1 00:00:00,000 --> 00:00:05,445 Sam Altman now joins us live on Bloomberg Television from Fort Mason in San 2 00:00:05,528 --> 00:00:08,640 Francisco with BTEC host Ed Ludlow. 3 00:00:11,840 --> 00:00:16,985 Thank you, Romaine. And we'll start with that transition to an always on agent, 4 00:00:17,068 --> 00:00:17,902 Dots. Sam, 5 00:00:18,120 --> 00:00:22,691 do you know what's interesting in the presentation is how internally open AI 6 00:00:22,774 --> 00:00:25,960 staff talk to Dots? Do you get what I mean? 7 00:00:26,290 --> 00:00:27,810 I think I do, but tell me. 8 00:00:28,670 --> 00:00:32,516 It isn't treating it like AI, like, hey, chat GPT. Yeah. Question. 9 00:00:32,599 --> 00:00:33,707 It's a conversation. 10 00:00:33,790 --> 00:00:37,375 Yeah. Yeah. Yeah. It's not I don't think it's, like, anthropomorphized. 11 00:00:37,458 --> 00:00:38,667 They don't we don't quite 12 00:00:38,750 --> 00:00:42,107 talk to it like a person, but there is a different style to it than traditional 13 00:00:42,190 --> 00:00:46,230 chat GPT. The reason I wanted to catch that first who's Dots for, really? 14 00:00:46,330 --> 00:00:49,837 It started as a premium product. We know it's you you know, we're putting our best 15 00:00:49,920 --> 00:00:51,117 model Astra in this. Premium 16 00:00:51,200 --> 00:00:55,117 kinda being prosumer. Sorry to interrupt. Or or it's an enterprise product? 17 00:00:55,200 --> 00:00:58,812 More like more like for getting work done, I think think is largely enterprise, 18 00:00:58,895 --> 00:00:59,327 but there's 19 00:00:59,410 --> 00:01:01,727 a lot of people that are gonna build their own startups that are sort of in 20 00:01:01,810 --> 00:01:06,176 the traditional prosumer category. But this is, like, a very capable model, 21 00:01:06,259 --> 00:01:07,007 and it's a ton 22 00:01:07,090 --> 00:01:10,607 of usage. So we're starting for something at the top end of the market, and you 23 00:01:10,690 --> 00:01:13,447 will see us over time make a mass market consumer version. 24 00:01:13,530 --> 00:01:19,877 How much do you and the team, various people at OpenAI benchmark against 25 00:01:19,960 --> 00:01:21,217 the agents that 26 00:01:21,300 --> 00:01:24,727 are already out there? So how much do you use a Grock Bot, a Muse, and 27 00:01:24,810 --> 00:01:27,946 Instinct from the product perspective? Because they they've been out there 28 00:01:28,029 --> 00:01:29,907 a while, and people are actually using them. 29 00:01:29,990 --> 00:01:33,241 I've tried them. I don't use them regularly. 30 00:01:33,325 --> 00:01:35,667 I have not found a sort of place 31 00:01:35,750 --> 00:01:39,107 for them in my life, but I, you know, try to try every major new AI 32 00:01:39,190 --> 00:01:42,078 product. Product. This is a totally different category. Like, 33 00:01:42,161 --> 00:01:43,637 the amount of work, the level of 34 00:01:43,720 --> 00:01:46,287 capability, the level of intelligence this can take on, 35 00:01:46,370 --> 00:01:48,980 I think this is just a completely new thing. 36 00:01:49,080 --> 00:01:53,369 Why is it different? Like, literally, are we talking about its capabilities, 37 00:01:53,452 --> 00:01:54,117 or is this an 38 00:01:54,200 --> 00:01:57,829 economic discussion of how you've priced it or where you've made it available 39 00:01:57,912 --> 00:01:58,797 in existing tiers? 40 00:01:58,880 --> 00:02:03,037 Well, I mean, it is priced higher than other things in the market, but the reason 41 00:02:03,120 --> 00:02:08,240 for that is, is because of the capability and the amount of compute worth putting 42 00:02:08,323 --> 00:02:11,207 into this, which is why I think it's useful in a new kind of way at a 43 00:02:11,290 --> 00:02:16,727 new level. You've launched, GBT 6.1 Soul at, like, Astra capability. 44 00:02:16,810 --> 00:02:17,857 And a fifth of the price. 45 00:02:17,940 --> 00:02:20,980 But the price is so different. And, like, I've been trying to understand, like, 46 00:02:21,063 --> 00:02:21,257 what is 47 00:02:21,340 --> 00:02:26,243 the big picture story? It seems like we're moving towards cheaper inference at a time 48 00:02:26,327 --> 00:02:26,537 where 49 00:02:26,620 --> 00:02:32,057 everyone's much more willing to pay for it. Well How did you arrive at the tiers 50 00:02:32,140 --> 00:02:32,657 that you've done? 51 00:02:32,740 --> 00:02:37,440 I mean, we've talked for a long time about our goal is to make increasingly better 52 00:02:37,530 --> 00:02:41,136 intelligence and to drive the price down and down. And the reason, 53 00:02:41,220 --> 00:02:42,097 that I think both 54 00:02:42,180 --> 00:02:45,767 of those are true is the better you make AI, the more people wanna use it. 55 00:02:45,850 --> 00:02:50,917 So people are clearly willing to spend more on AI than ever before, but they they 56 00:02:51,000 --> 00:02:54,437 wanna do so much with it. So so our belief is that if we continue to 57 00:02:54,520 --> 00:02:58,846 drive the quality up and the price down and really lead the Pareto frontier 58 00:02:58,930 --> 00:03:04,157 at every point on that curve, people will use our tools in tremendous ways. 59 00:03:04,240 --> 00:03:05,477 And just even here 60 00:03:05,560 --> 00:03:09,142 today, like, hearing from developers about what they're building with this and what 61 00:03:09,226 --> 00:03:12,757 what they can do if we continue to drive the price down is really quite remarkable. 62 00:03:12,840 --> 00:03:15,199 What is OpenAI's business momentum right now? 63 00:03:15,282 --> 00:03:18,370 How did you end the third quarter? What does that look like? 64 00:03:18,470 --> 00:03:20,947 I don't think we, like, put our specific numbers out, but 65 00:03:21,030 --> 00:03:22,787 numbers in the newspapers this morning. 66 00:03:22,870 --> 00:03:28,947 It is our, I mean, it's like a very steep growth period for us. I think 67 00:03:29,030 --> 00:03:33,853 what people love are models. We have really been able to make new products 68 00:03:33,936 --> 00:03:34,317 and new 69 00:03:34,400 --> 00:03:39,835 tools for developers, for enterprises. ChatGPT is growing phenomenally well. 70 00:03:39,918 --> 00:03:42,947 So across the board, consumers, developers, 71 00:03:43,030 --> 00:03:46,807 enterprises, all this new stuff that we're launching at Dev Day, we see a lot of 72 00:03:46,890 --> 00:03:51,854 business momentum. The the reported number was ARR 70,000,000,000, I think, 73 00:03:51,937 --> 00:03:55,367 a three q number. I don't know Is that meaningful to you? 74 00:03:55,450 --> 00:04:00,727 I don't know where these reported leaks or reported numbers come from. But I I am 75 00:04:00,810 --> 00:04:04,185 obviously not gonna comment either way on that, but I will say, like, 76 00:04:04,268 --> 00:04:05,990 very pleased with the growth. 77 00:04:06,090 --> 00:04:10,595 The interesting thing about Dots as a product and agent, but also, 78 00:04:10,679 --> 00:04:11,930 six point one Soul 79 00:04:12,110 --> 00:04:17,867 is you made the decision to release those. This week, you held back the release of 80 00:04:17,950 --> 00:04:22,667 Astra 6.1, and last week, you paused training on a separate model. 81 00:04:22,750 --> 00:04:24,267 Yeah. Explain that how 82 00:04:24,350 --> 00:04:30,890 you come to that decision. So SOL 6.1. Okay. We held back Astra 6.1. 83 00:04:32,170 --> 00:04:34,439 We've been talking about this for a few months now, 84 00:04:34,522 --> 00:04:39,983 but our we're clearly getting into a very steep part of the capabilities curve. 85 00:04:40,067 --> 00:04:43,510 And given that our whole mission is to 86 00:04:45,430 --> 00:04:48,073 build this for people and to empower people, and part of that safety, 87 00:04:48,157 --> 00:04:48,627 part of that's 88 00:04:48,710 --> 00:04:53,773 why they're distributing the benefits, when we need to make a decision about 89 00:04:53,857 --> 00:04:54,857 slowing down our 90 00:04:54,940 --> 00:04:59,107 training or our model release so that we can make more progress and put more of 91 00:04:59,190 --> 00:05:03,242 our attention into safety, alignment, monitoring, security, 92 00:05:03,326 --> 00:05:06,467 we will happily do that. Now, of course, we're 93 00:05:06,550 --> 00:05:11,980 gonna continue to make progress, but the world should always have confidence in 94 00:05:12,064 --> 00:05:13,250 our safety cases, 95 00:05:13,410 --> 00:05:17,133 in our safety claims, and that we are going to be responsible with 96 00:05:17,217 --> 00:05:18,287 this technology. And 97 00:05:18,370 --> 00:05:21,627 and we don't wanna do this thing where, like, oh, we need, you know, regulation or 98 00:05:21,710 --> 00:05:24,667 we need our competitors to do x, y, or z, we won't be safe. We are 99 00:05:24,750 --> 00:05:28,123 going to do this responsibly. We want the world to have that confidence. 100 00:05:28,206 --> 00:05:31,893 Now there are still things that we can very reasonably release. So, like, 101 00:05:31,977 --> 00:05:33,397 a faster version of Astra or 102 00:05:33,480 --> 00:05:39,115 a cheaper version of Astra Six, we can totally do on these wonderful new 103 00:05:39,198 --> 00:05:40,227 products. But, 104 00:05:40,310 --> 00:05:44,947 you know, I think, like, I understand why some people are just like, full YOLO pedal 105 00:05:45,030 --> 00:05:48,387 to the metal go, but we're gonna we're gonna have to be responsible as we push 106 00:05:48,470 --> 00:05:49,587 towards these new levels of capabilities. 107 00:05:49,670 --> 00:05:51,414 There are some difficult questions, posed. 108 00:05:51,497 --> 00:05:53,927 I'm very conscious of the developer audience and also like 109 00:05:54,010 --> 00:05:57,457 the users of the technology. You said that the the two main things that they want 110 00:05:57,540 --> 00:06:00,476 are speed and performance, and you you released, 111 00:06:00,560 --> 00:06:03,927 new mechanisms to achieve that with existing generation 112 00:06:04,010 --> 00:06:09,092 of model. Why is that better as a technique than releasing a more capable 113 00:06:09,175 --> 00:06:13,757 model? Well, I mean, they want that too. We will also do that. The the so I 114 00:06:13,840 --> 00:06:16,447 think all of those things are good. You know, speed, price, 115 00:06:16,530 --> 00:06:17,997 and capability are certainly three 116 00:06:18,080 --> 00:06:22,727 of the things people ask for the most often. And I think the answer is all. 117 00:06:22,810 --> 00:06:26,293 People want all of them in different different use cases. That said, 118 00:06:26,376 --> 00:06:27,437 ultrafast, I think, is 119 00:06:27,520 --> 00:06:30,877 a really special thing. I did not appreciate how much I was going to like it 120 00:06:30,960 --> 00:06:33,527 and how hard it is for me to go back. I should have known because, like, 121 00:06:33,610 --> 00:06:35,607 we all want everything to be faster, but it is awesome. 122 00:06:35,690 --> 00:06:40,090 Is it an economic proposition for OpenAI? Are the economics a bit good for you? 123 00:06:40,190 --> 00:06:45,627 Well, it's much more expensive to run, but we do charge more for it too. So 124 00:06:45,710 --> 00:06:48,267 we tried to make it something that, you know, it's like a 125 00:06:48,350 --> 00:06:50,597 We can talk to Sarah about the balance between 126 00:06:50,680 --> 00:06:54,747 Yeah. I mean, it's like, you know, I I think the market will decide, but I 127 00:06:54,830 --> 00:06:57,477 think people will be excited about it at this price, and it's reasonable for us. 128 00:06:57,560 --> 00:07:02,187 You you were just asked again about an IPO, and what you said is you want 129 00:07:02,270 --> 00:07:04,717 to be public one day. Yes. But, essentially, 130 00:07:04,800 --> 00:07:07,017 the duties of being a public company are 131 00:07:07,100 --> 00:07:10,620 incompatible with a dedication or focus on safety. Is that 132 00:07:10,780 --> 00:07:17,177 Not quite. To be a newly public company and for us to understand how to live 133 00:07:17,260 --> 00:07:23,011 life as a public company with, you know, pressure from investors and sort of just 134 00:07:23,094 --> 00:07:23,857 the 135 00:07:23,940 --> 00:07:27,010 all of the additional weight that comes with that while we are going through 136 00:07:27,094 --> 00:07:30,162 this period of adjusting to this new level of capability, 137 00:07:30,245 --> 00:07:32,797 and the new safety requirements for that. We 138 00:07:32,880 --> 00:07:37,597 just wanna get our feet under us, make sure we understand how to operate in this 139 00:07:37,680 --> 00:07:40,637 new way, be able to make some of these decisions in front of us without the 140 00:07:40,720 --> 00:07:45,537 pressure of being a newly public company. And then I think we'll be public at some 141 00:07:45,620 --> 00:07:48,727 point. There will be more new levels that come, but but it won't be all at 142 00:07:48,810 --> 00:07:49,537 the same time. 143 00:07:49,620 --> 00:07:55,217 Do you feel support from the existing cap table to just hold off 2028? Does it 144 00:07:55,300 --> 00:07:59,597 matter? Our investors seem, like, very happy with us and very patient. 145 00:07:59,680 --> 00:08:04,077 Let me ask you this to conclude. It's very rare people have said for you and 146 00:08:04,160 --> 00:08:07,019 Jerome Powell and Elon Musk to agree on something, 147 00:08:07,103 --> 00:08:09,187 this idea of pacing the frontier. And 148 00:08:09,270 --> 00:08:14,138 OpenAI has disclosed quite clearly, like, internal frameworks of what you plan 149 00:08:14,221 --> 00:08:15,037 to do. I think 150 00:08:15,120 --> 00:08:18,161 people are still trying to understand what the collective effort looks like. 151 00:08:18,244 --> 00:08:22,244 If you collaborate with your peers or you work with the administration, 152 00:08:22,327 --> 00:08:25,870 what does that actually result in on an AI safety 153 00:08:25,970 --> 00:08:30,607 I think one thing that would be good would be shared standards for safety cases about 154 00:08:30,690 --> 00:08:36,767 when you know, what you need at each level to proceed, without undue risk and then 155 00:08:36,850 --> 00:08:41,150 some sort of shared methods for ensuring that everybody is doing a good job 156 00:08:41,233 --> 00:08:44,237 in implementing those. And that could be independent evaluators. 157 00:08:44,321 --> 00:08:46,817 That could be government review. That could be companies 158 00:08:46,900 --> 00:08:49,109 checking each other's work. There have been many proposals there. 159 00:08:49,192 --> 00:08:50,577 But I think those two things would be good. 160 00:08:50,660 --> 00:08:55,237 How do you feel about the disclosure of the risk? You talked to the United Nations 161 00:08:55,320 --> 00:08:58,952 about it doesn't matter the percentage risk. Still too great. Yeah. 162 00:08:59,035 --> 00:09:00,117 The news cycle had in 163 00:09:00,200 --> 00:09:05,557 a tropics prospectus, the the existential threat to humanity disclosed in black 164 00:09:05,640 --> 00:09:06,797 and white. Is that 165 00:09:06,880 --> 00:09:10,571 incumbent on you as a frontier lab? I appreciate you felt confidentially, 166 00:09:10,654 --> 00:09:12,930 but do you just need to put it out there? 167 00:09:13,200 --> 00:09:16,157 I think this has nothing to do with going public. This is just, I think, our 168 00:09:16,240 --> 00:09:19,167 responsibility for our mission and what we're working on To do it anyway. 169 00:09:19,250 --> 00:09:22,190 Explain to the world the risks and benefits that we see in the sort 170 00:09:22,274 --> 00:09:24,350 of shifting landscape as things evolve. 171 00:09:24,450 --> 00:09:28,447 Sam Altman, OpenAI CEO, having us here at Dev Day. Thank you very much. 172 00:09:28,530 --> 00:09:30,230 Thank you for coming. Indeed.