WEBVTT - Smart Talks with IBM: The Debating AI 0:00:00.080 --> 0:00:02.719 In this episode, we'll be focusing on Project Debater, which 0:00:02.720 --> 0:00:06.800 is an AI system designed to process evidence and persuasive 0:00:06.880 --> 0:00:10.039 arguments and text so that it can ultimately understand and 0:00:10.080 --> 0:00:13.800 participate in human debate. To get to the heart of 0:00:13.840 --> 0:00:16.880 this effort, we're going to share two interviews we recorded 0:00:16.880 --> 0:00:20.799 with leaders at IBM. The first is with Noam slow Name, 0:00:21.120 --> 0:00:24.200 who is a distinguished engineer at IBM Research and founder 0:00:24.239 --> 0:00:27.600 of Project Debater, and the second chat will be with 0:00:27.680 --> 0:00:31.840 matdou Coachar, who is Vice President Offering Management for IBM 0:00:31.920 --> 0:00:34.680 Data and AI. So today's episode is going to be 0:00:34.760 --> 0:00:38.000 the third of four episodes in this series that Robert 0:00:38.040 --> 0:00:39.840 and I are releasing here on the Stuff to Blow 0:00:39.840 --> 0:00:42.720 Your Mind feed. If you'd like to hear more episodes, 0:00:42.800 --> 0:00:45.159 you can check out the ones labeled smart Talks that 0:00:45.200 --> 0:00:47.479 we've released over the past few weeks, and you can 0:00:47.520 --> 0:00:50.480 also listen to the first four episodes of smart Talks, 0:00:50.520 --> 0:00:52.840 which were released not on our show but in the feed. 0:00:52.880 --> 0:00:55.400 For the podcast Text Stuff. You can find them on 0:00:55.440 --> 0:00:57.720 the I Heart Radio app or wherever you get your podcast. 0:00:57.800 --> 0:01:00.000 Just look up text Stuff and click on the episode 0:01:00.040 --> 0:01:02.840 has labeled Smart Talks, and of course stay tuned for 0:01:02.840 --> 0:01:04.920 the one remaining episode in the series, which is going 0:01:05.000 --> 0:01:06.839 to be published in our feed in a couple of weeks. 0:01:07.200 --> 0:01:10.320 And now straight onto our conversation with no One Slowly, 0:01:12.920 --> 0:01:14.880 no One, thanks so much for joining us today. Can 0:01:14.920 --> 0:01:18.360 you start by introducing yourself and talking about your role 0:01:18.400 --> 0:01:22.880 at IBM? Sure, thank you for hosting me. So I'm 0:01:22.920 --> 0:01:26.200 no One Slownym. I'm a distinguished engineer at IBMI Research. 0:01:27.440 --> 0:01:30.920 I did my PhD in the Hebew University quite a 0:01:30.920 --> 0:01:36.400 few years ago walking on machine learning, staff and artificial intelligence, 0:01:37.040 --> 0:01:39.600 and then I did a past doc at Princeton University 0:01:39.680 --> 0:01:44.240 and I joined the IBM research in two thousand and seven, 0:01:45.319 --> 0:01:50.960 and uh in two thousand and eleven, I suggested the 0:01:50.960 --> 0:01:53.320 project that I guess we're going to talk about today, 0:01:53.520 --> 0:01:56.440 and of course that project was Project Debat, right do 0:01:56.480 --> 0:01:58.280 you do? You want to mention a little bit about 0:01:58.280 --> 0:02:01.640 the origins of that. In IBM research, we have this 0:02:02.120 --> 0:02:08.840 interesting tradition of grand challenges in artificial intelligence. Back in 0:02:08.919 --> 0:02:12.440 the nineties, idem introduced the Blue that was able to 0:02:12.480 --> 0:02:16.800 defeat Gary customers in chess, and in two thousand eleven 0:02:16.880 --> 0:02:19.720 id AM introduced Watson that was able to defeat the 0:02:19.760 --> 0:02:23.760 all time winners of the TV trivia game Jeopardy. And 0:02:23.919 --> 0:02:27.440 just a few days after this event, an email was 0:02:27.520 --> 0:02:31.120 sent to all the thousands of researchers in i DM 0:02:31.200 --> 0:02:35.880 across the globe, myself included, asking us what should be 0:02:35.919 --> 0:02:41.119 the next grand challenge for IDM research and uh I 0:02:41.160 --> 0:02:44.160 was intrigued by that, so I offered my office mate 0:02:44.600 --> 0:02:48.480 at the time to brainstone together, and this is what 0:02:48.520 --> 0:02:50.680 we did. We set in the office in Tel Aviv 0:02:50.880 --> 0:02:54.440 and we raised many different ideas that probably I should 0:02:54.480 --> 0:02:58.600 not share with you today, but at some point towards 0:02:58.639 --> 0:03:02.640 the end of the hour, well I suggested this notion 0:03:03.160 --> 0:03:08.320 of developing a machine that we'll be able to debate humans, 0:03:08.520 --> 0:03:11.399 and that this is how we will demonstrate the technology 0:03:11.480 --> 0:03:15.440 for a full life debate between this envisioned system and 0:03:15.520 --> 0:03:20.720 an expert human debate. And we submitted that the only 0:03:20.720 --> 0:03:23.240 guidance that we got from the management was really to 0:03:23.280 --> 0:03:27.200 submit the proposals in a single side so they will 0:03:27.240 --> 0:03:30.320 not be swamped with too many details. And we were 0:03:30.400 --> 0:03:34.400 able to helpfully follow these guidelines and we submitted a 0:03:34.440 --> 0:03:37.120 single slide. This was fair Boy in two thousand eleven, 0:03:37.840 --> 0:03:40.920 and this started a fairly long, and the thought review 0:03:41.000 --> 0:03:44.880 process that lasted for a year, and in February two 0:03:44.920 --> 0:03:48.280 thousand and twelve, this proposal was selected as the next 0:03:48.320 --> 0:03:52.520 Man Challenge for IBM research and we started to walk 0:03:52.560 --> 0:03:56.440 a few months later with a small team that gradually expanded, 0:03:57.480 --> 0:04:01.839 and we walked on that intensively for I would say 0:04:01.880 --> 0:04:06.160 six and a half yels dedicated solewly to dismission of 0:04:06.600 --> 0:04:09.880 developing a machine that will be able to debate humans. 0:04:10.960 --> 0:04:15.840 And eventually we demonstrated this system in a in a 0:04:15.880 --> 0:04:18.240 full life debate. It was a little bit more than 0:04:18.279 --> 0:04:21.680 a year ago, and it was a debate between this 0:04:21.760 --> 0:04:26.279 system now being called the project debate and one of 0:04:26.400 --> 0:04:31.880 the legendary debates in the history of university debate competitions, 0:04:32.000 --> 0:04:35.400 and it still Harris Naam. It was in San Francisco, 0:04:35.680 --> 0:04:39.240 and and it was a full life debate, surprisingly reminiscent 0:04:39.320 --> 0:04:42.640 to division that we had back in the office in 0:04:42.640 --> 0:04:46.960 Tel Aviv quite a few fields earlier in that single side. 0:04:47.360 --> 0:04:49.520 So the topic of debate brings with it a few 0:04:49.520 --> 0:04:52.920 different connotations, um, you know, and therefore the idea of 0:04:53.240 --> 0:04:56.000 AI entering the frame might might be a bit confusing 0:04:56.080 --> 0:04:58.679 for for some you know, we might imagine a computer 0:04:58.839 --> 0:05:01.840 designed to defeat play or or perhaps a robot that 0:05:01.920 --> 0:05:06.279 can shout louder and a televised US presidential debate to Daddy, 0:05:06.320 --> 0:05:09.360 and can you walk us through what Project Debater is 0:05:09.600 --> 0:05:14.160 and perhaps what it isn't. Yes, absolutely so. So first 0:05:14.160 --> 0:05:17.120 of all, it is worth explaining what we mean, indeed 0:05:17.120 --> 0:05:21.800 by a debate between an AI system like Project Debata 0:05:21.960 --> 0:05:26.960 and a human opponent. So the debate starts with with 0:05:27.080 --> 0:05:30.640 a motion in the debate jargon that defines what we're 0:05:30.680 --> 0:05:34.880 going to debate. And in the event in San Francisco, 0:05:35.000 --> 0:05:38.080 the topic was whether or not the government should subsidize 0:05:38.320 --> 0:05:42.159 the schools. Uh. There are many considerations around how this 0:05:42.320 --> 0:05:44.839 topic is being selected which we can skip, but the 0:05:44.920 --> 0:05:48.160 only thing we should really emphasize is that this topic 0:05:49.080 --> 0:05:52.279 is selected from a list of topics that were never 0:05:52.400 --> 0:05:57.039 included in the training of the system, So the system 0:05:57.120 --> 0:06:00.160 was never able to train on this particular topic. It 0:06:00.320 --> 0:06:03.880 was trying to debate a new topic from from the 0:06:04.240 --> 0:06:08.000 perspective of the machine. And then we are on the 0:06:08.000 --> 0:06:10.680 side of the governments of Project Debta is supporting the 0:06:10.760 --> 0:06:14.200 motion and how the issues on the opposition, and we 0:06:14.320 --> 0:06:18.039 have a full minutes opening speeches for each side and 0:06:18.160 --> 0:06:23.480 full minutely bottom speeches and two minutes closing statements. So 0:06:23.520 --> 0:06:26.400 all you know, we are talking about a little more 0:06:26.440 --> 0:06:29.760 than twenty to twenty five minutes of a discussion that 0:06:29.880 --> 0:06:34.800 we hope we will be a meaningful discussion between Project 0:06:34.839 --> 0:06:38.000 Debata and and and a human plish in these particularly 0:06:38.760 --> 0:06:42.240 so to clarify for people who might not be familiar 0:06:42.279 --> 0:06:45.920 with competitive debating. So competitive debating does not involve what 0:06:46.080 --> 0:06:48.640 people might be more familiar with, which is like passionately 0:06:48.760 --> 0:06:52.279 arguing your actual point of view. It involves having a 0:06:52.360 --> 0:06:56.080 position selected for you that you then must get up 0:06:56.120 --> 0:06:59.440 and defend in front of the judges. Correct, yes, this 0:06:59.520 --> 0:07:02.320 is called act and and this is indeed important to 0:07:02.360 --> 0:07:06.719 emphasize because you do not know in advance what is 0:07:06.720 --> 0:07:10.440 going to be your side. And and even if you 0:07:10.480 --> 0:07:12.080 know in advance that you are going to be on 0:07:12.120 --> 0:07:15.440 the side of the government, we should bear in mind 0:07:15.440 --> 0:07:19.360 the motion could have been phrased we should not subsidize previously, 0:07:20.280 --> 0:07:23.640 and then you should actually contest that. So you do 0:07:23.720 --> 0:07:25.800 not know in advance what is going to be your 0:07:25.880 --> 0:07:29.080 stance to the topic. This is true for Project Debata 0:07:29.160 --> 0:07:32.960 and also for the for the human opponent, and you 0:07:33.080 --> 0:07:36.240 have only ten to fifteen minutes to PerPell. You don't 0:07:36.240 --> 0:07:38.680 know the topic in advance. This is again true for 0:07:39.160 --> 0:07:44.240 project debata and for the human opponent, and uh, your 0:07:44.320 --> 0:07:47.280 goal is really to to persuade the audience. And this 0:07:47.440 --> 0:07:50.960 actually touches on an interesting question of how do you 0:07:51.200 --> 0:07:56.240 do you measure who won the debate? Because in chess 0:07:56.320 --> 0:07:59.040 and in other games this is very clear and and 0:07:59.240 --> 0:08:03.280 really part of the problem with with with debate in 0:08:03.320 --> 0:08:07.840 general and with developing artificial intelligence that is capable of 0:08:07.880 --> 0:08:11.640 debating in particular now is that it is very hard 0:08:12.040 --> 0:08:15.160 to to be fine who actually won the debate. Yeah, 0:08:15.200 --> 0:08:17.680 I know. There are a couple of different metrics. So 0:08:17.760 --> 0:08:20.200 of course one would just be like, what is the 0:08:20.360 --> 0:08:23.320 percentage of the audience that is convinced to either side? 0:08:23.320 --> 0:08:25.560 But that can be problematic because people come in with 0:08:25.600 --> 0:08:28.800 their own opinions already formed on an issue. So one 0:08:29.680 --> 0:08:33.480 metric I've seen is how much the percentages change. They 0:08:33.559 --> 0:08:37.920 ask people before and afterward what their positions are, and 0:08:37.960 --> 0:08:41.680 then after word they say, okay, which side has one 0:08:41.840 --> 0:08:45.199 over more people? Whatever the starting percentages were is, And 0:08:45.440 --> 0:08:48.600 I assume you all had a metric like that precisely so, 0:08:48.600 --> 0:08:51.800 so this is exactly the point, because if you simply 0:08:51.840 --> 0:08:54.520 ask people who is more convinced, you need somehow to 0:08:54.559 --> 0:08:58.600 take into account the opinions to begin with, and and 0:08:58.640 --> 0:09:01.800 the it is done exactly as as you described it. 0:09:01.880 --> 0:09:06.559 And all this event was in collaboration with with Intelligence 0:09:06.720 --> 0:09:10.000 as well, which is really I think the leading platform 0:09:10.160 --> 0:09:14.640 in the US for organizing such a high profile competitive debate. 0:09:15.480 --> 0:09:19.719 It was hosted, the moderator was the moderator Form Intelligence 0:09:19.800 --> 0:09:23.480 as well, John Dunvan, and and the voting was done 0:09:23.520 --> 0:09:27.840 exactly as you described and as being done with the 0:09:27.880 --> 0:09:30.760 show of Intelligence Square. That is, the audience is voting 0:09:31.240 --> 0:09:35.400 before the debate starts, and they vote again after the 0:09:35.440 --> 0:09:38.680 debate ends, and you win if you were able to 0:09:38.720 --> 0:09:41.640 move more people to to your side. Now I think 0:09:41.640 --> 0:09:43.720 a lot of people might be wondering, how on earth 0:09:43.720 --> 0:09:47.640 would you even begin to organize a persuasive argument from 0:09:47.679 --> 0:09:49.680 an AI point of view? Could you walk us through 0:09:49.720 --> 0:09:53.960 the technical specifics of how Project Debater would put together 0:09:54.040 --> 0:09:57.280 an argument. Yes, so we were asking ourselves the same 0:09:57.400 --> 0:10:03.240 question actually when when we started this project. And I 0:10:03.280 --> 0:10:07.480 think this is part of the of the nature of 0:10:07.640 --> 0:10:10.920 such a grand challenge that you do not really know 0:10:11.120 --> 0:10:15.800 how exactly you are going to to approach the problem. 0:10:15.840 --> 0:10:21.000 But we did what computer scientists often do, and this 0:10:21.080 --> 0:10:25.480 is to take this big and somewhat amorphic problem and 0:10:25.600 --> 0:10:31.320 break it into more modular and hopefully more tangible tasks. 0:10:31.480 --> 0:10:37.640 And so in general, the debated system had uh two 0:10:37.760 --> 0:10:41.520 major sources of information. One of them is the massive 0:10:41.600 --> 0:10:48.480 collection of around four hundred million newspaper articles, and when 0:10:48.480 --> 0:10:54.679 the debate starts, the system was using various AI artificial 0:10:54.720 --> 0:11:00.840 intelligence engines in order to try and pinpoint short pieces 0:11:00.880 --> 0:11:04.480 of text within this massive collection. We're talking about ten 0:11:04.600 --> 0:11:09.120 billion sentences, so we were trying to automatically pinpoint these 0:11:09.200 --> 0:11:14.880 short pieces of text that should satisfy three criteria. They 0:11:14.880 --> 0:11:19.200 should be relevant to the topic, they should be argumentative 0:11:19.240 --> 0:11:22.880 in nature, they should argue something about the topic, and 0:11:22.920 --> 0:11:26.720 they should support our side of the debate. And this 0:11:26.800 --> 0:11:30.280 is quite a formidable challenge. But assuming that you are 0:11:30.320 --> 0:11:33.360 capable of finding these short pieces of tax, the system 0:11:33.440 --> 0:11:38.040 is then using other AI capabilities in order to try 0:11:38.160 --> 0:11:42.559 and glue them together into a meaningful narrative. So this 0:11:42.679 --> 0:11:47.080 is one major source of information for the system. The 0:11:47.160 --> 0:11:50.600 second important source of information for the system was a 0:11:50.800 --> 0:11:58.480 unique collection of more principled arguments that were actually written 0:11:58.800 --> 0:12:03.320 by by humans, and we are talking about thousands of 0:12:03.480 --> 0:12:07.400 more principled arguments. And the role of the system was 0:12:07.440 --> 0:12:11.080 when the debate starts, was really to navigate within this 0:12:11.200 --> 0:12:14.880 collection and find the most relevant principled arguments and use 0:12:14.960 --> 0:12:17.280 them in the right timing. So so, to make this 0:12:17.360 --> 0:12:21.839 more concrete what we mean by a principal argument, imagine 0:12:21.880 --> 0:12:25.200 that we are debating whether or not to ban organ 0:12:25.280 --> 0:12:28.400 trade or whether or not to ban the sale of alcohol. 0:12:28.880 --> 0:12:32.040 In both cases, the opposition may argue that if you 0:12:32.120 --> 0:12:35.520 ban something, you are at the risk of the emergence 0:12:35.559 --> 0:12:38.120 of a black market. So a black market is a 0:12:38.160 --> 0:12:41.280 principled argument that can be used almost in the same 0:12:41.360 --> 0:12:46.400 way in many different contexts. So one may naively assume 0:12:47.360 --> 0:12:51.760 that this is kind of a simple keyword matching thing. 0:12:52.040 --> 0:12:55.560 If we ban something, then the opposition is going to 0:12:55.679 --> 0:12:58.240 use the black market argument, and we should be prepared 0:12:58.320 --> 0:13:02.000 for that. But obviously this is far from true. So 0:13:02.120 --> 0:13:07.679 imagine a debate about banning breastfeeding in public. Obviously there 0:13:07.800 --> 0:13:11.400 is little risk for a black market in this contract. 0:13:11.520 --> 0:13:15.000 Or imagine a debate about banning internet cookies. We're not 0:13:15.120 --> 0:13:18.320 going to tee a black market of internet cookies if 0:13:18.360 --> 0:13:22.560 we band these. So the system really needs to develop 0:13:22.640 --> 0:13:27.960 a more subtle understanding after human language in order to 0:13:28.080 --> 0:13:31.920 be able to identify the most relevant principle argument and 0:13:31.960 --> 0:13:35.400 need use them doing a debate. And and this is, 0:13:35.440 --> 0:13:40.199 by the way, just what all this description is before 0:13:40.320 --> 0:13:43.800 listening to the opponent. This is just what we're going 0:13:43.880 --> 0:13:47.840 to say on our side. And and the most the 0:13:47.920 --> 0:13:51.960 most challenging part is really too uh to listen to 0:13:52.000 --> 0:13:55.199 the opponent. And it's some kind of a battle to 0:13:55.320 --> 0:13:59.040 the arguments generated by the opponment raised by the And 0:13:59.200 --> 0:14:03.920 we do that you using uh an arsenal of technique 0:14:04.320 --> 0:14:07.240 that most of them rely on the same principle. We 0:14:07.360 --> 0:14:11.800 start by listening to the world articulated by the opponment, 0:14:11.920 --> 0:14:14.760 and for that we simply use what's on speech recognition 0:14:15.200 --> 0:14:17.839 capabilities out of the box. But of course we need 0:14:17.880 --> 0:14:20.360 to go to beyond the world, and we need to 0:14:20.440 --> 0:14:23.520 understand the gist of the arguments of the opponent. And 0:14:23.560 --> 0:14:27.640 in order to do that we try using various smackloads 0:14:27.680 --> 0:14:33.200 to anticipate in advance what kind of arguments the opposition 0:14:33.800 --> 0:14:38.520 mind you and then listen to determine whether he did 0:14:38.560 --> 0:14:43.720 the opposition was making these arguments and then responded cold yeah. 0:14:43.800 --> 0:14:47.920 That calls to mind the question of the difference between, say, 0:14:47.960 --> 0:14:51.720 what's a sound argument versus what's a persuasive argument? I mean, 0:14:51.960 --> 0:14:55.840 we know from reality that often the most persuasive appeals 0:14:55.840 --> 0:15:00.680 and debates rely on just straightforwardly false claims and logical fallacies, 0:15:00.840 --> 0:15:03.960 or even on little emotional cues that have little to 0:15:04.000 --> 0:15:06.680 do with the matter at hand. I was thinking about 0:15:06.680 --> 0:15:09.240 how in live debates, if you can get a laugh 0:15:09.400 --> 0:15:12.560 at your opponent's expense, that's worth you know, a dozen 0:15:13.640 --> 0:15:18.200 sound arguments or claims. So to what degree can AI 0:15:18.360 --> 0:15:21.840 understand these sorts of persuasive appeals that that go beyond 0:15:22.000 --> 0:15:24.560 just like what kind of evidence you can bring and 0:15:24.640 --> 0:15:29.880 the appeals based on style you're right in in in 0:15:29.880 --> 0:15:33.040 in debate and in the methods. We know already from 0:15:33.080 --> 0:15:38.320 the ancient weeks that that we have free elaps, we 0:15:38.440 --> 0:15:43.240 have logos, and we have ethos, and we have afforts, 0:15:43.280 --> 0:15:47.120 and humans are using a mixture of these pilas when 0:15:47.440 --> 0:15:51.600 they are debating one another. And just as a quick clarification, logos, 0:15:51.640 --> 0:15:55.080 pathos and ethos are the types of appeals that were 0:15:55.120 --> 0:15:58.840 identified in the study of classical rhetoric. Where logos is 0:15:58.920 --> 0:16:03.160 appeals based on our logical arguments and evidence, Pathos is 0:16:03.160 --> 0:16:06.200 the appeal to the emotions or the passions, and ethos 0:16:06.320 --> 0:16:09.440 is an appeal based on the credibility or authority of 0:16:09.480 --> 0:16:14.360 the speaker. I mean, as you know broadly understood and 0:16:14.360 --> 0:16:18.800 and the technology that we developed, and and by the way, 0:16:18.800 --> 0:16:23.600 it should be stated that there is a rapidly emerging 0:16:23.680 --> 0:16:28.880 community of scientists across the globe that are investigating this 0:16:29.080 --> 0:16:32.120 kind of topic. It is all under the umbrella of 0:16:32.240 --> 0:16:38.240 this emerging field, yeah, referred to as a computational argumentation. 0:16:38.960 --> 0:16:41.760 And when we started in two thousand and twelve, there 0:16:41.920 --> 0:16:46.720 was a handful of teams pursuing that, and we see 0:16:46.720 --> 0:16:50.160 a very dramatic increase in the result in these areas 0:16:50.200 --> 0:16:54.160 of the last few years is very I think from 0:16:55.000 --> 0:17:01.880 exacting and as I mentioned, the technology that we developed 0:17:01.880 --> 0:17:06.159 a most focused on logos, and you can see in 0:17:06.200 --> 0:17:09.640 the debate between proper Debate and Hali. By the way, 0:17:09.680 --> 0:17:13.879 this this debate is is fully available on YouTube, and 0:17:14.080 --> 0:17:19.080 you can see that indeed a woman is better in 0:17:19.240 --> 0:17:23.560 making in using path as and perhaps in using ethos 0:17:23.600 --> 0:17:27.040 and it is harder for the machine. And indeed most 0:17:27.040 --> 0:17:30.800 of the research being done by by the by the 0:17:30.880 --> 0:17:35.560 relevant research communities around logos, but there are already attempt 0:17:36.040 --> 0:17:40.320 trying to model and to capture additional aspect of path 0:17:40.359 --> 0:17:44.400 of and ethos in all the further enhanced this kind 0:17:44.400 --> 0:17:48.240 of technology. So another question I have is debater has 0:17:48.280 --> 0:17:52.879 to source claims and facts and arguments from existing written 0:17:52.880 --> 0:17:55.400 work produced by humans, which of course we know can 0:17:55.440 --> 0:17:58.280 be full of all sorts of flaws. Is there any 0:17:58.280 --> 0:18:01.600 way at this point for it to to have an 0:18:01.600 --> 0:18:05.960 analytical function to tell a say, factually true claim or 0:18:06.000 --> 0:18:09.960 a logically valid argument from just something that is wrong 0:18:10.080 --> 0:18:12.920 or dubious but repeated a lot in writing, or are 0:18:13.000 --> 0:18:19.200 we not there yet? This is a very kindly important 0:18:19.240 --> 0:18:24.000 and difficult problem, and that is receiving going attempting over 0:18:24.280 --> 0:18:30.639 over the previous teams and go to tackle that. This 0:18:30.840 --> 0:18:35.080 is certainly not bullet bof and and the problem is 0:18:35.080 --> 0:18:39.520 is quite complex because one may say, you know, okay, fine, 0:18:39.600 --> 0:18:43.600 maybe I should only take my argument from highly credibally 0:18:43.640 --> 0:18:49.760 so and by boxy I can assume that that these 0:18:49.880 --> 0:18:54.720 arguments are our valid. But this is not necessarily the case. Right. 0:18:54.800 --> 0:18:58.240 You can see you can lead an opinion article in 0:18:58.359 --> 0:19:05.240 a highly respectable newspaper which is actually quoting a false 0:19:05.359 --> 0:19:08.879 argument that was made as well, and if you're not 0:19:08.920 --> 0:19:13.440 careful enough, you you might be your system is going 0:19:13.480 --> 0:19:17.440 to pull this argument without understanding that something is happening. 0:19:18.119 --> 0:19:21.199 So we try to develop and we actually part of 0:19:21.240 --> 0:19:26.800 Project Debate included some kind of filtering mechanism in order 0:19:26.920 --> 0:19:30.119 to to filter out these kind of cases. And the 0:19:30.200 --> 0:19:34.359 way we did that was really once a specific claim 0:19:34.920 --> 0:19:37.800 was affected and by the way to being ordered, the 0:19:37.960 --> 0:19:41.200 claim is not a full sentence. A claim is often 0:19:41.600 --> 0:19:44.600 only a part of a tentence. Even if you were 0:19:44.680 --> 0:19:48.720 able to detect sentence that contains a claim relevant one 0:19:48.800 --> 0:19:51.800 that supportal side out of the billions of sentences in 0:19:51.840 --> 0:19:55.160 the popos, you still need to find the coret boundaries 0:19:55.560 --> 0:19:58.520 after claim within the sentence, and you have hundreds of 0:19:58.600 --> 0:20:02.440 options and only all of them is correct. So this 0:20:02.520 --> 0:20:05.320 is just going back why this this problem is it 0:20:05.440 --> 0:20:08.760 so talenting? But until you do that and found this 0:20:08.960 --> 0:20:12.119 claim and asked what is the stance of this claim, 0:20:12.440 --> 0:20:15.560 and if the stance is supporting your side, you can 0:20:15.600 --> 0:20:18.920 still ask what is the stance of the full sentence? 0:20:20.200 --> 0:20:22.439 And if the stance of the full sentences in the 0:20:22.480 --> 0:20:26.520 opposite direction, you may suspect that something is going on. 0:20:27.359 --> 0:20:31.280 And perhaps this this claim is quoted in order to 0:20:31.920 --> 0:20:35.680 contradict and not because it is true. And then perhaps 0:20:35.680 --> 0:20:39.879 it is there it is safer to avoid using it. 0:20:39.920 --> 0:20:44.120 But but this is just one safety mechanism, and and 0:20:44.200 --> 0:20:46.880 the problem that you raise is actually a much more 0:20:46.960 --> 0:20:52.080 beneval one, and and I think many teams are working 0:20:52.119 --> 0:20:55.359 on that, and we try to address that as well. 0:20:55.600 --> 0:21:00.280 And I think it has many interesting dimensions because it 0:21:00.400 --> 0:21:04.600 is not even just about the validity of the argument. Often, 0:21:04.680 --> 0:21:08.600 when when you show people to arguments, they will agree 0:21:08.640 --> 0:21:11.720 that one of them is better than the other. But 0:21:11.920 --> 0:21:15.920 what are the underlying mechanisms that I'd ask to the 0:21:16.160 --> 0:21:19.439 one argument over the other, And how do you train 0:21:19.800 --> 0:21:23.639 an artificial as in system to make the distinction. This 0:21:23.840 --> 0:21:27.240 is kind of another example of the problems that welcome 0:21:27.280 --> 0:21:30.440 to them. I have a question about what could come 0:21:30.520 --> 0:21:33.960 out of AI research like this, because I would say, 0:21:33.960 --> 0:21:36.879 from my personal perspective, I think studying rhetoric and debate 0:21:37.040 --> 0:21:43.040 is extremely important, but not necessarily because getting into debates 0:21:43.160 --> 0:21:45.760 is a good way to figure out what's true and 0:21:45.920 --> 0:21:48.040 establish you know, the right thing to do. I think 0:21:48.119 --> 0:21:50.720 one of the most important reasons to study rhetoric and 0:21:50.760 --> 0:21:54.800 debate is so that you can understand how other people's 0:21:54.960 --> 0:21:58.760 arguments and persuasive appeals are operating on you, or are 0:21:58.880 --> 0:22:02.280 designed to operate you. A clear understanding of rhetoric can 0:22:02.280 --> 0:22:04.959 be a kind of suit of armor for going into 0:22:05.160 --> 0:22:08.879 you know, the world and seeing how political actors and 0:22:08.960 --> 0:22:11.960 business actors and advertising and all that is trying to 0:22:12.040 --> 0:22:15.720 affect you. Do you see project debate or serving any 0:22:15.800 --> 0:22:19.080 kind of educational purpose like this in the world today. 0:22:19.560 --> 0:22:25.080 So there are several levels by which I can I 0:22:25.119 --> 0:22:30.680 can answer that. The first one is that this kind 0:22:30.680 --> 0:22:36.320 of technology is is definitely relevant and we believe highly 0:22:36.520 --> 0:22:42.200 valuable in the context of education. You can imagine using 0:22:42.240 --> 0:22:47.080 the technology in order to build better arguments and more 0:22:47.119 --> 0:22:53.640 of all, to perform a more analytical and perhaps more 0:22:53.680 --> 0:23:01.640 objective analysis off complex and controversial topics. This is one aspect. 0:23:02.560 --> 0:23:06.280 There is another aspect, but often when we debate is 0:23:06.920 --> 0:23:13.639 other humans. There are many layouts that that are involved 0:23:13.840 --> 0:23:16.560 in this discussion. In this debate. What all of them 0:23:16.560 --> 0:23:20.119 are related? To the facts and to the arguments that 0:23:20.200 --> 0:23:23.400 we are raising. Perhaps we have history with that Belton, 0:23:23.880 --> 0:23:28.040 Perhaps we have history with ourselves that actually impact our 0:23:28.119 --> 0:23:32.600 on part and decisions. Perhaps other people are listening and 0:23:32.680 --> 0:23:38.480 this actually improvides contact, uh that impact what is happening. 0:23:38.960 --> 0:23:42.720 And we are curious about this option of the dating 0:23:42.800 --> 0:23:47.680 with the machine in the privacy of your office. Maybe 0:23:47.720 --> 0:23:51.840 this is a different form of a discussion that to 0:23:52.000 --> 0:23:58.159 some extent is perhaps all free of of external biases 0:23:58.240 --> 0:24:04.280 and maybe will enable treat some people to identify situations 0:24:04.280 --> 0:24:08.000 where they have a blind book and to better listen 0:24:08.320 --> 0:24:12.159 to the other side. So I think in this case 0:24:12.280 --> 0:24:17.400 the whole of the technology could be quite instrumental and positive. 0:24:17.840 --> 0:24:21.880 The false business applications that are also very interesting from 0:24:21.920 --> 0:24:28.240 the IBM perspective and uh, and this is another another dimension, 0:24:28.520 --> 0:24:37.280 another level by which we can consider the technology as exacuable. Again, 0:24:37.280 --> 0:24:39.720 big thanks to No One slow name for taking time 0:24:39.760 --> 0:24:41.600 to chat with us. And now we're going to go 0:24:41.640 --> 0:24:49.240 straight into our second talk on the subject with Madu Matt. 0:24:49.520 --> 0:24:51.960 Thanks so much for joining us today. Could you start 0:24:52.000 --> 0:24:56.320 off by introducing yourself and talking about your role at IBM. Yeah, absolutely, 0:24:56.440 --> 0:24:59.920 and really nice to meet you. Robert and Joe uh 0:25:00.040 --> 0:25:06.040 maduco Chi, vice President Offering Management in Data and AI IBM, 0:25:06.080 --> 0:25:09.159 And the role of offering management is really all about 0:25:09.680 --> 0:25:14.040 laying down the strategy and then delivering and executing towards 0:25:14.119 --> 0:25:18.520 such strategy. And I'm based out of San Jose, Sunny, California, excellent. 0:25:19.359 --> 0:25:21.679 So just to kick things off here, um, you know 0:25:21.680 --> 0:25:24.600 we're gonna be talking a lot about AI here, and 0:25:25.440 --> 0:25:28.280 it makes sense to to to really get into what 0:25:28.359 --> 0:25:31.520 we mean when we're talking about AI for business. How 0:25:31.560 --> 0:25:35.560 does AI serve business compared to the way it serves consumers. 0:25:36.359 --> 0:25:39.520 That's a great question to get started on. UM so 0:25:40.160 --> 0:25:44.679 redeveloped a thesis a couple of years ago about really 0:25:44.720 --> 0:25:50.000 how AI for business would be different from consumer AI. 0:25:50.280 --> 0:25:53.159 Think of consumer AI, which we all know work with 0:25:53.200 --> 0:25:58.200 our smartphones, smart speakers, social media, photos, everything what it comes. 0:25:58.280 --> 0:26:01.440 But when it comes for AI for business, it's really 0:26:01.560 --> 0:26:06.880 very very different. AI for business is all about automation, 0:26:07.280 --> 0:26:12.560 optimization and making better predictions, and it requires really a 0:26:12.680 --> 0:26:16.000 very different set of technical capabilities, like you would have 0:26:16.040 --> 0:26:19.080 to understand how to deal with language, have to deal 0:26:19.119 --> 0:26:23.240 with what does automation means, and then be able to 0:26:23.680 --> 0:26:28.120 have the explainability and trust up AI. UM. So that's 0:26:28.119 --> 0:26:31.200 sort of the big difference between commercial AI and AI 0:26:31.280 --> 0:26:33.560 for business. So we know that one of the big 0:26:33.600 --> 0:26:36.280 AI projects at IBM is Watson. Could you tell us 0:26:36.320 --> 0:26:39.880