1 00:00:47,302 --> 00:00:50,403 Welcome back to the DeepView Conversations. Today for our 2 00:00:50,463 --> 00:00:53,712 second episode, I traveled all the way from 3 00:00:53,892 --> 00:00:57,974 New Jersey to the United Arab Emirates, specifically 4 00:00:58,034 --> 00:01:01,355 Abu Dhabi, which is the home of the Mohamed bin Zayed University of 5 00:01:01,395 --> 00:01:04,857 Artificial Intelligence. This is the world's first and only 6 00:01:05,717 --> 00:01:09,118 university that is focused exclusively on AI research. And 7 00:01:09,158 --> 00:01:12,460 right now I'm speaking with Dr. Eric Jing, who is the 8 00:01:12,540 --> 00:01:15,721 president of MBZ UAI. Eric, thanks so 9 00:01:17,657 --> 00:01:21,282 Well, thank you for traveling all the way here. I appreciate your diligence 10 00:01:21,603 --> 00:01:25,048 and I love to chat 11 00:01:28,162 --> 00:01:31,324 So I want to start with you, right? And as we've kind 12 00:01:31,344 --> 00:01:34,825 of talked about before, the cameras started rolling, right? Your background 13 00:01:34,885 --> 00:01:38,407 is in research, and you've been a professor, and you still kind 14 00:01:38,447 --> 00:01:42,149 of wear that research hat. How do you go from a professor 15 00:01:42,189 --> 00:01:46,271 of computer science to the president of MBZUAI 16 00:01:47,391 --> 00:01:50,793 Well, to be honest, I never thought I'd become a 17 00:01:50,873 --> 00:01:56,306 president of a university at any time. I'm 18 00:01:56,366 --> 00:02:00,249 still right now affiliated with Carnegie Mellon. I've been working 19 00:02:00,309 --> 00:02:04,112 as a professor in Carnegie Mellon since 2004. I 20 00:02:05,293 --> 00:02:08,455 advise students, I teach classes, I 21 00:02:08,515 --> 00:02:11,698 write grant proposals, and write 22 00:02:11,738 --> 00:02:15,220 papers. In fact, I'm still doing all this right now, except 23 00:02:15,701 --> 00:02:19,344 for classroom teaching. So research is very, very strong, very, 24 00:02:19,364 --> 00:02:22,645 very important for me. That's why I haven't 25 00:02:22,665 --> 00:02:27,286 thought about becoming a university administrator. During my career, when 26 00:02:27,306 --> 00:02:31,687 I became senior, I did receive numerous invitations 27 00:02:31,787 --> 00:02:35,388 and requests to be department chairs, deans, 28 00:02:35,988 --> 00:02:39,668 even in one case another president, in 29 00:02:39,688 --> 00:02:43,049 different institutions in the United States or in 30 00:02:43,329 --> 00:02:46,650 UK and other places in the world. And I 31 00:02:46,670 --> 00:02:50,998 didn't leave Carnegie Mellon. So 32 00:02:51,058 --> 00:02:54,662 when I got an invitation for interviewing 33 00:02:54,762 --> 00:02:58,565 here, I thought it's maybe 34 00:02:58,585 --> 00:03:02,029 a learning experience for me to see the 35 00:03:02,049 --> 00:03:06,433 place and also to maybe offer my my 36 00:03:06,593 --> 00:03:11,157 insights or maybe experiences or opinions to 37 00:03:11,858 --> 00:03:15,562 the leadership in here. During the interview, my interaction 38 00:03:15,662 --> 00:03:20,206 with the host here is mainly based 39 00:03:20,386 --> 00:03:23,650 on my identity as a researcher, as 40 00:03:23,730 --> 00:03:26,973 a developer, and as a scientist. So I give them 41 00:03:26,993 --> 00:03:30,296 basically the viewpoint you 42 00:03:30,316 --> 00:03:33,599 know, from that angle, you know, how a university should be 43 00:03:34,280 --> 00:03:38,524 built, should be operated, and what's 44 00:03:38,604 --> 00:03:43,088 the best function and reputation 45 00:03:43,348 --> 00:03:47,312 for university. And I also know that I'm 46 00:03:47,352 --> 00:03:51,476 not the only one they interviewed, you know, they interviewed quite a few people. All 47 00:03:51,516 --> 00:03:54,842 of them are much more senior and and 48 00:03:55,083 --> 00:03:58,729 recognized that me. So I wasn't taking it too seriously 49 00:03:59,010 --> 00:04:03,077 to be a real job interview. It's really an exchange, an 50 00:04:03,137 --> 00:04:08,434 intellectual exchange. So I went back home. But then After 51 00:04:08,534 --> 00:04:11,755 some time, to my surprise, they made an offer to 52 00:04:11,795 --> 00:04:14,956 me. I wasn't ready at 53 00:04:15,116 --> 00:04:19,638 all to take this offer because at 54 00:04:19,658 --> 00:04:23,379 that time I was still very busy working with my students 55 00:04:23,719 --> 00:04:27,100 and working on a few exciting projects. That was 56 00:04:27,240 --> 00:04:30,901 at the beginning of the COVID. In fact, the 57 00:04:30,941 --> 00:04:34,262 trip to UAE for the interview was the last trip I made 58 00:04:34,743 --> 00:04:38,211 before travel shut down. in the States and 59 00:04:38,291 --> 00:04:41,413 around the world. As we entered the 60 00:04:41,473 --> 00:04:44,715 COVID season, things changed a 61 00:04:44,855 --> 00:04:48,437 lot, even back home in the States. You 62 00:04:49,037 --> 00:04:52,219 are locking down at home. You don't really get to 63 00:04:52,259 --> 00:04:55,701 see your students so often. You teach through a camera, which 64 00:04:56,021 --> 00:04:59,463 all were very new to all of us. And you started 65 00:05:01,864 --> 00:05:05,709 getting yourself isolated from the rest of the world. Which 66 00:05:05,789 --> 00:05:09,551 is good and bad. I didn't like that experience. I feel research 67 00:05:11,092 --> 00:05:15,374 is important for science to be done collaboratively 68 00:05:15,474 --> 00:05:18,735 with colleagues, with students. So I missed that part. 69 00:05:19,176 --> 00:05:23,038 But on the other hand, it gave me the time to think. and 70 00:05:23,158 --> 00:05:27,601 to really ask myself again where 71 00:05:27,621 --> 00:05:30,763 I'm heading. So with this offer, which is 72 00:05:31,203 --> 00:05:34,565 definitely non-trivial, it is supposed to 73 00:05:34,605 --> 00:05:38,448 be bringing me to a place that I've never had previous 74 00:05:38,568 --> 00:05:41,890 interaction and experience with, and also it's a job I've not 75 00:05:41,931 --> 00:05:45,573 done before. So I consulted 76 00:05:45,713 --> 00:05:49,876 some of my colleagues, mentors, and 77 00:05:50,536 --> 00:05:54,038 friends. Professor Sir Michael Bradley, the former 78 00:05:54,158 --> 00:05:57,620 acting president of the university, he 79 00:05:57,680 --> 00:06:01,722 made me a very, very interesting case. He said, look, this 80 00:06:01,842 --> 00:06:06,184 is more than just a university administrator. It 81 00:06:06,304 --> 00:06:09,706 is about creating something from 82 00:06:12,267 --> 00:06:16,089 scratch with a clean slate. And also, it 83 00:06:16,209 --> 00:06:19,471 is in the region where no 84 00:06:19,531 --> 00:06:23,336 such university ever existed. And the 85 00:06:23,376 --> 00:06:27,683 country and the region needs such an institution to 86 00:06:28,424 --> 00:06:31,989 really be the beacon of a new culture. 87 00:06:33,010 --> 00:06:36,959 In a sense, I have the opportunity to to 88 00:06:37,059 --> 00:06:41,761 influence and to shape a new type of culture, to 89 00:06:41,901 --> 00:06:45,361 really help, you know, a 90 00:06:45,542 --> 00:06:49,202 whole people, a whole country, you know, to 91 00:06:49,543 --> 00:06:54,684 seek new directions, you know, or create new 92 00:06:55,584 --> 00:06:59,566 results and also to redefine you 93 00:06:59,586 --> 00:07:03,310 know, some of the decision process and the operation and 94 00:07:03,630 --> 00:07:06,733 even culture, you know, in the region. So in a sense, there is 95 00:07:06,773 --> 00:07:09,956 a chance for you to impact a country to 96 00:07:10,016 --> 00:07:13,479 reshape the culture. So that's quite exciting to 97 00:07:13,499 --> 00:07:16,983 me. I'm always surrounded by adventures in trying new things, and 98 00:07:17,023 --> 00:07:21,827 this is definitely very new. Then I asked another friend 99 00:07:22,047 --> 00:07:25,563 and mentor, Professor Michael Jordan, who has 100 00:07:25,623 --> 00:07:28,884 been my advisor back in UC 101 00:07:28,964 --> 00:07:32,064 Berkeley when I was doing my PhD. But then we kept the 102 00:07:32,104 --> 00:07:35,685 friendship all the way. And he 103 00:07:35,925 --> 00:07:39,606 also strongly suggested 104 00:07:39,626 --> 00:07:43,207 to me to consider this opportunity more seriously. He 105 00:07:43,247 --> 00:07:46,767 said, Hui, you are a professor 106 00:07:46,867 --> 00:07:50,228 and a researcher with a good track 107 00:07:50,268 --> 00:07:53,489 record of being productive and being creative on the 108 00:07:53,509 --> 00:07:56,751 scientific part. But what could be 109 00:07:56,771 --> 00:08:00,054 the next big stage? And that's what he asked, 110 00:08:00,214 --> 00:08:03,297 right? Yeah, he understands that I want to 111 00:08:03,317 --> 00:08:07,080 continue to do science, but bigger science also requires bigger 112 00:08:07,120 --> 00:08:10,643 resources. And also, from a pure 113 00:08:11,424 --> 00:08:15,187 career development standpoint, getting such 114 00:08:15,307 --> 00:08:18,929 an offer at a relatively young age really actually 115 00:08:18,989 --> 00:08:22,670 means that you have the 116 00:08:22,730 --> 00:08:27,011 opportunity to experiment and also to try 117 00:08:27,331 --> 00:08:31,292 different new ideas. In a sense, you have a chance to fail and 118 00:08:31,312 --> 00:08:34,593 do it again, do it differently and so forth. So 119 00:08:34,633 --> 00:08:38,554 he said this is a once in a life opportunity because 120 00:08:39,094 --> 00:08:42,634 not every country are trying to build a university of 121 00:08:42,714 --> 00:08:46,043 this and also not any time such 122 00:08:46,063 --> 00:08:49,726 thing can happen. This is maybe just a once-in-a-life opportunity. 123 00:08:50,467 --> 00:08:54,210 So these are the two strongest positive inputs 124 00:08:54,290 --> 00:08:57,874 I can remember. There are other inputs as well that 125 00:08:57,974 --> 00:09:01,978 are as positive. But of course, I also received a lot of warnings 126 00:09:02,218 --> 00:09:05,481 and negative feedbacks. It is very 127 00:09:05,521 --> 00:09:11,309 risky. It could ruin my career and reputation if I screw the job. or 128 00:09:11,750 --> 00:09:15,232 I may run into a very different culture that I 129 00:09:15,272 --> 00:09:18,613 just cannot understand and work with, 130 00:09:18,793 --> 00:09:22,956 all these pragmatic issues and other things. So 131 00:09:23,516 --> 00:09:26,578 I took a step back and I 132 00:09:26,618 --> 00:09:30,820 took advantage of being in the height of the COVID with 133 00:09:30,860 --> 00:09:34,142 very few people to talk to if I don't 134 00:09:34,322 --> 00:09:37,727 actively seek out and this summer, in that thinking. 135 00:09:37,987 --> 00:09:41,328 And in the end, I think I 136 00:09:41,368 --> 00:09:44,969 will be living in my house probably 137 00:09:45,029 --> 00:09:48,470 for the next God knows how many months and 138 00:09:48,530 --> 00:09:53,332 years. I wouldn't be able to achieve too many 139 00:09:53,452 --> 00:09:57,173 exciting things anyway. And this 140 00:09:57,273 --> 00:10:00,514 particular opportunity is indeed once in a life, right? 141 00:10:01,343 --> 00:10:04,930 And from a technical standpoint, I 142 00:10:04,990 --> 00:10:08,457 started to see AI as a rising 143 00:10:08,557 --> 00:10:13,484 sun on the horizon. the established 144 00:10:13,524 --> 00:10:17,465 university infrastructure is already having 145 00:10:17,526 --> 00:10:21,027 difficulties in coping with the new needs, 146 00:10:22,188 --> 00:10:25,489 the new culture, and 147 00:10:25,529 --> 00:10:30,072 the new evaluation and merit 148 00:10:30,152 --> 00:10:33,313 metrics in this new discipline. But it is 149 00:10:33,373 --> 00:10:37,555 very hard to change an established organization. So 150 00:10:38,036 --> 00:10:41,719 even from a self-professional growth 151 00:10:41,939 --> 00:10:46,523 and development point of view, it 152 00:10:46,583 --> 00:10:50,147 is an opportunity to take your fate into your own hands and 153 00:10:50,247 --> 00:10:53,710 also maybe create an environment that can help many of 154 00:10:53,890 --> 00:10:56,993 our colleagues, students, and of 155 00:10:57,053 --> 00:11:00,909 course the country and the people in 156 00:11:01,289 --> 00:11:06,291 this region to a brand new set 157 00:11:06,371 --> 00:11:10,033 up of such an institution. So I see that as a, 158 00:11:10,513 --> 00:11:13,795 I start to appreciate that's becoming a huge opportunity that I can really 159 00:11:13,835 --> 00:11:18,194 make a difference because I am an active researcher. and 160 00:11:18,675 --> 00:11:22,458 it is not so common for an active 161 00:11:22,578 --> 00:11:25,921 researcher and teacher to jump 162 00:11:26,001 --> 00:11:29,344 into a university professor role. Usually, the 163 00:11:29,364 --> 00:11:33,608 typical presidents you see are already somewhat 164 00:11:33,688 --> 00:11:37,171 removed from the daily scientific life and 165 00:11:37,231 --> 00:11:40,492 teaching life. Therefore, it 166 00:11:40,532 --> 00:11:44,013 can be difficult for them to appreciate, you know, from 167 00:11:44,153 --> 00:11:47,474 a close distance, the difficulties, the 168 00:11:47,534 --> 00:11:51,695 needs, and all that, you know, in programming 169 00:11:52,016 --> 00:11:55,197 and in organizing the university, right? So 170 00:11:55,277 --> 00:11:58,638 all this converged to, you know, a 171 00:11:59,858 --> 00:12:03,170 decision that, you know, I'm going to give it a try. And 172 00:12:03,570 --> 00:12:06,651 then, you know, it took me a 173 00:12:06,691 --> 00:12:10,332 year to make that decision. I actually felt 174 00:12:10,492 --> 00:12:14,253 really appreciative that the university was 175 00:12:14,413 --> 00:12:17,753 able to wait for me for that year. In fact, also the 176 00:12:17,853 --> 00:12:21,274 Board of Trustees, they had the patience and also the trust 177 00:12:21,834 --> 00:12:25,389 on me, you know, to onboard myself. I 178 00:12:25,429 --> 00:12:29,232 remember it took me some effort, took 179 00:12:29,292 --> 00:12:32,614 everybody a lot of effort, in fact, to even make the trip to here because all 180 00:12:32,754 --> 00:12:36,256 the flights were unpredictable. And even getting 181 00:12:36,276 --> 00:12:40,579 a COVID test certificate is also unpredictable. 182 00:12:40,979 --> 00:12:46,302 But you need that to board a plane, which is a 24-hour fresh kind of certificate. a 183 00:12:46,322 --> 00:12:49,744 lot of logistic challenges. So eventually I made 184 00:12:49,764 --> 00:12:53,607 it to Abu Dhabi one day before the first 185 00:12:53,727 --> 00:12:57,369 class of the students opened. I remember I got 186 00:12:57,409 --> 00:13:01,752 myself prepared in a temporary 187 00:13:02,072 --> 00:13:06,255 residence that I basically borrowed a 188 00:13:06,695 --> 00:13:10,358 room from a very kind host, and 189 00:13:10,599 --> 00:13:14,382 I delivered the opening speech of 190 00:13:14,442 --> 00:13:17,665 that class, again, from a Zoom 191 00:13:17,685 --> 00:13:21,385 meeting. Very 192 00:13:21,405 --> 00:13:24,666 simple, not as fancy as today, the recording devices, which is coming 193 00:13:24,706 --> 00:13:28,346 from my computer camera. So that's basically how the whole thing started. 194 00:13:28,966 --> 00:13:32,527 At that time, the university was really, really very young and weak. 195 00:13:33,687 --> 00:13:37,068 It has barely, I think, less 196 00:13:37,088 --> 00:13:40,268 than a dozen faculty and only one 197 00:13:40,328 --> 00:13:44,629 class of students, which was 198 00:13:44,669 --> 00:13:48,050 very good. But at least I didn't have 199 00:13:48,070 --> 00:13:51,342 any control or 200 00:13:51,442 --> 00:13:55,067 knowledge about that because I was not admitted under 201 00:13:55,248 --> 00:13:58,872 my kind of direction. So we literally walk 202 00:13:58,933 --> 00:14:02,537 into the unknown and that basically 203 00:14:05,057 --> 00:14:08,998 If you run a startup, you're going to want to pay attention to this. Startups, especially 204 00:14:09,118 --> 00:14:12,359 AI startups, need good cloud providers. Lucky for you, 205 00:14:12,559 --> 00:14:16,060 Google Cloud exists, and they offer an experience, the Google for Startups 206 00:14:16,100 --> 00:14:19,461 Cloud Program that's tailored specifically to you. The program 207 00:14:19,501 --> 00:14:22,902 offers eligible startups up to $200,000 in Google Cloud 208 00:14:22,942 --> 00:14:27,283 credits for up to two years. For AI startups, that number is $350,000. Plus, 209 00:14:27,423 --> 00:14:30,966 you get the mentorship and experience of Google Cloud's top-notch team. What 210 00:14:30,986 --> 00:14:35,170 are you waiting for? Start building today with Google Cloud at cloud.google.com slash 211 00:14:35,190 --> 00:14:39,233 startup. That's C-L-O-U-D dot G-O-O-G-L-E dot 212 00:14:44,877 --> 00:14:48,740 Yeah, it's quite a journey. Now, about AI specifically, 213 00:14:48,780 --> 00:14:52,403 right? And I know you've had an interesting journey with that. You started studying 214 00:14:52,443 --> 00:14:57,398 molecular biology, and you kind of jumped into machine learning. You've 215 00:14:57,438 --> 00:15:01,783 said of AI that it could usher in an age of empowerment. The 216 00:15:01,823 --> 00:15:05,404 printing press ushered in the Age of Enlightenment, and 217 00:15:05,564 --> 00:15:10,246 you've kind of equated the two as somewhat similar in how society-changing 218 00:15:10,786 --> 00:15:13,967 AI could be. What does that Age of Empowerment look like? And 219 00:15:14,247 --> 00:15:17,888 from where we're sitting now, what's the gap to get there? And 220 00:15:19,629 --> 00:15:24,070 I'm glad that you asked this question because, you know, many people would ask 221 00:15:24,130 --> 00:15:27,471 how to achieve artificial superintelligence and so on. 222 00:15:27,731 --> 00:15:32,016 And then what is that, right? So, yeah, I think I 223 00:15:32,076 --> 00:15:36,787 like to see any innovation and breakthrough from 224 00:15:36,807 --> 00:15:40,175 a utility standpoint, what 225 00:15:40,255 --> 00:15:44,175 function it delivers. and how such functionality 226 00:15:44,575 --> 00:15:48,076 change people's life, change society, and 227 00:15:48,136 --> 00:15:51,717 maybe even change civilization. So 228 00:15:51,777 --> 00:15:55,378 it is in that context I call AI 229 00:15:55,778 --> 00:15:59,339 to be equivalent to the 230 00:15:59,839 --> 00:16:03,880 breakthrough leading to the age of enlightenment, 231 00:16:04,040 --> 00:16:08,301 which is the printing press. There are many, many other inventions. The 232 00:16:08,361 --> 00:16:12,208 invention of the steam engine, for example, discovery of 233 00:16:12,509 --> 00:16:16,310 electricity, all these are huge. But I 234 00:16:16,370 --> 00:16:19,571 think AI and 235 00:16:19,651 --> 00:16:23,532 also printing press, all the invention of letters 236 00:16:23,632 --> 00:16:27,714 and writing are at a different level compared 237 00:16:27,734 --> 00:16:31,495 to all these technical innovations. Because if 238 00:16:31,535 --> 00:16:35,436 you look at, if you put AI 239 00:16:36,196 --> 00:16:39,380 and the printing press and maybe the 240 00:16:39,400 --> 00:16:42,562 invention of paper and writing next to each other, they have something in 241 00:16:42,662 --> 00:16:46,164 common. It is actually the backbone of human 242 00:16:46,544 --> 00:16:49,785 civilization in their experience with 243 00:16:49,925 --> 00:16:53,607 information. It's about how we 244 00:16:53,847 --> 00:16:57,029 process and store and use 245 00:16:57,049 --> 00:17:00,110 information in a 246 00:17:00,270 --> 00:17:03,552 massively different scope of 247 00:17:03,612 --> 00:17:08,030 magnitude. have 248 00:17:08,110 --> 00:17:11,914 the technology of writing, use letters and characters 249 00:17:11,954 --> 00:17:16,912 and papers. then 250 00:17:17,192 --> 00:17:20,653 the information can be recorded rather than through 251 00:17:20,953 --> 00:17:24,774 verbal communications. I 252 00:17:24,834 --> 00:17:28,855 remember, is that Plato? 253 00:17:30,756 --> 00:17:35,657 I think Plato is a student of Socrates. Socrates 254 00:17:35,837 --> 00:17:39,038 didn't actually believe that knowledge, one should write 255 00:17:39,078 --> 00:17:43,600 any book. He believed in in-person experience of 256 00:17:44,000 --> 00:17:47,142 teaching. But it is 257 00:17:48,083 --> 00:17:51,587 Plato who convinced him, or maybe who took 258 00:17:51,687 --> 00:17:55,712 on the job of taking down all Socrates' 259 00:17:55,752 --> 00:17:58,955 teaching and ideas into writing. And now we 260 00:17:58,995 --> 00:18:02,699 appreciate how important that is, because that knowledge can 261 00:18:02,899 --> 00:18:06,035 pass down to generations as it is. So in 262 00:18:06,075 --> 00:18:09,737 a sense, it is a embodiment of 263 00:18:10,117 --> 00:18:13,759 the antiquity time, information technology. But 264 00:18:14,579 --> 00:18:17,841 at that time, you need to copy the book. And the 265 00:18:17,881 --> 00:18:21,543 people who own the book need to be 266 00:18:21,643 --> 00:18:25,144 literate and also need to have that kind of status, have 267 00:18:25,224 --> 00:18:28,526 the opportunity to access to that. So knowledge belongs to a very few 268 00:18:28,566 --> 00:18:31,722 people. Printing press is 269 00:18:32,582 --> 00:18:35,743 a major advancement unlocking the knowledge from a few 270 00:18:35,803 --> 00:18:39,344 people to everyone because it allows every 271 00:18:39,384 --> 00:18:42,705 people to have their own copy of the book which began 272 00:18:42,745 --> 00:18:46,366 from the Bible and they can form their own interpretations and 273 00:18:46,466 --> 00:18:49,646 understanding and then after that people start to 274 00:18:49,706 --> 00:18:52,807 think basically based on that from those knowledges and 275 00:18:52,847 --> 00:18:56,308 then they write more books right so we have really 276 00:18:56,768 --> 00:18:59,849 even the university evolving from a theoretical school to you 277 00:18:59,869 --> 00:19:03,160 know study every topics Now, why 278 00:19:03,341 --> 00:19:07,248 AI is directly related 279 00:19:07,308 --> 00:19:10,694 to that train of evolution? If 280 00:19:10,734 --> 00:19:14,260 you look at large language model, for example, it is trained not 281 00:19:14,401 --> 00:19:17,591 on one book, It is not trained on 282 00:19:18,692 --> 00:19:22,132 one sector of knowledge. It is literally trained on 283 00:19:22,612 --> 00:19:26,053 everything ever written into language through 284 00:19:26,253 --> 00:19:30,154 civilization and also in any media. In 285 00:19:30,214 --> 00:19:34,275 a sense, you can imagine it is about turning all the 286 00:19:34,315 --> 00:19:37,916 libraries in the world into one 287 00:19:37,936 --> 00:19:41,076 device and then make it available next to your 288 00:19:41,096 --> 00:19:45,258 fingertip. You can basically now easily get knowledge and 289 00:19:45,358 --> 00:19:49,461 learn from it using this interface, which 290 00:19:49,621 --> 00:19:54,424 makes the knowledge broadcast 291 00:19:54,604 --> 00:19:57,766 and storage and accessibility to 292 00:19:57,806 --> 00:20:02,517 a different level. Then AI also not only retrieves 293 00:20:02,577 --> 00:20:05,679 knowledge for you, it actually started to be 294 00:20:05,739 --> 00:20:08,880 able to use the knowledge to solve problems, as we can see. They can 295 00:20:08,960 --> 00:20:12,441 at least answer many of your questions. And they can also 296 00:20:12,541 --> 00:20:16,403 start producing, for example, code 297 00:20:16,523 --> 00:20:20,064 and softwares. And nowadays, people 298 00:20:20,344 --> 00:20:23,566 are trying to develop a lot 299 00:20:23,626 --> 00:20:27,487 of applications in different sectors based on that, because 300 00:20:27,968 --> 00:20:32,265 knowledge can be actionated to solve problems. And 301 00:20:32,645 --> 00:20:36,006 in that sense, I call it an empowerment after 302 00:20:36,026 --> 00:20:39,308 enlightenment. It is not about just giving you the 303 00:20:39,348 --> 00:20:43,009 space to think. It is now giving you the tools to 304 00:20:43,509 --> 00:20:47,391 make you a more powerful enabler 305 00:20:47,491 --> 00:20:51,032 or actor in whatever occupation that you do. 306 00:20:51,192 --> 00:20:54,393 And then, in a sense, leave you out space to 307 00:20:54,433 --> 00:20:58,415 do maybe more challenging, intellectually 308 00:20:58,435 --> 00:21:01,946 challenging things. And so that empowerment comes 309 00:21:01,986 --> 00:21:05,588 together with even a next level of enlightenment 310 00:21:05,608 --> 00:21:08,990 where, you know, everybody can do 311 00:21:09,210 --> 00:21:12,611 very often time, you know, a very advanced level of 312 00:21:12,672 --> 00:21:16,193 work without necessarily spending 20 years of 313 00:21:16,233 --> 00:21:19,815 training, you know, in school. And then people should 314 00:21:19,875 --> 00:21:23,017 not start rethinking, you know, what I'm going to learn now, you know, 315 00:21:23,157 --> 00:21:26,398 in the in the next five years and what kind of problems I 316 00:21:26,559 --> 00:21:30,551 should tackle. So we can refocus on problems 317 00:21:31,432 --> 00:21:35,697 and maybe needs rather than just 318 00:21:36,037 --> 00:21:39,861 focusing on repeating, executing 319 00:21:44,239 --> 00:21:47,501 The difference to me between the large language models that 320 00:21:47,521 --> 00:21:51,124 we're kind of dealing with today, right, and something like the printing press is, 321 00:21:51,924 --> 00:21:55,026 you know, I think of all these technological innovations as levels of 322 00:21:55,086 --> 00:21:58,448 automation, right? The printing press, as you said, got, instead 323 00:21:58,468 --> 00:22:02,011 of having to handwrite books and pass them out, right, we could, we 324 00:22:02,031 --> 00:22:05,574 have copies of stuff. it created more accessibility. The 325 00:22:05,614 --> 00:22:09,118 jump to language models and 326 00:22:09,399 --> 00:22:12,782 the kind of generative AI we're seeing now, to me, what you're starting 327 00:22:12,803 --> 00:22:17,327 to do is you're starting to take the human out of the loop to a degree. Do 328 00:22:17,347 --> 00:22:21,508 you see some sort of risk in that environment where maybe 329 00:22:21,548 --> 00:22:25,569 for some people it might unlock them to explore greater problems? Is 330 00:22:25,589 --> 00:22:29,610 there a concern that for other people it just acts as a crutch for human 331 00:22:29,650 --> 00:22:33,251 intelligence, human creativity, because you have everything accessible, 332 00:22:35,372 --> 00:22:39,345 I don't deny that there is a risk. In fact, any new technology you 333 00:22:39,365 --> 00:22:42,608 know, comes together with both 334 00:22:42,668 --> 00:22:45,991 opportunities and risk. Just like, you know, even 335 00:22:46,031 --> 00:22:49,313 we talk about something that we are very familiar with, 336 00:22:49,353 --> 00:22:52,756 you know, medicine, for example, you know, there are always a side effect that 337 00:22:52,796 --> 00:22:55,879 we need to control. That's why FDA is formed. Genetic engineering, you 338 00:22:55,899 --> 00:22:59,562 know, we use them now to dramatically raise the productivity of 339 00:22:59,622 --> 00:23:03,285 crops. And we can engineer, you 340 00:23:03,305 --> 00:23:06,647 know, into the crops, you know, or the 341 00:23:06,707 --> 00:23:10,387 cattle's or the livestock's some nutritional deficiencies 342 00:23:10,467 --> 00:23:13,611 and so forth. But you can also use 343 00:23:13,651 --> 00:23:18,456 that technology to create very harmful viruses 344 00:23:19,197 --> 00:23:22,782 or other forms of toxicities. So 345 00:23:23,542 --> 00:23:26,806 it's really about how to mitigate and 346 00:23:26,966 --> 00:23:30,448 how to regulate those risks. I think AI 347 00:23:30,608 --> 00:23:33,870 is not an exception. But on the other hand, 348 00:23:34,230 --> 00:23:37,612 also not a unique 349 00:23:38,032 --> 00:23:41,614 threat that we didn't deal with before. In 350 00:23:42,515 --> 00:23:47,277 the past, people figured out that there is a clear distinction between 351 00:23:47,898 --> 00:23:51,920 the science, the technology, and the product, right? 352 00:23:52,343 --> 00:23:55,846 And where you set the brick and 353 00:23:55,926 --> 00:23:59,489 where you regulate is well 354 00:23:59,509 --> 00:24:03,192 understood, that you probably should focus more 355 00:24:03,632 --> 00:24:08,276 on regulating the users, on regulating the distributions, but 356 00:24:08,676 --> 00:24:11,819 on the intellectual part, which is the science, and 357 00:24:11,859 --> 00:24:15,702 maybe even the technology, which is the production part, you use 358 00:24:15,842 --> 00:24:20,512 a different type of management. Typically, 359 00:24:20,912 --> 00:24:24,695 for the science part, you don't even manage, or you'd manage very little, because 360 00:24:25,395 --> 00:24:29,779 the creativity, once you shut it down, the cost 361 00:24:29,819 --> 00:24:33,442 is too big to bear. In fact, many of the assumptions 362 00:24:33,682 --> 00:24:36,965 about the current rush into AI 363 00:24:37,005 --> 00:24:40,743 regulation discussion and even panic is 364 00:24:40,903 --> 00:24:45,344 also based on an unclear understanding 365 00:24:45,404 --> 00:24:49,965 about really the power of AI and how powerful 366 00:24:50,085 --> 00:24:53,606 and how potentially dangerous it can be. It 367 00:24:53,666 --> 00:24:57,166 is not well understood, especially by the general public. 368 00:24:58,146 --> 00:25:01,747 Sometimes, even from a tactical standpoint, even 369 00:25:01,947 --> 00:25:05,068 if there is a risk, how material that 370 00:25:05,108 --> 00:25:09,329 risk is, and it actually had a big impact 371 00:25:09,469 --> 00:25:12,812 on the risk and benefit calculations. So 372 00:25:12,832 --> 00:25:16,155 I want to say a few words about really where we are in AI in 373 00:25:16,195 --> 00:25:19,578 terms of its danger and its risk. Yeah, AI 374 00:25:19,698 --> 00:25:23,341 is more than automation for sure, because AI can 375 00:25:24,362 --> 00:25:27,825 automatically learn skills from training 376 00:25:27,845 --> 00:25:32,171 materials. So it is a one-step beyond 377 00:25:32,631 --> 00:25:36,575 just a programmed automation. 378 00:25:37,155 --> 00:25:40,799 It does have the ability to acquire knowledge and to 379 00:25:41,860 --> 00:25:45,563 become more and more powerful once 380 00:25:45,583 --> 00:25:49,106 you plug them into some data sources and 381 00:25:49,206 --> 00:25:53,637 also run an algorithm. But on the other hand, AI 382 00:25:53,697 --> 00:25:56,799 also has a lot of limitations. It 383 00:25:56,879 --> 00:26:02,724 is not a living creature that has self-identity, 384 00:26:03,504 --> 00:26:07,007 self-drive, agency, free 385 00:26:07,067 --> 00:26:10,509 will, and so forth. In fact, it is unclear to me at least where 386 00:26:10,630 --> 00:26:14,092 all those will come from based on the current 387 00:26:14,212 --> 00:26:18,135 architecture and our current mathematics of doing things. I 388 00:26:18,175 --> 00:26:22,500 think down the road, new architectures beyond transformers, new 389 00:26:22,580 --> 00:26:26,284 form of data beyond text and image will be feeding into 390 00:26:26,424 --> 00:26:29,888 the next generation systems to create more 391 00:26:29,948 --> 00:26:33,772 functions. But these functions actually are still quite 392 00:26:33,812 --> 00:26:37,376 a few steps away from the type of higher 393 00:26:37,416 --> 00:26:40,638 level intelligence or human-like intelligence I 394 00:26:40,658 --> 00:26:43,940 just mentioned, like an agency, like identity, like a will, and 395 00:26:44,040 --> 00:26:47,502 so forth. For example, how to even do more 396 00:26:47,542 --> 00:26:50,844 complex reasoning, multiple steps reasoning. You want to 397 00:26:52,165 --> 00:26:55,547 plan a shopping trip, you know, a travel, and 398 00:26:55,847 --> 00:26:59,009 purchase a flight ticket, you know, 399 00:26:59,049 --> 00:27:02,150 a plan on hotels. That requires multiple steps of 400 00:27:02,190 --> 00:27:05,571 reasoning. I don't think right now the current 401 00:27:05,612 --> 00:27:09,715 version of GPTs is able to do well on that job. Now, 402 00:27:10,316 --> 00:27:14,100 talking maybe a slightly more challenging task. How to run 403 00:27:14,120 --> 00:27:17,703 a store? How to run a company? How 404 00:27:17,763 --> 00:27:21,206 to win a campaign, a battle? 405 00:27:21,907 --> 00:27:24,990 These are reasonings which are typically a 406 00:27:25,030 --> 00:27:28,910 hundred steps. conditioning on real-world 407 00:27:28,950 --> 00:27:33,314 environments, and also conditioning on 408 00:27:33,755 --> 00:27:38,059 embodied experiences from sensory and from anything. And 409 00:27:38,099 --> 00:27:44,084 these are very difficult. In fact, you can probably already find 410 00:27:44,144 --> 00:27:47,347 out that you cannot ask, for example, a GPT or AI 411 00:27:47,407 --> 00:27:50,550 assistant to teach you how to swim, to teach you how to play the 412 00:27:50,570 --> 00:27:54,353 music. Because why even different 413 00:27:54,673 --> 00:27:58,235 music mentors, why 414 00:27:58,255 --> 00:28:01,736 we have master classes? You know, those musicians, when they teach you, 415 00:28:02,116 --> 00:28:05,557 it's not about reading from a book and say, hey, you apply 416 00:28:05,857 --> 00:28:09,339 two grams of force onto a key and hold for 417 00:28:09,679 --> 00:28:13,260 this many milliseconds and to produce a sound. It's 418 00:28:13,300 --> 00:28:16,903 a very different experience of music teaching. And 419 00:28:16,923 --> 00:28:20,686 I don't think right now our AI system is 420 00:28:20,786 --> 00:28:24,309 even designed to do that. And certainly they cannot do that. So 421 00:28:24,789 --> 00:28:28,092 I think the current AI system is still 422 00:28:28,192 --> 00:28:32,776 on the steep learning curve to gain more capability, whether 423 00:28:32,936 --> 00:28:36,518 it has the risk that people 424 00:28:36,979 --> 00:28:40,621 are now, you know, fear 425 00:28:40,701 --> 00:28:44,825 about. I think it's premature to say 426 00:28:44,865 --> 00:28:49,548 that. Now people say disinformation, deepfake, 427 00:28:50,529 --> 00:28:54,252 and all that, these are risks 428 00:28:55,688 --> 00:28:58,910 not coming from AI technology. It is coming from 429 00:28:59,010 --> 00:29:02,253 the users of the technology. It's just like a 430 00:29:02,313 --> 00:29:05,715 gun. You definitely can cause a lot of harm by using 431 00:29:05,735 --> 00:29:08,977 the gun. But is that the mistake of the gun? 432 00:29:09,597 --> 00:29:12,740 People are still making better guns these days, and 433 00:29:12,860 --> 00:29:16,462 for whatever specific purposes. But how 434 00:29:16,502 --> 00:29:19,724 to regulate its usage is the key. So I 435 00:29:19,784 --> 00:29:22,966 would like to remind policymakers and the 436 00:29:23,006 --> 00:29:26,177 general public to really be 437 00:29:26,237 --> 00:29:29,523 a little bit more concise and 438 00:29:29,563 --> 00:29:32,948 careful about where the focus is 439 00:29:33,129 --> 00:29:36,795 in such AI risk and AI regulation conversations. 440 00:29:37,208 --> 00:29:40,989 The AI regulation conversation is very broad, right? 441 00:29:41,849 --> 00:29:45,310 There are some instances that are trying to be very focused, such 442 00:29:45,390 --> 00:29:48,770 as with deepfakes and job loss related to 443 00:29:48,810 --> 00:29:51,931 deepfakes and the creative industries. There are others that 444 00:29:51,971 --> 00:29:56,412 are trying to mitigate and prepare for these hypothetical risks 445 00:29:56,492 --> 00:30:00,192 that we don't really know if they'll happen at all. And 446 00:30:03,113 --> 00:30:06,658 And I want to add one more thing. In fact, even if you want to regulate 447 00:30:06,778 --> 00:30:10,636 and counter that, you actually need AI to do it. Because like 448 00:30:10,676 --> 00:30:13,798 the deepfake images and 449 00:30:13,818 --> 00:30:17,040 the videos, the misinformation, it's just 450 00:30:17,100 --> 00:30:20,222 too much for people or for a 451 00:30:21,162 --> 00:30:24,985 manual devices or approach to discover 452 00:30:25,105 --> 00:30:28,287 and to contain all of them. Unless you shut down the 453 00:30:28,307 --> 00:30:32,109 computer, every computer in the world, they will be generated. And 454 00:30:35,485 --> 00:30:38,986 As you just said, the problem with this technology and 455 00:30:39,106 --> 00:30:42,727 regulating this technology is it requires 456 00:30:42,767 --> 00:30:46,428 a global effort. If one country regulates really 457 00:30:46,468 --> 00:30:50,929 hard, that's not going to stop China or other countries from 458 00:30:51,569 --> 00:30:54,850 not regulating at all. And I think there's some concern about that. 459 00:30:56,303 --> 00:31:00,027 the way you kind of framed the regulation question was to focus 460 00:31:03,030 --> 00:31:06,434 Exactly. Yeah. That's why I think, for example, you 461 00:31:06,454 --> 00:31:09,917 said about adversarial scenarios at certain countries 462 00:31:09,957 --> 00:31:13,559 like China, will be doing what they need to do anyway, 463 00:31:14,259 --> 00:31:19,201 and you cannot stop them. I think the best thing 464 00:31:19,241 --> 00:31:23,323 that one can do is to advance 465 00:31:23,363 --> 00:31:26,904 your own technology, in fact, to make AI even better. 466 00:31:28,364 --> 00:31:31,846 Even we are under a material risk, we have 467 00:31:31,986 --> 00:31:35,247 the means to take care of that, to counter that. 468 00:31:35,647 --> 00:31:40,512 And do you think it would be an achievable thing? There 469 00:31:40,532 --> 00:31:44,015 has been talk about the idea of international government, international 470 00:31:44,055 --> 00:31:47,658 governance of AI that kind of supersedes any 471 00:31:47,919 --> 00:31:51,482 individual efforts. You know, in the U.S. it's even more fragmented because 472 00:31:51,502 --> 00:31:54,965 you have states doing things and then the federal government is trying to figure stuff 473 00:31:55,105 --> 00:31:58,976 out. the EU is moving, and that'll impact where 474 00:31:59,316 --> 00:32:02,857 companies deploy. But the idea of getting 475 00:32:03,017 --> 00:32:06,598 every developer around the world from universities to 476 00:32:06,638 --> 00:32:10,760 corporations to agree to a set of guidelines that maybe 477 00:32:15,101 --> 00:32:18,663 Should it be feasible? I'm actually optimistic. It's absolutely 478 00:32:18,723 --> 00:32:22,165 achievable. But on the other hand, I have to also 479 00:32:22,225 --> 00:32:26,208 point out that if you look at the age of Enlightenment after 480 00:32:26,228 --> 00:32:30,691 the printing press, it took roughly 100 years. before 481 00:32:31,632 --> 00:32:35,415 all the warring states to eventually align 482 00:32:35,595 --> 00:32:39,258 on some common principles and also to learn 483 00:32:39,318 --> 00:32:44,302 how to live with this technology. In 484 00:32:44,342 --> 00:32:48,125 West Europe, different religious sectors 485 00:32:49,426 --> 00:32:52,929 fought to death just because a 486 00:32:53,189 --> 00:32:57,200 different interpretation of a Bible principle. you 487 00:32:57,220 --> 00:33:02,143 know, after all the people learn about it and have their own interpretations. But 488 00:33:02,223 --> 00:33:06,206 eventually, you know, they learn how to settle things. And 489 00:33:06,286 --> 00:33:10,569 also they learn how to live with a new society that 490 00:33:10,869 --> 00:33:15,071 all can read and have many books and so forth. And 491 00:33:15,332 --> 00:33:18,734 I can also name some counter examples where printing press 492 00:33:18,794 --> 00:33:22,336 was actually banned. And after some hundred years, 493 00:33:23,156 --> 00:33:26,469 those regions become much, much less developed. than 494 00:33:26,750 --> 00:33:29,985 the West European countries. Why I mention this? 495 00:33:31,032 --> 00:33:35,295 because I want to emphasize the importance of experimenting 496 00:33:35,355 --> 00:33:39,117 and also the time. We are just barely five 497 00:33:39,157 --> 00:33:42,299 years or maybe even three years into the age of 498 00:33:42,339 --> 00:33:45,601 large-language models, right? So the AI era 499 00:33:45,621 --> 00:33:49,223 has just started, and it is also evolving very fast. 500 00:33:49,603 --> 00:33:53,826 We actually don't even see, you know, where it will platoon and 501 00:33:54,026 --> 00:33:57,208 how a steady state would look like. And that's why I 502 00:33:57,248 --> 00:34:00,869 think at this point, rush 503 00:34:00,969 --> 00:34:04,411 into a premature set 504 00:34:04,471 --> 00:34:08,574 of very stiff laws or regulations isn't 505 00:34:08,614 --> 00:34:12,376 going to be very helpful because it may cause 506 00:34:12,696 --> 00:34:16,278 missed opportunities and also maybe 507 00:34:16,458 --> 00:34:20,040 an awkward misfit of all these 508 00:34:20,841 --> 00:34:24,580 policies. with 509 00:34:24,620 --> 00:34:27,965 respect to the real world, how they evolve. So I would say that 510 00:34:28,245 --> 00:34:32,331 first, from a scientific standpoint, 511 00:34:32,391 --> 00:34:36,316 we need to collect enough data points and 512 00:34:36,696 --> 00:34:40,228 make careful observations and analysis. about 513 00:34:40,248 --> 00:34:43,752 the opportunity and the risk, and 514 00:34:43,932 --> 00:34:48,356 also the trajectory of the technological development. Thinking 515 00:34:48,436 --> 00:34:52,040 ahead of the curve and be able to predict where 516 00:34:52,060 --> 00:34:55,223 that technology is evolving, where I should put a 517 00:34:56,664 --> 00:35:00,548 gate or a checkpoint down the road, rather 518 00:35:00,568 --> 00:35:03,999 than passively you know, chase, you 519 00:35:04,040 --> 00:35:08,506 know, a running kind of objects 520 00:35:09,548 --> 00:35:12,993 and interfere with some of the, you know, 521 00:35:13,133 --> 00:35:16,879 very organic and natural and creative and productive kind 522 00:35:16,919 --> 00:35:20,282 of activities of scientific 523 00:35:20,302 --> 00:35:23,384 advancements. The trick is just trying to figure out where it's going to go so 524 00:35:23,405 --> 00:35:26,627 you can get ahead of it. Yeah, and it will take some time. But of 525 00:35:26,687 --> 00:35:30,551 course, I'm optimistic that it wouldn't take 100 years. It wouldn't be as terrible 526 00:35:30,831 --> 00:35:34,314 and as costly as the religious wars 527 00:35:40,598 --> 00:35:43,739 Now, you talked a minute ago about we don't want to miss the 528 00:35:43,759 --> 00:35:47,141 opportunity, right? And we've talked 529 00:35:47,181 --> 00:35:50,902 about how impactful this technology can be. I've 530 00:35:50,942 --> 00:35:54,203 been obsessed with this idea of trying to narrow down on 531 00:35:54,243 --> 00:35:57,524 what the true promise of AI is, because like you said, I think a lot 532 00:35:57,545 --> 00:36:00,706 of people misunderstand what the technology is itself. Part 533 00:36:00,726 --> 00:36:03,907 of that is the fault of the kind of hype of corporations that 534 00:36:03,927 --> 00:36:07,128 we've been seeing and social media is not helping. Part 535 00:36:07,148 --> 00:36:10,590 of that is the fact that for a lot of subject 536 00:36:10,630 --> 00:36:13,992 areas within AI, we don't have unified definitions, or people say 537 00:36:14,072 --> 00:36:17,595 one thing when they're talking about something else. But 538 00:36:17,975 --> 00:36:21,917 trying to narrow in on the idea of what is the promise of this technology and 539 00:36:21,997 --> 00:36:25,219 how it can help people, because I think in 540 00:36:25,259 --> 00:36:29,312 a lot of cases, people are afraid. People 541 00:36:29,352 --> 00:36:32,976 are afraid about losing their jobs. People are afraid about existential 542 00:36:33,096 --> 00:36:36,540 risk scenarios, which get kind of thrust 543 00:36:36,600 --> 00:36:40,104 upon them from some of the companies developing this 544 00:36:40,144 --> 00:36:43,688 technology. And it's set up as something that seems 545 00:36:43,768 --> 00:36:46,951 to have so many, it's set up as something that's 546 00:36:46,991 --> 00:36:50,334 dual use, but it seems to have so many negative risks that 547 00:36:50,354 --> 00:36:53,937 I think for a lot of people, there is a question of why 548 00:36:53,997 --> 00:36:57,460 develop this at all, if the risks can range across 549 00:36:57,520 --> 00:37:00,763 this realm of severity. So to you, 550 00:37:04,835 --> 00:37:08,498 Yeah, that has a close connection to really the 551 00:37:09,658 --> 00:37:13,261 level of functionality that the AI will 552 00:37:13,281 --> 00:37:16,563 be achieving step by step. And at 553 00:37:16,703 --> 00:37:22,067 every functionality, you can accordingly kind 554 00:37:22,087 --> 00:37:25,570 of define the cost and also 555 00:37:25,870 --> 00:37:29,412 the opportunity. For example, Henry 556 00:37:29,432 --> 00:37:32,895 Ford used to say that every American need to own a car. And 557 00:37:33,296 --> 00:37:37,000 I think it's very bad news to the train 558 00:37:37,020 --> 00:37:40,344 drivers and maybe to the carriage drivers. But 559 00:37:40,884 --> 00:37:44,108 look, if every American needs to own a car, the number of 560 00:37:44,168 --> 00:37:47,411 workers to produce those cars will be humongous. In fact, it 561 00:37:47,491 --> 00:37:51,194 ends up creating a job. Now imagine that in our household, 562 00:37:51,615 --> 00:37:54,876 we want to have a robotic maid to just take 563 00:37:54,916 --> 00:37:58,538 care of all the household 564 00:37:58,599 --> 00:38:03,501 businesses. Wouldn't that be an amazing product 565 00:38:03,721 --> 00:38:07,123 that everybody wouldn't object to? I don't think right now people object to 566 00:38:07,824 --> 00:38:11,026 washing machines and so on. So that's a 567 00:38:11,086 --> 00:38:14,588 huge market. And the key technology to 568 00:38:14,668 --> 00:38:17,930 get there is AI. You need to basically have a 569 00:38:18,270 --> 00:38:21,845 robot that is able to plan, able to act, 570 00:38:22,746 --> 00:38:26,329 be able to understand, communicate with you, and also 571 00:38:27,830 --> 00:38:31,794 get things done. So we are actually, honestly, many 572 00:38:31,834 --> 00:38:35,298 years away even from achieving that. But in achieving 573 00:38:35,338 --> 00:38:38,421 that, I would imagine that everybody wants to be in 574 00:38:38,441 --> 00:38:41,784 that business and join the production line and 575 00:38:41,924 --> 00:38:45,291 make a lot of money out of it. Yeah, you will be losing, you know, of course, 576 00:38:45,311 --> 00:38:48,334 you know, the clean ladies may be losing their job, but they will get a 577 00:38:48,354 --> 00:38:51,436 better job elsewhere, because you will be joining the 578 00:38:51,616 --> 00:38:55,159 production line in different steps of producing that amazing machine. 579 00:38:55,600 --> 00:38:59,983 And also, you know, we talk about AI being 580 00:39:02,225 --> 00:39:05,748 replacing some even white collar fancy jobs, like 581 00:39:06,028 --> 00:39:10,211 a accountant, like a legal 582 00:39:10,251 --> 00:39:13,714 worker. because we see HHBT are 583 00:39:13,754 --> 00:39:19,918 passing the bar examinations, passing the medical examinations. Wouldn't 584 00:39:20,098 --> 00:39:24,281 it be nice that it freed up these very smart people to 585 00:39:24,581 --> 00:39:28,084 think about new practices in 586 00:39:28,124 --> 00:39:32,827 those domains? So I think Sam 587 00:39:32,907 --> 00:39:36,070 Altman, from his business standpoint, defined a 588 00:39:36,110 --> 00:39:39,780 few tiers of AI. I 589 00:39:39,820 --> 00:39:43,965 don't want to comment on whether those steps are too ambitious or 590 00:39:45,206 --> 00:39:48,990 maybe less realistic. But 591 00:39:49,591 --> 00:39:53,595 I think the way of thinking is very clear. It is aiming 592 00:39:53,655 --> 00:39:57,980 for higher and higher functionality. And what 593 00:39:58,280 --> 00:40:01,555 I can say from a technological standpoint, It 594 00:40:01,775 --> 00:40:05,276 also will be dependent on 595 00:40:05,937 --> 00:40:09,838 multiple generations of data, architecture, 596 00:40:10,338 --> 00:40:13,600 and algorithmic evolutions. We're not there. 597 00:40:13,840 --> 00:40:17,241 In fact, we don't have the tool right now to achieve even his second and third goals. 598 00:40:17,781 --> 00:40:21,282 And within these second, third, and fifth goals, in fact, I can define 599 00:40:21,562 --> 00:40:24,884 multiple concrete reasoning and 600 00:40:25,064 --> 00:40:29,159 also acting tasks that requires 601 00:40:29,179 --> 00:40:32,508 a lot of innovations. So at this university, in 602 00:40:32,649 --> 00:40:36,699 fact, we were involved in some of this work, such as developing the 603 00:40:36,779 --> 00:40:40,262 post-LLM models, including the word models, 604 00:40:40,342 --> 00:40:43,625 that can even do a trivial thing of 605 00:40:43,825 --> 00:40:47,088 simulating the word. If you cannot simulate the word, how 606 00:40:47,108 --> 00:40:50,270 can you even teach a guy to pour a glass of water? That's actually not 607 00:40:50,310 --> 00:40:53,633 written in the book. There are many other things that require 608 00:40:53,653 --> 00:40:57,596 you to be able to simulate real-world experiences, and also generating 609 00:40:57,616 --> 00:41:00,799 the next word based on your action on the real world, and 610 00:41:00,819 --> 00:41:04,421 so on and so forth. And only in that way, you 611 00:41:04,441 --> 00:41:07,865 can start to do thought experiments. You can do a new form of reasoning by 612 00:41:07,925 --> 00:41:10,989 simulation, which could overcome some of 613 00:41:11,029 --> 00:41:14,187 the limitations of the current language models. Then once you 614 00:41:14,227 --> 00:41:17,568 have the word model, you can go one step further into 615 00:41:17,628 --> 00:41:20,969 an agent model, where he has the word model as the brain, 616 00:41:21,309 --> 00:41:24,970 but also he needs to have other mental devices, such 617 00:41:25,150 --> 00:41:28,711 as perception, and 618 00:41:28,991 --> 00:41:32,192 goal, and belief, and all these sort 619 00:41:32,212 --> 00:41:35,653 of things. Each of them requires another 620 00:41:35,893 --> 00:41:39,313 model architectures. So I can go on and speak maybe 621 00:41:39,333 --> 00:41:42,934 for an hour about the technical steps that is required, even 622 00:41:42,974 --> 00:41:47,235 to achieve what we probably can see as a very normal 623 00:41:47,395 --> 00:41:51,056 and tedious level of intelligence. Then 624 00:41:51,136 --> 00:41:54,657 I want to also say a few words about the other opportunities that is beyond 625 00:41:54,797 --> 00:41:58,516 our daily life. which is the AI for science. Doing 626 00:41:58,536 --> 00:42:02,061 scientific research is always a function of a technology environment. 627 00:42:02,322 --> 00:42:06,468 Science requires more than brain, it requires the tool. Microscope, 628 00:42:06,648 --> 00:42:10,274 telescope actually play the bigger role in advancing biology and 629 00:42:10,294 --> 00:42:14,057 physics. AI is actually one such tool. Right now 630 00:42:14,458 --> 00:42:17,781 we have, for example, in the space of medicine and 631 00:42:17,801 --> 00:42:21,604 biology, there's a huge amount of genomic data, proteomic 632 00:42:21,644 --> 00:42:25,447 data, and all sorts of orbit data that 633 00:42:25,547 --> 00:42:28,710 is not meant even for people to analyze because they are too big and 634 00:42:28,750 --> 00:42:32,430 too complex. Before the modern AI 635 00:42:32,731 --> 00:42:36,075 of large language models and foundation models, we 636 00:42:36,856 --> 00:42:39,980 can do very little in understanding and utilizing the 637 00:42:40,040 --> 00:42:43,424 data. And now there is a new 638 00:42:45,306 --> 00:42:48,991 movement, in fact, in computational biology of building foundation 639 00:42:49,011 --> 00:42:52,634 models to distill in an unprecedented 640 00:42:52,654 --> 00:42:56,858 way, you know, new form of knowledges and also predictability and 641 00:42:56,978 --> 00:43:00,721 the simulation or capability of biological 642 00:43:00,841 --> 00:43:04,124 systems. Imagine that. What if you can now 643 00:43:04,304 --> 00:43:07,747 use an AI model to make proposals in 644 00:43:07,787 --> 00:43:10,990 a very, very short amount of time of hundreds, if not thousands, of 645 00:43:11,130 --> 00:43:14,253 drug design that can be applied to cure a 646 00:43:14,273 --> 00:43:17,594 particular disease. And then you don't have to even apply the 647 00:43:17,614 --> 00:43:21,296 drug to a real biological system 648 00:43:21,336 --> 00:43:24,738 like a cell or a model organism. You can actually 649 00:43:24,878 --> 00:43:28,140 also do the same in a word model of 650 00:43:28,160 --> 00:43:32,082 the biology, which simulates all biological 651 00:43:32,202 --> 00:43:35,444 phenomenons and the responses up to perturbation. You 652 00:43:35,464 --> 00:43:39,226 can pretty much do this whole loop of innovation and 653 00:43:39,326 --> 00:43:42,928 design and experimentation on a computer. 654 00:43:43,642 --> 00:43:47,503 And that actually is now becoming a very, very exciting 655 00:43:47,563 --> 00:43:50,704 goal for many of my colleagues, including students and 656 00:43:50,784 --> 00:43:54,065 the professors in this university. Likewise, you can 657 00:43:54,985 --> 00:43:58,826 draw the same experiences or expand the same experience in physics, in 658 00:43:59,306 --> 00:44:02,927 agriculture, in many other sciences. So I think AI will 659 00:44:02,967 --> 00:44:06,628 become, again, the modern day new microscope, which 660 00:44:07,268 --> 00:44:10,829 tells you how to, helps you how to understand the complex data, 661 00:44:11,069 --> 00:44:14,913 just like the classical microscope did. which is also about helping 662 00:44:14,933 --> 00:44:18,195 you to understand information that you cannot see too small or 663 00:44:19,816 --> 00:44:23,558 So you were talking about limitations of the current architecture. And 664 00:44:23,878 --> 00:44:27,160 something else you mentioned that I want to tie into this as well is 665 00:44:27,200 --> 00:44:32,103 that you refer to AI as our 666 00:44:32,163 --> 00:44:35,384 new age microscope in the ways it'll 667 00:44:35,444 --> 00:44:38,811 advance science. and not 668 00:44:38,951 --> 00:44:42,314 as a species, right? And I think a lot of the source 669 00:44:42,354 --> 00:44:46,157 of where the fear is coming from and where the hype is coming from, and I 670 00:44:46,177 --> 00:44:50,060 think those two are very interrelated, is this language around 671 00:44:50,080 --> 00:44:53,143 AI. And some researchers have taken issue with the 672 00:44:53,183 --> 00:44:56,440 term AI at all because of this. as 673 00:44:56,621 --> 00:45:00,204 the creation of a new species, as a species to supplant the 674 00:45:00,244 --> 00:45:03,728 human species, as the kind of next thing. 675 00:45:03,768 --> 00:45:07,392 And we are creating this being thing, replicable, 676 00:45:07,432 --> 00:45:11,997 whatever that is. And it's vague if you try to pin it down, but the 677 00:45:12,037 --> 00:45:15,100 way you're talking about it, with the limitations and the 678 00:45:15,160 --> 00:45:18,360 focus on it as a specific tool, It's 679 00:45:18,460 --> 00:45:22,081 very different. And so, you know, one, I 680 00:45:22,121 --> 00:45:25,262 want to start with what those limitations are. You know, 681 00:45:25,302 --> 00:45:29,143 we talk about hallucination, we talk about biases and 682 00:45:29,183 --> 00:45:32,504 these kinds of inherent issues to the architecture. What's also 683 00:45:32,604 --> 00:45:36,045 the importance of the language and how 684 00:45:37,942 --> 00:45:41,224 Whether AI is a species and so on, these kinds of 685 00:45:41,264 --> 00:45:44,865 discussions rarely happen among scientists 686 00:45:44,946 --> 00:45:51,049 and engineers. I call them the consequence 687 00:45:51,189 --> 00:45:55,371 of people living in different dimensions. Imagine 688 00:45:55,431 --> 00:45:59,253 that there is a tribe somewhere in the Amazon jungle that 689 00:45:59,273 --> 00:46:02,874 has never seen an aircraft or 690 00:46:02,954 --> 00:46:06,207 maybe a car. and then all of a sudden they see 691 00:46:06,227 --> 00:46:10,557 a person landing from 692 00:46:10,577 --> 00:46:14,642 an aircraft. It is not it's 693 00:46:14,662 --> 00:46:18,443 going to be surprising that they take that as a new species. In 694 00:46:18,463 --> 00:46:23,303 fact, if we look at the ancient mosaics 695 00:46:23,423 --> 00:46:27,604 and the pictures and sculptures in Egypt 696 00:46:27,704 --> 00:46:31,265 and in other ancient civilizations, you actually see 697 00:46:31,505 --> 00:46:35,025 a lot of such depictions of a god-like 698 00:46:35,085 --> 00:46:39,486 figure or magical species, which, 699 00:46:39,606 --> 00:46:42,847 I don't know, may actually represent a machinery. That's very 700 00:46:42,907 --> 00:46:46,828 possible. So there is a key 701 00:46:47,868 --> 00:46:50,889 distinction between technical and a 702 00:46:50,929 --> 00:46:54,931 scientific discussion from a maybe 703 00:46:55,251 --> 00:46:59,112 romantic and humanity discussions. For 704 00:46:59,132 --> 00:47:03,294 example, this very word of intelligence. Intelligence is 705 00:47:03,494 --> 00:47:06,995 not a very scientific design. I regret actually our 706 00:47:07,355 --> 00:47:10,623 discipline is called artificial intelligence. because the other 707 00:47:10,723 --> 00:47:14,645 fields are called physics and biology. It's a noun. Intelligence 708 00:47:15,385 --> 00:47:18,606 is a noun, but it is derived from an adjunct. It's, 709 00:47:19,306 --> 00:47:23,448 you are intelligent, and so on. It's a very subjective thing. For 710 00:47:23,468 --> 00:47:27,009 example, imagine that I have in my hand a 711 00:47:27,089 --> 00:47:30,570 calculator, but I didn't tell you. In fact, 712 00:47:31,491 --> 00:47:35,892 especially for people who didn't know the existence of calculator, he would be amazed how 713 00:47:36,712 --> 00:47:40,292 fast and how well I can do not just the trivial 714 00:47:40,312 --> 00:47:44,173 calculations, but also square roots and the integrations and everything. I 715 00:47:44,193 --> 00:47:48,015 can be called very intelligent, but it is a calculator. Once 716 00:47:48,055 --> 00:47:51,537 you are in the same dimension, you see how it is made and 717 00:47:51,737 --> 00:47:55,339 how it is supposed to function, you would be less impressed 718 00:47:55,559 --> 00:47:59,015 and nervous. For example, large language 719 00:47:59,055 --> 00:48:02,276 model. We built it. In fact, we are building it here as 720 00:48:02,336 --> 00:48:05,637 well in MBCI. We are among the few universities, if not the only one in the world, 721 00:48:05,957 --> 00:48:09,898 who actually pre-train large language model of our own. It 722 00:48:09,998 --> 00:48:13,298 is a model. It is a language model doing 723 00:48:13,418 --> 00:48:16,859 one thing, which is the next world prediction. So 724 00:48:17,079 --> 00:48:20,780 in that function, many of the utilities somehow 725 00:48:21,380 --> 00:48:24,543 can be encapsulated and dispatched. For 726 00:48:24,584 --> 00:48:27,907 example, answering all your questions in 727 00:48:27,927 --> 00:48:31,090 a certain scope. Next World 728 00:48:31,110 --> 00:48:34,975 Prediction is like a linguistic simulator of 729 00:48:35,095 --> 00:48:38,558 simulating how a person would 730 00:48:38,659 --> 00:48:42,222 react to a prompt or a question through language. 731 00:48:43,118 --> 00:48:46,139 It is literally a definable and a 732 00:48:46,219 --> 00:48:49,680 measurable function. But human beings may 733 00:48:50,820 --> 00:48:54,081 respond in a certain way. In fact, even between different human 734 00:48:54,121 --> 00:48:57,401 beings, we do different. Some people recite some text. They 735 00:48:57,441 --> 00:49:00,662 need to have a script to read from it. Some other 736 00:49:00,702 --> 00:49:04,903 people improvise this. Some people are half-half. So 737 00:49:05,383 --> 00:49:08,684 the approach of achieving that to a 738 00:49:08,744 --> 00:49:11,805 receiver is actually less relevant. They draw on 739 00:49:11,825 --> 00:49:15,308 their experiences. Now you sit on the end of the language 740 00:49:15,368 --> 00:49:18,691 model, you see all your difficult questions, your difficult math 741 00:49:19,032 --> 00:49:22,675 being answered. It doesn't prove that the machine actually understands 742 00:49:22,756 --> 00:49:25,959 it, you know, or really, really digests it, or do it 743 00:49:25,979 --> 00:49:29,363 the way a human would do. They can do it in a very different 744 00:49:29,423 --> 00:49:33,227 way, and that way, by now, happens to be next-world prediction. 745 00:49:34,370 --> 00:49:38,052 So that's why among scientists, among engineers, we 746 00:49:38,133 --> 00:49:41,935 tend to be less amazed by this technology. And 747 00:49:42,015 --> 00:49:45,277 we knew it is going to, for example, I'm not surprised at all that in one 748 00:49:45,317 --> 00:49:48,699 day, in February soon, that a large-language model could 749 00:49:49,199 --> 00:49:52,521 get a gold medal in the International Math Olympiad. Because 750 00:49:52,761 --> 00:49:56,424 that type of reasoning, that type of problem-solving is 751 00:49:56,784 --> 00:49:59,906 doable through a natural prediction model. As long as 752 00:49:59,966 --> 00:50:03,593 you have a big enough brain to remember zillions 753 00:50:03,653 --> 00:50:07,624 of words. you know, a prefix, 754 00:50:07,884 --> 00:50:12,026 the sentence, and then the same amount next to that. But 755 00:50:12,066 --> 00:50:15,367 I also know that there are some questions that they cannot answer, even through a 756 00:50:15,447 --> 00:50:19,009 very, very simple interface. I've tried that. I reserved that 757 00:50:19,449 --> 00:50:22,851 question list. If you are interested, I can send it to you. But I tested even 758 00:50:22,911 --> 00:50:27,413 the latest strawberry. They still cannot answer. But 759 00:50:27,553 --> 00:50:31,134 to a general public, they need education, I think. Right now, the 760 00:50:31,294 --> 00:50:34,881 AI literacy is very, very different from 761 00:50:35,342 --> 00:50:38,807 how much we know about algebra. Some people are very poor at algebra, but 762 00:50:38,827 --> 00:50:41,972 they know what that is. They wouldn't be surprised to 763 00:50:42,052 --> 00:50:46,203 see a calculator. But on the other hand, AI 764 00:50:46,303 --> 00:50:50,086 literacy is much worse right now. When people see something amazing, 765 00:50:50,466 --> 00:50:54,209 they literally get scared. I think there is a need to enhance 766 00:50:54,229 --> 00:50:58,131 the education of AI. It's not about just math and programming. There 767 00:50:58,211 --> 00:51:02,534 are other dimensions regarding AI utility, AI 768 00:51:02,594 --> 00:51:06,176 measurements. and business and products that 769 00:51:06,336 --> 00:51:09,838 people should know. Then they can have a much, much, 770 00:51:10,378 --> 00:51:13,999 hopefully, authentic and realistic view about AI, even 771 00:51:14,059 --> 00:51:18,141 without knowing the mathematics and programming. You don't need to be a mathematician or 772 00:51:20,363 --> 00:51:23,466 That's the fascinating thing to me about AI is that it 773 00:51:23,526 --> 00:51:26,909 is possibly one of the most multidisciplinary technologies that 774 00:51:26,989 --> 00:51:30,592 I can think of. You have the technical aspects, 775 00:51:30,652 --> 00:51:34,335 the computer science and the math, but there's also a lot of philosophy, human 776 00:51:34,375 --> 00:51:37,598 psychology, neuroscience, linguistics, the 777 00:51:37,638 --> 00:51:41,240 list kind of goes on. in terms of what you're talking about and the 778 00:51:41,300 --> 00:51:44,663 core of what language models are and what they do, 779 00:51:44,963 --> 00:51:48,185 which is a point that, you know, I appreciate you for bringing up because it's not 780 00:51:48,225 --> 00:51:51,467 talked about often enough. And this kind 781 00:51:51,507 --> 00:51:55,550 of race to intelligence and what intelligence is and what human intelligence is. 782 00:51:55,610 --> 00:51:59,192 And the fascinating thing to me is the kind of stark differences between 783 00:51:59,492 --> 00:52:02,614 language and intelligence and how these two 784 00:52:02,674 --> 00:52:06,117 are not necessarily related. And 785 00:52:06,517 --> 00:52:09,627 in interacting with these chatbots, right, I feel 786 00:52:09,647 --> 00:52:12,909 like we, as a species, communicate our intelligence, our 787 00:52:12,969 --> 00:52:16,332 thought, through language. This is our medium. This 788 00:52:16,452 --> 00:52:20,074 is what sets us apart. And when we see something talking 789 00:52:20,234 --> 00:52:23,356 to us in this medium of how we 790 00:52:23,396 --> 00:52:26,658 convey our intelligence, it's easy for us to assume that it 791 00:52:29,981 --> 00:52:33,263 That extrapolation is very natural for humans to make. 792 00:52:34,313 --> 00:52:37,615 But even so, I guess it is not difficult 793 00:52:37,655 --> 00:52:41,018 for us to realize that even if we don't speak, 794 00:52:41,418 --> 00:52:45,281 we still function and we can still communicate. You have gestures, 795 00:52:45,821 --> 00:52:49,123 you have many sensories that 796 00:52:49,163 --> 00:52:53,006 can sense visual, audio, 797 00:52:53,786 --> 00:52:56,968 temperature, all sorts of things. And not 798 00:52:57,088 --> 00:53:00,989 all of this actually were described in language. In fact, language can 799 00:53:01,710 --> 00:53:05,151 not be used to describe everything. There are many, many reasoning that 800 00:53:05,251 --> 00:53:08,513 is existing outside of language. So 801 00:53:08,533 --> 00:53:11,954 I would really draw a difference 802 00:53:12,014 --> 00:53:16,056 between the functionality of communicating and 803 00:53:16,756 --> 00:53:20,537 functioning linguistically versus 804 00:53:20,758 --> 00:53:23,999 being intelligent. Because intelligence is, again, a very, 805 00:53:24,019 --> 00:53:28,021 very subjective experience. It's a user experience. It's not a measurable 806 00:53:30,423 --> 00:53:33,604 I think the fascinating thing about what we're seeing with 807 00:53:33,624 --> 00:53:36,946 generative AI is that- There is one more thing, which 808 00:53:37,066 --> 00:53:41,029 is oftentimes we call the soul. you 809 00:53:41,049 --> 00:53:44,252 know, the spirit of a creature. I 810 00:53:44,312 --> 00:53:47,975 have to say that I am very limited in my knowledge in understanding that 811 00:53:47,995 --> 00:53:51,538 part. I don't know what that exactly is. Can we assume 812 00:53:51,578 --> 00:53:54,660 that just by becoming knowledgeable enough, I 813 00:53:54,800 --> 00:53:58,083 automatically become soulful? I'm not sure. 814 00:53:58,703 --> 00:54:02,546 I think there's a difference between a machine and a biological creature. 815 00:54:03,067 --> 00:54:06,510 And that difference is actually at least not well understood at this point. 816 00:54:06,830 --> 00:54:10,212 I think the older generation philosophers had time to 817 00:54:10,252 --> 00:54:13,855 think about it. They actually debated 818 00:54:14,695 --> 00:54:19,058 on what is the status of human 819 00:54:19,098 --> 00:54:22,720 reasoning and human morality and 820 00:54:23,541 --> 00:54:27,303 will in the definition 821 00:54:27,343 --> 00:54:31,165 of human versus machineries. I 822 00:54:31,205 --> 00:54:34,688 think that discussion and that debate and thinking 823 00:54:35,468 --> 00:54:39,129 should continue or restart because now we 824 00:54:39,169 --> 00:54:42,450 have a new technology, which is AI. At that time, 825 00:54:43,070 --> 00:54:47,072 they had this discussion because of the 826 00:54:47,792 --> 00:54:52,034 latest invention, which is the calculus, the the 827 00:54:52,114 --> 00:54:55,256 orbiting planets and so forth, which was fascinating. It may give you the 828 00:54:55,276 --> 00:54:59,339 feeling that, wow, I'm like a god. I can predict the position 829 00:54:59,399 --> 00:55:03,121 of the planets. I can predict the ellipse and so forth. And 830 00:55:03,161 --> 00:55:06,523 then people start to think, oh, what is knowledge? What 831 00:55:06,643 --> 00:55:10,446 is reasoning? What is conscious? So forth. I 832 00:55:10,486 --> 00:55:13,688 think now it's time to rethink about that. Because with AI, which we 833 00:55:13,728 --> 00:55:17,370 actually can make in a mechanical way using a 834 00:55:17,450 --> 00:55:21,232 very, very simple training kind 835 00:55:21,272 --> 00:55:24,434 of strategy, say, next world prediction, how 836 00:55:24,474 --> 00:55:27,655 can we redefine what is ration and what 837 00:55:27,736 --> 00:55:31,257 is mind and so forth? But I would 838 00:55:31,397 --> 00:55:34,899 caution people from 839 00:55:35,700 --> 00:55:40,862 making the jump of directly connecting 840 00:55:41,403 --> 00:55:45,145 a very, very capable and seemingly 841 00:55:45,245 --> 00:55:48,665 smart artificial system to 842 00:55:49,245 --> 00:55:55,127 a human-like creature which has the 843 00:55:55,227 --> 00:55:58,708 ego and also maybe 844 00:55:58,828 --> 00:56:02,169 even the malicious intent to come after us. I think 845 00:56:02,249 --> 00:56:05,910 there are a lot of steps, even philosophical ones and cognitive 846 00:56:06,010 --> 00:56:10,093 ones, between them. And drawing an automatic extrapolation 847 00:56:11,614 --> 00:56:16,018 is not very convincing, at least to a 848 00:56:17,039 --> 00:56:20,281 The point you brought up about the soul, I find that 849 00:56:20,301 --> 00:56:23,744 fascinating. And that's the kind of these differences 850 00:56:23,804 --> 00:56:27,207 between knowledge and having been trained on every human work 851 00:56:27,267 --> 00:56:30,491 ever written. the different facets that 852 00:56:30,511 --> 00:56:34,052 make up what human intelligence is. And it goes beyond 853 00:56:34,813 --> 00:56:38,454 knowing things. The stuff that makes humans human 854 00:56:39,535 --> 00:56:43,137 I had one interesting debate with a colleague. For example, we 855 00:56:43,157 --> 00:56:46,758 train generative AI, such as the Zara, for producing movies. Now, 856 00:56:46,798 --> 00:56:50,340 let's imagine that the system watch a lot of Animal Kingdom 857 00:56:50,380 --> 00:56:53,621 movies, like animals fight each other, eat 858 00:56:53,661 --> 00:56:57,683 each other, and so forth, so that they can produce similar movies. Along 859 00:56:57,723 --> 00:57:01,206 the way, are they going to be learning the will of 860 00:57:01,246 --> 00:57:04,728 killing and of predator and prey 861 00:57:04,948 --> 00:57:08,290 type of mindset? Well, nobody actually 862 00:57:08,310 --> 00:57:11,832 knows, at least at this point. But it's a fascinating question to 863 00:57:13,433 --> 00:57:17,156 Intention is a big thing here that's missing. 864 00:57:17,416 --> 00:57:20,978 But there's also a lot that's not known about language models. 865 00:57:21,179 --> 00:57:25,061 The real risk I do see in a potentially 866 00:57:25,221 --> 00:57:28,505 out of control AI system is not about 867 00:57:28,525 --> 00:57:32,151 that they kill people and so forth. It's about their 868 00:57:33,311 --> 00:57:36,573 environmental footprint. Imagine that if 869 00:57:36,613 --> 00:57:40,415 you do have an AI system that is told to 870 00:57:40,595 --> 00:57:44,357 infinitely iterate on some very, 871 00:57:44,397 --> 00:57:48,039 very difficult objectives that actually can be measured, but 872 00:57:48,059 --> 00:57:51,581 they can go very, very big. What if the machine decides 873 00:57:51,681 --> 00:57:54,902 to keep training themselves and burn a 874 00:57:54,942 --> 00:57:58,624 heck of energy and so forth, and causing really environmental 875 00:57:58,644 --> 00:58:01,966 disasters? To me, that's a real risk we have to be careful 876 00:58:01,986 --> 00:58:06,008 about. There has to be checkpoints and the bricks, so 877 00:58:06,028 --> 00:58:09,629 that at least a person should be allowed to turn it off. If 878 00:58:09,669 --> 00:58:13,131 you want to build a system that is never having a switch, probably 879 00:58:13,191 --> 00:58:16,352 no good. But these are something you 880 00:58:16,392 --> 00:58:20,275 can already bear in your mind when you do the design. And 881 00:58:20,295 --> 00:58:24,116 we could call it also a regulation in the science space. But 882 00:58:24,936 --> 00:58:28,217 I don't want to overgeneralize that. It's not like 883 00:58:28,317 --> 00:58:31,658 everywhere. It's only a very, very 884 00:58:33,918 --> 00:58:37,259 definable and measurable risk that is indeed very, 885 00:58:37,319 --> 00:58:40,860 very grave. But on the other hand, I think 886 00:58:42,265 --> 00:58:46,148 Right. And a lot of these risks come from putting 887 00:58:46,168 --> 00:58:49,972 these systems in a place where they can make decisions without human oversight. And 888 00:58:49,992 --> 00:58:53,294 I guess the immediate and obvious remedy is don't do 889 00:58:53,334 --> 00:58:56,837 that. Maintain human oversight. You won't have those problems. 890 00:58:58,078 --> 00:59:01,481 Yeah, yeah. That's true. So I 891 00:59:01,802 --> 00:59:05,204 think it's an ongoing discussion. But again, there is still 892 00:59:05,305 --> 00:59:08,687 a difference between the scientific kind 893 00:59:08,727 --> 00:59:12,033 of brain work, or maybe even 894 00:59:12,053 --> 00:59:15,900 the technological experimental work from 895 00:59:16,160 --> 00:59:19,425 the production and deployments and 896 00:59:19,445 --> 00:59:23,093 the usage. I would imagine 897 00:59:23,133 --> 00:59:26,515 that it's more effective and productive to focus on 898 00:59:26,535 --> 00:59:29,776 the latter. Also, one more thing. I 899 00:59:29,796 --> 00:59:33,758 think AI, compared to other sciences, 900 00:59:34,618 --> 00:59:37,940 has the advantage that it is still one step 901 00:59:38,040 --> 00:59:41,481 away from the physical world. Imagine that you are a biologist, you 902 00:59:41,521 --> 00:59:45,623 do virology and genetic 903 00:59:45,703 --> 00:59:49,625 engineering. Well, the products of 904 00:59:49,645 --> 00:59:53,394 your experimentation is a physical being. If 905 00:59:53,474 --> 00:59:57,978 you somehow smash the glass, or you somehow cause 906 00:59:58,058 --> 01:00:01,480 a lab leak, it actually does enter the 907 01:00:01,901 --> 01:00:05,343 environment. And they will stay there, or they could spread, 908 01:00:06,324 --> 01:00:09,967 just like what we see in COVID. So these are very mature 909 01:00:10,027 --> 01:00:14,370 risks. And therefore, it requires rather 910 01:00:14,471 --> 01:00:20,365 stringent regulations and control. AI maybe 911 01:00:20,565 --> 01:00:24,528 conceptually have that kind of same risk, but 912 01:00:24,708 --> 01:00:28,331 it is fortunately still in a digital 913 01:00:28,351 --> 01:00:31,533 and a virtual environment. In the worst case, you just 914 01:00:31,553 --> 01:00:34,735 throw off your iPhones and shut down your machines. You go back 915 01:00:34,775 --> 01:00:38,137 to the cage, the caves. We will still 916 01:00:38,458 --> 01:00:41,800 survive, right? And without, you know, for 917 01:00:41,820 --> 01:00:46,603 example, computer. Of course, I know it's very difficult these days, but I 918 01:00:46,663 --> 01:00:51,308 mean, it's a different level of risk. That 919 01:00:51,388 --> 01:00:55,491 existential risk is different from a biological disaster, 920 01:00:55,851 --> 01:00:59,513 a chemical skill, a nuclear kind of explosion, 921 01:01:01,014 --> 01:01:04,276 Part of the regulation question, and all 922 01:01:04,316 --> 01:01:07,597 these topics are so passionately debated in the public space. 923 01:01:08,698 --> 01:01:11,859 is the idea of open source versus a lot of what 924 01:01:11,879 --> 01:01:15,839 we see from the companies, which is super, super closed source. 925 01:01:16,039 --> 01:01:19,220 We're not even going to tell you what we trained on. That has to come 926 01:01:19,280 --> 01:01:22,540 out through reports from investigative journalists. And 927 01:01:22,580 --> 01:01:26,681 it trickles out and is fought over in court because of copyright concerns. 928 01:01:27,381 --> 01:01:30,662 But the debate hinges on this idea 929 01:01:30,682 --> 01:01:34,703 that science needs to be open. Science 930 01:01:34,743 --> 01:01:37,843 needs to be explainable and replicable. And if 931 01:01:38,224 --> 01:01:42,287 it's not open, it's not really science. And it's the distinction between product 932 01:01:42,607 --> 01:01:46,050 and science. But there's also concerns about 933 01:01:46,590 --> 01:01:50,153 what open source could mean for security risks. And you 934 01:01:50,173 --> 01:01:54,157 were talking about chemical and biological situations 935 01:01:54,217 --> 01:01:57,759 and how AI open source can maybe help bad 936 01:01:57,860 --> 01:02:01,262 actors in those realms. So as 937 01:02:01,282 --> 01:02:05,586 it relates to that and making sure that nothing bad happens, what 938 01:02:08,291 --> 01:02:11,535 Well, I don't 939 01:02:11,555 --> 01:02:14,939 want to reiterate the philosophy in science 940 01:02:14,979 --> 01:02:18,483 that there is a fundamental value of share 941 01:02:18,543 --> 01:02:22,248 and open and so forth. Definitely, I'm for that. More 942 01:02:22,288 --> 01:02:25,471 than that, I actually don't think secrecy is 943 01:02:25,611 --> 01:02:28,759 a recipe for security. I 944 01:02:28,799 --> 01:02:31,902 think there's a fundamental issue. People thought by keeping things secret, you 945 01:02:31,942 --> 01:02:35,145 become safer. I doubt it. I 946 01:02:35,165 --> 01:02:39,249 don't think history gave us too many examples of that. Because 947 01:02:39,689 --> 01:02:43,101 eventually, you will lose the secrecy. I think having 948 01:02:44,082 --> 01:02:47,565 more people to understand the technology, to be able 949 01:02:47,665 --> 01:02:51,469 to investigate and experiment creates 950 01:02:52,129 --> 01:02:56,553 a much larger pool of people 951 01:02:56,854 --> 01:03:00,237 and also a large pool of knowledge to contain 952 01:03:00,257 --> 01:03:03,740 the risk if that risk happens. That's kind of, again, maybe 953 01:03:03,900 --> 01:03:08,865 a more philosophical and abstract argument. But 954 01:03:09,025 --> 01:03:12,688 I can give you a few examples. Right 955 01:03:12,728 --> 01:03:16,491 now, we don't actually know exactly how 956 01:03:16,952 --> 01:03:20,215 GPTs are trained. Do 957 01:03:20,255 --> 01:03:24,218 you think that makes the technology safer or less safe? Well, 958 01:03:24,498 --> 01:03:28,020 I don't know. I feel less comfortable 959 01:03:28,600 --> 01:03:31,821 if I'm using something that I have 960 01:03:31,901 --> 01:03:35,722 no knowledge from. Yes, I'm not going to break my car and 961 01:03:36,162 --> 01:03:42,065 do it myself or understand how 962 01:03:42,145 --> 01:03:45,706 it is made. But at least I know that there 963 01:03:45,786 --> 01:03:49,327 are more than one manufacturer who knows how to do it. If 964 01:03:49,447 --> 01:03:52,708 my car breaks down, like I have to drive a Mercedes, I don't have to go back 965 01:03:52,728 --> 01:03:55,949 to Mercedes. I can go to any car shop to have it fixed. That 966 01:03:55,969 --> 01:04:00,490 makes me feel safe. So under the same principle, I think important 967 01:04:00,510 --> 01:04:03,771 technologies, I'm not even arguing a commercial reason. I think many of 968 01:04:03,811 --> 01:04:07,552 the arguments was because of a commercial interest and other 969 01:04:07,652 --> 01:04:12,613 geopolitical interest. But from a security standpoint, open 970 01:04:12,653 --> 01:04:15,774 source has the advantage of 971 01:04:18,530 --> 01:04:22,212 sharing the knowledge to more 972 01:04:22,272 --> 01:04:25,434 people, and also keeping the 973 01:04:25,474 --> 01:04:29,696 technology transparent for scrutiny, and therefore, for 974 01:04:29,996 --> 01:04:33,438 a better opportunity to regulate it. So I think that 975 01:04:33,458 --> 01:04:37,419 part is very important, even from a pure security 976 01:04:37,460 --> 01:04:41,587 standpoint. Now, of course, there are even more 977 01:04:41,747 --> 01:04:45,109 arguments about the economical and the scientific kind of 978 01:04:45,149 --> 01:04:48,350 value from it. Economy is obvious. You don't want 979 01:04:48,870 --> 01:04:52,131 a lucrative and important technology to be 980 01:04:52,312 --> 01:04:55,833 monopolized by one of your companies. That's why in 981 01:04:55,933 --> 01:04:59,354 the US, you have the antitrust laws and the anti-monopoly laws, 982 01:04:59,474 --> 01:05:02,856 right? So if your company, Rockefeller, once 983 01:05:02,876 --> 01:05:06,277 they own all the petroleum production, the 984 01:05:06,317 --> 01:05:09,778 government orders it to break it. There is always an interest 985 01:05:09,898 --> 01:05:14,020 of healthy competition among entities for 986 01:05:14,060 --> 01:05:18,241 key technology and key products to actually better the 987 01:05:18,641 --> 01:05:21,822 quality of technology itself and also to make it affordable and 988 01:05:21,882 --> 01:05:25,443 accessible to people. So that's a commercial argument I want to make. Open 989 01:05:25,503 --> 01:05:29,344 source gives startups, even 990 01:05:29,444 --> 01:05:34,692 universities, a share of their impact 991 01:05:34,992 --> 01:05:39,753 and also their ability to 992 01:05:39,833 --> 01:05:43,754 innovate in this market. That's scientifically even 993 01:05:43,834 --> 01:05:47,194 more important. I think no science actually gets 994 01:05:47,434 --> 01:05:51,275 developed very well if it's kept in secret. So 995 01:05:51,775 --> 01:05:55,636 for the sake of advancing the science of artificial 996 01:05:55,676 --> 01:05:59,496 intelligence, especially in the United States, definitely the 997 01:05:59,776 --> 01:06:04,011 U.S. is right now the current leader. I 998 01:06:04,051 --> 01:06:07,374 think by having the open 999 01:06:07,414 --> 01:06:11,418 source community very robustly 1000 01:06:11,438 --> 01:06:15,122 grounded in the United States, actually end up helping the 1001 01:06:15,142 --> 01:06:18,325 United States to be the leader for 1002 01:06:18,345 --> 01:06:21,808 the past many decades in IT technology. Look at, for 1003 01:06:21,828 --> 01:06:24,971 example, the development of 1004 01:06:25,051 --> 01:06:28,535 operating systems like Linux, the search engines. 1005 01:06:29,118 --> 01:06:32,741 and many of the novel IT innovations 1006 01:06:33,221 --> 01:06:37,965 is because there is this openness of 1007 01:06:38,606 --> 01:06:43,890 technology to be played and experimented and 1008 01:06:44,050 --> 01:06:47,173 advanced, not just by a designated player, but 1009 01:06:47,253 --> 01:06:50,876 by the entire academic and 1010 01:06:51,056 --> 01:06:55,400 industrial community at large. I think that capability and 1011 01:06:55,960 --> 01:06:59,462 freedom is nowhere to see, in fact, elsewhere 1012 01:06:59,502 --> 01:07:03,644 in the world, not even in Europe. In 1013 01:07:03,684 --> 01:07:07,326 fact, that's part of the reason I become excited to take 1014 01:07:07,406 --> 01:07:10,767 on this job in here, because I do like to 1015 01:07:10,927 --> 01:07:14,629 see such a mindset and such a value, 1016 01:07:14,829 --> 01:07:18,957 a principle in education, in technology, to 1017 01:07:19,357 --> 01:07:24,139 impact regions which are yet to be impacted. So 1018 01:07:24,179 --> 01:07:27,460 here is a young country that 1019 01:07:27,540 --> 01:07:31,941 wants to have a voice and have a stake in the global technological 1020 01:07:32,001 --> 01:07:35,202 advancements. And they have the 1021 01:07:35,242 --> 01:07:38,783 resource and also the ambition and also the enthusiasm of 1022 01:07:38,843 --> 01:07:41,984 being a leader. And absolutely, I think by 1023 01:07:47,895 --> 01:07:51,397 you know, moving the, 1024 01:07:52,018 --> 01:07:55,260 maybe I'm lacking the right word, by learning or 1025 01:07:55,300 --> 01:07:58,461 maybe by advocating and advancing, amplifying in that 1026 01:07:58,502 --> 01:08:02,444 kind of openness, that kind of transparency, you 1027 01:08:02,504 --> 01:08:06,146 know, from the very beginning of the foundation of this university and 1028 01:08:06,306 --> 01:08:09,548 of the talent we are training, I believe is 1029 01:08:09,688 --> 01:08:13,671 at the end of the day helping, you know, to grow this technology. And 1030 01:08:13,691 --> 01:08:17,113 that's also why in our university, yeah, we indeed take a very, very 1031 01:08:18,116 --> 01:08:22,238 open, but of course, compliant and 1032 01:08:22,478 --> 01:08:25,900 regulated approach to develop our AI. 1033 01:08:26,521 --> 01:08:30,062 You may have heard about, of course, everybody heard about Lama from Meta, 1034 01:08:30,363 --> 01:08:34,325 right? And it's a huge benefit 1035 01:08:34,545 --> 01:08:37,866 and support to 1036 01:08:37,886 --> 01:08:41,588 the community. And they benefit a lot, a lot of researchers working 1037 01:08:41,648 --> 01:08:45,390 on that and maybe inspired 1038 01:08:45,410 --> 01:08:48,964 a lot of new advancements. In here, we 1039 01:08:49,024 --> 01:08:52,647 have a project called LLM360, which actually went 1040 01:08:53,148 --> 01:08:56,912 even multiple steps further. We not only open-sourced the 1041 01:08:57,312 --> 01:09:00,636 model weights, we open-sourced the 1042 01:09:00,696 --> 01:09:03,959 checkpoints and also the code that 1043 01:09:03,999 --> 01:09:07,222 you use to produce the model. And our goal is 1044 01:09:07,282 --> 01:09:10,684 to let researchers to be able to reproduce the result, 1045 01:09:10,764 --> 01:09:14,767 good and bad, because otherwise you can only use 1046 01:09:14,847 --> 01:09:18,089 it and augment it. But you 1047 01:09:18,109 --> 01:09:21,832 still have this dependency of waiting for the next version, you 1048 01:09:21,872 --> 01:09:25,374 know, for you to rebuild everything from scratch. And 1049 01:09:25,414 --> 01:09:28,976 in our case, We want to make the community even bigger 1050 01:09:29,837 --> 01:09:32,938 with this higher degree of transparency. And I 1051 01:09:32,978 --> 01:09:36,260 think, does that give us any risk? Well, 1052 01:09:36,720 --> 01:09:40,722 you can never rule out the possibility of a bad player somewhere 1053 01:09:40,783 --> 01:09:44,705 in the world to take that and turn it into somewhat 1054 01:09:44,885 --> 01:09:47,986 a malicious tool. I 1055 01:09:48,026 --> 01:09:52,067 don't think, you know, that's going 1056 01:09:52,087 --> 01:09:55,308 to be stopped if we just close source it. People will 1057 01:09:55,348 --> 01:09:58,990 figure out from else places. But on the other hand, by 1058 01:09:59,491 --> 01:10:02,793 opening the technology in such a way, 1059 01:10:03,473 --> 01:10:07,095 we end up having more people understanding how 1060 01:10:07,675 --> 01:10:11,537 to augment and how to manage 1061 01:10:11,657 --> 01:10:14,899 and contain the risk from this machine and how to make 1062 01:10:14,919 --> 01:10:18,701 the machine better. So again, it's a debatable topic, but 1063 01:10:18,941 --> 01:10:22,123 I just want to share with you my view from a developer and 1064 01:10:22,163 --> 01:10:25,365 a researcher who makes these kind of technologies, I think 1065 01:10:28,647 --> 01:10:32,088 And we talked about this a little bit, but the thing that I think is important 1066 01:10:32,189 --> 01:10:36,171 in the realm of AI literacy and helping ordinary 1067 01:10:36,211 --> 01:10:39,573 people who don't spend their time reading 1068 01:10:39,633 --> 01:10:45,455 research papers or writing research papers about AI, What 1069 01:10:45,475 --> 01:10:48,962 are the biggest misconceptions that you see people 1070 01:10:49,022 --> 01:10:52,329 make about artificial intelligence in general 1071 01:10:55,322 --> 01:10:59,945 Well, all the things that we just discussed, like taking 1072 01:11:00,925 --> 01:11:05,468 AI as some sort of a creature that 1073 01:11:05,788 --> 01:11:09,210 could spin out of control from human hands, is definitely 1074 01:11:09,330 --> 01:11:12,651 one misperception. And 1075 01:11:12,751 --> 01:11:16,914 also exaggerating the power 1076 01:11:17,054 --> 01:11:20,576 of AI in real human 1077 01:11:20,676 --> 01:11:26,084 life, and also in scientific exploration, In 1078 01:11:26,104 --> 01:11:30,486 my opinion, it's also a misperception because AI 1079 01:11:30,686 --> 01:11:33,947 is very powerful, but 1080 01:11:34,067 --> 01:11:37,428 it is powerful in certain dimensions. For example, it is very powerful now 1081 01:11:37,488 --> 01:11:40,569 in playing the game of chess and Go. But it 1082 01:11:40,609 --> 01:11:44,111 doesn't mean that it plays other things well, like planning 1083 01:11:44,131 --> 01:11:48,512 your trips. So it takes time to get there. Another 1084 01:11:48,612 --> 01:11:51,954 misperception is that AI is 1085 01:11:52,154 --> 01:11:55,956 only about large language models and the GPTs. In 1086 01:11:55,996 --> 01:12:01,505 fact, AI is a very, very broad technology that 1087 01:12:01,625 --> 01:12:04,990 is more than actually we 1088 01:12:05,070 --> 01:12:08,275 expected to be coexisting with 1089 01:12:08,356 --> 01:12:11,835 us already. Our search engine is also an AI tool. And 1090 01:12:11,895 --> 01:12:15,216 I don't think people want to ban such engine. And 1091 01:12:15,916 --> 01:12:19,676 also, now everybody is driving 1092 01:12:20,136 --> 01:12:23,337 a car with a GPS or 1093 01:12:23,357 --> 01:12:26,778 with your iPhone next to you to 1094 01:12:26,978 --> 01:12:30,178 show you the road. I can tell you that I'm old enough that I live 1095 01:12:30,198 --> 01:12:33,419 in an age where I use a paper map to drive. It's painful. I don't want 1096 01:12:33,439 --> 01:12:36,819 to go back that age. People don't realize that that function is also 1097 01:12:37,039 --> 01:12:41,304 from AI. So AI is already everywhere. So 1098 01:12:42,125 --> 01:12:48,428 using the GPT as the 1099 01:12:48,648 --> 01:12:51,790 main or the only representation of AI is 1100 01:12:51,970 --> 01:12:56,192 itself a misperception. And therefore, the discussion around 1101 01:12:56,352 --> 01:13:00,455 AI sometimes is 1102 01:13:00,495 --> 01:13:03,696 very, very confused and ambiguous. Maybe we 1103 01:13:03,776 --> 01:13:07,538 should say that, how about we discuss large language models? That's actually 1104 01:13:07,578 --> 01:13:11,060 a lot more precise. and 1105 01:13:11,100 --> 01:13:15,023 better scoped. But when you say AI, literally 1106 01:13:15,043 --> 01:13:19,246 you are talking about how about we take a Google 1107 01:13:19,266 --> 01:13:22,888 map from your iPhone. That's a part of AI. And 1108 01:13:23,388 --> 01:13:27,751 who knows what will come along that line, the next product. I think 1109 01:13:27,911 --> 01:13:31,718 self-driving is one of the major things That 1110 01:13:31,798 --> 01:13:35,480 is coming very close to what 1111 01:13:35,500 --> 01:13:38,741 it's already realized. It's just like right now the policy is 1112 01:13:39,641 --> 01:13:43,202 lagging behind for its deployment. In 1113 01:13:43,222 --> 01:13:46,403 fact, I think a policy is more needed over 1114 01:13:46,483 --> 01:13:50,485 there in terms of deployment. Because if you look at the statistics, the 1115 01:13:50,625 --> 01:13:53,826 accidents committed by an autonomous driving car actually is 1116 01:13:53,946 --> 01:13:58,168 lower than people. And the 1117 01:13:58,228 --> 01:14:01,590 reason why it is not widely adopted is a policy and 1118 01:14:01,710 --> 01:14:05,091 infrastructure issue, not necessarily a technological issue. 1119 01:14:06,212 --> 01:14:09,593 And so I think there are a lot of examples in 1120 01:14:09,633 --> 01:14:13,055 such a space where especially decision makers 1121 01:14:13,295 --> 01:14:16,937 and policy makers will be benefiting from 1122 01:14:17,237 --> 01:14:20,514 having scientists and developers next to 1123 01:14:20,574 --> 01:14:23,756 them to be part of the discussion. I have to say 1124 01:14:23,776 --> 01:14:29,081 that I've been to events in various organizations 1125 01:14:29,241 --> 01:14:33,284 in the government bodies where a scientist 1126 01:14:33,304 --> 01:14:37,647 like myself could amount to 1127 01:14:37,647 --> 01:14:41,230 5% of the participants, or even less. So 1128 01:14:41,350 --> 01:14:45,253 in that environment, it is amplifying, in 1129 01:14:48,387 --> 01:14:51,788 We talk about people kind of trying to educate themselves better. And 1130 01:14:52,248 --> 01:14:55,489 like you were saying, it's really difficult. A 1131 01:14:55,529 --> 01:14:58,770 lot of this is ambiguous. A lot of the language changes depending on who you 1132 01:14:58,790 --> 01:15:02,211 talk to. What should people 1133 01:15:02,271 --> 01:15:06,712 be doing about AI right now? How 1134 01:15:06,772 --> 01:15:10,193 should they be kind of going about thinking 1135 01:15:10,253 --> 01:15:13,333 about it? I think there's a 1136 01:15:13,373 --> 01:15:16,737 degree that some people are afraid that it's too complex, too 1137 01:15:16,857 --> 01:15:20,061 out of reach to have an opinion on it or to think 1138 01:15:20,101 --> 01:15:23,965 about it too much. What would you tell them they 1139 01:15:28,533 --> 01:15:31,694 Good question. I 1140 01:15:31,754 --> 01:15:34,935 have to say that I don't want 1141 01:15:34,975 --> 01:15:38,135 to push the blame to the public just because they don't do 1142 01:15:38,155 --> 01:15:42,256 enough. I think the general public, even 1143 01:15:42,356 --> 01:15:45,817 including decision makers, policy makers, I 1144 01:15:45,917 --> 01:15:49,158 genuinely believe that they wanted to know more, but they don't know where 1145 01:15:49,198 --> 01:15:52,870 to learn and how to learn. So our educational system 1146 01:15:52,890 --> 01:15:58,037 should be the starting point that our current curriculum regarding 1147 01:15:58,198 --> 01:16:01,616 AI or computer science probably it's time to revisit 1148 01:16:01,636 --> 01:16:05,597 them. In fact, for the past, I 1149 01:16:05,637 --> 01:16:09,238 would say, several decades, if you look at the curriculum, 1150 01:16:10,118 --> 01:16:13,679 there hasn't been a fundamental change. It's literally just about how 1151 01:16:13,759 --> 01:16:17,160 much I teach along the same line. Which program 1152 01:16:17,180 --> 01:16:20,561 language? Okay, I can start it from Fortran in the 50s, 1153 01:16:20,701 --> 01:16:23,921 and then C, and then Java, and now Python, and so forth. But 1154 01:16:23,941 --> 01:16:27,566 it's still about teaching a different language. mathematics, well 1155 01:16:27,586 --> 01:16:30,928 there are new material coming in, let's retire the older ones 1156 01:16:31,008 --> 01:16:34,449 and bring in new ones. But yeah, you need to know linear algebra, 1157 01:16:34,489 --> 01:16:37,831 you need to know calculus, and so forth. So it's still the 1158 01:16:38,331 --> 01:16:41,653 very, very conventional way of thinking, and the statistics, for example, 1159 01:16:41,713 --> 01:16:46,530 are still using a year to teach regressions and so forth. With 1160 01:16:46,610 --> 01:16:50,131 now AI and with functions like 1161 01:16:50,171 --> 01:16:54,172 GPT or search engine and so forth, in 1162 01:16:54,192 --> 01:16:57,593 fact, we already are in the crisis where some people are using 1163 01:16:57,653 --> 01:17:01,674 GPTs to write scientific reviews of the submitted 1164 01:17:01,754 --> 01:17:05,335 publications. And people are getting very frustrated, 1165 01:17:05,415 --> 01:17:08,836 but you cannot stop it. It's happening, right? So then your attitude towards 1166 01:17:08,996 --> 01:17:12,116 even published papers may need to be augmented. Are 1167 01:17:12,156 --> 01:17:15,700 they really the product of the the researchers or 1168 01:17:15,760 --> 01:17:19,449 it's a mix and so forth. So I think this awareness of 1169 01:17:19,529 --> 01:17:22,977 the changing technology should be first taught to the public, 1170 01:17:23,358 --> 01:17:26,803 you know, in some curriculum. And that material does not require you 1171 01:17:26,863 --> 01:17:30,326 to know calculus and know linear algebra and know all this fancy 1172 01:17:30,686 --> 01:17:34,048 mathematical stuff. And in fact, it could be taught even 1173 01:17:34,088 --> 01:17:38,171 in elementary school and high school. So I think bringing 1174 01:17:38,471 --> 01:17:42,534 the AI into the background of 1175 01:17:43,235 --> 01:17:47,037 all teaching materials makes them not a specialty, 1176 01:17:47,437 --> 01:17:51,712 but a very, very horizontal, fundamental 1177 01:17:51,772 --> 01:17:55,073 science is probably where we want to begin from. Start early 1178 01:17:55,834 --> 01:17:59,535 and start very, very simple but broad to 1179 01:17:59,695 --> 01:18:03,016 create awareness about what we 1180 01:18:03,096 --> 01:18:06,518 already are using AI to do, and how 1181 01:18:06,738 --> 01:18:10,821 well they can do, and what else we set 1182 01:18:10,901 --> 01:18:14,324 forth for them to do. That kind of high-level knowledge 1183 01:18:14,784 --> 01:18:18,347 can be already taught to the general public 1184 01:18:18,567 --> 01:18:22,530 without any emphasis on technical specialty. 1185 01:18:23,551 --> 01:18:27,093 Then you can talk about the connections of 1186 01:18:27,313 --> 01:18:31,396 AI with all disciplines, not just AI 1187 01:18:31,436 --> 01:18:36,611 for science or AI making a product. For 1188 01:18:36,631 --> 01:18:40,276 example, artists are already using generative 1189 01:18:40,376 --> 01:18:43,860 AI to create visual arts, 1190 01:18:45,022 --> 01:18:49,347 music, and many other things. How 1191 01:18:49,367 --> 01:18:52,771 do we appreciate and interpret these 1192 01:18:52,871 --> 01:18:56,956 kinds of outcomes? and how can we redefine copyrights 1193 01:18:57,957 --> 01:19:02,724 and maybe authorship 1194 01:19:02,824 --> 01:19:06,407 and all that kind of things. I think this kind 1195 01:19:06,427 --> 01:19:09,731 of issues are already happening, but 1196 01:19:09,891 --> 01:19:14,256 very few people really put time 1197 01:19:14,316 --> 01:19:17,879 to think about it, or not even maybe knowing it. So 1198 01:19:17,899 --> 01:19:22,144 I think the AI literacy is really about awareness 1199 01:19:22,805 --> 01:19:26,028 of such an entity, such a 1200 01:19:26,509 --> 01:19:30,312 capability. We all know, for example, the capability of linear 1201 01:19:30,412 --> 01:19:33,615 algebra, or not linear algebra, just algebra. That's why 1202 01:19:33,635 --> 01:19:37,658 we are not afraid of a calculator, because we actually know algebra, even 1203 01:19:37,698 --> 01:19:41,021 though we don't know how to do algebra, but we know what is algebra. Right 1204 01:19:41,061 --> 01:19:44,844 now, we don't have such a literacy in AI. And 1205 01:19:45,024 --> 01:19:48,407 that's where, from the educational standpoint, I 1206 01:19:48,447 --> 01:19:51,629 would like to see universities and even 1207 01:19:52,811 --> 01:19:56,173 going further earlier, high schools, elementary schools, start 1208 01:19:56,193 --> 01:19:59,635 thinking about that. My son is in high school right now and 1209 01:20:00,516 --> 01:20:04,018 I sometimes look into 1210 01:20:04,098 --> 01:20:07,741 his curriculum and try to understand how 1211 01:20:07,781 --> 01:20:12,264 much their teachers and principals actually know 1212 01:20:12,344 --> 01:20:16,246 about all this. I think it's fun to keep an eye on it. Then 1213 01:20:16,286 --> 01:20:19,462 given that's happening, I have more to say to 1214 01:20:19,482 --> 01:20:22,825 the general public. Hey, go read 1215 01:20:23,626 --> 01:20:27,630 such material. Go watch a few informative 1216 01:20:27,990 --> 01:20:31,113 videos on YouTube and other places. But on 1217 01:20:31,133 --> 01:20:34,716 the other hand, go ignore certain kind of information that 1218 01:20:34,776 --> 01:20:39,741 you get from all different sources. There is definitely a 1219 01:20:39,821 --> 01:20:43,184 difference of good and bad information. And there 1220 01:20:43,224 --> 01:20:46,767 is also a judgment that you have to grow to be 1221 01:20:46,827 --> 01:20:51,172 able to tell what is really 1222 01:20:52,933 --> 01:20:58,539 valuable and authentic versus 1223 01:20:58,559 --> 01:21:02,483 what may be something suspicious. I think, again, 1224 01:21:02,503 --> 01:21:06,360 I'm not a fan of just 1225 01:21:07,240 --> 01:21:10,822 cleaning up the things to make it like a 1226 01:21:10,882 --> 01:21:14,964 hospital and a clean room. In fact, if 1227 01:21:14,984 --> 01:21:18,086 you look at evolution, both in 1228 01:21:18,126 --> 01:21:21,968 terms of a biological entity or even evolution 1229 01:21:22,008 --> 01:21:25,450 of our brain, You evolve and 1230 01:21:25,490 --> 01:21:29,252 you become healthier because you have adversaries. You 1231 01:21:29,292 --> 01:21:33,534 have all these viruses, you have all these pathogens, you 1232 01:21:33,574 --> 01:21:37,576 have actually all these misinformations from 1233 01:21:37,956 --> 01:21:41,438 your world. And you co-evolve with it. You 1234 01:21:41,498 --> 01:21:45,139 end up developing your own immune to be stronger. And 1235 01:21:45,260 --> 01:21:48,621 keeping you away from all this isn't going to make you healthier. In 1236 01:21:48,641 --> 01:21:53,458 fact, it makes you weaker. I don't think the politicians 1237 01:21:53,678 --> 01:21:56,979 and the general public appreciate that. They want to be as 1238 01:21:57,039 --> 01:22:00,361 safe as possible, away from all risk and all 1239 01:22:00,801 --> 01:22:03,882 kind of pathogens, psychological pathogens, you 1240 01:22:03,902 --> 01:22:07,223 know, or biological pathogens. I think it's going to weaken 1241 01:22:07,343 --> 01:22:10,665 our civilization. And then when you have an adversary, we 1242 01:22:10,685 --> 01:22:14,506 talk about China, we talk about some other countries, what if they don't do that? You 1243 01:22:14,566 --> 01:22:17,927 will eventually run into an adversary that is healthier than 1244 01:22:17,967 --> 01:22:21,330 you, and that will be even worse. So I would really 1245 01:22:21,430 --> 01:22:24,615 advocate a culture of openness and 1246 01:22:24,655 --> 01:22:29,061 also of brevity to embrace this 1247 01:22:29,141 --> 01:22:32,385 reality and then try to take on it. Thanks so much,