WEBVTT 00:00:05.000 --> 00:00:08.917 Good afternoon, everyone, and welcome to the OpenAI forum. I'm Chris Nicholson, 00:00:09.000 --> 00:00:11.477 and I'm glad to be here with all of you today. 00:00:11.560 --> 00:00:16.157 The forum is a place for serious conversation about how AI is being used 00:00:16.240 --> 00:00:20.437 in the world, what we're learning from that, and how more people can help shape 00:00:20.520 --> 00:00:22.837 its trajectory. Today's conversation 00:00:22.920 --> 00:00:24.917 focuses on one of the biggest questions 00:00:25.000 --> 00:00:30.237 in technology, what it will mean as AI systems grow dramatically more capable, 00:00:30.320 --> 00:00:31.837 and how we should think about their 00:00:31.920 --> 00:00:34.917 implications for science, work, our life 00:00:35.000 --> 00:00:39.237 together, and governance. To discuss that, I'm joined by Sam 00:00:39.320 --> 00:00:43.517 Altman, co-founder and CEO of OpenAI, Josh Acklin, 00:00:43.600 --> 00:00:48.797 OpenAI's chief futurist, and Adrian Akyol, a long-time researcher here. So, let's 00:00:48.880 --> 00:00:51.477 get started. Sam, 00:00:51.560 --> 00:00:56.037 the blueprint we released this morning talks a lot about superintelligence. Um, 00:00:56.120 --> 00:01:00.237 a big question on my mind is, why are we doing that now, and what are 00:01:00.320 --> 00:01:04.317 some things that you can see from the inside that you wish everybody knew? 00:01:04.400 --> 00:01:08.197 >> The The biggest reason is simply that the rate of progress is continuing to 00:01:08.280 --> 00:01:12.357 accelerate, and we believe we are very close now. 00:01:12.440 --> 00:01:16.037 And this won't be a one-time thing. This will be a you know, 00:01:16.120 --> 00:01:19.677 over the next few years to powerful models that will impact the world in 00:01:19.760 --> 00:01:24.877 important ways. The researchers that are working on these models that did and 00:01:24.960 --> 00:01:28.277 I think an incredible job with this set of ideas, and these are meant to be 00:01:28.360 --> 00:01:30.997 early ideas to start a discussion. 00:01:31.080 --> 00:01:33.637 You know, I'm sure we'll get to much better ideas as the world debates all of 00:01:33.720 --> 00:01:36.397 these. Are staring down the models that are 00:01:36.480 --> 00:01:39.517 coming, the pipeline in front of us. 00:01:39.600 --> 00:01:42.877 And you know, we may be wrong. We may hit 00:01:42.960 --> 00:01:45.397 some wall. We We are imperfect, but our 00:01:45.480 --> 00:01:47.437 But given what we see, 00:01:47.520 --> 00:01:50.677 we expect to be in a world of extremely 00:01:50.760 --> 00:01:53.397 capable models quite soon, and then for 00:01:53.480 --> 00:01:56.197 the the ramp of the capability to 00:01:56.280 --> 00:01:58.117 continue to increase. I think this will 00:01:58.200 --> 00:02:03.597 have huge impact on the economy, on the way we live, on what we can do. And 00:02:03.680 --> 00:02:05.037 one thing I've observed from watching 00:02:05.120 --> 00:02:06.997 the world go through some number of 00:02:07.080 --> 00:02:11.197 transitions is that the more time the 00:02:11.280 --> 00:02:15.517 the public, our leaders, the political system has to debate ideas before you 00:02:15.600 --> 00:02:19.597 really have to make a decision, the more likely you are to make a good decision. 00:02:19.680 --> 00:02:23.877 So, starting this now, given what we see coming, so that as this becomes 00:02:23.960 --> 00:02:26.637 a very large issue of public debate, I think it's important. 00:02:26.720 --> 00:02:31.237 >> For sure. Um and speaking of debate, we brought in a lot of researchers in 00:02:31.320 --> 00:02:34.997 during very early in this process. I think it was a 00:02:35.080 --> 00:02:38.717 maybe a unique exercise here in how many researchers were working 00:02:38.800 --> 00:02:42.037 with the folks who also think about policy a lot. 00:02:42.120 --> 00:02:46.117 It's closely tied. What was that like for you and for the research group as a 00:02:46.200 --> 00:02:49.597 whole? >> Yeah, you know, I think it was Well, it 00:02:49.680 --> 00:02:52.757 was a very interesting experience. Like it was my first time 00:02:52.840 --> 00:02:56.997 working like actively on a on a policy doc. I think it was a little bit like 00:02:57.080 --> 00:03:01.357 humbling in some cases, you know, uh uh sometimes 00:03:01.440 --> 00:03:05.437 as researchers, we can have these abstract ideas of oh, we should really 00:03:05.520 --> 00:03:09.837 be thinking about the economic impact of this or or about like policy for for 00:03:09.920 --> 00:03:12.597 safety, but you know, it's it's one thing to think about this and another 00:03:12.680 --> 00:03:17.277 thing to like actually put, you know, pen to paper and um and think of 00:03:17.360 --> 00:03:22.197 concrete uh policy ideas that are going to be debated by by your your peer 00:03:22.280 --> 00:03:23.997 uh your your peer researchers. Um so, it 00:03:24.080 --> 00:03:25.717 was an interesting experience for me. 00:03:25.800 --> 00:03:29.557 Um I I hope there was 00:03:29.640 --> 00:03:31.717 uh you know, very helpful for the final 00:03:31.800 --> 00:03:36.877 product uh that we had the researchers involved so early and so and so deeply 00:03:36.960 --> 00:03:40.597 like for for something like this, uh especially since it's so 00:03:40.680 --> 00:03:44.437 forward-looking. Uh you kind of need to to Sam's point, 00:03:44.520 --> 00:03:48.677 like you need people who are um who are dealing with this technology 00:03:48.760 --> 00:03:52.237 every day, who know how to build it, know how the safety stack works, 00:03:52.320 --> 00:03:54.637 and also are seeing like 00:03:54.720 --> 00:03:58.077 the speed of progress. Um and one thing 00:03:58.160 --> 00:04:01.997 that, you know, that I remember as part of this process in the in the past few 00:04:02.080 --> 00:04:07.157 months that we were were working on it, like a lot of researchers went through a a 00:04:07.240 --> 00:04:10.957 transition uh from, you know, writing most of their 00:04:11.040 --> 00:04:13.677 own code to having AI most write most of 00:04:13.760 --> 00:04:17.957 their code. And I think that I think to some extent led them to bring a lot of 00:04:18.040 --> 00:04:19.557 the urgency of like 00:04:19.640 --> 00:04:23.997 this technology is is real and important 00:04:24.080 --> 00:04:26.997 moving fast in a way that maybe um 00:04:27.080 --> 00:04:30.397 not everyone can can see. Uh and and 00:04:30.480 --> 00:04:34.397 that's part of also kind of your your earlier question about why now, like 00:04:34.480 --> 00:04:36.837 because we are seeing uh this urgency. 00:04:36.920 --> 00:04:40.717 >> Can I tell a quick story that reminds me of? Um 00:04:40.800 --> 00:04:45.557 There was a night in 00:04:45.640 --> 00:04:49.797 I guess it would've been late January, early February of 2020, 00:04:49.880 --> 00:04:54.997 when uh the OpenAI researchers kind of got obsessed with COVID before the rest 00:04:55.080 --> 00:04:57.277 of the world did. And we were talking about it all the time, and we were 00:04:57.360 --> 00:04:59.677 watching the numbers every day, and we're like, "This is going to happen." 00:04:59.760 --> 00:05:04.157 >> Yeah. >> And, you know, we were like making plans to go uh 00:05:04.240 --> 00:05:07.397 work from home, and there was this like some article that came out like mocking 00:05:07.480 --> 00:05:10.557 us because they're like these crazy people at OpenAI. We had put like some 00:05:10.640 --> 00:05:14.157 copper or something on some of the door handles because people >> I remember that. 00:05:14.240 --> 00:05:16.477 >> You remember that? And some journalist wrote about this, 00:05:16.560 --> 00:05:21.560 and there's all all these things going on. And I 00:05:21.720 --> 00:05:26.197 And you know, we we kind of like made our plans, and we assumed a 00:05:26.280 --> 00:05:28.117 shutdown was going to happen. We're like, "There's this thing coming. The 00:05:28.200 --> 00:05:32.037 world's not paying attention." And for whatever reason, uh something about like 00:05:32.120 --> 00:05:34.517 working on exponential makes you understand these things better. So, I 00:05:34.600 --> 00:05:38.717 think we were like a group of people that were primed to do it. Um 00:05:38.800 --> 00:05:41.517 And then there was this one night, it was like a very cold night. I lived in 00:05:41.600 --> 00:05:45.397 the mission at the time. And I was like, "I'm about to get locked in my house for 00:05:45.480 --> 00:05:46.757 a while, and I'm going to go for a walk. 00:05:46.840 --> 00:05:49.357 I want to one more time I want to like, you know go for a walk cuz who knows 00:05:49.440 --> 00:05:52.277 what what's going to happen. And I went through this long walk 00:05:52.360 --> 00:05:55.797 through the city, hours, cold cold night. And I was like watching people, 00:05:55.880 --> 00:05:58.277 you know, breathing in each other's faces in restaurants and bars through 00:05:58.360 --> 00:06:01.557 the windows. And I was like wearing my mask looking all looking crazy. And 00:06:01.640 --> 00:06:04.237 there was one other dude out wearing a mask and kind of we nodded at each 00:06:04.320 --> 00:06:06.757 other. And that was But other than that it was just like 00:06:06.840 --> 00:06:11.797 life felt totally normal. I have not felt that so acutely as I do 00:06:11.880 --> 00:06:13.637 again in this moment. >> Yeah. 00:06:13.720 --> 00:06:17.957 >> Um where it's like there is this crazy change. The change 00:06:18.040 --> 00:06:21.157 has already happened. Like the models have already hit some level. Society has 00:06:21.240 --> 00:06:23.397 not digested them yet. >> Yeah. 00:06:23.480 --> 00:06:27.317 >> We feel like we see it clearly. We are trying to tell the world that it's 00:06:27.400 --> 00:06:31.517 going to happen. It is hard to get this across. But it feels like that night in 00:06:31.600 --> 00:06:33.397 the very beginning of COVID walking through the streets again. 00:06:33.480 --> 00:06:37.397 >> That's really interesting. And and you've talked a bit about the upsides to 00:06:37.480 --> 00:06:39.877 unlock and a lot about science as a 00:06:39.960 --> 00:06:43.077 means to do that. Um so so if we're 00:06:43.160 --> 00:06:45.877 going to like contrast this with COVID, 00:06:45.960 --> 00:06:51.677 what are some of the hugely positive things in the pipeline for us? 00:06:51.760 --> 00:06:54.517 >> I will answer that. I think the positives are so positive we should talk 00:06:54.600 --> 00:06:58.597 more about the other things that we're thinking about here. But like it it you 00:06:58.680 --> 00:07:01.717 mentioned science. If we can really go make a 00:07:01.800 --> 00:07:05.237 decades worth of scientific progress in a year, if we can go cure a ton of 00:07:05.320 --> 00:07:08.597 diseases, if we can come up with personalized medicine for people, find 00:07:08.680 --> 00:07:13.277 new materials to sort of make cheap safe energy 00:07:13.360 --> 00:07:16.157 there's if we can make it such that anybody who 00:07:16.240 --> 00:07:19.397 can come up with an idea for a startup can have the AI implement it, write the 00:07:19.480 --> 00:07:22.557 piece of software they want, uh you know, have a custom video game 00:07:22.640 --> 00:07:28.277 that's like the most fun for you to play. Uh this stuff is all wonderful. Now, 00:07:28.360 --> 00:07:30.757 you know, part of the reason for putting this out with urgency is there are going 00:07:30.840 --> 00:07:34.557 to come there are a lot of things coming that we'll need to mitigate. 00:07:34.640 --> 00:07:37.837 I assume we'll figure it out. I I'm always an optimistic person by nature. 00:07:37.920 --> 00:07:42.677 But also the benefits of this this has change the 00:07:42.760 --> 00:07:44.037 option space in front of us and what we 00:07:44.120 --> 00:07:47.517 can do that one of 00:07:47.600 --> 00:07:51.877 my first reaction to reading the blueprint after the researchers wrote it 00:07:51.960 --> 00:07:55.877 was like it is awesome but also a little crazy that we can credibly be talking 00:07:55.960 --> 00:08:00.717 about these kinds of things. Um so we do have like if 00:08:00.800 --> 00:08:02.917 the AI is as good as we think >> Yeah. 00:08:03.000 --> 00:08:06.677 >> and all these wonderful things happen, we also have an incredible new tool at 00:08:06.760 --> 00:08:09.077 our disposal to help us mitigate the potential downsides. 00:08:09.160 --> 00:08:13.397 >> Yeah. And it's a tool for everyone. Uh Josh, I'd like to ask you 00:08:13.480 --> 00:08:17.277 So So I I think we've talked about this as social infrastructure that changes 00:08:17.360 --> 00:08:19.637 the way people work and learn and 00:08:19.720 --> 00:08:24.477 participate in society. Um what 00:08:24.560 --> 00:08:28.997 what responsibilities do we have and do institutions have to help society get 00:08:29.080 --> 00:08:31.397 ready for that and how how do you think about it? 00:08:31.480 --> 00:08:34.237 >> Yeah. Um you know, one of the things that I think about here, this does kind 00:08:34.320 --> 00:08:37.037 of come back to the broad benefits for everyone thing. 00:08:37.120 --> 00:08:39.317 Uh for a long time in society, we've had 00:08:39.400 --> 00:08:44.357 this great aspiration that everyone could have, you know, food, shelter, 00:08:44.440 --> 00:08:48.837 electricity, uh health care and we've we've always 00:08:48.920 --> 00:08:53.197 wanted to see there be some kind of 00:08:53.280 --> 00:08:55.877 new step in the direction of providing these sorts of things for everyone so 00:08:55.960 --> 00:08:59.437 that everyone can focus on the things that are most important in their lives. 00:08:59.520 --> 00:09:03.357 Um and we've often been told over and over again like, you know, actually we 00:09:03.440 --> 00:09:05.037 can't have that, it's too expensive. 00:09:05.120 --> 00:09:09.437 There's no way to pay for it. Um like it's it's nice but it would be too 00:09:09.520 --> 00:09:11.917 burdensome to society to do all of that. 00:09:12.000 --> 00:09:14.117 And I think what AI and 00:09:14.200 --> 00:09:16.877 superintelligence will unlock is the 00:09:16.960 --> 00:09:21.197 freedom to do all of it at a much lower cost than has ever been possible. And 00:09:21.280 --> 00:09:24.397 correspondingly for the folks of us who are sort of the stewards of the 00:09:24.480 --> 00:09:28.357 technology, we have a special responsibility to make sure that we 00:09:28.440 --> 00:09:30.837 actually do fully realize those benefits 00:09:30.920 --> 00:09:36.357 and be a part of building policies and systems that help working people, middle 00:09:36.440 --> 00:09:39.397 class people, people in low-income countries, really make sure that it 00:09:39.480 --> 00:09:43.677 benefits everyone and not just the very wealthy. So, I'm those are the things 00:09:43.760 --> 00:09:47.197 that I think of as key responsibilities here, and I'm very excited about it. 00:09:47.280 --> 00:09:50.397 Plus, the downside risks are also quite 00:09:50.480 --> 00:09:54.557 serious, and there are systemic shocks that could happen. We have to prepare. 00:09:54.640 --> 00:09:56.517 We have to be thoughtful. We have to be honest about them. 00:09:56.600 --> 00:10:00.397 >> Yeah. Yeah, so that sounds like a resilience question. Um 00:10:00.480 --> 00:10:04.077 and and when I look at the document and and speak with 00:10:04.160 --> 00:10:08.917 you folks, it sounds like you think of resilience in layers, right? There's um 00:10:09.000 --> 00:10:12.397 kind of and and a lot of it is not before the AI goes out, but it's also 00:10:12.480 --> 00:10:17.197 after the AI goes out in terms of um our responses to AI and how people are are 00:10:17.280 --> 00:10:20.317 prepared for it. I'd actually like to ask each of you how you think about 00:10:20.400 --> 00:10:23.717 resilience. Um Adrian, why don't we why don't we start with you? 00:10:23.800 --> 00:10:28.037 >> Yeah. You know, I I think maybe you want one distinction 00:10:28.120 --> 00:10:31.477 that I would draw with uh you know, classically we've thought 00:10:31.560 --> 00:10:33.677 about about safety and about making sure 00:10:33.760 --> 00:10:38.517 that we that we run, you know, safety evals, that we red team our models, um 00:10:38.600 --> 00:10:41.397 that we implement mitigations. And I think like that's, you know, to your 00:10:41.480 --> 00:10:45.677 point of resilience is a layer. This is like a very important layer that we 00:10:45.760 --> 00:10:50.117 should, you know, still keep doing and keep expanding, but um 00:10:50.200 --> 00:10:53.357 at the same time uh 00:10:53.440 --> 00:10:55.997 you want society to 00:10:56.080 --> 00:11:00.517 uh be prepared for the possibility that well, maybe there will be some actors 00:11:00.600 --> 00:11:02.917 who do less safety testing, and like 00:11:03.000 --> 00:11:07.597 what what happens what happens then? How is society resilient to risks from AIs 00:11:07.680 --> 00:11:11.197 uh in these cases? Um maybe there will be 00:11:11.280 --> 00:11:15.837 uh incidents in spite of our safety testing or near misses uh in the 00:11:15.920 --> 00:11:20.077 in the uh blueprint, we talk, for instance, about um 00:11:20.160 --> 00:11:23.677 incident reporting that's modeled a little bit of after um 00:11:23.760 --> 00:11:26.557 how the uh aviation industry uh does 00:11:26.640 --> 00:11:30.757 things when whenever there's kind of like a near miss or any incidents, like 00:11:30.840 --> 00:11:35.717 however minor, that kind of like gets reported to um to a database so that, 00:11:35.800 --> 00:11:39.277 you know, all the companies can kind of know, okay, well, this is a risk and 00:11:39.360 --> 00:11:42.597 perhaps uh you know, here are uh mitigations that that you could 00:11:42.680 --> 00:11:47.357 implement. And so, there's a lot of things that could happen at a society-wide uh 00:11:47.440 --> 00:11:51.037 level. Um another thing would be um 00:11:51.120 --> 00:11:55.077 you know, defending against uh against risks. Uh 00:11:55.160 --> 00:11:57.517 we're talking about models that that can 00:11:57.600 --> 00:11:58.997 code um 00:11:59.080 --> 00:12:02.637 a lot better now. Uh that also implies 00:12:02.720 --> 00:12:04.917 uh cyber capabilities. Like, could they 00:12:05.000 --> 00:12:08.637 help uh bad actors uh run cyber attacks? 00:12:08.720 --> 00:12:12.397 Uh and I think part of resilience is ensuring 00:12:12.480 --> 00:12:14.317 that we uh 00:12:14.400 --> 00:12:19.477 that we make our software systems uh more secure uh and use AI to do these 00:12:19.560 --> 00:12:22.277 things. So, so again, there's as you said, all these layers. So, that's kind 00:12:22.360 --> 00:12:25.277 of how I take the the resilience question. 00:12:25.360 --> 00:12:29.037 >> Sam, before the show, we were talking about how prosperity can be emergent 00:12:29.120 --> 00:12:32.237 when everybody has access to AI. It also strikes me that resilience can be 00:12:32.320 --> 00:12:34.717 emergent. How How do you think about it? 00:12:34.800 --> 00:12:39.597 >> Yeah, I I think our original AI safety thinking 00:12:39.680 --> 00:12:42.757 and the field as in general, like I'll call it classical AI safety thinking, 00:12:42.840 --> 00:12:47.437 was that there are going to be a very tiny number of AIs in the world. The 00:12:47.520 --> 00:12:51.317 only thing that matters is making sure those AIs do the right thing and that as 00:12:51.400 --> 00:12:52.957 long as you align them and they don't do 00:12:53.040 --> 00:12:56.397 unsafe things, the world will be okay. 00:12:56.480 --> 00:13:00.797 I think the picture now is actually more stable but more complex 00:13:00.880 --> 00:13:05.957 and there are going to be many AIs in the world uh and the 00:13:06.040 --> 00:13:09.517 it will not be enough to just say like, you know, 00:13:09.600 --> 00:13:12.517 this one company is going to make sure the AI never does something it shouldn't 00:13:12.600 --> 00:13:14.717 do. Um but there will need to be an 00:13:14.800 --> 00:13:19.917 emergent response across society. Uh you know, 00:13:20.000 --> 00:13:23.117 Adrian talked a little bit about some of this, but if we just take a few examples 00:13:23.200 --> 00:13:25.277 of threats that we expect to be coming. 00:13:25.360 --> 00:13:29.037 Um cybersecurity is definitely going to become a huge issue. AI will be 00:13:29.120 --> 00:13:33.357 incredibly good at finding vulnerabilities in software, and I think 00:13:33.440 --> 00:13:36.517 the world will find that their software is much more brittle and much less 00:13:36.600 --> 00:13:39.677 secure than we thought. And you know, humans just have limited 00:13:39.760 --> 00:13:45.037 capacity to find the exploits. Um, it is not enough to say that, you know, 00:13:45.120 --> 00:13:48.117 one or two or three model providers are just going to make sure that their 00:13:48.200 --> 00:13:52.397 systems won't do this because code is like, you know, the same thing that 00:13:52.480 --> 00:13:55.237 code is very useful to be good at, and being good at writing code can also help 00:13:55.320 --> 00:13:58.277 find security problems. And even if all 00:13:58.360 --> 00:14:02.637 of us somehow could prevent our models from ever being used for this, there 00:14:02.720 --> 00:14:05.477 will be open-source models coming soon that are that are good at code and thus 00:14:05.560 --> 00:14:08.957 good at security exploits. So, what has to happen is 00:14:09.040 --> 00:14:11.597 the world has to use these models, and there can be, you know, differential 00:14:11.680 --> 00:14:14.797 access. You can give it to good known trusted defenders first. We have a 00:14:14.880 --> 00:14:18.677 program for that. Other companies will do similar things. Um, 00:14:18.760 --> 00:14:22.197 and you have to empower the companies that defend software because there will 00:14:22.280 --> 00:14:26.277 be some power plant that no one has to the software for 20 years, and no one 00:14:26.360 --> 00:14:28.317 can patch it, and there's a big problem with it, and you have to do something 00:14:28.400 --> 00:14:30.917 about that. Um, but 00:14:31.000 --> 00:14:35.677 a resilience approach here is, okay, there's this new thing in the world. 00:14:35.760 --> 00:14:39.597 There's AI that is really good at exploiting computer systems. Um, let's 00:14:39.680 --> 00:14:42.437 use AI to defend it, and that is not a one-company thing. That's going to 00:14:42.520 --> 00:14:47.237 require this huge effort. Um, if we go a little bit further, I I think 00:14:47.320 --> 00:14:51.517 there will be a uh bio version of this where, 00:14:51.600 --> 00:14:54.277 you know, classically people have said, well, we're just going to restrict our 00:14:54.360 --> 00:14:57.557 models from being able to develop pathogens. Someone at some point is 00:14:57.640 --> 00:15:01.877 going to use some model to develop a pathogen. And the world needs defense 00:15:01.960 --> 00:15:05.277 shields against that, detection systems, rapid response treatments, a whole bunch 00:15:05.360 --> 00:15:08.957 of other things. And this is not 00:15:09.040 --> 00:15:12.397 This doesn't get us off the hook in any way of aligning our systems and building 00:15:12.480 --> 00:15:14.277 safe systems. We still have to do that. 00:15:14.360 --> 00:15:16.557 Like, we get a time advantage there as 00:15:16.640 --> 00:15:18.517 long as we stay at the frontier. 00:15:18.600 --> 00:15:23.517 But, we do need the world to like do its thing, society to have this emergent 00:15:23.600 --> 00:15:27.197 magic and build these layers of defensive shields. Um there are many 00:15:27.280 --> 00:15:30.237 threats besides those two, but that's enough given we have short time. 00:15:30.320 --> 00:15:34.317 >> For sure. Um Josh, you have shared some interesting 00:15:34.400 --> 00:15:38.957 thoughts with me about how each huge technological shift has produced new 00:15:39.040 --> 00:15:41.557 institutions and new democratic mechanisms. 00:15:41.640 --> 00:15:43.757 >> Mhm. >> And also you've uh been thinking a lot 00:15:43.840 --> 00:15:45.597 about new institutions that might emerge 00:15:45.680 --> 00:15:48.437 now. Um how are you 00:15:48.520 --> 00:15:51.077 What are What are your kind of 00:15:51.160 --> 00:15:55.517 What are your most exciting ideas and what what do you really like to think 00:15:55.600 --> 00:15:59.037 about in terms of the collective response to superintelligence? 00:15:59.120 --> 00:16:01.357 >> Yeah, certainly. Um so so one thing I'll 00:16:01.440 --> 00:16:05.277 say first on the subject of resilience just as grounding, a lot of the problems 00:16:05.360 --> 00:16:09.797 that we're concerned about AI creating new externalities for are problems and 00:16:09.880 --> 00:16:13.677 vulnerabilities that exist in the world regardless of whether AI is present and 00:16:13.760 --> 00:16:16.237 AI just increases the urgency of action. 00:16:16.320 --> 00:16:20.757 But coming back to COVID, we found back then that everyone had a much deeper 00:16:20.840 --> 00:16:25.117 dependency on the functioning of supply chains than most people were previously 00:16:25.200 --> 00:16:28.437 conscious of. Um supply chains are super important, 00:16:28.520 --> 00:16:31.197 supply chains for food, for goods, for for everything that sustains 00:16:31.280 --> 00:16:36.157 civilization. Um and also of course uh you know, for for other types of 00:16:36.240 --> 00:16:39.837 vulnerabilities like democracy. We've been debating for years how um there are 00:16:39.920 --> 00:16:44.117 real risks when people are influencing society sort of inappropriately. 00:16:44.200 --> 00:16:48.397 And we're we're worried that AI is going to make these things potentially easier 00:16:48.480 --> 00:16:52.117 to attack in the near term because there will be tools at people's disposal that 00:16:52.200 --> 00:16:55.037 they haven't had before. Um I'm excited 00:16:55.120 --> 00:16:59.477 about the possibility that we can build new institutions and state capacities to 00:16:59.560 --> 00:17:02.997 use AI to rectify some of these vulnerabilities. Um we can 00:17:03.080 --> 00:17:06.957 systematically identify them with AI in ways we couldn't in the past. We can 00:17:07.040 --> 00:17:09.957 systematically close them in ways we couldn't in the past and we can use AI 00:17:10.040 --> 00:17:14.357 to scale up efforts to combat certain types of issues uh in in ways where we 00:17:14.440 --> 00:17:16.517 can potentially make it too expensive 00:17:16.600 --> 00:17:18.117 for attackers to really do something. 00:17:18.200 --> 00:17:21.397 So, uh you know, on on cyber and bio, I 00:17:21.480 --> 00:17:23.917 am optimistic that maybe we can build an 00:17:24.000 --> 00:17:27.117 ecosystem of defenders that all together 00:17:27.200 --> 00:17:31.157 can make it so expensive for attackers to try to, you know, run a cyber attack 00:17:31.240 --> 00:17:35.437 that they there just won't be that much of an incentive to do it. And if we can 00:17:35.520 --> 00:17:39.517 fully implement the the bio resilient side of things, um not just for 00:17:39.600 --> 00:17:42.637 pathogens that might impact humans, but uh especially for the food supply chain. 00:17:42.720 --> 00:17:44.757 This is one where I have like a hobby horse on this and I'm going to talk 00:17:44.840 --> 00:17:47.677 about it every chance I get. I think people under attend the food supply 00:17:47.760 --> 00:17:52.077 chain bio risks. Um but we can use AI to make that resilient at scale in a way 00:17:52.160 --> 00:17:55.797 that today is cost prohibitive. Um so I'm very excited about, you know, the 00:17:55.880 --> 00:18:00.437 things we can build there. >> Yeah. Neat. Um one more question. 00:18:00.520 --> 00:18:04.797 When we talk about kind of the individual transition for many people, um 00:18:04.880 --> 00:18:08.157 what work looks like, where value shifts to. I know you've been thinking about 00:18:08.240 --> 00:18:09.597 that a lot. Um 00:18:09.680 --> 00:18:12.797 how how what do you see happening as 00:18:12.880 --> 00:18:15.677 folks kind of transition to other ways 00:18:15.760 --> 00:18:18.597 of creating value in an AI economy? 00:18:18.680 --> 00:18:21.797 >> Oh. Oh goodness. That's very broad. Um 00:18:21.880 --> 00:18:24.837 I I think people will have 00:18:24.920 --> 00:18:29.197 a lot more uh opportunity to exercise agency. Um 00:18:29.280 --> 00:18:33.917 you know, if you can uh start a new business and have a team 00:18:34.000 --> 00:18:37.117 of AIs that handle all of the functions of putting the business together that 00:18:37.200 --> 00:18:40.437 you have no expertise in yourself. Um you can get something off the ground an 00:18:40.520 --> 00:18:43.037 awful lot easier. And so, you know, there there are a lot of ways that the 00:18:43.120 --> 00:18:46.997 economy just fundamentally changes when you get people access and tools to help 00:18:47.080 --> 00:18:48.997 them do more things than they could have before. 00:18:49.080 --> 00:18:53.077 >> Yeah, for sure. Um Sam, similar question. You have so much 00:18:53.160 --> 00:18:56.877 experience with startups. Running them, managing Y Combinator, or 00:18:56.960 --> 00:18:59.317 just being in the ecosystem for so long. 00:18:59.400 --> 00:19:04.597 What um how do you see startups changing and like our potential to realize new 00:19:04.680 --> 00:19:07.477 ideas changing with AI? >> I'm obsessed with trying to explore the 00:19:07.560 --> 00:19:11.557 space. I don't know exactly what it's going to look like, but I this idea of 00:19:11.640 --> 00:19:16.477 one person or a very small team being able to create an entire startup 00:19:16.560 --> 00:19:18.757 as Josh was talking about pretty quickly. 00:19:18.840 --> 00:19:23.717 Um all my instincts are there's something 00:19:23.800 --> 00:19:25.637 deep and important to figure out here. 00:19:25.720 --> 00:19:30.957 Every time in our industry that the 00:19:31.040 --> 00:19:33.877 friction, cost, whatever you want to call it of starting the startup has come 00:19:33.960 --> 00:19:37.717 down a lot. Amazing new things have happened. Um you 00:19:37.800 --> 00:19:42.077 know, I remember the transition The one of these transitions that I was doing uh 00:19:42.160 --> 00:19:43.477 a startup during 00:19:43.560 --> 00:19:47.717 uh was when like basically AWS came out 00:19:47.800 --> 00:19:51.037 and all of a sudden there was like an idea of a cloud and a small startup 00:19:51.120 --> 00:19:53.677 didn't have to go do all the crazy things you used to have to do. 00:19:53.760 --> 00:19:55.797 >> Manage his racks in the closet. Yeah. 00:19:55.880 --> 00:20:01.077 >> And that was like an amazing change to what you could do as a small startup 00:20:01.160 --> 00:20:03.877 with a few people. This one that's coming is much bigger and there have 00:20:03.960 --> 00:20:07.077 been several in between. But uh 00:20:07.160 --> 00:20:10.917 I really want to find out what it looks like when a startup is 00:20:11.000 --> 00:20:14.157 you know, two or three people and a ton of GPUs and you can just >> Yeah. 00:20:14.240 --> 00:20:17.557 >> you can really democratize who can start a startup. 00:20:17.640 --> 00:20:19.477 >> Yeah. Yeah. Yeah, it's that 00:20:19.560 --> 00:20:21.317 democratization aspect and it gets down 00:20:21.400 --> 00:20:25.517 to the widespread availability of AI. Um 00:20:25.600 --> 00:20:27.517 how do you What's What's the best frame 00:20:27.600 --> 00:20:34.360 for thinking about bringing more people in, democratizing AI access? 00:20:35.360 --> 00:20:37.477 >> I think when people talk about democratization of AI, they they mean 00:20:37.560 --> 00:20:41.437 two different things. One is share the access 00:20:41.520 --> 00:20:45.797 and making sure that everybody gets to use sufficient AI to 00:20:45.880 --> 00:20:48.157 improve their own lives, build things for other people, all that. And the 00:20:48.240 --> 00:20:51.357 other is a sort of a voice in where it's all going to go. 00:20:51.440 --> 00:20:55.917 Uh I think both are very important. Um part of why we do things like this 00:20:56.000 --> 00:21:00.397 blueprint are if we're not debating the issues as a 00:21:00.480 --> 00:21:02.917 society. Also part of the reason we actually release products in the first 00:21:03.000 --> 00:21:06.557 place. People don't have a feel for this if we're not talking about this. 00:21:06.640 --> 00:21:10.597 Um that is kind of a prerequisite to people being able to have input, um, but 00:21:10.680 --> 00:21:13.757 it's not enough. You also have to have a way that people's input that we listen 00:21:13.840 --> 00:21:16.477 and the input is captured back into the into the system. 00:21:16.560 --> 00:21:20.997 Um, so that's one that I think is really important and the other is 00:21:21.080 --> 00:21:23.237 we need to put 00:21:23.320 --> 00:21:25.957 not just services like ChatGPT, but the 00:21:26.040 --> 00:21:28.437 real deal high compute valuable services 00:21:28.520 --> 00:21:33.237 where people can start a startup or make a scientific discovery or whatever in 00:21:33.320 --> 00:21:34.717 broadly into many people's hands. 00:21:34.800 --> 00:21:37.717 >> Yeah. Yeah, and that's going to take new 00:21:37.800 --> 00:21:42.837 economic models or much cheaper uh, inference to to get it 00:21:42.920 --> 00:21:45.757 wider spread, right? And that's an infrastructure problem among many other 00:21:45.840 --> 00:21:49.317 problems to solve. >> You know, we we used to talk for years 00:21:49.400 --> 00:21:51.677 in OpenAI about when we were going to get through the compute crunch, when we 00:21:51.760 --> 00:21:54.517 were going to build like enough compute that we wouldn't be so strapped. I don't 00:21:54.600 --> 00:21:58.717 think we ever get out of it. If we do our job, if we keep driving the cost of 00:21:58.800 --> 00:22:02.957 intelligence down and the capabilities of intelligence up, I think uh, you 00:22:03.040 --> 00:22:06.637 know, effectively if that happens, the demand is uncapped. 00:22:06.720 --> 00:22:08.277 And 00:22:08.360 --> 00:22:11.397 the worlds where 00:22:11.480 --> 00:22:15.597 we don't build enough infrastructure, 00:22:15.680 --> 00:22:20.197 I think you get a crazy concentration of power and concentration of compute 00:22:20.280 --> 00:22:23.717 because people will just bid the price up and up and up. So, the only thing 00:22:23.800 --> 00:22:28.277 that I really believe in as a long-term democratization strategy is to just make 00:22:28.360 --> 00:22:32.997 so much AI infrastructure available and make the models so good and so capable 00:22:33.080 --> 00:22:35.877 um, that we hope to get to a world where people are like, 00:22:35.960 --> 00:22:38.357 "I need help coming up with ideas for what to use all this compute for." I 00:22:38.440 --> 00:22:39.557 don't think we will in practice. 00:22:39.640 --> 00:22:44.797 >> Yeah. >> Um, but I certainly think that if compute is very limited, 00:22:44.880 --> 00:22:47.517 you know, the richest people and the richest companies in the world will just 00:22:47.600 --> 00:22:50.277 sort of bid up the price to a kind of extreme degree. 00:22:50.360 --> 00:22:52.557 >> It'll be another kind of scarcity that 00:22:52.640 --> 00:22:55.437 that is um, monopolized. Where So, this 00:22:55.520 --> 00:22:59.717 is a more data centers is actually a very egalitarian 00:22:59.800 --> 00:23:04.837 um, initiative in the sense that they can make AI access more widespread. 00:23:04.920 --> 00:23:08.277 >> I I would look at many examples throughout history. You know, one of the 00:23:08.360 --> 00:23:12.557 best things that we ever did for really increasing people's quality of life was 00:23:12.640 --> 00:23:14.997 to drive the price of electricity way down. 00:23:15.080 --> 00:23:17.117 >> Mhm. >> Price of energy broadly way down. 00:23:17.200 --> 00:23:20.037 >> Mhm. >> Um energy correlates incredibly well 00:23:20.120 --> 00:23:22.557 with quality of life, or at least it has for a long time. Maybe now it'll be more 00:23:22.640 --> 00:23:26.037 about AI. Um and 00:23:26.120 --> 00:23:31.597 by making energy abundant and shockingly cheap relative to what it cost 50, 200 00:23:31.680 --> 00:23:35.637 years ago, uh that has done like quite a bit for 00:23:35.720 --> 00:23:38.317 lifting the entire world up. I think we need to do the same thing with AI. 00:23:38.400 --> 00:23:40.277 >> Mhm. >> And I think that means you need a lot of 00:23:40.360 --> 00:23:43.237 it, and just like with energy, you need to innovate new ways to make it much 00:23:43.320 --> 00:23:48.797 more affordable. >> Yeah. Can I give give a thought on on this whole 00:23:48.880 --> 00:23:51.357 issue as well? I think we've talked a little bit about 00:23:51.440 --> 00:23:55.557 uh you know, the broad access to AI, which I think is very important and and 00:23:55.640 --> 00:23:58.197 can create like you know, potentially a lot of new 00:23:58.280 --> 00:24:01.077 products that will be um very helpful 00:24:01.160 --> 00:24:02.997 for the world, very useful for people. 00:24:03.080 --> 00:24:05.437 Um I think one thing that, you know, we 00:24:05.520 --> 00:24:07.397 try to think about as well 00:24:07.480 --> 00:24:12.157 in in writing this blueprint is 00:24:12.240 --> 00:24:15.717 you know, ensuring that, you know, ordinary people who maybe, you know, 00:24:15.800 --> 00:24:19.837 aren't going to uh start start start up necessarily um all of them like aren't 00:24:19.920 --> 00:24:22.197 left behind by this technology. Um I 00:24:22.280 --> 00:24:26.957 think this is related to to uh something that that we might talk about, which is 00:24:27.040 --> 00:24:28.197 kind of 00:24:28.280 --> 00:24:31.437 how does AI change the composition of 00:24:31.520 --> 00:24:35.437 the economy? Does it uh move it more towards 00:24:35.520 --> 00:24:37.917 uh capital or towards like 00:24:38.000 --> 00:24:40.397 labor, but labor that that is done by by 00:24:40.480 --> 00:24:43.117 AI really. Um and and one of the things 00:24:43.200 --> 00:24:48.077 that we talk about in the blueprint is uh how do you monitor modernize the the 00:24:48.160 --> 00:24:52.997 the tax base for an economy that is like this uh uh and how do you distribute 00:24:53.080 --> 00:24:56.357 kind of uh this prosperity that, you know, will 00:24:56.440 --> 00:24:59.877 be created by this technology? Uh how do you make sure that this is prosperity 00:24:59.960 --> 00:25:02.637 for everyone and not like wealth for, 00:25:02.720 --> 00:25:05.477 you know, a relatively few few people. 00:25:05.560 --> 00:25:09.397 And so, this is something that we try to address in in the blueprint, but I'm 00:25:09.480 --> 00:25:12.597 curious if we can talk about it here as well. 00:25:12.680 --> 00:25:16.797 >> we totally should. A couple of the big ideas that I saw were 00:25:16.880 --> 00:25:20.837 have workers co-authoring how AI is deployed in their workspaces. I'd like 00:25:20.920 --> 00:25:24.397 to to ask you about that, Josh, and I'd also ask I'd like to ask all you about 00:25:24.480 --> 00:25:29.037 more broadly how society can capture the upside. What are the institutional forms 00:25:29.120 --> 00:25:31.917 that'll that'll take, but let's start with you, Josh. 00:25:32.000 --> 00:25:34.877 >> Sure. Just on the on this compute question for 1 second, I want to say the 00:25:34.960 --> 00:25:37.757 compute allocation problem, figuring out 00:25:37.840 --> 00:25:42.077 what things we use compute to help people do with AI will probably one of 00:25:42.160 --> 00:25:45.597 the most important society-wide questions to navigate over the next few 00:25:45.680 --> 00:25:49.117 years while compute is relatively scarce, and we should try to boost the 00:25:49.200 --> 00:25:52.997 amount of compute in the world as much and as quickly as we possibly can so we 00:25:53.080 --> 00:25:55.917 don't have to have painful trade-off questions where there's some 00:25:56.000 --> 00:25:59.357 extraordinary good we could provide to everyone, and we're stuck with a hard 00:25:59.440 --> 00:26:01.677 question where someone says, "Well, how are you going to pay for that?" because 00:26:01.760 --> 00:26:05.997 the cost of compute is so high. Um on getting workers involved in AI, 00:26:06.080 --> 00:26:09.757 um I I actually I kind of want to back up and just acknowledge an elephant in 00:26:09.840 --> 00:26:13.197 the room, which is that a lot of workers are concerned about AI. They're worried 00:26:13.280 --> 00:26:17.077 about what AI means for them. They are not immediately excited at the prospect 00:26:17.160 --> 00:26:19.717 of of figuring out, all right, how are we going to use AI in our workplace? And 00:26:19.800 --> 00:26:25.037 they're thinking, "Oh my gosh, is the AI going to replace me?" Um and what I what 00:26:25.120 --> 00:26:27.797 I think is sort of the important first step here is that those of us who are 00:26:27.880 --> 00:26:30.637 working on AI and who are kind of the stewards for this 00:26:30.720 --> 00:26:34.037 have to be putting out things like this this blueprint document where we say, 00:26:34.120 --> 00:26:36.997 "Well, here's how we're going to advocate for policies to make sure that 00:26:37.080 --> 00:26:41.517 the economy is fair and that you are supported, no matter what." And then, 00:26:41.600 --> 00:26:46.637 given that we've made this level of safety net, now we can talk about, you know, 00:26:46.720 --> 00:26:49.477 something where where you have confidence in in us, we can talk 00:26:49.560 --> 00:26:53.837 together, have a good conversation, Um, we should figure out how do we empower 00:26:53.920 --> 00:26:58.117 unions to make wise choices about where and how to use AI. Um, how do we empower 00:26:58.200 --> 00:27:01.317 workers to participate in conversations about the acceptable use of AI in the 00:27:01.400 --> 00:27:04.957 workplace. I think a lot of folks are rightly concerned about AI surveillance 00:27:05.040 --> 00:27:07.637 in the workplace, making sure that workers are a part of the decisions 00:27:07.720 --> 00:27:08.917 about those types of things feels very 00:27:09.000 --> 00:27:11.877 important. And um, a big push on AI 00:27:11.960 --> 00:27:15.397 literacy to ensure that folks um, 00:27:15.480 --> 00:27:19.077 you know, get the tools they need to use AI to make their lives better, uh, to 00:27:19.160 --> 00:27:22.677 start small businesses, to to do all the kinds of things that, you know, will 00:27:22.760 --> 00:27:24.117 help them realize their potential. 00:27:24.200 --> 00:27:28.077 >> For sure. Um, what are some institutions that you've 00:27:28.160 --> 00:27:33.237 considered that allow everyone to capture some of the upside, Josh? 00:27:33.320 --> 00:27:36.117 >> Um, so 00:27:36.200 --> 00:27:40.357 let's see. For for new kinds of institutions, I think state capacity to 00:27:40.440 --> 00:27:43.957 measure pieces of the economy in, you know, greater granular depth so that 00:27:44.040 --> 00:27:46.637 there can be responses if there's, you know, like an economic shift in one 00:27:46.720 --> 00:27:49.477 place or another. >> yeah. >> Um, I think AI's actually a very 00:27:49.560 --> 00:27:52.357 exciting tool for making that more scalable and less expensive than it 00:27:52.440 --> 00:27:53.837 would have been in the past. >> Yeah. 00:27:53.920 --> 00:27:57.637 >> Uh, I I also think that there can be more institutions that are kind of in 00:27:57.720 --> 00:28:02.397 between corporations and governments, which have very very different levels of 00:28:02.480 --> 00:28:05.437 accountability governance-wise. There maybe is like a need for something with 00:28:05.520 --> 00:28:09.317 like more in between level of accountability. Um, 00:28:09.400 --> 00:28:14.677 that can provide uh, you know, like social safety net type 00:28:14.760 --> 00:28:18.557 services or >> Something between a corporate board and the regulator. 00:28:18.640 --> 00:28:22.637 >> Yeah, yeah, yeah. Like there's um, 00:28:22.720 --> 00:28:25.757 like this can't all just happen in private companies that have very minimal 00:28:25.840 --> 00:28:29.717 governance. Uh, and we also don't expect that government, which moves fairly 00:28:29.800 --> 00:28:32.597 slowly, is going to do all of it quite quickly. And so we sort of maybe need 00:28:32.680 --> 00:28:36.157 some in-between institutions that can help us prototype things. Um, this is 00:28:36.240 --> 00:28:39.637 like a little bit spitballing. Uh, and I I I don't think this is like quite 00:28:39.720 --> 00:28:42.797 covered in the blueprint, but uh, you know, you asked and so that's I think in 00:28:42.880 --> 00:28:44.917 the innovative new institutions that I think we could do. 00:28:45.000 --> 00:28:48.197 >> Cool. Um, Sam, institutionally, how are 00:28:48.280 --> 00:28:53.920 you envisioning that people might broadly participate in the upside? 00:28:56.000 --> 00:28:59.597 >> I think we talked earlier about very broad 00:28:59.680 --> 00:29:04.557 access and giving a lot of people a lot of compute. Um, I 00:29:04.640 --> 00:29:08.357 you you hear these ideas like universal basic compute or other things with nice 00:29:08.440 --> 00:29:13.477 branding, but really what they mean is just like uh 00:29:13.560 --> 00:29:16.557 instead of the traditional thinking of we're going to give people, you know, a 00:29:16.640 --> 00:29:19.797 monthly stipend or money or whatever when AI does all the jobs, I think it's 00:29:19.880 --> 00:29:23.157 way better to say, "Actually, people are pretty good about knowing what they need 00:29:23.240 --> 00:29:26.197 and pretty creative about figuring out how to use things." But if people are 00:29:26.280 --> 00:29:30.237 boxed out of access to this resource, uh that will be 00:29:30.320 --> 00:29:32.717 a a challenge. I do suspect that we're 00:29:32.800 --> 00:29:37.477 going to have to make changes to how we tax like in a world where AI is doing 00:29:37.560 --> 00:29:40.917 most of the intellectual work in the world or at least of the work of today. 00:29:41.000 --> 00:29:43.077 Um, you know, we probably are going to need 00:29:43.160 --> 00:29:46.757 to explore some way to tax that instead of taxing human income in the 00:29:46.840 --> 00:29:50.197 traditional sense. Um, I suspect that we will need to 00:29:50.280 --> 00:29:53.997 provide new kinds of transition assistance, unemployment insurance, 00:29:54.080 --> 00:29:57.997 things like that. Um, and I suspect that eventually we will need to think about 00:29:58.080 --> 00:30:02.717 how people get to be an owner in the upside of all of this 00:30:02.800 --> 00:30:06.317 in new ways. Um you know, capitalism is dependent on a 00:30:06.400 --> 00:30:10.077 certain balance between labor and capital and if that gets totally out of 00:30:10.160 --> 00:30:13.037 whack, then the current system is not going to work and there'll have to be 00:30:13.120 --> 00:30:15.557 some sort of evolution. What that is I 00:30:15.640 --> 00:30:19.397 think is a very open question and again 00:30:19.480 --> 00:30:22.717 part of the goal here is to throw out some ideas, but there's many more. 00:30:22.800 --> 00:30:27.557 Um and I will always leave some room and say maybe we're wrong and maybe no 00:30:27.640 --> 00:30:30.877 change at all is required and somehow this just works differently than we think. 00:30:30.960 --> 00:30:34.517 But again, in a spirit of trying to use 00:30:34.600 --> 00:30:38.757 the time we have to think and debate, it seems like a good time to start putting 00:30:38.840 --> 00:30:41.077 ideas here. >> Can I say something on this maybe we're 00:30:41.160 --> 00:30:43.237 wrong aspect? Like I I think one thing 00:30:43.320 --> 00:30:45.757 that you know in the in the blueprint we 00:30:45.840 --> 00:30:50.677 we have you know proposals around like modernizing the the tax base and around 00:30:50.760 --> 00:30:54.597 like uh we talk about maybe even a 32-hour work 00:30:54.680 --> 00:30:56.677 week and and these types of things. I 00:30:56.760 --> 00:31:01.717 think there's there's something um important here about 00:31:01.800 --> 00:31:04.757 uh trying to create counter-cyclical 00:31:04.840 --> 00:31:10.717 measures where you know conditional on disruption from uh from AI we have kind 00:31:10.800 --> 00:31:15.237 of uh additional like unemployment insurance uh we have these measures like 00:31:15.320 --> 00:31:19.917 the um 32-hour work week but that we um 00:31:20.000 --> 00:31:24.877 it's I think fairly important to me that you know some of these measures uh 00:31:24.960 --> 00:31:27.797 maybe are maybe some of these measures are good in the current world but I 00:31:27.880 --> 00:31:30.677 think many of them would be quite disruptive and we're really talking 00:31:30.760 --> 00:31:35.877 about a world uh that uh that that changes a lot and so like to me 00:31:35.960 --> 00:31:37.397 institutionally something that we need 00:31:37.480 --> 00:31:41.557 is kind of um some thinking about 00:31:41.640 --> 00:31:45.757 what are potential disruptions that could occur from AI and 00:31:45.840 --> 00:31:50.277 what are things that we can implement as we as we see those coming 00:31:50.360 --> 00:31:54.037 uh to kind of uh counter the disruptions and uh 00:31:54.120 --> 00:31:58.637 distribute benefits uh broadly uh at that time. Um 00:31:58.720 --> 00:32:01.357 >> Yeah and I saw one of the ideas in the report was portable benefits since 00:32:01.440 --> 00:32:04.957 benefits are so linked to employers in America now. It sounded like there were 00:32:05.040 --> 00:32:07.197 a lot of >> Correct. Yeah yeah. And this was this 00:32:07.280 --> 00:32:11.757 one I think an example of the one of the more kind of uh US-focused uh proposals 00:32:11.840 --> 00:32:16.157 of course uh but um uh but yeah >> it's a great idea. I think it's insane 00:32:16.240 --> 00:32:20.597 we don't already have that and that like the way that the US benefits world has 00:32:20.680 --> 00:32:22.957 evolved is really bad but I think it's a great idea. 00:32:23.040 --> 00:32:26.037 >> No no one should lose their health care if they lose their job. Like that just 00:32:26.120 --> 00:32:30.477 shouldn't happen. >> Agreed. Um so Adrian from inside the 00:32:30.560 --> 00:32:34.437 research organization you're seeing this since acceleration it's real for you you 00:32:34.520 --> 00:32:39.157 can see some some um scientific um kind of 00:32:39.240 --> 00:32:41.957 progress that's being made with these models. 00:32:42.040 --> 00:32:45.517 Um what's what's our We want institutions. We want 00:32:45.600 --> 00:32:47.317 society to keep pace with technology. 00:32:47.400 --> 00:32:50.397 What what do you think the window is for adaptation? 00:32:50.480 --> 00:32:54.797 >> So, you know, it's it's a Well, I was going to say it's hard to 00:32:54.880 --> 00:32:58.877 put a number. I would say that there's a lot of uncertainty about these numbers. 00:32:58.960 --> 00:33:01.757 Uh we've talked uh I think about 00:33:01.840 --> 00:33:06.917 uh having an automated researcher in 2028 or late 2028. Uh and I I want 00:33:07.000 --> 00:33:08.637 >> 2028 is the official goal. >> What's that? 00:33:08.720 --> 00:33:12.997 >> March of 2028 is the >> March, thank you. Um uh and 00:33:13.080 --> 00:33:17.757 I I think one useful thing to to think about here is once you have this 00:33:17.840 --> 00:33:20.797 automated researcher, which is an automated like AI researcher capable of 00:33:20.880 --> 00:33:23.557 doing AI research, you potentially have kind of like a 00:33:23.640 --> 00:33:27.397 double whammy of disruption, I would say. Like, first of all, you have a 00:33:27.480 --> 00:33:30.437 model that's capable of advanced 00:33:30.520 --> 00:33:34.517 cognitive work, clearly, which which AI research is. And so, that in itself is a 00:33:34.600 --> 00:33:37.037 disruption. It may accelerate 00:33:37.120 --> 00:33:40.517 uh further AI progress. And so, um 00:33:40.600 --> 00:33:44.517 you know, I I can't tell you, you know, after that point exactly how much 00:33:44.600 --> 00:33:48.477 progress we'll have made like a year from then, but it's probably, you know, 00:33:48.560 --> 00:33:52.557 more than the the pace of progress that we've been making so far. And so, uh 00:33:52.640 --> 00:33:55.477 this is kind of the the type of window that uh that we're talking about, I 00:33:55.560 --> 00:33:58.517 mean, you. >> Yeah, thank you. Okay, so we are at the 00:33:58.600 --> 00:34:02.437 point in the conversation where we open up questions. It's not just me anymore. 00:34:02.520 --> 00:34:05.117 It's uh members of the community. So, 00:34:05.200 --> 00:34:08.757 I've got one here. Um it's from Svetlana 00:34:08.840 --> 00:34:11.677 Romanova. As AI becomes more capable, 00:34:11.760 --> 00:34:15.797 what which human qualities do you think will matter most in the future? I'd like 00:34:15.880 --> 00:34:18.477 to ask each of you one by one, Josh. 00:34:18.560 --> 00:34:22.117 >> Um you know, I think there are human qualities that are timeless and that 00:34:22.200 --> 00:34:27.797 will always matter. Uh character, commitment, effort, um compassion. 00:34:27.880 --> 00:34:29.117 I I think they're going to matter a lot. 00:34:29.200 --> 00:34:33.437 >> Mhm. Yeah. Sam. >> All of those, creativity, understanding 00:34:33.520 --> 00:34:36.117 what other people want, but I will share a recent anecdote that really struck 00:34:36.200 --> 00:34:40.637 home for me that things don't quite go the same way. I went to my first robot cafe 00:34:40.720 --> 00:34:43.437 for the first like I was so excited to try. I thought I was going to love it. 00:34:43.520 --> 00:34:46.357 >> Mhm. >> And it was the most underwhelming 00:34:46.440 --> 00:34:49.957 experience. And I thought I was someone that did not need like the barista at 00:34:50.040 --> 00:34:52.957 Starbucks to smile at me and say hi and ask how my day was going. I really 00:34:53.040 --> 00:34:57.117 thought I was like didn't care about that. It turns out that I really want 00:34:57.200 --> 00:34:59.997 that and walking in to like push on the screen >> Mhm. 00:35:00.080 --> 00:35:02.917 >> and have the robot do the thing and give you a delicious cup of coffee 00:35:03.000 --> 00:35:07.317 >> Yeah. >> was like a deeply unfulfilling I don't want this experience. 00:35:07.400 --> 00:35:09.357 >> Yeah. >> And I thought I wouldn't have cared. 00:35:09.440 --> 00:35:12.837 >> I totally just those small interactions throughout the day. I 00:35:12.920 --> 00:35:15.437 I really appreciate them, too. Okay. 00:35:15.520 --> 00:35:18.957 Here Here's one. Should those >> is >> Oh, that's fine. That's fine. 00:35:19.040 --> 00:35:20.757 >> No, go for it. >> Well, I I mean I think what I was going 00:35:20.840 --> 00:35:23.797 to say is a little bit along the lines of of what Sam said. And And I believe 00:35:23.880 --> 00:35:27.917 it's in the blueprint like uh humans 00:35:28.000 --> 00:35:29.997 need each other. Like I think we care 00:35:30.080 --> 00:35:35.877 about other humans. Um in fact, you know, to some extent uh one of the most 00:35:35.960 --> 00:35:41.357 uh uh scary things with the development of of technology and and social media, 00:35:41.440 --> 00:35:46.237 video games, and stuff like that is like maybe uh that is gotten us to lose a 00:35:46.320 --> 00:35:48.917 little bit of the connection that that we used to have. But I I I still think 00:35:49.000 --> 00:35:52.477 that it's something that matters tremendously to people and that they 00:35:52.560 --> 00:35:55.517 will recognize this more and more in fact uh 00:35:55.600 --> 00:36:00.437 as as AI becomes uh becomes more advanced and can do some of some of 00:36:00.520 --> 00:36:04.277 these other tasks that don't require human connection. And so uh that's like 00:36:04.360 --> 00:36:07.277 one big thing that I believe will um 00:36:07.360 --> 00:36:09.517 will expand. It seems like a good thing 00:36:09.600 --> 00:36:14.677 to me, right? That like this is um this can be of course like a type of work, 00:36:14.760 --> 00:36:18.877 right? You know, nursing and and uh and all 00:36:18.960 --> 00:36:23.197 these these types of works in the care economy, teaching. Um 00:36:23.280 --> 00:36:27.157 and but you know, it's also just important to to our lives, right? And I 00:36:27.240 --> 00:36:31.717 think that's to me like going to be the the big quality, like how good you are 00:36:31.800 --> 00:36:35.957 as a person to other people. >> Yes, I agree. Um okay, so in America 00:36:36.040 --> 00:36:37.517 things like health care, child care, and 00:36:37.600 --> 00:36:40.677 elder care are very expensive. Um here's 00:36:40.760 --> 00:36:43.397 a question. How can AI expand access to 00:36:43.480 --> 00:36:47.237 those for everyday people? Josh? 00:36:47.320 --> 00:36:50.837 >> Um one thing I am so excited about is 00:36:50.920 --> 00:36:55.597 the way that AI can help provide the best uh health care in the world to 00:36:55.680 --> 00:36:58.557 everybody. Um I've heard a lot of stories from 00:36:58.640 --> 00:37:02.397 folks navigating the health care system that they didn't know where to go, um 00:37:02.480 --> 00:37:04.557 how to navigate the insurance, 00:37:04.640 --> 00:37:09.437 uh what specialist to talk to, and you hear stories of people bouncing around 00:37:09.520 --> 00:37:12.517 for a diagnosis for years and years and not getting anywhere. You hear about 00:37:12.600 --> 00:37:16.117 folks who are who are stuck in a health care system where even if you have the 00:37:16.200 --> 00:37:20.437 most caring, compassionate nurses and doctors, there's not enough time to give 00:37:20.520 --> 00:37:23.717 everyone the best quality care. And I think that AI is going to make it 00:37:23.800 --> 00:37:27.637 possible to deliver the best quality care at scale. I don't think it's going 00:37:27.720 --> 00:37:31.037 to replace doctors. I think it's going to make their workloads manageable, and 00:37:31.120 --> 00:37:33.557 I think it's going to make it possible for patients to get the best experience 00:37:33.640 --> 00:37:35.557 possible. >> Yeah. I agree with that. A lot of 00:37:35.640 --> 00:37:39.157 patient empowerment out there. So, yeah. 00:37:39.240 --> 00:37:43.877 >> Uh we have all told stories many times of things we've seen on uh social media 00:37:43.960 --> 00:37:47.117 about people having this like amazing health care experience. Uh 00:37:47.200 --> 00:37:49.637 I don't have anything that amazing, but I'll tell my own because at least it 00:37:49.720 --> 00:37:53.757 happened to me. Uh I recently got a blood test. Uh there were like nothing 00:37:53.840 --> 00:37:57.317 seriously wrong, but a few markers are just like kind of out of their range, 00:37:57.400 --> 00:37:59.037 and you know, you kind of scan down. 00:37:59.120 --> 00:38:01.917 There's like 100 things, and you just look at the list of new tests and the 00:38:02.000 --> 00:38:04.877 ones that are like a little bit out of range. I asked my doctor, he was like, 00:38:04.960 --> 00:38:07.637 uh you know, probably those are all close enough, it's fine. I put it in 00:38:07.720 --> 00:38:10.557 ChatGPT, it said, yeah, you're fine, but like here's what's going on, take this 00:38:10.640 --> 00:38:12.957 one supplement, get your blood test again in a month. 00:38:13.040 --> 00:38:15.877 >> Mhm. >> You should be okay. And it did. Again, I 00:38:15.960 --> 00:38:20.397 wasn't like that sick, I wasn't but like the fact that I could just like upload 00:38:20.480 --> 00:38:24.197 my blood test and instantly get the right answer. For like a kind of complex thing. 00:38:24.280 --> 00:38:27.837 >> Yeah. >> Was an amazing experience. Uh and I think there will be many things like 00:38:27.920 --> 00:38:31.757 that across health care education and all these other areas, elder care. You 00:38:31.840 --> 00:38:34.397 know, that's one where I think we want people doing that. 00:38:34.480 --> 00:38:37.197 >> Very much. Yeah. >> But there will be a lot of ways we can 00:38:37.280 --> 00:38:38.997 really drive down the cost of >> Yeah. 00:38:39.080 --> 00:38:40.797 >> um health care and education and things like that. 00:38:40.880 --> 00:38:43.077 >> For sure. >> And even personalized learning like it 00:38:43.160 --> 00:38:47.557 can still be humans but you can figure out how to teach individuals, right? Um 00:38:47.640 --> 00:38:48.837 you know, how do you think about it? 00:38:48.920 --> 00:38:52.837 >> Yeah, I mean, you know, I would say similar things of course. I mean, 00:38:52.920 --> 00:38:56.517 another basic thing in terms of health care, you know, that I hope at least is 00:38:56.600 --> 00:39:01.517 is simply that, you know, AI will be able to help medical research and that, 00:39:01.600 --> 00:39:04.957 you know, that that's like a basic thing that that would make 00:39:05.040 --> 00:39:09.637 health care better for people. But also, you know, I just talked about 00:39:09.720 --> 00:39:13.637 the care economy and and I think to some extent, you know, to the extent that AI 00:39:13.720 --> 00:39:15.717 can um 00:39:15.800 --> 00:39:18.597 facilitate some of the of the more, you 00:39:18.680 --> 00:39:20.877 know, bureaucratic, frankly, aspects of 00:39:20.960 --> 00:39:24.437 of health care and and help us have more 00:39:24.520 --> 00:39:27.637 people actually, you know, providing the 00:39:27.720 --> 00:39:31.597 the actual the actual health care 00:39:31.680 --> 00:39:34.757 to to individuals. That that seems like 00:39:34.840 --> 00:39:38.957 like a a positive thing. Like it's not AI would be providing 00:39:39.040 --> 00:39:44.397 would be doing this. We would be freeing up people to to actually do this work. 00:39:44.480 --> 00:39:46.797 >> Yeah. >> We can hope. >> We we've talked a lot about the 00:39:46.880 --> 00:39:50.797 capability overhang this year. How it AI has 00:39:50.880 --> 00:39:54.917 capable of so many things that most folks are not leaning into it for. And I 00:39:55.000 --> 00:39:58.237 sometimes suspect that it's because they're resigned to the world like it 00:39:58.320 --> 00:40:02.277 is, right? I think PG calls it schlepp blindness, right? So they're they're 00:40:02.360 --> 00:40:06.797 they have ceased to consider that something better is possible. Um 00:40:06.880 --> 00:40:08.637 so I think they're going to start 00:40:08.720 --> 00:40:14.077 leaning in more, right? Um and and I think when they lean in that 00:40:14.160 --> 00:40:17.197 their behavior is actually what will change society as much as any 00:40:17.280 --> 00:40:22.117 technological breakthrough. Just that moment of adoption. Um as you in your 00:40:22.200 --> 00:40:24.917 lives what are the moments where you've seen 00:40:25.000 --> 00:40:26.677 somebody kind of light up and realize, 00:40:26.760 --> 00:40:33.477 oh I I can do this. I don't have to live resigned anymore to the old ways? 00:40:33.560 --> 00:40:35.957 >> Um wait, I've taken a lot of questions first. Adrian, do you want to take a 00:40:36.040 --> 00:40:39.237 crack at this? >> Now you get time to think about it. 00:40:39.320 --> 00:40:42.797 >> Oh my god. Wow. Well, >> Yeah, putting you on the spot. 00:40:42.880 --> 00:40:46.557 >> Yeah, maybe I maybe I maybe I do want time to think about this, but 00:40:46.640 --> 00:40:50.480 you know, I mean to your point, I think it 00:40:50.840 --> 00:40:54.797 I think it's definitely the case that uh that people take a while to adapt to 00:40:54.880 --> 00:40:56.037 new technologies. I think to some 00:40:56.120 --> 00:40:59.197 extent, you know, the the capability 00:40:59.280 --> 00:41:04.157 overhang might be large in terms of capabil- capabilities versus how much 00:41:04.240 --> 00:41:05.757 people are actually using them. 00:41:05.840 --> 00:41:08.757 Um I think there is an extent to which 00:41:08.840 --> 00:41:10.517 that's a function 00:41:10.600 --> 00:41:13.517 of how fast the capabilities are 00:41:13.600 --> 00:41:18.477 improving versus how fast people are used to, you know, things getting uh 00:41:18.560 --> 00:41:21.077 things getting better. >> I have a favorite to mention. 00:41:21.160 --> 00:41:24.117 >> Yeah. >> My favorite There are many options here, 00:41:24.200 --> 00:41:25.797 but my favorite of all is watching 00:41:25.880 --> 00:41:31.157 parents who are coders by training watch watching the parent watch their kids use 00:41:31.240 --> 00:41:33.357 Codex >> Mhm. >> for the first time. 00:41:33.440 --> 00:41:38.277 >> Mhm. >> And uh a kid who has a bunch of ideas and no idea about like what the 00:41:38.360 --> 00:41:41.757 traditional limits are or what would be hard or easy, just start describing a 00:41:41.840 --> 00:41:44.757 video game and having Codex make it. 00:41:44.840 --> 00:41:49.117 Uh And the the kind of like creative journey the kid goes through. I mean, 00:41:49.200 --> 00:41:51.237 often you see the kid doing this mostly by voice. 00:41:51.320 --> 00:41:52.957 >> Yeah. >> And the parent is just like, that's not 00:41:53.040 --> 00:41:55.757 going to work, and then it works, and then they're like, wow, my kid is going 00:41:55.840 --> 00:42:00.037 to grow up in a world where he or she just like expects this this happens. 00:42:00.120 --> 00:42:03.197 Like and I kind of still like I wouldn't even thought to try that cuz I would've 00:42:03.280 --> 00:42:04.597 been so certain it didn't work. 00:42:04.680 --> 00:42:07.477 >> Yeah. >> So like watching watching it through the parent's eyes 00:42:07.560 --> 00:42:10.117 the kid do it for the first time, especially if the parent is a software 00:42:10.200 --> 00:42:12.037 engineer by training, is awesome. 00:42:12.120 --> 00:42:15.197 >> Yeah, especially. I mean, this is a tale as old as time, right? Like the kids 00:42:15.280 --> 00:42:18.957 always know how to use the the VCR in the '80s or whatever, right? 00:42:19.040 --> 00:42:23.037 Uh and yeah. Interesting. >> That's That's funny. 00:42:23.120 --> 00:42:25.877 >> Um I just on the capability overhang and 00:42:25.960 --> 00:42:30.077 how fast people notice when a capability has arrived, there's an interesting time 00:42:30.160 --> 00:42:34.797 scale mismatch where people have like a who who aren't super in the know on AI 00:42:34.880 --> 00:42:37.357 have this like distant awareness that something is happening. They know 00:42:37.440 --> 00:42:40.797 there's a product out there. Once every few months, they might check it out. 00:42:40.880 --> 00:42:44.157 They don't immediately and instinctively probe it to the maximal extent of its 00:42:44.240 --> 00:42:47.237 capabilities. Uh and they often don't put it on the 00:42:47.320 --> 00:42:49.757 thinking setting. Like they don't know that reasoning models have happened. 00:42:49.840 --> 00:42:53.517 They stay on the default chat model that's that's right out there. And so 00:42:53.600 --> 00:42:55.917 they wind up with this misperception that things aren't moving as fast as 00:42:56.000 --> 00:42:58.997 they are. And you hear people talk about, well, there's hallucinations, 00:42:59.080 --> 00:43:02.637 it's slop, it's making mistakes, it's inaccurate. Why would they ever tell us 00:43:02.720 --> 00:43:03.957 that it's going to do, you know, these 00:43:04.040 --> 00:43:06.917 great things? And uh this like visceral 00:43:07.000 --> 00:43:11.397 belief gap is I think an issue that will get overcome when they start to see 00:43:11.480 --> 00:43:15.357 other folks and institutions very successfully use AI at sort of the 00:43:15.440 --> 00:43:20.197 maximal reasoning settings, at the most capable settings, in ways that are 00:43:20.280 --> 00:43:23.877 shocking to them. Um like this video game example, but sort of like at scale 00:43:23.960 --> 00:43:28.317 for society as a whole. Uh they're going to see a lot of people get diagnoses 00:43:28.400 --> 00:43:32.037 that, you know, they wouldn't have expected that someone could get quickly. 00:43:32.120 --> 00:43:34.557 And it's going to update them. And I think I think that's just like an 00:43:34.640 --> 00:43:38.317 interesting phenomenon that the the time scale for AI progress is weeks and 00:43:38.400 --> 00:43:41.757 months. And the time scale for people to currently checking back is like every 00:43:41.840 --> 00:43:45.677 half year or something. And uh yeah, some big change will happen when 00:43:45.760 --> 00:43:48.717 people um realize the maximal extent of capabilities today. 00:43:48.800 --> 00:43:52.717 >> Yeah. Um I agree. The uh the point you brought up, Sam, 00:43:52.800 --> 00:43:57.317 about the kids really illustrates for me what an advantage 00:43:57.400 --> 00:44:01.237 creative people have. Like there are some folks that are just kind of a font 00:44:01.320 --> 00:44:04.597 of ideas. Many of those ideas have never been realized. Some are scientists, some 00:44:04.680 --> 00:44:07.917 are artists, many are children. Um and 00:44:08.000 --> 00:44:11.557 it feels like the floodgates are opening 00:44:11.640 --> 00:44:14.317 for them to realize more things. 00:44:14.400 --> 00:44:18.437 I think you mentioned that you would actually burn through your Codex list, 00:44:18.520 --> 00:44:21.157 and you're not having it run all night anymore. 00:44:21.240 --> 00:44:24.437 >> We just need to make a model that helps you come up with good new ideas. I 00:44:24.520 --> 00:44:28.437 actually think this will be one of the most exciting things to do. I don't 00:44:28.520 --> 00:44:31.077 think we're that far away from a model where you can say, "Go look through all 00:44:31.160 --> 00:44:34.997 my text messages, all my email, look at my entire computer, 00:44:35.080 --> 00:44:39.317 anything you can find about me, and just suggest ideas that I've like gestured at 00:44:39.400 --> 00:44:40.997 as fragments or that might be interesting to me." 00:44:41.080 --> 00:44:42.357 >> Yeah. >> And then I'll build those. 00:44:42.440 --> 00:44:45.917 >> Yeah. Yeah, agreed. A thought partner, I see a lot of people using it like that. 00:44:46.000 --> 00:44:47.637 Well, I think we're wrapping up here. 00:44:47.720 --> 00:44:50.797 So, thank you all for joining us today, 00:44:50.880 --> 00:44:55.557 and thank you, Sam, Josh, Adrian. This has been pretty cool. 00:44:55.640 --> 00:44:56.517 >> This is a lot of fun. Thank you. 00:44:56.600 --> 00:45:00.437 >> Yeah. So, the ideas discussed may sound 00:45:00.520 --> 00:45:04.797 ambitious to you, and they're meant to be. 00:45:04.880 --> 00:45:08.837 But, we know that they're also early and exploratory. We're offering them 00:45:08.920 --> 00:45:11.237 not as a final plan, but as a starting 00:45:11.320 --> 00:45:14.677 point in a very public conversation 00:45:14.760 --> 00:45:16.837 with policy makers and everyone in 00:45:16.920 --> 00:45:19.637 society to encourage more discussion and 00:45:19.720 --> 00:45:23.717 research and debate. OpenAI wants this 00:45:23.800 --> 00:45:28.637 conversation to continue. The company, we are inviting feedback 00:45:28.720 --> 00:45:34.680 through a new email address. It's newindustrialpolicy.openai.com. 00:45:36.040 --> 00:45:38.597 Please send us your best ideas. 00:45:38.680 --> 00:45:40.917 We're launching a pilot program of 00:45:41.000 --> 00:45:43.997 fellowships and focused research grants 00:45:44.080 --> 00:45:46.437 for up to $100,000 in funding and up to 00:45:46.520 --> 00:45:48.477 a million in API credits. And we're 00:45:48.560 --> 00:45:51.477 convening further discussions at the new 00:45:51.560 --> 00:45:56.277 OpenAI workshop opening in Washington, D.C. in May. So, thank you again for 00:45:56.360 --> 00:45:59.957 being part of this forum. We appreciate the time and thought that everyone is 00:46:00.040 --> 00:46:04.760 bringing to this conversation, and we look forward to continuing it.