WEBVTT 00:00:09.449 --> 00:00:11.869 This episode of the Khronos Fusion Energy podcast 00:00:11.869 --> 00:00:15.369 features Jack Dungara, one of the foundational 00:00:15.369 --> 00:00:18.089 architects of modern high -performance scientific 00:00:18.089 --> 00:00:22.050 computing and the 2021 winner of the Turing Award, 00:00:22.350 --> 00:00:25.589 the highest honor in computer science, described 00:00:25.589 --> 00:00:30.070 as the Nobel Prize of Computing. Jack is internationally 00:00:30.070 --> 00:00:33.450 recognized for his pioneering work in numerical 00:00:33.450 --> 00:00:36.469 linear algebra and high -performance computing. 00:00:37.149 --> 00:00:39.490 including the development of software libraries 00:00:39.490 --> 00:00:42.350 that underpin much of today's scientific computing 00:00:42.350 --> 00:00:45.210 ecosystem and form the mathematical backbone 00:00:45.210 --> 00:00:49.649 of tools such as MATLAB. His career spans decades 00:00:49.649 --> 00:00:52.689 at the intersection of mathematics, computer 00:00:52.689 --> 00:00:55.570 science, and large -scale scientific simulations, 00:00:56.049 --> 00:00:59.490 with appointments at institutions including Argonne 00:00:59.490 --> 00:01:02.049 National Laboratory, the Los Alamos National 00:01:02.049 --> 00:01:05.909 Lab, and Oak Ridge National Laboratories. as 00:01:05.909 --> 00:01:09.129 well as longstanding leadership roles. In this 00:01:09.129 --> 00:01:12.609 conversation, Jack shares his personal and professional 00:01:12.609 --> 00:01:16.310 journey from early academic challenges to becoming 00:01:16.310 --> 00:01:20.250 one of the world's influential figures in computational 00:01:20.250 --> 00:01:23.409 science, while offering deep insight into how 00:01:23.409 --> 00:01:26.670 computation has evolved from supporting infrastructure 00:01:26.670 --> 00:01:31.129 into a true scientific instrument. We explore 00:01:31.129 --> 00:01:34.170 the limits of modern supercomputing design, why 00:01:34.170 --> 00:01:36.709 most systems fail to achieve their theoretical 00:01:36.709 --> 00:01:39.969 performance, and why co -design, bringing together 00:01:39.969 --> 00:01:42.870 applications, algorithms, and hardware architecture, 00:01:43.230 --> 00:01:46.230 is essential for the next generation of scientific 00:01:46.230 --> 00:01:50.489 breakthroughs. As a scientific advisor at Kronos 00:01:50.489 --> 00:01:53.489 Fusion Energy, Jack plays a critical role in 00:01:53.489 --> 00:01:56.189 shaping how large -scale physics -based simulations 00:01:56.189 --> 00:01:59.909 are designed, validated, and trusted. especially 00:01:59.909 --> 00:02:02.969 in the pursuit of a proper digital twin for fusion 00:02:02.969 --> 00:02:07.090 energy systems. His perspective directly informs 00:02:07.090 --> 00:02:10.389 Chronos' approach to verification, validation, 00:02:10.889 --> 00:02:14.009 uncertainty quantification, and the responsible 00:02:14.009 --> 00:02:16.990 integration of AI with first principles numerical 00:02:16.990 --> 00:02:21.349 methods. Together, we discuss the future of high 00:02:21.349 --> 00:02:23.830 -performance computing, the complementary role 00:02:23.830 --> 00:02:27.840 of AI, the energy costs of next -generation supercomputers, 00:02:27.900 --> 00:02:30.599 and what it will take to build computing systems 00:02:30.599 --> 00:02:33.800 capable of supporting predictive, safety -critical 00:02:33.800 --> 00:02:37.460 simulations for fusion energy and beyond. This 00:02:37.460 --> 00:02:40.360 episode is both a masterclass in scientific computing 00:02:40.360 --> 00:02:44.319 and a deeply human story about curiosity, resilience, 00:02:44.659 --> 00:02:49.199 and the people behind the code. Here's Jack Dungara. 00:02:50.379 --> 00:02:53.939 All right, so let's see. Let me go back quite 00:02:53.939 --> 00:02:58.039 a bit. quite a ways here so um this my story 00:02:58.039 --> 00:03:01.919 is something like this my grandparents so all 00:03:01.919 --> 00:03:05.219 of my grandparents as well as my father were 00:03:05.219 --> 00:03:12.479 born in sicily so um italian is my heritage my 00:03:12.479 --> 00:03:15.780 grandparents come from two small villages in 00:03:15.780 --> 00:03:18.740 the middle of sicily one's called villa rosa 00:03:18.740 --> 00:03:21.479 and the other is called santa catarina And I 00:03:21.479 --> 00:03:24.879 had an opportunity to go visit those places a 00:03:24.879 --> 00:03:28.939 number of years ago. And they're isolated. It's 00:03:28.939 --> 00:03:32.120 mountainous. It's rocky. And I'm sure it was 00:03:32.120 --> 00:03:38.919 a very poor existence. My grandfather, my father's 00:03:38.919 --> 00:03:43.539 father, worked in a sulfur mine in Sicily. And 00:03:43.539 --> 00:03:45.639 I guess he was looking for a little bit better. 00:03:47.639 --> 00:03:54.580 And what they did is they decided to leave Sicily. 00:03:54.599 --> 00:04:00.439 And that was back in 1929. My grandfather was 00:04:00.439 --> 00:04:04.340 42 years old. And he took the family and went 00:04:04.340 --> 00:04:08.580 to the U .S. and arrived at Ellis Island. He 00:04:08.580 --> 00:04:11.900 had $25 in his pocket. I'm sure he was filled 00:04:11.900 --> 00:04:16.600 with a lot of hope for a new life. My father 00:04:16.600 --> 00:04:19.579 was 10 years old at that point when he made the 00:04:19.579 --> 00:04:24.800 journey. So they immigrated, they settled in 00:04:24.800 --> 00:04:27.060 Chicago, they knew some people there and they 00:04:27.060 --> 00:04:31.620 made their way to Chicago and settled in in the 00:04:31.620 --> 00:04:36.160 Italian neighborhood. My parents and my father 00:04:36.160 --> 00:04:40.480 met my mother in Chicago and she has a similar 00:04:40.480 --> 00:04:44.860 background. My grandparents on that side of the 00:04:44.860 --> 00:04:48.740 family also come from Sicily, of course. So my 00:04:48.740 --> 00:04:53.459 parents had a very basic education and they actually 00:04:53.459 --> 00:04:56.500 didn't finish high school. So they did have an 00:04:56.500 --> 00:04:59.459 understanding about education and college and 00:04:59.459 --> 00:05:02.980 knew those things were important. They also understood 00:05:02.980 --> 00:05:07.480 and modeled the importance of hard work. So I 00:05:07.480 --> 00:05:10.829 went to school. all through elementary school 00:05:10.829 --> 00:05:15.589 and high school, I sort of struggled with reading 00:05:15.589 --> 00:05:19.050 and spelling. I think I accelerated in math and 00:05:19.050 --> 00:05:22.810 science, but I was really having a hard time 00:05:22.810 --> 00:05:27.889 with reading, spelling, and everything that went 00:05:27.889 --> 00:05:31.829 along with that. And as an adult, I learned I 00:05:31.829 --> 00:05:35.160 have dyslexia. So that was probably one reason 00:05:35.160 --> 00:05:38.819 why I was not quite at the same level as my other 00:05:38.819 --> 00:05:45.639 peers. And I wanted to be a high school teacher. 00:05:45.879 --> 00:05:49.300 So that was my ambition when I graduated high 00:05:49.300 --> 00:05:52.899 school and decided I would go to college. So 00:05:52.899 --> 00:05:56.920 I was the first person in my family to go to 00:05:56.920 --> 00:05:59.660 college. I have two younger brothers, and my 00:05:59.660 --> 00:06:03.120 younger brothers are electricians. And they, 00:06:03.120 --> 00:06:06.300 you know, they have a trade. They went to college 00:06:06.300 --> 00:06:09.240 maybe for a year or two and decided it wasn't 00:06:09.240 --> 00:06:12.620 for them. And they decided to drop out and get 00:06:12.620 --> 00:06:16.620 a career as electricians. And they formed a company 00:06:16.620 --> 00:06:20.860 and were very successful in that. And I went 00:06:20.860 --> 00:06:24.920 on to college. I went to a place in Chicago where 00:06:24.920 --> 00:06:27.920 they trained the teachers for the Chicago public 00:06:27.920 --> 00:06:31.899 school system. And I did okay. I was probably 00:06:31.899 --> 00:06:35.139 an average student. I was still struggling with 00:06:35.139 --> 00:06:38.540 reading and spelling and writing. But I did excel 00:06:38.540 --> 00:06:43.139 in math and science. And I guess I stood out 00:06:43.139 --> 00:06:48.800 and my professors at college. encouraged me to 00:06:48.800 --> 00:06:52.720 apply to a program. And the program was to spend 00:06:52.720 --> 00:06:56.319 a semester with a scientist. So I was in the 00:06:56.319 --> 00:06:59.379 Chicago area and we have a Department of Energy 00:06:59.379 --> 00:07:02.639 laboratory in Chicago called Argonne National 00:07:02.639 --> 00:07:07.300 Lab. And I applied for this program and they 00:07:07.300 --> 00:07:10.240 accepted about 50 students from around the country. 00:07:10.579 --> 00:07:13.079 And I was one of those students. So I got to 00:07:13.079 --> 00:07:16.990 spend roughly 16 weeks. working at Argonne National 00:07:16.990 --> 00:07:22.050 Lab. Transformational in the sense that, you 00:07:22.050 --> 00:07:24.970 know, I discovered things that I had never really 00:07:24.970 --> 00:07:28.910 seen before. I was working with scientists. I 00:07:28.910 --> 00:07:35.410 was engaging in a form of research. And I found 00:07:35.410 --> 00:07:39.189 my passion in some sense. I found that that's 00:07:39.189 --> 00:07:43.089 really something I wanted to do. So I went on 00:07:43.089 --> 00:07:49.730 to... to get a master's degree. After the master's 00:07:49.730 --> 00:07:51.970 degree, I worked at that place, Argonne National 00:07:51.970 --> 00:07:55.129 Lab. They hired me full -time. I wasn't ready 00:07:55.129 --> 00:07:58.750 for a PhD. And over the course of working at 00:07:58.750 --> 00:08:02.490 Argonne, we had many visitors and those visitors 00:08:02.490 --> 00:08:08.850 came. I interacted with them and they were encouraging 00:08:08.850 --> 00:08:12.269 me to go back to school. So I had an interest 00:08:12.269 --> 00:08:16.290 in linear algebra. And I took up with one of 00:08:16.290 --> 00:08:19.829 the visitors from the University of New Mexico. 00:08:20.290 --> 00:08:24.410 And I was encouraged by Argonne National Lab 00:08:24.410 --> 00:08:27.449 to go to school. They paid me a little bit to 00:08:27.449 --> 00:08:32.389 do that. And I went and got a Ph .D. at the University 00:08:32.389 --> 00:08:34.629 of New Mexico. And while I was there, I was working 00:08:34.629 --> 00:08:38.350 at Los Alamos National Lab, which is close to 00:08:38.350 --> 00:08:43.950 Albuquerque. and had just a really door -opening 00:08:43.950 --> 00:08:47.309 experience with, again, another set of scientists 00:08:47.309 --> 00:08:52.870 working on high -performance computing. I was 00:08:52.870 --> 00:08:55.269 doing linear algebra, so I was trying to implement 00:08:55.269 --> 00:08:58.409 and design things that would run fast on those 00:08:58.409 --> 00:09:03.029 computers that we had at Los Alamos. And we had 00:09:03.029 --> 00:09:06.210 really some of the fastest computers at that 00:09:06.210 --> 00:09:09.960 time. were in use at Los Alamos and I got to 00:09:09.960 --> 00:09:14.340 really get a first hand first exposure to those 00:09:14.340 --> 00:09:17.500 to those computers. So that was sort of my path 00:09:17.500 --> 00:09:22.759 to where I ended up. That was my lead into what 00:09:22.759 --> 00:09:27.340 I would call my experience with with numerical 00:09:27.340 --> 00:09:30.700 computing and also high performance computing. 00:09:32.519 --> 00:09:34.639 That's pretty cool, Jack. Are these programs 00:09:34.639 --> 00:09:37.799 still available for people to like apply to out 00:09:37.799 --> 00:09:41.200 of high school and undergrad? Oh, yeah. Yeah, 00:09:41.240 --> 00:09:45.799 absolutely. And I I I encourage students to do 00:09:45.799 --> 00:09:49.320 that. I take in students. So I worked for many 00:09:49.320 --> 00:09:51.639 years at Argonne National Lab and I would have 00:09:51.639 --> 00:09:54.840 students interacting with me from that program. 00:09:55.100 --> 00:10:00.100 And when I moved to Tennessee, I I worked. at 00:10:00.100 --> 00:10:02.940 Oak Ridge National Lab, another Department of 00:10:02.940 --> 00:10:06.539 Energy laboratory, and also took in students 00:10:06.539 --> 00:10:11.960 in working with that. So I view it as a very 00:10:11.960 --> 00:10:18.279 engaging, very interesting way to open new minds 00:10:18.279 --> 00:10:21.879 and to expose people to some of the ideas that 00:10:21.879 --> 00:10:24.980 we're doing and having them sort of participate 00:10:24.980 --> 00:10:28.179 and contribute to the overall cause. So it's 00:10:28.179 --> 00:10:32.429 a very good exposure there is um uh you know 00:10:32.429 --> 00:10:35.309 a relationship that gets formed and you know 00:10:35.309 --> 00:10:37.789 i've i've been in contact with some of the students 00:10:37.789 --> 00:10:43.570 i helped mentor uh for many years very cool um 00:10:43.570 --> 00:10:46.649 i kind of wanted to tell you my my car license 00:10:46.649 --> 00:10:49.909 plate says math lab because like the program 00:10:49.909 --> 00:10:53.090 math lab has been so central in like everything 00:10:53.090 --> 00:10:57.159 i've done in life but i know Behind, I'd love 00:10:57.159 --> 00:10:59.720 to, I'll send you a picture someday, but behind 00:10:59.720 --> 00:11:05.080 that is a stack of mathematical computing, like 00:11:05.080 --> 00:11:08.720 baseline applications that you designed, right, 00:11:08.860 --> 00:11:11.360 Jack? Like how did you guys do that? How did 00:11:11.360 --> 00:11:13.960 you guys put that together? So I'm stunned that 00:11:13.960 --> 00:11:17.879 your license plate says MATLAB. I've been pulled 00:11:17.879 --> 00:11:21.139 over twice where they thought they said MATLAB. 00:11:21.159 --> 00:11:27.039 I see. yeah yeah i'm breaking bad right well 00:11:27.039 --> 00:11:29.580 that's that's so interesting so um so just to 00:11:29.580 --> 00:11:32.179 go back a little bit uh where i was telling the 00:11:32.179 --> 00:11:38.440 story um my phd advisor at new mexico um uh was 00:11:38.440 --> 00:11:41.919 his name is cleve moeller and cleve is the author 00:11:41.919 --> 00:11:46.120 of matlab so um so i have a very close relationship 00:11:46.120 --> 00:11:51.100 to um to the math works also to matlab i was 00:11:51.100 --> 00:11:55.320 at ground zero when cleve was creating it matlab 00:11:55.320 --> 00:11:58.740 uses the software that we've written over the 00:11:58.740 --> 00:12:05.100 years and it's integral to to how matlab does 00:12:05.100 --> 00:12:08.220 its computation so that's that's cool that you 00:12:08.220 --> 00:12:11.940 have a matlab matlab license plate my early license 00:12:11.940 --> 00:12:17.620 plate in new mexico on my car now we're talking 00:12:17.620 --> 00:12:21.919 back in the late 70s, was LINPACK. So LINPACK 00:12:21.919 --> 00:12:25.840 was the project that myself and Cleve Moeller 00:12:25.840 --> 00:12:28.919 and a couple other people, Pete Stewart and Jim 00:12:28.919 --> 00:12:31.620 Bunch, worked on. It's a package of linear algebra 00:12:31.620 --> 00:12:34.879 software for solving systems of linear equations. 00:12:35.240 --> 00:12:39.240 And that LINPACK became the basis of MATLAB. 00:12:39.539 --> 00:12:43.360 Later, it was transitioned to a much more updated 00:12:43.360 --> 00:12:47.620 version. called LAPAC, which I was responsible 00:12:47.620 --> 00:12:52.480 for. And so every time I touch MATLAB, I know 00:12:52.480 --> 00:12:57.240 I'm using my software and hopefully it's working 00:12:57.240 --> 00:13:00.879 correctly. You've been using it here for the 00:13:00.879 --> 00:13:03.000 last, I've been using it for the last decade 00:13:03.000 --> 00:13:05.340 and trying to apply it to fusion energy and it's 00:13:05.340 --> 00:13:09.799 been going great. Yeah, I'll show it to you sometimes. 00:13:09.980 --> 00:13:15.879 And I feel Like, I should not have this license 00:13:15.879 --> 00:13:18.259 plate, so if either of you want it, I will transfer 00:13:18.259 --> 00:13:22.700 ownership. And I absolutely love that we both 00:13:22.700 --> 00:13:25.840 express our love for things through license plates. 00:13:26.059 --> 00:13:29.379 That's terrific. That's really great. That's 00:13:29.379 --> 00:13:33.460 a rare thing to have in common. So where are 00:13:33.460 --> 00:13:36.340 we going with simulations, do you think, in terms 00:13:36.340 --> 00:13:40.480 of large systems that use supercomputing, parallel 00:13:40.480 --> 00:13:45.090 computing, you know? nanotechnology when it comes 00:13:45.090 --> 00:13:49.529 to microchips and data collection. Where are 00:13:49.529 --> 00:13:51.509 we, Jack, and where are we going, you think? 00:13:51.970 --> 00:13:54.649 Well, just at the beginnings in some sense, you 00:13:54.649 --> 00:14:00.250 know, we treat computing, many people treat computing 00:14:00.250 --> 00:14:03.049 as infrastructure. And, you know, it's the thing 00:14:03.049 --> 00:14:05.710 behind the scenes that lets us solve equations 00:14:05.710 --> 00:14:10.210 faster. And, you know, the shift to seeing computing 00:14:10.210 --> 00:14:14.769 as a scientific instrument. is happening gradually. 00:14:15.250 --> 00:14:18.210 And there's a number of things that could realize. 00:14:18.350 --> 00:14:21.049 First, you know, computing isn't neutral. The 00:14:21.049 --> 00:14:23.850 algorithms, the discretization, the numerical 00:14:23.850 --> 00:14:28.490 stability, these things, these are choices that 00:14:28.490 --> 00:14:31.529 shape what we observe. And in some sense, the 00:14:31.529 --> 00:14:36.590 code and the algorithms are... part of the instrument 00:14:36.590 --> 00:14:40.809 so if we solve something that's unstable or frowning 00:14:40.809 --> 00:14:43.309 errors dominate you're not measuring the phenomenon 00:14:43.309 --> 00:14:46.129 you thought you were measuring you're measuring 00:14:46.129 --> 00:14:49.549 artifacts of the method and that's um that's 00:14:49.549 --> 00:14:53.129 exactly the same thing as a miscalibrated i don't 00:14:53.129 --> 00:14:56.409 know sensor in some sense and you know the biggest 00:14:56.409 --> 00:15:00.429 scientific payoffs come when computation enables 00:15:00.429 --> 00:15:04.870 new questions not just faster answers, when simulation 00:15:04.870 --> 00:15:09.909 starts to replace or complement experiments in 00:15:09.909 --> 00:15:14.649 the early design stages. And when it becomes 00:15:14.649 --> 00:15:18.610 predictive and not merely descriptive, that's 00:15:18.610 --> 00:15:22.289 when it feels like an instrument. So a telescope 00:15:22.289 --> 00:15:24.889 just doesn't confirm what we already know. It 00:15:24.889 --> 00:15:28.299 lets you see what you couldn't see before. And 00:15:28.299 --> 00:15:32.379 large -scale computing began to do that for complex 00:15:32.379 --> 00:15:35.779 systems. So once you accept the instrument framing, 00:15:36.059 --> 00:15:38.840 the priorities change. You start thinking about 00:15:38.840 --> 00:15:42.500 calibration, about error bars, about reproducibility, 00:15:42.740 --> 00:15:46.440 about controlling experiments and software. You 00:15:46.440 --> 00:15:52.720 build test suites that sort of help you understand 00:15:52.720 --> 00:15:57.600 and calibrate runs. You preserve. reference solutions, 00:15:57.679 --> 00:16:00.620 so you can do checking. You design algorithms 00:16:00.620 --> 00:16:04.679 to behave reliably under stress, and you become 00:16:04.679 --> 00:16:09.419 very cautious about claiming results without 00:16:09.419 --> 00:16:14.500 really understanding the numerical sensitivities. 00:16:14.740 --> 00:16:18.200 So from the instrument mind sense, it also forces 00:16:18.200 --> 00:16:24.649 us to think. Not just ask, did it run, but ask, 00:16:24.750 --> 00:16:27.649 you know, did it measure something real? And 00:16:27.649 --> 00:16:30.889 with AI, you know, that's really complementing 00:16:30.889 --> 00:16:34.309 what we do today in terms of the numerics. It 00:16:34.309 --> 00:16:38.789 allows us to really go beyond that. AI is really 00:16:38.789 --> 00:16:43.830 a very important tool that we will use in helping 00:16:43.830 --> 00:16:46.070 us solve problems. It's not going to replace 00:16:46.070 --> 00:16:49.929 the numerical methods. It's going to be used 00:16:49.929 --> 00:16:52.389 in conjunction with those with those methods 00:16:52.389 --> 00:16:55.210 and do something. I think that'll really help 00:16:55.210 --> 00:16:59.169 us transform The way in which we approach and 00:16:59.169 --> 00:17:01.950 tackle some of the most challenging problems 00:17:01.950 --> 00:17:06.089 that we have I like that you said telescopes 00:17:06.089 --> 00:17:08.930 because I was reading an article about in video 00:17:08.930 --> 00:17:12.589 being heavily involved in the Vera Rubin Observatory 00:17:12.589 --> 00:17:17.490 and I was well, what is the big What is the big 00:17:17.490 --> 00:17:20.569 data processing component of that? Is it putting 00:17:20.569 --> 00:17:23.509 thousands of pictures together? What are they 00:17:23.509 --> 00:17:26.109 doing there that requires the supercomputing? 00:17:26.130 --> 00:17:28.750 Because that's a really reliable application, 00:17:28.970 --> 00:17:33.009 it seems. Oh, yeah, it's tremendous. Tremendous 00:17:33.009 --> 00:17:36.210 capabilities are being given to us. You know, 00:17:36.210 --> 00:17:38.369 when you think about astronomy and think about 00:17:38.369 --> 00:17:42.230 looking at changes and how those changes are 00:17:42.230 --> 00:17:46.599 affected over. over time, we can scoop up all 00:17:46.599 --> 00:17:50.619 of this data and quickly analyze and see differences 00:17:50.619 --> 00:17:53.880 between things and understand things at a very 00:17:53.880 --> 00:17:56.759 fine level that we couldn't really do before. 00:17:57.039 --> 00:18:01.019 So it allows us to really carry out these experiments 00:18:01.019 --> 00:18:05.140 in a way that, in a much faster way, in a more 00:18:05.140 --> 00:18:08.779 inductive way than we could have from before. 00:18:09.019 --> 00:18:13.079 But we have to be cautious. So when we use AI 00:18:13.079 --> 00:18:17.859 to help us navigate through this data, one thing 00:18:17.859 --> 00:18:21.299 AI is good at is predicting the next thing based 00:18:21.299 --> 00:18:23.799 on all of the information that it's seen before. 00:18:24.119 --> 00:18:26.700 But if it sees something that it hasn't seen 00:18:26.700 --> 00:18:29.700 before, it has a hard time figuring out what 00:18:29.700 --> 00:18:33.000 it should do next. And that's where we're often 00:18:33.000 --> 00:18:40.630 led into an area where it starts to... hallucinate 00:18:40.630 --> 00:18:44.690 and perhaps get into trouble. So, you know, building 00:18:44.690 --> 00:18:48.730 things in conjunction with the standard way in 00:18:48.730 --> 00:18:53.109 which we model stuff really helps us in getting 00:18:53.109 --> 00:18:55.690 past some of those difficulties. So applying 00:18:55.690 --> 00:18:59.049 the physical laws which may help us predict those 00:18:59.049 --> 00:19:03.730 maybe uncertain things or rare events allows 00:19:03.730 --> 00:19:07.930 us to work in conjunction with of the AI. So 00:19:07.930 --> 00:19:10.410 I think it's going to play an important role. 00:19:10.650 --> 00:19:13.470 We see that in many areas. Astronomy is one area, 00:19:13.690 --> 00:19:16.690 but there are many more areas where this will 00:19:16.690 --> 00:19:20.990 help us and drive us forward and put us on a 00:19:20.990 --> 00:19:24.910 much better footing, get us to solutions or getting 00:19:24.910 --> 00:19:27.630 us to a point where we can ask questions at a 00:19:27.630 --> 00:19:33.170 deeper level and hopefully come up with new answers 00:19:33.170 --> 00:19:37.000 to the problems. What do you think we're going 00:19:37.000 --> 00:19:41.400 to solve with high -performance computing in 00:19:41.400 --> 00:19:44.680 the next decade that nobody thinks, it's not 00:19:44.680 --> 00:19:48.559 in anybody's purview? Like, what is that? Yeah, 00:19:48.559 --> 00:19:50.900 what is that indeed? So if we had some of those 00:19:50.900 --> 00:19:53.660 answers, we would know. I hope it's communication 00:19:53.660 --> 00:19:57.420 with animals. That's what I'm rooting for, and 00:19:57.420 --> 00:20:01.059 nobody's talking about this. That's my secret 00:20:01.059 --> 00:20:04.500 wish. okay well okay that's uh that's something 00:20:04.500 --> 00:20:07.160 that perhaps is tractable with the use of high 00:20:07.160 --> 00:20:09.359 performance computing and uh you know trying 00:20:09.359 --> 00:20:12.599 to understand better how those patterns can fit 00:20:12.599 --> 00:20:14.940 together and how we can match those patterns 00:20:14.940 --> 00:20:18.839 to something that we might be able to express 00:20:18.839 --> 00:20:22.779 and hopefully from that exchange information 00:20:22.779 --> 00:20:26.539 with animals. So yes, so perhaps that's a possible 00:20:26.539 --> 00:20:31.599 solution to some of these problems. But there 00:20:31.599 --> 00:20:34.779 are so many important issues that we need to 00:20:34.779 --> 00:20:40.799 address from that standpoint. um that the high 00:20:40.799 --> 00:20:43.839 performance computing allow us to really uh to 00:20:43.839 --> 00:20:46.700 get to so high performance computing you know 00:20:46.700 --> 00:20:51.339 today we think of it as a way to um to come up 00:20:51.339 --> 00:20:54.319 with the solution faster and come up with the 00:20:54.319 --> 00:20:57.400 solution with greater fidelity and come up with 00:20:57.400 --> 00:21:01.869 a way of optimizing the solutions that we have. 00:21:02.109 --> 00:21:05.710 So we really use high performance computing to 00:21:05.710 --> 00:21:09.269 do those three things. So when I say optimizing, 00:21:09.670 --> 00:21:16.269 we think about running multiple versions of an 00:21:16.269 --> 00:21:20.210 application, an ensemble calculation where we 00:21:20.210 --> 00:21:23.549 make small changes to the parameters and look 00:21:23.549 --> 00:21:26.829 at what the outcome is. And from that, we can 00:21:26.829 --> 00:21:30.230 perhaps choose a design of something which is 00:21:30.230 --> 00:21:33.150 optimal for the case that we have and that's 00:21:33.150 --> 00:21:40.369 used in many cases it's used for optimizing something 00:21:40.369 --> 00:21:44.269 to make it safer for instance a car so we're 00:21:44.269 --> 00:21:47.430 gonna we're gonna make many runs of this model 00:21:47.430 --> 00:21:51.069 of a car crashing into a wall and we adjust the 00:21:51.069 --> 00:21:55.859 parameters so that the car survives the crash 00:21:55.859 --> 00:21:58.339 in a way that allows the human in the car to 00:21:58.339 --> 00:22:02.079 walk away. And through that, we can come up with 00:22:02.079 --> 00:22:06.099 better methods. So that's just one way in which 00:22:06.099 --> 00:22:11.980 these are being used for optimizing and making 00:22:11.980 --> 00:22:15.380 things much more safe and perhaps making us more 00:22:15.380 --> 00:22:23.660 productive. Is there, like, a dream theory for 00:22:23.660 --> 00:22:26.779 a supercomputer that we have not built yet? Like, 00:22:26.839 --> 00:22:30.400 there are a few, I would think. What is, like, 00:22:30.880 --> 00:22:35.140 the one that if we, other than money, what is 00:22:35.140 --> 00:22:40.039 the holdup in building that one? Right. So, you 00:22:40.039 --> 00:22:43.640 know, so today we have this really, I think, 00:22:43.660 --> 00:22:47.480 unfortunate situation. So when we buy a supercomputer. 00:22:48.359 --> 00:22:51.460 We have a certain amount of money and we have 00:22:51.460 --> 00:22:54.920 a certain target performance that we would like 00:22:54.920 --> 00:22:59.039 to see. So those are the two poles in the tent, 00:22:59.140 --> 00:23:03.359 if you will. And those poles will dictate the 00:23:03.359 --> 00:23:06.019 kind of computer that we get. So we make a request 00:23:06.019 --> 00:23:08.839 for proposals. We go out to vendors and say, 00:23:08.880 --> 00:23:10.599 we have this much money and we want this much 00:23:10.599 --> 00:23:13.140 performance. And the vendors cobble together 00:23:13.140 --> 00:23:15.660 a machine. And what they do is they get a machine 00:23:15.660 --> 00:23:19.900 which... uses commodity parts, commodity parts 00:23:19.900 --> 00:23:22.400 meaning standard off -the -shelf processors, 00:23:22.859 --> 00:23:26.400 standard off -the -shelf interconnect networks, 00:23:26.700 --> 00:23:30.660 and put them together in a box. And that box 00:23:30.660 --> 00:23:33.740 has a price tag where it matches what we have 00:23:33.740 --> 00:23:37.420 in terms of the funds available, and it'll match 00:23:37.420 --> 00:23:42.579 the peak performance of what we're asking for. 00:23:42.920 --> 00:23:47.359 So those two things are good. But unfortunately, 00:23:47.720 --> 00:23:50.240 that peak performance is going to be hard to 00:23:50.240 --> 00:23:54.200 realize in many applications. So, you know, our 00:23:54.200 --> 00:23:57.980 dirty secret in terms of high performance computing 00:23:57.980 --> 00:24:01.339 or scientific computing is that we typically 00:24:01.339 --> 00:24:06.900 achieve less than or equal to 10 % of the theoretical 00:24:06.900 --> 00:24:09.720 peak performance of these computers. That is 00:24:09.720 --> 00:24:13.650 the high performance computing. The supercomputers 00:24:13.650 --> 00:24:17.269 have a rated peak performance, and our applications, 00:24:17.470 --> 00:24:20.690 when implemented on them, have achieved only 00:24:20.690 --> 00:24:25.309 a small fraction of that potential. And the reason 00:24:25.309 --> 00:24:29.329 for that is the machines are basically misdesigned. 00:24:29.329 --> 00:24:32.849 They're misdesigned because they're designed 00:24:32.849 --> 00:24:34.930 to give us high performance. High performance 00:24:34.930 --> 00:24:38.589 means, in this context, floating point performance, 00:24:39.269 --> 00:24:42.880 how many flops we get. floating point operations 00:24:42.880 --> 00:24:47.940 per second. And the reality is that on our computers 00:24:47.940 --> 00:24:52.160 today, we are over provisioned for flops. So 00:24:52.160 --> 00:24:56.240 the flop count is very high. But what's missing 00:24:56.240 --> 00:24:59.400 is the part where we move data to the part of 00:24:59.400 --> 00:25:01.299 the computer where we can do the operations. 00:25:01.480 --> 00:25:05.099 So data movement is very expensive on our computers. 00:25:05.380 --> 00:25:09.299 So we are not building systems which have enough. 00:25:09.769 --> 00:25:13.470 bandwidth to meet the demands of the applications. 00:25:13.930 --> 00:25:18.470 And the result is that we achieve only 10 % of 00:25:18.470 --> 00:25:21.670 that peak performance in our computing systems. 00:25:22.170 --> 00:25:25.269 So that's a problem which is very hard to solve. 00:25:25.490 --> 00:25:28.210 We've invested in floating point operations, 00:25:28.490 --> 00:25:31.589 but we've not done this similar investment in 00:25:31.589 --> 00:25:34.690 terms of data movement, which perhaps is more 00:25:34.690 --> 00:25:38.519 expensive. And the result is this mismatch. So 00:25:38.519 --> 00:25:41.319 if I was designing a computer, I would design 00:25:41.319 --> 00:25:46.440 it with co -design in mind. So co -design is 00:25:46.440 --> 00:25:50.000 where we get the application people together 00:25:50.000 --> 00:25:53.119 with the architects, together with the algorithms 00:25:53.119 --> 00:25:55.839 people, together with the software designers 00:25:55.839 --> 00:25:59.859 and build a system. which matches the needs of 00:25:59.859 --> 00:26:02.900 the applications rather than matches the needs 00:26:02.900 --> 00:26:05.420 of the money that we have and the expectation 00:26:05.420 --> 00:26:09.680 for this large peak performance. So building 00:26:09.680 --> 00:26:12.140 the system based on that would certainly give 00:26:12.140 --> 00:26:16.279 us a much better return. And what's important 00:26:16.279 --> 00:26:18.920 in scientific computing is the scientist's time. 00:26:19.059 --> 00:26:22.380 It's not the time on the computer. So we really 00:26:22.380 --> 00:26:25.579 don't care about how much computing time it takes 00:26:25.579 --> 00:26:29.339 compared to how much time we can save by having 00:26:29.339 --> 00:26:32.140 something which is much simpler to use and which 00:26:32.140 --> 00:26:35.859 provides us with a solution in a much better 00:26:35.859 --> 00:26:39.519 fashion. So this co -design is something that 00:26:39.519 --> 00:26:42.420 we really need to follow, and we are not doing 00:26:42.420 --> 00:26:45.980 it in terms of scientific computing. And today, 00:26:46.259 --> 00:26:50.819 you know, with AI and with the AI companies, 00:26:51.480 --> 00:26:54.799 having this tremendous capital that they can 00:26:54.799 --> 00:26:59.779 they can spend on hardware to satisfy the applications 00:26:59.779 --> 00:27:03.640 that are being driven by ai we see a situation 00:27:03.640 --> 00:27:09.839 where the ai uh super kings are are able to design 00:27:09.839 --> 00:27:13.579 their hardware to do the applications that they 00:27:13.579 --> 00:27:16.619 need so they are doing co -design they are designing 00:27:16.619 --> 00:27:19.799 their own hardware so you take a look at amazon 00:27:19.799 --> 00:27:23.710 they have their own hardware design google has 00:27:23.710 --> 00:27:28.430 their own systems the tpus microsoft has their 00:27:28.430 --> 00:27:32.910 own systems they're all designing equipment hardware 00:27:32.910 --> 00:27:36.150 to match the needs of their of their application 00:27:36.150 --> 00:27:38.650 putting them in a much better position but they 00:27:38.650 --> 00:27:41.309 have the resources to do that you know we see 00:27:41.309 --> 00:27:45.390 companies that are on the four or five trillion 00:27:45.390 --> 00:27:49.809 dollars of uh of capital of the of their market 00:27:49.809 --> 00:27:52.940 capitalization So they have tremendous resources 00:27:52.940 --> 00:27:55.900 that they can invest in the hardware to get them 00:27:55.900 --> 00:27:58.740 to a place where they can actually design their 00:27:58.740 --> 00:28:01.559 own hardware. In the scientific area, we don't 00:28:01.559 --> 00:28:04.920 have quite the same set of resources. So we're 00:28:04.920 --> 00:28:08.779 in some sense stuck. And there's a new model 00:28:08.779 --> 00:28:12.539 that's being formed. It's a public -private partnership 00:28:12.539 --> 00:28:16.319 that's being encouraged by the government. So 00:28:16.319 --> 00:28:20.000 the president had an executive order in November. 00:28:20.619 --> 00:28:24.660 which said that there would be this partnership 00:28:24.660 --> 00:28:28.019 between the Department of Energy and companies 00:28:28.019 --> 00:28:33.079 to let companies like Cisco deploy an AI system 00:28:33.079 --> 00:28:38.259 on government land. The government would supply 00:28:38.259 --> 00:28:44.859 the power, the electric power, and Cisco is going 00:28:44.859 --> 00:28:49.119 to supply the hardware and have that in existence 00:28:49.119 --> 00:28:53.430 to help. meet some of the needs of the Department 00:28:53.430 --> 00:28:58.069 of Energy need for high performance computing. 00:28:58.390 --> 00:29:02.750 And Cisco would then also be selling AI services 00:29:02.750 --> 00:29:06.990 to the rest of the community. So it's a new kind 00:29:06.990 --> 00:29:10.750 of relationship that's being formed. This is 00:29:10.750 --> 00:29:14.769 a private public partnership. It's being exploited 00:29:14.769 --> 00:29:18.890 in Europe. They have these AI gigafactories, 00:29:18.910 --> 00:29:23.109 which are similar. in terms of the ability to 00:29:23.109 --> 00:29:27.029 solve large complex problems that are being used 00:29:27.029 --> 00:29:33.549 both for ai purposes solely as well as for meeting 00:29:33.549 --> 00:29:39.349 the needs of the scientific community right i 00:29:39.349 --> 00:29:44.670 like that um so basically build for for a purpose 00:29:44.670 --> 00:29:48.450 and then not be limited by buying dollar amounts 00:29:48.450 --> 00:29:52.440 and you can build teams that are capable of great 00:29:52.440 --> 00:29:55.859 things yeah um so kind of just backwards engineer 00:29:55.859 --> 00:30:00.619 like game it not from the budget but i see i 00:30:00.619 --> 00:30:04.380 feel like um i feel like private companies i 00:30:04.380 --> 00:30:09.920 guess have more more spend money spending money 00:30:09.920 --> 00:30:14.940 to do stuff like that um i'm i would be surprised 00:30:14.940 --> 00:30:18.940 if uh nvidia wasn't on top of such things Well, 00:30:18.940 --> 00:30:21.019 NVIDIA is supplying the hardware, of course, 00:30:21.079 --> 00:30:23.680 for that. Ultimately, they're going to be using 00:30:23.680 --> 00:30:29.700 accelerators, GPUs to really push. So there's 00:30:29.700 --> 00:30:32.440 not a big difference between the kind of computers 00:30:32.440 --> 00:30:35.539 we use in the scientific area and the kind of 00:30:35.539 --> 00:30:40.960 computers that are needed for AI. And that's 00:30:40.960 --> 00:30:45.009 what we're seeing. is this balancing act between 00:30:45.009 --> 00:30:47.710 between them the government doesn't have enough 00:30:47.710 --> 00:30:51.190 funding to buy the kind of computers that those 00:30:51.190 --> 00:30:56.130 commercial companies can invest in and the result 00:30:56.130 --> 00:31:02.609 is this kind of partnership yeah um you you are 00:31:02.609 --> 00:31:07.720 kind of a regulations guy if I got that right, 00:31:07.839 --> 00:31:12.099 but how do you feel AI is being regulated now 00:31:12.099 --> 00:31:14.960 versus like, what should we do? Should we let 00:31:14.960 --> 00:31:19.259 her lose? Should we put like sport leashes on 00:31:19.259 --> 00:31:23.380 these things? I don't know. Well, AI is such 00:31:23.380 --> 00:31:29.200 a tremendous tool and as a tool that can be used 00:31:29.200 --> 00:31:32.940 for good and bad things as many. uh many things 00:31:32.940 --> 00:31:36.500 that we create that has both pluses and minuses 00:31:36.500 --> 00:31:40.740 and we need some guard rails or controls so that 00:31:40.740 --> 00:31:44.180 we don't we don't depart in a substantial way 00:31:44.180 --> 00:31:48.319 but you know it needs to it needs to evolve it 00:31:48.319 --> 00:31:51.799 needs to be we need to understand what the capabilities 00:31:51.799 --> 00:31:55.279 are what the limitations are we need to understand 00:31:55.279 --> 00:31:58.200 better you know where we can get in trouble and 00:31:58.200 --> 00:32:01.779 how those issues could be resolved and what we 00:32:01.779 --> 00:32:05.059 can do to, I won't say control, but to put some 00:32:05.059 --> 00:32:12.319 regulation, some kind of check and balance on 00:32:12.319 --> 00:32:16.859 how these tools get used and deployed. Without 00:32:16.859 --> 00:32:19.380 that, you know, I think we'll just have chaos 00:32:19.380 --> 00:32:24.680 ultimately. So we do need some oversight in these 00:32:24.680 --> 00:32:29.750 things. I'm assuming that when you were back 00:32:29.750 --> 00:32:34.029 in Los Alamos back when you were working on nuclear 00:32:34.029 --> 00:32:36.930 simulations, that's what that's what I feel like 00:32:36.930 --> 00:32:39.609 you'd be where you don't have to tell me. But 00:32:39.609 --> 00:32:46.190 I feel how what is our accuracy now when it comes 00:32:46.190 --> 00:32:50.930 to simulations of large. Complex multivariant 00:32:50.930 --> 00:32:56.559 systems. Right, so so we have. uh you know we 00:32:56.559 --> 00:33:00.740 have a number of things simulation the physics 00:33:00.740 --> 00:33:04.519 is grounded and and you know we have conservative 00:33:04.519 --> 00:33:08.440 laws the conservation laws that is we have known 00:33:08.440 --> 00:33:14.559 constraints we have tested hypothesis and you 00:33:14.559 --> 00:33:17.960 know we have to we have to put everything together 00:33:17.960 --> 00:33:21.619 ai gives us speed and pattern extraction and 00:33:21.619 --> 00:33:25.289 powerful approximations uh when we need to explore 00:33:25.289 --> 00:33:29.309 you know large design spaces um you know we have 00:33:29.309 --> 00:33:33.349 a number of things that ai can help with in terms 00:33:33.349 --> 00:33:38.230 of surrogate models that approximate expensive 00:33:38.230 --> 00:33:43.650 simulations for rapid exploration we have reduced 00:33:43.650 --> 00:33:46.930 order models that capture the domain behavior 00:33:46.930 --> 00:33:50.609 we have inverse problems that we need to solve 00:33:51.119 --> 00:33:55.960 And we can extract parameters and AI can accelerate 00:33:55.960 --> 00:34:00.519 and search some of these things and get the correct 00:34:00.519 --> 00:34:06.259 fittings. We have control in real time decision 00:34:06.259 --> 00:34:10.639 support where we need fast prediction. We have 00:34:10.639 --> 00:34:14.860 guided simulation campaigns where we decide what 00:34:14.860 --> 00:34:18.210 to do based on the current. uh situation and 00:34:18.210 --> 00:34:21.849 the uncertainty or the expected information gain 00:34:21.849 --> 00:34:24.269 so you know there are many things that that go 00:34:24.269 --> 00:34:27.510 into this and and somehow we we have to put them 00:34:27.510 --> 00:34:30.210 together and come up with solutions that are 00:34:30.210 --> 00:34:34.909 meaningful in the long run and allow us to have 00:34:34.909 --> 00:34:40.349 a have faith in the methods that we that we're 00:34:40.349 --> 00:34:44.230 using to solve these problems. So there are really 00:34:44.230 --> 00:34:50.590 many areas and many things that we need to refine. 00:34:51.190 --> 00:34:56.769 And we really need to fall back on our principles 00:34:56.769 --> 00:35:02.070 of investigation and looking at how things fit 00:35:02.070 --> 00:35:06.110 together and understanding when we really can 00:35:06.110 --> 00:35:11.030 get into trouble. We need some way to help us 00:35:11.030 --> 00:35:16.210 navigate when we get a solution to a problem, 00:35:16.329 --> 00:35:19.809 that that solution is, in fact, something we 00:35:19.809 --> 00:35:22.909 can believe in or trust. So, you know, there 00:35:22.909 --> 00:35:24.849 are many, many things that we have to do along 00:35:24.849 --> 00:35:28.110 the way to get to that point. Really, too far 00:35:28.110 --> 00:35:34.389 off. What would say 99 .5 % reliability? What 00:35:34.389 --> 00:35:38.940 is it like? if we were to pick a low -hanging 00:35:38.940 --> 00:35:46.480 fruit application, what would it be? Sorry, for 00:35:46.480 --> 00:35:50.800 what? For what purpose? Yeah, so let's say the 00:35:50.800 --> 00:35:55.300 telescope that we spoke of before, what would 00:35:55.300 --> 00:36:00.199 that... What's the, if you were to simulate something 00:36:00.199 --> 00:36:02.980 like that or the next generation of the VR laboratory 00:36:02.980 --> 00:36:06.000 or something like that, the telescope, what is 00:36:06.000 --> 00:36:09.119 our level of simulation accuracy? Would it be 00:36:09.119 --> 00:36:15.119 like 99 .5 %? Is it at 95 %? Where are we? How 00:36:15.119 --> 00:36:18.159 accurately can we predict these, like a digital 00:36:18.159 --> 00:36:23.679 twin predict an actual, a real model? So it's 00:36:23.679 --> 00:36:26.380 all going to depend on the number of... a number 00:36:26.380 --> 00:36:29.019 of issues which are complicated which relate 00:36:29.019 --> 00:36:33.199 to the um to the problem itself so you know i 00:36:33.199 --> 00:36:36.059 like to think of it in terms of a framework so 00:36:36.059 --> 00:36:39.760 there's verification so we need to understand 00:36:39.760 --> 00:36:42.900 the the mathematics and the implications that 00:36:42.900 --> 00:36:46.659 we have uh are we solving the correct set of 00:36:46.659 --> 00:36:49.599 equations are the discretizations implemented 00:36:49.599 --> 00:36:54.460 as intended uh do we have some kind of convergence 00:36:54.460 --> 00:36:58.760 uh under refinement and you know where are we 00:36:58.760 --> 00:37:03.079 stable uh under some perturbations so you know 00:37:03.079 --> 00:37:05.159 there are a number of things that uh that come 00:37:05.159 --> 00:37:09.420 along that really give us confidence in the solutions 00:37:09.420 --> 00:37:12.139 are we getting what we expect in terms of the 00:37:12.139 --> 00:37:15.719 order of accuracy verification is where you know 00:37:15.719 --> 00:37:20.460 we use manufacturing solutions uh analytic benchmarks 00:37:20.460 --> 00:37:24.500 and code to code comparisons to get there and 00:37:24.500 --> 00:37:26.760 you know beyond that there's validation you know 00:37:26.760 --> 00:37:29.579 it's about the physics are these the right equations 00:37:29.579 --> 00:37:32.800 and models for this kind of regime that we're 00:37:32.800 --> 00:37:36.019 looking at do the results match the experiments 00:37:36.019 --> 00:37:39.760 and the data that we have are the boundary conditions 00:37:39.760 --> 00:37:43.980 and the materials modeling realistic you know 00:37:43.980 --> 00:37:47.420 validation is hard because the experiments have 00:37:47.420 --> 00:37:51.480 uncertainty too and you know, the whole regime 00:37:51.480 --> 00:37:54.539 where we're looking at may be unacceptable. And 00:37:54.539 --> 00:37:57.980 then comes uncertainty quantification, where 00:37:57.980 --> 00:38:01.360 even if the model is right, what's the uncertainty 00:38:01.360 --> 00:38:04.019 from the parameters, the initial conditions, 00:38:04.360 --> 00:38:07.539 the discretizations that were chosen, and the 00:38:07.539 --> 00:38:12.199 subgrid that we have. So, you know, modern simulations, 00:38:12.199 --> 00:38:16.800 this pipeline needs error bars in the same way 00:38:16.800 --> 00:38:20.320 experiments need. uh physical experiments need 00:38:20.320 --> 00:38:23.719 error bars so you know there's um there's a number 00:38:23.719 --> 00:38:27.579 of things that come into the uh analysis that 00:38:27.579 --> 00:38:31.780 would lead to giving us that number 99 .5 percent 00:38:31.780 --> 00:38:36.500 uh effectiveness and it really depends on a number 00:38:36.500 --> 00:38:38.980 of things along the way so i can't really quite 00:38:38.980 --> 00:38:43.800 quantify it for that telescope but um you know 00:38:43.800 --> 00:38:48.030 there's there's things that people have to put 00:38:48.030 --> 00:38:53.090 together to help us in getting to that solution. 00:38:53.809 --> 00:38:59.710 So people often underestimate how often we lose 00:38:59.710 --> 00:39:04.150 track of the truth because the runs can't be 00:39:04.150 --> 00:39:07.889 reproduced or because of tiny changes in the 00:39:07.889 --> 00:39:12.369 version or the compilers or the numerical results 00:39:12.369 --> 00:39:16.269 have shifted. So if the simulation that we're 00:39:16.269 --> 00:39:19.510 doing is going to replace the experiment in the 00:39:19.510 --> 00:39:23.429 early design, then that simulation pipeline must 00:39:23.429 --> 00:39:26.849 behave like an instrument. It has to be calibrated, 00:39:26.929 --> 00:39:30.909 documented, and we need some way to audit what's 00:39:30.909 --> 00:39:37.170 going on. Right. Yeah. And there are lots of 00:39:37.170 --> 00:39:40.670 applications that are built in that kind of keep 00:39:40.670 --> 00:39:44.150 it honest. you know my math lab simulations that 00:39:44.150 --> 00:39:50.750 i built but um so in terms of like just if we 00:39:50.750 --> 00:39:53.789 had a magical supercomputer 25 years from now 00:39:53.789 --> 00:39:55.769 what is like the big thing that we would have 00:39:55.769 --> 00:39:59.530 solved for humanity if we got it today how would 00:39:59.530 --> 00:40:03.019 that translate into a solution 25 years from 00:40:03.019 --> 00:40:06.300 now right so you know there's so many ways these 00:40:06.300 --> 00:40:08.539 computers are used it's hard to pinpoint one 00:40:08.539 --> 00:40:11.039 thing okay so if we're going to design something 00:40:11.039 --> 00:40:13.199 we would want to design a drug we want to be 00:40:13.199 --> 00:40:16.760 able to have drug discovery so it's being you 00:40:16.760 --> 00:40:19.199 know we're using that today for you know coming 00:40:19.199 --> 00:40:22.510 up with new drugs that can effectively combat 00:40:22.510 --> 00:40:25.969 something. So if we had a wish, we would wish 00:40:25.969 --> 00:40:30.329 that we can come up with drugs that would effectively 00:40:30.329 --> 00:40:36.929 combat cancer. So cancer is something which has 00:40:36.929 --> 00:40:39.550 many forms, and it's not going to be one drug. 00:40:39.710 --> 00:40:42.449 It's going to be multiple drugs that go after 00:40:42.449 --> 00:40:46.949 those multiple forms of cancer. So in 25 years, 00:40:47.070 --> 00:40:49.949 if we had the right stuff, maybe we would be 00:40:49.949 --> 00:40:55.159 in a better position to to detect, treat, and 00:40:55.159 --> 00:40:59.539 to overcome, you know, a disease like that. But 00:40:59.539 --> 00:41:02.179 we can point to many diseases, cancer being one 00:41:02.179 --> 00:41:06.659 that perhaps is the most maybe widespread and 00:41:06.659 --> 00:41:13.099 affects many, many people here. I hear 2026 is 00:41:13.099 --> 00:41:16.690 the year of singularity. Where are you on this? 00:41:16.750 --> 00:41:21.510 How optimistic are you? I'm not a believer. Well, 00:41:21.610 --> 00:41:24.630 I don't say believer. There seems to be a limit 00:41:24.630 --> 00:41:28.610 to everything. And are we going to get to a point 00:41:28.610 --> 00:41:31.730 where we can get past that limit? I don't think 00:41:31.730 --> 00:41:35.670 it's going to happen in 2026. And I'm not sure 00:41:35.670 --> 00:41:38.349 it's going to happen in my lifetime anyway. So 00:41:38.349 --> 00:41:40.630 I don't think I'm going to benefit from that 00:41:40.630 --> 00:41:44.239 singularity. um you know maybe my my predictive 00:41:44.239 --> 00:41:47.780 capabilities are incorrect but uh i'm not holding 00:41:47.780 --> 00:41:52.780 out for that uh for that point oh no oh no okay 00:41:52.780 --> 00:41:56.880 that's too bad um what do you think about quantum 00:41:56.880 --> 00:41:59.079 computing jack like what is it going to do to 00:41:59.079 --> 00:42:02.239 all of the uh high performance computing from 00:42:02.239 --> 00:42:09.269 now is it going to 10 exit 100 exit yeah i think 00:42:09.269 --> 00:42:11.730 you know quantum computing is a wonderful research 00:42:11.730 --> 00:42:15.030 area so you know if i was younger i might i might 00:42:15.030 --> 00:42:18.730 be thinking about going into that area it's very 00:42:18.730 --> 00:42:22.059 ripe there's a lot of there's a lot of interesting 00:42:22.059 --> 00:42:25.420 research problems and a number of things that 00:42:25.420 --> 00:42:29.460 need to be explored and pushed on to understand 00:42:29.460 --> 00:42:34.099 better. Today, I think quantum computing is overhyped. 00:42:34.239 --> 00:42:37.659 So we have a number of companies that are making 00:42:37.659 --> 00:42:42.739 quantum computing computers and have overblown 00:42:42.739 --> 00:42:48.860 the expectations, if you will, for these computers. 00:42:49.739 --> 00:42:52.440 And I'm afraid what's going to happen is we're 00:42:52.440 --> 00:42:55.860 going to go through a quantum winter where things 00:42:55.860 --> 00:42:59.480 don't pan out, people lose interest, and those 00:42:59.480 --> 00:43:03.139 companies will go away. And we'll survive from 00:43:03.139 --> 00:43:06.159 that. We've seen a quantum winter happen with 00:43:06.159 --> 00:43:09.199 AI, and we've recovered nicely, I would say, 00:43:09.239 --> 00:43:12.760 from that. And today, AI is viewed with a tremendous 00:43:12.760 --> 00:43:15.599 amount of respect and is being used to help. 00:43:16.010 --> 00:43:18.969 solve some real problems. One of the issues with 00:43:18.969 --> 00:43:22.210 quantum computing is we have only a very few, 00:43:22.269 --> 00:43:25.989 a handful of algorithms which can benefit from 00:43:25.989 --> 00:43:28.789 quantum computing. And they could be very important 00:43:28.789 --> 00:43:33.909 methods, but there's only a handful. So we're 00:43:33.909 --> 00:43:36.170 not gonna replace our conventional computing 00:43:36.170 --> 00:43:38.769 with quantum computing. That's not gonna happen. 00:43:39.449 --> 00:43:42.190 What we might do in the near future, and people 00:43:42.190 --> 00:43:45.699 are doing it today. is to augment our computing 00:43:45.699 --> 00:43:49.559 systems with a quantum device. So I think of 00:43:49.559 --> 00:43:52.800 it almost like the way we think of GPU. So a 00:43:52.800 --> 00:43:55.860 GPU is something we connect to a conventional 00:43:55.860 --> 00:44:00.400 computer and benefit from it by offloading parts 00:44:00.400 --> 00:44:04.440 of the computation that can be done by that device, 00:44:05.000 --> 00:44:09.199 by that GPU, much faster. So in a quantum situation, 00:44:09.639 --> 00:44:12.500 we might offload. the the part of the computation 00:44:12.500 --> 00:44:14.960 that could be done by the quantum computing much 00:44:14.960 --> 00:44:18.659 much faster and then get back to our our conventional 00:44:18.659 --> 00:44:22.599 thing so i see a think of a computing system 00:44:22.599 --> 00:44:25.039 of the future as composed of our conventional 00:44:25.039 --> 00:44:28.139 digital computers our our accelerators in the 00:44:28.139 --> 00:44:32.000 form of gpus perhaps a quantum computer maybe 00:44:32.000 --> 00:44:34.739 an optical computer optical computing is quite 00:44:34.739 --> 00:44:38.360 intriguing you know we shine light which is digitized 00:44:38.360 --> 00:44:42.539 into some into some media and the results is 00:44:42.539 --> 00:44:45.480 maybe a multiplication takes place. And it happens 00:44:45.480 --> 00:44:48.440 really fast. So we end up with something which 00:44:48.440 --> 00:44:51.739 can do something very quickly. Or maybe neuromorphic 00:44:51.739 --> 00:44:54.760 computer, a computer that acts like our brains 00:44:54.760 --> 00:44:59.059 in terms of how it does things, in terms of spiky 00:44:59.059 --> 00:45:03.760 information. and networks that are involved in 00:45:03.760 --> 00:45:08.039 that neuromorphic computer. Or maybe going back 00:45:08.039 --> 00:45:11.400 to analog computing could be part of this ensemble 00:45:11.400 --> 00:45:15.579 of computing devices that we deploy, each one 00:45:15.579 --> 00:45:18.460 having its own characteristic and its own ability 00:45:18.460 --> 00:45:21.619 to solve something much better than the others 00:45:21.619 --> 00:45:25.260 can, and it's used when it's appropriate to. 00:45:25.820 --> 00:45:28.239 to use that. And we might build a computing device 00:45:28.239 --> 00:45:31.739 which has multiple sets of these things to help 00:45:31.739 --> 00:45:34.000 meet the needs of the problem that we're trying 00:45:34.000 --> 00:45:36.940 to solve. So that could be a computing system 00:45:36.940 --> 00:45:40.780 of the future. And as these things go, they get 00:45:40.780 --> 00:45:43.380 integrated into other parts to make them run 00:45:43.380 --> 00:45:45.659 really fast. So we're seeing this integration 00:45:45.659 --> 00:45:48.719 happen today between what I would call the conventional 00:45:48.719 --> 00:45:52.639 computer and the GPU. So they're now being made 00:45:52.639 --> 00:45:57.219 as a as a chip in itself and that chip has the 00:45:57.219 --> 00:46:00.019 ability to do both of those components and it's 00:46:00.019 --> 00:46:02.099 doing it much better because the communication 00:46:02.099 --> 00:46:07.900 happens in a shorter span of of distance and 00:46:07.900 --> 00:46:11.780 that that translates into a lot faster and a 00:46:11.780 --> 00:46:15.579 lot less energy being exploited in terms of coming 00:46:15.579 --> 00:46:17.900 up with the solution so it's all going to amount 00:46:17.900 --> 00:46:21.219 to energy and how much energy is expended in 00:46:21.219 --> 00:46:24.139 it you know our computers today You know, I think 00:46:24.139 --> 00:46:27.019 about our big supercomputers that we use in scientific 00:46:27.019 --> 00:46:30.780 computing. They have energy needs on the order 00:46:30.780 --> 00:46:35.099 of, we're talking about megawatts, tens of megawatts 00:46:35.099 --> 00:46:37.719 of power. So the big computer we have here at 00:46:37.719 --> 00:46:42.460 Oak Ridge National Lab has an energy requirement 00:46:42.460 --> 00:46:45.860 of something on the order of 30 megawatts. So 00:46:45.860 --> 00:46:49.539 30 million watts. And think of a megawatt, one 00:46:49.539 --> 00:46:52.400 megawatt. If I use one megawatt at my home here 00:46:52.400 --> 00:46:55.909 in... in Tennessee for a year, I'll get a bill 00:46:55.909 --> 00:46:58.429 from the electric company for a million dollars. 00:46:58.550 --> 00:47:02.230 So a megawatt year is a million dollars. So that 00:47:02.230 --> 00:47:05.150 computer at Oak Ridge costs roughly $30 million 00:47:05.150 --> 00:47:08.769 just to turn it on for the year. Now that computer 00:47:08.769 --> 00:47:11.849 costs $600 million, to put it in perspective. 00:47:12.210 --> 00:47:14.710 But the computers that are being built today 00:47:14.710 --> 00:47:20.360 for AI are going to gigawatts of power. So now 00:47:20.360 --> 00:47:24.820 we're talking about tremendous amounts of energy 00:47:24.820 --> 00:47:30.119 necessary to fire up these computers and a lot 00:47:30.119 --> 00:47:33.300 of cooling that has to take place. And we're 00:47:33.300 --> 00:47:37.139 talking about billions of dollars, really, to 00:47:37.139 --> 00:47:39.860 put that machine, just to operate the machine 00:47:39.860 --> 00:47:43.860 for a year period of time. The actual hardware 00:47:43.860 --> 00:47:46.820 is going to cost at least a billion dollars. 00:47:47.449 --> 00:47:50.730 for those AI computers. So you see the need for 00:47:50.730 --> 00:47:53.909 this private public kind of arrangement where 00:47:53.909 --> 00:47:58.510 the public with its tremendous resources in terms 00:47:58.510 --> 00:48:02.849 of the companies like Amazon, Google, Microsoft 00:48:02.849 --> 00:48:06.630 can invest in these machines and allow them to 00:48:06.630 --> 00:48:11.199 be used in the scientific context. Yeah, they're 00:48:11.199 --> 00:48:14.940 doing a lot. I heard 10 gigawatt data centers 00:48:14.940 --> 00:48:19.260 being built in Tennessee, I think, near Oak Ridge. 00:48:19.360 --> 00:48:23.599 Google is building one with a nuclear fission 00:48:23.599 --> 00:48:27.639 company. The guys over in Memphis are building 00:48:27.639 --> 00:48:30.659 something for Elon Musk with his company that 00:48:30.659 --> 00:48:36.050 are on that order today. Yeah, yeah. We're hoping 00:48:36.050 --> 00:48:38.690 to use Fusion Energy to run some of these things 00:48:38.690 --> 00:48:41.650 within the next decade or so. We're hoping we're 00:48:41.650 --> 00:48:46.329 a viable option. But I hear you. In terms of 00:48:46.329 --> 00:48:49.809 quantum computing, though, I've plugged into 00:48:49.809 --> 00:48:53.150 Microsoft's quantum computer through their API. 00:48:53.349 --> 00:48:56.730 It's called Quixit. And then I've used Google 00:48:56.730 --> 00:49:00.210 DeepMind. So I've plugged into some of these. 00:49:01.449 --> 00:49:04.809 run quantum algorithms for like, you know, fuel 00:49:04.809 --> 00:49:08.889 systems and kind of test test things. We bought 00:49:08.889 --> 00:49:13.710 10 hours for the whole year in 2024 and we bought 00:49:13.710 --> 00:49:17.590 10 hours and we used about eight of it and we 00:49:17.590 --> 00:49:22.550 did so much work. It was a tremendous amount 00:49:22.550 --> 00:49:26.119 of work. It was. 10 15 years worth of work but 00:49:26.119 --> 00:49:29.199 what you're saying is that we could have a quantum 00:49:29.199 --> 00:49:32.460 computing like a quantum setup that's commercially 00:49:32.460 --> 00:49:35.280 available that i could have downstairs in the 00:49:35.280 --> 00:49:39.460 warehouse uh that my company could then use is 00:49:39.460 --> 00:49:41.980 is what you're saying like this is this is possible 00:49:41.980 --> 00:49:44.420 is what you're saying Well, I would say that, 00:49:44.440 --> 00:49:47.219 yes, it's possible. I would also say that, you 00:49:47.219 --> 00:49:49.940 know, think about the cloud as you're using that 00:49:49.940 --> 00:49:52.539 Microsoft system. That's often the cloud and 00:49:52.539 --> 00:49:55.000 you're using it remotely. And that's a fine way 00:49:55.000 --> 00:49:57.739 to use it rather than to have to provide that 00:49:57.739 --> 00:50:01.059 whole infrastructure in your basement. Using 00:50:01.059 --> 00:50:03.739 it someplace else is certainly an attractive 00:50:03.739 --> 00:50:11.659 arrangement. Yeah, because I mean, for us, the 00:50:11.659 --> 00:50:16.179 big application is plasma control. And that has 00:50:16.179 --> 00:50:19.280 these mitigation algorithms to keep them away 00:50:19.280 --> 00:50:21.860 from hitting the walls and like, and there's 00:50:21.860 --> 00:50:24.880 just so much in it that if there was something 00:50:24.880 --> 00:50:27.199 that was scalable and commercially available 00:50:27.199 --> 00:50:31.179 and like a unique setup, I would like to dedicate 00:50:31.179 --> 00:50:34.320 one such setup to like each generator just to 00:50:34.320 --> 00:50:37.989 keep that thing just so precisely running. So 00:50:37.989 --> 00:50:41.989 that would be our goal. And something like that 00:50:41.989 --> 00:50:45.550 takes such a tremendous amount of computing and 00:50:45.550 --> 00:50:49.030 so many parallel things running at the same time 00:50:49.030 --> 00:50:53.610 and keeping it in sync. I don't know. I guess 00:50:53.610 --> 00:50:55.909 nothing would be possible without some of the 00:50:55.909 --> 00:51:02.909 things that you have built. That's crazy. I was 00:51:02.909 --> 00:51:06.309 writing, I was helping my... buddy here in la 00:51:06.309 --> 00:51:11.409 make a movie on um on basically a quantum computer 00:51:11.409 --> 00:51:16.750 uh crafting the code to santoshi nakamoto's bitcoin 00:51:16.750 --> 00:51:20.590 uh he has about uh apparently the legend says 00:51:20.590 --> 00:51:24.369 you know uh 20 of all bitcoin is kind of in his 00:51:24.369 --> 00:51:27.409 wallet and and we can see it we know that we 00:51:27.409 --> 00:51:32.110 can see it um it's worth about 70 billion dollars 00:51:32.110 --> 00:51:36.389 in today's today's money and so can a quantum 00:51:36.389 --> 00:51:39.269 computer crack it and this is this is this is 00:51:39.269 --> 00:51:41.389 the theme of the movie script right jack and 00:51:41.389 --> 00:51:45.210 so i'm uh so so we we theorize that it's possible 00:51:45.210 --> 00:51:48.150 and we kind of you know without giving away the 00:51:48.150 --> 00:51:53.130 ending you know hopefully um what do you think 00:51:53.130 --> 00:51:56.010 about that can we can we crack this bitcoin thing 00:51:56.010 --> 00:51:58.070 can we make it go away with the quantum with 00:51:58.070 --> 00:52:02.920 a couple of people well okay so let me say that 00:52:02.920 --> 00:52:06.159 quantum computing is good at solving problems 00:52:06.159 --> 00:52:09.780 which are very hard and in computing we characterize 00:52:09.780 --> 00:52:14.019 those problems as being non -polynomial time 00:52:14.019 --> 00:52:18.440 to solution so we think about np non -polynomial 00:52:18.440 --> 00:52:22.320 time and quantum computing is good at making 00:52:22.320 --> 00:52:26.820 inroads into NP problems. So what you're saying 00:52:26.820 --> 00:52:30.619 is you're going to crack a code, and that's something 00:52:30.619 --> 00:52:33.880 which is an NP problem. And we have a computer 00:52:33.880 --> 00:52:38.079 that potentially does that. So using that computer 00:52:38.079 --> 00:52:42.159 does make sense if you have the right algorithm 00:52:42.159 --> 00:52:47.599 and that it's going to be able to uh uh to uh 00:52:47.599 --> 00:52:50.360 to be run and to be implemented and and to do 00:52:50.360 --> 00:52:52.420 all the things it needs to do to crack the code 00:52:52.420 --> 00:52:56.800 i'm not an expert on on on cryptography uh but 00:52:56.800 --> 00:52:59.699 you know sure if we're dreaming if we're making 00:52:59.699 --> 00:53:02.420 writing a science fiction script i would say 00:53:02.420 --> 00:53:04.900 that's a good thing to go on to go after is exactly 00:53:04.900 --> 00:53:07.860 that use a quantum computer to crack the code 00:53:07.860 --> 00:53:13.420 and to unleash those bitcoins and to produce 00:53:13.420 --> 00:53:16.400 billions of dollars that that'd be great right 00:53:16.400 --> 00:53:19.719 like yeah it's a small robin hood group of people 00:53:19.719 --> 00:53:23.539 that's right yeah um so i don't know how far 00:53:23.539 --> 00:53:26.119 are we from that for a willing group of like 00:53:26.119 --> 00:53:29.320 12 12 really smart folks would you say we're 00:53:29.320 --> 00:53:34.400 like a decade away from that or so cracking cracking 00:53:34.400 --> 00:53:37.380 bitcoin is a little bit more difficult uh it's 00:53:37.380 --> 00:53:42.349 not like a uh public private key encryption uh 00:53:42.349 --> 00:53:45.090 it's uh it's slightly different than that i'm 00:53:45.090 --> 00:53:47.530 not an expert so i don't know all the ins and 00:53:47.530 --> 00:53:50.349 outs but um i think you're right you know a quantum 00:53:50.349 --> 00:53:54.010 computer has this incredible ability to uh come 00:53:54.010 --> 00:53:58.170 up with solutions uh in a very rapid way we gotta 00:53:58.170 --> 00:53:59.909 we gotta implement something so there has to 00:53:59.909 --> 00:54:02.710 be an algorithm uh quantum algorithm and then 00:54:02.710 --> 00:54:04.610 an implementation there has to be some software 00:54:04.610 --> 00:54:07.250 that runs there has to be some way to verify 00:54:07.250 --> 00:54:09.250 the solution you know with quantum computing 00:54:10.010 --> 00:54:12.409 you don't get one solution you get an approximation 00:54:12.409 --> 00:54:15.170 to a solution and you're going to be getting 00:54:15.170 --> 00:54:17.110 the number you have to make a number of tries 00:54:17.110 --> 00:54:19.730 at the solution because you're getting a statistical 00:54:19.730 --> 00:54:23.130 distribution for where the solution might be 00:54:23.130 --> 00:54:25.789 so you know we have to take all of those things 00:54:25.789 --> 00:54:28.750 into account and you know as we know quantum 00:54:28.750 --> 00:54:32.570 computing are very sensitive so that um you know 00:54:32.570 --> 00:54:35.429 a truck going by will disrupt the quantum computer 00:54:35.429 --> 00:54:38.269 and we have to start all over again and you know 00:54:38.269 --> 00:54:40.849 there are many things which are getting in the 00:54:40.849 --> 00:54:43.690 way and many things which are quite different 00:54:43.690 --> 00:54:47.150 than what we think of this in terms of our standard 00:54:47.150 --> 00:54:52.369 conventional computational resources so computers 00:54:52.369 --> 00:54:56.969 make mistakes that happens all the time and we 00:54:56.969 --> 00:55:01.090 hope it doesn't happen when we're doing the computation 00:55:01.090 --> 00:55:03.530 and we hope we hope that there's some way to 00:55:03.530 --> 00:55:06.750 mitigate those those failures so you know one 00:55:06.750 --> 00:55:09.869 of the big problems with quantum computing is 00:55:09.869 --> 00:55:14.690 being able to do fault tolerant, that is to check 00:55:14.690 --> 00:55:19.610 and to correct errors when they happen and do 00:55:19.610 --> 00:55:23.309 it at the gate level, do it at the very smallest 00:55:23.309 --> 00:55:26.949 level inside of the computer. So that's one of 00:55:26.949 --> 00:55:29.989 the big challenges, I would say, to making quantum 00:55:29.989 --> 00:55:34.289 computing a reality today is having fault tolerance 00:55:34.289 --> 00:55:36.960 at the gate level. And, you know, I think that's 00:55:36.960 --> 00:55:40.400 an interesting, again, research area to be involved 00:55:40.400 --> 00:55:43.239 in. There's a number of, you know, people looking 00:55:43.239 --> 00:55:47.400 at that, of course, and hopefully we'll overcome 00:55:47.400 --> 00:55:50.960 that and quantum computing could become a real 00:55:50.960 --> 00:55:57.920 reality. Yeah, I'm hopeful. I'm hopeful for that. 00:55:58.059 --> 00:56:00.739 I mean, the big task would be like something 00:56:00.739 --> 00:56:03.480 like sweeping the whole internet and making it 00:56:03.480 --> 00:56:06.539 all. like taking all the lies away or like, you 00:56:06.539 --> 00:56:12.599 know, some like huge thing would be game changer. 00:56:12.699 --> 00:56:15.920 I was at a talk once in Boston and someone said 00:56:15.920 --> 00:56:19.619 the way to find out if somebody successfully 00:56:19.619 --> 00:56:22.280 built a quantum computer in their basement is 00:56:22.280 --> 00:56:26.880 when everybody's bank account goes to zero. and 00:56:26.880 --> 00:56:29.179 we're like oh and then the next day the banks 00:56:29.179 --> 00:56:31.960 reset it and they get new encryption and whatever 00:56:31.960 --> 00:56:34.539 and they try to or they pay off the people that 00:56:34.539 --> 00:56:37.820 did this what however this works out but but 00:56:37.820 --> 00:56:40.159 that was going to be the way that people let 00:56:40.159 --> 00:56:42.679 the world know that we have a quantum computer 00:56:42.679 --> 00:56:48.260 um what are other criminal ways jack can we apply 00:56:48.260 --> 00:56:53.440 quantum computing to just for crime since we're 00:56:53.440 --> 00:56:56.510 on topic wow yeah so i've never been faced with 00:56:56.510 --> 00:56:59.449 that question how can we use uh computing to 00:56:59.449 --> 00:57:04.389 uh uh to create a to do a crime yeah so i i i'm 00:57:04.389 --> 00:57:08.110 not uh yeah i'm not i haven't thought about it 00:57:08.110 --> 00:57:11.989 i'm sure there i'm sure there are ways uh you 00:57:11.989 --> 00:57:15.429 know we could um we could sabotage uh things 00:57:15.429 --> 00:57:19.369 and and make things of a fail uh you know we're 00:57:19.369 --> 00:57:23.050 looking at uh so so okay so where does computing 00:57:24.119 --> 00:57:28.059 affects everybody in terms of weather forecasting. 00:57:28.719 --> 00:57:32.199 So if we were able to mispredict intentionally 00:57:32.199 --> 00:57:36.639 the weather, we can cause chaos at some level. 00:57:36.739 --> 00:57:39.920 And that chaos may translate into some criminal 00:57:39.920 --> 00:57:43.760 activity to do something nefarious. We could 00:57:43.760 --> 00:57:45.780 write a script, I'm sure, around that premise. 00:57:48.909 --> 00:57:51.650 I was going to say like steal from the Louvre, 00:57:51.750 --> 00:57:54.610 but that happened a few months ago. It feels 00:57:54.610 --> 00:57:56.670 like it's not fiction anymore. And it didn't 00:57:56.670 --> 00:57:59.889 require any sort of sophisticated computing or... 00:57:59.889 --> 00:58:04.530 No. I think it was a ladder. A ladder and a big 00:58:04.530 --> 00:58:06.570 saw to cut through glass. That was what they 00:58:06.570 --> 00:58:10.969 needed. Yeah. He stole stolen stuff. That's pretty 00:58:10.969 --> 00:58:14.650 cool. So if you were 25 today, you would get 00:58:14.650 --> 00:58:16.650 into quantum computing. Is that where we're at? 00:58:17.920 --> 00:58:22.639 Well, if I was 25 today, I would consider quantum. 00:58:23.000 --> 00:58:27.039 I would consider going into AI. AI is where all 00:58:27.039 --> 00:58:30.539 of the action is today and trying to use it to 00:58:30.539 --> 00:58:33.519 help solve some of the problems that we have. 00:58:33.639 --> 00:58:36.860 So those are the two big areas, I would say. 00:58:37.719 --> 00:58:43.639 You know, I tell my students, you know, the most 00:58:43.639 --> 00:58:48.039 important thing to do is to... follow your passion 00:58:48.039 --> 00:58:51.320 you know pick a problem that you can't stop thinking 00:58:51.320 --> 00:58:57.260 about and you know coming up with coming up with 00:58:57.260 --> 00:59:00.300 ways to overcome some of those dead ends that 00:59:00.300 --> 00:59:05.079 you encounter confusing results that happen papers 00:59:05.079 --> 00:59:08.019 that have been rejected codes that won't run 00:59:08.019 --> 00:59:10.760 you know that you have to overcome passion is 00:59:10.760 --> 00:59:13.760 what keeps you learning and when nobody is really 00:59:13.760 --> 00:59:19.099 giving you a pat on the back. It's, you know, 00:59:19.099 --> 00:59:23.539 spend time early, sampling broadly, go to courses, 00:59:23.579 --> 00:59:27.539 seminars, and do things. You know, aim high is 00:59:27.539 --> 00:59:31.739 another thing I recommend. Don't optimize for 00:59:31.739 --> 00:59:35.599 incremental wins. You know, choose questions 00:59:35.599 --> 00:59:39.980 that matter scientifically, from a social standpoint, 00:59:40.059 --> 00:59:43.139 or from an architectural standpoint you know 00:59:43.139 --> 00:59:46.079 define success in a way that still would be meaningful 00:59:46.079 --> 00:59:51.320 if you only get halfway let's say aim to work 00:59:51.320 --> 00:59:54.480 aim for work that changes how the field measures 00:59:54.480 --> 00:59:58.539 progress i also tell people in research you should 00:59:58.539 --> 01:00:02.260 expect to fail failure isn't an accident in science 01:00:02.260 --> 01:00:06.639 it's the default output from exploration most 01:00:06.639 --> 01:00:10.900 times we fail you know they don't survive Our 01:00:10.900 --> 01:00:14.460 ideas don't survive. Our hypotheses don't survive. 01:00:14.760 --> 01:00:18.500 And the trick is to fail cheaply and informatively. 01:00:18.699 --> 01:00:22.420 Design experiments that can teach you something 01:00:22.420 --> 01:00:26.480 even when something doesn't work. And you'll 01:00:26.480 --> 01:00:28.500 be in a better situation. And the final thing 01:00:28.500 --> 01:00:32.059 I tell people, my students, is to build a network 01:00:32.059 --> 01:00:34.880 of colleagues. Now, your career will be shaped 01:00:34.880 --> 01:00:41.679 by how much by people as by ideas. Build relationships 01:00:41.679 --> 01:00:47.380 before you need them. Give talks in seminars. 01:00:47.539 --> 01:00:51.179 Ask thoughtful questions. Share code and credit. 01:00:51.440 --> 01:00:54.239 Be reliable. Follow through on what you're doing. 01:00:54.739 --> 01:00:59.039 A good network isn't transactional. It's a community 01:00:59.039 --> 01:01:06.139 where it makes you smarter and it builds collaborations 01:01:06.139 --> 01:01:09.760 and complements your skills. So all of those 01:01:09.760 --> 01:01:14.679 things, I think, go into helping whatever area 01:01:14.679 --> 01:01:19.760 people go into in making it really an important 01:01:19.760 --> 01:01:28.900 step. That's awesome. Wow. Did your parents know 01:01:28.900 --> 01:01:31.579 that you were this genius child when you were 01:01:31.579 --> 01:01:35.440 8 or 10 years old? When did they know what they 01:01:35.440 --> 01:01:39.070 had running around their house? Well, they never 01:01:39.070 --> 01:01:42.730 did. So here's the story I sometimes tell. So 01:01:42.730 --> 01:01:46.349 I mentioned earlier that my brothers have a trait. 01:01:46.570 --> 01:01:49.329 You know, they were electricians. And my mother 01:01:49.329 --> 01:01:53.090 knew exactly what they did. I would try to explain 01:01:53.090 --> 01:01:56.130 to my mother, you know, what I did. And she never 01:01:56.130 --> 01:01:58.289 really got it. And she was always, you know, 01:01:58.309 --> 01:02:01.230 a bit concerned that I wasn't doing very well 01:02:01.230 --> 01:02:04.389 financially. You know, she was, I'm a teacher. 01:02:04.530 --> 01:02:06.630 Okay, how much do I teach? I teach one course 01:02:06.630 --> 01:02:09.300 a year. you don't teach very much and i said 01:02:09.300 --> 01:02:11.059 you know what you know she asked what do i do 01:02:11.059 --> 01:02:13.860 well i think i come up with a problem i write 01:02:13.860 --> 01:02:17.380 a proposal 10 pages and i send it into an agency 01:02:17.380 --> 01:02:19.920 and if they like it they they send me millions 01:02:19.920 --> 01:02:22.719 of dollars so i can carry out those ideas and 01:02:22.719 --> 01:02:25.039 that was a that was a confusing uh confusing 01:02:25.719 --> 01:02:27.760 story. You know, she couldn't understand any 01:02:27.760 --> 01:02:30.000 part of that story. You know, sitting around 01:02:30.000 --> 01:02:32.280 thinking about things, the government giving 01:02:32.280 --> 01:02:34.519 you money so that you can carry out what you 01:02:34.519 --> 01:02:37.739 wanted to do. That didn't make any sense at all. 01:02:38.019 --> 01:02:41.840 And, you know, she would visit us and, you know, 01:02:41.840 --> 01:02:44.260 we had a slightly different way of existing from 01:02:44.260 --> 01:02:48.719 what she learned and how she carried out her 01:02:48.719 --> 01:02:52.920 life. An example of that is my wife. You know, 01:02:52.960 --> 01:02:57.369 we had three kids. my wife would um whenever 01:02:57.369 --> 01:03:01.050 we had a birthday from one of them my wife would 01:03:01.050 --> 01:03:03.449 make a cake she would she would and she would 01:03:03.449 --> 01:03:05.030 ask the kids you know what kind of cake do you 01:03:05.030 --> 01:03:07.269 want oh give me a dinosaur cake or give me a 01:03:07.269 --> 01:03:09.989 uh give me a you know an animal of some kind 01:03:09.989 --> 01:03:12.469 you know give me something in the shape of that 01:03:12.469 --> 01:03:15.489 so my my wife would go off and and make this 01:03:15.489 --> 01:03:18.010 cake and my mother was shocked at this you know 01:03:18.010 --> 01:03:20.650 she she thought it was a a sign that we didn't 01:03:20.650 --> 01:03:22.909 have enough money to go out and buy a cake from 01:03:22.909 --> 01:03:25.610 the bakery because that's that's the mode that 01:03:25.610 --> 01:03:27.510 she would have she would have done she wouldn't 01:03:27.510 --> 01:03:29.929 go out and make it she would she would go out 01:03:29.929 --> 01:03:32.550 and buy it so there was this you know quite uh 01:03:32.550 --> 01:03:35.389 so she never really did understand what i what 01:03:35.389 --> 01:03:39.389 i did and um even when i you know received this 01:03:39.389 --> 01:03:42.949 award she uh she wasn't quite sure uh what it 01:03:42.949 --> 01:03:45.250 was her her comment when it appeared when my 01:03:45.250 --> 01:03:47.130 name appeared in my picture appeared in the new 01:03:47.130 --> 01:03:51.030 york times i tried to explain what this was about 01:03:51.030 --> 01:03:53.750 and she says oh you got a nice sweater on in 01:03:53.750 --> 01:03:56.050 the picture so that was her that was her take 01:03:56.050 --> 01:03:59.050 on the whole on the whole thing oh it was sort 01:03:59.050 --> 01:04:03.730 of funny oh no jack i i that that was so that 01:04:03.730 --> 01:04:07.269 hit me in a funny way because my mom uh she sees 01:04:07.269 --> 01:04:10.349 me working on fusion energy she she's just like 01:04:10.349 --> 01:04:13.909 i have why don't you have a regular job like 01:04:13.909 --> 01:04:17.719 how do you get a paycheck and health insurance 01:04:17.719 --> 01:04:20.679 and she doesn't know and i tell her you know 01:04:20.679 --> 01:04:23.079 because i'm kind of like that kid like i'm an 01:04:23.079 --> 01:04:25.880 only child and i'm very communicative so like 01:04:25.880 --> 01:04:28.480 i tell her everything and she's just kind of 01:04:28.480 --> 01:04:33.639 like whatever but last year my cousin saw some 01:04:33.639 --> 01:04:36.659 me doing something and then my aunt called her 01:04:36.659 --> 01:04:40.380 and said you know if if what priyanka is doing 01:04:40.380 --> 01:04:43.800 works out it could change the world like real 01:04:43.800 --> 01:04:49.010 things She comes to me about six months ago and 01:04:49.010 --> 01:04:52.829 she goes, by the way, I spoke to your aunt and 01:04:52.829 --> 01:04:56.349 she's saying that you did that. My God, I tell 01:04:56.349 --> 01:05:02.690 you every day. I tell you every day. Maybe you 01:05:02.690 --> 01:05:04.570 should have had someone else explain it to her. 01:05:04.769 --> 01:05:10.110 That's right. But wow, this was awesome. Thank 01:05:10.110 --> 01:05:12.070 you so much, Jack. My pleasure.