Kai Fu Lee on the Future of Artificial Intelligence

thank you so much for joining us from beijing so an evening hello to you uh kaifu a little introduction for our uh our guest uh probably doesn't need that much but he is one of the world's leading ai experts he founded microsoft asia's research lab and he's trained chief technology officers and ai heads at baidu tencent alibaba and huawei he was president of google china where he helped to establish the company in the chinese market now he is ceo of synovation ventures he's invested in china's high-tech sector which gives kaifu you a unique perspective on how ai is going to develop in all the different geographies around the world it must be one of the most important subjects of the 21st century is how the interface between artificial intelligence and us mere human beings actually works we're going to chat this event will run for an hour kaifu and i will have a good chat for the first part of this meeting then as connor says do put your questions in the q a and i'll come to them uh in the second half of this hour-long event so be great for you all to get involved so kaifu artificial intelligence i've been fortunate enough to cover uh this subject for some years as a journalist and your book ai 2041 i've sent a lot of books any other people who have seen me do intelligence squared events before will know that um not all books are at the same standard this which i finished last night was brilliant it was exciting it was thrilling and i want to ask you about the trick initially as your first question that you pulled with this book is that you have written it with a famous fiction science fiction writer chen chu fan and i just wondered kaifu could you kick us off first with your idea in the book there have been a lot of books about ai but this is very different why approach it with a fiction writer because i've met so many people who tell me they're intimidated by the complex rocket science math behind ai they don't get it so they don't bother to understand it and therefore they are sometimes misled by statements that or comments they hear from other people on social media etc and yet i think ai is such an important technology that everybody should understand and the gap between ai and their understanding is just someone who can tell it in an interesting engaging and maybe even entertaining way and i know i can't do that so i reached out to my friend uh chopin and we actually both used to work at google and said hey you've now become a science fiction writer you're the president of the chinese science fiction association you've written best-selling science fiction books how about if we work together and and here's what i would like to see if you're willing to do constrain your imagination and don't put anything in the stories that i don't think is feasible to develop in 20 years and actually to my surprise he quickly agreed and then we moved on to uh to write the book it's a really fascinating way to as you say tell the story or look into the future around a.i and does it in a way that allows mere mortals like me to understand many of the concepts that ai is grappling with the ai development is grappling with i just wondered just to kick us off on the actual discussion about ai itself this is a complicated at times thrilling at times intimidating as you say sometimes sounds scary you quote at the first um in the preface to the book amara's law we tend to overestimate the effect of a technology in the short run and under estimate the effects in the long run give us a big overview of your sense of where a i is taking us are you net positive or net negative when you look to the future uh i do think it's so true that we overestimate the short term i would say three to five years and underestimate for the long term a case in point is when i worked on speech recognition in the 80s and 90s and whenever i worked at apple microsoft and you know our ceo would ask hey when will this really work and i would say definitely in five years and i was always wrong until the last maybe 10 years or so but had i answered 20 years i would have been right all the time so we we technologists sometimes just think hey it's demonstrated it seems to work but we haven't figured out that you require so much more computing and figuring out the user interface and getting market acceptance and potentially legal issues etc before it gets accepted so 20 years is really a long time a lot of things that seems maybe a little bit like science fiction can really work and we can extrapolate it based on various types of trends like what percentages are speech recognition machine translation image recognition improving how fast are autonomous vehicles improving one can really make a trajectory for these and as far as the net positive or not question you asked i believe that history has shown us that big technology platforms like electricity internet computers have definitely proven themselves to be by far more positive than negative for humanity yet at the time that they were introduced there were periods when people may have been disillusioned or felt it was hype or failed is too difficult to build or overcome but eventually if the technology is truly magical and useful enough it does prove itself in the long term so i'm going to use that as my basis to to as a foundation why i feel positively over time i've also realized that today when you look at ai there are many uh issues that could be very troubling that could be very challenging and some could actually become existential problems but but electricity internet had those challenges also so i want to remain optimistic and i think history is um i think is providing evidence uh that we can be optimistic now kaifu lots of people have described you know the impact that ai may have in various terms describing it comparing it with the internet electricity as you as you did cinder pichai um chief executive alphabet um said that ai was as significant as the invention of fire or the discovery of fire sorry and how to um actually manipulate fire and be able to control it in some way i mean is that the significance that really the public need to understand about how ai can affect almost every process that we're involved in yes i think fire would be a suitable uh comparison um i like electricity and um on internet better because they were actually invented through critical thinking and scientific process as opposed to fire which was presumably discovered um as an accident but i think the impacts are similar that these are platforms they're not one invention that solves one problem it's not like the elevator or uh you know the switchboard those are uh or the typewriter these were good inventions but they basically solved one problem but electricity internet uh ai and i think fire were became a platform that enabled many many things and many other inventions to be built on top of that so i i think those are probably four of the a handful of key breakthroughs that changed the future for for the human race let's go through some of the key concepts uh that surround ai and as you say i think can make it feel intimidating for people trying to speak about its relevance to them and how things might develop and in some sense allows some of the more scary um um uh possible uses to to over index for many members of the public let's kick off with uh deep learning and you do a lot of work on deep learning and how vision object recognition for computers has been hugely helped by deep learning processes talk us through deep learning as a as a subset of the ai world because people can confuse terms and not quite know which part of ai they're trying to talk about and given your point about it's actually a platform for many different applications where does deep learning sit in that group okay the proper definition of ai is study of all technologies and things that resemble intelligence human or superhuman and underneath that study machine learning is a particular way for algorithms to to learn to become intelligent and usually using mathematical uh mathematically sound algorithms to do that from data and deep learning is a particular branch of machine learning that uses deep neural networks and you expose a lot of data to it you don't actually program it to do things you tell it the goal you want to accomplish show it a lot of data and then it learns from the data the structure of the data in a way that is suitable for the computational model that structure almost definitely does not match the human structure because humans are good at some things like analysis and inferencing and creativity and and rules and contexts and things like that but machines just look at huge amount of data in a complex deep network and trains a mathematical model that optimizes a certain decision-making process so that is the process that's being used whether it's in speech recognition machine translation a bank trying to decide whether to give someone a loan would actually collect data from lots of people all this all the features they can imagine the salary the address the zip code and also previous purchases and whatever data we can get their hands on and and all the and then labeled with whether that person when borrowed so much money whether they defaulted or returned it and then it learns to make a very complex function that separates the the people who returned it from the people who defaulted and that decision surfaced in a thousand dimensional space potentially is able to look at any new person who contributes these answers to the questions and and gets mapped to the space so that's why uh deep learning is able to work so effectively uh by looking at huge amounts of data and it and what and the key property several key properties it possesses one is that it gets better with more data evidence seems to show that the performance of a system continues to improve if you throw more data at it with along with more processing and maybe some tweaks to the model so that's something that's really amazing because your excel and your photoshop does not automatically improve just with a faster processor or more pictures but this algorithm actually does so that's pretty amazing the other thing is that it can be targeted personally so the fact that tick tock shows me a bunch of videos i can't help but looking at the same videos probably won't work on my daughter or on my wife or on my friend it's basically learning each person and showing that person videos if you don't use tik tok youtube does the same thing showing you videos that you can't help but watch and and that ability to customize for each person is something static web pages or humans cannot do there are other capabilities but these capabilities really set it apart learn growth capabilities as more data is presented and ability to personally target each individual with a something perfect for that person so so that's what distinguishes deep learning from other machine learning algorithms and also why the deep learning has become the superstar the big invention that's propelled artificial intelligence forward that said a lot of people including myself sometimes use artificial intelligence machine learning and deep learning interchangeably yeah so a lot in there neural networks kaifu such an important part of how the ai approach uh will develop can you take us through in in lay persons terms neural networks because sometimes i find talking to people about it and as i said i've covered this subject for a while they then think that is actually sort of mimicking the way the brain thinks what does neural networks actually mean and then i think we'll get on to some of those issues around the personal targeting okay so a deep neural network would be like thousands of layers and the layers are connections and nodes so uh connections have weights so the way you train a neural network through deep learning is to show it lots and lots of pictures so let's say you want to separate pictures of dogs from cats so you feed it a photograph of a dog and then um into the network and then the layers are are numbers that keep multiplying uh two so so the the the connections going to a node are are multiplied by the weights and added together so think of it as just flowing forward uh internally structures that um propagate these weights from the photo of the dog or cat itself then it flows all the way to the end and the last note either says dog or cat and and if there's a picture of a dog and it says a dog then you want to go back and tell the network europe you basically gave the right answer uh if it if you gave the wrong answer then it would ask the network how would you want to readjust your weights in a way that the next time you might get the right answer so it's like you know teaching a child saying this is a dog that's a cat that's a dog that's a cat what's this and if they got it wrong then that uh that one whether it's right or wrong becomes reinforced um by by learning by saying okay you figure out whatever it is that makes this picture a dog yet you thought it's a cat whatever made you think is a cat it's a little bit wrong whatever makes you think is not a dog is a little bit wrong tweak your tweak your weights and then this weight tweaking is the mathematical um magic called back propagation uh or gradient descent that uh that solves the problem of course it actually has to go through these um potentially millions of iterations for the gradient descent to adjust the weights so that when i presented let's say a million picture of dogs and cats after many many iterations maybe it'll categorize 99.9 accurately and it can't just can't whatever it does it can't tweak the weights anymore to get it more accurate then we feel that is converged and then we take this network and show it more pictures of cats and dogs and hopefully it'll perform close to 99.9 fantastic that's a super helpful way of walking us through the neural network uh system just on that personalization issue you raised um around ticktock youtube of course there's that kind of slight clarkson moment then for some members you know some people that's the problem with tick tock and with youtube it serves me certain things in certain ways to to trap me some critics say on the site for commercial reasons is ai simply a super boost away of very big companies making even more money out of their ability to keep us trapped in their ecosystems it is i wouldn't say the internet companies are evil in the way that you may imply uh they're just they just want to make money and i don't think they want to cause any of us any harm but um they're the basically first set of companies that started using ai because they're the first ones who have so much data and then the first ones who realize uh ai applied to large amounts of data can maxim you can maximize whatever you want think about it if you're a ceo of amazon google or facebook all of a sudden you have a tool at your disposal you can basically turn the knob and say i want some more uh revenue i want some more usage i want some more profit and whatever it is you want it will show people uh whatever it is that will get you the most of whatever it is you want so a company in the growth stage might want more users then more revenue then more profits and if your quarter is challenging you tweak for a little more profit and if you're if you want to go for more growth you tweak for revenue it's almost like a magical uh thousands of employees who will obey your exact direction so this comes as a huge gift and the result of which is first these internet companies made a lot of money at least half of their market cap is based on the power of these technologies combined with data secondly it reinforced their market leadership or you want to call it monopoly you can uh but um because previously uh powerful market positions were created by brand good products and the users and loyalty and maybe your data stuck with it or something like that but now it's adds on top of that the total vast amount of data it collects from all the people on which it can train to target any of the all of the people and that power cannot be replicated by a competitor who has one-tenth the market position so it will reinforce monopolies um as well so these are the two powerful things uh and and really i think um the big mismatch is that they didn't realize hopefully they're beginning to realize is that they are using these this double edged sword called ai optimization maniacally to help them make more profit without realizing the externalities that it may have on the users so the externalities were not intentionally created our addiction or our inclination to watching more violent or more extremist content uh they're they're basically ai uh maybe one step at a time making us a little bit more extremist a little more violent or a little bit um uh interested in inappropriate content uh just because it gets more eyeballs and minutes for the website so i think people now understand the danger the question is how do the internet companies find a way to build and rebuild and fix their ai so that they're really more in alignment with our the user's interest not not just their interests and and that is still not yet happening but it's clear we need to get to that goal otherwise they will continue to be criticized by users by media uh and by books doesn't that suggest um that uh kaifu that the issue around regulation public debate which i know your book challenges head on and says there is a need for this is really necessary the length of time we have been speaking about this and the possible rocket boosters that ai could put um under the business models of the big tech companies who have as you say the access to these huge data sets allowing them to carry on uh monetizing or building successful businesses i i didn't mean to infer that they're evil for doing this their businesses that that's as you say is is is is a corporate goal and a perfectly reasonable one but i suppose the critics would say that that regulation is often discussed but is so difficult to make happen whether it's governments nation to nation global regulation regulation the tech companies keep saying well please send us new rules we'll abide by the new rules but so slow the engagement on regulation how could ai help us regulate these processes because frankly getting human beings to agree whether they're in the u.s or china or in the uk or in south korea or australia or sudan ai must be able to help here surely give us a solution i don't think it's so simple uh it is clear to me that breaking these companies up isn't the right solution neither is take all the data away from them because then we'll also be will be gone all the conveniences that we have through these apps so uh efforts to push them closer to interest alignment should be the solution so one punitive way to do that is to uh publicly uh publicize how well they do with fake news or some other metric and and measure them on a monthly basis if they don't perform well maybe there's a fine if they do particularly poorly on some others some aspect maybe there's an ai audit and if they do well then maybe they qualify for some kind of corporate esg investment threshold and people will buy their stock so just make make their reputation more linked by quantitative metrics on how they're doing that might be one possible path a another path would be make very clear what things are not acceptable because if you're going to do audits we need to decide if it's fairness we're after or addiction or showing fake news have to decide what is the metric and and i think uh their usual answer is well we can't possibly catch every fake news because there's a lot of areas of gray but i think we have to figure out how to have regulations that have degrees of gray that is your fake new score must be 80 or better whatever that means and then we're going to call it right sometimes call it wrong sometimes but if you're doing a great job overall we might make a few mistakes but you still score 95 and we set the bar at 80. this is the way ai works ai is never certain whether the picture is a dog or cat it's just got some likelihood so why not apply some kind of a score and say you just got to score better than that yes i know not every question is scored perfectly but most of them are scored pretty well and if you do a good job we know you can get over 80. so set some pretty tough rules and be comfortable that your point your point correction won't be absolutely perfect and my favorite one is to come up with a whole new ecosystem where every app builds stuff that are consistent with our interest that's not at all impossible think about our interest what do we really want for ourselves we don't want to be addicted to three hours of youtube videos we don't want to be shown things that make us angry uh we we want what we want are things like uh become smarter have more knowledge become more liked and respected and become happier these are all goals so can we somehow measure apps in the way that they give us these truly valuable indisputable universal goals um not measured in one day like how many clicks but measure maybe in a month maybe in a year uh the closest approximation commercially i can think of is netflix because it's not ever it's not supported by advertising it's not trying to trick more eyeballs per day it's trying to create high quality content and and based on our rating and based on our willingness to watch 20 episodes in one season so as a result most people would agree content on netflix is is professional and it's suitable for different people uh and it's not it doesn't as at least um doesn't come anywhere close to advertising based businesses in showing us by the biased addictive short-term addictive content so i think it points the way there are such services that people will pay money for that make themselves have fun or grow or become happy over time so promoting these business models with whatever policies or investment models that are possible would probably be a good way to permanently get out of the current situation that's really interesting is that a global how do you set that type of process so that globally all the players can be clear on what that set of rules or guidance i think that's a wonderful way of putting it it actually works who who sort of i suppose the question is sometimes if you think about the as we would say in the uk the person on what we call the kappa mommy bus which is kind of just the person going to go to work on the bus might just say who's in charge of that bit of it of making sure that we're all a bit more like netflix than we are like some of the uh rather more criticized um platforms yeah yeah um i think people who invest like us i can do a better job in investing things that are i know are for longer term good uh governments can give incentives to investors and entrepreneurs who will do things uh inves investing things in that direction uh entrepreneurs should be informed about do you want to be regarded like a a company helping people or do you want to kind of be always mentioned by media as a public enemy and i think people will gravitate in the right direction uh there's not a one person responsible it's there's a bits of it that all of us can do and and i think once once um there is momentum then there will be a large shift it's like the first advertising funded website whatever it was yahoo or netscape or whatever uh brought in the avalanche of the rest of the uh business so every for for a long period of time every internet company when they're raising money and asks what is your business model the default answer is through advertising only recently i think we are creating momentum against it we're hearing more companies doing subscription whether it's enterprise software or cloud solutions or netflix-like products so i think a shift is kind of happening for business reasons but i think we can maybe add a little more uh fuel to that uh move away from advertising uh towards subscription funded then that will naturally create a better ecosystem and it could also be actually become part of sort of esg your sort of environment social governance model maybe should be referring to these types of issues and should be part of your of your of your way of operating as a corporate well right right esg probably shouldn't penalize companies for being advertising funded no you couldn't penalize companies for having too much fake news or too much uh unfairness or too many or for many consumer complaints or something like that yeah we're getting some questions in which are brilliant i'm going to come to them in a few minutes but this is a fascinating conversation thanks so much uh kaifu uh about your book ai 2041. i don't want to appear negative because there's much much uh positive in your book around healthcare for example around education around personalization of education but i do want to ask you about the what might be described as the fairness and bias issue around big data that it doesn't become clear until you get to the outputs that because of the way data is collected and frankly because of who leads on the coding but also who leads on the use of digital platforms particularly where lots of data will come from can lead to biases in the outputs issues around gender issues around race how do we build systems that are able before you get to the output moment where there's a problem and it over recognizes and over indexes people for biased reasons how do we capture that before it creates the types of controversies that we've seen there are a lot of simple obvious errors that can be fixed for example a large american company trained their hr ai system on a lot more men than women then their their resume filtering uh ended up giving a bias for men naturally so that kind of problem can be addressed by better training for all the ai engineers that they bear responsibility to have balanced data especially regarding gender race and some of the key issues that people care deeply about and furthermore there can be tools that will raise a big red flag and says hey you just compiled an ai program that got all this 90 men 10 percent women we don't recommend launching it launch it at your um your responsibility so those are things that can be done at a reasonably simple level of balance but um a lot of the more complex issues the first story in ai 2041 is actually one of my favorite even though it's a fairly simple example of a powerful insurance company that owns a lot of apps social e-commerce and everything it ends up getting in the way of the love life of a girl and trying to prevent her from dating a a man due to racial reasons even though all traces of the race have been erased from the system because ai is so smart it can infer that based on not just you declaring what your what's your race or in this case cast part of the caste system it can figure out based on your name where you live how how much money you appear to have and your zip code and so on so so the power of ai is very strong that even if you go out of your way to balance your data to to remove elements that you think may cause a gender or racial bias ai may still figure it out so i think that requires a lot more studies and i think it's at at this point i would say that we should re-examine how biased humans are i think if we can do what i mentioned earlier very good ai engineering education try to balance our data be conscientious to the problem have good tools i would argue that the ai at that point in time is going to be dramatically fairer than most people because people can't remove their bias and ai actually can remove most of it by balancing its data most people have are biased and don't even know it or don't admit it uh you know as an example in israel israel they did a study that judges a set of judges gave much tougher sentences before lunch than after lunch just because they were hungry so so that level of bias and irrationality and being overrun by our emotions is something that's um i think quite terrible with humans and i i do think if we do the things i mentioned we will already reach a good state that should be better than a great majority of the people it's interesting that you talk about humans and bias what about the concerns that have been raised that the great powers within ai if you imagine them in the hands of bad actors whether they be terrorist organizations or even more particularly actually governments states facial recognition you write about that the inference can be made by ai techniques about someone's sexual orientation for example possibly their religious background how can the public really trust that in the hands of bad actors and governments and states this material will not simply be misused well if there are people with a male intention who want to do harm to society that is the biggest danger behind this and every other technology right those bad actors could have used um electricity uh or the internet to do very bad things and those would have to be basically uh uh laws to protect people and enforcement that is um sets an example and create some deterrence enough punishment to deter people from doing that there's nothing else we can do uh in fact in the ai 2041 uh probably the the most negative story was one called quantum genocide and that is a uh a a terrorist who's like una bomber in the us and he just wants to kill all the elites and he in fact trains a a large number of drones which are very very small by 2041 they can be very tiny very tiny drones that just have a little bit of dynamite in each one and that has facial recognition gps tagging and and and tracking of people and when you fly these 10 000 drones out they will identify and shoot and to kill in point-blank range and that power of building autonomous weapons that can kill massive number of people in a particular type in this case the elites or gender or race uh to do it in a massive way and without leaving traces or leaving yourself immediately vulnerable like a you know if you were to tie a bomb and go bomb a a school or church or something you would be killing a lot of people doing terrible things but you would kill yourself too when you have that autonomous weapon you could potentially get away with murder without adding any more risk so i do think in this kind of area is where there needs to be international collab cooperation uh whether it's by treaties or ban uh to really uh ensure that these non-state actors or terrorist groups aren't able to get away with this whether it's through the sales of equipment or the occupation of airspace or some other mechanism or very severe punishment if this doesn't get started soon we're going to see a lot more targeted assassinations that are enabled by these thousand dollar drones that can be more accurate and uh doesn't create uh vulnerability or risk for the for the bad guys before i go to questions let's finish on uh an up note your the ending to the book is really fascinating that you do leave us with this it's really i don't want to say utopian because that's not that's not the right use of the word but you around the issues of happiness and what you call plenitude which is the economics of scarcity could be solved by a i could you just talk us through this very optimistic ending to the book and why you have left why you leave the reader with that notion that although as you say of course bad actors terrorist groups states governments whoever they may be can do bad things why you feel that happiness and plenitude is actually is likely to be the outcome of well-used ai right actually i think the last three stories were alternate endings to the human race right yeah there's there's kind of one negative one and two positive ones yeah that roughly represents my belief in the odds so um the last chapter is based on some facts and some speculation the facts are that in 20 years we should be able to use ai automation robots to do most of the routine work that we do today thereby dramatically reducing the cost of labor in the production of goods at the same time the cost of materials will come down through advances in synthetic biology life sciences and also the cost of energy will come way down by the use of distributed um battery more advanced than lithium storage using solar as the primary and wind as a secondary maybe there's hydrogen and other things there's just so many things being worked on towards the goal of a carbon neutral that i feel it's um not at all outlandish to predict 90 reduction in the cost of energy in 20 years so if you think about the production of anything labor's reduced maybe eighty ninety percent materials some percent fifty percent let's say and the energy by ninety percent we are really reaching uh a level of uh low cost and uh uh plummeting of the prices of many most goods that we have today so if we were building a new earth from the beginning with that's zero based we surely can use this new materials new energy and new robots to create enough vegetation and farms and food and housing because all of that can be manufactured for all the people with enough surplus so i think that i believe that is factual uh whether we will get there i think is incredibly difficult because you're to me companies will find ways to trick us to pay more money and people want to accumulate wealth wealth inequality countries aren't equal so i actually don't think we can get there but um the the so i think the last chapter is meant for us to think if in 20 or maybe 30 years we could get to a state when we can feed everyone bring everyone out of poverty let everyone on earth living a life of comfort yet we're not likely to do that where are we screwing up in our economic system in our permitting human greed and selfishness and vanity to get ahead of doing something that should be fairly basic to provide sustenance and respectable living for all what are we doing wrong as a human race and that's really meant to to be the question asked it's framed in a little bit more positive light but um i'm not optimistic we can solve it in 20. i hope we can solve it in 40 or 60 years kai for you thank you so much it's a it is a fantastic ending to the book and very very much food for thought let's go to some of the questions that you've sent in thank you so much all the audience for joining us on this lunch time and and evening chatting in beijing lunchtime in london and joining us from around the world so kaifu let's try and rattle through these um as much as we can for for our audience um first question and you touch on this in the book with some detail what do you think about of the rise of crypto currencies and ai use in the financial system is this an area where you see big change uh i believe blockchain is a fundamentally valuable technology it can do a lot of things in the book i describe several uses one of which is to guarantee authenticity of all the video audio images captured by all the devices thereby allowing us to authenticate the courtroom evidence and data and also get rid of problems from deep fakes so that's kind of one use of authentication another example is that the cryptocurrency uh one of the key things in ensuring that our cryptocurrency stored safely in a distributed manner is that we have a strong encryption security system but uh we should be aware that with quantum computers some bad person can attack a certain percentage of the bitcoin wallets and take the money away and and and yet at the same time it is quantum computing that will reinvent the future of the ultimately safe encryption system so those were the key technology points that i point out uh with respect to today's cryptocurrency uh i'm afraid i still see a bit of a hype and participants whose intentions are not completely pure and it's still quite a bit speculative uh i'm quite disappointed that there's not enough people who are really working on hard technology-based solutions to solving the pow problem that is using up so much energy uh to protect the bitcoins and other cryptocurrencies and and not solving uh the energy problem which is uh environmentally problematic and and also arguably not generating a lot of value for the society i think if we solve that one then many uses of blockchain and cryptocurrency will become possible i hope more computer scientists will try to figure out a better solution than the current pow solution to prove authenticity i think it is should be possible to get a sizable increase uh but i just don't see enough serious efforts on this and that's a little disappointing to me thank you sarah asks um you mentioned personalization with apps uh like youtube is this something that should be regulated and if so how obviously you touched on this guy but i know if there's more more to add around that personalization specific that sarah asks around about yeah um i don't really know how to regulate that because it's well it because you know in the car oh now i i've lost uh kaifu i don't know if anyone else has i don't know if we're trying to re uh connect with him so just bear with us while we're just trying to fix uh the connection to beijing hello can you can you hear me now i feel you're back hello i'm back okay we just lost a little bit about uh regulation and youtube we just lost you for 30 seconds there yeah right right uh yeah i think it's very it's a double-edged sword uh the personalization is what makes a lot of good content i i don't think we should let the last 40 minutes make us think everything from companies are terrible most of it is quite good it's the things that i want but sometimes it can do bad things too so in this case i don't i don't really i can't think of any ways right now how to only deliver good from bad because that is so hard to to to separate and if you get rid of personalization you get rid of the power of ai so i would say you know don't throw out the baby with the bathwater figure out some way to salvage the situation for now if this really angry angers you then and then don't use the services don't don't watch anything recommended for now i'm pretty sure it gives give the technologist and the business community five to ten years uh something something that's uh much better than today will uh will come up with something like that thank you another question do you think initiatives like the eu's gdpr are good protection for consumers or does it just make european companies less competitive the answer is yes it makes them less competitive yes but it does many things first i think it's very good in that it raised the whole world's awareness that people care about personal information and privacy and something needs to be done about it and i think it's set a gold standard that the world is actually emulating now so i think that's good um but i don't think it's all that effective uh you know because when i go to all these websites windows pop up and i just click yes yes yes yes so it's kind of transferred responsibilities to me without me not knowing with me not knowing what i signed up for so it hasn't really achieved that purpose i think it sort of pushed the websites and the apps to say i better be careful otherwise the big fine is coming that part is good and the way it's implemented i'm not sure it's all that protecting because individual users either don't have the patience time or technology background to judge what part of the data to send to what websites so maybe an ai will be needed one day to to judge for me because i i don't think if i can't do that i think a lot of other people also can't can't do it and i think if the other thing is that i think generally speaking not gdpr per se but generally speaking the eu approach has the following downside i guess uk follows it so i should say eu and the uk approach has the following downside i think it asks too much to have a human in the loop because human in the loop causes ai to be ineffective the whole point of aii is not to have a human in the loop and and that's that becomes a problem and also i think having um uh i think the the wish that let's all i think the idea that we own our data is absolutely correct but let's the actual action or hope that all the internets and apps internet companies will take all the data ship it back to us and won't store it anymore that's something that just won't and can't happen i think it's just too optimistic to hope for that and if that even if that were to be possible the ai will no longer be that effective so i don't think that's a good idea and i think the regulators also have a strong belief that it is a mechanism gdpr and other regulations are are american are set of mechanisms intended to slow down ai and i think that i can see why you feel that way but if you feel that way and implement it and enforce it you will be behind countries that don't have policies that are designed to slow down ai in fact u.s and china are trying to accelerate ai and and that would hold back eu on the uh realizing the commercial potential for the uh ai companies and industries and kaifu given what you've just said it would appear sensible under for you that the uk withdrew from gdpr now that it is outside the european union and went to more what might be described as the american route i actually think the american uh i mean i i would think the american policies are ought to be culturally acceptable to the uk i haven't studied the details of the differences but i know europe your eu is tougher u.s is lighter and and i don't know enough difference between the culture uk and u.s but i would say for if you want to study to help accelerate your businesses uh without going all the way to abandoning this direction then it's reasonable to consult the american regulations to see if they're a better fit thank you climate change one of the biggest issues of course or the biggest issue facing the planet a question about that do you have insight on how a.i will impact climate change does it have the answers and you have a question mark and two exclamation marks so i'm not sure that the uh the audience member thinks it will but it's an interesting and exciting question yes i think ai can make significant contributions and anything with data is something ai can digest and propose solutions today we probably don't in understand the cause and effect of all the climate change and i think ai can help us figure out what are the problem areas and and with people symbiotically find solutions i think that is an effort that that should be done ai can also be used on more specific level uh we invested in a company that builds an ai based simulator for new materials used for batteries and then it simulates the battery and uh and so so it can improve uh the energy storage by ins uh using ai to emulate it in software and accelerates the discovery process it's it's uh i think you know climate change is like a human body if you will like our human body is a microcosm of climate change we're if we can use ai to to diagnose human illnesses and improve them and to invent drugs for humans it would seem uh the climate change is just the more complex version of that that that people should definitely work on but i don't think it's a short-term proposition you know healthcare ai or probably take another five years to take off climate change will also take some time and we don't have any more time to waste so i hope more countries and and more companies will uh will will look at ways in which that can be can be integrated you touch on healthcare interestingly and there's a lot of material in your book about healthcare but of course we've gone through or are still going through um the the covid emergency the disaster for many millions of people around the world and um sarah asks we are still in a pandemic can ai help us with developing faster vaccines for new variants on covid19 or other diseases that may emerge in the future ai based drug discovery is absolutely one of the most exciting and near-term possibilities alpha fold is a huge step in being able to fold proteins we have invested in a company called in silicon medicine that has extended that by looking for targets within pathogens and then finding small molecules that will fit into the target effectively finding the drug that's once the protein has been folded so combining these discoveries together we can significantly accelerate drug um discovery and you still need a human to watch the process and select from the candidates so we've demonstrated within silicon medicine that it can actually get two uh two drugs in a clinical trial at this stage uh and we're already in human trials so it's a proven technology so i think this will have the effect of reducing the cost of discovering a drug by as much as a factor of 10. and when that happens more rare diseases will be treatable the speed of discovery can be as much as one-third uh the total length not counting the clinical trials so it will have faster lower cost drugs drugs treating potentially rare diseases so all of that will be great for humanity the one part of the question that i'm not sure is whether it would apply equally to vaccines i haven't studied that i haven't seen any work to demonstrate it uh there are ai tools that will help drug discovery but not to the extent that i talked about my speculation is there's probably not enough time for ai to make a big contribution to vaccines in the next year or two which may be the most critical time for vaccines so probably not this time around but um but certainly when and if there's another pandemic i would think um ai based drug discovery ought to carry over to vaccines thank you two minutes to go in this wonderful session that we've had with you kai food could i ask you just for very quick answers to the two questions that are left from our audiences so firstly what effect on unemployment will a i have and specifically do you believe in universal basic income or some or another initiative to soften the blow of displacement uh yes i i think ai be given so good at optimizing quantitative routine tasks will displace a lot of jobs uh probably 40 to 50 percent of the existing jobs are rooting in nature and poised for uh and vulnerable to ai displacement over the next 15 or so years so next 15 years will be a tough time we are watching in the companies we invest in that ai is now doing customer service now it's doing telemarketing it's doing some types of assembly line it's doing forklifts so we're watching this um really unfold before our eyes so it is really happening i do believe ai will also create many new jobs and for example data labeling and ai programmers and robotic robot repair and so on and and also many jobs that we can't anticipate today so over a long period of time let's say 30 30 years i think ai will create more jobs than it displaces but in the 15-year time frame more displacement than creation that would be my prediction in my book i talk a lot in a lot more detail about about how this comes about and what are things we could do i believe universal basic income is an approach that can be helpful it addresses the wealth in the quality side it taxes the rich to give the poor a a buffer and i think that's useful but uh keep remind remember though it's not just about not having enough money it's about how to redirect some of that money to retrain and gain the skills that will not make someone who's lost the job once to lose another job so it's about having spending it on retraining investing in oneself and giving the guidance on what what to learn what to retrain so that is arguably even more important than the universal basic income and also lastly the universal basic income should not become a crutch uh something that people can say okay i no longer have to work it there needs to be some motivation for people to have the drive to move forward again so maybe better to cover uh health shelter living expenses and and things like that rather than just give people money because we have seen that historically when people are in destitute positions and given a lot of money they may just spend it on alcohol and drugs and not really use it for self-enrichment which they badly need final question and thanks for your patience uh typhoon thanks audience for hanging on just past the hour mark but which companies do you think are best positioned to take advantage of ai and that's a question from diana company's most advantaged uh to do that yeah well what whichever company has the most data and i would qualify that by saying structured data that has accurate labels that connects to a business metric because if you possess that then you can create a closed loop in which your data will improve your business metrics and then to become more specific the companies that have already benefited are the internet companies the amazon google facebook tests and alibaba by dance and so on uh the next set of companies will be companies that are traditional but have data nevertheless the most prominent of which will be financial companies banks insurance companies investment companies um and then i think there will be also opportunities for disruptors some industries are poised for disruption and sai scan ai learns to see and read and understand and machine translate and transportation i think all of these are poised for potential disruptions but in the in the near term whoever has the data has the biggest advantage in the long term you can take advantage of ai by either being a disruptor using it to create to do something that nobody can do and change an entire industry or an industry early adopter that embraces ai combined with your traditional industry to be a step ahead of the competition so it's a huge opportunity kind of like whoever embraced the internet back 20 years ago got a huge huge leg up on the competition just finally talking about employment in 2041 will an ai human robot be able to do this interview better than me uh no i think i have to say no but actually just uh just to be polite no seriously no i don't think so i i do think they can do a reasonable job i think they can do a reasonable job but just by training it on all the interviews they should be able to do a decent job but i think some some parts will be hard you know humor and and subtlety will be difficult but just um asking questions and um doing follow-ups those are things that uh ai will will be able to do uh but probably the part that's hardest to emulate is that um as as a good interviewer you have your fan base and you have people who want a human in in the process and they want to know your background they want to get to know you so even if an ai approximates your capabilities i'm sure uh intelligence square was still much preferred to have you kaifu thank you very much very good line for my cv kaifu says i should have a job in 20 years still thank you for that kaifu what a wonderful book what a wonderful hour we have spent with you thank you so much for your time thank you to the intelligence squared audience to intelligence squared themselves for hosting such fascinating debates do go online and have a look at what's coming up as well as have the opportunity to discount to buy kaifu's

Channel: Intelligence Squared

Views: 23,316

Keywords: Intelligence squared, intelligence squared debates, artificial intelligence, ai, kai fu lee, kamal ahmed, chee ufan, ai 2041, china, machine learning, technology, robots

Id: 0tCwkSXQzO0

Length: 63min 16sec (3796 seconds)

Published: Thu Oct 07 2021