Eric Schmidt (Former Google CEO): A Global Perspective on AI With

thank you so much for sitting down with us eric uh we're super excited to be chatting with you and uh really glad you're taking the time to discuss ai in the future with us and alex thank you for all the help you've given me in strategy and ai over the last few years well so i want to really uh dive into two main topics the first is ai in the new world order and the second is the ai industrial revolution but before we get there you know you you were at google through many formative years of ai development many modern techniques of ai were actually sort of invented or strongly developed at google in your time there and i'm kind of curious um you know the past decade or decade plus has obviously been this incredible boom for ai and its emergence but what was what was your first aha moment with ai when did you first realize the potential the technology and and what the the sort of like massive potential in the world could be well it was really 2011 what happened was a group of people including some of the founders i think of this movement did an experiment where they analyzed youtube trying to figure out what they could find and they discovered the concept of cats using essentially unsupervised learning as we know it now and it was quite interesting because we always thought there were a lot of cat pictures in youtube but the fact that that's what he would discover after looking at the corpus of youtube information was rather disturbing but all of a sudden that started the prospect a group inside of x google x created a group called google brain and that brain that group began to build what you know today as bert and transformers and the other systems google had been using various forms of machine learning for advertising for a long time but the fact that we then had a language model and we had an ability to do predictive text and we could actually look at binary synonyms and so forth allowed us to materially improve both search and advertising today my opinion i'm no longer there is that much of the gains that they're seeing in revenue and in search quality are actually coming from the technologies built starting with google x and discovery of cats in youtube 10 years ago at this at a similar time we also purchased a company called deepmind which seemed impossibly expensive when we bought it and today looks like one of the smartest decisions ever made because of course they're the leader in reinforcement learning globally yeah no it's it's uh it's absolutely incredible i think i think to your point the the impacts of these technologies they're so deeply embedded now within these these large technology uh products that it's it's almost impossible to imagine what the gains will look like without these incredible technologies well i can give you a simple way of thinking about it which is ai companies are defined as companies that do something involving learning and if you think about it if you're a consumer tech company the more clicks you have the more learning opportunities you have so here you have a situation where every action at google you take you have a learning opportunity for both the quality of the answer but also the quality of the ad now the goal 20 years ago was to end up with one search result one ad the perfect search result and the perfect ad well this technology may get us close to that in the sense that the understanding that you get from learning what works in this situation in this context is so powerful and the interesting story inside of google was that in both cases the traditional teams the normal teams if you will who were doing traditional algorithmic programming who were brilliant i might add right he brought the company to this point we created a competitor in the form of a machine learning model typically built on top of what you would know as tensorflow using the same amount of database and so forth and so on and on our fast tpus and we would see what would happen often the interesting thing is that the computer did not come up with a very different result but it got there in a very different way so what would happen is you'd run the tests and then the traditional team would look at what the ai was sampling and would say it never occurred to us that there was a correlation between a and b so if you believe at its core in the last decade ai was good at looking for patterns that humans couldn't see just mathematically it couldn't see them then that produced a lot of the gains now the same is actually true in inverse which is now generative models which is what everyone is doing have the property that they can both be used in things like gans but you can also use them to generate something new which is the next day so think of that as prediction so we went from deep understanding and because we had the understanding we could predict the likely next outcome with some certainty well that's very powerful too so if you go back to the consumer tech companies just to close it out why have the consumer tech companies gotten so excited about this because it directly improves customer quality directly improves uh customer revenue and it also allows individual targeting to the person or the cohort that the person appears to be part of without having to profile them now it's important to say that it's not perfect they make mistakes and so forth but that's driving this huge increase across the board yeah diane i actually really want to this will be a discussion topic later because one thing that i want to dive in with you on is what are what is the fundamental change to business models that that ai helps enable so um but uh to kind of dive into our first primary topic around ai in the new world order you know one of the things that we've spent a lot of time talking about is what does ai do to governments how governments need to interact with one another how they interact with their citizens and and there's a lot of pretty pretty deep sweeping changes and so i'm taking a big step back what geopolitical trends do you see that will ultimately impact the long-term benefits of ai and and what are sort of the important shifts that we should be paying attention to at kind of like the global level today today well i would say to start with that the most important competition is not going to be google versus apple or whatever it's going to be china versus the united states and let me make the case for it we all understand how the power of america our innovation model where the the envy of the world this was an accomplishment from people over you know 70 years especially since vanderbar bush maybe 80 years to build a system of research scientists innovation and so forth they brought us gps it brought us semiconductors it brought us the internet all of the things that brought us social networks etc everything that you could imagine came out of that maelstrom china has a different model but an equally powerful one they've got four times as many engineers they've got a huge number of companies as you know because we've looked at this the companies work nine nine six nine a.m to nine pm six six days a week by the way that's illegal in the united states don't do it here um so you have a work ethic you have a profit motive and you have a scale of platforms that china can exploit they'll be exploited differently but the important thing is that both the us and china are large enough markets to make their own weather right they're so big they generate their own storms around each other they create platforms that are unique within their own domain that is a stable structure for the rest of our lives and unless something changes materially i don't think we're going to see one internet i think we're going to see this splintering at least at the applications layer and the fundamental reason this is occurring in retrospect is the internet when i was doing this at google was essentially optional you know you could do everything without it you could do it with it people like you and me thought it was really cool it worked really well today the internet is no longer optional the internet is fundamental to everything so every form of human activity occurs on the internet including the ones you don't like and if you look at china the chinese model the chinese have solved the internet problem in their own way they've criminalized or made it impossible to be anonymous on the internet well that cuts down a whole bunch of stuff so all of your speech is recorded and you can be again uh met with the police can prosecute you and using their laws which are quite vague in these areas so it's a very different internet they also don't allow the american companies to operate and we typically don't allow them to operate in our country with the notable exception of tick tock and so the result is two different information spaces and those information spaces will shape the outcomes because people take cues of the information space in the tech industry and you and i have done this for a very long time the tech industry has its own sort of social sociology it's kind of a certain world view it's somewhat libertarian it's quite liberal in terms of personal behavior it's relatively conservative financially um it's it has its own little political zeitgeist and it's big now but that's not but i learned a while ago and i'll say it's everyone we're not the same as everyone else the fact that you and i and everyone watching this are tech people we are a minority compared to most societies probably thank goodness by the way and we need to respect the fact that they operate in other ways i think that to answer this to be one level more specific china is building an internet that makes sure that china remains in power that it maintains what it considers an appropriate level of control over its citizens they see this in the surveillance and in social credit systems and all of these things that are at various levels of deployment now i'm not a chinese expert and you know more about it than i do but i would bet that their strategy is going to work that that the technology that they're deploying will be successful i'm not endorsing it by my saying that please don't take me out of context i'm simply saying i think that they have a strategy they have the resources then they're directive the last time i two times ago when i was in china i was with the minister who regulated us and he gave a speech which i attended where he explained that the only solution to the problems on the internet was regulation they're not hiding behind what they're doing they know what they're doing and it means that i'll give you another example it's very hard to get a vpn now because the vpns use machine learning the system the great firewall uses the various forms of machine learning to try to find them right so that's another example of a small thing but it added all up they're very different that so having said that the next question is what do the other 995 out of 197 countries do and a reasonable prediction is the democracies will form around this u.s western consensus and the democracies by the way include japan south korea india is a democracy and the countries that are fundamentally authoritarian or they're so weak they can't tell what they are are highly likely to be influenced by the chinese architecture because they benefit from it yeah well you know one thing that's pretty curious is that there's you know china is the biggest lender to a huge percentage of the countries uh in the world they're also the biggest trade partner to many many countries the biggest lender to it's something like 70 of of countries or whatnot um what what do you think happens to countries that are that are sort of in the middle you know ones that are highly dependent on china for economic activity um but also ideologically are more aligned with uh you know democratic values in the united states well a good example is australia and australia over the last year has taken a very tough stand against the demands of the chinese to essentially quelch certain kinds of descent and other if you will interference in australian affairs and they've done that knowing that they're by far their largest trading partner is china and china is central to their future growth so far that strategy has worked i think it's reasonable to expect that the bri countries the so-called belt and road initiative countries will all become essentially a clients if you will of the chinese information space so it'll have the chinese information rules which include surveillance and some forms of censorship because literally they're going to get their technology from china so they sort of get it for free to me the hard question is what does germany do number one business partner of germany is not the u.s it's china right the country the companies that i have spoken with in europe most recently are that have their number one their number one international market is not the us is china and their number one supply base the things they buy come from china not the us now i think they're gonna stay in the western fold for a hundred zillion reasons but they're in a tough spot as this division occurs i'm not suggesting that we're going to fully decouple what i'm suggesting is that we're going to have an incredibly uncomfortable competition with china where china will uncomfortably force people into these spots and they'll have to make tough decisions so if you go to the leading countries the leading western partners we're going to keep them you start to wonder about countries like hungary as an example which is having trouble deciding how close he wants to follow eu norms i think they'll probably stay in the west but let's pick my favorite example tunisia which is a fantastic country they don't make much it's very nice people very well run by comparison to its peers it will have more money and more opportunities with china than the us how does it decide right right and you know especially when we think about how this pertains to ai and we've talked about this there's sort of there are sort of two ways to do ai there's one way in which you do ai which is you know sort of authoritarian in uh in nature you know your focus on massive data centralization uh in a relatively uh sort of um unambiguous way and then there's ways to do aim which are far more mindful of bias and ethics and privacy and all the things that i think you know in the western world we think are really important one one interesting front of this battle has been that china has been leading the creation of of global global norms on ai and there's certainly this sort of conflict around uh the standards of ai and and one thing i'm curious about is how do you think this ultimately will develop you think this is just yet another front of a general bifurcation and and what can be done to sort of avoid a a paradigm where the sort of maybe the more authoritarian regime of ai becomes the dominant one globally well i think the fight is underway as you said and everyone focuses on the fact that china has more data and less privacy rules i will tell you that the commission that i had the privilege of leading which finished its work a few months ago called the national security commission for ai appointed by congress found that algorithms are just as important as data and that we recommend a number of things perhaps the most important of them for universities in addition to doubling research funding which is sort of a no-brainer is to build a national research network and the idea is that the big companies google being an obvious example have tremendous hardware tremendous scalability but little startups smaller research groups those sorts of things don't have it so a number of universities have put together a proposal i'm familiar with the one from stanford and we're working hard to try to get that funding into the ndaa the ndaa is the yearly approach appropriations bill for the military and national security that number not along with a number of other things but what i would say is we can't win by adopting somebody else's good practices we have to take our good practices and make them stronger how did we get here first we allowed high high skills high value immigration sort of a no-brainer most people would prefer to work here than in china if they're not from either they prefer to work in the west why don't we let them in why don't we work hard to increase our research funding and also work hard to get more talent into our government government has a lot to say about how these things are doing remember the american government is very complicated all the states all the regul all the federal regulatory bodies which are allegedly independent and then you have the white house and the military and so forth these are huge operations they have relatively little technical talent to even understand these debates most of the ai work that you see coming out of them has been done by volunteers outsiders i worked for the defense department for five years as a consultant and as part of that we produced an ai ethics proposal for the dod which it adopted which is one of our better wins so it can be done but it requires all of us let me just get on my soapbox for a sec sorry this is a national security challenge for the united states if you want for the next 20 or 30 years for american technology american values american startups to be global platforms we need to get our act together now because our competitor by the way they're not our enemy but they're our competitor china is busy doing that exactly they're doing it in energy transportation electronic commerce where they're already leading uh surveillance which they're already leading they're working very hard in quantum where we're still leading i think and they're working very hard to catch up an ai i'll give an example in march we said that we were one to two years ahead of china in ai in june they demonstrated a universal model of a size similar to that of gpt3 open ai's gpt3 which is a significant accomplishment on china's part now maybe it's not as good but the important point is they know what they're doing and they're on their way yeah you know one one question that naturally comes out of this i think you've kind of said in the past hey the government is not prepared for the ways in which digital technologies and ai are going to completely change the way that systems work and i think to your point i think it's very important that governments have more technical talent so they're able to sort of see around the corner and understand what are the the overall implications around how the system will evolve in tandem with these digital technologies what do you think are some of the key ways in which you know that the government needs to be thinking about in which in which the system will evolve and therefore what are the sort of the key initiatives that matter a lot outside of just ensuring that we have an innovative environment where we're building the great technologies well let's start by saying we need to have an innovative environment where we're building the great technologies let me observe that the vast part of the innovation is occurring in private companies 50 years ago the vast amount of this kind of innovation was being done in government labs and in universities so it's crucial that the tech industry be allowed if you will to go as far as it can with these technologies and one of the issues is people like to prematurely regulate things that have not occurred yet so why we wait until something bad happens and then we can figure out how to regulate it otherwise you're going to slow everybody down trust me china's not busy starts stopping things because of regulation they're starting new things so that's the first point i think the second point is that we have to remember that in a national security situation like this we actually have to have a coherent national plan if you take a look at operation warp speed you had a situation where universities invented mrna private sector built the vaccines and the government guaranteed the market whether it worked or not i don't know whether you want to call that industrial policy you call it what you wanted but it was in and under trump and it was a national emergency where we were very very and correctly worried about it and we innovated and by the way biointec which is the source of pfizer is a european actually german company so there's lots of examples where if we put our mind to it we can do this now it requires uh leadership at the presidential level it also requires shifting resources what i found in my work with the government is that everyone gives speeches all day about what the government should do but the government only does what it's supposed to be doing at the moment and it tends to self-propagate so if you want to change the a company you change from customers and bottom up if you want to change the government you have to start from the very top and go it's very directed another example with the dod was they have a process which is internally known as the palm process the funding for ai was proposed and you have to wait two years for the money to show up and then you have to get the deployment plan and so forth it's built around 15-year weapon systems now we managed to find a way to get around that but it's a good example of how the systems aren't aren't capable of rapid change and if ai is anything it's something which normalizes an awful lot of data and provides new insights and we should spend a minute and talk about how powerful ai will be for example for science right it will transform our understanding of biology and chemistry and material science and all of those kinds of things those are the basis of the next trillion dollar industries you know the trillion dollar industries that exist today are in software um thank goodness that was you know what i was doing and the next generation will be in the application of digital technology mai in these other industries which are huge right so think about drug discovery health care uh 18 of gdp today anything that you can do that materially affects that is a huge company yeah no i'm really excited to talk to you about it and in a sec here we'll we'll talk about sort of the uh the ai industrial revolution um and i think it'll be a great topic but just kind of motivate um what you're saying right now you know as part of the national security commission on ai you all you came up with a list of recommendations and i think one of the things that you just noted which i'm very sensitive to is that the timelines upon which governments implement these sorts of recommendations can be quite long it could be years uh years at a time it's not months or or faster and and what do you think is the what do you think the urgency here is what is the time window in which we need to be operating in to be successfully competitive well i'll give you an example of tick tock so tick tock is taking the world by storm it's extremely popular in the united states president trump using a series of tactics try to get it changed try to get it to be the u.s operation hosted and owned most recently on a on oracle platform none of that actually happened so tick tock is a good example of the first real breakout platform from china and by the way it's a high quality platform and much of its um apparent success is because it has a different ai algorithm for matching it actually matches not to who your friends are but rather to what your interests are and using a very very special algorithm so that's an example where i would have told you that would not occur for another five years so we have relatively little time maybe a year or two not five or ten to get ourselves organized around the initiatives that we raised to repeat more money for research building a national research network working with our allies establishing guidelines and ethics rules that apply to everything that are consistent with american values the hiring people into the government we make a set of in our report we make a set of very specific recommendations for how the defense department and the intelligence communities should work and they're typically of the form take this function and make it more senior and give it more resources we also make a couple of other suggestions including creating a civilian university for technical talent it would be free in return for up to five years of work in any form of government not just the military we also have a proposal for a reserve corps where people would spend it's it's built it's modeled basically on rotc where people could spend up to 30 days inside the government helping them and then go back to their jobs in a legally supported and promoted way there are plenty of people who want to help our government get to the right outcomes here they want national security they care about doing things the right way they want to do it with right ethics they want to be involved we can do this yeah and and then you know just uh um as a closing thought here i know like i agree with you i think the urgency is is really there it is sort of a a question of the next few years not a question of the next decade um in terms of the the initiatives and what we need to drive what are some of the the most critical and existential problems that could arise due to ai technology there being a mismatch in ai technology between the us and china for example you know it's not a fun day if china builds an ai that's able to silently execute complex and difficult to track cyber attacks and so yeah i i think there are simple ones that are obvious um it's a balance of power in terms of cyber so you could imagine an offensive cyber weapon or a defensive cyber weapon that was stronger than anyone else's and that's kind of an obvious argument one of the ways the military thinks about this is they think okay we'll take the people who build that and we'll put it in the equivalent of los alamos and we'll keep it a secret but one of the things that's different about ai is that there's almost no containment of ai the technology leaks literally the ideas leak so fast that you don't have much advantage on one side of the other which which is sort of a new grand strategy stable paradigm problem that's the way i would describe that let me give you an example of of agi which we'll talk about but let's imagine that one of the well let's assume that there are 10 10 countries that working on agi and that you end up with three or four in china three or four in the u.s and a couple sprinkled around including one in israel and a couple of in europe and so forth maybe one in russia you get the idea what happens if one of them invents something that the rest of them wants that's really hard to steal right so uh i don't know not a good example but they figure out how to cure cancer in a way that nobody else can well that would be bad well now let's take it one more to the extreme let's imagine that one of them builds a system that's so dangerous that we don't want even that country to use it so for example i can answer questions like how do i kill a million people tomorrow it's hugely dangerous you clearly don't want to make that happen it's reasonable to expect again in our lifetimes that there will be the equivalent of nuclear non-proliferation discussions under a new regime where we're trying to say these things can be built but we don't want very many of them and we want to keep them under some form of guard and we don't want extreme terrorists using them and we only want them to use be used in certain situations even by their owner governments so i think that's the thing that's changed in the last 10 years is the aggressive use by governments of cyber and influence and the russian interference in 2016 that's all new it didn't happen 10 years ago if you take it to its logical extreme then you're going to end up with things which are as dangerous as nuclear or close to it and we don't have any language by which to discuss what does balance of power mean how do we keep it under control how do we make sure that the equivalent of uh uranium plutonium doesn't get leaked right there is no analogy we haven't figured out that doctrine yet and if we don't then you're going to end up in a situation where china has a poor level of security around its agi somebody copies the network inside of it they copy it onto the equivalent of a usb stick and they take it over to another country where they're not subject by that rule and something bad happens right so we have to have this conversation i'd rather have it now yeah i think it's a it's a great call to action i mean aji and ai in general as a technology that's as powerful potentially as nuclear weapons and sort of the previous age but with very different properties it's very easy to replicate it's hard to contain and these these cause like very real challenges for the for the new world order you know with that um i think it's a good i i wanted to segue into this topic that i could tell that you're really excited about which is kind of the the application of ai to to all industries in the world and i you know we like to use this term the sort of ai industrial revolution and one of the incredible things about ai is just how broadly applicable the technology is as you mentioned you know can be applied to most industries uh in some way that is that is deeply transformative and b it can be used across most these domains to really great benefit uh almost very uniquely from any other technology you know if you think about the internet or software those were most of the value that was actually generated by creating almost new domains or new ecosystem environments but ai has this ability to be maybe more foundational maybe more cross-cutting into many of the existing industries so what what technology analogy do you think is best suited for thinking through ai and its implications do you think it's more like sort of the computer more like the internet more like electricity what do you think that the right analogy there is it's hard to know i mean electricity was pretty important too and i lived through the personal you know the mainframe revolution the personal computer revolution and each revolution has seemed bigger and more impactful than the last so it's fair to say that ai and ml are world changing norms changing and especially because they tend to work on information spaces so one of the ways to understand how society works is society has morals more a's excuse me and values that are embedded in information space which each of us is raised in and lives in so let's imagine a situation where you have ai systems that are shifting it how do we want to deal with that a simple example is that you have a two-year-old you get a two-year-old toy that can speak to it at three the three-year-old gets a smarter toy by the time the kid is 10 this toy is that's been upgraded of course is by far his or her best friend what happens when the best friend tells it something wrong or encourages bad behavior or observes bad behavior we've never had a situation where we've had a human kind of intelligence that's on par with how humans raise their children operates their society and so forth there are obvious examples from misinformation so today we already have a misinformation problem but because of targeting ai systems will be able to learn which human biases these are recency bias and you know so forth and so on to exploit to get you even more passionate about something which is false and that cannot be good so in the same sense that that you have the damage that's possible in the information space because of this targeting and because we live in an information space you have this extraordinary gains that will occur with savants that can help you understand specific fields let's look at synthetic biology so synthetic biology can be understood as like ecad in the way back when and that you're basically building biological organisms but you're not making copies right they work differently but biology is a lot about coming to the lo to the i may not see this correctly to the lowest energy state in any of the formulations of of the atoms and molecules and the things that are synthesized around of them and machine learning is very very good at doing that so machine learning is very very good at guessing which set of compounds for example will improve or make worse this outcome and i'll give you an example there's a drug in mit called hallison it's a combination between synthetic biologists and computer scientists and the synthetic biologist said we want to build a new general-purpose antibiotic right like the ones we already have there hasn't been one in decades and people have obviously been looking so what they did is collectively they organized the system to generate as many compounds as they could that had some kind of antibiotic resistance and then they did a further network that looked for the ones that were farthest away from the current ones and they came up with a compound which is now in various forms of trial called hallison now that's something that humans couldn't do and i think it's going to work you know because you know the story of alphago and alpha chess that not only did they beat the humans which was a surprise in the case of alphago in 2016 and caused a massive massive reaction to that in china but more importantly they discovered new moves in games that are 2000 years old now that's something that's extraordinary so example after example where the simple ones are parts management in your inventory right that's an easy one that's essentially a prediction but why don't we do one where um when i go to the hospital it predicts what i'm showing up in the hospital for and let's see how well that works that'll help the doctors right because i am although i think of myself as distinct i'm genetically very similar to 999 other humans i just don't know them but the computer can find them and say well they all have this problem america has the same problem totally yeah i want to dive into into two pieces that you had just mentioned here and so the first is the scientific and the second is is the economic implications you know what one of the things that um i'm equally excited about is how ai is actually being successfully applied to science uh whether it be biology drug discovery material science et cetera and to exactly what you just mentioned it allows us ai has enabled us to do significantly more of the science digitally um and very very efficiently on computer which takes out you know many fields like biology or physics for material science or whatnot takes out a lot of the the costly components of actually doing the science which is the alternative is you have a bunch of scientists and they have to do work in a lab and it's just very very costly it's very it's meaningfully different what do you think is what do you think are the scientific implications of this you know um one thing that's been that's been noted is that over the past few decades scientific discovery is all has actually slowed um over time do you think that this results in a sort of like renaissance or a re-acceleration of scientific discovery it should be a renaissance i'll give you an example um i'm one of the funders of an important project at caltech which is trying to do essentially a new forecasting model for climate change and this particular group focused on clouds and i was not aware that it's impossible to simulate clouds using navier stokes because of the number of equations and computation and computers not even all the computers in world will be fast enough to really simulate them but it turns out that you can learn an approximation of how clouds work that's quite workable i was in another meeting where the there was a need to understand it's hard to describe whether a mouse was sleeping or not and i won't bore you with why they need to know this so and it's very difficult to tell if a mouse is sleeping so they didn't have very good labeled data so what they did is they built a natural model of the physics of sleeping for a mouse and then they generated the training data for synthetic mice sleeping and they built successfully a mouse a model that would tell them whether the mouse was asleep or not now this is humorous but it's really quite an accomplishment they knew the laws of physics they had no training data right they generated the training data and they were able to do it so a simple formula for you is it sometimes in science ai is used to approximate a function we don't know the function at all or we can't compute it quantum chromodynamics is another good example they just there's just not enough computers in the world and nor will there ever be to get the stuff right and so you need an approximation and the approximation works really well the other aspect of science is this generative part where you can generate new things and you can try them they're both very important most science progress seems to me to occur after the development of a new instrument you know a new microscope if you will and the problem with things like spectrographs and so forth is that they're very very expensive to operate at scale so with these techniques we can use existing data existing databases and natural natural data physics simulations and so forth and we can really break through that limit so that's why we should imagine that what you said is true and that you'll see these breakthroughs and if you care about climate change which i do a lot this will probably be the thing that will allow us to address it through innovation because the current approaches are just not working well enough yeah and you had alluded to this um before where you know you'd mentioned hey if if ai is actually able to insert into each one of these industries in each one of these scientific scientific pursuits and drive breakthroughs that's going to result in trillions of dollars of value over the next few decades and i'm curious to to dig into that model a little bit more deeply what do you think are the economic implications of ai being scalably applied to every single industry and scientific area so let's talk about industries that are not regulated or lightly regulated those industries will fairly quickly be disrupted by these techniques because either an existing company or a new company will adopt them and they really will solve the problem better and that will then create a crisis for the number two number three number four and that's called capitalism when that happens there's an awful lot of destruction of jobs and of shareholder wealth as well as winning and so forth and so the question is in aggregate do you create more jobs or less jobs and unless we make a few changes in the way we operate we'll probably end up with less at least high quality jobs and the reason is that the winners tend to be concentrated and they tend to get more of the spoils there are other jobs available but they're not very interesting ones and so i think that along the way with this technology is super important that we figure out a way to use ai to solve some of the problems that have bedeviled us i'm familiar with a company that is using ai to try to determine why certain parole officers put everyone that they see into jail for parole violations and others don't that's an example of a real human crisis for those people involved that's an inefficiency in our in our market so overall i'm just using that as an example there are a thousand such examples we need to figure out a way to use these tools to produce a better educated and more empowered and higher income workforce if we don't the ginny coefficient literally the gap between the richest and the poorest in every society is going to increase and that's clearly not good yeah and what do you think are some of the the you know one of the very interesting things about many of the recent advancements in ai is that if you look at where the advancements are happening it's actually cognitive work that is being more successfully automated you know you look at the codex or copilot systems out of out of open ai and microsoft you look at alpha fold coming out deep mind and these are these are esoteric skills or skills that are difficult to train humans on um but they're being automated very well partially because they're so digital in nature um what what do you think of the implications of this where at the same time sort of more manual skills are actually much harder to automate you know robotics has been sort of this devilishly hard problem um for a very long time self-driving has been a hard problem how do you think this plays out over the course of the next five to ten years well i think it's reasonable to expect we're going to see in purely digital industries extremely rapid change from the ai platforms and the universal models if you take a look at gpt3 and then the products that microsoft is offering now around coding that's a good example where in programming you get a return signal you get a reward signal did they did they like my suggestion so it's reasonably obvious that if you have a big enough model and you have a big enough training set you should be able to build quite a good product and i'm sure that that's their strategy i also know that they have competitors coming so not only do you have them but you have you know three or four competitors who will be well funded that collectively will move that industry forward and who wouldn't want the computer to help them write their code anyway that would benefit everybody so i think you're going to see that much more quickly than you'll see some of these other industries the industries that move the slowest are the ones that are not subject to commercial or regulatory pressure most regulated industries the companies and the regulators essentially have worked it out and so they tend to agree on everything and they tend to be difficult to for new entrants and to enter so if the result is a partitioning between regulated industries where it's very hard to enter and then this incredibly fast evolution on the unregulated side that will create other problems and other disparities the simplest way to think about it is if you're not to offering the equivalent of an android and iphone app to do anything what's wrong with you right so today someone your age and your generation says like why do i have to carry a passport right why do i have to go to a bank uh why do i have a vaccination card these all seem dinosaur strategies and yet they made sense 20 years ago so i think you're also seeing a situation where the rate of change is outstripping humans ability to change it my own view is that the rate of change in the next 10 to 20 years will be much faster than the last 10 years and boy what a 10 years that has been because of the compounding you have the situation where it's combinatorial innovation each layer builds on the top and those layers are getting filled in very very quickly so if you start with the presumption of in every business and in every task there's some learning function we just have to find it then you can probably build a killer app the most obvious one that i would like to emphasize here is education you have all the things you need you have a lot of students you have a lot of learning you have a lot of quick clicks or equivalent of clicks why have we not figured out the optimal way literally the scientifically optimal way to teach english math physics science and so forth i'm busy funding some of this activity but it's amazing to me that you have this immense education industry right which is many percentage points of gdp and almost no innovation in the underlying science literally the computer science of learning uh in our in the field and it's an oversight that we need to correct yeah 100 you know i if you um so i think i agree with you which is that the sort of the one of the fundamental shifts at a sort of physics of business perspective you know one of the things that happened when we had software was that all of a sudden you had this thing it was free to replicate that was like a big change you know it changed how we thought about many things um one thing that you're mentioning is now with ai or with machine learning algorithms you have this new physics that is occurring which is systems that just get better very very very quickly that have compounding improvement to quality and and uh and uh effectiveness and uh and functionality um so one of the things so that's that's sort of that maybe the the underlying node that's changing but can i just let me just add that when we started all of this you know i've been doing this now for 45 plus years it never occurred to me that we would end up with this amount of concentration of power in countries in the form of china and in companies in the form of the u.s leading companies the entire vision of technology which of course all all of this came the internet all came out of vietnam in the vietnam anti-war movement was decentralized control not centralized control freedom of the individuals remember the end-to-end principle in the internet so it's important to understand that the structure we have now where you have this concentration of power which is both economic social moral regulated and so forth again using china in the u.s that may not ultimately be the state this may be a situation where the technologies go from centralized to decentralized to centralized to decentralized and it's presumed that the chinese model of authoritarian control is going to be the dominant one but i can imagine that with the empowerment of these models giving each person their own super computer identity human friend they're going to be a lot more powerful too and i don't think we understand this and it's canonic to say everything will be structured everything will be hierarchical and these big companies will be formed but let's imagine that in your next company you found you found one that builds the assistant that helps everyone get through life is that company going to be a bigger company than the current companies if you pulled it off yeah you would be and you would also be regulated to death right now what happens instead is let's say that you do this after your current company and there are a hundred and no single one becomes dominant because they're all specialized right that in fact people are different enough that there's no single market i just don't think we know yeah well this this uh is a great segue to i think one of the last questions which is if we think really far out what are the what are the changes to how business is done and you know if you really look at it if you you know my view is if you look at um sort of the last generations of technology the internet the personal computer the phone et cetera there's effectively in three business models that that over time dominated there's ads which you know a lot about there's enterprise software which you also know a lot about and and e-commerce which we have in the united states it's huge in china massive massive in china and those have been sort of like the three dominant business models that technology has really enabled do you think what do you think are the the the new sorts of business models that uh potentially or wouldn't exist because of the advent of ai or these new technologies that are being integrated well we've talked about for decades the um that there will eventually be micropayments of one kind or another and we're still waiting for those micropayments it makes sense to me that advertising sponsorships subsidies of all of those will be part of it but at the end of the day i have a different way of thinking about it which is just build a product that's 10 times better than the incumbents now when i say i didn't say 10 i said 10 times if it's 10 times better then none of these things matter because it'll all come to you now if it's 10 the incumbents have such strong incumbency benefits including regulatory capture brands and so forth and so on that you it that the standard is high i think that with ai you have an opportunity across most fields to build a better mousetrap for every sorry to use mice as an example in pretty much every industry and it's not just ai most industries do not do digital design so for example there's a concept called digital twinning where you in build an entire software version of the thing you're going to build this is now done in the car industry for example most of the manufacturing industries are not doing it most industries still do things the way they were done 10 or 20 years ago at the human level there for every one of those there is a new way of coming in with a sharp technical team using the data and learning the outcome quicker tech is the first one because we're the tip of the spear we have the best training environment we don't have the regulatory requirement we don't have the capital cost but the same the same applies to all these others i'm not worried about the monetization i'm worried about getting users if you can basically get users and get a growing business trust me we can make money we can make money by licensing selling transferring the technology we can build widgets we can build another widget we can sell that and so forth people always blame the revenue because they didn't have a good revenue plan why don't you just build a great product right if you have a great product your customers will come to you yep and uh this has been this has been super wonderful i i want to close on on uh one important question one that i know that you've thought a lot about and uh this could be just almost like a call to action to the audience we have today you know the common moonshot associated with ai is often agi which you know could be very very far in the future and who knows exactly what it means anyway um but uh but one of the things that i i'm really excited about is like what are the other grand challenges of ai that might be sooner that we can all get really excited about so you mentioned climate change earlier which is one that you're personally very excited about what do you think are the the big grand challenges of ai that are soon but also deeply important well what i've noticed is that the the simple formulation of basically traditional learnings self-supervised learning partially supervised learning and so forth those were last year's scenario what people are doing now is they're building extremely sophisticated multi-model reasoning systems they generate a set of candidates and then they get rid of those candidates in some other way and then they do something else so there's typically a two or three stage pipeline and getting that pipeline right is the job of whether we like it or not phds in those areas because they really have to understand at a very deep level what this network is doing i wish the network could sort of figure that itself out but we don't know how to do that yet so it seems to me that the greatest short-term opportunity is to build enough of an infrastructure that these powerful models can be done by master students instead of phd students historically in tech this stuff started by phd students building it and then it became common you won't remember this but there was a time when email was considered a vertical right it was sort of something that you added you had to buy right and then it became we used to say it was vertical and went to horizontal it became part of the platform i want in software for the tools that we're describing now to be so commonplace that people of let's just say relatively normal technical education can master them the problem in our field today is it fundamentally takes a phd in math or physics or computer science to understand what these things are doing and that's not a good long-term situation the digitization the process of digitizing the world is a massive business it's a massive calling there's lots and lots of opportunities but we're too reliant now on these very rare specialists men and women who are really really good at this and we need to make it more common it's the old thing of we talk about the top universities and they certainly have a big impact but the the universities that have the biggest impact are the large state schools that generate so many people that fill the companies of the country so let's focus on that too let's focus on making the tools for them such that they can really do sexy stuff right but they don't have to understand it when i when i started as a computer scientist the first thing they taught me was bubble sort my guess is that there is about a zillion github bubble sort and sorting algorithms and why do i really need to understand that except that i had to get a good grade in my class right why why do i need to understand how to make modems work which i used to do right all of that stuff should be elided and so the the notion of progress is is platform progress um we used to worry about things like language scanners and translation and so forth and so on all of that stuff should be excerpted so that the programmers that that we work with are working on the really hard problems and furthermore ideally with with these universal programming models coding models that the computer will help you figure out how to solve the problem yeah this is a great it's a great call to action it's something that i think many people in the community are really passionate about is how do we democratize ai and machine learning and make it accessible to all well that's it i mean just just take take the stuff you're doing and show it to your roommate cool go like what what are you doing i mean try to get try to get it so that you can even explain to a reasonably normal intelligent person how these systems work and then figure out a way to build models that that and that's always how we work and by the way i mentioned open source and github i was part of the open source movement when it was started and the open source is critical for the knowledge sharing that goes on because people share the stuff and they really we really do move faster because of sharing so if you go back to your earlier point about the scientific discoveries are slowing down and so forth how do we accelerate that we work together to build very powerful knowledge platforms that the next generation which in your industries every year or two can build the next generation of app on and that app solves a really important problem amazing call out well thank you so much for taking the time today eric um this was a really interesting conversation we covered a lot um and uh and excited to chat next okay thanks alex and i'll see you soon [Music] you

Channel: Scale AI

Views: 30,564

Keywords:

Id: CmBpsw1ORQ0

Length: 58min 4sec (3484 seconds)

Published: Fri Dec 03 2021