WEBVTT - The Godfather of AI is Worried About AI 0:00:04.440 --> 0:00:12.920 Welcome to tech Stuff, a production from iHeartRadio. Heydarren, Welcome 0:00:12.960 --> 0:00:16.200 to tech Stuff. I'm your host, Jonathan Strickland. I'm an 0:00:16.239 --> 0:00:20.800 executive producer with iHeartRadio and how the tech are you? So? 0:00:21.320 --> 0:00:25.080 Last week I said I would do an episode about 0:00:25.079 --> 0:00:30.440 doctor Jeffrey Hinton, the so called godfather of AI. My 0:00:30.600 --> 0:00:33.360 dog is very interested in this, as he winds in 0:00:33.400 --> 0:00:37.960 the background, and as I published this. We're into the 0:00:38.000 --> 0:00:40.720 fifth month of twenty twenty three, and I still feel 0:00:40.720 --> 0:00:43.960 pretty good about calling this the Year of AI. While 0:00:44.080 --> 0:00:49.159 artificial intelligence has obviously been a discipline for decades, with 0:00:49.880 --> 0:00:54.680 lots of impressive displays and exhibitions and developments throughout the years, 0:00:55.280 --> 0:00:59.240 the buzz around and attention to aifields has really hit 0:00:59.240 --> 0:01:02.840 a high point this year, largely driven by stuff like 0:01:03.320 --> 0:01:07.960 large language models LMS, as well as the chatbots built 0:01:08.000 --> 0:01:12.600 on top of them that seem to be pretty knowledgeable, 0:01:12.680 --> 0:01:17.080 almost human in their capabilities. Plus you throw in some 0:01:17.319 --> 0:01:20.520 image and video and audio capabilities that allow us to 0:01:20.600 --> 0:01:23.319 use a machine to create all sorts of stuff, and 0:01:23.360 --> 0:01:26.120 you got yourself something that the average person can at 0:01:26.240 --> 0:01:30.600 least recognize as AI. A lot of AI applications historically 0:01:31.240 --> 0:01:36.040 have been so far behind the scenes that you might 0:01:36.080 --> 0:01:39.280 not even recognize it as artificial intelligence, or you might 0:01:39.319 --> 0:01:41.800 not think of it in that context. But now we're 0:01:41.800 --> 0:01:45.880 getting to a point where there's at least the appearance 0:01:46.040 --> 0:01:51.280 of machines behaving similarly to people within certain contexts, and 0:01:51.320 --> 0:01:54.840 it becomes way easier for the average person to say, Wow, 0:01:55.560 --> 0:01:58.360 hang on, what's going on now. I say that because, 0:01:58.760 --> 0:02:00.360 as I have pointed out in this show so many 0:02:00.360 --> 0:02:03.560 many times, there are lots of different aspects of artificial intelligence, 0:02:03.960 --> 0:02:07.160 some of which have been around for many years, and 0:02:07.240 --> 0:02:10.480 some of them have even been causing problems for many years. 0:02:11.000 --> 0:02:15.480 See also facial recognition technology and the fact that bias 0:02:15.560 --> 0:02:19.840 and systems can lead to really terrible consequences in the 0:02:19.880 --> 0:02:22.880 real world. And today I wanted to talk about how 0:02:22.919 --> 0:02:24.720 some of the folks in the AI field are voice 0:02:24.720 --> 0:02:29.040 and concerns that they have around AI and AI's evolution 0:02:29.320 --> 0:02:32.840 and also its deployment and how it could be a 0:02:32.880 --> 0:02:35.919 destructive tool in the future. Now, if you've been listening 0:02:35.960 --> 0:02:38.440 to me for a while, you know I try to 0:02:38.480 --> 0:02:42.280 take a very thoughtful approach to this. I think it's 0:02:42.320 --> 0:02:46.200 important to understand the capabilities of AI, and it's also 0:02:46.280 --> 0:02:49.920 important to understand the potential misuses of AI, or at 0:02:50.000 --> 0:02:54.480 least the unintended consequences of using AI. But I also 0:02:54.520 --> 0:02:59.080 want to try to avoid fud that is, fear, uncertainty 0:02:59.240 --> 0:03:03.000 and doubt that can air on the side of being 0:03:03.080 --> 0:03:07.040 an alarmist. So I think that we should be concerned, 0:03:07.440 --> 0:03:10.640 But so far I haven't been ready to push the 0:03:10.720 --> 0:03:14.760 panic button just yet for AI. But maybe that's about 0:03:14.800 --> 0:03:17.800 to change, because while I've been trying to wrap my 0:03:17.880 --> 0:03:21.959 brain around this, a person like doctor Jeffrey Hinton has 0:03:22.000 --> 0:03:24.480 come forward with his own concerns about AI. And if 0:03:24.480 --> 0:03:29.119 doctor Hinton is concerned, I should probably listen. And that's 0:03:29.160 --> 0:03:32.080 because doctor Hinton has been at the cutting edge of 0:03:32.160 --> 0:03:37.320 AI development for years for decades, particularly in fields like 0:03:37.560 --> 0:03:41.920 artificial neural networks and deep neural networks. In particular, he 0:03:42.000 --> 0:03:45.040 recently resigned from his position at Google, where he had 0:03:45.080 --> 0:03:48.480 been working in AI research, at age seventy five. He's 0:03:48.480 --> 0:03:50.720 certainly at a point in his life where retirement would 0:03:50.720 --> 0:03:53.440 seem pretty natural. You would just think that he would 0:03:53.440 --> 0:03:55.680 come to the conclusion of, yes, it's time for me 0:03:55.800 --> 0:03:58.680 to rest. But his decision was made at least in 0:03:58.760 --> 0:04:03.000 part so that he could speak out about AI and 0:04:03.400 --> 0:04:08.120 the dangers he considers to be important without considering how 0:04:08.160 --> 0:04:12.760 it would impact Google. And that's from doctor Hinton himself. 0:04:12.800 --> 0:04:15.480 He posted that on Twitter, where he said he was 0:04:15.960 --> 0:04:19.400 doing this without considering how it would impact on Google. 0:04:19.839 --> 0:04:22.320 He was addressing a New York Times article that implied 0:04:22.360 --> 0:04:25.040 he had left Google so that he could criticize Google 0:04:25.080 --> 0:04:27.760 in particular. He was quick to say that he felt 0:04:27.800 --> 0:04:31.679 Google had been pretty responsible in its pursuit of AI, 0:04:31.880 --> 0:04:36.599 at least arguably until relatively recently. So let's learn a 0:04:36.640 --> 0:04:40.360 bit about doctor Hinton and his background, the work that 0:04:40.400 --> 0:04:44.120 he pursued, and what his concerns around AI actually cover, 0:04:44.560 --> 0:04:47.800 and maybe along the way we'll figure out some questions 0:04:47.839 --> 0:04:52.120 that we need to answer at some point, implications that 0:04:52.200 --> 0:04:56.120 need to be considered, and perhaps choices we absolutely should 0:04:56.160 --> 0:04:59.039 not make if we want to create helpful AI that 0:04:59.080 --> 0:05:02.240 provides a net benefit rather than something that you know, 0:05:03.120 --> 0:05:07.160 creates the terminator or how or whatever. And I am 0:05:07.200 --> 0:05:10.200 being a bit flippant, but there are reasons we should 0:05:10.279 --> 0:05:14.279 have some concerns, even if they don't involve single minded 0:05:14.440 --> 0:05:20.119 cyborg soldiers. Jeffrey Hinton was born in nineteen forty seven 0:05:20.279 --> 0:05:24.800 in London, England. He attended the University of Edinburgh and 0:05:24.880 --> 0:05:28.400 graduated with a degree in psychology in nineteen sixty nine, 0:05:28.440 --> 0:05:31.240 which is an interesting starting point for someone who had 0:05:31.240 --> 0:05:35.080 become deeply involved in computer science, and the background in 0:05:35.120 --> 0:05:38.400 psychology is probably an important component for someone who would 0:05:38.440 --> 0:05:42.560 contribute to the advancement of artificial intelligence in general and 0:05:42.720 --> 0:05:46.560 neural networks in particular. In nineteen seventy eight, he earned 0:05:46.560 --> 0:05:50.479 a PhD in artificial intelligence at the University of Sussex. 0:05:50.920 --> 0:05:54.880 He transitioned into being an AI researcher, but it was 0:05:54.960 --> 0:05:57.440 kind of a tough go in the UK. There just 0:05:57.520 --> 0:06:01.320 really wasn't that much support and funding for AI research 0:06:01.480 --> 0:06:04.400 over in the UK, so it was hard for him 0:06:04.440 --> 0:06:07.480 to make much progress. So he decided to immigrate to 0:06:07.520 --> 0:06:10.680 the United States, where he first worked as a researcher 0:06:10.720 --> 0:06:13.880 at the University of California in San Diego, and then 0:06:13.920 --> 0:06:18.039 he moved on to Carnegie Mellon University and he worked 0:06:18.080 --> 0:06:20.960 at Carnegie Mellon as a professor from nineteen eighty two 0:06:21.080 --> 0:06:24.360 to nineteen eighty seven, but by the late eighties he 0:06:24.440 --> 0:06:29.200 made a decision to relocate to Canada and you might say, well, 0:06:29.200 --> 0:06:32.039 why would you go to Canada when you were already 0:06:32.040 --> 0:06:34.240 working in AI in the United States. I mean, the 0:06:34.320 --> 0:06:39.320 US has spent billions of dollars in research and development 0:06:39.440 --> 0:06:44.400 in the technology field. Well, his primary reason was because 0:06:44.480 --> 0:06:47.640 the main source for research funding in AI at the 0:06:47.680 --> 0:06:51.400 time came from the Department of Defense, and doctor Hinton 0:06:51.600 --> 0:06:54.520 wasn't comfortable with the idea of working on machine intelligence 0:06:55.160 --> 0:06:59.960 that was through military backing, because the presumption is whatever 0:07:00.000 --> 0:07:03.279 our work you create is ultimately going to be put 0:07:03.320 --> 0:07:06.200 to use by the Department of Defense, and it's reasonable 0:07:06.240 --> 0:07:09.920 to assume that at least some of those uses could 0:07:09.920 --> 0:07:13.720 be weaponized, and Hinton didn't want to contribute to work 0:07:13.760 --> 0:07:17.400 that could later be used to harm or kill others, 0:07:17.880 --> 0:07:21.720 so he would rather sidestep that and get funding from 0:07:21.760 --> 0:07:26.720 other sources. So he settled in Toronto, Canada. He took 0:07:26.760 --> 0:07:30.600 on more academic roles. He continued to be professor. He 0:07:30.680 --> 0:07:33.040 also continued to work in the field of AI research, 0:07:33.160 --> 0:07:37.600 specifically in artificial neural networks and deep learning approaches, and 0:07:37.640 --> 0:07:40.320 we will talk more about those in just a little bit. 0:07:40.800 --> 0:07:44.239 In twenty twelve, he co founded a company with two 0:07:44.360 --> 0:07:49.080 of his students after publishing a paper on deep learning. 0:07:49.440 --> 0:07:53.320 So his paper got the attention of some really smart 0:07:53.360 --> 0:07:57.000 people around the world, and before Hinton knew it, he 0:07:57.120 --> 0:08:00.640 was being courted by some really big companies, companies that 0:08:00.680 --> 0:08:04.520 had super deep pockets and wanted to hire him and 0:08:04.560 --> 0:08:07.560 a couple of his students on to work in the 0:08:07.560 --> 0:08:11.120 field of AI research. So he got one offer from 0:08:11.200 --> 0:08:14.080 the Chinese company Baidu that would have had him and 0:08:14.160 --> 0:08:16.760 his two students work for the company for a few 0:08:16.840 --> 0:08:21.360 years in return for twelve million smackaroo's worth of compensation. 0:08:22.120 --> 0:08:26.280 That's a healthy salary. But Hinton also had other folks 0:08:26.480 --> 0:08:29.000 who were potentially interested in his work, and he also 0:08:29.080 --> 0:08:33.000 figured it would be far more lucrative if he created 0:08:33.080 --> 0:08:37.120 a company with his students, if they made a company 0:08:37.160 --> 0:08:41.200 together that could then be acquired, so instead of getting 0:08:41.320 --> 0:08:44.520 hired as individuals, they would have a company that would 0:08:44.520 --> 0:08:46.800 have to be purchased. And so that's when he and 0:08:46.920 --> 0:08:52.520 these two students incorporated into DNN Research. The DNN stands 0:08:52.520 --> 0:08:57.760 for deep Neural Networks. Hinton then took this brand new 0:08:57.800 --> 0:09:02.640 company which really just had three employees, including himself, and 0:09:02.760 --> 0:09:06.800 had no products and no services, no business plan, nothing 0:09:06.840 --> 0:09:09.760 other than the fact that it was incorporated, and then 0:09:09.760 --> 0:09:13.560 he put it up for auction. The actual auction took 0:09:13.600 --> 0:09:18.560 place in Lake Tahoe during a conference on AI and 0:09:18.640 --> 0:09:24.120 machine learning, and by Do participated, but so did Microsoft, Google, 0:09:24.320 --> 0:09:27.960 and an AI research company called DeepMind, which a couple 0:09:28.000 --> 0:09:32.600 of years later would become a Google subsidiary of its own. Now, 0:09:32.640 --> 0:09:36.280 DeepMind was the first company to bow out of the auction. 0:09:36.400 --> 0:09:40.040 It just did not have the resources of these three 0:09:40.240 --> 0:09:45.360 giant tech companies. Microsoft then followed. Actually, Microsoft kind of 0:09:45.360 --> 0:09:47.120 bounced in and out of the auction a couple of 0:09:47.160 --> 0:09:50.959 times before finally throwing in the towel, and the bidding 0:09:51.040 --> 0:09:53.360 war came down to Google versus by Do, and it 0:09:53.440 --> 0:09:56.280 just kept going and going and going. Once the price 0:09:56.400 --> 0:10:00.640 hit an astounding forty four million dollars. And keep in 0:10:00.679 --> 0:10:04.520 mind DNN Research had only been around for a very 0:10:04.559 --> 0:10:07.640 short while and had no products or services to its name, 0:10:08.320 --> 0:10:11.240 doctor Hinton called the auction closed and the company went 0:10:11.280 --> 0:10:14.720 to its new owner, that of Google. Reportedly, the people 0:10:14.760 --> 0:10:18.079 at Google were actually surprised that he stopped the auction 0:10:18.200 --> 0:10:21.400 at that point, because they figured that he was leaving 0:10:21.440 --> 0:10:23.960 millions of dollars on the table that the bidding war 0:10:24.080 --> 0:10:27.160 would have continued between Google and by Doo, and he 0:10:27.200 --> 0:10:30.320 could have gotten more for it, but doctor Hinton was 0:10:30.440 --> 0:10:33.439 more concerned with working for Google rather than for by 0:10:33.520 --> 0:10:36.000 Do and felt that forty four million dollars was more 0:10:36.080 --> 0:10:40.760 than enough. So that's a pretty you know, mature approach 0:10:40.840 --> 0:10:44.000 as opposed to let's take every cent we can grab. 0:10:44.559 --> 0:10:48.200 Also in twenty twelve, he won an award when he 0:10:48.320 --> 0:10:52.280 co invented a deep learning model called alex Net, named 0:10:52.280 --> 0:10:56.840 after the other co creator, his student, Alex Krzewski. This 0:10:57.000 --> 0:10:59.200 was bigger than just a two person operation. By the way, 0:10:59.240 --> 0:11:03.480 it's not like Alex Krzewski and doctor Hinton were the 0:11:03.480 --> 0:11:05.560 only two to work on it, but they were the 0:11:05.640 --> 0:11:10.080 leads on this project and it was named after Alex. Alex, 0:11:10.080 --> 0:11:11.720 by the way, was also one of the two students 0:11:11.720 --> 0:11:15.080 who was part of DNN research, the other being a 0:11:15.120 --> 0:11:20.800 student named Iliya Sutzkever. And my apologies for the butchering 0:11:20.840 --> 0:11:26.120 of pronunciation, but the learning model Alex Natt focused quite literally, 0:11:26.360 --> 0:11:30.080 i guess you could say, on image recognition and participated 0:11:30.120 --> 0:11:32.720 in a competition in which the model proved to have 0:11:32.760 --> 0:11:36.440 an eighty five percent accuracy rate. And while it's a 0:11:36.440 --> 0:11:38.480 trivial thing for a human to look at a photo 0:11:38.600 --> 0:11:41.560 and say something like that's a bunny rabbit. It's not 0:11:41.720 --> 0:11:43.960 so trivial to create a way for computers to be 0:11:44.040 --> 0:11:47.559 able to do the same thing. So this eighty five 0:11:47.600 --> 0:11:52.040 percent accuracy rate was like that was like a stake 0:11:52.120 --> 0:11:56.000 in the ground saying we have made a massive leap 0:11:56.000 --> 0:11:58.920 ahead with machine learning and artificial intelligence. It was one 0:11:58.920 --> 0:12:03.000 of the reasons why DNA research was so highly sought after, 0:12:03.559 --> 0:12:06.839 and alex Nett wasn't just an impressive approach toward machine learning. 0:12:08.280 --> 0:12:11.200 It really got enough buzz that money began to pour 0:12:11.280 --> 0:12:16.200 into deep learning projects everywhere, not just with doctor Hinton 0:12:16.240 --> 0:12:20.080 and his students, Like we literally started to see more 0:12:20.120 --> 0:12:24.240 development in the discipline as a whole because this was 0:12:24.240 --> 0:12:29.000 such an impressive display. From twenty twelve until just this year, 0:12:29.760 --> 0:12:32.679 doctor Hinton worked in AI research and deep learning, in 0:12:32.720 --> 0:12:37.280 particular over at Google. One of his two students, Ilijas Skiver, 0:12:37.640 --> 0:12:41.160 actually would leave Google to join a little AI nonprofit 0:12:41.280 --> 0:12:45.079 called open Ai. I mean, originally it was a nonprofit, 0:12:45.120 --> 0:12:48.320 and technically the nonprofit part of open ai is still 0:12:48.320 --> 0:12:52.720 a parent organization, but really the for profit arm of 0:12:52.720 --> 0:12:56.079 open AI is in the news way more frequently these days. 0:12:56.559 --> 0:12:59.560 In twenty eighteen, doctor Hinton was a co recipient of 0:12:59.600 --> 0:13:03.640 the award. This is a prestigious honor for those in 0:13:03.679 --> 0:13:06.600 the computing field. Some people even refer to it as 0:13:06.640 --> 0:13:11.080 being the equivalent to a Nobel Prize. And now he's 0:13:11.280 --> 0:13:15.520 stepping forward with concerns relating to the work he dedicated 0:13:15.840 --> 0:13:19.000 his life too. Now we're going to take a quick break. 0:13:19.000 --> 0:13:22.640 When we come back, we're going to talk about deep 0:13:22.720 --> 0:13:26.240 neural networks and what they do within the realm of 0:13:26.360 --> 0:13:40.040 machine learning. But first these messages. Okay, we're back. Let's 0:13:40.120 --> 0:13:43.439 talk about doctor Hinton's work and deep neural networks. Now, 0:13:43.480 --> 0:13:47.640 as you might imagine, this subject gets really complicated. It's 0:13:47.679 --> 0:13:51.480 really nuanced, it's technical, and as I'm sure you have 0:13:51.559 --> 0:13:55.120 no need to imagine, my understanding of deep neural networks 0:13:55.200 --> 0:13:57.959 is pretty limited. I mean, you could call it surface 0:13:58.040 --> 0:14:00.440 level and I wouldn't be able to disagree. So we're 0:14:00.440 --> 0:14:04.720 going to paint this topic in broad strokes. And I'm 0:14:04.720 --> 0:14:08.160 doing this not to dumb it down, but rather to 0:14:08.200 --> 0:14:10.360 do my best to kind of get across the general 0:14:10.440 --> 0:14:14.199 way it works without making too many egregious errors. Along 0:14:14.240 --> 0:14:18.800 the way. So first up, the goal of a deep 0:14:18.960 --> 0:14:23.600 neural network is to provide a learning mechanism that mimics 0:14:23.640 --> 0:14:27.920 the human brain, but using a computer rather than a 0:14:27.960 --> 0:14:31.520 human brain. So, for the purposes of an overly simple 0:14:31.560 --> 0:14:35.720 thought experiment, imagine you've got a black box. It's an 0:14:35.760 --> 0:14:39.320 opaque black box. Now, one side of the box allows 0:14:39.360 --> 0:14:42.000 you to put something in, and the other side of 0:14:42.000 --> 0:14:46.640 the box allows stuff to come out. And let's say 0:14:46.640 --> 0:14:51.920 that you are putting in one thing and it transforms 0:14:51.960 --> 0:14:53.800 in some way inside the box and comes out as 0:14:53.800 --> 0:14:57.520 something else. That's the general thought here, because I'm feeling 0:14:57.800 --> 0:15:01.800 a little peckish and a little puckish. Let's say that 0:15:01.880 --> 0:15:06.000 you decide to put in the inputs as the ingredients 0:15:06.000 --> 0:15:08.080 you would need for a pizza, and you're shoving that 0:15:08.120 --> 0:15:10.920 into the box. So we're talking stuff like pizza dough 0:15:11.400 --> 0:15:15.400 and some sauce and some cheese and any toppings you like, 0:15:16.040 --> 0:15:18.280 and you shove that into the input in the box, 0:15:18.280 --> 0:15:21.600 and then the output shoots out a cowl zone. Well shucks, 0:15:22.080 --> 0:15:24.720 you think, unless you're Ben Wyatt from Parks and rec 0:15:24.760 --> 0:15:27.400 in which case you celebrate because you think col zones 0:15:27.440 --> 0:15:30.040 are superior to pizza in every way. But assuming you're 0:15:30.120 --> 0:15:33.720 not Ben Wyatt, you say, that's not what I wanted. 0:15:33.880 --> 0:15:36.400 I wanted to get a pizza, not a calzone. So 0:15:36.440 --> 0:15:38.680 you have to open up the box, you have to 0:15:38.720 --> 0:15:41.560 adjust some stuff inside it, you have to close it 0:15:41.600 --> 0:15:44.040 all up and try it again, and you keep doing 0:15:44.080 --> 0:15:46.600 this over and over until you get a pizza, a 0:15:46.720 --> 0:15:50.680 properly cooked and prepared pizza. This is sort of similar 0:15:50.720 --> 0:15:56.320 to how computer scientists perform supervised learning with artificial neural networks, 0:15:56.760 --> 0:16:01.800 because that box represents what we call hidden layers. They 0:16:01.840 --> 0:16:05.000 could be lots of hidden layers, and these are layers 0:16:05.080 --> 0:16:10.160 of computer nodes that serve as artificial neurons, and pathways 0:16:10.280 --> 0:16:14.760 form between these different nodes as they process information. So 0:16:15.080 --> 0:16:18.680 when you put input into the system, that input goes 0:16:18.720 --> 0:16:21.240 to a node and it begins to sort the data 0:16:21.400 --> 0:16:24.560 based on some criteria that the system has been trained on. 0:16:24.640 --> 0:16:28.000 Whatever the purpose of the system actually is. It's kind 0:16:28.040 --> 0:16:30.880 of like, you know, let's say it's for recognizing bunnies, 0:16:30.920 --> 0:16:33.880 since we use that example earlier. So you feed it 0:16:33.960 --> 0:16:37.240 a whole bunch of images, and the node takes the 0:16:37.320 --> 0:16:40.360 data and passes the data to another node. A layer 0:16:40.440 --> 0:16:43.600 down and it does it. It chooses the node based 0:16:43.680 --> 0:16:48.720 on some transformational function at that artificial neuron, right, so 0:16:48.760 --> 0:16:52.320 you can think data comes in, the neuron, performs a 0:16:52.440 --> 0:16:55.800 transformational function on this data. Based on that result, it 0:16:55.840 --> 0:16:59.640 goes to one node or another, and then the process 0:16:59.680 --> 0:17:02.200 repeat and it does this again and again until it 0:17:02.240 --> 0:17:06.480 comes out the output side. Where like in our example, 0:17:07.480 --> 0:17:10.360 we figure out whether the machine is able to recognize 0:17:10.400 --> 0:17:12.639 if a picture has a bunny in it or not. 0:17:13.359 --> 0:17:18.439 So you feed millions, tens of millions, hundreds of millions 0:17:18.480 --> 0:17:22.879 of pictures to this system to train it. When you 0:17:22.920 --> 0:17:26.280 start off, you might be doing this with a bunch 0:17:26.320 --> 0:17:29.080 of images that you've already determined whether or not there 0:17:29.080 --> 0:17:33.360 are bunnies in them. So you've got to control amount 0:17:33.359 --> 0:17:36.040 of data that you're feeding just for the purposes of 0:17:36.080 --> 0:17:40.159 training your system. And at the end, after it's sorted 0:17:40.200 --> 0:17:43.320 through those images, you evaluate the system to see how 0:17:43.359 --> 0:17:46.639 well it did in figuring out whether an image had 0:17:46.640 --> 0:17:48.359 a bunny in it or not. Maybe in some of 0:17:48.400 --> 0:17:51.560 the pictures it misses a bunny. Maybe in some pictures 0:17:51.640 --> 0:17:54.360 it thinks there's a bunny there and there's not. And 0:17:54.400 --> 0:17:57.560 then you might go in and start to adjust the 0:17:57.600 --> 0:18:01.520 weights on those artificial neurons. This is the thing that 0:18:01.800 --> 0:18:06.520 creates that transformational function. You might tweak those transformational functions 0:18:06.560 --> 0:18:09.680 a little bit. You might start closest to the output 0:18:09.720 --> 0:18:12.760 and work your way back. That's called back propagating. And 0:18:13.000 --> 0:18:16.920 what you're trying to do is adjust all these settings 0:18:17.240 --> 0:18:19.800 so that it is more accurate the next time you 0:18:19.840 --> 0:18:22.200 feed all the images through, and you do it again, 0:18:22.840 --> 0:18:26.520 and you might do this dozens or hundreds of times 0:18:26.600 --> 0:18:29.239 in an effort to really refine your model so that 0:18:29.280 --> 0:18:32.560 it gets better and better at identifying the pictures that 0:18:32.600 --> 0:18:36.000 have bunnies in them. And then ideally you get the 0:18:36.040 --> 0:18:37.879 system to a point where you could just feed it 0:18:38.080 --> 0:18:40.960 raw data, like you haven't even looked at these images. 0:18:41.000 --> 0:18:45.080 You're just dumping millions of images in and you're letting 0:18:45.160 --> 0:18:47.680 it sort it through. And because it has reached the 0:18:47.800 --> 0:18:51.800 level of accuracy that it's at, because you've trained it 0:18:51.840 --> 0:18:54.119 for so long, you don't even have to worry so 0:18:54.280 --> 0:18:56.880 much about whether or not it caught all the images 0:18:57.200 --> 0:19:00.359 or if it misidentified some there's probably going to be 0:19:00.400 --> 0:19:02.920 some error in there, but if your accuracy level is 0:19:02.960 --> 0:19:06.399 high enough, then it's possibly good enough for whatever purpose 0:19:06.480 --> 0:19:09.679 you've built it for. And yeah, the more data you 0:19:09.840 --> 0:19:13.439 use to train your machine learning model in general, the 0:19:13.560 --> 0:19:17.240 better it will perform, because it'll start to eliminate things 0:19:17.280 --> 0:19:20.879 like outliers. And while image recognition is just one of 0:19:20.920 --> 0:19:23.880 the more famous uses for deep neural networks in machine learning, 0:19:23.920 --> 0:19:26.760 it is clearly not the only one. The one we've 0:19:26.760 --> 0:19:30.840 been hearing about a lot lately involves large language models 0:19:30.960 --> 0:19:33.200 or llms, like I mentioned at the top of the show. 0:19:33.280 --> 0:19:38.720 So imagine feeding millions or even billions of documents to 0:19:38.800 --> 0:19:43.560 a neural network that's trained to recognize patterns in language. 0:19:44.000 --> 0:19:46.679 So you're feeding all sorts of stuff to this model, 0:19:47.240 --> 0:19:51.720 and as you do, the system quote unquote learns how 0:19:52.000 --> 0:19:55.639 words follow each other, like which words are likely to 0:19:55.680 --> 0:19:58.359 follow other words. You probably wouldn't go so far as 0:19:58.359 --> 0:20:03.240 to say the system under stands a language like English, 0:20:03.280 --> 0:20:06.320 but it does have an incredibly sophisticated statistical model that 0:20:06.400 --> 0:20:09.680 breaks down how likely one word is to follow another. 0:20:10.520 --> 0:20:12.240 So you can think of it a little bit like 0:20:12.280 --> 0:20:15.520 a word association game. You've probably played something like this 0:20:15.560 --> 0:20:17.760 at some point or another. Someone gives you a word 0:20:18.160 --> 0:20:20.120 and you're supposed to say the first word that comes 0:20:20.119 --> 0:20:22.480 to your mind. So if I were to say the 0:20:22.520 --> 0:20:27.720 word nuclear, you might think power or bomb or radiation. 0:20:28.760 --> 0:20:34.960 You probably wouldn't think penguin or Chesterfield sofa just not 0:20:35.160 --> 0:20:38.000 likely to pop up statistically, it's unlikely. Well, you can 0:20:38.080 --> 0:20:40.640 kind of think of the large language model as being 0:20:41.200 --> 0:20:45.119 an enormous version of that. So as these large language 0:20:45.119 --> 0:20:49.000 models process increasing amounts of information, and as the neural 0:20:49.080 --> 0:20:54.879 network experiences refinement over countless learning runs, you end up 0:20:54.880 --> 0:20:57.320 with a system that is capable of doing some pretty 0:20:57.400 --> 0:21:00.760 extraordinary things, at least on the surface life. It can 0:21:00.840 --> 0:21:06.000 pull information together to answer questions about practically any topic. Unfortunately, 0:21:06.760 --> 0:21:11.000 it can also invent answers by following a statistical probability 0:21:11.400 --> 0:21:14.720 when it doesn't actually contain the answers to the question 0:21:14.760 --> 0:21:17.080 you asked. This means you can end up with an 0:21:17.119 --> 0:21:21.000 answer that isn't accurate at all, but it follows a 0:21:21.040 --> 0:21:24.919 statistical model where each word is from a probability standpoint, 0:21:25.480 --> 0:21:29.800 the perfect word to go in that point in a sentence, 0:21:29.840 --> 0:21:31.639 which is a weird thing to think about, right, Like, 0:21:31.920 --> 0:21:35.560 the answer you get isn't right, but each word is, 0:21:35.760 --> 0:21:40.040 statistically speaking, the best one to put in that place 0:21:40.280 --> 0:21:45.680 in lieu of any actual information. Now here's another thing 0:21:45.720 --> 0:21:49.159 to consider. These tools can do stuff like build code. 0:21:49.880 --> 0:21:53.239 This code isn't always reliable, it's not always right, but 0:21:53.640 --> 0:21:56.760 sometimes it is. So maybe you use a tool like 0:21:56.880 --> 0:22:00.879 GPT to look over the code that was made by 0:22:00.920 --> 0:22:03.119 a group of engineers, and you do it to search 0:22:03.119 --> 0:22:06.680 for errors, like you're using this to look for mistakes 0:22:06.680 --> 0:22:09.080 that were made in the code. Or maybe you use 0:22:09.160 --> 0:22:11.280 it to see if there's a way to make the 0:22:11.280 --> 0:22:16.679 code that was written more elegant or efficient. Maybe you 0:22:16.720 --> 0:22:19.080 figure you've reached a point where you don't even need 0:22:19.440 --> 0:22:23.520 human engineers because the AI agent performs at a standard 0:22:23.600 --> 0:22:27.440 that's high enough to replace them. Maybe you think it's 0:22:27.440 --> 0:22:30.359 even better than what human engineers can do, and that 0:22:30.400 --> 0:22:33.520 it's far faster, and that you can therefore develop and 0:22:33.600 --> 0:22:37.080 deployee software at a pace that you couldn't before. So 0:22:37.160 --> 0:22:41.280 the IT industry is in a particularly delicate place as 0:22:41.320 --> 0:22:45.880 companies begin to explore how AI could augment or potentially 0:22:46.440 --> 0:22:51.200 replace people. I go back to what IBM's CEO recently said. 0:22:51.480 --> 0:22:55.280 He said that for nearly eight thousand job offerings that 0:22:55.320 --> 0:22:58.960 the company has now put on a hiring freeze, he 0:22:59.119 --> 0:23:02.560 might never hire a human to take one of those jobs. Instead, 0:23:02.600 --> 0:23:06.439 he might rely on automation and AI to cover that job. 0:23:06.840 --> 0:23:10.240 So it's not quite the same thing as firing someone 0:23:10.280 --> 0:23:12.720 and then replacing him with a robot, but it is 0:23:12.800 --> 0:23:16.000 given a robot a job instead of a human being. Okay, 0:23:16.640 --> 0:23:19.960 let's switch gears. Let's talk about AI in the arts, 0:23:20.040 --> 0:23:24.119 because that's also a really relevant conversation right now. So 0:23:24.359 --> 0:23:26.919 last year we already started to see debates about the 0:23:27.000 --> 0:23:32.760 validity of AI generated images. Should an AI generated image 0:23:32.800 --> 0:23:38.600 be considered art? We saw people submit AI generated paintings 0:23:38.680 --> 0:23:44.040 into competitions, some of which ended up receiving awards, and 0:23:44.080 --> 0:23:47.280 then we're subsequently either stripped of those awards or you know, 0:23:47.320 --> 0:23:50.040 people got in trouble for using AI even when they 0:23:50.040 --> 0:23:53.560 were you know, admitting to it in an effort to say, hey, 0:23:53.640 --> 0:23:56.199 we're trying to start a conversation about AI and its 0:23:56.280 --> 0:24:00.720 role in arts. So is art actually art if the 0:24:00.800 --> 0:24:03.760 image is a product of a complex series of decisions 0:24:04.080 --> 0:24:08.640 that aren't driven by imagination or creativity, but rather some 0:24:09.880 --> 0:24:14.200 really weird statistical model that's so complicated that no one 0:24:14.280 --> 0:24:18.920 really understands it. Or is it just a meaningless image? 0:24:19.240 --> 0:24:23.080 You know, maybe it's an image that mimics specific artists, 0:24:23.160 --> 0:24:26.119 but in itself it's nothing more than just a picture. 0:24:26.520 --> 0:24:29.879 I mean, you know, drawing a perfect circle freehand with 0:24:30.000 --> 0:24:32.560 no tools is really really hard for a human to do, 0:24:32.800 --> 0:24:34.520 but it's a piece of cake for a computer. So 0:24:34.520 --> 0:24:38.000 should we be astounded by a computer's ability to generate 0:24:38.000 --> 0:24:41.800 a perfect circle? What about a computer's ability to mimic 0:24:42.040 --> 0:24:47.840 the style of say, you know, Picasso or Dali. Beyond 0:24:47.920 --> 0:24:51.200 visual arts, there are examples like writing and music. There's 0:24:51.240 --> 0:24:53.520 the case of the song Hard on My Sleeve that 0:24:53.720 --> 0:24:57.000 features the deep fake voices of Drake and the Weekend, 0:24:57.720 --> 0:25:00.680