Go is an example of a machine algorithm that was to play against the ancient game of go. Its vastly more complicated back in 1997. There are more possible moves in the game of go then there are atoms in the universe and that means the traditional technique of trying to search exhaustively , even a tiny fraction of the moves is completely infeasible. For that reason, as recently as 2016, computer scientists were surveyed when they thought machines would be able to beat humans at the game of go and they predicted probably around 2027. That was the median estimate. As some of you in the audience know, they did beat the chip World Champion and gold last year end they did it by using a new technique by looking at the board and seeing patterns. Thats the way humans do it. If you ask people how they play it, they can explain it. They say they look at the board when they see a move that seems right but they can explain how. They cant describe strategy or tell you exactly what theyre doing. You can imagine, if you cant explain it, how would you code it. With this new approach to Machine Learning and successes and failures, they can figure out the pattern on their own by applying reinforcement learning p they were babe able to very quickly learn the game of go to where they could beat humans at the game, and now human champions, when they see with the machines are doing, they say we have not even touched the edge of the game of go. We have been studying this for 3000 years but the machines are so far ahead of us that we realized we have no conception of what is possible. The reason we think we are still underestimating the impact in the power of Machine Learning, despite the fact that we all read about it all the time is exactly what he just said. Here is the domain of intent human study for three millennium we built up this body of what we thought was really exhaustive knowledge about how to play this game. This piece of Machine LearningTechnology Comes out of nowhere and shows us how much more headroom there is in this particular domain. Thats pretty amazing. We think about all of the other other incredibly complex things that were trying to get good at, whether that is solving the mystery of cancer or figuring out what to invest in or viewing the human genome, we now have a colleague for that work and it might be able to show us some ways forward. This last point is really key. Sometimes its called the second wave of the second machine age. Thats the idea that the machine starts doing mental tasks. The fast past few decades weve programmed them. Weve codified our own knowledge and included it so the machine to do it. That tells us how too do tax preparation or spreadsheets but its all stuff we knew exactly what needed to be done and we had to be very, very precise and the machine would do the same thing. Thats great. Theres so many problems that it sounds paradoxical. I wouldnt know how to recognize my mothers or ride a bicycle. These are just things we can do intuitively. They are giving them example of what works and what doesnt work this is the word yes or no in french or english in the machine , if you give them enough examples, they will figure out the patterns on their own. Its a remarkable transition that has been important in the last few years. There was an example that went viral on twitter a while back. It was, could the machine tell if it was a blueberry muffin or a dog. There was another one that came out the same time. There was the lab or doodle or the kentucky fried chicken. This is a story that really looks works better if you look at the picture because if you dont see someone similarly looking at the pictures they can be hard to tell apart, but we humans can often do it and was remarkable is that seven years ago, machines had about a 30 air rate and that was using the best technique. Now its down lower than 5 which, importantly is less than the air rate of humans on that same data set. Weve kind of crossed a point. Is not just a matter of degree but when you raise the temperature of water past the boiling point its a fundamental transformation. If a machine can do a task better than a human, then if youre an entrepreneur or or manager and you have to make a decision, who do i sign, youre in the one i can do the job better. The one thing i was hoping we could talk about is the idea that in the future or now, people will be learning from machines in addition to teaching them. One thing i liked about the go story is that after the World Champion was defeated, he said so beautiful, so beautiful, about the moon because it wasnt something hed ever conceptualized before not something we could see in a lot of fields. And were seeing it over and over and youre right about people earning side sidebyside with machines. You talk about the green guy with a widely televised matched. Its part of the reason there was a downward turn in the Green Economy for a while, ever but he to pay attention. He was beating about 75 of human opponents before he played apple go. After he had done and dealt with that technology for a while he was beating about 90 of his human opponents. The other thing that happened is the sales of go boards in korea and ages asia skyrocketed and little kids started playing the games and much higher volumes. One thing we hear is isnt it terrible that were no longer the best of this activity, doesnt that reduce human value or dignity. I think thats dead flat wrong. If we do ourselves by the things that were best in the universe, thats a pretty silly thing to do. We now have this neat thing happening that people want to be a part of. If technology is part of that, i think its good. What do you think. I think its good. Let me under one of the points he made. This machine came at the game from a fundamentally different way. When they saw some of the moves the machine made, the go expert said oh my gosh, that is idiotic every beginner knows not to do what the machine just did. A few minutes later, guess what. The machine one. Then they were go, i see what it did there. That is really innovative. For 3000 years theyve been learning from the master and if they made that move their master would say stupid human, dont do that, but the machine came out with a different way of thinking when they came up with an innovative approach and its not just true in the game of go but its true in all kinds of problems. Theyre not just better in certain ways, but more importantly, theyre different in certain ways and they come out with a different perspective thats hoping that human machines can work together. Whats cool about that story, lots of things are cool. You probably needed to be a player somewhere near the level of lisa to even get the insight of why they moved the goal that they did. You have to be able to unpack to why the machine was behaving that way. We need really highlevel humans to look at that and say i think i know whats going on, lets go test that idea. Where are some companies, parts of the economy where were arty seeing this, instead of a replacement of people were seeing like a mutual collaboration and learning. Let me describe, is not in the book, its something i learned afterwards but its in the spirit of the book. One of our friends with the company you daphne, its an Online Learning company that has a bunch of things you can learn and 90 degrees, 90 days so people come to the website and they have a with sales reps saying is this course right for me, what shall i be doing. So when they call the human sales representative, some of them are better than others and some of them are able to answer the questions effectively, come for the people and commence them its a good course to take, others not so much. Its a Machine Learning expert and along with a grad student they recognize if they take the transcript of the chat, that becomes the data for whats called supervised learning. Thats basically giving examples of successes and failures to the machine and let the machine figure out what matters. They gave these transcripts and the machine a logarithm started seeing patterns that the successful sales reps had different ways of answering questions. At that point, they did not turn it over to the boss because there were still too many unstructured questions that the machine would be good clueless about but there are some very common questions that account for a certain percentage of the inquiries. They still have the humans doing the interactions with their incoming traffic, but the bot would watch them, and one of these common phrases or questions came up, the bar would say heres a phrase you might consider using. Its worked really well in the past. Then the humans would pick up on that. Especially the not so good sales representative, their performance improved dramatically and sales went up to about 50 . It was a tremendous improvement and it was the ability to combine humans and machines in a way that neither one would have been successful on their own. And we should be clear, in some cases we are going to automate work and thats the appropriate thing to do, but its naive to think that thats the unidirectional, the only force going on. We love that story because it shows the ends of technology and automation realizing they still need humans involved in this process, but we could kind of amplify their abilities with a lot of Smart Technology in the background. In the broader lesson thats been applied is that each of those pairs, were not thing time turn the dial all the way to one extreme, worsening thinking about the balance. Sometimes it makes sense to go automated but in most cases you need to find balance in the balance is more toward machines platforms and cloud that i think more individuals realize, but its not as simple. This question comes up a lot. What you see is areas that are most resistant to the advent of machines and computers . We were talking earlier about a part in your book for you mention this famous saying, would you like to quote it. Andrew bird set a while back that the only thing, the only difference between a completely unlinked series of observation and a hit song is excessive confidence on the part of the songwriter. I think thats a lovely way to look at it, but hes lowballing himself because it would be trivially easy to program a computer to spit out observations that might rhyme and might have words associated with love and lost in them. You can program a computer pretty easily but we are incredibly confident that the lyrics would probably not make a lot of sense and would not activate any emotions on the part of the rest of us. What great human lyricist do, they have a fairly linked observation that resonates with other human beings, and thats a skill we have not seen Computers Come up with yet. Its a common set comsat engine concept from linguistics. What it means is if i look at almost any length of can plex sentence in english, i can immediately tell if its canonically correct or not before my brain has processed its meaning or what its trying to convey. My bet brain is trying to determine if its kosher or not. I spent speak some french but i cant do that to save my life. English as second language individuals cant do that anywhere near as well as i can or any other native speaker. We have intuition about this physical world that we move through, trying to keep teach computers to have that, if we ever do that, i believe thats decades away. I have a lot of confidence that the human ability to invoke an emotional response via words and language, in fact its automated quickly, im actually going to be surprised. But at the same time we actually have a logarithms that are starting to compose music. Its hard to tell apart from the music of famous composers. Not hard, actually impossible so the interesting question is, what is the difference between music and lyrics and what about that makes it automated. This is getting a far away from my area of expertise, but the aesthetic of music is well known. We know what progression sounds good to our ear and which sounds like sirens on the new york street. We want to do more the former and less of the latter. You can encode that stuff and you know the rules, you know these kind of things so its relatively easy to hit the button and send those things out and he goes beyond that. The machines that are doing a really good job, theyre not just encoded with the rules of a few, but there are patterns that emerge. You dont have to know anything about the structure meaning of the music. Machines are actually quite good at the shallow pattern recognitions, but the lyrics, to really be proactive you need to understand the meaning. Its not just the two words that often appear next to each other in sequence. If you want to go beyond that you need to understand in the machines, although they can recognize patterns and words, they still do not really understand what its doing. They can translate one language to another but theyre not actually understanding whats going on. Thats very different. And machines can be very good at certain kinds of video games that are quick reactions shooter games, great at that, but if they cant involve some kind of deeper planning and structure, they are hopeless at that. I think language is a little more like complicated and music is more like the pattern matching. We included in our book, and exert from a system that got trained on jane austen novels. Its not even gibberish. It actually makes your head hurt it is are really unpleasant too so if we can make that go 180 degrees difference, i dont think that will happen anytime soon. Theyve often said when you look at jobs and what job services are likely to become automated that its not the heiskell low scale as much is repeatable and highfrequency test versus variable and unpredictable tasks. The broader lesson here is that machines are not just good, theyre superhuman at certain types of tasks. They been superhuman at arithmetic for decades so we shouldnt be surprised. Now theyre getting superhuman that pattern recognition. When things are repeatable, you can get the data to be able to train the system. If they involve more complex planning, creativity, those are things that are much harder to have a machine. One thing we want to stress as were talking about whats available 2017, theres a bunch of very bright people working on all sorts of things that may be breakthroughs in 2018 or 2080, were not sure because they havent happened yet but based on current technology, theyre basically taking a bunch of input and mapping them into a bunch of outputs and figuring out the mapping in the really good at that if theres not too much complexity between the act s and why. Think that framework is rapidly becoming out of date. Id never wouldve thought of medical diagnosis as any type of routine, repeatable work that he would hand off to machines. Its an incredibly important and difficult and subtle and complex pattern matching exercise. Machines are better at those things now that we are and make come out of essentially nowhere and have reached superhuman level of performance. I think medical diagnosis is squarely in the sight of these automatable tasks and one we should apply a lot more machines to but that did not appear to be the case five or ten years ago. If you look at image recognition, we talked about recognizing dogs and cats and faces, but guess what camille can also have some medical images. Heres pictures of lungs. Which have cancers in which dont. Pathologists and radiologist took years to learn what it took for look for. You train a machine and they can barely very quickly learn the markers taught in medical school and other markers that no human has ever noticed before. What is the accuracy . You also talk about how machines are rapidly eclipsing human experts in their own subject. This is a number that is changing as we speak. Even since we finished the book there has been significant article published in february of this year. That example we gave of giving 30 was an example for me. In Voice Recognition, understanding speech, it went from 8. 5 error rate to 4. 9 error rate, its not just the improvement but that improvement was done over the past ten months, 11 months since july 2016. As we speak, they are improving it more. A lot of it just comes from taking the same set of a and a lot more data and a lot more computational power. With medicine, the studies weve come across our little conservative. They generally say properly done Machine Learning system will be at human level performance. I think thats probably lowballing it, i think theyre probably better than the average doctor in a lot of cases today. So where are most people going to first really encounter this kind of technology. Is it alexa, is she saying buy me whole foods. You do get those kinds of errors, but i think a lot of it, one example is Voice Recognition its clearly not perfect yet, it still makes those kinds of errors but its been measurably better to the point where i now dictate rather than type a lot of the texts that i write and theres a study of the stanford that dictating opposed to typing is about three times faster and more accurate for the applications they looked at. Weve encountered some of them, probably the leading example, the one most obviously profitable has been deciding which ads to serve up next to your searches and if they can figure out what youre likely to click done versus what youre not likely to click on, that is worth billions upon billions of dollars. That is one that we actually encounter quite a bit. The insights that we get are being translated over into all sorts of other domains. The main way were coming across these cuttingedge technology is via devices like these, speaking into them, having them understand what we want, getting decent quality translation, type out a few letters and give them to you. These are all Artificial Intelligence. You may notice, those of you who use facebook that it will offer to label the names of the faces of your friends. That is in the capital market. And that can be a more e efficient way those who have cars most of them are sitting idle maybe 50 percent . That is 10 times improvement. One of the weird outcomes is said go home in cambridge because i have a big expensive honk of metal now wishes to a colossal wave of opportunity to take bad infrastructure that is out there in the world to the warehouse capacity to make a much more efficient and you can be underwhelmed by that we should all care about efficiency if you want to make better use of our resources we will not walk away to be more efficient with their use but those are unparalleled assets to make greater use of those assets of the infrastructure up there in the world. Tehs anyone have questions . I will ask a hard question we have known each other for a while and there is that dystopia claim. Messines and i suspect you personally know the editorial did you see that . And he ran various pieces of google or microsof