1 00:00:00,240 --> 00:00:02,637 Well, >> without our first guest of the show, 2 00:00:02,720 --> 00:00:06,477 >> Sam Alman, the CEO of OpenAI. He's in the ream waiting room. Welcome to the 3 00:00:06,560 --> 00:00:07,997 show, Sam. How are you doing? >> Welcome back. 4 00:00:08,080 --> 00:00:09,517 >> Good. Thank you guys for having me back. 5 00:00:09,600 --> 00:00:10,877 >> Thanks so much. >> Big day. 6 00:00:10,960 --> 00:00:15,037 >> Big day. Kick us off with where should we start? Should we start with uh the 7 00:00:15,120 --> 00:00:19,117 model or Frontier? >> Can we start with the model just cuz I'm having fun with it? 8 00:00:19,200 --> 00:00:21,517 >> Yeah, absolutely. Break it down. What' you launch? 9 00:00:21,600 --> 00:00:25,197 >> Uh so we launched 5.3 Codeex. It is I think the best coding model in the 10 00:00:25,280 --> 00:00:29,277 world. We took a lot of the feedback that people had about 5.2 uh 5.2 two 11 00:00:29,360 --> 00:00:33,037 codecs and got it into one model. It is much smarter at programming, but it's 12 00:00:33,120 --> 00:00:36,637 also way faster. You can interact with it mid turn. I think it's got a much 13 00:00:36,720 --> 00:00:39,357 better personality. >> Um, it's really good at computer use. 14 00:00:39,440 --> 00:00:42,957 Uh, so it feels like a very big step forward. It was funny as we were uh 15 00:00:43,040 --> 00:00:45,437 deploying it this morning. A couple of 16 00:00:45,520 --> 00:00:49,997 very uh like extremely 17 00:00:50,080 --> 00:00:52,557 experts at using these models noticed and said, "Man, something's really 18 00:00:52,640 --> 00:00:55,197 different with codeex." And they like caught it mid deploy. So, I think you 19 00:00:55,280 --> 00:00:57,517 can really feel it quickly. >> That's great. Oh, you're saying people 20 00:00:57,600 --> 00:00:59,437 outside of OpenAI just everyday use. 21 00:00:59,520 --> 00:01:02,717 >> Yeah, I like the shower that we put it out before we, you know, 22 00:01:02,800 --> 00:01:06,397 >> talk about uh interacting with it mid turn. How does that work? Why is that 23 00:01:06,480 --> 00:01:08,477 important? What does that unlock? 24 00:01:08,560 --> 00:01:09,917 >> So, people are starting to use these 25 00:01:10,000 --> 00:01:13,277 these tools for very long pieces of work 26 00:01:13,360 --> 00:01:15,757 at one time, you know, multi-hour tasks. 27 00:01:15,840 --> 00:01:18,717 And sometimes you don't specify it correctly, right? Sometimes something's 28 00:01:18,800 --> 00:01:22,477 not set up right, something just screws up. the ability they they can do amazing 29 00:01:22,560 --> 00:01:25,597 things with no steering but they can do much more amazing things if you steer 30 00:01:25,680 --> 00:01:29,677 them along the way. Uh so this has been this is one of the things I felt most 31 00:01:29,760 --> 00:01:33,277 new about this model. >> So talk about orchestration how this 32 00:01:33,360 --> 00:01:36,397 fits into Frontier because I >> imagine one second is it's notable like 33 00:01:36,480 --> 00:01:38,797 if you see a co-orker making a mistake 34 00:01:38,880 --> 00:01:43,437 and you don't interrupt them that's rude right like it's it's it's deeply 35 00:01:43,520 --> 00:01:46,557 inefficient. >> It is incredible what these models can 36 00:01:46,640 --> 00:01:51,117 do without any feedback. Like if you think about a new co-orker especially um 37 00:01:51,200 --> 00:01:53,837 you know you train them and you give them a lot of feedback early on and they 38 00:01:53,920 --> 00:01:57,357 learn the job and you correct them and they they kind of get practice and the 39 00:01:57,440 --> 00:02:01,437 models they will soon do that but right now they don't do that. So we just rely 40 00:02:01,520 --> 00:02:05,277 on either they get it right one shot or we collect them correct them along the way. 41 00:02:05,360 --> 00:02:09,437 >> Yeah. So uh you know I think there's a lot of people that are running multiple 42 00:02:09,520 --> 00:02:12,637 agents and multiple tabs. They're starting to think about orchestration. 43 00:02:12,720 --> 00:02:17,037 Feels like Frontier is a piece of that, but if you're interacting with a model 44 00:02:17,120 --> 00:02:19,117 that's running midturn, like how does 45 00:02:19,200 --> 00:02:21,357 the user experience change for 46 00:02:21,440 --> 00:02:24,397 developers with 5.3 and then what will 47 00:02:24,480 --> 00:02:27,277 it look like in the frontier world? 48 00:02:27,360 --> 00:02:31,357 I I think we will be heading towards a workflow where a lot of people just feel 49 00:02:31,440 --> 00:02:33,997 like they're managing a team of agents and they'll keep as the agents get 50 00:02:34,080 --> 00:02:39,117 better, they'll keep operating at a higher and higher level of abstraction, which 51 00:02:39,200 --> 00:02:42,237 at least watching what's happening so far is a is a jump that people are going 52 00:02:42,320 --> 00:02:46,877 to make >> pretty well. Um the 53 00:02:46,960 --> 00:02:48,877 the models are so good now there's such 54 00:02:48,960 --> 00:02:51,677 a capability overhang that building 55 00:02:51,760 --> 00:02:55,677 better tools to let people um do that 56 00:02:55,760 --> 00:02:59,677 which the Codex app that we launched on Monday was a great step forward for will 57 00:02:59,760 --> 00:03:04,717 be very very important but you will be managing very complex workflows the 58 00:03:04,800 --> 00:03:07,757 agents will keep getting better so you'll keep working at the maximum of 59 00:03:07,840 --> 00:03:11,357 your management bandwidth or cognitive ability to keep track of all the stuff 60 00:03:11,440 --> 00:03:15,037 um and the tools to make that easy to do will matter 61 00:03:15,120 --> 00:03:18,237 uh I think more than intelligence for a little while because there's such an 62 00:03:18,320 --> 00:03:19,917 intelligence overhang already. 63 00:03:20,000 --> 00:03:21,597 >> What's the role of a forward deployed 64 00:03:21,680 --> 00:03:25,037 engineer today at OpenAI towards the end 65 00:03:25,120 --> 00:03:27,677 of the year? Capability overhang that 66 00:03:27,760 --> 00:03:30,477 feels like raw meat for a for a forward 67 00:03:30,560 --> 00:03:32,317 deployed engineer. It's like the like 68 00:03:32,400 --> 00:03:34,957 they solve that problem, right? 69 00:03:35,040 --> 00:03:38,877 >> Yeah. I mean look eventually the models will get so good that they'll help 70 00:03:38,960 --> 00:03:41,997 companies deploy themselves and the four deployed engineers will again get to 71 00:03:42,080 --> 00:03:45,997 work at a higher level of abstraction >> but for now you go into a company that 72 00:03:46,080 --> 00:03:50,477 is not AI native and say uh okay you've said you want us 73 00:03:50,560 --> 00:03:53,997 >> you know they say they want to deploy AI um they really are not sure what to do 74 00:03:54,080 --> 00:03:56,637 how do I hook this up to my systems do I 75 00:03:56,720 --> 00:04:01,197 need to fine-tune a model on my codebase uh how do I think about orchestrating 76 00:04:01,280 --> 00:04:04,477 agents and using things from different companies most of all or at least what 77 00:04:04,560 --> 00:04:09,037 we hear most frequently is how do I think about security of my data and how 78 00:04:09,120 --> 00:04:12,717 do I know that these like AI co-working agents are not going to go access a 79 00:04:12,800 --> 00:04:15,997 bunch of information and share it in ways they shouldn't or get you know a 80 00:04:16,080 --> 00:04:17,997 context exploit or something like that. 81 00:04:18,080 --> 00:04:20,877 So the forward engineers take this 82 00:04:20,960 --> 00:04:25,357 incredible new technology and uh a platform like Frontier and say we will 83 00:04:25,440 --> 00:04:29,597 connect your company to an AI platform so that you can use all these agents and 84 00:04:29,680 --> 00:04:31,197 workflows and everything else you want. 85 00:04:31,280 --> 00:04:35,917 How important are these metaphors or how temporary are they? I was very 86 00:04:36,000 --> 00:04:39,437 interested in reading about Gas Town and you have these pole cats and it's this 87 00:04:39,520 --> 00:04:44,877 whole Mad Max world and uh that feels like maybe just a temporary aberration 88 00:04:44,960 --> 00:04:48,717 where you're setting up agents for specific tasks but also that could be 89 00:04:48,800 --> 00:04:52,397 incredibly valuable in explaining to a large corporation of how they're going 90 00:04:52,480 --> 00:04:55,517 to integrate AI across the whole organization. 91 00:04:55,600 --> 00:04:58,877 Yeah, I 92 00:04:58,960 --> 00:05:02,557 I suspect like everything else that's happening 93 00:05:02,640 --> 00:05:07,037 when an industry is moving so fast, all of this is somewhat temporary on like a 94 00:05:07,120 --> 00:05:08,397 long enough time scale. 95 00:05:08,480 --> 00:05:11,357 >> Um, and 96 00:05:11,440 --> 00:05:16,637 you as these models become more capable and these agents are operating on very 97 00:05:16,720 --> 00:05:19,917 long time horizons with the ability to just kind of figure it out and our trust 98 00:05:20,000 --> 00:05:23,997 and their robustness keeps going up. Um, then maybe you don't need a lot of the 99 00:05:24,080 --> 00:05:25,837 abstractions we need today. You know, 100 00:05:25,920 --> 00:05:28,397 maybe you just like have a a single AI 101 00:05:28,480 --> 00:05:31,917 bot that runs at your company and you 102 00:05:32,000 --> 00:05:35,517 can say, "Hey, I want to like launch this new product and it does everything 103 00:05:35,600 --> 00:05:38,957 an ambitious person would do." Um, but that's not where we are today. So, today 104 00:05:39,040 --> 00:05:42,077 we have to >> yeah, >> put in a little more work to get the pieces put together. 105 00:05:42,160 --> 00:05:46,237 >> Yeah. How are you thinking about the the the meter benchmark for long task 106 00:05:46,320 --> 00:05:50,717 horizon? Uh, you're at the top of the charts. Uh, at the same time, it feels 107 00:05:50,800 --> 00:05:55,117 like we might need a new chart if we're talking about agent swarms because 108 00:05:55,200 --> 00:05:59,837 they'll be able to do things that go for weeks, but they will subdivide the work. 109 00:05:59,920 --> 00:06:03,117 There's some subdivision that happens within a reasoning model, but it doesn't 110 00:06:03,200 --> 00:06:06,557 truly parallelize, at least that I'm aware of. So, what does it look like in 111 00:06:06,640 --> 00:06:11,437 a world where you go to a model, but now it's spinning up a whole bunch of 112 00:06:11,520 --> 00:06:15,197 different models underneath? >> I think you have there's two of of the 113 00:06:15,280 --> 00:06:18,237 key insights of the whole field are in this question. Number one, uh the 114 00:06:18,320 --> 00:06:20,157 implication that no chart in AI lasts 115 00:06:20,240 --> 00:06:25,437 more than a few years is right. And like this one kind of, you know, we'll see 116 00:06:25,520 --> 00:06:26,877 how much longer it's really useful. 117 00:06:26,960 --> 00:06:32,397 >> Yeah. >> The second is >> a lot of people thought that, okay, 118 00:06:32,480 --> 00:06:35,677 we're going to just need a super long task and a super long task horizon, so 119 00:06:35,760 --> 00:06:39,757 we need a super long context. And definitely what people have already seen 120 00:06:39,840 --> 00:06:42,637 with coding agents is by agents breaking 121 00:06:42,720 --> 00:06:46,797 up work, orchestrating it well, farming it off to sub agents. even with the 122 00:06:46,880 --> 00:06:50,477 current limitations of the technology, we can do something which should not be 123 00:06:50,560 --> 00:06:54,317 surprising because it's similar to how people do things and get amazing amounts 124 00:06:54,400 --> 00:07:00,920 of work done. So that's been cool to watch. I think that will keep going. Um 125 00:07:01,520 --> 00:07:05,517 a a joke that some people at OpenAI make is that soon the chart that matters is 126 00:07:05,600 --> 00:07:07,437 just going to be GDP impact. >> Yeah. 127 00:07:07,520 --> 00:07:10,877 >> And then the question is what's the one that comes after that? But everything 128 00:07:10,960 --> 00:07:14,237 else, these lot of these proxy metrics, there's now so much economic value in 129 00:07:14,320 --> 00:07:16,637 what the models are doing. What do you think could come after? 130 00:07:16,720 --> 00:07:18,317 >> I have no idea. Do you have an opinion? 131 00:07:18,400 --> 00:07:20,397 >> Happiness. >> I don't know. When you look When you 132 00:07:20,480 --> 00:07:23,677 look back at some of your when you look back at some of your blog posts from 10 133 00:07:23,760 --> 00:07:26,717 years ago, your predictions were usually pretty on point. 134 00:07:26,800 --> 00:07:28,317 >> Yeah. >> Maybe it's hard to predict. 135 00:07:28,400 --> 00:07:33,437 >> Thank you. Thank you. >> I uh >> specifically the specifically the merge 136 00:07:33,520 --> 00:07:36,237 basically. There's a bunch that are good. Like right now we're getting you 137 00:07:36,320 --> 00:07:38,317 know thousands of messages in the chat 138 00:07:38,400 --> 00:07:41,437 about 40 and you predicted in 2016 that 139 00:07:41,520 --> 00:07:44,557 people could uh would would become uh 140 00:07:44,640 --> 00:07:46,877 you know very attached to a chatbot. 141 00:07:46,960 --> 00:07:49,197 >> Yeah. I'm working on like a big 142 00:07:49,280 --> 00:07:51,997 prediction blog post for the next 10 143 00:07:52,080 --> 00:07:57,117 years seems too far, but the next five uh but because it's like I'm sure a lot 144 00:07:57,200 --> 00:08:01,517 of it'll be wrong and it >> you know it's it's still fun to try. Um 145 00:08:01,600 --> 00:08:03,837 the 146 00:08:03,920 --> 00:08:08,637 >> the the sort of like relationships with chatbots clearly that's something now we 147 00:08:08,720 --> 00:08:10,237 got to worry about more and is no longer 148 00:08:10,320 --> 00:08:14,317 a abstract concept. Um the even the 149 00:08:14,400 --> 00:08:17,277 question of what comes after GDP like 150 00:08:17,360 --> 00:08:19,117 one reason I think that's interesting is 151 00:08:19,200 --> 00:08:24,397 the way we measure GDP now could start going down even though quality of life 152 00:08:24,480 --> 00:08:27,997 goes way up. >> Sure. >> And we don't have a lot of practice with 153 00:08:28,080 --> 00:08:32,637 things like that. But the the the massively deflationary 154 00:08:32,720 --> 00:08:37,437 >> not just Europe we want quality of life going up too just going down you know. 155 00:08:37,520 --> 00:08:41,117 >> Yeah. Yeah. Um, >> switching gears, uh, what do you think 156 00:08:41,200 --> 00:08:45,997 about the work that that the Neol boom, the research efforts that are happening 157 00:08:46,080 --> 00:08:50,317 all over Silicon Valley? It feels like there's there's an acknowledgment that 158 00:08:50,400 --> 00:08:54,157 there's research breakthroughs that need to happen and everyone's taking 159 00:08:54,240 --> 00:08:56,157 different uh, shots at those. Do you 160 00:08:56,240 --> 00:09:01,277 think that those companies will just find a breakthrough and join a lab or 161 00:09:01,360 --> 00:09:04,557 launch their own products? >> I think it's fantastic. First of all, 162 00:09:04,640 --> 00:09:08,317 um, one of the meta things that we wanted to do when we started with OpenAI, 163 00:09:08,400 --> 00:09:11,437 >> yeah, >> is >> when we started OpenAI is like 164 00:09:11,520 --> 00:09:14,477 >> there had been a period where the technology industry and Silicon Valley 165 00:09:14,560 --> 00:09:17,437 in particular was amazing at new research labs or just doing new research 166 00:09:17,520 --> 00:09:20,477 in industry in general >> and then it kind of fell apart and there 167 00:09:20,560 --> 00:09:23,597 hadn't been a good one in a while and part of what we hope to show and this 168 00:09:23,680 --> 00:09:27,037 was not only us like a lot of people were excited about new research labs 169 00:09:27,120 --> 00:09:30,877 >> is that industry could do research again. So seeing that now become like 170 00:09:30,960 --> 00:09:34,557 fashionable in all of these new labs, I think it's totally awesome. Some will 171 00:09:34,640 --> 00:09:38,557 succeed, some will fail, some will going to go into some other effort, but uh 172 00:09:38,640 --> 00:09:43,117 having industry support research in, you know, startup style, I think it's 173 00:09:43,200 --> 00:09:46,157 wonderful. >> Over the next two years, would you 174 00:09:46,240 --> 00:09:48,157 expect to acquire more individual 175 00:09:48,240 --> 00:09:53,480 product companies or more research labs? 176 00:09:54,880 --> 00:09:57,037 >> Good question. 177 00:09:57,120 --> 00:09:59,037 I don't have a strong opinion there. I 178 00:09:59,120 --> 00:10:01,997 would bet 179 00:10:02,080 --> 00:10:06,717 well I'd say the very best ones will often look like a mixture of both. Um 180 00:10:06,800 --> 00:10:09,917 like the the the one that I have in mind right now is something that very much 181 00:10:10,000 --> 00:10:12,077 looks like a a mixture of both. So maybe 182 00:10:12,160 --> 00:10:17,197 that's may maybe the shape of things to come is the really truly extraordinary 183 00:10:17,280 --> 00:10:20,157 product work will have more and more of a research component 184 00:10:20,240 --> 00:10:22,397 >> and it'll be kind of a more of a hybrid thing. 185 00:10:22,480 --> 00:10:26,717 >> Is data the new oil? Is there value? We were joking a couple days ago. 186 00:10:26,800 --> 00:10:29,517 >> They have a bunch of data, but they don't understand AI. They don't know how 187 00:10:29,600 --> 00:10:33,037 to monetize. >> Well, and that phrase was effectively wasted a decade ago. 188 00:10:33,120 --> 00:10:35,117 >> Yeah. >> And so to say it now sounds really 189 00:10:35,200 --> 00:10:38,157 silly, but it feels like it could be more true now than ever. 190 00:10:38,240 --> 00:10:42,477 >> Yeah. >> Um, you know, certainly. Yeah, man. They 191 00:10:42,560 --> 00:10:45,117 really did waste it a decade ago. I was just thinking of the kind of people that 192 00:10:45,200 --> 00:10:46,717 used to say that came out of it. 193 00:10:46,800 --> 00:10:48,637 >> TED talks, you you're not supposed to 194 00:10:48,720 --> 00:10:49,917 call them out. 195 00:10:50,000 --> 00:10:53,837 >> They're nice people over at TED. Uh, 196 00:10:53,920 --> 00:10:55,677 you know, definitely like the the sort 197 00:10:55,760 --> 00:11:00,397 of the magic relationship of this last 198 00:11:00,480 --> 00:11:01,517 eight years, whatever you want to call 199 00:11:01,600 --> 00:11:04,557 it, has has been like, you know, that we 200 00:11:04,640 --> 00:11:08,077 can put in more and more resources, 201 00:11:08,160 --> 00:11:10,317 compute, data, 202 00:11:10,400 --> 00:11:13,197 new ideas, whatever into creating uh an 203 00:11:13,280 --> 00:11:17,917 artifact and it gets like the log of it gets better. Mhm. 204 00:11:18,000 --> 00:11:21,277 >> Um, so you can throw in, you know, that's why we have this huge exponential 205 00:11:21,360 --> 00:11:24,557 increase in resources, but we keep getting better and better models. And 206 00:11:24,640 --> 00:11:28,957 for all of the concern people have about it's it's going to like top out or it's 207 00:11:29,040 --> 00:11:32,317 slowing down or whatever, like no one's been right about that. I mean, sometimes 208 00:11:32,400 --> 00:11:35,757 they it looked like they were for a couple of months as we digested a new 209 00:11:35,840 --> 00:11:38,317 model or came to a new form factor, but 210 00:11:38,400 --> 00:11:41,757 it has been an incredibly smooth last 211 00:11:41,840 --> 00:11:46,317 six or eight years of this. uh what those resources are. There can be some 212 00:11:46,400 --> 00:11:50,077 trade-off between, you know, sometimes it's better to spend your money on 213 00:11:50,160 --> 00:11:55,037 better data, sometimes on more compute, sometimes something else. On the whole, 214 00:11:55,120 --> 00:11:59,357 uh compute power is the new oil is the 215 00:11:59,440 --> 00:12:02,157 statement that feels closest to true to me, 216 00:12:02,240 --> 00:12:04,077 >> but there will be other parts too. 217 00:12:04,160 --> 00:12:08,157 >> Makes sense. Is software dead? 218 00:12:08,240 --> 00:12:12,237 >> It's different. It's definitely not dead. Um, but what software, like how 219 00:12:12,320 --> 00:12:16,717 you create it, how you're going to use it, how much you're going to have written 220 00:12:16,800 --> 00:12:21,437 for you each time you need it versus how much you'll want sort of a consistent UX. 221 00:12:21,520 --> 00:12:24,797 >> Yeah. >> Uh, that's all going to change. Um, you 222 00:12:24,880 --> 00:12:29,197 know, there have been a number of these like big sell-offs of SAS stocks over 223 00:12:29,280 --> 00:12:34,157 the last few years as these models have rolled out. Um, there I expect there 224 00:12:34,240 --> 00:12:37,437 will continue to be more. I expect there will be big booms in software. I think 225 00:12:37,520 --> 00:12:39,997 it's just going to be volatile for a while. Mhm. 226 00:12:40,080 --> 00:12:43,517 >> As people figure out what this um looks 227 00:12:43,600 --> 00:12:47,597 like, the statement that uh someone said to me that has stuck in my mind most 228 00:12:47,680 --> 00:12:51,437 these last couple of weeks is that every company is an API company now whether 229 00:12:51,520 --> 00:12:53,037 they want to be or not. >> Oh yeah. 230 00:12:53,120 --> 00:12:56,797 >> Um because agents are just going to be able to >> write a CS. 231 00:12:56,880 --> 00:13:00,317 >> Yeah. We had we had D we had D from Uber on yesterday and it he had a pretty 232 00:13:00,400 --> 00:13:04,077 refreshing kind of approach. We were asking about integrating agents with 233 00:13:04,160 --> 00:13:08,637 Uber and and he recognized that yeah, the ad business could potentially be 234 00:13:08,720 --> 00:13:13,597 threatened if you can order an Uber and chatbt, but he basically said 235 00:13:13,680 --> 00:13:15,517 >> like you have to think of the consumer. 236 00:13:15,600 --> 00:13:19,757 >> The consumer wants to order an Uber via their preferred agent. You should let 237 00:13:19,840 --> 00:13:23,357 them otherwise you're you're going to have other problems. 238 00:13:23,440 --> 00:13:27,277 >> Yeah, that that is the right take for sure. And I don't or I think so at least. 239 00:13:27,360 --> 00:13:29,917 >> And >> you know, we've been through platform 240 00:13:30,000 --> 00:13:33,037 shifts like this before where you I mean Uber wouldn't have existed without one. 241 00:13:33,120 --> 00:13:35,837 It wasn't until the iPhone where you could like have it make sense to order 242 00:13:35,920 --> 00:13:38,797 an Uber to right where you were uh as you're out in the world. 243 00:13:38,880 --> 00:13:40,797 >> Um so I think there will be totally new 244 00:13:40,880 --> 00:13:42,397 things that happen. other things you'll 245 00:13:42,480 --> 00:13:46,077 use in new ways. But definitely as I've 246 00:13:46,160 --> 00:13:49,517 started using Codex, my 247 00:13:49,600 --> 00:13:53,357 excitement about having agents go off and do things for me and still use other 248 00:13:53,440 --> 00:13:56,477 services, pay other services. I'm sure we'll need to figure out new business 249 00:13:56,560 --> 00:14:00,237 models and how revenue gets shared around, but that will happen. 250 00:14:00,320 --> 00:14:01,837 >> Yeah. Talk more about desktop. 251 00:14:01,920 --> 00:14:05,117 >> One more question on on SAS. Have any uh 252 00:14:05,200 --> 00:14:10,237 public market SAS companies tried to get a soft landing with OpenAI? M and do you 253 00:14:10,320 --> 00:14:14,557 think there's any there's there's any value just let's say 254 00:14:14,640 --> 00:14:18,397 >> that I'm no one has no public SAS companies that I'm aware of have tried 255 00:14:18,480 --> 00:14:22,477 that >> with open AI um look >> but some of them some of them some of 256 00:14:22,560 --> 00:14:24,077 them will certainly be 257 00:14:24,160 --> 00:14:26,877 >> durable and and are on sale right now 258 00:14:26,960 --> 00:14:29,757 and and potentially just need new energy 259 00:14:29,840 --> 00:14:34,237 and and uh need kind of a new an entirely new approach and maybe OpenAI 260 00:14:34,320 --> 00:14:37,517 could provide that. >> Yeah. >> Yeah. I think some will be incredibly 261 00:14:37,600 --> 00:14:42,557 valuable. Some do feel like a thinner layer now. Um, but I don't know like I 262 00:14:42,640 --> 00:14:46,877 was talking recently to a bunch of SAS companies and 263 00:14:46,960 --> 00:14:50,477 they do not they do not feel unexited like they're like we're going to go 264 00:14:50,560 --> 00:14:54,157 through a big transformation here and you know yeah sure other people can 265 00:14:54,240 --> 00:14:57,677 instantly write software now but so can we and we got a great system of record 266 00:14:57,760 --> 00:15:01,197 and seems reasonable. Some won't make it of course. 267 00:15:01,280 --> 00:15:03,517 >> Yeah. Yeah. Uh yeah. Talk more about the 268 00:15:03,600 --> 00:15:05,917 the Codex desktop roll out. It feels 269 00:15:06,000 --> 00:15:09,757 like um uh you know successful amount of 270 00:15:09,840 --> 00:15:13,277 downloads but but key like a shift for 271 00:15:13,360 --> 00:15:16,797 you know people who are maybe lightly technical but don't have time to set up 272 00:15:16,880 --> 00:15:19,437 an IDE and configure an environment to 273 00:15:19,520 --> 00:15:21,997 actually start writing software. I want 274 00:15:22,080 --> 00:15:24,317 to know about plans to integrate to the 275 00:15:24,400 --> 00:15:26,717 phone. That was a big moment I think for 276 00:15:26,800 --> 00:15:29,197 a lot of people with the the clawedbot, 277 00:15:29,280 --> 00:15:32,317 moltbot, open claw thing was like, oh, I 278 00:15:32,400 --> 00:15:36,237 can text something and it will go and write code and that's valuable and that 279 00:15:36,320 --> 00:15:38,637 unlocks a new agentic experience. Like 280 00:15:38,720 --> 00:15:43,517 where do you see the codeex desktop ecosystem going? 281 00:15:43,600 --> 00:15:47,197 >> I am so 282 00:15:47,280 --> 00:15:51,757 codex desktop has been a a somewhat of a surprise to me um in terms of how much 283 00:15:51,840 --> 00:15:55,837 people love it, including how much I love it myself. you know, the I think 284 00:15:55,920 --> 00:15:59,837 it's a great example of 10% of polish of the experience of using these models, 285 00:15:59,920 --> 00:16:00,797 especially when there's so much 286 00:16:00,880 --> 00:16:04,317 capability overhang goes an extremely 287 00:16:04,400 --> 00:16:09,677 long way to what you can build and how how you interact with this stuff. Um, of 288 00:16:09,760 --> 00:16:12,717 course, we should have an ability to kick off new tasks from mobile and we'll 289 00:16:12,800 --> 00:16:16,237 do that. I mean I I mean really what you want is like your single AI that's 290 00:16:16,320 --> 00:16:20,077 working for you on a unified backend, access to all of your data and your 291 00:16:20,160 --> 00:16:24,157 ideas and your stuff and your memory and your all the context and the ability to 292 00:16:24,240 --> 00:16:28,557 work across a lot of surfaces and often you'll be at your desktop. Um often 293 00:16:28,640 --> 00:16:30,237 you'll be on your phone and you just 294 00:16:30,320 --> 00:16:33,197 want to add something in. Um 295 00:16:33,280 --> 00:16:36,557 but it is a pretty 296 00:16:36,640 --> 00:16:40,717 profound shift in my own workflow, not just for coding tasks, but more general 297 00:16:40,800 --> 00:16:44,157 purpose tasks. It's still kind of hard to use if you're not at least reasonably 298 00:16:44,240 --> 00:16:48,077 technical. But obviously, we'll find a version of this product that can do 299 00:16:48,160 --> 00:16:51,917 other knowledge work tasks and control your computer and things like that where 300 00:16:52,000 --> 00:16:55,997 you don't have to be and we'll and it'll bring the magic of building stuff really 301 00:16:56,080 --> 00:16:59,917 to a lot of people because even if you never look at code, you'll be able to 302 00:17:00,000 --> 00:17:03,677 build something reasonably sophisticated. One of the things that I 303 00:17:03,760 --> 00:17:06,717 have built when I was playing around with the new codeex app is this thing I 304 00:17:06,800 --> 00:17:11,037 had always wanted just like this magic autocompleting to-do list. 305 00:17:11,120 --> 00:17:15,197 I I really work with to-do lists and this idea that I could just put tasks in 306 00:17:15,280 --> 00:17:17,917 and it would try to go do them. If it could complete them, it could complete 307 00:17:18,000 --> 00:17:21,037 them. If it needed questions, it would ask me questions. If it needed, you 308 00:17:21,120 --> 00:17:25,917 know, if I had to do something, I could still do it the oldfashioned way. But an 309 00:17:26,000 --> 00:17:28,877 interface like that where, you know, all the stuff you want to do, you just sort 310 00:17:28,960 --> 00:17:32,957 of explain to a computer or your AI and it tries to go off and do them. And 311 00:17:33,040 --> 00:17:36,797 sure, if you're on your phone, you're going to just add a a task on your phone 312 00:17:36,880 --> 00:17:39,357 or, you know, if you want to easily import something from email, you're 313 00:17:39,440 --> 00:17:41,677 going to do that. like felt feels really 314 00:17:41,760 --> 00:17:44,717 good. So, I'm excited about all of the 315 00:17:44,800 --> 00:17:48,957 ways that this will just become a general knowledge work agent. 316 00:17:49,040 --> 00:17:49,357 >> Yeah. 317 00:17:49,440 --> 00:17:53,837 >> Were you uh unsurprised to see a product 318 00:17:53,920 --> 00:17:57,437 like Open Claw come from open source? 319 00:17:57,520 --> 00:18:01,837 because it's certainly that kind of user experience is not I would imagine this 320 00:18:01,920 --> 00:18:06,637 is something that you knew would be a thing and yet I think part of the the 321 00:18:06,720 --> 00:18:11,277 magic of open claw is that it would be very very difficult for a large tech 322 00:18:11,360 --> 00:18:13,597 company >> Peter didn't make many phone calls to 323 00:18:13,680 --> 00:18:18,397 hyperscalers to say hey I'm going to be integrating your API >> it just went 324 00:18:18,480 --> 00:18:22,477 >> and you guys you know and and when I think back on like the sky acquisition 325 00:18:22,560 --> 00:18:26,157 this this kind of experience was probably very top of mind and things 326 00:18:26,240 --> 00:18:28,477 that you're working towards internally. 327 00:18:28,560 --> 00:18:31,677 >> I love the spirit of everything about OpenClaw. 328 00:18:31,760 --> 00:18:33,837 >> Yeah. >> And you are totally right that it's much 329 00:18:33,920 --> 00:18:37,917 easier to imagine a one person open source project doing something like that 330 00:18:38,000 --> 00:18:40,637 than a company who is going to be afraid 331 00:18:40,720 --> 00:18:43,117 of lawsuits and data privacy and 332 00:18:43,200 --> 00:18:47,677 everything else. Um you know there like I think this is kind of how innovation 333 00:18:47,760 --> 00:18:51,357 works there. Something like that starts is clearly amazing. There will be a way 334 00:18:51,440 --> 00:18:55,117 to make a mass market version of that product, but letting the builders build 335 00:18:55,200 --> 00:18:59,197 the equivalent of the homerew homebrew computer club spirit go here is so 336 00:18:59,280 --> 00:19:02,557 important. >> Yeah, totally. Uh, can we switch to 337 00:19:02,640 --> 00:19:05,677 social? I feel like if I Google Sam 338 00:19:05,760 --> 00:19:09,037 Alman social, I get pure AI and Sora and 339 00:19:09,120 --> 00:19:11,357 then also demand or or predictions about 340 00:19:11,440 --> 00:19:14,477 a humanly social network. Where do you 341 00:19:14,560 --> 00:19:18,717 see social going broadly? How do you want to integrate it with it and power 342 00:19:18,800 --> 00:19:20,637 it in the future? 343 00:19:20,720 --> 00:19:23,597 I the mold book thing was like a very 344 00:19:23,680 --> 00:19:25,917 interesting social experiment to watch. 345 00:19:26,000 --> 00:19:29,677 Uh and I think points to 346 00:19:29,760 --> 00:19:31,837 agents interacting in some sort of 347 00:19:31,920 --> 00:19:37,197 social space hopefully on behalf of >> people at least in some degrees uh could 348 00:19:37,280 --> 00:19:40,797 be quite interesting. I don't think we know what to do there yet but it feels 349 00:19:40,880 --> 00:19:43,437 like social is going to change 350 00:19:43,520 --> 00:19:48,637 >> a lot. Um, and I am interested in the space of what a social experience can 351 00:19:48,720 --> 00:19:52,477 look like when you know your agent is talking to my agent and coming up with 352 00:19:52,560 --> 00:19:54,957 new stuff. Um, clearly putting like a 353 00:19:55,040 --> 00:19:59,597 lot of AI bots on the existing social platforms is just making everyone crazy 354 00:19:59,680 --> 00:20:01,917 and not that fun. So that's not the right answer. Yeah. 355 00:20:02,000 --> 00:20:05,837 >> But >> I think we can design something new for 356 00:20:05,920 --> 00:20:10,317 what this technology is capable of that will feel good and useful. 357 00:20:10,400 --> 00:20:14,717 >> Yeah. Is there is there a solution to the bot problem that's just all the labs 358 00:20:14,800 --> 00:20:18,877 sort of integrating with all the other platforms and and even if you can't 359 00:20:18,960 --> 00:20:23,277 detect it's AI generated you can literally say we just we just generated 360 00:20:23,360 --> 00:20:26,157 those tokens like those exact tokens are in our database 361 00:20:26,240 --> 00:20:29,837 >> you can't do that because their open source models are like good enough to 362 00:20:29,920 --> 00:20:33,037 write at this point >> um 363 00:20:33,120 --> 00:20:37,277 >> I am excited about sort of like assertion of humanity 364 00:20:37,360 --> 00:20:41,437 >> instead of trying to like detection of AI as a as a thing here. I don't know if 365 00:20:41,520 --> 00:20:44,877 the social platforms it's like in their interest to solve this. It is cause it's 366 00:20:44,960 --> 00:20:47,677 creating like at least in the short term it creates like a lot of engagement and 367 00:20:47,760 --> 00:20:52,077 increased usage. So, um I believe they could solve it if they wanted to. I'm 368 00:20:52,160 --> 00:20:55,917 not sure it's in their interest. I'm actually not even like it is I don't 369 00:20:56,000 --> 00:20:59,437 like it, but it is some people do seem to enjoy it. 370 00:20:59,520 --> 00:21:01,277 >> Yeah. Can you talk a little bit more 371 00:21:01,360 --> 00:21:04,797 about where Sora as as a video 372 00:21:04,880 --> 00:21:07,837 generation model is going? It feels like 373 00:21:07,920 --> 00:21:11,277 tool use is maybe under discussed. You 374 00:21:11,360 --> 00:21:13,437 know, adding adding reasoning. It's not 375 00:21:13,520 --> 00:21:16,237 just the diffusion model. It's giving 376 00:21:16,320 --> 00:21:21,197 these models the ability to make linear cuts and overlay motion graphics. And 377 00:21:21,280 --> 00:21:23,197 when I scroll, the Instagram reels that 378 00:21:23,280 --> 00:21:28,317 I see, they're like vibe reels with cut cutouts and it flips negative and it's 379 00:21:28,400 --> 00:21:30,957 all color graded and it's stuff that like, yeah, you could probably diffuse 380 00:21:31,040 --> 00:21:35,837 it all, but it's pretty cool just to teach a model to also use After Effects 381 00:21:35,920 --> 00:21:39,917 or whatever video, you know, m motion graphics suite you want to use. Uh, is 382 00:21:40,000 --> 00:21:43,197 that an interesting unlock? What do you see going on? 383 00:21:43,280 --> 00:21:47,517 >> Um, so all of that stuff will happen and I agree with you, the models will get 384 00:21:47,600 --> 00:21:52,077 really great at doing that. People love generating videos. I would say people, 385 00:21:52,160 --> 00:21:55,677 we have not yet found a way that people really love watching other people's 386 00:21:55,760 --> 00:21:59,117 videos. This is true for a lot of other AI like they love to, 387 00:21:59,200 --> 00:22:02,077 >> you know, people love talking to chatbt or whatever. 388 00:22:02,160 --> 00:22:03,677 >> It's not that compelling for most people 389 00:22:03,760 --> 00:22:06,077 to like read other people's generation. 390 00:22:06,160 --> 00:22:07,117 So, I think there is something. 391 00:22:07,200 --> 00:22:10,877 >> But isn't that for all writing and all video? 392 00:22:10,960 --> 00:22:14,477 >> Yeah, it seems stronger to me in this case than the general case. But maybe 393 00:22:14,560 --> 00:22:16,877 you're right. Maybe this is the >> Yeah, but if but if if somebody says, 394 00:22:16,960 --> 00:22:20,957 "Hey, I generated a 15-minute video. I I'm really excited for you to watch it." 395 00:22:21,040 --> 00:22:25,357 And you watch the first 10 seconds and you're not that captured by it. I don't 396 00:22:25,440 --> 00:22:27,517 care that it was humanmade. I don't necessarily 397 00:22:27,600 --> 00:22:31,197 >> Yeah, may may right and it this is it's not a it's not a special case. 398 00:22:31,280 --> 00:22:34,317 >> Do you see that in the data with uh with Sora downloads? because I've noticed 399 00:22:34,400 --> 00:22:38,237 that uh I'll generate stuff on Sora, download it and share it to a group chat 400 00:22:38,320 --> 00:22:42,877 and then it's this little in joke that me and five other people get uh and we 401 00:22:42,960 --> 00:22:46,077 see this like family group chats of oh it's our dog and our kids but like 402 00:22:46,160 --> 00:22:49,757 there's not really like okay yeah I this is a business you know everyone likes 403 00:22:49,840 --> 00:22:53,357 this >> absolutely that I would say that the 404 00:22:53,440 --> 00:22:57,117 most common use case is something like that you know like memes on group chats 405 00:22:57,200 --> 00:23:01,757 is a real killer use case of sort >> how is the Disney rollout going I was 406 00:23:01,840 --> 00:23:05,437 super excited about it Jordy was extreme extremely bullish on it from a business 407 00:23:05,520 --> 00:23:07,917 strategy perspective. >> Yeah. When you look at how image models 408 00:23:08,000 --> 00:23:13,197 have grown various LLMs historically and now you're going to have an image image 409 00:23:13,280 --> 00:23:16,717 and video model that can do something that no other LLM can do at least 410 00:23:16,800 --> 00:23:17,437 legally. 411 00:23:17,520 --> 00:23:20,717 >> Yeah. And is Bob Iger joining OpenAI? 412 00:23:20,800 --> 00:23:24,797 >> Uh Bob Iger I would love it. No credit for that. 413 00:23:24,880 --> 00:23:28,157 >> He's going to be looking for a job and free a year. He's a free agent. Pick him 414 00:23:28,240 --> 00:23:29,517 up. >> Hit him up for us. You know, do some 415 00:23:29,600 --> 00:23:31,437 recruiting. That'd be that'd be great. 416 00:23:31,520 --> 00:23:36,397 Um the I think that generating 417 00:23:36,480 --> 00:23:40,077 >> characters in images and videos is going 418 00:23:40,160 --> 00:23:43,037 to be very important to people and they really like that. 419 00:23:43,120 --> 00:23:46,797 >> Uh like other like we were saying otherwise I don't think many people like 420 00:23:46,880 --> 00:23:49,597 want to watch me and >> some Star Wars character doing something 421 00:23:49,680 --> 00:23:50,717 together but I might think it's cool. 422 00:23:50,800 --> 00:23:53,597 >> Yeah. Um, it's, you know, there there's like a real trend going on right now 423 00:23:53,680 --> 00:23:57,837 with chat GPT where it's make a caricature of me and my job based off of 424 00:23:57,920 --> 00:23:59,197 everything you know about me. Yeah. 425 00:23:59,280 --> 00:24:02,957 >> And those kinds of things people actually do like looking at other people's 426 00:24:03,040 --> 00:24:07,757 >> Yeah. >> Um, media. >> Yeah. It's almost like a like a face 427 00:24:07,840 --> 00:24:10,557 filter or something. There's like enough of it's the studio Giblly moment. Like 428 00:24:10,640 --> 00:24:14,877 there's enough of the human still in there that it's not it's not you can't 429 00:24:14,960 --> 00:24:18,477 yours is not the same as mine. So, it's still personalized. It's personalized 430 00:24:18,560 --> 00:24:22,717 and it says something about you and you know the the the like a lot of these 431 00:24:22,800 --> 00:24:26,157 things. >> A lot of what's gone viral before with 432 00:24:26,240 --> 00:24:29,677 image gen. I think it's like if you can make people look a little bit more 433 00:24:29,760 --> 00:24:32,317 attractive or cool than they look in than they are in real life. 434 00:24:32,400 --> 00:24:34,717 >> Yeah. >> Without sort of having to ask for that. 435 00:24:34,800 --> 00:24:36,637 >> Yeah. >> How are you thinking about the actual 436 00:24:36,720 --> 00:24:39,277 the actual roll out? We were we were debating between like open the 437 00:24:39,360 --> 00:24:43,357 floodgates, you can generate any Disney property versus like it's it's 438 00:24:43,440 --> 00:24:47,277 Spider-Man week and everyone's posting Spider-Man and then it's it's you know 439 00:24:47,360 --> 00:24:50,397 Mickey Mouse week and there's another viral moment. 440 00:24:50,480 --> 00:24:53,357 >> I'm not sure what the team is planning there. I know Disney's had some 441 00:24:53,440 --> 00:24:55,917 different opinions about how they what they want to do and we'll try to be a 442 00:24:56,000 --> 00:24:58,717 good partner there. But I'd be excited to open the floodgates personally. 443 00:24:58,800 --> 00:25:01,677 >> Oh, that'd be fun. Cool. Uh talk about 444 00:25:01,760 --> 00:25:06,477 your first speaking of video your first Super Bowl ad. Uh, it felt like not 445 00:25:06,560 --> 00:25:10,717 generated, lots of motion graphics, the black dots coming together, like what 446 00:25:10,800 --> 00:25:13,357 was the what was the goal with that ad? 447 00:25:13,440 --> 00:25:17,437 Who were you trying to speak to? It didn't feel like a direct response QR 448 00:25:17,520 --> 00:25:20,877 code, download the app. No. Uh, what was the mission? 449 00:25:20,960 --> 00:25:23,677 >> Um, I love that ad. Uh, I think that was 450 00:25:23,760 --> 00:25:27,917 such a cool one. It was it clearly not meant to be like a mass market or direct 451 00:25:28,000 --> 00:25:28,877 response ad. 452 00:25:28,960 --> 00:25:33,037 >> Yeah. But you know speaking to the like 453 00:25:33,120 --> 00:25:37,357 people who are at the center of this revolution and just trying to like 454 00:25:37,440 --> 00:25:40,717 celebrate everything that has come before and everything that will come after. 455 00:25:40,800 --> 00:25:45,197 >> Yeah. >> Um it we didn't hear a lot about it from 456 00:25:45,280 --> 00:25:48,237 like average users of chatbt but we heard a lot about it from like 457 00:25:48,320 --> 00:25:52,237 researchers in the field and a lot of resonance there. Um it was definitely 458 00:25:52,320 --> 00:25:55,517 not generated. It was done the oldfashioned way. Uh, and you know, it 459 00:25:55,600 --> 00:25:59,517 had like a lot of people loved it, a lot of people hated it, and then many people 460 00:25:59,600 --> 00:26:03,117 in the middle didn't get it, and I felt okay about that. Um, 461 00:26:03,200 --> 00:26:05,517 >> it's a great example. >> I like our ad for Sunday. 462 00:26:05,600 --> 00:26:08,637 >> Okay. >> Uh, it's about Codex, no surprise. But 463 00:26:08,720 --> 00:26:10,957 >> yeah, talk about the evolution of the 464 00:26:11,040 --> 00:26:15,117 advertising to be more uh just clear 465 00:26:15,200 --> 00:26:18,797 about the actual use case, the value, like what are you trying to say with 466 00:26:18,880 --> 00:26:20,557 your advertising strategy now as it 467 00:26:20,640 --> 00:26:23,437 refers as it relates to like video? 468 00:26:23,520 --> 00:26:26,520 Um 469 00:26:27,760 --> 00:26:31,357 well the thing I would most like us to say and I think this is a new challenge 470 00:26:31,440 --> 00:26:37,357 given where the models are is to teach people what they can go do with 471 00:26:37,440 --> 00:26:40,557 >> AI. Yeah. >> I think AI is now unbelievably capable 472 00:26:40,640 --> 00:26:45,357 in most of the world. It's still like asking it basic questions on Chad GBT. 473 00:26:45,440 --> 00:26:48,157 >> Everyone can go build amazing things now. Everyone can go do all kinds of 474 00:26:48,240 --> 00:26:52,557 work. Scientists are going to make new discoveries. Um, and I'd like to, you 475 00:26:52,640 --> 00:26:56,717 know, to the degree that advertising we do can teach people how to use this, I I 476 00:26:56,800 --> 00:26:59,197 think that'd be awesome. >> Yeah. So, the KPI is like reduce the 477 00:26:59,280 --> 00:27:02,557 capability overhang broadly. 478 00:27:02,640 --> 00:27:06,717 >> I think that's should be a general KPI for us, not just of our ads. Yeah. 479 00:27:06,800 --> 00:27:10,077 >> Um, the products that we build, how we teach people to use those products, like 480 00:27:10,160 --> 00:27:13,277 the the >> that feels very important. 481 00:27:13,360 --> 00:27:17,437 >> Yeah. Uh, Anthropic also has a bunch of ads in the Super Bowl. Seems like run a 482 00:27:17,520 --> 00:27:20,077 ton. What do you think that they're 483 00:27:20,160 --> 00:27:23,197 getting wrong about uh about their 484 00:27:23,280 --> 00:27:25,517 characterization of how ads will roll 485 00:27:25,600 --> 00:27:29,357 out in chat apps? 486 00:27:29,440 --> 00:27:32,237 >> Uh well, it's just wrong. Um like the 487 00:27:32,320 --> 00:27:37,517 the the main thing that I think is we are not stupid. We respect our users. We 488 00:27:37,600 --> 00:27:39,117 understand that if we did something like 489 00:27:39,200 --> 00:27:41,117 what those ads depict, people would 490 00:27:41,200 --> 00:27:45,277 rightfully stop using the product. Um, 491 00:27:45,360 --> 00:27:48,157 no one like our our first principle with 492 00:27:48,240 --> 00:27:49,757 ads is that we're not going to put stuff 493 00:27:49,840 --> 00:27:53,037 into the LLM stream. Um, that would feel 494 00:27:53,120 --> 00:27:56,157 crazy dystopic like bad sci-fi movie. 495 00:27:56,240 --> 00:27:57,597 Um, so the main thing that's wrong with 496 00:27:57,680 --> 00:28:00,477 the ads is like using a deceptive ad to 497 00:28:00,560 --> 00:28:05,517 criticize deceptive ads feels I don't know, something doesn't sit 498 00:28:05,600 --> 00:28:07,277 right with me about that. I asked I 499 00:28:07,360 --> 00:28:11,997 asked Claude uh what uh if if what the 500 00:28:12,080 --> 00:28:16,557 definition of playing dirty >> and it said uh what did it say? 501 00:28:16,640 --> 00:28:20,317 Misleading others about your intentions, hiding information, or creating false 502 00:28:20,400 --> 00:28:24,957 impressions. >> Yeah. >> Thought it was a little dirty. I thought 503 00:28:25,040 --> 00:28:28,477 it was well played, but it was it was >> it was it was well played for sure. And 504 00:28:28,560 --> 00:28:33,437 it was a funny ad and and they you know like the sort of 505 00:28:33,520 --> 00:28:35,357 >> the stuff about the chat GBT personality 506 00:28:35,440 --> 00:28:39,037 that most annoys me which we'll fix very soon. I thought they nailed in the ad. 507 00:28:39,120 --> 00:28:40,717 So that part was funny. >> Yeah. 508 00:28:40,800 --> 00:28:45,437 >> Um >> but I don't know you know like I also 509 00:28:45,520 --> 00:28:48,317 like I think it's great for them not to do ads. We have a different shaped 510 00:28:48,400 --> 00:28:51,917 business. I did notice that they said in their thing like we may later revise 511 00:28:52,000 --> 00:28:55,197 this decision and we'll explain why. So, um, 512 00:28:55,280 --> 00:28:57,997 >> yeah, the the blog post the blog post 513 00:28:58,080 --> 00:29:03,277 was uh kind of did a good job of disarming the pro ad people gave 514 00:29:03,360 --> 00:29:06,557 themselves an out >> in the future. 515 00:29:06,640 --> 00:29:07,997 >> Uh, do you think they care? 516 00:29:08,080 --> 00:29:11,117 >> It doesn't matter. Like I think it's a sideeshow. You know, people are excited 517 00:29:11,200 --> 00:29:13,917 for a food fight and uh between 518 00:29:14,000 --> 00:29:17,597 companies, but like the 519 00:29:17,680 --> 00:29:22,317 the amazing capabilities of these models, the product, the kind of like 520 00:29:22,400 --> 00:29:26,637 the ground swell of excitement around codecs, that feels way more important to 521 00:29:26,720 --> 00:29:29,437 me. >> How do you stop the pausing that happens 522 00:29:29,520 --> 00:29:32,797 in voice mode? Do you need new hardware for that or is it a model capability 523 00:29:32,880 --> 00:29:34,717 thing? >> We need new model. Uh we may need some 524 00:29:34,800 --> 00:29:38,477 new hardware too, but mostly we just need a new model. I think we will have a 525 00:29:38,560 --> 00:29:42,237 great voice mode. >> Yeah. >> By the end of this year. 526 00:29:42,320 --> 00:29:45,677 >> What's the bigger bottleneck? Energy or chips? 527 00:29:45,760 --> 00:29:48,237 >> It goes back and forth. Uh right now again it's chips. 528 00:29:48,320 --> 00:29:51,357 >> Chips. Is there anything >> but it it you know be different different times. 529 00:29:51,440 --> 00:29:54,637 >> Is there anything we like society America should be doing more 530 00:29:54,720 --> 00:29:58,557 aggressively to increase the supply of fabs? 531 00:29:58,640 --> 00:30:04,520 >> Um >> or fab. >> Yeah. I think it is 532 00:30:04,880 --> 00:30:07,277 well it may get solved on its own. It 533 00:30:07,360 --> 00:30:11,997 may like normal capitalism may solve it but I think somehow deciding as a 534 00:30:12,080 --> 00:30:15,517 society that we are going to increase the wafer capacity of the world and 535 00:30:15,600 --> 00:30:19,837 we're going to fund that uh and we're going to get you know the whole supply 536 00:30:19,920 --> 00:30:22,477 chain and the talented people we need to make that happen would be a very good 537 00:30:22,560 --> 00:30:24,957 thing to do. >> Do you think there's an upper bound on 538 00:30:25,040 --> 00:30:27,917 model IQ? Like the the race right now is 539 00:30:28,000 --> 00:30:30,797 is you're smart, but you're not smart 540 00:30:30,880 --> 00:30:33,917 for days. You're smart for hours. Can 541 00:30:34,000 --> 00:30:37,917 you go much farther and get much smarter? 542 00:30:38,000 --> 00:30:41,437 >> It seems certain. >> Upper bound. I don't know. Uh I don't 543 00:30:41,520 --> 00:30:43,917 know how to think about that question yet. 544 00:30:44,000 --> 00:30:47,517 >> Yeah. I can't even reason about what 2,000 IQ looks like, you know? Like I 545 00:30:47,600 --> 00:30:50,157 don't even I don't know what that means. 546 00:30:50,240 --> 00:30:53,597 >> It's funny you say. I mean, I I I can't reason about what it means to think 547 00:30:53,680 --> 00:30:57,917 about a problem for like 10,000 human years. That's another good one. Yeah. 548 00:30:58,000 --> 00:31:02,557 >> Yeah, >> that's crazy. >> But maybe IQ is going to feel even even weirder. I I 549 00:31:02,640 --> 00:31:06,477 >> Yeah, >> I don't know. I feel like 550 00:31:06,560 --> 00:31:10,397 >> I somehow feel like this isn't going to feel as strange as it sounds. 551 00:31:10,480 --> 00:31:13,357 >> And the like for a bunch of reasons. 552 00:31:13,440 --> 00:31:16,477 We're so focused on other people. We're so focused on our own lives. We're so 553 00:31:16,560 --> 00:31:20,557 focused. We have such a humanentric nature that like, okay, this thing is 554 00:31:20,640 --> 00:31:23,437 really smart. It's inventing new science for us. It's running companies for us. 555 00:31:23,520 --> 00:31:27,517 It's doing all this stuff. And 556 00:31:27,600 --> 00:31:30,477 that sounds like it should be impossibly weird and I think it will just be very 557 00:31:30,560 --> 00:31:33,517 weird. >> Do you think space data centers will 558 00:31:33,600 --> 00:31:38,397 provide a meaningful amount of compute for open AAI in the next two to three 559 00:31:38,480 --> 00:31:42,957 years, five years? >> No. >> 10 years. 560 00:31:43,040 --> 00:31:43,997 >> Um, 561 00:31:44,080 --> 00:31:47,597 >> you just keep going 10,000 years. 562 00:31:47,680 --> 00:31:50,717 >> I I wish Ethan luck. 563 00:31:50,800 --> 00:31:53,517 >> Okay. Uh the the the funny thing about 564 00:31:53,600 --> 00:31:57,837 the whole like back and forth about ads is that uh in in our world the criticism 565 00:31:57,920 --> 00:32:02,077 is that you didn't launch ads early enough. Is there a world where you wish 566 00:32:02,160 --> 00:32:04,157 you launched earlier? How is the actual 567 00:32:04,240 --> 00:32:07,597 roll out going? Are advertisers happy? 568 00:32:07,680 --> 00:32:10,957 Do you have like a really long roadmap or do you think you'll be faster at 569 00:32:11,040 --> 00:32:13,677 catching up to sort of what's frontier and ad? 570 00:32:13,760 --> 00:32:17,997 >> We haven't even started the test yet. We start the test soon. Um but you know 571 00:32:18,080 --> 00:32:21,677 we're going to it's going to take us some number of iterations to figure out 572 00:32:21,760 --> 00:32:25,677 the right ad unit the right kind of the right way this all works. Um do I wish 573 00:32:25,760 --> 00:32:28,317 we had started earlier? 574 00:32:28,400 --> 00:32:32,797 >> We have gone from like not a company you know 3 years and 3 months ago or 575 00:32:32,880 --> 00:32:34,877 something like that. We were like a research lab 576 00:32:34,960 --> 00:32:37,597 >> and now we are like a pretty big company 577 00:32:37,680 --> 00:32:41,597 with a lot of products. So there's many things I wish we had done faster. I 578 00:32:41,680 --> 00:32:44,717 think we were correct on the trade-off here of how we balance things that we 579 00:32:44,800 --> 00:32:47,917 need to do. Um I wish you know we launched this very cool enterprise 580 00:32:48,000 --> 00:32:50,957 platform this morning. >> Wish we had done that earlier too but 581 00:32:51,040 --> 00:32:54,477 had to like deal with the monstrous growth of CHPT and Codex and all other stuff. 582 00:32:54,560 --> 00:32:56,397 >> Good problems to have. Last question uh 583 00:32:56,480 --> 00:32:59,357 for me uh what happened to that internal 584 00:32:59,440 --> 00:33:01,517 writing model that you used to write the 585 00:33:01,600 --> 00:33:04,317 essay. Uh that feels like something that 586 00:33:04,400 --> 00:33:08,717 was really cool but we never really saw the light of day. 587 00:33:08,800 --> 00:33:13,037 >> Uh we're going to get a lot of that spirit into a future model. Again, it's 588 00:33:13,120 --> 00:33:16,477 like >> there's so much stuff happening. 589 00:33:16,560 --> 00:33:19,197 >> Yeah, >> we have to make these hard prioritization decisions. 590 00:33:19,280 --> 00:33:21,757 >> Sure. >> Um I would love a cool writing model, 591 00:33:21,840 --> 00:33:24,157 >> not as much as I would love a cool coding model. 592 00:33:24,240 --> 00:33:28,237 >> Um and it's what is possible now for 593 00:33:28,320 --> 00:33:31,517 coding for science. That's the thing I'm most excited about. 594 00:33:31,600 --> 00:33:35,917 >> Um for accelerating all kinds of research AI and otherwise for really 595 00:33:36,000 --> 00:33:40,637 accelerating the economy. Uh I think that's like the right thing for us to 596 00:33:40,720 --> 00:33:45,757 most prioritize in terms of new capabilities. But yeah, well like you 597 00:33:45,840 --> 00:33:49,837 want you want a model that can write beautifully because it means it well you 598 00:33:49,920 --> 00:33:53,197 want to write a model that can write beautifully if it can also think very 599 00:33:53,280 --> 00:33:54,717 clearly and express that very clearly. 600 00:33:54,800 --> 00:33:56,317 That's just useful in normal work. 601 00:33:56,400 --> 00:33:57,597 >> Yeah, that makes a ton of sense. 602 00:33:57,680 --> 00:34:00,157 >> Last question for me. How how have 603 00:34:00,240 --> 00:34:03,357 conversations been with the broad open 604 00:34:03,440 --> 00:34:07,197 AI leadership team? You guys are in a position where any single word or 605 00:34:07,280 --> 00:34:09,517 sentence you say in any situation can be 606 00:34:09,600 --> 00:34:14,877 spun into a headline immediately and then you guys have to go on damage 607 00:34:14,960 --> 00:34:17,917 control kind of correcting the narrative. But of course the original 608 00:34:18,000 --> 00:34:21,997 message is often or at least the original news is often seen more broadly 609 00:34:22,080 --> 00:34:23,677 than the than the correction and it 610 00:34:23,760 --> 00:34:25,677 seems like an interesting challenge. 611 00:34:25,760 --> 00:34:29,277 >> It it is a strange way to live. Um it 612 00:34:29,360 --> 00:34:31,837 and I don't like I don't know of any 613 00:34:31,920 --> 00:34:34,637 private company that has ever like 614 00:34:34,720 --> 00:34:36,797 >> been so in the news and so under a 615 00:34:36,880 --> 00:34:39,917 microscope and it you know at some level 616 00:34:40,000 --> 00:34:43,997 it's frustrating and you know we're so 617 00:34:44,080 --> 00:34:47,597 squarely in the sights of everybody's anxieties and every competitor trying to 618 00:34:47,680 --> 00:34:51,677 take us down and everybody's like just what is going to happen with AI to their 619 00:34:51,760 --> 00:34:56,960 part of the business or their own lives um that there's like a lot 620 00:34:58,160 --> 00:34:59,757 plasma looking for an instability to 621 00:34:59,840 --> 00:35:02,477 collapse on. Um in some other sense 622 00:35:02,560 --> 00:35:05,917 though the the subjective experience of 623 00:35:06,000 --> 00:35:10,557 it is we are so busy on so much exciting stuff that it often feels like there's 624 00:35:10,640 --> 00:35:12,877 this in crazy hurricane turning around 625 00:35:12,960 --> 00:35:15,997 us and when we sit it's like fairly calm 626 00:35:16,080 --> 00:35:20,077 you know the media or Twitter goes insane about something one day they're 627 00:35:20,160 --> 00:35:23,837 talking about a crazy meltdown we're like that is insane like okay and people 628 00:35:23,920 --> 00:35:27,597 talk about it all day and then later find out it's wrong and sort of seemed 629 00:35:27,680 --> 00:35:30,797 like a lot of wasted energy but we're just like we have this great new model 630 00:35:30,880 --> 00:35:34,877 coming. People are building incredible stuff. Companies are transforming. We're 631 00:35:34,960 --> 00:35:37,437 trying to like figure out how to get more compute and deal with this compute 632 00:35:37,520 --> 00:35:41,917 crunch. And we just kind of like keep going and we're busy. And then if we 633 00:35:42,000 --> 00:35:43,917 like open Twitter, pop up our heads and 634 00:35:44,000 --> 00:35:45,997 look at the news, it's like, wow, that 635 00:35:46,080 --> 00:35:49,037 is an insane crazy thing happening. 636 00:35:49,120 --> 00:35:53,677 Completely divorced from reality or 99% divorced from reality. And like, okay, 637 00:35:53,760 --> 00:35:55,037 someone will correct it. But then we get 638 00:35:55,120 --> 00:35:57,357 back to work and people flip out again. 639 00:35:57,440 --> 00:36:01,037 It's it is much it is weird to watch 640 00:36:01,120 --> 00:36:04,957 when we look outside, but it is uh 641 00:36:05,040 --> 00:36:09,277 >> it is less chaotic internally than I think you would imagine from reading the 642 00:36:09,360 --> 00:36:12,397 media reports. >> Yep, that makes sense. Well, thank you 643 00:36:12,480 --> 00:36:16,237 so much for taking the time to come chat with us. Congrats on the 644 00:36:16,320 --> 00:36:18,557 >> I'm excited to see the Codeex ad. 645 00:36:18,640 --> 00:36:23,037 >> Please try it. The It is the app and 53 have been like uh I think the coolest 646 00:36:23,120 --> 00:36:25,357 thing we've done in a while. >> Yes. With one prompt, I rebuilt the 647 00:36:25,440 --> 00:36:30,397 tbpn.com homepage to look exactly like uh Bergkshire Hathaway. And it was just 648 00:36:30,480 --> 00:36:33,197 immediate. It was very fun. >> Interesting choice. 649 00:36:33,280 --> 00:36:35,037 >> Plain text. It was very easy. 650 00:36:35,120 --> 00:36:37,757 Immediately oneshot it. Did not really push it to its limits. But I'm having 651 00:36:37,840 --> 00:36:40,317 fun. So, thank you so much for coming on the show. We'll talk to you soon. 652 00:36:40,400 --> 00:36:41,037 >> Thank you. Great to catch up. 653 00:36:41,120 --> 00:36:43,840 >> Goodbye. lectures.