1 00:00:00,480 --> 00:00:03,677 All right, everyone. Um, so how many of 2 00:00:03,760 --> 00:00:06,317 you uh are Stanford students? Raise your 3 00:00:06,400 --> 00:00:08,957 hand. Okay, pretty good. We got a lot. 4 00:00:09,040 --> 00:00:14,157 Um, how many of you in your like years at Stanford lived in Donner? 5 00:00:14,240 --> 00:00:17,917 >> We also got a few. >> I think we got [laughter] a few. 6 00:00:18,000 --> 00:00:20,637 >> Well, Sam was actually a Donner resident 7 00:00:20,720 --> 00:00:24,157 and uh yeah, so why don't we start there? Do you remember moving in on your 8 00:00:24,240 --> 00:00:25,597 first day at Stanford over here? 9 00:00:25,680 --> 00:00:28,717 >> I do. I actually I haven't been back to campus in so long. I'm feeling inspired 10 00:00:28,800 --> 00:00:30,717 to like go walk by after this cuz I have 11 00:00:30,800 --> 00:00:35,997 like such fond memories of that first day. Um it was like, you know, I'm sure 12 00:00:36,080 --> 00:00:39,277 like for all of you it was a very strange experience, but it was like 13 00:00:39,360 --> 00:00:41,037 awesome as soon as I got here. 14 00:00:41,120 --> 00:00:45,197 >> Do you have any favorite memories? Any building climbing, tunnel diving? 15 00:00:45,280 --> 00:00:50,157 >> Uh yeah, I I was mentioning this backstage. I like it's sort of, you 16 00:00:50,240 --> 00:00:52,477 know, when you go to a place you haven't been in a while, all these memories that 17 00:00:52,560 --> 00:00:55,917 you forgot about come come back. Uh, and 18 00:00:56,000 --> 00:01:00,077 very near here, uh, like early in freshman year, a bunch of us went steam 19 00:01:00,160 --> 00:01:01,197 tunneling, which I don't know if people 20 00:01:01,280 --> 00:01:03,197 still do, but was like a cool right of 21 00:01:03,280 --> 00:01:07,277 passage at the time, uh, one of the 22 00:01:07,360 --> 00:01:11,357 people that we went with was like not told you had to put the cover back on. 23 00:01:11,440 --> 00:01:14,797 So, they kind of left the cover off. We went down and looked around and whatever 24 00:01:14,880 --> 00:01:16,957 and came back and there was a car, 25 00:01:17,040 --> 00:01:22,317 professor's car that had like fallen into the open manhole. Uh, I thought I 26 00:01:22,400 --> 00:01:24,957 was going to get like, you know, kicked out. it was like first few weeks of 27 00:01:25,040 --> 00:01:27,117 school. Um, but I had totally forgotten 28 00:01:27,200 --> 00:01:29,357 about that and it's fun to see. Um, my 29 00:01:29,440 --> 00:01:32,637 memories are like Stanford is awesome. 30 00:01:32,720 --> 00:01:34,877 Uh, and I just have like very happy 31 00:01:34,960 --> 00:01:39,917 memories of that time. Uh, and kind of 32 00:01:40,000 --> 00:01:44,157 even the parts that I didn't think I was going to like at the time like staying 33 00:01:44,240 --> 00:01:48,317 up all night working in like the basement of Gates are very happy memories now. 34 00:01:48,400 --> 00:01:51,437 >> I think that's like a core CS student type of thing. I have also pulled all 35 00:01:51,520 --> 00:01:54,717 nighters in Gates. It's >> there. It's a good memory in retrospect. 36 00:01:54,800 --> 00:01:59,037 >> Yeah. And every time I walk by Gates, I just think back to that all nighter and 37 00:01:59,120 --> 00:02:04,077 uh you know, Gates Gates comes with certain memories now. [laughter] >> Yep. 38 00:02:04,160 --> 00:02:07,757 >> Did you have any specific clubs that you were a part of here at Stanford? 39 00:02:07,840 --> 00:02:12,477 >> Um I did a few, but the honestly the most 40 00:02:12,560 --> 00:02:16,477 the best thing I did was just like hacking on projects with friends. Like 41 00:02:16,560 --> 00:02:19,837 that was if I could go back and do it, I would have done like less clubs and more 42 00:02:19,920 --> 00:02:22,477 just building stuff. Yeah. All right. 43 00:02:22,560 --> 00:02:25,277 Um, well, we went through the archives 44 00:02:25,360 --> 00:02:28,077 uh of Sam's time here at Stanford, and I 45 00:02:28,160 --> 00:02:32,877 don't know if you guys know the, uh, club faces. It's the student 46 00:02:32,960 --> 00:02:34,877 entrepreneurship club here on campus. 47 00:02:34,960 --> 00:02:40,637 And they found this document, the business symposium. Um, whoa. And it was 48 00:02:40,720 --> 00:02:43,517 for the next [laughter] generation of business leaders. So, this is from 49 00:02:43,600 --> 00:02:45,837 spring 2005. This is like right before 50 00:02:45,920 --> 00:02:50,637 you dropped out. Uh and in this document 51 00:02:50,720 --> 00:02:54,077 there's this one entry uh and it says 52 00:02:54,160 --> 00:02:58,557 Vienno and Viendo is a service for cell phones that provides personalized 53 00:02:58,640 --> 00:03:02,717 information about hot venues, a map of locations of friends nearby and 54 00:03:02,800 --> 00:03:07,197 compatible new friends and dates in the area. And keep in mind this is 2005 so 55 00:03:07,280 --> 00:03:08,797 this is like before the iPhone came out. 56 00:03:08,880 --> 00:03:11,357 So this is kind of like a crazy idea at 57 00:03:11,440 --> 00:03:14,877 the time. Uh, and it's co-founded by 58 00:03:14,960 --> 00:03:16,957 four students as you can see here. Um, 59 00:03:17,040 --> 00:03:22,957 Peter, Alec, Nick, and, uh, this guy named Sam up there as well. Um, is this 60 00:03:23,040 --> 00:03:23,517 looped? 61 00:03:23,600 --> 00:03:27,677 >> Yeah. Uh, I forgot about this also, but 62 00:03:27,760 --> 00:03:31,517 the the important thing here is not 63 00:03:31,600 --> 00:03:33,437 about building stuff, but it's about how 64 00:03:33,520 --> 00:03:37,037 much Y Combinator did to change what it 65 00:03:37,120 --> 00:03:41,437 took to get a project funded. Um, I was like, you know, I came here, I was not 66 00:03:41,520 --> 00:03:45,037 that interested in startups and then it kind of like I was like it was in the 67 00:03:45,120 --> 00:03:49,437 water and I got excited about startups, but it was a super different time then. 68 00:03:49,520 --> 00:03:52,717 And you would like talk to people and they'd be like, "Oh, you need like a 69 00:03:52,800 --> 00:03:56,797 business person. You can't just you got to like get these connections to VCs. It 70 00:03:56,880 --> 00:04:00,077 was like, you know, you got to like get serious. You need like a bunch of MBAs. 71 00:04:00,160 --> 00:04:03,997 You got to write a business plan. And, you know, it's like you need to like 72 00:04:04,080 --> 00:04:06,237 raise this big round and you're going to have these VCs. They're going to tell 73 00:04:06,320 --> 00:04:10,237 you what to do and whatever." And I was like, "Okay, I guess I don't know any 74 00:04:10,320 --> 00:04:12,797 better." And these smart people are saying that's what you have to do, so 75 00:04:12,880 --> 00:04:16,397 like I'll believe it. Um, this was right 76 00:04:16,480 --> 00:04:19,037 around this time is when Y Combinator 77 00:04:19,120 --> 00:04:22,637 launched. Uh, and I'm sure you all 78 00:04:22,720 --> 00:04:26,797 already think this from just hearing from Gary, but like if you want to start 79 00:04:26,880 --> 00:04:28,157 a startup, you should definitely do Y 80 00:04:28,240 --> 00:04:32,317 Combinator. Uh, but and it's an amazing 81 00:04:32,400 --> 00:04:35,357 thing. Uh but before YC like in the dark 82 00:04:35,440 --> 00:04:40,157 ages, this is what you had to do. Like you had to go do business plan 83 00:04:40,240 --> 00:04:43,277 competitions which are a business plans are a waste of time. Business plan 84 00:04:43,360 --> 00:04:44,957 competitions even worse waste of time. 85 00:04:45,040 --> 00:04:48,157 Um but this is kind of what you were stuck with. Like this was how you could 86 00:04:48,240 --> 00:04:51,677 get things out in the world. And in the 87 00:04:51,760 --> 00:04:56,317 20 years since uh things have changed like unbelievably for the better. 88 00:04:56,400 --> 00:04:59,677 >> Yeah. So how long were you working on this idea before like you discovered Y 89 00:04:59,760 --> 00:05:03,517 Combinator? Uh I was like hacking on like a project. That was another thing 90 00:05:03,600 --> 00:05:07,517 like it didn't this like many other things did not start intending to be a 91 00:05:07,600 --> 00:05:10,877 company in any way. This was like a sort of late night dorm project. 92 00:05:10,960 --> 00:05:14,237 >> Oh wow. >> But over like a while and then it got serious when YC 93 00:05:14,320 --> 00:05:16,237 >> you remember if you won this competition. 94 00:05:16,320 --> 00:05:19,117 >> We did. Yeah. >> That's pretty good. 95 00:05:19,200 --> 00:05:23,117 >> Um so then how did you end up getting in contact with like YC? You were in the 96 00:05:23,200 --> 00:05:26,237 first batch, right? >> Um 97 00:05:26,320 --> 00:05:27,917 you just like applied online. I actually 98 00:05:28,000 --> 00:05:30,637 I think it was maybe it was like you 99 00:05:30,720 --> 00:05:35,197 filled out this weird form that PG had that you had to like 100 00:05:35,280 --> 00:05:37,037 paste into an email and not change any 101 00:05:37,120 --> 00:05:43,560 formatting because he had some weird parser for it. I think it was like that. 102 00:05:44,000 --> 00:05:46,397 >> So all three of us are sophomores right 103 00:05:46,480 --> 00:05:48,397 now. Um could you walk us through a day 104 00:05:48,480 --> 00:05:51,597 in your life back then? 105 00:05:51,680 --> 00:05:56,897 >> Um I mean it was really boring like there like 106 00:05:56,980 --> 00:05:59,437 >> [laughter] >> I don't know. I like went to class or 107 00:05:59,520 --> 00:06:01,757 didn't and I like hung out with my 108 00:06:01,840 --> 00:06:06,557 friends and I built stuff >> like 109 00:06:06,640 --> 00:06:11,517 Yeah. Again, very fun memories, but it was like very like simple, lovely, 110 00:06:11,600 --> 00:06:15,277 boring time. >> Yeah. >> So, when you were considering dropping 111 00:06:15,360 --> 00:06:18,397 out, did you feel like it was it was risky? I mean, were you excited? I mean, 112 00:06:18,480 --> 00:06:22,717 it sounds like school maybe wasn't the whole uh environment that you were 113 00:06:22,800 --> 00:06:27,037 looking for. What was kind of like your your process behind that? 114 00:06:27,120 --> 00:06:31,357 Uh, look, I feel a little bad saying this 115 00:06:31,440 --> 00:06:35,437 because obviously it worked out for me and I don't, you know, survivorship 116 00:06:35,520 --> 00:06:37,517 survivorship bias is a real thing. 117 00:06:37,600 --> 00:06:41,917 >> Uh, I didn't think at the time and I don't think now that dropping out is 118 00:06:42,000 --> 00:06:44,317 like a particularly risky thing to do. 119 00:06:44,400 --> 00:06:46,877 >> Uh, I don't know if Stanford still has this, but at the time it was like you 120 00:06:46,960 --> 00:06:51,037 could stop out for eight years and come back kind of no questions asked. um 121 00:06:51,120 --> 00:06:53,117 which I think is a great a great policy 122 00:06:53,200 --> 00:06:55,357 and it certainly made it easier to get 123 00:06:55,440 --> 00:06:58,157 my parents slightly on board with the 124 00:06:58,240 --> 00:07:01,597 idea. Um 125 00:07:01,680 --> 00:07:04,637 I think you should just work on whatever 126 00:07:04,720 --> 00:07:07,357 the sort of most interesting highest 127 00:07:07,440 --> 00:07:12,317 rate of learning highest impact thing you can is. And if that's being in 128 00:07:12,400 --> 00:07:15,117 school and learning a ton of stuff and being around interesting people and 129 00:07:15,200 --> 00:07:18,637 working on interesting stuff, great. If it's doing a startup, great. if it's 130 00:07:18,720 --> 00:07:20,797 dropping out to do AI research or 131 00:07:20,880 --> 00:07:23,677 whatever. Also great. Um 132 00:07:23,760 --> 00:07:26,077 I think one of the trends of the last 133 00:07:26,160 --> 00:07:28,237 few decades, couple decades at least, 134 00:07:28,320 --> 00:07:31,117 has been that um 135 00:07:31,200 --> 00:07:35,837 the default path has gotten 136 00:07:35,920 --> 00:07:39,757 relatively less good. And the option space of what you can do and still be 137 00:07:39,840 --> 00:07:43,597 really successful and happy and have a great impact on the world has increased 138 00:07:43,680 --> 00:07:45,357 and increased and increased. And as the 139 00:07:45,440 --> 00:07:48,557 society gets richer and freer and more 140 00:07:48,640 --> 00:07:51,597 capability with a better technology tree 141 00:07:51,680 --> 00:07:53,197 and 142 00:07:53,280 --> 00:07:56,077 like I think that will just keep going. 143 00:07:56,160 --> 00:07:59,597 So uh it didn't feel like a crazy or 144 00:07:59,680 --> 00:08:04,077 risky thing to me at the time. Uh I think if I were a student today it would 145 00:08:04,160 --> 00:08:07,117 feel like even more of a that's totally something I could consider. 146 00:08:07,200 --> 00:08:09,997 >> Yeah. >> My parents didn't like it. 147 00:08:10,080 --> 00:08:14,477 >> Yeah, that makes sense. That makes sense. Wow. So you raised over like $30 148 00:08:14,560 --> 00:08:17,917 million for this idea which you know is like kind of considered like seed round 149 00:08:18,000 --> 00:08:19,997 today. How >> I mean I was going to say that's like a 150 00:08:20,080 --> 00:08:23,197 pre pre [laughter] like it was a lot of money at the time. 151 00:08:23,280 --> 00:08:26,557 >> Yeah. Um how how did you do this? I mean what what did 152 00:08:26,640 --> 00:08:30,157 >> 17 GPUs now or something? >> Wait. No exactly right. 153 00:08:30,240 --> 00:08:34,717 >> Well one thing that was is cool about 154 00:08:34,800 --> 00:08:38,077 Stanford and particularly so at the time is it had like a good on-ramp. It was 155 00:08:38,160 --> 00:08:41,277 like very well connected to the whole like VC machinery. 156 00:08:41,360 --> 00:08:45,037 >> Yeah. and Y cominator also was helpful there. 157 00:08:45,120 --> 00:08:47,997 >> Yeah. >> Um 158 00:08:48,080 --> 00:08:50,317 it was maybe there were less VC firms 159 00:08:50,400 --> 00:08:53,357 and less money at the time but they were 160 00:08:53,440 --> 00:08:55,437 extremely focused on Stanford as like a 161 00:08:55,520 --> 00:08:57,997 source of ideas and people. 162 00:08:58,080 --> 00:09:00,637 >> Do you remember any like stories uh from 163 00:09:00,720 --> 00:09:02,877 that YC batch or from founding? Like did 164 00:09:02,960 --> 00:09:09,917 it go really well all at first or like what was that process like? Um, 165 00:09:10,000 --> 00:09:11,197 actually it felt like still being in 166 00:09:11,280 --> 00:09:13,757 college. Everybody was like pretty young 167 00:09:13,840 --> 00:09:17,357 at the time. Um, we kind of like would 168 00:09:17,440 --> 00:09:19,757 stay up late talking or working on 169 00:09:19,840 --> 00:09:24,877 computers and it it it felt like it felt shockingly not that different. 170 00:09:24,960 --> 00:09:28,157 And I've heard a lot of people say this that have done YC or not that have 171 00:09:28,240 --> 00:09:30,157 dropped out to do startups. Like if 172 00:09:30,240 --> 00:09:32,637 you're a college student kind of like if 173 00:09:32,720 --> 00:09:37,837 you're like a CSish college student and then you do a startup the fir like the 174 00:09:37,920 --> 00:09:40,477 last year that you did school and the first year of doing a startup eventually 175 00:09:40,560 --> 00:09:43,917 it gets very different. They're like shockingly the same. 176 00:09:44,000 --> 00:09:45,517 >> So I mean we have a lot of students here 177 00:09:45,600 --> 00:09:48,637 from Stanford but we also have a lot of non-stamford students and I agree with 178 00:09:48,720 --> 00:09:52,237 you. I do think Stanford has a pipeline into like this PC culture and all this 179 00:09:52,320 --> 00:09:55,277 money, right? I mean what would you recommend to students coming from other 180 00:09:55,360 --> 00:09:59,117 schools if they were interested in going into this the same process? That was the 181 00:09:59,200 --> 00:10:02,317 case at the time. I think that's no longer the case. I think the sort of 182 00:10:02,400 --> 00:10:06,317 like Stanford special access I mean maybe there's still some of it but I 183 00:10:06,400 --> 00:10:09,757 think it's like mostly over maybe totally over maybe even slight negative 184 00:10:09,840 --> 00:10:11,277 selection for Stanford students at this 185 00:10:11,360 --> 00:10:15,757 point. Um the I think YC did a lot for 186 00:10:15,840 --> 00:10:19,197 this but also like VC just matured as an 187 00:10:19,280 --> 00:10:22,797 industry and there's so much money in the world that wants to fund startups 188 00:10:22,880 --> 00:10:26,717 now. Um that I think it's no longer the case at all that it's like a very 189 00:10:26,800 --> 00:10:30,477 Stanfordy thing. Okay. So, fast forward a few years and 190 00:10:30,560 --> 00:10:32,877 you're now the president of YC and then 191 00:10:32,960 --> 00:10:35,837 all of a sudden, boom, open AI is born. 192 00:10:35,920 --> 00:10:38,077 Um, what happened? How did you guys come 193 00:10:38,160 --> 00:10:39,197 up with the idea? 194 00:10:39,280 --> 00:10:42,877 >> Uh, so I was studying AI when I was here 195 00:10:42,960 --> 00:10:47,437 and it was not working at the time at all. Like I thought AI would be the 196 00:10:47,520 --> 00:10:51,917 coolest, most important thing, but it was I worked in the AI lab the summer 197 00:10:52,000 --> 00:10:53,917 between uh my freshman and sophomore 198 00:10:54,000 --> 00:10:59,277 year and I was like, okay, these are this is not it. Um, and I got distracted 199 00:10:59,360 --> 00:11:01,837 and did a startup with something else and then I became an investor for a 200 00:11:01,920 --> 00:11:05,757 while. But I always thought that uh AI 201 00:11:05,840 --> 00:11:08,157 was at least one of the two or three 202 00:11:08,240 --> 00:11:10,317 most important things that could happen. 203 00:11:10,400 --> 00:11:13,757 Um, and 204 00:11:13,840 --> 00:11:17,357 you know, I was like paying attention to it. I This is like a a thing I recommend 205 00:11:17,440 --> 00:11:21,277 and something I have heard many people describe over the course of your career, 206 00:11:21,360 --> 00:11:25,757 which is you get conviction on the things that you care about early, a 207 00:11:25,840 --> 00:11:28,157 super set of them. Some you'll lose 208 00:11:28,240 --> 00:11:30,557 interest in, some will develop more than 209 00:11:30,640 --> 00:11:33,917 you thought, but the the landscape of 210 00:11:34,000 --> 00:11:35,757 the world just is going to keep changing 211 00:11:35,840 --> 00:11:41,037 as you get more experience. And continuing to focus and like stick your 212 00:11:41,120 --> 00:11:45,597 head up and look and not get into tunnel vision on one thing seems like it's 213 00:11:45,680 --> 00:11:47,677 really important and really correlates 214 00:11:47,760 --> 00:11:50,317 with success over a long period of time. 215 00:11:50,400 --> 00:11:52,797 And you know, so I was like paying 216 00:11:52,880 --> 00:11:53,997 attention to all the areas I was 217 00:11:54,080 --> 00:11:57,677 interested in. In 2012, 218 00:11:57,760 --> 00:12:01,197 uh, the Alex Net paper happened and I, like everybody else, was like, "Oh, 219 00:12:01,280 --> 00:12:04,797 that's kind of cool or very cool, but it didn't update me as much as it should 220 00:12:04,880 --> 00:12:07,517 have." Like, you know, strong priors are 221 00:12:07,600 --> 00:12:11,517 a real factor. I I then watched over the 222 00:12:11,600 --> 00:12:15,117 next couple of years as 223 00:12:15,200 --> 00:12:20,397 it deep learning models kept getting better with scale and I still didn't 224 00:12:20,480 --> 00:12:25,437 update enough but now I was at least like this is like an asteroid coming and 225 00:12:25,520 --> 00:12:30,077 a crazy thing that people are not paying more attention to and so in 2014 we 226 00:12:30,160 --> 00:12:34,717 started saying you know we should really do something about this. Um it was 227 00:12:34,800 --> 00:12:38,397 incredibly unpopular at the time. Like we had said we want to make an AGI lab 228 00:12:38,480 --> 00:12:41,517 and all the like elders of the field were like you're insane. You're a 229 00:12:41,600 --> 00:12:45,437 scammer. This is like you know AGI is 100 years away. None of these approaches 230 00:12:45,520 --> 00:12:48,397 are going to work. And we really said 231 00:12:48,480 --> 00:12:52,557 let's just push on this one idea that scaling deep learning seems to matter. 232 00:12:52,640 --> 00:12:56,077 We didn't know at the time how beautifully predictable the scaling laws 233 00:12:56,160 --> 00:13:00,077 were, but it was at least clear back in those days that if you threw more 234 00:13:00,160 --> 00:13:03,037 compute at something, you got better results most of the time. 235 00:13:03,120 --> 00:13:05,917 Um, and we kind of just decided we were 236 00:13:06,000 --> 00:13:10,237 going to push on it as far as we could and see what happened. Uh, we were also 237 00:13:10,320 --> 00:13:13,357 very worried about what it would mean for like how we were going to solve the 238 00:13:13,440 --> 00:13:17,517 alignment problem, what it would mean for society. We didn't know that we were 239 00:13:17,600 --> 00:13:19,757 going to make a product. We just thought we were going to be a research lab. So, 240 00:13:19,840 --> 00:13:22,477 we kind of had a weird start. Um, but it 241 00:13:22,560 --> 00:13:25,597 was really this 242 00:13:25,680 --> 00:13:27,757 it was this mindset of if AI works, it 243 00:13:27,840 --> 00:13:29,917 will be unbelievably important and we 244 00:13:30,000 --> 00:13:32,637 finally have an attack vector that maybe 245 00:13:32,720 --> 00:13:35,517 can go very far. It felt and still feels 246 00:13:35,600 --> 00:13:37,437 kind of crazy to me that it can go as 247 00:13:37,520 --> 00:13:41,197 far as it's turned out to go. Um, deep 248 00:13:41,280 --> 00:13:43,917 learning, 249 00:13:44,000 --> 00:13:48,557 the fact that these models can sort of do unsupervised learning in particular 250 00:13:48,640 --> 00:13:50,557 still seems sort of magical to me. The 251 00:13:50,640 --> 00:13:53,277 fact that next word prediction 252 00:13:53,360 --> 00:13:57,757 can be used basically to discover new 253 00:13:57,840 --> 00:14:01,757 scientific ideas. It makes like intellectual sense, but to 254 00:14:01,840 --> 00:14:06,077 see it work is still like a very crazy thing. Um, 255 00:14:06,160 --> 00:14:08,477 but that was it. It was we have this one idea, let's push on it. 256 00:14:08,560 --> 00:14:11,837 >> So, you see all these ideas now, but as you mentioned, it was incredibly 257 00:14:11,920 --> 00:14:15,037 unpopular at the time. I'm really curious to know what the first pitch for 258 00:14:15,120 --> 00:14:16,637 OpenAI looked like. 259 00:14:16,720 --> 00:14:19,997 >> Um, I love the unpopular ideas. I I I 260 00:14:20,080 --> 00:14:23,117 have always liked ideas that are big if 261 00:14:23,200 --> 00:14:26,397 they work, difficult to get to work, and 262 00:14:26,480 --> 00:14:29,677 unpopular. Um, my career as an investor 263 00:14:29,760 --> 00:14:34,397 was basically betting on those. Uh, for whatever like reason that people think 264 00:14:34,480 --> 00:14:37,037 one thing is fun and not another, this is something that has always seemed 265 00:14:37,120 --> 00:14:40,957 particularly fun to me. So, even if it weren't that I loved AI so much, I would 266 00:14:41,040 --> 00:14:43,037 have thought the fact that this seems 267 00:14:43,120 --> 00:14:47,117 really undervalued and promising uh was 268 00:14:47,200 --> 00:14:51,917 was cool. And that was basically the pitch. The pitch was uh something 269 00:14:52,000 --> 00:14:56,557 important is happening with deep learning and scale. Uh we don't know 270 00:14:56,640 --> 00:14:59,037 what we're going to do with it. We don't know where it's going to go. We had a 271 00:14:59,120 --> 00:15:03,117 bunch of ideas that turned out to be wrong on the hops along the way, but 272 00:15:03,200 --> 00:15:07,117 we're going to push on this and something good is going to come out of 273 00:15:07,200 --> 00:15:09,917 it and other people are undervaluing 274 00:15:10,000 --> 00:15:14,717 this. So it was a it was like that was really the pitch. It was not a 275 00:15:14,800 --> 00:15:18,397 great pitch. [laughter] Well, it worked. Um, how did you meet 276 00:15:18,480 --> 00:15:20,157 your co-founders? 277 00:15:20,240 --> 00:15:25,277 >> Uh, I had known Greg forever. Um, 278 00:15:25,360 --> 00:15:28,477 he I was an early investor in Stripe. I think he was the first employee and I 279 00:15:28,560 --> 00:15:30,717 was like brought in to help close him. 280 00:15:30,800 --> 00:15:32,397 Um, and we had stayed in touch for the 281 00:15:32,480 --> 00:15:34,957 sort of decade or whatever in between. 282 00:15:35,040 --> 00:15:39,517 Um, Ilia is the sort of genius 283 00:15:39,600 --> 00:15:41,757 and godfather of the field and I stalked 284 00:15:41,840 --> 00:15:44,877 him at a conference. Uh and then later 285 00:15:44,960 --> 00:15:47,757 we started talking about um doing a 286 00:15:47,840 --> 00:15:51,917 startup. Uh Voytech I met through kind 287 00:15:52,000 --> 00:15:55,517 of the like weird AGI nerds network of 288 00:15:55,600 --> 00:15:59,037 the Bay Area at the time and John 289 00:15:59,120 --> 00:16:01,197 one of the others connected me to and 290 00:16:01,280 --> 00:16:06,850 Elon uh I guess yeah. [laughter] 291 00:16:08,080 --> 00:16:12,157 So, so how did founding feel on this time with OpenAI compared to your 292 00:16:12,240 --> 00:16:14,717 experience with Looped before? 293 00:16:14,800 --> 00:16:17,677 >> Uh, I mean this was really a weird one. 294 00:16:17,760 --> 00:16:22,957 Like I the degree to which OpenAI was not meant 295 00:16:23,040 --> 00:16:27,277 to be a company is like very hard to overstate. We had no idea that we were 296 00:16:27,360 --> 00:16:29,837 going to ever do a product. >> Yeah, >> we thought we were just going to do 297 00:16:29,920 --> 00:16:32,957 research papers. Um, 298 00:16:33,040 --> 00:16:37,677 as we got going, it became clear that a there was going to be a lot more 299 00:16:37,760 --> 00:16:40,957 economic value along the way than we thought. We kind of had this conception 300 00:16:41,040 --> 00:16:43,837 that you were going to build this super intelligence and you're going to have 301 00:16:43,920 --> 00:16:47,357 this like crazy thing happen all at once. And we learned that was not quite 302 00:16:47,440 --> 00:16:50,477 the right picture. We also realized that instead of needing the few billion 303 00:16:50,560 --> 00:16:52,797 dollars we thought we might need, you're going to need like hundreds of billions 304 00:16:52,880 --> 00:16:57,917 of dollars. >> And so it evolved, but at the start it 305 00:16:58,000 --> 00:17:00,797 was like we're going to be a research lab. actually one of my most vivid 306 00:17:00,880 --> 00:17:03,357 memories of OpenAI. 307 00:17:03,440 --> 00:17:06,317 Very first day been all this drama to 308 00:17:06,400 --> 00:17:11,517 get it started. Uh very first day we're together in Greg Brockman's apartment 309 00:17:11,600 --> 00:17:13,037 which was where we worked until we got 310 00:17:13,120 --> 00:17:14,797 our first office a few weeks or maybe a 311 00:17:14,880 --> 00:17:20,157 month later. And it's like, you know, 9:30, 10 in the 312 00:17:20,240 --> 00:17:23,117 morning. Everybody rolls in. We're sitting on Greg's couch. Maybe there's 313 00:17:23,200 --> 00:17:26,477 10 or 12 of us. Maybe not even that many. Maybe eight or 10. And we look 314 00:17:26,560 --> 00:17:30,157 around and say like, "Okay, what are we going to do?" 315 00:17:30,240 --> 00:17:32,557 >> Yeah. >> And someone said, "We should write some 316 00:17:32,640 --> 00:17:37,117 papers." And someone said, "Okay, we need a whiteboard to start talking." 317 00:17:37,200 --> 00:17:40,717 >> And like someone orders like a whiteboard on, you know, Amazon. 318 00:17:40,800 --> 00:17:42,477 >> Yeah. >> Instant or something. 319 00:17:42,560 --> 00:17:47,997 >> And and I had this real moment of like, what? This is not this is not good. Like 320 00:17:48,080 --> 00:17:50,717 I don't know what this is, but this is not a startup or any kind of thing 321 00:17:50,800 --> 00:17:54,077 that's going to be effective. Um, but 322 00:17:54,160 --> 00:17:56,877 you know, I like I am a big believer 323 00:17:56,960 --> 00:18:02,077 that in moments like that, you just like breathe, trust that if it's if you have 324 00:18:02,160 --> 00:18:04,557 good people around, it's like going to work out. And we just kind of went with 325 00:18:04,640 --> 00:18:08,077 it. Sometime in that first week, most of 326 00:18:08,160 --> 00:18:13,117 the ideas that turned out to be the core ideas for the um first four years of 327 00:18:13,200 --> 00:18:16,637 OpenAI were written down on that whiteboard. Even though at the time we 328 00:18:16,720 --> 00:18:19,837 didn't realize that and we're kind of just like, eh, this is probably not. We 329 00:18:19,920 --> 00:18:21,757 went in some other directions. 330 00:18:21,840 --> 00:18:25,517 And you know, we said we need to get some 331 00:18:25,600 --> 00:18:28,877 GPUs, we need to do this, we need to do that. Uh we need to like pick some 332 00:18:28,960 --> 00:18:32,477 projects to see if this is really working. 333 00:18:32,560 --> 00:18:34,557 >> And stuff just kind of kept going. 334 00:18:34,640 --> 00:18:38,637 >> Wow. >> Yeah. So you're uh you were in this 335 00:18:38,720 --> 00:18:41,437 place where you're like, "Okay, what do we do? Let's start doing some research 336 00:18:41,520 --> 00:18:45,037 papers." like when was the moment after that where things started to click for 337 00:18:45,120 --> 00:18:49,197 the first time and you're like uh things started to feel like uh you guys really 338 00:18:49,280 --> 00:18:55,800 had discovered something or something was like big was about to happen. Um 339 00:19:01,120 --> 00:19:02,637 there were a few moments I think it was 340 00:19:02,720 --> 00:19:07,597 not just one but the the felt sense was 341 00:19:07,680 --> 00:19:10,637 it was like miracle after miracle and so 342 00:19:10,720 --> 00:19:14,717 at some point more than any one thing that worked. It was like, man, something 343 00:19:14,800 --> 00:19:19,357 about the underlying approach here feels like we've discovered like a new law of 344 00:19:19,440 --> 00:19:24,077 physics or something. And you know, it was like great to get Dota working or 345 00:19:24,160 --> 00:19:27,997 the robot hand working and then eventually the the first language model. 346 00:19:28,080 --> 00:19:32,157 But the the biggest takeaway was just like there is something about this 347 00:19:32,240 --> 00:19:34,077 underlying approach that really feels 348 00:19:34,160 --> 00:19:37,917 like we're on a trajectory to the degree 349 00:19:38,000 --> 00:19:40,797 that I had like a single moment. It was 350 00:19:40,880 --> 00:19:44,797 the GPT series. Um like by GPT2 it was 351 00:19:44,880 --> 00:19:49,517 clear to me like man this is gonna this is something important is about to 352 00:19:49,600 --> 00:19:52,877 happen. >> Do you remember that day the GP G2 launch? 353 00:19:52,960 --> 00:19:55,597 >> I do. Uh I don't remember the launch day 354 00:19:55,680 --> 00:19:57,037 but I remember the day that like I got 355 00:19:57,120 --> 00:19:59,997 to first talk to the model and I was 356 00:20:00,080 --> 00:20:03,197 just like wow this is 357 00:20:03,280 --> 00:20:07,517 this is great. And actually I think it was technically a model that was like a 358 00:20:07,600 --> 00:20:11,597 little bit beyond what we released as GPT2. Um, 359 00:20:11,680 --> 00:20:16,157 but like I knew GPT3 would be amazing and GPT4 would be amazing and they were 360 00:20:16,240 --> 00:20:19,757 and they were amazing in new ways. But that first night that I got to play with 361 00:20:19,840 --> 00:20:22,957 the model and it was like doing 362 00:20:23,040 --> 00:20:26,797 something I never had seen a computer do before. I was like, "Okay, this is and 363 00:20:26,880 --> 00:20:31,517 and we had the builtup like confidence of deep learning working." 364 00:20:31,600 --> 00:20:36,237 >> Yeah. And you initially withheld that model GBT2, right? What at what point 365 00:20:36,320 --> 00:20:40,557 did you feel like we were ready or or like what made you decide that the world 366 00:20:40,640 --> 00:20:43,997 was ready for these models? >> Uh 367 00:20:44,080 --> 00:20:48,157 look, clearly that looked like silly in retrospect. I think the principle of 368 00:20:48,240 --> 00:20:50,477 precaution with each new level of AI 369 00:20:50,560 --> 00:20:53,277 capability is reasonable. Alignment has 370 00:20:53,360 --> 00:20:55,757 been easier to make progress on than 371 00:20:55,840 --> 00:21:00,157 many of us in the field thought. Uh 372 00:21:00,240 --> 00:21:04,877 but the societal impacts have been 373 00:21:04,960 --> 00:21:08,957 or are getting big. Um, and 374 00:21:09,040 --> 00:21:13,837 you know, if we bias like slightly too much towards the 375 00:21:13,920 --> 00:21:15,917 cautionary side, I think that's good. 376 00:21:16,000 --> 00:21:20,477 Uh, too much would be bad. I I'm also worried about like a mega AI capability 377 00:21:20,560 --> 00:21:22,157 overhang building up. Like the way I 378 00:21:22,240 --> 00:21:24,717 would explain this is if companies don't 379 00:21:24,800 --> 00:21:27,677 adopt AI fast enough then they will lose 380 00:21:27,760 --> 00:21:32,797 to fully autonomous completely AI run and started companies and that will be a 381 00:21:32,880 --> 00:21:34,477 crazy destabilizing thing for society. 382 00:21:34,560 --> 00:21:36,877 So you can't be that cautious either. Um 383 00:21:36,960 --> 00:21:42,840 but you know like we'll go a little bit on that side of the line. 384 00:21:43,600 --> 00:21:45,597 >> So what do you think the next chat GPT 385 00:21:45,680 --> 00:21:47,757 moment will be? 386 00:21:47,840 --> 00:21:50,877 Uh well I think the next JBT moment 387 00:21:50,960 --> 00:21:53,757 already was coding agents. Um and I 388 00:21:53,840 --> 00:21:55,517 think the next one after this that has 389 00:21:55,600 --> 00:22:00,077 not come yet will be that same level of 390 00:22:00,160 --> 00:22:03,357 ability to get stuff done for all knowledge work which I don't think is 391 00:22:03,440 --> 00:22:05,037 that far away. 392 00:22:05,120 --> 00:22:10,637 >> Wow. So I mean we're all sophomores here. Um there's a lot that's going to 393 00:22:10,720 --> 00:22:13,837 change in the next few years. I mean before we graduate I think. Do I mean do 394 00:22:13,920 --> 00:22:16,717 you think AGI will be >> Yeah, I think if you are a sophomore 395 00:22:16,800 --> 00:22:19,197 now, you will graduate to a world with 396 00:22:19,280 --> 00:22:24,237 AGI in it. Uh and it will be much, you 397 00:22:24,320 --> 00:22:28,877 know, in some ways it'll like be the same. You'll go move and get a job and 398 00:22:28,960 --> 00:22:31,917 think about having a family and whatever else like the human, you never bet 399 00:22:32,000 --> 00:22:36,477 against like sort of the human drives and what we care about staying the same. 400 00:22:36,560 --> 00:22:38,477 But in some other sense, it's just going 401 00:22:38,560 --> 00:22:40,797 to be wildly different. uh you know I 402 00:22:40,880 --> 00:22:44,637 think we'll have AI uh improving at an 403 00:22:44,720 --> 00:22:47,197 incredibly rapid clip. Um society will 404 00:22:47,280 --> 00:22:50,077 be coming along uh at a slower clip but 405 00:22:50,160 --> 00:22:55,837 still very fast. Science will be getting automated. Uh what it means to start a 406 00:22:55,920 --> 00:22:58,397 startup will be totally different. What it means to go work at a company will be 407 00:22:58,480 --> 00:23:02,200 totally different. Uh 408 00:23:02,400 --> 00:23:06,077 there's going to be a lot of change in the next couple of years. Uh you are 409 00:23:06,160 --> 00:23:09,597 lucky to be a student right now because you can go do kind of whatever. uh and 410 00:23:09,680 --> 00:23:12,237 you can try a lot of things quickly. Um 411 00:23:12,320 --> 00:23:14,557 but and you're also lucky in that this 412 00:23:14,640 --> 00:23:17,837 is probably like 413 00:23:17,920 --> 00:23:20,637 the most interesting time to be starting 414 00:23:20,720 --> 00:23:26,077 off like adult life or a career. Um 415 00:23:26,160 --> 00:23:28,637 but I think a lot of the traditional 416 00:23:28,720 --> 00:23:30,477 advice is not quite going to work. 417 00:23:30,560 --> 00:23:34,717 >> Yeah. So speaking of execution, um in 418 00:23:34,800 --> 00:23:39,197 your first ever blog post on blog.samalttman.com, sam alman.com. You 419 00:23:39,280 --> 00:23:43,597 say that knowing is easy and it's the doing that's difficult. Um, is there a 420 00:23:43,680 --> 00:23:46,717 truth that you think you know now about the future where the missing piece is 421 00:23:46,800 --> 00:23:52,200 execution? >> Um, 422 00:23:55,360 --> 00:23:57,197 well, I think that's just generally 423 00:23:57,280 --> 00:24:01,837 going to continue to be true. Uh, I 424 00:24:01,920 --> 00:24:05,357 mean, knowing knowing is easier and easier like you can you get the help of 425 00:24:05,440 --> 00:24:10,077 these AI tools. Um, getting stuff done is easier too, but it it's easier for 426 00:24:10,160 --> 00:24:12,717 everybody. And so you have to like compete with everybody else trying to 427 00:24:12,800 --> 00:24:16,717 get stuff done in the world. Um, the 428 00:24:16,800 --> 00:24:18,877 it is for sure true that what one 429 00:24:18,960 --> 00:24:23,117 programmer can do now uh is like 430 00:24:23,200 --> 00:24:26,637 mindbendingly different than what you could do a year 431 00:24:26,720 --> 00:24:28,797 or two years ago, a year ago, let's say. 432 00:24:28,880 --> 00:24:31,917 Um, it is also true that the people that 433 00:24:32,000 --> 00:24:34,637 I know that are using these tools the 434 00:24:34,720 --> 00:24:39,517 most effectively say uh it has never been harder to do their work. And I 435 00:24:39,600 --> 00:24:43,677 suspect this trend is going to keep going. But like man, watching these 436 00:24:43,760 --> 00:24:48,157 people who are really good at using AI coding tools have like 12 agents running 437 00:24:48,240 --> 00:24:50,317 all the time and like these complicated 438 00:24:50,400 --> 00:24:51,837 custom harnesses that they're constantly 439 00:24:51,920 --> 00:24:55,597 evolving like the the yeah, the tools 440 00:24:55,680 --> 00:25:01,437 have gotten unbelievably good. Um, and the ability to use them to be at the top 441 00:25:01,520 --> 00:25:03,197 of the game I think has never been harder. 442 00:25:03,280 --> 00:25:06,157 >> And you think this will apply across like all verticals, not just coding? 443 00:25:06,240 --> 00:25:09,677 >> Uh, I mean, look, eventually you will just 444 00:25:09,760 --> 00:25:13,437 say like, "Hey, Mr. AI, please like make me my whole startup and do everything 445 00:25:13,520 --> 00:25:18,397 and do a great job. You know, I'll help you with whatever you need." But 446 00:25:18,480 --> 00:25:22,157 between now and then, yeah, I think it'll apply to all all verticals. 447 00:25:22,240 --> 00:25:26,477 Given that uh these models are becoming more and more powerful then and you you 448 00:25:26,560 --> 00:25:29,277 we have people that are running this many coding agents. There's definitely 449 00:25:29,360 --> 00:25:33,677 some of y'all in the audience here. Um what should people be working on and 450 00:25:33,760 --> 00:25:36,637 focusing on then? Like what do you think is the most important problem in the 451 00:25:36,720 --> 00:25:41,720 world right now? >> Uh 452 00:25:45,200 --> 00:25:49,437 the the degree to which it feels like 453 00:25:49,520 --> 00:25:53,117 everything is up for a rethink. I'm not like that interested. People are like, 454 00:25:53,200 --> 00:25:55,117 "Oh, all SAS is dead. I'm gonna like, 455 00:25:55,200 --> 00:25:58,237 you know, make a new startup to make 456 00:25:58,320 --> 00:26:03,117 SAS company X better." Like, you know what? That will probably work. But I 457 00:26:03,200 --> 00:26:04,477 think that's missing the point of what's 458 00:26:04,560 --> 00:26:08,717 possible. Now, the degree to which 459 00:26:08,800 --> 00:26:12,477 things are just completely fundamentally 460 00:26:12,560 --> 00:26:16,557 reimaginable, uh, 461 00:26:16,640 --> 00:26:18,717 and the space feels super open on that. 462 00:26:18,800 --> 00:26:21,277 like you know if you like just to pick a 463 00:26:21,360 --> 00:26:26,557 random example uh social products have not changed that 464 00:26:26,640 --> 00:26:30,237 much in the last and I'm not even counting like video scrolling apps just 465 00:26:30,320 --> 00:26:33,117 like an actual connect to people and 466 00:26:33,200 --> 00:26:35,277 like look at information about them 467 00:26:35,360 --> 00:26:38,397 style apps have not changed that much. I 468 00:26:38,480 --> 00:26:41,597 think uh moldbook points to the first 469 00:26:41,680 --> 00:26:45,437 interesting new thing where you can imagine a bunch of agents on behalf of 470 00:26:45,520 --> 00:26:48,877 their users off in a virtual space talking to each other like exchanging 471 00:26:48,960 --> 00:26:52,877 new information completely changing what it means to 472 00:26:52,960 --> 00:26:56,557 interact with other people on the internet like that was just not possible 473 00:26:56,640 --> 00:27:00,317 at all until very recently and now it feels like the space of social for the 474 00:27:00,400 --> 00:27:04,477 first time is totally open for some startup to do something completely new. 475 00:27:04,560 --> 00:27:07,677 Um, you know, a lot of people are like 476 00:27:07,760 --> 00:27:12,637 bolting AI into some productivity software and you know, like I don't 477 00:27:12,720 --> 00:27:16,477 really need something in Google Docs trying to like summarize it or write a 478 00:27:16,560 --> 00:27:19,277 sentence for me. I mean, that's like fine, but probably the whole thing of 479 00:27:19,360 --> 00:27:23,117 what it means to write a document is 480 00:27:23,200 --> 00:27:26,797 going to completely change. And there's a whole new kind of thing to imagine 481 00:27:26,880 --> 00:27:30,557 there. And this will apply not just traditional 482 00:27:30,640 --> 00:27:33,597 software startups, but most companies in 483 00:27:33,680 --> 00:27:39,757 the world pretty soon. You know, robots won't be that far off, whatever. Uh, 484 00:27:39,840 --> 00:27:41,837 I love times like this. They're rare. 485 00:27:41,920 --> 00:27:45,277 The iPhone when iPhone launched was one of them. I don't think we've really had 486 00:27:45,360 --> 00:27:48,637 one since. And the fact that this time, 487 00:27:48,720 --> 00:27:52,637 not only can you use the technology to do these things that you could never do 488 00:27:52,720 --> 00:27:56,477 before, but the fact that you can use technology to like make the product or 489 00:27:56,560 --> 00:28:01,037 make the company so quickly and with so few people, 490 00:28:01,120 --> 00:28:02,477 I think it's going to mean a lot more 491 00:28:02,560 --> 00:28:08,557 startups getting super successful really fast. And 492 00:28:08,640 --> 00:28:10,797 a lot of startups that have like one 493 00:28:10,880 --> 00:28:12,477 person or six people or something like 494 00:28:12,560 --> 00:28:15,837 that that get as impactful and as much 495 00:28:15,920 --> 00:28:18,157 revenue as some of the you know a 496 00:28:18,240 --> 00:28:21,837 medium-sized company does today. 497 00:28:21,920 --> 00:28:24,637 Do you think then that uh given you were 498 00:28:24,720 --> 00:28:29,037 talking about uh like Mbook these agents that like once AGI happens it will then 499 00:28:29,120 --> 00:28:32,237 surprise us and just all of a sudden like all of everything we've done all 500 00:28:32,320 --> 00:28:35,197 these new companies are popping up and creating themselves and integrating into 501 00:28:35,280 --> 00:28:39,517 our world and competing with things in a way like that or like how will it be 502 00:28:39,600 --> 00:28:44,600 introduced into our world then >> uh 503 00:28:45,440 --> 00:28:46,557 so I think I think that what it's going 504 00:28:46,640 --> 00:28:50,157 to feel like in the next few years let's 505 00:28:50,240 --> 00:28:55,037 say two years is stuff's just going to happen faster. You know, startups are 506 00:28:55,120 --> 00:28:57,597 going to do new things faster. Existing companies are going to ramp their 507 00:28:57,680 --> 00:29:00,317 revenue faster. It will feel like we've 508 00:29:00,400 --> 00:29:04,957 broken the laws of traditional speed and product development and 509 00:29:05,040 --> 00:29:08,077 business growth rates and and things like that. And, you know, companies will 510 00:29:08,160 --> 00:29:11,677 start to hire AI co-workers um and do 511 00:29:11,760 --> 00:29:13,997 amazing things with fewer people. Uh, I 512 00:29:14,080 --> 00:29:15,437 I think we'll continue to see AI 513 00:29:15,520 --> 00:29:19,757 research go much faster. Um, and we'll 514 00:29:19,840 --> 00:29:22,477 see scientific research happen much faster, but I don't think it'll feel 515 00:29:22,560 --> 00:29:25,837 like totally autonomous companies 516 00:29:25,920 --> 00:29:27,277 transacting with each other in an 517 00:29:27,360 --> 00:29:31,837 environment sort of moldbook style. Uh, 518 00:29:31,920 --> 00:29:33,997 eventually that probably does happen. 519 00:29:34,080 --> 00:29:37,837 Uh, and you know, one of one of the sort of 520 00:29:37,920 --> 00:29:42,237 like central observations I've had in life is that 521 00:29:42,320 --> 00:29:44,637 people are unbelievably adaptable. 522 00:29:44,720 --> 00:29:48,557 Societyy's unbelievably adaptable. So that sounds like a totally crazy 523 00:29:48,640 --> 00:29:52,397 destabilizing thing. And I bet it's not going to be as weird 524 00:29:52,480 --> 00:29:55,597 to live through as it sounds or only for a day. 525 00:29:55,680 --> 00:29:57,997 >> So I kind of wanted to go back to that thing that you'd mentioned about 526 00:29:58,080 --> 00:30:01,197 societal impacts. I mean, how do you think this is going to impact like 527 00:30:01,280 --> 00:30:04,877 politics like religion? like just expanding more just outside of like the 528 00:30:04,960 --> 00:30:06,557 startup space. How's society going to be 529 00:30:06,640 --> 00:30:09,517 impacted by this? Do you feel 530 00:30:09,600 --> 00:30:13,597 >> one of the um everyone's got a different definition 531 00:30:13,680 --> 00:30:17,197 for AGI, different definition for super intelligence. Uh you know, I kind of 532 00:30:17,280 --> 00:30:20,877 think like the most important definitions are something about the 533 00:30:20,960 --> 00:30:23,197 ability to self-improve. But one of the 534 00:30:23,280 --> 00:30:25,197 ones I like as a sort of example for a 535 00:30:25,280 --> 00:30:28,797 way to think about this is if 536 00:30:28,880 --> 00:30:32,637 if we had an AI system that could do a 537 00:30:32,720 --> 00:30:35,197 better job at making decisions 538 00:30:35,280 --> 00:30:37,117 than anyone we could possibly have be 539 00:30:37,200 --> 00:30:40,877 say president of the United States. Uh 540 00:30:40,960 --> 00:30:44,157 would we and like confidently much better. 541 00:30:44,240 --> 00:30:48,077 >> Um would we politically or society want that? 542 00:30:48,160 --> 00:30:51,517 >> It's very unclear. uh you know most people I think today at least would say 543 00:30:51,600 --> 00:30:54,157 probably not. Um maybe as things evolve 544 00:30:54,240 --> 00:30:56,637 more they do or you know like an AI CEO 545 00:30:56,720 --> 00:30:58,557 of a big company would would you rather 546 00:30:58,640 --> 00:31:02,077 have uh I think it'd be awesome for open 547 00:31:02,160 --> 00:31:04,557 AI to have an AI CEO but I think a lot 548 00:31:04,640 --> 00:31:10,077 of people be like ah that's a little that's a little scary. Um, 549 00:31:10,160 --> 00:31:13,517 if though AI is better at doing some of 550 00:31:13,600 --> 00:31:17,917 these things, I think there will be like many things that people are better at 551 00:31:18,000 --> 00:31:19,677 than AI long term, or at least people 552 00:31:19,760 --> 00:31:21,917 really care about other people doing, 553 00:31:22,000 --> 00:31:25,277 but I have no delusions that an AI will 554 00:31:25,360 --> 00:31:27,517 not be a better CEO of OpenAI than me at 555 00:31:27,600 --> 00:31:31,757 some reasonably soon time. Um, if 556 00:31:31,840 --> 00:31:34,477 if that's true and if some companies or 557 00:31:34,560 --> 00:31:39,197 countries don't do that and others do and AI is much better like that will 558 00:31:39,280 --> 00:31:41,997 have a real impact in economic competitiveness, national 559 00:31:42,080 --> 00:31:45,917 competitiveness, things like that. And and so I think one of the things that 560 00:31:46,000 --> 00:31:48,077 will happen about how this is going to 561 00:31:48,160 --> 00:31:54,120 impact us economically, society, politically is 562 00:31:54,560 --> 00:31:56,557 if the AI is just much better, I think 563 00:31:56,640 --> 00:31:58,477 people will feel increasingly like we 564 00:31:58,560 --> 00:32:01,917 either do worse or we put the AI in more 565 00:32:02,000 --> 00:32:06,157 and more influential positions. I think people will always like, you know, if 566 00:32:06,240 --> 00:32:09,197 you think of like an AI CEO of a company, I think people will always want 567 00:32:09,280 --> 00:32:10,557 human board members and that's probably 568 00:32:10,640 --> 00:32:14,717 a really great idea. Um, but 569 00:32:14,800 --> 00:32:17,837 I kind of think that 570 00:32:17,920 --> 00:32:21,197 people may have not quite 571 00:32:21,280 --> 00:32:23,997 I certainly have not really wrestled 572 00:32:24,080 --> 00:32:28,237 with like is there a way that this 573 00:32:28,320 --> 00:32:29,997 doesn't happen and what all of the 574 00:32:30,080 --> 00:32:32,557 political, economic, societal impacts 575 00:32:32,640 --> 00:32:36,477 are. Uh, I I still believe that although 576 00:32:36,560 --> 00:32:40,557 it sounds like a completely crazy thing to say, 577 00:32:40,640 --> 00:32:43,837 we're very adaptable. We really care about other people. You know, maybe 578 00:32:43,920 --> 00:32:47,917 there's like a AI CEO, but you just care about like the 579 00:32:48,000 --> 00:32:51,197 company's spokesperson speaking to you in these 580 00:32:51,280 --> 00:32:53,597 like 8second video clips or whatever. 581 00:32:53,680 --> 00:32:56,680 >> Um, 582 00:32:59,360 --> 00:33:02,637 >> I think I'm not a jobs doomer at all. I mean a lot of the current jobs are going 583 00:33:02,720 --> 00:33:05,917 to go away or be heavily changed but 584 00:33:06,000 --> 00:33:08,717 human desire for 585 00:33:08,800 --> 00:33:13,117 being useful to each other to compete with each other uh to like create 586 00:33:13,200 --> 00:33:15,517 differential status to do something 587 00:33:15,600 --> 00:33:18,637 fulfilling to express creativity that is 588 00:33:18,720 --> 00:33:21,917 not going to go anywhere. And maybe 589 00:33:22,000 --> 00:33:25,677 maybe the jobs of 100 years from now look nothing like the jobs of today. 590 00:33:25,760 --> 00:33:28,397 >> Yeah. nor do that many of the jobs of today 591 00:33:28,480 --> 00:33:32,397 look much like the jobs of a few hundred years ago. Um, 592 00:33:32,480 --> 00:33:36,477 but I'm confident we'll have stuff to do. I'm confident we'll really care 593 00:33:36,560 --> 00:33:38,317 about other people and less about the 594 00:33:38,400 --> 00:33:42,957 machines. And then beyond that, uh, 595 00:33:43,040 --> 00:33:46,957 yeah, I think we're just like AI is going to get ever more competent and 596 00:33:47,040 --> 00:33:50,397 we're going to have we're gonna want to use it for more and more things. 597 00:33:50,480 --> 00:33:54,237 >> Awesome. All right, so we're gonna go into a quick Q&A segment now uh, with 598 00:33:54,320 --> 00:33:56,957 questions from the audience. I would love like hard questions that I've never 599 00:33:57,040 --> 00:34:00,557 heard before. We have a Slack channel, Sam Alman Q&A. 600 00:34:00,640 --> 00:34:05,597 >> Sam Alman Q&A, guys, and a lot of you have put in questions. Um, 601 00:34:05,680 --> 00:34:08,317 >> so we're going to we're going to start off by reading some of the questions 602 00:34:08,400 --> 00:34:12,157 that were already submitted, I think, since we posted it yesterday. But we 603 00:34:12,240 --> 00:34:15,757 encourage you guys to submit more questions and then if you really really 604 00:34:15,840 --> 00:34:17,997 like a certain question, we want you to 605 00:34:18,080 --> 00:34:22,877 just uh reply with some sort of emoji so that we can see it and we can ask Sam. 606 00:34:22,960 --> 00:34:24,237 Awesome. All right. You want to start off 607 00:34:24,320 --> 00:34:25,837 >> to get started? Uh, you're talking about 608 00:34:25,920 --> 00:34:29,037 like OpenAI potentially having a AI CEO. 609 00:34:29,120 --> 00:34:32,797 So, if you ask a question that's kind of related to this about like say OpenAI 610 00:34:32,880 --> 00:34:34,877 becomes this like end- all beall company 611 00:34:34,960 --> 00:34:37,597 that can do anything because it's uh CEO 612 00:34:37,680 --> 00:34:40,557 is an AI. How do you compete with OpenAI 613 00:34:40,640 --> 00:34:45,517 then in in the future? Um, 614 00:34:45,600 --> 00:34:48,237 when we started OpenAI, everyone was like, you can't do this. You can't 615 00:34:48,320 --> 00:34:50,557 compete with Google. It's impossible. 616 00:34:50,640 --> 00:34:54,477 And there will someday be much bigger and more successful companies than 617 00:34:54,560 --> 00:34:58,557 OpenAI. They will not do exactly what OpenAI does. Uh you may steal our users. 618 00:34:58,640 --> 00:35:01,277 You may do some other thing. You may like have a replacement product. It will 619 00:35:01,360 --> 00:35:07,400 not very unlikely to be exactly what Open AAI does. Uh 620 00:35:08,320 --> 00:35:12,397 Google screwed up, didn't react enough to our approach early enough. They're, 621 00:35:12,480 --> 00:35:15,757 you know, doing a very nice nice job now, but had Google had their act 622 00:35:15,840 --> 00:35:19,597 together, we would not have been able to get off the ground at all. Um and they 623 00:35:19,680 --> 00:35:24,557 had very smart people. It's just big companies have big company problems. Uh 624 00:35:24,640 --> 00:35:26,237 someone in this room I hope will do the 625 00:35:26,320 --> 00:35:29,597 same thing to OpenAI. 626 00:35:29,680 --> 00:35:34,397 We are in we sell intelligence for what 627 00:35:34,480 --> 00:35:37,677 I think of as a very low price. Um you 628 00:35:37,760 --> 00:35:39,277 have 629 00:35:39,360 --> 00:35:42,637 probably the biggest option space of any 630 00:35:42,720 --> 00:35:45,837 startup of any of startups that have at 631 00:35:45,920 --> 00:35:48,877 any time in history. If if you try to 632 00:35:48,960 --> 00:35:52,317 make like an exact copy of chatbt or 633 00:35:52,400 --> 00:35:55,357 codecs or you know frontier models in an 634 00:35:55,440 --> 00:35:59,277 API probably not going to work. There's 635 00:35:59,360 --> 00:36:03,437 so much other stuff to do. Uh and the 636 00:36:03,520 --> 00:36:08,877 best thing of all is to figure out you know we talked about this earlier 637 00:36:08,960 --> 00:36:12,157 like things that are just going to totally change now. Take your pick. 638 00:36:12,240 --> 00:36:16,317 We're not going to be able to go do all those things. Uh, and neither is any 639 00:36:16,400 --> 00:36:20,077 other company that is not a startup and that has the sort of like freedom to be 640 00:36:20,160 --> 00:36:23,517 a little bit weird and decisive and pull together a band of rebels. Just go do 641 00:36:23,600 --> 00:36:28,317 that. It always sounds really scary. I remember how many like very wise people 642 00:36:28,400 --> 00:36:30,477 said you can't start an AI company. 643 00:36:30,560 --> 00:36:32,397 Google's just going to crush you. Uh, 644 00:36:32,480 --> 00:36:34,237 maybe they still do. I don't know. Um, 645 00:36:34,320 --> 00:36:39,117 but just just like go at it. 646 00:36:39,200 --> 00:36:42,877 >> Um, okay. Well, I can ask the next question. Um, this might be a burning 647 00:36:42,960 --> 00:36:47,357 question. Let me see. All right. News reports say OpenAI is burning cash very 648 00:36:47,440 --> 00:36:51,117 fast and could face liquid liquidity pressure in the future. How do you plan 649 00:36:51,200 --> 00:36:55,117 to handle this? And is the situation really as serious as you think the media 650 00:36:55,200 --> 00:36:56,317 reports it to be? 651 00:36:56,400 --> 00:36:58,557 >> Uh we are burning cash very fast. It is 652 00:36:58,640 --> 00:37:02,237 not a serious situation like the you 653 00:37:02,320 --> 00:37:06,637 know if we can spend a billion dollars this year to make $3 654 00:37:06,720 --> 00:37:10,317 billion next year there's a lot of capital in the world that wants to do 655 00:37:10,400 --> 00:37:12,397 that. So we will raise a lot of money 656 00:37:12,480 --> 00:37:18,920 equity and debt. we will spend a lot of money. We'll make a lot of money. Uh the 657 00:37:19,040 --> 00:37:21,677 it's weird. The economy hasn't seen 658 00:37:21,760 --> 00:37:26,237 something quite like this before, but given what we see about revenue growth 659 00:37:26,320 --> 00:37:29,357 and model growth and how much model growth correlates with future revenue 660 00:37:29,440 --> 00:37:32,877 growth, we're like, 661 00:37:32,960 --> 00:37:36,797 yeah, super confident in where things are going to go. 662 00:37:36,880 --> 00:37:37,837 >> Yeah. 663 00:37:37,920 --> 00:37:39,437 >> So, we've got another question here. 664 00:37:39,520 --> 00:37:44,480 It's um Sam, are you really happy 665 00:37:45,010 --> 00:37:48,077 >> [laughter] >> I I am actually one of the happiest 666 00:37:48,160 --> 00:37:50,477 people I know. Uh but I didn't used to 667 00:37:50,560 --> 00:37:52,957 be. I used to be a very anxious person. 668 00:37:53,040 --> 00:37:57,117 Um the [snorts] 669 00:37:57,200 --> 00:38:01,437 a thing people say to me all the time is like you have a miserable job. You have 670 00:38:01,520 --> 00:38:03,917 all the stress. You have all these people coming after you. You have like 671 00:38:04,000 --> 00:38:06,077 crazy bad You know, so many things 672 00:38:06,160 --> 00:38:07,997 happen to you a day. Like how do you not 673 00:38:08,080 --> 00:38:12,397 have a mental breakdown? Um, and 674 00:38:12,480 --> 00:38:15,997 my answer to that is like I just figured out what it was going to take for me 675 00:38:16,080 --> 00:38:19,197 like in my 20s when I realized I was not particularly happy and certainly not 676 00:38:19,280 --> 00:38:22,717 particularly calm. I figured out what it was going to take for me to get happy 677 00:38:22,800 --> 00:38:25,837 and calm. Um, I think people find their 678 00:38:25,920 --> 00:38:28,397 own paths there. But, 679 00:38:28,480 --> 00:38:30,877 uh, meditation was a big part of it for 680 00:38:30,960 --> 00:38:33,917 me. Uh, and it's kind of like a onetime 681 00:38:34,000 --> 00:38:38,317 thing. And now like even when stuff is 682 00:38:38,400 --> 00:38:42,717 really hard, like really painful, um, 683 00:38:42,800 --> 00:38:46,157 one of a thing I kind of 684 00:38:46,240 --> 00:38:48,317 learned in meditation that always stuck 685 00:38:48,400 --> 00:38:53,277 with me is that everybody thinks the opposite of a bad experience is a good 686 00:38:53,360 --> 00:38:56,877 experience. So if I'm having a bad experience, then I'm not happy because I 687 00:38:56,960 --> 00:38:58,397 wish I were having a good experience. 688 00:38:58,480 --> 00:38:59,997 And the kind of like a big mental shift 689 00:39:00,080 --> 00:39:02,637 for me was to reframe it as the opposite 690 00:39:02,720 --> 00:39:05,917 of a bad experience is no experience. 691 00:39:06,000 --> 00:39:09,757 And someday that'll happen to you and you'll miss the bad experiences too. And 692 00:39:09,840 --> 00:39:12,717 you know like I have a lot of days where it's very stressful and very hard and a 693 00:39:12,800 --> 00:39:16,557 lot of things go wrong. Uh I'm still grateful for them. 694 00:39:16,640 --> 00:39:19,357 >> Adding on to this um someone asks what 695 00:39:19,440 --> 00:39:22,637 does a day in your life look like then? 696 00:39:22,720 --> 00:39:28,317 >> Um absolute chaos. Uh I I like wake up 697 00:39:28,400 --> 00:39:30,717 early. I work for a few hours. I can I 698 00:39:30,800 --> 00:39:32,957 kind of have control of my life until 699 00:39:33,040 --> 00:39:37,277 like 8 o'clock in the morning. Uh 700 00:39:37,360 --> 00:39:40,557 and I can like get something done. I can try to drive whatever I'm want to get 701 00:39:40,640 --> 00:39:44,317 done. Then I play with my kid for an hour. Then I go into the office. And 702 00:39:44,400 --> 00:39:47,197 then it's just like insane. 703 00:39:47,280 --> 00:39:49,117 Whatever is going to happen happens like 704 00:39:49,200 --> 00:39:55,160 crazy. I don't think any company has been as 705 00:39:55,760 --> 00:39:58,157 fastm moving, chaotic in everybody's 706 00:39:58,240 --> 00:40:02,317 crosshairs as us. And so it's just yeah, 707 00:40:02,400 --> 00:40:05,837 unpredictable chaos all day. >> Yeah, a lot of us definitely don't wake 708 00:40:05,920 --> 00:40:08,077 up at uh before 8 in the morning. 709 00:40:08,160 --> 00:40:12,637 >> I did not used to wake up early. This this this job is the thing that made me 710 00:40:12,720 --> 00:40:16,077 start doing it. >> Awesome. >> Should we do two more questions and then 711 00:40:16,160 --> 00:40:18,317 >> let's do two more questions and then we'll wrap up. 712 00:40:18,400 --> 00:40:21,037 >> We are at time. 713 00:40:21,120 --> 00:40:23,037 I I have another question from someone. 714 00:40:23,120 --> 00:40:25,517 Um, what AI sub field today feels like 715 00:40:25,600 --> 00:40:30,077 open AI in 2016 a little undervalued or underestimated but could shoot out of 716 00:40:30,160 --> 00:40:33,597 nowhere? >> Uh, that's a good that's a great 717 00:40:33,680 --> 00:40:37,277 question there. This is the kind of question that I would like spend a lot 718 00:40:37,360 --> 00:40:40,237 of time trying to think about. Uh, I'm 719 00:40:40,320 --> 00:40:41,917 not sure if you mean it as like a 720 00:40:42,000 --> 00:40:46,397 research thing or a company thing. Uh, 721 00:40:46,480 --> 00:40:49,037 I'll answer both and I'll do it quickly 722 00:40:49,120 --> 00:40:50,397 because I know we're almost out of time. 723 00:40:50,480 --> 00:40:53,997 Um, on the research 724 00:40:54,080 --> 00:40:56,237 perspective, 725 00:40:56,320 --> 00:40:58,877 I bet there is another new architecture 726 00:40:58,960 --> 00:41:01,197 to find that is going to be like as big 727 00:41:01,280 --> 00:41:04,557 of a gain as transformers were over 728 00:41:04,640 --> 00:41:08,317 LSTMs. And I think you finally have models that 729 00:41:08,400 --> 00:41:11,597 are smart enough to help do that kind of research. 730 00:41:11,680 --> 00:41:14,637 So I would I would go look for like 731 00:41:14,720 --> 00:41:17,837 where can I find a mega 732 00:41:17,920 --> 00:41:23,117 breakthrough uh and I would use the models to help me on the product side. I 733 00:41:23,200 --> 00:41:26,557 guess we talked about this a little bit earlier, but I would I would just pick a 734 00:41:26,640 --> 00:41:29,117 big area and say what is possible now 735 00:41:29,200 --> 00:41:32,397 with AI that wasn't possible at all 736 00:41:32,480 --> 00:41:34,317 before. like where can I totally redo 737 00:41:34,400 --> 00:41:37,357 something that's like AI is the absolute 738 00:41:37,440 --> 00:41:42,637 core to the the interaction working and 739 00:41:42,720 --> 00:41:45,437 look I don't want to say this is like a once in a-lifetime opportunity because I 740 00:41:45,520 --> 00:41:48,477 hope the world keeps getting dramatically better every few years and 741 00:41:48,560 --> 00:41:51,677 that if we all came back here or actually when some of you come back here 742 00:41:51,760 --> 00:41:53,597 and sit up here and talk to the then 743 00:41:53,680 --> 00:41:58,157 current students um AGI will uh look 744 00:41:58,240 --> 00:42:01,277 like just a warm-up for whatever the next important thing was and that'll 745 00:42:01,360 --> 00:42:04,637 keep going for the rest of history. Um, 746 00:42:04,720 --> 00:42:07,677 but this is at least the best time ever 747 00:42:07,760 --> 00:42:11,960 so far. And what 748 00:42:13,200 --> 00:42:16,157 what you can do now is so crazily 749 00:42:16,240 --> 00:42:20,157 exciting that I think you should probably try to figure out how to take 750 00:42:20,240 --> 00:42:23,117 advantage of this moment. >> Yeah. 751 00:42:23,200 --> 00:42:26,317 >> And one last question. Um, a lot of labs 752 00:42:26,400 --> 00:42:30,397 like XAI and Google have their own data centers and even custom silicon for 753 00:42:30,480 --> 00:42:34,637 inference. Does OpenAI have any interest in developing hardware as well or stick 754 00:42:34,720 --> 00:42:36,877 to external partners? Why or why not? 755 00:42:36,960 --> 00:42:39,117 >> Uh we have a big custom silicon program. 756 00:42:39,200 --> 00:42:42,157 Um we 757 00:42:42,240 --> 00:42:44,717 we're very excited about our uh our 758 00:42:44,800 --> 00:42:47,677 inference chip and clearly the world 759 00:42:47,760 --> 00:42:52,077 wants a lot of cheap tokens and I think what we're working on there will be 760 00:42:52,160 --> 00:42:54,877 great. We have a very deep partnership with Nvidia. We do all of our training 761 00:42:54,960 --> 00:42:56,317 on Nvidia. We also use them for a lot of 762 00:42:56,400 --> 00:42:59,757 inference. Um but 763 00:42:59,840 --> 00:43:02,877 we are excited to be doing uh our our 764 00:43:02,960 --> 00:43:05,837 chip in our system. Um building data 765 00:43:05,920 --> 00:43:08,397 centers I would love to not have to do 766 00:43:08,480 --> 00:43:10,877 that. [laughter] Um if we have to at 767 00:43:10,960 --> 00:43:14,557 some point we will. Uh 768 00:43:14,640 --> 00:43:18,477 that's you know like you can't do everything well as a company and that's 769 00:43:18,560 --> 00:43:22,477 like a big intensive thing to do. Um if we can just design a great rack and have 770 00:43:22,560 --> 00:43:25,677 other people operate those and build the crazy buildings around them that'd be 771 00:43:25,760 --> 00:43:28,637 awesome. Thank you, Sam. Everyone, let's give it 772 00:43:28,720 --> 00:43:32,880 up for Sam. >> Thank you. [applause]