1 00:00:00,000 --> 00:00:02,757 I think it's extremely clear that we are 2 00:00:02,840 --> 00:00:07,037 going to have AGI within the next couple of years [music] in a way that 3 00:00:07,120 --> 00:00:11,877 it's still going to be jagged, but that the the floor of task will just be 4 00:00:11,960 --> 00:00:15,557 almost for any intellectual task of how [music] you use your computer. The AI 5 00:00:15,640 --> 00:00:19,517 will be able to do that. The scariest moment at OpenAI was actually after we 6 00:00:19,600 --> 00:00:23,497 launched ChatGPT. And I remember being at the holiday party and just feeling 7 00:00:23,580 --> 00:00:28,757 [music] this vibe of we won. I have never felt that. I was like, no. That we 8 00:00:28,840 --> 00:00:31,637 we are the underdog. And we always have been. From the moment we launched 9 00:00:31,720 --> 00:00:35,317 ChatGPT, I remember talking with my team having [music] this exact conversation. 10 00:00:35,400 --> 00:00:38,477 I said, "How much compute should we buy?" I said, "All of it." I said, "No, 11 00:00:38,560 --> 00:00:42,717 no, no, really. How much compute should we buy?" I said, "No matter how much we 12 00:00:42,800 --> 00:00:47,077 try to build, I know we're not going to be able to keep up with the demand." 13 00:00:47,160 --> 00:00:50,877 >> OpenAI co-founder and president Greg [music] Brockman joins us to talk about 14 00:00:50,960 --> 00:00:55,797 AI's most promising opportunities, how OpenAI plans to capitalize on them, and 15 00:00:55,880 --> 00:00:57,437 what the super [music] app is all about. 16 00:00:57,520 --> 00:01:01,077 And Greg is with us here in studio today. Greg, great to see you. 17 00:01:01,160 --> 00:01:02,877 >> Thank you for having me. 18 00:01:02,960 --> 00:01:04,917 >> Well, we're speaking at a time where 19 00:01:05,000 --> 00:01:08,677 OpenAI is shutting down video generation 20 00:01:08,760 --> 00:01:11,237 and focusing its energies on a super 21 00:01:11,320 --> 00:01:15,877 app, which is going to combine business and coding use cases. 22 00:01:15,960 --> 00:01:19,717 And I think from the outside, those of us watching this are like, including 23 00:01:19,800 --> 00:01:25,517 myself, OpenAI's winning in consumer, and now it's shifting its resources. What is 24 00:01:25,600 --> 00:01:28,397 happening? >> Well, the way I would think about this 25 00:01:28,480 --> 00:01:32,677 is that we have been in a world where we're developing this technology, deep 26 00:01:32,760 --> 00:01:35,317 learning, to really see can it 27 00:01:35,400 --> 00:01:38,317 have the positive impact that we have 28 00:01:38,400 --> 00:01:42,597 always pictured. Can it build can it be used to build 29 00:01:42,680 --> 00:01:45,117 applications that help people that help 30 00:01:45,200 --> 00:01:50,317 them in their lives. And we've separately had a arm that's 31 00:01:50,400 --> 00:01:53,597 saying, "Let's actually try to deploy this technology, whether that's to help 32 00:01:53,680 --> 00:01:57,797 sustain the business, to start getting some uh practice with getting real-world 33 00:01:57,880 --> 00:02:01,437 impact, those kinds of things, for the time when this technology actually comes 34 00:02:01,520 --> 00:02:05,917 to fruition, that it actually becomes the uh everything that we that we've 35 00:02:06,000 --> 00:02:09,357 imagined, that we started this company to try to to have. And I think that 36 00:02:09,440 --> 00:02:12,797 we're at a moment now where we've really seen this technology, it's going to 37 00:02:12,880 --> 00:02:15,197 work. And that we're moving out of 38 00:02:15,280 --> 00:02:17,797 testing on benchmarks and sort of these 39 00:02:17,880 --> 00:02:23,077 almost cerebral demonstrations of capability to it actually being the case 40 00:02:23,160 --> 00:02:26,637 that for us to develop it further, we need to see it in the real world and get 41 00:02:26,720 --> 00:02:30,637 feedback from how people are using it in knowledge work and various applications. 42 00:02:30,720 --> 00:02:33,117 And so, the way I think about it is that this is a 43 00:02:33,200 --> 00:02:37,477 bigger strategic shift because of the phase of the technology. 44 00:02:37,560 --> 00:02:41,477 And it's not so much that we're saying we're moving from consumer to B2B. It's 45 00:02:41,560 --> 00:02:45,317 really what we're saying is that what are the most important applications that 46 00:02:45,400 --> 00:02:48,437 we can focus on because we can't focus on everything, right? But what are the 47 00:02:48,520 --> 00:02:52,237 things that we can bring to life that will actually synergize together as we 48 00:02:52,320 --> 00:02:55,797 build them and that will deliver meaningful impact and help elevate 49 00:02:55,880 --> 00:03:00,117 everyone. And when we look at the list, so there's 50 00:03:00,200 --> 00:03:04,237 consumer, you can think of it as many things, but there's a personal 51 00:03:04,320 --> 00:03:06,917 assistant, right? Something that knows you, that's aligned with your goals, 52 00:03:07,000 --> 00:03:09,557 that's going to help you achieve whatever whatever it is that you want in 53 00:03:09,640 --> 00:03:13,597 your life. There's also creative expression and entertainment and many 54 00:03:13,680 --> 00:03:18,357 other applications. On the business side, maybe you can if 55 00:03:18,440 --> 00:03:21,797 you zoom out, it looks more like one thing of just you have a hard task, can 56 00:03:21,880 --> 00:03:25,077 AI go do it? Does it have all the context to do all these things? 57 00:03:25,160 --> 00:03:27,717 And for us, it's very clear that there 58 00:03:27,800 --> 00:03:32,397 the stack rank includes two things at the top. One is the personal assistant, 59 00:03:32,480 --> 00:03:36,197 the other is the AI that can go and solve hard problems for you. 60 00:03:36,280 --> 00:03:40,157 And when we look at the compute we have, we are not even going to have enough 61 00:03:40,240 --> 00:03:43,997 compute to fund those two things. And then once we start adding in many other 62 00:03:44,080 --> 00:03:48,077 applications, many other things that AI is going to be very useful for, is going 63 00:03:48,160 --> 00:03:50,997 to help people with, we just can't possibly get to all of them. And so, I 64 00:03:51,080 --> 00:03:55,517 think that this is a recognition of the maturation of the technology and the 65 00:03:55,600 --> 00:03:59,357 incredible impact it's going to have very quickly and our need to to 66 00:03:59,440 --> 00:04:04,517 prioritize and to actually pick the set of applications that we want to shine 67 00:04:04,600 --> 00:04:06,317 and to really bring to the world. 68 00:04:06,400 --> 00:04:09,437 >> And when I've heard you talk about OpenAI's various bets, one of the ways 69 00:04:09,520 --> 00:04:13,117 that you've described it is that OpenAI can be a version of Disney or a like 70 00:04:13,200 --> 00:04:17,157 Disney where you have this core compelling advantage at the center and 71 00:04:17,240 --> 00:04:18,717 then you farm it out in different ways. 72 00:04:18,800 --> 00:04:22,837 So, Disney has Mickey Mouse and then it can do the movies and the theme park and 73 00:04:22,920 --> 00:04:25,397 Disney Plus. And for OpenAI, it's the 74 00:04:25,480 --> 00:04:29,957 model and you can do video generation and be this assistant and then help with 75 00:04:30,040 --> 00:04:34,037 enterprise and work. So, is it no longer possible then to have 76 00:04:34,120 --> 00:04:38,477 that sort of central advantage and then be able to farm it out in all sorts of 77 00:04:38,560 --> 00:04:41,837 ways? Like, have you decided Have you come to this realization that basically 78 00:04:41,920 --> 00:04:44,037 like it's time to pick or choose? 79 00:04:44,120 --> 00:04:46,757 >> Well, I actually think that in some ways 80 00:04:46,840 --> 00:04:50,557 that story is even more true than it's been, but the thing that's important to 81 00:04:50,640 --> 00:04:54,957 realize is technologically that the Sora models, which are incredible models, by 82 00:04:55,040 --> 00:04:57,917 the way, are a different branch of the tech tree 83 00:04:58,000 --> 00:05:01,757 than the core reasoning GPT series. Mhm. 84 00:05:01,840 --> 00:05:05,317 They're just built in a very different way. And to some extent, we're really 85 00:05:05,400 --> 00:05:09,797 saying that pursuing both branches is very hard for us to do for these 86 00:05:09,880 --> 00:05:13,957 applications. Now, we we are actually continuing the Sora research program in 87 00:05:14,040 --> 00:05:18,357 the context of robotics, right? Which I think is very clearly going to be a 88 00:05:18,440 --> 00:05:21,317 transformative application, which is still a little bit in the research 89 00:05:21,400 --> 00:05:25,077 phase, right? That robotics is not really yet mature and deployed in the 90 00:05:25,160 --> 00:05:29,117 way that we're going to see this this real takeoff of this technology in 91 00:05:29,200 --> 00:05:32,477 knowledge work over the next year. And so, it's a recognition of for this 92 00:05:32,560 --> 00:05:34,597 moment, we really need to put the 93 00:05:34,680 --> 00:05:37,917 primary focus on developing the GPT 94 00:05:38,000 --> 00:05:41,077 series and that doesn't just mean text. 95 00:05:41,160 --> 00:05:42,717 It doesn't just mean cerebral things. 96 00:05:42,800 --> 00:05:46,397 Like, for example, I bidirectional communication, having a great 97 00:05:46,480 --> 00:05:49,517 speech-to-speech interface, that is something that also is going to make 98 00:05:49,600 --> 00:05:52,477 this technology very usable and very useful, 99 00:05:52,560 --> 00:05:55,837 but it's not a different branch of the tech tree. It's all kind of one model, 100 00:05:55,920 --> 00:05:58,077 and we just sort of tweak that in slightly different ways, kind of like 101 00:05:58,160 --> 00:06:00,597 you described. And so, I think there's something about if you're if you branch 102 00:06:00,680 --> 00:06:04,677 too far and you have two different artifacts, that is very hard to sustain 103 00:06:04,760 --> 00:06:07,997 in a world where there is limited compute. And the reason there's limited 104 00:06:08,080 --> 00:06:09,837 compute is cuz there's so much demand. 105 00:06:09,920 --> 00:06:13,157 There's so much people want to do with every single model that we create. 106 00:06:13,240 --> 00:06:14,637 >> Okay, so talk a little bit then about 107 00:06:14,720 --> 00:06:17,477 why your bet is not on this seems like 108 00:06:17,560 --> 00:06:21,677 world model version where the um you know, the video understands where things 109 00:06:21,760 --> 00:06:23,597 go. It's it's obviously useful for 110 00:06:23,680 --> 00:06:26,637 robotics. Why is your bet on the GPT 111 00:06:26,720 --> 00:06:31,557 reasoning model tree as opposed to this area which you've you had been seeing 112 00:06:31,640 --> 00:06:35,517 real progress with Sora. I mean, to see the progress of video generation, you 113 00:06:35,600 --> 00:06:38,437 know, generation 1, 2, 3 was enormous. 114 00:06:38,520 --> 00:06:40,797 So, why is your bet where it is? 115 00:06:40,880 --> 00:06:44,237 >> So, the problem in this field is too much opportunity. 116 00:06:44,320 --> 00:06:48,797 Right? It's the thing The thing that we observe very early on in OpenAI is that 117 00:06:48,880 --> 00:06:52,197 everything we could imagine works. Now, there's different levels of friction 118 00:06:52,280 --> 00:06:54,957 associated with it, different amounts of engineering effort, different compute 119 00:06:55,040 --> 00:06:59,477 requirements, all those things, but every single different idea, as long as 120 00:06:59,560 --> 00:07:02,317 it's kind of mathematically sound, you actually can start getting some pretty 121 00:07:02,400 --> 00:07:06,197 good results. And I think that shows you the power of the underlying technology 122 00:07:06,280 --> 00:07:08,557 of deep learning, the ability to really 123 00:07:08,640 --> 00:07:13,277 take any sort of problem and to get to the meat of it, to have an AI that 124 00:07:13,360 --> 00:07:16,637 really understands the underlying rules that generated the data. So, it's not 125 00:07:16,720 --> 00:07:19,637 about the data itself, it's about understanding the underlying process, 126 00:07:19,720 --> 00:07:22,477 and then be able to apply to new contexts. So, you can do that in world 127 00:07:22,560 --> 00:07:26,197 models, you can do that in scientific discovery, you can do that in coding. 128 00:07:26,280 --> 00:07:28,117 And I think that where we are, as we 129 00:07:28,200 --> 00:07:30,557 think about the rollout of this 130 00:07:30,640 --> 00:07:34,637 technology, is again that the 131 00:07:34,720 --> 00:07:36,557 you know, there's been this debate of 132 00:07:36,640 --> 00:07:41,797 how far will the text models go? How far can text intelligence go? Can you have a 133 00:07:41,880 --> 00:07:45,037 real conception of how the world operates? And I think that we have 134 00:07:45,120 --> 00:07:49,277 definitively answered that question of it's it is going to go to AGI. Like we 135 00:07:49,360 --> 00:07:51,877 see line of sight and that it is at this 136 00:07:51,960 --> 00:07:56,477 point we have line of sight to these much better models that are coming this 137 00:07:56,560 --> 00:08:01,237 year and the the the amount of pain within OpenAI 138 00:08:01,320 --> 00:08:05,517 that we've had to decide how to allocate compute, that goes up not down over 139 00:08:05,600 --> 00:08:09,997 time. And so I think that maybe the core of it is that we have 140 00:08:10,080 --> 00:08:14,997 a it's about sequencing and timing and that in this moment the kinds of 141 00:08:15,080 --> 00:08:18,517 applications that we've always dreamed of are starting to come into reach. I, 142 00:08:18,600 --> 00:08:23,077 for example, solving unsolved physics problems, right? We had this result 143 00:08:23,160 --> 00:08:27,637 recently where a physicist had been working on a problem for some time. He 144 00:08:27,720 --> 00:08:31,317 gave it to our model. 12 hours later, we have a solution and he said this is the 145 00:08:31,400 --> 00:08:34,757 first time he's seen a model where he felt like he was thinking, that it felt 146 00:08:34,840 --> 00:08:37,957 like this is a problem that maybe humanity would never solve and our AI 147 00:08:38,040 --> 00:08:42,157 solved it. When you see something like that, like you have to double down, you 148 00:08:42,240 --> 00:08:46,917 have to triple down because we can really unlock all of this potential for 149 00:08:47,000 --> 00:08:51,197 humanity. And so I I think for me it's not about relative importance of these 150 00:08:51,280 --> 00:08:55,077 things. It's more about what is OpenAI's mission of delivering 151 00:08:55,160 --> 00:08:58,477 AGI to the world, our vision of how it can benefit everyone and the fact that 152 00:08:58,560 --> 00:09:03,397 we have a tech tree, that we see how to just push it, how to do the engineering, 153 00:09:03,480 --> 00:09:07,357 do the further science and research to then have that come to fruition. 154 00:09:07,440 --> 00:09:09,917 >> Okay, so I do want to come back to the next line of 155 00:09:10,000 --> 00:09:13,277 >> models that you're anticipating, but I want to press you on this for a moment. 156 00:09:13,360 --> 00:09:17,917 I was speaking with Demis Hassabis from Google DeepMind earlier this year and 157 00:09:18,000 --> 00:09:19,957 interestingly he said that the thing 158 00:09:20,040 --> 00:09:25,277 closest that feels closest to AGI for him was Nano Banana, the image generator 159 00:09:25,360 --> 00:09:26,917 that they have. And the reason is 160 00:09:27,000 --> 00:09:28,957 because for an image generator or a 161 00:09:29,040 --> 00:09:32,317 video generator to create the images and 162 00:09:32,400 --> 00:09:36,037 the videos that it makes, it does have to understand the interaction between 163 00:09:36,120 --> 00:09:40,317 objects and have at least some conception of how the world works. So, 164 00:09:40,400 --> 00:09:45,157 is this a potential I mean it's a big bet, but does OpenAI potentially miss 165 00:09:45,240 --> 00:09:49,117 something by doubling down on the other tree if that's the case? 166 00:09:49,200 --> 00:09:51,637 >> So, two answers. >> One is absolutely. Yeah. 167 00:09:51,720 --> 00:09:55,357 >> Right, there still is not like in this field you do have to 168 00:09:55,440 --> 00:09:58,277 make choices, right? You have to make a bet. And that's actually where OpenAI 169 00:09:58,360 --> 00:10:03,117 started is we really said, "What is the path to AGI that we believe in?" And 170 00:10:03,200 --> 00:10:05,837 really focused hard on that. >> Right, the sum of random vectors is 171 00:10:05,920 --> 00:10:09,917 zero, but if you align your vectors, then you can go in a direction. 172 00:10:10,000 --> 00:10:14,557 But the second point is it's actually ImageGen is something that has been 173 00:10:14,640 --> 00:10:17,797 very, very popular within ChatGPT, and that's something we're continuing to 174 00:10:17,880 --> 00:10:21,397 invest in, continuing to prioritize. And the reason we're able to do that is 175 00:10:21,480 --> 00:10:26,397 because it's not actually on the on the world model like diffusion model tech 176 00:10:26,480 --> 00:10:30,677 tech branch, it's actually based on the GPT architecture. And so there, even 177 00:10:30,760 --> 00:10:33,277 though it's a different data distribution, 178 00:10:33,360 --> 00:10:38,197 the actual core technology, the core stack, it's all one thing. And that is 179 00:10:38,280 --> 00:10:42,037 that is actually the pretty wild thing about what AGI is is that sometimes 180 00:10:42,120 --> 00:10:45,197 these very different looking applications between speech to speech, 181 00:10:45,280 --> 00:10:48,637 image generation, text, and text is by 182 00:10:48,720 --> 00:10:51,357 the way itself many faceted of like 183 00:10:51,440 --> 00:10:55,997 science and coding and personal like you know, wellness 184 00:10:56,080 --> 00:10:57,957 information, those kinds of things. 185 00:10:58,040 --> 00:11:02,117 Uh all of that you can do in one technological envelope. And so a lot of 186 00:11:02,200 --> 00:11:04,357 what I'm looking at and what we as a company are looking at from a 187 00:11:04,440 --> 00:11:08,317 technological perspective is how to have as much unification of our of our 188 00:11:08,400 --> 00:11:12,157 efforts because we really see this technology as being something that's 189 00:11:12,240 --> 00:11:15,397 going to uplift and empower the whole economy. The whole economy is a massive 190 00:11:15,480 --> 00:11:19,717 thing, and so we can't possibly do all of it, but we can do our part. 191 00:11:19,800 --> 00:11:21,917 >> That's the general part in artificial general intelligence. 192 00:11:22,000 --> 00:11:23,797 >> That's the G. >> That's what we're talking about that. 193 00:11:23,880 --> 00:11:27,117 >> It really is. >> Speaking of unifying things, uh what is 194 00:11:27,200 --> 00:11:31,077 this super app going to be? >> So, the way I think about the super app 195 00:11:31,160 --> 00:11:33,917 is so it it's going to bring together coding, 196 00:11:34,000 --> 00:11:36,477 browser, and ChatGPT. That's right. So, 197 00:11:36,560 --> 00:11:38,917 what we want is to build an endpoint 198 00:11:39,000 --> 00:11:42,197 application for you that really lets you 199 00:11:42,280 --> 00:11:44,957 experience the power of AGI. So, the 200 00:11:45,040 --> 00:11:49,517 generality. And so, if that's you think about what chat is today, I 201 00:11:49,600 --> 00:11:53,517 think chat is really going to become your personal assistant, your personal 202 00:11:53,600 --> 00:11:57,077 AGI, right? An AI that's looking out for you, that knows a lot about you, that's 203 00:11:57,160 --> 00:12:00,557 aligned with your goals, that that's trustworthy, that kind of 204 00:12:00,640 --> 00:12:03,397 represents you in this like digital world. 205 00:12:03,480 --> 00:12:08,277 Codex, you can think of as right now it's been a tool that we built 206 00:12:08,360 --> 00:12:11,757 for software engineers, but it's becoming Codex for everyone. 207 00:12:11,840 --> 00:12:14,397 That anyone who wants to build can use 208 00:12:14,480 --> 00:12:17,677 Codex and to produce to get the computer 209 00:12:17,760 --> 00:12:21,957 to go do the thing that they want. And it's not just about the actual software 210 00:12:22,040 --> 00:12:24,357 anymore. It's really about almost the 211 00:12:24,440 --> 00:12:29,677 use of computer, whether it's to set up like I used to set settings on my 212 00:12:29,760 --> 00:12:32,837 laptop. Like I forget how how does that the hot corners? You just ask Codex to 213 00:12:32,920 --> 00:12:35,517 do it, it just does it. Right? That's what computers were always supposed to 214 00:12:35,600 --> 00:12:38,997 be is contort to the human rather than me contort to them. And so, imagine one 215 00:12:39,080 --> 00:12:43,157 application that anything you want your computer to do, you can ask it. And so, 216 00:12:43,240 --> 00:12:48,117 there's like a that there's computer use browsing built in for an AI to be able 217 00:12:48,200 --> 00:12:51,397 to actually use web browser and for you to be able to oversee what the AI is 218 00:12:51,480 --> 00:12:55,437 doing, that all of your conversations, regardless of application, whether it's 219 00:12:55,520 --> 00:12:58,557 for chat or whether it's for code, whether it's for general knowledge work, 220 00:12:58,640 --> 00:13:00,597 that's all unified in one way, that the 221 00:13:00,680 --> 00:13:03,877 AI has memory, knows about you. So, that 222 00:13:03,960 --> 00:13:08,717 is what we are building. But, it's really an iceberg because that's the tip. 223 00:13:08,800 --> 00:13:11,197 What to me is actually much more important is the technological 224 00:13:11,280 --> 00:13:14,917 unification. And we talked about it a little bit in the case of the underlying 225 00:13:15,000 --> 00:13:19,317 models, but the thing that's really changed over the past couple years has 226 00:13:19,400 --> 00:13:23,117 been that it's no longer just about the model, it's about the harness, it's 227 00:13:23,200 --> 00:13:24,917 about how does the model get context? 228 00:13:25,000 --> 00:13:29,317 How is it connected to the world? What actions can it take? How does the 229 00:13:29,400 --> 00:13:33,637 actual, as you get new context, how does the loop of interacting with the model 230 00:13:33,720 --> 00:13:37,157 work? All of that was something that we had multiple implementations of or 231 00:13:37,240 --> 00:13:40,197 slightly different and we're converging it. We're going to have one version of 232 00:13:40,280 --> 00:13:43,237 that and almost end up with this AI 233 00:13:43,320 --> 00:13:47,757 layer that can be pointed at specific applications in a very thin way. So you 234 00:13:47,840 --> 00:13:51,477 can build a little plug-in, a little skill, a little UI if you really want 235 00:13:51,560 --> 00:13:53,477 something that's great for finance, if you want something that's great for 236 00:13:53,560 --> 00:13:57,357 legal. But you generally won't have to because there'll be this one super app 237 00:13:57,440 --> 00:14:00,837 that will be very broad. >> This app is for business use cases, 238 00:14:00,920 --> 00:14:02,077 personal use cases? 239 00:14:02,160 --> 00:14:04,797 >> So both. And And that is that is really 240 00:14:04,880 --> 00:14:09,757 the core is that just like a computer, like your laptop, is it for personal, is 241 00:14:09,840 --> 00:14:13,237 it for business? Right? Well, both. Both. And it's for 242 00:14:13,320 --> 00:14:15,637 you. It's your personal machine that 243 00:14:15,720 --> 00:14:18,357 gives you a interface to this digital 244 00:14:18,440 --> 00:14:20,557 world. And that's what we want to build. 245 00:14:20,640 --> 00:14:22,757 >> So just talk a little bit about 246 00:14:22,840 --> 00:14:28,037 from a non-business standpoint, I'm using the super app in my personal life. 247 00:14:28,120 --> 00:14:29,997 What am I using it for? How does my life change? 248 00:14:30,080 --> 00:14:34,637 >> So I would think of it as So personal life, just the way that you use ChatGPT, 249 00:14:34,720 --> 00:14:36,117 right? How do you use ChatGPT right now? 250 00:14:36,200 --> 00:14:37,997 And people use it for such a diversity 251 00:14:38,080 --> 00:14:40,117 of really amazing applications. 252 00:14:40,200 --> 00:14:44,637 Sometimes that's just asking for I'm going to give a speech at a wedding, can 253 00:14:44,720 --> 00:14:48,637 you help me with drafting it? Can you give me some feedback on this idea that 254 00:14:48,720 --> 00:14:52,397 I have? I'm working on a small business, can you give me some ideas there? Which 255 00:14:52,480 --> 00:14:55,077 maybe starts to bridge between personal and work. 256 00:14:55,160 --> 00:14:58,677 There's Any of those questions should be things 257 00:14:58,760 --> 00:15:02,717 that you can go to the super app for and it answers. But that if you think about 258 00:15:02,800 --> 00:15:06,517 what ChatGPT has been, it's already been evolving. It used to not have any 259 00:15:06,600 --> 00:15:10,877 memory, right? It's just the same AI for everyone starting from scratch. 260 00:15:10,960 --> 00:15:12,717 It's almost like talking to a stranger. 261 00:15:12,800 --> 00:15:16,677 It's way more powerful if it remembers, remembers the interactions you've had. 262 00:15:16,760 --> 00:15:20,037 It's way more powerful if it has access to context, right? That if it's hooked 263 00:15:20,120 --> 00:15:22,757 up to your email and to your calendar 264 00:15:22,840 --> 00:15:25,357 and really knows your preferences and 265 00:15:25,440 --> 00:15:28,957 has this this almost deeper set of 266 00:15:29,040 --> 00:15:32,077 of just you know past experiences with 267 00:15:32,160 --> 00:15:36,317 you that it's able to leverage to achieve your goals and to 268 00:15:36,400 --> 00:15:40,197 you look at things like pulse is a feature in ChatGPT right now where every 269 00:15:40,280 --> 00:15:44,637 day it surfaces for you things that you might be interested in based on what 270 00:15:44,720 --> 00:15:46,677 ChatGPT knows about you. So I say that 271 00:15:46,760 --> 00:15:49,677 in the personal capacity that the super 272 00:15:49,760 --> 00:15:54,917 app will be doing all of that and will be doing it a much deeper and richer way. 273 00:15:55,000 --> 00:15:56,037 >> When are you planning to ship it? 274 00:15:56,120 --> 00:16:00,077 >> Uh so the way to think about it is we should 275 00:16:00,160 --> 00:16:03,437 we're taking incremental steps to get there over the next couple months. We 276 00:16:03,520 --> 00:16:07,197 should have shipped the complete 277 00:16:07,280 --> 00:16:09,397 vision of what we're talking about here, 278 00:16:09,480 --> 00:16:11,277 but it's going to come in pieces and the 279 00:16:11,360 --> 00:16:14,637 place that we're starting is with 280 00:16:14,720 --> 00:16:17,357 for example the Codex app today is 281 00:16:17,440 --> 00:16:19,917 something which is a it's really two 282 00:16:20,000 --> 00:16:23,677 things in one. It's a general agent 283 00:16:23,760 --> 00:16:26,877 harness that can use tools and it's also 284 00:16:26,960 --> 00:16:30,997 a agent that knows how to write software. 285 00:16:31,080 --> 00:16:34,957 That general agent harness that can be used for so many different things. You 286 00:16:35,040 --> 00:16:37,517 hook it up to spreadsheets, you hook it up to 287 00:16:37,600 --> 00:16:40,837 to word documents. It's able to help you with knowledge work. And so we're going 288 00:16:40,920 --> 00:16:45,477 to make the Codex app just so much more usable for general knowledge work 289 00:16:45,560 --> 00:16:48,637 because it already what we've seen within OpenAI is all this organic 290 00:16:48,720 --> 00:16:51,717 adoption of people using it for that. So that'll be the first step and there are 291 00:16:51,800 --> 00:16:53,637 many to come. >> I was speaking with one of your 292 00:16:53,720 --> 00:16:55,277 colleagues yesterday taking a look at 293 00:16:55,360 --> 00:16:58,277 Codex and he mentioned that someone 294 00:16:58,360 --> 00:17:01,317 using Codex had built had had instructed 295 00:17:01,400 --> 00:17:03,237 Codex to help them with video editing. 296 00:17:03,320 --> 00:17:06,517 It built a plugin for Adobe Premiere, 297 00:17:06,600 --> 00:17:11,157 started separating it into chapters and started the edit. 298 00:17:11,240 --> 00:17:14,197 That's what we're looking at. >> I I love hearing that. That's exactly 299 00:17:14,280 --> 00:17:17,437 exactly the kinds of things that we want 300 00:17:17,520 --> 00:17:20,637 this system to be useful for. And it's 301 00:17:20,720 --> 00:17:25,517 been really interesting seeing like the Codex app itself was originally built 302 00:17:25,600 --> 00:17:27,597 for software engineers, and that it's 303 00:17:27,680 --> 00:17:30,717 almost like the current usability of it 304 00:17:30,800 --> 00:17:35,397 for non-software engineers is actually quite low because there's a bunch of 305 00:17:35,480 --> 00:17:39,477 little things where when you set things up, you run into some error that a 306 00:17:39,560 --> 00:17:43,157 developer knows what it means, knows how to fix it. It's just kind of what we're 307 00:17:43,240 --> 00:17:47,317 used to. But if you're not a developer, you're like, 308 00:17:47,400 --> 00:17:50,157 "What is this? Like this this is not something that I that I've encountered 309 00:17:50,240 --> 00:17:54,637 before." And despite that, we are seeing people start to use this who have never 310 00:17:54,720 --> 00:17:58,157 programmed before to be able to build websites, to be able to do exactly the 311 00:17:58,240 --> 00:18:01,637 kinds of things you you said of like be able to automate different their 312 00:18:01,720 --> 00:18:04,677 interactions with different pieces of software, be able to get lots lots of 313 00:18:04,760 --> 00:18:08,597 leverage. Like someone on our communications team used it to it's 314 00:18:08,680 --> 00:18:10,797 hooked up to Slack, to their email. 315 00:18:10,880 --> 00:18:13,797 They're able to go through a bunch of feedback, be able to synthesize it very 316 00:18:13,880 --> 00:18:17,397 well. So these kinds of tasks, people 317 00:18:17,480 --> 00:18:21,397 who are very motivated can jump through the hoops and then get great return from 318 00:18:21,480 --> 00:18:25,917 it. And so some extent we did the super hard part of an AI that is really smart, 319 00:18:26,000 --> 00:18:30,037 capable, can actually accomplish your task. Now we have to do the much easier 320 00:18:30,120 --> 00:18:34,677 part in some sense of make it broadly useful and to remove these barriers to 321 00:18:34,760 --> 00:18:37,037 entry. >> And just looking at the competitive 322 00:18:37,120 --> 00:18:41,117 landscape, I mean Anthropic they have the Claude app. You can use Claude the 323 00:18:41,200 --> 00:18:44,037 chatbot, Claude co-write, Claude code. 324 00:18:44,120 --> 00:18:46,557 Uh so they have a version of a super app 325 00:18:46,640 --> 00:18:49,077 of their own. I'm curious what you think 326 00:18:49,160 --> 00:18:53,837 Anthropic saw that got them to this position earlier, and 327 00:18:53,920 --> 00:18:56,077 what do you think your chances are of catching up there? 328 00:18:56,160 --> 00:18:59,877 >> Well, I think that if you 329 00:18:59,960 --> 00:19:03,637 rewind 12 18 always been focused on 330 00:19:03,720 --> 00:19:08,597 coding as a domain. We always had the best numbers on 331 00:19:08,680 --> 00:19:11,837 different programming competitions, these very cerebral things. But the 332 00:19:11,920 --> 00:19:15,837 thing that we didn't invest in as much was that last mile of usability. Of 333 00:19:15,920 --> 00:19:18,757 really trying to think about, okay, this this AI is so smart, it can solve all 334 00:19:18,840 --> 00:19:22,357 these great programming competitions, but it's never seen someone's real-world 335 00:19:22,440 --> 00:19:26,877 codebase, which is messy and not quite as pristine as the world that it it sort 336 00:19:26,960 --> 00:19:30,517 of has experienced. And I think that is something that we were behind on. But 337 00:19:30,600 --> 00:19:35,157 about, you know, maybe mid last year is when we got very serious about that. And 338 00:19:35,240 --> 00:19:38,157 that we had a team very focused on what are all 339 00:19:38,240 --> 00:19:41,157 the the gaps, what are all the kind of messiness the real world would have and 340 00:19:41,240 --> 00:19:44,237 we haven't encountered. How do we actually get 341 00:19:44,320 --> 00:19:48,917 training data that that build training environments that let the AI experience 342 00:19:49,000 --> 00:19:52,037 what it's like to actually do software engineering, be interrupted in weird 343 00:19:52,120 --> 00:19:55,437 ways, all those things. And I'd say at this point we are caught up. When people 344 00:19:55,520 --> 00:19:59,357 go head-to-head, for us versus competitors, the people tend to prefer 345 00:19:59,440 --> 00:20:03,397 us. We do know where we're lagging in front end, we're going to fix that, but 346 00:20:03,480 --> 00:20:06,637 this is the general motion that we've been we've been we've been taking is to 347 00:20:06,720 --> 00:20:09,637 say that that usability of thinking 348 00:20:09,720 --> 00:20:14,197 about the product end-to-end, not just a model, and then build a separate thing, 349 00:20:14,280 --> 00:20:17,517 right? Really think of it as one product. When we're doing the research, 350 00:20:17,600 --> 00:20:20,557 we're thinking about how will it be used. That has been a motion that we've 351 00:20:20,640 --> 00:20:23,277 been changing within OpenAI. And so, I 352 00:20:23,360 --> 00:20:26,117 think that the way I would look at it is 353 00:20:26,200 --> 00:20:29,877 that we have incredible step-up models 354 00:20:29,960 --> 00:20:33,637 coming. Like this whole year I look at the road map, it's truly inspiring. What 355 00:20:33,720 --> 00:20:35,717 will be possible. And then we've been 356 00:20:35,800 --> 00:20:38,797 really focusing now on let's also get 357 00:20:38,880 --> 00:20:40,997 the last mile usability. 358 00:20:41,080 --> 00:20:43,477 >> So since 2022, OpenAI has been like the 359 00:20:43,560 --> 00:20:47,597 undisputed leader and uh obviously now 360 00:20:47,680 --> 00:20:51,197 the competition is intense. Like you just used the word we're caught the 361 00:20:51,280 --> 00:20:54,637 phrase we're caught up. Is there a different vibe within the company where 362 00:20:54,720 --> 00:20:59,437 it's like now instead of the one that's like far ahead on something like ChatGPT 363 00:20:59,520 --> 00:21:03,197 in a real in a real fight? I mean, you're seeing it come out of some of the 364 00:21:03,280 --> 00:21:07,837 reporting on what's happening within the company, the fact that there are no more 365 00:21:07,920 --> 00:21:12,197 there's been meetings, there's no more side quests at OpenAI. It's all focus on 366 00:21:12,280 --> 00:21:14,997 this. How's the environment? Has the vibe changed here? 367 00:21:15,080 --> 00:21:18,957 >> Well, I would say that for me personally, yeah, the scariest moment at 368 00:21:19,040 --> 00:21:23,157 OpenAI was actually after we launched ChatGPT. And [clears throat] I remember 369 00:21:23,240 --> 00:21:26,837 being at the holiday party and just feeling this vibe of we won. I have 370 00:21:26,920 --> 00:21:31,557 never felt that. I was like, no, that we we are the underdog and we always have 371 00:21:31,640 --> 00:21:35,637 been, right? The the competitors in this space, established companies that have 372 00:21:35,720 --> 00:21:38,157 just sort of much more capital, much 373 00:21:38,240 --> 00:21:40,797 more, you know, human resources, data, 374 00:21:40,880 --> 00:21:42,997 the whole the whole thing. Why is OpenAI 375 00:21:43,080 --> 00:21:48,037 able to compete at all? And to some extent, the answer is only because we 376 00:21:48,120 --> 00:21:52,317 never feel complacent, where we always feel like we are the challenger. And it 377 00:21:52,400 --> 00:21:56,237 actually for me has been a very healthy thing to see 378 00:21:56,320 --> 00:22:00,077 us start to see that in the marketplace, to see other competitors emerge and do a 379 00:22:00,160 --> 00:22:04,477 good job. And that that is, you know, in my mind, you can never fix it on your 380 00:22:04,560 --> 00:22:08,117 competitors. If you focus on where they are, then you'll be where they are and 381 00:22:08,200 --> 00:22:11,557 they'll already have moved. And I think that that's what's been happening in the 382 00:22:11,640 --> 00:22:12,757 other direction, right? A lot of people 383 00:22:12,840 --> 00:22:14,797 have focused on exactly where we are and 384 00:22:14,880 --> 00:22:18,637 we get to move. And I think that the 385 00:22:18,720 --> 00:22:23,437 it almost gives us and this alignment, this unification of the company. And I 386 00:22:23,520 --> 00:22:26,957 kind of described how we almost thought of research and deployment as separate 387 00:22:27,040 --> 00:22:31,237 things. And now we really want to integrate them. Like, that to me is such 388 00:22:31,320 --> 00:22:33,077 a wonderful thing. And so, 389 00:22:33,160 --> 00:22:38,997 I'd say that the world that we're in is one where I've never felt like we were 390 00:22:39,080 --> 00:22:41,357 you know, you're never as good as they say you are, you're never as bad as they 391 00:22:41,440 --> 00:22:46,597 say you are. I think it's just been very steady. And that the core of the model 392 00:22:46,680 --> 00:22:50,517 production, that is something where I actually feel extremely extremely 393 00:22:50,600 --> 00:22:54,077 confident in our road map, the research investments we've been made making. And 394 00:22:54,160 --> 00:22:57,237 I think on the product side, we have such great energy that's all coming 395 00:22:57,320 --> 00:22:59,557 together to deliver this to the world. 396 00:22:59,640 --> 00:23:04,437 >> You foreshadowed a couple times already that you have some good models on the way. 397 00:23:04,520 --> 00:23:06,037 What is Spud? 398 00:23:06,120 --> 00:23:09,077 Uh the information said uh that you 399 00:23:09,160 --> 00:23:11,797 finished uh pre-training Spud, and Sam 400 00:23:11,880 --> 00:23:14,557 Altman, the CEO of OpenAI, has told the 401 00:23:14,640 --> 00:23:18,957 staff that uh they should expect to have a very strong model in a few weeks. This 402 00:23:19,040 --> 00:23:22,077 was a few weeks ago. And the team believes it can really 403 00:23:22,160 --> 00:23:23,917 accelerate the economy, and things are 404 00:23:24,000 --> 00:23:26,517 moving faster than many of us expect 405 00:23:26,600 --> 00:23:29,477 expected. So, what's Spud? 406 00:23:29,560 --> 00:23:33,757 >> It's a good model. Um but I think that it's really not about any one model. 407 00:23:33,840 --> 00:23:37,757 Okay, right? The the the way that our development process works is you have 408 00:23:37,840 --> 00:23:41,917 pre-training, so you produce a new base model that then is the foundation that 409 00:23:42,000 --> 00:23:45,357 we build further improvements on top of. And that 410 00:23:45,440 --> 00:23:49,277 that is always a huge effort across many people in the company, and that's where 411 00:23:49,360 --> 00:23:52,597 I've actually been spending most of my efforts over the past 18 months has been 412 00:23:52,680 --> 00:23:56,637 really focused on our GPU infrastructure, on supporting the 413 00:23:56,720 --> 00:23:59,717 teams that do all of the training frameworks to scale up at these big 414 00:23:59,800 --> 00:24:03,877 runs. But then there's a reinforcement learning process. So, you take this AI 415 00:24:03,960 --> 00:24:07,837 that has learned lots of things about the world, and it applies that 416 00:24:07,920 --> 00:24:10,197 knowledge, and then we do a post-training process where you really 417 00:24:10,280 --> 00:24:13,517 say, "Okay, now you know how to solve problems. You practice it in in all 418 00:24:13,600 --> 00:24:16,437 these different contexts." And then here's kind of the last mile of of 419 00:24:16,520 --> 00:24:18,357 behavior and usability. So, I think of 420 00:24:18,440 --> 00:24:21,837 Spud as a new base, as a new pre-train, 421 00:24:21,920 --> 00:24:25,957 and that we have had this uh I I'd say 422 00:24:26,040 --> 00:24:30,317 it's like we have maybe 2 years' worth of research that is coming to fruition 423 00:24:30,400 --> 00:24:33,517 in this model. It's going to be very exciting. 424 00:24:33,600 --> 00:24:37,157 And I think that the way that the world will experience it is just improved 425 00:24:37,240 --> 00:24:41,957 capabilities. And that for me, it's never about any one release, because as 426 00:24:42,040 --> 00:24:45,997 soon as we have this one release, it'll be an early version of what we have 427 00:24:46,080 --> 00:24:50,477 coming. We'll do much more of each of these steps of the improvement process. 428 00:24:50,560 --> 00:24:54,837 And so, I think that where we're going is this almost just we have this engine 429 00:24:54,920 --> 00:24:58,957 of progress that just moves faster and faster. And this one is just one step 430 00:24:59,040 --> 00:25:01,877 along the way. And so, what do you think it'll be able to do that today's models 431 00:25:01,960 --> 00:25:04,597 can't? I So, I think it's going to be 432 00:25:04,680 --> 00:25:07,237 able to solve both much harder problems. 433 00:25:07,320 --> 00:25:09,397 I think it will be much more nuanced. 434 00:25:09,480 --> 00:25:12,077 It'll understand instructions better. 435 00:25:12,160 --> 00:25:15,317 It'll understand the context much better. That there's this thing called 436 00:25:15,400 --> 00:25:18,637 big model smell that people talk about where it's just like there's something 437 00:25:18,720 --> 00:25:22,997 about like when these models are just actually just much smarter, much more 438 00:25:23,080 --> 00:25:28,117 capable, that they bend to you much more. You feel it. 439 00:25:28,200 --> 00:25:31,397 Right? When you ask a question and the AI doesn't quite get it, it's always so 440 00:25:31,480 --> 00:25:34,237 disappointing. Right? When you have to like explain and you're just like, "You 441 00:25:34,320 --> 00:25:37,357 really should be able to figure this out." And so, I would just think of it 442 00:25:37,440 --> 00:25:42,317 as in some ways just qualitatively, there will be 443 00:25:42,400 --> 00:25:44,197 but quantitatively lots of shifts. 444 00:25:44,280 --> 00:25:46,837 Right? And qualitatively, there will just be new things where you would be 445 00:25:46,920 --> 00:25:50,037 frustrated before, you never use an AI for it, and now you just use it without 446 00:25:50,120 --> 00:25:52,837 without thinking very much. And I think that that that is what we're going to 447 00:25:52,920 --> 00:25:57,237 see across the board. I'm super excited to see how it raises the ceiling. Right? 448 00:25:57,320 --> 00:25:59,837 We've already seen these physics applications, things like that. And I 449 00:25:59,920 --> 00:26:04,157 think we will be able to just solve like way more open-ended problems, way longer 450 00:26:04,240 --> 00:26:08,677 time horizons. And then also very excited to see how it raises the floor, 451 00:26:08,760 --> 00:26:12,237 where just for anything you want to do, it's just so much useful more useful for 452 00:26:12,320 --> 00:26:14,317 you. >> It can be kind of tough for everyday 453 00:26:14,400 --> 00:26:17,797 users to really feel the change. Like there was uh talk about a lot of 454 00:26:17,880 --> 00:26:22,357 build-up before GPT-5 came out, and then it came out, and actually the initial 455 00:26:22,440 --> 00:26:26,357 reaction was somewhat disappointment among the public. But then, I think 456 00:26:26,440 --> 00:26:30,237 people realized that for certain tasks it was really good. Or 457 00:26:30,320 --> 00:26:35,197 with these next series of models, do you expect that it'll really be felt sort of 458 00:26:35,280 --> 00:26:38,917 um in the trenches in certain occupations, or do you think it'll be a 459 00:26:39,000 --> 00:26:43,037 broadly uh tangible improvement for everyone? 460 00:26:43,120 --> 00:26:47,277 >> I think that it will be a similar story where you when you 461 00:26:47,360 --> 00:26:50,277 release it, there will be people who will try it and be like, "This is a 462 00:26:50,360 --> 00:26:53,757 night and day different than anything I've seen. And then there will be some 463 00:26:53,840 --> 00:26:56,957 applications where we weren't necessarily intelligence bottlenecked. 464 00:26:57,040 --> 00:26:59,357 And so if you have a model that's more intelligent, maybe you won't feel it 465 00:26:59,440 --> 00:27:02,717 right there. But I think over time that 466 00:27:02,800 --> 00:27:07,437 you will feel it because the fundamental thing that shifts is how much do you 467 00:27:07,520 --> 00:27:10,917 rely on the system? Like if you think about the way we all interact with AI, 468 00:27:11,000 --> 00:27:14,277 we have some mental model for what we think it can do. And that mental model 469 00:27:14,360 --> 00:27:17,557 shifts actually fairly slowly. Right? As you get more experience, it does 470 00:27:17,640 --> 00:27:20,917 something magical for you. You're like, "Oh, wow, it can do that? I never 471 00:27:21,000 --> 00:27:26,317 imagined that." And we see this for example in applications like 472 00:27:26,400 --> 00:27:27,917 access to health information, right? 473 00:27:28,000 --> 00:27:30,877 That we see people who 474 00:27:30,960 --> 00:27:33,237 You know, like I I have a a friend who 475 00:27:33,320 --> 00:27:38,437 used ChatGPT to understand different treatments for his cancer. And that he 476 00:27:38,520 --> 00:27:40,757 was told by doctors that he was terminal, that there's nothing they 477 00:27:40,840 --> 00:27:43,677 could do for him. He used ChatGPT to actually research 478 00:27:43,760 --> 00:27:46,957 a bunch of different ideas and he was able to get treatment that way. And 479 00:27:47,040 --> 00:27:50,957 that's something where you need to have some level of belief 480 00:27:51,040 --> 00:27:54,437 that the AI is going to be helpful in that application for you to really put 481 00:27:54,520 --> 00:27:57,477 in the effort to get something out of the machine. And I think what we're 482 00:27:57,560 --> 00:28:01,797 going to see is that for any application like that, it's going to become so much 483 00:28:01,880 --> 00:28:06,357 more evident to everyone that the AI can help you. 484 00:28:06,440 --> 00:28:09,757 And so I think it's a little bit of the technology getting better, but it's 485 00:28:09,840 --> 00:28:11,997 also our understanding of the technology 486 00:28:12,080 --> 00:28:14,037 shifting and and catching up to that. 487 00:28:14,120 --> 00:28:16,797 And you'll be relying on it more. Inside 488 00:28:16,880 --> 00:28:20,077 OpenAI, you have uh an automated AI 489 00:28:20,160 --> 00:28:21,797 researcher in the works that's supposed 490 00:28:21,880 --> 00:28:23,837 to come out this fall. 491 00:28:23,920 --> 00:28:27,837 What is that? So, the the direction of 492 00:28:27,920 --> 00:28:32,117 travel right now, we are in this early phase of takeoff of this technology. 493 00:28:32,200 --> 00:28:36,317 What does takeoff mean? Takeoff is as the AI gets better and better on this 494 00:28:36,400 --> 00:28:38,957 exponential. And 495 00:28:39,040 --> 00:28:43,957 in part because we can use the AI to make the AI better. Mhm. So, our 496 00:28:44,040 --> 00:28:47,557 development process speeds up. But I also think when I think of takeoff, it's 497 00:28:47,640 --> 00:28:49,837 also about real world impact. And in 498 00:28:49,920 --> 00:28:54,957 some ways we've been, you know, every technology is is an S curve. 499 00:28:55,040 --> 00:28:58,717 Or if you zoom out, some of S curves that end up being an exponential. And I 500 00:28:58,800 --> 00:29:02,077 think that that's what we're encountering right now. So, it's the 501 00:29:02,160 --> 00:29:04,717 tech technology development is moving 502 00:29:04,800 --> 00:29:07,477 with increasing speed and it's this 503 00:29:07,560 --> 00:29:10,437 engine that's picking up momentum. 504 00:29:10,520 --> 00:29:12,957 But it's also in the world, there's all 505 00:29:13,040 --> 00:29:16,197 of these tailwinds because there's chip 506 00:29:16,280 --> 00:29:20,237 developers that are getting more resourcing into their programs. There's 507 00:29:20,320 --> 00:29:23,197 this economy of people who are building on top of it, trying to figure out how 508 00:29:23,280 --> 00:29:26,717 it fits into every different application. And all of that energy is 509 00:29:26,800 --> 00:29:29,037 just accumulating more and more into 510 00:29:29,120 --> 00:29:34,037 this this takeoff phase of the AI becoming just a kind of sideshow to 511 00:29:34,120 --> 00:29:37,877 being the main driver of economic growth. And I think that that is 512 00:29:37,960 --> 00:29:42,037 something that it's not just about what we're doing in these walls. It's about 513 00:29:42,120 --> 00:29:45,677 how the whole world, the whole economy comes together and are the to push 514 00:29:45,760 --> 00:29:48,757 forward this technology and its usefulness together. And the researcher, 515 00:29:48,840 --> 00:29:50,597 well, then what will it do exactly? 516 00:29:50,680 --> 00:29:55,117 Also, the researcher will be a moment where the AI which we're building that 517 00:29:55,200 --> 00:29:59,037 right now it's taking a you know, it's doing a larger percentage of tasks, that 518 00:29:59,120 --> 00:30:02,277 we should be able to let it run autonomously. And that I think there's a 519 00:30:02,360 --> 00:30:05,797 lot of thought that goes into what that means. And that it doesn't necessarily 520 00:30:05,880 --> 00:30:09,957 mean that we just let it off on its own and then come back later and see if it 521 00:30:10,040 --> 00:30:13,357 did something good. I think that we are going to be very involved in managing 522 00:30:13,440 --> 00:30:17,277 it. Right? Just like right now if you have a junior researcher, 523 00:30:17,360 --> 00:30:19,837 if you leave them on their own too long, they're probably going to go down a path 524 00:30:19,920 --> 00:30:23,597 that's not very useful. But if you have a senior researcher or someone who has a 525 00:30:23,680 --> 00:30:27,477 vision, they don't even necessarily need to know the mechanical skills, 526 00:30:27,560 --> 00:30:31,237 they will be able to provide feedback, review the plots that this person's that 527 00:30:31,320 --> 00:30:36,317 the you know, the interns producing and to provide direction in terms of the 528 00:30:36,400 --> 00:30:39,597 vision of what is it that I want you to accomplish. And so I think of this as a 529 00:30:39,680 --> 00:30:42,997 system that we're going to build that will massively accelerate our ability to 530 00:30:43,080 --> 00:30:47,397 produce models, to make new research breakthroughs happen, to be able to make 531 00:30:47,480 --> 00:30:50,957 these models more useful and usable in the real world. 532 00:30:51,040 --> 00:30:53,117 >> Right. >> And to do that at increasing speed. 533 00:30:53,200 --> 00:30:56,717 >> So, sorry. What's it going to do? Are you going to say go find AGI and it will 534 00:30:56,800 --> 00:30:59,077 just try to >> I I think I think the way I think of it 535 00:30:59,160 --> 00:31:01,117 is is something like that to first order. 536 00:31:01,200 --> 00:31:03,237 >> Okay. >> And at a practical level, I think I 537 00:31:03,320 --> 00:31:06,597 would view it as is taking the full 538 00:31:06,680 --> 00:31:10,637 end-to-end of what one of our research scientists does 539 00:31:10,720 --> 00:31:12,277 and be able to do that in silicon. 540 00:31:12,360 --> 00:31:16,437 >> Another way to think about takeoff is their progress in AI goes from 541 00:31:16,520 --> 00:31:18,597 incremental to 542 00:31:18,680 --> 00:31:23,477 gathering momentum and then sort of this unstoppable march to an intelligence 543 00:31:23,560 --> 00:31:26,117 that's smarter than humans. 544 00:31:26,200 --> 00:31:30,677 Um do you do you worry that 545 00:31:30,760 --> 00:31:34,957 there's just as there's possibilities for things to go right on that front, 546 00:31:35,040 --> 00:31:36,317 there's also possibilities for that 547 00:31:36,400 --> 00:31:39,757 progress that process to go wrong? 548 00:31:39,840 --> 00:31:42,677 >> I mean, I think that that's absolutely 549 00:31:42,760 --> 00:31:47,157 yes. I think that the way to get the benefits of this technology is also to 550 00:31:47,240 --> 00:31:48,917 really think about the risks. >> Mhm. 551 00:31:49,000 --> 00:31:52,757 >> And if you look at how we've approached technology development from a technical 552 00:31:52,840 --> 00:31:57,477 perspective, we invest a lot in safety, security. 553 00:31:57,560 --> 00:31:59,837 Good example of this is prompt injections. All right, if you're going 554 00:31:59,920 --> 00:32:04,957 to have an AI that is very smart, very capable, hooked up to lots of tools, 555 00:32:05,040 --> 00:32:07,357 you want to make sure that it can't be subverted by someone giving it a weird 556 00:32:07,440 --> 00:32:09,997 instruction. And that's something we've invested in 557 00:32:10,080 --> 00:32:12,837 quite a lot and I think have really incredible results, have an incredible 558 00:32:12,920 --> 00:32:18,037 team working on. And it's interesting to think about some of these problems where 559 00:32:18,120 --> 00:32:22,437 you can make analogies to humans. Like humans are also susceptible to fishing 560 00:32:22,520 --> 00:32:25,957 attacks, to being deceived in different ways, to not really understanding the 561 00:32:26,040 --> 00:32:27,677 full context of what they're working on. 562 00:32:27,760 --> 00:32:31,557 And we bring those analogies into our development process and think about this 563 00:32:31,640 --> 00:32:35,717 whenever we release a model, develop a model, how do we ensure that it's going 564 00:32:35,800 --> 00:32:41,477 to be aligned with people and be able to actually be helpful. And 565 00:32:41,560 --> 00:32:44,317 that is something that we care quite a lot about. I think that there are bigger 566 00:32:44,400 --> 00:32:47,957 questions about the world, the economy, how does everything change, how does 567 00:32:48,040 --> 00:32:49,797 everyone benefit from this technology. 568 00:32:49,880 --> 00:32:53,957 They're not purely technical, not purely something that OpenAI on our own will be 569 00:32:54,040 --> 00:32:58,597 able to solve. But yes, I I think quite a lot about not just pushing forward the 570 00:32:58,680 --> 00:33:02,597 technology, but also really about how do we ensure that we have the positive 571 00:33:02,680 --> 00:33:07,037 impact that is its potential. >> The worry though is that this is a race 572 00:33:07,120 --> 00:33:11,157 and what's being done within these walls where at OpenAI 573 00:33:11,240 --> 00:33:15,277 headquarters is also being copied by many of the open source players, which 574 00:33:15,360 --> 00:33:18,597 have much less bare boundaries and barriers and 575 00:33:18,680 --> 00:33:22,997 protection on on on the safety side of things. And if you said this months that 576 00:33:23,080 --> 00:33:27,357 it takes, you know, people getting a lot of things right to be creative and sort 577 00:33:27,440 --> 00:33:31,077 of one person with bad intent to be destructive. 578 00:33:31,160 --> 00:33:35,517 And that's sort of where the concern lies for me at least is just 579 00:33:35,600 --> 00:33:39,397 when this this is it's clearly a race, it's going fast. Many of your 580 00:33:39,480 --> 00:33:42,477 counterparts have said if everybody agrees 581 00:33:42,560 --> 00:33:47,117 to stop it, we'll stop it and it doesn't seem like it's going to slow 582 00:33:47,200 --> 00:33:50,437 it all. >> So is the reward worth the risk basically? 583 00:33:50,520 --> 00:33:52,037 >> I think I think the reward is worth the 584 00:33:52,120 --> 00:33:55,317 risk, but I think that that is too too 585 00:33:55,400 --> 00:33:59,357 coarse grained of a of a answer in some sense. Okay, like the way that I think 586 00:33:59,440 --> 00:34:03,277 about it is that we've asked from the beginning of OpenAI 587 00:34:03,360 --> 00:34:07,437 how what does a great future look like? How 588 00:34:07,520 --> 00:34:11,517 can this technology really be something that uplifts everyone? 589 00:34:11,600 --> 00:34:15,637 And you can think of there almost being two different angles. One is the 590 00:34:15,720 --> 00:34:19,797 centralization view of saying that well, the way to make this technology safe is 591 00:34:19,880 --> 00:34:23,437 that you have only one actor building it and so then you don't have any 592 00:34:23,520 --> 00:34:26,277 pressures, right? You can really think about getting it right and you know, 593 00:34:26,360 --> 00:34:29,317 then figure out how to roll it out to everyone when it's ready, those kinds of 594 00:34:29,400 --> 00:34:33,477 things. That's a pretty tough pill in some ways 595 00:34:33,560 --> 00:34:36,917 and think that there's a lot of properties that you can think instead 596 00:34:37,000 --> 00:34:40,997 think about approaching differently, which we prefer to as resilience. To 597 00:34:41,080 --> 00:34:44,677 think of it as this open system where there's lots of players who are 598 00:34:44,760 --> 00:34:47,997 developing the technology. It's not not just about the technology, it's about 599 00:34:48,080 --> 00:34:50,197 building societal infrastructure that 600 00:34:50,280 --> 00:34:52,437 helps this technology really go well. 601 00:34:52,520 --> 00:34:56,597 And if you think about how electricity has developed, that's something where 602 00:34:56,680 --> 00:34:59,797 lots of people produce it, that it actually has dangers and risks, but we 603 00:34:59,880 --> 00:35:03,397 also build our safety infrastructure in 604 00:35:03,480 --> 00:35:07,597 a diversity of different ways around safety standards for electricity, around 605 00:35:07,680 --> 00:35:10,437 different ways of harnessing it, about how you scale it, that there's 606 00:35:10,520 --> 00:35:14,037 regulations when you're at these massive scales that lots of people are able to 607 00:35:14,120 --> 00:35:17,957 use in a democratized fashion, there's inspectors. Like there's a whole system 608 00:35:18,040 --> 00:35:21,437 that's been built around the needs of that technology, the proclivities of 609 00:35:21,520 --> 00:35:23,557 that specific technology. And I think 610 00:35:23,640 --> 00:35:27,157 that one thing that we have 611 00:35:27,240 --> 00:35:31,997 really I think seen with AI is that it is 612 00:35:32,080 --> 00:35:35,517 something where we need this broad conversation. We need lots of people to 613 00:35:35,600 --> 00:35:39,917 be aware if this technology is going to come and change everything for everyone, 614 00:35:40,000 --> 00:35:42,997 people need to participate in that. It can't be something that's done often 615 00:35:43,080 --> 00:35:45,557 secret by just one, you know, sort of 616 00:35:45,640 --> 00:35:48,797 centralized group. And so this has been 617 00:35:48,880 --> 00:35:53,157 to me a very core question to how this 618 00:35:53,240 --> 00:35:56,237 technology should play out and something we really believe in is this resilience 619 00:35:56,320 --> 00:35:59,957 ecosystem that should emerge around the development of this technology. 620 00:36:00,040 --> 00:36:03,757 >> So you said we're uh in takeoff, in the middle of a takeoff process and we I 621 00:36:03,840 --> 00:36:07,197 guess see all of humanity are experiencing this. 622 00:36:07,280 --> 00:36:10,077 Uh Nvidia CEO Jensen Huang said recently 623 00:36:10,160 --> 00:36:14,357 that he believes that AGI has been achieved. Do you agree? 624 00:36:14,440 --> 00:36:15,997 >> I think that AGI has a different 625 00:36:16,080 --> 00:36:18,757 definition to many people and I think 626 00:36:18,840 --> 00:36:23,477 that there are many people who would say that what we have right now is AGI. I 627 00:36:23,560 --> 00:36:27,117 think you can debate it. 628 00:36:27,200 --> 00:36:31,357 But I think that maybe the thing that's interesting is that AGI, like the 629 00:36:31,440 --> 00:36:35,637 technology we have right now, is very jagged. Like it is absolutely superhuman 630 00:36:35,720 --> 00:36:40,197 at many tasks. When it comes to writing code, those kinds of things, 631 00:36:40,280 --> 00:36:43,797 AI can just do it, right? And it really removes a lot of the friction to 632 00:36:43,880 --> 00:36:45,517 creating things. But there's some very 633 00:36:45,600 --> 00:36:48,277 basic tasks that a human can do that our 634 00:36:48,360 --> 00:36:53,117 AI still struggle with. And so, it's almost to say that 635 00:36:53,200 --> 00:36:56,677 where do you draw the cutline? Mhm. It's a little bit more of a vibe than a than 636 00:36:56,760 --> 00:37:00,917 a feeling than it is a uh than it is science at the moment. And 637 00:37:01,000 --> 00:37:05,277 so, I think for myself, we're definitely going through that moment. And if you 638 00:37:05,360 --> 00:37:08,597 were to show me 5 years ago the systems we have today, I'd have to go, "Oh, 639 00:37:08,680 --> 00:37:10,557 yeah. That's what we're talking about." 640 00:37:10,640 --> 00:37:13,117 But it's just different. It's so different from anything we ever 641 00:37:13,200 --> 00:37:16,797 pictured. And so, I think we need to adjust our mental models appropriately. 642 00:37:16,880 --> 00:37:18,557 So, you're not there yet. I think I 643 00:37:18,640 --> 00:37:23,957 think that I'd say I'm I'm I'm basically like 70, 80% there. So, I 644 00:37:24,040 --> 00:37:27,917 think we're we're quite close. And I think it's extremely clear that we are 645 00:37:28,000 --> 00:37:32,237 going to have AGI within the next couple of years in a way that 646 00:37:32,320 --> 00:37:37,037 is still going to be jagged, but that the the floor of task will just be 647 00:37:37,120 --> 00:37:40,877 almost for any intellectual task of how you use your computer, that the AI will 648 00:37:40,960 --> 00:37:45,437 be able to do that. And I think that that yeah, right now I have to give a 649 00:37:45,520 --> 00:37:48,917 little bit of a uncertain answer because there's some there's some it's almost 650 00:37:49,000 --> 00:37:52,717 like a like a uncertainty principle kind of thing. They you can you can you can 651 00:37:52,800 --> 00:37:56,717 debate it. For my own personal definition, I think we're almost there, 652 00:37:56,800 --> 00:38:00,197 and with maybe a little bit more, we will absolutely be. 653 00:38:00,280 --> 00:38:03,757 >> Okay. Well, we're we're going to go to a break, but as long as we're on the way 654 00:38:03,840 --> 00:38:06,757 to the break, I want to let folks watching at home know that you and I are 655 00:38:06,840 --> 00:38:09,277 going to be talking again June 18th here 656 00:38:09,360 --> 00:38:13,877 in San Francisco at SFJAZZ. So, I will put some information if you want to come 657 00:38:13,960 --> 00:38:17,197 join that conversation in the show notes, and I do hope you sign up. All 658 00:38:17,280 --> 00:38:19,397 right. We'll be back right after this. 659 00:38:19,480 --> 00:38:22,957 And we're back here on Big Technology Podcast with OpenAI co-founder and 660 00:38:23,040 --> 00:38:25,397 president Greg Brockman. Greg, let me 661 00:38:25,480 --> 00:38:27,997 just ask you, what happened in December 662 00:38:28,080 --> 00:38:30,077 2025? Because it seems like it was an 663 00:38:30,160 --> 00:38:32,957 inflection point where all this idea of 664 00:38:33,040 --> 00:38:37,837 uh letting the machine code for hours uninterrupted what went from theory to a 665 00:38:37,920 --> 00:38:41,597 moment where everyone said, "I think I can trust this to keep going 666 00:38:41,680 --> 00:38:45,117 for a while." Uh so, what exactly happened? 667 00:38:45,200 --> 00:38:47,437 >> So, new model releases 668 00:38:47,520 --> 00:38:49,197 really went from the AI being able to do 669 00:38:49,280 --> 00:38:53,837 like 20% of your tasks like 80%. 670 00:38:53,920 --> 00:38:57,917 And that was this massive shift because it went from being kind of a yeah, it's 671 00:38:58,000 --> 00:39:01,997 a nice thing to do to you absolutely need to retool your workflow around 672 00:39:02,080 --> 00:39:07,237 these AIs. And for myself, I very much had this moment where 673 00:39:07,320 --> 00:39:10,397 I have a test prompt that I've been using for years of build a website for 674 00:39:10,480 --> 00:39:13,757 me. I'd actually built this website back when I was learning to code, took me 675 00:39:13,840 --> 00:39:17,477 months. Used to be over the course of 25 that 676 00:39:17,560 --> 00:39:21,037 you know, it'd take like 4 hours, bunch of different prompts to get it right. 677 00:39:21,120 --> 00:39:24,917 In December, one shot. Just asked the AI one time and 678 00:39:25,000 --> 00:39:26,997 it produced it and did a great job. 679 00:39:27,080 --> 00:39:29,197 >> So, how did those models make the leap? 680 00:39:29,280 --> 00:39:34,037 >> Well, a lot of it is about the better base models. 681 00:39:34,120 --> 00:39:38,277 >> Mhm. >> The one thing about OpenAI is that we've been working on improving our 682 00:39:38,360 --> 00:39:39,677 pre-training technology for quite some 683 00:39:39,760 --> 00:39:42,957 time and that in that moment we got to 684 00:39:43,040 --> 00:39:45,437 see a little taste of what is going to 685 00:39:45,520 --> 00:39:48,237 be coming for for the rest of this year. 686 00:39:48,320 --> 00:39:52,437 But it's also really about not any one thing. It's about we're constantly 687 00:39:52,520 --> 00:39:55,597 pushing on every single axis of innovation. 688 00:39:55,680 --> 00:40:00,637 And the thing that's very interesting about these models is in some ways you get 689 00:40:00,720 --> 00:40:04,117 these leaps. In some ways it's all continuous. All right, it didn't go from 690 00:40:04,200 --> 00:40:08,797 0% to 80%. It went from 20% to 80%. And so, in some ways it's it just got 691 00:40:08,880 --> 00:40:13,917 better. And I think that we've actually seen this improvement continue with every 692 00:40:14,000 --> 00:40:15,677 single point release that we've had. 693 00:40:15,760 --> 00:40:20,637 Like between 5.2 and 5.3, one of my engineers I work with very closely 694 00:40:20,720 --> 00:40:24,237 went from he couldn't get it to do the like low-level hardcore systems engineer 695 00:40:24,320 --> 00:40:26,637 he does to it absolutely being graded. 696 00:40:26,720 --> 00:40:29,237 He gives it a design doc and actually implements it, adds metrics, 697 00:40:29,320 --> 00:40:34,277 observability, runs the profiler, improves it to the point that it's the 698 00:40:34,360 --> 00:40:37,677 exact thing that he was hoping to produce. And so I think that the way to 699 00:40:37,760 --> 00:40:41,957 think about it is it's almost a sort of slowly slowly slowly all at 700 00:40:42,040 --> 00:40:46,837 once. But it is all indicated by what's kind 701 00:40:46,920 --> 00:40:50,077 of working right now. Certainly within a year, sometimes much 702 00:40:50,160 --> 00:40:52,277 sooner, is going to be incredibly reliable. 703 00:40:52,360 --> 00:40:55,997 >> And it surprised you because I heard you talking on an interview not long ago 704 00:40:56,080 --> 00:40:58,077 about how Codex, right, this autonomous 705 00:40:58,160 --> 00:41:01,517 coder was just for software developers. 706 00:41:01,600 --> 00:41:05,637 And earlier this conversation you said actually everyone can use this stuff. 707 00:41:05,720 --> 00:41:08,157 Yes, what led to the fact that you sort 708 00:41:08,240 --> 00:41:10,597 of changed your perspective on? 709 00:41:10,680 --> 00:41:13,957 >> Well, I think I'd been focusing on Codex 710 00:41:14,040 --> 00:41:18,837 and it's got the code in it, right? So it's really been for coders. And 711 00:41:18,920 --> 00:41:21,957 thinking about people within OpenAI because many of us are software 712 00:41:22,040 --> 00:41:24,637 engineers building for ourselves, 713 00:41:24,720 --> 00:41:29,117 it's very natural to think that way. But as this technology has been progressing, 714 00:41:29,200 --> 00:41:33,677 we've started to realize that the underlying technology we produced is 715 00:41:33,760 --> 00:41:37,437 mostly not about code at all. It's mostly about solving problems. It's 716 00:41:37,520 --> 00:41:41,357 mostly about being able to manage contexts and harnesses and think about 717 00:41:41,440 --> 00:41:43,877 how an AI should integrate and do work. 718 00:41:43,960 --> 00:41:48,797 And that's something that becomes both even for code 719 00:41:48,880 --> 00:41:51,837 suddenly anyone can have access because you can manage something that's going to 720 00:41:51,920 --> 00:41:55,157 go do work, right? If you have a vision, you have something you want to 721 00:41:55,240 --> 00:41:59,437 accomplish, you can describe your intent, the AI can execute, can get that 722 00:41:59,520 --> 00:42:03,397 done. But then it also starts to go, "Well, why am I just focused on coding?" 723 00:42:03,480 --> 00:42:07,957 Like there's so much just very mechanical skill associated with Excel 724 00:42:08,040 --> 00:42:12,597 spreadsheets, with presentations. And if the AI has the context, it has the raw 725 00:42:12,680 --> 00:42:16,517 intelligence now to be able to do these things at a great level. So, if we can 726 00:42:16,600 --> 00:42:20,717 just make it more accessible, suddenly goes from Codex is for coders to Codex 727 00:42:20,800 --> 00:42:24,637 is for everyone. And soon after this moment where we saw all this 728 00:42:24,720 --> 00:42:27,277 improvement, there was another so on 729 00:42:27,360 --> 00:42:31,917 phenomenon in Silicon Valley, which was open claw. Right, which is and maybe 730 00:42:32,000 --> 00:42:34,957 it's the broader tech community where where people started to trust it in ways 731 00:42:35,040 --> 00:42:38,317 that you suggested giving the an AI bot 732 00:42:38,400 --> 00:42:41,837 access to their their desktop or getting 733 00:42:41,920 --> 00:42:46,237 a Mac Mini and giving it access to like their mail and calendar there and their 734 00:42:46,320 --> 00:42:51,157 files and then just kind of letting it go run their life. And then OpenAI 735 00:42:51,240 --> 00:42:53,117 brought the founder of open claw 736 00:42:53,200 --> 00:42:55,637 in-house. So, you talked a little bit 737 00:42:55,720 --> 00:43:01,197 more about the AI as something that will help run your life for you in a way. 738 00:43:01,280 --> 00:43:04,197 Um is that the vision by bringing the open 739 00:43:04,280 --> 00:43:06,277 claw team in-house? 740 00:43:06,360 --> 00:43:09,597 >> Well, I say that the core thing about 741 00:43:09,680 --> 00:43:14,317 this technology is that figuring out how it's useful, how people want to use it. 742 00:43:14,400 --> 00:43:17,637 What is the vision for agents? How is it going to slide into people's lives? That 743 00:43:17,720 --> 00:43:21,237 is a hard problem and that one thing I've seen across many generations of 744 00:43:21,320 --> 00:43:24,757 this technology is the people who really lean in, who have a lot of curiosity, 745 00:43:24,840 --> 00:43:28,037 who have a lot of vision. That's a real skill and that's an emerging very 746 00:43:28,120 --> 00:43:32,997 valuable skill in this new economy that is emerging. And Peter who is the open 747 00:43:33,080 --> 00:43:35,997 claw founder is I think someone who's got incredible vision, incredible 748 00:43:36,080 --> 00:43:39,797 creativity. And so, to some extent it's about the specific technology, but to 749 00:43:39,880 --> 00:43:43,197 some extent it's not at all. It's really about the how do we take these 750 00:43:43,280 --> 00:43:47,637 capabilities and figure out how they slide into people's lives. And so, I 751 00:43:47,720 --> 00:43:51,677 think as a technologist, it's very exciting, but as a someone who is 752 00:43:51,760 --> 00:43:55,877 focused on bringing utility to people, that's something that we are doubling 753 00:43:55,960 --> 00:43:57,917 down on and investing quite a lot. 754 00:43:58,000 --> 00:43:59,597 >> You had a pretty interesting quote about 755 00:43:59,680 --> 00:44:04,837 this recently. Uh talking about getting these autonomous AI agents to work on 756 00:44:04,920 --> 00:44:08,357 your behalf. You said, um you become when you do it, you become 757 00:44:08,440 --> 00:44:12,917 this CEO of a fleet of hundreds of thousands of agents that are completing 758 00:44:13,000 --> 00:44:16,477 your objectives, your goals, your vision, and you're not in the weeds on 759 00:44:16,560 --> 00:44:18,557 exactly how different things are solved. 760 00:44:18,640 --> 00:44:22,077 And in some ways, this new way of work can make you feel like you're losing 761 00:44:22,160 --> 00:44:25,717 your pulse on the problem. Is that good? 762 00:44:25,800 --> 00:44:28,837 >> I think I think that there's a mixed bag. And so, I think that what we need 763 00:44:28,920 --> 00:44:30,957 to do is acknowledge the strengths of 764 00:44:31,040 --> 00:44:33,597 what these tools can deliver 765 00:44:33,680 --> 00:44:39,077 and mitigate the weaknesses. And so, giving people leverage, agency, making 766 00:44:39,160 --> 00:44:42,197 it so that if you have a vision, something you want to accomplish, that 767 00:44:42,280 --> 00:44:45,717 you can have a fleet of agents that will go do it for you. 768 00:44:45,800 --> 00:44:49,557 But, if you think about how the world works, that 769 00:44:49,640 --> 00:44:51,317 at the end of the day, there's an 770 00:44:51,400 --> 00:44:54,877 accountable party. Right? If you're 771 00:44:54,960 --> 00:44:59,757 trying to build a website and your agent messes it up, and your user is affected, 772 00:44:59,840 --> 00:45:03,917 it's not really the agent's fault, it's your fault. And so, you need to care. 773 00:45:04,000 --> 00:45:07,197 And I think that for people to use these tools right, you need to realize that 774 00:45:07,280 --> 00:45:11,117 human agency, human accountability, that's a core part of the system. How 775 00:45:11,200 --> 00:45:15,917 the human uses the AI, that's something that is deeply fundamental. And so, I 776 00:45:16,000 --> 00:45:19,477 think the important thing is that as a user of these agents, and we do this 777 00:45:19,560 --> 00:45:23,117 within Open AI, you cannot abdicate responsibility. You cannot just say, 778 00:45:23,200 --> 00:45:24,757 "Ah, the AI is just going to do stuff." 779 00:45:24,840 --> 00:45:26,077 >> Of course, but you said feel you're like 780 00:45:26,160 --> 00:45:30,357 you're losing your pulse on the problem itself. That's different then 781 00:45:30,440 --> 00:45:33,757 accountability layer to it. >> Well, to me, they actually are linked 782 00:45:33,840 --> 00:45:37,237 together because the point is that if you if you're a CEO and you're too far 783 00:45:37,320 --> 00:45:40,157 from the details, right? If you're running this company, you're running 784 00:45:40,240 --> 00:45:44,957 this this team, and that you've lost your finger on the pulse, that is 785 00:45:45,040 --> 00:45:48,157 something that's not going to lead to great results. And so, the point that I 786 00:45:48,240 --> 00:45:51,877 was trying to make there is that not that it's a desirable thing for humans 787 00:45:51,960 --> 00:45:55,797 to not have to to to know about what's going on. There's some details that 788 00:45:55,880 --> 00:45:59,277 because you can trust like if you are working with a team like a general 789 00:45:59,360 --> 00:46:02,797 contractor to build a house, there's a bunch of details there that 790 00:46:02,880 --> 00:46:05,437 you probably don't need to worry about because you can trust that they'll be 791 00:46:05,520 --> 00:46:08,837 taken care of. But at the end of the day, if there are details that are 792 00:46:08,920 --> 00:46:11,757 wrong, you should care about it. You 793 00:46:11,840 --> 00:46:14,237 should be aware. And so, this is, I 794 00:46:14,320 --> 00:46:16,997 think, an important nuance of you cannot 795 00:46:17,080 --> 00:46:21,717 just blindly say, "I'm okay with who's my finger on the pulse." That we need to 796 00:46:21,800 --> 00:46:25,117 lean in and say, "I need to keep it there to really understand the strengths 797 00:46:25,200 --> 00:46:28,877 and weaknesses." And that as you disengage from some of these details, 798 00:46:28,960 --> 00:46:33,117 these lower-level mechanical things, you should do it because you have 799 00:46:33,200 --> 00:46:37,157 built trust with a system that it will do a good job. 800 00:46:37,240 --> 00:46:40,677 >> One last question about the models. You talked a little bit about the evolutions 801 00:46:40,760 --> 00:46:42,157 that the models have gone through. 802 00:46:42,240 --> 00:46:46,237 Pre-training and fine-tuning, reinforcement learning that gets it uh 803 00:46:46,320 --> 00:46:48,917 more equipped to solve problems 804 00:46:49,000 --> 00:46:53,557 step-by-step and go out in the on the internet and do things. And now we're in 805 00:46:53,640 --> 00:46:57,437 this moment where they're the models have learned through 806 00:46:57,520 --> 00:46:59,197 that process to use tools. And correct 807 00:46:59,280 --> 00:47:02,277 me if I'm wrong on this one. What 808 00:47:02,360 --> 00:47:04,797 What is next in that progression? 809 00:47:04,880 --> 00:47:07,997 >> Well, I think that the 810 00:47:08,080 --> 00:47:13,117 world that we're in is one of this increasing capability and depth of what 811 00:47:13,200 --> 00:47:16,757 the machine can do. And some of this is about 812 00:47:16,840 --> 00:47:20,837 got this tool use, but now we also need to actually build really great tools. 813 00:47:20,920 --> 00:47:24,277 You think about something like computer use in AI that can actually use a desktop, 814 00:47:24,360 --> 00:47:28,437 then it is really able to do anything that you can do. But we also have to 815 00:47:28,520 --> 00:47:32,477 build a little bit for the machine think about how does in the enterprise 816 00:47:32,560 --> 00:47:36,037 credentialing work? How does How do audit trails and observability work? 817 00:47:36,120 --> 00:47:40,157 >> Mhm. >> There's a lot of technology to build to catch up with what the core model 818 00:47:40,240 --> 00:47:44,357 capability is. And I think the overall direction of travel includes things like 819 00:47:44,440 --> 00:47:48,637 a really great speech interface. So, you can just talk to your computer naturally 820 00:47:48,720 --> 00:47:51,837 and just as natural as this conversation, and it understands you. It 821 00:47:51,920 --> 00:47:53,917 does what you need. Has good advice. 822 00:47:54,000 --> 00:47:57,357 It's able to surface that I've been working on this thing. I have a problem. 823 00:47:57,440 --> 00:48:00,717 Here's you wake up in the morning and says, here's your daily report of how 824 00:48:00,800 --> 00:48:03,557 much progress your agents made overnight. Maybe it's running a business 825 00:48:03,640 --> 00:48:06,237 for you, which I think is going to be huge application of this technology. The 826 00:48:06,320 --> 00:48:09,677 democratization of entrepreneurship is absolutely coming. It'll say, here is 827 00:48:09,760 --> 00:48:12,197 these problems. There's this customer that's upset, you know, that they want 828 00:48:12,280 --> 00:48:15,157 to talk to a real human. Like you should go talk to them. Like all of that's 829 00:48:15,240 --> 00:48:18,357 going to happen. And then, I think that 830 00:48:18,440 --> 00:48:20,957 the raising of the ceiling of ambition 831 00:48:21,040 --> 00:48:26,157 of challenges humanity can solve. That is also a next step for 832 00:48:26,240 --> 00:48:29,917 this technology. And we're seeing the leading edges of it. The thing that I am 833 00:48:30,000 --> 00:48:34,117 just very excited to see is almost, if you remember AlphaGo move 37, right? 834 00:48:34,200 --> 00:48:36,117 This move that no human ever would have come up with. 835 00:48:36,200 --> 00:48:38,557 >> It's creative. >> Creative. And it changed humanity's 836 00:48:38,640 --> 00:48:40,557 understanding of the game. That is going 837 00:48:40,640 --> 00:48:43,197 to happen in every single domain. It 838 00:48:43,280 --> 00:48:47,677 will happen in science, in math, in physics, in chemistry. It's going to 839 00:48:47,760 --> 00:48:50,077 happen in material science, going to happen in biology, it's going to happen 840 00:48:50,160 --> 00:48:52,477 in healthcare, drug discovery. But it 841 00:48:52,560 --> 00:48:55,397 may also even happen in literature, in 842 00:48:55,480 --> 00:48:57,837 poetry, in a bunch of other fields that 843 00:48:57,920 --> 00:49:03,437 going to unlock human creative understanding and ideation in ways we 844 00:49:03,520 --> 00:49:05,597 can't imagine right now. >> Why do you think that hasn't happened 845 00:49:05,680 --> 00:49:08,757 yet, given how strong you say the models are? 846 00:49:08,840 --> 00:49:11,557 >> Well, I think that there is an overhang 847 00:49:11,640 --> 00:49:16,237 where of what the models are capable of and how people are using them. So, the 848 00:49:16,320 --> 00:49:20,237 application Well, yeah, it's it's almost our understanding of what is in these 849 00:49:20,320 --> 00:49:22,397 models. >> Okay. >> That's something that I think is still 850 00:49:22,480 --> 00:49:25,557 emerging. So, I think that even with no further progress, there's still a 851 00:49:25,640 --> 00:49:29,477 massive shift that will happen. The economy being powered by computing AI is 852 00:49:29,560 --> 00:49:32,477 still going to happen. But I think there's also something where what we've 853 00:49:32,560 --> 00:49:34,837 gotten very good at is training models 854 00:49:34,920 --> 00:49:40,477 on tasks that could be measured. And so, what we started with was math problems. 855 00:49:40,560 --> 00:49:43,597 Programming problems where you have a perfect verifier. 856 00:49:43,680 --> 00:49:47,437 And a lot of what the progress has been in bringing this to more open-ended 857 00:49:47,520 --> 00:49:51,557 problems has been expanding the space of what can be created. And the itself can 858 00:49:51,640 --> 00:49:54,397 really help with that. If the AI is smart and understands things, you give 859 00:49:54,480 --> 00:49:56,957 it a rubric for how well a task goes. 860 00:49:57,040 --> 00:50:00,357 And of course, for things like creative writing, like is this a good poem? 861 00:50:00,440 --> 00:50:05,077 That's a much harder thing to grade. And so, we've had less ability to teach the 862 00:50:05,160 --> 00:50:09,957 AI and for it to experience and try things out. But all of that is changing and 863 00:50:10,040 --> 00:50:12,157 something that we we have a lot of sight for. 864 00:50:12,240 --> 00:50:13,477 >> Now, it's interesting reflecting on that 865 00:50:13,560 --> 00:50:16,757 Peter Thiel has mentioned, pretty sure this is what he said, that 866 00:50:16,840 --> 00:50:20,997 if you're a math person, you're probably in deeper trouble in terms of these 867 00:50:21,080 --> 00:50:24,797 models coming for what you do than if you're a words person. And you were a 868 00:50:24,880 --> 00:50:29,397 member of math club back in the day. Uh are you not concerned about that? 869 00:50:29,480 --> 00:50:33,797 >> Well, I think that it's much easier to see what we lose than what we gain, 870 00:50:33,880 --> 00:50:36,637 right? Because we have a deep understanding of I used to do things 871 00:50:36,720 --> 00:50:40,077 this way. I used to do this math competition. Now the AI can do the math 872 00:50:40,160 --> 00:50:44,037 competition. But it was never really about the math competition. 873 00:50:44,120 --> 00:50:47,237 Right. Right. That's not really the thing that drives humanity. And if you 874 00:50:47,320 --> 00:50:50,757 think about the way that we do work right now of 875 00:50:50,840 --> 00:50:52,797 there's a box and we type behind a box. 876 00:50:52,880 --> 00:50:54,237 We weren't doing that 100 years ago. 877 00:50:54,320 --> 00:50:57,957 That's not natural. That's not this digital world that we all got kind of 878 00:50:58,040 --> 00:51:02,717 sucked into. That's not really what being human's about. Being a human's 879 00:51:02,800 --> 00:51:05,997 about being here, being present, connecting with other humans. And I 880 00:51:06,080 --> 00:51:07,797 think that what we're going to see is 881 00:51:07,880 --> 00:51:10,797 that AI is going to free up so much time 882 00:51:10,880 --> 00:51:15,357 to increase human connection, to build more bonds across people. 883 00:51:15,440 --> 00:51:18,197 And that's something I'm extremely excited about. 884 00:51:18,280 --> 00:51:20,957 >> Okay. And then as we shift well, as you 885 00:51:21,040 --> 00:51:24,117 shift really to uh these more 886 00:51:24,200 --> 00:51:27,237 agentic use cases, there's been 887 00:51:27,320 --> 00:51:32,397 discussion about whether the bigger training runs really need to happen. And 888 00:51:32,480 --> 00:51:36,237 you know, especially if you like get the model good enough, uh then you could 889 00:51:36,320 --> 00:51:39,157 sort of let it go out in the world and then 890 00:51:39,240 --> 00:51:43,557 you can uh effectively get much of the uplift in 891 00:51:43,640 --> 00:51:47,757 areas that aren't the pre-training, which is what these big data centers are 892 00:51:47,840 --> 00:51:51,957 needed for. So, you you were worked out you work on scaling here. 893 00:51:52,040 --> 00:51:54,237 Uh lead that lead that process. 894 00:51:54,320 --> 00:51:56,157 Uh what do you think about that argument? 895 00:51:56,240 --> 00:51:59,717 >> Well, I think it misses something very important for how the technology 896 00:51:59,800 --> 00:52:04,237 development goes because it is absolutely the case that every single 897 00:52:04,320 --> 00:52:08,677 step of the model production pipeline multiplies. 898 00:52:08,760 --> 00:52:10,797 And so, you want to improve all of them. 899 00:52:10,880 --> 00:52:14,597 And the thing that we see as we improve the pre-training, it makes all the other 900 00:52:14,680 --> 00:52:18,357 steps much easier and it makes sense because it's a model 901 00:52:18,440 --> 00:52:21,997 is able to learn faster. It's a model that is 902 00:52:22,080 --> 00:52:25,637 because it already is like more capable to start when it's trying out different 903 00:52:25,720 --> 00:52:29,797 ideas and learning from its own mistakes, that process just is faster. 904 00:52:29,880 --> 00:52:33,837 It needs to make fewer mistakes. And so, I think that the big shift has been from 905 00:52:33,920 --> 00:52:39,477 thinking of it as just it's you're just training this cerebral system on its own 906 00:52:39,560 --> 00:52:42,517 and you just make it bigger and bigger to it's also about trying things out. 907 00:52:42,600 --> 00:52:45,597 It's also about understanding how people are using it in the real world and 908 00:52:45,680 --> 00:52:49,077 connecting that back into your training, but it doesn't remove the value and the 909 00:52:49,160 --> 00:52:52,717 importance of continuing that that that research. And the thing that I think has 910 00:52:52,800 --> 00:52:56,357 also shifted is we used to really just focus on the raw pre-training 911 00:52:56,440 --> 00:52:59,957 capability, but not think as much about the inference ability and that's been a 912 00:53:00,040 --> 00:53:02,197 big change over the past 24 months to 913 00:53:02,280 --> 00:53:05,917 realize that it's a balance between you 914 00:53:06,000 --> 00:53:10,077 can have this model that has all those great properties in the base, but then 915 00:53:10,160 --> 00:53:12,797 you really need it to be able to be inferencable because you need to 916 00:53:12,880 --> 00:53:16,397 reinforcement learning, you need to serve it to the world. And that that 917 00:53:16,480 --> 00:53:20,077 means that you don't necessarily go as big as you possibly could because you 918 00:53:20,160 --> 00:53:23,357 also really think about there's going to be all this downstream use and you 919 00:53:23,440 --> 00:53:28,757 really want the thing that has the best intelligence times the that that cost 920 00:53:28,840 --> 00:53:31,237 and to to optimize those two things together. 921 00:53:31,320 --> 00:53:35,037 >> Do you still need the Nvidia GPU if things move mostly to inference? 922 00:53:35,120 --> 00:53:36,837 >> We absolutely do. Yes. >> Why? 923 00:53:36,920 --> 00:53:39,117 >> Well, because the 924 00:53:39,200 --> 00:53:41,477 there's multiple reasons, um but one is 925 00:53:41,560 --> 00:53:44,437 that even as the balance of how much of 926 00:53:44,520 --> 00:53:47,517 inference versus training changes, that 927 00:53:47,600 --> 00:53:52,157 you cannot get massive scale training 928 00:53:52,240 --> 00:53:56,157 through any other way besides this concentration 929 00:53:56,240 --> 00:54:00,757 of compute on one problem. And so, I think that the the thing that I think 930 00:54:00,840 --> 00:54:04,637 will happen is there's some amount of the the the deployment footprint goes up 931 00:54:04,720 --> 00:54:08,677 quite a lot, but that sometimes there will be you have a particular massive 932 00:54:08,760 --> 00:54:12,437 pre-training run, and you really want to concentrate a bunch in there. Um, I also 933 00:54:12,520 --> 00:54:15,957 think that the NVIDIA team is just incredible and does really, really 934 00:54:16,040 --> 00:54:18,477 amazing work. And so, yeah, we we partner very closely with them. 935 00:54:18,560 --> 00:54:21,117 >> Isn't there going to be a time where people just say, "We've pre-trained 936 00:54:21,200 --> 00:54:23,197 enough. The models are smart enough?" 937 00:54:23,280 --> 00:54:27,637 >> I think that that's a little bit like once humanity has solved all problems in 938 00:54:27,720 --> 00:54:30,437 front of us, then maybe we can we can say that. 939 00:54:30,520 --> 00:54:34,677 Right. But I think that the ceiling of what we want to accomplish, I think that 940 00:54:34,760 --> 00:54:39,237 that there's just so much ambition that that maybe we've over the past 50 years 941 00:54:39,320 --> 00:54:42,557 or so just sort of backed off from, right? You think about I mean, even 942 00:54:42,640 --> 00:54:46,717 problems that seem very clear like can we have health care for everyone that is 943 00:54:46,800 --> 00:54:50,037 not just that's that's actually preventative, not just 944 00:54:50,120 --> 00:54:53,517 targeting when people have a problem, but really think about the lifestyle and 945 00:54:53,600 --> 00:54:57,717 how to really help people early detect potential diseases before they 946 00:54:57,800 --> 00:55:02,117 happen. Like that's a problem that I think we can actually achieve through 947 00:55:02,200 --> 00:55:05,797 more intelligent models. And there's probably some level where you can 948 00:55:05,880 --> 00:55:08,997 totally solve that problem, and then you say, "Well, do I need a model that's two 949 00:55:09,080 --> 00:55:10,997 times smarter?" But there's other 950 00:55:11,080 --> 00:55:14,717 problems that are going to demand that. 951 00:55:14,800 --> 00:55:16,317 >> Let's talk about the math about building 952 00:55:16,400 --> 00:55:20,077 these data centers. You raised 110 billion 953 00:55:20,160 --> 00:55:23,797 uh, earlier this year. Is what's the math behind that? Does 954 00:55:23,880 --> 00:55:25,677 that money go right into data centers? 955 00:55:25,760 --> 00:55:29,557 How do you think about how you're going to return that money to investors? Talk 956 00:55:29,640 --> 00:55:33,317 about those calculations. >> Yeah, so I think it's as simple as the 957 00:55:33,400 --> 00:55:35,397 massive expense we see in front of us is 958 00:55:35,480 --> 00:55:38,437 compute. But you can think of compute 959 00:55:38,520 --> 00:55:42,437 not as a cost center, but as a revenue center. Think of it a little bit like 960 00:55:42,520 --> 00:55:45,277 hiring sales people. Right? How many sales people do you want 961 00:55:45,360 --> 00:55:48,077 to hire? As long as you can sell your product, as long as you have a scalable 962 00:55:48,160 --> 00:55:52,037 way to sell that product, then the more sales people you have, the more revenue 963 00:55:52,120 --> 00:55:55,597 you will make. And I think the world that we're in is we have continually 964 00:55:55,680 --> 00:55:59,917 found we cannot build compute fast enough to keep up with demand. And I see 965 00:56:00,000 --> 00:56:03,397 this very concretely. Right? Right now we have to make very painful decisions 966 00:56:03,480 --> 00:56:08,157 about what we're launching, about where the compute goes, and that I think we're 967 00:56:08,240 --> 00:56:11,357 going to experience this more broadly within the economy as we shift to this 968 00:56:11,440 --> 00:56:15,117 AI-powered economy. The question will be what problems are going to get that 969 00:56:15,200 --> 00:56:19,517 massive compute. How do you scale so everyone can have personal agent running 970 00:56:19,600 --> 00:56:22,837 for them? How can everyone be using systems like Codex? Like there just 971 00:56:22,920 --> 00:56:27,077 isn't enough compute in the world to be able to do that. 972 00:56:27,160 --> 00:56:30,237 And so, we're trying to get ahead of that problem. But, it is a new category. 973 00:56:30,320 --> 00:56:34,477 >> Right? So, I you're doing it with real confidence. I mean, in sums of money the 974 00:56:34,560 --> 00:56:37,917 world has never seen put towards a project like this. 975 00:56:38,000 --> 00:56:41,397 When you're building a new category, how do you do it with certainty that it's 976 00:56:41,480 --> 00:56:44,477 going to work out? >> Well, I think there's several components 977 00:56:44,560 --> 00:56:48,877 that go into it. So, the first is there is historical precedent at this point. 978 00:56:48,960 --> 00:56:52,117 From the moment we launched ChatGPT, I remember talking with my team having 979 00:56:52,200 --> 00:56:55,397 this exact conversation. They said, "How much compute should we buy?" I said, 980 00:56:55,480 --> 00:56:58,277 "All of it." They said, "No, no, no, really. How much compute should we buy?" 981 00:56:58,360 --> 00:57:01,117 I said, "No matter how much we try to 982 00:57:01,200 --> 00:57:05,437 build, I know we're not going to be able to keep up with the demand." And that 983 00:57:05,520 --> 00:57:09,237 has been true, and that has been true every year since then. And the the 984 00:57:09,320 --> 00:57:13,077 challenge is that these compute purchases, you have to lock them in 18 985 00:57:13,160 --> 00:57:16,757 months, sometimes 24 months, sometimes longer 986 00:57:16,840 --> 00:57:20,477 in advance of the machines being delivered, which means you really need 987 00:57:20,560 --> 00:57:24,597 to project forward. And I think that the world that we're moving towards is one 988 00:57:24,680 --> 00:57:29,197 where to date most of our revenue has come from consumer subscriptions, and that 989 00:57:29,280 --> 00:57:31,677 will always be very important. There's other revenue streams we have emerging 990 00:57:31,760 --> 00:57:35,277 as well, but the the opportunity that clearly is emerging now is knowledge 991 00:57:35,360 --> 00:57:39,077 work. And we're seeing this very concretely across every single 992 00:57:39,160 --> 00:57:43,397 enterprise realizing this technology it actually really works and to be 993 00:57:43,480 --> 00:57:47,517 competitive they need to adopt it. And you can see this organic energy of all 994 00:57:47,600 --> 00:57:51,597 these software engineers using it and then we're starting to see the 995 00:57:51,680 --> 00:57:54,237 percolation of people using it for 996 00:57:54,320 --> 00:57:59,237 various knowledge work inside of the enterprises and the willingness to pay 997 00:57:59,320 --> 00:58:03,357 and the revenue growth that you're seeing in this industry is very clear. 998 00:58:03,440 --> 00:58:07,757 Right? It's very clearly happening right now and just project that forward. And 999 00:58:07,840 --> 00:58:11,237 we look like one thing we get to see that maybe the world doesn't is the line 1000 00:58:11,320 --> 00:58:15,757 of sight to how these models will improve. And all of this together says 1001 00:58:15,840 --> 00:58:18,197 that the economy, which is a massive thing, 1002 00:58:18,280 --> 00:58:22,077 right? The economy is just so large it's it's almost incomprehensible. 1003 00:58:22,160 --> 00:58:26,437 All of the growth, like the the the highest order bit on how this economy 1004 00:58:26,520 --> 00:58:30,437 grows from here will be about AI, how well you can leverage AI and the 1005 00:58:30,520 --> 00:58:32,917 computational power you have available to power it. 1006 00:58:33,000 --> 00:58:34,397 >> You said consumer subscriptions are your 1007 00:58:34,480 --> 00:58:38,157 biggest source of revenue right now. Is the projection that will flip and that 1008 00:58:38,240 --> 00:58:40,197 business will be the biggest source? 1009 00:58:40,280 --> 00:58:41,877 >> I think well, I think that that it is 1010 00:58:41,960 --> 00:58:47,557 very like very clear how quickly the uh the the enterprise it's not just 1011 00:58:47,640 --> 00:58:50,677 enterprise cuz I think enterprise is also changing what it means. So really 1012 00:58:50,760 --> 00:58:53,997 people using it for productive knowledge work for those kinds of things. And I 1013 00:58:54,080 --> 00:58:57,597 think that as we think about pricing, one thing if you look at how Codex works 1014 00:58:57,680 --> 00:59:01,957 right now is if you have a ChatGPT consumer subscription, you can use 1015 00:59:02,040 --> 00:59:05,717 Codex. And so I think it's not going to be as well defined as there's this 1016 00:59:05,800 --> 00:59:07,357 category, that category. I think it will 1017 00:59:07,440 --> 00:59:10,757 really be about you as a user are going 1018 00:59:10,840 --> 00:59:12,957 to have just again like your laptop this 1019 00:59:13,040 --> 00:59:17,957 portal to the digital world and that is what the the revenue fundamentally will 1020 00:59:18,040 --> 00:59:21,917 will will come from. >> Dario said, "I think about you." 1021 00:59:22,000 --> 00:59:25,917 There are some players who are yoloing who pull the wrist dial too far and I'm 1022 00:59:26,000 --> 00:59:30,637 very concerned. I think he's referencing your infrastructure bets there. 1023 00:59:30,720 --> 00:59:33,557 What do you think about that? >> Well, I just disagree. 1024 00:59:33,640 --> 00:59:38,317 I think we've been very thoughtful and very much seeing 1025 00:59:38,400 --> 00:59:40,477 what is coming and I think that we will 1026 00:59:40,560 --> 00:59:43,997 see even this year how everyone who is 1027 00:59:44,080 --> 00:59:48,437 participating is going to be compute strapped. And I think we have been the 1028 00:59:48,520 --> 00:59:50,077 most forward in realizing that this is 1029 00:59:50,160 --> 00:59:54,117 coming and building in anticipation of 1030 00:59:54,200 --> 00:59:58,437 how this technology is playing out. And I think that what we have seen is that 1031 00:59:58,520 --> 01:00:02,637 for other players that they kind of realized that probably late last year 1032 01:00:02,720 --> 01:00:05,677 and started scrambling to see what compute is available and there really 1033 01:00:05,760 --> 01:00:09,677 wasn't any. And so I think that even as people it's very easy to to make 1034 01:00:09,760 --> 01:00:13,317 statements like that, but I think that everyone has kind of realized that 1035 01:00:13,400 --> 01:00:16,837 this technology it's working, it's here, it's real, right? Software engineering 1036 01:00:16,920 --> 01:00:21,317 is just the first example of it. And that we are fundamentally limited by the 1037 01:00:21,400 --> 01:00:23,037 computational power available. 1038 01:00:23,120 --> 01:00:27,437 >> And you said that also that if he's off by with his prediction by a little bit 1039 01:00:27,520 --> 01:00:31,117 then uh his the company could potentially go bankrupt. Is that the 1040 01:00:31,200 --> 01:00:35,117 same case for you? >> I think that Look, I I think that that 1041 01:00:35,200 --> 01:00:37,277 there's actually more degrees of off-ramp here. 1042 01:00:37,360 --> 01:00:38,957 >> Okay. >> If you start to worry about the downside 1043 01:00:39,040 --> 01:00:42,237 case, which I think is a very reasonable question, right? But to some extent what 1044 01:00:42,320 --> 01:00:46,597 I think the bet is on isn't about anyone company. It's really about the sector. 1045 01:00:46,680 --> 01:00:49,877 It's really about do you believe this technology can be produced and can 1046 01:00:49,960 --> 01:00:53,677 deliver this massive amount of value that we see coming. 1047 01:00:53,760 --> 01:00:57,397 And again, I'll point to proof points, right? That software engineering it's 1048 01:00:57,480 --> 01:00:59,757 just like the degree to which if you're not a software engineer, you haven't 1049 01:00:59,840 --> 01:01:03,837 tried code apps the degree to which it's different, like it's just hard to 1050 01:01:03,920 --> 01:01:07,957 describe. And I think that people will experience it very quickly. Like, you 1051 01:01:08,040 --> 01:01:12,837 know, 6 months ago I think that for us we we saw this internally, but there 1052 01:01:12,920 --> 01:01:16,077 were less proof points out there. Now there's proof points out there. 6 months 1053 01:01:16,160 --> 01:01:19,517 from now, I think that everyone will feel it and I think that we will all 1054 01:01:19,600 --> 01:01:25,037 feel the pain of there's an awesome model and there's just no availability because 1055 01:01:25,120 --> 01:01:28,477 there's not enough compute. >> Yeah, but as we were looking at our 1056 01:01:28,560 --> 01:01:33,437 predictions for 2026 on this show, we had a conversation uh towards the end of 1057 01:01:33,520 --> 01:01:38,077 the year last year where Runa John Roy was on with us was like 2026 is going to 1058 01:01:38,160 --> 01:01:42,077 be the year where everybody uses agents and I said, "Yeah, well, I'll believe 1059 01:01:42,160 --> 01:01:45,917 that when I see it." And I'm using the agents. So. 1060 01:01:46,000 --> 01:01:47,797 >> Here we are. Here we go. 1061 01:01:47,880 --> 01:01:49,757 Um what what what do you use it for? 1062 01:01:49,840 --> 01:01:53,237 >> I I use it to build um build tools 1063 01:01:53,320 --> 01:01:55,157 internally for my 1064 01:01:55,240 --> 01:01:58,037 uh for the people who I work with to 1065 01:01:58,120 --> 01:02:01,797 sort of get on the same page about um 1066 01:02:01,880 --> 01:02:05,757 about when videos are coming and what the thumbnails need to look like and I'm 1067 01:02:05,840 --> 01:02:10,917 also integrating things from uh from YouTube and so we can basically then 1068 01:02:11,000 --> 01:02:15,677 rank how the videos are doing based off of thumbnail and like a custom-built piece 1069 01:02:15,760 --> 01:02:18,837 of software that I never would have paid for. And that's one of the things that I 1070 01:02:18,920 --> 01:02:20,237 think is interesting about this moment, 1071 01:02:20,320 --> 01:02:22,837 I guess, is that software it's it 1072 01:02:22,920 --> 01:02:28,357 scales, used by the masses, but um when you use it, therefore, there's going to 1073 01:02:28,440 --> 01:02:30,677 be so many things that are not made for you. 1074 01:02:30,760 --> 01:02:35,277 And maybe what this does is it allows us to interact with software in a way 1075 01:02:35,360 --> 01:02:39,037 that's much more natural. >> I think that is the key. And again, I 1076 01:02:39,120 --> 01:02:41,797 just think a lot about the fact that 1077 01:02:41,880 --> 01:02:46,997 the way we've built computers has really pulled us in into this digital world. 1078 01:02:47,080 --> 01:02:50,037 You think about how much time you just spend scrolling through your phone. 1079 01:02:50,120 --> 01:02:52,157 >> Yep. >> Right? The amount of time that you spend 1080 01:02:52,240 --> 01:02:55,437 clicking different buttons and trying to like connect this thing to that thing. 1081 01:02:55,520 --> 01:02:59,797 Like, why? Like, why do you have to do that? Instead the AI being about 1082 01:02:59,880 --> 01:03:02,957 bringing the machine closer to you, personalizing to you, understanding what 1083 01:03:03,040 --> 01:03:07,157 you're trying to accomplish. And that we have all of this 1084 01:03:07,240 --> 01:03:10,437 pop culture of just computers you can talk to and that they go and do stuff 1085 01:03:10,520 --> 01:03:13,557 for you and it's starting to become real. It's starting to become the thing 1086 01:03:13,640 --> 01:03:19,037 that you can actually do. And I think that the amazingness of that is something where 1087 01:03:19,120 --> 01:03:22,557 you just have to try it to really understand. So, I I definitely think 1088 01:03:22,640 --> 01:03:25,437 it's a it's a very special moment we're in. Yeah. 1089 01:03:25,520 --> 01:03:27,517 >> Then I want to know why is AI so 1090 01:03:27,600 --> 01:03:30,837 unpopular uh with the public? Um YouGov, 1091 01:03:30,920 --> 01:03:35,037 for instance, says uh three times as many Americans expect the effects of AI 1092 01:03:35,120 --> 01:03:39,677 on society to be negative as they expect it to be positive. 1093 01:03:39,760 --> 01:03:43,317 I mean, what do you think the reasoning is behind that? And are you concerned 1094 01:03:43,400 --> 01:03:45,517 about AI's brand? 1095 01:03:45,600 --> 01:03:47,597 >> Well, I think that there is something 1096 01:03:47,680 --> 01:03:50,957 that we need to show the country of why 1097 01:03:51,040 --> 01:03:55,357 AI is good for them. Not just for the broad economy, for 1098 01:03:55,440 --> 01:03:58,877 growing the GDP and things like that, but how does it help them in their 1099 01:03:58,960 --> 01:04:01,237 lives? And I think there are actually 1100 01:04:01,320 --> 01:04:03,557 many very concrete stories that I hear 1101 01:04:03,640 --> 01:04:06,597 every day. For example, there's a family 1102 01:04:06,680 --> 01:04:10,517 where their child was having some some 1103 01:04:10,600 --> 01:04:14,797 headaches, some some medical issues, was denied an MRI, and they researched the 1104 01:04:14,880 --> 01:04:18,997 symptoms with ChatGPT and realized that they could make an argument to insurance 1105 01:04:19,080 --> 01:04:21,917 to actually get the MRI. They did that, turns out he had a brain tumor, they 1106 01:04:22,000 --> 01:04:25,757 were able to save his life because they used ChatGPT to get access to the right 1107 01:04:25,840 --> 01:04:27,957 information. That's just one story. 1108 01:04:28,040 --> 01:04:32,397 There are so many more just like that of people who have been deeply, profoundly, 1109 01:04:32,480 --> 01:04:36,077 their lives have been improved or saved through their use of this technology, 1110 01:04:36,160 --> 01:04:38,677 through partnering with the technology in a real way. 1111 01:04:38,760 --> 01:04:42,637 And so, that is a story I don't think gets out there. I think that this is 1112 01:04:42,720 --> 01:04:47,237 happening in so many people's lives, but somehow the story is not yet oh is not 1113 01:04:47,320 --> 01:04:50,757 yet told. And one thing I notice is that there's you know, so much a lot of pop 1114 01:04:50,840 --> 01:04:53,997 culture from, you know, the '90s, from from the 1115 01:04:54,080 --> 01:04:57,597 historical context that we have, that's very negative on AI, that worries about 1116 01:04:57,680 --> 01:05:01,957 what could go wrong. But when people actually use AI, that they find utility 1117 01:05:02,040 --> 01:05:05,197 in it, they find value in it. And so, I think that I am definitely very 1118 01:05:05,280 --> 01:05:09,597 concerned about us not having successfully helped people 1119 01:05:09,680 --> 01:05:13,637 understand why this technology wave is something that will improve their lives. 1120 01:05:13,720 --> 01:05:16,797 It will help improve human connection, and that is something that's a big focus 1121 01:05:16,880 --> 01:05:19,877 in my mind. And if you think about the 1122 01:05:19,960 --> 01:05:21,637 opportunity here and why AI is so 1123 01:05:21,720 --> 01:05:27,237 important, I think this will be the source of economic and national security 1124 01:05:27,320 --> 01:05:30,957 going forward. I think it's going to be about national competitiveness, and that 1125 01:05:31,040 --> 01:05:36,077 there are other countries like China where AI pulls in the exact opposite direction. 1126 01:05:36,160 --> 01:05:38,717 >> And so I ate there. >> Yes, I think it's very, very important 1127 01:05:38,800 --> 01:05:42,317 that we acknowledge that and we really understand how to get the benefits for 1128 01:05:42,400 --> 01:05:45,077 everyone. >> But we also are in a time that's like 1129 01:05:45,160 --> 01:05:50,277 politically unstable. There's concerns about uh about work. People are Every 1130 01:05:50,360 --> 01:05:52,837 time I speak with someone about AI, they're like, "How long do I have left 1131 01:05:52,920 --> 01:05:55,837 to work in my job?" And then when I 1132 01:05:55,920 --> 01:06:00,117 think about the data centers, I mean, the polling is even worse than 1133 01:06:00,200 --> 01:06:04,517 um you know, AI in general. This is from Pew. Far more data Far more people say 1134 01:06:04,600 --> 01:06:08,677 data centers are mostly bad than good for the environment, uh home energy 1135 01:06:08,760 --> 01:06:13,757 costs, and quality of life of those nearby. So, we are at this moment where 1136 01:06:13,840 --> 01:06:16,757 good jobs are tough to come by, and people see these data centers come into 1137 01:06:16,840 --> 01:06:20,597 their communities, and they say not good for the environment, home energy, and 1138 01:06:20,680 --> 01:06:23,757 quality of life. I mean, are they Are they wrong? 1139 01:06:23,840 --> 01:06:26,477 >> Well, I think there's definitely a lot of misinformation about data centers. 1140 01:06:26,560 --> 01:06:28,917 Good example is water usage. 1141 01:06:29,000 --> 01:06:34,037 That if you actually look at our Abilene facility, which is 1142 01:06:34,120 --> 01:06:37,477 one of the if not the biggest supercomputers in the world, 1143 01:06:37,560 --> 01:06:40,197 the amount of water it uses 1144 01:06:40,280 --> 01:06:43,757 is the same as a household 1145 01:06:43,840 --> 01:06:45,477 over the course of a year. Right? So, 1146 01:06:45,560 --> 01:06:49,117 it's really negligible water use. And 1147 01:06:49,200 --> 01:06:52,717 yet that there's a lot of misinformation that these data centers 1148 01:06:52,800 --> 01:06:55,197 consume a lot. And so, similar on on 1149 01:06:55,280 --> 01:06:59,917 power, we have a commitment that we are going to pay our own way to not drive up 1150 01:07:00,000 --> 01:07:03,557 energy prices for people. That's something that as an industry now that 1151 01:07:03,640 --> 01:07:06,357 people are making these commitments because it is very important that we 1152 01:07:06,440 --> 01:07:10,437 improve local communities, and when we build data centers, we really try to go 1153 01:07:10,520 --> 01:07:14,197 into those local communities, understand what's happening on the ground, how we 1154 01:07:14,280 --> 01:07:18,677 can help. There's tax revenue that are associated with these data centers, and 1155 01:07:18,760 --> 01:07:22,317 I think that there's jobs that they create. There's a lot of benefits that 1156 01:07:22,400 --> 01:07:23,957 come from them. 1157 01:07:24,040 --> 01:07:29,397 And so I think that's one thing where it is about how we show up, and that's a 1158 01:07:29,480 --> 01:07:32,637 responsibility that we take very seriously. 1159 01:07:32,720 --> 01:07:35,997 >> Okay, but also like if their power costs are not going to go up, you have to 1160 01:07:36,080 --> 01:07:39,957 bring in power, which means potentially more pollution. 1161 01:07:40,040 --> 01:07:42,077 Is that not a concern? >> Well, I think that I think that there's 1162 01:07:42,160 --> 01:07:46,037 much more nuance in terms of not driving up energy costs. If you look at how the 1163 01:07:46,120 --> 01:07:50,237 grid works today, that there's actually a a lot of just stranded power, power 1164 01:07:50,320 --> 01:07:53,397 that is there that is not being utilized, and that you need to upgrade 1165 01:07:53,480 --> 01:07:56,797 the transmission systems, and again, that's something where putting that on 1166 01:07:56,880 --> 01:07:59,717 us rather than putting it on the rate payers is very important. Right. There's 1167 01:07:59,800 --> 01:08:03,397 lots of places where they have clean power that you can that is actually 1168 01:08:03,480 --> 01:08:06,917 being underutilized and just being kind of thrown away. And so, there's 1169 01:08:07,000 --> 01:08:10,357 a lot of benefit that comes from having 1170 01:08:10,440 --> 01:08:14,797 real reasons for that grid, which is aging and obsolete in many places, to 1171 01:08:14,880 --> 01:08:18,517 upgrade. And that's something actually has real benefits to the community. Like 1172 01:08:18,600 --> 01:08:20,877 we've seen, for example, in North Dakota 1173 01:08:20,960 --> 01:08:22,917 that people's rates have gone down 1174 01:08:23,000 --> 01:08:27,917 because a data center has shown up and has helped with improving the utilities 1175 01:08:28,000 --> 01:08:30,037 for everyone. >> All right, one last question on the 1176 01:08:30,120 --> 01:08:35,477 politics. Uh you gave $25 million to Magna Inc., which is uh PAC that uh 1177 01:08:35,560 --> 01:08:39,077 pro-Trump um You spoke with Wired about it, and you 1178 01:08:39,160 --> 01:08:42,837 said, "Anything I can do to support this technology benefiting everyone is a 1179 01:08:42,920 --> 01:08:46,597 thing I will do." And, you know, if that makes you a 1180 01:08:46,680 --> 01:08:49,637 one-issue voter or one-issue uh 1181 01:08:49,720 --> 01:08:54,237 you know, political supporter, I'll I'll share the one thing I always wonder when 1182 01:08:54,320 --> 01:08:58,237 it comes to just this the one-issue camp is 1183 01:08:58,320 --> 01:09:03,237 um ultimately, doesn't a stronger country make you, 1184 01:09:03,320 --> 01:09:07,717 uh you know, sort of make your goals much more feasible. Even if a candidate 1185 01:09:07,800 --> 01:09:11,397 isn't, you know, fully in support of what you're doing. Like, shouldn't a 1186 01:09:11,480 --> 01:09:13,157 stronger country, no matter what, be the 1187 01:09:13,240 --> 01:09:15,997 North Star of of any political activity? 1188 01:09:16,080 --> 01:09:18,797 Um and if it's the if that's the case, 1189 01:09:18,880 --> 01:09:23,997 you know, then is that is that part of the donation? 1190 01:09:24,080 --> 01:09:26,837 >> So, the way I look at this, so my wife and I 1191 01:09:26,920 --> 01:09:31,157 made that donation. We've donated to bipartisan super PACs as well. I think 1192 01:09:31,240 --> 01:09:35,877 this technology is one where it's coming quickly. Within 1193 01:09:35,960 --> 01:09:38,637 the next couple years, it's really going to transform everything. 1194 01:09:38,720 --> 01:09:42,677 Going to be the underpinning of the economy, and it's not popular. And we 1195 01:09:42,760 --> 01:09:45,517 really want to support politicians who 1196 01:09:45,600 --> 01:09:49,757 really lean into this technology, really engage with it. And so, I think that 1197 01:09:49,840 --> 01:09:53,157 certainly this technology is about up- uplifting 1198 01:09:53,240 --> 01:09:56,037 us as a country. I, you know, I am a 1199 01:09:56,120 --> 01:10:00,717 one-issue donor. This is something where I feel like I've unique contribution to 1200 01:10:00,800 --> 01:10:04,877 to make. But, it's really about just expressing support for this technology 1201 01:10:04,960 --> 01:10:07,717 is something that we should be leaning into as a country. 1202 01:10:07,800 --> 01:10:10,237 >> What would you tell someone who's who's scared of AI? I mean, if you have a 1203 01:10:10,320 --> 01:10:13,077 moment here where you can speak directly to them. They might think it's going to 1204 01:10:13,160 --> 01:10:16,677 take my job. It's going to pollute my community. It will change the world too 1205 01:10:16,760 --> 01:10:19,037 fast. What's your message to that? 1206 01:10:19,120 --> 01:10:21,757 >> Number one thing is try the tools. 1207 01:10:21,840 --> 01:10:28,440 Because, to really understand what it can do for you is something that 1208 01:10:28,560 --> 01:10:31,917 only by experiencing the AI as it exists 1209 01:10:32,000 --> 01:10:35,797 now will that really hit home. 1210 01:10:35,880 --> 01:10:39,237 And we see so much opportunity and potential and empowerment coming from 1211 01:10:39,320 --> 01:10:43,277 this technology today. You talked a little bit about what you can build now. 1212 01:10:43,360 --> 01:10:44,957 >> Right. >> People who have never built a website 1213 01:10:45,040 --> 01:10:48,637 before can build a website. If you want to start a small business, and you're 1214 01:10:48,720 --> 01:10:52,077 thinking about all the back-end processing and how to actually manage 1215 01:10:52,160 --> 01:10:56,357 it, all those things, the AI can help you with that right now. And so, I think 1216 01:10:56,440 --> 01:11:00,597 that in your life, thinking about how it can help you with your health, how it 1217 01:11:00,680 --> 01:11:04,077 can help your loved ones, how it can help you make money, how it can help you 1218 01:11:04,160 --> 01:11:05,437 save money, 1219 01:11:05,520 --> 01:11:08,557 these are all going to be on the table. 1220 01:11:08,640 --> 01:11:11,917 And I think it is much easier to see 1221 01:11:12,000 --> 01:11:17,117 what's going to change than it is to see what you're going to gain. 1222 01:11:17,200 --> 01:11:20,757 But I think that it's worth giving it a fair shot of really trying to understand 1223 01:11:20,840 --> 01:11:23,757 both sides of the equation. That's the one thing that doesn't get talked about 1224 01:11:23,840 --> 01:11:28,237 in the polling data, by the way. It's the people that have seen it used but 1225 01:11:28,320 --> 01:11:31,757 haven't tried it themselves or the people that have never tried AI are much 1226 01:11:31,840 --> 01:11:35,877 more negative. And then you get to the power users and are even people that use 1227 01:11:35,960 --> 01:11:38,677 it casually and they're generally pretty positive 1228 01:11:38,760 --> 01:11:41,877 about the technology. For for myself, we've been thinking about this 1229 01:11:41,960 --> 01:11:45,757 technology for a long time. What I see playing out in front of us is more 1230 01:11:45,840 --> 01:11:50,397 amazing, more beneficial, and going to really have a much more positive impact 1231 01:11:50,480 --> 01:11:54,277 than we ever imagined. So last one for you, how would you invite advise someone 1232 01:11:54,360 --> 01:11:55,837 to prepare themselves for the future? 1233 01:11:55,920 --> 01:11:59,677 And and it has to be more than just getting the tools. I mean, I have friends 1234 01:11:59,760 --> 01:12:03,477 who are who come to me and they say, I I don't know what's going to happen with 1235 01:12:03,560 --> 01:12:07,957 my job or the world and I I just need to know what to do with this. 1236 01:12:08,040 --> 01:12:12,477 I do think that the number one thing is about understanding the technology. One 1237 01:12:12,560 --> 01:12:16,557 thing we've seen is the people who get the most out of the technology and you 1238 01:12:16,640 --> 01:12:19,517 have to approach it with a curiosity trying to really try it in your 1239 01:12:19,600 --> 01:12:23,197 workflows, really be able to get over that initial hump of you have a blank 1240 01:12:23,280 --> 01:12:25,037 box, what do I do with a blank box? 1241 01:12:25,120 --> 01:12:27,637 Right? To really develop this sense of 1242 01:12:27,720 --> 01:12:32,197 agency, this sense of I can be the manager. I can set the 1243 01:12:32,280 --> 01:12:34,477 direction. I can 1244 01:12:34,560 --> 01:12:39,477 delegate, I can provide oversight and to really develop that skill because that 1245 01:12:39,560 --> 01:12:41,997 is something that's going to be fundamental. Is we're building this 1246 01:12:42,080 --> 01:12:46,397 technology for humans, to help humans foster more human connection, for humans 1247 01:12:46,480 --> 01:12:49,797 to be able to spend more time doing what they want. And so the question is, well, 1248 01:12:49,880 --> 01:12:53,797 what do you want? And really trying to crystallize that and trying to realize 1249 01:12:53,880 --> 01:12:57,717 that with help of this technology is going to be the most important thing. 1250 01:12:57,800 --> 01:12:59,517 Greg, thanks so much for coming on the show. 1251 01:12:59,600 --> 01:13:01,797 >> Thank you for having me. >> All right. Thank you everybody for 1252 01:13:01,880 --> 01:13:07,680 listening and watching and we'll see you next time on Big Technology Podcast.