1 00:00:00,720 --> 00:00:04,317 Open AAI founder and CEO Sam Alman 2 00:00:04,400 --> 00:00:06,637 sounded very enthusiastic during the 3 00:00:06,720 --> 00:00:09,117 Cisco AI summit about the launch [music] 4 00:00:09,200 --> 00:00:12,317 of Codex, the software development agent 5 00:00:12,400 --> 00:00:15,037 that helps developers write, understand, 6 00:00:15,120 --> 00:00:17,677 debug, and refactor code by 7 00:00:17,760 --> 00:00:20,397 understanding natural language prompts. 8 00:00:20,480 --> 00:00:22,957 Altman feels like the launch of Codex 9 00:00:23,040 --> 00:00:25,757 marks another CH GBT moment that can 10 00:00:25,840 --> 00:00:30,797 lead to full AI companies. In these highlights, he explains why 11 00:00:30,880 --> 00:00:33,597 >> this is going to change how all of OpenAI works and change how other 12 00:00:33,680 --> 00:00:37,597 companies work. Um, it yeah, the model's 13 00:00:37,680 --> 00:00:40,877 really hit some threshold and I think now the interface and harness has caught 14 00:00:40,960 --> 00:00:44,157 up and 15 00:00:44,240 --> 00:00:47,917 I think this is the first this you're talking about chatbt moments. I think 16 00:00:48,000 --> 00:00:51,997 this is like the first time I felt another chat GPT moment 17 00:00:52,080 --> 00:00:55,917 >> for here is a a clear clear glimpse of the 18 00:00:56,000 --> 00:01:00,477 future of knowledge work and how enterprises and individual people are 19 00:01:00,560 --> 00:01:02,237 going to use AI to work in a completely 20 00:01:02,320 --> 00:01:05,357 different way. The idea that a coding 21 00:01:05,440 --> 00:01:09,917 model can create a full complex piece of software but also interact with the rest 22 00:01:10,000 --> 00:01:14,077 of the real world to build a company around it is a very big deal. And then 23 00:01:14,160 --> 00:01:17,117 this this notion of what's happening right now, I'm sure you're tracking this 24 00:01:17,200 --> 00:01:21,837 with Moldbook and um what happened with Claudebot. Is it just a passing fad or 25 00:01:21,920 --> 00:01:25,277 do you think that there's something that we should take away from that on um 26 00:01:25,360 --> 00:01:27,197 >> No, I think it is definitely not a 27 00:01:27,280 --> 00:01:30,157 passing well mold book maybe. I don't 28 00:01:30,240 --> 00:01:34,077 know. But uh Open Claw is not. Um, 29 00:01:34,160 --> 00:01:36,637 I 30 00:01:36,720 --> 00:01:40,637 I think this idea that code is really 31 00:01:40,720 --> 00:01:43,277 powerful, but code plus generalized 32 00:01:43,360 --> 00:01:46,397 computer use is even much more powerful 33 00:01:46,480 --> 00:01:50,477 is here to stay. The fact I 34 00:01:50,560 --> 00:01:53,597 when I initially installed Codex, I said I was never going to give Codex full 35 00:01:53,680 --> 00:01:57,517 control of my computer with without checking what it was doing. And that 36 00:01:57,600 --> 00:02:01,597 lasted about like two hours because it was so useful. And then I and then I 37 00:02:01,680 --> 00:02:05,517 kind of got persuaded by people that I really shouldn't have it running like 38 00:02:05,600 --> 00:02:08,637 that. So now I have two laptops until I figure out how this is all going to 39 00:02:08,720 --> 00:02:12,317 work. But giving an AI agent full use of 40 00:02:12,400 --> 00:02:14,317 your computer and your web browser with 41 00:02:14,400 --> 00:02:16,797 all your sessions leads to incredible 42 00:02:16,880 --> 00:02:19,997 stuff. And that seems here to stay. the 43 00:02:20,080 --> 00:02:22,237 the the 44 00:02:22,320 --> 00:02:24,557 idea that we can take 45 00:02:24,640 --> 00:02:27,197 a codeex style workflow and expand it to 46 00:02:27,280 --> 00:02:30,157 all knowledge work. Um so that whatever 47 00:02:30,240 --> 00:02:33,037 you're doing codecs can be or what some 48 00:02:33,120 --> 00:02:34,477 version of that can be using your 49 00:02:34,560 --> 00:02:37,997 computer and using the web for you and 50 00:02:38,080 --> 00:02:41,677 editing documents and 51 00:02:41,760 --> 00:02:46,157 whatever other kind of work you need. um 52 00:02:46,240 --> 00:02:48,477 feels like a genuine transformation in 53 00:02:48,560 --> 00:02:51,677 how knowledge work will happen and that 54 00:02:51,760 --> 00:02:55,517 I think OpenClaw did an incredible job 55 00:02:55,600 --> 00:02:58,157 of 56 00:02:58,240 --> 00:03:00,717 bringing many ideas together to make 57 00:03:00,800 --> 00:03:05,117 that feel usable and real and that seems 58 00:03:05,200 --> 00:03:07,597 certain to be part of our future. 59 00:03:07,680 --> 00:03:10,877 I think moldbook is cool. Uh, and I 60 00:03:10,960 --> 00:03:14,157 think it points to something that will 61 00:03:14,240 --> 00:03:16,397 be real and maybe it will be mold book. 62 00:03:16,480 --> 00:03:20,557 Uh, but there will be new kinds of 63 00:03:20,640 --> 00:03:22,797 social interaction where you have many 64 00:03:22,880 --> 00:03:25,757 agents in a space interacting with each 65 00:03:25,840 --> 00:03:29,917 other on behalf of people. Uh, leading to all sorts of new things. It's like 66 00:03:30,000 --> 00:03:34,237 you you can imagine a totally new kind of social network where everybody makes 67 00:03:34,320 --> 00:03:37,437 an agent or many agents and puts them in there and the agents are talking and 68 00:03:37,520 --> 00:03:40,797 doing stuff and finding them people and information and collaborating with other 69 00:03:40,880 --> 00:03:44,797 people's agents to come up with new ideas that I'm sure there will be an 70 00:03:44,880 --> 00:03:49,037 interesting I think the future of social may look something like that very 71 00:03:49,120 --> 00:03:52,717 different than today whether it's a mult book or not uncertain I think the 72 00:03:52,800 --> 00:03:55,517 obvious constraints are the biggest ones 73 00:03:55,600 --> 00:03:58,637 uh still energy 74 00:03:58,720 --> 00:04:00,237 manufacturing enough hardware, all that 75 00:04:00,320 --> 00:04:02,957 kind of stuff. The non-obvious ones that 76 00:04:03,040 --> 00:04:05,677 are most top of mind for me, one, how 77 00:04:05,760 --> 00:04:09,997 are we going to balance the sort of 78 00:04:10,080 --> 00:04:12,557 security and data access versus the 79 00:04:12,640 --> 00:04:14,477 utility of all of these models? 80 00:04:14,560 --> 00:04:17,597 >> Um, I don't think anyone has a great 81 00:04:17,680 --> 00:04:19,117 answer to this yet. I it feels to me 82 00:04:19,200 --> 00:04:23,197 like there is a new kind of security or 83 00:04:23,280 --> 00:04:28,077 data access paradigm that needs to be invented for this. Um, another is how 84 00:04:28,160 --> 00:04:32,397 are we going to rewrite all software to be equally usable by humans and AIs? 85 00:04:32,480 --> 00:04:35,917 There's like a bunch of weird quirks right now about trying to do that with 86 00:04:36,000 --> 00:04:38,157 the way software work. >> And does that change the architecture of 87 00:04:38,240 --> 00:04:41,277 the software itself where you're going to optimize it for agents more so than 88 00:04:41,360 --> 00:04:45,757 humans? And so it fundamentally changes how you build software. 89 00:04:45,840 --> 00:04:48,157 >> Yes. And there's like big examples of 90 00:04:48,240 --> 00:04:49,677 that and then there's like dumb examples 91 00:04:49,760 --> 00:04:55,197 of that. So, for example, uh I 92 00:04:55,280 --> 00:04:58,077 I would love my agent to be able to use 93 00:04:58,160 --> 00:05:02,237 Slack on my behalf because I hate drowning in Slack and I think it's like 94 00:05:02,320 --> 00:05:04,717 this chaotic mess for me and but it's 95 00:05:04,800 --> 00:05:08,397 important. Uh but the way it works right 96 00:05:08,480 --> 00:05:10,877 now is my agent can use the Slack web 97 00:05:10,960 --> 00:05:15,037 interface and go read all my threads and 98 00:05:15,120 --> 00:05:18,877 do something for me. But then it has marked a bunch of stuff as red in the 99 00:05:18,960 --> 00:05:21,597 process of doing that and it's broken my workflows. 100 00:05:21,680 --> 00:05:24,397 >> Yeah. >> So I don't like that's just a silly 101 00:05:24,480 --> 00:05:28,877 example of how software is not like a lot of software is not quite meant for 102 00:05:28,960 --> 00:05:30,557 an AI and a person to be using it 103 00:05:30,640 --> 00:05:34,237 together. Um maybe we'll want AIs to 104 00:05:34,320 --> 00:05:38,477 like have different sort of user accounts in some ways using 105 00:05:38,560 --> 00:05:40,397 the same kinds of things. maybe a lot of 106 00:05:40,480 --> 00:05:43,997 software will get rewritten so that it's 107 00:05:44,080 --> 00:05:47,677 primarily or largely used by AI but also still works for people using it the 108 00:05:47,760 --> 00:05:50,877 oldfashioned way. Um, another kind of 109 00:05:50,960 --> 00:05:55,117 non-obvious block is how like one of the most powerful things about AI is you 110 00:05:55,200 --> 00:05:59,357 could do this sort of always on computing where you could have 111 00:05:59,440 --> 00:06:03,757 an AI um listening to your meeting or watching your meeting and you know 112 00:06:03,840 --> 00:06:07,597 watching what you're doing on your computer and 113 00:06:07,680 --> 00:06:12,557 uh then just like add a lot of value and do stuff for you. We don't even our like 114 00:06:12,640 --> 00:06:15,917 existing computer hardware is not really meant for that. our permissioning system 115 00:06:16,000 --> 00:06:18,717 and how we think about what an AI gets to see and do stuff with and what it 116 00:06:18,800 --> 00:06:23,517 gets to keep is not really meant for that. Um, our legal system doesn't 117 00:06:23,600 --> 00:06:27,277 really support that. Well, you'd like to be able to, you know, record a meeting 118 00:06:27,360 --> 00:06:31,917 and learn something from it and delete the recording. Uh, so I think there's a 119 00:06:32,000 --> 00:06:33,677 lot of just usability things like that. 120 00:06:33,760 --> 00:06:36,637 The the capability overhang feels to me 121 00:06:36,720 --> 00:06:40,957 like the biggest it has ever felt. I used to say like a few months ago maybe 122 00:06:41,040 --> 00:06:43,917 I would have said it feels bigger than any time except right when CHBT right 123 00:06:44,000 --> 00:06:47,597 before CHBT launched. It now feels even bigger than that. Even though people are 124 00:06:47,680 --> 00:06:50,397 using AI for a lot of stuff. Um the fact 125 00:06:50,480 --> 00:06:55,117 that AI can make small but increasingly big scientific 126 00:06:55,200 --> 00:06:58,637 discoveries. The fact that AI can write full pieces of software. The fact that 127 00:06:58,720 --> 00:07:00,477 soon AI can do more generalized 128 00:07:00,560 --> 00:07:04,637 knowledge work. Um those are all huge 129 00:07:04,720 --> 00:07:06,877 huge things like we have always said 130 00:07:06,960 --> 00:07:10,237 that you know we're going to automate 131 00:07:10,320 --> 00:07:14,637 research use that to be able to automate the economy use that to be able to 132 00:07:14,720 --> 00:07:16,477 deliver incredible new value to people 133 00:07:16,560 --> 00:07:19,597 and share all these benefits and this 134 00:07:19,680 --> 00:07:23,677 whole new technological world. Uh and we always said that as a sort of like 135 00:07:23,760 --> 00:07:25,757 abstract sometime in the future thing. 136 00:07:25,840 --> 00:07:28,237 And yet now AI is doing research and AI 137 00:07:28,320 --> 00:07:33,277 is able to do huge amounts of economic work. And 138 00:07:33,360 --> 00:07:36,637 this, as you said, has happened many times in the past, but living through it 139 00:07:36,720 --> 00:07:40,477 definitely feels like, huh, this is not quite how it seems like it should go. 140 00:07:40,560 --> 00:07:43,197 The the diffusion, the absorption is so slow. 141 00:07:43,280 --> 00:07:46,557 >> Yeah. >> Um, >> is it slower than you thought it would be? 142 00:07:46,640 --> 00:07:48,637 >> Yes. But I think I was just naive and 143 00:07:48,720 --> 00:07:52,077 didn't think about it that hard. And in 144 00:07:52,160 --> 00:07:54,157 retrospect and looking at the history, 145 00:07:54,240 --> 00:07:56,477 it shouldn't be surprising. It feels 146 00:07:56,560 --> 00:07:59,357 fast in some ways uh relative to other 147 00:07:59,440 --> 00:08:03,037 things like you know CHBT grew crazily 148 00:08:03,120 --> 00:08:05,037 quickly uh relative to any other piece 149 00:08:05,120 --> 00:08:07,917 of software before. Um 150 00:08:08,000 --> 00:08:10,157 Codex is growing extremely quickly. I 151 00:08:10,240 --> 00:08:12,877 expect the general purpose computer use 152 00:08:12,960 --> 00:08:17,197 knowledge work to grow quickly and yet 153 00:08:17,280 --> 00:08:19,357 looking at what's capable looking at 154 00:08:19,440 --> 00:08:21,917 what's possible it does feel sort of 155 00:08:22,000 --> 00:08:27,800 surprisingly slow uh in terms of advice 156 00:08:29,200 --> 00:08:31,517 figuring out how to set up enterprises 157 00:08:31,600 --> 00:08:34,557 such that they can quickly absorb these 158 00:08:34,640 --> 00:08:38,557 new tools and not have a year or 159 00:08:38,640 --> 00:08:40,717 multiple years of how are we going to 160 00:08:40,800 --> 00:08:45,677 think about the sort of data access and security questions that 161 00:08:45,760 --> 00:08:47,277 are currently blocking people even at 162 00:08:47,360 --> 00:08:49,917 the level today of adopting things like 163 00:08:50,000 --> 00:08:54,717 codecs that feels very important and I don't want to make a too dramatic of a 164 00:08:54,800 --> 00:08:56,717 prediction but I think the companies 165 00:08:56,800 --> 00:09:01,277 that are not set up to be able to adopt 166 00:09:01,360 --> 00:09:05,357 let's call them AI co-workers very quickly um will be at a huge 167 00:09:05,440 --> 00:09:09,597 disadvantage and it is going to take a lot of work and some risk to be able to adopt 168 00:09:09,680 --> 00:09:14,877 The Codex app is the first time to me it has truly felt like interacting with a 169 00:09:14,960 --> 00:09:18,797 teammate. >> Yeah. >> Um it and I I think like one of the 170 00:09:18,880 --> 00:09:23,117 lessons there is even if you have amazing technology which the models got 171 00:09:23,200 --> 00:09:24,957 I think 52 was the model that got super 172 00:09:25,040 --> 00:09:27,757 good but even once you get there's still 173 00:09:27,840 --> 00:09:30,317 so much value in how you package it, how 174 00:09:30,400 --> 00:09:33,757 you how you have users interact with it, 175 00:09:33,840 --> 00:09:36,477 how easy you can make it. But this it 176 00:09:36,560 --> 00:09:41,197 really does to me now feel in a way that talking to chatbt has always clearly 177 00:09:41,280 --> 00:09:44,637 felt like a tool that I am like working with a collaborator now. 178 00:09:44,720 --> 00:09:47,997 >> Right. Right. >> And 179 00:09:48,080 --> 00:09:53,000 that seems very much like the shape of things to come. 180 00:09:53,520 --> 00:09:55,517 All the evidence we see right now is 181 00:09:55,600 --> 00:09:57,677 that AI models will continue to get much 182 00:09:57,760 --> 00:10:01,117 more capable and also that they will get 183 00:10:01,200 --> 00:10:03,997 much less expensive and use many fewer 184 00:10:04,080 --> 00:10:09,117 resources per task. And all of the history we've seen of 185 00:10:09,200 --> 00:10:10,557 every time that happens is people want 186 00:10:10,640 --> 00:10:13,837 to use them much much more. So we are 187 00:10:13,920 --> 00:10:17,677 planning for a world where 188 00:10:17,760 --> 00:10:21,037 AI usage grows at an accelerated pace 189 00:10:21,120 --> 00:10:24,797 each year. Do you think people are underestimating the capacity that's 190 00:10:24,880 --> 00:10:30,157 going to be needed even now after everything that's being built out? 191 00:10:30,240 --> 00:10:33,037 >> Right now they're saying like $5 trillion will be spent on this over the 192 00:10:33,120 --> 00:10:35,117 course of >> Yeah, maybe if that really gets spent 193 00:10:35,200 --> 00:10:37,757 quickly uh that will be enough. And I 194 00:10:37,840 --> 00:10:39,597 and I think it's also possible that 195 00:10:39,680 --> 00:10:42,477 along the way we have some supply gluts 196 00:10:42,560 --> 00:10:46,317 temporarily. Um but over a period of 197 00:10:46,400 --> 00:10:49,197 decades it seems certain to me that the 198 00:10:49,280 --> 00:10:53,917 world is going to need a lot more tokens even though the and we'll make you know 199 00:10:54,000 --> 00:10:57,917 each token way more efficient. You'll all have a device in your pocket running 200 00:10:58,000 --> 00:11:02,637 super powerful models off of a battery and what but you know there will still 201 00:11:02,720 --> 00:11:04,557 be this drive for more and more and 202 00:11:04,640 --> 00:11:07,117 more. Um 203 00:11:07,200 --> 00:11:10,317 I think the world has come to realize 204 00:11:10,400 --> 00:11:14,637 this and capitalism is doing its thing 205 00:11:14,720 --> 00:11:17,677 and supply chains are reconfiguring. Uh 206 00:11:17,760 --> 00:11:20,237 policy is changing and we are going to 207 00:11:20,320 --> 00:11:22,317 build an incredible amount of 208 00:11:22,400 --> 00:11:26,157 infrastructure. Now is it enough? Uh 209 00:11:26,240 --> 00:11:28,477 people always talk about the total 210 00:11:28,560 --> 00:11:31,917 market demand for AI. It feels to me 211 00:11:32,000 --> 00:11:37,117 something like the market demand for electricity or energy. You can't you 212 00:11:37,200 --> 00:11:39,357 can't talk about that as a general thing. You can say how much demand there 213 00:11:39,440 --> 00:11:43,997 will be at different price levels. Um and in this case you can say at 214 00:11:44,080 --> 00:11:46,157 different price levels for different 215 00:11:46,240 --> 00:11:48,477 quality like how smart it is or how fast 216 00:11:48,560 --> 00:11:53,197 it is or thing like things like that. Um but 217 00:11:53,280 --> 00:11:56,717 if we continue to make AI really capable and really cheap, there will be a ton of 218 00:11:56,800 --> 00:12:01,117 demand at some price. If it's more expensive, there will be uh you know 219 00:12:01,200 --> 00:12:05,197 less demand. But I would like the world to just get 220 00:12:05,280 --> 00:12:09,517 to use a ton of it. Uh we're in this capability over now as we mentioned 221 00:12:09,600 --> 00:12:12,637 where people are like oh you know can use it for chat or maybe some people 222 00:12:12,720 --> 00:12:16,237 understand you can use it for code. I think this will just be like how we do 223 00:12:16,320 --> 00:12:18,957 stuff, how companies run, how scientific 224 00:12:19,040 --> 00:12:22,797 discovery happens, how we use most 225 00:12:22,880 --> 00:12:26,957 software personally in our lives and 226 00:12:27,040 --> 00:12:30,797 making a lot of it if we can have it at a reasonable price seems like a very 227 00:12:30,880 --> 00:12:32,877 good bet. 228 00:12:32,960 --> 00:12:36,397 >> Do you worry about um 229 00:12:36,480 --> 00:12:39,437 the fact that the US has not had as much 230 00:12:39,520 --> 00:12:41,357 of a lead on the open source side? Talk 231 00:12:41,440 --> 00:12:43,597 to us a little bit about that. like 232 00:12:43,680 --> 00:12:48,760 >> I do worry about that. Um 233 00:12:49,600 --> 00:12:52,877 >> yeah, I think we should >> we should do more. 234 00:12:52,960 --> 00:12:56,717 >> Yes. >> Uh so what's what's stopping you from doing it? 235 00:12:56,800 --> 00:13:00,317 >> Focus and time, but I think we need to solve that somehow. To 236 00:13:00,400 --> 00:13:01,837 >> to be clear, I think it's most important 237 00:13:01,920 --> 00:13:03,677 that we lead on frontier models and I 238 00:13:03,760 --> 00:13:06,237 expect those to be accessed via APIs and 239 00:13:06,320 --> 00:13:09,517 other products. So I it would be okay 240 00:13:09,600 --> 00:13:11,517 but not great if we didn't also have the 241 00:13:11,600 --> 00:13:16,637 open source lead. Um people want their own models. People want control of their 242 00:13:16,720 --> 00:13:19,677 own models. Uh people want to run models locally. Especially 243 00:13:19,760 --> 00:13:23,117 >> if you think about a world where you have a model that is going to see your 244 00:13:23,200 --> 00:13:27,117 whole life. You're going to have a new kind of device that's 245 00:13:27,200 --> 00:13:30,157 sort of always on and kind of keeping track of everything and adds huge value 246 00:13:30,240 --> 00:13:32,957 to you. I at least would really like 247 00:13:33,040 --> 00:13:36,317 that running on inference eye control. 248 00:13:36,400 --> 00:13:39,437 Um and and so I think people are going 249 00:13:39,520 --> 00:13:43,917 to need this. Uh you know again if we get behind obviously it's not the end of 250 00:13:44,000 --> 00:13:47,037 the world but I would really like us to lead there and I think there will be 251 00:13:47,120 --> 00:13:49,757 increasing demand for locally running 252 00:13:49,840 --> 00:13:52,557 private models. 253 00:13:52,640 --> 00:13:55,117 We have kind of 254 00:13:55,200 --> 00:13:58,797 two large products and now we have some 255 00:13:58,880 --> 00:13:59,837 others that are coming up. We have 256 00:13:59,920 --> 00:14:04,877 chatbt and our API business. Um and 257 00:14:04,960 --> 00:14:07,117 now we have codeex which looks like it's 258 00:14:07,200 --> 00:14:09,757 going to get very big uh and a few other 259 00:14:09,840 --> 00:14:13,037 things. The and in the future we'll have 260 00:14:13,120 --> 00:14:15,037 consumer devices and robots and all 261 00:14:15,120 --> 00:14:18,317 sorts of other parts of this too. Uh 262 00:14:18,400 --> 00:14:22,797 people do seem very willing to pay for AI as a subscription service. Uh not 263 00:14:22,880 --> 00:14:26,957 everybody but a lot of people are more than we thought would. Um, businesses of 264 00:14:27,040 --> 00:14:29,757 course are willing to pay for things like Chachet Enterprise, 265 00:14:29,840 --> 00:14:31,437 >> but even consumers, you've reconditioned 266 00:14:31,520 --> 00:14:34,397 the consumers quite a bit. Like >> consumers, I mean, we have I don't know 267 00:14:34,480 --> 00:14:37,597 the exact number off the top of my head, but you know, many many tens of millions 268 00:14:37,680 --> 00:14:39,597 of consumers paying a subscription fee. 269 00:14:39,680 --> 00:14:42,717 >> Yeah. >> Uh, and 270 00:14:42,800 --> 00:14:45,837 that was a surprise to me to be honest, 271 00:14:45,920 --> 00:14:49,357 a happy one. Um, so that I think can go 272 00:14:49,440 --> 00:14:50,957 further and as we add things like 273 00:14:51,040 --> 00:14:52,957 codecs, people are willing to pay much 274 00:14:53,040 --> 00:14:56,717 much more there. um advertising 275 00:14:56,800 --> 00:15:01,517 for mass scale consumer businesses does seem like a good model. Uh although I 276 00:15:01,600 --> 00:15:03,277 think we'll have to be very careful in 277 00:15:03,360 --> 00:15:07,277 how we do that. Uh and then 278 00:15:07,360 --> 00:15:11,197 businesses increasingly want 279 00:15:11,280 --> 00:15:14,557 something like an AI cloud subscription >> where they're like I want to partner 280 00:15:14,640 --> 00:15:15,997 with an AI company. I want you to like 281 00:15:16,080 --> 00:15:19,837 handle security and context linking and 282 00:15:19,920 --> 00:15:23,837 access and I want to be able to run lots of agents on it. I want like a general 283 00:15:23,920 --> 00:15:26,477 purpose platform there. I want some agents from you, some agents from other 284 00:15:26,560 --> 00:15:30,237 people. I may even want to run other people's models. I want a Chach 285 00:15:30,320 --> 00:15:35,117 enterprise license. I want a ton of API access. Uh you know, things like that. 286 00:15:35,200 --> 00:15:38,557 So, I think there will be a a model there as well. Clear. We don't want to 287 00:15:38,640 --> 00:15:42,077 take a share of people that are just like paying us for the API and making 288 00:15:42,160 --> 00:15:46,557 you know, discoveries like that's great and that's theirs. Uh I can imagine 289 00:15:46,640 --> 00:15:48,717 though that like if there is a future 290 00:15:48,800 --> 00:15:51,917 where we can spend billions of dollars 291 00:15:52,000 --> 00:15:54,637 on inference and cure an important 292 00:15:54,720 --> 00:15:57,597 disease. Uh 293 00:15:57,680 --> 00:15:59,597 we may explore partnerships there where 294 00:15:59,680 --> 00:16:04,637 we pay for that cost in partnership with the drug company and then kind of get 295 00:16:04,720 --> 00:16:09,197 some royalty on it. This is not something we're doing now, but 296 00:16:09,280 --> 00:16:10,957 I think the frontier of scientific 297 00:16:11,040 --> 00:16:13,597 discovery with AI will require so much 298 00:16:13,680 --> 00:16:16,317 capital that maybe we think of ourselves 299 00:16:16,400 --> 00:16:19,037 as an investor in some of those cases. I 300 00:16:19,120 --> 00:16:21,597 can imagine like billions of humanoid 301 00:16:21,680 --> 00:16:26,237 robots building more data centers and mining for material and building more 302 00:16:26,320 --> 00:16:30,397 power plants and all of that. Um, I can imagine just the economy growing at an 303 00:16:30,480 --> 00:16:33,917 unprecedented rate as there's all sorts of 304 00:16:34,000 --> 00:16:36,157 incredible new services and scientific 305 00:16:36,240 --> 00:16:40,477 discoveries happening. Um, I can imagine 306 00:16:40,560 --> 00:16:45,037 the Bonoyman probes launching and then beyond that, I got no idea. And the 307 00:16:45,120 --> 00:16:49,037 models are going to get so much better quickly. like I have we've been trying 308 00:16:49,120 --> 00:16:52,797 to figure out how we communicate about what we think is happening that does not 309 00:16:52,880 --> 00:16:55,597 seem like hype or 310 00:16:55,680 --> 00:16:59,117 um you know crazy but the models are 311 00:16:59,200 --> 00:17:01,117 just going to get 312 00:17:01,200 --> 00:17:04,237 very very good this year and that will 313 00:17:04,320 --> 00:17:06,717 that is a tailwind that I think we can 314 00:17:06,800 --> 00:17:13,480 build incredible things with headwind uh 315 00:17:13,600 --> 00:17:16,717 yeah probably some sort of like global 316 00:17:16,800 --> 00:17:20,797 destabilization, mega supply chain disruption, something like that. 317 00:17:20,880 --> 00:17:23,597 >> And and on the on the on the tailwind 318 00:17:23,680 --> 00:17:25,037 side on the models getting infinitely 319 00:17:25,120 --> 00:17:28,557 better, do you think in 26 we see a 10x 320 00:17:28,640 --> 00:17:32,797 improvement, a 100x improvement, a 5x improvement? 321 00:17:32,880 --> 00:17:35,597 >> What do you mean by like a what is a 10x improvement? Like 322 00:17:35,680 --> 00:17:37,197 >> things that you can problems you can 323 00:17:37,280 --> 00:17:40,717 solve that you you were 324 00:17:40,800 --> 00:17:44,557 >> Yeah, I would bet it subjectively feels something like 10x the end of this year. 325 00:17:44,640 --> 00:17:48,237 Like that's a I don't know how to like put an exact metric on that. 326 00:17:48,320 --> 00:17:52,760 >> Yeah. >> But that feels reasonable.