1 00:00:00,280 --> 00:00:02,840 Speaker 1: Brought to you by the reinvented two thousand twelve camera. 2 00:00:03,160 --> 00:00:08,920 Speaker 1: It's ready. Are you get in touch with technology with 3 00:00:09,039 --> 00:00:17,880 Speaker 1: tech Stuff from how stuff works dot com. Hello again, everyone, 4 00:00:17,920 --> 00:00:20,119 Speaker 1: Welcome to tech Stuff. My name is Chris Poulette and 5 00:00:20,120 --> 00:00:22,320 Speaker 1: I am an editor at how stuff works dot Com. 6 00:00:22,320 --> 00:00:25,760 Speaker 1: Sitting across from me, as always, is senior writer Jonathan Strickland. 7 00:00:25,920 --> 00:00:32,080 Speaker 1: The game is afoot Okay. This episode is about a 8 00:00:32,240 --> 00:00:37,320 Speaker 1: system created by IBM as a scientific experiment to determine 9 00:00:37,360 --> 00:00:41,640 Speaker 1: whether a computer can beat a human in a game 10 00:00:41,680 --> 00:00:47,680 Speaker 1: of skill and intelligence. Jonathan, what is Watson? That is correct? 11 00:00:49,640 --> 00:01:00,040 Speaker 1: I like all I would too? And Big Bucks? Are you? 12 00:01:00,400 --> 00:01:06,080 Speaker 1: Are you a giant computer? Sorry's really reaching back now. 13 00:01:06,120 --> 00:01:08,280 Speaker 1: I would like to tell you my sob story about 14 00:01:08,319 --> 00:01:11,000 Speaker 1: my life so I can win a new refrigerator. There. 15 00:01:11,120 --> 00:01:13,440 Speaker 1: That's reaching back, and it's really obscure. If you know 16 00:01:13,480 --> 00:01:16,200 Speaker 1: what I'm referring to with that particular game show, let 17 00:01:16,280 --> 00:01:18,480 Speaker 1: me know, sadly I do. So I'm just gonna stay 18 00:01:18,480 --> 00:01:20,520 Speaker 1: out at this. I'm not eligible to win. I read 19 00:01:20,520 --> 00:01:24,560 Speaker 1: the rules. So we're gonna talk today about the Watson computer. 20 00:01:24,680 --> 00:01:26,800 Speaker 1: We actually had a lot of listeners right in about 21 00:01:26,800 --> 00:01:30,320 Speaker 1: this because The announcement of the Watson computer came shortly 22 00:01:30,400 --> 00:01:33,800 Speaker 1: after we are episode on. Actually, I think it might 23 00:01:33,840 --> 00:01:36,760 Speaker 1: have even been just before our episode about Computers Versus 24 00:01:36,840 --> 00:01:40,080 Speaker 1: Humans published, So of course it looked like we had 25 00:01:40,080 --> 00:01:46,440 Speaker 1: a glaring omission. Yes, but by in our defense, we 26 00:01:46,480 --> 00:01:49,360 Speaker 1: didn't know about it yet. Yes, actually we mentioned one 27 00:01:49,400 --> 00:01:59,320 Speaker 1: of Watson's cousins predecessors, is probably a predecessor of processors. Yeah. Actually, um, 28 00:01:59,440 --> 00:02:02,480 Speaker 1: deep Blue, I'm sorry, Deep Blue, Deep Blue, Big Blue 29 00:02:02,480 --> 00:02:05,200 Speaker 1: would be the company that made it. But the the 30 00:02:07,640 --> 00:02:10,640 Speaker 1: we're talking about IBM, and IBM does this thing occasionally 31 00:02:10,720 --> 00:02:13,600 Speaker 1: where they issue Yeah, well it is a thing, I mean, 32 00:02:13,639 --> 00:02:16,880 Speaker 1: it's it's because it's not just Deep Blue, it's not 33 00:02:16,960 --> 00:02:21,600 Speaker 1: just Watson. They issue what they call grand challenges their 34 00:02:21,639 --> 00:02:25,160 Speaker 1: engineering teams. Yes, they've had a series of these, and 35 00:02:25,400 --> 00:02:29,320 Speaker 1: some of them are are more noticeable to the public, 36 00:02:29,360 --> 00:02:32,120 Speaker 1: I guess, and others. Deep Blue would definitely be one 37 00:02:32,160 --> 00:02:34,959 Speaker 1: of those because that made headlines. In the nineties. Deep 38 00:02:34,960 --> 00:02:39,040 Speaker 1: Blue was of course the computer that challenged Gary Kasparov, 39 00:02:39,280 --> 00:02:44,600 Speaker 1: the chess grand Master um to a series of games. 40 00:02:44,880 --> 00:02:48,840 Speaker 1: In the first series of games, Kasparov was emerged victorious, 41 00:02:48,840 --> 00:02:52,320 Speaker 1: and in the second Deep blue one, and so that 42 00:02:52,440 --> 00:02:54,240 Speaker 1: was one of those things that kind of propelled the 43 00:02:54,240 --> 00:02:58,120 Speaker 1: whole idea of computers being able to outwit humans, to 44 00:02:58,120 --> 00:03:02,200 Speaker 1: be able to outperform humans in certain tasks. But there 45 00:03:02,200 --> 00:03:05,320 Speaker 1: were other tasks that humans were still much much more 46 00:03:05,440 --> 00:03:10,280 Speaker 1: capable of completing than computers. And UM, as it turns out, 47 00:03:10,280 --> 00:03:13,440 Speaker 1: Watson is a grand challenge. To answer one of those, 48 00:03:14,040 --> 00:03:16,320 Speaker 1: so to speak, or maybe question one of those would 49 00:03:16,320 --> 00:03:17,520 Speaker 1: be better because you have to put it in the 50 00:03:17,520 --> 00:03:20,320 Speaker 1: form of a question, right, That's that's correct. UM. I 51 00:03:20,320 --> 00:03:23,480 Speaker 1: would imagine that Watson does this flawlessly. But we could 52 00:03:23,480 --> 00:03:26,080 Speaker 1: talk about the differences in a human opponent and a 53 00:03:26,120 --> 00:03:28,760 Speaker 1: computer opponent in a little bit. UM. I wanted to 54 00:03:28,760 --> 00:03:31,040 Speaker 1: get into some of the details. Watson is not actually 55 00:03:31,400 --> 00:03:35,920 Speaker 1: a single computer as I typically think about it. UM. 56 00:03:35,960 --> 00:03:38,880 Speaker 1: It is made of ten racks of IBM power, seven 57 00:03:38,880 --> 00:03:42,560 Speaker 1: fifty servers using the Linux operating system. How many cores 58 00:03:42,720 --> 00:03:46,520 Speaker 1: does it have? Two thousand, eight hundred eight processor cores 59 00:03:46,800 --> 00:03:50,880 Speaker 1: wholly free holies? Have you thought your quad core processor 60 00:03:51,120 --> 00:03:55,040 Speaker 1: was the bees knees? I also thought my, uh my 61 00:03:55,160 --> 00:04:00,360 Speaker 1: computers for gigabytes of RAM were pretty much for what 62 00:04:00,400 --> 00:04:04,520 Speaker 1: I'm doing. But Watson has fifteen terabytes of RAM. A 63 00:04:04,640 --> 00:04:10,000 Speaker 1: terabyte is one thousand, twenty four gigabytes, that's right. Also, 64 00:04:10,880 --> 00:04:14,040 Speaker 1: it computes eight at the rate of eighty tarra flops, 65 00:04:14,120 --> 00:04:19,240 Speaker 1: which is eighty trillion calculations per second. And in fact, 66 00:04:20,200 --> 00:04:23,960 Speaker 1: I understand from reading IBM's website about Watson that it 67 00:04:24,080 --> 00:04:29,120 Speaker 1: has somewhere in the neighborhood of two million books essentially. 68 00:04:29,160 --> 00:04:30,960 Speaker 1: I mean, that's it's it's kind of hard to say 69 00:04:31,000 --> 00:04:34,720 Speaker 1: how much information is in a book, but um more 70 00:04:34,760 --> 00:04:36,640 Speaker 1: or less two million books, and it can scan the 71 00:04:36,839 --> 00:04:42,000 Speaker 1: entirety of information on all of those hard drives in 72 00:04:42,040 --> 00:04:46,760 Speaker 1: that machine in roughly two to three second. Right. The 73 00:04:46,800 --> 00:04:49,679 Speaker 1: idea here is that they needed to create a computer. 74 00:04:49,960 --> 00:04:51,800 Speaker 1: You have, the whole the whole challenge here was to 75 00:04:51,800 --> 00:04:54,200 Speaker 1: create a computer that could compete in a game of 76 00:04:54,279 --> 00:04:57,800 Speaker 1: Jeopardy and compete on a championship level. Yeah. And as 77 00:04:57,839 --> 00:04:59,960 Speaker 1: a matter of fact, when we talked about the computer 78 00:05:00,120 --> 00:05:06,000 Speaker 1: to versus person challenge in that podcast, we were discussing how, 79 00:05:06,440 --> 00:05:09,040 Speaker 1: you know, computers do some things really really well and 80 00:05:09,120 --> 00:05:11,160 Speaker 1: some things they don't do so well. And ib AM 81 00:05:11,200 --> 00:05:14,360 Speaker 1: freely admitted that this was a real toughie. Yeah, because 82 00:05:14,400 --> 00:05:16,200 Speaker 1: as it turns out one of the things computers do 83 00:05:16,240 --> 00:05:19,200 Speaker 1: really well. They do well with things like like logical problems, 84 00:05:19,800 --> 00:05:23,480 Speaker 1: you know, because you follow a very set a series 85 00:05:23,480 --> 00:05:28,440 Speaker 1: of steps, things that that obey specific rules. The English 86 00:05:28,560 --> 00:05:35,480 Speaker 1: language does not obey rules as strictly as a mathematical formula. Yes, 87 00:05:36,160 --> 00:05:38,400 Speaker 1: as a matter of fact, we we sort of go 88 00:05:38,560 --> 00:05:42,200 Speaker 1: with with things that might be tricky for computers to 89 00:05:42,240 --> 00:05:45,000 Speaker 1: understand all the time because we constantly on this show 90 00:05:45,040 --> 00:05:49,360 Speaker 1: do wordplay and puns, um, and computers may not necessarily 91 00:05:49,440 --> 00:05:52,680 Speaker 1: understand the nuances of such things, or or slang, or 92 00:05:52,760 --> 00:05:57,120 Speaker 1: metaphors or metaphors. Um. There's a lot of elements to 93 00:05:57,200 --> 00:06:00,640 Speaker 1: human speech that we naturally understand as we develop our 94 00:06:00,720 --> 00:06:03,599 Speaker 1: language skills. Right speak for yourself, I have no idea 95 00:06:03,640 --> 00:06:07,279 Speaker 1: how this thing works, okay, but most of us figure 96 00:06:07,279 --> 00:06:11,280 Speaker 1: out how to determine what someone is talking about based 97 00:06:11,360 --> 00:06:15,280 Speaker 1: on contextual clues and our knowledge of things like wordplay 98 00:06:15,320 --> 00:06:18,400 Speaker 1: and metaphors. So as we build our vocabulary, as we 99 00:06:18,440 --> 00:06:23,040 Speaker 1: build our ability to create sentences, as we understand concepts 100 00:06:23,080 --> 00:06:28,000 Speaker 1: that are not necessarily concrete, then we are able to 101 00:06:28,080 --> 00:06:33,240 Speaker 1: communicate in a more ambiguous way than a computer would 102 00:06:33,279 --> 00:06:37,640 Speaker 1: necessarily be capable of on any normal computer. That is, So, 103 00:06:37,680 --> 00:06:39,919 Speaker 1: what are you trying to say, Johnny get Yeah, what 104 00:06:39,960 --> 00:06:41,640 Speaker 1: I'm trying to say is that I'm trying to say 105 00:06:41,680 --> 00:06:45,160 Speaker 1: is that the depending on the way you word a sentence, Uh, 106 00:06:45,240 --> 00:06:48,080 Speaker 1: a human might be able to determine immediately what the 107 00:06:48,120 --> 00:06:51,240 Speaker 1: significance is of the sentence. You know, what you just said. 108 00:06:51,240 --> 00:06:53,640 Speaker 1: They'd be able to understand it. A computer, depending upon 109 00:06:53,640 --> 00:06:56,159 Speaker 1: the wording, may not be able to interpret it properly 110 00:06:56,240 --> 00:06:59,839 Speaker 1: because you know, you didn't necessarily say like, the ball 111 00:07:00,120 --> 00:07:03,280 Speaker 1: is blue. You know, you might have used a much 112 00:07:03,320 --> 00:07:06,160 Speaker 1: more poetic way of saying it that a computer just 113 00:07:06,200 --> 00:07:09,479 Speaker 1: can't you know, the computer can't equate that as being 114 00:07:09,560 --> 00:07:11,880 Speaker 1: the ball is blue. But any human listener would be 115 00:07:12,760 --> 00:07:15,040 Speaker 1: able to understand what you were getting at and say, oh, 116 00:07:15,080 --> 00:07:17,400 Speaker 1: it's a blue ball. It was just a really fancy, 117 00:07:17,520 --> 00:07:21,680 Speaker 1: flowery way of saying that. Yes, Um, I watched a 118 00:07:21,760 --> 00:07:24,880 Speaker 1: number of videos on the IBM site and some of 119 00:07:24,880 --> 00:07:29,160 Speaker 1: them are quite amusing. Actually, uh, because the early versions 120 00:07:29,160 --> 00:07:32,600 Speaker 1: of Watson just didn't get it. Yeah, they weren't. They 121 00:07:32,600 --> 00:07:37,080 Speaker 1: weren't the most um accurate. And what what's funny about 122 00:07:37,160 --> 00:07:39,760 Speaker 1: is not that the computer didn't get it. But the 123 00:07:39,880 --> 00:07:43,600 Speaker 1: looks on the engineer's faces and as they were going, yeah, okay, no, 124 00:07:43,760 --> 00:07:45,920 Speaker 1: maybe not not so much. We have to go back 125 00:07:45,960 --> 00:07:50,440 Speaker 1: to the drawing board. But Dr Chris Welty was saying 126 00:07:50,640 --> 00:07:53,600 Speaker 1: the point of this exercise is to do the science 127 00:07:54,120 --> 00:07:57,560 Speaker 1: behind this and and they specifically we're looking forward to 128 00:07:57,640 --> 00:08:01,880 Speaker 1: the challenge of Jeopardy and UM. You know, if you 129 00:08:02,520 --> 00:08:05,040 Speaker 1: if you're unfamiliar with the show UM, which some of 130 00:08:05,080 --> 00:08:08,520 Speaker 1: you maybe uh a lot of the questions. Of course, 131 00:08:08,560 --> 00:08:12,840 Speaker 1: the the the answers are presented first. UH. The contestants 132 00:08:12,840 --> 00:08:15,520 Speaker 1: are given the opportunity to choose one of six categories 133 00:08:15,520 --> 00:08:19,600 Speaker 1: that are on the board at different values UH monetary 134 00:08:19,640 --> 00:08:24,000 Speaker 1: values UM. And so you can expect in these categories 135 00:08:24,040 --> 00:08:27,160 Speaker 1: that the the answers UH you are actually supposed to 136 00:08:27,160 --> 00:08:28,920 Speaker 1: give the question if you are contestant on the game. 137 00:08:29,080 --> 00:08:33,400 Speaker 1: The answers can fall within a certain domain of knowledge UM. 138 00:08:33,440 --> 00:08:38,040 Speaker 1: For example, the infamous Potent Potables category UM is about 139 00:08:38,200 --> 00:08:41,480 Speaker 1: alcoholic drinks, and you can expect that if you are 140 00:08:41,640 --> 00:08:44,560 Speaker 1: fairly knowledgeable about different kinds of drinks that you might 141 00:08:44,840 --> 00:08:47,839 Speaker 1: do well or poorly in the category. So you should 142 00:08:47,840 --> 00:08:52,920 Speaker 1: either choose questions or answers from the category or not. Um. Well, 143 00:08:53,280 --> 00:08:55,240 Speaker 1: you know, if no one has bothered to program that 144 00:08:55,280 --> 00:08:59,440 Speaker 1: information into Watson, Uh, then Watson will do poorly in 145 00:08:59,440 --> 00:09:03,920 Speaker 1: that category. But some of the categories on Jeopardy are 146 00:09:03,920 --> 00:09:05,959 Speaker 1: written with a lot of word smithing involved, so you 147 00:09:06,040 --> 00:09:09,640 Speaker 1: might have to supply an answer that rhymes or unscramble 148 00:09:09,840 --> 00:09:13,120 Speaker 1: the war letters to do to form another word. Now, 149 00:09:13,120 --> 00:09:17,240 Speaker 1: the unscrambling thing might come very easy to a computer, um, 150 00:09:17,280 --> 00:09:20,960 Speaker 1: but the rhyming answer, you'd have to go over a 151 00:09:20,960 --> 00:09:23,360 Speaker 1: lot of synonyms in your head to try to find. Okay, well, 152 00:09:23,400 --> 00:09:25,960 Speaker 1: I know the answer to this question, but it obviously 153 00:09:26,000 --> 00:09:29,600 Speaker 1: isn't going to rhyme right. So um. Dr Welty said, 154 00:09:29,720 --> 00:09:31,000 Speaker 1: you know, this is one of the things that we 155 00:09:31,000 --> 00:09:33,920 Speaker 1: were really looking forward to. We wanted, we wanted to challenge. 156 00:09:33,920 --> 00:09:36,439 Speaker 1: We wanted the computer to be answered able to answer 157 00:09:36,559 --> 00:09:40,839 Speaker 1: questions or question answers that the computer normally wouldn't be 158 00:09:40,880 --> 00:09:43,200 Speaker 1: able to. So they were really looking forward to cracking 159 00:09:43,200 --> 00:09:46,640 Speaker 1: this nut, so to speak. Um. They talked about there 160 00:09:46,640 --> 00:09:50,400 Speaker 1: being five major areas that they had to concentrate on 161 00:09:50,480 --> 00:09:54,280 Speaker 1: in order to make Watson work based upon the way 162 00:09:54,360 --> 00:09:59,600 Speaker 1: Jeopardy works, because again they designed this project with a 163 00:09:59,679 --> 00:10:02,640 Speaker 1: very specific application in mind. It helped give them direction 164 00:10:02,720 --> 00:10:04,640 Speaker 1: as opposed to it just being I just want to 165 00:10:04,679 --> 00:10:08,520 Speaker 1: make a computer that is able to analyze semantics and 166 00:10:08,520 --> 00:10:11,439 Speaker 1: and respond. Um. That's you know, that's a much more 167 00:10:11,480 --> 00:10:14,240 Speaker 1: general approach. By giving them the fact that, okay, well, 168 00:10:14,280 --> 00:10:16,240 Speaker 1: our goal is to be able to create a computer 169 00:10:16,280 --> 00:10:21,200 Speaker 1: that can compete and potentially beat champions in Jeopardy, Uh, 170 00:10:21,360 --> 00:10:24,480 Speaker 1: it provided more focus. So with Jeopardy in mind, they 171 00:10:24,480 --> 00:10:26,360 Speaker 1: said the five things they needed to concentrate on was 172 00:10:26,360 --> 00:10:30,560 Speaker 1: that Jeopardy creates a broad and open domain, which means 173 00:10:30,600 --> 00:10:34,200 Speaker 1: that you don't just get questions about one subject. Yes, 174 00:10:34,280 --> 00:10:35,760 Speaker 1: you're not going to have to know everything there is 175 00:10:35,800 --> 00:10:38,000 Speaker 1: to know about alcoholic drinks and that's the only thing 176 00:10:38,040 --> 00:10:39,760 Speaker 1: you were going to be asked about. Right There might 177 00:10:39,800 --> 00:10:46,120 Speaker 1: be politics, pop culture, sports, literature, all sorts of categories 178 00:10:46,120 --> 00:10:49,080 Speaker 1: that you could potentially come up against. So with that 179 00:10:49,120 --> 00:10:51,720 Speaker 1: in mind, the computer had to be able to answer 180 00:10:51,760 --> 00:10:56,360 Speaker 1: those things. Uh. There were as Chris was saying, there 181 00:10:56,440 --> 00:11:01,560 Speaker 1: was an element of complex language. Jeopardy answers can be tricky. 182 00:11:02,000 --> 00:11:04,520 Speaker 1: They're not necessarily straightforward. It's kind of like the New 183 00:11:04,559 --> 00:11:07,360 Speaker 1: York Times crossword puzzle. If you read the clues to 184 00:11:07,400 --> 00:11:11,760 Speaker 1: that crossword puzzle, they aren't necessarily straightforward. They require you 185 00:11:11,840 --> 00:11:15,160 Speaker 1: to make some You have to bridge some gaps in 186 00:11:15,240 --> 00:11:18,760 Speaker 1: order to get to the right answer yes. And in fact, 187 00:11:18,840 --> 00:11:22,040 Speaker 1: they will ask you even in clues for for that puzzle. 188 00:11:22,120 --> 00:11:24,680 Speaker 1: They will ask you for things in poetic language, and 189 00:11:24,720 --> 00:11:26,760 Speaker 1: you'll have to think about things in a completely different 190 00:11:26,760 --> 00:11:30,880 Speaker 1: way than you might have otherwise. The next area that 191 00:11:30,920 --> 00:11:33,560 Speaker 1: they had to focus on was high precision, so you 192 00:11:33,600 --> 00:11:37,559 Speaker 1: had to be able to narrow down your choices and 193 00:11:37,640 --> 00:11:41,679 Speaker 1: find out which of your potential answers would be the most, 194 00:11:42,240 --> 00:11:46,160 Speaker 1: the most accurate, or the best one to choose. Along 195 00:11:46,200 --> 00:11:49,000 Speaker 1: with that was accurate confidence, which means that the computer 196 00:11:49,040 --> 00:11:52,000 Speaker 1: itself has to be able to determine how likely is 197 00:11:52,080 --> 00:11:56,000 Speaker 1: this answer? How likely is this the right answer? Yes? Right, 198 00:11:56,400 --> 00:11:58,760 Speaker 1: and um. And then the last one was high speed. 199 00:11:58,760 --> 00:12:01,200 Speaker 1: It had to be a really really fast computer in 200 00:12:01,280 --> 00:12:04,880 Speaker 1: order to compete against people, because if you know something, 201 00:12:05,120 --> 00:12:07,920 Speaker 1: you just you just spout it out right, you know, 202 00:12:07,960 --> 00:12:11,559 Speaker 1: you you buzz and you say, who is Marshall brain? 203 00:12:12,200 --> 00:12:14,840 Speaker 1: You know? And then you've got the answer, who is 204 00:12:15,280 --> 00:12:18,120 Speaker 1: Marshal brain? I think only one person can answer that question, 205 00:12:18,360 --> 00:12:21,680 Speaker 1: and he is not in the studio today. UM. But yeah, 206 00:12:21,720 --> 00:12:26,559 Speaker 1: you have to have computers capable of of accessing all 207 00:12:26,600 --> 00:12:29,560 Speaker 1: this information and picking it out as quickly as a 208 00:12:29,600 --> 00:12:32,480 Speaker 1: human would be able to. UM. In fact, I saw 209 00:12:32,520 --> 00:12:36,280 Speaker 1: on one of these videos that uh, if you had 210 00:12:36,960 --> 00:12:42,680 Speaker 1: a two point six giga Hurts core processor a computer 211 00:12:42,760 --> 00:12:45,280 Speaker 1: running one of those Okay, posably, I do own a 212 00:12:45,280 --> 00:12:47,719 Speaker 1: computer with a two point six gigga Hurts process right, 213 00:12:47,800 --> 00:12:49,920 Speaker 1: so you know, kind of a middle of the road 214 00:12:50,000 --> 00:12:53,880 Speaker 1: computer right now. But but two point six gigga Hurts computer. 215 00:12:53,920 --> 00:12:57,960 Speaker 1: If you were to try and answer one question uh, 216 00:12:57,960 --> 00:13:02,679 Speaker 1: and you were going to go through all of Watson's 217 00:13:03,000 --> 00:13:06,319 Speaker 1: UH data in order to find that question, the answer 218 00:13:06,360 --> 00:13:08,839 Speaker 1: to that question and compare all the answers and come 219 00:13:08,920 --> 00:13:12,120 Speaker 1: up with the best result and then presented, it would 220 00:13:12,120 --> 00:13:15,959 Speaker 1: take you two hours for that one computer. It doesn't 221 00:13:15,960 --> 00:13:19,199 Speaker 1: surprise me much. So that's why you have that two 222 00:13:19,240 --> 00:13:23,840 Speaker 1: thousand eight processor. You know that with all the different 223 00:13:23,920 --> 00:13:26,800 Speaker 1: uh the web servers running, you have to have those 224 00:13:26,840 --> 00:13:30,280 Speaker 1: core processors running so that you can solve these questions 225 00:13:30,320 --> 00:13:33,719 Speaker 1: in parallel. Excuse me, And you probably remember us talking 226 00:13:33,760 --> 00:13:37,600 Speaker 1: about parallel computing and other podcasts. That's the idea that 227 00:13:37,640 --> 00:13:39,680 Speaker 1: you try and solve a problem by working on parts 228 00:13:39,720 --> 00:13:42,080 Speaker 1: of the problem all at the same time. In this case, 229 00:13:42,640 --> 00:13:47,679 Speaker 1: Watson gets the the answer from Jeopardy and then goes 230 00:13:47,760 --> 00:13:51,520 Speaker 1: through and tries to process all the potential questions that 231 00:13:51,600 --> 00:13:54,640 Speaker 1: would be the correct response to that answer, and then 232 00:13:54,640 --> 00:13:56,920 Speaker 1: it has to evaluate them and choose the right one, 233 00:13:57,360 --> 00:13:59,520 Speaker 1: and has to do this in just a couple of seconds. 234 00:14:00,960 --> 00:14:06,720 Speaker 1: It's a pretty cool idea. The the challenges are not trivial, 235 00:14:09,679 --> 00:14:14,560 Speaker 1: the answers are, but not the the challenges um and 236 00:14:14,600 --> 00:14:16,800 Speaker 1: like you were saying, the early tests were very amusing 237 00:14:16,840 --> 00:14:19,880 Speaker 1: because Watson just didn't get it. It would it would 238 00:14:19,880 --> 00:14:23,560 Speaker 1: give answers that were obviously related to the question, or 239 00:14:23,600 --> 00:14:26,760 Speaker 1: at least related to words that were within the question, 240 00:14:26,800 --> 00:14:29,760 Speaker 1: but we're not the right answer. It's kind of like 241 00:14:29,800 --> 00:14:31,920 Speaker 1: if you were ever using a search engine and you 242 00:14:32,000 --> 00:14:35,360 Speaker 1: put in certain terms and the results you're getting back 243 00:14:35,920 --> 00:14:38,240 Speaker 1: are related to the terms you put in, but not 244 00:14:38,280 --> 00:14:42,080 Speaker 1: to the subject matter you wanted, because it's maybe using hominem's, 245 00:14:42,240 --> 00:14:45,960 Speaker 1: or it's using synonyms, or it's or maybe you misspelled 246 00:14:46,000 --> 00:14:48,400 Speaker 1: something or whatever. But anyway, you're getting the wrong kind 247 00:14:48,400 --> 00:14:53,840 Speaker 1: of responses, same sort of thing. Yep. And speaking of trivial, 248 00:14:54,160 --> 00:14:56,800 Speaker 1: I did want to point out to that Dr Kelly, 249 00:14:56,920 --> 00:14:59,040 Speaker 1: Dr John E. Kelly the third He is a senior 250 00:14:59,120 --> 00:15:01,760 Speaker 1: vice president of ib i'm in the director of IBM Research. 251 00:15:02,480 --> 00:15:06,480 Speaker 1: Um this the project itself, you know, Yes, they're building 252 00:15:06,520 --> 00:15:10,680 Speaker 1: a computer to win a trivia contest, so that might 253 00:15:10,800 --> 00:15:16,360 Speaker 1: seem trivial. Yes, However, the point is, you know, Dr 254 00:15:16,440 --> 00:15:20,800 Speaker 1: Kelly was saying, Look, the amount of information that is 255 00:15:20,960 --> 00:15:27,840 Speaker 1: being created today is rapidly uh, overcoming our ability to 256 00:15:29,160 --> 00:15:31,960 Speaker 1: identify it, process it, makes sense of it, and and 257 00:15:31,960 --> 00:15:34,840 Speaker 1: and derive knowledge from it. Yeah. In fact, I think 258 00:15:34,880 --> 00:15:38,520 Speaker 1: it is a fifteen petabytes of data raw data get 259 00:15:38,560 --> 00:15:41,240 Speaker 1: generated every day, not just not just from people but 260 00:15:41,280 --> 00:15:44,520 Speaker 1: from machines as well. But that's that's an insane amount 261 00:15:44,520 --> 00:15:47,000 Speaker 1: of information. Yes, yes, now, I mean, the human mind 262 00:15:47,040 --> 00:15:49,200 Speaker 1: is a remarkable thing, and if you have systems in place, 263 00:15:49,240 --> 00:15:53,240 Speaker 1: you can help manage that. But at some point, uh, 264 00:15:53,280 --> 00:15:55,160 Speaker 1: you know, even even people can't keep up with that. 265 00:15:55,200 --> 00:15:59,800 Speaker 1: Even there are remarkable computing machines and our skulls. So uh, 266 00:15:59,840 --> 00:16:03,200 Speaker 1: the idea is to build a tool that can actually 267 00:16:03,400 --> 00:16:06,960 Speaker 1: help people. There will be a tool for people to 268 00:16:06,960 --> 00:16:10,880 Speaker 1: help people make sense of this vast amount of information 269 00:16:11,360 --> 00:16:13,800 Speaker 1: and and to overcome that and get get real help 270 00:16:13,840 --> 00:16:19,720 Speaker 1: I guess from machines and and help people understand or 271 00:16:19,840 --> 00:16:23,920 Speaker 1: navigate the world of information that is rapidly creating. UM. 272 00:16:24,120 --> 00:16:26,400 Speaker 1: One of the cooler videos on this site I think 273 00:16:27,400 --> 00:16:29,520 Speaker 1: was the one where they were explaining, look, there there's 274 00:16:29,560 --> 00:16:33,640 Speaker 1: always been this interconnected system of information going on all 275 00:16:33,680 --> 00:16:36,560 Speaker 1: over the world, but we didn't really understand it nearly 276 00:16:36,600 --> 00:16:40,560 Speaker 1: as well. Until the Internet came around. We could actually 277 00:16:40,640 --> 00:16:43,560 Speaker 1: see what was going on, you know, in seconds, rather 278 00:16:43,600 --> 00:16:47,120 Speaker 1: than you know, having it take hours or days or 279 00:16:47,160 --> 00:16:52,120 Speaker 1: weeks or months or even years in many many years past. UM, 280 00:16:52,440 --> 00:16:56,240 Speaker 1: and it's it's just enabled this and is accelerating the problem. 281 00:16:56,320 --> 00:17:00,400 Speaker 1: So UM, the challenge of creating the computer to play 282 00:17:00,400 --> 00:17:03,440 Speaker 1: the game, well, this is basically, I guess an exercise 283 00:17:03,600 --> 00:17:06,359 Speaker 1: to see can we really do this? Can we create 284 00:17:07,080 --> 00:17:12,600 Speaker 1: uh reasonably intelligent computer that can help us, you know, 285 00:17:12,640 --> 00:17:16,119 Speaker 1: figure out what's going on and where the the answers 286 00:17:16,119 --> 00:17:18,919 Speaker 1: are to our questions? Can can we create a computer 287 00:17:19,000 --> 00:17:23,960 Speaker 1: that can understand natural language so that that you challenge it, right, 288 00:17:24,160 --> 00:17:27,000 Speaker 1: It's it's not it's not that you have to tailor 289 00:17:27,080 --> 00:17:29,480 Speaker 1: your language to the computer so that it understands I mean, 290 00:17:29,560 --> 00:17:31,720 Speaker 1: we were familiar with that. You know, we talked about 291 00:17:31,720 --> 00:17:34,880 Speaker 1: Boollyan logic before, about how if you want to do 292 00:17:35,160 --> 00:17:38,600 Speaker 1: really effective search terms, you need to understand how Booleyan 293 00:17:38,640 --> 00:17:41,280 Speaker 1: logic works so that you can. Because search engines don't 294 00:17:41,359 --> 00:17:45,080 Speaker 1: understand natural language, they'll do their best to try and 295 00:17:45,119 --> 00:17:48,520 Speaker 1: match your query with the right result, but they don't 296 00:17:48,600 --> 00:17:53,040 Speaker 1: understand it. They aren't able to analyze the information. One 297 00:17:53,080 --> 00:17:56,520 Speaker 1: of the concepts that it was really important with Watson 298 00:17:56,920 --> 00:17:58,879 Speaker 1: is one that's going to be very important if we 299 00:17:58,960 --> 00:18:02,160 Speaker 1: ever are to have us semantic web, which is the 300 00:18:02,200 --> 00:18:06,080 Speaker 1: idea that you could talk to your computer, whether you're 301 00:18:06,240 --> 00:18:09,240 Speaker 1: actually speaking or typing or whatever. You you can communicate 302 00:18:09,240 --> 00:18:11,840 Speaker 1: with your computer in a natural way, and the computer 303 00:18:11,880 --> 00:18:14,880 Speaker 1: will be able to understand, at least on some level. 304 00:18:15,040 --> 00:18:17,160 Speaker 1: It may not be a deep level, but be able 305 00:18:17,160 --> 00:18:20,480 Speaker 1: to interpret what you're saying and give you the right result. 306 00:18:21,000 --> 00:18:24,600 Speaker 1: Uh in response, that's right. It just it depends on 307 00:18:24,680 --> 00:18:29,159 Speaker 1: a system of contexts, and without those contexts, and the 308 00:18:29,160 --> 00:18:32,080 Speaker 1: computer has to be able to interpret that well, um, 309 00:18:32,800 --> 00:18:36,000 Speaker 1: you're you know, it's it's not nearly as effective as 310 00:18:36,040 --> 00:18:39,360 Speaker 1: it could be um, So this is this is definitely 311 00:18:39,359 --> 00:18:41,920 Speaker 1: a step in the right direction. Yeah, I think it's 312 00:18:41,920 --> 00:18:44,679 Speaker 1: pretty fascinating the way it talked about how or the 313 00:18:44,680 --> 00:18:48,240 Speaker 1: way the the engineers talked about how the computer comes 314 00:18:48,320 --> 00:18:50,600 Speaker 1: up with its answers. So what it does is it 315 00:18:50,600 --> 00:18:54,480 Speaker 1: will it comes up with candidate answers. This is part 316 00:18:54,480 --> 00:18:58,159 Speaker 1: of that parallel processing where all the potential answers to 317 00:18:58,200 --> 00:19:01,159 Speaker 1: a question pop up, and then it turns each of 318 00:19:01,200 --> 00:19:06,160 Speaker 1: those answers into a hypothesis and then examines each hypothesis 319 00:19:06,240 --> 00:19:10,399 Speaker 1: to determine how likely that hypothesis is in fact the 320 00:19:10,520 --> 00:19:13,480 Speaker 1: right answer, and if it doesn't meet a certain level 321 00:19:13,560 --> 00:19:18,159 Speaker 1: of confidence, then then Watson won't buzz in. So Watson 322 00:19:18,240 --> 00:19:20,200 Speaker 1: is not going to buzz in on every question because 323 00:19:20,200 --> 00:19:22,119 Speaker 1: occasionally there's gonna be a question it's gonna be worded 324 00:19:22,119 --> 00:19:24,960 Speaker 1: in such a way that Watson is not really able 325 00:19:25,000 --> 00:19:28,520 Speaker 1: to interpret what what the answer is or just doesn't 326 00:19:28,560 --> 00:19:31,120 Speaker 1: have the information and database. That's another thing we should 327 00:19:31,119 --> 00:19:35,000 Speaker 1: point out. Watson is completely self contained. Yes, it is 328 00:19:35,080 --> 00:19:37,480 Speaker 1: not hooked up to the Internet, so lest you think 329 00:19:37,560 --> 00:19:40,119 Speaker 1: it is searching on Google, it is not. Right. So 330 00:19:40,600 --> 00:19:43,520 Speaker 1: all the information that Watson has available to it is 331 00:19:43,720 --> 00:19:47,320 Speaker 1: self contained. It doesn't. It cannot get more information during 332 00:19:47,320 --> 00:19:51,239 Speaker 1: the course of a game. Now, in between games, um, 333 00:19:51,920 --> 00:19:55,160 Speaker 1: the people ib folks at IBM where it would update Watson, 334 00:19:55,359 --> 00:19:58,400 Speaker 1: especially with things like pop culture references, so that pop 335 00:19:58,480 --> 00:20:01,560 Speaker 1: so that Watson would be able to interpret questions that 336 00:20:01,680 --> 00:20:04,000 Speaker 1: revolved around pop culture and be able to respond to 337 00:20:04,040 --> 00:20:06,840 Speaker 1: them U or news items, things that just happened in 338 00:20:06,880 --> 00:20:09,080 Speaker 1: the news that would have they'd have to update Watson 339 00:20:09,119 --> 00:20:11,760 Speaker 1: with that information as well. But yeah, the key was 340 00:20:11,840 --> 00:20:15,640 Speaker 1: to be able to let Watson break down a sentence 341 00:20:15,680 --> 00:20:18,800 Speaker 1: and really understand what the sentence was saying, not just 342 00:20:19,040 --> 00:20:21,640 Speaker 1: you know this this must be the object and this 343 00:20:21,720 --> 00:20:24,560 Speaker 1: is the the subject and this is the verb, but 344 00:20:24,640 --> 00:20:28,360 Speaker 1: to really understand what it was saying because uh, context, 345 00:20:28,400 --> 00:20:30,840 Speaker 1: as you were pointing out, is so important. One of 346 00:20:30,880 --> 00:20:35,760 Speaker 1: the elements that they talked about was temporal reasoning. Temporal 347 00:20:35,840 --> 00:20:39,320 Speaker 1: reasoning meaning that, uh, there are different ways of saying 348 00:20:39,320 --> 00:20:44,359 Speaker 1: the same thing. For instance, I could say, uh that, um, 349 00:20:44,400 --> 00:20:49,320 Speaker 1: I graduated twenty years ago, or I could say I graduated, 350 00:20:51,280 --> 00:20:53,679 Speaker 1: or I could say the twenty high school reunion is 351 00:20:53,720 --> 00:20:56,280 Speaker 1: coming up for me. All of those things essentially give 352 00:20:56,320 --> 00:21:00,120 Speaker 1: you the same information. By the way I did not graduate. Um. 353 00:21:00,160 --> 00:21:03,480 Speaker 1: But all that all that information, all those those phrases 354 00:21:03,480 --> 00:21:08,200 Speaker 1: give you the same information that I graduated high school. UM, 355 00:21:08,200 --> 00:21:10,520 Speaker 1: but it's different ways of saying it, and a computer 356 00:21:10,800 --> 00:21:14,600 Speaker 1: does not necessarily know that each of those different sentences 357 00:21:14,640 --> 00:21:17,080 Speaker 1: means the same thing. So they had to find a 358 00:21:17,119 --> 00:21:21,080 Speaker 1: way for Watson to learn that, to learn that there 359 00:21:21,119 --> 00:21:25,120 Speaker 1: are many different ways of conveying the same information using 360 00:21:25,240 --> 00:21:29,000 Speaker 1: totally different sentences. And you'll actually be able to see 361 00:21:29,000 --> 00:21:32,119 Speaker 1: that on on February fourteenth, if you tune in to 362 00:21:32,160 --> 00:21:35,280 Speaker 1: watch the show. That's when it's scheduled to air here 363 00:21:35,280 --> 00:21:38,320 Speaker 1: in the United States. Um. And we we know that, 364 00:21:38,520 --> 00:21:41,480 Speaker 1: we know that it performed pretty well already at least, 365 00:21:41,680 --> 00:21:44,359 Speaker 1: let's kind of get into that. Okay, Sorry, No, I 366 00:21:44,440 --> 00:21:46,560 Speaker 1: just figured after after we you know, we could talk 367 00:21:46,560 --> 00:21:48,960 Speaker 1: about the actual show. It's coming up there. I think 368 00:21:49,000 --> 00:21:52,280 Speaker 1: actually the show itself, uh, this particular episode is going 369 00:21:52,320 --> 00:21:54,399 Speaker 1: to be interesting. But well, I was gonna mention that 370 00:21:54,440 --> 00:21:59,320 Speaker 1: a minute, Okay, uh no, basically one of the things 371 00:21:59,320 --> 00:22:01,639 Speaker 1: that I think is really kind of cool. You're not 372 00:22:01,680 --> 00:22:04,119 Speaker 1: going to be just sitting there watching a box and 373 00:22:04,200 --> 00:22:07,160 Speaker 1: to human opponents, they actually made They actually made an 374 00:22:07,160 --> 00:22:11,160 Speaker 1: interface for people to watch, which I think was probably 375 00:22:11,200 --> 00:22:14,000 Speaker 1: key for Jeopardy because I imagine they would actually want 376 00:22:14,040 --> 00:22:15,560 Speaker 1: to see It's like, well, how do we know what 377 00:22:15,560 --> 00:22:18,600 Speaker 1: it's doing? Um, it could be brewing coffee for all 378 00:22:18,640 --> 00:22:22,959 Speaker 1: we know, um, mr coffee. It has an avatar, then 379 00:22:23,000 --> 00:22:24,639 Speaker 1: you'll see it. It looks kind of like a planet 380 00:22:24,640 --> 00:22:27,119 Speaker 1: with a little uh, I don't know, thought wigglies. What 381 00:22:27,240 --> 00:22:33,159 Speaker 1: do you call those? Illustrated I'd call that Doug's hair. Um. Basically, 382 00:22:33,400 --> 00:22:36,879 Speaker 1: if the computer is feeling I put this in quotes, 383 00:22:36,920 --> 00:22:39,960 Speaker 1: if you don't mind confident, the avatar that you see 384 00:22:40,040 --> 00:22:43,280 Speaker 1: is green, so it has it's feeling pretty sure that 385 00:22:43,400 --> 00:22:47,160 Speaker 1: it's got an answer it can use to to buzz in. However, 386 00:22:47,240 --> 00:22:50,440 Speaker 1: if it doesn't have the correct answer, it will be orange, 387 00:22:51,160 --> 00:22:53,760 Speaker 1: so you will be able to see what's going on, 388 00:22:53,880 --> 00:22:55,520 Speaker 1: and you will also be able to see it thinking 389 00:22:55,720 --> 00:22:58,800 Speaker 1: because as the algorithms are processing information to try to 390 00:22:58,840 --> 00:23:02,679 Speaker 1: find an uh A correct question. It's so weird to 391 00:23:02,680 --> 00:23:06,080 Speaker 1: say in this context, um, the avatar is going to flicker, 392 00:23:06,240 --> 00:23:08,800 Speaker 1: so you'll actually be able to see it in the 393 00:23:08,840 --> 00:23:12,640 Speaker 1: process of trying to determine an answer for itself. Um. Now, 394 00:23:12,680 --> 00:23:15,320 Speaker 1: and in two thousand seven, they started building Watson, which, 395 00:23:15,320 --> 00:23:18,440 Speaker 1: by the way, we didn't mention, I don't think uh uh, 396 00:23:18,440 --> 00:23:21,160 Speaker 1: this is named after IBMS founder Thomas J. Watson nine 397 00:23:21,160 --> 00:23:26,000 Speaker 1: after the h Sir Arthur Arthur Conan Doyle character. Right, 398 00:23:26,119 --> 00:23:31,440 Speaker 1: he's not a doctor who who served in India. Um. 399 00:23:31,480 --> 00:23:34,439 Speaker 1: But yeah, that they actually started working on this problem 400 00:23:34,520 --> 00:23:37,720 Speaker 1: and our project in two thousand seven and didn't really 401 00:23:37,720 --> 00:23:40,520 Speaker 1: have a candidate until that. They were ready to share 402 00:23:40,520 --> 00:23:44,320 Speaker 1: with the Jeopardy producers until late two thousand nine. Now. UM, 403 00:23:44,359 --> 00:23:46,359 Speaker 1: one of the videos, or a couple of videos that 404 00:23:46,400 --> 00:23:49,879 Speaker 1: I saw on the website interviewed one of the producers 405 00:23:50,119 --> 00:23:53,880 Speaker 1: of Jeopardy UM and I had his name, Harry Friedman, 406 00:23:54,000 --> 00:23:58,320 Speaker 1: Executive producer. Uh. And he said, basically, you know, we 407 00:23:58,320 --> 00:23:59,760 Speaker 1: were interested in it, but we didn't want it to 408 00:23:59,760 --> 00:24:03,240 Speaker 1: come off as some kind of stunt. Um. And I 409 00:24:03,720 --> 00:24:05,880 Speaker 1: understand that the Jeopardy has sort of a cache as 410 00:24:05,920 --> 00:24:08,320 Speaker 1: being Uh yes, it's a trivia show. But these people 411 00:24:08,359 --> 00:24:11,520 Speaker 1: are seriously intelligent and they have a lot of domain 412 00:24:11,680 --> 00:24:15,359 Speaker 1: you know, cross domain knowledge. Celebrity Jeopardy accepted, of course, 413 00:24:16,960 --> 00:24:23,320 Speaker 1: we won't go there. Um. Actually some of them are anyway. UM. So, 414 00:24:23,720 --> 00:24:26,320 Speaker 1: but that's always entertaining to there there's an element of entertainment, 415 00:24:26,359 --> 00:24:30,120 Speaker 1: but they also have a certain um cash A yes, 416 00:24:30,440 --> 00:24:32,879 Speaker 1: it's like, yeah, we have seriously smart people on this show. 417 00:24:32,920 --> 00:24:36,200 Speaker 1: We don't we don't want to devolve and cheap in 418 00:24:36,240 --> 00:24:38,520 Speaker 1: the show UM. So they showed it to the producers 419 00:24:38,520 --> 00:24:40,960 Speaker 1: in late two thousand nine, and they have video of 420 00:24:40,960 --> 00:24:44,720 Speaker 1: the producers watching Watson perform in a contest with some 421 00:24:45,080 --> 00:24:48,639 Speaker 1: IBM employees and they seemed pretty impressed. Obviously, they're impressed 422 00:24:48,720 --> 00:24:50,960 Speaker 1: enough to actually go forward with the with the show 423 00:24:51,960 --> 00:24:55,040 Speaker 1: UM now to recruit. They recruited two of the very 424 00:24:55,080 --> 00:24:59,440 Speaker 1: best Jeopardy champions for show UM. You probably have heard 425 00:24:59,480 --> 00:25:02,720 Speaker 1: of both of them. One as Ken Jennings who won 426 00:25:02,880 --> 00:25:06,480 Speaker 1: seventy four games a few years ago one two point 427 00:25:06,520 --> 00:25:09,000 Speaker 1: four million dollars on the show, and Brad Rutter, who 428 00:25:09,040 --> 00:25:11,560 Speaker 1: is the all time money champion who won three million, 429 00:25:11,600 --> 00:25:16,520 Speaker 1: two hundred fifty five thousand, hundred two dollars UM. And 430 00:25:16,600 --> 00:25:21,000 Speaker 1: they stand to win one million dollars. Whomever takes home 431 00:25:21,080 --> 00:25:23,680 Speaker 1: first place will take home a million dollars. Second place 432 00:25:23,720 --> 00:25:25,960 Speaker 1: is good for three hundred thousand dollars, and third is 433 00:25:26,000 --> 00:25:29,480 Speaker 1: to two hundred thousand now that the human contestants I 434 00:25:29,480 --> 00:25:32,320 Speaker 1: have agreed to UH to donate half of that charity, 435 00:25:32,359 --> 00:25:35,119 Speaker 1: and I V will donate all of its prize winnings 436 00:25:35,119 --> 00:25:37,440 Speaker 1: to charity, no matter what place it comes in. Yeah, 437 00:25:37,480 --> 00:25:40,919 Speaker 1: that's pretty phenomenal when you consider how much time and 438 00:25:41,000 --> 00:25:44,960 Speaker 1: effort and money must have been put into this project. Yes, now, 439 00:25:45,000 --> 00:25:48,040 Speaker 1: as Jonathan said, these three have already gone at it 440 00:25:48,080 --> 00:25:52,639 Speaker 1: for a a prep round and Watson did pretty well. Yeah. 441 00:25:52,800 --> 00:25:54,840 Speaker 1: Actually I was doing really really well in the first 442 00:25:54,840 --> 00:25:59,959 Speaker 1: half of the game. It ended up winning. Um. And uh, 443 00:26:00,000 --> 00:26:02,880 Speaker 1: actually they asked Brad Rudder. I read an article in 444 00:26:02,880 --> 00:26:07,639 Speaker 1: in Wired magazine UM by Sam Gustin who who was 445 00:26:07,680 --> 00:26:10,919 Speaker 1: writing who talked to Brad Rudder and said, uh, you 446 00:26:10,960 --> 00:26:13,439 Speaker 1: know that He said, are you scared to be going 447 00:26:13,520 --> 00:26:16,240 Speaker 1: up against his computers? Or nervous? He said, and not 448 00:26:16,359 --> 00:26:18,840 Speaker 1: and this is a quote, not nervous, But I will 449 00:26:18,880 --> 00:26:21,280 Speaker 1: be when Watson's progeny comes back from the future to 450 00:26:21,359 --> 00:26:24,400 Speaker 1: kill me. Yeah. There's been a lot of Skynet jokes 451 00:26:24,400 --> 00:26:28,000 Speaker 1: about this, and how jokes as well. UM, but yeah, 452 00:26:28,040 --> 00:26:30,320 Speaker 1: you know we That's one of the other things that's 453 00:26:30,320 --> 00:26:33,720 Speaker 1: really cool about uh Watson is that you know, I 454 00:26:33,800 --> 00:26:36,960 Speaker 1: mentioned a little bit that it kind of thinks thanks 455 00:26:37,000 --> 00:26:42,880 Speaker 1: being yeah, taken in context, folks. Um, No, that Watson 456 00:26:43,680 --> 00:26:46,040 Speaker 1: looks for answers the same way we do, and that 457 00:26:46,560 --> 00:26:49,400 Speaker 1: it has all this information that's been stored in its database. 458 00:26:49,440 --> 00:26:51,159 Speaker 1: But it's all been stored like in the form of 459 00:26:51,280 --> 00:26:54,280 Speaker 1: books and plays and poems and things like that. Right, Yes, 460 00:26:54,760 --> 00:26:59,400 Speaker 1: So it's not organizing all its information and tables, which 461 00:26:59,440 --> 00:27:02,080 Speaker 1: is typic lee how you would do that in a database, 462 00:27:02,720 --> 00:27:06,520 Speaker 1: you know, it's it's actually searching through contextually, which to 463 00:27:06,600 --> 00:27:08,520 Speaker 1: me is phenomenal. That's one of the reasons why. But 464 00:27:08,560 --> 00:27:10,639 Speaker 1: it's also whether reasons why it does so well because 465 00:27:10,640 --> 00:27:14,080 Speaker 1: it's not looking for specific patterns, it's it's looking through 466 00:27:14,200 --> 00:27:18,359 Speaker 1: the actual information. Um. And it was no small feat 467 00:27:18,760 --> 00:27:23,240 Speaker 1: to design this computer. They had several teams working at IBM. 468 00:27:23,280 --> 00:27:25,639 Speaker 1: Actually I've got I've written down the different teams here 469 00:27:25,680 --> 00:27:29,720 Speaker 1: they had. They had an algorithms team that fifteen people 470 00:27:29,720 --> 00:27:31,919 Speaker 1: on it. By the way, some of these teams had 471 00:27:32,080 --> 00:27:35,280 Speaker 1: just had shared members, like there there would be someone 472 00:27:35,280 --> 00:27:38,280 Speaker 1: who be on more than one team. So in total 473 00:27:38,320 --> 00:27:40,600 Speaker 1: it was around twenty five people who worked on this project, 474 00:27:41,280 --> 00:27:44,200 Speaker 1: but fifteen of them were working on algorithms, and these 475 00:27:44,200 --> 00:27:47,560 Speaker 1: were the ones that would identify the context created by 476 00:27:47,560 --> 00:27:51,760 Speaker 1: the question and and look for the available sources UH 477 00:27:52,000 --> 00:27:55,760 Speaker 1: for answers. UM there was a strategy team, and the 478 00:27:55,800 --> 00:27:59,920 Speaker 1: strategy team actually was in charge of designing Watson's game 479 00:28:00,080 --> 00:28:04,920 Speaker 1: play and betting strategies. Well, that's important, that's um. Yeah again, 480 00:28:04,960 --> 00:28:07,800 Speaker 1: if you haven't watched the show, UH, you know, as 481 00:28:07,840 --> 00:28:11,119 Speaker 1: you go on, you either make money when you answer 482 00:28:11,240 --> 00:28:14,520 Speaker 1: questions correctly, get nothing if you don't answer at all, 483 00:28:15,600 --> 00:28:17,880 Speaker 1: but lose money if you And at the final round, 484 00:28:17,880 --> 00:28:20,640 Speaker 1: there are two rounds of regular questioning and once that's done, 485 00:28:20,680 --> 00:28:23,960 Speaker 1: there's what they call Final jeopardy, which is UH a 486 00:28:24,119 --> 00:28:28,280 Speaker 1: last question on which you are shown the category. So 487 00:28:28,400 --> 00:28:31,000 Speaker 1: you have the domain from which this question is being pulled, 488 00:28:31,200 --> 00:28:33,760 Speaker 1: but you don't know what the answer will be for 489 00:28:33,840 --> 00:28:35,560 Speaker 1: you to come up with a question, so you have 490 00:28:35,640 --> 00:28:39,120 Speaker 1: to bet based on what the other two contestants have 491 00:28:39,400 --> 00:28:44,160 Speaker 1: on on their boards versus what you have earned over 492 00:28:44,160 --> 00:28:46,400 Speaker 1: the course of the game. And if if they both 493 00:28:46,440 --> 00:28:49,640 Speaker 1: have fifteen dollars each then and you have ten thousand, 494 00:28:49,720 --> 00:28:51,959 Speaker 1: then you don't have to worry about your betting strategy. Right. 495 00:28:52,000 --> 00:28:54,920 Speaker 1: If your neck and neck you have to figure out, well, 496 00:28:55,000 --> 00:28:58,240 Speaker 1: do I know enough to answer this question or question 497 00:28:58,320 --> 00:29:01,160 Speaker 1: this answer it really is? Or do I do I 498 00:29:01,360 --> 00:29:04,160 Speaker 1: wager that they don't know what it is, and therefore 499 00:29:04,200 --> 00:29:07,000 Speaker 1: I keep my bets small, hoping that they're going to 500 00:29:07,080 --> 00:29:10,080 Speaker 1: bet big and lose enough money so that I win anyway? 501 00:29:10,200 --> 00:29:12,560 Speaker 1: Or am I in the lead? Do I? Am I 502 00:29:12,560 --> 00:29:14,680 Speaker 1: in the lead enough where I can bet a smaller 503 00:29:14,720 --> 00:29:17,520 Speaker 1: amount just so that in case either of them double up, 504 00:29:17,520 --> 00:29:20,480 Speaker 1: they still don't overtake me. Yeah, there's a lot of 505 00:29:20,480 --> 00:29:23,720 Speaker 1: betting strategy involved. Or you could cliff clayvin it and 506 00:29:23,800 --> 00:29:26,480 Speaker 1: just bet the whole thing, even though you are hopelessly 507 00:29:26,720 --> 00:29:28,440 Speaker 1: in the lead. I mean, there's like no way you 508 00:29:28,440 --> 00:29:30,800 Speaker 1: could lose. You bet the whole thing and then you lose. 509 00:29:31,720 --> 00:29:36,080 Speaker 1: Who are seven people who have never been in my kitchen? Uh? 510 00:29:36,080 --> 00:29:37,960 Speaker 1: So Yeah, the strategy team, they were in charge of 511 00:29:38,440 --> 00:29:42,120 Speaker 1: the game playing betting strategies. Then you had the systems team, 512 00:29:42,240 --> 00:29:46,560 Speaker 1: um and uh they were the ones who helped design 513 00:29:46,640 --> 00:29:49,600 Speaker 1: the way that Watson would interpret a question across thousands 514 00:29:49,640 --> 00:29:52,920 Speaker 1: of different cores, you know. So then you've got the 515 00:29:52,920 --> 00:29:55,040 Speaker 1: speech team. So that's the team that actually worked on 516 00:29:55,080 --> 00:29:58,120 Speaker 1: that text to speech capability so that Watson talks too. 517 00:29:58,600 --> 00:30:00,840 Speaker 1: In the game. You don't just see words appear on 518 00:30:00,840 --> 00:30:03,680 Speaker 1: the screen. Watson actually has a voice. It does not 519 00:30:03,760 --> 00:30:06,640 Speaker 1: always pronounce everything correctly, but they worked very hard to 520 00:30:06,680 --> 00:30:10,240 Speaker 1: try and give him a pretty wide range of pronunciations 521 00:30:10,240 --> 00:30:14,360 Speaker 1: because Jeopardy tends to use lots of fancy words. Um. 522 00:30:14,480 --> 00:30:17,880 Speaker 1: There was an annotations team which built the taxonomy for 523 00:30:17,960 --> 00:30:23,480 Speaker 1: the search databases. That's interesting to all our librarians out there. Yes, 524 00:30:23,680 --> 00:30:26,720 Speaker 1: taxonomies are important. I mean, that's how you find information, 525 00:30:26,720 --> 00:30:28,120 Speaker 1: and of course you have to design in such a 526 00:30:28,160 --> 00:30:30,520 Speaker 1: way so that the computer can hit the most likely 527 00:30:30,560 --> 00:30:33,000 Speaker 1: sources first so you can come up with the answer 528 00:30:33,040 --> 00:30:36,760 Speaker 1: as quickly as possible. Uh. There are also teams in China, 529 00:30:36,880 --> 00:30:40,960 Speaker 1: Tokyo and Haifa. Uh. There was a project management team 530 00:30:41,040 --> 00:30:44,240 Speaker 1: which was sort of the liaison between Jeopardy and IBM. 531 00:30:44,320 --> 00:30:46,760 Speaker 1: And then there was an applications team, and that's the 532 00:30:46,800 --> 00:30:50,240 Speaker 1: one that I think is really the most interesting moving forward, 533 00:30:50,280 --> 00:30:53,640 Speaker 1: no matter whether Watson wins on the fourteenth or not. 534 00:30:54,840 --> 00:30:57,800 Speaker 1: The applications team, that's the group that's looking at ways 535 00:30:57,880 --> 00:31:01,760 Speaker 1: to use this kind of capability. Be yawned. The Jeopardy 536 00:31:01,840 --> 00:31:06,360 Speaker 1: scenario so some of the examples I heard were included, 537 00:31:06,400 --> 00:31:08,440 Speaker 1: Like the one that they spent the most time on 538 00:31:08,560 --> 00:31:13,600 Speaker 1: was a diagnostics like medical diagnoses. Yeah, the idea being 539 00:31:13,640 --> 00:31:18,360 Speaker 1: that you could input your doctors could use this when 540 00:31:18,640 --> 00:31:23,280 Speaker 1: seeing patients who are giving, you know, interesting symptoms, something 541 00:31:23,280 --> 00:31:26,840 Speaker 1: that maybe was contradictory, and you would use a computer 542 00:31:26,960 --> 00:31:32,280 Speaker 1: that could could essentially reference the world's information on medical 543 00:31:32,720 --> 00:31:37,760 Speaker 1: knowledge and come up with the most likely of diagnoses, 544 00:31:38,240 --> 00:31:42,240 Speaker 1: which is pretty interesting. But I've also seen other potential 545 00:31:42,320 --> 00:31:44,960 Speaker 1: uses of government and law were two that were mentioned 546 00:31:45,000 --> 00:31:46,640 Speaker 1: as well, which is kind of interesting where you know, 547 00:31:46,680 --> 00:31:49,880 Speaker 1: you start looking for a precedent maybe for a law 548 00:31:49,920 --> 00:31:54,520 Speaker 1: case or something along those lines. So, um, yeah, there's 549 00:31:54,560 --> 00:31:59,240 Speaker 1: there's definitely uses for this beyond just hitting that daily double. 550 00:32:00,080 --> 00:32:02,760 Speaker 1: That's true. That's true. You know, I was just thinking 551 00:32:02,760 --> 00:32:05,960 Speaker 1: about it, uh too. I was reversing in my head 552 00:32:06,000 --> 00:32:09,920 Speaker 1: the betting strategy because when you when you mentioned whether 553 00:32:10,600 --> 00:32:13,600 Speaker 1: Watson wins or not, I started thinking, what if you're 554 00:32:14,040 --> 00:32:17,000 Speaker 1: Brad Rutter or Ken Jennings and you're trying to devise 555 00:32:17,040 --> 00:32:19,440 Speaker 1: a betting strategy and you're like, well, I know he's 556 00:32:19,480 --> 00:32:21,840 Speaker 1: going to do this because I've seen him. I mean, 557 00:32:21,840 --> 00:32:24,240 Speaker 1: both of these guys have played Jeopardy enough times where 558 00:32:24,240 --> 00:32:27,400 Speaker 1: the other one probably knows how they're going to bet. 559 00:32:27,840 --> 00:32:31,440 Speaker 1: But how do you devise a betting strategy against the computer, 560 00:32:31,760 --> 00:32:34,200 Speaker 1: especially a computer that seems to jump all over the board. 561 00:32:34,480 --> 00:32:36,719 Speaker 1: Did you watch any of the things where like there 562 00:32:36,800 --> 00:32:39,880 Speaker 1: was one there was one video in particular where Watson 563 00:32:39,920 --> 00:32:42,040 Speaker 1: got someone went went for like one of the two 564 00:32:42,120 --> 00:32:45,480 Speaker 1: hundred dollar questions, which is the lowest level, right right, 565 00:32:45,800 --> 00:32:48,000 Speaker 1: and uh, and Watson got it right. And then Watson 566 00:32:48,000 --> 00:32:51,440 Speaker 1: went immediately for the thousand or two thousand whatever the 567 00:32:51,440 --> 00:32:53,640 Speaker 1: top level question is now on on that board, it's 568 00:32:53,680 --> 00:32:56,000 Speaker 1: a thousand, okay, So he went right for the like 569 00:32:56,760 --> 00:32:59,400 Speaker 1: in the category had been untouched, so all of the 570 00:32:59,640 --> 00:33:04,520 Speaker 1: all of the versions were available, every single variation of 571 00:33:04,800 --> 00:33:07,200 Speaker 1: however much. I can't even remember how they go anymore 572 00:33:07,200 --> 00:33:09,640 Speaker 1: because I haven't watched it so long. The first round 573 00:33:09,680 --> 00:33:12,080 Speaker 1: of Jeopardy is two hundred four six eight hundred and 574 00:33:12,120 --> 00:33:14,120 Speaker 1: a thousand dollar questions for each kid right, and then 575 00:33:14,200 --> 00:33:17,400 Speaker 1: it doubles four. And I remember when it was one 576 00:33:17,760 --> 00:33:21,440 Speaker 1: d two or three hundred four and oh my god, 577 00:33:21,160 --> 00:33:24,280 Speaker 1: we're old. I think there are people who remember when 578 00:33:24,280 --> 00:33:32,640 Speaker 1: it was um, yeah, Serony San Francisco treat. Uh, I'm 579 00:33:32,680 --> 00:33:34,880 Speaker 1: sorry that was that was I lost on Jeopardy by 580 00:33:35,000 --> 00:33:37,800 Speaker 1: weird al Yankovic. I remember that too. Yeah, I also 581 00:33:37,960 --> 00:33:40,960 Speaker 1: remember when that came out on three D. I think 582 00:33:41,920 --> 00:33:44,880 Speaker 1: I think this is gonna be a fun exper I'm 583 00:33:44,880 --> 00:33:46,560 Speaker 1: sure it's It's been fun for the people who've been 584 00:33:46,560 --> 00:33:50,120 Speaker 1: working on and extremely challenging. Um. I'm interested to see 585 00:33:50,120 --> 00:33:53,400 Speaker 1: how it turns out and whether or not IBM will 586 00:33:53,440 --> 00:33:55,920 Speaker 1: be up for a rematch. Depending on how it goes, 587 00:33:55,960 --> 00:33:58,400 Speaker 1: will they be able to improve it enough, and will 588 00:33:58,440 --> 00:34:00,720 Speaker 1: they convinced the Jeopardy producers to them back on. But 589 00:34:00,960 --> 00:34:03,120 Speaker 1: I think it's gonna be fun. It'll be fun to watch, yeah, 590 00:34:03,160 --> 00:34:07,160 Speaker 1: even if even if it loses. It's such a phenomenal 591 00:34:07,320 --> 00:34:12,520 Speaker 1: achievement to create the algorithms and the database necessary to 592 00:34:12,560 --> 00:34:15,720 Speaker 1: be able to navigate natural language. I mean, that really 593 00:34:15,880 --> 00:34:20,760 Speaker 1: is I did not expect to see it this early, 594 00:34:21,360 --> 00:34:23,560 Speaker 1: you know, I thought that might be a thing, not 595 00:34:23,680 --> 00:34:27,400 Speaker 1: a not a twenty eleven thing. It's it's extremely difficult 596 00:34:27,440 --> 00:34:30,719 Speaker 1: to do. As you can the aforementioned librarians will tell 597 00:34:30,760 --> 00:34:35,720 Speaker 1: you or the catalogs to process natural language questions English English, 598 00:34:35,800 --> 00:34:38,360 Speaker 1: majors will tell you that the language is very difficult 599 00:34:38,360 --> 00:34:41,840 Speaker 1: as well. And you know, so my hat is off 600 00:34:41,880 --> 00:34:46,040 Speaker 1: to to IBM and those those engineers and employees who 601 00:34:46,120 --> 00:34:49,520 Speaker 1: all work together to bring this this technology to life 602 00:34:49,520 --> 00:34:53,359 Speaker 1: because um, like you know, even the applications they were 603 00:34:53,400 --> 00:34:57,160 Speaker 1: talking about, that's just the beginning. We had talked about 604 00:34:57,160 --> 00:35:00,839 Speaker 1: the semantic web before. Um, this is really kind of 605 00:35:00,880 --> 00:35:04,239 Speaker 1: what the semantic web is promising, is as this this 606 00:35:04,400 --> 00:35:07,640 Speaker 1: web experience, uh not grant again. Watson is not a 607 00:35:07,680 --> 00:35:09,879 Speaker 1: web based experience, but a web experience where it can 608 00:35:09,960 --> 00:35:14,319 Speaker 1: understand what you're saying and give you the right response. Oh, yeah, 609 00:35:14,360 --> 00:35:17,359 Speaker 1: I know what you mean. You're looking for this right, right? Yeah, like, 610 00:35:17,520 --> 00:35:19,920 Speaker 1: and I mean it's amazing. You could think in a 611 00:35:19,960 --> 00:35:21,920 Speaker 1: few years you could have a computer that can understand 612 00:35:21,920 --> 00:35:26,360 Speaker 1: a joke. Supposedly it made a joke and yeah. And 613 00:35:26,560 --> 00:35:31,000 Speaker 1: when one of the preliminary games, supposedly it said something 614 00:35:31,040 --> 00:35:33,600 Speaker 1: that caused the entire audience to laugh, and it was 615 00:35:34,000 --> 00:35:35,680 Speaker 1: that it was I think it was Fox News that 616 00:35:35,719 --> 00:35:38,440 Speaker 1: was reporting it, and they did not go into detail 617 00:35:38,520 --> 00:35:41,680 Speaker 1: about what this thing was, but they said that it 618 00:35:41,800 --> 00:35:43,600 Speaker 1: was at the end of one of the like Watson 619 00:35:43,640 --> 00:35:48,399 Speaker 1: got something right and then said something that made people laugh. Now, 620 00:35:48,400 --> 00:35:50,280 Speaker 1: whether or not it was a joke in the sense 621 00:35:50,360 --> 00:35:55,279 Speaker 1: that the computers somehow manifested this desire to make a joke, 622 00:35:55,400 --> 00:35:57,959 Speaker 1: I don't know, because clearly we're not talking about saying 623 00:35:58,000 --> 00:36:02,040 Speaker 1: that's actually alive. If answer is correct and next next 624 00:36:02,120 --> 00:36:08,640 Speaker 1: question has not been asked, say yeah, people on that show, um, 625 00:36:08,680 --> 00:36:11,640 Speaker 1: just follow that logic. So and I'm also looking forward 626 00:36:11,640 --> 00:36:14,080 Speaker 1: to the segment before the second round begins where they 627 00:36:14,080 --> 00:36:16,640 Speaker 1: start asking you about your background. Right, well, Alex, I 628 00:36:16,680 --> 00:36:18,560 Speaker 1: was born four years ago. Right, Well, I don't know 629 00:36:18,560 --> 00:36:21,640 Speaker 1: if you could say born right, And I like computing, 630 00:36:21,800 --> 00:36:24,279 Speaker 1: reading and long walks on the beach. But yeah, the 631 00:36:24,280 --> 00:36:26,160 Speaker 1: other the other side of this that we haven't really 632 00:36:26,200 --> 00:36:28,080 Speaker 1: touched on, and I think it's a good place to 633 00:36:28,080 --> 00:36:32,200 Speaker 1: wrap up. It really shows you how remarkable human beings are. Yeah, 634 00:36:32,800 --> 00:36:35,440 Speaker 1: because look at what has to happen. In order for 635 00:36:35,520 --> 00:36:38,440 Speaker 1: a machine to compete against humans. You have to have 636 00:36:39,440 --> 00:36:43,440 Speaker 1: two thousand, eight h eight cores processors, you have to 637 00:36:43,480 --> 00:36:46,799 Speaker 1: have fifteen terabytes of RAM. You have to have this 638 00:36:47,080 --> 00:36:50,399 Speaker 1: computer that has the equivalent of two million books worth 639 00:36:50,440 --> 00:36:54,319 Speaker 1: of information stored on it. In order to compete with 640 00:36:54,920 --> 00:36:58,600 Speaker 1: humans and in order to even come close right too, 641 00:36:58,640 --> 00:37:01,160 Speaker 1: I mean if if it doesn't win. So that's really 642 00:37:01,280 --> 00:37:04,759 Speaker 1: kind of a testament to how amazing people are, not 643 00:37:04,800 --> 00:37:08,240 Speaker 1: just how amazing the technology is. And I I also 644 00:37:08,400 --> 00:37:12,000 Speaker 1: think it's nice that IBM found a way to do 645 00:37:12,040 --> 00:37:15,319 Speaker 1: this experiment in a way that will actually make people interested, 646 00:37:15,920 --> 00:37:18,000 Speaker 1: right and it building some interesting and I'm glad that 647 00:37:18,000 --> 00:37:22,440 Speaker 1: that Sony Uh Entertainment has found a way to uh, 648 00:37:22,480 --> 00:37:26,879 Speaker 1: you know, use this to their advantage to to show off, um, 649 00:37:26,920 --> 00:37:29,839 Speaker 1: you know, how cool they are essentially, you know, and 650 00:37:29,840 --> 00:37:33,560 Speaker 1: and give IBM an opportunity to play. It's definitely a nice, 651 00:37:33,920 --> 00:37:37,200 Speaker 1: a nice uh event to see. I mean the fact 652 00:37:37,239 --> 00:37:41,200 Speaker 1: that it's going to promote this idea of of the 653 00:37:41,239 --> 00:37:45,920 Speaker 1: semantic computing and artificial intelligence in a way that is 654 00:37:46,280 --> 00:37:50,680 Speaker 1: both entertaining and and really informative. It's it was clever. 655 00:37:50,719 --> 00:37:55,960 Speaker 1: It's a very clever approach. Definitely, So kudos IBM, kudos Jeopardy. 656 00:37:56,760 --> 00:38:00,160 Speaker 1: And with that we're going to wrap this up. You 657 00:38:00,280 --> 00:38:03,520 Speaker 1: have any suggestions for topics or you want to chime 658 00:38:03,560 --> 00:38:06,520 Speaker 1: in on our discussion about Watson, you can let us 659 00:38:06,520 --> 00:38:09,319 Speaker 1: know on Twitter or Facebook. Are handled. There is tech 660 00:38:09,400 --> 00:38:12,600 Speaker 1: stuff hs W or you can write us an email 661 00:38:12,640 --> 00:38:16,000 Speaker 1: and that address is tech stuff at how stuff works 662 00:38:16,000 --> 00:38:17,760 Speaker 1: dot com and Chris and I will talt you again 663 00:38:18,680 --> 00:38:24,160 Speaker 1: really soon. Boop For more on this and thousands of 664 00:38:24,200 --> 00:38:26,640 Speaker 1: other topics. Is it how stuff works dot com. So 665 00:38:26,760 --> 00:38:29,600 Speaker 1: learn more about the podcast clock on the podcast icon 666 00:38:29,719 --> 00:38:32,920 Speaker 1: in the upper right corner of our homepage. The How 667 00:38:33,000 --> 00:38:36,719 Speaker 1: Stuff Works iPhone app has arrived. Download it today on iTunes, 668 00:38:41,600 --> 00:38:44,200 Speaker 1: brought to you by the reinvented two thousand twelve camera. 669 00:38:44,480 --> 00:38:45,719 Speaker 1: It's ready. Are you