WEBVTT 00:00:00.000 --> 00:00:05.397 and innovators like Alex himself who were at the frontier of developing this 00:00:05.480 --> 00:00:08.717 technology and I thought they needed to be in the 00:00:08.800 --> 00:00:12.837 room not least because it's businesses like yours that are have the resources 00:00:12.920 --> 00:00:14.437 to invest in this just to give you a bit 00:00:14.520 --> 00:00:17.237 of context this year you know the the 00:00:17.320 --> 00:00:20.757 large AI companies are going to spend 20 00:00:20.840 --> 00:00:24.717 times more on developing this technology than the 00:00:24.800 --> 00:00:28.917 US did for the Manhattan Project. Can you think about seminal moments where 00:00:29.000 --> 00:00:33.997 the government mobilized to do something 20 times more just this year alone. So I 00:00:34.080 --> 00:00:37.637 thought look it's vital that we do this in a different way not like the G7 and 00:00:37.720 --> 00:00:40.917 G20 we bring industry innovators and and 00:00:41.000 --> 00:00:44.917 government together and that's that's what we did at Bletchley Park a few 00:00:45.000 --> 00:00:46.877 years ago. The summit series went to 00:00:46.960 --> 00:00:49.357 Paris, South Korea and now brilliantly 00:00:49.440 --> 00:00:52.557 in India a country with huge digital 00:00:52.640 --> 00:00:56.397 ambitions and indeed capacity. 00:00:56.480 --> 00:00:58.757 But I think over those years and Alex 00:00:58.840 --> 00:01:03.437 let's start with this. I think actually the the the debate on AI has shifted 00:01:03.520 --> 00:01:07.837 from kind of technology to strategy. 00:01:07.920 --> 00:01:12.157 You know from what these tools can do to what governments are choosing to do with 00:01:12.240 --> 00:01:15.557 them. And I wanted to maybe just start there. 00:01:15.640 --> 00:01:19.037 Is if you think about the last few years you are someone who is at the frontier 00:01:19.120 --> 00:01:22.357 of developing this technology. You know that the journey that we've been on in 00:01:22.440 --> 00:01:24.557 the last few years you know help us 00:01:24.640 --> 00:01:27.717 understand that where have we come 00:01:27.800 --> 00:01:30.037 what is it that is exciting at this this 00:01:30.120 --> 00:01:35.157 moment and just ground us in in where we are so that we can get get on to some of 00:01:35.240 --> 00:01:38.037 these questions about what we do with it. 00:01:38.120 --> 00:01:39.837 Yeah I think um 00:01:39.920 --> 00:01:42.637 I mean we are sitting here right now at 00:01:42.720 --> 00:01:45.677 such an incredibly exciting moment for 00:01:45.760 --> 00:01:50.437 the technology. I think we are um at the beginning of 00:01:50.520 --> 00:01:55.277 uh you know of a true acceleration in the technology and you know in many ways 00:01:55.360 --> 00:01:56.917 it's felt like this very clear 00:01:57.000 --> 00:01:59.677 acceleration even until now, but I think 00:01:59.760 --> 00:02:02.717 we're going to see sort of 2026 is an 00:02:02.800 --> 00:02:04.717 inflection point in many ways. I think 00:02:04.800 --> 00:02:09.877 to contextualize this over the past few years, I think from 00:02:09.960 --> 00:02:11.557 2018 00:02:11.640 --> 00:02:14.677 through 2024 00:02:14.760 --> 00:02:17.957 really, which is a good 6-year period, 00:02:18.040 --> 00:02:20.477 we were writing we were in the era of 00:02:20.560 --> 00:02:25.437 pre-training. So we were writing one very large exponential curve towards 00:02:25.520 --> 00:02:27.197 improving the performance of the models. 00:02:27.280 --> 00:02:30.797 And that curve was very predictable and it 00:02:30.880 --> 00:02:34.637 sort of had clear characteristics where the more resources you put in, the more 00:02:34.720 --> 00:02:38.117 results you would get out and there was a lot of questions at the time of, you 00:02:38.200 --> 00:02:41.277 know, whether or not this curve would continue returns and I think this was 00:02:41.360 --> 00:02:43.757 sort of one of the 00:02:43.840 --> 00:02:46.357 one of the more recent eras of AI 00:02:46.440 --> 00:02:50.397 skepticism. And then in 2024, 00:02:50.480 --> 00:02:55.237 like towards the end of 2024, began the era of reinforcement learning. 00:02:55.320 --> 00:02:57.317 So this was sort of the the next major 00:02:57.400 --> 00:02:59.757 wave that stacked on top of the previous 00:02:59.840 --> 00:03:02.397 wave and this is when we saw the models 00:03:02.480 --> 00:03:07.157 begin to learn to reason. And you know, there's always this meme of 00:03:07.240 --> 00:03:09.277 what did Ilya see? Well, Ilya saw 00:03:09.360 --> 00:03:11.597 reasoning from the models and it was 00:03:11.680 --> 00:03:14.997 very clearly this this next mega wave of 00:03:15.080 --> 00:03:20.317 of the technology and even reinforcement learning had, you know, clearly some 00:03:20.400 --> 00:03:24.117 limitations and there were all these questions around when that was going to, 00:03:24.200 --> 00:03:25.757 you know, how long that era was going to 00:03:25.840 --> 00:03:27.677 last and maybe there were limitations in 00:03:27.760 --> 00:03:31.157 that era. And then at the end of 2025, 00:03:31.240 --> 00:03:32.917 so just a few months ago, I think we've 00:03:33.000 --> 00:03:35.077 really begun 00:03:35.160 --> 00:03:38.677 an entirely new paradigm or era of the 00:03:38.760 --> 00:03:40.997 technology, which is 00:03:41.080 --> 00:03:45.757 really beginning the era of recursive self-improvement. And 00:03:45.840 --> 00:03:49.917 I think we're really starting to see this in the development of our models. I 00:03:50.000 --> 00:03:52.597 I you know, most of the major labs are seeing this in the development of their 00:03:52.680 --> 00:03:54.397 models, which is that the models 00:03:54.480 --> 00:03:57.957 themselves have now become instrumental 00:03:58.040 --> 00:04:02.877 in accelerating the process of producing the next AIs. And so, you know, the the 00:04:02.960 --> 00:04:06.477 sort of from the outside what you see is maybe an acceleration of overall 00:04:06.560 --> 00:04:10.677 development velocity, an acceleration in at which the speed at which the the new 00:04:10.760 --> 00:04:14.117 models are being released. Um internally, what we see is just dramatic 00:04:14.200 --> 00:04:16.717 speed ups. I think the sort of um 00:04:16.800 --> 00:04:18.877 the productivity of an individual 00:04:18.960 --> 00:04:21.757 researcher has grown dramatically. Um 00:04:21.840 --> 00:04:24.677 and we don't expect that to stop. In 00:04:24.760 --> 00:04:29.637 fact, we expect it to kind of smoothly continue accelerating um where the 00:04:29.720 --> 00:04:32.957 output of an individual researcher, an individual engineer, well, I think 00:04:33.040 --> 00:04:36.357 that itself will continue growing exponentially as the technology 00:04:36.440 --> 00:04:38.557 improves. And um 00:04:38.640 --> 00:04:43.637 the other mega wave that is somewhat coincident with this new paradigm within 00:04:43.720 --> 00:04:46.637 the sort of development of the technology of uh 00:04:46.720 --> 00:04:50.117 recursive self-improvement is really the 00:04:50.200 --> 00:04:54.277 arrival of agents in earnest. And um 00:04:54.360 --> 00:04:56.597 this is something that has been talked 00:04:56.680 --> 00:04:59.957 about since I honestly want to say 2023 00:05:00.040 --> 00:05:02.517 with a lot of fanfare. And for a long 00:05:02.600 --> 00:05:05.477 long time, it was sort of um one of 00:05:05.560 --> 00:05:09.677 these hype uh sort of uh buzzwords that 00:05:09.760 --> 00:05:13.637 you know, never actually lived up to the expectations. And I think also at the 00:05:13.720 --> 00:05:16.677 end of 2025, we and toward you know, 00:05:16.760 --> 00:05:19.517 middle to end of 2025, we started to see 00:05:19.600 --> 00:05:21.917 agents actually work. And this is you 00:05:22.000 --> 00:05:26.557 know, these are where the models can actually start taking um actions um and 00:05:26.640 --> 00:05:28.637 start automating entire workflows and 00:05:28.720 --> 00:05:32.037 being um sort of dramatically more uh 00:05:32.120 --> 00:05:35.157 productive as actual sort of entities 00:05:35.240 --> 00:05:39.477 unto themselves versus assistants. And um 00:05:39.560 --> 00:05:42.997 and this era is you know, it started with coding agents. We're starting to 00:05:43.080 --> 00:05:46.317 see it um really percolate through personal agents. This is a big part of 00:05:46.400 --> 00:05:49.237 what uh we believe in at Meta. 00:05:49.320 --> 00:05:52.277 And I think over the course of 2026, we 00:05:52.360 --> 00:05:56.077 will see large-scale agent deployments 00:05:56.160 --> 00:05:58.717 in many areas of the economy, many areas 00:05:58.800 --> 00:06:01.797 of the world. And 00:06:01.880 --> 00:06:07.597 the sort of, you know, if the the GDP of AI, so to speak, is going to grow 00:06:07.680 --> 00:06:11.077 exponentially. >> Well, that that is very welcome to to 00:06:11.160 --> 00:06:15.917 hear. I think this you hit on something that is important because 00:06:16.000 --> 00:06:19.277 I remember when I first started doing this in the job, the conversations were 00:06:19.360 --> 00:06:22.117 all around these scaling laws and would they continue? And you mentioned the 00:06:22.200 --> 00:06:24.157 skepticism. I mean, not only have they 00:06:24.240 --> 00:06:26.877 continued and continued to deliver this 00:06:26.960 --> 00:06:31.437 exponential growth. It feels very much like the last few months have seen an 00:06:31.520 --> 00:06:35.677 acceleration. That's what you touched on. It's what Dario and and Demis have 00:06:35.760 --> 00:06:40.077 also said to me and and publicly, driven by this recursive self-improvement that 00:06:40.160 --> 00:06:44.397 the models themselves are now able to help us develop the new models quicker. 00:06:44.480 --> 00:06:48.277 So, you know, we're at this quite exciting but also a moment of some 00:06:48.360 --> 00:06:51.277 anxiety because of that accelerating 00:06:51.360 --> 00:06:55.637 development time scale. So, I wanted to just then talk about trust. 00:06:55.720 --> 00:06:58.077 Because I I tell you give you a story. 00:06:58.160 --> 00:07:02.717 My I have two, you know, young girls and you know, as they're using their 00:07:02.800 --> 00:07:06.517 chatbots and things and they're always like, "Please, you know, so-and-so and 00:07:06.600 --> 00:07:10.197 so-and-so." And "Thank you." when they finish writing their request or they get 00:07:10.280 --> 00:07:14.157 the answer back. There's lots of pleases and and thank yous everywhere. And I and 00:07:14.240 --> 00:07:16.637 I I thought that was you know, interesting. I said to them, I 00:07:16.720 --> 00:07:20.317 said, "Look, this, you know, you don't need to do that. It's not not a person. 00:07:20.400 --> 00:07:24.357 It doesn't It doesn't matter. And And by the way, you know, it costs a fortune 00:07:24.440 --> 00:07:26.477 when we have to do the compute for that." 00:07:26.560 --> 00:07:30.357 And And they they said to me, "You know what, Daddy?" They said, "If If AI takes 00:07:30.440 --> 00:07:34.037 over the world, we want to have been kind to the AI." 00:07:34.120 --> 00:07:37.717 And I thought it was good good insurance policy. But it kind of you know, it 00:07:37.800 --> 00:07:39.877 touches on this anxiety. >> Very wise. 00:07:39.960 --> 00:07:43.397 >> Yeah. >> [laughter] >> In that maybe. 00:07:43.480 --> 00:07:46.637 You know, across the world, we're seeing these different attitudes 00:07:46.720 --> 00:07:51.117 uh towards AI. In countries like India, where we are, there's enormous optimism 00:07:51.200 --> 00:07:54.637 and and trust. And then actually, in Western countries, you know, we're 00:07:54.720 --> 00:07:56.997 seeing that anxiety is still the kind of 00:07:57.080 --> 00:07:59.597 dominant feeling towards AI. Now, I 00:07:59.680 --> 00:08:03.237 think closing that confidence gap is as 00:08:03.320 --> 00:08:07.477 much a policy task as it is a a technical one, but I I'd love to just 00:08:07.560 --> 00:08:11.077 get your thoughts. I mean, obviously, you are at a you know, a B2C company 00:08:11.160 --> 00:08:13.917 dealing with, you know, billions of consumers every 00:08:14.000 --> 00:08:17.717 day. You know, how do you think about this trust question? You know, how do we 00:08:17.800 --> 00:08:23.037 get trust up in places where it is not for something that, you know, you believe, I 00:08:23.120 --> 00:08:27.197 believe, and everyone in this room probably believes can do untold good. 00:08:27.280 --> 00:08:30.957 Yeah, this is this is one of the most important questions. And actually, 00:08:31.040 --> 00:08:34.437 perhaps even more so than the technology and the development of the technology, 00:08:34.520 --> 00:08:38.157 um, maybe will govern how how 00:08:38.240 --> 00:08:40.637 uh successful the technology diffuses, 00:08:40.720 --> 00:08:44.157 um, more so than anything. And, um, 00:08:44.240 --> 00:08:46.877 you know, we we think deeply about this 00:08:46.960 --> 00:08:49.637 cuz our vision for the technology is 00:08:49.720 --> 00:08:51.957 that it is um we can deploy personal 00:08:52.040 --> 00:08:56.277 agents that uh know you better than most 00:08:56.360 --> 00:09:00.037 people know you, know about your health, know about your goals, know about your 00:09:00.120 --> 00:09:04.077 relationships, know about your family and your friends, and um can really help 00:09:04.160 --> 00:09:08.517 you live your life more effectively, you know, can help you be healthier, can 00:09:08.600 --> 00:09:12.437 help you accomplish projects that you would never would have had time for, can 00:09:12.520 --> 00:09:13.597 help you build the business you always 00:09:13.680 --> 00:09:16.157 wanted, um, can help you explore your 00:09:16.240 --> 00:09:19.757 interests. And to properly build that 00:09:19.840 --> 00:09:22.757 requires an immense amount of trust from 00:09:22.840 --> 00:09:26.357 consumers, from governments, from um, 00:09:26.440 --> 00:09:30.677 from effectively everyone, from all organizations. And so, you know, this is 00:09:30.760 --> 00:09:34.837 one of these interesting moments where, um, 00:09:34.920 --> 00:09:37.637 I I because the the sort of 00:09:37.720 --> 00:09:39.917 the form that the of the technology that 00:09:40.000 --> 00:09:42.877 we wanted to take is so intimate and is 00:09:42.960 --> 00:09:46.357 so deeply tied with what we think is 00:09:46.440 --> 00:09:48.757 really the the future of, you know, the 00:09:48.840 --> 00:09:51.877 consumer life and the consumer, um, you 00:09:51.960 --> 00:09:53.557 know, is at the core of the relationship 00:09:53.640 --> 00:09:56.197 with technology. Um, 00:09:56.280 --> 00:09:59.237 the figuring out ways to develop the the 00:09:59.320 --> 00:10:03.317 technology in partnership with governing bodies and in partnership with, um, 00:10:03.400 --> 00:10:07.717 nonprofits and other organizations to garner that trust is very important. 00:10:07.800 --> 00:10:11.877 This is something we've seen at Meta, actually, in, um, with WhatsApp. You 00:10:11.960 --> 00:10:15.957 know, one of the we were talking about this actually backstage, but, um, 00:10:16.040 --> 00:10:18.197 WhatsApp has become, uh, incredibly 00:10:18.280 --> 00:10:20.517 ubiquitous, um, and a real, I think, 00:10:20.600 --> 00:10:23.077 utility for much of the world, um, in 00:10:23.160 --> 00:10:28.517 large part due to, uh, its commitment to privacy and the the sort of trust that 00:10:28.600 --> 00:10:32.437 we've built around WhatsApp as an organization. Um, oftentimes at the, uh, 00:10:32.520 --> 00:10:35.837 sort of, uh, you know, not necessarily for the 00:10:35.920 --> 00:10:39.957 optimization of our business objectives, but for the optimization of trust in 00:10:40.040 --> 00:10:42.157 WhatsApp as a platform. I think a 00:10:42.240 --> 00:10:45.317 similar journey will be necessary for 00:10:45.400 --> 00:10:49.197 the AI products of the modern age, and we've seen, um, 00:10:49.280 --> 00:10:53.197 I mean, we've already seen various uh, kerfuffles or sort of jabs back and 00:10:53.280 --> 00:10:57.357 forth between the existing AI, you know, products and labs around, you know, 00:10:57.440 --> 00:11:00.677 whether their actions are truly garnering, you know, truly building 00:11:00.760 --> 00:11:04.877 this, uh, these levels of trust. But, um, and then I think the, you know, the 00:11:04.960 --> 00:11:08.557 topic that you really kicked off at Bletchley 00:11:08.640 --> 00:11:13.557 around, um, in a sort of like global way around safety, I think, is more 00:11:13.640 --> 00:11:14.717 important than ever. I think we're 00:11:14.800 --> 00:11:17.237 already seeing that, um, 00:11:17.320 --> 00:11:20.877 the models are, uh, 00:11:20.960 --> 00:11:23.637 are are achieving levels of capability 00:11:23.720 --> 00:11:26.237 which cause some pause around some of 00:11:26.320 --> 00:11:29.797 the, sort of like, risks that, uh, we 00:11:29.880 --> 00:11:32.197 um, monitor for these models as it comes 00:11:32.280 --> 00:11:34.997 to various national security risks or 00:11:35.080 --> 00:11:38.357 various cybersecurity risks. And so 00:11:38.440 --> 00:11:40.957 the the importance of properly testing 00:11:41.040 --> 00:11:45.877 and then building the right systems to monitor and deploy these models safely 00:11:45.960 --> 00:11:47.197 is is of paramount importance. 00:11:47.280 --> 00:11:49.957 >> And that you and I have been talking about the same topic for some years. And 00:11:50.040 --> 00:11:52.277 that was key to us at 00:11:52.360 --> 00:11:57.077 Bletchley was establishing the AI security institute in the UK to do 00:11:57.160 --> 00:12:00.357 exactly that work. And the And the thought was there was a gap where the 00:12:00.440 --> 00:12:04.437 technical capability didn't exist amongst governments to evaluate some of 00:12:04.520 --> 00:12:08.717 these risks. So how could we in the UK hopefully take a lead, but in sense 00:12:08.800 --> 00:12:13.397 provide a global public good where we work in partnership with 00:12:13.480 --> 00:12:16.997 companies like yours and the other frontier labs to do pre-deployment 00:12:17.080 --> 00:12:21.397 testing on these models. And my thought was always that by doing that, and 00:12:21.480 --> 00:12:24.677 that's not regulation, but it gives citizens confidence that the technology 00:12:24.760 --> 00:12:28.357 is safe cuz ultimately people don't adopt a technology that they don't 00:12:28.440 --> 00:12:32.317 trust. And I So what I my sense hearing you is that there's a role for the 00:12:32.400 --> 00:12:34.557 companies. I thought the the privacy 00:12:34.640 --> 00:12:36.237 example with WhatsApp is a good one is 00:12:36.320 --> 00:12:39.997 how do they design AI products that will 00:12:40.080 --> 00:12:43.797 will engender trust. And then there's a role, it sounds like you agree, for 00:12:43.880 --> 00:12:48.277 governments to do some of that technical evaluation. And is your sense that we 00:12:48.360 --> 00:12:51.637 need, you know, more cooperation between governments on that, more interaction 00:12:51.720 --> 00:12:52.997 with the labs? Is that working as well 00:12:53.080 --> 00:12:56.397 as it should at the moment? Um 00:12:56.480 --> 00:13:01.237 I think the interaction is working well, but I think this is one area where 00:13:01.320 --> 00:13:04.797 you know, uh we always can be more thoughtful. And 00:13:04.880 --> 00:13:09.917 especially as we enter this phase of dramatically increased AI progress, 00:13:10.000 --> 00:13:13.157 I think, you know, we constantly need to be thinking about how we can do better. 00:13:13.240 --> 00:13:15.877 I I think one of the other areas where 00:13:15.960 --> 00:13:20.397 this debate on trust will be won or lost is in the public sector. I think it's 00:13:20.480 --> 00:13:24.397 citizens that you experience that health care, more 00:13:24.480 --> 00:13:27.597 efficient government services, a state that is 00:13:27.680 --> 00:13:32.277 quicker at dealing with you. I think this debate goes from abstract to real 00:13:32.360 --> 00:13:36.437 when you have a real way to demonstrate benefit. And do you have a a sense from 00:13:36.520 --> 00:13:38.037 where you sit about where the most 00:13:38.120 --> 00:13:41.157 obvious use cases or deployment cases 00:13:41.240 --> 00:13:44.637 are in the public sector for AI that we 00:13:44.720 --> 00:13:49.077 should be excited about and that governments should be thinking about? 00:13:49.160 --> 00:13:51.357 I do. I think that um 00:13:51.440 --> 00:13:53.717 one of the first 00:13:53.800 --> 00:13:57.077 you know, if you think governments as um 00:13:57.160 --> 00:13:59.157 service providers in a sense to their 00:13:59.240 --> 00:14:01.797 citizens, I think there's there's a huge 00:14:01.880 --> 00:14:06.397 opportunity for all governments to be able to deliver their services more 00:14:06.480 --> 00:14:10.437 effectively. We're already seeing that actually in India 00:14:10.520 --> 00:14:13.157 on top of WhatsApp funnily enough where 00:14:13.240 --> 00:14:18.037 there are in certain states they actually deliver fast majority of the 00:14:18.120 --> 00:14:19.717 government services directly through 00:14:19.800 --> 00:14:22.877 WhatsApp. And then the for transit a 00:14:22.960 --> 00:14:28.437 nation actually able to a true what's up in India. 00:14:28.520 --> 00:14:32.237 last year in some 00:14:32.320 --> 00:14:36.077 what's up That's not AI 00:14:36.160 --> 00:14:41.277 the future opportunity there that is quite exciting is actually deployment of 00:14:41.360 --> 00:14:44.237 uh uh of AI to 00:14:44.320 --> 00:14:47.037 greater increase the ease and speed at 00:14:47.120 --> 00:14:48.357 which these government services are 00:14:48.440 --> 00:14:51.597 deployed and provided to citizens. 00:14:51.680 --> 00:14:54.477 I think that you know there's one 00:14:54.560 --> 00:14:56.877 term that I've always I've used and 00:14:56.960 --> 00:14:59.437 and I like this concept of the magic 00:14:59.520 --> 00:15:02.077 quadrant, right? And it really is you 00:15:02.160 --> 00:15:06.557 know if you think about how should governments be 00:15:06.640 --> 00:15:08.997 um how should the state be reimagined? 00:15:09.080 --> 00:15:12.437 Um you know I think we need to move away from 00:15:12.520 --> 00:15:16.517 large-scale bureaucratic services and um 00:15:16.600 --> 00:15:21.077 and these organizations and think about how can the government move as quickly 00:15:21.160 --> 00:15:25.837 as possible to deliver using a combination of technology and humans. verify 00:15:25.920 --> 00:15:29.077 one of those agents the fastest possible public public 00:15:29.160 --> 00:15:31.877 services for citizens in a 00:15:31.960 --> 00:15:36.877 country or a state to develop um and I think this cuts through civil 00:15:36.960 --> 00:15:41.197 services, I think it cuts through health care you mentioned and certainly has a 00:15:41.280 --> 00:15:45.917 has a large role to play in security and in 00:15:46.000 --> 00:15:51.077 in government this government So 00:15:51.160 --> 00:15:54.557 I think it's quite a problem here. Yeah, I think you were 00:15:54.640 --> 00:15:59.560 you were contacted by the Chinese government 00:16:01.120 --> 00:16:05.557 I need one more one more mic. I wanted to suggest I pick up on 00:16:05.640 --> 00:16:10.317 something you talked about one of the benefits of of leaving office 00:16:10.400 --> 00:16:13.237 is you have Thank you. A lot of benefits of leaving 00:16:13.320 --> 00:16:17.637 office is you have more more time at your desk and when I left someone gave 00:16:17.720 --> 00:16:20.957 me a great book by a guy called Jeffrey Professor Ding and it's called 00:16:21.040 --> 00:16:23.117 Technology and and Great Powers I'm sure 00:16:23.200 --> 00:16:25.397 you you know the book and and in it he 00:16:25.480 --> 00:16:30.317 makes a really compelling argument about technology diffusion essentially looking 00:16:30.400 --> 00:16:33.677 at the history of general purpose technologies going all the way back to 00:16:33.760 --> 00:16:37.477 the printing press and making the argument that you don't need to be the 00:16:37.560 --> 00:16:41.557 country that invents the technology to be the country that gets the most 00:16:41.640 --> 00:16:44.397 benefit from it and I think you know you 00:16:44.480 --> 00:16:47.717 touched on India India has recognized 00:16:47.800 --> 00:16:50.357 that technology leadership exactly does 00:16:50.440 --> 00:16:52.397 not depend on just inventing the 00:16:52.480 --> 00:16:57.477 technology it's about how effectively you deploy it and they here are focused 00:16:57.560 --> 00:17:00.917 on mass broad adoption and obviously 00:17:01.000 --> 00:17:03.637 supported by a few critical attributes. 00:17:03.720 --> 00:17:06.077 You have a very AI 00:17:06.160 --> 00:17:11.437 talent deep talent pool here. You talked about the rails of distribution 00:17:11.520 --> 00:17:16.317 obviously the digital public infrastructure here are there the ID UPI 00:17:16.400 --> 00:17:20.757 the payments infrastructure and then most recently Ayushman Bharat health 00:17:20.840 --> 00:17:25.837 accounts provide these infrastructure rails to distribute apps to over a 00:17:25.920 --> 00:17:29.477 billion people. And they also have a population that supportive of that. 00:17:29.560 --> 00:17:34.317 And I think focusing on that has allowed India to position itself as a leader 00:17:34.400 --> 00:17:39.477 by by pursuing the deployment and adoption strategy. But 00:17:39.560 --> 00:17:44.477 but thinking about that, you know, what are those critical attributes that 00:17:44.560 --> 00:17:48.757 government has to get right if we want to turn your aspiration into a reality 00:17:48.840 --> 00:17:52.397 about a gentle government, about delivering for our citizens. You know, I 00:17:52.480 --> 00:17:54.997 think government does need to think about things differently and needs to 00:17:55.080 --> 00:17:58.397 build the the architecture around 00:17:58.480 --> 00:18:02.077 deployment and diffusion, whether it's skills, whether it's compute, whether 00:18:02.160 --> 00:18:06.077 it's regulation. But crucially probably also data, 00:18:06.160 --> 00:18:09.717 which is something that you know better than most. So maybe if we start there 00:18:09.800 --> 00:18:13.797 and then broaden it out, if I'm, you know, sitting as a government leader as 00:18:13.880 --> 00:18:16.917 many people in the audience are, I'm trying to figure out what do I need 00:18:17.000 --> 00:18:21.597 to get right? What are the building blocks to deliver this thing properly? 00:18:21.680 --> 00:18:24.597 Let's start with data and what would your advice be there and then we'll 00:18:24.680 --> 00:18:26.437 we'll broaden it out a little bit. 00:18:26.520 --> 00:18:29.877 Yeah, there there is uh 00:18:29.960 --> 00:18:31.077 data 00:18:31.160 --> 00:18:33.677 is is so important and I think over the 00:18:33.760 --> 00:18:37.277 past few years many people have I think 00:18:37.360 --> 00:18:41.077 you know realized this as we sort of look at the power of these models and we 00:18:41.160 --> 00:18:44.837 look behind the curtains and you realize that all the powers and beauty of these 00:18:44.920 --> 00:18:49.717 models are actually imbued by data. And if you think about 00:18:49.800 --> 00:18:53.117 So if you think about data is like in some ways this sort of new oil, this 00:18:53.200 --> 00:18:57.437 sort of like base material that is incredibly important for the future 00:18:57.520 --> 00:18:59.237 of of AI and the development of these 00:18:59.320 --> 00:19:04.717 models. Um then I think as like governments should think about what are 00:19:04.800 --> 00:19:07.957 their, you know, what are their reserves? What are their data reserves? 00:19:08.040 --> 00:19:12.277 What are their data assets? And Um, and to 00:19:12.360 --> 00:19:15.277 in many ways build strategies where they have 00:19:15.360 --> 00:19:18.957 sovereignty over some set of data or or 00:19:19.040 --> 00:19:23.397 they're able to build certain certain pockets of data sets that enable the the 00:19:23.480 --> 00:19:27.237 country to perform the best. I think one of the areas that is the two areas that 00:19:27.320 --> 00:19:30.957 is clear that are clearest for this are one is health care. 00:19:31.040 --> 00:19:33.877 I think many countries um 00:19:33.960 --> 00:19:36.277 uh already are and other countries need 00:19:36.360 --> 00:19:38.397 to be thinking I think about um 00:19:38.480 --> 00:19:41.237 very effective mechanisms for 00:19:41.320 --> 00:19:44.157 uh uh for health care data within the 00:19:44.240 --> 00:19:46.957 country to enable ultimately over time 00:19:47.040 --> 00:19:51.357 there to be AI models and AI agents that are built on top of that to deliver 00:19:51.440 --> 00:19:54.597 better uh health services to their citizens. And then the second is 00:19:54.680 --> 00:19:57.717 national security. And I think that is um you know, in many ways pretty 00:19:57.800 --> 00:20:02.037 self-explanatory, but I think you know, countries need to be thinking about that 00:20:02.120 --> 00:20:06.917 um uh quite deeply as to what are the sort of like core data assets. Then I think 00:20:07.000 --> 00:20:10.877 if you if you sort of build around data and you look at the 00:20:10.960 --> 00:20:13.277 next few layers, um the next one is is 00:20:13.360 --> 00:20:15.517 infrastructure. Um, so you know, we 00:20:15.600 --> 00:20:18.677 think about the power, the um 00:20:18.760 --> 00:20:23.437 the computational infrastructure, um and and what does that story look like? I 00:20:23.520 --> 00:20:27.397 don't think every country needs to build large-scale AI data centers in their 00:20:27.480 --> 00:20:31.637 country, but they do need to have a strategy of over time how are they going 00:20:31.720 --> 00:20:33.557 to have access to that infrastructure. 00:20:33.640 --> 00:20:35.717 And so whether this is in the form of 00:20:35.800 --> 00:20:39.557 alliances of various formats or um or uh 00:20:39.640 --> 00:20:43.957 partnerships even with with uh corporate entities, the large cloud providers, I 00:20:44.040 --> 00:20:45.557 think that's an important element to 00:20:45.640 --> 00:20:47.237 think through. Um, 00:20:47.320 --> 00:20:51.637 then comes I think the sort of uh the um 00:20:51.720 --> 00:20:53.597 the innovation ecosystem. I do think 00:20:53.680 --> 00:20:55.357 that, you know, we're in a moment of 00:20:55.440 --> 00:20:59.837 momentous technological change and um it 00:20:59.920 --> 00:21:03.917 is I think incredibly important for I think there's 00:21:04.000 --> 00:21:05.917 near boundless uh 00:21:06.000 --> 00:21:11.317 entrepreneurial and economic opportunity Um, and I think it is the responsibility 00:21:11.400 --> 00:21:15.997 of of governments and countries to think through how they can build ecosystems 00:21:16.080 --> 00:21:19.717 that support that. I think India is actually in many ways a great, 00:21:19.800 --> 00:21:23.557 very positive case study in large part due to the due to the talent pool, but 00:21:23.640 --> 00:21:28.357 you know, I was at a dinner with a number of Indian founders and venture 00:21:28.440 --> 00:21:30.277 capitalist last night and and the stat 00:21:30.360 --> 00:21:32.717 was I think there are more 00:21:32.800 --> 00:21:35.837 consumer AI companies in India consumer 00:21:35.920 --> 00:21:39.957 AI startups in India than in the United States. And so, there really are some of 00:21:40.040 --> 00:21:41.877 these like, you know, shining examples 00:21:41.960 --> 00:21:47.237 of of incredible development of these ecosystems. And then, 00:21:47.320 --> 00:21:49.557 and then lastly, I think are the government services which we talked 00:21:49.640 --> 00:21:53.037 about previously. So, we've got I think the elements of almost a a playbook 00:21:53.120 --> 00:21:57.197 emerging, right? You need to have the right data and make sure it's it's used 00:21:57.280 --> 00:22:00.117 effectively. You need some degree of compute. 00:22:00.200 --> 00:22:03.997 You've got appropriate regulation as we've talked about to build to build 00:22:04.080 --> 00:22:09.077 trust. And then we've got some public sector use cases as well and an innovation 00:22:09.160 --> 00:22:12.677 ecosystem. So, these are the kind of ingredients where you can start to turn 00:22:12.760 --> 00:22:15.877 these ambitions into reality and and 00:22:15.960 --> 00:22:19.317 actual impact on people. I I just 00:22:19.400 --> 00:22:24.157 staying on that theme, you know, you you work at an incredible company. You work 00:22:24.240 --> 00:22:28.517 with other enterprises. You know, I talked to a lot of CEOs and 00:22:28.600 --> 00:22:33.597 and I'm struck by how they're thinking about deploying AI in their businesses 00:22:33.680 --> 00:22:37.597 and thinking about the parallel between that and and what governments should be 00:22:37.680 --> 00:22:41.157 doing. And you know, I think there is a kind of clear parallel. I'm I'm 00:22:41.240 --> 00:22:43.717 intrigued. I think CEOs are you know, 00:22:43.800 --> 00:22:48.317 kind of constantly thinking about AI. I think almost any large company CEO is 00:22:48.400 --> 00:22:51.877 spending their time figuring out, am I going to get disrupted, disintermediated 00:22:51.960 --> 00:22:55.997 by AI, how do I use it to stay ahead, where do I use it to to get more 00:22:56.080 --> 00:23:01.917 efficiency and more growth. And they are driving projects from the top down. And 00:23:02.000 --> 00:23:06.437 they are saying, "Look, this is something that is, you know, not 00:23:06.520 --> 00:23:10.277 a topic that could just stay with the IT department, right? This is something 00:23:10.360 --> 00:23:12.757 that has to be driven by the CEO, where 00:23:12.840 --> 00:23:17.397 we pick the use cases, we have accountability, and we start piloting 00:23:17.480 --> 00:23:19.957 these and scaling them up quite quickly." 00:23:20.040 --> 00:23:22.597 I you know, I my sense is we haven't 00:23:22.680 --> 00:23:27.397 quite seen that that same transition in governments yet, but I thought it's a 00:23:27.480 --> 00:23:30.397 good analogy in that look, this has to be driven by the person at the top. At 00:23:30.480 --> 00:23:33.637 the end of the day, if you're a prime minister, and I'm sure Tony would say 00:23:33.720 --> 00:23:37.717 the same thing, and indeed has, you can only do a few things that you drive 00:23:37.800 --> 00:23:41.277 personally from the top, and this really has to be one of them. You know, what 00:23:41.360 --> 00:23:43.757 are your kind of thoughts or advice for 00:23:43.840 --> 00:23:46.117 I get both CEOs and and governments 00:23:46.200 --> 00:23:48.957 about how to kind of drive the machine. 00:23:49.040 --> 00:23:53.637 What does it take to get execution on this? I mean, you've had to do that at 00:23:53.720 --> 00:23:56.557 Meta yourself, right? So, you'll have a sense from that. 00:23:56.640 --> 00:24:01.957 Yeah, I think it's a mix of of two things. I think first, um at 00:24:02.040 --> 00:24:05.597 uh at the highest level, you need to make 00:24:05.680 --> 00:24:09.237 it very clear that the organization will be an AI-first organization. And what 00:24:09.320 --> 00:24:10.757 that means for every organization is 00:24:10.840 --> 00:24:14.317 slightly different, but it it it um like 00:24:14.400 --> 00:24:18.517 the way that truly I think about it is that uh the success or failure of the 00:24:18.600 --> 00:24:21.077 organization actually entirely rests on 00:24:21.160 --> 00:24:26.157 its ability to properly embrace AI. And so, that's sort of the tops-down um 00:24:26.240 --> 00:24:29.637 component of it. And you know, there's sort of a strategy that must be designed 00:24:29.720 --> 00:24:32.917 more thoughtfully. And actually, honestly, you can probably ask AI uh to 00:24:33.000 --> 00:24:36.397 help you design that strategy now, and it'll probably do a pretty good job. And 00:24:36.480 --> 00:24:39.357 then you need um a a 00:24:39.440 --> 00:24:42.957 bed of bottoms-up adoption and 00:24:43.040 --> 00:24:46.837 bottoms-up um success stories. And for 00:24:46.920 --> 00:24:50.997 uh for us at Meta, like the example of this are, you know, we have these 00:24:51.080 --> 00:24:52.717 engineers throughout the company who are 00:24:52.800 --> 00:24:56.117 able to use AI to uh be you know, 10 to 00:24:56.200 --> 00:24:59.597 100 times more productive. And um those 00:24:59.680 --> 00:25:05.357 are just such incredible examples that where all of a sudden everyone else in 00:25:05.440 --> 00:25:08.597 the organization looks around and realizes, wow, this is really something 00:25:08.680 --> 00:25:12.717 that must be taken very seriously. It's not just something that, you know, um 00:25:12.800 --> 00:25:16.797 the CEO or the C-level individuals are talking about and, you know, that 00:25:16.880 --> 00:25:20.797 they're so detached from reality. It's like, no, it's very real on the ground 00:25:20.880 --> 00:25:24.717 um in a very big way. And so, I think you need this sort of mix of what are 00:25:24.800 --> 00:25:28.757 the success stories that cause everyone to realize how big of a shift it 00:25:28.840 --> 00:25:33.277 actually is, and then also ensure that there's a continual um there's a very 00:25:33.360 --> 00:25:36.957 clear top-down direction that this transition is is mandatory. 00:25:37.040 --> 00:25:38.717 >> It's mandatory. Very helpful advice. 00:25:38.800 --> 00:25:41.237 We're almost out of time. So, maybe the 00:25:41.320 --> 00:25:44.477 last final point uh to close on is, you 00:25:44.560 --> 00:25:49.397 know, you're looking ahead, you know, what are you most excited about uh in 00:25:49.480 --> 00:25:52.317 the development of this of this technology? You've touched on personal 00:25:52.400 --> 00:25:55.157 agents, other scientific breakthroughs. 00:25:55.240 --> 00:25:59.237 You know, what where is this going, and what should governments be doing now, 00:25:59.320 --> 00:26:00.877 you know, so that they are prepared and 00:26:00.960 --> 00:26:03.797 and not playing catch-up? 00:26:03.880 --> 00:26:08.597 I think the things that um you can depend on, I think the the piece 00:26:08.680 --> 00:26:10.197 of technology development will continue 00:26:10.280 --> 00:26:16.117 accelerating. Um I think this this era of recursive self-improvement will 00:26:16.200 --> 00:26:21.157 uh feel like everything is speeding up, and I think that um the technology will 00:26:21.240 --> 00:26:22.997 continue developing very quickly. 00:26:23.080 --> 00:26:25.477 Um I also think that the the diffusion 00:26:25.560 --> 00:26:30.117 of the technology is is going to accelerate. I think there's been this 00:26:30.200 --> 00:26:33.717 question for many for the for the past many years, which is everyone's talking 00:26:33.800 --> 00:26:35.757 about AI, but where is AI in my daily 00:26:35.840 --> 00:26:38.517 life? I think that um that really is 00:26:38.600 --> 00:26:42.837 going to accelerate. Maybe not um you know, it's not going to go be 100 00:26:42.920 --> 00:26:46.837 times faster, but maybe it'll be, you know, two or three times faster every 00:26:46.920 --> 00:26:48.637 year for the you know, for the next many 00:26:48.720 --> 00:26:50.357 years. Um 00:26:50.440 --> 00:26:54.757 and and I think the thing that uh 00:26:54.840 --> 00:26:57.277 where where I really am excited about 00:26:57.360 --> 00:26:59.757 this going is you know, we talk about 00:26:59.840 --> 00:27:04.597 this idea of personal super intelligence within meta which is, you know, how does 00:27:04.680 --> 00:27:08.437 you know, how does each person have an agent that helps them accomplish the 00:27:08.520 --> 00:27:12.197 things that they're you know, ex- extremely excited about. 00:27:12.280 --> 00:27:15.477 How can that help them, you know, explore their passions, build a 00:27:15.560 --> 00:27:19.917 business, make a discovery, invent a thing, you know, 00:27:20.000 --> 00:27:21.757 you know, achieve the most maximal 00:27:21.840 --> 00:27:25.157 version of their own lives and I think 00:27:25.240 --> 00:27:28.357 where this is going even beyond then is 00:27:28.440 --> 00:27:32.397 sort of this concept of interpersonal super intelligence which is 00:27:32.480 --> 00:27:36.237 I think we already see in human society sometimes the thing that is necessary 00:27:36.320 --> 00:27:40.677 for big things to happen are just for the right set of people to be organized 00:27:40.760 --> 00:27:43.917 in the right way. You see this with companies, you see this with nonprofits, 00:27:44.000 --> 00:27:46.317 you see this with large movements, you 00:27:46.400 --> 00:27:48.997 see this even with with governments when 00:27:49.080 --> 00:27:52.117 they when they reorganize and I think 00:27:52.200 --> 00:27:53.597 there's an incredible opportunity for 00:27:53.680 --> 00:27:57.597 the technology of AI to enable humans to 00:27:57.680 --> 00:28:00.797 organize in ways that are even better 00:28:00.880 --> 00:28:04.917 than we do today and all of a sudden maybe we'll be able to organize as 00:28:05.000 --> 00:28:08.797 communities or as groups or as new companies or who knows what it'll look 00:28:08.880 --> 00:28:12.637 like to to solve many of our greatest challenges and so 00:28:12.720 --> 00:28:16.357 you know, it sounds a bit high falutin but but it's something I'm genuinely quite 00:28:16.440 --> 00:28:18.557 excited about. >> Yeah, I can see that and it's a very a 00:28:18.640 --> 00:28:21.877 very compelling vision. Look, this has been fascinating conversation. It's 00:28:21.960 --> 00:28:25.957 clear that the technology is is not just keeping pace that it's accelerating. 00:28:26.040 --> 00:28:29.317 It's clear that the debate is moving from, you know, just talking about the 00:28:29.400 --> 00:28:33.237 technology to as you said, how is it making a a difference in people's lives 00:28:33.320 --> 00:28:37.157 and being deployed. And that one of one of my concerns and maybe we'll finish on 00:28:37.240 --> 00:28:39.797 this is that I think some not all but 00:28:39.880 --> 00:28:44.517 some political leaders think that AI is is going to be tomorrow's issue when I 00:28:44.600 --> 00:28:49.477 think they need to recognize that it's it's an action this day issue. And this 00:28:49.560 --> 00:28:51.237 needs to go from, you know, AI needs to 00:28:51.320 --> 00:28:53.437 go from being a kind of specialized 00:28:53.520 --> 00:28:57.517 topic to a centralized responsibility of 00:28:57.600 --> 00:29:01.877 government. And it needs to have that same urgency and and action attached to 00:29:01.960 --> 00:29:05.197 it so we can realize all the benefits that Alex is talking about. But this has 00:29:05.280 --> 00:29:08.317 been a great discussion. Thanks for all your advice and tips everyone in the 00:29:08.400 --> 00:29:10.917 audience and giving us an insight into where things are going. Thanks very 00:29:11.000 --> 00:29:15.470 much. Thank you so much. >> [applause] 00:29:19.240 --> 00:29:21.317 >> Thank you so much to our speakers. 00:29:21.400 --> 00:29:24.677 Earlier, can we have another round of applause, please, for Rishi Sunak and 00:29:24.760 --> 00:29:28.837 Alexander Wang? >> [applause] >> Thank you. 00:29:28.920 --> 00:29:33.357 >> When I became Prime Minister, it's been an exact AI obsessed because of my time 00:29:33.440 --> 00:29:37.117 in in California. You you get the job in Downing Street and one of the earliest 00:29:37.200 --> 00:29:41.677 meetings you have is with the the Foreign Affairs team. And they sit down 00:29:41.760 --> 00:29:45.597 with you. I can I can see someone who will remember this in the audience. And 00:29:45.680 --> 00:29:48.117 they go over your international calendar 00:29:48.200 --> 00:29:53.397 for the next 12 months. And they they want to kind of figure out, you know, 00:29:53.480 --> 00:29:56.557 what you should be doing. And it turns out there's an international summit for 00:29:56.640 --> 00:30:00.637 basically every single thing that you can imagine. Turns out there were about 00:30:00.720 --> 00:30:05.157 150. And the British Foreign Office being the British Foreign Office, they wanted me 00:30:05.240 --> 00:30:08.637 to go to all of them. And although I did I did say I'm going 00:30:08.720 --> 00:30:12.237 to as a British Prime Minister, I think I can draw the line at the United 00:30:12.320 --> 00:30:17.197 Nations high-level meeting on sandstorms. That's where we agreed to disagree. 00:30:17.280 --> 00:30:20.597 But I noticed that there was no gathering to bring together people to 00:30:20.680 --> 00:30:24.357 talk about what I thought was the most transformative technology of our time. 00:30:24.440 --> 00:30:28.437 And I think we're all fortunate to be living at a time where AI is a 00:30:28.520 --> 00:30:32.037 general-purpose technology. These don't come along all that often 00:30:32.120 --> 00:30:35.837 and when they do, they had the opportunity in the past and will in the 00:30:35.920 --> 00:30:38.277 future to change every aspect of our 00:30:38.360 --> 00:30:42.957 economies, of our societies, of our lives. And that's why I thought it was 00:30:43.040 --> 00:30:48.237 necessary that leaders came together to discuss this at a regular basis. But, I 00:30:48.320 --> 00:30:50.957 also thought it needed to be done slightly differently. 00:30:51.040 --> 00:30:52.557 In contrast to previous technology 00:30:52.640 --> 00:30:56.077 cycles, it's companies like Alex's and 00:30:56.160 --> 00:31:00.317 innovators like Alex himself who were at the frontier of developing this 00:31:00.400 --> 00:31:03.637 technology. And I thought they needed to be in the 00:31:03.720 --> 00:31:07.757 room. Not least because it's businesses like yours that are have the resources 00:31:07.840 --> 00:31:12.037 to invest in this. Just to give you a bit of context. This year, you know, the 00:31:12.120 --> 00:31:14.157 the large AI companies are going to 00:31:14.240 --> 00:31:17.597 spend 20 times more 00:31:17.680 --> 00:31:22.397 on developing this technology than the US did for the Manhattan Project. If you 00:31:22.480 --> 00:31:26.477 think about seminal moments where the government mobilized to do something, 20 00:31:26.560 --> 00:31:30.677 times more just this year alone. So, I thought, "Look, it's vital that we do 00:31:30.760 --> 00:31:32.477 this in a different way, not like the G7 00:31:32.560 --> 00:31:35.517 and G20. We bring industry, innovators, 00:31:35.600 --> 00:31:39.677 and and government together." And that's that's what we did at Bletchley Park a 00:31:39.760 --> 00:31:41.837 few years ago. The summit series went to 00:31:41.920 --> 00:31:44.317 Paris, South Korea, and now brilliantly 00:31:44.400 --> 00:31:47.477 in India, a country with huge digital 00:31:47.560 --> 00:31:51.357 ambitions and indeed capacity. 00:31:51.440 --> 00:31:53.717 But, I think over those years, and Alex 00:31:53.800 --> 00:31:55.557 will start with this, I think actually 00:31:55.640 --> 00:31:59.917 the debate on AI has shifted from 00:32:00.000 --> 00:32:02.757 kind of technology to strategy. 00:32:02.840 --> 00:32:06.957 You know, from what these tools can do to what governments are choosing to do 00:32:07.040 --> 00:32:10.517 with them. And I wanted to maybe just start there. 00:32:10.600 --> 00:32:13.397 Is if you think about the last few years, you are someone who is at the 00:32:13.480 --> 00:32:15.837 frontier of developing this technology. 00:32:15.920 --> 00:32:19.117 You know, the the journey that we've been on in the last few years, you know, 00:32:19.200 --> 00:32:22.677 help us understand that. Where have we come? 00:32:22.760 --> 00:32:24.997 What is it that is exciting at this this 00:32:25.080 --> 00:32:30.077 moment? And just ground us in in where we are so that we can get get on to some 00:32:30.160 --> 00:32:32.997 of these questions about what we do with it. 00:32:33.080 --> 00:32:34.797 Yeah, I think um 00:32:34.880 --> 00:32:37.637 I I we are sitting here right now at 00:32:37.720 --> 00:32:40.637 such an incredibly exciting moment for 00:32:40.720 --> 00:32:45.437 the technology. I think we are um at the beginning of 00:32:45.520 --> 00:32:50.277 uh you know, of a true acceleration in the technology. And, you know, in many ways 00:32:50.360 --> 00:32:51.837 it's felt like this very clear 00:32:51.920 --> 00:32:54.597 acceleration even until now, but I think 00:32:54.680 --> 00:32:57.637 we're going to see sort of 2026 is uh an 00:32:57.720 --> 00:33:02.117 inflection point in many ways. I think to contextualize this over the past few 00:33:02.200 --> 00:33:04.877 years, I think from um 00:33:04.960 --> 00:33:06.477 2018 00:33:06.560 --> 00:33:10.237 through 2024 00:33:10.320 --> 00:33:12.877 really, which is a good 6-year period, 00:33:12.960 --> 00:33:15.397 we were writing we were in the era of 00:33:15.480 --> 00:33:20.357 pre-training. So, we were writing um one very large exponential curve towards 00:33:20.440 --> 00:33:22.117 improving the performance of the models. 00:33:22.200 --> 00:33:25.717 And um that curve was very predictable and it 00:33:25.800 --> 00:33:29.557 sort of had clear characteristics where the more resources you put in, the more 00:33:29.640 --> 00:33:33.077 results you would get out. And there was a lot of questions at the time of, you 00:33:33.160 --> 00:33:36.597 know, whether or not this curve would continue. And even to the degree it did 00:33:36.680 --> 00:33:40.157 continue, were there diminishing marginal returns? And I think this was sort of um 00:33:40.240 --> 00:33:41.277 one of the 00:33:41.360 --> 00:33:43.317 uh one of the more recent sort of 00:33:43.400 --> 00:33:47.597 eras of AI skepticism. And then in 2024, 00:33:47.680 --> 00:33:50.437 um like towards the end of 2024, uh 00:33:50.520 --> 00:33:52.837 began the era of reinforcement learning. 00:33:52.920 --> 00:33:54.957 So, this was sort of the the next major 00:33:55.040 --> 00:33:57.397 wave that stacked on top of the previous 00:33:57.480 --> 00:34:00.037 wave. And this is when we saw the models 00:34:00.120 --> 00:34:04.877 begin to learn to reason. And um you know, there's always this meme of 00:34:04.960 --> 00:34:06.917 what did Ilya see? Uh well, Ilya saw 00:34:07.000 --> 00:34:09.237 reasoning uh from the models. And it was 00:34:09.320 --> 00:34:12.637 very clearly this this next mega wave of 00:34:12.720 --> 00:34:17.357 of the technology. And even reinforcement learning had, you know, 00:34:17.440 --> 00:34:20.757 clearly some limitations. And there were all these questions around when that was 00:34:20.840 --> 00:34:24.677 going to, you know, how long that era was going to last and maybe there were 00:34:24.760 --> 00:34:27.517 limitations in that era. And then at the 00:34:27.600 --> 00:34:30.157 end of 2025, so just a few months ago, I 00:34:30.240 --> 00:34:34.597 think we really begun um an entirely new 00:34:34.680 --> 00:34:38.677 paradigm or era of the technology, which is 00:34:38.760 --> 00:34:41.237 really beginning the era of recursive 00:34:41.320 --> 00:34:46.317 self-improvement. And I think we're really starting to see this in the 00:34:46.400 --> 00:34:49.477 development of our models. I think you know, most of the major labs are seeing 00:34:49.560 --> 00:34:53.277 this in the development of their models, which is that the models themselves have 00:34:53.360 --> 00:34:56.437 now become instrumental in accelerating 00:34:56.520 --> 00:34:58.717 the process of producing the next AIs. 00:34:58.800 --> 00:35:02.877 And so, you know, the the sort of from the outside what you see is maybe an 00:35:02.960 --> 00:35:06.917 acceleration of overall development velocity, an acceleration in which the 00:35:07.000 --> 00:35:10.557 speed at which the the new models are being released. Internally, what we see 00:35:10.640 --> 00:35:14.317 is just dramatic speedups. I think the sort of 00:35:14.400 --> 00:35:16.437 the productivity of an individual 00:35:16.520 --> 00:35:19.317 researcher has grown dramatically. 00:35:19.400 --> 00:35:22.277 And we don't expect that to stop. In 00:35:22.360 --> 00:35:27.637 fact, we expect it to kind of smoothly continue accelerating where the output 00:35:27.720 --> 00:35:31.197 of an individual researcher, an individual engineer, well, I think that 00:35:31.280 --> 00:35:36.157 itself will continue growing exponentially as the technology improves. And 00:35:36.240 --> 00:35:41.237 the other mega wave that is somewhat coincident with this new paradigm within 00:35:41.320 --> 00:35:44.237 the sort of development of the technology of 00:35:44.320 --> 00:35:47.237 a recursive self-improvement is really 00:35:47.320 --> 00:35:51.877 the arrival of agents in earnest. And 00:35:51.960 --> 00:35:54.197 this is something that has been talked 00:35:54.280 --> 00:35:57.597 about since I honestly want to say 2023 00:35:57.680 --> 00:36:00.117 with a lot of fanfare. And for a lot 00:36:00.200 --> 00:36:02.837 long time, it was sort of 00:36:02.920 --> 00:36:06.637 one of these hype sort of buzzwords 00:36:06.720 --> 00:36:08.557 that, you know, never actually lived up 00:36:08.640 --> 00:36:11.037 to the expectations. And I think also at 00:36:11.120 --> 00:36:14.317 the end of 2025, we and toward you know, 00:36:14.400 --> 00:36:17.117 middle to end of 2025, we started to see 00:36:17.200 --> 00:36:19.557 agents actually work. And this is, you 00:36:19.640 --> 00:36:21.437 know, these are where the models can 00:36:21.520 --> 00:36:24.797 actually start taking actions and start 00:36:24.880 --> 00:36:27.157 automating entire workflows and being 00:36:27.240 --> 00:36:29.717 sort of dramatically more 00:36:29.800 --> 00:36:32.797 productive as actual sort of entities 00:36:32.880 --> 00:36:37.117 onto themselves versus assistants. And 00:36:37.200 --> 00:36:41.557 in this era is you know it started with coding agents. We're starting to see it 00:36:41.640 --> 00:36:44.837 really percolate through personal agents. This is a big part of what we 00:36:44.920 --> 00:36:46.877 believe in at Meta. 00:36:46.960 --> 00:36:49.957 And I think over the course of 2026 we 00:36:50.040 --> 00:36:53.757 will see large-scale agent deployments 00:36:53.840 --> 00:36:56.357 in many areas of the economy, many areas 00:36:56.440 --> 00:36:59.437 of the world. And 00:36:59.520 --> 00:37:05.237 the sort of you know if the the GDP of AI so to speak is going to grow 00:37:05.320 --> 00:37:08.757 exponentially. >> Well that that is very welcome to to 00:37:08.840 --> 00:37:13.557 hear. I I think this you you hit on something that is important because 00:37:13.640 --> 00:37:16.957 I remember when I first started doing this in the job the conversations were 00:37:17.040 --> 00:37:19.797 all around these scaling laws and would they continue. And you mentioned the 00:37:19.880 --> 00:37:21.837 skepticism. I mean not only have they 00:37:21.920 --> 00:37:24.557 continued and continued to deliver this 00:37:24.640 --> 00:37:29.117 exponential growth it feels very much like the last few months have seen an 00:37:29.200 --> 00:37:31.357 acceleration. It's what you touched on. 00:37:31.440 --> 00:37:36.357 It's what Dario and Demis have have also said to me in in publicly driven by this 00:37:36.440 --> 00:37:39.717 recursive self-improvement that the models themselves are now able to help 00:37:39.800 --> 00:37:42.597 us develop the new models quicker. So 00:37:42.680 --> 00:37:46.797 you know we're at this quite exciting but also a moment of some anxiety 00:37:46.880 --> 00:37:50.797 because of that accelerating development time scale. So I wanted to 00:37:50.880 --> 00:37:53.317 just then talk about trust 00:37:53.400 --> 00:37:55.437 because I tell you 00:37:55.520 --> 00:37:58.557 so I might I have two young girls and 00:37:58.640 --> 00:38:02.557 you know as they're using their chatbots and things and they're always like 00:38:02.640 --> 00:38:06.757 please you know so and so and so and so and thank you when they finish writing 00:38:06.840 --> 00:38:10.037 their request or they get the answer back. There's lots of pleases and and 00:38:10.120 --> 00:38:13.837 thank yous everywhere. And I and I I thought that was you know interesting. I 00:38:13.920 --> 00:38:16.717 said to them I said look there's you know you don't need to do that. There's 00:38:16.800 --> 00:38:21.237 not not a person. It doesn't it doesn't matter. And and by the way you know it 00:38:21.320 --> 00:38:25.197 it a fortune when we have to do the compute for that. Um and and they they 00:38:25.280 --> 00:38:29.277 said to me, "You know what, Daddy?" They said, "If if AI takes over the world, we 00:38:29.360 --> 00:38:31.717 want to have been kind to the AI." 00:38:31.800 --> 00:38:35.037 Uh and I thought that was good good insurance policy. But it it kind of you 00:38:35.120 --> 00:38:37.077 know, it touches on this anxiety. 00:38:37.160 --> 00:38:41.037 >> wise. >> Yeah. >> [laughter] >> Um in that maybe. 00:38:41.120 --> 00:38:44.237 You know, across the world, we're seeing these different attitudes 00:38:44.320 --> 00:38:48.757 uh towards AI. In countries like India, where we are, there's enormous optimism 00:38:48.840 --> 00:38:52.277 and and trust. And then actually in Western countries, you know, we're 00:38:52.360 --> 00:38:54.597 seeing that anxiety is still the kind of 00:38:54.680 --> 00:38:57.197 dominant feeling towards AI. Now, I 00:38:57.280 --> 00:39:00.677 think closing that confidence gap is uh 00:39:00.760 --> 00:39:05.117 as much a policy task as it is a a technical one. But I I'd love to just 00:39:05.200 --> 00:39:08.637 get your thoughts. I mean, obviously, you are at a you know, B2C company 00:39:08.720 --> 00:39:11.557 dealing with you know, billions of consumers every 00:39:11.640 --> 00:39:15.357 day. You know, how do you think about this trust question? You know, how do we 00:39:15.440 --> 00:39:20.717 get trust up in places where it is not for something that you know, you believe, I 00:39:20.800 --> 00:39:24.837 believe, and everyone in this room probably believes can do untold good. 00:39:24.920 --> 00:39:28.597 Yeah, this is this is one of the most important questions. And actually, 00:39:28.680 --> 00:39:30.277 perhaps even more so than the technology 00:39:30.360 --> 00:39:32.797 and development of the technology, um 00:39:32.880 --> 00:39:35.797 maybe will govern how how 00:39:35.880 --> 00:39:40.757 uh successfully the technology diffuses um more so than anything. And 00:39:40.840 --> 00:39:44.517 um you know, we we think deeply about this 00:39:44.600 --> 00:39:47.277 cuz our vision for the technology is 00:39:47.360 --> 00:39:49.597 that it is um we can deploy personal 00:39:49.680 --> 00:39:53.917 agents that uh know you better than most 00:39:54.000 --> 00:39:57.677 people know you, know about your health, know about your goals, know about your 00:39:57.760 --> 00:40:01.717 relationships, know about your family and your friends, and um can really help 00:40:01.800 --> 00:40:06.157 you live your life more effectively, you know, can help you be healthier, can 00:40:06.240 --> 00:40:10.077 help you accomplish projects that you would never would have had time for, can 00:40:10.160 --> 00:40:11.237 help you build the business you always 00:40:11.320 --> 00:40:13.797 wanted, um can help you explore your 00:40:13.880 --> 00:40:17.397 interests. And to properly build that 00:40:17.480 --> 00:40:20.437 requires an immense amount of trust from 00:40:20.520 --> 00:40:23.997 consumers, from governments, from 00:40:24.080 --> 00:40:28.277 from effectively everyone, from all organizations. And so it you know, this 00:40:28.360 --> 00:40:32.517 is one of these interesting moments where 00:40:32.600 --> 00:40:35.277 I think because the the sort of 00:40:35.360 --> 00:40:37.717 the form that of the technology that we 00:40:37.800 --> 00:40:40.517 wanted it to take is so intimate and is 00:40:40.600 --> 00:40:43.997 so deeply tied with what we think is 00:40:44.080 --> 00:40:49.477 really the the future of you know, the consumer life and the consumer 00:40:49.560 --> 00:40:52.757 you know, is at the core of the relationship with technology. 00:40:52.840 --> 00:40:56.877 Um The figuring out ways to develop the 00:40:56.960 --> 00:41:00.957 technology in partnership with governing bodies and in partnership with 00:41:01.040 --> 00:41:05.397 nonprofits and other organizations to garner that trust is very important. 00:41:05.480 --> 00:41:08.517 This is something we've seen at Meta actually in 00:41:08.600 --> 00:41:12.237 with WhatsApp. You know, one of the we were talking about this actually 00:41:12.320 --> 00:41:15.197 backstage, but WhatsApp has become 00:41:15.280 --> 00:41:18.117 incredibly ubiquitous and a real I think 00:41:18.200 --> 00:41:20.957 utility for much of the world in large 00:41:21.040 --> 00:41:25.477 part due to its commitment to privacy and the the 00:41:25.560 --> 00:41:28.997 sort of trust that we've built around WhatsApp as an organization. 00:41:29.080 --> 00:41:33.477 Often times at the sort of you know, not necessarily for the 00:41:33.560 --> 00:41:37.637 optimization of our business objectives, but for the optimization of trust in 00:41:37.720 --> 00:41:39.837 WhatsApp as a platform. I think a 00:41:39.920 --> 00:41:45.317 similar journey will be necessary for the AI products of the modern age and 00:41:45.400 --> 00:41:48.837 we've seen I mean we've already seen various 00:41:48.920 --> 00:41:53.197 kerfuffles or sort of jabs back and forth between the existing AI you know, 00:41:53.280 --> 00:41:56.277 products and labs around you know, whether their actions are truly 00:41:56.360 --> 00:42:01.197 garnering you know, truly building this these levels of trust. But 00:42:01.280 --> 00:42:04.117 and then I think the you know, the topic that 00:42:04.200 --> 00:42:06.157 you really kicked off at Bletchley 00:42:06.240 --> 00:42:08.277 around in a sort of like global way 00:42:08.360 --> 00:42:13.357 around safety, I think is more important than ever. I think we're already seeing 00:42:13.440 --> 00:42:17.317 that um the models are 00:42:17.400 --> 00:42:21.317 uh are are achieving levels of capability 00:42:21.400 --> 00:42:23.877 which cause some pause around some of 00:42:23.960 --> 00:42:27.717 the sort of like risks that uh we 00:42:27.800 --> 00:42:30.357 monitor for these models as it comes to 00:42:30.440 --> 00:42:32.637 various national security risks or uh 00:42:32.720 --> 00:42:35.997 various cybersecurity risks. And so um 00:42:36.080 --> 00:42:38.597 the the importance of properly testing 00:42:38.680 --> 00:42:43.077 and then building the right systems to uh monitor and deploy these models 00:42:43.160 --> 00:42:44.837 safely is is a paramount importance. 00:42:44.920 --> 00:42:47.557 >> And that you and I have been talking about the same topic for some years. And 00:42:47.640 --> 00:42:51.437 that was key to us at uh Bletchley was 00:42:51.520 --> 00:42:55.837 establishing the AI Security Institute in the UK to do exactly that work. And 00:42:55.920 --> 00:42:59.357 the and the thought was there was a gap where the technical capability didn't 00:42:59.440 --> 00:43:03.917 exist amongst governments to evaluate some of these risks. So how could we in 00:43:04.000 --> 00:43:06.077 the UK hopefully take a lead, but in 00:43:06.160 --> 00:43:11.037 sense provide a global public good uh where we're working partnership with 00:43:11.120 --> 00:43:14.637 companies like yours and the other frontier labs to do pre-deployment 00:43:14.720 --> 00:43:19.037 testing on these models. And my thought was always that by doing that, and 00:43:19.120 --> 00:43:22.317 that's not regulation, but it gives citizens confidence that the technology 00:43:22.400 --> 00:43:25.997 is safe cuz ultimately people don't adopt a technology that they don't 00:43:26.080 --> 00:43:29.997 trust. And it's what I my sense hearing you is that there's a role for the 00:43:30.080 --> 00:43:33.877 companies, I thought the the privacy example with WhatsApp is a good one, is 00:43:33.960 --> 00:43:37.637 how do they design AI products that will um 00:43:37.720 --> 00:43:41.437 will engender trust. And then there's a role, it sounds like you agree, for 00:43:41.520 --> 00:43:45.917 governments to do some of that technical evaluation. And is your sense that we 00:43:46.000 --> 00:43:49.277 need, you know, more cooperation between governments on that, more interaction 00:43:49.360 --> 00:43:50.597 with the labs? Is that working as well 00:43:50.680 --> 00:43:54.037 as it should at the moment? Um 00:43:54.120 --> 00:43:58.877 I think the interaction is working well, um but I think this is one area where 00:43:58.960 --> 00:44:02.437 you know uh we always can be more thoughtful, and 00:44:02.520 --> 00:44:07.637 especially as we enter this phase of dramatically increased AI progress. Um I 00:44:07.720 --> 00:44:10.837 think, you know, we constantly need to be thinking about how we can do better. 00:44:10.920 --> 00:44:13.517 I I think one of the other areas where 00:44:13.600 --> 00:44:18.117 this debate on trust will be won or lost is in the public sector. I think it's 00:44:18.200 --> 00:44:22.077 this is the If you experience that health care, more 00:44:22.160 --> 00:44:25.277 efficient government services, a state that is 00:44:25.360 --> 00:44:29.957 quicker at dealing with you, I think this debate goes from abstract to 00:44:30.040 --> 00:44:32.597 you have a real way to demonstrate benefit. 00:44:32.680 --> 00:44:34.637 Do you have a a sense from where you sit 00:44:34.720 --> 00:44:37.517 about where the most obvious use cases 00:44:37.600 --> 00:44:40.917 or deployment cases are in the public 00:44:41.000 --> 00:44:44.837 for AI that we should be excited about and that governments should be thinking 00:44:44.920 --> 00:44:48.997 about? I I do. I think that um 00:44:49.080 --> 00:44:53.757 one of the first you know, if you think of governments as 00:44:53.840 --> 00:44:56.797 um uh service providers in a sense to their 00:44:56.880 --> 00:44:59.477 citizens, I think there's there's a huge 00:44:59.560 --> 00:45:04.077 opportunity for all governments to um be able to deliver their services more 00:45:04.160 --> 00:45:08.717 effectively. Um we're already seeing that uh actually in India um on top of 00:45:08.800 --> 00:45:12.997 WhatsApp, funnily enough, where uh there are in certain states they actually 00:45:13.080 --> 00:45:16.557 deliver uh fast majority of the government services 00:45:16.640 --> 00:45:19.197 directly through WhatsApp. And then the 00:45:19.280 --> 00:45:24.917 uh for transit a nation actually able to uh through what in India. 00:45:25.000 --> 00:45:29.157 Uh we last year 100 million sub 00:45:29.240 --> 00:45:33.757 uh in the What? And that's not AI per se, but the 00:45:33.840 --> 00:45:37.477 the future opportunity that is quite exciting is actually the 00:45:37.560 --> 00:45:39.717 deployment of uh 00:45:39.800 --> 00:45:44.597 uh of AI to greatly increase the ease and and speed 00:45:44.680 --> 00:45:46.077 at which these government services are 00:45:46.160 --> 00:45:49.237 deployed and provided assistance. Um 00:45:49.320 --> 00:45:52.117 I think that you know, there's one uh 00:45:52.200 --> 00:45:54.397 term that I've always I've used and and 00:45:54.480 --> 00:45:56.837 I like this concept of proactive 00:45:56.920 --> 00:45:59.717 government, right? And it really is, you 00:45:59.800 --> 00:46:02.517 know, if you think about how should my 00:46:02.600 --> 00:46:07.237 governments be how should the state be imagined? 00:46:07.320 --> 00:46:10.117 You know, I think we need to move away from 00:46:10.200 --> 00:46:14.237 large-scale bureaucratic services 00:46:14.320 --> 00:46:18.557 and support organizations and think about how can the government deliver 00:46:18.640 --> 00:46:23.477 quickly and promptly deliver using computer people and make it easy to verify 00:46:23.560 --> 00:46:27.677 one of those agencies as fast as possible government services 00:46:27.760 --> 00:46:32.357 so that citizens can steer the state to develop solutions. 00:46:32.440 --> 00:46:35.917 Um and I think this cuts through civil services I think it cuts through health 00:46:36.000 --> 00:46:42.600 care as you mentioned and certainly has a has a large role to play in security 00:46:42.960 --> 00:46:47.677 and in the government with the government 00:46:47.760 --> 00:46:50.517 so I think it's quite a problem. 00:46:50.600 --> 00:46:57.080 I think you were you were talking about the concept of jet technology 00:46:57.400 --> 00:46:58.677 and that sort 00:46:58.760 --> 00:47:04.840 I need one more like one more like Uh I want to suggest something 00:47:05.440 --> 00:47:09.117 One of the benefits of leaving office is you have 00:47:09.200 --> 00:47:13.597 I think A lot of benefits leaving office you have a lot more time as I said 00:47:13.680 --> 00:47:17.637 and when I left someone gave me a great book by a guy called Jeffrey Ding, 00:47:17.720 --> 00:47:21.477 Professor Ding and technology and and great powers I'm sure you you know the 00:47:21.560 --> 00:47:26.237 book and and in it he makes a really compelling argument about technology 00:47:26.320 --> 00:47:30.477 diffusion essentially looking at the history of general purpose these going 00:47:30.560 --> 00:47:34.077 all the way back to the printing press and making the argument that you don't 00:47:34.160 --> 00:47:36.317 need to be the country that invents the 00:47:36.400 --> 00:47:38.397 technology to be the country that gets 00:47:38.480 --> 00:47:41.557 the most benefit from it and I think you 00:47:41.640 --> 00:47:44.237 know you you touched on India and India 00:47:44.320 --> 00:47:46.357 has recognized that technology 00:47:46.440 --> 00:47:49.277 leadership exactly does not depend on 00:47:49.360 --> 00:47:54.317 just inventing the technology it's about how effectively you deploy it and they 00:47:54.400 --> 00:47:58.597 here have focused on mass broad adoption and obviously 00:47:58.680 --> 00:48:01.317 supported by a few critical attributes. 00:48:01.400 --> 00:48:05.997 You have a very deep talent pool 00:48:06.080 --> 00:48:09.957 You talked about the rails of distribution. Obviously the digital 00:48:10.040 --> 00:48:12.397 public infrastructure here. 00:48:12.480 --> 00:48:15.277 The ID, UPI, the payments infrastructure 00:48:15.360 --> 00:48:18.797 and then most recently health accounts 00:48:18.880 --> 00:48:23.757 provide these infrastructure rails to distribute apps to over a billion 00:48:23.840 --> 00:48:27.197 people. And they also have a population that's supportive of that. 00:48:27.280 --> 00:48:31.917 And I think focusing on that has allowed India to position itself as a leader 00:48:32.000 --> 00:48:37.117 by by pursuing the deployment and adoption strategy. 00:48:37.200 --> 00:48:42.077 But thinking about that, you know, what are those critical attributes that 00:48:42.160 --> 00:48:46.397 government has to get right? If we want to turn your aspiration into a reality 00:48:46.480 --> 00:48:49.997 about agentic government, about delivering for our citizens. You know, I 00:48:50.080 --> 00:48:52.637 think government does need to think about things differently. It needs to 00:48:52.720 --> 00:48:56.597 build the architecture around deployment 00:48:56.680 --> 00:48:59.957 and diffusion, whether it's skills, whether it's compute, whether it's 00:49:00.040 --> 00:49:03.717 regulation. But crucially probably also data, 00:49:03.800 --> 00:49:07.357 which is something that you know better than most. So maybe if we start there 00:49:07.440 --> 00:49:11.437 and then broaden it out. If I'm you know, sitting as a government leader as 00:49:11.520 --> 00:49:14.677 many people in the audience are and trying to figure out what do I need to 00:49:14.760 --> 00:49:19.477 get right? What are the building blocks to deliver this thing properly? Let's 00:49:19.560 --> 00:49:22.397 start with data and what would your advice be there and then we'll we'll 00:49:22.480 --> 00:49:26.837 broaden it out a little bit. Yeah, there there is 00:49:26.920 --> 00:49:28.757 uh data 00:49:28.840 --> 00:49:31.277 is is so important. I think over the 00:49:31.360 --> 00:49:34.877 past few years many people have I think 00:49:34.960 --> 00:49:38.677 you know, realized this as we sort of look at the power of these models and we 00:49:38.760 --> 00:49:42.477 look behind the curtains and you realize that all the powers and beauty of these 00:49:42.560 --> 00:49:47.357 models are actually imbued by data. And if you think about 00:49:47.440 --> 00:49:50.917 So if you think about data is like in some ways the sort of new oil, the sort 00:49:51.000 --> 00:49:55.117 of like base material that is incredibly important for the future 00:49:55.200 --> 00:49:58.437 of of AI and the development of these models. 00:49:58.520 --> 00:50:03.437 Then I think as like governments should think about what are their, you know, 00:50:03.520 --> 00:50:06.197 what are their reserves? What are their data reserves? What are their data 00:50:06.280 --> 00:50:11.597 assets? And and to in many ways 00:50:11.680 --> 00:50:16.557 build strategies where they have sovereignty over some set of data or or 00:50:16.640 --> 00:50:21.037 they're able to build certain certain pockets of data sets that enable the the 00:50:21.120 --> 00:50:24.877 country to perform the best. I think one of the areas that is the two areas that 00:50:24.960 --> 00:50:27.077 is clear that are clearest for this are 00:50:27.160 --> 00:50:31.957 one is health care. I think many countries 00:50:32.040 --> 00:50:35.997 already are and other countries need to be thinking I think about 00:50:36.080 --> 00:50:39.437 very effective mechanisms for uh 00:50:39.520 --> 00:50:42.397 for health care data within the country 00:50:42.480 --> 00:50:44.837 to enable ultimately over time there to 00:50:44.920 --> 00:50:49.637 be AI models and AI agents that are built on top of that to deliver better 00:50:49.720 --> 00:50:53.157 health services to their citizens. And then the second is national security. 00:50:53.240 --> 00:50:56.797 And I think that is you know, in many ways pretty self-explanatory, but I 00:50:56.880 --> 00:51:00.677 think you know, countries need to be thinking about that 00:51:00.760 --> 00:51:04.557 quite deeply as to what are their sort of like core data assets. Then I think 00:51:04.640 --> 00:51:08.557 if you if you sort of build around data and you look at the 00:51:08.640 --> 00:51:10.957 next few layers, the next one is is 00:51:11.040 --> 00:51:13.357 infrastructure. So, you know, we think 00:51:13.440 --> 00:51:16.317 about the power, the 00:51:16.400 --> 00:51:21.037 the computational infrastructure and and what does that story look like? 00:51:21.120 --> 00:51:24.957 I don't think every country needs to build large-scale AI data centers in 00:51:25.040 --> 00:51:29.157 their country, but they do need to have a strategy of over time how are they 00:51:29.240 --> 00:51:32.797 going to have access to that infrastructure. And so whether this is 00:51:32.880 --> 00:51:36.197 in the form of alliances of various formats or 00:51:36.280 --> 00:51:41.597 or partnerships even with with corporate entities, the large cloud providers, I 00:51:41.680 --> 00:51:44.877 think that's an important element to think through. 00:51:44.960 --> 00:51:48.357 Then comes I think the sort of the 00:51:48.440 --> 00:51:51.277 the innovation ecosystem. I do think 00:51:51.360 --> 00:51:52.997 that, you know, we're in a moment of 00:51:53.080 --> 00:51:57.357 momentous technological change, and 00:51:57.440 --> 00:52:01.597 it is, I think, incredibly important for I think there is 00:52:01.680 --> 00:52:04.917 near boundless entrepreneurial and economic 00:52:05.000 --> 00:52:10.477 opportunity, and I think it is the responsibility of 00:52:10.560 --> 00:52:13.597 governments and countries to think through how they can build ecosystems 00:52:13.680 --> 00:52:17.277 that support that. I think India is actually, in many ways, a great 00:52:17.360 --> 00:52:21.117 very positive case study, in large part due to the due to the talent pool, but 00:52:21.200 --> 00:52:25.917 you know, I was at a dinner with a number of Indian founders and venture 00:52:26.000 --> 00:52:27.837 capitalists last night, and and the stat 00:52:27.920 --> 00:52:30.277 was, I think, there are more 00:52:30.360 --> 00:52:33.437 consumer AI companies in India consumer 00:52:33.520 --> 00:52:37.557 AI startups in India than in the United States. And so, there really are some of 00:52:37.640 --> 00:52:39.477 these like, you know, shining examples 00:52:39.560 --> 00:52:45.197 of incredible development of these ecosystems. And then, um And then, 00:52:45.280 --> 00:52:47.397 lastly, I think are the government services, which we talked about 00:52:47.480 --> 00:52:50.677 previously. So, we've got, I think, the elements of almost a a playbook 00:52:50.760 --> 00:52:54.837 emerging, right? You need to have the right data and make sure it's it's used 00:52:54.920 --> 00:52:57.757 effectively. You need some degree of compute. 00:52:57.840 --> 00:53:01.637 You've got appropriate regulation, as we've talked about, to build to build 00:53:01.720 --> 00:53:04.557 trust. And then, we've got some public sector 00:53:04.640 --> 00:53:08.477 use cases, as well, and an innovation ecosystem. So, these are the kind of 00:53:08.560 --> 00:53:10.317 ingredients where you can start to turn 00:53:10.400 --> 00:53:13.517 these ambitions into reality and and 00:53:13.600 --> 00:53:16.957 actual impact on people. I I just 00:53:17.040 --> 00:53:21.757 staying on that theme, you know, you you work at an incredible company. You work 00:53:21.840 --> 00:53:26.117 with other enterprises. You know, I talked to a lot of CEOs, and 00:53:26.200 --> 00:53:31.197 and I'm struck by how they're thinking about deploying AI in their businesses, 00:53:31.280 --> 00:53:35.237 and thinking about the parallel between that and and what governments should be 00:53:35.320 --> 00:53:39.437 doing. And you know, I think there is a kind of clear parallel. I'm intrigued. I 00:53:39.520 --> 00:53:43.637 think CEOs are, don't, know, kind of constantly thinking about AI. I think 00:53:43.720 --> 00:53:47.957 almost any large company CEO is spending their time figuring out, am I going to 00:53:48.040 --> 00:53:52.237 get disrupted, disintermediated by AI, how do I use it to stay ahead, where do 00:53:52.320 --> 00:53:54.317 I use it to to get more efficiency and 00:53:54.400 --> 00:53:57.037 more growth. And they are driving 00:53:57.120 --> 00:54:01.997 projects from the top down. And they are saying, look, this is something that is, 00:54:02.080 --> 00:54:06.517 you know, not not a topic that can just stay with the IT 00:54:06.600 --> 00:54:10.397 department, right? This is something that has to be driven by the CEO, where 00:54:10.480 --> 00:54:13.957 we pick the use cases, we have accountability, 00:54:14.040 --> 00:54:17.597 and we start piloting these and scaling them up quite quickly. 00:54:17.680 --> 00:54:20.237 I, you know, I my sense is we haven't 00:54:20.320 --> 00:54:25.037 quite seen that that same transition in governments yet, but I thought it's a 00:54:25.120 --> 00:54:28.037 good analogy in that look, this has to be driven by the person at the top. At 00:54:28.120 --> 00:54:31.277 the end of the day, if you're a prime minister, and I'm sure Tony would say 00:54:31.360 --> 00:54:35.357 the same thing, and indeed has, you can only do a few things that you drive 00:54:35.440 --> 00:54:38.917 personally from the top, and this really has to be one of them. You know, what 00:54:39.000 --> 00:54:41.357 are your kind of thoughts or advice for, 00:54:41.440 --> 00:54:43.757 I guess, both CEOs and and governments 00:54:43.840 --> 00:54:46.637 about how to kind of drive the machine. 00:54:46.720 --> 00:54:51.197 What does it take to get execution on this? And I mean, you've had to do that 00:54:51.280 --> 00:54:54.197 at Meta yourself, right? So, you'll have a sense from that. 00:54:54.280 --> 00:54:59.517 Yeah, I think it's a mix of of two things. I think first, um at 00:54:59.600 --> 00:55:03.197 uh at the highest level, you need to make 00:55:03.280 --> 00:55:06.877 it very clear that the organization will be an AI-first organization. And what 00:55:06.960 --> 00:55:08.357 that means for every organization is 00:55:08.440 --> 00:55:11.957 slightly different, but it it it um like 00:55:12.040 --> 00:55:16.117 the way that truly I think about it is that uh the success or failure of the 00:55:16.200 --> 00:55:18.677 organization actually entirely rests on 00:55:18.760 --> 00:55:23.757 its ability to properly embrace AI. And so, that's sort of the tops-down um 00:55:23.840 --> 00:55:27.277 component of it. And you know, there's sort of a strategy that must be designed 00:55:27.360 --> 00:55:30.557 more thoughtfully. And actually, honestly, you can probably ask AI uh to 00:55:30.640 --> 00:55:33.997 help you design that strategy now, and it'll probably do a pretty good job. And 00:55:34.080 --> 00:55:37.157 then you need um a a 00:55:37.240 --> 00:55:40.637 bed of bottoms-up adoption and 00:55:40.720 --> 00:55:43.357 bottoms-up, um, success stories. And 00:55:43.440 --> 00:55:48.477 for, uh, for us at Meta, like the example of this are, you know, we have 00:55:48.560 --> 00:55:50.237 these engineers throughout the company 00:55:50.320 --> 00:55:53.157 who are able to use AI to, uh, be, you 00:55:53.240 --> 00:55:55.597 know, 10 to 100 times more productive. 00:55:55.680 --> 00:55:59.517 And, um, those are just such incredible 00:55:59.600 --> 00:56:03.637 examples that where all of a sudden everyone else in the organization looks 00:56:03.720 --> 00:56:07.037 around and realizes, "Wow, this is really something that must be taken very 00:56:07.120 --> 00:56:11.797 seriously. It's not just something that, you know, um, the CEO or the C-level 00:56:11.880 --> 00:56:15.117 individuals are talking about and, you know, that they're so detached from 00:56:15.200 --> 00:56:19.477 reality." It's like, "No, it's very real on the ground, um, in a very big way." 00:56:19.560 --> 00:56:23.397 And so, I think you need this sort of mix of what are the success stories that 00:56:23.480 --> 00:56:27.997 cause everyone to realize how big of a shift it actually is. And then also 00:56:28.080 --> 00:56:31.997 ensure that there's a continual, um, there's a very clear top-down direction 00:56:32.080 --> 00:56:34.717 that this transition is, is mandatory. 00:56:34.800 --> 00:56:37.237 >> very helpful advice. Look, we're almost 00:56:37.320 --> 00:56:39.877 out of time. So, maybe a last final 00:56:39.960 --> 00:56:42.517 point, uh, to close on is, you know, 00:56:42.600 --> 00:56:47.117 you're looking ahead, you know, what are you most excited about, uh, in the 00:56:47.200 --> 00:56:48.917 development of this, of this technology? 00:56:49.000 --> 00:56:52.917 You've touched on personal agents, other scientific breakthroughs. You know, 00:56:53.000 --> 00:56:56.957 what, where is this going and what should governments be doing now, you 00:56:57.040 --> 00:56:58.677 know, so that they are prepared and and 00:56:58.760 --> 00:57:01.317 not playing catch-up? 00:57:01.400 --> 00:57:06.357 I think the things that, um, you can depend on, I think the, the pace of 00:57:06.440 --> 00:57:07.797 technology development will continue 00:57:07.880 --> 00:57:13.757 accelerating. Um, I think this, this era of recursive self-improvement will, 00:57:13.840 --> 00:57:16.957 uh, feel like everything is speeding up and 00:57:17.040 --> 00:57:21.357 I think that, um, the technology will continue developing very quickly. Um, I 00:57:21.440 --> 00:57:25.997 also think that the, the diffusion of the technology is, is going to 00:57:26.080 --> 00:57:29.717 accelerate. I think there's been this question for many, for the, for the past 00:57:29.800 --> 00:57:33.437 many years, which is, everyone's talking about AI, but where is AI in my daily 00:57:33.520 --> 00:57:36.237 life? I think that um that really is 00:57:36.320 --> 00:57:40.757 going to accelerate. Maybe not um you know, it's not going to go be 100 times 00:57:40.840 --> 00:57:42.797 faster, but maybe it'll be you know, two 00:57:42.880 --> 00:57:45.517 or three times faster every year for the 00:57:45.600 --> 00:57:47.997 you know, for the next many years. Um 00:57:48.080 --> 00:57:53.117 and and I think the thing that uh where 00:57:53.200 --> 00:57:57.557 where I really am excited about this going is um you know, we talk about this 00:57:57.640 --> 00:58:02.237 idea of personal superintelligence within Meta, which is you know, how does 00:58:02.320 --> 00:58:06.077 you know, how does each person have an agent that helps them accomplish the 00:58:06.160 --> 00:58:09.837 things that they're you know, ex- extremely excited about. 00:58:09.920 --> 00:58:13.117 How can they help them, you know, explore their passions, build a 00:58:13.200 --> 00:58:15.437 business, make a discovery, invent a 00:58:15.520 --> 00:58:18.557 thing, um you know, um you know, achieve 00:58:18.640 --> 00:58:20.957 the most maximal version of their own 00:58:21.040 --> 00:58:24.517 lives. And I think where um this is 00:58:24.600 --> 00:58:28.197 going even beyond then is sort of this concept of interpersonal 00:58:28.280 --> 00:58:31.877 superintelligence, which is um I think we already see in human society 00:58:31.960 --> 00:58:36.597 sometimes the thing that is necessary for big things to happen are just for 00:58:36.680 --> 00:58:39.797 the right set of people to be organized in the right way. You see this with 00:58:39.880 --> 00:58:43.837 companies, you see this with nonprofits, you see this with large movements, um 00:58:43.920 --> 00:58:47.757 you see this even with with um governments when they uh when they 00:58:47.840 --> 00:58:50.037 reorganize. And I think there's an 00:58:50.120 --> 00:58:51.357 incredible opportunity for the 00:58:51.440 --> 00:58:55.237 technology of AI to enable humans to 00:58:55.320 --> 00:58:58.437 organize in ways that are even better um 00:58:58.520 --> 00:59:02.557 than we do today. And all of a sudden maybe we'll be able to organize as 00:59:02.640 --> 00:59:06.437 communities or as groups or as new companies or who knows what it'll look 00:59:06.520 --> 00:59:10.517 like to uh to solve many of our greatest challenges. And so um you know, it 00:59:10.600 --> 00:59:13.997 sounds a bit highfalutin, but um but it's something I'm genuinely quite 00:59:14.080 --> 00:59:16.477 excited about. >> No, I can see that. It's a very a very 00:59:16.560 --> 00:59:19.757 compelling vision. But this has been fascinating conversation. It's clear 00:59:19.840 --> 00:59:23.557 that the technology is is not just keeping pace, that it's accelerating. 00:59:23.640 --> 00:59:27.037 It's clear that the debate is moving from you know, just talking about the 00:59:27.120 --> 00:59:30.917 technology to as you said, how is it making a a difference in people's lives 00:59:31.000 --> 00:59:34.757 and being deployed. And that one of one of my concerns, and maybe we'll finish 00:59:34.840 --> 00:59:38.717 on this, is that, you know, I think some, not all, some political leaders 00:59:38.800 --> 00:59:42.637 think that AI is is going to be tomorrow's issue, when I think they need 00:59:42.720 --> 00:59:47.717 to recognize that it's it's an action this day issue. And this needs to go 00:59:47.800 --> 00:59:49.317 from, you know, AI needs to go from 00:59:49.400 --> 00:59:52.477 being a kind of specialized topic to a 00:59:52.560 --> 00:59:57.117 centralized responsibility of government. Uh and it needs to have that 00:59:57.200 --> 01:00:00.877 same urgency and and action attached to it so we can realize all the benefits 01:00:00.960 --> 01:00:04.557 that Alex is talking about. But this has been a great discussion. Thanks for all 01:00:04.640 --> 01:00:07.397 your advice and tips for everyone in the audience and giving us an insight into 01:00:07.480 --> 01:00:09.197 how things are going. Thanks very much. 01:00:09.280 --> 01:00:13.080 Thank you so much. >> [applause] 01:00:16.880 --> 01:00:18.997 >> Thank you so much to our speakers. 01:00:19.080 --> 01:00:22.317 Earlier, can we have another round of applause, please, for Rishi Sunak and 01:00:22.400 --> 01:00:26.357 Alexander Wa When I became Prime Minister, I'm an AI obsessed, AI 01:00:26.440 --> 01:00:29.837 obsessed because of my time in in California. You you get the job in 01:00:29.920 --> 01:00:33.037 Downing Street and one of the earliest meetings you have is with the the 01:00:33.120 --> 01:00:37.877 Foreign Affairs team. And they sit down with you. I can I can see someone who 01:00:37.960 --> 01:00:39.757 will remember this in the audience. 01:00:39.840 --> 01:00:41.917 And they go over your international 01:00:42.000 --> 01:00:47.517 calendar for the next 12 months. And they they want to kind of figure out, 01:00:47.600 --> 01:00:50.357 you know, what you should be doing. And it turns out there's an international 01:00:50.440 --> 01:00:54.677 summit for basically every single thing that you can imagine. Uh turns out there 01:00:54.760 --> 01:00:58.277 were about 150. Uh and the British Foreign Office being 01:00:58.360 --> 01:01:01.037 the British Foreign Office, they wanted me to go to all of them. 01:01:01.120 --> 01:01:04.157 Uh and although I did I did say I'm going to as a British Prime Minister, I 01:01:04.240 --> 01:01:07.877 think I can draw the line at the United Nations high-level meeting on 01:01:07.960 --> 01:01:12.357 sandstorms. That's where we agreed to disagree. Um but I noticed that there 01:01:12.440 --> 01:01:16.157 was no gathering to bring together people to talk about what I thought was 01:01:16.240 --> 01:01:19.317 the most transformative technology of our time. And I think we're all 01:01:19.400 --> 01:01:22.317 fortunate to be living at a time where 01:01:22.400 --> 01:01:24.717 AI is a general purpose technology. 01:01:24.800 --> 01:01:28.597 These don't come along all that often and when they do, they had the 01:01:28.680 --> 01:01:32.597 opportunity in the past and will in the future to change every aspect of our 01:01:32.680 --> 01:01:37.197 economies, of our societies, uh of our lives. And that's why I thought it was 01:01:37.280 --> 01:01:42.397 necessary that leaders came together to discuss this at a regular basis. But, uh 01:01:42.480 --> 01:01:45.957 I also thought it needed to be done slightly differently. Um in contrast to 01:01:46.040 --> 01:01:47.517 previous technology cycles, it's 01:01:47.600 --> 01:01:50.397 companies like Alex's and 01:01:50.480 --> 01:01:53.117 innovators like Alex himself who were at 01:01:53.200 --> 01:01:57.477 the frontier of developing this technology. Uh and I thought they needed 01:01:57.560 --> 01:02:01.397 to be in the room. Not least because it's businesses like yours that are have 01:02:01.480 --> 01:02:05.877 the resources to invest in this. Just to give you a bit of context, this year, 01:02:05.960 --> 01:02:08.157 you know, the the large AI companies are 01:02:08.240 --> 01:02:11.917 going to spend 20 times more 01:02:12.000 --> 01:02:16.717 on developing this technology than the US did for the Manhattan Project. And if 01:02:16.800 --> 01:02:19.597 you think about seminal moments where the government mobilized to do 01:02:19.680 --> 01:02:23.717 something, 20 times more just this year alone. Um so, I thought, look, it's 01:02:23.800 --> 01:02:25.637 vital that we do this in a different 01:02:25.720 --> 01:02:28.477 way, not like the G7 and G20. We bring 01:02:28.560 --> 01:02:32.837 industry, innovators, and and government uh together and that's uh that's what we 01:02:32.920 --> 01:02:35.117 did at Bletchley Park a few years ago. 01:02:35.200 --> 01:02:36.917 The summit series went to Paris, South 01:02:37.000 --> 01:02:39.197 Korea, and now, brilliantly, in India, a 01:02:39.280 --> 01:02:43.317 country with huge digital ambitions and 01:02:43.400 --> 01:02:46.357 uh indeed capacity. Um but I think over 01:02:46.440 --> 01:02:48.557 those years, and Alex, let's start with 01:02:48.640 --> 01:02:50.677 this, I think actually the the the 01:02:50.760 --> 01:02:54.237 debate on AI has shifted from 01:02:54.320 --> 01:02:57.157 kind of technology to strategy. Uh you 01:02:57.240 --> 01:03:01.437 know, from what these tools can do to what governments are choosing to do with 01:03:01.520 --> 01:03:04.797 them. And I wanted to maybe just start there. 01:03:04.880 --> 01:03:07.717 Is if you think about the last few years, you are someone who is at the 01:03:07.800 --> 01:03:10.157 frontier of developing this technology. 01:03:10.240 --> 01:03:13.397 You know, the the journey that we've been on in the last few years, you know, 01:03:13.480 --> 01:03:16.997 help us understand that. Where have we come? 01:03:17.080 --> 01:03:19.357 What is it that is exciting at this this 01:03:19.440 --> 01:03:24.517 moment and just ground us in in where we are so that we can get get on to some of 01:03:24.600 --> 01:03:27.237 these questions about what we do with it. 01:03:27.320 --> 01:03:29.117 Yeah, I think 01:03:29.200 --> 01:03:31.957 I mean we are sitting here right now at 01:03:32.040 --> 01:03:34.957 such an incredibly exciting moment for 01:03:35.040 --> 01:03:37.637 the technology. I think we are 01:03:37.720 --> 01:03:42.677 at the beginning of uh you know, of a true acceleration in 01:03:42.760 --> 01:03:46.197 the technology and you know, in many ways it's felt like this very clear 01:03:46.280 --> 01:03:48.917 acceleration even until now, but I think 01:03:49.000 --> 01:03:53.757 we're going to see sort of 2026 is an inflection point in many ways. I 01:03:53.840 --> 01:03:55.517 think to contextualize this over the 01:03:55.600 --> 01:03:59.117 past few years, I think from 01:03:59.200 --> 01:04:00.797 2018 01:04:00.880 --> 01:04:04.597 through 2024 01:04:04.680 --> 01:04:07.197 really, which is a good 6-year period, 01:04:07.280 --> 01:04:09.717 we were writing we were in the era of 01:04:09.800 --> 01:04:14.637 pre-training. So, we were writing one very large exponential curve towards 01:04:14.720 --> 01:04:16.477 improving the performance of the models. 01:04:16.560 --> 01:04:20.037 And that curve was very predictable and it 01:04:20.120 --> 01:04:23.877 sort of had clear characteristics where the more resources you put in, the more 01:04:23.960 --> 01:04:27.397 results you would get out and there was a lot of questions at the time of, you 01:04:27.480 --> 01:04:30.917 know, whether or not this curve would continue and even to the degree it did 01:04:31.000 --> 01:04:34.517 continue, were there diminishing marginal returns? And I think this was sort of 01:04:34.600 --> 01:04:35.677 one of the 01:04:35.760 --> 01:04:37.637 one of the more recent sort of 01:04:37.720 --> 01:04:42.357 eras of AI skepticism. And then in 2024, 01:04:42.440 --> 01:04:47.157 like towards the end of 2024, began the era of reinforcement learning. 01:04:47.240 --> 01:04:49.277 So, this was sort of the the next major 01:04:49.360 --> 01:04:51.677 wave that stacked on top of the previous 01:04:51.760 --> 01:04:54.357 wave and this is when we saw the models 01:04:54.440 --> 01:04:59.117 begin to learn to reason. And you know, there's always this meme of 01:04:59.200 --> 01:05:01.237 what did Ilya see? Well, Ilya saw 01:05:01.320 --> 01:05:03.557 reasoning from the models. And it was 01:05:03.640 --> 01:05:06.957 very clearly this this next mega wave of 01:05:07.040 --> 01:05:12.277 of the technology and even reinforcement learning had, you know, clearly some 01:05:12.360 --> 01:05:16.117 limitations and there were all these questions around when that was going to, 01:05:16.200 --> 01:05:17.757 you know, how long that era was going to 01:05:17.840 --> 01:05:19.677 last and maybe there were limitations in 01:05:19.760 --> 01:05:23.157 that era. And then at the end of 2025, 01:05:23.240 --> 01:05:24.877 so just a few months ago, I think we 01:05:24.960 --> 01:05:27.077 really began 01:05:27.160 --> 01:05:30.677 an entirely new paradigm or era of the 01:05:30.760 --> 01:05:32.997 technology, which is 01:05:33.080 --> 01:05:35.557 really beginning the era of recursive 01:05:35.640 --> 01:05:40.677 self-improvement. And I think we're really starting to see this in the 01:05:40.760 --> 01:05:43.797 development of our models. I think, you know, most of the major labs are seeing 01:05:43.880 --> 01:05:47.557 this in the development of their models, which is that the models themselves have 01:05:47.640 --> 01:05:50.717 now become instrumental in accelerating 01:05:50.800 --> 01:05:53.037 the process of producing the next AIs. 01:05:53.120 --> 01:05:57.237 And so, you know, the the sort of from the outside what you see is maybe an 01:05:57.320 --> 01:06:01.117 acceleration of overall development velocity, an acceleration in at which 01:06:01.200 --> 01:06:04.797 the speed at which the the new models are being released. Internally, what we 01:06:04.880 --> 01:06:08.637 see is just dramatic speed-ups. I think the sort of 01:06:08.720 --> 01:06:10.797 the productivity of an individual 01:06:10.880 --> 01:06:13.677 researcher has grown dramatically, 01:06:13.760 --> 01:06:16.637 and we don't expect that to stop. In 01:06:16.720 --> 01:06:21.957 fact, we expect it to kind of smoothly continue accelerating where the output 01:06:22.040 --> 01:06:25.117 of an individual researcher, an individual engineer, well, I think 01:06:25.200 --> 01:06:28.317 that itself will continue growing exponentially as the technology 01:06:28.400 --> 01:06:30.517 improves. And 01:06:30.600 --> 01:06:35.597 the other mega wave that is somewhat coincident with this new paradigm within 01:06:35.680 --> 01:06:38.597 the sort of development of the technology of 01:06:38.680 --> 01:06:41.597 a recursive self-improvement is really 01:06:41.680 --> 01:06:46.237 the arrival of agents in earnest. And 01:06:46.320 --> 01:06:48.557 this is something that has been talked 01:06:48.640 --> 01:06:51.917 about since I honestly want to say 2023 01:06:52.000 --> 01:06:54.437 with a lot of fanfare, and for a long 01:06:54.520 --> 01:06:57.157 long time it was sort of 01:06:57.240 --> 01:07:02.197 one of these hype sort of buzzwords that, you know, never 01:07:02.280 --> 01:07:04.117 actually lived up to the expectations. 01:07:04.200 --> 01:07:07.797 And I think also at the end of 2025, we 01:07:07.880 --> 01:07:12.357 and toward you know, middle to end of 2025, we started to see agents actually 01:07:12.440 --> 01:07:16.357 work. And this is, you know, these are where the models can actually start 01:07:16.440 --> 01:07:19.197 taking um actions um and start 01:07:19.280 --> 01:07:21.517 automating entire workflows and being um 01:07:21.600 --> 01:07:23.837 sort of dramatically more 01:07:23.920 --> 01:07:27.157 uh productive as actual sort of entities 01:07:27.240 --> 01:07:31.437 onto themselves versus assistants. And um 01:07:31.520 --> 01:07:34.957 And this era is, you know, it started with coding agents. We're starting to 01:07:35.040 --> 01:07:36.797 see it um really percolate through 01:07:36.880 --> 01:07:38.277 personal agent. This is a big part of 01:07:38.360 --> 01:07:42.077 what uh we believe in at Meta. Um and I 01:07:42.160 --> 01:07:44.397 think over the course of 2026, we will 01:07:44.480 --> 01:07:48.797 see large-scale agent deployments um in 01:07:48.880 --> 01:07:50.957 many areas of the economy, many areas of 01:07:51.040 --> 01:07:54.957 of the world, and um the sort of, you 01:07:55.040 --> 01:07:58.677 know, if the the GDP of AI, so to speak, 01:07:58.760 --> 01:08:00.277 uh is going to grow exponentially. 01:08:00.360 --> 01:08:03.037 >> Well, that that is uh very welcome to to 01:08:03.120 --> 01:08:07.797 hear. I I think this you you hit on something that is important because 01:08:07.880 --> 01:08:11.237 I remember when I first started doing this in the job, the conversations were 01:08:11.320 --> 01:08:14.077 all around these scaling laws and would they continue? And you mentioned the 01:08:14.160 --> 01:08:16.117 skepticism. I mean, not only have they 01:08:16.200 --> 01:08:18.837 continued and continue to deliver this 01:08:18.920 --> 01:08:23.397 exponential growth, it feels very much like the last few months have seen an 01:08:23.480 --> 01:08:25.637 acceleration. It's what you touched on. 01:08:25.720 --> 01:08:30.317 It's what Dario and and Demis have have also said to me and and publicly, driven 01:08:30.400 --> 01:08:33.757 by this recursive self-improvement, that the models themselves are now able to 01:08:33.840 --> 01:08:36.357 help us develop the new models quicker. 01:08:36.440 --> 01:08:40.197 So, you know, we're at this quite exciting but also a moment of some 01:08:40.280 --> 01:08:43.237 anxiety because of that accelerating uh 01:08:43.320 --> 01:08:47.637 development time scale. So, I wanted to just then talk about trust 01:08:47.720 --> 01:08:50.037 because I I tell you give you a story. 01:08:50.120 --> 01:08:54.677 My I have two, you know, young girls and you know, as they're using their 01:08:54.760 --> 01:08:58.397 chatbots and things and they're always like, "Please, you know, if so-and-so 01:08:58.480 --> 01:09:02.237 and so-and-so." And thank you when they finish writing their request or they get 01:09:02.320 --> 01:09:06.157 the answer back. There's lots of pleases and and thank yous everywhere. And I and 01:09:06.240 --> 01:09:08.797 I I thought that was, you know, interesting. But I said to them, I said, 01:09:08.880 --> 01:09:12.677 "Look, there's, you know, you don't need to do that. There's not not a person it 01:09:12.760 --> 01:09:16.357 doesn't it doesn't matter. And and by the way, you know, it costs a fortune 01:09:16.440 --> 01:09:18.277 when we have to do the compute for that. 01:09:18.360 --> 01:09:21.997 Um and and they they said to me, "You know what, Daddy?" They said, "If if AI 01:09:22.080 --> 01:09:25.997 takes over the world, we want to have been kind to the AI." 01:09:26.080 --> 01:09:29.237 Uh and I thought that was good good insurance policy. But it it it kind of 01:09:29.320 --> 01:09:32.317 it, you know, it touches on this anxiety. Very wise. 01:09:32.400 --> 01:09:35.357 >> [laughter] >> Um in that, maybe. 01:09:35.440 --> 01:09:38.557 You know, across the world, we're seeing these different attitudes 01:09:38.640 --> 01:09:43.077 uh towards AI. In countries like India, where we are, there's enormous optimism 01:09:43.160 --> 01:09:46.557 and and trust. And then actually in Western countries, you know, we're 01:09:46.640 --> 01:09:48.917 seeing that anxiety is still the kind of 01:09:49.000 --> 01:09:51.557 dominant feeling towards AI. Now, I 01:09:51.640 --> 01:09:54.997 think closing that confidence gap is uh 01:09:55.080 --> 01:09:59.437 as much a policy task as it is a a technical one. But I I'd love to just 01:09:59.520 --> 01:10:02.957 get your thoughts. And then obviously you are at a, you know, B2C company 01:10:03.040 --> 01:10:05.877 dealing with, you know, billions of consumers every 01:10:05.960 --> 01:10:09.677 day. You know, how do you think about this trust question? You know, how do we 01:10:09.760 --> 01:10:13.517 get trust up in places where it is not? For 01:10:13.600 --> 01:10:16.037 something that, you know, you believe, I believe, and everyone in this room 01:10:16.120 --> 01:10:19.197 probably believes can do untold good. 01:10:19.280 --> 01:10:22.917 Yeah, this is this is one of the most important questions. And actually, 01:10:23.000 --> 01:10:26.397 perhaps even more so than the technology and the development of the technology, 01:10:26.480 --> 01:10:30.197 um maybe will govern how how 01:10:30.280 --> 01:10:35.117 uh successfully the technology diffuses um more so than anything. And 01:10:35.200 --> 01:10:38.837 um you know, we we think deeply about this 01:10:38.920 --> 01:10:41.637 cuz our vision for the technology is 01:10:41.720 --> 01:10:43.917 that it is um we can deploy personal 01:10:44.000 --> 01:10:48.237 agents that uh know you better than most 01:10:48.320 --> 01:10:51.997 people know you, know about your health, know about your goals, know about your 01:10:52.080 --> 01:10:56.037 relationships, know about your family and your friends, and um can really help 01:10:56.120 --> 01:11:00.517 you live your life more effectively, you know, can help you be healthier, can 01:11:00.600 --> 01:11:04.517 help you accomplish projects that you never would have had time for, can help 01:11:04.600 --> 01:11:05.597 you build the business you always 01:11:05.680 --> 01:11:08.117 wanted, um can help you explore your 01:11:08.200 --> 01:11:11.717 interests. And to properly build that 01:11:11.800 --> 01:11:14.717 requires an immense amount of trust from 01:11:14.800 --> 01:11:18.037 consumers, from governments, from 01:11:18.120 --> 01:11:20.357 um from effectively everyone, from all 01:11:20.440 --> 01:11:25.117 organizations. And so it you know, this is one of these interesting moments 01:11:25.200 --> 01:11:29.597 where um I think because the the sort of 01:11:29.680 --> 01:11:31.877 the form that the of the technology that 01:11:31.960 --> 01:11:34.837 we wanted to take is so intimate and is 01:11:34.920 --> 01:11:38.317 so deeply tied with what we think is 01:11:38.400 --> 01:11:40.717 really the the future of, you know, the 01:11:40.800 --> 01:11:45.517 consumer life and the consumer um you know, is at the core of the relationship 01:11:45.600 --> 01:11:48.037 with technology. Um 01:11:48.120 --> 01:11:51.197 The figuring out ways to develop the the 01:11:51.280 --> 01:11:55.317 technology in partnership with governing bodies and in partnership with um 01:11:55.400 --> 01:11:59.717 nonprofits and other organizations to garner that trust is very important. 01:11:59.800 --> 01:12:04.197 This is something we've seen at Meta actually in um with WhatsApp. You know, 01:12:04.280 --> 01:12:06.557 one of the we were talking about this actually 01:12:06.640 --> 01:12:09.637 backstage, but um WhatsApp has become 01:12:09.720 --> 01:12:12.277 uh incredibly ubiquitous um and a real I 01:12:12.360 --> 01:12:14.877 think utility for much of the world um 01:12:14.960 --> 01:12:20.397 in large part due to uh its commitment to privacy and the the sort of trust 01:12:20.480 --> 01:12:21.677 that we've built around WhatsApp as an 01:12:21.760 --> 01:12:24.557 organization. Um oftentimes at the sort 01:12:24.640 --> 01:12:27.797 of uh you know, not necessarily for the 01:12:27.880 --> 01:12:31.957 optimization of our business objectives, but for the optimization of trust in 01:12:32.040 --> 01:12:34.157 WhatsApp as a platform. I think a 01:12:34.240 --> 01:12:37.277 similar journey will be necessary for 01:12:37.360 --> 01:12:39.597 the AI products of the modern age and 01:12:39.680 --> 01:12:42.037 we've seen um I mean we've already seen 01:12:42.120 --> 01:12:44.877 various uh kerfuffles or sort of jabs 01:12:44.960 --> 01:12:49.157 back and forth between the existing AI, you know, products and labs around, you 01:12:49.240 --> 01:12:53.237 know, whether their actions are truly garnering, you know, truly building this 01:12:53.320 --> 01:12:57.277 uh these levels of trust. But and then I think the you know, the topic 01:12:57.360 --> 01:13:00.557 that you really kicked off at Bletchley 01:13:00.640 --> 01:13:02.637 around in a sort of like global way 01:13:02.720 --> 01:13:07.677 around safety. I think is more important than ever. I think we're already seeing 01:13:07.760 --> 01:13:11.677 that the models are 01:13:11.760 --> 01:13:15.637 uh are are achieving levels of capability 01:13:15.720 --> 01:13:18.197 which cause some pause around some of 01:13:18.280 --> 01:13:22.757 the sort of like risks that we monitor 01:13:22.840 --> 01:13:24.917 for these models as it comes to various 01:13:25.000 --> 01:13:27.677 national security risks or various cyber 01:13:27.760 --> 01:13:30.277 security risks. And so 01:13:30.360 --> 01:13:32.877 the the importance of properly testing 01:13:32.960 --> 01:13:37.797 and then building the right systems to monitor and deploy these models safely 01:13:37.880 --> 01:13:39.317 is is of paramount importance. 01:13:39.400 --> 01:13:42.637 >> you and I have been talking about the same topic for some years. And that was 01:13:42.720 --> 01:13:47.037 key to us at Bletchley was establishing the AI 01:13:47.120 --> 01:13:50.517 security institute in the UK to do exactly that work. And the And the 01:13:50.600 --> 01:13:54.517 thought was there was a gap where the technical capability didn't exist 01:13:54.600 --> 01:13:58.637 amongst governments to evaluate some of these risks. So how could we in the UK 01:13:58.720 --> 01:14:03.317 hopefully take a lead but in sense provide a global public good 01:14:03.400 --> 01:14:08.800 where we work in partnership with companies like yours and the other