China leads the U.S. in the race to implement artificial intelligence
![China leads the U.S. in the race to implement artificial intelligence China leads the U.S. in the race to implement artificial intelligence](/pic/china_leads_the_u_s_in_the_race_to_implement_artificial_intelligence/RllJVlg1c0ZlWlk_.jpeg)
You. Hi. Everyone. Thank. You guys so much for coming today my name is Michelle Lee I am a second, year PhD student at the Stanford, AI lab and I'm one of the organizers. Of the AI salon, and today's, salon moderator. To. Just give a little bit of background on, AI salon. ASL. On was started four years ago by dr. Faye Faye Lee it is, one of the most beloved and iconic, student-faculty. Events, at the Stanford, Stanford AI lab at. AI salon we have had. Many speakers. Such, as Jitendra Malik Elon Musk Young Laocoon and Jenson Huang and we talked about big picture topics, and AI that, goes beyond just talking about code, and algorithms. Today. We have three exceptional, speakers, for the salon, we. Have dr. Kai Fuli, we, have Professor Susan a fee and we, have Eric a professor, Eric Bernie Olsen who, are here to discuss on an incredibly important, and pertinent topic, AI, in, the future of work but. Before we start a salon we, want to welcome dr., Kiley we'll be introducing his book AI superpowers. Which, is New York Times Wall Street Journal and, USA, Today, bestseller. Dr.. Kevin Lee is a venture capitalist, technology. Executive, writer, an artificial, intelligence, expert. He, is the founder of sin Ovation ventures a leading VC firm focusing, on developing, Chinese high-tech companies prior. To founding sin Ovation ventures lead was the vice president of Google, founding. President of Google China and founding, director of Microsoft, Research Asia he, was named one of the hundred, most influential, people in the world by Time magazine, Lee. Earned his PhD scmu, in 1988. Where he developed one of the first continuous, speech, recognition. System, he, has authored ten US, patents, in over a hundred journal, and conference papers and is the author of AI superpowers. Let's. Welcome dr. Kevin Lee to the stage. Thank. You it's, great to be here with such a distinguished. People in the audience but. I thought I would bring even, someone, more distinguished, to, say, something about AI. It's. A great thing to build, a better world with artificial, intelligence. You're. Going to knock Jones a guy be insured yeah. I'm. Not sure if he came himself he, would get that much applause but. But. That, wasn't President Trump talking that, was speech. Synthesis, system, built, using deep learning to, train. His. Voice and it's. A Chinese company, called I fly tech so, I think in one example we see the power of machine, learning and also, the, progress. That China has made so. Most, of you probably know AI very well so I won't go into any description. Of technology, but. I would just say for if there are a few people in the audience who are not familiar. Machine. Learning is the core. Part of AI and deep, learning is the most advanced, technology, that is working and, think. Of it as in, a single domain when, you have a huge amount of training data it, can do things much, much better than people but. Only in a single domain with objective, functions, so. That's. All I'm gonna say about technology, but, I want to talk a bit about applications. Because. We are a venture capitalist, firm we've been thinking about what are the areas to invest in and from our perspective there. Are really four waves of a I'm now, given AI, requires, a huge amount of training data the, first natural place is of course internet applications where.
We Are all. Generators. Of data and free. Guinea-pigs. Labeling. For the likes of Amazon Facebook, and, Google, every day every, time you make. A click buy, something, you, are create data that, makes the system smarter, and able, to guess what you might want to buy next, time and send, the right ads to you and what's, more important, is that AI. Gives, very powerful, knobs to the CEOs of these companies so. That they. Can optimize. User, minutes. Per user and that's. How Facebook got into trouble actually, by. Optimizing, towards, one goal, or. You, can maximize, revenue. Per, day or per month or you can maximize profit. So, each. Of the different objective. Function will, cause the company to display, different ads. And, products, and choices to you in order to maximize, that function, and that I think is the big power of AI and that's, why pretty. Much all the, giant. AI companies, are first. Internet companies that, includes, the Giants, here Amazon. Facebook Google. Microsoft but. Also the Chinese giant's, Alibaba. Tencent, and Baidu, they. Have a huge, amount of data and they use it to extract value and make money so. That's kind, of the first wave the, second wave you might ask is who else has a lot of data and those. Are going to be companies. That, used to consider. The, storage, of data and the data center as a cost center and the, requirement. To store the data as, as. A, something. You you, have to do in order for archival, purposes let's. Say a bank used, to have store all the customers transactions, because you don't know when a customer, what might want, it want to see it but, now with AI all, of, those transactions. And data becomes. Mountains. Of gold that. Can. Be built used to, do things like customers. Of a loan, determination. Credit, card fraud detection. Asset. Allocation, and also deciding. What, product. To try to sell to each customer, and estimate. Each customers net worth and so on and. That's what one of our investments, for our fourth, paradigm does and that of course is not just for banks but also insurance, companies, on. Financial. Investors, pretty. Much any company, that has a large amount of data to. Give you an idea the. Why, AI is so powerful, I'll pick, one example, in. A company that we invested, in called, smart finance what. They do is loans, so basically, micro loans imagine. Along. With five hundred dollars for. Something.
Like Six, months and at. Let's say credit card level rates, but. It's done through an app so, all you have to do is download, the app fill. Out the usual things you'll fill out with loan applications. Your, name, address, where you work how much you make rent. By place, and things like that but, also it, asks for your consent, to send, information from, your phone at. The same level other apps, send. Information to, the. Likes of Facebook and. Amazon. And, snapchat, and it, takes all that information, into. Actually. A deep, learning, function. That, the determines. Whether to lend. The money to you so now could you imagine going, outside Stanford, with. $500,000. For those of you who have it and, for. The first 3,000. People that comes to you you pick you picked 1,000 of the, 3,000, and hand, each person, five hundred dollars so your $500,000. Is gone what, do you think the default rate will be eighty. Percent ninety percent what's. The likelihood a stranger, off. The street will pay you the money back very low right but, the default rate. For, this app is, three percent so. How do we manage that we, manage that because of all the information that comes in that no human, loan, officer, could possibly consider, so, it would have things like, your. Name your address there's. A match on, the internet how long did it take for you to type your address if, it takes too long maybe you were making, it up or maybe you're copying it off something right it would also, have. Your contact list which is submitted. Like. Just like you do too Facebook and snapchat and the, contact list can, be verified well. Who's, the person you call mom it is that person in fact your mom and, things, like that can be double-checked and also. What apps you have installed do. You have a lot of gaming apps gambling. Apps or. Serious. Apps and. That all play, a role and also. A kind of a phone you have what's the model of the phone is it a newest, iPhone or, some old old, phone, and, we. And what, day of the week is it what day of the month is it why, is that important, well is it before payday or is it after payday you're on the borrow the money before payday very reasonable. Borrow the money after you've just been paid that. May be a negative. Signal and. Then just for fun we went to see the 3,000, features that, were extracted, and the least important, feature turns. Out to be the. Battery level. Why. Does that matter well. If you think about it if. You have OCD and charge, your phone all the time it's, probably a little bit correlated, with someone who tends to repay. Right and if you are kind of irresponsible. Letting your battery run out well maybe. That's a little bit correlated, with someone who defaults, of course, this is a very unimportant. Feature but, it has some contribution, nevertheless so. No human could possibly scan. And combine, these 3,000, features you're, probably wondering, well how did they train the system, well, of course based on actual, outcome whether, the person returned the loan or not so, now you're starting to figure, out wait a minute and so when, you have no data they, have no data to begin with well that's where venture capitalists, start for and, we. Give them the money they lose it at the 20 percent default rate they, come back for more money our default rates now down to 14 percent can, we have another 30 million okay and, had kept going it, kept going until they got to about 7 percent at. Which point they could just. Borrow. Money from banks and be, assured that they're going to make more money from that without using, VC, money so. You see how terrible. Our business, is so, that's an interesting example and, the third is the third wave is. What we call a perception, and that is. Essentially. Digitizing, the physical world with. Cameras. Sensors. Microphones. And so on. An. Example. Of that, examples, of that include Amazon.
Echo. Includes. The. Autonomous, stores, and. Of course it includes the, very, controversial, application. Face recognition just. In the headlines, today I'll. Use that as an example not, to endorse, the, use of face recognition but, just as an example of how powerful it can be, recently. There is a very famous Chinese singer. Jacky, Cheung any. Of you know John. Shea oh ok very famous, I see some older people here so you you know you would know him very famous, singer so, he was giving, concerts in China he gave I, think four concerts and then, after that he got a nickname policeman. Cheung because. For. The safety of those stadiums, in which he, gave the concerts, as. People went in there. Was face recognition and, the, face recognition was, connected, to the most wanted criminal, list and then. People were apprehended, when suspicious. People. Recognitions. Were made and then, in, maybe, 70%. Of cases it, was a false alarm so, there sorry here's. Your ID you're not a criminal but. Then 30% of cases people, were actually from, the most wanted list so, imagine, just just on the technical level how. This could possibly be done without. AI in, Mac with any police might be able to remember a hundred thousand, faces of course not so now, we can see these, applications. Are not just a human, level or not but there can. Be dramatically, better than people and, then the fourth layer is what, I call autonomous, AI and that's basically, robots. And, autonomous. Vehicles. We. Have made a number of investments in, this area including. Robots. That pick, fruits, robots. That wash dishes do you want one, if. You want it's right right out. We can take orders later it's some only. $300,000. Each. There's. These robots, actually you you would put everything off the table, into the machine and it. Actually put separates, them into piles and cleans them so it's not a dishwasher, so how, can they sell any at three hundred thousand well everything comes down right, with volume, cost will come down so eventually you can you can get one and. So. Robotics. In in. The use of autonomous stores, so, we have an investment in China called f5 future store that. Is an autonomous, fast-food. You, can have a bowl of beef. Noodles of, very. Authentic. Cantonese. Style for. About a dollar fifty so. That's going to give. McDonald's. Run for it for its money because, it's much cheaper and, it's completely, autonomous. No no humans at all and autonomous. Convenience, stores and so on and finally there's the autonomous, vehicles. Which we we all know a lot about so, I won't won't go into that so, these are the waves I think, will really. Revolutionize. Almost. Every imaginable, industry. Well. There are only be four waves surely, not if, you ask me 20 years ago what, are the ways of the internet I might. Have told you there. Are waves related. To websites, and.
Browsers. And search engines, or, something like that but. We. Would not be able to predict all the other things that have that happen later. For example. Share. A sharing, economy ecommerce. Social. Mobile, so those are many more waves of the internet and AI I think, will be similar. To that. So. Again. AI is really, about, a single, a large amount of data in a single domain with, label and then, with, a fair, amount of compute and some experts that's these, are the magic ingredients. That makes AI work and. All. The famous, scientists, are Americans. And. Canadians. So, you would think us. Is by far the leader in the well, in research that is the case but, in implementation, I would, argue not, and I'll. Show you why, that is so. This book I'm. Sorry this this another. One shows the h-index, of, the. Top 1000, researchers, u.s. has 68, percent and then. China is only six percent so again demonstrating. Us is well ahead but, these are the three issues that one has to consider about. AI implementation. Application. And monetization. First. Is well how many breakthroughs have there been people. Ask me what the y-axis, is I said I made, it up, this. Basically shows the magnitude, of the, various, the magnitude. And the importance, of various. Types of innovations. And people, can draw their own chart, but the idea here is that there. Has only been one single. Big, breakthrough, and that, was nine, years ago so, it is probably. A big. Question. Whether there will be a lot more breakthroughs, in the next decade and if. There are no big, breakthroughs, it's hard for us to, retain. Its leadership because. The. AI technologies. Are reasonably. Well understood because. Of all. The open. Source and the, sharing, and people, published online. Immediately. So there's not even a latency in the journal. Papers so, all. Countries, are more or less in, terms of implementation on, an, even, playing field, and. Given, that I would argue that we, are now not, no longer in the early, phase of of. AI. In. Entrepreneurship. Where. Expertise. Is the king but, today we're more in implementation. So it's a question of who can build, faster, and run, faster, and who has more data so. So. The on the last point I, would, argue for many applications you, really don't need these super AI experts, young. AI engineers, will suffice. Especially. In waves 1 & 2 so, given, this. As oh as an example we, run a training camp for. Form. For, AI every, every, summer last, year we had 300, sorry, this year we had 300 students next year we'll do a thousand, and these, are largely undergraduate. Students, who have had maybe one course in AI and then, we basically have, an industry. Project. Leader that. After. One week of core courses that we teach, we. Have basically. Industry, people from autonomous, vehicle companies speech. Company's, vision companies, lead, teams of eight to, build projects. And actually. Just in one week of lecture four weeks of, implementation. They're, able to build, things like in this case it's an autonomous, vehicle it's. A toy car but, it has a real camera it, mapped, the campus, of Beijing University and, is. Able to drive. By itself from, any building to another building and this was done by eight students. So, it goes to show you that the. Barrier. Of entry is really not that high anymore so. A little bit more detail about China's. Position first. There. Are a lot of young, Chinese. AI. Engineers. People, are rushing into it the upper right shows you the picture on the. Bottom is a lecture. That I give and. A little more people than this room and and. Then the, left side we see the. Number. Of articles written, in AI at all actually Chinese authors, are about 40% so, much larger than even China as a population. So, it's just that they haven't become, famous. Yet. They're entering at the bottom of the pyramid and they're growing so, I, think to, the extent that, AI. Implementation. Just requires young hard-working. Engineers China has them secondly. Chinese companies are innovative. So, this slide shows China.
Began As a copycat this, was only ten years ago on the on the leftmost side. Basically. Every Chinese company, was, the. Google, of China the Amazon of China the Yahoo of China and so on but, very quickly China, began to, learn the product market fit see, in Silicon Valley I think people really from upon, copying, not talking about IP theft I'm talking about. What, Facebook, did to snapchat for example okay and, that happens, everywhere, in in China and. Actually. We, learn a lot of things by copying that we not learn music, and art by first copying, and then, a small percentage, become, innovative and of, course if you're forever a copycat you'll get nowhere so, in the second. Phase you see that the Chinese, companies. Actually, are as good in some cases better than the American companies the. Ones that inspired them for example we chat is better, than what's up. Many. Of you probably use both of them and and. You wouldn't you would know and Weibo. Is better than Twitter at, least as a product, maybe not in the diversity. Of content. That's. A different story, and. Then in the the third phase actually these are really, innovative, applications, all invented. In China I can't. Even begin to explain each, of them maybe. I'll try one there, are quite, show and the tic-tock are the, top, apps in. China together. They, have, 220. Million daily, active users. This is a phenomenal. Number and what they are is, basically. Video. Based, social network, something. That doesn't exist at all in the US in fact there was a review of tik-tok, and. I think it was in Wall Street Journal from a few days ago so feel, free to check it out and the total number that. There, are seven I think there are seven or eight apps here and a, total valuation of, these brand new Chinese, apps is, about, 300 billion dollars and. This. Goes to show you that China is becoming, innovative. And. Today, we really have a parallel universe with, us. Basically. Running. On these apps and China running on those apps and these. Are parallel, universes so a lot of people keep asking me can Google go back to China and succeed well it's very hard to traverse, the, parallel, universe. It's just like a Chinese company probably cannot come here and succeed so, the, Chinese, apps are really every, bit as good and the. China also has great entrepreneurs, and many of them are going to go into AI I won't go into details, here but I'll just say that the China. Approach, to building, a company is very different, than the Silicon Valley approach, here. I think is about vision changing. The world technology. Centric, can, make. It light non, capital, intensive, China, is almost the opposite it's just building it fast. Tenaciously. Iterate, quickly and, execute. And the. One very unique aspect of Chinese entrepreneurship. Is that the Chinese entrepreneurs. Because. There's so many copycats around, the. Only way you can win is. Winner-take-all. Not, only do you have to win everyone, else has to lose but. Also on top of that you. Have have, to build a business model that is uncopyable, otherwise, someone will take it so, how do you do that well, you make something really incredibly. Complex, hairy, expensive. That's how you do it so mate wine is an example in the middle is the. GrubHub of of China or the Yelp of China but, what they do is they deliver, food, to every. Person. In in. A city within. 30 minutes for, the costs, of at, 70. Cents per delivery and that. Has changed the way the Chinese people eat but, how do you do that they have six, hundred thousand. People essentially. On uber, imagine. They're on the uber like network so, and and there's reverse search pricing, inviting. Them to deliver, and also. There are very cheap electrical. Mopeds. That, have to have batteries replaced so, they have to build a giant algorithm, and bring.
In And train 600,000. People pay them minimally and of, course many will turn over and leave you have to bring more in so, that, is the complexity, of making. This 600,000. Over like delivery. Network that makes their business model impregnable. So the Chinese entrepreneurs. Are really good, at this and. Of, course the Chinese. A lot, of the money is flowing into AI even, more than the US last, year, 2017. 48. Percent of the world's venture, capital went into China AI for. AI and 38, percent in, in, the u.s. the, equivalent, company is actually, rising. Faster, in market, valuation, in China, compared, to here this. Example is I fly tech and newest and, just. Sign of Asian ventures alone our investments. We already have made five, unicorns, in AI and the. Youngest, of these companies is only less, than two years old so, it's very very fast and they really, execute, and deliver. And, of course the, amount of data is very important, and. The. Right is a generic, graph that shows the more data you have no matter what algorithm is, also reasonable, more, data gives you better results so. In, the era of AI if, data, is the new oil then. China is the new OPEC. Used. To say Saudi Arabia, but not. A good analogy. Anymore. And. Why. This trying to have more data a lot of people say oh it's because there's no privacy there everybody gets exchanges, data that is not true the, Chinese companies, get data the same way as the American companies think of them as getting, data Facebook, like Facebook, and Google, China. Has more data for two reasons one is breath, just more people and one, is depth because, the usage is much, deeper the Chinese users using. The example I gave earlier with Mae to an order. Takeout. Fifth. Ten times more than in the US shared. Bicycle, three hundred times more than the US and every usage is a data, point that's used to train AI and the, most important, is of course the mobile payments. In. China, mobile, payment is used 50, times more in the US and some. Might say oh yeah we've got Apple pay just takes some time but that is not true because these. Are mobile payments, unattached, to credit cards Apple. Pay PayPal, still, largely, connect, to credit cards and as, long as you connect to credit cards. It's basically, tack in the economy, at 2%, or. Something like that so, in China the, use, of mobile payment, has become so convenient, that you. See here in the farmers market and, my wife last month saw a beggar on in. Beijing and he was holding up a sign as I am hungry, scam me so, I. Would, never joke about that this is seriously, because, nobody. Has changed. No. Cash no credit cards, and all, of that becomes data used, to train to train the AI it, also will contribute to make China go, from a spending. Savings. Economy, to a spending economy, also makes entrepreneurship. A lot easier because you can monetize. Users from day one you don't have to wait until you have you, know a million. Users so, this is another huge benefit, and of, course lastly, Chinese. Government, strongly, supports, AI but. Unlike what most people read China. The government came in a little bit late all, the unicorns. Were made without. Any government, support, they became unicorns, on their. Own capabilities. With, private capital, however, once, the government saw the importance. Of AI I think, there are a couple of really important, things that Chinese government, does one. Is a techno, utilitarian. Policy, what, that means is let, the technology be. Implemented. And see, how it goes if there problems, then regulated, so. That is how the, mobile. Payment became. Pervasive, imagine. In. The u.s. if, if. You. Know Facebook. Announced we're going to have a new payment, method I think, immediately Visa, MasterCard. Will complain and say oh software, companies they're. Not reliable, there could be hacks and David could leak and fraud, and all that stuff right and then there, might be regulation. Or. Sasori or checks to check to slow them down but in China when Tenzin and Alibaba, were proven competent. They are allowed to go forward, of course technique utilitarian. Doesn't, mean everything, is allowed cryptocurrency. For example is not legal, in China. And of course the state document. It really sets the tone about the importance, of of, AI. That. Actually has no budget associated with it the budget is determined, by each, reader. Of the document, for example in. One of our investments, in AI for banking, after this. Document, came out the banks were much more open to buying, AI software, so that was helpful to us in, the city of Nanjing that. City government, said well we have great universities, so let's do an AI park and China, is building a new, highway in Georgian province, with, sensors, to, help the safety of autonomous, vehicles and the new city called. Xuan which is the size of Chicago is being built, with, two layers top. Layer is not not the whole city to downtown the top layer is for pedestrians and, pets, and bicycles, and the, bottom layer is for, cars, and that, very. Expensively. But nicely avoids, the kind of problem that we saw in Phoenix, with uber, autonomous, because, the largest problem, of autonomous, vehicle is when you hit a human and.
That's The highest likelihood of casualty, by separating, the flow you. Eliminated, that possibility also the cars driving underneath, will. Have controlled, lighting so, also avoiding, problems that we saw in Tesla autopilot because you have a fixed lighting in the b1. Level of roads so, that kind. Of huge infrastructure. Will spend I think will accelerate our, China's, development of AI. So. Where. Does China stand in terms of AI implementation. Again this is not research this is implementation, I think China came from way behind to, a little roughly, caught up to. Probably a little bit ahead going. Forward and this, is my my projection. However. I do want to say this is not a zero-sum game because. Chinese companies, right. Now only sell to Chinese customers, so, their success, do not come at the expense of, an American, company but. People, want to know where things stand, and this is my, my, estimates, so. With two. Engines now driving AI, forward, not one engine and all, so lots of money pouring into it and, Giants. Training the experts, and lots, of open source. And cloud. Technologies, AI will create a huge amount of value PwC. Estimates, about 16 trillion dollars. By in next 11 years net. Increment, to the GDP McKenzie. Estimates, about 13 trillion, so this is the size of the GDP of China plus, India, it was huge and this. Will do a lot of things, make. A lot of money for people it, could be used to reduce, hunger and poverty but there are also a lot of problems. In AI I know, we're gonna have a discussion on this so, I'll just talk. More quickly over this part a lot, of the top issues are discussed a lot in the US the, future of work is something that, maybe not discussed, as much so, I want to spend just a few minutes on that. Because. AI is. Single. Domain lots, of data and it, does superhuman. Capabilities, so, what that means is the types of jobs that, can, displace will, increase. Over time because, many jobs are routine. An hour repetitive. And those, jobs with, improvements, of AI algorithms, will, be better done by the Machine. Where. Some, of us are still safe, you. Know professors, researchers. Are. Creative, and CEOs. Deal with complex, problems so. AI single, domain and non creative, so there's some safety there, this. Is happening at white collar and blue collar I personally believe white, collar routine, jobs will come we'll be hurt first because, that's just software, robotics. Don't yet have the dexterity of, you, know putting an iPhone together and probably won't for a long time so. Here you see examples in the white collar city has announced half, of their operational. Back-office will be replaced, by automation here. You see the example of the fast food I was telling you about that, is basically. They're, not going to displace any jobs one on one but, if they have 50% market share then, McDonald's, and KFC well we'll have a in force and then the last one shows you a basic. Autonomous. Cashier. That, we have that is in. Use in Beijing if you visit, this pastry shop you, pay yourself, just by with, computer vision scanning, and it's only like two thousand dollars, one-one, cashier, displacement, so, this is really, happening very quickly we. Should be worried about large number of jobs being displaced, there. Are many, people who would argue hey. Every technology revolution. Creates more jobs than, it this, disrupts, which, may be the case with AI as well but, the problem is we, don't know what those jobs are nor, do we know when they'll be created, for, example when internet was started, no one had any idea so, many uber, jobs would become available and, it was impossible, to predict so same with AI but, I would say you. Know Ober Jobs didn't get created, until 20, years after the internet invention, we can't wait that long also. AI. Any, job. Ai creates, will tend to be non routine jobs because, if it were a routine, job then AI would just do it itself, so, even, if the total number of jobs remains. Constant or even increases, the, people, who are displaced let's, say 40%, of the workforce, over the next 20 years they. Will not be able to, easily. Transition, into, from an on route from a routine job into, a non routine job so, so. What jobs would, can be offered it's going to be a significant, challenge, one.
Thing I believe. Gives us hope is that if we think about what, AI cannot do one. Is that it cannot create as we mentioned, it's not not creative, not. Strategic. And the, other is that it has no empathy or, love and has. No compassion, and there, are many jobs that we don't want a robot. To do that we want to interact with a human so, if we draw these two dimensions. Changing. It from a two dimension, one dimensional, to a two dimensional picture x axis. Showing creativity. And why showing, compassion, or empathy. Then. We can move these jobs around and, will see actually, there are many jobs in the upper, left that will. Become possible, for. People, who, may lose the jobs on the lower left to shift into as an, example in, the area of health care us, will need 2.3. Million people. In. Additional. In health care and that includes, nursing, and. Home. Care elderly. Care and so on in, the next six years and those, jobs aren't easily getting filled because, the, job of an elderly, care. Taker is about, $19, an hour and it's, pays about, half of them less, than half of a truck driver or heavy, machine operator, yet. The, heavy. Machine operators, job may disappear, but, the elderly care taker will, actually need more and more as. We, live longer our, longevity, increases. People. Over 80 need. Five. Times as much care as people, between 60 and 80 and that will create an opportunity for, more jobs and we can even imagine many, jobs that are not being, paid today for example someone. Who homeschools. His, or her children. Or hot, line volunteer, elderly, companion, these, are jobs that may be unpaid, or voluntary, one, could imagine that if people from the lower-left were to lose the jobs these, could become paid jobs I think this approach would be much better than universal. Basic income which, gives money to everyone, whether they need it or not but. This will give targeted, training for people to move into more, compassionate. Jobs a very, good example is, Amazon. Which announced, this April that it will offer, $12,000. Reimbursement. For taking, classes. Up to four years so so. For. Taking classes, into new professions. That are, in. Big need or less. Likely to be displaced, by AI you've, got to think Jeff Bezos thinking. About, his. Whole. Food cashiers, thinking. About his warehouse, Pickers. You, know how amazon, has Kiva moved large. Things. And then the people actually, does the peak well, we all know from our robotic, professors, over the next five years a lot, of that will be done by by. By by autumn play become automated so, this is actually very interesting. And generous, move by Amazon, and. If. Many. Of you probably don't. Know but the average salary. At Facebook, and Google is about two hundred thousand dollars and at, Amazon is $28,000, sorry the median the median, 28,000. So offering, $12,000. Of reimbursable, training, on the, $28,000, salary that's a pretty generous, step and Jeff. Bezos has said that he would like to, feel. His responsibility. Isn't just with shareholders but, also with employees, by ensuring that they are still, reemployed, employable. Even if they're displaced, even, if the job isn't available in. Amazon the, types of classes he offers are. Aeronautics. Repair, something. That can can be understandably. Trained if someone took the class who was a warehouse, picker, or a nurse someone. Who could probably also be trained if if, he or she was a cashier, so, these are really. Steps. That we need to take so. In terms of whether a I displaces. Jobs, actually. Lower left side they will be displaced, largely, by AI but. The lower right side are where, AI can be a tool to make the scientist more creative, to help, invent more drugs and. For. The same amount of time and to. Alleviate. Pain and cure diseases and, then, on the upper left is a different, type of human-machine. Symbiosis, where. AI, might, perform, most of the analytical, part of a task where, human.
Adds The warmth for, example a doctor might. Use more, and more AI tools, and diagnostics, to determine, what's wrong with the patient and what to do but, the it, is the doctor, who who would communicate. With the patient tease. Out all the issues problems family history, inputted. Into the computer, and then, explain. It to the, patient, in a way that comfort, the patient gives the patient confidence maybe, with home visits, and so on and and, we know that through the placebo effect if, the patient has higher confidence there's, a higher likelihood of, recuperation. Or survival, and then, on the upper right, is really, where. Humans. Really. Shine with both our compassion, and creativity, so. This is the, blueprint. Of coexistence. Of human, and AI. So. In, my talk I've talked about the opportunities, and challenges in the next 15 to 20 years but. I firmly, believe, if we look a little bit beyond that for the students, in the room when. You're when, your children enter, the workforce I, think, when 30. 40 years from now when, they look back and, think what a I meant for Humanity I think, what they will think is that really, two. Things first, AI is, serendipity. Because, it liberates us, from routine. Jobs so, that we can spend more time with our loved ones we, can do, things, we're passionate about and we can have time to, think about what it means to be human and, also. If people, are worried about ai ai, is just a tool and we. Are the only ones with, the free will and we, set the goals for AI so, we humans, are going to write the ending of the story of a yawn thank you. Thank, you so much doctor. Carefully, I, want to welcome to the stage, Professor. Susan Athey and professor. Erik Brynjolfsson. To. Start our discussion. So. We have a tradition, at AI salon where, we. Always. Use, our hourglass. To. Identify. The time because, we'd use no technology, during AI salon this is an Enlightenment, era salon, where, technology, doesn't exist yet so. In. Light of that we will also be using our hourglass to, track the time and, so, this. Will be an hour for a discussion I just. Want to do a quick introduction of our two other guests here, Professor. Soon ëthey is a professor, at the Stanford Graduate. School of Business she, received her bachelor's degree, from Duke University and her PhD from Stanford. Her. Current research focuses on, the economics, of did I say ssin marketplace, design in the intersection, of economic. Econometrics, and machine learning. She's. One of the first tech economists, and served as a consulting. Chief economist, for Microsoft. Corporation for six years and in. 2007. Professor Athey became the first female winner of the John Bates Clark medal won the most prestigious, award, in the field economics. Professor. Erik Brynjolfsson, is a director, of the MIT initiative, on digital economy and, a professor at the MIT Sloan, School his, research examines, the effects of information. Technologies, on business, strategy productivity. And performance on, digital. Commerce and intangible, assets he. Was one of the first researchers to measure the productivity and contributions, of IT and his research has appeared, in leading economics. Management and science journals and recognized. With 10 best paper Awards and five patents so, thank, you guys so much for being here today, I, wanted. To start, out the. Salon by kind of thinking, about the applications, of AI so. Kai, foo here talked a lot about many. Different uses such as loans, or we, can think of things in the news like alphago, forgo. Autonomous. Driving but, how generally, applicable is, this technology, was warning of err key and talk more about this well, I think chi foo is exactly, right that these are some amazing technologies. But, they're quite narrow. In many ways I think this is one of the biggest misunderstandings. In the popular press especially with Hollywood where, there's a lot of sort of impression, of that we're close to what, a lot of people call artificial, general intelligence or, AGI and, and. We're really very far from it there are more breakthroughs we don't know how many are needed but. Deep. Learning by itself is quite remarkable but but can't I think, most people would agree get us to hei that said, we don't make the opposite mistake and under play how, remarkable, these breakthroughs are so in certain specific areas the, image recognition is particularly one Feifei with, or with the image net set, off an explosion of work -, showing how rapidly, you could using. Deep learning techniques, get, to recognize, images voice. Recognition, credit. Decisions and cuff you gave lots of other examples and each of those are in their own narrow way not just human but superhuman.
The. Issue is that when when a human is able to do something extremely, well you know speak Chinese, you assume they also know, something about which Chinese restaurants, are good or which you know something about Chinese culture AI, you. Be. Mistake to take extreme. Competence, in one domain and generalize. It to other areas I see. So kind, of going off of that Susan. You just gave a tutorial, and Europe's, about. Causal, inference and you, kind of talked about a lot of unanswered, questions in, the field of AI and a lot of work that we still have to do can you talk a little bit more about the research and Genda that your group and other researchers, are working on to, make AI more generally. Applicable yeah. So first I would just say I really agree with the perspective in the book and also the way that eric, has nuanced, it that we. Have this. Incredible. Revolution, and, and the the big breakthrough of neural, nets allowed us to solve problems that, we couldn't solve before but, yet those problems, still fall into fairly narrow classes, so, you, know trying to understand, sometimes, alphago, is a very hard game because, it has a very big state space but, it's still a game where, if I have two strategies I can, play them against each other and see what one is better and so the computer, can generate, massive, amounts of training data the actual. Algorithms, used, in an alphago are very similar to what we've been using for decades in, in economics, to try to learn. From either, human, behavior, or firm optimisation, decisions, or firm equilibria, about. Their payoffs, and to try to simulate what, would happen in a different world and that's sort of a form of counterfactual, inference trying to understand what would happen if something. Changed, a bit but. Even. Though the the, sort of conceptual. Approaches, are similar. The. Neural nets don't necessarily, make a big breakthrough for. Those problems, I'm not going back to revisit those problems, using, all I've learned about machine learning because, the problem there was actually just lack of data if.
We're Trying to study you know where Walmart and Target. Should put their stores or you, know how much firms should invest, in. Or. You know even how a human. Should, just. Make dynamic decisions. About. Training. Or unemployment. Those. Decisions. We still have like a relatively, small data set to study them and the real problem is kind of data sparsity, and also, lack. Of enough sort of experimental. Variation, and the data to really learn about cause, and effect so. Even. Though these breakthroughs are huge it's not to say that oh well we tackled, go, so next step is to replace the economists -, noted that economists, wouldn't be replaced, I was very happy he, also, noted, that we don't have a lot of compassion which, sadly. Is. True. Of my colleagues hopefully not of me. But. We. You know so so I think it's it's really just as Eric said it's not like it's just one short leap from, the problems that we've solved even though they sound hard to other types of business, problems, and so then, in terms of the research agenda though, I think there there is actually a really exciting research, agenda. And I think for the next generation, of students now that we've sort of ingested. AI and as Typhoo talked about it's become more incremental I think we can go back to some of the techniques that we used in the more small data world because, inside, every AI, is, it. Is an agent, that you're trying to make decisions they're. Doing counterfactual. Reasoning you know how should I climb the wall you're what kind of recommendation, should I make these are decisions, and so, you, have to do, counterfactual. Reasoning you have to think well what would happen if I made this decision what would happen if I make that decision and inside. The agent is sort, of a statistician trying, to use a small amount of data to, figure out what decision, there is to make that, is a hard, problem and it's an especially hard problem, in a data poor environment, where maybe you're in a situation you haven't been in before and you need to still reason, about what comes next that's really a more human my creasing, and one, way to make, it better is to put. Some structure on the problem don't, just learn from the data and a completely unstructured way, but, actually, use what you know about the environment, use the domain knowledge. About about the system that you're in can make it much more make you much more efficient, with the limited data that you have but. That's not been the focus, of the, last ten years of, machine, learning and so I think it's actually really exciting going. Forward to think about marrying those and that's what I'm doing in my lab but you know just on a small number of problems there's really a much, broader set, of questions, and then the last theme that I focused, on in my tutorial, was how thinking. About things this way putting a little bit more of a structured, way of thinking on things can actually solve, a lot of the problems, that AI has had in the implementation, phase so, when you go out to implement, you know the if you want to get humans to listen to you the, human you you have to be interpreted, by they have people have to believe you they have to trust you you have to actually be able to tell a human that you're making a recommendation to do, I know the answer to this maybe I'm uncertain over here maybe I'm more, certain over here maybe my algorithm might be biased over here but over here I've got plenty of data and I think you should trust it we, need to be we need to deal with these issues of stability and, trustworthiness. And so that, also. Really. Requires a clear conceptual. Framework, and in. Layering. That conceptual, framework on top of all of the algorithmic innovations, that we've had so I'm really, excited about the, next ten years of, basic, AI research, and. It's Stanford we're, really putting a lot of emphasis on that human centered AI is something, that you know Feifei is helping. Lead us on and and and I'm really excited about what we're gonna be able to do in that area to make a the domains of AI more applicable, yeah. So. I feel. Like in the past ten.
Years Speaking of these great. Innovations, and breakthroughs within, the eye there kind of has, been two. Extreme, responses. To, the breakthroughs, on one hand you have, optimists. Like Ray, Kurzweil who says AI will, create. Paradise, for. Humans on the other hand you have doomsayers, like Elon Mustang, that AI will, create. Robot. Killers that will kill us all. So. If. We see, that as like a spectrum, which of course it may not be one but if it is a spectrum, I'm just curious to hear, where, all of you are on that spectrum do you believe that you. Know a is paradise, or is, AI going to kill us all. Shaking. Your head it's going to kill us all. No. III, am, very frustrated, that that is so much of the discussion. Out there I mean first getting back to what we were saying earlier we're. Still quite far from AG I know Ray Kurzweil, has a different, view on that but most, a AI researchers, that I've spoken to would, would disagree with him in terms of how, close we are to that but, there's a more fundamental point. And it's really a key, point that Kaku makes in his book which you actually look at and he made made his talk and I want to hit on it again which, is that the, technology is a tool so the right question isn't, what, is AI going to do to us is it going to give us you know Nirvana is, it going to solve all of our problems for us is your going to kill us all, those, are both they. Both both those questions not opposite of the spectrum they're really doing, saying the same thing which, is treating, AI as if it's the one that makes the decisions, and the reality. Is is that technology. Is a tool a hammer. Is a tool AI, is a tool and those tools can be used in lots of different ways, they can be used to do, constructive. Things they can be used to do destructive. Things and the, real, important. Reason, that we have to understand that is too many people I think are being passive, about what's going on and we have to recognize, that we are agents, as how food was saying we, can make these decisions and we have to decide, how we want to what kind of society do we want to live in do we want to have one with the kind of compassion that, Caillou was calling for what, does that mean in terms of the policies, we put in place what does it mean in terms of us as workers. As CEOs as citizens, we, have to take agency, and use these tools in different ways and then, the question isn't which, of those is going to happen it's which one do we want to have happen and what steps are we going to take to achieve that well, I just want to kind, of then bring that to our topic, at hand today which is you know AI in the future work. It. Is true that AI is definitely, a tool but you know they're projections, Oh in fact Typhoo writes, in his book that perhaps.
40 To 50 percent of jobs will be replaced, you. Know in the next 20 30 years and that is a huge. Huge. Obstacle. Society I'm just, wondering if you, guys can talk a little bit about what, kind of impact this, will on the world will, have on the workforce and how can we prepare for that -, want you elaborate okay sure. Yeah. There have been different, types, of estimates, from a lot of different studies my my. Numbers are more aggressive. Mackenzie. And. Oacd. And others have come out with different numbers generally. The. The, in terms of the numbers people, are believing. That we should look at how many tasks, can be done, by AI, not the jobs however. When. You have a job half, of the tasks can be done there's, going to be 50%, people who will probably not be working right, so. So. On the one hand I think is a scary. Large number, but. Also if, we look at agriculture. To manufacturing transition. Their, numbers are even even. Larger than that so the issue really, isn't how many jobs are displaced, many. Of the jobs that when. You graduate from here our jobs that didn't exist five or ten years ago so, having, new jobs and having jobs go away is, has. Been the, de. Facto right the status quo always, happens I think the issue, really is when, we went from agriculture, to manufacturing, the. People from the farms were able to now go to the factories, because, of the relative, unskilled. Low. Training required, to, be on the assembly line for example the big problem now is I think AI, is displacing. Mostly, the routine work and the. People who are displaced. Really. Need to, find. Their place in a. New place and it's not just a loss, of income issue, but. It's a loss of meaning that, people. Attach. Their meaning to the work that they do and, yet. When. Most, or all routine, jobs or tasks are gone people, have to really. Be trained to do the more complex. Announ, routine tasks, so I think the, big issue is a train is one about training, and, Amazon has shown one way of training. You. Know share the positive, example, but that's because they're almost a trillion dollar company, they can afford it Walgreen, has the same workforce that will be this place I don't think they can afford the, training program so I think we really have to think about how. Corporations. And. Governmental. Programs, can, be applied for example, in, the US Congress there, are a number of bills. Being discussed. Such. As giving human, resource, training credits. Back to companies, so. I think it's going to be moves like that not. Universal basic income that. Starts, to move the dial forward, last. Comment I'll make is that China, and US are quite different, I think. China is a very. Decisive. Execution. Oriented. Form, of government, and. In. The last phase of the transition to, from agriculture, to manufacturing the. Government, played a very strong role of, saying okay, we're shutting these down you, guys move over there you're going to become that in a new job so, a very very top-down. Organization. Which. In the case of a crisis, may, be more effective, so, I think actually us probably, should be a little bit more concerned. Because, you, know with the. Government. Not being able to move. People from job to job and. And, also the, and. Also, I think currently, unemployment. Numbers are at historic, low I worry. Whether US. Government, is going. To do anything, so. Susan do you want to elaborate, or do you have a different, opinion on this well so I first. Let me agree, with everything that have, you said and I think actually I'm also concerned, about the u.s. policy in the sense that like our it's not like our government has gotten more functional, in the last few years and yet, we we. May need to be preparing, for something really important, and at the moment you know we don't actually know, if you, wrote me a big check I don't know how to spend it in terms of making workers better off like we actually just don't have that muscle that capability, we don't know how to retrain people we don't know what to retrain them for so, I'm concerned, that the time will come when we will need that and we're not going to be ready for it and actually that's a big, emphasis that I'm moving into my own research is to try to you, know work with other, scholars and students here at Stanford to try to you. Know prototype. Digital, and AI driven, work or training and also, worker recommendations. To try to help people understand. And make better decisions so I completely agree it's an important, problem and then, even with the displacement, coming relatively, soon you know I advise a lot of companies, banks. All over the world have call centers in poor parts of their countries, and they're. Those call centers really will be gone in, you, know they're really not economical, today they're gonna be shutting down over the next years and there's gonna be these, these concentrated. Hits to, regions that already, lost, manufacturing, they already lost a bunch of stuff and they're on their last legs and then the call centers will go, too and I think that's gonna be problematic, and we've had some interesting research by drone awesome igloo at MIT and co-authors, kind, of showing how you, know when you get these concentrated, hits like from robots in Detroit that those areas, really can can spiral down, on.
The Other hand now I'm economist. I have two hands, on. The other hand there's. Also some, reasons to be less. Concerned so how very ins an eminent economist, and he's been giving a really nice talk about this recently and his talk is called BOTS versus Tots and so, he takes the most aggressive, numbers, about, worker, displacement, but then he looks at demographic, trends so it's a little bit hard to predict their future in a lot of ways but demographics, are actually pretty easy to predict like we kind of know how, many people will be 50 30 years from now okay so that's a that's something that's easier to predict, and in, fact in developed, countries with following birth rates and and also China is very. Harmful. One-child. Policy ends. Up with this aging workforce I would I really like from Caiphas discussion, is how old people actually need a lot of care and we're. Gonna have all these people consuming. But not working. And. So we're actually without big, changes in immigration policy, we're actually gonna have worker shortages, over the next 30 years and how. Argues, it's not my research but how argue that those effects are bigger than, the most aggressive, job loss effects from. AI, so. If you put those together though, you know what's, right and what's wrong it's hard to do but I think that pushes, you in a couple of directions the. The not. So much universal basic income but instead why don't we be thinking about how we're gonna take care of the elderly which. Could be augmented by, AI and, in. Both physical robots, as well as monitoring. And and so on and decision, assistance, you. Could probably have one human per old person, and like. If anybody's cared for an aging, relative, you know that's actually pretty labor-intensive you, know and so you actually could employ a lot of people in these service jobs and so I think we should be looking at sort of labor augmenting. Technology. And if the government's gonna do something they can train people to work in those things and also they can subsidize, the services, but, in the end there's, plenty of work you know we could have one one, worker per every preschool, child and one worker per, every older. Sick person and kind. Of employed everybody. So, I'm not worried about not enough jobs but, I am worried about how big I you're ready to yeah I just, want to very, much underscore, what Susan just said that that.
Kathy, Was right many, jobs are going to be Lamia maybe 30 40 50 % but. I don't think that there's a shortage of work that needs to be done in our society, that only humans can do because. Machines, can't do the whole spectrum of things even, within particular tasks, there are a lot of things that require. Humans that I think we value a lot taking, care of the elderly taking, care of kids. Cleaning. The environment a lot of creative, work, arts. Entertainment. That. Are. Inherently, require. Humans at least with existing, technologies. Nodes that will be for some time so, our challenge I don't think is so much the end of work, our. Challenge, is this transition that Chi flow alluded to and the sudras owners going and how do we how do we get, people to shift I'm a little bit more optimistic than Susan, about that we do know some of the things that can be done I think that that, if you take Chi Foods points about creativity, and compassion as being important, things I think we can do more to have education, that that supports, creativity, that supports interpersonal. Skills and and in, fact when you think about it right now a lot of schools. 19th. And 20th century schools were designed to stamp those out like, keep, you know don't be doodling, you know don't, play but if you put put a pile, of blocks in front of a kid the first thing they're gonna do is they just want to start building things so inherently I think we like being creative, and if our school didn't stop us we'd probably be even more creative and we like interacting, with other people we like playing we like teamwork. So, we could do more in our education, to support that and those are the kinds of skills that are going to be in more, demand the other side of it is entrepreneurship. On, one hand we want to have the skills in the workforce on the other hand we, need people, who can figure out how can we, combine those skills with. Technology. To solve existing problems well we have a whole class of people that are that's, their job is to make those new combinations, we call them entrepreneurs, they're not usually professors, or policymakers, and. The. Surprising, statistics, that I saw were that I'll. Not. In Silicon Valley but in the United States as a whole entrepreneurship. Is down that, we have less creative, destruction less, invention. Less new business formation fewer, businesses, that are five years old less. Turmoil. Mark, in the market as companies. Reconfigure. Than we did 10. 20 or 30 years ago so, we need to do more also to make our economy, the United States more, dynamic, I think that I'm. Also more optimistic, than Chi foo that a a, entrepreneurial. Decentralized. Economy, can respond, to these kinds of transitions, if we make it a little bit easier. To do it but ultimately the challenge again is not that. Lots. Of people simply. Won't have jobs it's that those. People won't be transitioning, to the new kinds of jobs that are needed and when. A. Technology. Automates. Part of the task and this is some work I did with Tom Mitchell. When. It automates part of the job like a radiologist, there are 26, different tasks, that radiologists, do, it. Could automate some of those but there are other ones that actually become more important, so the re-engineering is will be the big. Challenge for us going forward both, at the task and occupation, level at the, industry, and firm level and at the societal, level well. One comment on another. Dimension is, other, countries other than us in China mm-hmm, I think, there's, going to be major, potentially. More major problems, there are especially. Countries, that have been hoping to use the China model or the India model to, climb out of poverty right, because China, did use. The lower cost labor workforce. To, outsource, manufacturing. India. Uses, english-speaking, population to. Take care of call centers and, IT. Outsourcing etc. But these jobs are the ones that are going to be displaced so. China. And US can absorb the because there are all these value, creation engines, entrepreneurs. And and and, and big companies, what.
About The poorer countries that have been hoping, to use the China or India model, and on, top of that they, don't have the revenue drivers. The big AI companies, or tech companies and on, top of that the. The. Workforce is relatively. Less trained so, I think that presents a bigger issue for a, lot of other countries well you're talking about the very poor countries, I also have concerns about Europe, for. Different, reasons you, know but I guess and that really gets back to one of your boxes that you didn't have time to talk about in, your talk, which is sort of market power and you. Know from. Its, completely, good. Thing for China in the sense that they're like you know Google, isn't the, only search engine in the whole world and of course being where I used to work, but. You. Know and it's, you know it's great I think it's really important for the world that there's more than one tech company, doing each particular thing, but, a lot of these things do end up being, very concentrated, and, so, especially. For AI where, exactly. There's Ada and then it eventually. Becomes monopolies, how do we how do we deal with exactly, and I spent years trying to help Microsoft, search engine compete with Google search engines so I've spent a lot of time thinking about how important, data is and how hard it is to be a number-two company to, a number one company that has more data and the answer is it's hard, and. And. So you know that that is something that we need to really consider and, we haven't seen as many European, tech companies, which means that the whole community and capability. Hasn't, built up as much there, to participate, in this and so, we already have trouble redistributing, within a country we're, really bad at redistributing. Across, countries, and and. So you know these these countries, are gonna have challenges, sort of keeping. Up and especially, given that you, know at least maybe in the Western world we're going to have you. Know a lot of concentration, and so I think that gets back to the inequality question. And another concern, that I have is that we. Have to think about how the benefits, of productivity get passed along to consumers, so, just roughly. If you talk to a macroeconomist. The. Old-fashioned. Simplistic, one-on-one, models, would say that like how could a I be, bad if you, can make more outputs, with, less inputs, you. Have to be better off and if like we're all identical, and we all own an equal share in like the one factory, that, produces our outputs then making it more productive it just has to be good sort of tautologically. You know how could we be worried, about being able to make more stuff with, less inputs. But. Of course, when. You have a real economy where we are not all owning, shares equally. And everything then this distribution comes in and then there's a question of market power so we could replace all of our workers with robots that in principle, could. Lower the marginal cost of output which, lowers, the marginal cost of living right. Because if the if the if the cost goes down the prices can go down and then it's actually cheaper to live so you, could you might have lower wages but, also lower, costs, of products and everybody's fine but. If there's a lot of market power then. The company might actually keep, a lot of that as profits, they. Can lower their costs, and of course you know some people share in that but that may be the workers don't, and so I think we're gonna need to be thinking, a lot about making, sure that all of these productivity, benefits, actually get, you know passed on to consumers so. That's one component of cost of living health, is another one, transportation. And housing is another really big one and so, actually, getting back, to like what can we do we. Need to think about how, our policy, towards, things like transportation. In. The advent of autonomous. Vehicles or changes in transportation, actually. Affect. The, ability of people to get service jobs in cities here, in the Bay Area you know you can't hire a service worker very easily people. Might have to commute two hours to live cheaply that's. A totally solvable problem, and especially. With improvements, in technology. You, know we should be able to have. A much wider set of land where people could live and work in the cities but, making, that actually, work for people that is making. These changes. Two great. Public transportation and, low-cost transportation, for the poor rather than just a bunch of rich people riding, around in their autonomous
2018-12-22 02:37
Other news
![](/pic/tech_stocks_flirt_with_worst_week_since_2022_bloomberg_technology/SzhfQS1zUGh0b00_.jpeg)
![](/pic/abc_news_prime_race_to_the_white_house_holiday_travel_new_tech_protecting_teens_on_social_media/NlJhY0UxdWpLWEk_.jpeg)
![](/pic/anticipating_nvidia_earnings_and_telegram_ceo_faces_potential_charges_bloomberg_technology/SExYbV9iZm5ZOEE_.jpeg)