YouTube2Text

Jensen Huang: The Doomer Hoax, Superintelligence is Here, and The Future of AI (ft. President Trump) — Transcript

by All-In Podcast · 9,096 words · 1,503 segments · language en · Watch on YouTube

Full transcript

  1. 0:00Some people call it vision. Vision's an
  2. 0:03awfully big word to me because I I
  3. 0:04believe, first of all, vision matters.
  4. 0:07>> [music]
  5. 0:09>> We preempted the weekly show, and
  6. 0:12there's only three people we preempt the
  7. 0:13show for.
  8. 0:15President Trump, Jesus, and Jensen.
  9. 0:17[laughter]
  10. 0:18>> The number one podcast [music]
  11. 0:19in the world.
  12. 0:20>> That's Jensen Huang. He's the founder,
  13. 0:22president, and CEO of Nvidia.
  14. 0:23>> Whether you know it or not, his
  15. 0:24decisions are shaping your future.
  16. 0:26>> Nvidia is the most important stock in
  17. 0:28this market, and Jensen is arguably the
  18. 0:29most
  19. 0:29>> best executive in history.
  20. 0:31>> Revenue [music] exploded 97%
  21. 0:34year-over-year.
  22. 0:34>> Not only is demand already strong, it's
  23. 0:36actually accelerating.
  24. 0:38>> Nvidia is the only computing platform
  25. 0:40that is a full-stack AI factory. [music]
  26. 0:42A GPU is like a time machine
  27. 0:44because it lets you see the future
  28. 0:46sooner. And if we could see the future,
  29. 0:48and we can predict the future, then we
  30. 0:50have a [music] better chance of making
  31. 0:52that future the best version of it.
  32. 0:55>> Please welcome Jensen Huang.
  33. 0:58>> [music]
  34. 1:01>> Oh, we got a stand-to-go on the way in.
  35. 1:04>> Oh, come on.
  36. 1:05>> Stand-to-go.
  37. 1:07Stand-to-go on the way in.
  38. 1:10That's our guy.
  39. 1:12>> [laughter]
  40. 1:12>> Ladies and gentlemen, GPU Jesus.
  41. 1:19They love you.
  42. 1:20>> They love you.
  43. 1:20>> Thank you. I love you back.
  44. 1:22>> We like the new jacket.
  45. 1:22>> Number one podcast in the world.
  46. 1:26>> Wow.
  47. 1:27>> We like the new jacket.
  48. 1:28>> Well,
  49. 1:29>> You auctioned the old
  50. 1:30>> I just I felt you guys needed some
  51. 1:32energy.
  52. 1:33>> Yes. This is the right time.
  53. 1:34>> talking about serious stuff here, but we
  54. 1:36need to talk about it with energy.
  55. 1:38>> Yes. Let's let's start with this
  56. 1:41essay from this weekend.
  57. 1:43>> Which one?
  58. 1:44>> [laughter]
  59. 1:46>> Let's start with Dario's essay because I
  60. 1:47>> Was Hemingway involved?
  61. 1:49>> [laughter]
  62. 1:50>> No.
  63. 1:50>> Actually, did anybody run it through
  64. 1:51Pangram? I don't even know how much of
  65. 1:53it was AI helped, but that was a pretty
  66. 1:57incredible thing, and then, I think what
  67. 1:59a lot of people were surprised by was
  68. 2:01the coalescing of the Frontier Labs
  69. 2:03around the SA itself. Just, Jensen,
  70. 2:05unpack what happened, how you read it,
  71. 2:08how you interpreted it, and then we'll
  72. 2:10get into some details that were inside
  73. 2:11of it. But, maybe just the high-level
  74. 2:13thoughts to kick it off.
  75. 2:14>> Well, first of all, there were a lot of
  76. 2:15stuff in there.
  77. 2:16>> Yeah.
  78. 2:17>> And and
  79. 2:19uh the first there is a there's a part
  80. 2:20about about safety, which we have to
  81. 2:23take very seriously. Safety's paramount,
  82. 2:25obviously. Um uh safety and leadership
  83. 2:28are not false they're false choices.
  84. 2:31You're you're able to innovate quickly.
  85. 2:34You're able to execute quickly. And and
  86. 2:36America's able to lead and to do it
  87. 2:39safely. I think those are those are
  88. 2:40false choices, but safety's obviously
  89. 2:42important. Uh there's a matter of of
  90. 2:46internal control that I think he was
  91. 2:48speaking to. Uh obviously, the the
  92. 2:50Coxswain uh whistleblower is very
  93. 2:52serious matter. When whenever you have a
  94. 2:55whistleblower, you have to take it very
  95. 2:56seriously. I thought
  96. 2:58Coxswain [clears throat] had great
  97. 2:58courage uh to uh put out put out
  98. 3:02uh what his concerns were. Um and even
  99. 3:04then, there were some issues that were
  100. 3:06kind of conflated within that. Uh I
  101. 3:08think the the the whistleblowing is is
  102. 3:10fine. I think the the uh scientific
  103. 3:13prediction
  104. 3:15uh uh about the future uh is less fine
  105. 3:18because it's not grounded on science,
  106. 3:20obviously. And um uh it it was expressed
  107. 3:23by a scientist, but it was obviously not
  108. 3:24grounded on science. And and so, I take
  109. 3:26I take issue with that. But, obviously,
  110. 3:28the whistleblower part of it, uh
  111. 3:30>> [snorts]
  112. 3:30>> you know, I I think there's just a whole
  113. 3:31bunch of stuff. Uh
  114. 3:33pausing uh
  115. 3:35uh pacing uh
  116. 3:38uh those are all the voluntary things
  117. 3:39that they could do if they feel that
  118. 3:41their company is out of control.
  119. 3:43Uh if Coxswain saw something, uh you
  120. 3:45know, obviously, we didn't we don't know
  121. 3:47what Coxswain saw. But, he if he saw
  122. 3:49that the company was out of control
  123. 3:52and maybe it's a transition from uh
  124. 3:56research to engineering. As you know,
  125. 3:58these labs are research are
  126. 4:00transitioning from research to
  127. 4:01engineering. Extraordinary talent,
  128. 4:03extraordinary engineering. Um but
  129. 4:05obviously engineering is different than
  130. 4:06research. Maybe that transition is
  131. 4:08clumsy. You know, I don't We don't know
  132. 4:10what what he saw and ultimately only he
  133. 4:12knows. Um but if there was a matter of
  134. 4:15lack of control, that's a different
  135. 4:16topic. Um how should the government deal
  136. 4:18with it? Now all of a sudden uh
  137. 4:20regulation and I mean it just covers
  138. 4:23everything in one block.
  139. 4:24>> just help us sort of unpack? We we play
  140. 4:26We tried to play this game actually this
  141. 4:27week on the pod and it was difficult,
  142. 4:29which is
  143. 4:30how do you describe like, you know, my
  144. 4:32mom calls me and she's like, "Chamath,
  145. 4:33what is this whole
  146. 4:35civilizational death thing?" I don't
  147. 4:37know how to explain it to her. So, when
  148. 4:39you have very smart people like that,
  149. 4:41quantize it and quantify it. I think
  150. 4:43that's probably what's perturbing to
  151. 4:45some people. They're like, "What does
  152. 4:46that mean? 10% of extinction?"
  153. 4:48Nobody knows how to explain that to the
  154. 4:50average person, how that's even
  155. 4:52possible.
  156. 4:53>> Well, first of all, we shouldn't.
  157. 4:55Uh because it's made up. Uh first of
  158. 4:57all, I THINK THAT
  159. 4:59>> [laughter]
  160. 5:01[applause]
  161. 5:03>> WE SHOULDN'T BECAUSE IT'S MADE UP. AND
  162. 5:05THESE ARE These are well-educated.
  163. 5:08Uh they're called researchers. Um
  164. 5:11obviously they're working a lab. And so,
  165. 5:14the confluence of these words and then
  166. 5:17and then and then the prediction is
  167. 5:18alarming and troubling and it shouldn't
  168. 5:20be done. It's It's It's irresponsible.
  169. 5:23Now, the fact of the matter is, let's go
  170. 5:24back and look at the real facts. The
  171. 5:25facts are uh there was a prediction that
  172. 5:28in 5 years' time, radiology will be
  173. 5:31completely taken over by artificial
  174. 5:32intelligence and there'll be no
  175. 5:33radiologists in the world. That has
  176. 5:35proven to be exactly the opposite. We
  177. 5:37need more radiologists than ever in the
  178. 5:39world. However, AI has taken over
  179. 5:41radiology completely, which is great. It
  180. 5:43has automated scan reading, which is
  181. 5:45great. Um uh, there was a prediction
  182. 5:48that within 6 to 12 months, wasn't it
  183. 5:50just last year, within 6 to 12 months,
  184. 5:53uh, 90% of code would already be
  185. 5:56generated, uh, by AI. That has turned
  186. 5:59out to be wrong. Uh, within 6 to 9
  187. 6:01months, that was predicted last year,
  188. 6:0350% of entry jobs will be wiped out.
  189. 6:05That has proven to be wrong.
  190. 6:07Uh, let's see, what else what else has
  191. 6:08proven to be wrong? I mean, all of these
  192. 6:10predictions have been wrong.
  193. 6:12>> Right.
  194. 6:13>> And if you're a
  195. 6:14>> that GPT-2 would be too unsafe to
  196. 6:16release, that Llama 3 would be too
  197. 6:18unsafe to release.
  198. 6:20>> Yeah, we've heard these.
  199. 6:21>> Half of white-collar jobs would be gone
  200. 6:23next year.
  201. 6:23>> The jobs apocalypse, yeah.
  202. 6:25>> Yeah.
  203. 6:25>> We have to take accountability. We have
  204. 6:28to take account for all of the stupid
  205. 6:30predictions that were made.
  206. 6:32>> Right.
  207. 6:33>> Somebody has Somebody has to take
  208. 6:36>> No.
  209. 6:36>> [applause]
  210. 6:37>> And
  211. 6:38And so, we ought to just keep track of
  212. 6:39all that, and of course, people do, and
  213. 6:42remind us that those those predictions
  214. 6:45are inconsistent
  215. 6:47with
  216. 6:48ultimately, America winning the AI race.
  217. 6:50>> The short form for that is some people
  218. 6:52are saying, you know, they say, "Trust
  219. 6:54the experts." [clears throat] And they
  220. 6:55use the analog of COVID, which again,
  221. 6:58started with people that were
  222. 6:59researchers, educated people that had an
  223. 7:01asymmetric awareness of the thing that
  224. 7:03the rest of us did not, saying things
  225. 7:05that ultimately turned out, we find out
  226. 7:07in facts, uh, not to be true. Um, and
  227. 7:10so, there's this war that's happening
  228. 7:12right now between the trust the experts
  229. 7:15movement and the, you know, "Well, let's
  230. 7:17just look at the actual history of these
  231. 7:18predictions, and let's just think more
  232. 7:20methodically."
  233. 7:22Where is this coming from? Because it's
  234. 7:24coming from inside the places that's
  235. 7:26actually making it. Like, what do you
  236. 7:27think is the psychological makeup, or
  237. 7:29what is the real incentive? Maybe it's a
  238. 7:31business incentive. Maybe it's a
  239. 7:32political incentive. Can you just maybe
  240. 7:34guess, or how do you how do you think
  241. 7:36about what why they're doing this?
  242. 7:38>> Well, first of all, I got to tell you,
  243. 7:39these [clears throat] are some of the
  244. 7:40most consequential companies in history.
  245. 7:43Uh
  246. 7:43uh, extraordinary engineers,
  247. 7:45extraordinary researchers, really
  248. 7:47fantastic work. Um
  249. 7:49>> [clears throat]
  250. 7:50>> Uh on the one hand, uh I work very
  251. 7:52closely with them as companies to
  252. 7:53companies.
  253. 7:54Uh
  254. 7:55on the other hand, uh we have to have
  255. 7:57conversations like this in public. And
  256. 8:00it's really unfortunate. And I I think
  257. 8:02that that these these companies um
  258. 8:05really ought to be built the way that we
  259. 8:07used to build companies, which is in
  260. 8:09silence.
  261. 8:10>> Right.
  262. 8:11>> You know, and
  263. 8:13>> So wait, Jensen, you don't allow anybody
  264. 8:16in your organization to speak for the
  265. 8:17entire organization, especially when
  266. 8:19they're having like a bad weekend or
  267. 8:21they rage quit? They're they're not
  268. 8:22allowed to tweet on your behalf and the
  269. 8:24organization's behalf?
  270. 8:25>> Uh no, because well, that's that's what
  271. 8:28they decided when they came to work for
  272. 8:29us. And we told them, uh these are this
  273. 8:32is the way you behave when you work in
  274. 8:33our company. And and if you would like
  275. 8:36if you like the culture of our company,
  276. 8:38um uh which as you know, the Nvidia
  277. 8:41culture and the Nvidia employee base,
  278. 8:43incredibly happy.
  279. 8:44>> Yeah.
  280. 8:44>> Uh they like the fact that the company
  281. 8:46is consistent, that we're stable, that
  282. 8:48our core values are consistent with
  283. 8:51taking care of the families and creating
  284. 8:52the conditions by which they can do
  285. 8:54their life's work. Um
  286. 8:56that we do meaningful work. We do it We
  287. 8:58do it as quietly as we can. And we
  288. 9:01contribute to everybody else's success,
  289. 9:02which we're very proud of. And so those
  290. 9:05kind of core values people are attracted
  291. 9:06to. Um but when you come in work in our
  292. 9:09company, there are also some things that
  293. 9:11we don't appreciate that you do. Like
  294. 9:12for example, we don't welcome uh
  295. 9:15political discourse in our inside our
  296. 9:17company.
  297. 9:18Take it home. You guys talk about
  298. 9:19politics outside the company. Um we
  299. 9:23>> Yeah.
  300. 9:23>> [applause]
  301. 9:27>> Uh
  302. 9:27we we are
  303. 9:29the company is an apolitical company.
  304. 9:31You know, we're bipartisan. We want
  305. 9:33America to succeed.
  306. 9:34And and um uh we want we want uh
  307. 9:38whatever uh uh government is in place uh
  308. 9:41we'll do everything in our power to help
  309. 9:42America succeed. And so so that the
  310. 9:45discourse
  311. 9:46uh about about uh about race and
  312. 9:49religion and politics and
  313. 9:53all of that stuff, we tell people do it
  314. 9:56outside the company. It's not not for
  315. 9:57us.
  316. 9:58>> In terms of uh maybe AI regulation then
  317. 10:00more narrowly, um Satya was here this
  318. 10:02morning and what he said is, you know,
  319. 10:03before we talk about regulation that
  320. 10:06could really stymie things, why don't we
  321. 10:07just get some basics right? Why don't we
  322. 10:09get measurement right? Why don't we get
  323. 10:10standardization right, right? Um
  324. 10:13where do you land on Get engineering
  325. 10:14right?
  326. 10:15>> Get the engineering right, right.
  327. 10:16Translate the research in a more
  328. 10:17predictable way so that we're not
  329. 10:18fear-mongering. Keep it inside until
  330. 10:20we're ready to expose it. Um what do you
  331. 10:22think the right response is? You know,
  332. 10:24Demis had a proposal which was sort of
  333. 10:25this more FINRA-like organization.
  334. 10:28It's not clear what Dario wants, this
  335. 10:30transnational mutated thing that has
  336. 10:33some sort of control. Where do you land
  337. 10:35on this, the sort of perspective of what
  338. 10:37what do we need right now?
  339. 10:38>> You know, regulation should solve actual
  340. 10:41problems.
  341. 10:43And so the question is, what actual
  342. 10:44problems have we enjoyed?
  343. 10:45>> Right.
  344. 10:46>> And and um
  345. 10:48uh if you look at look at the actual
  346. 10:50problems, um all of the actual problems
  347. 10:52so far have come from the labs.
  348. 10:55And the reason for that The reason for
  349. 10:57that and and and just in their defense,
  350. 10:59the reason for that is because they have
  351. 11:00the most compute.
  352. 11:01>> Right.
  353. 11:02>> And the reason for that is because
  354. 11:03they're trying to solve uh the frontier
  355. 11:05problems. And so in their defense, and
  356. 11:08so it's sensible that that um the labs,
  357. 11:13the frontier labs, will be where the
  358. 11:15most danger come from.
  359. 11:17It is unlikely that a high school
  360. 11:19student
  361. 11:20uh did something because they just
  362. 11:22simply won't have enough compute.
  363. 11:23>> Right.
  364. 11:24>> And so
  365. 11:25uh it's unlikely that a startup will be
  366. 11:27the reason because they won't have
  367. 11:28enough compute. It's it's uh they you
  368. 11:30know, in fact, you you look across the
  369. 11:32planet and everybody won't have enough
  370. 11:34compute with the exception of the
  371. 11:35frontier labs.
  372. 11:36And so, so now the question is if you
  373. 11:38look at what actually happened. Um and
  374. 11:41they're doing pioneering work. It's
  375. 11:42really very hard. Um they're
  376. 11:44transitioning from research to
  377. 11:45engineering. Um I could imagine and
  378. 11:48they're they're they're they're
  379. 11:49obviously building some of the most
  380. 11:51consequential technology and companies
  381. 11:53in the world. Uh they're building their
  382. 11:55company, they're building their culture,
  383. 11:56they're building the technology, they're
  384. 11:58building engineering, they're building
  385. 11:59products all at the same time. And so, I
  386. 12:01I can understand it's a little bit hair
  387. 12:03on fire.
  388. 12:04Um uh but nonetheless,
  389. 12:06uh the four incidents from one lab, the
  390. 12:09one giant incident from the other lab,
  391. 12:11um the first thing that you have to do
  392. 12:13is just root cause the problem from an
  393. 12:15engineering perspective. What happened?
  394. 12:18What could we have done differently?
  395. 12:20And what are we going to to implement
  396. 12:23and institutionalize, whether it's
  397. 12:24technology or methods or processes, and
  398. 12:27make sure that we don't let it happen
  399. 12:28again. Now, I would bet you money that
  400. 12:31in every single one of those cases is
  401. 12:33within their control
  402. 12:35in the future to prevent it.
  403. 12:39Because the alternative it if it's not
  404. 12:42in their control, and I'm sure that they
  405. 12:43are. I'm sure I'm sure that I'm sure
  406. 12:45those four four incidents won't happen
  407. 12:47again. Um they I'm sure they root caused
  408. 12:49it and fixed it. I'm sure uh they have
  409. 12:52now technology much better technology
  410. 12:54for, you know, sandboxes and runtimes
  411. 12:57and monitors and continuous in
  412. 12:59continuous monitors. And you know, and
  413. 13:01so, I'm I'm certain they have much much
  414. 13:03better technology now.
  415. 13:05The alternative is also
  416. 13:07unlikely, which is for them to say,
  417. 13:10"Look, we had these
  418. 13:11incidents.
  419. 13:13After we're done analyzing it, we came
  420. 13:15to the conclusion we don't know anything
  421. 13:17that happened,
  422. 13:18and we have no idea how to control it,
  423. 13:21and we're asking society for help."
  424. 13:23>> Yeah.
  425. 13:23>> Now, if that's the case, then we ought
  426. 13:25to, you know, a bunch of a bunch of
  427. 13:27companies with engineers ought to send
  428. 13:29engineers in. I mean
  429. 13:30and we should advise them if we can, but
  430. 13:32I doubt it. I think they they have
  431. 13:34extraordinary people they got this
  432. 13:35handled.
  433. 13:35>> are we're not operating in a vacuum.
  434. 13:37[clears throat] David, last night you
  435. 13:38informed me that there is a Chinese lab,
  436. 13:40the makers of GLM, who are going to put
  437. 13:423 billion towards a recursive
  438. 13:44self-improvement run. So, maybe you
  439. 13:46could tee that up for Jensen.
  440. 13:48>> what was announced. Yeah, zhipu.com the
  441. 13:50founder just raised 5 billion and said
  442. 13:52that one of their priorities going to be
  443. 13:54trying to get
  444. 13:55recursive, you know,
  445. 13:56AI that trains the next AI and to try
  446. 14:00and automate as much of that as
  447. 14:01possible.
  448. 14:03Yeah, I think that I mean
  449. 14:04>> Well, this is the new sexy phrase, but
  450. 14:06as you guys [snorts] know, RSI is a
  451. 14:09combination of a system of ideas. It's
  452. 14:13it starts everything with
  453. 14:15in context stuff. It starts with skills.
  454. 14:17It starts with reflection. It starts
  455. 14:19with, you know, reinforcement learning
  456. 14:21and synthetic data generation. And these
  457. 14:23are all very sensible ideas that causes
  458. 14:26AI to get better at solving a problem,
  459. 14:30you know, over time. And you could also
  460. 14:32have low rank
  461. 14:35All of that stuff doesn't include the
  462. 14:36weights. You could actually improve the
  463. 14:38weights and it's called Laura.
  464. 14:40Laura could be could be improved in
  465. 14:42synthetic data generation, reinforcement
  466. 14:45learning, enhance it without training
  467. 14:47the the base model itself. And then over
  468. 14:49time, you could train the base model
  469. 14:51again with all of that experience. And
  470. 14:53so I I think I think it's a sensible
  471. 14:55thing that that you're going to use the
  472. 14:58technology
  473. 15:00to enhance productivity of all kinds of
  474. 15:02tasks including building AI. I think
  475. 15:05that's a very logical idea. And and I'm
  476. 15:07I'm certain that everybody is using it
  477. 15:09in some degree. It's just this phrase is
  478. 15:12now being used
  479. 15:14to weaponize the technology in some way
  480. 15:17and maybe to turn the the technology
  481. 15:19>> going to spiral out out control is the
  482. 15:22impression they're trying to give, but
  483. 15:23you don't believe that's real.
  484. 15:25>> No, no, of course not. And the reason
  485. 15:26for that is because you could RSI all
  486. 15:28day long inside your company, but when
  487. 15:31you release a product, you've got to
  488. 15:33evaluate it, don't you? You have to test
  489. 15:34it again, don't you? You have to make
  490. 15:36sure that there's no regression, right?
  491. 15:38And so the basic process of control
  492. 15:43These labs are going to as they move
  493. 15:44from labs to engineering, they will have
  494. 15:46much, much better control.
  495. 15:48>> Right.
  496. 15:49>> And when they have much better control,
  497. 15:50and that that control comes from methods
  498. 15:52and knowledge and practice and tools and
  499. 15:55technology. All of those things that
  500. 15:56leads to better control, verification
  501. 15:59and eval.
  502. 16:00It's going to enable RSI to be done
  503. 16:03inside the company and for good products
  504. 16:04to be released outside the company.
  505. 16:06>> open source for a second. I mean, this
  506. 16:07hugging face
  507. 16:09we we were communicating about this and
  508. 16:10I said it's going to be one of the most
  509. 16:11consequential
  510. 16:13um
  511. 16:15acquisitions. I don't even want to call
  512. 16:16it a transaction cuz I think it's more
  513. 16:18important than that.
  514. 16:20Give us your first principles
  515. 16:22explanation of open source versus closed
  516. 16:24source versus open weights and how the
  517. 16:26ecosystem should fit together over time.
  518. 16:28>> The world needs both closed models and
  519. 16:30open models. Um you want you want to use
  520. 16:33I use as much closed models as I can.
  521. 16:35This weekend I I used four of them. And
  522. 16:38and [clears throat] they work
  523. 16:38terrifically. They're frontier, they're
  524. 16:40great experience. They right, they work
  525. 16:42incredibly well. They're getting better
  526. 16:43all the time.
  527. 16:44Uh and and the way I think about closed
  528. 16:46closed closed models is kind of like
  529. 16:49bottled water,
  530. 16:50you know?
  531. 16:52Water is free, you guys. I don't know if
  532. 16:54I've told you guys, but water is free. I
  533. 16:56I don't want to, you know, burst
  534. 16:57everybody's bubble, but water is free.
  535. 16:59And this morning I used a lot of free
  536. 17:00water taking a shower. And so
  537. 17:03you use the right water in the right
  538. 17:05places. And this is no different than
  539. 17:07electricity. This is, you know, this is
  540. 17:09no different than all kinds of
  541. 17:10commodities that we use in the world.
  542. 17:12You need both. Now, in the case of open,
  543. 17:14the The why you you need it is because
  544. 17:17it could be for sovereignty reasons,
  545. 17:19privacy reasons,
  546. 17:20uh proprietary technology reasons. Look
  547. 17:22at the facts. The facts are in the last
  548. 17:256 months
  549. 17:27$400 billion
  550. 17:28of venture funding went into AI native
  551. 17:31companies.
  552. 17:3280% of them use open models. If not for
  553. 17:35open models, how could they build their
  554. 17:38dream?
  555. 17:39>> Right.
  556. 17:39>> Because their dream could be different.
  557. 17:41Obviously, it'll be different than the
  558. 17:43labs, the frontier labs dreams. I mean,
  559. 17:45there's America has so many different
  560. 17:47ways to innovate. That's one of our core
  561. 17:49strengths.
  562. 17:50Great ideas are just coming out of the
  563. 17:51fountain. And And so, open models
  564. 17:54enables that. Open models enables every
  565. 17:56single If we want to win the AI race,
  566. 17:59it's not about a few technology
  567. 18:01companies winning the AI race. It's
  568. 18:03about every company in in America. Every
  569. 18:06comp Every company, every industry,
  570. 18:08every
  571. 18:09researcher, every
  572. 18:11teacher, every student, every startup,
  573. 18:14everybody wins. Some of them will use
  574. 18:17closed models. A lot of them will use
  575. 18:19open models. There's 10 million open
  576. 18:21source
  577. 18:22>> Does it matter if the breakdown Well,
  578. 18:23let me just ask, does it matter if the
  579. 18:24model the open models come from China or
  580. 18:27the US?
  581. 18:28>> Well, we're doing everything we can
  582. 18:30um to make a contribution in open
  583. 18:32models.
  584. 18:34However, the moment you download
  585. 18:37Like for example,
  586. 18:39probably the vast majority of the
  587. 18:41world's contribution to open source
  588. 18:43today is coming from China. They just
  589. 18:45have a lot more engineers. They produce
  590. 18:47everything in large scale because it's a
  591. 18:49larger country. And so, they produce
  592. 18:51science and math students in volume.
  593. 18:53Right. That's one of our disadvantages.
  594. 18:55They're manufacturing them through
  595. 18:57amazing universities like Tsinghua
  596. 18:59University in high volume. Well, they
  597. 19:01contribute to open source today. We
  598. 19:03download Linux. We download Kubernetes.
  599. 19:05We download all the software. A lot of
  600. 19:08it has been touched by Chinese.
  601. 19:11Once you download it, it's yours. We
  602. 19:13fork it, we improve it, we make it ours.
  603. 19:15And so, we
  604. 19:17when you download one of these Chinese
  605. 19:19models, it just happens to be made by
  606. 19:21some really great researchers in China,
  607. 19:23but it's now yours.
  608. 19:25>> So, what exactly
  609. 19:25>> Whatever you want to do with it.
  610. 19:26>> So, what exactly is the race?
  611. 19:29>> The race?
  612. 19:30>> Yeah.
  613. 19:31>> I think that's That's a really good
  614. 19:32point. My point is, the race is really
  615. 19:35about who exploits the technology best.
  616. 19:40You know, in the last industrial
  617. 19:41revolution, all of the inventors were
  618. 19:44Maxwell, Volta, Ampere. None of them
  619. 19:46were American.
  620. 19:48They were Right? The last industrial
  621. 19:50revolution came from Europe, but we
  622. 19:52exploited it. We took advantage of it
  623. 19:55socially better than anybody else in the
  624. 19:57world, and look how it turned out for
  625. 19:59us. I want to make sure that this next
  626. 20:01generation happens just like this.
  627. 20:03>> Yeah, yeah. So, why why are the
  628. 20:05communists getting their message out so
  629. 20:07successfully here right now?
  630. 20:11>> [laughter]
  631. 20:14[clears throat]
  632. 20:14>> You know, I I think first of all, the
  633. 20:16narrative is much more practical.
  634. 20:19The narrative is much more practical.
  635. 20:21Nobody in China is saying that there's
  636. 20:24end of this and end of that and
  637. 20:27you know, cataclysmic this and
  638. 20:30you know,
  639. 20:30>> Doomer that.
  640. 20:31>> Doomer that.
  641. 20:32>> They're just much more pragmatic about
  642. 20:33it. They see AI as a technology that's
  643. 20:35going to advance their economy, advance
  644. 20:37their society. And they don't have these
  645. 20:40these groups who are basically saying
  646. 20:41it's going to end civilization.
  647. 20:43>> And we're making it up. The part that is
  648. 20:45frustrating is if it was true, if it was
  649. 20:47true, then we ought to talk about it and
  650. 20:49go do something about it.
  651. 20:50>> Right.
  652. 20:51>> Even even if it's true,
  653. 20:53we ought to spend more time doing
  654. 20:55something about it than worrying a bunch
  655. 20:57of people who can't do anything about
  656. 20:58it. It's our job to build it right.
  657. 21:00>> And that's Has there ever been a point
  658. 21:02in history where so many people have so
  659. 21:05vehemently said something that is so
  660. 21:06untrue?
  661. 21:08>> And they're measurably they're they're
  662. 21:09actually demonstrably untrue. And it
  663. 21:12actually makes sense as untrue.
  664. 21:15It's not based on science. It's not
  665. 21:16based on research. Everything that's
  666. 21:18based on science and research proves
  667. 21:19otherwise.
  668. 21:20>> of the frontier. Humans have never been
  669. 21:22there. We've never seen it. Therefore,
  670. 21:23we're scared of it. And therefore, it's
  671. 21:25easy to tell everyone to be scared of
  672. 21:26it.
  673. 21:26>> it could be life experience as well,
  674. 21:28David. Um so let let me give you an
  675. 21:30example.
  676. 21:31When I first graduated from school, I
  677. 21:33was an engineer and I didn't do that
  678. 21:35much typing.
  679. 21:36And the reason for for that is because I
  680. 21:38was the first generation before software
  681. 21:40software became popular. We had to go
  682. 21:42build the computers to make software pop
  683. 21:44possible. Could you imagine
  684. 21:47in this generation every single engineer
  685. 21:49who came into the world of engineering,
  686. 21:51you spend all your time typing.
  687. 21:54Literally, that's what you do.
  688. 21:56When you get a job, they give you a
  689. 21:57laptop, they give you a chair, and you
  690. 21:59start typing. You you type all day long.
  691. 22:02You type from the moment you wake up to
  692. 22:03the
  693. 22:04Well, there was engineering before
  694. 22:05typing.
  695. 22:06>> Right.
  696. 22:07>> And so so can you imagine that the world
  697. 22:10has a amount of engineering work to do
  698. 22:12where most of it is not typing anymore?
  699. 22:15Sure.
  700. 22:16We were We had busy engineers before
  701. 22:18typing. I think we're going to do a lot
  702. 22:20of great engineering after typing.
  703. 22:22>> Yeah.
  704. 22:22>> When I say typing, I mean coding. I mean
  705. 22:25yeah. And so even at Nvidia, when
  706. 22:27software engineers talk to me, I I tell
  707. 22:29them
  708. 22:30you're just typing. I've been saying
  709. 22:32that forever, but obviously for fun. And
  710. 22:35I I told I tell them my favorite key is
  711. 22:38backspace.
  712. 22:40And and the reason for that is because
  713. 22:42the best software is the smallest
  714. 22:43software. Yeah. So I I want you to use
  715. 22:46backspace as often as
  716. 22:47>> actually talk about Nvidia. I Let's
  717. 22:50Let's do a little teardown of Nvidia. So
  718. 22:53teardown meaning just explain the pieces
  719. 22:54cuz there's a lot of strategy at play.
  720. 22:57Let's start at the absolute bottom. So
  721. 22:58>> Oh, no.
  722. 23:03>> This is not planned, but we know who it
  723. 23:04is.
  724. 23:05>> Oh, no.
  725. 23:07No.
  726. 23:09Mr. President.
  727. 23:12>> Oh, yes, sir.
  728. 23:14Um I got to tell you something.
  729. 23:16It
  730. 23:16I w-
  731. 23:17I uh if it wasn't because of you
  732. 23:19calling, I would I'm on stage with the
  733. 23:22besties. I'm on stage with the besties.
  734. 23:24>> [laughter]
  735. 23:25>> I'm on stage with the besties. I'm on
  736. 23:27stage with Sachs and
  737. 23:29>> Oh my god.
  738. 23:30>> You kn- you know the whole group.
  739. 23:33Yeah, Jason's here, Chamath's here,
  740. 23:35David and David's here.
  741. 23:37Yeah, I'm sitting in front of a few
  742. 23:38thousand people.
  743. 23:41And we're talking as it turns out we
  744. 23:42were we were talking about you.
  745. 23:45>> [laughter]
  746. 23:48>> Good job, sir. Good job. The fact that
  747. 23:50you saw through all of that, I mean,
  748. 23:51there's a lot of complexity and the fact
  749. 23:53that it matters you saw through all of
  750. 23:55that and and I you know, we're all just
  751. 23:57really grateful.
  752. 24:00>> Tell him I said hi.
  753. 24:02>> [laughter]
  754. 24:04>> So, you want to say
  755. 24:05hi to the crowd?
  756. 24:06>> Jason would like Jason would like to put
  757. 24:08you on the
  758. 24:08>> Even Jason.
  759. 24:09>> Come on, speaker speaker mode.
  760. 24:12>> How do you put it How do we put on
  761. 24:13POTUS?
  762. 24:15On speaker? Put him on speaker? Speaker?
  763. 24:16Yeah.
  764. 24:17Right into the microphone.
  765. 24:18>> Yeah, we're going to get a mic.
  766. 24:19>> Hang on a second, sir.
  767. 24:19>> Hold on, sir. We're getting a
  768. 24:20microphone.
  769. 24:21>> Mr. President.
  770. 24:22>> Yes.
  771. 24:23>> You're now talking to the planet.
  772. 24:25>> See, the great thing about life is that
  773. 24:28Jensen can develop the most complex
  774. 24:30computer chip in the world that nobody
  775. 24:32can copy for 10 years, but he can't
  776. 24:34figure out to put me ON SPEAKER. THAT'S
  777. 24:36>> [laughter]
  778. 24:39>> THIS THING, WE HAVE TO REMEMBER THIS
  779. 24:40ONE.
  780. 24:41SO,
  781. 24:42>> ROASTED.
  782. 24:43>> the AI, it's almost a conspiracy.
  783. 24:46And the happiest group is China, and
  784. 24:49China's very happy. And I can even say
  785. 24:51in the country, a lot of states are
  786. 24:54happy that weren't going to get anything
  787. 24:56because they're being
  788. 24:57uh inundated by people that want to be
  789. 24:59there. But now all of a sudden you see
  790. 25:01they're building in Finland. They want
  791. 25:02to build one Google wants to build a big
  792. 25:04one in Finland, which I'm not happy
  793. 25:06about because they were unable to get
  794. 25:08permitting. And I'm telling you it's all
  795. 25:10a hoax. The data centers are great and
  796. 25:13they make people wealthy and they make
  797. 25:15states wealthy and it's the oil of the
  798. 25:17next 20 25 years. It's bigger than the
  799. 25:19internet and the AI,
  800. 25:21you know, much more so.
  801. 25:23And they're just playing right into the
  802. 25:26hands of a lot of people that don't want
  803. 25:28to see it happen and that could be
  804. 25:30political people and could also be
  805. 25:32China.
  806. 25:33And we're not going to let that happen.
  807. 25:34It's a it's a hoax. And
  808. 25:37>> You're right. We're not going to let
  809. 25:38that happen, sir.
  810. 25:39>> No, we're not going to let it happen.
  811. 25:41The The robots are not going to be
  812. 25:43taking over the world and that's not
  813. 25:46going to happen. You know, my uncle was
  814. 25:48a the top probably maybe the best of all
  815. 25:51time frankly but professors at MIT for
  816. 25:5441 42 years
  817. 25:57and can known as being one of the most
  818. 25:59brilliant men and he was he was there
  819. 26:02for 41 years as
  820. 26:04the top. He was like at the top top of
  821. 26:06the ladder top of did many things.
  822. 26:08Jensen knows all about it but did many
  823. 26:10things. So I have a little genetic
  824. 26:13a little genetic strength if you believe
  825. 26:15in the resource [laughter] theory but I
  826. 26:16do. I have genetics I
  827. 26:18>> That That explains why you know so much
  828. 26:20about AI.
  829. 26:22>> Well, I know about AI. I know I also
  830. 26:24have common sense about AI.
  831. 26:26The robots will not be taking over
  832. 26:28the AI will not be taking over the rest
  833. 26:31of the world. The whole thing is a hoax.
  834. 26:33Now with that we have to be a little bit
  835. 26:35careful. We have to very be you know, we
  836. 26:38have to do things and we have to do them
  837. 26:40prudently but that doesn't mean we're
  838. 26:42going to stop an industry because you
  839. 26:44know, as we work on the next 10 years
  840. 26:46about how to destroy it. So I'm with you
  841. 26:48all the way. I didn't even know how you
  842. 26:50felt about it. I assumed you felt the
  843. 26:51same way as me.
  844. 26:52>> Yes, sir.
  845. 26:53>> And we If we're going to lead, and I
  846. 26:55have an expression, it's whoever wins AI
  847. 26:57wins. That's how big it is. It's bigger
  848. 26:59than the internet. And whoever wins AI
  849. 27:02wins, and we can't let this kind of
  850. 27:03stuff happen. And that includes very
  851. 27:06much includes data centers. There are
  852. 27:07communities that were dying that have
  853. 27:10data centers right now, and now they're
  854. 27:11wealthy communities.
  855. 27:13>> Yeah.
  856. 27:13>> Really wealthy communities.
  857. 27:14>> We're We're going to make sure that
  858. 27:16We're going to make sure that everybody
  859. 27:17We're going to make sure that everybody
  860. 27:19wins in the AI race in America. Every
  861. 27:21industry, every company, every state,
  862. 27:23every people.
  863. 27:25>> Good. Well, I feel strongly about it,
  864. 27:27and I have the position that can do
  865. 27:28something about it. We're not going to
  866. 27:29let that stuff happen. So, I have no
  867. 27:32idea who's at the meeting. I have no
  868. 27:33idea who the hell I'm talking to, but
  869. 27:35I'll say
  870. 27:35>> [laughter]
  871. 27:38>> Did you Did you hear that? Did you hear
  872. 27:40that? Thousands of people are clapping
  873. 27:43for you, sir.
  874. 27:44>> All I know if you're there
  875. 27:46>> [applause]
  876. 27:47>> they'll listen to Jensen, but uh he's
  877. 27:50done an amazing job, and David has done
  878. 27:52an amazing job, and and good luck to
  879. 27:54everybody, and uh
  880. 27:56we're going to stay with the future. The
  881. 27:57country has never done better. We have
  882. 27:59$20 trillion of investment coming into
  883. 28:01the country, and that's as opposed to
  884. 28:04much less than $1 trillion under Sleepy
  885. 28:07Joe Biden. And that was 4 [laughter]
  886. 28:08years. This is in 1 year.
  887. 28:11So, you know, it's It's really the
  888. 28:12country has has never The country has
  889. 28:14never seen anything like it, and we're
  890. 28:16going to keep it going. And so, thank
  891. 28:18you all very much.
  892. 28:19>> Thank you, Mr.
  893. 28:20>> Thank you, Mr. President.
  894. 28:21>> Okay. I'll call you back later. Thank
  895. 28:24you, Mr. President. Thank you.
  896. 28:26Um
  897. 28:27I
  898. 28:27>> Well, that was unique.
  899. 28:28>> I thought it was a bit
  900. 28:30>> Did you know [laughter] that was
  901. 28:30happening?
  902. 28:31>> I thought it was a bit.
  903. 28:32>> You thought it was a bit?
  904. 28:33>> You thought it was a bit?
  905. 28:34>> No, it was real.
  906. 28:35>> I thought it was a bit at first, and I
  907. 28:36was like, put him on speakerphone.
  908. 28:38>> [laughter]
  909. 28:38>> Wow.
  910. 28:40And he calls [clears throat] you
  911. 28:41>> How do you How do you think
  912. 28:42>> He calls you any hour of the night,
  913. 28:43right?
  914. 28:43>> Well, we we were we were in the uh we
  915. 28:45were in the Oval that time when he
  916. 28:47called you.
  917. 28:47>> I was I was
  918. 28:48>> was I was in the Oval.
  919. 28:49>> You were asleep and he like said, "Wake
  920. 28:50him up." And he I felt I felt so bad
  921. 28:53because he's like, "Who's coming to this
  922. 28:54dinner?" And we go through the list and
  923. 28:55he's like, "Well, what about Jensen?" I
  924. 28:56said, "No, sir. We I he's on vacation
  925. 28:58cuz he had to postpone this vacation for
  926. 29:005 years."
  927. 29:02And he's like, "Get him on the phone."
  928. 29:03>> [laughter]
  929. 29:04>> What's vacation?
  930. 29:05>> But what do you Why do you think he sees
  931. 29:07through the hoax? It's It's This is the
  932. 29:08thing I was asking you
  933. 29:09>> quite an extraordinary thing.
  934. 29:10>> It was It's polling minus 80.
  935. 29:12>> Yeah.
  936. 29:13>> So, for anyone else that's sitting in
  937. 29:14the Oval Office, you're going to do
  938. 29:16what's popular. You're representing the
  939. 29:18people. This is what everyone wants.
  940. 29:20They want to shut down the data centers,
  941. 29:21end AI. It seems to be the popular thing
  942. 29:24in the moment. But he says it's a hoax,
  943. 29:26and he calls it. How does he do that?
  944. 29:29>> I got to tell you, I'm not sure. And the
  945. 29:30reason for that is because a lot of
  946. 29:31people are falling for it. And so, the
  947. 29:33fact of the matter is it's complicated.
  948. 29:35You know, at first, I mean, if you look
  949. 29:37at the story if you look at the stories,
  950. 29:39it's all anchored on two things. The
  951. 29:41first thing that it was anchored on was
  952. 29:42national security. And recently, that
  953. 29:44was all blown blown to bits.
  954. 29:46>> Right.
  955. 29:46>> And so, no no that story is no longer
  956. 29:48anchored on national security. Now it's
  957. 29:50anchored on safety. Now,
  958. 29:52if you want AI to be safe, um the first
  959. 29:55thing is we need to make sure that the
  960. 29:57the labs that are building it are in
  961. 29:59control. That there there good tests for
  962. 30:01them. Uh if we would like to have third
  963. 30:03parties uh
  964. 30:05uh to to um uh make sure that a third
  965. 30:08party evaluate third party evaluators
  966. 30:10are available. That's no different than
  967. 30:11financial control. You guys know we have
  968. 30:13auditors.
  969. 30:14And the auditors are quite quite um they
  970. 30:17don't have to be as expert as we are in
  971. 30:19our business, but they just have to ask
  972. 30:21the right questions. And um I I think I
  973. 30:23heard somebody say that it's good to
  974. 30:24have uh independent auditors or
  975. 30:27evaluators, but they just have to have
  976. 30:29multiple. I agree with that, too. Just
  977. 30:31as there's multiple evaluated and
  978. 30:33auditors, it makes sure that one company
  979. 30:35doesn't become, you know, pill or
  980. 30:37somehow influenced um for for whatever
  981. 30:40reason. And so, you you know, there's a
  982. 30:41lot of different ways that you could
  983. 30:42solve this. Um and so, that I think the
  984. 30:44number one thing is let's build the
  985. 30:46technology safely.
  986. 30:49Let's make sure that the testing of it
  987. 30:51is safe. And I I recognize completely
  988. 30:54that that what what is being built is
  989. 30:56extraordinary. Um but, these are
  990. 30:57extraordinary companies and and we ought
  991. 30:59to hold them to to uh extraordinary
  992. 31:01standards. Um and that they want to be.
  993. 31:03And they want to be that
  994. 31:04>> I wanted to go back to open source for a
  995. 31:06second. [clears throat]
  996. 31:07Um
  997. 31:08a year ago,
  998. 31:09we were taking it very seriously. It was
  999. 31:122 years, 18 months behind.
  1000. 31:13>> The one thing that, you know, one one of
  1001. 31:15the as you guys know, one of the
  1002. 31:16challenges
  1003. 31:17when you're on the call with President
  1004. 31:19Trump, it's hard to
  1005. 31:20say something.
  1006. 31:22Um
  1007. 31:23>> [laughter]
  1008. 31:24>> I'm going to get in trouble for that,
  1009. 31:25I'm sure. He's going to call me up up on
  1010. 31:26that. But, anyhow,
  1011. 31:28I what I was going to tell him and and
  1012. 31:30and all of you is that AI is creating
  1013. 31:32enormous number of jobs. The the the the
  1014. 31:35thing that he wanted more than anything
  1015. 31:36at the beginning of the the
  1016. 31:37administration and that my first phone
  1017. 31:39call with him, my first time I met him,
  1018. 31:41is that he wants to create jobs in
  1019. 31:44America. He wants to re-industrialize
  1020. 31:46the United States. He wants to make sure
  1021. 31:48that United States has the energy to
  1022. 31:49support the next industrial revolution.
  1023. 31:52Without energy, there's no industrial
  1024. 31:53growth. And so, he wants to make sure
  1025. 31:56that there's energy growth, that there's
  1026. 31:58job growth, that the re-indust- we're
  1027. 32:00re-industrializing the supply chain.
  1028. 32:02Look at everything that we're doing
  1029. 32:03right now. All of it is happening right
  1030. 32:05now as we speak. We're creating more
  1031. 32:06jobs than ever. We're creating software
  1032. 32:08jobs. Uh we were just talking about
  1033. 32:10earlier, 400 billion dollars of venture
  1034. 32:13financing went into the AI industry just
  1035. 32:16recently. 6 months. Well, that's created
  1036. 32:19a ton of jobs. That's created a ton of
  1037. 32:21jobs. Um it's created, you know,
  1038. 32:23obviously enormous amount of demand for
  1039. 32:25compute, which I'm which I'm I'm happy
  1040. 32:27about. Um which is also which is also
  1041. 32:29creating a lot of demand for data
  1042. 32:31centers. And we had to talk about that.
  1043. 32:32I think I was just I was talking to um
  1044. 32:35uh Governor Abbott uh
  1045. 32:37uh of uh Texas, and he was he was uh he
  1046. 32:40wants to appeal to the industry to make
  1047. 32:41sure that we are we are empathetic to
  1048. 32:44the small communities as we're building
  1049. 32:46data centers all over all across
  1050. 32:48America, just to be better listeners.
  1051. 32:50>> Let's actually talk about that for a
  1052. 32:51second. Let's [clears throat]
  1053. 32:53What's incredible about Nvidia, if you
  1054. 32:54if you break down the component parts is
  1055. 32:57you've effectively had to become the
  1056. 33:00bank of AI to get the ecosystem going.
  1057. 33:03And you've had to do it at all levels.
  1058. 33:05You know, you just did this thing with
  1059. 33:06Clover Leaf where you're doing Lam Power
  1060. 33:07Shell.
  1061. 33:08You did this great thing with BlackRock
  1062. 33:11and Goldman and all these folks to to
  1063. 33:13essentially create the financing
  1064. 33:14capability. Walk us through your capital
  1065. 33:17allocation strategy. Like, what has to
  1066. 33:19happen
  1067. 33:20to get a broader ecosystem folks to be
  1068. 33:23able to come in and underwrite this next
  1069. 33:24phase?
  1070. 33:25>> Well, we're we're creating As you guys
  1071. 33:27know, this is a new industrial
  1072. 33:28revolution. And and um every aspect of
  1073. 33:31it is true. Um this new industry
  1074. 33:34requires manufacturing just as just as
  1075. 33:37uh the the the the um electricity,
  1076. 33:39internet, and now AI.
  1077. 33:41We power anything.
  1078. 33:43We can
  1079. 33:45find anything. Now with AI, we can ask
  1080. 33:49and know anything. Isn't that right? And
  1081. 33:51so that's our future. We're tapping to
  1082. 33:52the ether and we can ask it of anything
  1083. 33:54we want and it could explain it to us.
  1084. 33:56Now, in order for that to happen, it's
  1085. 33:58got to produce the intelligence. And so
  1086. 34:00that's a production process, which is
  1087. 34:02the reason why this infrastructure has
  1088. 34:03to get built. But once you get the
  1089. 34:05infrastructure built, the question is um
  1090. 34:07what about all of the other layers
  1091. 34:09across the United States?
  1092. 34:10Uh this industry isn't just about the
  1093. 34:13model. It's not just about the chips.
  1094. 34:15It's mostly about the applications on
  1095. 34:18top. It's mostly about the
  1096. 34:20infrastructure layer, the data centers,
  1097. 34:22and all the infrastructure, the the the
  1098. 34:24the construction, the electricity, the
  1099. 34:26power generation. That all of that is
  1100. 34:29involved. And so I look across the
  1101. 34:31entire ecosystem and look for
  1102. 34:32bottlenecks. And if there are places
  1103. 34:34where extraordinary companies are being
  1104. 34:36built
  1105. 34:37>> Constraints.
  1106. 34:38>> Constraints. Extraordinary companies
  1107. 34:39being built. Maybe it's
  1108. 34:43supply chain that has to get scaled up
  1109. 34:46so that when we're ready to deploy
  1110. 34:49compute that they'll be ready for us.
  1111. 34:50Land power shell. And so, this is no
  1112. 34:52different than looking at the supply
  1113. 34:54chain upstream. You know, I I probably
  1114. 34:57think about the long-term supply chain
  1115. 35:00more than most because our company's
  1116. 35:01really large and and in order for us to
  1117. 35:04succeed, a whole bunch of companies has
  1118. 35:06to support me. You know, it's got to
  1119. 35:08Corning has to you know, Wendel
  1120. 35:11Corning has to support me, Lumentum and
  1121. 35:14you know, TSMC of course and memory
  1122. 35:16companies and and so, we started working
  1123. 35:18with all of these companies long before
  1124. 35:20the revolution
  1125. 35:21the growth came so that the growth could
  1126. 35:24happen. Now, I'm going to Now, I'm doing
  1127. 35:26a downstream.
  1128. 35:26>> I I have a cycle tends to be that that
  1129. 35:28the earnings over time over long
  1130. 35:30stretches of time tends to move up the
  1131. 35:32stack right towards the application
  1132. 35:34layer where you can over earn for larger
  1133. 35:36periods of time.
  1134. 35:37Um
  1135. 35:38I mean, you bought Hugging Face. Now,
  1136. 35:40you're sort of in the actively in the
  1137. 35:42serving business. I mean, it seems
  1138. 35:44pretty natural that
  1139. 35:46products like Open Router make a lot of
  1140. 35:48sense. It seems pretty
  1141. 35:50obvious that, you know, there are better
  1142. 35:51versions of ways to build things like
  1143. 35:53Bedrock. I'm sure you think about it.
  1144. 35:55What's the natural conclusion because it
  1145. 35:57seems like the folks up here have no
  1146. 35:59issue trying to move down.
  1147. 36:01>> Mhm.
  1148. 36:02>> And you have the best balance sheet,
  1149. 36:03these incredible engineers, and you have
  1150. 36:05the proven experience to make it right
  1151. 36:08and engineer the product and get it out.
  1152. 36:10So, how do you think about looking up
  1153. 36:12and saying, I could probably do that?
  1154. 36:14>> The The reason why Nvidia runs every
  1155. 36:16single model in the world we we're the
  1156. 36:19only it's incredible. Last year, about a
  1157. 36:21year and a half ago, the only thing we
  1158. 36:23ran was Open AI.
  1159. 36:24>> Yeah.
  1160. 36:25>> And now look at amazing models are
  1161. 36:27available. The Meta Muse is available.
  1162. 36:29You got Grok is available. Grok bot's
  1163. 36:31incredible.
  1164. 36:32Um we now run Gemini uh and uh Anthropic
  1165. 36:35is is uh scaling up on our platform as
  1166. 36:37well. Uh since a year and a half ago,
  1167. 36:40you got all these frontier AI models
  1168. 36:41that are now open, that are available.
  1169. 36:44So, the number of models that are are
  1170. 36:45are growing. Um there's a whole bunch of
  1171. 36:47companies that I won't mention that are
  1172. 36:49building uh frontier models as well.
  1173. 36:51And the the number of AI labs are
  1174. 36:53growing. Yeah, the the Ineffables, the
  1175. 36:57uh the Reflections, the right? The the
  1176. 37:00list goes on. The Physical Intelligence,
  1177. 37:01the list goes on. Okay? And so, all of
  1178. 37:04these labs are building on Nvidia. And
  1179. 37:06the reason for that is because
  1180. 37:08as a company, I'd rather for us to help
  1181. 37:11everybody succeed instead of taking a
  1182. 37:14slice out.
  1183. 37:16And so, we would go up as far as we need
  1184. 37:19to,
  1185. 37:20but as low as possible.
  1186. 37:22>> Huh.
  1187. 37:22>> Our strategy is go up as far as we need
  1188. 37:25to and as low as possible. And the
  1189. 37:26reason for that is because if I do that,
  1190. 37:29if I solved if it if not for Nvidia
  1191. 37:31creating cuDNN, all of the frameworks
  1192. 37:34wouldn't exist. If not for us creating
  1193. 37:36Megatron uh make Megatron core, uh then
  1194. 37:40all of the large-scale training wouldn't
  1195. 37:41have happened.
  1196. 37:42>> Training wouldn't exist.
  1197. 37:42>> Um so, we we go and we invent all the
  1198. 37:44technology necessary as far as we need
  1199. 37:46to and then we let a thousand flowers
  1200. 37:49bloom. And so, that posture allows us to
  1201. 37:52be, quite frankly, the only
  1202. 37:54>> be honest. I I agree with you.
  1203. 37:57The pushback would be that
  1204. 37:59it really would be great to have more
  1205. 38:00competition at the hyperscale layer.
  1206. 38:03And I think you've done a great job
  1207. 38:04supporting the neo clouds. There's some
  1208. 38:06I'm by the way, I think you introduced
  1209. 38:07me to Nimbix. Superb. Great. Everything
  1210. 38:09there amazing. But we need like 50 of
  1211. 38:11these guys. We need a hundred of them.
  1212. 38:13We need a thousand of them. And they
  1213. 38:15just may take
  1214. 38:16some
  1215. 38:17>> Yeah. Yeah, you know, Chamath, it's just
  1216. 38:20I'm surprisingly uncompetitive.
  1217. 38:23>> Really?
  1218. 38:24>> Yeah.
  1219. 38:25>> [laughter]
  1220. 38:25>> That's not my thing. You know, my thing
  1221. 38:27is kind of like for example, I'd be more
  1222. 38:30than happy with five hyperscalers.
  1223. 38:32However,
  1224. 38:33um the reason noticed the early
  1225. 38:36customers of all the neo clouds, all the
  1226. 38:38what we call NCPs, all the early
  1227. 38:40customers were the hyperscalers.
  1228. 38:42>> Exactly.
  1229. 38:42>> And the reason for that is because the
  1230. 38:44hyperscalers plan once a year,
  1231. 38:47but the market dynamics is so volatile
  1232. 38:49right now that they're always almost
  1233. 38:51wrong.
  1234. 38:52And so with all these regional clouds
  1235. 38:56who are agile and they can move fast, um
  1236. 38:59they know their state or they know their
  1237. 39:00country, they know their region, they're
  1238. 39:02securing land, power, and shell in a way
  1239. 39:05that's hard for somebody who sits in
  1240. 39:07Seattle or sits in Palo Alto to be able
  1241. 39:09to see the planet.
  1242. 39:10>> Mhm, right.
  1243. 39:10>> And so we now have basically a
  1244. 39:12large-scale distributed network of
  1245. 39:14companies that are building, securing
  1246. 39:16land, power, shell for us.
  1247. 39:18And [snorts] um
  1248. 39:19and and now countries realize it's
  1249. 39:21strategic.
  1250. 39:22>> Yeah.
  1251. 39:22>> So many countries are saying, "I'm going
  1252. 39:24to take my power and only give it to my
  1253. 39:26own companies."
  1254. 39:27>> Right.
  1255. 39:28>> Well, Nvidia's in that country as well
  1256. 39:30and we could help the neo clouds in that
  1257. 39:32country grow and and so whether it's
  1258. 39:34whether it's Fermus in Australia, we
  1259. 39:36just did a whole bunch of stuff in
  1260. 39:37Australia, um brought on two more
  1261. 39:39gigabytes, uh Southeast Asia of course,
  1262. 39:42uh IOH and others bring on a few
  1263. 39:45gigabytes. And so we're building a
  1264. 39:46gigawatts, so we're building, you know,
  1265. 39:48we're
  1266. 39:49we're scaling up, you know,
  1267. 39:51>> We're getting there though. I just want
  1268. 39:52to get this one thing in. It's pretty
  1269. 39:53clear that you're going pretty high up
  1270. 39:56and getting very focused on open source.
  1271. 39:58Obviously, you have your Neumotrons
  1272. 40:00doing exceptionally well. I use them
  1273. 40:02often. Hugging Face, Poolside, and
  1274. 40:04Laguna, uh very very solid product that
  1275. 40:07you're now uh aqua-hiring, hiring,
  1276. 40:09whatever it is.
  1277. 40:10Um and then you have your open source
  1278. 40:12stack for self-driving, also uh very
  1279. 40:15disruptive.
  1280. 40:16>> So we are the frontier model in five
  1281. 40:18domains.
  1282. 40:19>> Yeah.
  1283. 40:19>> And so,
  1284. 40:21you don't seem to build products to get
  1285. 40:23the silver medal. You seem to go for the
  1286. 40:25gold. So, are you going for the gold,
  1287. 40:28and will you have the best hands-down
  1288. 40:30open-source model?
  1289. 40:32And then,
  1290. 40:33part B to that is, can open source catch
  1291. 40:36up to frontier models, and are you the
  1292. 40:38person to do it?
  1293. 40:39>> So, the logic the logic, Jason, is that
  1294. 40:41that we will build it because one,
  1295. 40:45uh we can. We have the skills to do it.
  1296. 40:48And because our customers need us to do
  1297. 40:50it.
  1298. 40:51>> Right.
  1299. 40:52>> So, for example, Alpaca Myo is the
  1300. 40:54world's first thinking self-driving car.
  1301. 40:57And by by thinking, by reasoning, you
  1302. 40:59don't need as much data as
  1303. 41:01you know, you don't have to train on a
  1304. 41:02few billion
  1305. 41:04hours of road data. Because you can
  1306. 41:06reason about it. Break down the problem
  1307. 41:08into oh, I've seen this before. It's not
  1308. 41:09exactly the same, but it's largely the
  1309. 41:11same as that. Okay? And so, so Alpaca
  1310. 41:14Myo, why is it necessary? Well, there's
  1311. 41:15a whole bunch of car companies. Every
  1312. 41:17car in the world is going to be
  1313. 41:18autonomous. But beyond that,
  1314. 41:21every
  1315. 41:22ag tech, every truck, every van, and
  1316. 41:25most of them aren't big enough in scale
  1317. 41:27to be able to build that whole stack.
  1318. 41:29So, I'll build an extraordinary stack
  1319. 41:31for them. They do last mile adapting for
  1320. 41:33their application. Now, everything that
  1321. 41:36moves in the future could be autonomous.
  1322. 41:38If not for us building
  1323. 41:40some of the some of the biology models,
  1324. 41:42the world wouldn't have it. The the ESM2
  1325. 41:45protein found language model, we created
  1326. 41:47that. ESM fold, open fold, AlphaFold2.
  1327. 41:51All this stuff with with co-equivariant.
  1328. 41:53All of that stuff technology wouldn't
  1329. 41:55have existed if we didn't build it.
  1330. 41:58One of my favorites,
  1331. 41:59protein complexa,
  1332. 42:01you know, synthesizing next-generation
  1333. 42:03proteins and its binding. It's
  1334. 42:05groundbreaking stuff. We built that. And
  1335. 42:07so, we'll build that because a Lilly
  1336. 42:08needs it, and and you know, Merck needs
  1337. 42:11it, and others need it, and they don't
  1338. 42:12have the capability to do it or they
  1339. 42:14they're not yet there. And so we can
  1340. 42:15make a real contribution. So I do
  1341. 42:17everything out of need. I'm not trying
  1342. 42:19to disrupt. I mean, we don't wake up in
  1343. 42:21the morning try to disrupt anybody. We
  1344. 42:24just wake up in the morning try to help
  1345. 42:25everybody.
  1346. 42:25>> Jensen, what about what about
  1347. 42:27competitive threats that might be
  1348. 42:28emerging to your core business? Can you
  1349. 42:30just comment
  1350. 42:31>> so nice.
  1351. 42:32>> Yes. Well, I know this is well, I
  1352. 42:34actually want to I want to just get your
  1353. 42:36let's just call it a take. What's your
  1354. 42:38take on Terafab 100 million square foot
  1355. 42:40facility Elon's announced? And
  1356. 42:43>> If anybody could do it, he can. And the
  1357. 42:45two of us were on a flight together to a
  1358. 42:48country.
  1359. 42:49And um
  1360. 42:51>> With a person
  1361. 42:54who sometimes calls you on the phone.
  1362. 42:56>> It was a It was a nice plane.
  1363. 42:57And and
  1364. 42:59And we had like, you know,
  1365. 43:01and
  1366. 43:02you know, Elon likes to talk about these
  1367. 43:03things and and so uh we spent a lot of
  1368. 43:06time talking about it.
  1369. 43:08I I
  1370. 43:08>> Is that
  1371. 43:08>> If anybody could do it, he could do it.
  1372. 43:10>> Cuz I mean, you could you design chips,
  1373. 43:12you don't fab them.
  1374. 43:13Could your chips be fab there or is it
  1375. 43:15>> Well, we know we have we know a lot
  1376. 43:17about process technology because we're
  1377. 43:18pushing the limits of everything.
  1378. 43:20>> Right.
  1379. 43:20>> And you know, [clears throat] because we
  1380. 43:22scale at such large scale,
  1381. 43:24uh we have incredible memory technology
  1382. 43:26inside the company. We're the world's
  1383. 43:27best 30s company. You know, we got lots
  1384. 43:29of amazing tech.
  1385. 43:30>> So your take is you've talked a lot
  1386. 43:31about it.
  1387. 43:32>> So we could just Yeah, we could talk
  1388. 43:33about it and and um you can't discourage
  1389. 43:36Elon from doing it, which is one of his
  1390. 43:38incredible That's his superpower. And
  1391. 43:40once he decides to go do something, it's
  1392. 43:42hard to stop him. And so I
  1393. 43:43>> Can can you give us your take on where
  1394. 43:44China is with advanced lithography
  1395. 43:46systems
  1396. 43:47um native grown?
  1397. 43:48>> They're going to get there by 2030.
  1398. 43:50>> By 2030? [clears throat]
  1399. 43:51>> Yeah. And 2030 is just around the
  1400. 43:53corner.
  1401. 43:53>> Yeah. Also, that's Sorry, how long will
  1402. 43:55we'll be dead at that time? So
  1403. 43:58And does And And for China, does that
  1404. 43:59mean the switch is flipped and then
  1405. 44:01that's all going to go into um
  1406. 44:03mainland fabs?
  1407. 44:05Almost immediately.
  1408. 44:06>> You know, the the way to think about
  1409. 44:07China is they're
  1410. 44:08good at high volume production.
  1411. 44:12And this is just a matter of time.
  1412. 44:15>> Yeah.
  1413. 44:16>> And I you know, I I think in
  1414. 44:18I I think in decades as well. You know,
  1415. 44:21I've been around a long time and
  1416. 44:23you know, for Nvidia I got to think
  1417. 44:25about what happens next decade and
  1418. 44:26decade after that. So, two or three
  1419. 44:28years is is just a click. It's nothing.
  1420. 44:31And so, as far as they're concerned,
  1421. 44:32they're already there.
  1422. 44:33>> They're already there.
  1423. 44:34>> Yeah.
  1424. 44:34>> Jason, Elon and Gwynne.
  1425. 44:37>> We've got to run.
  1426. 44:39America, we got to run.
  1427. 44:40>> Yeah.
  1428. 44:40>> Speed run. We got to speed run. Slowing
  1429. 44:43down is definitely the wrong strategy.
  1430. 44:45>> Well, I mean
  1431. 44:47it [clears throat] it feels apparent I
  1432. 44:48think to most of us in the industry that
  1433. 44:50we're kind of in the AGI moment and it's
  1434. 44:53a definition obviously, as smart as any
  1435. 44:55other human.
  1436. 44:55>> we're already there.
  1437. 44:56>> We're there, right? And so, then super
  1438. 44:58intelligence is the next waypoint. Based
  1439. 45:01on what you see, based on your customer
  1440. 45:02base, based on your history here.
  1441. 45:04>> And Jason, I think we're there, too.
  1442. 45:05>> You think we're at super intelligence?
  1443. 45:07>> When you when you when you take a narrow
  1444. 45:09segment,
  1445. 45:11a narrow segment, I mean my my
  1446. 45:13self-driving car, I don't want you to
  1447. 45:14make me an omelet. I just want you to
  1448. 45:16drive the car.
  1449. 45:17>> Right.
  1450. 45:17>> That is super intelligence.
  1451. 45:19>> It's better it's better than a human.
  1452. 45:20>> Yeah, yeah.
  1453. 45:21>> Yeah. Like 1/10 the the accident rate.
  1454. 45:23>> Exactly.
  1455. 45:24Uh synthesizing proteins, you know, uh
  1456. 45:27doing virtual screening of proteins.
  1457. 45:29We're already there.
  1458. 45:30>> Yeah.
  1459. 45:30>> Yeah, yeah.
  1460. 45:31>> Are you having fun being on the frontier
  1461. 45:32of humanity?
  1462. 45:36>> I like it.
  1463. 45:37>> Yeah.
  1464. 45:37>> [laughter]
  1465. 45:40>> Ladies and gentlemen,
  1466. 45:41ladies and
  1467. 45:42>> I like it. I like it. And then guys,
  1468. 45:44guys, it's it's it's great there.
  1469. 45:47The future is great and we want to get
  1470. 45:49there. The listen,
  1471. 45:51>> Ride the wave.
  1472. 45:51>> a lot of us don't have to work, but I
  1473. 45:54got to tell you, it's too good not to
  1474. 45:56be.
  1475. 45:56>> So fun.
  1476. 45:56>> Right? And so, so I want every we I want
  1477. 45:59to be there. I want all of you guys
  1478. 46:01there with me. We're all going to be
  1479. 46:03there. We're going to be enormously
  1480. 46:05successful together as humanity. And and
  1481. 46:08in the meantime,
  1482. 46:10we got to encourage them, urge them on.
  1483. 46:12They're doing really, really important
  1484. 46:13work as you guys know. And I want them
  1485. 46:15to succeed. I also would love for us to
  1486. 46:19tone down the the the drama. And and
  1487. 46:22most importantly, we need all of America
  1488. 46:25to come with us. That's how we make it.
  1489. 46:27>> Ladies and gentlemen, Jensen Huang.
  1490. 46:30>> [applause]
  1491. 46:31>> Thanks, [music] man.
  1492. 46:32Appreciate you. Thank you.
  1493. 46:35>> Thank you.
  1494. 46:35>> Thank you. Thank you.
  1495. 46:36>> That was great.
  1496. 46:36>> awesome.
  1497. 46:37>> [music]
  1498. 46:38>> Following your party.
  1499. 46:39>> awesome.
  1500. 46:39>> That was great.
  1501. 46:41>> Thanks, guys.
  1502. 46:42>> Thank you.
  1503. 46:43That was great, huh?

About this transcript

This page contains the full transcript of Jensen Huang: The Doomer Hoax, Superintelligence is Here, and The Future of AI (ft. President Trump) by All-In Podcast, generated from the public captions YouTube serves with the video. The transcript has 9,096 words across 1,503 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.