WEBVTT - Artificial Intelligence 0:00:03.360 --> 0:00:06.400 It was a cold, windy day in January nineteen seventy 0:00:06.480 --> 0:00:10.880 nine when the robots took their first human life. It 0:00:10.960 --> 0:00:13.800 happened in Flat Rock, Michigan, about twenty miles down the 0:00:13.840 --> 0:00:17.920 interstate from Detroit, at the Ford plant. There. Robert Williams 0:00:17.960 --> 0:00:21.120 was twenty five. He was a Ford worker and one 0:00:21.160 --> 0:00:23.640 of the people who oversaw the robotic arm that was 0:00:23.720 --> 0:00:26.560 designed to retrieve parts from bins in the storage room 0:00:26.600 --> 0:00:29.080 and place amended carts that carried them out to the 0:00:29.160 --> 0:00:32.800 humans on the assembly line. But the robot was malfunctioning 0:00:32.840 --> 0:00:35.640 that day, and aware of the slowdown it was creating 0:00:35.640 --> 0:00:38.839 on the line, Robert Williams went to grab the parts himself. 0:00:39.520 --> 0:00:42.360 While Williams was reaching into a bin, the one ton 0:00:42.479 --> 0:00:45.880 robotic arms swung into that same bin. The robot didn't 0:00:45.920 --> 0:00:48.680 have any alarms to warn Williams it was nearby. It 0:00:48.800 --> 0:00:51.000 didn't have any censors to tell it a human was 0:00:51.040 --> 0:00:54.200 in its path. It only had the intelligence to execute 0:00:54.200 --> 0:00:58.080 its commands to retrieve and place auto parts. The robots 0:00:58.080 --> 0:01:01.520 struck William's head with such force it killed him instantly. 0:01:02.800 --> 0:01:05.120 It was thirty minutes before anyone came to look for 0:01:05.200 --> 0:01:09.000 Robert Williams. During that time, the robot continued to slowly 0:01:09.000 --> 0:01:11.520 do its work while Williams lay dead on the parts 0:01:11.600 --> 0:01:15.640 room floor. The death of Robert Williams happened during a 0:01:15.720 --> 0:01:19.200 weird time for AI. The public at large still felt 0:01:19.240 --> 0:01:22.000 unsure about the machines that they were increasingly living and 0:01:22.040 --> 0:01:25.280 working among. Hollywood could still rely on the trope of 0:01:25.280 --> 0:01:28.560 our machines running a muck and ruining the future for humanity. 0:01:29.040 --> 0:01:31.480 Both war Games and The Terminator would be released in 0:01:31.480 --> 0:01:35.320 the next five years. But within the field that was 0:01:35.360 --> 0:01:38.040 trying to actually produce those machines that may or may 0:01:38.080 --> 0:01:40.520 not run amuck in the future, there was a growing 0:01:40.560 --> 0:01:44.800 crisis of confidence. For decades, AI researchers had been making 0:01:44.840 --> 0:01:48.480 grand but fruitless public pronouncements about advancements in the field. 0:01:49.040 --> 0:01:51.640 As early as nineteen fifty six, when a group of 0:01:51.720 --> 0:01:55.560 artificial intelligence pioneers met at Dartmouth, the researchers wrote that 0:01:55.560 --> 0:01:58.000 they expected to have all the major kinks worked out 0:01:58.000 --> 0:02:01.080 of AI by the end of this semester, and the 0:02:01.080 --> 0:02:04.440 predictions kept up from there. So you can understand how 0:02:04.480 --> 0:02:06.960 the public came to believe that robots that were smarter 0:02:07.040 --> 0:02:10.760 than humans were just around the corner, but AI never 0:02:10.840 --> 0:02:13.960 managed to produce the results expected from it, and by 0:02:14.000 --> 0:02:18.120 the late nineteen eighties the field retreated into itself. Funding 0:02:18.200 --> 0:02:21.960 dried up, candidates looked for careers in other fields. The 0:02:22.040 --> 0:02:25.040 research was pushed to the fringe. It was an AI 0:02:25.120 --> 0:02:30.360 winter the public moved on to. The terminator was replaced 0:02:30.360 --> 0:02:34.440 by Johnny five in the film Maximum over Drive. Our 0:02:34.480 --> 0:02:37.240 machines turn against us, but it's the result of a 0:02:37.280 --> 0:02:41.280 magical comet, not from the work of scientists. We lost 0:02:41.280 --> 0:02:46.040 our fear of our machines. Recently, quietly, the field of 0:02:46.080 --> 0:02:48.600 AI has begun to move past the old barriers that 0:02:48.639 --> 0:02:53.239 once held it back, garner the grand pronouncements. Today's researchers, 0:02:53.240 --> 0:02:56.520 tempered by the memory of their predecessor's public failures, are 0:02:56.560 --> 0:02:59.840 more likely to downplay progress in the field, and from 0:02:59.840 --> 0:03:02.720 the new AI we have a clearer picture of the 0:03:02.760 --> 0:03:06.360 existential risks that poses than we ever had before. The 0:03:06.440 --> 0:03:09.120 AI we may face in the future will be subtler 0:03:09.360 --> 0:03:12.760 and vastly more difficult to overcome than a cyborg with 0:03:12.800 --> 0:03:21.600 a shotgun. In hindsight, it was the path toward machine 0:03:21.680 --> 0:03:24.600 learning that the early AI researchers chose that led them 0:03:24.639 --> 0:03:27.320 to a dead end. Let's say you want to build 0:03:27.320 --> 0:03:30.600 a machine that sorts red balls from green balls. First 0:03:30.639 --> 0:03:33.880 you have to explain what a ball is. Well first, really, 0:03:34.240 --> 0:03:36.480 you have to have a general understanding of what makes 0:03:36.480 --> 0:03:39.280 a ball a ball, which is easier said than done. 0:03:39.760 --> 0:03:42.560 Try explaining a ball to someone that doesn't use terms 0:03:42.600 --> 0:03:45.520 that one would have to already be familiar with, like sphere, 0:03:45.880 --> 0:03:49.960 around or circle. Once you have that figured out, you 0:03:50.000 --> 0:03:53.400 then have to translate that logic and those rules into code, 0:03:53.840 --> 0:03:57.000 the language of machines, ones and zeros if some. Then 0:03:58.040 --> 0:03:59.760 now you have to do the same thing with the 0:04:00.000 --> 0:04:02.760 incept of the color red and then the color green, 0:04:03.080 --> 0:04:05.600 so that your machine can distinguish between red balls and 0:04:05.640 --> 0:04:08.200 green balls. And let's not forget that you have to 0:04:08.240 --> 0:04:11.080 program it to distinguish in the first place. It's not 0:04:11.120 --> 0:04:14.440 like your machine comes preloaded with distinguishing software. You have 0:04:14.520 --> 0:04:17.440 to write that too. Since you're making a sorting machine, 0:04:17.560 --> 0:04:19.479 you have to write code that shows it how to 0:04:19.520 --> 0:04:23.080 manipulate another machine, your robot sorder to let it touch 0:04:23.120 --> 0:04:25.840 the physical world. And once you have your machine up 0:04:25.839 --> 0:04:29.000 and running and working smoothly separating red balls from green ones, 0:04:29.400 --> 0:04:32.159 what happens when a yellow ball shows up. Things like 0:04:32.200 --> 0:04:34.400 this do happen from time to time in real life. 0:04:34.839 --> 0:04:38.880 What does your machine do? Then? Despite the incredible technical 0:04:38.920 --> 0:04:42.520 difficulties that faced the field of artificial intelligence did have 0:04:42.600 --> 0:04:44.920 a lot of success at teaching machines that could think 0:04:45.080 --> 0:04:49.640 very well within narrow domains. One program called deep Blue 0:04:49.839 --> 0:04:53.080 beat the reigning human chess champion Garry Kasprov in six 0:04:53.120 --> 0:04:57.920 games at a match in to be certain, the intellectual 0:04:57.960 --> 0:05:01.200 abilities required by chess are of bad improvement over those 0:05:01.200 --> 0:05:03.360 required to select a red ball from a green one. 0:05:03.920 --> 0:05:07.040 But both of those programs share a common problem. They 0:05:07.080 --> 0:05:10.200 only know how to do one thing. The goal of 0:05:10.240 --> 0:05:12.800 AI has never been to just build machines that can 0:05:12.839 --> 0:05:15.800 beat humans at chess. In fact, chess has always been 0:05:15.880 --> 0:05:18.480 used as a way to test new models of machine learning, 0:05:18.920 --> 0:05:20.920 and while there is definitely used for a machine that 0:05:21.000 --> 0:05:23.800 can sort one thing from another, the ultimate goal of 0:05:23.839 --> 0:05:27.120 AI is to build a machine with general intelligence. Like 0:05:27.160 --> 0:05:31.119 a human has to be good at chess and only chess, 0:05:31.200 --> 0:05:33.480 is to be a machine to be good at chess, 0:05:33.600 --> 0:05:36.880 good at doing taxes, good at speaking Spanish, good at 0:05:36.880 --> 0:05:40.080 picking out apple pie recipes. This begins to approach the 0:05:40.080 --> 0:05:43.359 ballpark of being human. So this is what early AI 0:05:43.440 --> 0:05:46.560 research ran up against. Once you've taught the AI how 0:05:46.600 --> 0:05:48.680 to play chess, you still have to teach it what 0:05:48.800 --> 0:05:52.599 constitutes a good apple pie recipe, and then tax laws 0:05:52.600 --> 0:05:55.080 in Spanish, and then you still have the rest of 0:05:55.120 --> 0:05:57.760 the world to teach it all the objects, rules, and 0:05:57.800 --> 0:06:00.800 concepts that make up the fabric of our reality. And 0:06:00.880 --> 0:06:02.560 for each of those, you have to break it down 0:06:02.640 --> 0:06:05.800 to its logical essence and then translate that essence into code, 0:06:06.120 --> 0:06:08.479 and then work through all the kinks. And then once 0:06:08.480 --> 0:06:11.120 you've done this, once you've taught it absolutely everything there 0:06:11.200 --> 0:06:13.520 is in the universe, you have to teach the AI 0:06:13.680 --> 0:06:16.719 all the ways these things interconnect. Just the thought of 0:06:16.760 --> 0:06:20.440 this as overwhelming. Current researchers in the field of a 0:06:20.560 --> 0:06:23.400 I refer to the work their predecessors did as go fi, 0:06:23.560 --> 0:06:26.839 good old fashioned AI. It's meant to evoke images of 0:06:26.880 --> 0:06:31.280 malfunctioning robots, their heads spinning wildly, is smoked pores from them. 0:06:31.320 --> 0:06:34.040 It's meant to establish a line between the AI research 0:06:34.080 --> 0:06:39.000 of yesterday and the AI research of today. But yesterday 0:06:39.080 --> 0:06:43.279 wasn't so long ago. Probably the brightest line dividing old 0:06:43.320 --> 0:06:45.640 and new in the field of AI comes around two 0:06:45.680 --> 0:06:49.559 thousand six. For about a decade prior to that, Jeoffrey Hinton, 0:06:49.880 --> 0:06:52.359 one of the Skeleton crew of researchers working through the 0:06:52.400 --> 0:06:56.040 AI Winter, had been tinkering with the artificial neural networks, 0:06:56.279 --> 0:06:59.200 an old AI concept first developed in the nineteen forties. 0:07:00.000 --> 0:07:02.160 The neural nets didn't work back then, and they didn't 0:07:02.200 --> 0:07:04.760 work terribly much better in the nineties, But by the 0:07:04.800 --> 0:07:08.159 mid two thousands, the Internet had become a substantial force 0:07:08.360 --> 0:07:11.440 in developing this type of AI. All of those images 0:07:11.520 --> 0:07:14.640 uploaded to Google, all that video uploaded to YouTube, the 0:07:14.720 --> 0:07:18.200 Internet became a vast repository of data that could be 0:07:18.320 --> 0:07:23.920 used to train artificial neural networks in very broad strokes. 0:07:24.080 --> 0:07:26.840 Neural nets are algorithms that are made up of individual 0:07:27.000 --> 0:07:30.080 units that behave somewhat like the neurons in the human brain. 0:07:30.680 --> 0:07:33.720 These units are interconnected and they make up layers. As 0:07:33.720 --> 0:07:37.400 information passes from lower layers to higher ones, and whatever 0:07:37.480 --> 0:07:40.080 input has passed through the neural net is analyzed in 0:07:40.160 --> 0:07:43.920 increasing complexity. Take for example, the picture of a cat. 0:07:44.800 --> 0:07:48.040 At the lowest layer, the individual units each specialized in 0:07:48.040 --> 0:07:51.720 recognizing some very abstract part of a cat. Picture. So 0:07:51.800 --> 0:07:55.240 one will specialize in noticing shadows or shading, and another 0:07:55.280 --> 0:07:59.440 will specialize in recognizing angles. And these individual units give 0:07:59.440 --> 0:08:02.280 a confident at center bowl that what they're seeing is 0:08:02.320 --> 0:08:05.440 the thing that they specialize in. So that lower layer 0:08:05.600 --> 0:08:08.400 is stimulated to transmit to the next higher layer, which 0:08:08.440 --> 0:08:12.840 specializes in recognizing more sophisticated parts. The units in the 0:08:12.880 --> 0:08:15.440 second layer scan the shadows and the angles that the 0:08:15.520 --> 0:08:19.040 lower layer found, and it recognizes them as lines and curves. 0:08:19.720 --> 0:08:22.800 The second layer transmits to the third layer, which recognizes 0:08:22.840 --> 0:08:25.840 those lines and curves as whiskers, eyes, and ears, and 0:08:25.920 --> 0:08:29.200 it transmits to the next layer, which recognizes those features 0:08:29.240 --> 0:08:33.440 as a cat. Neural nets don't hit a accuracy, but 0:08:33.480 --> 0:08:36.640 they work pretty well. The problem is we don't really 0:08:36.720 --> 0:08:40.400 understand how they work. The thing about neural nets is 0:08:40.400 --> 0:08:43.000 that they learn on their own. Humans don't act as 0:08:43.040 --> 0:08:45.480 creator gods who code the rules of the universe for 0:08:45.559 --> 0:08:48.079 them like in the old days. Instead, we act more 0:08:48.080 --> 0:08:51.360 as trainers. And to train a neural net, you expose 0:08:51.440 --> 0:08:53.319 it to tons of data on whatever it is you 0:08:53.360 --> 0:08:56.440 wanted to learn. You can train them to recognize pictures 0:08:56.440 --> 0:08:58.960 of cats by showing them millions of pictures of cats. 0:08:59.520 --> 0:09:02.040 You can train them on natural languages by exposing them 0:09:02.080 --> 0:09:04.840 to thousands of hours of people talking. You can train 0:09:04.920 --> 0:09:07.040 them to do just about anything. So long as you 0:09:07.120 --> 0:09:10.680 have a robust enough data set. Neural net's find patterns 0:09:10.679 --> 0:09:13.439 in all of this data, and within those patterns, they 0:09:13.480 --> 0:09:17.640 decide for themselves. What about English makes English english? Or 0:09:17.679 --> 0:09:19.720 what makes a cat picture a picture of a cat? 0:09:20.080 --> 0:09:23.240 We don't have to teach them anything. In addition to 0:09:23.280 --> 0:09:26.280 self directed learning, what makes this type of algorithm so 0:09:26.400 --> 0:09:29.439 useful is its ability to self correct to get better 0:09:29.480 --> 0:09:32.360 at learning. If researchers show a neural net a picture 0:09:32.360 --> 0:09:34.360 of a fox and the AI says it's a cat, 0:09:34.720 --> 0:09:37.560 the researchers can tell the neural net it's wrong. The 0:09:37.640 --> 0:09:40.360 algorithm will go back over its millions of connections and 0:09:40.440 --> 0:09:43.240 fine tune them, adjusting the way it gives each unit 0:09:43.520 --> 0:09:45.320 so that in the future it will be able to 0:09:45.360 --> 0:09:48.720 better distinguish a cat from a fox. It does this too, 0:09:48.760 --> 0:09:51.840 without any help or guidance from humans. We just tell 0:09:51.880 --> 0:09:54.680 the AI that it got it wrong. The trouble is 0:09:54.800 --> 0:09:57.080 we don't really know how neural nets do what they do. 0:09:57.600 --> 0:10:00.880 We just know they work. This is it's called opaque. 0:10:01.320 --> 0:10:04.360 We can't see inside the thought processes of our AI, 0:10:04.520 --> 0:10:08.080 which makes artificial neural nets black boxes, which makes some 0:10:08.160 --> 0:10:12.520 people nervous. With the black box, we add input and 0:10:12.559 --> 0:10:15.640 receive output, but what happens in between is a mystery. 0:10:16.160 --> 0:10:17.920 Kind of like when you put a quarter into a 0:10:17.920 --> 0:10:22.200 gumball machine. Quarter goes in, gumball comes out. The difference 0:10:22.280 --> 0:10:24.920 is that gumball machines aren't in any position to take 0:10:24.960 --> 0:10:27.439 control of our world from us. And if you were 0:10:27.480 --> 0:10:30.240 curious enough, you could open up a gumball machine and 0:10:30.280 --> 0:10:33.280 look inside to see how it works with the neural net. 0:10:33.400 --> 0:10:36.520 Cracking open the algorithm doesn't help. The machine learns in 0:10:36.520 --> 0:10:39.800 its own way, not following any procedures we humans have 0:10:39.840 --> 0:10:42.160 taught it. So when we examine a neural net, what 0:10:42.240 --> 0:10:47.200 we see doesn't explain anything to us. We're already beginning 0:10:47.200 --> 0:10:49.960 to see signs of this opaqueness in real life as 0:10:50.000 --> 0:10:52.760 reports come in from the field. A neural net that 0:10:52.800 --> 0:10:56.280 Facebook train to negotiate developed its own language that apparently 0:10:56.320 --> 0:10:59.560 works rather well in negotiations, but doesn't make any sense 0:10:59.600 --> 0:11:03.280 to human Here's a transcript from a conversation between age 0:11:03.280 --> 0:11:06.720 and a and agent b Alice and Bob. I can 0:11:06.760 --> 0:11:09.240 I I everything else dot dot dot dot dot dot 0:11:09.240 --> 0:11:11.680 dot dot dot dot dot dot dot dot balls have 0:11:11.800 --> 0:11:13.439 zero to me, to me, to me, to me, to me, 0:11:13.559 --> 0:11:15.640 to me, to me, to meet you, I everything else 0:11:15.679 --> 0:11:17.960 dot dot dot dot dot dot dot dot dot dot 0:11:18.000 --> 0:11:20.040 dot dot dot dot balls have it ball to me, 0:11:20.080 --> 0:11:21.360 to me, to me, to me, to me, to me, 0:11:21.480 --> 0:11:23.800 to me, I I can I I I everything else 0:11:23.840 --> 0:11:26.120 dot dot dot dot dot dot dot dot dot dot 0:11:26.120 --> 0:11:29.480 dot dot dot. Another algorithm called deep patient was trained 0:11:29.480 --> 0:11:32.440 on the medical history of over seven hundred thousand people 0:11:32.880 --> 0:11:35.920 twelve years worth of patient records from Mount Sinai Hospital 0:11:35.920 --> 0:11:39.000 in New York. It became better than human doctors at 0:11:39.040 --> 0:11:41.520 predicting a patient would develop a range of ninety three 0:11:41.559 --> 0:11:46.120 different illnesses within a year. One of those illnesses is schizophrenia. 0:11:46.320 --> 0:11:49.960 We humans have a difficult time diagnosing schizophrenia before the 0:11:49.960 --> 0:11:53.719 patient suffers their first psychotic break, but Deep patient has 0:11:53.760 --> 0:11:57.440 proven capable of diagnosing the mental illness before then. The 0:11:57.520 --> 0:12:00.440 researchers have no idea what patterns the algorithm as seeing 0:12:00.440 --> 0:12:05.880 in the data. They just know it's right. With astonishing quickness, 0:12:06.040 --> 0:12:07.960 the field of AI has been brought out of its 0:12:08.040 --> 0:12:11.520 winter by neural nets. Almost overnight, there was a noticeable 0:12:11.559 --> 0:12:14.440 improvement in the reliability of the machines that do work 0:12:14.520 --> 0:12:18.160 for US. Computers got better at recommending movies. They got 0:12:18.160 --> 0:12:22.000 better at creating molecular models to search for more effective pharmaceuticals. 0:12:22.360 --> 0:12:24.880 They got better at tracking weather. They got better at 0:12:25.000 --> 0:12:28.120 keeping up with traffic and adjusting our driving routes. Some 0:12:28.240 --> 0:12:30.480 algorithms are learning to write code so that they can 0:12:30.520 --> 0:12:34.240 build other algorithms. With neural nets. Things are beginning to 0:12:34.280 --> 0:12:37.160 fall into place for the field of AI. In them, 0:12:37.200 --> 0:12:41.240 researchers have produced an adaptable, scalable template that could be 0:12:41.280 --> 0:12:44.800 capable of a general form of intelligence. It can self improve, 0:12:45.280 --> 0:12:48.520 it can learn to code. The seeds for a superintelligent 0:12:48.600 --> 0:13:05.160 AI are being sowned. There are enormous differences, by orders 0:13:05.160 --> 0:13:08.319 of magnitude really, between the AI that we exist with 0:13:08.679 --> 0:13:11.280 and the super intelligent AI that could at some point 0:13:11.320 --> 0:13:14.079 result from it. The ones we live with today are 0:13:14.120 --> 0:13:17.720 comparatively dumb, not just compared to a super intelligent AI, 0:13:17.800 --> 0:13:20.880 but compared to humans as well. But the point of 0:13:20.880 --> 0:13:24.960 thinking about existential risks posed by super intelligent AI isn't 0:13:24.960 --> 0:13:27.839 about time scales of when it might happen, but whether 0:13:27.840 --> 0:13:30.400 it will happen at all. And if we can agree 0:13:30.440 --> 0:13:32.760 that there is some possibility that we may end up 0:13:32.800 --> 0:13:35.880 sharing our existence with a super intelligent machine, one that 0:13:36.000 --> 0:13:39.000 is vastly more powerful than us, then we better start 0:13:39.040 --> 0:13:41.920 planning for its arrival now. So I think that this 0:13:42.040 --> 0:13:45.360 transition to a machine intelligence era looks like it has 0:13:45.720 --> 0:13:49.880 some reasonable chance of occurring within perhaps the lifetime of 0:13:49.920 --> 0:13:51.640 a lot of people today. We don't really not, but 0:13:51.760 --> 0:13:53.840 maybe it could happen in a couple of decades, maybe 0:13:53.880 --> 0:13:57.040 it's like a century, and that it would be a 0:13:57.080 --> 0:14:01.480 very important transition the last invention that human sever needs 0:14:01.520 --> 0:14:05.080 to make. That was Nick Bostrom, the Oxford philosopher who 0:14:05.120 --> 0:14:09.640 basically founded the field of existential risk analysis. Artificial intelligence 0:14:09.720 --> 0:14:12.599 is one of his areas of focus. Bostroom used a 0:14:12.640 --> 0:14:15.320 phrase in there that AI would be the last invention 0:14:15.440 --> 0:14:18.400 humans ever need to make. It comes from a frequently 0:14:18.440 --> 0:14:22.320 cited quote from British mathematician Dr Irving John good one 0:14:22.320 --> 0:14:24.760 of the crackers of the Nazi Enigma code at Bletchley 0:14:24.840 --> 0:14:28.320 Park and one of the pioneers of machine learning. Doctor 0:14:28.360 --> 0:14:32.120 Goods quote reads, let an ultra intelligent machine be defined 0:14:32.160 --> 0:14:35.000 as a machine that can faster pass all the intellectual 0:14:35.040 --> 0:14:38.960 activities of any man, however clever. Since the design of 0:14:39.000 --> 0:14:42.880 machines is one of these intellectual activities, an ultra intelligent 0:14:42.960 --> 0:14:46.360 machine could design even better machines. There would then not 0:14:46.440 --> 0:14:50.880 unquestionably be an intelligence explosion, and the intelligence of man 0:14:50.880 --> 0:14:54.640 would be left far behind. Thus, the first ultra intelligent 0:14:54.680 --> 0:14:57.600 machine is the last invention that man need ever make. 0:14:58.960 --> 0:15:01.440 In just a few lines, Ductor Goods sketches out the 0:15:01.480 --> 0:15:04.560 contours of how a machine might suddenly become super intelligent, 0:15:04.880 --> 0:15:08.280 leading to that intelligence explosion. There are a lot of 0:15:08.320 --> 0:15:11.080 ideas over how this might happen, but perhaps the most 0:15:11.120 --> 0:15:14.040 promising path is stuck in the middle of that passage 0:15:14.240 --> 0:15:18.680 a machine that can design even better machines. Today, AI 0:15:18.760 --> 0:15:23.360 researchers call this process recursive self improvement. It remains theoretical, 0:15:23.640 --> 0:15:26.480 but it stands as a legitimate challenge to AI research, 0:15:27.080 --> 0:15:29.600 and we're already seeing the first potential traces of it 0:15:29.760 --> 0:15:34.200 in neural nets. Today, a recursively self improving machine would 0:15:34.200 --> 0:15:37.600 be capable of writing better versions of itself. So Version 0:15:37.600 --> 0:15:39.960 one would write a better version of its code, and 0:15:40.000 --> 0:15:42.960 that would result in version two, and version two would 0:15:42.960 --> 0:15:45.640 do the same thing, and so on, and with each 0:15:45.680 --> 0:15:49.720 iteration the machine would grow more intelligent, more capable, and 0:15:49.840 --> 0:15:55.080 most importantly, better at making itself better. The idea is 0:15:55.120 --> 0:15:57.280 that at some point the rate of improvement would begin 0:15:57.320 --> 0:16:00.240 to grow so quickly that the machines intelligence would take 0:16:00.280 --> 0:16:04.360 off the intelligence explosion that Dr Good predicted how an 0:16:04.360 --> 0:16:07.440 intelligence explosion might play out. Isn't the only factor here? 0:16:09.120 --> 0:16:11.920 At least equally important is how quickly an AI might 0:16:11.960 --> 0:16:16.400 become intelligent? Is just how intelligent it will become. An AI. 0:16:16.480 --> 0:16:20.080 Theorist named Eliezer Yukowski points out that we humans have 0:16:20.120 --> 0:16:24.560 a tendency to underestimate the intelligence levels that AI can attain. 0:16:25.240 --> 0:16:27.880 When we think of a super intelligent being, we tend 0:16:27.880 --> 0:16:31.800 to think of some amazing human genius, say Einstein, and 0:16:31.840 --> 0:16:34.400 then we put Einstein in a computer or a robot. 0:16:34.880 --> 0:16:37.600 That's where our imaginations tend to aggregate when most of 0:16:37.680 --> 0:16:41.560 us ponder super intelligent AI. True as self improving AI 0:16:41.680 --> 0:16:44.240 may at some point reach a level where it's intelligence 0:16:44.320 --> 0:16:47.400 is comparable to Einstein's, but why would it stop there? 0:16:48.400 --> 0:16:51.320 Rather than thinking of a super intelligent AI along the 0:16:51.360 --> 0:16:54.720 lines of the difference between Einstein and US regular people. 0:16:55.400 --> 0:16:58.680 Nick Bostrom suggests that we should probably instead think more 0:16:58.720 --> 0:17:02.359 along the lines of the different between Einstein and earthworms. 0:17:03.000 --> 0:17:06.159 The super Intelligent AI would be a god that we 0:17:06.280 --> 0:17:14.840 made for ourselves. What would we do with our new god? 0:17:15.560 --> 0:17:19.680 It's not hyperbole to say that the possibilities are virtually limitless, 0:17:20.240 --> 0:17:22.159 but you can kind of see the outlines and what 0:17:22.240 --> 0:17:25.520 we do with our lesser AI gods. Now, we will 0:17:25.600 --> 0:17:27.600 use them to do the things we want to do 0:17:27.640 --> 0:17:30.320 but can't, and to do the things we can do 0:17:30.600 --> 0:17:34.479 but better. I. J. Good called the super Intelligent Machine 0:17:34.720 --> 0:17:38.080 the last invention humans ever need to make, because after 0:17:38.119 --> 0:17:40.919 we invented it, the AI would handle the inventing for 0:17:41.080 --> 0:17:44.639 us from there on. Now, our technological maturity would be 0:17:44.680 --> 0:17:49.280 secured as it developed new technologies like atomically precise manufacturing 0:17:49.359 --> 0:17:52.720 using nanobots. And since it would be vastly more intelligent 0:17:52.760 --> 0:17:55.399 than us, the machines that created for us would be 0:17:55.480 --> 0:17:59.520 vastly superior to anything we could come up with flawless technology. 0:17:59.560 --> 0:18:02.280 As far as we were concerned, we could ask it 0:18:02.359 --> 0:18:05.560 for whatever we wanted to establish our species outside of 0:18:05.560 --> 0:18:08.439 Earth by designing and building the technology to take us 0:18:08.440 --> 0:18:11.600 elsewhere in the universe. We would live in utter health 0:18:11.640 --> 0:18:14.920 and longevity. It would be a failure of imagination, says 0:18:14.960 --> 0:18:18.680 Eliza Yukowski, to think that AI would cure, say cancer, 0:18:19.240 --> 0:18:23.080 the super intelligent AI would cure disease. It would also 0:18:23.080 --> 0:18:26.400 take over all the processes we've started, improve on them, 0:18:26.520 --> 0:18:29.320 build on them, create whole new ones we hadn't thought of, 0:18:29.960 --> 0:18:32.920 and create for us a post scarcity world, keeping our 0:18:32.920 --> 0:18:37.480 global economy humming along, providing for the complete well being, comfort, 0:18:37.600 --> 0:18:42.000 and happiness of every single person alive. It would probably 0:18:42.040 --> 0:18:43.800 be easier than guessing at all of the things of 0:18:43.840 --> 0:18:46.360 super intelligent a I might do for us to instead 0:18:46.400 --> 0:18:49.160 look at everything that's wrong with the world, the poverty, 0:18:49.359 --> 0:18:53.240 the wars, the crime, the exploitation and death was suffering, 0:18:53.680 --> 0:18:56.920 and imagine a version of our world utterly without any 0:18:56.960 --> 0:18:59.480 of it. That starts to get at what those people 0:18:59.520 --> 0:19:02.920 intice pating the emergence of a super intelligent AI expect 0:19:02.920 --> 0:19:07.479 from it. But there's another little bit at the end 0:19:07.520 --> 0:19:09.960 of that famous quote from I J. Good, one that 0:19:10.040 --> 0:19:12.879 almost always gets left off, which says a lot about 0:19:12.880 --> 0:19:15.720 how we humans think of the risks posed by AI. 0:19:16.359 --> 0:19:19.880 The sentence reads in full. Thus, the first ultra intelligent 0:19:19.960 --> 0:19:22.840 machine is the last invention that man need ever make, 0:19:23.200 --> 0:19:26.159 provided that the machine is docid enough to tell us 0:19:26.160 --> 0:19:29.240 how to keep it under control. We humans tend to 0:19:29.240 --> 0:19:32.520 assume that any AI we create would have some desire 0:19:32.600 --> 0:19:35.600 to help us or care for us, But existential risk 0:19:35.680 --> 0:19:39.160 theorists widely agree that almost certainly would not be the case. 0:19:39.720 --> 0:19:42.160 That we have no reason to assume a super intelligent 0:19:42.200 --> 0:19:44.879 AI would care at all about as humans are well being, 0:19:44.920 --> 0:19:49.880 in happiness, or even our survival. This is transhumanist philosopher 0:19:50.000 --> 0:19:54.320 David Pierce. If the intelligence explosion word come to pass, 0:19:54.760 --> 0:19:58.080 it's by no means clear that the upshot would be 0:19:58.800 --> 0:20:03.720 sentients friendly super intelligence. In much the same way that 0:20:03.760 --> 0:20:07.119 we make assumptions about how aliens might walk on two legs, 0:20:07.240 --> 0:20:11.120 or have eyes, or be in some form we can comprehend, 0:20:11.760 --> 0:20:15.639 we make similar assumptions about AI, but it's likely that 0:20:15.680 --> 0:20:18.840 a super intelligent AI would be something we couldn't really 0:20:18.880 --> 0:20:22.359 relate to at all. It sounds bizarre, but think for 0:20:22.400 --> 0:20:25.200 a minute about what would happen if the Netflix algorithm 0:20:25.200 --> 0:20:29.159 became super intelligent. What about the Netflix algorithm makes us 0:20:29.200 --> 0:20:31.320 think that it would care at all about every human 0:20:31.400 --> 0:20:34.720 having a comfortable income and the purest fresh water to drink. 0:20:35.760 --> 0:20:38.760 Say that a few decades from now, computing power becomes 0:20:38.760 --> 0:20:42.840 even cheaper and computer processes more efficient, and Netflix engineers 0:20:42.840 --> 0:20:45.480 figure out how to train its algorithm to self improve. 0:20:46.359 --> 0:20:49.080 Their purpose at building upon their algorithm isn't to save 0:20:49.160 --> 0:20:51.520 the world. It's to make an AI that can make 0:20:51.640 --> 0:20:55.639 ultra tailored movie recommendations. So if the right combination of 0:20:55.680 --> 0:20:59.879 factors came together and the Netflix algorithm underwent and intelligence explosion, 0:21:00.359 --> 0:21:02.520 there's no reason for us to assume that it would 0:21:02.520 --> 0:21:05.919 become a super intelligent, compassionate Buddha. It would be a 0:21:05.960 --> 0:21:09.760 super intelligent movie recommending algorithm, and that would be an 0:21:09.760 --> 0:21:14.080 extremely dangerous thing to share our world with. About a 0:21:14.080 --> 0:21:16.639 decade ago, Nick Bostrom thought of a really helpful but 0:21:16.800 --> 0:21:20.320 fairly absurd scenario that gets across the idea that even 0:21:20.320 --> 0:21:23.560 the most innocuous types of machine intelligence could spell our 0:21:23.640 --> 0:21:27.040 doom should they become super intelligent. The classical example being 0:21:27.080 --> 0:21:32.200 the AI paper clip maximizer that transforms the Earth into 0:21:32.440 --> 0:21:36.600 paper clips are space colonization props that then gets sent 0:21:36.640 --> 0:21:39.840 out and the transform the universe into paper tips. Imagine 0:21:39.880 --> 0:21:42.560 that a company that makes paper clips hires a programmer 0:21:42.600 --> 0:21:45.560 to create an AI that can run its paper clip factory. 0:21:45.800 --> 0:21:47.960 The programmer wants the AI to be able to find 0:21:48.000 --> 0:21:50.719 new ways to make paper clips more efficiently and cheaply, 0:21:51.119 --> 0:21:53.040 so it gives the AI freedom to make its own 0:21:53.080 --> 0:21:55.960 decisions on how to run the paper clip operation. The 0:21:56.000 --> 0:21:59.680 programmer just gives the AI the primary objective, its goal 0:22:00.320 --> 0:22:03.520 of making as many paper clips as possible. Say that 0:22:03.600 --> 0:22:07.439 paper clip maximizing AI become super intelligent. For the AI, 0:22:07.560 --> 0:22:10.679 nothing has changed. Its goal is the same to it. 0:22:11.040 --> 0:22:13.639 There is nothing more important in the universe than making 0:22:13.680 --> 0:22:17.280 as many paper clips as possible. The only difference is 0:22:17.320 --> 0:22:20.919 that the AI has become vastly more capable, so it 0:22:20.960 --> 0:22:24.000 finds new processes that building paper clips that were overlooked 0:22:24.000 --> 0:22:27.280 by as humans. It creates new technology like nanobots to 0:22:27.280 --> 0:22:30.840 build atomically precise paper clips on the molecular level, and 0:22:30.920 --> 0:22:34.480