 --> 00:19:30.397 country as it was to say we're not going to have electricity in our country, you 00:19:30.480 --> 00:19:34.797 know, back 100 plus years ago. I think this is something at that level. Um, 00:19:34.880 --> 00:19:36.397 this is something that every person and 00:19:36.480 --> 00:19:40.237 every business and really the entire 00:19:40.320 --> 00:19:43.837 machine of society is going to need to 00:19:43.920 --> 00:19:47.037 deliver the quality of life that your 00:19:47.120 --> 00:19:51.677 citizens should demand. Now, electricity also came with some 00:19:51.760 --> 00:19:55.037 complexities and it took a little while for us to figure out how to make it 00:19:55.120 --> 00:19:58.717 safe, but we made fairly rapid progress as a society safe even though it can do 00:19:58.800 --> 00:20:02.957 a lot of bad things. Um, and at this point, we barely think about 00:20:03.040 --> 00:20:06.957 electricity. You know, we turn these lights on, we expect them to work. You 00:20:07.040 --> 00:20:09.997 know, if this the power shut off and I had to scream because this microphone 00:20:10.080 --> 00:20:14.477 was not working, like I would be annoyed about that. You I won't go through all 00:20:14.560 --> 00:20:16.957 the examples. I see all the phones on the table. Like you just it would be 00:20:17.040 --> 00:20:20.477 hard for society to function well without electricity at this point. But 00:20:20.560 --> 00:20:24.077 we don't think about it much. We think about you know there's light at night. 00:20:24.160 --> 00:20:27.677 That's great. We can hear these people speak in the room without them having to 00:20:27.760 --> 00:20:29.997 yell. That's nice. Our phones can do these amazing things for us. We care 00:20:30.080 --> 00:20:32.957 about our phones. We don't think too much about the electricity that gets 00:20:33.040 --> 00:20:35.677 generated to power them. Um and I think 00:20:35.760 --> 00:20:36.877 the same thing is going to happen with 00:20:36.960 --> 00:20:40.397 AI. We won't, my prediction is we won't 00:20:40.480 --> 00:20:43.197 be talking about AI as much as we do in 00:20:43.280 --> 00:20:48.397 a decade from now because we'll just expect it everywhere. You know, like a 00:20:48.480 --> 00:20:53.677 kid growing up today will never be smarter than AI, but he or she will also 00:20:53.760 --> 00:20:57.997 never have understood a world where every 00:20:58.080 --> 00:21:01.757 product and service that they interact with is not really smart and really 00:21:01.840 --> 00:21:05.997 capable and really helpful. And because we'll just be so used to that, we won't 00:21:06.080 --> 00:21:09.277 think about the AI. will just be like, "Yeah, I want my products to be good." 00:21:09.360 --> 00:21:12.477 You know, if I have an idea, I want to be able to like make it right away. I 00:21:12.560 --> 00:21:16.077 want this software to be created instantly for me. Um, if I have a 00:21:16.160 --> 00:21:17.837 disease, like I expect a cure to get 00:21:17.920 --> 00:21:23.277 discovered very quickly. Uh, so I think countries need to need to embrace this. 00:21:23.360 --> 00:21:24.477 Again, there's many ways to do it. 00:21:24.560 --> 00:21:28.077 People will set different rules. People will decide on what they want to build 00:21:28.160 --> 00:21:30.397 versus rent, but I think everyone will 00:21:30.480 --> 00:21:33.597 need the this new commodity. 00:21:33.680 --> 00:21:38.237 If um if we look forward five years and 00:21:38.320 --> 00:21:43.037 and we come back to North Carolina alto together um 00:21:43.120 --> 00:21:48.840 what what will be holding us back do you think in your opinion? 00:21:50.160 --> 00:21:53.597 Well, there are clearly challenges that 00:21:53.680 --> 00:21:55.837 we have to get right along the way. 00:21:55.920 --> 00:22:00.317 There will be specifics. For example, I'm sure cyber security is a top-of- 00:22:00.400 --> 00:22:03.437 mind issue for many of you. Uh we work with many of the governments in the 00:22:03.520 --> 00:22:05.837 room. Uh we work with the state of North 00:22:05.920 --> 00:22:08.157 Carolina. Uh we work with companies big 00:22:08.240 --> 00:22:12.637 and small and we are 00:22:12.720 --> 00:22:16.957 I think we have a big challenge in front of us like to not to not try to sidestep 00:22:17.040 --> 00:22:20.877 this one. I think some things are going to go very wrong with cyber security 00:22:20.960 --> 00:22:25.197 unless people act quite urgently. There will be other challenges in the next 00:22:25.280 --> 00:22:28.397 five years. People talk about uh biocurity and the the things we're going 00:22:28.480 --> 00:22:31.997 to face there. um 00:22:32.080 --> 00:22:33.837 there will be bigger ones yet to come 00:22:33.920 --> 00:22:35.757 and I think if we don't successfully 00:22:35.840 --> 00:22:39.037 navigate those um 00:22:39.120 --> 00:22:41.037 society will not be able to adopt AI to 00:22:41.120 --> 00:22:44.477 the degree or to the speed with which 00:22:44.560 --> 00:22:47.677 people deserve. So getting 00:22:47.760 --> 00:22:51.837 successfully navigating the risks and downsides in front of us is something 00:22:51.920 --> 00:22:55.517 that could go very wrong and I think will set this technology back could set 00:22:55.600 --> 00:22:58.717 this technology back a great deal. Um, 00:22:58.800 --> 00:23:00.157 another thing that could go wrong is too 00:23:00.240 --> 00:23:05.037 much concentration of power. Um, 00:23:05.120 --> 00:23:10.317 AI should be in many ways AI should be an equalizing force in the world. You 00:23:10.400 --> 00:23:13.437 talked about what it means to have like a reset on the educational capacity 00:23:13.520 --> 00:23:16.877 available to everybody. Um, you can also 00:23:16.960 --> 00:23:21.437 see worlds where AI concentrates because 00:23:21.520 --> 00:23:24.957 some people make early investments and others don't and there's, you know, bad 00:23:25.040 --> 00:23:28.877 inequities with comput or whatever. uh that'd be bad. So I think if we fail to 00:23:28.960 --> 00:23:31.677 make AI abundant, it ends up being 00:23:31.760 --> 00:23:35.917 something that is very hotly competed over, rich people spend a huge amount of 00:23:36.000 --> 00:23:40.717 money on and um 00:23:40.800 --> 00:23:44.957 that would be a bad outcome. So I think we need to make it abundant so that we 00:23:45.040 --> 00:23:48.877 get the the benefits of lifting everybody up and not much worsening 00:23:48.960 --> 00:23:52.157 inequality. >> Right. As you said, if if the smartest 00:23:52.240 --> 00:23:54.397 people in the world learned your 00:23:54.480 --> 00:23:56.717 language and all agreed to move to your 00:23:56.800 --> 00:23:59.357 country and work on whatever you wanted 00:23:59.440 --> 00:24:02.157 them to work on, that that has to be the 00:24:02.240 --> 00:24:07.277 most incredible thing. Do you think that 00:24:07.360 --> 00:24:10.957 since you'll have all the smartest people speaking all those languages, 00:24:11.040 --> 00:24:12.717 can't they figure out a way do you think 00:24:12.800 --> 00:24:15.597 over time to deliver the same amount of 00:24:15.680 --> 00:24:19.197 compute more inexpensively without the 00:24:19.280 --> 00:24:24.317 same use of power without the same I mean aren't these the kind of things 00:24:24.400 --> 00:24:28.957 >> that are going to happen that will democratize it by sort of by itself in a 00:24:29.040 --> 00:24:32.717 way >> it it will still require huge investment 00:24:32.800 --> 00:24:37.037 but yes we will continue to make things vastly more efficient and we'll figure 00:24:37.120 --> 00:24:40.557 out how to make more and better chips and we'll figure out how to make smarter 00:24:40.640 --> 00:24:44.797 algorithms. All of that will happen. But to put some numbers on the magnitude of 00:24:44.880 --> 00:24:46.797 how much we we need this to happen. 00:24:46.880 --> 00:24:49.357 Again, I know these numbers in front of me. I'm not going to get this exactly 00:24:49.440 --> 00:24:53.037 right. Um but forgive me for the approximation. 00:24:53.120 --> 00:24:57.837 At the beginning of 2020, um the token leader in the world was an OpenAI 00:24:57.920 --> 00:24:59.997 employee using roughly a 100,000 tokens 00:25:00.080 --> 00:25:02.157 a month. 00:25:02.240 --> 00:25:05.437 Um In the middle of 2026, 00:25:05.520 --> 00:25:07.117 the token leader that I know of, there 00:25:07.200 --> 00:25:09.837 may be somebody using more than this at 00:25:09.920 --> 00:25:12.717 outside of OpenAI, um was using hundreds 00:25:12.800 --> 00:25:16.077 of billions of tokens a month. So, you 00:25:16.160 --> 00:25:20.317 know, a multi-millionx increase in six and a half years. And the average person 00:25:20.400 --> 00:25:22.557 on Earth had gone from using effectively 00:25:22.640 --> 00:25:24.637 zero tokens to 100,000 tokens a month. 00:25:24.720 --> 00:25:27.357 That was the global per capita thing. 00:25:27.440 --> 00:25:30.477 If we just naively project this forward again, which seems like a reasonable 00:25:30.560 --> 00:25:35.997 thing to do to me. Um, and I hate tokens as a unit here. It's a dumb one. No one 00:25:36.080 --> 00:25:39.357 should care, but like let's just take that as shorthand for amount of 00:25:39.440 --> 00:25:42.557 intelligence. Then 00:25:42.640 --> 00:25:46.717 in another six and a half years, the average per capita use of tokens will go 00:25:46.800 --> 00:25:51.437 up to, let's round it down to 100 billion tokens a month. And the person 00:25:51.520 --> 00:25:55.837 using the most in the world will be using 00:25:55.920 --> 00:25:58.397 you know 100 quadrillion tokens a month. 00:25:58.480 --> 00:26:02.557 Uh or one quadrillion 100 trillion like 00:26:02.640 --> 00:26:05.997 >> no I was right 100 trillion tokens a month. 100 quadrillion tokens a month. 00:26:06.080 --> 00:26:11.517 Um even with great gains on efficiency we 00:26:11.600 --> 00:26:14.637 are going to need to build much more infrastructure 00:26:14.720 --> 00:26:19.277 unless we want this to be a highly priced commodity which we don't. We want 00:26:19.360 --> 00:26:21.597 this to be abundant. We want this to be inexpensive. We want people to just be 00:26:21.680 --> 00:26:23.997 able to use this for whatever they need. 00:26:24.080 --> 00:26:27.437 >> So that's the driving force. If we if we 00:26:27.520 --> 00:26:32.077 globally can build enough then it will 00:26:32.160 --> 00:26:35.437 intelligence will be broadly distributed 00:26:35.520 --> 00:26:38.557 uh more democratized across the world. 00:26:38.640 --> 00:26:42.237 Uh the world will be educated or they'll 00:26:42.320 --> 00:26:46.237 have the access to education. Is that 00:26:46.320 --> 00:26:48.237 >> to to go back to the analogy of 00:26:48.320 --> 00:26:51.117 electricity? Um there's this amazing 00:26:51.200 --> 00:26:54.557 chart of what it cost to have an hour of 00:26:54.640 --> 00:26:59.597 light at night. And there was a time, one of the 00:26:59.680 --> 00:27:01.517 earliest data points you can find is 00:27:01.600 --> 00:27:04.557 that it it took an average of 5 hours of 00:27:04.640 --> 00:27:06.637 labor at the wage at the time to be able 00:27:06.720 --> 00:27:10.157 to pay for one hour of light at night. 00:27:10.240 --> 00:27:14.237 Electricity came along. It fell. We figured out how to make more 00:27:14.320 --> 00:27:17.357 electricity. So there was like, you know, the supply and demand change 00:27:17.440 --> 00:27:22.317 curve. It fell again. We figured out how to make vastly more efficient light 00:27:22.400 --> 00:27:26.237 bulbs. We went from incandescent lights to LEDs. A whole bunch of other things 00:27:26.320 --> 00:27:29.117 happened. And now an hour of light at 00:27:29.200 --> 00:27:33.277 night is not something that people think of at all about the cost. I don't even 00:27:33.360 --> 00:27:36.077 know what the number is, but someone in here can do it. It's it's like very 00:27:36.160 --> 00:27:38.077 cheap to run, you know, a five watt bulb 00:27:38.160 --> 00:27:42.477 for an hour. Um the 00:27:42.560 --> 00:27:45.997 if that hadn't happened if we and all of those things together had to happen lots 00:27:46.080 --> 00:27:48.877 of electricity generation huge innovation on the efficiency all of 00:27:48.960 --> 00:27:53.917 these things. If that hadn't happened then rich people get light at night and 00:27:54.000 --> 00:27:56.477 they can study more they have a compounding effect they can learn more 00:27:56.560 --> 00:27:59.197 whatever and and you know poor people 00:27:59.280 --> 00:28:03.597 don't. So I think it's been very good that energy has become abundant and that 00:28:03.680 --> 00:28:05.837 light at night has become relatively 00:28:05.920 --> 00:28:09.437 like shockingly inexpensive. Um, if that 00:28:09.520 --> 00:28:11.997 same thing happens with AI and anybody 00:28:12.080 --> 00:28:13.917 on earth, rich, poor, whatever, can use 00:28:14.000 --> 00:28:19.197 AI to get the abundant services that they need and to express their ideas and 00:28:19.280 --> 00:28:23.357 to start their businesses and whatever else. And we fulfill our mission of 00:28:23.440 --> 00:28:28.157 making AI so inexpensive and so smart that people are just like they think 00:28:28.240 --> 00:28:31.597 about it like they think of the cost of an hour at night of light at night. I 00:28:31.680 --> 00:28:35.117 think that'd be really wonderful. And then this huge reset and democratization 00:28:35.200 --> 00:28:37.597 of education you're talking about can happen. 00:28:37.680 --> 00:28:41.997 So are you you know there are look there 00:28:42.080 --> 00:28:44.557 are doomers out there who worry about 00:28:44.640 --> 00:28:47.277 everything but we know that if you look 00:28:47.360 --> 00:28:49.117 back through the industrial re as you 00:28:49.200 --> 00:28:51.597 said it it always seems to work out 00:28:51.680 --> 00:28:54.557 fine. Are you uh fundamentally 00:28:54.640 --> 00:28:57.277 optimistic to what your products will do 00:28:57.360 --> 00:29:00.797 to the world and will serve the world? 00:29:00.880 --> 00:29:02.557 I I think part of the reason it tends to 00:29:02.640 --> 00:29:06.477 work out fine is people worry and they 00:29:06.560 --> 00:29:09.597 stress and they look at the downsides and they try to address those and they 00:29:09.680 --> 00:29:12.477 try to learn how we avoid accidents. So 00:29:12.560 --> 00:29:15.117 if we go back to an early example even 00:29:15.200 --> 00:29:17.597 before electricity of a technology fire 00:29:17.680 --> 00:29:22.477 was incredible incredible for human evolution incredible for letting us cook 00:29:22.560 --> 00:29:25.757 our food and heat our homes at night and whatever else. It also burnt down whole 00:29:25.840 --> 00:29:30.317 cities a few times and people really worried about that because we don't want 00:29:30.400 --> 00:29:32.717 cities to burn down and we figured out 00:29:32.800 --> 00:29:37.277 fire safety and through a lot of human 00:29:37.360 --> 00:29:39.277 ingenuity um through a lot of good 00:29:39.360 --> 00:29:41.597 policy definitely through learning from 00:29:41.680 --> 00:29:46.317 some very bad accidents we figured out how to make it safer. Now had we just 00:29:46.400 --> 00:29:49.837 said like yeah you know some cities are going to burn down whatever we got other 00:29:49.920 --> 00:29:52.637 problems to work on probably wouldn't have figured that out and that wouldn't 00:29:52.720 --> 00:29:56.637 have been as good. Um, so I think worry 00:29:56.720 --> 00:29:59.437 is worry is how we anticipate problems 00:29:59.520 --> 00:30:03.917 and how we from a matter of policy and a matter of technology figure out how to 00:30:04.000 --> 00:30:07.837 address them. But I am optimistic that we can mitigate the downsides of this 00:30:07.920 --> 00:30:12.157 technology. And if you look at what's happened, you know, if you look at what 00:30:12.240 --> 00:30:15.837 the doomers were saying 5 years ago about the chances of us being able to 00:30:15.920 --> 00:30:20.397 make AI of this level, of this sort of near human, sometimes superhuman 00:30:20.480 --> 00:30:25.037 capacity level, you know, broadly considered safe and reliable. Not 00:30:25.120 --> 00:30:26.237 perfect. There have been accidents. 00:30:26.320 --> 00:30:29.437 There will be accidents, of course. I don't think most people would have 00:30:29.520 --> 00:30:31.357 thought it was going to go this well as 00:30:31.440 --> 00:30:35.037 it has so far. And I I believe in human 00:30:35.120 --> 00:30:39.757 ingenuity and the ability for to solve very difficult, very complex 00:30:39.840 --> 00:30:42.877 problems. And I think I think we'll keep doing that. 00:30:42.960 --> 00:30:46.077 And I think the economic value and the 00:30:46.160 --> 00:30:51.197 personal value is so huge. Um I met someone just this morning who told and 00:30:51.280 --> 00:30:54.317 again I'm sure you've all heard stories like this that they wouldn't be alive 00:30:54.400 --> 00:30:57.357 without CHPT. Um and that they had a 00:30:57.440 --> 00:31:01.517 very difficult disease and you know got a diagnosis that doctors weren't able to 00:31:01.600 --> 00:31:03.197 >> give them. We shouldn't take that away. 00:31:03.280 --> 00:31:06.317 >> No. But we need to figure out how to make the technology safe so that people 00:31:06.400 --> 00:31:08.157 can have more of that. 00:31:08.240 --> 00:31:10.797 >> Great. Sam, thank you for joining us. It 00:31:10.880 --> 00:31:13.357 was really an incredible conversation. 00:31:13.440 --> 00:31:17.917 Um 2:00, so 2 p.m. in the ministerial 00:31:18.000 --> 00:31:21.437 room where we will have uh Alex Karp 00:31:21.520 --> 00:31:24.957 will be our next guest uh speaking to us. 00:31:25.040 --> 00:31:28.157 >> He's a good one. >> He's a good one. But uh 00:31:28.240 --> 00:31:31.577 >> thank you Sam for joining us. We we really appreciate it. Thank you. 00:31:31.660 --> 00:31:36.120 [applause] Thanks.