1 00:00:00,000 --> 00:00:10,759 Okay, we have about 15 minutes, so I'm going to go through this more quickly, but I just 2 00:00:10,759 --> 00:00:15,839 want to talk a little bit more about the history of AI, because that's the hot topic these 3 00:00:15,839 --> 00:00:19,300 days, and I think we should all learn a little bit more about what it is instead of using 4 00:00:19,300 --> 00:00:19,980 it as a buzzword. 5 00:00:21,800 --> 00:00:22,199 Cool. 6 00:00:23,440 --> 00:00:24,780 All right, so what is AI, right? 7 00:00:24,780 --> 00:00:30,520 All right, AI is just a machine, like the intelligence of machines or programs. 8 00:00:31,560 --> 00:00:35,700 So that sounds simple, but let's unpack that, right? 9 00:00:36,140 --> 00:00:37,719 So there's multiple types of AI. 10 00:00:37,939 --> 00:00:40,799 There's narrow AI, and there's general AI. 11 00:00:40,960 --> 00:00:50,880 So narrow AI is you specialize in a specific topic, and general AI is you can learn to 12 00:00:50,880 --> 00:00:53,180 do and solve any problems, just like a human. 13 00:00:54,780 --> 00:01:04,599 So here are some examples of AI, self-driving, go-playing or chess-playing board games, and 14 00:01:04,599 --> 00:01:08,099 finally, robots that can walk and run, just like humans. 15 00:01:10,299 --> 00:01:11,200 Okay, great. 16 00:01:11,859 --> 00:01:15,700 So I'm going to skip some of this, because this is some technicalities, but basically, 17 00:01:16,359 --> 00:01:22,680 the precursors of AI is how do we formalize human thoughts? 18 00:01:22,680 --> 00:01:23,939 How do we formalize? 19 00:01:24,780 --> 00:01:29,900 Human thought in a mathematical way, and so there are many theorems that came up in 20 00:01:29,900 --> 00:01:41,299 the 1930s that formalized the math of human thought, and in 1945, we got the first computer 21 00:01:41,299 --> 00:01:46,219 that was programmable and electronic and digital, so it was this big. 22 00:01:46,859 --> 00:01:51,260 So think like this whole classroom, that's how big it was, and you had many people who 23 00:01:51,260 --> 00:01:54,060 had to operate this machine in 1945. 24 00:01:54,780 --> 00:01:55,780 Great. 25 00:01:55,780 --> 00:02:01,780 So now let's talk about some of the fundamentals of AI. 26 00:02:01,780 --> 00:02:08,780 So people started talking about how we can create an artificial brain, and so there's 27 00:02:08,780 --> 00:02:09,780 a Turing test. 28 00:02:09,780 --> 00:02:15,439 Maybe you've heard of this, but this is just something, a test that can tell if a computer 29 00:02:15,439 --> 00:02:23,340 can say intelligent things like a human. 30 00:02:23,340 --> 00:02:24,659 And so we also got some really cool programs. 31 00:02:24,659 --> 00:02:33,659 Like the first neural network machine built by Marvin Minsky, a MIT professor for a long 32 00:02:33,659 --> 00:02:40,659 time, and some other programs of AI, for example, logic theorists that proved many math theorems, 33 00:02:40,659 --> 00:02:43,659 which is kind of insane to think about back in the 50s. 34 00:02:46,659 --> 00:02:47,659 Great. 35 00:02:47,659 --> 00:02:48,659 So this is where it all started. 36 00:02:48,659 --> 00:02:49,659 The birth of AI. 37 00:02:49,659 --> 00:02:51,659 1956, the Dartmouth workshop. 38 00:02:51,659 --> 00:02:53,659 Dartmouth is a university in the United States. 39 00:02:53,659 --> 00:03:01,659 This is where AI got its name, got its first success, and these ten folks were some of 40 00:03:01,659 --> 00:03:06,659 the people who attended, and they became famous for their contributions to AI. 41 00:03:06,659 --> 00:03:14,659 We saw earlier Marvin Minsky, and Cloud Shannon is also very famous for his contributions. 42 00:03:14,659 --> 00:03:18,659 And everyone is, but those are the two that I personally know the most because they're 43 00:03:18,659 --> 00:03:19,659 from MIT. 44 00:03:19,659 --> 00:03:20,659 Great. 45 00:03:20,659 --> 00:03:21,659 Thank you. 46 00:03:21,659 --> 00:03:22,659 Thank you. 47 00:03:22,659 --> 00:03:23,659 Great. 48 00:03:23,659 --> 00:03:28,659 So in this time range, we had many advances. 49 00:03:28,659 --> 00:03:35,659 So computers were able to solve algebra problems, geometry problems, learn English, just some 50 00:03:35,659 --> 00:03:37,659 crazy things that were going on. 51 00:03:37,659 --> 00:03:43,659 And so a lot of these universities like MIT, CMU Stanford, they received millions of dollars 52 00:03:43,659 --> 00:03:46,659 from the government to research. 53 00:03:46,659 --> 00:03:51,659 And so we were thinking, or not we, they were thinking that in 20 years, we would have a 54 00:03:51,659 --> 00:03:53,659 computer as smart as a human. 55 00:03:53,659 --> 00:03:54,659 And so, yeah. 56 00:03:54,659 --> 00:04:01,659 Like, DARPA was just feeding money to these universities. 57 00:04:01,659 --> 00:04:09,659 And so maybe not to get too complicated here, but back in the 50s, the AI was more about 58 00:04:09,659 --> 00:04:10,659 searching. 59 00:04:10,659 --> 00:04:19,659 So you have some big space of possible choices or possible solutions, and you have to search 60 00:04:19,660 --> 00:04:22,240 and you have to search it to find the best one. 61 00:04:26,860 --> 00:04:29,860 And we also got the first kinds of neural networks, 62 00:04:30,060 --> 00:04:30,780 if you've heard of that. 63 00:04:30,860 --> 00:04:33,700 Those are the most popular type of AI today. 64 00:04:34,840 --> 00:04:38,640 And so we had very small neural networks, 65 00:04:38,740 --> 00:04:40,379 very simple ones, back in the 50s. 66 00:04:40,379 --> 00:04:41,280 This is where it all started. 67 00:04:43,379 --> 00:04:46,360 And we also got the first chatbot created by, 68 00:04:46,660 --> 00:04:47,980 yes, another guy from MIT. 69 00:04:47,980 --> 00:04:53,980 It was created to be some sort of psychotherapist, 70 00:04:54,620 --> 00:04:55,180 in some sense. 71 00:04:55,680 --> 00:04:58,460 So let's look a little bit more at the text here. 72 00:05:00,920 --> 00:05:04,080 I'll leave this later, but just read it briefly. 73 00:05:05,040 --> 00:05:08,600 And it's quite impressive to me, 74 00:05:08,879 --> 00:05:12,220 this chatbot that was made 50 years ago. 75 00:05:12,480 --> 00:05:16,560 It's giving some very human-like responses. 76 00:05:16,560 --> 00:05:17,220 Okay. 77 00:05:17,980 --> 00:05:21,280 So, all right. 78 00:05:21,420 --> 00:05:25,620 Of course, we did not make the general AI in 20 years. 79 00:05:25,960 --> 00:05:27,600 Instead, we got the first AI winter. 80 00:05:28,340 --> 00:05:32,060 So there were some fundamental barriers 81 00:05:32,060 --> 00:05:33,980 to actually implement AI. 82 00:05:35,580 --> 00:05:37,400 So, for example, there's something called 83 00:05:37,400 --> 00:05:38,640 combinatorial explosion. 84 00:05:38,860 --> 00:05:41,740 That's just a bunch of fancy words that really means 85 00:05:41,740 --> 00:05:45,560 when your problem becomes complicated, 86 00:05:45,560 --> 00:05:47,700 the solution space becomes, 87 00:05:47,980 --> 00:05:49,360 really, really big. 88 00:05:54,700 --> 00:05:57,200 And so another problem was the computers in the 50s, 89 00:05:57,200 --> 00:05:59,980 70s, excuse me, were terrible. 90 00:05:59,980 --> 00:06:02,180 Like, look at the Cray 2 supercomputer, 91 00:06:02,180 --> 00:06:05,879 super computer, back in 1985. 92 00:06:05,879 --> 00:06:08,780 And look at an Apple iPhone 6 smartphone. 93 00:06:08,780 --> 00:06:10,200 The comparison is astounding. 94 00:06:10,200 --> 00:06:13,220 Our Apple iPhone 6 is hundreds of times better 95 00:06:13,220 --> 00:06:16,460 than this super computer back in 1985. 96 00:06:17,980 --> 00:06:22,220 And finally, most programs in the 70s 97 00:06:22,220 --> 00:06:24,580 didn't have common sense. 98 00:06:24,580 --> 00:06:29,580 So they didn't understand the things that a baby human would, 99 00:06:30,140 --> 00:06:33,500 like, how do you throw a ball? 100 00:06:33,500 --> 00:06:34,900 How do you look at digits? 101 00:06:34,900 --> 00:06:36,700 They just did not have common sense. 102 00:06:39,140 --> 00:06:41,980 So, of course, everyone lost their funding. 103 00:06:41,980 --> 00:06:44,620 The government said, you guys are not doing well. 104 00:06:44,620 --> 00:06:47,960 And so, unfortunately, also, you know, 105 00:06:47,980 --> 00:06:50,600 the networks, the research there and funding there 106 00:06:50,600 --> 00:06:54,840 also got cut off because Marvin Minsky, 107 00:06:54,840 --> 00:06:57,600 same guy from earlier, and another person named Poppert, 108 00:06:57,600 --> 00:07:00,300 both MIT folks, they published a book that said, 109 00:07:00,300 --> 00:07:02,660 hey, neural networks don't work. 110 00:07:02,660 --> 00:07:06,960 And so they basically just, yeah, it was bad. 111 00:07:06,960 --> 00:07:11,960 And so now after the first decline, the first winter, 112 00:07:13,120 --> 00:07:15,140 we had actually an AI boom. 113 00:07:15,140 --> 00:07:17,379 So thankfully, things recovered in the 80s. 114 00:07:17,980 --> 00:07:22,980 And so there's these things called expert systems. 115 00:07:23,080 --> 00:07:26,300 So people started working on expert systems 116 00:07:26,300 --> 00:07:29,460 where instead of trying to solve all of AI, 117 00:07:29,460 --> 00:07:31,000 they started working on narrow AI, 118 00:07:31,000 --> 00:07:34,160 just solving a very small portion of AI problems. 119 00:07:34,160 --> 00:07:36,620 And so an expert system would become an expert 120 00:07:36,620 --> 00:07:38,340 on a small topic. 121 00:07:38,340 --> 00:07:40,520 And then other people, not experts, 122 00:07:40,520 --> 00:07:44,000 could ask the expert system, how do I do this? 123 00:07:44,000 --> 00:07:45,540 Can you help me do this, right? 124 00:07:45,540 --> 00:07:47,500 Kind of like ChatGPT, 125 00:07:47,500 --> 00:07:50,420 but just for a very small dominion. 126 00:07:50,420 --> 00:07:53,639 So like just about pasta or just about making pizza 127 00:07:53,639 --> 00:07:55,779 instead of the whole world. 128 00:08:00,639 --> 00:08:05,639 So then knowledge became the primary focus of research. 129 00:08:06,860 --> 00:08:09,100 How can we gather information 130 00:08:09,100 --> 00:08:13,379 and how can a robot synthesize information and understand it? 131 00:08:15,060 --> 00:08:16,879 And of course, funding came back. 132 00:08:17,500 --> 00:08:20,160 So Japan started first, and then the United Kingdom 133 00:08:20,160 --> 00:08:22,699 and the US followed with millions of dollars of funding. 134 00:08:25,699 --> 00:08:28,740 Also, neural networks, they came back 135 00:08:28,740 --> 00:08:31,439 because a very smart physicist demonstrated 136 00:08:31,439 --> 00:08:34,480 that neural networks can actually work. 137 00:08:34,480 --> 00:08:35,759 They can converge. 138 00:08:35,759 --> 00:08:39,559 And they can, another very important computer scientist, 139 00:08:39,559 --> 00:08:43,279 Hinton and Rumelhart, they popularized something 140 00:08:43,279 --> 00:08:47,480 called back propagation that is today still one of the most popular ways to do this. 141 00:08:47,500 --> 00:08:50,340 The most popular way for the world to train AI. 142 00:08:51,460 --> 00:08:53,620 And of course that didn't last too long. 143 00:08:53,620 --> 00:08:56,840 Another second winter came, unfortunately. 144 00:08:56,840 --> 00:09:01,059 What happened was there was a collapse in the market 145 00:09:01,059 --> 00:09:03,259 for specialized AI hardware, 146 00:09:03,259 --> 00:09:07,820 because Apple and IBM, they started selling computers 147 00:09:07,820 --> 00:09:11,639 that beat the AI machines. 148 00:09:11,639 --> 00:09:13,779 So if you remember in 1987, 149 00:09:13,779 --> 00:09:16,500 that was when the Apple II personal computer was released. 150 00:09:16,500 --> 00:09:17,460 And that was a huge big deal. 151 00:09:17,500 --> 00:09:26,620 revolution and the personal computing space and also AI expert systems they work sometimes but 152 00:09:26,620 --> 00:09:32,639 usually they stop working after you push them to their limits and so this is the Apple computer 153 00:09:32,639 --> 00:09:41,100 Apple too and so the United States government said we don't like AI anymore we're going to 154 00:09:41,100 --> 00:09:49,480 take your money and put it elsewhere and so also a lot of AI companies they lost they lost their 155 00:09:49,480 --> 00:09:56,259 funding they closed or they were sold to other companies now something bright that came out of 156 00:09:56,259 --> 00:10:02,120 this was people realized maybe there's actually we're doing this the wrong way maybe instead of 157 00:10:02,120 --> 00:10:08,220 thinking about how smart a human is we should think about how good a human as as understanding 158 00:10:08,220 --> 00:10:10,980 the world so instead of building brains we should 159 00:10:10,980 --> 00:10:11,080 you know we should think about how smart a human is we should think about how good a human is 160 00:10:11,080 --> 00:10:20,600 build bodies so some smart researchers thought that okay instead of building the brain of the 161 00:10:20,600 --> 00:10:26,580 machine let's just build a machine that can learn how to perceive the world instead of just looking 162 00:10:26,580 --> 00:10:33,500 at words but instead of thinking about how do I hold a tennis racket how do I look at the room 163 00:10:33,500 --> 00:10:40,600 and understand that I'm in a room so basically advanced robotics cool 164 00:10:41,080 --> 00:10:47,860 all right now narrow AI started having some success in the 90s and the early 2000s for 165 00:10:47,860 --> 00:10:55,660 example Deep Blue it beat the chess champion Garry Kasparov in 1997 a very famous example 166 00:10:55,660 --> 00:11:04,900 of machine beating human also some successes in autonomous driving some robots from Stanford and 167 00:11:04,900 --> 00:11:11,060 CMU navigated a car with no driver in the early 2000s for hundreds of millions of dollars and 168 00:11:11,080 --> 00:11:21,160 kilometers and also if you know the quiz show trivia show jeopardy IBM Watson was able to beat 169 00:11:21,160 --> 00:11:33,040 the two best human champions of all time by $50,000 in 2011 of course these successes were not because 170 00:11:33,040 --> 00:11:38,860 humans made better algorithms it was actually because computers just got way better so there's 171 00:11:38,860 --> 00:11:40,960 something called Moore's law where there was a 172 00:11:40,960 --> 00:11:52,360 tremendous uh pattern of increase in computer capacity in these uh in these years this is an 173 00:11:52,360 --> 00:11:57,340 example of Moore's law you can take a look at it later but it just shows you that progress 174 00:11:57,340 --> 00:12:10,360 was tremendous it was exponentially good and so the AI had at this point AI had resolved many 175 00:12:10,960 --> 00:12:22,180 problems but it was still not considered uh a fun area because AI had two winters and so people 176 00:12:22,180 --> 00:12:27,340 didn't want to say that they were working at AI but its applications were all across the industries 177 00:12:32,440 --> 00:12:39,340 so basically AI very useful but nobody wants to call the work AI because it was not the cool topic 178 00:12:39,340 --> 00:12:40,360 it was the weird topic 179 00:12:40,960 --> 00:12:48,519 very different from now right okay now very recently you might remember the big data 180 00:12:49,120 --> 00:12:55,780 bubble or the hype basically big data is you have a lot of data and you need very big computers to 181 00:12:55,780 --> 00:13:01,240 work with a lot of data and we can do that because we have cheap computers lots of computers and lots 182 00:13:01,240 --> 00:13:10,600 of data so also we had some Advances in deep learning 183 00:13:10,960 --> 00:13:20,740 where we got some really uh crazy deep uh deep learning algorithms and also had the compute 184 00:13:20,740 --> 00:13:27,400 capacity to actually use them to solve problems like identify cats and dogs or identify what kind 185 00:13:27,400 --> 00:13:37,960 of object where the object is in an image great so let's talk about the present 2017 186 00:13:39,040 --> 00:13:39,879 our transformer 187 00:13:40,960 --> 00:13:52,300 um that was so we had this thing called GPT-3 that's the precursor of the um that's a big 188 00:13:52,300 --> 00:13:58,900 it's a large language model that was based on the transformer architecture and GPT-3 you see GPT 189 00:13:59,800 --> 00:14:07,420 well yeah chat GPT was based on GPT 3.5 um and that was launched in 2022 190 00:14:09,100 --> 00:14:10,900 and also uh now 191 00:14:10,960 --> 00:14:16,900 Microsoft is working on GPT-4 which is supposed to be even better and there are some chat GPT 192 00:14:16,900 --> 00:14:25,360 versions based on GPT-4 but you have to pay money to see them cool so um this is a small 193 00:14:25,360 --> 00:14:30,820 demonstration of some current research in AI some really smart Stanford researchers they made a 194 00:14:30,820 --> 00:14:39,879 robot that can cook food for you um to be fair this one is controlled by the human but their robot also 195 00:14:39,879 --> 00:14:45,280 cooks stuff on its own too autonomously um but yeah it's just cooking some uh Cantonese food 196 00:14:46,180 --> 00:14:51,520 um I don't know if it's any good or not but I'd be down to try it but we can skip it um 197 00:14:53,860 --> 00:14:59,560 all right cool all right okay let's talk about AI ethics right this is important because when you do 198 00:14:59,560 --> 00:15:04,900 you need to make sure that you are not harming people all right how do we 199 00:15:04,900 --> 00:15:11,760 eliminate bias in science examples Amazon tried to once use a AI system to 200 00:15:11,760 --> 00:15:18,820 recruit people and try to interview people and it's not using it why because 201 00:15:18,820 --> 00:15:23,480 the algorithm only chose men because mostly men work at Amazon so it assumed 202 00:15:23,480 --> 00:15:26,740 that all the best engineers are men which is obviously not true right 203 00:15:26,740 --> 00:15:33,460 another problem facial detection it had more data from white men and less data 204 00:15:33,460 --> 00:15:38,860 from darker woman so therefore it was worse on detecting darker woman so it's 205 00:15:38,860 --> 00:15:44,379 you know this is not okay right we need to have a unbiased system for 206 00:15:44,379 --> 00:15:55,480 classifying or detecting things another problem is job loss AI can do a lot of 207 00:15:55,480 --> 00:15:56,440 things that humans can do 208 00:15:56,440 --> 00:15:56,720 you 209 00:15:56,720 --> 00:16:02,960 better faster and people might lose jobs and so we need to make sure that people 210 00:16:02,960 --> 00:16:09,040 are able to still be able to make enough money to live if their job gets taken 211 00:16:09,040 --> 00:16:15,080 away or have an alternative job that they can do instead and finally this is 212 00:16:15,080 --> 00:16:17,740 a little bit far out there I don't think it's going to happen but some people do 213 00:16:17,740 --> 00:16:22,100 existential risk what if we make an AI that's so smart 214 00:16:22,100 --> 00:16:26,420 you can't stop it and it like kills all humans right that I don't ask what 215 00:16:26,420 --> 00:16:34,640 happened but hey I didn't say that all right so finally I like to show a 216 00:16:34,640 --> 00:16:38,240 timeline you can click on this link when I share the position afterwards and look 217 00:16:38,240 --> 00:16:44,180 at some of the advances that we've talked about and yeah so that was a 218 00:16:44,180 --> 00:16:47,380 little bit of an intro thank you again for listening and once again my email 219 00:16:47,380 --> 00:16:51,760 and any questions I'd be happy to take thank you 220 00:16:51,760 --> 00:16:53,820 you 221 00:16:53,820 --> 00:16:55,880 you