WEBVTT - Artificial Artificial Intelligence 0:00:04.480 --> 0:00:12.399 Welcome to tech Stuff, a production from iHeartRadio. Hey there, 0:00:12.440 --> 0:00:15.960 and welcome to tech Stuff. I'm your host Jonathan Strickland. 0:00:15.960 --> 0:00:19.000 I'm an executive producer with iHeart Podcasts and how the 0:00:19.079 --> 0:00:24.720 tech are you? So? In the late eighteenth century, there 0:00:24.760 --> 0:00:28.360 was a man named Wolfgang von Kimpland, and he had 0:00:28.400 --> 0:00:32.520 a clever idea. He really wanted to knock the proverbial 0:00:32.760 --> 0:00:38.640 socks off of Maria Theresa, the Impress of Austria. Moreover, 0:00:38.920 --> 0:00:43.479 he wanted to make a more spectacular display than an 0:00:43.560 --> 0:00:49.160 illusionist named Francois Pelletier who had performed for the Impress 0:00:49.280 --> 0:00:54.240 to great renown, and Kimpland was not impressed. He was like, huh, 0:00:54.280 --> 0:00:57.200 I'm gonna show Frank up. I'm gonna make something that's 0:00:57.240 --> 0:00:59.560 really gonna rub his face in it, and the Impress 0:00:59.640 --> 0:01:03.440 is gonna think I'm her favorite. So, fueled by a 0:01:03.560 --> 0:01:07.679 competitive and perhaps petty spirit, Kimpland came up with an 0:01:07.680 --> 0:01:13.200 invention that some would call the mechanical Turk. Now the 0:01:13.840 --> 0:01:17.759 machine hesitate to call it that, but the machine consisted 0:01:17.800 --> 0:01:21.559 of a large table like cabinet, and on the top 0:01:21.640 --> 0:01:26.040 of this cabinet was a chessboard and standing behind this 0:01:26.319 --> 0:01:30.920 cabinet was a mechanical man dressed in the Western European 0:01:31.040 --> 0:01:35.240 concept of traditional Turkish attire. If you were to open 0:01:35.319 --> 0:01:38.800 the cabinet doors, you would reveal a mass of gears 0:01:38.840 --> 0:01:42.080 and cogs and such, so it looked as though everything 0:01:42.360 --> 0:01:45.840 was mechanical. Kimplan claimed that this machine could play an 0:01:45.880 --> 0:01:49.120 expert game of chess against any opponent, and as it 0:01:49.160 --> 0:01:53.200 turned out, the machine performed very well and won more 0:01:53.240 --> 0:01:55.720 games than it lost. But it was all a trick. 0:01:56.000 --> 0:02:00.120 The machine wasn't really a machine, or at least it 0:02:00.160 --> 0:02:05.120 wasn't a machine that did any work. Instead, hidden inside 0:02:05.160 --> 0:02:09.639 the cabinet, concealed from exposure, hidden by these gears and cogs, 0:02:10.080 --> 0:02:15.600 was a cramped human chess player, and the player could 0:02:15.639 --> 0:02:18.720 manipulate the Turkish figure and was able to play chess 0:02:18.760 --> 0:02:21.720 from below the chess board. So it was an actual 0:02:21.800 --> 0:02:25.160 human being who was actually playing these games against people. 0:02:25.240 --> 0:02:30.040 It wasn't some mechanical construct. The Turk only seemed to 0:02:30.120 --> 0:02:33.480 be a chess playing machine. Now, this was way back 0:02:33.520 --> 0:02:38.600 in seventeen seventy. Today, in twenty twenty four, we still 0:02:38.600 --> 0:02:43.040 have to deal with companies and entrepreneurs peddling artificial intelligence 0:02:43.080 --> 0:02:46.440 that when you look at it more closely, is really 0:02:46.480 --> 0:02:52.840 relying on plain, old, reliable human intelligence. Why, Well, the 0:02:53.120 --> 0:02:57.400 short answer for that is money. It seems like, you know, 0:02:57.520 --> 0:03:00.800 not a day goes by in twenty twenty that doesn't 0:03:00.840 --> 0:03:04.600 include at least one news story about how artificial intelligence 0:03:04.680 --> 0:03:08.240 is going to completely change our lives. And the stories 0:03:08.320 --> 0:03:12.799 run the gamut of hyperbole, from doomsday prophecies about weaponized 0:03:12.840 --> 0:03:17.200 AI making battlefield decisions, to company executives as saying that 0:03:17.240 --> 0:03:21.560 AI programs are a viable alternative to hiring actual human beings, 0:03:21.800 --> 0:03:27.440 to optimists who describe a star trek like utopia in 0:03:27.480 --> 0:03:30.480 which AI handles all the dull stuff and it leaves 0:03:30.560 --> 0:03:33.760 us to experience the world as a never ending series 0:03:33.800 --> 0:03:36.960 of adventures. I'm not sure if any of those scenarios 0:03:37.000 --> 0:03:39.480 are what's actually in store for us, but I do 0:03:39.640 --> 0:03:41.920 know things are going to be messy for a good 0:03:42.240 --> 0:03:46.400 long while. But AI is such a buzzy term, and 0:03:46.520 --> 0:03:53.200 with big companies like Google, Microsoft, Amazon, Apple, Meta and 0:03:53.360 --> 0:03:59.640 more all stomping relentlessly forward to make AI the next 0:03:59.680 --> 0:04:04.440 big thing, there are literally billions of dollars pouring into 0:04:04.600 --> 0:04:09.840 various AI pursuits. Now, with that much money, and enthusiasm 0:04:09.920 --> 0:04:13.840 at play. It's no wonder that dozens of startups attempting 0:04:13.880 --> 0:04:16.640 to cash in on the gold rush have cropped up 0:04:17.000 --> 0:04:20.240 in recent years. And some of those companies might actually 0:04:20.320 --> 0:04:25.280 be making genuine strides toward advancing AI or implementing it 0:04:25.360 --> 0:04:28.560 in a useful way. Some might just be jumping on 0:04:28.600 --> 0:04:33.080 the opportunity to get some of that sweet, sweet VC cash. 0:04:33.200 --> 0:04:38.200 Since AI is the new metaverse slash NFT slash virtual 0:04:38.240 --> 0:04:42.080 reality slash three D technology thing, what I'm saying is 0:04:42.120 --> 0:04:46.520 that we've been through this hype cycle many many times before. 0:04:46.920 --> 0:04:53.480 The term AI itself is incredibly useful if you want to, 0:04:53.520 --> 0:04:57.599 you know, sell some snake oil, because AI as a 0:04:57.720 --> 0:05:02.640 term is still a bit vague. Like the term AI 0:05:03.200 --> 0:05:07.760 is seventy years old at this point, and yet we 0:05:07.880 --> 0:05:12.760 don't have an easy definition for what really is artificial intelligence. 0:05:12.800 --> 0:05:16.440 It's kind of like our definition for actual intelligence. We 0:05:16.560 --> 0:05:20.839 don't have a super great explanation for that either. We 0:05:20.960 --> 0:05:24.360 have ways of describing parts of it, but we don't 0:05:24.360 --> 0:05:28.560 really have a holistic, perfect encapsulation of what is intelligence. 0:05:28.600 --> 0:05:31.800 So how could we do that? For artificial intelligence? You 0:05:31.839 --> 0:05:35.000 don't even have to make an AI application or implementation. 0:05:35.200 --> 0:05:40.559 To take advantage of the opportunities that this vague state 0:05:40.600 --> 0:05:45.200 of affairs creates, you just call whatever thing you're trying 0:05:45.240 --> 0:05:48.120 to sell AI, and you let the hype do the 0:05:48.160 --> 0:05:52.200 work for you. Because people don't understand it fully, you 0:05:52.400 --> 0:05:54.760 probably aren't going to get called out on it unless 0:05:54.800 --> 0:05:58.479 you're really sloppy, which means you can make hay while 0:05:58.520 --> 0:06:00.480 the sun shines and then get up out of town 0:06:00.520 --> 0:06:03.080 when the clouds roll in. So today I thought we 0:06:03.120 --> 0:06:07.120 would talk a bit about fake artificial intelligence, or perhaps 0:06:07.120 --> 0:06:10.840 we should call it artificial artificial intelligence, which in a 0:06:10.880 --> 0:06:13.679 way comes back round to just plain old intelligence, because 0:06:13.680 --> 0:06:15.719 we're going to chat about some cases in which a 0:06:15.839 --> 0:06:19.240 person or group of people passed off stuff that isn't 0:06:19.279 --> 0:06:24.360 really AI, but was rather powered by human intelligence onto 0:06:24.680 --> 0:06:29.480 unsuspecting targets. First, however, let's do a quick refresher on AI, 0:06:29.880 --> 0:06:32.880 because I find that the term is so broad and 0:06:32.920 --> 0:06:36.440 it is so overused that it's really starting to lose 0:06:36.480 --> 0:06:40.279 its meaning. These days, as consumers, you and I, we 0:06:40.400 --> 0:06:45.279 are most likely to encounter AI as applied to generative AI. 0:06:45.640 --> 0:06:48.920 That's the hotness right now, And I know I'm old. 0:06:49.640 --> 0:06:54.279 I use phrases like the hotness. Sorry, but this is 0:06:54.520 --> 0:06:58.360 artificial intelligence that's capable of generating something. Thus you have 0:06:58.680 --> 0:07:03.080 generative AI. Now the something might be written text, it 0:07:03.160 --> 0:07:06.839 might be spoken words, it might be music, it might 0:07:06.920 --> 0:07:10.920 be a sketch or a painting. And there's no denying 0:07:11.280 --> 0:07:15.440 that generative AI can be really impressive when it works well. 0:07:15.640 --> 0:07:17.640 It seems to be able to do the same sort 0:07:17.680 --> 0:07:21.360 of things we humans can, though there remain questions regarding 0:07:21.400 --> 0:07:23.960 how much of that is thanks to the various AI 0:07:24.160 --> 0:07:28.360 programs cribbing from actual human work. Because you'll often hear 0:07:28.640 --> 0:07:34.360 human artists argue that generative AI borrows liberally or outright 0:07:34.440 --> 0:07:37.240 steals if we're being more forthright about all this, and 0:07:37.320 --> 0:07:41.080 does so from human artists. You know, AI doesn't magically 0:07:41.200 --> 0:07:44.640 know how to paint something in a specific style or 0:07:44.920 --> 0:07:49.320 maybe even more specifically, to mimic a particular artists style 0:07:49.440 --> 0:07:53.320 and technique. The AI quote unquote knows how to do 0:07:53.360 --> 0:07:56.080 this because it has been trained to do it on 0:07:56.360 --> 0:08:00.360 countless examples of actual human generated art. That's a real 0:08:00.400 --> 0:08:04.520 problem because it could mean that the AI is lifting 0:08:05.280 --> 0:08:10.560 from real artists and thus potentially putting real artists livelihoods 0:08:10.560 --> 0:08:14.080 at stake. But there are tons of other implementations that 0:08:14.120 --> 0:08:18.000 have nothing or at least little to do with generative AI. So, 0:08:18.120 --> 0:08:24.960 for example, facial recognition technology is a discipline under artificial intelligence. 0:08:25.320 --> 0:08:29.600 The basic task is to compare an incoming signal with 0:08:29.680 --> 0:08:34.400 a database of image records. This is relatively trivial if 0:08:34.440 --> 0:08:38.280 the incoming signal is one that matches the database record precisely. 0:08:38.559 --> 0:08:41.680 In other words, let's say that you've got an incoming 0:08:41.720 --> 0:08:45.800 signal where the camera angle, the lighting, the distance from 0:08:45.920 --> 0:08:48.920 the subject, all of that stuff is the same as 0:08:49.000 --> 0:08:52.440 the reference image that you have in your database. Then 0:08:52.520 --> 0:08:54.839 the computer can very quickly say, yes, these two are 0:08:54.880 --> 0:08:58.600 a match. Typically, it does get trickier if you are 0:08:58.760 --> 0:09:03.640 moving away from whatever types of faces the AI had 0:09:03.640 --> 0:09:07.960 been trained upon. But it gets even trickier if conditions 0:09:08.040 --> 0:09:11.840 are different between the incoming signal and the reference. So 0:09:12.080 --> 0:09:15.559 an example I often give, and this isn't facial recognition, 0:09:15.600 --> 0:09:19.640 this is image recognition. So imagine you have a coffee mug, 0:09:20.040 --> 0:09:22.959 and let's say that we first we have a picture 0:09:23.120 --> 0:09:26.079 of a coffee mug. It's sitting on a table. The 0:09:26.360 --> 0:09:29.680 mug's handle is pointing to the right, you know, to 0:09:29.840 --> 0:09:32.480 our right. As we look at the picture. The mug 0:09:32.720 --> 0:09:35.880 is dark red in color. The body of the mug 0:09:35.920 --> 0:09:38.800 is essentially just a you know, a simple cylinder. There's 0:09:38.840 --> 0:09:41.880 no writing on it or anything. This is the reference 0:09:41.920 --> 0:09:44.199 image that we are using. It's the one that's in 0:09:44.240 --> 0:09:48.440 our database. Now imagine that you've pointed a camera at 0:09:48.440 --> 0:09:52.760 a coffee mug, and this mug is an oversized coffee mug, 0:09:53.120 --> 0:09:56.240 and it's off white in color, and it has the 0:09:56.559 --> 0:10:00.720 words World's Greatest Podcaster on the mud ug and the 0:10:00.720 --> 0:10:03.280 handle's pointed to the left, not the right. And this 0:10:03.440 --> 0:10:06.760 mug actually isn't a perfect cylinder. Let's say that it 0:10:06.840 --> 0:10:10.160 kind of curves outward from the base just about, you know, 0:10:10.240 --> 0:10:13.160 sort of like a bowl more than a cylinder. And 0:10:13.360 --> 0:10:15.880 if I ask you what is this thing, you would 0:10:16.000 --> 0:10:18.080 quickly say, oh, that's a mug, or maybe that's a 0:10:18.080 --> 0:10:20.880 coffee mug. You'd say that right away. You would recognize this. 0:10:21.400 --> 0:10:26.760 But it doesn't match the reference picture in our database perfectly, right, Like, 0:10:27.480 --> 0:10:31.320 it doesn't look exactly like or even close to the 0:10:31.440 --> 0:10:34.800 red reference mug we have in our database. It's got 0:10:34.840 --> 0:10:37.160 features that make it a mug, and you, as a 0:10:37.200 --> 0:10:40.800 human being, can naturally apply your knowledge of those features 0:10:40.800 --> 0:10:43.480 to identify a coffee mug. Even if you've never seen 0:10:43.480 --> 0:10:46.480 that specific mug before, you immediately know, oh, that's a 0:10:46.480 --> 0:10:50.520 coffee mug. Even if the coffee mugs form deviates from 0:10:50.679 --> 0:10:53.320 others you've encountered in the past, You're able to apply 0:10:53.400 --> 0:10:56.360 your intelligence to say that's a coffee mug. But a 0:10:56.400 --> 0:10:59.360 computer cannot do that, not on its own. It has 0:10:59.400 --> 0:11:02.440 to be fed hundreds of thousands, or even millions of 0:11:02.480 --> 0:11:06.959 images of coffee mugs, all in various shapes and sizes 0:11:07.160 --> 0:11:11.840 and colors and orientations to the camera and more. Even then, 0:11:12.040 --> 0:11:14.120 there's no guarantee that the computer will be able to 0:11:14.160 --> 0:11:17.760 identify a new image of a coffee mug that deviates 0:11:17.760 --> 0:11:20.520 from this collection of reference material. So we can help 0:11:20.559 --> 0:11:24.680 computers by applying metadata to information. We might take a 0:11:24.679 --> 0:11:28.959 photo of a new coffee mug and apply metadata labels 0:11:29.280 --> 0:11:32.960 to this image so that a computer can quickly reference 0:11:33.000 --> 0:11:36.200 the metadata and then pull up our new photo of 0:11:36.240 --> 0:11:38.120 a coffee mug when we ask for it. But this 0:11:38.280 --> 0:11:41.040 is not the same thing as quote unquote knowing that 0:11:41.080 --> 0:11:43.920 it's a coffee mug. That would be more like using 0:11:44.240 --> 0:11:49.240 a reference index in order to pull up the matching image. 0:11:49.520 --> 0:11:52.360 It doesn't involve the image itself, It just involves the 0:11:52.360 --> 0:11:56.480 meta data about the image. So facial and image recognition 0:11:56.720 --> 0:12:01.040 are just one of thousands of different aimplmentations that have 0:12:01.120 --> 0:12:04.160 nothing to do or very little to do with generative AI. 0:12:04.400 --> 0:12:07.120 They might ultimately have stuff to do with generative AI, 0:12:07.200 --> 0:12:09.960 but that's because of convergence. It's not that they're the 0:12:10.000 --> 0:12:13.440 same thing. It's that things these different disciplines are converging 0:12:13.600 --> 0:12:19.360 into new implementations. Alan Turing, the great computer scientist, theorized 0:12:19.400 --> 0:12:22.360 that machines might one day be able to take all 0:12:22.440 --> 0:12:27.040 available information in a given situation and apply reasoning to 0:12:27.240 --> 0:12:30.960 that situation in order to reach conclusions similar to how 0:12:31.000 --> 0:12:33.920 we humans operate, and he wrote about it in a 0:12:33.920 --> 0:12:39.400 paper titled Computing, Machinery and Intelligence. It was all hypothetical 0:12:39.440 --> 0:12:42.559 at the time, since computers were still quite primitive back then. 0:12:42.840 --> 0:12:47.920 For one thing, they lacked the capability of persistent memory. 0:12:48.320 --> 0:12:51.320 I'll explain more in just a moment, but first let's 0:12:51.360 --> 0:13:03.679 take a quick break to thank our sponsors. Okay, we're back. 0:13:03.960 --> 0:13:07.160 What was I talking about? All right? Persistence of memory? 0:13:07.440 --> 0:13:10.480 And I should get myself some of that anyway, What 0:13:10.640 --> 0:13:13.120 I meant by that with Alan Turing and the lack 0:13:13.200 --> 0:13:17.280 of persistent memory, is that the computers of Turing's day 0:13:17.559 --> 0:13:21.400 could execute a command, but they couldn't quote unquote remember 0:13:21.840 --> 0:13:24.520 what it was they just did. They would just perform 0:13:24.520 --> 0:13:27.559 an operation, and they would continue to perform that operation 0:13:27.679 --> 0:13:32.240 on new incoming data until you changed all the factors 0:13:32.240 --> 0:13:36.359 of the computer, which often involved physical switches and cables 0:13:36.440 --> 0:13:39.800 and plugs and stuff like. This was a big deal 0:13:39.880 --> 0:13:43.720 to set a computer up to run calculations, so you 0:13:43.920 --> 0:13:49.600 couldn't naturally build upon an outcome and then do a 0:13:49.640 --> 0:13:51.720 new operation. You had to do a lot of work 0:13:51.920 --> 0:13:55.480 in order to make this happen. But Turing thought, there 0:13:55.520 --> 0:13:58.040 will come a day where computers will be able to 0:13:58.080 --> 0:14:02.040 do this. They'll be able to complete the task, create 0:14:02.240 --> 0:14:05.600 an outcome, and then take that outcome and then perform 0:14:06.120 --> 0:14:10.640 new tasks upon that outcome, all with the goal of 0:14:10.720 --> 0:14:14.000 some specific outcome further down the line, like ten or 0:14:14.040 --> 0:14:19.480 twenty steps further along. So it was only a whole 0:14:19.520 --> 0:14:23.160 bunch of different smarty pants is from different disciplines who 0:14:23.240 --> 0:14:26.680 were able to advance the technology of computing that machines 0:14:26.720 --> 0:14:30.960 could actually have something that resembled memory, let alone this 0:14:31.320 --> 0:14:36.160 capability of taking in information and then being able to reason. 0:14:36.640 --> 0:14:41.480 So from transistors to integrated circuits to computer languages, et cetera, 0:14:41.560 --> 0:14:44.240 a lot of different pieces had to come together in 0:14:44.360 --> 0:14:47.800 order to even make this a possibility. So in nineteen 0:14:47.880 --> 0:14:52.160 fifty six, the Dartmouth Summer Research Project and Artificial Intelligence 0:14:52.280 --> 0:14:55.680 saw buffins from across the young discipline of computer science 0:14:55.760 --> 0:14:59.960 gathered to talk about researching concepts relating to machine intelligence. 0:15:00.200 --> 0:15:03.320 This was the conference that serves as the official birthplace 0:15:03.440 --> 0:15:07.800 for the term artificial intelligence. While there were a lot 0:15:07.880 --> 0:15:10.960 of people really excited about the idea and many people 0:15:11.040 --> 0:15:14.000 attending the conference felt pretty sure that machines would one 0:15:14.080 --> 0:15:17.480 day reach a point where they could simulate human intelligence, 0:15:17.800 --> 0:15:20.760 there was no agreement on exactly how this would happen. 0:15:21.840 --> 0:15:26.200 No one proposed any standards or anything like that, and 0:15:26.800 --> 0:15:30.360 because of that, the following decades would see various researchers 0:15:30.440 --> 0:15:34.880 pursue their own pathways toward a common goal. So everyone 0:15:35.000 --> 0:15:37.520 kind of knew where they wanted to get, but they 0:15:37.960 --> 0:15:40.520 weren't in agreement as to how they were going to 0:15:40.560 --> 0:15:42.720 get there, and so there was a lot of different 0:15:42.720 --> 0:15:47.280 work being done in different approaches toward artificial intelligence. In 0:15:47.320 --> 0:15:51.320 the nineteen sixties, a computer scientist and programmer named Joseph 0:15:51.360 --> 0:15:57.000 Weisenbaum created an early chatbot called Eliza. So this chatbot 0:15:57.080 --> 0:16:01.320 is exceedingly primitive by today's standards, and it gave the 0:16:01.360 --> 0:16:05.080 illusion of understanding communications from a human being, But in fact, 0:16:05.240 --> 0:16:09.040 Eliza was really just spouting off responses using some rudimentary 0:16:09.160 --> 0:16:13.560 pattern recognition and substitution strategies. So in a very superficial 0:16:13.920 --> 0:16:19.000 and not particularly useful way, Eliza could chat with humans. Now, 0:16:19.040 --> 0:16:21.840 I think Eliza is a really important early step in 0:16:21.920 --> 0:16:26.960 artificial intelligence. I would also say it's not very intelligent. 0:16:27.440 --> 0:16:30.120 It's following a pretty simple set of rules in an 0:16:30.160 --> 0:16:35.200 effort to simulate conversation, and limited conversation at that And 0:16:35.280 --> 0:16:38.680 while we have much more sophisticated chatbots now, ones that 0:16:38.720 --> 0:16:42.400 can draw on immense libraries of information and use complicated 0:16:42.440 --> 0:16:46.520 statistics to select words and word order, ultimately they're kind 0:16:46.520 --> 0:16:49.520 of doing the same thing. They are using rules to 0:16:49.560 --> 0:16:53.640 create an illusion of intelligence. But the projects I really 0:16:53.680 --> 0:16:56.320 want to talk about don't necessarily even do that much. 0:16:56.640 --> 0:17:00.240 They are creating the illusion of artificial intelligence because us 0:17:00.440 --> 0:17:03.920 that's a field that's getting crazy amounts of investment. Sometimes 0:17:04.320 --> 0:17:07.320 these companies are doing it because they don't yet have 0:17:07.440 --> 0:17:10.760 the money to really dive into AI. So it's not 0:17:10.800 --> 0:17:14.480 that they want to deceive, it's that until they get 0:17:14.520 --> 0:17:17.000 the investment to do the thing they want to do, 0:17:17.440 --> 0:17:22.359 they can't do it. AI is expensive. The processing you 0:17:22.480 --> 0:17:27.879 need in order to run complicated AI implementations is considerable, 0:17:28.040 --> 0:17:30.480 and most people don't have access to that, especially if 0:17:30.480 --> 0:17:34.040 you're just starting up a company. So sometimes an AI 0:17:34.200 --> 0:17:39.080 startup is not using AI, not in an effort to 0:17:39.160 --> 0:17:44.240 deceive investors, but rather as a placeholder with the intent 0:17:44.680 --> 0:17:48.199 of using AI later on when it becomes feasible to 0:17:48.240 --> 0:17:51.480 do so. Sometimes the company just needs to do a 0:17:51.520 --> 0:17:55.400 lot of early work before it can launch whatever AI 0:17:55.440 --> 0:17:58.520 tool it has in mind, and that this early work 0:17:58.640 --> 0:18:00.679 needs to be done by humans. There might be a 0:18:00.720 --> 0:18:04.440 lot of generating training data or that kind of thing, 0:18:04.680 --> 0:18:07.600 and for that you might employ a bunch of people 0:18:07.640 --> 0:18:12.720 to do it, and eventually you will develop your AI tool. 0:18:12.800 --> 0:18:15.240 But again, it does not like AI tools just spring 0:18:15.440 --> 0:18:18.520 fully formed and you can make use of them. So 0:18:19.160 --> 0:18:23.360 again there are cases where an quote unquote AI startup 0:18:23.680 --> 0:18:27.800 isn't using AI, but it's not necessarily an attempt to 0:18:27.880 --> 0:18:31.680 mislead people, but sometimes it might just be a scam. 0:18:32.000 --> 0:18:35.040 It might just be an effort to tap into people's 0:18:35.480 --> 0:18:40.560 enthusiasm and excitement around a buzzy term, but have no 0:18:40.800 --> 0:18:45.159 intent on ever doing any significant work within the AI field. 0:18:45.520 --> 0:18:47.840 And this has been going on for quite a few 0:18:47.920 --> 0:18:52.359 years now. Back in twenty eighteen, a writer named Kian 0:18:53.160 --> 0:18:57.600 jen Cheng, and I apologize for my pronunciation. I am 0:18:57.680 --> 0:19:02.840 notoriously terrible about this way Xichang wrote a piece for 0:19:03.119 --> 0:19:09.080 sixth tone dot com and it's called AI company accused 0:19:09.240 --> 0:19:13.080 of using humans to fake its AI. So the company 0:19:13.119 --> 0:19:16.919 in question was iFly Tech that's little I, big f 0:19:17.800 --> 0:19:22.200 l y te k, and among other things, I fly 0:19:22.359 --> 0:19:27.000 Tech was offering AI powered interpretation services, or at least 0:19:27.359 --> 0:19:29.879 that seems to be what the claim was. So the 0:19:29.960 --> 0:19:34.879 product was supposed to provide real time interpretation and translation 0:19:35.080 --> 0:19:39.720 services and was demonstrated at you know, things like international events. 0:19:40.119 --> 0:19:43.439 But then a man named Belle Wong came forward and 0:19:43.440 --> 0:19:46.280 claimed that he was part of a group of interpreters 0:19:46.400 --> 0:19:51.280 who did the actual work and essentially they were posing 0:19:51.480 --> 0:19:56.879 as the interpretation software. So Wang's accusations centered around a 0:19:56.920 --> 0:20:01.960 symposium called the twenty eighteen International Forum on Innovation and 0:20:02.000 --> 0:20:07.359 Emerging Industries Development Catchy. At this forum, there was a 0:20:07.400 --> 0:20:11.399 professor from Japan who gave a presentation, and his presentation 0:20:11.640 --> 0:20:15.639 was in English, and his words were being transcribed in 0:20:15.720 --> 0:20:19.080 real time by a text to speech program from iFly 0:20:19.240 --> 0:20:23.680 Tech and displayed on a screen behind him. So as 0:20:23.720 --> 0:20:27.360 he spoke, his words in English were showing up behind him, 0:20:27.600 --> 0:20:32.720 but next to the English transcription, his words appeared as 0:20:32.960 --> 0:20:39.280 a Chinese transcription written in Chinese characters, also supposedly handled 0:20:39.320 --> 0:20:43.760 by iFly Tech's incredible technology. Now this is amazing, not 0:20:43.920 --> 0:20:46.280 just because you're talking about translation. I mean we have 0:20:46.480 --> 0:20:50.440 translation apps out there right, We've got translation tools where 0:20:50.440 --> 0:20:53.240 you can speak into a app and have it generate 0:20:53.880 --> 0:20:58.399 an actual response in another language. This was incredible because 0:20:58.440 --> 0:21:03.719 it's not just translation but in interpretation, meaning that the 0:21:03.760 --> 0:21:08.119 turns of phrase that the speaker used were being interpreted 0:21:08.520 --> 0:21:12.879 and then translated into Chinese so that the Chinese translation 0:21:13.480 --> 0:21:17.719 would make sense. Because obviously, like the sayings, the idioms 0:21:17.760 --> 0:21:20.879 that we use in one language do not necessarily translate 0:21:20.920 --> 0:21:24.160 to another. If I say it's raining cats and dogs, 0:21:24.520 --> 0:21:27.399 English speakers know what I mean is that it's raining 0:21:27.480 --> 0:21:31.359 really hard. Non English speakers, if they saw that translation, 0:21:31.480 --> 0:21:35.200 would wonder why are animals falling from the sky when 0:21:35.800 --> 0:21:38.199 that's not what I literally mean when I say it's 0:21:38.280 --> 0:21:42.040 raining cats and dogs. So interpretation requires an extra step. 0:21:42.400 --> 0:21:46.159 It's not just translating word for word, and in fact, 0:21:46.280 --> 0:21:49.719 what was appearing behind the speaker was an interpretation of 0:21:49.760 --> 0:21:53.840 the speaker's words. The problem was Belle Wong says it 0:21:53.880 --> 0:21:57.360 was really him and his colleagues who were doing the 0:21:57.400 --> 0:22:01.439 work of that interpretation and translation. So Wong pointed to 0:22:01.480 --> 0:22:05.600 the fact that the professor's accent was fairly strong. He 0:22:05.640 --> 0:22:09.280 had a Japanese accent as he was giving his English presentation, 0:22:09.680 --> 0:22:13.480 and the real time English text to speech program from 0:22:13.520 --> 0:22:17.560 iFly Tech got into some issues with this. The program 0:22:17.640 --> 0:22:21.560 would sometimes misinterpret what the professor was saying, and so 0:22:21.680 --> 0:22:24.879 the transcript had errors in it. If you were reading 0:22:24.920 --> 0:22:27.520 along while the professor was speaking, you would see, oh, 0:22:27.640 --> 0:22:30.359 the program thought he said this one thing, but in 0:22:30.400 --> 0:22:33.400 fact he was saying this other thing. But the Chinese 0:22:33.440 --> 0:22:37.679 interpretation of the professor's words did not include these mistakes. 0:22:38.040 --> 0:22:42.399 The Chinese translation was accurate, and that's because Wang and 0:22:42.440 --> 0:22:46.359 his colleagues were translating accurately. They were listening to what 0:22:46.520 --> 0:22:50.000 he was saying, interpreting it, translating it, and then putting 0:22:50.040 --> 0:22:54.440 it up in Chinese text, so they had the appropriate 0:22:55.480 --> 0:23:01.600 Chinese interpretations displaying not the mistaken text to speech stuff. 0:23:01.960 --> 0:23:04.840 Wong said that I fly Tech never really acknowledged the 0:23:04.960 --> 0:23:08.320 use of human interpreters at that event, and that the 0:23:08.359 --> 0:23:12.480 implication was the technology was doing all the heavy lifting. 0:23:12.800 --> 0:23:16.359 So Wang said this made him feel very uncomfortable to 0:23:16.400 --> 0:23:19.080 be part of what he felt was a deceptive presentation. 0:23:19.480 --> 0:23:23.679 And it's interesting because in that same piece in sixth tone, 0:23:24.119 --> 0:23:27.480 the piece quotes I fly Tech executives who have essentially 0:23:27.520 --> 0:23:30.840 said the machines are not a suitable replacement for human 0:23:30.880 --> 0:23:34.560 interpreters and that it's far more likely that the future 0:23:34.680 --> 0:23:39.400 of interpreting will involve humans and machines working together, rather 0:23:39.480 --> 0:23:43.080 the machines replacing humans. Outright now, perhaps the ie fly 0:23:43.200 --> 0:23:46.800 Tech representatives at the International Forum were a bit over 0:23:46.920 --> 0:23:49.840 zealous in promoting the work of their company. But it 0:23:49.880 --> 0:23:53.440 feels a lot like the mechanical turk. You know, at 0:23:53.440 --> 0:23:56.440 a casual glance, you have a machine that's doing this 0:23:56.920 --> 0:24:00.800 incredibly complex action but if you take a closer look, 0:24:00.880 --> 0:24:04.399 you see that humans are powering the real process behind 0:24:04.440 --> 0:24:09.320 the scenes. Then there's Olivia Salon's twenty eighteen piece in 0:24:09.359 --> 0:24:13.680 The Guardian titled the Rise of Pseudo AI and how 0:24:13.720 --> 0:24:17.240 tech firms quietly use humans to do bots work. Now. 0:24:17.280 --> 0:24:21.320 I love how Salon frames her piece by saying, quote, 0:24:21.560 --> 0:24:24.960 some startups have worked out it's cheaper and easier to 0:24:24.960 --> 0:24:27.960 get humans to behave like robots than it is to 0:24:28.000 --> 0:24:32.280 get machines to behave like humans. End quote that I 0:24:32.440 --> 0:24:35.199 feel is bang on the money. She did a great 0:24:35.400 --> 0:24:39.399 job with this article. We humans are really versatile. We 0:24:39.520 --> 0:24:42.439 have evolved to be that way like It's not that 0:24:42.480 --> 0:24:46.919 we're special. We have millions of years of evolution behind 0:24:47.000 --> 0:24:49.679 us that have shaped us to be like this. But 0:24:50.440 --> 0:24:53.800 we have to put that same work into machines in 0:24:53.880 --> 0:24:56.760 order to make machines perform in versatile ways, and that 0:24:56.880 --> 0:25:00.160 is a considerable amount of work. We haven't been working 0:25:00.200 --> 0:25:03.200 with computers for millions of years. We've only been doing 0:25:03.200 --> 0:25:06.080 it for a few decades. So companies like open Ai 0:25:06.240 --> 0:25:09.119 and Google and such are spending billions of dollars to 0:25:09.200 --> 0:25:12.480 achieve that goal. It is not at all easy and 0:25:12.520 --> 0:25:16.680 it doesn't always go smoothly. So some startups use humans 0:25:16.720 --> 0:25:18.880 in the early days almost as a way to show 0:25:18.960 --> 0:25:22.280 the proof of concept for their end product. So sure, 0:25:22.440 --> 0:25:25.080 right now, humans are the ones doing whatever it is, 0:25:25.160 --> 0:25:29.280 like the coding or the translating or whatever the startup 0:25:29.359 --> 0:25:32.600 AI is focused on. But further down the line, well 0:25:32.600 --> 0:25:35.439 that's going to be bots. Maybe in fact, it'll have 0:25:35.520 --> 0:25:38.320 to be bots because if the startup were to take 0:25:38.320 --> 0:25:41.560 off and become a big company, then it could become 0:25:41.640 --> 0:25:44.280 too expensive to rely on humans to do all the 0:25:44.320 --> 0:25:47.680 work that needs to be done when you're operating at scale. 0:25:48.080 --> 0:25:51.000 So there's a danger of doing this as a startup, 0:25:51.080 --> 0:25:53.560 right Like, if you're doing it early on, you're saying, 0:25:53.840 --> 0:25:56.520 I'm going to be transparent with you. Right now, we 0:25:56.600 --> 0:25:59.560 have human beings doing this work, but what we're working 0:25:59.600 --> 0:26:03.600 on is developing AI to do the work instead. And 0:26:04.359 --> 0:26:08.280 this is how we're presenting it to you, and we 0:26:08.320 --> 0:26:12.560 want you to be aware of our goals and our strategy. 0:26:13.000 --> 0:26:15.440 If it turns out that whatever they want to do 0:26:15.640 --> 0:26:18.080 is too hard to do by AI, like it's just 0:26:18.119 --> 0:26:21.359