Documentary filmmaker james barrat talks about runaway Artificial Intelligence. Mr. Barrat argues we are headed into the future in which machines will be able able to outthink humans and may eventually view is in the same way we view lower life forms today. This is about 50 minutes. [applause] hi. I am james barrat and i want to thank mary and jenice for inviting me here and i would like to thank cspan for being here. I think this is such a great bookstore. It is good to remember when books really were sold in bookstores and this was kind of a miracle because it manages to me, kind of captures the wild and mysterious bookstores from when i was a kid. They were just a little bit spooky. [laughter] but very rich and fascinating. That is what this one is and obviously its also such a great Meeting Place for people. I want to thank them for inviting me. I have written a book about Artificial Intelligence but my usual job is it documentary film producer. Ive made a lot of films you mightve seen on the National Geographic network and pbs. Some are available on netflix. It was through the documentary that i became interested in Artificial Intelligence and that is what im here to talk to you about tonight. Artificial intelligence what it is and why i think and a lot of ai researchers think its being developed in the wrong way. I hope to give you some things to think about because i really believe that this conversation is the most important conversation of our time. So lets begin with this. What is Artificial Intelligence . It is a theory and developmedevelopme nt of Computer Systems able to perform tasks that normally require human intelligence and such as visual perception etc. The human intelligence phrase or the whole idea of ai back to humans because by and large what we know much about his human intelligence and human intelligence with ai is both this study, the subject of study and the tool of which we try to penetrate what intelligence is. This is what makes ai fascinating to me. So most inward looking of any of the sciences. It involves psychology narrow science medicine statistics and a lot more on top of programming and Computer Science. It makes us ponder what it is we are looking for when we hope to mirror human cognition and machines. The science of ai asks us what to humans do . What are you . What is intelligence . There are a lot of definitions of intelligence in the ai research business. I like this concise one this is intelligence is the ability to achieve goals in a variety of novel environments and to learn and theres a lot packed into that definition. Its as intelligence is goal oriented so that the intelligence is not doing something its not displaying intelligence. Intelligent should be mobile and probably comes with a body although if you cant move around and adapt your intelligence may be poor quality and theres really no way of tee around you need some sort of oddly and you must learn from experience and this is a real important one for ai. Most animals come with all the abilities they will ever have. We can learn new languages jobs crafts etc. And of course other animals can learn that nothing like the scale of humans because we have intelligence. Ive i have been interested in ai for several decades but i really got bitten by the ai bug and was learning for the working for the learning channel i was making a film about Artificial Intelligence and i got to interview a man who was my hero at the time ray criswell. Ray criswell as you know as a pioneer of Speech Recognition Technology machines that read hooks to the blinded many other inventions. He has been called the Thomas Edison of our time. He is the man who coined the term singularity or rebranded the term singularity. He is in charge of the process to reengineer the brain and most researchers i spoken with think reverse engineering the brain is the fastest way to create artificial general intelligence which is human level intelligence and it might be something you want to look into. Its quite fascinating. I also got to interview another hero of mine back then rodney brooks. Rodney brooks is the foremost robotocized of our time. The company who founded it makes irobot and he has sold and moved on. This is a general purpose robot called baxter that is designed to learn to do things in your home or in factories. He imagines working on farms. But he also irobot makes robots but also a lot of battlefield robots that carry guns. Theres a debate going on right now thats a very important debate about whether or not at all filled robots whether they should make the kill decision without a human and the question of drones too and i will get into that later. Im taking to irobot robots. They are going to help us. Yeah they are going to help us. They are called first looks and they will help us go into a pyramid that hasnt been explored as got a lot of rockfall so all the passages inside we cant get through to. They will be excavated eventually while we are there but the first thing to do is put these robots in and get a sense of what is the fastest way to the burial chamber, what does the overall layout look like . And obviously dont let the title of my book mislead you because i really do like robots. Criswell and brooks were optimistic about the time that is coming when we will share the planet with machines that are smarter than we are. Criswell in his nonfiction books predicts ai will help us solve every medical problem facing us including the overall general problem of mortality. But after criswell and perks i interviewed martha c. Clarke. Most of the decisions affecting our lives will be made i machines. So i began asking followup questions will that transition be friendly. Will that be a handover handover or a takeover . Will we change ourselves to become machines with rain modifications which is criswells singularity or will be create machines smarter than ourselves and will while those machines somehow replace us . What i learned is if we see them on the course we are currently following and i want to explain why we will create Intelligent Machines that wont be benign or harmless. They will start out being tools to quickly we will become their tools if we continue to exist at all. My book is called our final invention Artificial Intelligence and the end of the human era. The books thesis is that we need to develop a science for understanding smarter human intelligence before we create it. The two years i spent writing the book drew me into a world of people who would i spend driven to create smart machines working at high levels on ai and knew they wanted to create smart rashean since they were teenagers and children. It also immersed me into the lives of all who are just as determined to stop the reckless use in the Reckless Development of advanced ai. The two years i spent writing the book were two years of the most intensely enjoyable but also the most harrowing because i went looking for, i got more than i bargained for. I went looking for a fish and i found a whale. I found more bad news than i wanted to find. So lets jump in. How did we get from smart ounce in our pockets to super Intelligent Machines that could threaten our existence . Let me ask you a question with a show of hands. Do you think scientists can make a machine as smart as a human . So its not than the problem is either too hard in an engineering sense or there is something about the human brain that defies engineering so who thinks the problem of intelligence is too to harden this is a legitimate, it could be a legitimate problem. It could be too hard. It may go to the next century. So the problem is either too hard or there something that is magical or mysterious about the human brain that cant be duplicated and who is on that side . Less than 15 of the professionals i spoke with or that polled believe the problem is too hard and none think theres anything magical about the brain. Being ai special as they are going to think that that i did a later poll of experts and nonexperts and combined them with a couple of graphs. My conclusions with them there is nothing unfathomable about the human brain. Given time we will create human level intelligence and then beyond. So in that case if you follow that path, you might not have been a where we just split off into different tribes i think that if you follow that, if you follow we can craft this problem and is just a matter of time, will it take 10 years or will it take 100 years . If intelligence is a problem that can be solved how long will it take . Ray criswell has been good at technological progress. The main date according to others i pulled was about 2045 which is in many of our lifetimes. The outside data from the specialists and nonspecialists was about 2200. Gary marcus of the new yorker magazine is a psychologist and neuroscientist and he was kind enough to review our final invention for the new yorker and he said this about how long it would take pity set a century from now nobody will care how long it took. What they will care about is what happened next. Its likely the machines will be smarter than us before the end of the century he said. So whats important says marcus is not when super intelligent shows up but what happens next. In other words will we be ready . Will we have prepared ourselves . Even ray criswell who is supremely optimistic believes Machine Intelligence will surpass our ability to understand it but how exactly so my question was how exactly will that happen . How will machines get smarter than us . Theres a Pretty Simple theorem put out by ig good in 1965 and there are are a couple of chapters about ig good. Hes a unknown scientist who worked as a crook code raider code breaker. Im going to give you time to read it. Im going to give you 20 seconds to read it. [laughter] i like the formula that i would like to put another way. We have created machines that are better than us at chess jeopardy now chess jeopardy and other tests like navigation and theorem proving and lots of other things. Probably within the next decade we will create machines that are better at ai research than we are. We should build su improve capabilities quickly. These machines will jump from human level intelligence to superhuman intelligence in a matter of days weeks or years. How close are we at software that improves his self . Now software exists and observes physics experiments derives hypotheses and make suggestions for further experimentation. Software to write software exists. Software the judges the quality of software exists. So a Software System that improves itself is within reach and their good attempt at doing that right now. If you think of the field of algorithms there is a lot that are capable of improving their own performance. Intelligence selfaware Software System is another thing. Thats general intelligence. When that cell phone grooving selfaware software exists it will share the planet with smarter machines. So that takes us back to the question how will we get along without . What makes us think we will understand machines to science. Watson beat the humans with an amazing collection of cognitive powers pattern recognition decisionmaking statistical reasoning hypothesis generation, hypothesihypothesi s generation is very important. We do pattern recognition when we pick out patterns in the crowd. We statistically weigh them all the time. For example how long will this guy talk on and on about Artificial Intelligence . You could hypothesize 10 minutes or 40 minutes and you can test the hypothesis. What is watson doing today . Watson is being trained to take the federal medical licensing exam so watson yeah. Watson is performing medical diagnostics and Drug Research so watson is beginning to do legal research. It wont be a consulting physician right away but it will it be a physicians aid and they want to license it so that they can avoid certain kinds of liability. Overall how good our ais cognitive cognitive functions in 2014 and this is what i bring up with people who say it hasnt gone anywhere since the achievements are few. You know that ai cognitive functions are pretty good when they start taking our jobs when they compete in the job market. Heres a shortlist of jobs were humans are being replaced by machines right now by ai and automation and automated intelligence. Sportswriters travel agents bank tellers manufacture jobs of all kinds postal workers Clerical Workers pharmacists. All of these are being computerized. Soon to be replaced by computers in medical diagnosticians. When selfdriving cars, about right now our are average, the average driver you grade them the messes c or a d. When we raise driving two b overall we will all be happier but we will all be drivers. Astor notts pilots soldiers software developers. A recent article in the m. I. T. Technology review said 45 of all jobs will be automated within the next 40 years. It i think thats a conservative how close is human intelligence and machine to being obtained . Is so close that reaching human intelligence is job number one for a lot of companies and governments. Why would a lot of companies and governments support millions of dollars in creating virtual brains . And the answer is because an artificial brain at the price of the computer would be the most lucrative product in the history of the world. An artificial brain at the price of the computer will be the most lucrative product in the history of the world. Imagine banks with 10,000 ph. D. Quality brains working 24 7 on things like Cancer Research climate modeling, Weapons Development. Imagine that product being offered by several companies that compete to drive the price down. Who wouldnt want that technology and who wouldnt want to be first to create that technology . This is a short list of the people who are going for agi and pouring billions of dollars into it. Companies like ibm google vicarious deep thought organizations like the department of defense nsa and darpa. China and israel want asi and it said as much. The European Union gave a billion euros to switch project called blue but rain ruined by gary martin which is reversing the brain. He gave the people who think about ai risk. Deep mind was bought by google for 400 million. The founders of deep mind and including shane legg whos been writing about ai for a long time before he became a millionaire said a condition of this the sale will be that google sets up a board for ethics and safety governing their technology. This is a giant milestone. They are technology and that advanced ai is risky and they are also setting a very high bar for future purchases. So once these guidelines get out the ones that deep mind is creating with google once those guidelines get out there will be Something Like the guidelines for dna and im going to get to that in a minute. The Industry Needs guidelines and it just happened two weeks ago. All of us thinking about these issues were just gob smacked and pleased. That is a great acknowledgments that these are issues that can hold up the 400 million sale and if google doesnt appoint this board or points aboard the doesnt do anything while all this show her holders from deep mind have a lawsuit in google is not afraid of lawsuits as you know but they will have to prove themselves in it court of Public Opinion about why they are not taking the ai risk seriously. Whats the one thing these groups have in common is they know the steampowered of the 19th century and electricity of the 20th this is ais century. So you might ask yourself what could possibly go wrong . Already we rely on machines for many things and smart machines and is working out okay. On wall street are carried out by algorithm. Our energy and transportation or infrastructure rely on an Automated System using ai. Our Banking System relies on systems using ai. Computers are everywhere and we depend on them but how do we jump to he is an ai make her who has created signs for understanding how artificial super intelligence will behave. His work is really important. I have two chapters about what he is doing. To analyze super intelligence he uses rational aging theory for economics. In brief rational aging theory posits that rational agents humans or machines which work to maximize preference is called utility functions. That makes them predictable. When economist posed this they quickly learn you can work for humans. We are not rational all the time time and we are impulse buyers. You cant base an Economic System on the logic of people but cant probably anticipate that smart machines will be logical and therefore rational in that economic sense. He argues that the rational systems that are selfaware and self improving will develop basic drives and these drives include resource acquisition selfprotection efficiency and creativity and how it works is like this. Self improving machines will pursue the goals they are created with. Whether that goal is to play chess or to pick stocks. To succeed they will need resources whether thats energy money hardware or whatever is most expedient. They wont be satisfied with trying to fulfill their goals. They will also seek to avoid Failure Modes like being turned off or unplugs. In other words they will protect themselves. They will be efficient and they wont squander resources. They will use their super intelligence to find creative ways to achieve their goals and since improving the software will be one winning route to the success they will grow their own intelligence at computer speeds. Now here is the rub. Being friendly isnt on the list of ai behaviors. Super intelligence doesnt apply. Being smart doesnt mean being kind. Super Intelligent Machines are dangerous because they will try to overthrow us but for any goal machines might have it will be useful for them to use all available resources to achieve that goal. All available resources includes virtually e