1 00:00:00,320 --> 00:00:04,557 I'm honored to be here and to be able to speak with you all at this important time. 2 00:00:04,640 --> 00:00:07,677 We've been talking about artificial intelligence for years, 3 00:00:07,760 --> 00:00:10,557 but it feels different in recent weeks and months. 4 00:00:10,640 --> 00:00:13,797 Rapid model progress has made the timeline feel more compressed, the 5 00:00:13,880 --> 00:00:18,397 upside more tangible, but also the stakes and the risks more immediate. 6 00:00:18,480 --> 00:00:20,677 We have a choice in front of us. 7 00:00:20,760 --> 00:00:25,117 AI can either be more like a new renaissance of creativity and discovery, 8 00:00:25,200 --> 00:00:29,477 or more like a new industrial revolution of upheaval and disarray. 9 00:00:29,560 --> 00:00:33,757 The renaissance was a time when new tools, new ideas, and new institutions 10 00:00:33,840 --> 00:00:36,957 expanded what people believed they could do. 11 00:00:37,040 --> 00:00:39,477 AI can play that role for our time. 12 00:00:39,560 --> 00:00:42,637 I believe that people are much more capable than they imagine themselves to 13 00:00:42,720 --> 00:00:47,197 be, but have been held back by limited technology that was meant to empower, 14 00:00:47,280 --> 00:00:50,837 but instead distracts and often steals attention. 15 00:00:50,920 --> 00:00:54,917 The best version of AI is not about making people cogs in a giant machine, 16 00:00:55,000 --> 00:00:58,877 or optimizing every part of life until the human parts disappear. 17 00:00:58,960 --> 00:01:02,677 It's about giving people more agency in an ever more complex world, more ability 18 00:01:02,760 --> 00:01:07,157 to learn, to create, discover, build, participate. 19 00:01:07,240 --> 00:01:10,637 There are many things that AI cannot and should not automate. 20 00:01:10,720 --> 00:01:14,157 AI should be built for people, not simply to turn the crank crank of the 21 00:01:14,240 --> 00:01:17,517 machines faster for their own sake. 22 00:01:17,600 --> 00:01:20,597 Some of the things people working to build AI have said are so dystopian that 23 00:01:20,680 --> 00:01:23,997 they sound like the plot of bad science fiction movies. 24 00:01:24,080 --> 00:01:27,877 Anything this power is also powerful is also intimidating. 25 00:01:27,960 --> 00:01:29,517 This moment is complicated as we feel 26 00:01:29,600 --> 00:01:32,197 both tremendous and very understandable 27 00:01:32,280 --> 00:01:36,677 anxiety at the same time. These models can help people discover 28 00:01:36,760 --> 00:01:40,317 new knowledge, new medicines, better health care, stronger education, and 29 00:01:40,400 --> 00:01:43,717 build more productive economies. They can be new tools for people to pursue 30 00:01:43,800 --> 00:01:47,157 dreams and ambition that today feel out of reach. 31 00:01:47,240 --> 00:01:51,517 The complexity of life has increased so dramatically for the average person. 32 00:01:51,600 --> 00:01:55,037 AI can be a defensive tool against this complexity as much as it can be this 33 00:01:55,120 --> 00:01:59,437 tool of unprecedented discovery and innovation. 34 00:01:59,520 --> 00:02:01,797 On the other hand, as AI systems become 35 00:02:01,880 --> 00:02:06,557 more capable and more autonomous, they could move faster than our institutions, 36 00:02:06,640 --> 00:02:08,437 concentrate power in too few too few 37 00:02:08,520 --> 00:02:10,877 hands, or make decisions that people no 38 00:02:10,960 --> 00:02:15,917 longer understand or control. That concern becomes especially 39 00:02:16,000 --> 00:02:19,637 important as we approach systems that can improve themselves and future 40 00:02:19,720 --> 00:02:24,237 versions of themselves. Often called recursive self-improvement. 41 00:02:24,320 --> 00:02:28,397 As the process of building AI becomes more automated, the pace of AI progress 42 00:02:28,480 --> 00:02:30,237 could accelerate rapidly. 43 00:02:30,320 --> 00:02:33,477 This moment calls for extreme care. 44 00:02:33,560 --> 00:02:36,037 I largely agree with Professor Bengio. 45 00:02:36,120 --> 00:02:39,117 Uh although I'll say two instead of three. I believe there are two ways that 46 00:02:39,200 --> 00:02:42,837 AI progress could go very badly that we have to avoid. 47 00:02:42,920 --> 00:02:47,437 First, we could lose control of the future to AI. 48 00:02:47,520 --> 00:02:50,237 The risk is that it moves so fast that people can no longer follow what's 49 00:02:50,320 --> 00:02:54,797 happening or intervene when needed. This would obviously be terrible. 50 00:02:54,880 --> 00:02:58,037 The industry must not accept too much technological risk just because the 51 00:02:58,120 --> 00:03:01,637 benefits are too great and that they feel too important to slow down. 52 00:03:01,720 --> 00:03:05,437 Beating companies in competitive race is not a reason to make rash decisions, nor 53 00:03:05,520 --> 00:03:08,277 do we believe we are locked in a race where we are unable to do that. We have 54 00:03:08,360 --> 00:03:11,877 unilaterally slowed down in the past. We will do so in the future. We do not want 55 00:03:11,960 --> 00:03:14,397 to build systems. We think it's important that everyone avoids building 56 00:03:14,480 --> 00:03:18,157 systems where we cannot make the alignment, monitorability, and safety 57 00:03:18,240 --> 00:03:21,557 guarantees that we must. We need to understand what these systems 58 00:03:21,640 --> 00:03:25,437 are doing and have strong evidence that they will do what people intend even as 59 00:03:25,520 --> 00:03:30,397 they get very, very smart. Actually, especially as they get very, very smart. 60 00:03:30,480 --> 00:03:34,357 It doesn't matter whether people put the risk of catastrophe at 10% or 1% or 12% 61 00:03:34,440 --> 00:03:38,597 or 0.1% None of these levels are remotely acceptable and we should not 62 00:03:38,680 --> 00:03:41,917 train models that we cannot make an extremely strong case they will be able 63 00:03:42,000 --> 00:03:46,997 to keep under human control. So, that's one way things can go wrong. 64 00:03:47,080 --> 00:03:50,157 In the other direction, these systems could concentrate too much power in too 65 00:03:50,240 --> 00:03:54,077 few hands. No one person or company or country 66 00:03:54,160 --> 00:03:58,197 should be able to use the most powerful AI models to impose their worldview on 67 00:03:58,280 --> 00:04:01,637 everyone else. A company or country that believes only 68 00:04:01,720 --> 00:04:05,197 it can be trusted with this technology can use that belief to justify almost 69 00:04:05,280 --> 00:04:08,317 anything else. We have to reject that logic even when 70 00:04:08,400 --> 00:04:11,037 it's convenient for anyone actor. 71 00:04:11,120 --> 00:04:14,717 I believe there is a pragmatic path through both of these. 72 00:04:14,800 --> 00:04:18,397 At OpenAI, we approach these questions with a clear goal. 73 00:04:18,480 --> 00:04:22,157 AI should give people more power over their own lives, more freedom to live 74 00:04:22,240 --> 00:04:26,077 lives the way they want, more ability to solve problems big or small, and to 75 00:04:26,160 --> 00:04:30,477 improve their communities. We don't want to fall into the trap of blind optimism. 76 00:04:30,560 --> 00:04:33,557 We also don't want to fall into the trap of doomerism, either. 77 00:04:33,640 --> 00:04:36,997 Getting pulled into either extreme will not help us successfully figure out the 78 00:04:37,080 --> 00:04:40,157 challenges in front of us, especially at a time when the world and 79 00:04:40,240 --> 00:04:45,157 its people face such enormous challenges that this technology can help solve. 80 00:04:45,240 --> 00:04:48,717 So, we try to walk the middle path between these two poles. 81 00:04:48,800 --> 00:04:50,957 For us, that means three principles. 82 00:04:51,040 --> 00:04:56,077 First, as AI becomes more capable, people must remain at the center of AI 83 00:04:56,160 --> 00:04:59,277 decision-making. Alignment is not an abstract research 84 00:04:59,360 --> 00:05:03,157 question. It is the work of ensuring that these systems reliably remain under 85 00:05:03,240 --> 00:05:06,557 human control, reflect human values, and help people 86 00:05:06,640 --> 00:05:08,957 guide the next stages of development. 87 00:05:09,040 --> 00:05:13,517 We are not trying to and must not automate human judgment or human values. 88 00:05:13,600 --> 00:05:17,677 Second, the benefits of scientific progress and economic growth must be by 89 00:05:17,760 --> 00:05:20,557 people and for people. AI should help researchers make 90 00:05:20,640 --> 00:05:24,077 discoveries, treat patients, teachers teach students, entrepreneurs start 91 00:05:24,160 --> 00:05:28,077 companies, communities solve problems that felt out of reach, and much more. 92 00:05:28,160 --> 00:05:31,677 The continuous story of human progress is not the tools 93 00:05:31,760 --> 00:05:35,317 replace human agency, it is that the right tools make people more capable, 94 00:05:35,400 --> 00:05:39,157 and I believe this is going to go much further than we believe possible. 95 00:05:39,240 --> 00:05:42,957 Third, this technology must empower people individually. 96 00:05:43,040 --> 00:05:46,237 A great future will come from people realizing their potential and building 97 00:05:46,320 --> 00:05:50,597 value for themselves, their communities, their countries, and the broader world. 98 00:05:50,680 --> 00:05:53,637 We might be with the builders, one of the builders of this technology, but we 99 00:05:53,720 --> 00:05:55,397 are not the heroes of this story. 100 00:05:55,480 --> 00:05:56,997 Our role is to enable people all over 101 00:05:57,080 --> 00:06:02,877 the world and to trust in the magic of humanity's skill and faculties. 102 00:06:02,960 --> 00:06:07,117 We now have models capable of discovering knowledge that we have never had before. 103 00:06:07,200 --> 00:06:09,677 Consider the progress we've seen in math. 104 00:06:09,760 --> 00:06:13,397 Three summers ago, our models were okay at grade school math. 105 00:06:13,480 --> 00:06:17,237 Two summers ago, they were pretty good at high school math competitions. 106 00:06:17,320 --> 00:06:20,837 Last summer, we reached gold level a gold level medal in the most prestigious 107 00:06:20,920 --> 00:06:22,997 international math competition. 108 00:06:23,080 --> 00:06:26,157 And just a few weeks ago, this summer, one of our models solved one of the 109 00:06:26,240 --> 00:06:29,917 millennium prize problems, the Navier-Stokes equations. 110 00:06:30,000 --> 00:06:33,157 These equations are used for aircraft design, weather forecasting, and the 111 00:06:33,240 --> 00:06:36,037 study of blood flow. We are not solving math problems for 112 00:06:36,120 --> 00:06:39,957 their own sake, but to empower people to discover new knowledge and to use it to 113 00:06:40,040 --> 00:06:44,077 improve health care, raise standards of livings of living, and expand everyone's 114 00:06:44,160 --> 00:06:48,557 potential around the world. There remains so much to be discovered. 115 00:06:48,640 --> 00:06:52,757 But it's very important that companies not substitute for the democratic process. 116 00:06:52,840 --> 00:06:56,397 If AI is to be democratic, the most important decisions cannot be made by 117 00:06:56,480 --> 00:07:01,077 labs in San Francisco alone. They must be shaped through democratic processes 118 00:07:01,160 --> 00:07:04,597 and by governments accountable to the people that they serve. 119 00:07:04,680 --> 00:07:08,517 At the international level, this will require cooperation. 120 00:07:08,600 --> 00:07:11,637 In our history, there have been times where countries who compete and don't 121 00:07:11,720 --> 00:07:15,277 always like each other very much still come together for shared interests and 122 00:07:15,360 --> 00:07:18,717 the collective good in the face of a powerful new technology. 123 00:07:18,800 --> 00:07:22,277 We believe this must be one of those times. 124 00:07:22,360 --> 00:07:25,637 We need a mechanism for complementary national and international frontier AI 125 00:07:25,720 --> 00:07:30,517 standards. Standards for measuring capabilities, assessing risks, 126 00:07:30,600 --> 00:07:33,877 determining whether safeguards are sufficient, and preserving meaningful 127 00:07:33,960 --> 00:07:37,597 human oversight as systems become more autonomous. 128 00:07:37,680 --> 00:07:40,597 We need common standards so countries can compare evidence, 129 00:07:40,680 --> 00:07:44,397 verify compliance, and have a shared language and understanding about what is 130 00:07:44,480 --> 00:07:48,477 happening. We need to accurate and speedy incident reporting, 131 00:07:48,560 --> 00:07:52,237 classification and reporting protocols, so the world can learn from failures 132 00:07:52,320 --> 00:07:54,917 before they become catastrophes. 133 00:07:55,000 --> 00:07:57,997 And we need secure channels among governments, critical infrastructure 134 00:07:58,080 --> 00:08:01,677 operators, and technical experts to share emerging vulnerabilities and new 135 00:08:01,760 --> 00:08:05,077 threats. Obviously, these standards should not 136 00:08:05,160 --> 00:08:08,917 lock in incumbents or favor one business model over another. 137 00:08:09,000 --> 00:08:12,397 They must support open and closed model developers, new entrants and established 138 00:08:12,480 --> 00:08:16,397 labs, national authority, and international cooperation. 139 00:08:16,480 --> 00:08:19,317 Each government should decide how to incorporate standards into its own legal 140 00:08:19,400 --> 00:08:23,517 system, but we will all be better off if we can agree on what good evidence, good 141 00:08:23,600 --> 00:08:27,957 safeguards, and good oversight look like on the global stage. 142 00:08:28,040 --> 00:08:31,797 We do not have all the answers, and we're very open to any better ideas. 143 00:08:31,880 --> 00:08:34,637 We're ready to work with you all and try our hardest to figure out the path 144 00:08:34,720 --> 00:08:38,637 forward. We are at a crossroads. 145 00:08:38,720 --> 00:08:42,997 One path leads to a world where power is concentrated, decisions are opaque, and 146 00:08:43,080 --> 00:08:45,797 people feel that the most important technology of their lifetime is 147 00:08:45,880 --> 00:08:50,157 something being done to them. The other path leads to a world where AI 148 00:08:50,240 --> 00:08:54,517 works for people, because people, through democratic institutions, help 149 00:08:54,600 --> 00:08:58,397 decide how it should develop and guide it. Where scientific progress improves 150 00:08:58,480 --> 00:09:01,917 real lives, where more people can participate in the future, and where 151 00:09:02,000 --> 00:09:05,357 this technology continues to give individuals and communities greater 152 00:09:05,440 --> 00:09:07,557 power to live life as they choose. 153 00:09:07,640 --> 00:09:11,117 This is a future of realized potential, excitement, and wonder, and it is the 154 00:09:11,200 --> 00:09:12,517 future we believe is worth building. 155 00:09:12,600 --> 00:09:15,400 Thank you very much.