WEBVTT - From Unsung Science with David Pogue: The Man Who Stopped the Spammers 0:00:15.316 --> 0:00:20.836 Pushkin. Hi, It's Jacob Goldstein and I'm here today with 0:00:20.916 --> 0:00:23.596 another podcast I think you might like. The show is 0:00:23.636 --> 0:00:27.636 called Unsung Science and it's hosted by David Pogue. You 0:00:27.716 --> 0:00:31.036 might know David from CBS Sunday Morning, where he's a 0:00:31.116 --> 0:00:35.276 correspondent covering topics like science, tech, and innovation, topics like 0:00:35.316 --> 0:00:37.716 the ones we talk about here on What's Your Problem. 0:00:37.796 --> 0:00:40.196 In the episode You're about to hear, David chats with 0:00:40.356 --> 0:00:44.036 Luis Vonon, the founder and CEO of the popular language 0:00:44.036 --> 0:00:47.036 app Duolingo. You might recall I talked with Louise earlier 0:00:47.036 --> 0:00:50.676 this year about Duolingo and language and the current limits 0:00:50.676 --> 0:00:53.636 of artificial intelligence. But the show you're about to hear 0:00:53.796 --> 0:00:57.276 is about what Louise did before he started Duolingo. He 0:00:57.316 --> 0:01:00.436 invented this thing called Capsha. Capsha is that test that 0:01:00.476 --> 0:01:02.396 you have to take all the time on the Internet 0:01:02.436 --> 0:01:05.556 to prove that you're not a robot. And yes, Louise 0:01:05.676 --> 0:01:08.196 knows that the test is super annoying. But the story 0:01:08.236 --> 0:01:11.156 of capsha and what happened with it is really interesting. 0:01:11.916 --> 0:01:17.916 It's got some great twists. By the year two thousand, 0:01:18.116 --> 0:01:22.556 the Internet was already becoming a cesspool software bots were 0:01:22.596 --> 0:01:26.156 signing up for millions of fake email accounts for sending 0:01:26.156 --> 0:01:31.436 out spam. Luis Vaughan stopped them. He invented the CAPTA, 0:01:31.956 --> 0:01:35.436 the website login test where you have to decipher the 0:01:35.476 --> 0:01:38.436 distorted image of a word, where you have to find 0:01:38.556 --> 0:01:41.556 the traffic lights in a grid of nine blurry photos. 0:01:42.156 --> 0:01:45.356 The only problem we hate that test. I would be 0:01:45.396 --> 0:01:47.036 at a party and you know, people would ask me 0:01:47.036 --> 0:01:48.316 what I did, and I would tell them that I 0:01:48.316 --> 0:01:50.076 helped invent that thing, and people would tell me, oh, 0:01:50.116 --> 0:01:54.076 I hate you. I'm David Pogue And this is Unsung Science, 0:02:01.476 --> 0:02:06.356 Season one, episode fourteen, The man who stopped the spammers. 0:02:08.236 --> 0:02:10.676 In his forty three years on this earth, so far, 0:02:11.316 --> 0:02:17.516 Luis vonn has had three ingenious innovative world changing ideas. 0:02:18.796 --> 0:02:22.996 I guarantee that you've encountered his second one, probably hundreds 0:02:22.996 --> 0:02:27.196 of times. Actually, most of us have zero world changing ideas. 0:02:27.596 --> 0:02:33.996 Occasionally somebody has won, but three times. His first idea 0:02:34.076 --> 0:02:37.836 came to him in Guatemala, where he grew up. They 0:02:37.916 --> 0:02:41.516 wanted to start a gym where instead of charging people 0:02:41.876 --> 0:02:44.116 to show up, let people just show up for free. 0:02:44.316 --> 0:02:46.716 We're going to connect all the machines to kind of 0:02:46.756 --> 0:02:48.916 the power grid, and we're going to use the kinetic 0:02:48.996 --> 0:02:53.076 energy that people had whenever they were exercising to generate power. 0:02:53.716 --> 0:02:55.276 And I thought we could make a lot of money 0:02:55.276 --> 0:02:57.356 from that. Now you will note that I did not 0:02:57.476 --> 0:03:00.956 say that all three of his world changing ideas actually 0:03:00.996 --> 0:03:03.836 succeeded in changing the world. I thought I was the 0:03:03.836 --> 0:03:05.396 first person to have this idea. It turns out it's 0:03:05.396 --> 0:03:07.716 a very old idea. It also turns out it doesn't work. 0:03:07.836 --> 0:03:11.756 That's right, the pedal power Jim idea flopped. It turns 0:03:11.756 --> 0:03:13.836 out this is not a good idea for many reasons, 0:03:13.836 --> 0:03:15.436 the biggest one of which is that humans are just 0:03:15.516 --> 0:03:18.316 not very good at creating energy. Oh, you just just 0:03:18.356 --> 0:03:20.116 don't make a lot of money from this. There's another 0:03:20.156 --> 0:03:22.396 reason why this doesn't work a lot. It turns out 0:03:22.476 --> 0:03:24.476 Jim's make most of their money from people who don't 0:03:24.476 --> 0:03:26.956 show up. Of course, here you kind of need people 0:03:26.996 --> 0:03:30.596 to show up to be fair. He was pretty new 0:03:30.636 --> 0:03:33.676 at the game when he had this first idea. And 0:03:33.796 --> 0:03:36.196 how old were you at this point, twelve years old, 0:03:36.196 --> 0:03:39.396 eleven years old. Things started going better six years later, 0:03:39.676 --> 0:03:42.996 when he came to the United States to attend Duke University. 0:03:43.636 --> 0:03:46.716 As the year two thousand dawned. Luis was at Carnegie 0:03:46.756 --> 0:03:49.476 Mellon in his first year of working toward a PhD 0:03:49.796 --> 0:03:53.916 in computer science, and one fateful day he went to 0:03:53.996 --> 0:03:58.236 a talk by an Israeli computer scientist named Udi Manber, 0:03:58.676 --> 0:04:01.916 who at this point was the chief scientist at Yahoo. 0:04:02.076 --> 0:04:03.756 By the way, at end the year two thousand, Yahoo 0:04:03.836 --> 0:04:07.076 was the biggest biggest tech company in the words like 0:04:07.076 --> 0:04:09.996 the Google of today. And you know, he was giving 0:04:10.276 --> 0:04:12.756 a talk about ten problems that they didn't know how 0:04:12.756 --> 0:04:15.916 to solve inside inside the company. And one of those 0:04:16.156 --> 0:04:18.996 ten problems that the greatest minds at Yahoo could not 0:04:19.116 --> 0:04:24.356 solve was automated software spam bots signing up for free 0:04:24.436 --> 0:04:28.916 Yahoo mail accounts. By the millions. Yahoo gave up free 0:04:28.916 --> 0:04:30.796 email accounts, and there were people who wanted to send 0:04:30.836 --> 0:04:34.636 spam from Yahoo accounts. But each Yahoo account only allowed 0:04:34.636 --> 0:04:37.316 you to send like five hundred messages a day. If 0:04:37.316 --> 0:04:39.796 you wanted to send millions of emails spam emails per day, 0:04:40.636 --> 0:04:42.996 then what these people did is they wrote programs to 0:04:43.036 --> 0:04:46.636 obtain millions of Yahoo accounts every day, and they didn't 0:04:46.676 --> 0:04:47.996 know how to solve that problem, how to stop that. 0:04:48.436 --> 0:04:51.116 So I started talking about it with a person who 0:04:51.236 --> 0:04:54.276 had just become my PhD advisor. His name was Manuel 0:04:54.316 --> 0:04:57.516 Blum or is Manolum. He's still he's most definitely still alive. 0:04:57.836 --> 0:05:00.196 And you know, we started thinking, and this is where 0:05:00.196 --> 0:05:02.756 this idea of a capture came up. The idea was this, 0:05:03.556 --> 0:05:06.556 anytime you tried to sign up for a Yahoo Mail account, 0:05:06.956 --> 0:05:10.596 you'd encounter a little puzzle, something easy for a person 0:05:10.636 --> 0:05:13.796 to solve, but hard for a spambot. The way to 0:05:13.876 --> 0:05:16.676 stop these spammers was to have a test that can 0:05:16.756 --> 0:05:20.036 distinguish between whether you're a human or a computer. If 0:05:20.116 --> 0:05:23.196 you are a human, then presumably you can't get millions 0:05:23.236 --> 0:05:26.276 of email accounts because you get bored, whereas if you're 0:05:26.276 --> 0:05:28.076 a computer, you can get millions. So if the only 0:05:28.276 --> 0:05:30.796 entity is that we were giving email accounts to where humans, 0:05:31.036 --> 0:05:34.436 then that would stop the spam. KAPTA, the name he 0:05:34.476 --> 0:05:38.556 gave his online mini puzzle, is an acronym. It stands 0:05:38.596 --> 0:05:44.636 for completely automated public touring tests to tell computers and 0:05:44.836 --> 0:05:49.476 humans apart more or less. Not sure if you've heard 0:05:49.516 --> 0:05:52.436 of the touring test, but it is incredibly famous among 0:05:52.476 --> 0:05:57.196 computer scientists. It's this experiment proposed by British mathematician and 0:05:57.236 --> 0:06:01.036 computer scientist Alan Turing, who's known as the father of 0:06:01.156 --> 0:06:04.836 artificial intelligence. There was actually a movie about him called 0:06:04.836 --> 0:06:09.796 The Imitation Game, where Benedict's Cumberbatch played Alan Tour. Would 0:06:09.836 --> 0:06:14.996 you like to play play? It's a game, a test 0:06:15.036 --> 0:06:20.636 of songs for determining whether something is machine or a 0:06:20.756 --> 0:06:24.796 human being. Anyway, the Touring test is intended to set 0:06:24.796 --> 0:06:29.156 a standard for determining if a computer has achieved true 0:06:29.476 --> 0:06:33.436 artificial intelligence. When can we tell that a computer is 0:06:33.436 --> 0:06:35.996 actually intelligent. This is kind of like a philosophical test 0:06:35.996 --> 0:06:38.756 that said, like, look, we're going to have a human 0:06:38.876 --> 0:06:42.316 judge ask questions to two entities. One is the computer, 0:06:42.396 --> 0:06:44.916 one is the human. The computer and the human are 0:06:44.996 --> 0:06:48.876 hidden behind two curtains. The judge can't see them. The 0:06:48.996 --> 0:06:51.876 judge types in questions and then looks at the text 0:06:51.916 --> 0:06:55.956 of the responses. If it's impossible to tell which answer 0:06:55.996 --> 0:06:58.676 came from the person in which from the computer, the 0:06:58.676 --> 0:07:02.316 computer has passed the Turing test. The judge can just 0:07:02.356 --> 0:07:04.676 ask whatever questions they want, and if we really can't 0:07:04.676 --> 0:07:07.236 distinguish them, we'll say the computer is really intelligent. To 0:07:07.356 --> 0:07:09.716 this day, we have not made a computer that can 0:07:09.716 --> 0:07:13.276 actually pass the turning test successfully. It's just it's just 0:07:13.316 --> 0:07:15.756 too hard. The funny thing is, if you really think 0:07:15.756 --> 0:07:19.476 about it, the capture problem is the opposite of the 0:07:19.516 --> 0:07:23.196 touring test. The touring test is successful if the judge 0:07:23.516 --> 0:07:26.476 can't tell the difference between a person and a machine. 0:07:27.036 --> 0:07:29.796 The whole point of Louis Vanan's project was to create 0:07:29.836 --> 0:07:33.596 a test that can tell the difference. There's another difference 0:07:33.596 --> 0:07:36.596 between the two tests too. Here's the key. In this case, 0:07:36.636 --> 0:07:40.116 the judge was a human. In our case for the capture, 0:07:40.236 --> 0:07:41.716 what we needed to do is we needed the judge 0:07:41.756 --> 0:07:44.276 to be a computer because we need we need the 0:07:44.276 --> 0:07:46.636 computer to determine whether it's talking to a human our computer, 0:07:46.796 --> 0:07:49.436 which is which is much harder in some sense, at 0:07:49.476 --> 0:07:51.316 least for to grade it. So I think the hardest 0:07:51.316 --> 0:07:53.916 thing was just coming up with this general idea that like, okay, 0:07:54.036 --> 0:07:56.036 what we need is a test that can assume shroom 0:07:56.116 --> 0:07:59.796 some computers, but that computers need to be able to grade. 0:07:59.916 --> 0:08:02.116 Then after that we started coming up with like, okay, 0:08:02.116 --> 0:08:03.756 what I think the computers are not very good at. 0:08:04.316 --> 0:08:07.996 In the year two thousand, the answer was obvious, computers 0:08:08.036 --> 0:08:12.636 are not very good at identifying what's in pictures. We 0:08:12.876 --> 0:08:15.756 quickly owned in on images and just doing you know, 0:08:15.916 --> 0:08:19.316 images of text, images of flowers, images of stuff. And 0:08:19.356 --> 0:08:21.716 then after a while, the images of text were the 0:08:21.716 --> 0:08:23.556 ones that seemed like the best idea. And then I 0:08:23.596 --> 0:08:27.276 just went and developed a program that distorted random text 0:08:27.916 --> 0:08:29.916 and that was the first version of a cap chow. 0:08:30.396 --> 0:08:33.236 That's right. The test they came up with presents you 0:08:33.236 --> 0:08:36.396 with the image of a typed word, but the letters 0:08:36.396 --> 0:08:40.516 are all like twisted, bent and distorted, as though the 0:08:40.556 --> 0:08:44.716 typist were severely drunk and typing on saran wrap. You 0:08:44.756 --> 0:08:47.996 are supposed to interpret what that word is and type 0:08:47.996 --> 0:08:52.476 it into a box on the website. Actually, computers in 0:08:52.516 --> 0:08:56.076 the early two thousands were pretty good at OCR. That's 0:08:56.196 --> 0:09:00.596 optical character recognition, meaning looking at a picture of text 0:09:00.836 --> 0:09:03.556 and figuring out what the letters are. But the added 0:09:03.636 --> 0:09:07.796 challenge of the twisty distortion really threw those OCR programs 0:09:07.876 --> 0:09:11.396 off the track. Behind the scenes, I mean, what is it. 0:09:11.436 --> 0:09:13.596 I mean there's got to be some I don't know, 0:09:13.636 --> 0:09:18.236 sequel database or massive bank of little images. I mean, 0:09:18.396 --> 0:09:20.916 actually there was no database at first. We would just 0:09:21.276 --> 0:09:23.116 write a program that what it would do is it 0:09:23.116 --> 0:09:26.436 would pick some first random characters, would put them on 0:09:26.436 --> 0:09:28.316 an image, then it would distore him, and then we 0:09:28.356 --> 0:09:31.236 would save that image. And then we just had I 0:09:31.236 --> 0:09:32.956 don't know, a couple of million of those saved, not 0:09:32.996 --> 0:09:34.676 even in a sequel database. Is just they were there, 0:09:34.716 --> 0:09:39.596 so save as files. It worked brilliantly. The spambots didn't 0:09:39.636 --> 0:09:43.676 have a chance. At the time. Vonn had no idea 0:09:43.796 --> 0:09:46.836 if his invention would be of any commercial use. But 0:09:46.956 --> 0:09:50.916 one guy he knew would be interested, Oodi Manber, that 0:09:51.156 --> 0:09:54.396 Yahoo chief scientist who'd given the talk that started this 0:09:54.436 --> 0:09:57.116 whole affair. We sent them an email saying, hey, we 0:09:57.196 --> 0:10:00.156 think we can solve your problem, and he said, oh, 0:10:00.196 --> 0:10:01.996 that that seems like it solves the problem. And then 0:10:02.036 --> 0:10:03.956 in fact, pretty soon after that it was being used 0:10:03.956 --> 0:10:06.836 by Yahoo, and then basically every website started using it, 0:10:07.676 --> 0:10:11.476 and you know, there was millions of websites out there. 0:10:11.276 --> 0:10:18.516 We're using it. Well, how wonderful Luis Spawn's ingenuity. One spammers, 0:10:18.636 --> 0:10:23.436 zero Internet saved. And at first I was very kind 0:10:23.476 --> 0:10:25.636 of proud of myself because, okay, look at the impact 0:10:25.676 --> 0:10:28.716 that my work has had. Basically we stopped spams being 0:10:28.796 --> 0:10:31.796 used by a lot of people. There was only one problem. Now, 0:10:32.956 --> 0:10:36.596 people hated his invention. How many of you have had 0:10:36.636 --> 0:10:38.156 to fill out some sort of web form where even 0:10:38.156 --> 0:10:40.756 has to read a distorted sequence of characters like this, Yeah, 0:10:40.836 --> 0:10:43.876 how many of you found it really really annoying. Okay, 0:10:43.916 --> 0:10:48.876 that's standing. So I invented that. That's how he introduces 0:10:48.956 --> 0:10:52.196 himself in a twenty eleven TEDx talk at Carnegie Mellon. 0:10:52.476 --> 0:10:54.036 I would be at a party, and you know, people 0:10:54.036 --> 0:10:55.436 would ask me what I did, and I would tell 0:10:55.476 --> 0:10:57.236 them that I helped invent that thing. And people would 0:10:57.236 --> 0:10:59.796 tell me, oh, I hate you. That's right. The inventor 0:10:59.836 --> 0:11:03.316 of kapta is fully aware that people hate the thing. 0:11:03.836 --> 0:11:08.716 I say, either well, I'm sorry, or I find it 0:11:08.716 --> 0:11:12.156 annoying too. You've heard it right here, folks. Even he 0:11:12.316 --> 0:11:15.956 finds them annoying. In fact, Louise can tell you exactly 0:11:16.036 --> 0:11:18.916 how much of your time they waste. I did a 0:11:18.956 --> 0:11:21.356 little back of the envelope calculation at the time, about 0:11:21.356 --> 0:11:23.836 two hundred million times a day somebody type one of 0:11:23.836 --> 0:11:28.316 these captures two hundred million times times ten seconds, which 0:11:28.396 --> 0:11:30.076 is how long it takes to type one of these. 0:11:30.556 --> 0:11:32.876 Humanity as a whole was wasting about five hundred thousand 0:11:32.916 --> 0:11:36.116 hours every day typing these annoying captures. Great. So I 0:11:36.116 --> 0:11:39.436 started feeling bad about that, and that's when I started thinking, Okay, 0:11:39.436 --> 0:11:41.636 can we do something good with that time? See, the 0:11:41.676 --> 0:11:44.636 thing is kind of similar to the gym idea. Can 0:11:44.676 --> 0:11:48.116 we get millions of people to do something during that 0:11:48.156 --> 0:11:51.316 time that is actually valuable. I'll give you a hint. 0:11:51.716 --> 0:11:54.956 We're only at the halfway point in this story. After 0:11:54.996 --> 0:11:56.876 the break, we'll tell you what he came up with 0:11:57.076 --> 0:12:00.676 to make those half a million hours every day useful 0:12:00.716 --> 0:12:11.796 to humanity. And one more plug here. I'm the author 0:12:11.836 --> 0:12:14.316 of a book called How to Prepare for Climate Change. 0:12:14.596 --> 0:12:17.716 It's a six hundred page paperback that's designed to be 0:12:17.756 --> 0:12:20.556 a field guide to the new climate. It tells you 0:12:20.596 --> 0:12:23.636 where to live, where to invest, what to grow, how 0:12:23.636 --> 0:12:26.556 to reinforce your home, how to insure, how to talk 0:12:26.596 --> 0:12:30.716 to your kids, and how to ride out wildfires, hurricanes, heatwaves, 0:12:30.716 --> 0:12:33.356 and so on. If you live in a state whose 0:12:33.436 --> 0:12:36.996 name contains a vowel, then you've been affected by climate 0:12:37.076 --> 0:12:39.756 change already, and you should check out this book to 0:12:39.796 --> 0:12:42.996 protect your health, your family, your home, and your finances. 0:12:43.316 --> 0:12:46.556 It's How to Prepare for Climate Change. The book that's 0:12:46.636 --> 0:12:54.036 exactly what it sounds like. Welcome back. By two thousand 0:12:54.076 --> 0:12:57.916 and five, Louis vaughan An's invention the captcha test was 0:12:57.956 --> 0:13:02.436 a huge hit. It reduced the world scumbag spammers to 0:13:02.556 --> 0:13:06.716 blubbering losers. No longer could they bombard websites with phony 0:13:06.796 --> 0:13:13.196 sign ups for the purpose of pursuing their pathetic spanny schemes. Unfortunately, 0:13:13.436 --> 0:13:18.316 he had achieved this success by transferring the burden onto us, 0:13:18.356 --> 0:13:21.156 treating us as though we were guilty until proven innocent. 0:13:21.876 --> 0:13:25.236 Now we were the ones being challenged. We were losing 0:13:25.436 --> 0:13:30.556 ten seconds per website typing in those stupid distorted letters. Now. 0:13:30.716 --> 0:13:33.636 To be fair, history is full of examples like that, 0:13:34.036 --> 0:13:37.636 where the actions of a few selfish, greedy idiots wind 0:13:37.676 --> 0:13:41.196 up inconveniencing billions of innocent people for the rest of 0:13:41.196 --> 0:13:44.796 our lives. You know, some dirtbag tries to put poison 0:13:44.876 --> 0:13:47.476 into drug store tile on all bottles, and now the 0:13:47.476 --> 0:13:51.076 rest of us are stuck with frustrating, plastic, wasteful bottle 0:13:51.116 --> 0:13:54.316 lids forever. Some delinquent tries to blow up a plane 0:13:54.316 --> 0:13:56.436 with a shoe bomb, and now we all have to 0:13:56.476 --> 0:14:00.076 walk through the TSA scanners in our socks. Louise felt 0:14:00.116 --> 0:14:04.116 bad that his hacker blockade wasted everybody's time, but at 0:14:04.196 --> 0:14:06.836 least he could do something about it. So that's a 0:14:06.916 --> 0:14:09.076 very valuable time, So can we use it for something? 0:14:09.236 --> 0:14:11.396 And then I ended up coming up with this idea 0:14:11.596 --> 0:14:14.356 that while you were typing a capture, you could be 0:14:14.396 --> 0:14:18.916 helping digitize books. And here's here's kind of how that works. 0:14:19.036 --> 0:14:20.756 So at the time, this is the year maybe two 0:14:20.756 --> 0:14:22.316 thousand and five, two thousand and six, there were a 0:14:22.356 --> 0:14:24.116 lot of projects trying to digitize all of the world's 0:14:24.156 --> 0:14:26.196 books where where you know. The way that worked is 0:14:26.356 --> 0:14:28.156 you start with a physical book and you want to 0:14:28.156 --> 0:14:30.156 put it on the internet. And the way you do 0:14:30.196 --> 0:14:33.476 that is you basically take a digital photograph of every 0:14:33.476 --> 0:14:36.396 page of the book. Now these are pictures of text. 0:14:37.076 --> 0:14:38.836 The next step in the process is that the computer 0:14:38.916 --> 0:14:41.716 needs to decipher what's the text in there. In other words, 0:14:42.036 --> 0:14:45.236 computers had to perform come on, you know, this term 0:14:45.596 --> 0:14:52.836 ocr optical character recognition, and unfortunately, for books that are 0:14:52.836 --> 0:14:55.916 older where maybe the ink has faded, computers could not 0:14:55.916 --> 0:14:59.436 recognize many of the words. So the thought, the idea was, 0:14:59.516 --> 0:15:01.476 let's take all those words that the computers could not 0:15:01.516 --> 0:15:04.356 recognize while books are being digitized, and let's get people 0:15:04.396 --> 0:15:06.516 to read them for us while they're typing a capture. 0:15:06.596 --> 0:15:08.796 So what we started giving people where these words that 0:15:08.796 --> 0:15:11.996 they con computer was not able to digitize and or 0:15:12.116 --> 0:15:15.876 to recognize. So yeah, all this time you thought you 0:15:15.916 --> 0:15:20.156 were typing random words. In fact, you were helping companies 0:15:20.356 --> 0:15:24.436 digitize old books and articles and, by the way, helping 0:15:24.516 --> 0:15:27.876 Luise's little company make money. The ideas we made a capture, 0:15:28.076 --> 0:15:30.596 a system, a whole system that would help your website 0:15:30.716 --> 0:15:33.316 be protected against BAM, and we gave that away for free. 0:15:33.356 --> 0:15:36.716 And for example, Facebook use our capture and we gave 0:15:36.716 --> 0:15:39.236 it away for free, etc. But always with a caveat 0:15:39.276 --> 0:15:42.316 that if they are going to do that, then we 0:15:42.356 --> 0:15:44.956 can see the answers that users are typing, so that 0:15:44.956 --> 0:15:47.996 we helped digitize something. And the way we made money 0:15:48.116 --> 0:15:52.356 is by charging people who needed digitization stuff. For example, 0:15:52.476 --> 0:15:54.996 the New York Times was our client. The New York 0:15:54.996 --> 0:15:57.636 Times had this old archive of all the editions of 0:15:57.676 --> 0:15:59.156 the New York Times from you know, one hundred and 0:15:59.236 --> 0:16:00.756 thirty years of the New York Times or something like that, 0:16:00.796 --> 0:16:05.116 from the eighteen hundreds, and they needed this to help 0:16:05.156 --> 0:16:07.316 digitize their whole archive. They were sending us all the 0:16:07.396 --> 0:16:10.356 scans they had scanded already and we were sending them. 0:16:10.476 --> 0:16:12.636 We were taking all the words that computer could not recognize, 0:16:12.716 --> 0:16:15.156 and we were getting through the captures people who were, 0:16:15.156 --> 0:16:18.036 for example, signing up for Facebook or Twitter or a 0:16:18.036 --> 0:16:19.676 lot of websites that we're using our capture. They were 0:16:19.676 --> 0:16:21.516 helping us digitize the New York Times, and we would 0:16:21.516 --> 0:16:23.996 make money from The New York Times. It became very successful, 0:16:24.276 --> 0:16:26.956 and then Google bought it to help their book digitization 0:16:27.116 --> 0:16:31.276 whole project. The new system called recapture became an even 0:16:31.356 --> 0:16:34.396 bigger hit. Here's how we described the aftermath in his 0:16:34.516 --> 0:16:37.276 TEDx talk. So every time you buy tickets on Ticketmaster, 0:16:37.396 --> 0:16:39.796 you hope to digitize a book. Facebook, every time you 0:16:39.796 --> 0:16:41.876 add a friend, you help to digitize a book. Twitter, 0:16:42.236 --> 0:16:44.396 and about three hundred and fifty thousand other sites are 0:16:44.436 --> 0:16:46.516 all using recapture. And in fact, the number of sites 0:16:46.556 --> 0:16:48.436 that are using recaptures so high that the number of 0:16:48.516 --> 0:16:51.196 words that we're digitizing per day is really really large. 0:16:51.276 --> 0:16:53.116 It's about one hundred million a day, which is the 0:16:53.156 --> 0:16:56.396 equivalent of about two and a half million books a year. 0:16:56.796 --> 0:16:58.476 And this is all being done one word at a 0:16:58.516 --> 0:17:05.076 time by just people tapping captures on the Internet. There 0:17:05.116 --> 0:17:07.276 are some people who are a little nervous about Google 0:17:07.956 --> 0:17:11.916 being the owner of one of the most widely used 0:17:12.556 --> 0:17:15.996 captive systems. I'm sure you've then asked about that. Yeah, 0:17:15.996 --> 0:17:17.436 there are people who are nervous about that. I mean, 0:17:17.476 --> 0:17:20.596 I understand, I think you know this is these are 0:17:20.676 --> 0:17:25.316 very very tricky questions. I mean, personally, I think the 0:17:25.396 --> 0:17:28.916 privacy fight US is over. I mean I I've given 0:17:28.996 --> 0:17:31.676 up on my privacy against large companies a while ago. Wow. 0:17:31.996 --> 0:17:35.236 Not only that, I also think after having been inside Google, 0:17:35.276 --> 0:17:37.796 I saw with how much respect they treat user data 0:17:37.836 --> 0:17:40.956 because they know that they are, you know, a few 0:17:40.996 --> 0:17:45.036 scandals away from being in deep trouble, so they take 0:17:45.076 --> 0:17:47.716 it with a lot of care, I think. And we 0:17:47.716 --> 0:17:50.796 should point out that Google has said we do not 0:17:50.956 --> 0:17:54.876 use data collected for advertising purposes. Yeah, that's the case, 0:17:54.916 --> 0:17:57.716 and so and I actually believe them. Now. Remember Louise 0:17:57.756 --> 0:18:00.116 said that the hard part was finding a test that 0:18:00.276 --> 0:18:03.276 was too hard for a computer to pass, but easy 0:18:03.356 --> 0:18:06.996 enough for a computer to judge whether the test had 0:18:07.036 --> 0:18:11.956 been passed. That's been bugging me. If the computer chooses 0:18:12.156 --> 0:18:15.156 a word that's so distorted that it itself cannot do 0:18:15.196 --> 0:18:19.196 the ocr then how does it know if we're right. Yeah, 0:18:19.196 --> 0:18:21.676 that's a great question. When we try to digitize books, 0:18:22.076 --> 0:18:24.476 Here's here's what we do. We take a word that 0:18:24.476 --> 0:18:28.156 the computer does not know. We actually pair it with 0:18:28.196 --> 0:18:30.236 another word for which the computer does know the answer, 0:18:30.436 --> 0:18:32.596 and we actually give people both words, and we say 0:18:33.196 --> 0:18:35.116 please type both, and we don't tell them which ones which, 0:18:35.116 --> 0:18:37.116 We just say, hey, please type both. If they type 0:18:37.116 --> 0:18:39.396 the word for which we know the answer, if they 0:18:39.396 --> 0:18:42.516 type that one correctly, we assume that they're human, and 0:18:42.556 --> 0:18:44.636 we also get some confidence that they type the other 0:18:44.676 --> 0:18:47.476 word correctly, and then what we do is okay, so 0:18:47.556 --> 0:18:49.316 now we have a guess for what that other word is. 0:18:49.436 --> 0:18:51.956 We give it to like ten other different people and 0:18:51.996 --> 0:18:53.916 we see if they type the same thing, and if 0:18:53.916 --> 0:18:55.676 they all type the same thing, we get with very 0:18:55.716 --> 0:18:59.756 high accuracy what that word really is, and that works. 0:19:00.396 --> 0:19:03.516 One hallmark of the recapture system in other words, is 0:19:03.556 --> 0:19:06.996 that you have to type in two words. There's sometimes 0:19:07.156 --> 0:19:11.036 also funny words that a funny combinations that happen, especially 0:19:11.036 --> 0:19:14.236 because we are showing two words at a time. Oh boy, 0:19:14.676 --> 0:19:16.636 I mean, you know, there's been all kinds of really 0:19:16.676 --> 0:19:19.636 funny examples where it's just like, you know, a website 0:19:19.636 --> 0:19:23.316 of a church that says like bad Christians and it's 0:19:23.356 --> 0:19:26.396 just but these are just two randomly chosen words, so 0:19:26.436 --> 0:19:30.276 we shouldn't infer any evil on your part. No, they're random. Now, 0:19:30.276 --> 0:19:33.796 a lot has happened since two thousand when capture came along, 0:19:34.196 --> 0:19:37.796 and since two thousand and six when you started unsuspectingly 0:19:37.836 --> 0:19:41.276 helping Google in the New York Times digitize their old pages. 0:19:41.756 --> 0:19:43.596 You know, early on in the first version of a 0:19:43.676 --> 0:19:47.556 cap shop, computers were pretty bad at recognizing distorted text, 0:19:47.756 --> 0:19:50.116 so they didn't have to be that distorted. But you know, 0:19:50.236 --> 0:19:52.716 over time, computers got better and better, and in fact, 0:19:52.716 --> 0:19:55.916 by now computers are in many cases about as good 0:19:55.916 --> 0:19:58.196 as humans. Because of that, we have to make them 0:19:58.196 --> 0:20:01.236 harder and harder. A lot of times, the puzzles are 0:20:01.316 --> 0:20:05.156 so hard that even the human can't pass the challenge. 0:20:05.236 --> 0:20:08.356 I'm sure you've been sent screenshots of words that are 0:20:08.436 --> 0:20:11.756 so much no one can tell where it is. Yes, 0:20:12.396 --> 0:20:17.756 that happens, I mean, it's rare that that happens, and 0:20:17.876 --> 0:20:21.996 that's why the capture itself in true arms race fashion 0:20:22.476 --> 0:20:26.436 has evolved. So what has happened is that for the 0:20:26.516 --> 0:20:30.036 more secure things, the captures have moved away from these 0:20:30.076 --> 0:20:33.036 distorted characters. And what is being used now are these 0:20:33.396 --> 0:20:35.676 the puzzles are now things like you see a bunch 0:20:35.676 --> 0:20:37.196 of pictures and you have to click the ones that 0:20:37.316 --> 0:20:41.956 contain a stop sign right the traffic lights, the fire hydrants. Yeah, 0:20:42.036 --> 0:20:45.116 it's exactly the same idea as recapture, except we're not 0:20:45.156 --> 0:20:47.436 the story. We're not trying to digitize books. This a 0:20:47.436 --> 0:20:49.996 lot of times comes from things like all the all 0:20:49.996 --> 0:20:53.996 the mapping cars or the self driving cars. Basically, these 0:20:54.076 --> 0:20:56.956 are cars that are driving around that are capturing images 0:20:56.956 --> 0:20:58.876 of the whole world. They're trying to figure out what's 0:20:58.916 --> 0:21:01.796 around them. Sometimes they cannot recognize what's in an image. 0:21:01.876 --> 0:21:04.356 So it's a similar case. It takes the things like 0:21:04.476 --> 0:21:06.316 is this a stopting I'm not sure, Okay, send it 0:21:06.316 --> 0:21:08.316 to a human, and then when you get it and 0:21:08.356 --> 0:21:10.916 you click on the store sign, you're actually helping either 0:21:10.956 --> 0:21:14.036 the self driving car or the mapping software or whatever 0:21:14.196 --> 0:21:16.356 know that there is actually a stop sign right here. 0:21:16.596 --> 0:21:18.876 Oh so we're still doing good for the world as 0:21:18.876 --> 0:21:21.156 we do this, still doing good for the world, or 0:21:21.316 --> 0:21:23.956 for a company or for a company, but maybe not 0:21:23.956 --> 0:21:26.516 digitizing books. But it's a similar ideas thing that a 0:21:26.556 --> 0:21:30.036 computer cannot do. You've just solved a mystery for hundreds 0:21:30.036 --> 0:21:33.716 of millions of people. Why it's always traffic lights and 0:21:33.756 --> 0:21:37.436 fire hydrants we're supposed to choose and not bananas and puppies, 0:21:37.596 --> 0:21:40.316 or it has to do with both self driving cars 0:21:40.316 --> 0:21:43.556 and also mapping software. Okay, so now we kind of 0:21:43.596 --> 0:21:45.636 get why we have to put up with these challenges, 0:21:46.276 --> 0:21:50.036 or we did twenty years ago, but really nothing better 0:21:50.076 --> 0:21:53.916 has come along since. Are we sure that there's nothing 0:21:54.116 --> 0:21:58.316 less annoying that we could do to thwart these spammers? Yes, 0:21:58.396 --> 0:22:00.676