Want and ultimately enable us to create a future with ai that is positive for humans and machines. We are fortunate that these conversations are happening more now and that two of the press spot leaders driving these conversations are our special guest tonight. First, we will hear from eric, director of the mit initiative on the Digital Economy and author with andrew, bestselling author of two books. Next we will hear from max tegmark, cofounder of the nature of life and author of the new book just coming out, life 3. 0 being human in the age of Artificial Intelligence. I understand its in its second week on the New York Times bestseller list. Congratulations. Then the two gentlemen will have a conversation together and finally there will be time for questions with the audience. For we conclude with the reception in a book signing. Please join me with a warm welcome for eric and max. [applause] thank you. I am so delighted to be here. Its a pleasure to be able to have a chance to share some of the research we have been doing and some of the work weve been doing at the mit initiative on the Digital Economy about Artificial Intelligence and how it has change society. I want to add my thanks to jack and susie reno for supporting this and for inviting me and a special thanks to max tegmark. Weve only known each other about three years but hes become one of my best friends. He is, as i say in the back cover of my book, absolutely joyful mind and when he said lets do this, i jumped at the opportunity. Mainly because its fun to talk with him but will have a fun conversation where we interact and im also looking forward to hearing all of your questions and comments because we will open it up and everyone will participate in it. We live in some very unusual times right now. As you may have read and seen cars are beginning to drive themselves and people are walking down the streets and talking to their phones and are not talking to another person but expecting the phone to understand what they are saying and talk back to them. Still bumpy and theyre not really that good at it but theyre beginning to talk to us and were in this idea of a ten year period where were going from machine mostly not understanding us to machines understanding us pretty well and that is a unique time in history and were very lucky to be able to participate in that and to be part of it. It opens some amazing possibilities. These could be the next ten years could be the best decade in the history of humanity or it could be one of the worst because the power being unleashed by Artificial Intelligence is unlike anything we have seen for. Let me set the stage by defining what we are talking about. You can think of Artificial Intelligence as the set of techniques to imitate the human mind, the classic is the [inaudible] test which was proposed by the famous computer scientist, alan turing, could you speak to a machine and not know whether or not you were interacting with a human or a machine. Machines arent passing that test yet but they are getting closer and closer, especially in certain narrow areas. Within that is a category called Machine Learning and this is what is driving a lot of the excitement recently is that good oldfashioned ai was an area where we would teach the machines what to do and we would write down symbols and say this is how you play checkers and this is how you play chess and these are the rules and this is how you prepare taxes and machines would follow those instructions. Now, the Machine Learning revolution is taking over and a set of us humans having to tell the machine stepbystep what to do, which frankly didnt work that well. It was okay but iran into a lot of barriers now, the machines are learning themselves how to solve problems. They are figuring it out in the way they do that is mostly by different techniques but the main approach is that we give them lots of examples and say this is a dog, this is a cat. This is a dog. This is a cat. I wont do it as many times as a due to the machines but they use usually do 10000 or 1 million times and eventually the machine starts to think oh, i see the pattern here and it will start learning. The nice thing is we have so much digital data now that we can show them lots and lots of examples of broad or successful go moves or of spaces that eventually they start to learn these patterns. And a particular subcategory of that we are really the biggest part of the breakthrough especially in the path of fiveeight years is in deep learning or deep neural networks. Is loosely based on the human brain. Let me show you some examples so you can keep learning in a subcategory called reinforce learning. These machines can learn new strategies on their own so one example is a group of people at a Company Called deep mine, google bought deep mind and what they did is on the cutter under cover of nature which is a Science Magazine and they gave this Machine Learning algorithm the pixels for atari game. Have anyone ever played Space Invaders contract what about breakout . I will show you the game of breakout and they gave the machine the raw pixels and they didnt say this is the paddle, this is the block and this is the ball the machine had to figure it out and learn it on its own. They give at the raw pixels, they gave it a controller that they could move left or right and they gave it a score and said here is the score, look at the score in your job is to move around the paddle or to move around the controller, they didnt tell it was a paddle but adjust the controller and make the score as high as possible. So, at first the machine wasnt all that good. It would sometimes get lucky and hit the ball and other times it would completely miss it and it was basically randomly moving it around but whenever it was successful and got it the score would go higher and it realized it had to do more of that. Reinforcement learning, getting feedback and learning what to do more and more. After about 300 games it was pretty good and almost never missing and is good as a good teenager and playing along successfully. They wanted to let it keep running for a while and the guys at google deep mind dont pray to break out a lot and they didnt know there was a strategy. It appeared out how to send the ball behind and they were like whoa, we didnt know you could do that. [laughter] so, the machine had not only learn how to play but learned how to play better than its designer imagine a newborn baby being born in the hospital and you handed a game and by the end of the day its beating all the surgeons and doctors at the game. That is how fast it is learning. Thats a cool little example in a scientific progress and by the way, the same technique works for a whole set of atari game and it did work on Space Invaders and eventually on pacman and other games. It didnt work on all of them but when it had a quick feedback loop and the score would change quickly it was able to learn on its own pretty quickly with no control programming. Now, you can use the same techniques for other things but not just games. You can think of a data center where google has all of their Computers Running as a big videogame and all this data is coming in and temperatures and the score is try to make it as efficient as possible, lets lower the cooling bill as much as possible and your control as you adjust the valve and move them up to write. They had a bunch of smart phds on optimizing this so they thought they very much had it running as efficiently as possible once they put it on the Machine Learning algorithm it got dramatically better. This is before the Machine Learning algorithm was on and it got about 40 more efficient and they turned it off again and it went back to the way it was before. The machine had figured out how to run their Data Center Better than all of these geniuses at google were running it. It doesnt take much imagination to say lets not just to it that one data center but all the data centers and why dont we do it for all kinds of factories. Steel fishing lines and makers of any kind of an object and there is room to apply these things to prove all sorts of categories. In fact theres three big breakthroughs where Machine Learning has made a big difference that works important as recently ten or 15 years ago and we write about it in my book to some extent but vision and language and interaction with physical world and problemsolving. For instance, you may want to get snacks afterward and be careful what you are reaching for but there are some others here and not all of them are muffins but sometimes we can make mistakes when we are seeing things and at stanford they have developed a large database with image net with 14 million images and each of them have been painstakingly labeled by humans as to what they are and back in 2010 when they tried that machine to see what they were they were not very good in the wrong 30 of the time. Today they are wrong about 2. 6 of the time so they can dramatically better and they started using these deep neural net algorithms. As a reference point, humans are about 5 and they havent improved a lot so we pretty much have the same hardware and software and the machines have crossed thresholds because what that means is now tasks that would be better to have humans do them but its more accurate to have a machine do them now. That shows up in a number of areas you can help diagnose diseases and use that same kind of algorithm and you show examples of patients with cancer in patients who do not have cancer and the machine starts here now as well as or better than the human pathology. [inaudible] this is just happening this past few months and past two years. I mentioned Voice Recognition and you can see the progress there, 8. 5 to [inaudible] and thats past ten years and thats the past year since july 1616. Humans are about 5 error rate so its in that ballpark right now and not quite better than humans but that is opening up a lot of economic possibilities. Interacting with physical world so once you can see and recognize things versus recognizing a pedestrian or bicyclist it starts to become visible to control the car through a machine and when we first started doing these we made errors for every 30 frames, one every 30 seconds which is not what you want but now its every one for 30 million frames and thats years you go about making a mistake. Better than humans. Very soon we will see more of these and i had the pleasure of writing them and i feel quite comfortable driving having driven in the road making a left turn through traffic, waiting and ultimately it will be much safer and there were 30000 deaths by humans drivers today and we can drop that by 90 or 99 when the machines are not a 100 so well have to face some ethical issues when machines drivers still make mistakes but it will be dramatically different than what we have today. They are beginning to work in factories, rod brooks who used to be at the Computer Science lab at mit now has a company over in boston since called rethink robotics and for about 4 an hour doing a simple task and you dont have to do Computer Programming and you show faster what you wanted to do, pick us up, put in the box and after a couple examples it understands you wanted to do and it does that task and can work seven days a week. I was just on thursday watching another robot like this and sorting different soft objects like clothing and that will replace a great deal of work in those areas. Last but not least, all sorts of problemsolving like medical diagnosis already showed in the legal area talking about guys with j. P. Morgan and they see a lot of relatively routine legal work where it says 36000 hours worth of legal work. What does this mean for the economy prospectively briefly touch on that before i briefly handed over to max. First off, there is good news but there are big challenges. It makes the pie bigger but there is no law or economic guarantee that everyone will benefit. It is possible for some people to get none of the share or even to be made worse off than they were before and sadly, that is part of what has been happening the past decade. It could get worse if youre not careful because productivity has continued to grow and gdp is at a new alltime high but the median Family Income is lower now than it was in 1990. Good news is they had reports opportunity 16 in the last year and there was an uptick over 3 in last year and depending on how you to the adjustment it may have matched the previous time although if you normalize it is still lower than the previous high back in 1997. We are roughly flat during that period. How can median be so much lower than the. Person and that is because median, in the name of science is the 50th percentile and its not the average but its a person right in the middle and half the people are higher in half the people are lower so the median can stay flat and if you have a whole bunch of wealth going to the top 1 or the top one tenth of 1 and that is basically what happened as computers have kicked in there has been biased technical change and you heard about the 1 and 1 have their own 1 and the. 1 and the share of income going to them is that a new record high and it was close during the Great Depression if thats any consolation. We are having an economic challenge of a pie getting bigger but the distribution is becoming more and more skewed and there is many reasons for that. Part of that has to do with tax policy and natural trade and most economists including me see the Way Technology is being used as the number one driver of that. Now, that is not inevitable. Ultimately, we have an opportunity to rethink how we organize our economy and as the pie is getting bigger and were creating more wealth that is a series for everyone to get richer at the same time and we can make the rich richer and the middle class richer and the poor richer and we could all be better off at the same time that mathematically adds up but there is some choices we will have to make as a society of what we want to do in terms of taking advantage of some of the bounty. This is as usual wont solve the problem but a number of things that mit will address it and were trying to understand the drivers of this and do research on it and weve also organize something called the inclusive innovation challenge and i invite you to another event that if you dont get tired out by this one on october 12 at Boston City Hall plaza the governor, eric schmidt, a whole bunch of other people will be coming to talk about how we can use technology to create shared prosperity for the many and not just for the few. With that, let me leave it with the closing thought that these technologies are certainly wondrous but they give us all sorts of opportunities. They can be used for good and to create past 12 but they dont automatically lead to distributional makes everyone else better off. Its important and im glad that you are here because of support for us to think hard about what we can do to change the kind of society we have towards a better one and what we want to do to use these technologies for broadly shared disparity. Thanks very much and with that let me turn it over to max and eight. [applause] you so much for inviting me here. Thank you, eric, for your friendship and your all too kind introduction and for setting me up so wonderfully here. Let me see if the technology cooperates rightly here. I will continue further forward in time and talk about what will happen if ai keeps getting even smarter and what it will be like to be human in the age of ai is what it should be like. First, lets go far back and look at the big picture. Millions of years ago, i have to start here since im a physicist. The universe was very boring with a uniform plasma almost everywhere and no one there to witness it or enjoy. Gradually, the laws of physics morphed into galaxies and stars and the first light appeared here on earth and satellites was done though. I call it life one point oh and couldnt learn anything in its lifetime. Life twopoint oh which is what i call us and if we use to bring the metaphor of a computer of sorts then learning corresponds to a ploy to new software into our and we heard that traditionally from a the world chess champion people taking their on intelligence simply because it could remember more. They have been driven by Machine Learning with a very Simple Machine in you just train them with massive amounts of data. And then this is the image they gets even more striking as we saw in this example if it could turn to ruth kleve party game . Match already toeshoes there is a lot of room for growth if you are a robot and then you are rewarded for certain things. And then they trained the robots to see if they could learn to walk but theyve never seen a video of somebody walking they did us in malaysia because it was cheaper but the idea is the same. Nobody taught it to do this. Theyll learn to run and they learn to jump. Any thing almost whether the stock market so where does this take us . So with those intellectual tasks where the high bid is so hard and that ocean level solatium then chess playing skills really even more than market calculators in occurred to them to be on the of boundary but this is lovell is rising because machines keep Getting Better. Something they can never do but others think they could do everything in 30year 40 years. And then what . We have very interesting choices to make i feel that we should not make them deliberately. Standup. Also someone we are very happy to have on board. And what we need to do to get things right. And that where we manage the technology . To change the of strategies ever relearn from mistakes. All in all it was a good strategy but eventually will reach a threshold where technology is so powerful you want to get things right the first time. I would argue Nuclear Weapons is a net category is super human intelligence and shifting from being reactive to proactive some people call it the scaremonger. And with the First Mission to the moon. And the stress of the mission. What is my a suggestion . First of all, we should try very hard with of trees so the biologists and chemists because of i ask them with interfirst . Letter about chemistry . Rather than chemical weapons. Because they persuaded the of politicians of the world. The physicists have a scorecard and to feel very strongly to be like a biologist and the chemist and keeping the power of a. I. Doing all that wonderful stuff rather than making it dramatically cheaper to murder people. But if you take Something Like that and drive the price at zero that is number one on my list. And then to keep us focused because the superpowers have a lot to lose. But this tide is to make everybody better off. And i think we have to invest in a. I. Safety research. There are a lot of technical problems we need to solve raise your hand if your computer has ever crashed. That is a lot. How does that feel . Not good for go frustrating for pro edison the worthy would use if a. I. Was controlling the u. S. Nuclear arsenal so it is incredibly important and the other key challenges the goal that the machines have but they are lined with hours. It is frightening to be in the presence because we have all done it. We are in the of presents of those entities. But if you tell them to take you to load again as fast as possible then you say no and though that is not what i asked for any begin to appreciate what it is to understand our goals is because they understand is not mean they will adopt our goals we know what thats like with children. My kids are much less interested in lagos now a free program them so theres some technical challenges their. So to summarize why they are smarter than us . I will summarize this and a very short video. Will Artificial Intelligence ever replace humans . Some claim they will been super intelligence to destroy humanity others say dont worry a. I. Is something we can control coming computers. So we have the collective takeaways and will help separate the myth from the fact is. They have long been better but those are from repetitive and mechanical operations. Making videos or consoling a friend . It can only exist in biological organisms. It is information and processing and reacting. Theres no way to make that better than humans already do. Many of them were already better than human. They were on track to a mob baffle intelligence. How to beekeepers those on the right side of the flourish for the founder balance what should we be concerned about . Super intelligence does not share our goal. No need to worry. But it is what they heat seeking missiles does not what it doesnt do. Super intelligence a. I. Is the most important thing is to align with hours. We want to but a bill that hydroelectric dam perot but they have done a great job to space their goals. I can help thinking but kittens are the cutest but if super intelligence would be better off or figure out how it is our goals and set the other way around. When we need to start panicking . First of all, it doesnt have to be negative. Everything i love about civilization is that amplifies our collective intelligence then humanity may flourish second most things super intelligence is decades away. Akio also take decades we need to figure out to adopt these goals for themselves should redo whenever the president once . And a very real way what kind of future we want to create for humanity. So how do i get involved to make sure we dont have a dictatorship . I am very interested to continue the conversation with what kind of future you want to create that is not the question that should be left to the geeks. Thank you so much. [applause] so when will a. I. Be here and you kind of dodge the question. So you can be more precise kidding when we would have machines and what do the other experts say . First of all, what do we mean by intelligence . I mean to accomplish the goals. But the general intelligence is as broad as ours we have a bench of leading a. I. Researchers when did they think they could do that all . There was a violent disagreement which means we really dont know. Is somebody tells you for sure it will happen soon then they dont think it will happen for at least 100 years but if they say for sure not to in our lifetime they are exaggerating because we have a lot of researchers who say the gap have been 2040 or 2050 including those leaders who support the showcase. I think the average was 55 . To be precise about that. So in a matter of decades we put a lot of thought coming out of retirement. Well i heard what if alien intelligence lands on earth 2055 . In some ways it is alien intelligence. Except we are better off here because if they arrive from the of vegas system vega but we have that opportunity to decide what the intelligence will be. You convinced one really important person that was very charismatic elon musk thought it was an important challenge and said you are summoning the demons and that never goes well and tell us of fascinating story you dont have to do all the details but how you got elon musk to make a donation. It is fascinating. Not too long ago i cannot even dream so i start to realize he was thinking very seriously about this and doing the conference so i reached out to to him and asked if he would be interested in and it became very clear very quickly that that only did he really care but he is much more maligned in the media than the doomsday or scaremonger. You know, him with his incredible optimism that is why he spends his money on what other people think are impossible. To beyond the planet or all electric. And to think of that longerterm for future to take it very seriously and i tried very hard to persuade him that we need to get the a. I. Community engaged in this conversation that this is about a. I. Research talking about the power and the wisdom but that is due invest with those various kinds of Safety Research and the kindly agreed to be the first person to respond. But it was the first ever spent on this research for a witty and the world we have 300 team supplying applying. He almost did not make the announcement it was ted chin go. Touch and go. And having heart palpitations. What was the reason . The first ever landing of the rocket booster and this was a dree money is for so long that you cannot distract the media from this by having another big announcement the day before so finally our friend said we would make the announcement at the conference but you cannot tell the world until two days had passed. Because that number is the headline. Bat really transformed so whenever they go to conference they say this is cool that doesnt mean we can help. You also mention north korea. Natalie lethal autonomous weapons but a project to reduce nuclear proliferation. It is not nearly as hypothetical that to day then number three and sent a missile to japan . Do you have a concrete thing we can do about that . It is very much the same phenomenon that we use through science to amplify their own power more and more. View never walk into kindergarten here is a box of hammer and nails for go play with it. I read tweets by a certain person but take this box of 4,000 nukes and play with them even today that could be handled lot better. To hire that determines how many Nuclear Weapons do you need . One hundred . 500 . That went down a the list and then i was in massachusetts. Why do we have 7,000 . Because russia does. If they came together to say cut by 1,000 each the deterrence is unchanged would not reduce our deterrence against kim jong goon either but but with bad risks that technology poses. As physicist and technologist like elon musk keeps developing you can hurt a few people with a machine gun or a grenade but if you kill millions of people now the next wave literally could be existential. That their technologies that will destroy humanity please do not push. Pointing out to each one on the planet and how long can they survive . There were those during the stone age you thought it would be a good idea but not so many as this society we need to get together with these a. I. Type of weapons this is why we need to make sure we dont get into in arms race with a. I. Weapons. These will just cost to wonder dollars within amazon delivery anybody could just program them. Somebody to just put in the address and the photograph of their ex girlfriend is a horrible situation we just dont want to be at. That was ted weaken russia and the u. S. Pro. And one to make sure before we turn over so i like to ask how a. I. Has a responsibility but then he said shame on us but i would like to push too little the specifically how would regulate to see this happen . For anybody against wealth distribution. My long term is not as long as yours may be five for 10 years over 20 horror 30 there is no shortage of work but one example for the attitude do some of the new jobs. What most economists love is education and business spending more but reinventing. With Emotional Intelligence these are things we dont teach very well so we want to reinvent where people continuously learn new things. There is more of this conceptual change never raising fact or following instructions so with m. I. T. Doing more with people involved with projects but another leg is entreprenuership you may be surprised to learn actually it has gone down but technology and Silicon Valley talk about or entreprenuers but overall there is less new business formation less jobs created and that makes it difficult to invent the new jobs replacing the old ones that successful strategy is agriculture and many others we didnt say hang onto those jobs, we invented goods and services called Creative Destruction we have to do more to encourage that. Even more and more regulations and barriers with this negative with the special tax on uber or even toronto where was banned altogether but then we distribution we have a system right now that has shifted more money even though most of that wealth has gone to the top 1 percent and that was to exacerbate back in the 50s and 60s i would not recommend that but certainly ill be want to balance things and in other countries like in sweden they have been very successful to invest in health and education and infrastructure and child care. It isnt giving people money but making life a little better those of the three things we can do so i do share your optimism they could do everything humans could do people settle my god machines will take all the jobs but shame on us talking about wealth with no need to work is a bad thing . Now that means we can spend our time having discussions playing and doing sports but that is a social stories. I agree with all of the things that you said to have that but of course, we have to think i know a lot of people who never worked a day in their life. So i do agree it has evolved but can we do that . And do you think we can do this . Now with the latest tax reform going in the opposite direction . One of the reasons were having this discussion is it is about making choices that the way it works in a democracy more or less is we the people decide. And they say we like that but unless the voters are clamoring we have to listen to what they say. So were changing the conversation what do we want to do . And that is the choice. That is a good note my understanding is theres a couple of microphones . Please make sure they are questions. [laughter] good evening have a question of the economic consequences so right now lot of people will be out of a job including the Truck Drivers to say it is a Political Choice but on television you dont hear those difficult choices. What are your ideas . I think part of that political this function is driven by the economic trends. So half of the people were worse off. Federal blame people for being angry. Things are not Getting Better for them. But those Underlying Forces will get stronger and stronger. Tens of millions of people will lose their jobs but is it tens of millions of new jobs . So far we have done the okayed job but unemployment is higher now than it used to be but it is more about wages. And to create new and better jobs. With childcare and health care teaching and science the money is also there. We can also a list of private sector more. Over 300 companies and organizations doing things to create jobs with the government helps to support that there is something called the earned income tax credit Silicon Valley has proposed to back that but it is tied to work if you work more with the low income job they will supplement 8 an hour turns into drover 15 not by having the of minimumwage but the employer hires them like they did before now they get more money so people do feel there want to be a part of society but right now if you just write them and check middlefield satisfied. Some of my sociologists trends they went the way to contribute. I will stop here there is a lot more we could do but we would give other people a chance. Thanks for being here to you both. In recent years many people of taken to the streets to solve this problem by asking with every nonviolent revolution i see most technological advances presenting that what do you see as the most least violent way with the nature of the collection of power . Can we use Artificial Intelligence . Batted an interesting question. But certainly one could use positive social change but i feel what is happening the anchor is felt it is very real with major cuts in the availability of Higher Education they have not had the opportunity to learn the details that is true for any demagogue so what you see in recent elections they are hungry for change but obama is slow did was change also about was Donald Trumps and so was brexit i think it would be really helpful to build that social movement to show People Solutions that actually work. With that machine platform and crowd that they could use that to be disruptive and above them to transform society. Al least getting power to the crowd whether social media or other tools but i just want to push back from one part and not sure the answer is to overthrow but we have had some institutions that are incredibly valuable talking about those scandinavian countries but if we dont dont, if we fail to live up to the promise with these institutions were originally designed for then they would be overthrown with that revolution may have seen could have been here. So those people right now should be very aware of that then people will come after them with the pitch fork like this guy. [laughter]. A thing we can address those fears of Artificial Intelligence field the way to win is not to play. We know the secret is almost always not compete any intelligent computer will learn that . Bid is a very good question how to make them cooperate . I think one of the most powerful ways is to think about positive shared vision why do they agree to give up freedom to be married because think of the cool things marriage would enable. So waterweed doing as a society . If you go to the movies actually that is the opposite of what we have to do and then asking for career advice i say what do you want to be in the future . In 20 years we could plan the future this is where i want to me we as human kind need to do that why i wrote my book faugh what sorts of future you are excited about . To foster the collaboration that if we Work Together we can get there. That seems like the perfect way to end with cooperation and how do we Work Together to envision a positive future . So we will continue this conversation at the reception and book signing so you have to exit the theater then turn right follow the signs. There are amazing comments and concerns this has been very i opening. [applause] [inaudible conversations]