Jensen Huang of Nvidia on the Future of A.I. | DealBook Summit 2023

[Music] [Music] please welcome Andrew Ross sorin and his guest the founder and CEO of Nvidia Jensen [Music] Hong welcome back everybody Jensen is here of course the CEO of Nvidia as I mentioned at the top of the day uh this is the clear winner uh of every winner in the world of artificial intelligence uh thus far his company Powers everything from open AI uh Google's programs meta they're all Frenemies in some ways well talk about it uh he founded the company back in 1993 uh over breakfast uh at Denny's with two friends since then as CEO uh he's LED Nvidia to become the world's uh most valuable Semiconductor Company nvidia's stock has been on a tear up 240% this year reaching a trillion doll uh market cap and we are so grateful uh to have you here today as we all try to make sense of what is happening in the world of AI and I think in so many ways you saw this uh first and so I'm hoping to start with this and I I said you you power uh what open Ai and chat GPT has been we've all been reading about uh open Ai and all of the travails um inside that that company and nonprofit and we maybe talk about some of the governance issues there as well but you delivered I think this is this is back um I don't know what year we're talking about now but you delivered the first box the first chips to Elon Musk who was one of the founders of open AI only a couple of years ago what did you what what happened well I delivered to him uh the first AI supercomputer the world ever made it took us 5 years to make it it's called a dgx it's everywhere in the world today um people think that we build gpus uh but this GPU is is a 70 lb 35,000 Parts out of the 35,000 eight of those chips come from tsmc uh it is uh so heavy you need robots to build it it's like an electric car it consumes 10,000 amps uh we sell it for $250,000 it um uh it's a supercomputer so it takes another supercomputer to test it this this is a a computer first of its kind and we started working on it in 2012 uh it took me 5 years to build it at first I built it for our own engineers and I spoke about it at one of our conferences and Elon saw it he goes I want one of those and and he said he he told me about open AI uh I don't also knew uh Peter Beal uh who was a a Berkeley Professor he was one of the early people at open Ai and ilas Suk uh he I met him during the Alex net days 5 years earlier he's involved in all the drama that we've been reading about and so anyways I I delivered the world's first AI supercomputer open AI on that day and and uh people took pictures of it it's on the internet somewhere yeah um when you did that and you said you didn't do it originally for him what was it though that you saw at that point five years before you even delivered it in 2012 when when this all first started first happened Alex net did something remarkable uh here's a here's a um a neural network it's a it's a it's a software program where the way you programmed it was to show show it the results that you wanted which is the backwards of most programs up to then you know programs up to then were where Engineers would sit down and you would write software and then you would test it to see if it produced the outputs you wanted but here you showed it examples and you you um uh you taught it uh what outputs you wanted what outputs to expect and so when we first saw the results of it alexnet the results were so spectacular that Alex uh um keski and and ilas suser and of course Jeffrey Hinton they achieved results that of computer vision recognition that that no no computer vision expert were was able to achieve before that and so so the first the first observation was is how remarkable it was but then then we were we were U fortunate to have taken a step back and asked ourselves what is the implication of this uh to the future of of computers and and we drew the right conclusions that that this was going to change the way Computing was going to be done this was going to change the way software was going to be written and this was going to change the type of applications we could write did you get to was there any part of you that was scared when all this happened you just mentioned two names George hint Hinton as well uh you also mentioned Elia and and those are names by the way if you've been following what's happening they have been very outspoken about the dangers of AI very I want to get into to actually what you think happened at open AI in the past couple weeks uh but it may very well be that uh there may have been a new Step change in terms of what this technology was but was there ever part of you when you're seeing this all happen say oh my goodness not only we on the cusp of a revolution in a great way but that this is dangerous well no I I would say I would say 12 years AG nobody expected the results what we get and I think anybody who who who would have would have uh uh said so back then uh would have overe exagger at you know our understanding of of the of the the rate of progress um there's no question that the rate of progress is is high and what what we what we realize today is that that of course of course uh uh what we can do today with with uh these models and intelligence are related but not the same you know we're very good at perception today and we're very good at those oneshot knee-jerk reaction uh I recognize that that's a dog um I can I can finish that sentence uh but there's a whole bunch of things that we can't do yet we can't reason yet you know this multi-step reasoning that humans are very good at uh AI can't do and how far away do you think we are from that well we'll see uh we'll see I think that that um just about everybody's working on it and and all the researchers are working on it everybody's working on it we're trying to figure out you know how do you take a goal break it down into a whole bunch of steps and create a decision tree and then walk down the decision tree to figure out you know which one of the paths leads to the most optimal answer uh this is this is uh how we reason through things how we iterate through a problem uh today as you know but you're making bets now in terms of of technology that you have to build and Investments you have to make yeah on where we're going to be 5 years from now 10 years from now 10 years from now right so you know people talk about AGI yeah right artificial general intelligence do you think at 10 years from now we are there uh by depending on how you define it I think the answer is yes and so the question is what is Agi um a if we Define AGI as uh a piece of software a computer that can take a whole bunch of tests and these tests uh reflect uh basic intelligence um and uh uh and by achieving by by by completing those tests uh deliver results that are uh fairly fairly competitive to a normal human um I would I would would say that within the next 5 years you're going to see um obviously uh AIS that can that can achieve those tests and design the chips that you're making right now yeah yeah will you need to have the same staff that designs them uh in fact uh none of our chips are possible today without AI literally the h100s we're shipping today was designed with the assistance of a whole lot of AIS otherwise we wouldn't be able to cram so many transistors on a chip or optimize the algorithms to the level that we have and you know software be written without AI chips can't be designed without AI nothing could yeah nothing is possible we started by talking about open Ai and and everybody's focused on that what did you make of what happened the ousting of of Sam Alman the return of Sam Alman the all of it yeah well first of all I'm I'm happy that that they're settled and I hope they're settled uh it's a really great team and and they're doing important work and they've achieved great results and I'm just really happy that they're settled um you know it also it also bring brings to mind the importance of corporate governance I mean there's a you know Nvidia here 30 years after our founding uh we've gone through a lot of adversity uh if we didn't set up our company properly uh who knows what would have been who knows what would have done and so I think when you're when you're architecting an industry you know you want to apply some of that some of that wisdom to architecting a company and and so I'm I'm really proud of nvidia's corporate governance by the way and and uh uh if not for the architecture that we we established when I was 29 years old it'd be kind of you're a for-profit company though what's so interesting I think about this sort of dynamic is that that is a a firm that is effectively operated from a governance perspective as a not for-profit and one of the reasons that they set it up that way was because they did think it was dangerous Elon Musk said it was dangerous at the beginning Elia said it was dangerous and so the question is in the sort of multitude of these different um businesses that are in AI do you think you do need these not for-profits do you think that that the incentive system is just fundamentally off and should be a for-profit I mean a lot of people now think the capitalists have taken over well Regulators are not are not for profit and we should regulate these uh first of all just take a step back and think about what AI is AI is an autonomous system uh it's an autonomous system that's more sop it's an autonomous information system we have a a lot of autonomous systems today uh self-driving cars uh in in in factories within factories already exist robots are autonomous uh uh planes are autonomous um autopilot self-landing um all of those capabilities exist uh we we ought to make sure that we apply the first principles of autonomous systems in the same way uh we have to design it properly test it properly stress test it properly monitor it there's Inside Out safety there's outside in safety um the FAA uh flight uh um Air Traffic Control uh redundancy diversity um there's a whole bunch of different systems that we have to put in place for autonomous systems there's a lot of Industries to learn from um at the beginning of this I mentioned there's sort of a frery situation going on uh with a lot of companies that use your chips they're desperate for your chips they they they they they want your gpus and at the same time they're also uh trying to build their own frankly I'm curious since you've seen it all how you would stack rank the success of the various companies that are in this AI space uh we have somebody from Google deep Minds here uh today their coo I'm curious where you think open AI ranks in that there's inflection uh Amazon is trying to play in I'm not going to rank my friends no but but you but you have a sense of and and I part of the question Adit I want to but I'm not going to do it I'm just kidding but but there is a question about actually whether all of these things converge uhhuh meaning that that they all that just this all become some kind of commoditized no business no I don't think so I don't think so I I think what's going to happen is we're going to have we're going to have um off-the-shelf AIS and these off-the-shelf AIS are going to be really really good at at at um solving a lot of problems um but but you're going to have uh uh uh companies in healthcare are going to have supervised you know super tuned AIS that take these off the shelf AIS and make them super good at drug Discovery or super good at ship design and we just use our our company for example the vast majority of our company's value is in the data and the intelligence and the the knowhow the craft that's inside our company and none of that data is out on the internet you can't get an AI to go learn it and so I've got to take a really smart AI which is what we do we build a smart Ai and then we teach it how to design chips we teach it how to write software you'll teach it how to you know do drug Discovery you'll teach it how to do you know Radiology let me ask you a geopolitical question we're going to hear from the uh President of Taiwan uh just after this and there is a big debate as you know about chip Independence uh the big investment that we're making in chips to manufacture here in the United States um whether we should be exporting certain types of chips uh to China where are we on the Journey of being chip independent if you will and do you think that that is a worthy goal we are we are somewhere between a decade to two decades away from from Supply chain Independence um as I mentioned earlier our systems comes 35,000 parts and eight of them come from tsmc and the supply chain when you think through it they're in Taiwan of course there a lot in Taiwan they're all over the world but supply chain Independence is going to be really challenging you know we should try it we should Endeavor it I mean we we should absolutely go down the Journey of it but total independence of supply chain is is not a real practical thing for for a decade or two okay one of the other things that's happening as you know so well is that the US government has effectively told you you need to throttle uh the speed of the chips yeah uh that you are exporting to China yeah this is having an impact on the business itself but I'm curious how you think about that also uh geopolitically as a business the National Security concerns Jamie Diamond we were talking earlier uh about you know what companies you should do business with should you do business with people in China or not given all of the concerns that people have well on first principles we're a company that was built for business and so we try to do business with everybody we can um on the other hand on the other hand our national security matters and our national competitive competitiveness matters um somewhere between between the the between that makes sense and so uh our country of course wants our industry to to on the one hand be successful lead the world uh invent amazing technology have uh techn ology Independence on the one hand and and be the leader of the world in technology on the one hand on the other hand we need to make sure that we ensure National Security our regulations provide for that uh the most critical technology that we build uh the Leading Edge of it uh is not made available to China and so what our what we have to do a new regulation just came out uh one that came out a year ago one just came out this year and so we have to we have to come up with new chips that comply with the regulation um and once we comply with the regulation we'll go back to Market and and do the best do you think the regulation is a good idea cuz I have I have heard you say that you think potentially by throttling these chips we are just inspiring and creating competitors in places like China that you can't control that's what you know look there are always unintended consequences everything that we do uh in complicated systems if we want to want to limit them from access to technology like nvidias um maybe it doesn't really they find a way to get it or they find a way to uh Inspire their local industry there's some 50 companies are being built in China that that uh are going to go provide this technology so we you know it's it's a it's a complicated thing and so what can you do um well you can you can make your own choices but the the other thing that's happened literally in the past U couple months now is Huawei came out with a new phone yeah um and it surprised everybody in terms of the chips in phone in terms of uh being a 7 nanometer chip there was a view that China was never going to get there we were we had this sort of real uh uh real opportunity ahead of them by by by many years were you surprised by that uh the the rumors of it in in in the market has been around for a long time and so was it were we surprised I don't think so I don't think anybody in the industry was really surprised and and um is is it possible to take something that that said 16 nmet and shrink it to seven you know these are just numbers uh is it really seven um did they shrink it down to something that was sufficiently good that you can make a Pham from yeah I think so and and so so I think it you know there's no magic in these numbers as you know it's just seven the numberb but the question is what is our lead over them do you think in semiconductors yeah in semiconductors you know call it call it a decade um uh you know you could decide yeah call it a decade but um uh could you take the decade old technology and just squeeze The Living Daylights out of it until it produces something that's kind of like something from five years ago yeah probably and so so I think there's a lot of you know a lot of a lot of clever Engineers all over the world and they're trying to you know get the most out of let ask you a different compan they have there's a company called asml uh in the Netherlands uh that's basically responsible for every chip that everybody makes um some people might call them a monopoly how powerful are they in all of this and should we be worried about that power um well a lot of people depend on them to build the instrument and they do build very very good instruments and um uh the technology is very complicated it took a long time for them to build it uh there there's no reason why they they don't want to provide it to the world and so I I'm not so I'm not sure what what the question is but but I'm not concerned I didn't wake up this morning concerned about asml I think they're they're an excellent provider and they're they're they're motivated to supply to us and you know so I I think everybody's everybody's incentives are aligned um I want to ask you a management question um because it's just fascinating given the success of this of this company uh you constantly say even at this point in the ball game uh you say I do everything I can not to go out of business I do everything I can not to fail that that is like a mantra inside the company even at this point what is that about uh what is that about I I think I think when you when you build a company from the ground up and you've you experience real real adversity and uh uh and you really really experience nearly going out of business several times uh that that feeling stays with you um I wake up every morning in in in you know some condition of concern and and uh I I don't I don't wake up proud and confident I wake up worried and concerned about you know and so so it just depends on which side of the bed you get out on this is the Andy Grove only the paranoid survive well I think paranoia needs needs therapy I I don't I don't think I don't think people are trying to put me out of business I probably know they're trying to and so so I I that's different and so so I I live in this condition where where um uh partly partly you know partly desperate partly you know partly partly uh aspirational and okay well let me ask you then about this uh you said this to the New Yorker and um I found it fascinating again goes to this idea of failure or or worries about failure but you said this and I this is like news you this is a selfish question um you said I find that I think best when I under adversity and then you said my heart rate actually goes down when I'm under adversity my heart rate goes up by a lot uhhuh well my let's see well I I think I think um I you during adversity you're more focused and when you're more focused you could you perform better and I like I like you know the last last last uh 5 minutes before uh before something you're you're more focused and so you know I I like to live in that state where where uh uh we're about to perish and everything every you know and so so I I enjoy that condition and and uh I do my best work in that condition and I you know I like going home and telling my wife I saved the company today and and uh may maybe maybe it wasn't true but but I like to think so and so um another question we have a lot of Business Leaders and CEOs here and and I think they're going to be surprised to hear this you have 40 direct reports at the comp 50 50 direct reports 50 direct Reports most people say I don't know if we have any consultants in the room they'd say you know what half a dozen maybe 10 that should be the limit what's your what's your philosophy or Theory here well the people that report to the CEO should require the least amount of pampering and so I don't think they need life advice I don't think they need career guidance um they should be at the top of their game incredibly good at their craft uh um and unless they need my personal help uh you know they should require very little management and so so I think that one uh the more the more direct reports the CEO has the less layers are in the company and so so uh it allows us to keep information fluid it allows us to uh make sure that that everyone is imp EMP powered by information uh and uh our company that you know just performs better because everybody's aligned everybody's informed of what's going on uh I want to open up some questions in just a moment uh so please do raise your hand so I can I can find you uh but but I want to ask you this uh you did a podcast recently um and there were a lot of headlines about it and you said during the podcast if you could do it all over again meaning like if you could start Nvidia again yeah you wouldn't no what do you what did you mean why I mean you've done this amazing thing you're worth $40 billion per that wasn't what I meant first of all I you know I I think it would be disingenuous if I said that that um uh it wasn't quote worth it um I I I enjoy a lot a lot of good things in life I've got a great family U we build a great company uh it all of that is worth it that's that wasn't what I meant what I was if if people realized how hard something is and if I would have realized how hard it was how many times we're going to fail how the original business plan had no hope of succeeding that that almost the the early Founders that we built the whole company with um we we had to completely relearn just about everything we had to know if if I would have known all of the things that that I had to know in order to be a CEO uh everything that we had to solve in order to be where we are that mountain of work that mountain of you know challenges that mountain of adversity and setback and some amount of humiliation and a lot of embarrassment if you would have M piled all of that on in 1993 in you know on the table of a 29y old I I don't think I would have done it I would you know I would have said there's no way I would know all this there's no way I could learn all this there's no way we can overcome all this there's no way you know this is a game plan that that's not going to work and so that that's what I meant that that I think I think the ignorance of entrepreneurs the this attitude that and I try to to keep that today which is ask yourself how hard could it be you know you approach life with this attitude of how hard could it be if they could do it I could do it that attitude is completely helpful but it's also completely wrong it's very helpful because it give you courage but it's wrong because it is way hard harder than you think it is and and uh uh the amount of skill that is necessary the amount of knowledge that's that's necessary uh you know I think it's it it's one of those uh uh teenager attitudes and and I think I think we I try to keep that in the company that teenage attitude how hard can something can can something be um you know it gives you courage it gives you confidence U let's try to sneak in uh one question or two if we could I know I I Ron Conway had a question last time at a at a different moment I don't know if he's uh still in the room I felt like I should should give him an opportunity but I I see Gary Lauder there hey Gary so there are a lot of startups and not non-st startups doing AI chips op optimized for llms uh can you talk about uh and they claim to be dramatically more effective and energy efficient than gpus can you talk about what you're planning in these regards yeah um f first of all this is one of the great observations that we made you know we realized that that um deep learning in AI was was not a chip problem it's a reinvention of computing problem everything from how the computer works how computer software works the type of software that it was going to write the way that we write it the the way we develop software today using AI creating AI that method of software is fundamentally different than the the way we did it before so every aspect of computing is is changed and in fact one of the things that that people don't realize is the vast majority of computing today is a retrieval model meaning just all you have to do ask yourself what happens when you touch your phone um some of like you know some electrons go to a data center somewhere retrieves a file and brings it back to you in the future the vast majority of computing is going to be retrieval plus generation and so the way that Computing is done is fundament mentally changed now we we observed that and realized that about a decade and a half ago I think a lot of people are still trying to sort that out it is the reason why you know people say we're we're practically the only company doing it it's probably because we're the only company that got it and people are still trying to get it you can't you can't solve this new way of doing Computing by just designing a chip every aspect of the computer has fundamentally changed and so everything from networking to the switching to the way that computers are designed to the chips itself um all of the software that sits on top of it and the methodology that pulls it all together it it's um it's a big deal because it's a complete reinvention of the computer industry and now we have a trillion dollars worth of data centers in the world all of that's going to get retooled that's the amazing thing we've got we're in the beginning of a brand new generation of computing it hasn't been reinvented in 60 years this is this is why it's such a big deal it's it's hard for people to wrap their head around it um but that's that's the um that was the great observation that we made is it includes a chip but it's not about that chip Jensen Wong everybody thank you very very much thanks everybody

2023-12-08 12:25

Other news