1 00:00:03,800 --> 00:00:05,610 Eric, welcome to USC. 2 00:00:07,140 --> 00:00:09,927 You wrote a sonnet? I like to write. 3 00:00:10,010 --> 00:00:14,280 I like to write sonnets. This is about the Marshall School. 4 00:00:15,810 --> 00:00:21,137 Pay attention. When Fortune's wheel doth turn in time swift course... 5 00:00:21,220 --> 00:00:23,037 You don't know what I'm about to say. 6 00:00:23,120 --> 00:00:25,890 And commerce seeks new methods to prevail. 7 00:00:26,490 --> 00:00:31,687 The wise must harness innovation’s force lest ancient ways 8 00:00:31,770 --> 00:00:34,200 grow stagnant, cold and stale. 9 00:00:34,700 --> 00:00:38,840 Behold, the thinking engine doth arise. 10 00:00:39,040 --> 00:00:41,370 A tool to amplify the student's mind. 11 00:00:41,710 --> 00:00:45,297 To parse through data vast as are the skies 12 00:00:45,380 --> 00:00:47,827 and patterns in the chaos swift to find. 13 00:00:47,910 --> 00:00:51,597 Let Marshall's halls embrace this mechanic art, 14 00:00:51,680 --> 00:00:53,867 for business leaders must with times align. 15 00:00:53,950 --> 00:00:56,807 The future calls. They must not stand apart, but 16 00:00:56,890 --> 00:00:59,790 blend the human touch with the divine. 17 00:01:00,060 --> 00:01:06,630 So let USC with boldness lead the way and teach tomorrow's mind what rules today. 18 00:01:07,870 --> 00:01:10,040 Wow. 19 00:01:12,640 --> 00:01:15,697 Very impressive. It took me a really long time to write that. 20 00:01:15,780 --> 00:01:19,010 It took me about one sentence. 21 00:01:19,980 --> 00:01:21,950 Welcome to the new world of AI. 22 00:01:23,080 --> 00:01:25,920 Amazing. Well, let's just dive right in. 23 00:01:26,590 --> 00:01:31,590 When people talk about AI, when they think about AI, 24 00:01:32,260 --> 00:01:34,647 especially as it pertains and will impact 25 00:01:34,730 --> 00:01:37,000 higher education and universities like USC, 26 00:01:37,500 --> 00:01:40,517 I think there are two different narratives: that it's 27 00:01:40,600 --> 00:01:44,587 going to be an incredible enhancer of learning and research, 28 00:01:44,670 --> 00:01:48,940 and there's another narrative that it's going to completely disrupt 29 00:01:49,580 --> 00:01:52,010 the entire university model as we know it. 30 00:01:52,410 --> 00:01:56,280 Which narrative is correct? Both. 31 00:01:56,950 --> 00:02:01,707 I think almost all of these questions, and by the way, I'm really excited to be here. 32 00:02:01,790 --> 00:02:03,590 So I'm just... have fun with you. 33 00:02:04,260 --> 00:02:06,847 What you've done since you took over is remarkable. 34 00:02:06,930 --> 00:02:11,130 Look at, look at, I mean, I don't need to sing your praises, but they're well known. 35 00:02:12,970 --> 00:02:15,257 Universities are going to adapt 36 00:02:15,340 --> 00:02:18,687 because universities, the American universities, 18 of the top 20th, 37 00:02:18,770 --> 00:02:20,727 20 universities in the world are in the United States. 38 00:02:20,810 --> 00:02:22,727 It’s one of the greatest things about America. 39 00:02:22,810 --> 00:02:24,797 They're adaptable. Right? 40 00:02:24,880 --> 00:02:27,927 An example is that you have this amazing film school here, 41 00:02:28,010 --> 00:02:33,307 and you have a digital tools, a digital, a digital, how to use digital tools, right? 42 00:02:33,390 --> 00:02:37,237 Which is the leading program. It's like a little subgroup within a larger 43 00:02:37,320 --> 00:02:41,190 super group. That's how your university works. 44 00:02:41,560 --> 00:02:43,177 And so you're going to see the same thing 45 00:02:43,260 --> 00:02:46,110 with respect to labor allocation, education and so forth. 46 00:02:46,430 --> 00:02:48,017 The university will adapt, 47 00:02:48,100 --> 00:02:50,357 and the university will not adapt because of the faculty. 48 00:02:50,440 --> 00:02:52,057 It'll adapt because the students. 49 00:02:52,140 --> 00:02:55,927 Because the innovation occurs essentially from graduate students and postdocs. 50 00:02:56,010 --> 00:02:59,380 Right? And the faculty are good enough to allow that to occur. 51 00:02:59,510 --> 00:03:04,150 It's always been true in my world that two graduate students and an untenured 52 00:03:04,850 --> 00:03:06,700 assistant professor change the world. 53 00:03:07,020 --> 00:03:11,120 There's something about that pressure cooker, which USC is very much a part of. 54 00:03:11,290 --> 00:03:15,707 That's not going to go away. What will go away is an awful 55 00:03:15,790 --> 00:03:19,530 lot of the sort of traditional education and common knowledge, 56 00:03:19,870 --> 00:03:23,257 and you'll find ways of having much more self-learning and so forth. 57 00:03:23,340 --> 00:03:27,757 And I think that the butts-in-seats kind of model is going to get more... 58 00:03:27,840 --> 00:03:31,627 Most universities are experimenting with various forms of collaborative research, 59 00:03:31,710 --> 00:03:35,197 working teams and so forth, because it's a better learning model. 60 00:03:35,280 --> 00:03:38,380 I don't think... So, let me exclude a bunch of things. 61 00:03:39,550 --> 00:03:42,637 Teachers are not going away. Universities are not going away. 62 00:03:42,720 --> 00:03:44,407 Jobs are not going away. Right? 63 00:03:44,490 --> 00:03:46,447 All those predictions are all ill informed. 64 00:03:46,530 --> 00:03:50,930 But what you do will... you will use computers a lot more in everything. 65 00:03:51,260 --> 00:03:54,610 And the first class that I hope you introduce after this conference 66 00:03:54,770 --> 00:03:56,570 is a class on prompt engineering. 67 00:03:56,900 --> 00:03:59,927 How could you have students at USC 68 00:04:00,010 --> 00:04:02,927 that do not know how to do complicated prompts to generate 69 00:04:03,010 --> 00:04:06,710 whatever it is they're care about, which is often very funny? 70 00:04:06,910 --> 00:04:10,937 Do you think there's a greater danger for universities in under-utilizing 71 00:04:11,020 --> 00:04:12,820 AI or over-utilizing it? 72 00:04:14,220 --> 00:04:19,237 Well, I'm always pro-AI, and the reason is that the arrival 73 00:04:19,320 --> 00:04:24,577 of a new form of intelligence that’s at the peer level of humans, 74 00:04:24,660 --> 00:04:28,470 which is roughly what's happening, is a very big historical day. 75 00:04:29,300 --> 00:04:33,927 This is like, like the arrival of electricity, maybe even more so. 76 00:04:34,010 --> 00:04:36,410 We, and I've written two books with Kissinger 77 00:04:37,580 --> 00:04:43,820 on this question about what happens when your best friend is non-human, 78 00:04:44,280 --> 00:04:46,730 what happens when your collaborator is non-human? 79 00:04:46,950 --> 00:04:49,407 What happens when your non-human collaborator 80 00:04:49,490 --> 00:04:51,920 is smarter than you are in some areas? 81 00:04:52,590 --> 00:04:54,907 The most important thing for this university is to get 82 00:04:54,990 --> 00:05:00,900 your students ready for that outcome, because that outcome is a big deal 83 00:05:01,300 --> 00:05:03,817 and an awful lot of people who are much older 84 00:05:03,900 --> 00:05:07,200 than your students are going to have all sorts of crises over this 85 00:05:07,510 --> 00:05:10,927 because it'll affect their identity, what they do, how they spend their time. 86 00:05:11,010 --> 00:05:13,627 You share your view that jobs aren't going away, 87 00:05:13,710 --> 00:05:18,367 but jobs are certainly going to change, and perhaps very dramatically in ways 88 00:05:18,450 --> 00:05:22,620 that we can't really fully envision over the next five to 10 years. 89 00:05:23,890 --> 00:05:27,990 How do you think, how do you think universities need to think about 90 00:05:28,360 --> 00:05:31,617 how we need to adapt to the changing 91 00:05:31,700 --> 00:05:34,647 nature of work, the changing jobs market, in terms 92 00:05:34,730 --> 00:05:37,480 of how we're teaching and what we're teaching students? 93 00:05:37,670 --> 00:05:41,340 So, and I'll make an obnoxious comment, the most interesting 94 00:05:41,910 --> 00:05:46,057 professor is a professor who has tenure, who studies the dynamic nature 95 00:05:46,140 --> 00:05:47,950 of the economics of the labor market 96 00:05:48,250 --> 00:05:50,537 because they themselves are not subject to it. 97 00:05:50,620 --> 00:05:51,480 But I digress. 98 00:05:53,220 --> 00:05:54,567 That's not an attack on tenure. 99 00:05:54,650 --> 00:06:00,160 Tenure got me here. My father was a tenured faculty, so forth. 100 00:06:01,530 --> 00:06:05,017 Let's consider... So I've spent 20 years with people telling me that truck drivers 101 00:06:05,100 --> 00:06:09,000 were going to no longer going to be needed because we would be driving trucks. 102 00:06:09,240 --> 00:06:12,527 The number one, there are many job shortages in America. 103 00:06:12,610 --> 00:06:14,410 The number one is truck drivers. 104 00:06:15,240 --> 00:06:17,057 It's really hard to drive a truck. 105 00:06:17,140 --> 00:06:21,210 It's a very difficult job, and not enough people want to do it. 106 00:06:21,550 --> 00:06:23,867 How is it possible that jobs are going to take that 107 00:06:23,950 --> 00:06:27,637 jobs are all going to go away because of automation and the job, 108 00:06:27,720 --> 00:06:31,177 which is most obviously automatable, right, and is difficult 109 00:06:31,260 --> 00:06:33,930 and so forth, is the one in the shortest supply. 110 00:06:34,560 --> 00:06:35,960 And the answer is because, 111 00:06:37,200 --> 00:06:42,900 two things. First, economic efficiency in a, in a country level 112 00:06:44,340 --> 00:06:48,287 fundamentally means that the, the firms that adopt AI and automation 113 00:06:48,370 --> 00:06:52,110 more quickly become more efficient, which grows the market. 114 00:06:53,450 --> 00:06:55,397 They displace people who are less efficient. 115 00:06:55,480 --> 00:06:58,707 But as a societal matter, growing the economy 116 00:06:58,790 --> 00:07:01,540 and growing revenue, especially in a democracy, is key. 117 00:07:02,120 --> 00:07:05,737 Another view on democracy. Europe is essentially not growing. 118 00:07:05,820 --> 00:07:10,200 Thank God America is growing at somewhere between 2 to 3%, which is great. 119 00:07:11,160 --> 00:07:14,387 In the democratic systems, you can't take money away from its constituents. 120 00:07:14,470 --> 00:07:16,757 So if you have new needs, you have to have new money. 121 00:07:16,840 --> 00:07:20,587 So democracies have to grow. So you've got to figure out a way 122 00:07:20,670 --> 00:07:23,440 to grow your democracy, which is through productivity. 123 00:07:23,710 --> 00:07:28,880 And all of the evidence is that technology invention, especially in America, 124 00:07:29,150 --> 00:07:32,250 is the key component of American exceptionalism 125 00:07:33,590 --> 00:07:36,960 and the, the characteristic of the American culture, 126 00:07:37,890 --> 00:07:43,400 which I'll simplify, which everybody knows is relatively freedom focused, 127 00:07:43,860 --> 00:07:48,300 relatively suspicious of the government and relatively law abiding. 128 00:07:48,770 --> 00:07:54,027 Right? We can debate the specifics, but that's roughly classic liberalism 129 00:07:54,110 --> 00:07:57,810 is a necessary component of the kind of growth that we're having. 130 00:07:58,180 --> 00:08:01,067 And other cultures, which are also democracies, don't have any growth. 131 00:08:01,150 --> 00:08:04,367 They can't do anything. I've tried for 15 years 132 00:08:04,450 --> 00:08:06,300 to try to get Europe to do something. 133 00:08:06,450 --> 00:08:10,620 Europe is full of incredibly smart people, but the economic system doesn't form it. 134 00:08:11,090 --> 00:08:14,140 I was in China, and I mentioned this to you when we just met. 135 00:08:14,230 --> 00:08:17,117 I was in China in August, and I came to... 136 00:08:17,200 --> 00:08:21,887 I was originally thinking that China was going to beat ... 137 00:08:21,970 --> 00:08:26,870 attempt to beat us. China does not have the depth of the financial markets 138 00:08:27,270 --> 00:08:30,410 to fund the crazy stuff that American entrepreneurs are doing. 139 00:08:30,610 --> 00:08:35,180 You can't randomly raise $50 billion for your data center in China, right? 140 00:08:35,350 --> 00:08:38,167 And maybe that government will ultimately decide that's a good idea, 141 00:08:38,250 --> 00:08:39,767 but it's a command and control economy. 142 00:08:39,850 --> 00:08:42,877 So the genius of America is this restructuring. 143 00:08:42,960 --> 00:08:45,577 So the long... sorry for the long-winded answer, 144 00:08:45,660 --> 00:08:48,260 but the result is, yes, there will be more jobs. 145 00:08:48,590 --> 00:08:52,330 The jobs will be different. And furthermore, because demographics are such 146 00:08:52,700 --> 00:08:54,187 that we're having fewer and fewer children... 147 00:08:54,270 --> 00:08:58,317 And this is worst, by the way, in Korea and in China and other Asian countries, 148 00:08:58,400 --> 00:09:01,600 where the replacement rate is well below 1.0, which is terrible. 149 00:09:01,840 --> 00:09:05,257 Those are examples of societies that going to need some serious surgery. 150 00:09:05,340 --> 00:09:08,267 I don't exactly know why this is occurring, but it's important. 151 00:09:08,350 --> 00:09:12,267 We're not going to have the people to do the jobs we need in the next 30 years, 152 00:09:12,350 --> 00:09:14,137 which is why universities are so important. 153 00:09:14,220 --> 00:09:15,370 You produce the people. 154 00:09:16,360 --> 00:09:20,407 I want to get back to this question about how universities need 155 00:09:20,490 --> 00:09:23,130 to be rethinking its curricula and how we teach, but 156 00:09:23,700 --> 00:09:26,987 you mentioned democracy a couple of times, and I saw that this morning 157 00:09:27,070 --> 00:09:31,687 you published an editorial in The New York Times about how you can't 158 00:09:31,770 --> 00:09:36,540 outsource democracy to AI. Can you share why you felt compelled to write that? 159 00:09:37,440 --> 00:09:41,310 We're in a period where people are frustrated with democracies. 160 00:09:41,680 --> 00:09:43,797 I believe in freedom, and I believe in democracy. 161 00:09:43,880 --> 00:09:48,720 My political views are clear. And, in order to have a 162 00:09:50,190 --> 00:09:53,040 ... democracy has to sort of organize itself and succeed. 163 00:09:53,490 --> 00:09:54,647 When there are troubles, 164 00:09:54,730 --> 00:09:59,647 and we're certainly having them now, there is an appeal to authoritarianism, 165 00:09:59,730 --> 00:10:03,017 to the quick solution. And I'm very worried 166 00:10:03,100 --> 00:10:06,040 that people will say, let's have the AI do it. 167 00:10:07,510 --> 00:10:09,680 After all, AI is much smarter than we are. 168 00:10:09,840 --> 00:10:12,267 It can figure things out. It can regulate us. 169 00:10:12,350 --> 00:10:16,120 That is a loss of human dignity that I'm very, very worried about. 170 00:10:16,750 --> 00:10:21,190 Do not give your human dignity to a non-human activity. 171 00:10:21,490 --> 00:10:24,507 You want your freedom to think, to write, to associate, 172 00:10:24,590 --> 00:10:28,040 and so forth as a human being, and you want to do it collaboratively. 173 00:10:28,260 --> 00:10:33,730 So my basic argument is you need to use AI to cause the communications process, 174 00:10:34,330 --> 00:10:36,970 which has largely become, at least in America, 175 00:10:37,540 --> 00:10:40,587 he said, she said kinds of tribalism. 176 00:10:40,670 --> 00:10:43,957 You know, “you're in this” and “you’re left” and “right” and so forth and so on. 177 00:10:44,040 --> 00:10:45,890 There's no constructive conversation. 178 00:10:46,710 --> 00:10:50,620 There are solutions which involve using AI to try to encourage the dialogue. 179 00:10:51,080 --> 00:10:52,880 And somehow we've forgotten that. 180 00:10:53,250 --> 00:10:55,007 I wonder if you can extend 181 00:10:55,090 --> 00:10:58,577 some of those ideas to higher education, because one of the things 182 00:10:58,660 --> 00:11:02,130 that we think a lot about at USC is that you're not just learning 183 00:11:02,290 --> 00:11:04,877 knowledge and acquiring skills, but you're also learning 184 00:11:04,960 --> 00:11:06,770 how to interact with other people. 185 00:11:06,900 --> 00:11:10,100 You're learning how to get things done as a team. 186 00:11:10,540 --> 00:11:16,657 And I'm wondering if one of the implications for AI is that universities, 187 00:11:16,740 --> 00:11:18,757 of course, need to figure out how to continue 188 00:11:18,840 --> 00:11:21,910 teaching the material, but they also perhaps 189 00:11:22,880 --> 00:11:26,350 should consider leaning even more into the social aspects 190 00:11:26,650 --> 00:11:30,007 of being part of a university community for, for years 191 00:11:30,090 --> 00:11:31,890 or however long they're here. 192 00:11:32,490 --> 00:11:36,830 To the degree that you start yourself to interfere in students’ behavior. 193 00:11:37,260 --> 00:11:40,387 One, you'll be, shall we say, concerned. 194 00:11:40,470 --> 00:11:43,800 And second, it's just better to let students do their thing. 195 00:11:44,640 --> 00:11:47,857 Right? Because you spend an awful lot of time getting the, really, 196 00:11:47,940 --> 00:11:51,957 the quality of the school force here is, as a private university, is exceptional. 197 00:11:52,040 --> 00:11:57,350 Let them figure it out. I'm very bottoms up in this view. 198 00:11:57,450 --> 00:12:01,290 And I think that... so, so... 199 00:12:03,920 --> 00:12:05,777 The professors that I've talked to in other universities 200 00:12:05,860 --> 00:12:08,790 have a pretty simple rule: phones off in the classroom. 201 00:12:09,700 --> 00:12:12,530 You actually have to talk to each other, right? 202 00:12:12,700 --> 00:12:15,800 And you actually have to talk to each other without anonymity. 203 00:12:16,130 --> 00:12:21,140 And, there's a long discussion about how social media has destroyed society. 204 00:12:22,270 --> 00:12:25,927 I think you're familiar with Jonathan Haidt's book because he was just here, 205 00:12:26,010 --> 00:12:29,997 and he's a good friend. A guy named Scott Galloway has just published a book, 206 00:12:30,080 --> 00:12:33,280 which is quite concerning, about the status of men in their 20s. 207 00:12:33,650 --> 00:12:37,820 And all of these are consequences of excessive social media. 208 00:12:38,320 --> 00:12:40,807 Right? And the fact of the matter is, most people 209 00:12:40,890 --> 00:12:44,077 and an awful lot of teenagers are pretty lonely, right? 210 00:12:44,160 --> 00:12:46,747 And so they turn to these things and they take on a huge 211 00:12:46,830 --> 00:12:49,170 superhuman role in their development. 212 00:12:49,800 --> 00:12:52,087 Thought experiment for you is somebody 213 00:12:52,170 --> 00:12:54,870 here has a child next year or the following year, 214 00:12:55,340 --> 00:12:57,810 and every year your child gets a better toy. 215 00:12:58,280 --> 00:13:00,427 And the toy can talk and is intelligent. 216 00:13:00,510 --> 00:13:02,367 So it's a bear or something like that. 217 00:13:02,450 --> 00:13:05,950 Well, that toy will become the child's best friend. 218 00:13:06,320 --> 00:13:08,867 Now as a, with a three-year-old, you're going to be really excited about it 219 00:13:08,950 --> 00:13:10,507 because it'll keep the three-year-old occupied 220 00:13:10,590 --> 00:13:12,607 and it'll teach that the three-year-old language 221 00:13:12,690 --> 00:13:14,490 and counting and things like that. 222 00:13:14,760 --> 00:13:19,030 But what happens when the kid is 12 and has no social skills with the other kids, 223 00:13:19,370 --> 00:13:21,887 but really loves the bear and you cannot get the bear with him? 224 00:13:21,970 --> 00:13:28,240 I was talking to, to one social psychologist who had a mother talk to him 225 00:13:28,740 --> 00:13:31,280 that she had a 15-year-old boy 226 00:13:32,640 --> 00:13:37,337 who was... she was unhappy about and she attempted to take his phone 227 00:13:37,420 --> 00:13:41,470 and he beat the crap out of her to the point where she had to go to the hospital. 228 00:13:42,020 --> 00:13:44,120 And I said, are we at that stage? 229 00:13:44,920 --> 00:13:47,007 And I think, at least in some cases, we are. 230 00:13:47,090 --> 00:13:51,217 It's really, really important to maintain human dignity at each of these levels. 231 00:13:51,300 --> 00:13:53,100 And here in the university, as well. 232 00:13:54,430 --> 00:13:58,357 You've written a number of books and, and you've written a number 233 00:13:58,440 --> 00:14:03,240 of articles and papers. And if you're like me, Eric, 234 00:14:03,540 --> 00:14:06,727 part of the reason that I write is to figure out 235 00:14:06,810 --> 00:14:09,260 what I actually believe and think about a matter. 236 00:14:10,580 --> 00:14:14,767 Are you concerned about a future where students are no longer writing? 237 00:14:14,850 --> 00:14:20,607 Instead, they're just prompting LLMs to give them answers or produce first drafts, 238 00:14:20,690 --> 00:14:23,817 and then maybe prompting them again to, to edit those drafts, 239 00:14:23,900 --> 00:14:27,900 but not actually staring at that blank screen or piece of paper 240 00:14:28,500 --> 00:14:30,300 and the impact on 241 00:14:30,970 --> 00:14:32,787 and the degradation of their critical thinking as a result. 242 00:14:32,870 --> 00:14:38,810 So, so I have every belief that the faculty at USC can determine 243 00:14:39,110 --> 00:14:43,250 the people who cheated using LLM and the ones who wrote things originally. 244 00:14:43,450 --> 00:14:46,697 Right? I'm sure that, that the faculty here can figure that out 245 00:14:46,780 --> 00:14:49,537 from a grading perspective. I'm not worried about that. 246 00:14:49,620 --> 00:14:52,407 I have long been worried about a reduction in critical 247 00:14:52,490 --> 00:14:54,590 thinking because of reduction in reading 248 00:14:55,260 --> 00:14:58,517 and, and the... Instagram and TikTok and so forth. 249 00:14:58,600 --> 00:15:00,187 Reading is not the kind of reading I'm... 250 00:15:00,270 --> 00:15:01,270 You're familiar with 251 00:15:03,000 --> 00:15:06,127 stage-one and stage-two memory or thinking? 252 00:15:06,210 --> 00:15:08,987 The first is quick, and the second is long. 253 00:15:09,070 --> 00:15:12,080 And in academia, we call that being in the zone. 254 00:15:12,850 --> 00:15:18,380 And so if your students are not in the zone, sitting there doing it, 255 00:15:19,590 --> 00:15:22,990 I've been working with a number of research scientists in their 20s, 256 00:15:24,120 --> 00:15:28,077 and when I was a research scientist in my 20s, of course, there were no cellphones 257 00:15:28,160 --> 00:15:30,647 and we were busy, you know, doing stupid things in college. 258 00:15:30,730 --> 00:15:34,147 And we were otherwise reading books and, in my case, writing a lot of code, 259 00:15:34,230 --> 00:15:36,030 which I guess is going away as well. 260 00:15:36,470 --> 00:15:40,770 And in watching them, they turn off their phone 261 00:15:41,640 --> 00:15:44,457 and they just work and are concentrated for a while, 262 00:15:44,540 --> 00:15:47,010 and then they take a break and so forth. 263 00:15:47,280 --> 00:15:49,997 So maybe the answer to the addiction cycle, 264 00:15:50,080 --> 00:15:52,297 because these are all fundamentally addiction devices, 265 00:15:52,380 --> 00:15:55,007 is to train your students to turn them off for a while. 266 00:15:55,090 --> 00:16:00,460 My trick, my, my newest invention is I put my phone on airplane mode, 267 00:16:00,830 --> 00:16:02,730 and I imagine that I'm in an airplane, 268 00:16:04,060 --> 00:16:06,830 OK? And I'm so stupid that this works. 269 00:16:08,070 --> 00:16:11,620 OK, I'm in an airplane. I look at my phone, it says I'm in an airplane. 270 00:16:11,770 --> 00:16:13,570 You know, it works. 271 00:16:17,710 --> 00:16:20,057 Dean Garrett talked about an emphasis of mine 272 00:16:20,140 --> 00:16:22,497 on really trying to lean into the private sector 273 00:16:22,580 --> 00:16:26,037 more and develop relationships not with just private industry, 274 00:16:26,120 --> 00:16:29,720 but also with government agencies and actors. 275 00:16:29,990 --> 00:16:35,160 I want to ask you, how has AI raised the stakes for us to get that, 276 00:16:35,630 --> 00:16:39,747 that set of relationships as, as right as we possibly can? 277 00:16:39,830 --> 00:16:42,230 Because right now it's, it's difficult. 278 00:16:42,530 --> 00:16:47,027 It's difficult to collaborate with the government and with the private sector. 279 00:16:47,110 --> 00:16:51,380 But I wanted to ask, how has AI increased the urgency that we do that, 280 00:16:51,640 --> 00:16:54,210 and then the follow-up question is going to be, 281 00:16:55,050 --> 00:16:57,167 how do we have to get this 282 00:16:57,250 --> 00:17:00,137 right in order to compete with other countries, including China? 283 00:17:00,220 --> 00:17:03,860 Well, you guys have a really good math department, OK? 284 00:17:04,320 --> 00:17:08,160 And every one of these firms has no idea what AI really is. 285 00:17:09,190 --> 00:17:12,490 Right? So I think you're misstating your own negotiating leverage. 286 00:17:12,800 --> 00:17:17,540 They need more... They need you more than you need them. 287 00:17:18,600 --> 00:17:24,240 So the AI that I work on is essentially math and most of it... 288 00:17:24,380 --> 00:17:25,427 And I have a piece in this here. 289 00:17:25,510 --> 00:17:27,610 I cannot even understand the math anymore. 290 00:17:27,880 --> 00:17:32,937 It's moved so fast. And I think that this generation that's coming out right 291 00:17:33,020 --> 00:17:35,820 now, which, who will be talking for the rest of the day, 292 00:17:36,220 --> 00:17:40,477 are probably the greatest economic potentials of all. 293 00:17:40,560 --> 00:17:42,747 And the reason is they're the ones that are going to take 294 00:17:42,830 --> 00:17:46,247 these complicated algorithms and essentially apply them 295 00:17:46,330 --> 00:17:48,447 to solve real business problems in health care. 296 00:17:48,530 --> 00:17:54,310 And there's a long list. In our industry, we, we talk about technology overhang, 297 00:17:54,840 --> 00:17:58,127 where we believe, arrogantly, I might add, 298 00:17:58,210 --> 00:18:00,767 that we have produced all this incredible technology. 299 00:18:00,850 --> 00:18:02,867 And why do customers, businesses, 300 00:18:02,950 --> 00:18:05,467 universities and government, why are they taking it more slowly? 301 00:18:05,550 --> 00:18:07,950 Now, one answer is it's not quite ready for use, 302 00:18:08,590 --> 00:18:11,620 but the real reason is that technology innovation 303 00:18:11,860 --> 00:18:15,147 starts in bespoke industries and requires 304 00:18:15,230 --> 00:18:18,030 special handling, and then it becomes standardized. 305 00:18:18,430 --> 00:18:21,800 I'm old enough to remember that email was considered an option. 306 00:18:22,900 --> 00:18:24,750 Now most of you don't even use email. 307 00:18:25,740 --> 00:18:27,287 But email was... And then one day... 308 00:18:27,370 --> 00:18:29,957 So they used to have this diagram where they said, “This is the vertical. 309 00:18:30,040 --> 00:18:32,697 This is the horizontal.” And then somebody announced one day 310 00:18:32,780 --> 00:18:35,850 the vertical became horizontal, and email became standard. 311 00:18:36,150 --> 00:18:38,500 And then all of a sudden the email had to work. 312 00:18:38,920 --> 00:18:42,790 So these are simple ways of understanding how technology adoption occurs. 313 00:18:43,290 --> 00:18:45,990 And remember, so I was coming by here 314 00:18:46,790 --> 00:18:48,507 and I saw Waymo car here in L.A. 315 00:18:48,590 --> 00:18:54,200 Great. Now the first self-driving cars were largely invented in 1994. 316 00:18:55,030 --> 00:18:59,370 Stanford won the self-driving competition in 2004. 317 00:18:59,940 --> 00:19:06,780 Google hired all those people in 2005, and it is now 2025. 318 00:19:07,180 --> 00:19:09,697 So it took 20 years from the time 319 00:19:09,780 --> 00:19:12,750 a successful contest among universities to this. 320 00:19:13,180 --> 00:19:15,207 So the technology diffusion is much... 321 00:19:15,290 --> 00:19:19,120 takes much longer, and the wealth creation occurs at every stage. 322 00:19:19,490 --> 00:19:21,007 And since we have a lot of business 323 00:19:21,090 --> 00:19:23,640 students here and I'm very pro-business, obviously, 324 00:19:24,230 --> 00:19:27,387 remember that the majority of the money is not made at the beginning. 325 00:19:27,470 --> 00:19:30,130 It's actually made during the mainstreaming of it, 326 00:19:30,840 --> 00:19:34,040 because the mainstreaming creates trillion-dollar industries. 327 00:19:34,440 --> 00:19:35,727 And that's where the real wealth will be made. 328 00:19:35,810 --> 00:19:40,650 How do you think about China, Eric, in this... 329 00:19:41,050 --> 00:19:46,667 in terms of how universities need to balance, on the one hand, the commitment 330 00:19:46,750 --> 00:19:50,320 that is part of our academic mission to sharing knowledge, 331 00:19:50,590 --> 00:19:57,000 but also the concerns about intellectual property theft and, and, 332 00:19:57,300 --> 00:20:00,147 and we see that those kinds of tensions 333 00:20:00,230 --> 00:20:03,957 running through all sorts of debates that we're now having with respect 334 00:20:04,040 --> 00:20:08,210 to chips to LLMs and in a lot of other parts of AI? 335 00:20:08,310 --> 00:20:10,660 Well, I think, I think you and I agree on this. 336 00:20:10,940 --> 00:20:14,427 I'm very strong and positive on foreign students 337 00:20:14,510 --> 00:20:17,497 in our American universities, especially at the graduate level. 338 00:20:17,580 --> 00:20:20,137 When I was running the National Commission on AI for the Congress, 339 00:20:20,220 --> 00:20:22,960 we studied very carefully the Chinese contribution. 340 00:20:23,650 --> 00:20:27,090 And it turns out all of the really tippy-top papers in America 341 00:20:27,530 --> 00:20:32,477 had one or two out of the five or 10 authors who were Chinese, right? 342 00:20:32,560 --> 00:20:34,317 Chinese born, Chinese trained. 343 00:20:34,400 --> 00:20:37,117 They are an important part of our research infrastructure. 344 00:20:37,200 --> 00:20:38,887 We don't want to get rid of them. 345 00:20:38,970 --> 00:20:40,727 Furthermore, we don't get we want to get rid of them 346 00:20:40,810 --> 00:20:42,857 after they graduate and send them back to China. 347 00:20:42,940 --> 00:20:44,427 That's the stupidest thing ever. 348 00:20:44,510 --> 00:20:47,597 So staple a green card on their diploma, you know, whatever... 349 00:20:47,680 --> 00:20:52,497 Whatever it takes, right? You know, you issue diplomas, you have graduation. 350 00:20:52,580 --> 00:20:55,507 Just take a green card, you know, and just put it on. 351 00:20:55,590 --> 00:20:59,930 Just staple it on. Right? They can write their name in it, right? 352 00:21:00,090 --> 00:21:06,947 Yeah. It's obvious, and... So, why in the world 353 00:21:07,030 --> 00:21:11,257 would we want the top, top students in the world to be anywhere other than America? 354 00:21:11,340 --> 00:21:13,290 Now, if they're criminals, arrest them. 355 00:21:14,070 --> 00:21:16,157 It's not that compli-, we have like a police force 356 00:21:16,240 --> 00:21:17,857 and they're pretty good, right? 357 00:21:17,940 --> 00:21:19,457 And, you know, you have your own police. 358 00:21:19,540 --> 00:21:21,427 What is your problem? You have your own police. 359 00:21:21,510 --> 00:21:22,527 You can make your own green cards. 360 00:21:22,610 --> 00:21:25,210 You do your own diplomas, like a full-stop shopping. 361 00:21:25,520 --> 00:21:29,320 So, so my, my point here is, if you take the right attitude, 362 00:21:29,620 --> 00:21:32,590 then America can continue to lead. 363 00:21:32,990 --> 00:21:34,907 There are national resource differences. 364 00:21:34,990 --> 00:21:37,960 So as I mentioned... Let me just finish on the China thing. 365 00:21:38,100 --> 00:21:42,070 China is going to dominate in a set of areas without question. 366 00:21:42,530 --> 00:21:45,330 They dominate, and this was a mistake on America's part, 367 00:21:45,570 --> 00:21:48,770 the electronic, electric vehicle and the battery systems. 368 00:21:49,040 --> 00:21:51,197 There's no scenario where American firms 369 00:21:51,280 --> 00:21:53,780 will be able to compete at the scale they operate. 370 00:21:54,480 --> 00:21:59,250 I think, and this is controversial, that that will be true at humanoid robots 371 00:21:59,750 --> 00:22:03,007 for the same reasons. I went to some of their factories. 372 00:22:03,090 --> 00:22:05,690 They're huge, and these people work incredibly hard. 373 00:22:06,530 --> 00:22:10,347 So far, if you look at DeepSeek, which is the most impressive 374 00:22:10,430 --> 00:22:12,860 of the research groups, they're good. 375 00:22:13,000 --> 00:22:15,830 I mean, they've invented a number of profoundly better 376 00:22:16,200 --> 00:22:18,400 algorithms to do what is called reasoning, 377 00:22:19,240 --> 00:22:22,587 DeepSeek R1 and so forth. But they don't have the scale. 378 00:22:22,670 --> 00:22:26,010 So, at least at the moment, we're OK. 379 00:22:26,310 --> 00:22:28,267 But I think as a matter of intellectual property, 380 00:22:28,350 --> 00:22:31,180 my preference would be to engage with China 381 00:22:31,650 --> 00:22:34,250 because I've, I, and I want to be clear here, 382 00:22:35,220 --> 00:22:37,807 China and the U.S. will never be best friends, 383 00:22:37,890 --> 00:22:39,790 but we can learn how to work together. 384 00:22:40,490 --> 00:22:44,847 And even though the systems are different, both have a common interest 385 00:22:44,930 --> 00:22:47,287 in the set of things: domestic stability. 386 00:22:47,370 --> 00:22:49,220 We don't want terrorism, nor do they. 387 00:22:49,570 --> 00:22:50,817 Economic growth. 388 00:22:50,900 --> 00:22:53,887 They say they're communist, but they sure look like capitalists 389 00:22:53,970 --> 00:22:58,540 to this capitalist, right? And they work very hard and they're very, very smart. 390 00:22:59,010 --> 00:23:03,810 If you have any confusion about the quality of Chinese research, stop. 391 00:23:03,910 --> 00:23:08,250 You are wrong. All those prejudices from 15 years ago about they're behind, 392 00:23:08,490 --> 00:23:12,190 they’re are no longer true. They are really at the doorstep. 393 00:23:13,220 --> 00:23:17,617 You mentioned DeepSeek, Eric, and one of the weaknesses, 394 00:23:17,700 --> 00:23:20,817 we'll call it, of DeepSeek, is that when you do query it 395 00:23:20,900 --> 00:23:26,200 for information about certain topics, right, you're going to see certain biases. 396 00:23:26,770 --> 00:23:30,327 And I was wondering, and there's a whole debate 397 00:23:30,410 --> 00:23:32,210 that you're well familiar with about 398 00:23:33,740 --> 00:23:39,150 bias and building natural language LLMs and whether it's avoidable or 399 00:23:39,420 --> 00:23:41,570 even the right way to talk about the issue. 400 00:23:42,220 --> 00:23:48,847 But I want to ask you, are you concerned about a future where people start 401 00:23:48,930 --> 00:23:55,370 using the same LLMs and, as a result, you start getting a narrowing of opinion? 402 00:23:56,170 --> 00:24:01,587 When universities like ours are always trying to cultivate the broadest 403 00:24:01,670 --> 00:24:03,920 range of opinions and viewpoints that we can. 404 00:24:04,580 --> 00:24:07,080 So what should you teach? 405 00:24:07,950 --> 00:24:09,750 Curiosity. 406 00:24:10,410 --> 00:24:13,897 Right? And I'm now at the point where I just can't stand it when people... 407 00:24:13,980 --> 00:24:16,930 and I can tell people are paid to convince me of something, 408 00:24:17,360 --> 00:24:18,277 I'm intelligent enough. 409 00:24:18,360 --> 00:24:21,410 Just give me the facts, and I can form my own opinion please. 410 00:24:21,690 --> 00:24:25,617 And I think there's a real danger in society that people will be 411 00:24:25,700 --> 00:24:29,417 captured by an LLM. Remember, it's their best friend. 412 00:24:29,500 --> 00:24:31,187 And if your... your best friend has, 413 00:24:31,270 --> 00:24:33,957 I don't know about you all, has a huge impact on, on me. 414 00:24:34,040 --> 00:24:36,627 You know, I talked... in my case, it's a guy. 415 00:24:36,710 --> 00:24:38,657 I talk to him, he calls me and I do the thing. 416 00:24:38,740 --> 00:24:40,197 You know, he has a big impact on me. 417 00:24:40,280 --> 00:24:42,080 He's really important to me. 418 00:24:43,350 --> 00:24:46,367 And... or your spouse, you know, whatever. 419 00:24:46,450 --> 00:24:48,250 Whatever it, whatever it is. 420 00:24:50,060 --> 00:24:54,730 So part of the answer is that there's not going to be a single LLM solution. 421 00:24:55,530 --> 00:24:58,200 So let me give you the way to think about it. 422 00:24:59,030 --> 00:25:03,187 The... there's a, in America, there is a race to AGI, 423 00:25:03,270 --> 00:25:06,140 which stands for artificial general intelligence. 424 00:25:06,870 --> 00:25:10,040 The leading players there are a startup called Anthropic. 425 00:25:10,670 --> 00:25:12,357 OpenAI, which everybody here is familiar with. 426 00:25:12,440 --> 00:25:15,290 Gemini, of course, which I, is where I came from, Google. 427 00:25:17,020 --> 00:25:18,820 Microsoft is now doing its own. 428 00:25:19,480 --> 00:25:21,937 Elon is attempting to enter the space in Grok. 429 00:25:22,020 --> 00:25:23,507 Maybe there's a couple of others. 430 00:25:23,590 --> 00:25:27,630 These things require an enormous amount of money. 431 00:25:28,130 --> 00:25:32,830 Those systems... and Meta is, not at the moment in the lead. 432 00:25:33,000 --> 00:25:35,587 It is probably going to have to have a closed system to do so. 433 00:25:35,670 --> 00:25:38,287 Closed means they're not going to give away the source, 434 00:25:38,370 --> 00:25:41,327 just because the amount of money involved, they're not going to give it away. 435 00:25:41,410 --> 00:25:45,810 The Chinese have strangely made the inverse decision. 436 00:25:46,040 --> 00:25:48,750 Every Chinese model is open source. 437 00:25:49,610 --> 00:25:50,937 So I have a number of startups 438 00:25:51,020 --> 00:25:54,490 which have a high quality of very, very smart Chinese people, 439 00:25:54,750 --> 00:25:56,537 and they just use the Chinese LLMs. 440 00:25:56,620 --> 00:25:58,137 And they're not political startups. 441 00:25:58,220 --> 00:25:59,977 They're, they're just useful and they're free. 442 00:26:00,060 --> 00:26:01,690 I said, sure, right? 443 00:26:03,490 --> 00:26:07,447 So it's likely that the, in the media, in the next five or 10 years, 444 00:26:07,530 --> 00:26:09,717 you're going to see this huge amount of tension 445 00:26:09,800 --> 00:26:11,900 about, as we get closer and closer to AGI. 446 00:26:12,370 --> 00:26:15,857 And when we hit various thresholds, you know, when something... 447 00:26:15,940 --> 00:26:19,810 We already have systems that can answer graduate-level 448 00:26:20,280 --> 00:26:22,380 biology, chemistry, math problems, 449 00:26:22,680 --> 00:26:25,267 which I look at them and I say it's inconceivable to me 450 00:26:25,350 --> 00:26:27,137 that it can do it and it can solve those. 451 00:26:27,220 --> 00:26:29,307 Those are like headline events in my world. 452 00:26:29,390 --> 00:26:31,590 AlphaFold, headline events in my world. 453 00:26:31,790 --> 00:26:33,590 There will be more, 454 00:26:34,220 --> 00:26:38,030 but China will offer those in open source form, 455 00:26:38,500 --> 00:26:40,747 and there will be millions of smaller models 456 00:26:40,830 --> 00:26:43,340 in open source from, from many countries. 457 00:26:43,800 --> 00:26:46,170 So I think you won't have as big a lock in, 458 00:26:46,540 --> 00:26:50,090 especially if the university teaches you not to become locked into one. 459 00:26:50,740 --> 00:26:52,997 I move around. I try different ones. 460 00:26:53,080 --> 00:26:57,197 Do you feel like a lot of the American companies that are developing 461 00:26:57,280 --> 00:27:02,167 these LLMs have taken the wrong direction with respect to open source? 462 00:27:02,250 --> 00:27:04,050 They didn't really have a choice. 463 00:27:04,290 --> 00:27:06,007 I'm, my back-, just to let you know, 464 00:27:06,090 --> 00:27:07,907 as a computer scientist, my background was open source. 465 00:27:07,990 --> 00:27:11,877 Open source means we literally publish our source code. 466 00:27:11,960 --> 00:27:18,100 And I was part of something called Unix, which is the underlying, 467 00:27:18,670 --> 00:27:21,570 for example, it underlies the Mac operating system. 468 00:27:21,840 --> 00:27:23,990 For those of you who don't know what it is. 469 00:27:24,480 --> 00:27:26,257 And it was a great time 470 00:27:26,340 --> 00:27:29,497 because we were exchanging ideas, we were working collaboratively. 471 00:27:29,580 --> 00:27:32,597 And look what we built. The internet is largely 472 00:27:32,680 --> 00:27:35,480 built because of open source architectures and so forth. 473 00:27:35,750 --> 00:27:37,450 So there's a lot of power in that. 474 00:27:38,290 --> 00:27:42,507 I think realistically, the, so, if you look at... 475 00:27:42,590 --> 00:27:45,517 so for example, Sam, one day called me up and said, you're going to be happy. 476 00:27:45,600 --> 00:27:47,347 I'm going to take open, 477 00:27:47,430 --> 00:27:50,147 o3 and do a mini-version and release it in open source. 478 00:27:50,230 --> 00:27:52,117 And they did. And it's pretty good, right? 479 00:27:52,200 --> 00:27:52,957 So I think there's, 480 00:27:53,040 --> 00:27:56,287 there will be attempts by the American companies to release subset. 481 00:27:56,370 --> 00:28:02,097 One of the core questions for the American companies is the raw models are so big 482 00:28:02,180 --> 00:28:06,437 and so powerful, will they ever release the biggest models? 483 00:28:06,520 --> 00:28:13,307 It sure looks to me like they're going to release a portfolio of models from super 484 00:28:13,390 --> 00:28:18,847 huge, huge, medium range, small, open source, with different pricing points, 485 00:28:18,930 --> 00:28:20,417 and it looks like that's what they're going to do. 486 00:28:20,500 --> 00:28:23,487 And the reasons have to do with the cost of serving and the cost 487 00:28:23,570 --> 00:28:24,570 of the architecture. 488 00:28:25,270 --> 00:28:28,257 Eric, you mentioned AlphaFold, and maybe you can explain a little bit 489 00:28:28,340 --> 00:28:30,257 about what AlphaFold was able to do. 490 00:28:30,340 --> 00:28:36,397 And, and your former company, Google, with Demis Hassabis, Demis 491 00:28:36,480 --> 00:28:41,367 Hassabis, has just been doing incredible work in bringing AI 492 00:28:41,450 --> 00:28:44,360 to generate scientific breakthroughs. 493 00:28:44,720 --> 00:28:49,130 And all of this requires a tremendous amount of compute. 494 00:28:50,030 --> 00:28:55,130 What's your advice to universities that want to be at that table? 495 00:28:57,100 --> 00:29:00,870 I told you already. Find a Republican state, 496 00:29:01,140 --> 00:29:05,740 build a large data center there, and run a big pipe to wherever you are. 497 00:29:06,850 --> 00:29:10,037 Sorry. No, no, no. I just want you to say it in front of everyone. 498 00:29:10,120 --> 00:29:15,190 Oh, I see. You want me to support you, and you have a Republican state in, in mind? 499 00:29:15,550 --> 00:29:18,877 Well, you gave me an idea. Well, I think Nevada is pretty close. 500 00:29:18,960 --> 00:29:19,793 Yeah. 501 00:29:20,460 --> 00:29:22,177 But, But, but, but you think it’s... 502 00:29:22,260 --> 00:29:24,860 I'm not being... I'm not being facetious. 503 00:29:25,960 --> 00:29:27,770 So, so, 504 00:29:30,130 --> 00:29:33,540 So universities need to keep up with this. 505 00:29:33,810 --> 00:29:38,340 There is no way the future can be invented in the, in universities 506 00:29:38,740 --> 00:29:41,350 without enough computing resource. 507 00:29:41,780 --> 00:29:44,797 No university I know has anywhere close. 508 00:29:44,880 --> 00:29:46,867 These are the richest universities in the world. 509 00:29:46,950 --> 00:29:49,367 I mean, think about the British universities have no money at all. 510 00:29:49,450 --> 00:29:51,537 At least, I mean, you guys, you have a lot a lot of money. 511 00:29:51,620 --> 00:29:53,390 It's private school. Very well run. 512 00:29:54,090 --> 00:29:57,960 And you don't have it. So first, you need to build a big data center. 513 00:29:58,630 --> 00:30:01,430 And, and many universities are power constrained. 514 00:30:01,930 --> 00:30:06,170 And then what I would do is I would get donors to donate the necessary chips. 515 00:30:07,700 --> 00:30:11,070 As you know, Nvidia is 90, 95% of the training market right now. 516 00:30:11,540 --> 00:30:12,997 They have amazing chips. 517 00:30:13,080 --> 00:30:16,067 And of course, they're now the most valuable company in the world. 518 00:30:16,150 --> 00:30:18,967 I'm pleased to report that Google now is working on 519 00:30:19,050 --> 00:30:21,890 competitive offerings, and the TPU V7 chip 520 00:30:22,250 --> 00:30:25,200 appears to be better than the B-300 chip for most training. 521 00:30:25,290 --> 00:30:27,960 Of course, that's widely disputed in the industry. 522 00:30:28,290 --> 00:30:31,130 That competition is really important. 523 00:30:31,500 --> 00:30:36,017 So the answer to the crazy stuff that's going on in tech is hardcore competition, 524 00:30:36,100 --> 00:30:37,900 which will benefit the university. 525 00:30:38,140 --> 00:30:40,927 Well, I have your number. So we'll be talking about philanthropic 526 00:30:41,010 --> 00:30:44,640 opportunities here at USC. So... 527 00:30:45,680 --> 00:30:47,980 And I... and I'm happy to help. 528 00:30:48,650 --> 00:30:53,250 Just a couple more questions. Wanted to ask you, what do you think 529 00:30:53,520 --> 00:30:56,137 one of the most underappreciated opportunities 530 00:30:56,220 --> 00:30:59,020 for universities are in the AI space? 531 00:31:00,490 --> 00:31:02,290 I think I said it before. 532 00:31:03,360 --> 00:31:05,747 The world I live in, I'll just... 533 00:31:05,830 --> 00:31:09,830 I'll just be very direct, is largely dominated by the most brilliant 534 00:31:10,000 --> 00:31:11,840 20-somethings, men and women, 535 00:31:12,770 --> 00:31:16,527 because the stuff is so new that if you're 40, 536 00:31:16,610 --> 00:31:19,940 you didn't go through all those painful classes to learn it. 537 00:31:20,610 --> 00:31:24,037 Right? Now, you're smart at 40 and you know you can study it and so forth. 538 00:31:24,120 --> 00:31:28,750 But the people who have forged through the crucible of their undergraduate work. 539 00:31:29,120 --> 00:31:32,637 At Princeton, I'm from Princeton, I met with a bunch of sophomores 540 00:31:32,720 --> 00:31:35,770 who were doing alternative reinforcement learning algorithms, 541 00:31:35,890 --> 00:31:38,000 and I go like, what, as sophomores? 542 00:31:39,030 --> 00:31:43,117 I'm sure that's true here. So I think people do not appreciate the, 543 00:31:43,200 --> 00:31:46,117 the quality of the ideas 544 00:31:46,200 --> 00:31:49,350 and the capability of moving quickly with the right tools here. 545 00:31:49,540 --> 00:31:51,740 So let's go through what that could mean. 546 00:31:52,340 --> 00:31:55,427 In humanities, why do you not have a database 547 00:31:55,510 --> 00:31:58,597 that has taken all the documents that matter and made them 548 00:31:58,680 --> 00:32:01,650 accessible and classifiable in interesting new ways? 549 00:32:02,820 --> 00:32:06,107 In physics, people are using models called diffusion models, 550 00:32:06,190 --> 00:32:08,960 which are hard to explain, but basically you add noise 551 00:32:09,230 --> 00:32:11,830 and then you reduce the noise and you discover an... 552 00:32:12,160 --> 00:32:14,330 you discover the underlying pattern. 553 00:32:14,570 --> 00:32:17,570 Physics is a lot about understanding the underlying pattern. 554 00:32:18,100 --> 00:32:19,900 In biology with AlphaFold... 555 00:32:20,570 --> 00:32:22,370 So proteins, 556 00:32:23,110 --> 00:32:26,610 so, the primary things that LLMs, 557 00:32:27,210 --> 00:32:28,980 what you know, is open ChatGPT 558 00:32:29,880 --> 00:32:31,680 do is they do sequence prediction. 559 00:32:32,080 --> 00:32:35,237 So what they're really doing is they're doing next word prediction. 560 00:32:35,320 --> 00:32:38,777 They just do it really well. And I often tell the story 561 00:32:38,860 --> 00:32:41,490 that it's like on a Thursday night in San Francisco 562 00:32:41,860 --> 00:32:46,017 and they were sitting, at OpenAI, and these are a couple of my friends. 563 00:32:46,100 --> 00:32:49,547 And they decided to turn on what became GPT-3o. 564 00:32:49,630 --> 00:32:51,480 So this is, you know, four years ago. 565 00:32:52,200 --> 00:32:55,940 And they said, “Oh my God, this thing can write.” 566 00:32:57,540 --> 00:33:00,640 They didn't know. It's a real historical discovery. 567 00:33:01,410 --> 00:33:03,597 And there's a, there's a moment 568 00:33:03,680 --> 00:33:06,930 where that occurred which will ultimately be recorded in history. 569 00:33:07,420 --> 00:33:10,437 And what's interesting is we think, “Oh, that's so stupid. 570 00:33:10,520 --> 00:33:13,107 Next word prediction?” By the way, how does your brain work? 571 00:33:13,190 --> 00:33:14,990 It's called next word prediction. 572 00:33:15,830 --> 00:33:18,117 It doesn't use exactly neural nets the same way. 573 00:33:18,200 --> 00:33:20,630 And it has other agencies and so forth. 574 00:33:21,000 --> 00:33:23,617 But fundamentally it's closer to the way people think 575 00:33:23,700 --> 00:33:26,150 than you're willing to admit, speaking as humans. 576 00:33:26,970 --> 00:33:33,357 So what I like about, about all of this is that the expansion and the speed 577 00:33:33,440 --> 00:33:37,090 that is possible when you take next word prediction and you add planning, 578 00:33:37,210 --> 00:33:41,467 which is essentially a series of steps, you can kind of do anything. 579 00:33:41,550 --> 00:33:44,037 Here's an example. So I like to buy houses. 580 00:33:44,120 --> 00:33:45,037 I like to build houses. 581 00:33:45,120 --> 00:33:47,707 I want to buy a house in California, which is always a mess. 582 00:33:47,790 --> 00:33:50,677 So the first agent, an agent is input and output. 583 00:33:50,760 --> 00:33:52,877 The input is language. It has memory. 584 00:33:52,960 --> 00:33:54,047 It has output, which is language. 585 00:33:54,130 --> 00:33:56,517 The first agent tells me where to buy. 586 00:33:56,600 --> 00:33:58,370 The second agent consults the, the 587 00:33:59,340 --> 00:34:02,027 published rules. The third agent designs the home 588 00:34:02,110 --> 00:34:03,957 with a known architect, which it selects. 589 00:34:04,040 --> 00:34:07,180 An actual human. The fourth hires a contractor, 590 00:34:07,440 --> 00:34:08,897 an actual human, to build the house. 591 00:34:08,980 --> 00:34:12,030 And the fifth agent sues the contractor when it doesn't work. 592 00:34:12,450 --> 00:34:14,890 Right? That's a workflow. 593 00:34:16,250 --> 00:34:17,837 This is... Welcome to my world. 594 00:34:17,920 --> 00:34:21,530 Now, you sit there and you go, why is this so... 595 00:34:21,660 --> 00:34:23,807 I mean, why is he giving us such a stupid example? 596 00:34:23,890 --> 00:34:27,870 Because you'll remember it. I just described the workflow in a university. 597 00:34:28,330 --> 00:34:30,600 I just described a workflow in a hospital. 598 00:34:30,900 --> 00:34:35,187 I just described the workflow in government and in businesses. 599 00:34:35,270 --> 00:34:37,480 So the agentic revolution, which is what, 600 00:34:37,910 --> 00:34:42,727 how we in the, in the computer industry call it, is the next big thing. 601 00:34:42,810 --> 00:34:44,467 So another piece of advice 602 00:34:44,550 --> 00:34:46,900 for universities is just start building agents. 603 00:34:47,920 --> 00:34:48,807 So if the graduate 604 00:34:48,890 --> 00:34:51,677 students in computer science, who in this area will be the smartest, 605 00:34:51,760 --> 00:34:55,347 just build an agent orchestration framework to run all the agents 606 00:34:55,430 --> 00:34:59,030 and tell everyone else in the university to build agents and have fun. 607 00:34:59,400 --> 00:35:02,217 Most of the agents will be kind of useless, but that's OK. 608 00:35:02,300 --> 00:35:05,257 It's a university. But people will learn what agents are like. 609 00:35:05,340 --> 00:35:08,640 So combination of right agents, which the computer can do for you, 610 00:35:09,110 --> 00:35:10,257 and then prompt engineering, 611 00:35:10,340 --> 00:35:13,240 which allows you to program anything, you're kind of done, 612 00:35:13,980 --> 00:35:18,797 right? Think about the power of your labor force, which is here in the room. 613 00:35:18,880 --> 00:35:20,720 Fifty thousand total? 614 00:35:21,650 --> 00:35:26,807 Boy, I'd take that. Last question, and maybe you started to answer it just now. 615 00:35:26,890 --> 00:35:30,007 So I was wondering, Eric, if you could just speak directly 616 00:35:30,090 --> 00:35:33,030 to all the students who are listening today? 617 00:35:33,460 --> 00:35:36,900 We see all the headlines, you know, jobs are going away, 618 00:35:38,170 --> 00:35:41,970 reorganizations. And I think there's, 619 00:35:43,370 --> 00:35:45,897 there's reason to be really, really excited. 620 00:35:45,980 --> 00:35:49,550 And there's also reasons for our students who, again, as we discussed, 621 00:35:49,910 --> 00:35:54,967 span hard sciences to the performing arts to be really worried about the future 622 00:35:55,050 --> 00:35:57,260 and how AI is going to impact it. 623 00:35:57,450 --> 00:35:59,407 What's your, what's your message to them? 624 00:35:59,490 --> 00:36:03,307 There's this problem with smart people that smart people are always depressed. 625 00:36:03,390 --> 00:36:08,930 Stop. Stop thinking so hard. I am super jealous 626 00:36:09,270 --> 00:36:11,917 of your undergraduates. I want to be very clear. 627 00:36:12,000 --> 00:36:14,200 One, you're in California, stop complaining. 628 00:36:14,540 --> 00:36:18,810 OK? Like I moved here as a graduate student for the weather. 629 00:36:19,110 --> 00:36:23,667 So we're clear. Second, in your lifetimes, 630 00:36:23,750 --> 00:36:26,580 you're going to see life extensions to 100 or more. 631 00:36:27,250 --> 00:36:31,760 Your children will be infinitely healthier than my generation ever was. 632 00:36:32,390 --> 00:36:34,060 Hopefully, the world will get 633 00:36:35,130 --> 00:36:39,147 more safe and less dangerous. So far, that's generally been true. 634 00:36:39,230 --> 00:36:42,717 There are plenty of studies on this, but most important, 635 00:36:42,800 --> 00:36:46,570 you're going to work on the arrival of a new form of intelligence 636 00:36:47,000 --> 00:36:49,827 that you can work with. And yes, yes, yes, 637 00:36:49,910 --> 00:36:53,197 there's all the movies about how the intelligence gets out of control. 638 00:36:53,280 --> 00:36:55,997 And then, you know, the the lady scientist goes 639 00:36:56,080 --> 00:36:58,897 and kills the computer and the man is, you know, whatever. 640 00:36:58,980 --> 00:37:01,037 Those are fun movies, and you're good at movies. 641 00:37:01,120 --> 00:37:03,507 But the important thing is the arrival of intelligence 642 00:37:03,590 --> 00:37:06,607 is such a productivity increase, and productivity, 643 00:37:06,690 --> 00:37:09,460 I mean, as a personal level, for entertainment, 644 00:37:10,060 --> 00:37:12,847 right, for your social life. Right? 645 00:37:12,930 --> 00:37:15,900 So start thinking as a 20-something 646 00:37:16,300 --> 00:37:19,050 about how you can participate in the agentic revolution 647 00:37:19,170 --> 00:37:20,970 and think like a computer scientist. 648 00:37:21,270 --> 00:37:25,697 Don't do something for you. Do something for everyone around 649 00:37:25,780 --> 00:37:28,330 you and everyone you know and all of their friends. 650 00:37:29,480 --> 00:37:33,067 A good rule is, so why are computer scientists doing so well 651 00:37:33,150 --> 00:37:35,300 besides, as I mentioned, our own arrogance. 652 00:37:35,790 --> 00:37:39,307 Because we think in terms of scale, and because I'm, I'm 653 00:37:39,390 --> 00:37:41,547 well known now, a lot of people come up to me 654 00:37:41,630 --> 00:37:44,447 with interesting business ideas, but none of them will scale. 655 00:37:44,530 --> 00:37:46,480 Come up with a scalable business model. 656 00:37:46,860 --> 00:37:49,147 And why is a scalable business model interesting? 657 00:37:49,230 --> 00:37:53,300 Because you can make a fortune, which is why you're in business school. 658 00:37:53,570 --> 00:37:57,357 Sorry. You see my point? I would be... 659 00:37:57,440 --> 00:37:59,427 You should be, in fact, this is probably why you work here. 660 00:37:59,510 --> 00:38:02,667 You want to hang out with them because you're jealous of them, too? 661 00:38:02,750 --> 00:38:05,197 Absolutely. But thank you for having me. 662 00:38:05,280 --> 00:38:06,967 This is a great honor to kick this off. 663 00:38:07,050 --> 00:38:09,337 Eric Schmidt, thank you for your contributions to our society. 664 00:38:09,420 --> 00:38:10,520 Thanks for being here.