NVIDIA CEO Jensen Huang's Vision for the Future — Transcript

Full transcript

1. 0:00At some point, you have to believe something.  We've reinvented computing as we know it. What
2. 0:03is the vision for what you see coming next? We  asked ourselves, if it can do this, how far can
3. 0:08it go? How do we get from the robots that  we have now to the future world that you
4. 0:13see? Cleo, everything that moves will be  robotic someday and it will be soon. We
5. 0:17invested tens of billions of dollars before  it really happened. No that's very good, you
6. 0:22did some research! But the big breakthrough  I would say is when we...
7. 0:28That's Jensen Huang, and whether you know it or not his decisions are shaping your future. He's the CEO of
8. 0:36NVIDIA, the company that skyrocketed over the past few years to become one of the most valuable companies in
9. 0:41the world because they led a fundamental shift  in how computers work unleashing this current
10. 0:46explosion of what's possible with technology.  "NVIDIA's done it again!" We found ourselves being
11. 0:51one of the most important technology companies in  the world and potentially ever. A huge amount of
12. 0:56the most futuristic tech that you're hearing about in AI and robotics and gaming and self-driving
13. 1:01cars and breakthrough medical research relies on  new chips and software designed by him and his
14. 1:06company. During the dozens of background interviews  that I did to prepare for this what struck me most
15. 1:10was how much Jensen Huang has already influenced  all of our lives over the last 30 years, and how
16. 1:16many said it's just the beginning of something  even bigger. We all need to know what he's building
17. 1:22and why and most importantly what he's trying  to build next. Welcome to Huge Conversations...
18. 1:36Thank you so much for doing this. I'm so happy to do  it. Before we dive in, I wanted to tell you
19. 1:42how this interview is going to be a little bit  different than other interviews I've seen you
20. 1:45do recently. Okay! I'm not going to ask you any  questions about - you could ask - company finances,
21. 1:51thank you! I'm not going to ask you questions  about your management style or why you don't
22. 1:55like one-on ones. I'm not going to ask you  about regulations or politics. I think all
23. 2:01of those things are important but I think that our  audience can get them well covered elsewhere. Okay.
24. 2:06What we do on huge if true is we make optimistic  explainer videos and we've covered - I'm the worst
25. 2:13person to be an explainer video. I think you  might be the best and I think that's what I'm
26. 2:18really hoping that we can do together is make a  joint explainer video about how can we actually
27. 2:25use technology to make the future better. Yeah. And  we do it because we believe that when people see
28. 2:30those better futures, they help build them. So  the people that you're going to be talking to
29. 2:33are awesome. They are optimists who want to  build those better futures but because we
30. 2:39cover so many different topics, we've covered  supersonic planes and quantum computers and
31. 2:43particle colliders, it means that millions  of people come into every episode without
32. 2:48any prior knowledge whatsoever. You might be  talking to an expert in their field who doesn't
33. 2:53know the difference between a CPU and a GPU or a  12-year-old who might grow up one day to be you
34. 3:00but is just starting to learn. For my part,  I've now been preparing for this interview for
35. 3:06several months, including doing background  conversations with many members of your team
36. 3:11but I'm not an engineer. So my goal is to help that  audience see the future that you see so I'm going
37. 3:18to ask about three areas: The first is, how did we  get here? What were the key insights that led to
38. 3:23this big fundamental shift in computing that we're  in now? The second is, what's actually happening
39. 3:29right now? How did those insights lead to the world  that we're now living in that seems like so much
40. 3:34is going on all at once? And the third is, what is  the vision for what you see coming next? In order
41. 3:42to talk about this big moment we're in with AI  I think we need to go back to video games in the
42. 3:48'90s. At the time I know game developers wanted  to create more realistic looking graphics but
43. 3:56the hardware couldn't keep up with all of that  necessary math. NVIDIA came up with
44. 4:02a solution that would change not just games  but computing itself. Could you take us back
45. 4:09there and explain what was happening and what  were the insights that led you and the NVIDIA
46. 4:15team to create the first modern GPU? So in the  early '90s when we first started the company
47. 4:20we observed that in a software program inside  it there are just a few lines of code, maybe
48. 4:2710% of the code, does 99% % of the processing  and that 99% of the processing could be done
49. 4:33in parallel. However the other 90% of the code  has to be done sequentially. It turns out that
50. 4:40the proper computer the perfect computer is one  that could do sequential processing and parallel
51. 4:45processing not just one or the other. That was the  big observation and we set out to build a company
52. 4:52to solve computer problems that normal computers  can't. And that's really the beginning of NVIDIA.
53. 5:00My favorite visual of why a CPU versus a  GPU really matters so much is a 15-year-old
54. 5:05video on the NVIDIA YouTube channel where the  Mythbusters, they use a little robot shooting
55. 5:11paintballs one by one to show solving problems  one at a time or sequential processing on a
56. 5:16CPU, but then they roll out this huge robot  that shoots all of the paintballs at once
57. 5:24doing smaller problems all at the same  time or parallel processing on a GPU.
58. 5:30"3... 2... 1..." So Nvidia unlocks all of this new power for video games. Why gaming first? The video games
59. 5:41requires parallel processing for processing  3D graphics and we chose video games because,
60. 5:47one, we loved the application, it's a simulation  of virtual worlds and who doesn't want to go to
61. 5:52virtual worlds and we had the good observation  that video games has potential to be the largest
62. 5:58market for for entertainment ever. And it turned  out to be true. And having it being a large market
63. 6:04is important because the technology is complicated  and if we had a large market, our R&D budget could
64. 6:10be large, we could create new technology. And that  flywheel between technology and market and greater
65. 6:17technology was really the flywheel that  got NVIDIA to become one of the most important
66. 6:21technology companies in the world. It was all  because of video games. I've heard you say that
67. 6:25GPUs were a time machine? Yeah. Could you tell me  more about what you meant by that? A GPU is like a
68. 6:31time machine because it lets you see the future  sooner. One of the most amazing things anybody's
69. 6:37ever said to me was a quantum chemistry  scientist. He said, Jensen, because of NVIDIA's work,
70. 6:46I can do my life's work in my lifetime. That's time  travel. He was able to do something that was beyond
71. 6:52his lifetime within his lifetime and this is  because we make applications run so much faster
72. 7:00and you get to see the future. And so when you're  doing weather prediction for example, you're seeing
73. 7:05the future when you're doing a simulation  a virtual city with virtual traffic and we're
74. 7:11simulating our self-driving car through  that virtual city, we're doing time travel. So
75. 7:17parallel processing takes off in gaming and it's  allowing us to create worlds in computers that
76. 7:24we never could have before and and gaming is sort  of this this first incredible cas Cas of parallel
77. 7:30processing unlocking a lot more power and then  as you said people begin to use that power across
78. 7:37many different industries. The case of the of the  quantum chemistry researcher, when I've heard you
79. 7:42tell that story it's that he was running molecular  simulations in a way where it was much faster to
80. 7:49run in parallel on NVIDIA GPUs even then than it  was to run them on the supercomputer with the CPU
81. 7:56that he had been using before. Yeah that's true. So  oh my god it's revolutionizing all of these other
82. 8:00industries as well, it's beginning to change  how we see what's possible with computers and my
83. 8:07understanding is that in the early 2000s you  see this and you realize that actually doing
84. 8:14that is a little bit difficult because what that  researcher had to do is he had to sort of trick
85. 8:18the GPUs into thinking that his problem was a  graphics problem. That's exactly right, no that's
86. 8:23very good, you did some research. So you create  a way to make that a lot easier. That's right
87. 8:29Specifically it's a platform called CUDA which  lets programmers tell the GPU what to do using
88. 8:34programming languages that they already know like  C and that's a big deal because it gives way more
89. 8:39people easier access to all of this computing  power. Could you explain what the vision was that
90. 8:44led you to create CUDA? Partly researchers  discovering it, partly internal inspiration and
91. 8:53and partly solving a problem. And you know a  lot of interesting interesting ideas come out
92. 9:00of that soup. You know some of it is aspiration  and inspiration, some of it is just desperation you
93. 9:06know. And so in the case of CUDA is very  much this the same way and probably the first
94. 9:13external ideas of using our GPUs for parallel  processing emerged out of some interesting work
95. 9:19in medical imaging a couple of researchers  at Mass General were using it to do CT
96. 9:26reconstruction. They were using our graphics  processors for that reason and it inspired us.
97. 9:32Meanwhile the problem that we're trying to solve  inside our company has to do with the fact that
98. 9:36when you're trying to create these virtual worlds  for video games, you would like it to be beautiful
99. 9:41but also dynamic. Water should flow like water and  explosions should be like explosions. So there's
100. 9:50particle physics you want to do, fluid dynamics  you want to do and that is much harder to do if
101. 9:56your pipeline is only able to do computer graphics.  And so we have a natural reason to want to do it
102. 10:02in the market that we were serving. So  researchers were also horsing around with using
103. 10:08our GPUs for general purpose uh acceleration and  and so there there are multiple multiple factors
104. 10:13that were coming together in that soup, we  just when the time came and we decided
105. 10:20to do something proper and created a CUDA as  a result of that. Fundamentally the reason why
106. 10:25I was certain that CUDA was going to be successful  and we put the whole company behind it was
107. 10:31because fundamentally our GPU was going to be  the highest volume parallel processors built in
108. 10:38the world because the market of video games was so  large and so this architecture has a good chance
109. 10:43of reaching many people. It has seemed to me like  creating CUDA was this incredibly optimistic "huge
110. 10:51if true" thing to do where you were saying, if we  create a way for many more people to use much
111. 10:58more computing power, they might create incredible  things. And then of course it came true. They did.
112. 11:04In 2012, a group of three researchers submits an  entry to a famous competition where the goal is
113. 11:09to create computer systems that could recognize  images and label them with categories. And their
114. 11:14entry just crushes the competition. It gets way  fewer answers wrong. It was incredible. It blows
115. 11:20everyone away. It's called AlexNet, and it's a kind  of AI called the neural network. My understanding
116. 11:24is one reason it was so good is that they used  a huge amount of data to train that system
117. 11:29and they did it on NVIDIA GPUs. All of a sudden,  GPUs weren't just a way to make computers faster
118. 11:35and more efficient they're becoming the engines  of a whole new way of computing. We're moving from
119. 11:40instructing computers with step-by-step directions  to training computers to learn by showing them a
120. 11:47huge number of examples. This moment in 2012 really  kicked off this truly seismic shift that we're
121. 11:54all seeing with AI right now. Could you describe  what that moment was like from your perspective
122. 11:59and what did you see it would mean for all of  our futures? When you create something new like
123. 12:06CUDA, if you build it, they might not come. And that's always the cynic's perspective
124. 12:14however the optimist's perspective would say, but  if you don't build it, they can't come. And that's
125. 12:20usually how we look at the world. You know we  have to reason about intuitively why this would
126. 12:25be very useful. And in fac, in 2012 Ilya Sutskever,  and Alex Krizhevsky and Geoff Hinton in the University
127. 12:33of Toronto the lab that they were at they reached  out to a gForce GTX 580 because they learned about
128. 12:39CUDA and that CUDA might be able to to be used as  a parallel processor for training AlexNet and
129. 12:45so our inspiration that GeForce could be the the  vehicle to bring out this parallel architecture
130. 12:51into the world and that researchers would somehow  find it someday was a good was a good strategy. It
131. 12:57was a strategy based on hope, but it was also  reasoned hope. The thing that really caught
132. 13:03our attention was simultaneously we were trying  to solve the computer vision problem inside the
133. 13:07company and we were trying to get CUDA to  be a good computer vision processor and we
134. 13:13were frustrated by a whole bunch of early  developments internally with respect to our
135. 13:19computer vision effort and getting CUDA to be  able to do it. And all of a sudden we saw AlexNet,
136. 13:25this new algorithm that is completely  different than computer vision algorithms before
137. 13:31it, take a giant leap in terms of capability  for computer vision. And when we saw that it was
138. 13:38partly out of interest but partly because we were  struggling with something ourselves. And so we were
139. 13:43we were highly interested to want to see it work.  And so when we when we looked at AlexNet we were
140. 13:49inspired by that. But the big breakthrough I  would say is when we when we saw AlexNet, we
141. 13:57asked ourselves you know, how far can AlexNet  go? If it can do this with computer vision, how
142. 14:04far can it go? And if it if it could go to the  limits of what we think it could go, the type
143. 14:11of problems it could solve, what would it mean for  the computer industry? And what would it mean for
144. 14:16the computer architecture? And we were,  we rightfully reasoned that if machine learning,
145. 14:25if the deep learning architecture can scale,  the vast majority of machine learning problems
146. 14:30could be represented with deep neural networks. And  the type of problems we could solve with machine
147. 14:36learning is so vast that it has the potential of  reshaping the computer industry altogether,
148. 14:42which prompted us to re-engineer the entire  computing stack which is where DGX came from
149. 14:49and this little baby DGX sitting here, all of  this came from from that observation that we ought
150. 14:56to reinvent the entire computing stack layer by  layer by layer. You know computers, after 65 years
151. 15:03since IBM System 360 introduced modern general  purpose computing, we've reinvented computing as we
152. 15:09know it. To think about this as a whole story, so  parallel processing reinvents modern gaming and
153. 15:16revolutionizes an entire industry then that way  of computing that parallel processing begins to
154. 15:22be used across different industries. You invest  in that by building CUDA and then CUDA and the
155. 15:29use of GPUs allows for a a step change in neural  networks and machine learning and begins a sort
156. 15:38of revolution that we're now seeing only  increase in importance today... All of a sudden
157. 15:45computer vision is solved. All of a sudden speech  recognition is solved. All of a sudden language
158. 15:50understanding is solved. These incredible  problems associated with intelligence one
159. 15:54by one by one by one where we had no solutions  for in past, desperate desire to have solutions
160. 16:01for, all of a sudden one after another get solved  you know every couple of years. It's incredible.
161. 16:07Yeah so you're seeing that, in 2012 you're  looking ahead and believing that that's
162. 16:12the future that you're going to be living in now,  and you're making bets that get you there, really
163. 16:17big bets that have very high stakes. And then my  perception as a lay person is that it takes a
164. 16:22pretty long time to get there. You make these bets -  8 years, 10 years - so my question is:
165. 16:30If AlexNet that happened in 2012 and this audience  is probably seeing and hearing so much more about
166. 16:36AI and NVIDIA specifically 10 years later,  why did it take a decade and also because you
167. 16:43had placed those bets, what did the middle  of that decade feel like for you? Wow that's
168. 16:48a good question. It probably felt like today. You  know to me, there's always some problem and
169. 16:55then there's some reason to be to be  impatient. There's always some reason to be
170. 17:03happy about where you are and there's always  many reasons to carry on. And so I think as I
171. 17:09was reflecting a second ago, that sounds like this  morning! So but I would say that in all things that
172. 17:16we pursue, first you have to have core beliefs. You  have to reason from your best principles
173. 17:25and ideally you're reasoning from it from principles  of either physics or deep understanding of
174. 17:32the industry or deep understanding of the  science, wherever you're reasoning from, you
175. 17:38reason from first principles. And at some point you  have to believe something. And if those principles
176. 17:45don't change and the assumptions don't change, then you, there's no reason to change your
177. 17:50core beliefs. And then along the way there's always  some evidence of you know of success and
178. 17:59and that you're leading in the right  direction and sometimes you know you go a
179. 18:04long time without evidence of success and you  might have to course correct a little but
180. 18:08the evidence comes. And if you feel like you're  going in the right direction, we just keep on going.
181. 18:12The question of why did we stay so committed for  so long, the answer is actually the opposite: There
182. 18:19was no reason to not be committed because we are,  we believed it. And I've believed in NVIDIA
183. 18:28for 30 plus years and I'm still here working  every single day. There's no fundamental
184. 18:34reason for me to change my belief system and  I fundamentally believe that the
185. 18:39work we're doing in revolutionizing computing  is as true today, even more true today than it
186. 18:43was before. And so we'll stick  with it you know until otherwise. There's
187. 18:51of course very difficult times along the way. You  know when you're investing in something and nobody
188. 18:58else believes in it and cost a lot of money and  you know maybe investors or or others would rather
189. 19:05you just keep the profit or you know whatever it  is improve the share price or whatever it is.
190. 19:11But you have to believe in your future. You have to  invest in yourself. And we believe this so
191. 19:17deeply that we invested you know tens  of billions of dollars before it really
192. 19:25happened. And yeah it was, it was 10 long  years. But it was fun along the way.
193. 19:32How would you summarize those core beliefs? What  is it that you believe about the way computers
194. 19:38should work and what they can do for us that keeps  you not only coming through that decade but also
195. 19:44doing what you're doing now, making bets I'm sure  you're making for the next few decades? The first
196. 19:50core belief was our first discussion, was about  accelerated computing. Parallel computing versus
197. 19:56general purpose computing. We would add  two of those processors together and we would do
198. 20:00accelerated computing. And I continue to believe  that today. The second was the recognition
199. 20:06that these deep learning networks, these DNNs, that  came to the public during 2012, these deep neural
200. 20:13networks have the ability to learn patterns and  relationships from a whole bunch of different
201. 20:18types of data. And that it can learn more and  more nuanced features if it could be larger
202. 20:24and larger. And it's easier to make them larger and  larger, make them deeper and deeper or wider and
203. 20:29wider, and so the scalability of the architecture  is empirically true. The fact
204. 20:40that model size and the data size being larger  and larger, you can learn more knowledge is
205. 20:47also true, empirically true. And so if that's  the case, you could you know, what what are the
206. 20:55limits? There not, unless there's a physical limit  or an architectural limit or mathematical limit
207. 21:00and it was never found, and so we believe that you  could scale it. Then the question, the only other
208. 21:05question is: What can you learn from data? What  can you learn from experience? Data is basically
209. 21:11digital versions of human experience. And so what  can you learn? You obviously can learn object
210. 21:17recognition from images. You can learn speech  from just listening to sound. You can learn
211. 21:22even languages and vocabulary and syntax and  grammar and all just by studying a whole bunch
212. 21:27of letters and words. So we've now demonstrated  that AI or deep learning has the ability to learn
213. 21:33almost any modality of data and it can translate  to any modality of data. And so what does that mean?
214. 21:42You can go from text to text, right, summarize a  paragraph. You can go from text to text, translate
215. 21:49from language to language. You can go from text  to images, that's image generation. You can go from
216. 21:55images to text, that's captioning. You can even go  from amino acid sequences to protein structures.
217. 22:03In the future, you'll go from protein to words: "What  does this protein do?" or "Give me an example of a
218. 22:11protein that has these properties." You know  identifying a drug target. And so you could
219. 22:17just see that all of these problems are around  the corner to be solved. You can go from words
220. 22:24to video, why can't you go from words to action  tokens for a robot? You know from the computer's
221. 22:33perspective how is it any different? And so it  it opened up this universe of opportunities and
222. 22:40universe of problems that we can go solve. And  that gets us quite excited. It feels like
223. 22:48we are on the cusp of this truly enormous change.  When I think about the next 10 years, unlike the
224. 22:56last 10 years, I know we've gone through a lot of  change already but I don't think I can predict
225. 23:02anymore how I will be using the technology that is  currently being developed. That's exactly right. I
226. 23:07think the last 10, the reason why you feel that way  is, the last 10 years was really about the science
227. 23:12of AI. The next 10 years we're going to have plenty  of science of AI but the next 10 years is going to
228. 23:18be the application science of AI. The fundamental  science versus the application science. And so the
229. 23:24the applied research, the application side of AI  now becomes: How can I apply AI to digital biology?
230. 23:31How can I apply AI to climate technology? How can  I apply AI to agriculture, to fishery, to robotics,
231. 23:39to transportation, optimizing logistics? How can  I apply AI to you know teaching? How do I apply AI
232. 23:47to you know podcasting right? I'd love to  choose a couple of those to help people see how
233. 23:53this fundamental change in computing that we've  been talking about is actually going to change
234. 23:58their experience of their lives, how they're  actually going to use technology that is based
235. 24:02on everything we just talked about. One of the  things that I've now heard you talk a lot about
236. 24:07and I have a particular interest in is physical  AI. Or in other words, robots - "my friends!" - meaning
237. 24:16humanoid robots but also robots like self-driving  cars and smart buildings or autonomous warehouses
238. 24:23or autonomous lawnmowers or more. From what  I understand, we might be about to see a huge
239. 24:29leap in what all of these robots are capable of  because we're changing how we train them. Up until
240. 24:37recently you've either had to train your robot in  the real world where it could get damaged or wear
241. 24:43down or you could get data from fairly limited  sources like humans in motion capture suits. But
242. 24:50that means that robots aren't getting as many  examples as they'd need to learn more quickly.
243. 24:56But now we're starting to train robots in digital  worlds, which means way more repetitions a day, way
244. 25:03more conditions, learning way faster. So we could  be in a big bang moment for robots right now and
245. 25:11NVIDIA is building tools to make that happen. You  have Omniverse and my understanding is this is 3D
246. 25:19worlds that help train robotic systems so that  they don't need to train in the physical world.
247. 25:26That's exactly right. You just just announced  Cosmos which is ways to make that 3D universe
248. 25:34much more realistic. So you can get all kinds  of different, if we're training something on
249. 25:39this table, many different kinds of lighting on the  table, many different times of day, many different
250. 25:44you know experiences for the robot to go through  so that it can get even more out of Omniverse. As
251. 25:52a kid who grew up loving Data on Star Trek, Isaac  Asimov's books and just dreaming about a future with
252. 26:00robots, how do we get from the robots that we have  now to the future world that you see of robotics?
253. 26:08Yeah let me use language models maybe ChatGPT  as a reference for understanding Omniverse and
254. 26:17Cosmos. So first of all when ChatGPT first  came out it, it was extraordinary and
255. 26:24it has the ability to do to basically from  your prompt, generate text. However, as amazing as
256. 26:32it was, it has the tendency to hallucinate if it goes on too long or if it pontificates about
257. 26:40a topic it you know is not informed about, it'll  still do a good job generating plausible answers.
258. 26:46It just wasn't grounded in the truth. And so people called it hallucination. And
259. 26:55so the next generation shortly it was, it had  the ability to be conditioned by context, so
260. 27:03you could upload your PDF and now it's grounded  by the PDF. The PDF becomes the ground truth. It
261. 27:09could be it could actually look up search and  then the search becomes its ground truth. And
262. 27:14between that it could reason about what is how  to produce the answer that you're asking for. And
263. 27:21so the first part is a generative AI and the  second part is ground truth. Okay and so now let's
264. 27:28come into the the physical world. The world model, we need a foundation model just like
265. 27:35we need ChatGPT had a core foundation model  that was the breakthrough in order for robotics
266. 27:41to to be smart about the physical world. It has to  understand things like gravity, friction, inertia,
267. 27:50geometric and spatial awareness. It has to uh  understand that an object is sitting there even
268. 27:57when I looked away when I come back it's still  sitting there, object permanence. It has to
269. 28:02understand cause and effect. If I tip it, it'll  fall over. And so these kind of physical
270. 28:08common sense if you will has to be captured or  encoded into a world foundation model so that
271. 28:16the AI has world common sense. Okay and so we  have to go, somebody has to go create that, and
272. 28:23that's what we did with Cosmos. We created a world  language model. Just like ChatGPT was a language model,
273. 28:29this is a world model. The second thing we have to  go do is we have to do the same thing that we did
274. 28:35with PDFs and context and grounding it with  ground truth. And so the way we augment Cosmos
275. 28:42with ground truth is with physical simulations,  because Omniverse uses physics simulation which
276. 28:49is based on principled solvers. The mathematics  is Newtonian physics is the, right, it's the math we
277. 28:56know, all of the the fundamental laws of  physics we've understood for a very long
278. 29:02time. And it's encoded into, captured into Omniverse.  That's why Omniverse is a simulator. And using the
279. 29:09simulator to ground or to condition Cosmos, we can  now generate an infinite number of stories of the
280. 29:19future. And they're grounded on physical truth. Just  like between PDF or search plus ChatGPT, we can
281. 29:30generate an infinite amount of interesting things,  answer a whole bunch of interesting questions. The
282. 29:37combination of Omniverse plus Cosmos, you could  do that for the physical world. So to illustrate
283. 29:43this for the audience, if you had a robot in a  factory and you wanted to make it learn every
284. 29:49route that it could take, instead of manually  going through all of those routes, which could
285. 29:53take days and could be a lot of wear and tear on  the robot, we're now able to simulate all of them
286. 29:59digitally in a fraction of the time and in many  different situations that the robot might face -
287. 30:04it's dark, it's blocked it's etc - so the robot  is now learning much much faster. It seems to
288. 30:10me like the future might look very different than  today. If you play this out 10 years, how do you see
289. 30:17people actually interacting with this technology  in the near future? Cleo, everything that moves
290. 30:22will be robotic someday and it will be soon. You  know the the idea that you'll be pushing around
291. 30:28a lawn mower is already kind of silly. You know  maybe people do it because because it's fun but
292. 30:35but there's no need to. And every car is  going to be robotic. Humanoid robots, the technology
293. 30:44necessary to make it possible, is just around  the corner. And so everything that moves will be
294. 30:50robotic and they'll learn how to be  a robot in Omniverse Cosmos and we'll generate
295. 30:59all these plausible, physically plausible futures  and the the robots will learn from them and
296. 31:05then they'll come into the physical world and you  know it's exactly the same. A future where
297. 31:11you're just surrounded by robots is for certain.  And I'm just excited about having my own R2-D2.
298. 31:18And of course R2-D2 wouldn't be quite the can that  it is and roll around. It'll be you know R2-D2
299. 31:25yeah, it'll probably be a different physical  embodiment, but it's always R2. You know so my R2
300. 31:32is going to go around with me. Sometimes it's in my  smart glasses, sometimes it's in my phone, sometimes
301. 31:36it's in my PC. It's in my car. So R2 is with me  all the time including you know when I get home
302. 31:43you know where I left a physical version of R2. And  you know whatever that version happens to
303. 31:49be you know, we'll interact with R2. And so I  think the idea that we'll have our own R2-D2 for
304. 31:55our entire life and it grows up with us, that's  a certainty now yeah. I think a lot of news media
305. 32:05when they talk about futures like this they focus  on what could go wrong. And that makes sense. There
306. 32:10is a lot that could go wrong. We should talk about  what could go wrong so we could keep it from from
307. 32:14going wrong. Yeah that's the approach that we like  to take on the show is, what are the big challenges
308. 32:19so that we can overcome them? Yeah. What buckets do  you think about when you're worrying about this
309. 32:24future? Well there's a whole bunch of the  stuff that everybody talks about: Bias or toxicity
310. 32:30or just hallucination. You know speaking with  great confidence about something it knows nothing
311. 32:37about and as a result we rely on that information.  Generating, that's a version of generating
312. 32:45fake information, fake news or fake images  or whatever it is. Of course impersonation.
313. 32:50It does such a good job pretending to be a  human, it could be it could do an incredibly good
314. 32:56job pretending to be a specific human. And so the spectrum of areas we
315. 33:05have to be concerned about is fairly clear and  there's a lot of people who are
316. 33:11working on it. There's a some of the stuff,  some of the stuff related to AI safety requires
317. 33:18deep research and deep engineering and  that's simply, it wants to do the right thing it
318. 33:24just didn't perform it right and as a result hurt  somebody. You know for example self-driving car
319. 33:29that wants to drive nicely and drive properly  and just somehow the sensor broke down or it
320. 33:36didn't detect something. Or you know made it  too aggressive turn or whatever it is. It did
321. 33:41it poorly. It did it wrongly. And so that's a whole bunch of engineering that has to
322. 33:47be done to to make sure that AI safety is upheld  by making sure that the product functions properly.
323. 33:54And then and then lastly you know whatever what  happens if the system, the AI wants to do a good
324. 34:00job but the system failed. Meaning the AI wanted  to stop something from happening
325. 34:07and it turned out just when it wanted to do  it, the machine broke down. And so this is
326. 34:13no different than a flight computer inside  a plane having three versions of them and then
327. 34:19so there's triple redundancy inside the  system inside autopilots and then you have two
328. 34:25pilots and then you have air traffic control  and then you have other pilots watching out for
329. 34:31these pilots. And so that the AI safety  systems has to be architected as a community
330. 34:38such that such that these AIs one, work, function properly. When they don't
331. 34:47function properly, they don't put people in harm's  way. And that they're sufficient safety and
332. 34:52security systems all around them to make sure  that we keep AI safe. And so there's
333. 34:58this spectrum of conversation is gigantic and and  you know we have to take the parts, take the
334. 35:05parts apart and and build them as engineers. One  of the incredible things about this moment that
335. 35:11we're in right now is that we no longer have a  lot of the technological limits that we had in a
336. 35:17world of CPUs and sequential processing. And we've  unlocked not only a new way to do computing and
337. 35:28and but also a way to continue to improve. Parallel  processing has a a different kind of physics to it
338. 35:35than the improvements that we were able to make  on CPUs. I'm curious, what are the scientific or
339. 35:42technological limitations that we face now in  the current world that you're thinking a lot
340. 35:47about? Well everything in the end is about how much  work you can get done within the limitations of
341. 35:54the energy that you have. And so that's  a physical limit and the laws of
342. 36:02physics about transporting information and  transporting bits, flipping bits and transporting
343. 36:11bits, at the end of the day the energy it takes  to do that limits what we can get done. And the
344. 36:18amount of energy that we have limits what we can  get done. We're far from having any fundamental
345. 36:23limits that keep us from advancing. In the meantime,  we seek to build better and more energy efficient
346. 36:29computers. This little computer, the the  big version of it was $250,000 - Pick up? - Yeah
347. 36:38Yeah that's little baby DIGITS yeah. This is  an AI supercomputer. The version that I delivered,
348. 36:46this is just a prototype so it's a mockup. The very first version was DGX 1, I
349. 36:52delivered to Open AI in 2016 and that was $250,000.  10,000 times more power, more energy necessary
350. 37:03than this version and this version has six times  more performance. I know, it's incredible. We're
351. 37:09in a whole in the world. And it's only since 2016  and so eight years later we've in increased the
352. 37:16energy efficiency of computing by 10,000 times.  And imagine if we became 10,000 times more energy
353. 37:25efficient or if a car was 10,000 times more  energy efficient or electric light bulb was
354. 37:3110,000 times more energy efficient. Our light  bulb would be right now instead of 100 Watts,
355. 37:3810,000 times less producing the same illumination.  Yeah and so the energy efficiency of
356. 37:45computing particularly for AI computing that we've  been working on has advanced incredibly and that's
357. 37:51essential because we want to create you  know more intelligent systems and and we want to
358. 37:56use more computation to be smarter and so  energy efficiency to do the work is our number one
359. 38:03priority. When I was preparing for this interview, I  spoke to a lot of my engineering friends and this
360. 38:09is a question that they really wanted me to ask. So  you're really speaking to your people here. You've
361. 38:15shown a value of increasing accessibility  and abstraction, with CUDA and allowing more
362. 38:21people to use more computing power in all kinds of  other ways. As applications of technology get more
363. 38:28specific, I'm thinking of transformers in AI for  example... For the audience, a transformer is a very
364. 38:35popular more recent structure of AI that's now  used in a huge number of the tools that you've
365. 38:40seen. The reason that they're popular is because  transformers are structured in a way that helps
366. 38:45them pay "attention" to key bits of information and  give much better results. You could build chips
367. 38:51that are perfectly suited for just one kind of AI  model, but if you do that then you're making them
368. 38:56less able to do other things. So as these specific  structures or architectures of AI get more popular,
369. 39:03my understanding is there's a debate between how  much you place these bets on "burning them into the
370. 39:09chip" or designing hardware that is very specific  to a certain task versus staying more general and
371. 39:15so my question is, how do you make those bets? How  do you think about whether the solution is a car
372. 39:22that could go anywhere or it's really optimizing  a train to go from A to B? You're making bets
373. 39:28with huge stakes and I'm curious how you think  about that. Yeah and that now comes back
374. 39:33to exactly your question, what are your  core beliefs? And the question, the core
375. 39:41belief either one, that transformer is the last AI  algorithm, AI architecture that any researcher will
376. 39:52ever discover again, or that transformers  is a stepping stone towards evolutions of
377. 40:01transformers that are uh barely recognizable as a  transformer years from now. And we believe the
378. 40:08latter. And the reason for that is because you  just have to go back in history and ask yourself,
379. 40:14in the world of computer algorithms, in  the world of software, in the world of
380. 40:20engineering and innovation, has one idea stayed  along that long? And the answer is no. And so that's
381. 40:27kind of the beauty, that's in fact  the essential beauty of a computer that it's able
382. 40:34to do something today that no one even imagined  possible 10 years ago. And if you would have, if
383. 40:41you would have turned that computer 10 years ago  into a microwave, then why would the applications
384. 40:48keep coming? And so we believe, we believe in the  richness of innovation and the
385. 40:54richness of invention and we want to create an  architecture that let inventors and innovators
386. 40:59and software programmers and AI researchers  swim in the soup and come up with some amazing
387. 41:05ideas. Look at transformers. The fundamental  characteristic of a transformer is this idea
388. 41:10called "attention mechanism" and it basically says  the transformer is going to understand the meaning
389. 41:16and the relevance of every single word with every  other word. So if you had 10 words, it has to figure
390. 41:22out the relationship across 10 of them. But if you  have a 100,000 words or if your context is
391. 41:27now as large as, read a PDF and that read a whole  bunch of PDFs, and the context window is now like
392. 41:35a million tokens, the processing all of it across  all of it is just impossible. And so the way you
393. 41:42solve that problem is there all kinds of new ideas,  flash attention or hierarchical attention or you
394. 41:49know all the, wave attention I just read about  the other day. The number of different types of
395. 41:54attention mechanisms that have been invented  since the transformer is quite extraordinary.
396. 42:00And so I think that that's going to continue  and we believe it's going to continue and that
397. 42:06that computer science hasn't ended and that AI  research have not all given up and we haven't
398. 42:12given up anyhow and that having a  computer that enables the flexibility of
399. 42:21of research and innovation and new ideas is  fundamentally the most important thing. One of the
400. 42:29things that I am just so curious about, you design  the chips. There are companies that assemble the
401. 42:37chips. There are companies that design hardware to  make it possible to work at nanometer scale. When
402. 42:44you're designing tools like this, how do you think  about design in the context of what's physically
403. 42:51possible right now to make? What are the things  that you're thinking about with sort of pushing
404. 42:56that limit today? The way we do it is even  though even though we have things made like for
405. 43:05example our chips are made by TSMC. Even though  we have them made by TSMC, we assume that we need
406. 43:13to have the deep expertise that TSMC has. And so  we have people in our company who are incredibly
407. 43:19good at semiconductive physics so that we have a  feeling for, we have an intuition for, what are the
408. 43:25limits of what today's semiconductor physics  can do. And then we work very closely with them to
409. 43:32discover the limits because we're trying to push  the limits and so we discover the limits together.
410. 43:36Now we do the same thing in system engineering and  cooling systems. It turns out plumbing is really
411. 43:41important to us because of liquid cooling.  And maybe fans are really important to us
412. 43:44because of air cooling and we're trying to design  these fans in a way almost like you know they're
413. 43:49aerodynamically sound so that we could pass the  highest volume of air, make the least amount of
414. 43:54noise. So we have aerodynamics engineers in our company. And so even though even though we don't
415. 44:01make 'em, we design them and we have to deep  expertise of knowing how to have them made. And
416. 44:09and from that we try to push the  limits. One of the themes of this conversation is
417. 44:18that you are a person who makes big bets on the  future and time and time again you've been right
418. 44:25about those bets. We've talked about GPUs, we've  talked about CUDA, we've talked about bets you've
419. 44:30made in AI - self-driving cars, and we're going to  be right on robotics and - this is my question. What
420. 44:37are the bets you're making now? the latest bet we just described at the CES and I'm very very proud
421. 44:43of it and I'm very excited about it is the  fusion of Omniverse and Cosmos so that we have
422. 44:50this new type of generative world generation  system, this multiverse generation system. I
423. 44:59think that's going to be profoundly important in  the future of robotics and physical systems.
424. 45:06Of course the work that we're doing with human  robots, developing the tooling systems and the
425. 45:11training systems and the human demonstration  systems and all of this stuff that that you've
426. 45:17already mentioned, we're just seeing the  beginnings of that work and I think the
427. 45:23next 5 years are going to be very interesting in  the world of human robotics. Of course the work
428. 45:28that we're doing in digital biology so that  we can understand the language of molecules and
429. 45:34understand the language of cells and just as  we understand the language of physics and the
430. 45:39physical world we'd like to understand the language  of the human body and understand the language of
431. 45:44biology. And so if we can learn that, and we can  predict it. Then all of a sudden our ability to
432. 45:50have a digital twin of the human is plausible.  And so I'm very excited about that work. I love
433. 45:56the work that we're doing in climate science  and be able to, from weather predictions, understand
434. 46:03and predict the high resolution regional climates,  the weather patterns within a kilometer above
435. 46:10your head. That we can somehow predict that with  great accuracy, its implications is really quite
436. 46:17profound. And so the number of things that  we're working on is really cool. You know we
437. 46:24we're fortunate that we've created this  this instrument that is a time machine and
438. 46:37we need time machines in all of these areas that  we just talked about so that we can see
439. 46:43the future. And if we could see the future and  we can predict the future then we have a better
440. 46:48chance of making that future the best version  of it. And that's the reason why scientists
441. 46:53want to predict the future. That's the reason why,  that's the reason why we try to predict the future
442. 46:58and everything that we try to design so that we  can optimize for the best version. So if
443. 47:05someone is watching this and maybe they came into  this video knowing that NVIDIA is an incredibly
444. 47:12important company but not fully understanding why  or how it might affect their life and they're now
445. 47:18hopefully better understanding a big shift that  we've gone through over the last few decades in
446. 47:23computing, this very exciting, very sort of strange  moment that we're in right now, where we're sort
447. 47:30of on the precipice of so many different things.  If they would like to be able to look into the
448. 47:36future a little bit, how would you advise them to  prepare or to think about this moment that they're
449. 47:42in personally with respect to how these tools  are actually going to affect them? Well there are
450. 47:49several ways to reason about the future that  we're creating. One way to reason about it is,
451. 47:57suppose the work that you do continues to  be important but the effort by which you
452. 48:04do it went from you know being a week long  to almost instantaneous. You know that the
453. 48:15effort of drudgery basically goes to zero.  What is the implication of that? This is, this
454. 48:23is very similar to what would change if all  of a sudden we had highways in this country?
455. 48:30And that kind of happened you know in the last  Industrial Revolution, all of a sudden we have
456. 48:34interstate highways and when you have interstate  highways what happens? Well you know suburbs start
457. 48:40to be created and and all of a sudden you know  distribution of goods from east to west is
458. 48:48no longer a concern and all of a sudden gas  stations start cropping up on highways and
459. 48:55and fast food restaurants show up and you  know someone, some motels show up because people
460. 49:03you know traveling across the state, across the  country and just wanted to stay somewhere for a
461. 49:07few hours or overnight, and so all of a sudden  new economies and new capabilities, new economies.
462. 49:13What would happen if a video conference made  it possible for us to see each other without
463. 49:19having to travel anymore? All of a sudden  it's actually okay to work further away from
464. 49:24home and from work, work and live  further away. And so you ask yourself kind of
465. 49:32these questions. You know what would happen  if I have a software programmer with me
466. 49:40all the time and whatever it is I can dream up,  the software programmer could write for me. You
467. 49:46know what would, what would happen  if I just had a seed of an idea and
468. 49:54and I rough it out and all of sudden a you know  a prototype of a production was put in front
469. 50:01of me? And what how would that change my life and  how would that change my opportunity? And you
470. 50:07know what does it free me to be able to do and  and so on so forth. And so I think that the next
471. 50:13the next decade intelligence, not for everything  but for for some things, would basically become
472. 50:22superhuman. But I can tell  you exactly what that feels like. I'm surrounded
473. 50:31by superhuman people, super intelligence from  my perspective because they're the best in the
474. 50:38world at what they do and they do what they  do way better than I can do it. and I'm
475. 50:46surrounded by thousands of them and yet what it  it never one day caused me to to think all of a
476. 50:56son I'm no longer necessary. It actually empowers  me and gives me the confidence to go tackle more
477. 51:05and more ambitious things. And so suppose,  suppose now everybody is surrounded by these
478. 51:13super AIs that are very good at specific things  or good at some of the things. What would that
479. 51:20make you feel? Well it's going to empower you,  it's going to make you feel confident and
480. 51:25and I'm pretty sure you probably use ChatGPT and  AI and I feel more empowered today, more
481. 51:32confident to learn something today. The knowledge  of almost any particular field, the barriers to
482. 51:38that understanding, it has been reduced and I have  a personal tutor with me all of the time. And
483. 51:44so I think that that feeling should be universal. If there's one thing that I would
484. 51:50encourage everybody to do is to go get yourself  an AI tutor right away. And that AI tutor could
485. 51:56of course just teach your things, anything you  like, help you program, help you write,
486. 52:03help you analyze, help you think, help you reason,  you know all of those things is going to
487. 52:10really make you feel empowered and and I think  that going to be our future. We're
488. 52:16going to become, we're going to become super humans,  not because we have super, we're going to become
489. 52:21super humans because we have super AIs. Could you  tell us a little bit about each of these objects?
490. 52:27This is a new GeForce graphics card and yes and  this is the RTX 50 Series. It is essentially
491. 52:39a supercomputer that you put into your PC and we  use it for gaming, of course people today use it
492. 52:45for design and creative arts and it does amazing  AI. The real breakthrough here and this is
493. 52:52this is truly an amazing thing, GeForce  enabled AI and it enabled Geoff Hinton, Ilya Sutskever,
494. 52:59Alex Krizhevsky to be able to train AlexNet. We discovered AI and we advanced AI then AI came back
495. 53:07to GeForce to help computer graphics. And so here's  the amazing thing: Out of 8 million pixels or so in
496. 53:16a 4K display we are computing, we're processing  only 500,000 of them. The rest of them we use AI
497. 53:24to predict. The AI guessed it and yet the image is  perfect. We inform it by the 500,000 pixels that we
498. 53:32computed and we ray traced every single one and it's  all beautiful. It's perfect. And then we tell the
499. 53:38AI, if these are the 500,000 perfect pixels in this  screen, what are the other 8 million? And it goes it
500. 53:44fills in the rest of the screen and it's perfect. And if you only have to do fewer pixels, are you
501. 53:50able to invest more in doing that because you have  fewer to do so then the quality is better so the
502. 53:58extrapolation that the AI does... Exactly. Because  whatever computing, whatever attention you have,
503. 54:03whatever resources you have, you can place it into  500,000 pixels. Now this is a perfect example of
504. 54:11why AI is going to make us all superhuman, because  all of the other things that it can do, it'll do
505. 54:17for us, allows us to take our time and energy and  focus it on the really really valuable things that
506. 54:23we do. And so we'll take our own resource which is  you know energy intensive, attention intensive, and
507. 54:33we'll dedicated to the few 100,000 pixels and  use AI to superres, upres it you know to
508. 54:39everything else. And so this this graphics card  is now powered mostly by AI and the computer
509. 54:47graphics technology inside is incredible as  well. And then this next one, as I mentioned
510. 54:52earlier, in 2016 I built the first one for AI  researchers and we delivered the first one to Open AI
511. 54:58and Elon was there to receive it and this  version I built a mini mini version and the
512. 55:06reason for that is because AI has now gone from AI  researchers to every engineer, every student, every
513. 55:15AI scientist. And AI is going to be everywhere.  And so instead of these $250,000 versions we're
514. 55:21going to make these $3,000 versions and schools  can have them, you know students can have them, and
515. 55:28you set it next to your PC or Mac and all of  a sudden you have your own AI supercomputer. And
516. 55:36you could develop and build AIs. Build your own  AI, build your own R2-D2. What do you feel like is
517. 55:42important for this audience to know that I haven't  asked? One of the most important things I would
518. 55:48advise is for example if I were a student today  the first thing I would do is to learn AI. How do
519. 55:54I learn to interact with ChatGPT, how do I learn  to interact with Gemini Pro, and how do I learn
520. 56:00to interact with Grok? Learning how to interact with with AI is not unlike being
521. 56:10someone who is really good at asking questions.  You're incredibly good at asking questions and
522. 56:17and prompting AI is very very similar. You can't just randomly ask a bunch of questions
523. 56:23and so asking an AI to be assistant  to you requires some expertise and
524. 56:30artistry and how to prompt it. And so if I were,  if I were a student today, irrespective whether
525. 56:35it's for for math or for science or chemistry  or biology or doesn't matter what field of science
526. 56:40I'm going to go into or what profession, I'm  going to ask myself, how can I use AI to do my job
527. 56:46better? If I want to be a lawyer, how can I use  AI to be a better lawyer? If I want to be a better
528. 56:50do doctor, how can I use AI to be a better doctor?  If I want to be a chemist, how do I use AI to be
529. 56:55a better chemist? If I want to be a biologist, I how  do I use AI to be a better biologist? That question
530. 57:02should be persistent across everybody. And just as  my generation grew up as the first generation
531. 57:10that has to ask ourselves, how can we use computers  to do our jobs better? Yeah the generation before
532. 57:17us had no computers, my generation was the first  generation that had to ask the question, how do I
533. 57:23use computers to do my job better? Remember I came  into the industry before Windows 95 right, 1984
534. 57:32there were no computers in offices. And after that,  shortly after that, computers started to emerge and
535. 57:38so we had to ask ourselves how do we use computers  to do our jobs better? The next generation doesn't
536. 57:45have to ask that question but it has to ask  obviously next question, how can I use AI to
537. 57:49do my job better? That is start and finish I think  for everybody. It's a really exciting and scary and
538. 57:59therefore worthwhile question I think for everyone.  I think it's going to be incredibly fun. AI is
539. 58:04obviously a word that people are just learning  now but it's just you know, it's
540. 58:10made your computer so much more accessible. It is  easier to prompt ChatGPT to ask it anything you
541. 58:15like than to go do the research yourself. And so  we've lowered a barrier of understanding, we've
542. 58:22lowered a barrier of knowledge, we've  lowered a barrier of intelligence, and
543. 58:26and everybody really had to just go try  it. You know the thing that's really really crazy
544. 58:32is if I put a computer in front of somebody and  they've never used a computer there is no chance
545. 58:37they're going to learn that computer in a day. There's just no chance. Somebody really has to
546. 58:43show it to you and yet with ChatGPT if you  don't know how to use it, all you have to do is
547. 58:49type in "I don't know how to use ChatGPT, tell  me," and it would come back and give you some
548. 58:54examples and so that's the amazing thing. You know the amazing thing about intelligence is
549. 59:02it'll help you along the way and make you uh  superhuman you know along the way. All right I have
550. 59:08one more question if you have a second. This is  not something that I planned to ask you but on the
551. 59:13way here, I'm a little bit afraid of planes,  which is not my most reasonable quality, and
552. 59:21the flight here was a little bit bumpy mhm very  bumpy and I'm sitting there and it's moving and
553. 59:30I'm thinking about what they're going to say at my  funeral and after - She asked good questions, that's
554. 59:37what the tombstone's going to say - I  hope so! Yeah. And after I loved my husband and my
555. 59:44friends and my family, the thing that I hoped that  they would talk about was optimism. I hope that
556. 59:49they would recognize what I'm trying to do here.  And I'm very curious for you, you've you've been
557. 59:56doing this a long time, it feels like there's  so much that you've described in this vision
558. 1:00:00ahead, what would the theme be that you would  want people to say about what you're trying to do?
559. 1:00:14Very simply, they made an extraordinary impact.  I think that we're fortunate because of some
560. 1:00:23core beliefs a long time ago and sticking with  those core beliefs and building upon them
561. 1:00:32we found ourselves today being one of  the most, one of the many most important and
562. 1:00:42consequential technology companies in the world and potentially ever. And so
563. 1:00:49we take that responsibility very seriously. We work hard to make sure that
564. 1:00:56the capabilities that we've created are  available to large companies as well as
565. 1:01:03individual researchers and developers, across  every field of science no matter profitable or
566. 1:01:10not, big or small, famous or otherwise. And it's because of this understanding of
567. 1:01:21the consequential work that we're doing and the  potential impact it has on so many people
568. 1:01:27that we want to make make this capability  as pervasively as possible and I
569. 1:01:37do think that when we look back in a few  years, and I do hope that what the
570. 1:01:47next generation realized is as they, well  first of all they're going to know us because of
571. 1:01:53all the you know gaming technology we create. I do think that we'll look back and the whole
572. 1:01:59field of digital biology and life sciences has  been transformed. Our whole understanding of of
573. 1:02:06material sciences has completely been  revolutionized. That robots are helping
574. 1:02:13us do dangerous and mundane things all over the  place. That if we wanted to drive we can drive
575. 1:02:19but otherwise you know take a nap or enjoy  your car like it's a home theater of yours,
576. 1:02:26you know read from work to home and at that  point you're hoping that you live far
577. 1:02:31away and so you could be in a car for longer.  And you look back and
578. 1:02:37you realize that there's this company almost at  the epicenter of all of that and happens
579. 1:02:43to be the company that you grew up playing games with.
580. 1:02:46I hope for that to be what the next generation learn.
581. 1:02:50Thank you so much for your time. I enjoyed it, thank you! I'm glad!