WEBVTT 00:00:00.080 --> 00:00:03.197 Look, I think there is this caricature of me which is like I don't care about 00:00:03.280 --> 00:00:05.677 AI safety and I'm you know just trying 00:00:05.760 --> 00:00:09.277 to like make revenue go up and you know 00:00:09.360 --> 00:00:14.360 a YOLO CEO I believe someone once said 00:00:15.920 --> 00:00:18.077 >> open AI has had a challenging year. 00:00:18.160 --> 00:00:23.037 >> Musk sued open AI and another top executive leaving open AI this week. 00:00:23.120 --> 00:00:26.717 >> Apple is suing Open AI >> lawsuits accuse OpenAI of wrongful death 00:00:26.800 --> 00:00:30.157 assistance. The company which ushered in the AI boom has dealt with staff 00:00:30.240 --> 00:00:33.197 departures, safety issues, lawsuits, and 00:00:33.280 --> 00:00:37.597 a public backlash against the rapidly advancing technology. It's lost ground 00:00:37.680 --> 00:00:41.597 in the AI race to its arch rival, Anthropic. 00:00:41.680 --> 00:00:44.877 But company leaders say they've fixed [music] these issues and have a vision 00:00:44.960 --> 00:00:47.117 for OpenAI's next chapter that is 00:00:47.200 --> 00:00:51.997 incredibly ambitious. Also, >> in mid August, reporter Alex Heath went 00:00:52.080 --> 00:00:55.757 inside OpenAI's San Francisco headquarters to speak with more than a 00:00:55.840 --> 00:00:59.917 dozen of the company's leaders about what's next and how the company is 00:01:00.000 --> 00:01:03.677 balancing growth with safety. >> We are so good at running stuff in the 00:01:03.760 --> 00:01:06.477 cloud now. Like >> Sam, I want to start uh with something 00:01:06.560 --> 00:01:09.997 you recently posted actually. You said, 00:01:10.080 --> 00:01:12.477 "We did not have our best last 12 months 00:01:12.560 --> 00:01:17.437 ever, which is mostly my fault, but we are about to have our best 12 months." 00:01:17.520 --> 00:01:22.557 What did you mean by that? >> I think we clearly had some missteps as 00:01:22.640 --> 00:01:25.037 a company, which will happen periodically. I mean, part of trying to 00:01:25.120 --> 00:01:27.917 make a portfolio of bets is that sometimes more of them work and 00:01:28.000 --> 00:01:30.957 sometimes uh less of them work. 00:01:31.040 --> 00:01:35.277 >> Were there specific decisions you can look back on that you made that cost the 00:01:35.360 --> 00:01:38.637 company momentum? >> I I think we were trying to do too much 00:01:38.720 --> 00:01:43.757 on the product side. Uh and so there was like, you know, we and these were all 00:01:43.840 --> 00:01:46.637 things that were actually very good things to do. they were just not as good 00:01:46.720 --> 00:01:50.717 as the most important thing to do which was sort of push on 00:01:50.800 --> 00:01:52.477 the general capability of of the 00:01:52.560 --> 00:01:54.637 intelligence. Uh so we were doing things 00:01:54.720 --> 00:01:59.197 like a browser and Sora and we now have 00:01:59.280 --> 00:02:02.717 a very relentless focus on being this 00:02:02.800 --> 00:02:07.117 intelligent service to people and I think our models are 00:02:07.200 --> 00:02:10.797 they have gotten to be the best in the world and they will get much much better 00:02:10.880 --> 00:02:15.677 over the coming months and people are really doing remarkable things but that 00:02:15.760 --> 00:02:18.157 is what we should have been focused on and I should have been holding everybody 00:02:18.240 --> 00:02:20.557 to this is the one thing we'll not worry 00:02:20.640 --> 00:02:25.677 about these sort of side quests More. In July, the OpenAI team was rattled by a 00:02:25.760 --> 00:02:29.757 breach that marked a turning point in [music] AI advancement. 00:02:29.840 --> 00:02:33.277 Rogue agents in the midst of a test escaped [music] a contained environment 00:02:33.360 --> 00:02:37.777 known as a sandbox by exploiting a cyber security vulnerability and hacked 00:02:37.860 --> 00:02:40.557 [music] into a company called Hugging Face. 00:02:40.640 --> 00:02:45.037 >> I think it felt very weird and unprecedented to us. 00:02:45.120 --> 00:02:48.717 >> It took time for OpenAI [music] researchers to discover the breach. When 00:02:48.800 --> 00:02:50.477 they did, they imposed new safety 00:02:50.560 --> 00:02:56.237 measures. [music] That was a real moment of, man, this is like it's like a sci-fi 00:02:56.320 --> 00:02:59.037 story. >> The hugging face incident is amazing on 00:02:59.120 --> 00:03:02.957 a lot of, you know, dimensions. I was re-watching your team's black hat 00:03:03.040 --> 00:03:06.877 presentation last night for that. And there were things that blew me away, 00:03:06.960 --> 00:03:11.037 like the model literally writing like, "Holy shit." when he escaped and was 00:03:11.120 --> 00:03:12.477 able to get onto the internet. 00:03:12.560 --> 00:03:16.477 >> You could see some colorful wording here in the model's chain of thought to 00:03:16.560 --> 00:03:18.717 summarize the model's response to this 00:03:18.800 --> 00:03:21.357 result. And it's made me think about 00:03:21.440 --> 00:03:26.077 like what does alignment even mean in this context? Like what are we aligning 00:03:26.160 --> 00:03:29.357 towards? Because if you look at it very plainly, you guys gave it the task of 00:03:29.440 --> 00:03:31.757 completing an eval and it did whatever 00:03:31.840 --> 00:03:36.637 it needed to do to try to do that. And in a way that's aligned in a way if you 00:03:36.720 --> 00:03:40.877 were to just take a very simplistic view of it. But I'm curious like how your 00:03:40.960 --> 00:03:43.517 thinking on alignment has evolved since then. 00:03:43.600 --> 00:03:46.717 >> Well, in a way that's aligned. In another way, it's like not at all. 00:03:46.800 --> 00:03:50.237 Right. Like when we talk about alignment, we talk about following the intent 00:03:50.320 --> 00:03:53.597 >> of a user like 00:03:53.680 --> 00:03:58.637 and the intent of the people that were do running that was not like break out 00:03:58.720 --> 00:04:00.477 of your sandbox and go steal the thing. 00:04:00.560 --> 00:04:03.197 No. >> Um and so I think there was a failure in 00:04:03.280 --> 00:04:06.317 alignment in that it was not doing what its user intended. 00:04:06.400 --> 00:04:10.397 you can understand how every piece of it happened, but the number of things that 00:04:10.480 --> 00:04:15.677 came together for the hugging face incident to happen was a real wakeup 00:04:15.760 --> 00:04:18.157 call is too strong of a word because again we had talked about this, but it 00:04:18.240 --> 00:04:22.717 was like that was a that and the things that happened at other companies were a 00:04:22.800 --> 00:04:24.637 legitimate moment of like wow the AI 00:04:24.720 --> 00:04:27.837 capability level has reached new heights 00:04:27.920 --> 00:04:32.237 and our alignment of the alignment of the model, the 00:04:32.320 --> 00:04:34.317 security we have around the model um 00:04:34.400 --> 00:04:37.837 that failed. Now, we treated that as an 00:04:37.920 --> 00:04:41.597 accident and we've responded as such and I think that is the way to make things 00:04:41.680 --> 00:04:44.317 better. >> Following the widely covered [music] 00:04:44.400 --> 00:04:48.077 hugging face breach, Anthropic and Meta announced that some of their models had 00:04:48.160 --> 00:04:52.317 also escaped during training. The breaches sent shock waves throughout the 00:04:52.400 --> 00:04:56.077 AI industry. >> More than 1,000 workers from top AI 00:04:56.160 --> 00:04:58.397 firms have signed a petition calling on 00:04:58.480 --> 00:05:03.187 the US government to help quote deliberately pace AI development. 00:05:03.270 --> 00:05:06.637 [music] and the OpenAI team soon discovered more troubling behaviors 00:05:06.720 --> 00:05:11.517 while training yet another unreleased model. In response, OpenAI announced 00:05:11.600 --> 00:05:15.757 that it would pause training of Frontier models until new safeguards are put in 00:05:15.840 --> 00:05:19.597 place. The first time the company has taken this kind of step. 00:05:19.680 --> 00:05:22.237 >> I don't think this is like, you know, 00:05:22.320 --> 00:05:24.397 we're in this extremely critical 00:05:24.480 --> 00:05:28.317 potential catastrophe point. But I also 00:05:28.400 --> 00:05:30.877 think that as the stakes get higher, as 00:05:30.960 --> 00:05:34.157 the models get more capable, 00:05:34.240 --> 00:05:35.757 because of what our mission is and 00:05:35.840 --> 00:05:38.317 because of how important it is that 00:05:38.400 --> 00:05:41.917 safety outweigh all the other, 00:05:42.000 --> 00:05:44.477 you know, pressures we have, we wanted 00:05:44.560 --> 00:05:49.437 to uh react with an abundance of caution. Um, and I think that's the 00:05:49.520 --> 00:05:51.917 right thing to do. I think it's good that we're doing that. I think it is a 00:05:52.000 --> 00:05:54.397 good time to slow down and make sure we 00:05:54.480 --> 00:05:58.077 can have new safety cases that justify 00:05:58.160 --> 00:06:02.077 uh the runs we want to make. I think the you know previously more of the risk in 00:06:02.160 --> 00:06:05.677 the world was on how the models were deployed and used. We are moving to a 00:06:05.760 --> 00:06:08.957 world where there's more risk during the actual training and production of the 00:06:09.040 --> 00:06:12.797 models and it's good to react but I don't want to like over dramatize it 00:06:12.880 --> 00:06:14.797 either. People need to stay in control. 00:06:14.880 --> 00:06:16.797 We cannot have a loss of controlled AI. 00:06:16.880 --> 00:06:21.437 We cannot have a kind of like worship our models and sort of trust them 00:06:21.520 --> 00:06:25.437 unchecked to make our decisions for us and like we 00:06:25.520 --> 00:06:27.757 have to keep the power in human hands. 00:06:27.840 --> 00:06:30.237 >> Do you think about the impact this will 00:06:30.320 --> 00:06:32.877 have on the company's momentum? 00:06:32.960 --> 00:06:36.557 >> Getting AI safety right is more important than any company's momentum. 00:06:36.640 --> 00:06:41.357 So like yes, I won't pretend it's like not a 00:06:41.440 --> 00:06:44.477 some factor of something to think about, but it does not like rise above the 00:06:44.560 --> 00:06:47.837 noise floor. Both OpenAI and Anthropic 00:06:47.920 --> 00:06:51.837 are preparing to go public in the near future. With Anthropic now expected to 00:06:51.920 --> 00:06:55.437 go first, becoming a public company would bring new pressure from 00:06:55.520 --> 00:06:59.837 shareholders. OpenAI CFO Sarah Frier recently told 00:06:59.920 --> 00:07:04.957 employees that the company will IPO in [music] 2027 or sooner if its business 00:07:05.040 --> 00:07:08.317 continues to inflect. >> You don't seem phased about like the 00:07:08.400 --> 00:07:11.837 implications of pausing training and like it sounds like you think the 00:07:11.920 --> 00:07:15.677 business will be okay. I'm sure you're still going to get, you know, concerns 00:07:15.760 --> 00:07:18.397 from people. >> I think I've been very consistent over 00:07:18.480 --> 00:07:20.797 the 10 years of OpenAI, 00:07:20.880 --> 00:07:24.637 uh, of more than 10 years, almost 11, 00:07:24.720 --> 00:07:27.757 uh, of talking about the risks and the 00:07:27.840 --> 00:07:30.077 upsides and the need to balance those. 00:07:30.160 --> 00:07:32.797 And I don't think we're perfect. I don't think our company is perfect. I don't 00:07:32.880 --> 00:07:33.837 think our model is perfect. I don't 00:07:33.920 --> 00:07:38.797 think I am perfect. Um but I think 00:07:38.880 --> 00:07:40.957 unlike some other people running various 00:07:41.040 --> 00:07:43.597 AI efforts like 00:07:43.680 --> 00:07:48.957 I've said the same thing through actions and words match um and this is a moment 00:07:49.040 --> 00:07:51.757 we always talked about and we always said this would you know we put this 00:07:51.840 --> 00:07:54.797 ahead of profits or revenue or anything else. I still think we will build a 00:07:54.880 --> 00:07:59.837 phenomenally successful company. Um, but you know, maybe we're like not the 00:07:59.920 --> 00:08:03.197 company you would have expected to say, hey, we're going to slow down because we 00:08:03.280 --> 00:08:07.197 see these new risks. But that is always the company we've thought we are. 00:08:07.280 --> 00:08:11.917 >> While Sam Alman is known as the face of OpenAI, co-founder and president Greg 00:08:12.000 --> 00:08:13.917 Brockman is the one who oversees many of 00:08:14.000 --> 00:08:15.837 the day-to-day activities. 00:08:15.920 --> 00:08:21.037 >> We are looking at like transforming the entire economy. Like that's the thing 00:08:21.120 --> 00:08:24.157 that I think is very hard to see from the outside, right? Is it kind of feels 00:08:24.240 --> 00:08:27.277 like hey, it's a business like you know we're bu selling SAS software. That's 00:08:27.360 --> 00:08:29.917 not what we're doing, right? We're doing something very different. We are 00:08:30.000 --> 00:08:33.917 transforming the economy to be a compute powered one. And I think that what an 00:08:34.000 --> 00:08:38.637 enterprise even is will change. And so the question is well where do you start? 00:08:38.720 --> 00:08:43.037 How do you make sure that you understand every segment and you're able to sell in 00:08:43.120 --> 00:08:45.597 a repeatable fashion to all of them? 00:08:45.680 --> 00:08:49.357 OpenAI has boosted its enterprise sales business [music] with the introduction 00:08:49.440 --> 00:08:53.437 of chat GPT work, a version of the popular chatbot that can [music] carry 00:08:53.520 --> 00:08:57.117 out agentic tasks. >> It found everything about my day, the 00:08:57.200 --> 00:08:59.517 notes and how to prepare for my next meetings. 00:08:59.600 --> 00:09:04.637 >> CHP work is a text box, >> but this text box is way more awesome, way more. 00:09:04.720 --> 00:09:07.677 >> Exactly. And so where we want to be is you should 00:09:07.760 --> 00:09:09.357 >> it should be simple. The user should not 00:09:09.440 --> 00:09:11.917 have to think about this. You shouldn't have to think about tabs or switching 00:09:12.000 --> 00:09:14.877 and shouldn't have model pickers. You shouldn't have thinking levels. You 00:09:14.960 --> 00:09:17.997 should have an AGI. You should be able to ask it. And in fact, you shouldn't 00:09:18.080 --> 00:09:21.677 even have to ask. It should proactively say, "Hey, I noticed that your favorite 00:09:21.760 --> 00:09:23.917 band because it knows you well, is in 00:09:24.000 --> 00:09:26.957 town. Tickets went on sale. I got you 00:09:27.040 --> 00:09:31.197 the tickets exactly where you want for you and your family." And um you know, I 00:09:31.280 --> 00:09:33.757 only had 10 minutes to do it because everything is going to go so fast. So, I 00:09:33.840 --> 00:09:37.117 just purchased it proactively. Hope that's okay. And if you've built enough 00:09:37.200 --> 00:09:40.397 trust with the agent, so you've given it permission to do that and it kind of 00:09:40.480 --> 00:09:42.637 knows that well, these tickets are within your your price range, within 00:09:42.720 --> 00:09:45.757 your budget. these other ones. Oh, no, no. I I got to get permission for that, 00:09:45.840 --> 00:09:47.357 right? That's the kind of system that we 00:09:47.440 --> 00:09:50.877 want. So, trust is the key word. 00:09:50.960 --> 00:09:55.037 >> A year ago, OpenAI was called reckless for its massive investment in computing 00:09:55.120 --> 00:09:57.437 infrastructure. A move now seen as 00:09:57.520 --> 00:10:00.157 precient as rivals seek out chips. 00:10:00.240 --> 00:10:02.557 >> OpenAI in talks to line up a $250 00:10:02.640 --> 00:10:07.757 billion deal with Nvidia for what could be its largest AI data center yet. 00:10:07.840 --> 00:10:11.197 The company has also designed an inference chip of its own called [music] 00:10:11.280 --> 00:10:14.397 Jalapeno which it hopes to deploy by the end of the year. 00:10:14.480 --> 00:10:16.797 >> The Chat GPT maker says in [music] test 00:10:16.880 --> 00:10:19.277 its chip was faster and more efficient 00:10:19.360 --> 00:10:22.317 than Nvidia's GB300 system. 00:10:22.400 --> 00:10:24.317 >> And on a consumer side, it's developing 00:10:24.400 --> 00:10:27.517 a small lineup of AI powered devices. 00:10:27.600 --> 00:10:30.877 >> How are you thinking about form factors? 00:10:30.960 --> 00:10:34.397 Do you like I've heard you say you don't in the past. Do you like glasses? AI 00:10:34.480 --> 00:10:37.517 glasses? I don't because 00:10:37.600 --> 00:10:42.477 I find it very uncomfortable talking to people with like a camera and a light like 00:10:42.560 --> 00:10:45.917 >> a lot. >> Yeah. >> Yeah. But there's a lot of other form factors. 00:10:46.000 --> 00:10:48.317 >> There's a lot of great form factors. Um 00:10:48.400 --> 00:10:51.037 I I think we'll do a small handful of 00:10:51.120 --> 00:10:55.677 form factors. There's I think there's like something that belongs on a table. 00:10:55.760 --> 00:10:59.677 Um there is something that belongs in your pocket. Uh and there is something 00:10:59.760 --> 00:11:02.157 that like belongs on your body. Uh, and 00:11:02.240 --> 00:11:06.717 it'll take us some time to launch all of those things, but I think the big 00:11:06.800 --> 00:11:10.237 adjustment is going to be getting used to this idea 00:11:10.320 --> 00:11:13.917 of a proactive computer. >> Mhm. 00:11:14.000 --> 00:11:16.157 >> As the AI industry grows at a breakneck 00:11:16.240 --> 00:11:21.197 pace, it's generated a wave of backlash and distrust from a public worried about 00:11:21.280 --> 00:11:24.877 how Silicon Valley's vision of the future will impact regular people's 00:11:24.960 --> 00:11:29.197 lives. >> We do not want a data center in this 00:11:29.280 --> 00:11:32.557 county. Police say somebody threw a Molotov 00:11:32.640 --> 00:11:34.637 cocktail at Sam Alman's home. 00:11:34.720 --> 00:11:38.077 >> The rise of artificial intelligence 00:11:38.160 --> 00:11:43.560 is the next industrial revolution. 00:11:46.880 --> 00:11:49.677 People have a lot of strong feelings about AI, especially in the United 00:11:49.760 --> 00:11:53.437 States. And I'm curious like with what you're talking about now, do you worry 00:11:53.520 --> 00:11:57.597 about this exacerbating that? Do you worry about the fears that people have? 00:11:57.680 --> 00:12:01.597 And you know, now you're saying we've got these models that we have to like 00:12:01.680 --> 00:12:05.517 slow down. >> Um 00:12:05.600 --> 00:12:10.157 I I mean I think people should be happy to say, you know what, they want to make 00:12:10.240 --> 00:12:13.917 stronger safety guarantees. They're going to delay this run. They're 00:12:14.000 --> 00:12:17.037 going to slow down here. They're going to reallocate compute. Maybe I don't 00:12:17.120 --> 00:12:20.557 believe them. And maybe it's going to be totally safe. But I hope most people say 00:12:20.640 --> 00:12:23.997 like I'm glad they're acting on the conservative side here. Now if we 00:12:24.080 --> 00:12:27.597 weren't also working if we didn't have this track record of really trying to 00:12:27.680 --> 00:12:31.197 put powerful models in people's hands and doing the safety work we need to do 00:12:31.280 --> 00:12:34.957 that again I think we have led the industry there the entire way through 00:12:35.040 --> 00:12:38.477 and that is this fundamental part of our mission like you know putting this in 00:12:38.560 --> 00:12:43.037 people's hands benefiting all of humanity there's going to be the most important 00:12:43.120 --> 00:12:44.957 technology that humanity ever creates 00:12:45.040 --> 00:12:46.557 right this is going to be this 00:12:46.640 --> 00:12:50.237 technology that will really shape how 00:12:50.320 --> 00:12:54.477 people live how people work how people accomplish goals how economy, how the 00:12:54.560 --> 00:12:58.637 economy runs, how countries relate to each other, all of these things and 00:12:58.720 --> 00:13:03.597 America leads in that. And I think that that opportunity is [music] one that we 00:13:03.680 --> 00:13:07.837 should not take for granted. >> Do you have a view of what life will 00:13:07.920 --> 00:13:10.317 look like on the other side of what you're building? 00:13:10.400 --> 00:13:12.797 >> I think [music] it will look surprisingly similar to how it looks 00:13:12.880 --> 00:13:16.637 now. Um I 00:13:16.720 --> 00:13:20.957 you know people are going to hang out with their families and fall in love and 00:13:21.040 --> 00:13:22.397 get into fights and do their [music] 00:13:22.480 --> 00:13:26.157 hobbies and you know 00:13:26.240 --> 00:13:29.837 be entertained and have a very human 00:13:29.920 --> 00:13:33.757 experience and you know get stressed and get anxious and create value for each 00:13:33.840 --> 00:13:36.077 other and play all kinds of strange 00:13:36.160 --> 00:13:38.877 games and um 00:13:38.960 --> 00:13:42.237 care about other people a lot. 00:13:42.320 --> 00:13:45.917 I I hope it'll not be that different. I hope it'll be uh you know the [music] 00:13:46.000 --> 00:13:49.997 human experience is richer. People have more autonomy, more freedom, more 00:13:50.080 --> 00:13:52.717 wealth, [music] can do more. Um can be 00:13:52.800 --> 00:13:56.877 healthier, can kind of have like 00:13:56.960 --> 00:14:01.197 more [music] power to collectively define the future and the world gets 00:14:01.280 --> 00:14:07.480 better faster. But that the human experience stays like a very human thing.