Recursion x NVIDIA event at JPM2024 — Fireside Chat with Jensen Huang & Martin Chavez

[Applause] [Music] well Jensen thank you for joining us I uh count myself among your Fanboys um just GNA share a story so I met Jensen in 2017 uh I was CFO of a bank as it turns out and for whatever reason this bank was bringing the board of directors by Nvidia and I don't know if you even remember Jensen but in the room there were these huge monitors usually when there's a there are Bankers in a boardroom during a board meeting something bad has happened was I don't remember 2017 2017 no it was just a it was a demo more than anything else you had these big monitors up in the room and I no noticed that there was a very beautiful person on the Monitor and then about 30 seconds later there was another beautiful person and it just kept going like this and then at the end you said oh none of those people actually exist they're being generated by Ai and for me it had a big effect because I was so passionate about AI I went to an AI in medicine program in the '90s and then really gave up because AI was going nowhere and that's how I ended up on Wall Street and was there for a long time and then suddenly in your I really hope it's different this time and then in your office I I saw wow things have really progressed right since that time and uh so I had gotten out of healthc care and now I'm coming back to my roots and I think you at about the same time started Nvidia that I went to Wall Street when did you start 93 93 and so what brought you back to to healthcare can you talk about that sure um f first of all 2017 I I think it's really really instructive to just reflect quickly on that that moment but in the context of what happened starting in 2012 uh you guys everybody knows is well documented uh Alex Alex's uh model Alex net and his partner ilio susc and uh their their adviser Hinton uh came up with Alex net and it was able to achieve computer vision capabilities uh with with uh uh no specific human engineered algorithms uh for object recognition computer vision and recognizing objects is a is a first step in in some capability of intelligence and so it's really important in in artificial intelligence perception perception in general but 2012 uh happened uh shortly after we achieved um uh superhuman speech recognition superhuman uh uh object recognition and then and then um then went kind of quiet you guys probably uh this is kind of 2015 is it went kind of quiet 2018 comes along uh you saw the first version of Nvidia work in generative ganss uh Ian Ian Goodfellow invented Gans uh but we really took it we elevated to a great new level we we were able to control Gans uh meaning that we could say generate just nothing but human faces or we could generate um uh mountains lakes uh oceans trees clouds compose them all together into a scene uh so we we worked on uh what was what what is the first first indications of generative Ai and that came out in this model and it got triggered a lot of excitement it's called Gan shortly after that Transformers was invented uh Bert was uh created and then GPT came along in 2018 so so 2018 if you go back in time 2018 is as big of a deal as 2012 about 6 years time about 6 years time about five year five six years time well uh it's now four five six years uh after that M and we've now uh unified unified uh the ability to uh both uh understand the language recognize patterns of very very long sequences and very large uh dimensionality and uh uh understand and learn the representations of them learn a language of just about all kinds of information and because we understand all the information uh uh we can also uh generate information from that so we can now translate from text to text text to image image to text image to text would be captioning uh text image would be image Generation Um uh text to proteins would be properties to proteins um amino acids uh to protein would be uh structur generation so on so forth and so so if you just generalized what we have now achieved generally just just take a step back and says what what is a computer able to do now it can now Rec it can now recognize and learn the language of almost anything with structure and it can translate it to anything with structure and so text protein protein text amino acid protein so on so forth and so this is where we are now this is the generative AI Revolution and um I and so so the the time I'm I'm just super excited to be here first of all this is not my normal crowd you know biologists and scientists and you use things like you you it's such an Angry Crowd first can so can I first can I first just like usually when we think about things we think about you know creation and and we we would like to be able to improve and accelerate or um you you you use words like Target um inhibit you know the list of words that you guys use are just generally angry and and so anyways anyways you're not my normal crowd my normal crowd the creators designers they're artists you know they're they're people who are just generally happier but you're really on to something here was that me I think I think I I don't think so I you you guys are just in it so of you guys in it so much you just angry people with other angry people just before you know it you don't even realize you're angry you're doing it for a good cause but yeah you you're you're doing it in a really angry way and so so anyways I I just you said you said that moment I just just wanted to reflect reflect on it for all of us uh we are at that moment in computer science we are at that moment in information science let me give you one example of just the the the the the magic of it okay this is just just the extraordinary part of it let's let's say video conferencing let's apply some random example video conferencing in the past we would take a camera we would encode we would literally T every single Pixel uh we would find the entropy in it and we would encode it and we would send the encoded video on the other side we would decode it and the data rate the data rate is a few megabits per second is that right yeah few megabits per second it could be 10 megabits per second or something like that and and the compression ratio is incredible it's like 50 to1 we're so we're Amazed by it video video conferencing uh it was invented by AT&T in uh 1964 1964 is it turns out to a very important year in the world a year after I was born it was the year that the world that uh invented the IBM system system 360 uh computer science has largely remained the same since then yeah AT&T came out with the video conferencing system and it's largely been the same since let's pretend for a second that video conferencing system now looks like this so I go home and my wife says you know Lor goes how's your day uh and uh uh did you did you uh add value today that's first two questions okay first two questions first two questions does you add value today what's your hit rate and largely largely what I say fine I did no harm I that's 99% of the time sometimes I say I had you know I saved the company and and so and sometimes it's true sometimes it's true for example Tuesday um and and and sometimes I exaggerate I Sav the company but that's not true but if it made her felt good Po and so she yeah oh she will say thank goodness feed the dog yeah and so and so she'll say something like what happened today and I'll I'll say I'll say like I gave I gave a talk uh Martin and I gave a fireside chat there were a few hundred people in the room and um it was raining uh and and uh uh uh it was leaking um and and notice in just a few words the compression Rio Is kilobytes but she had reimagined that entire scene MH in the future conference is going to be like this it's going to take a picture that picture is going to Pi picture of our face and then after that it will it's it's perceived it after that it will reanimate it on the other side reimagine it on the other side the compression ratio is going to be like a million to One MH we are going to use artificial intelligence to exceed this limit in information Theory called Shannon's Theory the Shannon's limit of information Theory will be exceeded now how is it possible we exceeded information because we have priors we have priors we recognize what a person looks like yeah we recognize what happens we generally understand what happens when they're animating their face We Could reconstruct it on the other side and so so now the inverted problem for all of you all of you scientists in the room is in the absence of information how do we go find information in the absence noisy data in the absence of such low data rate how do we find Insight how do we find data how do we find the the embedded uh data that we need the information that we need well the reason for that the reason you're able to exceed what your common sense says is because there are so many other modalities of information we call them priors and so M multimodality I heard multimodality probably more often in this this today today's in several of my meetings than just about in any circumstance and and and I'm just super excited for all of you multimodality language models Foundation models are clearly doable today we know how to do it um uh it embeds a lot of priors uh and uh uh and and the auxiliary information that you're going to bring from all these different modalities is going to give you inside like you can't believe and so it just applied this this concept of video conferencing I just described uh and in fact today uh you could use uh an AI model from Nvidia it's called audio to face you apply it to a picture uh we call it live portrait and literally from the voice that is being transmitted and voice encoding is very high and the amount of data is you know very low and from the voice and the words that are spoken we can animate a face and it looks just perfect it looks just perfect and so you can have a video conference with not just you know a few kilobytes wow so you can go you could you could transmit enormous amount ounts of information with just very little amounts of data and we're in that era oh your question was was uh uh biology did you how did we get involved this was 19 uh we invented this Computing model called accelerated Computing and um there were several there were several uh events that and we invented this thing and I had no idea whether it was going to be useful mhm and this is kind of the the usually you have an intuition about why it's going to be useful but you're not exactly sure so we invented it two events happened that gave me a great deal of confidence two things and they were both in biology and so uh two uh two researchers at Mass General uh saw the work that we did and they applied it to uh CT reconstruction inverse physics inverse inverse inverse uh image processing basically and CT reconstruction and I saw their paper uh and I I try to you know even though I I probably have better things to do but but I enjoy I enjoyed reading some of the papers that that are published and and I noticed that they used our our our gpus which is designed for playing video games for uh for CT reconstruction I was so excited by that and uh uh um uh we flew out to go see them and and I kind of remember them saying you know why are you here and I said well I want to see what you guys are doing with these gpus and how are you using it for inverse you know CT reconstruction he goes and said they said uh we've never had a chip exec come and see us before and I just want you to know that if your trip was incredibly successful we would buy two gpus and I said I said well that's totally worth it that was that was the first and the second time the second time was around this is this the the when in when Cuda was invented uh was uh was a uiu where they used um uh our gpus for molecular Dynamics okay and and it g it it opened my mind towards hey look we could apply the same methodology that we use in designing chips computer AED chip design we might be able to help the world of drug Discovery go from drug Discovery computer AED drug Discovery to computer AED drug design and this is now about I guess about 15 years ago and when I saw Nami and you all of your molecular Dynamic simulation I said hey if we scale this up by a billion times we could like simulate biology and and I was so enthusiastic about it and I I said how hard could this be it turns out it's way harder than I thought but it got me into the journey how many years of mors law is that about 30 years well in the in that time in that time we've Advanced Computing Now by uh a billion times yeah yeah 15 years a billion times and we're still a billion times away and so but but the thing is the fact that we're only a billion times away says it's close right it's very close and and this is a big deal this a very big deal but but I think that the maybe the the the important message is is um uh it didn't take much to get get me excited about it uh because we had I think it was Aviv that she was saying something about you saw success somewhere else and you like you like the smell of that success and you want to you want to apply it to something something that you're working on I had the benefit of growing up in the in one computer Revolution that my entire career is defined by electronic design Automation and we we made my generation before my generation my boss spend all of his time his name is Mark Allen uh and and he uh he spent all of his time in a lab I was the first generation that spent less time in a lab our Engineers today never goes to the lab their entire reality is in is completely in simulation and we build giant comp incredible systems you guys uh it takes 20,000 engineering years to build one generation of our systems and when that system they tell me it's we call a tape out when that system is ready for tape out when I press that button it is a first of all $500 million button wow and and I launch I kick off5 billion dollar worth of engineering subsequent engineering and I press it and I know it's perfect I know it's perfect it better be perfect because otherwise and so I know it's perfect I know it's perfect because if it's not we're in trouble and and and so we we literally prefetch all of the experimentation all the complex experimentation of the future and bring it all into en silico now we couldn't have done that in in one step but it took us 40 years to get here I have every confidence this is going to happen here now it's 40 years after our journey 40 years after our journey which is basically about a trillion times more Computing capability necessary to do what you need to do but over the course of the next 40 years you know I think over the course of the next 10 years frankly um almost everything will largely start in in silico largely end in in silico and and um and I'm hoping that computer AED drug design will be the way we talk about it so so Jensen on that point of course I share that view but questions very much on my mind I think many of the people here I'll just say there's two extremes and these are caricatures but one extreme is we've got a lot of data about human biology it might be in disperate places and non-standard forms hard to Corral but there's a ton of data we have all the data we need and now we just need really great algorithms and llms and they're going to train on all that data and then we're going to learn all of human biology that's one extreme The Other Extreme is we don't have anywhere near the right amount of curated reliable reproducible data and we've got to do Millions maybe billions of experiments highly reproducible to get that data and this is as you know something recursions working on so do we have enough data for the llms to just deduce biology uh probably not and it doesn't even matter if it's true it that we that we know that there is enough data if you told me that the world has enough data I will still do what recursion is doing systematically create data so that we can um learn what we need to learn in a systematic way that's called the engineering method sure the engineering method suggests that that there's a there's a structured process a repeatable process um we augment it we augment it with domain randomization but we don't start our life in domain randomization we don't start our life by by un by wandering around the universe um exhaustively and and so so I do think that that you'll just you're going to do all of it you're going to do a systematic uh data generation uh you're going to do synthetic data Generation Um uh you're going to uh uh uh you know of course learn from all the data available in the world and you're going to connect a fly a flywheel together uh that allows you to uh learn from reinforcement learning and other methods and so I think the answer is yes in all of those cases MH but the one case that the if I were to start from nothing I would do it the way recursion does it the systematic way of generating uh I I think it's an excellent method which is the reason why we're we're an investor I think it's smart approach and so that means it's probably a while before we can be like your engineers operating entirely in silico well we we and we operate entirely in silico but um we design it completely in silico simulate we simulate in multiple levels uh transistor level which is physics physics level it's like close to it's not quite uh uh not quite Str equations but um uh but electromagnetics equations okay so principal simulations you can't you can't scale that up very far uh logical simulations timing simulations functional simulations behavioral simulations look at how many abstractions I've already captured uh system level emulation and then we do all of that connected together when we tape out then we have a prototype that prototype we got to bring it up in the lab that's and so uh and whenever we make a mistake uh the more the more the more launches rocket launches we do the better we're going to be as a company and so we launch a lot of rockets uh every time we launch the rocket we learn something new even though uh the chip is good enough for you know for production uh it there are probably some areas that are slightly different than our expectation we take that put put it back into the model so that the next chip improves from that and so so I I think our our methodology of Designing chips is not unlike recursions methodology of Designing drugs and discovering drugs um uh we just had the benefit of of having a simpler problem a far simpler problem for example here here's something that we have the advantage of that you don't uh if we don't understand a transistor very well we make the transistor different that's cheating which would you guys agree like for example if you don't understand a protein very well you don't get to make a different no so we go hey you know what the simulation capability of our of our company is not at the limits of the size of the transistor or the shape or whatever let's just change the shape let's let's design the transistor so that we can understand it it's not like the human body is going to shape itself so that you can understand it and so you have a much harder problem you're also not regulated the same way what's that you're also not regulated the same way all true all true all true anyhow anyhow you have a much a much harder problem uh but but one of the the advantages is that it's now 40 years after our starting point and the technology that that um uh that's available is is is genuinely miraculous and if uh if you haven't had a chance to engage the technology please do and the the the the attitude that I would would use um uh if you if you if you wouldn't mind me just saying so is the same attitude that I approached uh the the the coming into this industry is how hard could it be yeah and it turns out it's much harder than you think right but at least it gets you on the journey that's probably the the wisdom so so Jensen stepping back how do you see the role of of of huge technology companies such as yours in Pharmaceutical and biofarma are you are you providing tools like like Levis and pickaxes or do you think that it evolves in a different direction from that I'm hoping hoping uh dear God we end up better than Levis uh I'm sure they're they're great are they still around oh yeah okay all right right here in San when I was in college when I was in college I went I went to school when I was 16 and uh and and I I had five pairs of Levis 501s sure I had no socks uh my mom didn't pack me with any socks 501s anyways I I haven't seen 501s are they still around I don't okay no so so uh I hope we're going to be around in 4 years uh our contribution is this three layers the first layer is is if you are doing if you if you want to simulate uh if you want to do your drug design and Drug Discovery in silico it is very likely that you have to process an enormous amount of data it's likely it's multi multimodal it's likely that it's very very long sequence longitudinal um it is very it is likely uh going to uh require state-of-the-art AI models uh there are several ways that we can help you and we can partner with you one is the Computing platform second is the algorithm the mathematics that are sitting on top of these Computing platforms we're quite special at that we're very good at that uh um and so this area and then the third of course uh you know we we're we are we are uh enthusiastic uh we're passionate and um we are determined uh to work with you to advance this field we believe in this uh very very few companies can say that they believed in this from the very beginning and we're you know still here 15 years later uh working with all of you and so we we deeply believe that this is going to be the future of the way that drugs will be discovered and designed and so so there's there several things that we can offer and of course uh we're also uh fairly Adept investors and so uh we would love to invest in in amazing companies like recursion uh and there are several other amazing companies that we've invested in the audience and and so so uh I we'd love to partner with you to uh create this future together so we can work with you in a whole lot of different levels and you know you know please if you have a hard time with computation or artificial intelligence you know just send us an email we're here for you I think there's so much to learn from Nvidia on computerated design right there's got to be so many analogies there I wonder if some of those simulations if you look at our if you look at our pipeline in Clara from cryo electron microscopy to x-ray crystallography to um uh uh uh you know Gene sequencing to uh amino acid to uh structure prediction uh all the way to Virtual screening of just about every single algorithm you can imagine uh first of all I don't mean to show off but what other chip CEO would talk like [Music] that can't think of any Jensen exactly yeah that's nasty talk that is nasty talk you know in our industry it was just like blah blah blah blah blah and right here everybody goes I know exactly what you know and so so um and so along the entire pipeline um of the discovery of of uh of of drugs and Medicine we we are we have algorithms and we have mathematics and we have expertise that we can be a partner to you okay so so please do reach out to us and and we' love to be a service to you so I cannot let you off the stage without asking for some predictions and how about this I will I will allow you to choose the time frame of your prediction uh but maybe something shortterm like this year more or less and something a few years out um so so short term short term one of the one of the observations that that everybody's going to realize is that artificial intelligence sounds complicated but it's supposed to make computers easier to use the single greatest contribution we have made to society is that we finally made it possible so that all of our children don't have to learn C++ y hey what a ridiculous idea a nightmare there was a time you guys know there was a time when Tech CEOs were oh yeah in the future everybody should learn how to program and I was thinking I don't want anybody to program right why why is everybody programming the computer makes no sense the computer should do what you intended to do and so um we have we have the first time because of artificial intelligence and the work that the groundbreaking work that our industry has now done we have closed the technology divide in a dramatic way everybody's a programmer and the programming language of the future is called human you could talk to it you could have incomplete sentences it could be half with gesture you could be Italian and program a computer you could you could be German French and Italians and Americans can equally program computers and because the computer is multimodal and understands your intention and understand waves and gestures and things like that and so so I I think this is our great contribution what does that mean to you it means that for the very first time our computers and the power of a computer is much much easier to access than ever in history you just have to engage it go take a step I believe that this year um every industry will become a technology industry the the beginning of a journey of a technology industry let me tell you what it means to to be a technology industry um you guys know there's a fundamental difference between selling a phone and building and Merchant uh and selling iPhones one of them selling a phone is neanderthals iPhone is a technology industry and the reason for that is because it's software defined and that device that device um offers a platform by which uh Apple can continue to offer services and and goodness and joy and all of those things to the world for a very very long time well uh you could you're now seeing this the fundamental difference between Tesla the car company and you know other companies uh car companies one of them is building a thing and one of them is Building Technology I fundamentally believe your industry is going to get revolutionized because of your transformation to become softwar designed softwar driven and artificial intelligence driven that is going to revolutionize your industry a medical instrument is never going to be the same again ultrasound systems CT systems you name it all kinds of instruments they're always going to be a device plus a whole bunch of AIS and so uh all the modern uh uh Tech bio companies are starting to think in this way and I think that that that the value that you'll create um uh the the opportunities you'll create uh are going to be incredible and so I I I think this is going to be one of the world's great future Industries it's going to be a technology industry and uh we're you know we're here to serve you my last question I have two young kids seven and nine I do not want them to be programmers in the way I was a programmer but they should play video games oh well they certainly do that and I do I I wonder wonder what you think about this I do tell them I do think you need to know about the algorithmic datadriven approach to problem solving yeah I think we still need to know that even in the age of AI what do you say I love it uh Marty I I think that everybody should learn philosophy 101 which is basically logic and and philosophy 101 which is basically logic is the is the foundations of programming right so go learn that yes and and that that is is Ample Ample um please learn algebra you know that that's how about linear algebra yeah I mean I I would push them until they get to calculus but but you know differential equations you know I I don't know you know shouldn't everybody learn linear algebra though yeah yeah you know multivariable systems is good to to be able to command but but you know you got an AI to do that okay yeah yeah I I think it's okay I I think that anybody who could who could solve um here we go you just have to you have to learn how to do this you have a wedding party of 300 people uh how would you seat these 300 people in tables of 12 so that everybody would have a harmonious and happy dinner as you go as you know the combinations are more than the number of atoms in the universe and that this is a problem only a mother-in-law can solve and we do not need a quantum computer you see so there so whatever it is that they use to solve this problem go learn that okay on that note Jensen thank you so [Applause] [Music] much amazing that was amazing

Channel: Recursion

Views: 37,955

Keywords:

Id: ytZcvwZxkrg

Length: 33min 14sec (1994 seconds)

Published: Mon Feb 05 2024