WEBVTT 00:00:00.500 --> 00:00:01.964 [SQUEAKING] 00:00:01.964 --> 00:00:04.419 [RUSTLING] 00:00:04.419 --> 00:00:05.892 [CLICKING] 00:00:09.843 --> 00:00:11.260 GARY GENSLER: On Monday, we talked 00:00:11.260 --> 00:00:13.540 about financial technology as a whole, 00:00:13.540 --> 00:00:16.210 and what we're going to do in these next four classes 00:00:16.210 --> 00:00:20.590 is really talk about three major technologies. 00:00:20.590 --> 00:00:24.250 The broad subject of artificial intelligence, machine learning, 00:00:24.250 --> 00:00:27.100 and deep learning, we'll talk about today, 00:00:27.100 --> 00:00:30.760 and next Monday, we'll move on and talk about the marketing 00:00:30.760 --> 00:00:35.080 channels, what I'll broadly call OpenAPI and some 00:00:35.080 --> 00:00:39.640 of those conversational agents and the relationship that 00:00:39.640 --> 00:00:41.210 is happening there. 00:00:41.210 --> 00:00:45.930 And then next Wednesday's class, I might be off. 00:00:45.930 --> 00:00:47.830 It's the following Monday's class, 00:00:47.830 --> 00:00:50.680 we'll talk about blockchain technology and cryptocurrency. 00:00:50.680 --> 00:00:53.920 So we're taking three big slices of technology, 00:00:53.920 --> 00:00:56.190 and then we're going to go into the sectors. 00:00:56.190 --> 00:01:00.430 And so what I thought of about in structuring this class, 00:01:00.430 --> 00:01:03.790 as we've already discussed, is I take a broad view 00:01:03.790 --> 00:01:08.470 of what the subject of fintech is, that it's technologies 00:01:08.470 --> 00:01:16.080 of our time that are potentially materially affecting finance. 00:01:16.080 --> 00:01:18.525 So in any decade, any five year, it 00:01:18.525 --> 00:01:20.010 might be different technologies. 00:01:20.010 --> 00:01:22.680 It's technologies of a certain time that are materially 00:01:22.680 --> 00:01:25.800 affecting finance, and it's the broad-- 00:01:25.800 --> 00:01:27.750 any competitor can use fintech. 00:01:27.750 --> 00:01:30.810 Now, I recognize that a lot of people 00:01:30.810 --> 00:01:33.270 use the subject or fintech, and they narrow it down, 00:01:33.270 --> 00:01:36.420 and they just use those terms for the disruptors, 00:01:36.420 --> 00:01:38.190 the startups. 00:01:38.190 --> 00:01:39.690 And there were really good questions 00:01:39.690 --> 00:01:42.990 of why I think of it more broadly. 00:01:42.990 --> 00:01:47.010 I just think that the incumbents are very much involved 00:01:47.010 --> 00:01:49.260 in this fintech wave, and to ignore 00:01:49.260 --> 00:01:54.640 that would be at your peril if you're starting a new business, 00:01:54.640 --> 00:01:56.520 if you're an entrepreneur. 00:01:56.520 --> 00:02:01.170 And of course, if you're Jamie Dimon running JPMorgan Chase, 00:02:01.170 --> 00:02:03.877 or you're running a big bank in China or in Europe, 00:02:03.877 --> 00:02:06.210 you would be at your peril if you didn't think about it. 00:02:06.210 --> 00:02:11.020 And big tech, of course, has found their way in here. 00:02:11.020 --> 00:02:15.403 The Alibaba example's in China, but all around the globe, 00:02:15.403 --> 00:02:17.070 whether it's Brazil, whether it's India, 00:02:17.070 --> 00:02:21.780 whether it's-- your big tech has found their way in here. 00:02:21.780 --> 00:02:24.300 And so I've also decided to sort of structure it 00:02:24.300 --> 00:02:28.470 around these three big thematic technologies, 00:02:28.470 --> 00:02:32.040 and then we'll start to dive into the sectors themselves 00:02:32.040 --> 00:02:34.680 and take a look at four or five sectors 00:02:34.680 --> 00:02:40.410 before we call it a day at the end of the half semester. 00:02:40.410 --> 00:02:42.780 So I'm about to pull up some slides, but, Romain, 00:02:42.780 --> 00:02:45.802 are there any broad questions? 00:02:45.802 --> 00:02:47.010 ROMAIN: Nothing so far, Gary. 00:02:47.010 --> 00:02:48.177 GARY GENSLER: OK, thank you. 00:02:51.070 --> 00:02:55.480 So just give me a second to make sure 00:02:55.480 --> 00:03:01.585 I pull up the right set of slides. 00:03:08.031 --> 00:03:24.880 I think that-- all right, so two readings today. 00:03:24.880 --> 00:03:27.980 I don't know if you were able to take a look at them. 00:03:27.980 --> 00:03:31.240 I hope you had the time to take a look at them. 00:03:31.240 --> 00:03:33.400 And these were really just a chance 00:03:33.400 --> 00:03:38.710 to sort of grab hold of a broad discussion of what's going on. 00:03:38.710 --> 00:03:41.080 Again, I went to the Financial Stability Board. 00:03:41.080 --> 00:03:44.220 It's a little dated because it's 2017, 00:03:44.220 --> 00:03:46.220 but I thought the executive summaries 00:03:46.220 --> 00:03:48.530 and these various sections were helpful. 00:03:48.530 --> 00:03:52.750 And then a shorter medium post really on six examples, 00:03:52.750 --> 00:03:55.210 and today, we're going to talk about 00:03:55.210 --> 00:03:58.780 what is artificial intelligence, what is machine learning 00:03:58.780 --> 00:04:03.820 and deep learning, and then what are the eight or 10 major areas 00:04:03.820 --> 00:04:06.505 in finance that it's being used. 00:04:06.505 --> 00:04:07.880 And then next Monday, we're going 00:04:07.880 --> 00:04:11.350 to talk about a lot of the challenges 00:04:11.350 --> 00:04:16.450 and go a little deeper with regard to this. 00:04:16.450 --> 00:04:18.910 Now, I said that I had study questions. 00:04:18.910 --> 00:04:21.149 We're going to see if we can get this to work, 00:04:21.149 --> 00:04:23.620 and I'm going to ask for volunteers to speak up. 00:04:23.620 --> 00:04:27.230 If not, Romain might just unmute everybody, 00:04:27.230 --> 00:04:31.690 and then we'll see who I can cold call on or something. 00:04:31.690 --> 00:04:34.430 But who would like to take a crack? 00:04:34.430 --> 00:04:38.140 And remember, we're all pass emergency and no credit 00:04:38.140 --> 00:04:40.180 emergency, so this is just about trying 00:04:40.180 --> 00:04:42.340 to get the conversation going. 00:04:42.340 --> 00:04:44.410 But who would want to answer the question, what 00:04:44.410 --> 00:04:47.300 is artificial intelligence, machine learning, deep 00:04:47.300 --> 00:04:47.800 learning? 00:04:47.800 --> 00:04:50.290 And you don't need to be a computer scientist, 00:04:50.290 --> 00:04:52.570 but these terms are really important 00:04:52.570 --> 00:04:55.450 if you're going to be an entrepreneur and do a startup, 00:04:55.450 --> 00:04:58.030 or if you're going to be in a big incumbent or big tech 00:04:58.030 --> 00:05:01.690 company, just to have this conceptual framework 00:05:01.690 --> 00:05:05.963 and understanding of artificial intelligence, machine 00:05:05.963 --> 00:05:07.130 learning, and deep learning. 00:05:07.130 --> 00:05:10.690 So I'm going to wait for Romain to either find 00:05:10.690 --> 00:05:14.080 somebody who's raised their blue hand, or, Romain, you 00:05:14.080 --> 00:05:17.440 get to help me cold call if you wish. 00:05:17.440 --> 00:05:21.970 But hopefully, somebody wants to just dive into this. 00:05:21.970 --> 00:05:24.550 ROMAIN: I'm still waiting let's see who will be-- ah, 00:05:24.550 --> 00:05:26.200 we have our first volunteer of the day. 00:05:26.200 --> 00:05:28.493 Thank you, Michael. 00:05:28.493 --> 00:05:29.410 GARY GENSLER: Michael. 00:05:32.530 --> 00:05:35.140 STUDENT: Sorry, I forgot to unmute. 00:05:35.140 --> 00:05:39.970 So my understanding-- artificial intelligence just is more 00:05:39.970 --> 00:05:43.980 of an over encompassing term, being-- 00:05:43.980 --> 00:05:47.130 just computer is kind of mimicking human behavior 00:05:47.130 --> 00:05:50.327 and thought, so that's more-- 00:05:50.327 --> 00:05:51.660 GARY GENSLER: Let's pause there. 00:05:51.660 --> 00:05:55.540 And a very good answer, so computers 00:05:55.540 --> 00:05:57.490 mimicking human behavior. 00:05:57.490 --> 00:05:58.750 And, Michael, just a sense-- 00:05:58.750 --> 00:06:01.000 do you have a sense of how-- when did this come about? 00:06:01.000 --> 00:06:04.720 Was this in the last five years, or was it 00:06:04.720 --> 00:06:08.050 a longer time ago that somebody came up with this term 00:06:08.050 --> 00:06:10.000 "artificial intelligence?" 00:06:10.000 --> 00:06:14.210 STUDENT: More like the early to mid 1900s. 00:06:14.210 --> 00:06:15.340 It's been a while. 00:06:15.340 --> 00:06:16.798 GARY GENSLER: So it's been a while. 00:06:16.798 --> 00:06:18.820 It's actually way back in the 1950s, 00:06:18.820 --> 00:06:21.610 "artificial intelligence--" the concept 00:06:21.610 --> 00:06:24.580 of a computer mimicking humans. 00:06:24.580 --> 00:06:29.140 In fact, at MIT, we have the Computer Science and Artificial 00:06:29.140 --> 00:06:31.210 Intelligence Lab. 00:06:31.210 --> 00:06:34.950 It's not a creature that was just invented 00:06:34.950 --> 00:06:36.940 in the 20-teens or 20-naughts. 00:06:36.940 --> 00:06:38.680 It goes back decades. 00:06:38.680 --> 00:06:40.760 It was a merger of two earlier labs, 00:06:40.760 --> 00:06:43.420 but we've had an artificial intelligence lab at MIT 00:06:43.420 --> 00:06:46.810 for multiple decades. 00:06:46.810 --> 00:06:50.880 So who wants to say what machine learning is? 00:06:50.880 --> 00:06:51.500 Romain? 00:06:51.500 --> 00:06:53.018 I hand it off you to find another-- 00:06:53.018 --> 00:06:55.060 ROMAIN: Who will be the next volunteer for today? 00:06:58.870 --> 00:07:00.670 STUDENT: I think the machine [INAUDIBLE],, 00:07:00.670 --> 00:07:05.860 there's very limited or even no intervention with human, 00:07:05.860 --> 00:07:09.370 and learning are you solve problems step 00:07:09.370 --> 00:07:11.680 by step in a sequential manner. 00:07:11.680 --> 00:07:13.960 So that is like the machine solve 00:07:13.960 --> 00:07:18.185 problems step by step without much intervention by humans. 00:07:18.185 --> 00:07:19.810 GARY GENSLER: So I think you're raising 00:07:19.810 --> 00:07:22.320 two points, is the machine solving 00:07:22.320 --> 00:07:25.510 a problem without intervention, which is good. 00:07:25.510 --> 00:07:27.550 And then you said briefly, learning. 00:07:27.550 --> 00:07:29.020 Do you want to say something more, 00:07:29.020 --> 00:07:30.728 or anybody want to say, what does it mean 00:07:30.728 --> 00:07:32.455 that the machine is learning? 00:07:39.090 --> 00:07:41.892 ROMAIN: We have Rakan who raised his hand. 00:07:41.892 --> 00:07:43.350 STUDENT: Yes, well, essentially, it 00:07:43.350 --> 00:07:48.330 means that you're feeding the machine data, 00:07:48.330 --> 00:07:50.415 and as you feed the machine data, 00:07:50.415 --> 00:07:53.370 the machine learns not to do specific tasks. 00:07:53.370 --> 00:07:55.770 And you get better results as you feed 00:07:55.770 --> 00:07:58.050 it more data and more and more. 00:07:58.050 --> 00:08:00.870 GARY GENSLER: All right, so the concept of machine learning, 00:08:00.870 --> 00:08:02.670 again, is not that new. 00:08:02.670 --> 00:08:07.200 It was first written about in the 1980s and 1990s. 00:08:07.200 --> 00:08:08.740 It had a little bit of a wave-- 00:08:08.740 --> 00:08:12.590 you might call it a boomlet, a little bit of a hype cycle-- 00:08:12.590 --> 00:08:14.660 and then it sort of tamped down. 00:08:14.660 --> 00:08:20.170 But the conceptual framework is that the computer, machines, 00:08:20.170 --> 00:08:25.250 with data, are actually learning, that they actually-- 00:08:25.250 --> 00:08:29.270 whatever algorithms or decision making or pattern recognition 00:08:29.270 --> 00:08:31.730 that they have gets better. 00:08:31.730 --> 00:08:35.490 So machine learning is a subset of artificial intelligence. 00:08:35.490 --> 00:08:37.520 Artificial intelligence is this concept 00:08:37.520 --> 00:08:41.380 of computers, a form of machines, 00:08:41.380 --> 00:08:45.790 mimicking human intelligence, but it could mimic it 00:08:45.790 --> 00:08:46.870 without learning. 00:08:46.870 --> 00:08:51.460 It could just replicate or automate what we are doing. 00:08:51.460 --> 00:08:54.960 Machine learning was a new concept-- 00:08:54.960 --> 00:08:56.040 it didn't take off. 00:08:56.040 --> 00:08:59.250 It didn't dramatically change our lives at first-- 00:08:59.250 --> 00:09:02.670 where, actually, the computers could adapt 00:09:02.670 --> 00:09:06.960 and change based upon analysis of the data, 00:09:06.960 --> 00:09:09.900 and that their algorithms and their pattern recognition 00:09:09.900 --> 00:09:11.830 could shift. 00:09:11.830 --> 00:09:14.430 Does anybody want to take a crack at deep learning? 00:09:20.830 --> 00:09:22.112 ROMAIN: Anyone? 00:09:22.112 --> 00:09:23.070 Albert raised his hand. 00:09:25.590 --> 00:09:29.670 STUDENT: So deep learning involves 00:09:29.670 --> 00:09:34.170 using large neural networks to perform machine learning, 00:09:34.170 --> 00:09:36.120 so it's sort of a subset of machine learning. 00:09:36.120 --> 00:09:39.127 But it can be very powerful, and most of the time, 00:09:39.127 --> 00:09:41.460 it just gets better and better as you feed it more data. 00:09:41.460 --> 00:09:42.835 Other machine learning algorithms 00:09:42.835 --> 00:09:46.680 tend to have sort of a plateau and can't improve no matter 00:09:46.680 --> 00:09:48.690 how much data you feed them. 00:09:48.690 --> 00:09:53.700 GARY GENSLER: So important concept there in a phrase 00:09:53.700 --> 00:09:57.510 was "neural networks." 00:09:57.510 --> 00:09:58.890 Think of our brain's-- 00:09:58.890 --> 00:10:01.290 neural, neurology. 00:10:01.290 --> 00:10:04.380 It's about our brains. 00:10:04.380 --> 00:10:06.360 Early computer scientists started 00:10:06.360 --> 00:10:09.780 to think can we learn something from how the human brain 00:10:09.780 --> 00:10:16.560 works, which is in essence, a biological computer that takes 00:10:16.560 --> 00:10:21.120 electrical pulses and stores data, analyzes data, 00:10:21.120 --> 00:10:23.760 recognizes patterns. 00:10:23.760 --> 00:10:25.830 Even all of us right now are recognizing 00:10:25.830 --> 00:10:27.960 voice and visual patterns. 00:10:27.960 --> 00:10:29.960 That's in our human brain. 00:10:29.960 --> 00:10:33.000 So when looking at the brain, the conceptual framework 00:10:33.000 --> 00:10:36.000 is could we build a network similar to the brain, 00:10:36.000 --> 00:10:40.380 and thus, using these words, neural networks. 00:10:40.380 --> 00:10:43.560 In deep learning and machine learning, 00:10:43.560 --> 00:10:45.930 there's pattern recognition, but deep learning 00:10:45.930 --> 00:10:48.990 is a subset of machine learning that has 00:10:48.990 --> 00:10:53.440 multiple layers of connections. 00:10:53.440 --> 00:10:58.950 And in machine learning, you can take a base layer of data 00:10:58.950 --> 00:11:02.190 and try to find patterns, but deep learning 00:11:02.190 --> 00:11:04.290 finds patterns in the patterns. 00:11:04.290 --> 00:11:08.760 And you can think of it as putting it through layers. 00:11:08.760 --> 00:11:12.120 Now, if you're in Sloan, and you're deeply involved 00:11:12.120 --> 00:11:17.340 in computer science, and you also enjoy the topic, 00:11:17.340 --> 00:11:18.780 you can go much further. 00:11:18.780 --> 00:11:22.530 But in this class, we're not trying to go there. 00:11:22.530 --> 00:11:28.530 The importance of deep learning is that it can find and extract 00:11:28.530 --> 00:11:32.490 patterns even better than machine learning, 00:11:32.490 --> 00:11:36.960 but it takes more computational power and often more data. 00:11:36.960 --> 00:11:41.460 Deep learning is more an innovation of the 20-teens, 00:11:41.460 --> 00:11:46.390 and by 2011 and 2012, it really started to change things. 00:11:46.390 --> 00:11:48.570 And in the last eight years, we've 00:11:48.570 --> 00:11:51.590 seen dramatic advancements even in deep learning. 00:11:51.590 --> 00:11:56.070 It's a conceptual framework of taking a pool of data, 00:11:56.070 --> 00:11:59.630 looking at its patterns, and going a little higher. 00:11:59.630 --> 00:12:01.760 I'm going to talk about an example a little bit 00:12:01.760 --> 00:12:04.850 later in this discussion, but please 00:12:04.850 --> 00:12:11.760 bring me back to talk about deep learning in facial recognition, 00:12:11.760 --> 00:12:14.730 deep learning in autonomous vehicles 00:12:14.730 --> 00:12:16.230 and just thinking about it there, 00:12:16.230 --> 00:12:18.000 and then we'll pull it back to finance. 00:12:18.000 --> 00:12:18.992 ROMAIN: Gary? 00:12:18.992 --> 00:12:20.200 GARY GENSLER: Yes, questions? 00:12:20.200 --> 00:12:23.050 ROMAIN: We have two questions-- one from Rosana, who's asking, 00:12:23.050 --> 00:12:25.230 what is the difference between representation 00:12:25.230 --> 00:12:27.150 learning and deep learning? 00:12:27.150 --> 00:12:29.430 And then we'll give the floor to Pablo. 00:12:29.430 --> 00:12:30.930 GARY GENSLER: So very good question. 00:12:30.930 --> 00:12:33.510 Representation learning, you can think almost as 00:12:33.510 --> 00:12:35.580 in between machine learning and deep learning. 00:12:35.580 --> 00:12:41.010 If we said, this is almost like those Russian dolls. 00:12:41.010 --> 00:12:43.770 The AI is the big vessel. 00:12:43.770 --> 00:12:46.890 Machine learning is a subset, and deep learning 00:12:46.890 --> 00:12:48.420 was a subset of machine learning. 00:12:48.420 --> 00:12:52.320 Representational learning is a subset, in this context, 00:12:52.320 --> 00:12:58.560 of machine learning, and it's basically extracting features. 00:12:58.560 --> 00:13:01.830 So think of machine learning that's looking for patterns. 00:13:01.830 --> 00:13:05.340 It's extracting features out of data. 00:13:05.340 --> 00:13:11.130 So photo recognition, our standard Facebook photo 00:13:11.130 --> 00:13:15.750 recognition that might recognize Kelly versus Romain 00:13:15.750 --> 00:13:19.910 is extracting certain features of Kelly's face. 00:13:19.910 --> 00:13:21.530 I'm sorry if I'm picking on you Kelly. 00:13:21.530 --> 00:13:24.610 You just happen to be on my screen. 00:13:24.610 --> 00:13:27.650 But it's extracting features, so some people 00:13:27.650 --> 00:13:31.820 call it feature learning or representational learning, 00:13:31.820 --> 00:13:36.570 but extracting features as opposed to specific data. 00:13:36.570 --> 00:13:38.390 And then deep learning is generally 00:13:38.390 --> 00:13:40.670 thought of as a subset even though there 00:13:40.670 --> 00:13:46.150 is some debate as to how you would categorize these. 00:13:46.150 --> 00:13:48.920 And you said there was another question? 00:13:48.920 --> 00:13:49.920 STUDENT: Yeah, hi, Gary. 00:13:49.920 --> 00:13:52.070 This is Pablo. 00:13:52.070 --> 00:13:54.490 So I just wanted to verify, because my understanding 00:13:54.490 --> 00:13:57.220 is that, also, one of the big differences between machine 00:13:57.220 --> 00:13:59.920 learning and deep learning, or artificial intelligence 00:13:59.920 --> 00:14:01.390 in general besides deep learning, 00:14:01.390 --> 00:14:04.300 is that deep learning used unstructured data, 00:14:04.300 --> 00:14:06.610 whereas typically, for machine learning models, 00:14:06.610 --> 00:14:10.390 you need basically tables of data that you have organized 00:14:10.390 --> 00:14:12.645 and labeled and everything. 00:14:12.645 --> 00:14:14.020 And just the check whether that's 00:14:14.020 --> 00:14:18.640 the difference besides the additional complexity 00:14:18.640 --> 00:14:20.270 and multiple layers. 00:14:20.270 --> 00:14:23.940 GARY GENSLER: So let me step back and share 00:14:23.940 --> 00:14:26.620 the important vocabulary, and again this 00:14:26.620 --> 00:14:29.470 is important vocabulary well beyond being a computer 00:14:29.470 --> 00:14:30.150 scientist. 00:14:30.150 --> 00:14:33.400 It's important vocabulary if you're running a business, 00:14:33.400 --> 00:14:38.290 and you're trying to get the most out of data and your data 00:14:38.290 --> 00:14:39.710 analytics team. 00:14:39.710 --> 00:14:42.700 So I'm going to assume we're talking as if you are now going 00:14:42.700 --> 00:14:45.610 to be in a C-suite, and you want to get the most out 00:14:45.610 --> 00:14:48.340 of your data analytics teams and so forth-- 00:14:48.340 --> 00:14:53.620 is this concept of structured data versus unstructured data, 00:14:53.620 --> 00:14:57.490 and then I'll go to that specific question. 00:14:57.490 --> 00:15:06.360 Data that we see all the time with our eyes, that we read, 00:15:06.360 --> 00:15:08.200 that comes into us, you can think 00:15:08.200 --> 00:15:11.020 of as unstructured sometimes because it 00:15:11.020 --> 00:15:13.910 doesn't have a label on it. 00:15:13.910 --> 00:15:17.200 But if it has a label on it, all of a sudden, 00:15:17.200 --> 00:15:20.270 people call it structured data. 00:15:20.270 --> 00:15:22.300 So machine learning, conceptually, 00:15:22.300 --> 00:15:30.120 is that the machines are getting better at recognizing patterns. 00:15:30.120 --> 00:15:32.980 It's primarily about pattern recognition, 00:15:32.980 --> 00:15:37.990 that the machines are getting better recognizing patterns off 00:15:37.990 --> 00:15:39.110 of the data. 00:15:39.110 --> 00:15:41.410 And the question here is, is machine learning 00:15:41.410 --> 00:15:44.050 always structured learning? 00:15:44.050 --> 00:15:46.390 And structured learning means that the data 00:15:46.390 --> 00:15:49.510 is labeled, that you have a whole data set, 00:15:49.510 --> 00:15:52.420 and it's labeled. 00:15:52.420 --> 00:15:54.580 An example of labeling I will give you, 00:15:54.580 --> 00:15:58.720 that we all live with in our daily lives-- how many people-- 00:15:58.720 --> 00:16:01.390 just use the blue hands if you wish. 00:16:01.390 --> 00:16:06.190 How many people have ever been asked in a computer, 00:16:06.190 --> 00:16:08.830 will you please look at this picture 00:16:08.830 --> 00:16:13.210 and tell us whether there are any traffic 00:16:13.210 --> 00:16:14.243 lights in the picture? 00:16:14.243 --> 00:16:15.910 We want to make sure you're not a robot. 00:16:20.020 --> 00:16:23.680 All right, so here, I'm just going to-- 00:16:23.680 --> 00:16:26.290 Romain, just cold call somebody that put their hand up. 00:16:26.290 --> 00:16:28.510 I want to ask them a question. 00:16:28.510 --> 00:16:30.500 ROMAIN: Sophia, please? 00:16:30.500 --> 00:16:33.100 GARY GENSLER: All right, Sophia, are you are unmuted? 00:16:33.100 --> 00:16:34.160 STUDENT: Yes I'm on. 00:16:34.160 --> 00:16:36.327 GARY GENSLER: All right, so, Sophia, when you go in, 00:16:36.327 --> 00:16:38.270 and you tell the computer that you're 00:16:38.270 --> 00:16:43.160 not a machine, why do you think they ask you that question? 00:16:43.160 --> 00:16:45.380 STUDENT: To make sure that there aren't any bots who 00:16:45.380 --> 00:16:48.090 are trying to take advantage of any service 00:16:48.090 --> 00:16:50.840 that the machine is trying to offer. 00:16:50.840 --> 00:16:52.280 GARY GENSLER: Really good answer. 00:16:52.280 --> 00:16:55.550 You're correct, but you're not completely correct. 00:16:55.550 --> 00:16:57.200 What's the other reason that they're 00:16:57.200 --> 00:17:00.755 asking you whether that's a traffic light or not? 00:17:00.755 --> 00:17:02.130 STUDENT: And also to collect data 00:17:02.130 --> 00:17:04.270 so that they can use that data for labeling 00:17:04.270 --> 00:17:06.020 for future purposes as well. 00:17:06.020 --> 00:17:08.450 GARY GENSLER: So they are labeling data. 00:17:08.450 --> 00:17:14.310 They're using Sophia, if I might say, as free labor. 00:17:14.310 --> 00:17:20.940 Sophia is training the data that Google or whomever 00:17:20.940 --> 00:17:22.569 is putting together there. 00:17:22.569 --> 00:17:23.670 And thank you, Sofia. 00:17:23.670 --> 00:17:26.940 You're labeling that data so that autonomous vehicles 00:17:26.940 --> 00:17:28.680 will work better in the future. 00:17:28.680 --> 00:17:31.280 You're also frankly labeling data 00:17:31.280 --> 00:17:34.530 to put millions of people out of jobs as truckers, 00:17:34.530 --> 00:17:36.030 but I don't want you to get sort of 00:17:36.030 --> 00:17:41.040 wrapped up into those sort of social and public policy 00:17:41.040 --> 00:17:43.300 debates, but that's what's happening. 00:17:43.300 --> 00:17:45.270 So back to the earlier question. 00:17:45.270 --> 00:17:47.820 Data can be labeled by us. 00:17:47.820 --> 00:17:50.490 An earlier form of labeling was labeling 00:17:50.490 --> 00:17:53.220 what is an A, what is a B, what is a C, what 00:17:53.220 --> 00:17:56.280 are all the letters of the alphabet 00:17:56.280 --> 00:17:58.680 so that our postal services now can 00:17:58.680 --> 00:18:04.140 use a form of machine learning to read all of our written 00:18:04.140 --> 00:18:06.090 scratch on envelopes. 00:18:06.090 --> 00:18:08.520 If we address an envelope, it can all 00:18:08.520 --> 00:18:11.490 be read by computers rather than humans. 00:18:11.490 --> 00:18:12.750 To your earlier question-- 00:18:12.750 --> 00:18:16.230 I went a long way around this, but machine learning 00:18:16.230 --> 00:18:19.740 can be both unstructured and structured. 00:18:19.740 --> 00:18:23.100 The question was, is machine learning always labeled? 00:18:23.100 --> 00:18:25.290 Is machine learning always structured? 00:18:25.290 --> 00:18:26.880 And the answer's no. 00:18:26.880 --> 00:18:31.670 Machine learning can also be unstructured and unlabeled. 00:18:31.670 --> 00:18:34.640 Deep learning can be both labeled, which 00:18:34.640 --> 00:18:38.230 is structured, or unlabeled. 00:18:38.230 --> 00:18:41.350 Some of the economics and some of the computer science 00:18:41.350 --> 00:18:43.570 are worthwhile to understand. 00:18:43.570 --> 00:18:47.530 Labeled data can be trained faster. 00:18:47.530 --> 00:18:54.580 Label data can, in many regards, lower your error rates faster 00:18:54.580 --> 00:18:56.620 and have a certain-- 00:18:56.620 --> 00:19:02.800 extract correlations better, but it comes with a cost. 00:19:02.800 --> 00:19:05.170 You need to label the data, and so there's 00:19:05.170 --> 00:19:10.900 some tradeoff of getting Sophia to label data or other humans 00:19:10.900 --> 00:19:16.660 to label data versus unlabeled data. 00:19:16.660 --> 00:19:20.320 Think about radiology in the practice of medicine, 00:19:20.320 --> 00:19:25.030 and looking at body scans or mammograms 00:19:25.030 --> 00:19:28.960 or any form of radiology to identify 00:19:28.960 --> 00:19:30.910 whether there's an anomaly. 00:19:30.910 --> 00:19:34.330 There's something that needs to have further investigation 00:19:34.330 --> 00:19:38.890 to see whether it's a tumor or not. 00:19:38.890 --> 00:19:42.130 Radiology is dramatically changing in the last three 00:19:42.130 --> 00:19:46.040 or five years based upon machine learning and deep learning. 00:19:46.040 --> 00:19:47.470 Remember, deep learning just means 00:19:47.470 --> 00:19:50.640 there's multiple layers of pattern recognition 00:19:50.640 --> 00:19:52.810 in these neural networks. 00:19:52.810 --> 00:19:57.520 Labeled radiology, labeled mammograms, or labeled 00:19:57.520 --> 00:20:04.730 MRIs will train the machines faster, 00:20:04.730 --> 00:20:11.170 but unstructured, unlabeled data can also be used. 00:20:11.170 --> 00:20:13.930 You need bigger data sets. 00:20:13.930 --> 00:20:15.550 So I went I went off a little bit, 00:20:15.550 --> 00:20:17.050 but I hope that that's helpful. 00:20:19.580 --> 00:20:21.080 Other questions, Romain? 00:20:21.080 --> 00:20:22.910 ROMAIN: Yes, we have one from Victor. 00:20:22.910 --> 00:20:24.164 GARY GENSLER: Please. 00:20:24.164 --> 00:20:25.460 STUDENT: Hi, professor. 00:20:25.460 --> 00:20:27.746 I just wanted to double click on the-- 00:20:27.746 --> 00:20:30.530 take a step back in the initial definitions between machine 00:20:30.530 --> 00:20:32.780 learning and deep learning, because we were discussing 00:20:32.780 --> 00:20:35.960 that deep learning had the feature that it 00:20:35.960 --> 00:20:38.970 keeps learning despite the data growing exponentially. 00:20:38.970 --> 00:20:40.310 It doesn't plateau. 00:20:40.310 --> 00:20:44.900 But I don't fully understood the differentiation between two 00:20:44.900 --> 00:20:46.670 concepts beyond that. 00:20:46.670 --> 00:20:50.930 GARY GENSLER: So I didn't disagree or agree 00:20:50.930 --> 00:20:53.030 with that comment, and I apologize, 00:20:53.030 --> 00:20:55.790 I can't remember who said that deep learning keeps 00:20:55.790 --> 00:20:57.770 to grow exponentially. 00:20:57.770 --> 00:21:02.000 I think both machine learning and deep learning-- 00:21:02.000 --> 00:21:05.720 both machine learning and deep learning learn from data, 00:21:05.720 --> 00:21:10.943 and this word "learn" should be explored a little bit more. 00:21:10.943 --> 00:21:12.860 What machine learning and deep learning can do 00:21:12.860 --> 00:21:15.770 is extract correlations. 00:21:15.770 --> 00:21:18.200 I hope nearly everybody in this class 00:21:18.200 --> 00:21:20.150 has taken some form of statistics 00:21:20.150 --> 00:21:21.510 at some point in time. 00:21:21.510 --> 00:21:25.270 You might have hated statistics, but we all took it, 00:21:25.270 --> 00:21:27.530 and some of us took more advanced statistics 00:21:27.530 --> 00:21:30.420 where you use linear algebra and the like. 00:21:30.420 --> 00:21:34.450 But just thinking about a standard regression analysis-- 00:21:34.450 --> 00:21:38.420 a standard regression analysis finds a pattern, 00:21:38.420 --> 00:21:42.170 generally a linear pattern, or a quadratic pattern 00:21:42.170 --> 00:21:45.200 if you move on. 00:21:45.200 --> 00:21:47.810 Machine learning and deep learning find pattern, 00:21:47.810 --> 00:21:52.850 and they're really remarkable tools to extract correlations. 00:21:52.850 --> 00:21:55.370 And one of the features of both machine learning 00:21:55.370 --> 00:21:58.130 and deep learning is they look at error rates, 00:21:58.130 --> 00:22:03.410 particularly versus data sets that have been labeled. 00:22:03.410 --> 00:22:07.790 And traditionally, what you do is you have a big data set-- 00:22:07.790 --> 00:22:11.690 maybe it's millions of pieces of data that you're training on, 00:22:11.690 --> 00:22:14.300 and you take a random sample of it 00:22:14.300 --> 00:22:17.810 and put it to the side, a random sample on the side 00:22:17.810 --> 00:22:19.730 that you label. 00:22:19.730 --> 00:22:22.580 And then you compare what comes out 00:22:22.580 --> 00:22:27.740 of the machine learning with the test data on the side 00:22:27.740 --> 00:22:29.240 and see what's the error rate. 00:22:29.240 --> 00:22:32.960 And this labeled set on this side might say, these are men, 00:22:32.960 --> 00:22:35.600 these are women, this is a stoplight, 00:22:35.600 --> 00:22:40.700 this is a traffic light, whatever the labeled data is, 00:22:40.700 --> 00:22:45.620 and you see the predictive model, what's the error rate. 00:22:45.620 --> 00:22:48.020 Both machine learning and deep learning 00:22:48.020 --> 00:22:52.130 continue to do quite well. 00:22:52.130 --> 00:22:55.080 And I apologize I cannot remember who said it earlier, 00:22:55.080 --> 00:22:59.390 which student said, deep learning continues to grow 00:22:59.390 --> 00:23:01.830 further than machine learning. 00:23:01.830 --> 00:23:05.720 It can, but I wouldn't accept that machine learning can't get 00:23:05.720 --> 00:23:08.180 better and lower error rates. 00:23:08.180 --> 00:23:10.730 Now, once you get down to very low error rates, 00:23:10.730 --> 00:23:15.645 that's another circumstance altogether. 00:23:15.645 --> 00:23:18.020 The difference between deep learning and machine learning 00:23:18.020 --> 00:23:20.540 is that-- 00:23:20.540 --> 00:23:26.340 I'm going to use photo recognition software. 00:23:26.340 --> 00:23:29.850 If you put a photograph into a computer, where does it start? 00:23:29.850 --> 00:23:34.070 Does anybody-- what does it see at the very beginning, 00:23:34.070 --> 00:23:35.650 it's base level of data? 00:23:39.050 --> 00:23:41.550 STUDENT: The pixel? 00:23:41.550 --> 00:23:43.470 GARY GENSLER: What did I hear? 00:23:43.470 --> 00:23:44.720 STUDENT: I said, the pixels. 00:23:44.720 --> 00:23:45.950 GARY GENSLER: Pixels. 00:23:45.950 --> 00:23:48.700 So the only thing a computer can read is pixels. 00:23:48.700 --> 00:23:51.840 It has to start with the pixels and build up, 00:23:51.840 --> 00:23:54.400 and the next layer at most-- 00:23:54.400 --> 00:23:58.120 and again, I once pretended to know something 00:23:58.120 --> 00:23:59.420 about computer science. 00:23:59.420 --> 00:24:03.820 But I programmed in Fortran and APL years ago, 00:24:03.820 --> 00:24:07.440 and that was before many of you were born. 00:24:07.440 --> 00:24:11.860 But I guess I used to know how to program something. 00:24:11.860 --> 00:24:14.320 But if you read the pixels, the next layer 00:24:14.320 --> 00:24:17.080 up to find a pattern in the pixels 00:24:17.080 --> 00:24:20.850 is just small changes of shade, and then 00:24:20.850 --> 00:24:23.700 you can think of the next layer up from those little-- 00:24:23.700 --> 00:24:26.130 you can see edges. 00:24:26.130 --> 00:24:30.060 So the computer has to sort of go through layers 00:24:30.060 --> 00:24:34.620 from the pixels up to this is a traffic light versus a stop 00:24:34.620 --> 00:24:40.920 sign, and so deep learning interposes multiple layers 00:24:40.920 --> 00:24:44.000 of pattern recognition. 00:24:44.000 --> 00:24:46.430 Some would say that you need many layers, 00:24:46.430 --> 00:24:48.590 and then other research shows that, no, 00:24:48.590 --> 00:24:50.950 once you get to about three layers, 00:24:50.950 --> 00:24:55.250 there's less and less return on this. 00:24:55.250 --> 00:24:56.570 So let me sort of-- 00:24:56.570 --> 00:24:59.285 unless, Romain, is there other questions, or can I-- 00:24:59.285 --> 00:25:00.410 ROMAIN: No, we're all good. 00:25:00.410 --> 00:25:03.650 GARY GENSLER: All right, so again, this 00:25:03.650 --> 00:25:05.150 is just a broad thing, and we're not 00:25:05.150 --> 00:25:06.410 going to spend as much time. 00:25:06.410 --> 00:25:09.020 But what's natural language processing? 00:25:09.020 --> 00:25:11.550 Does anybody want to-- 00:25:11.550 --> 00:25:14.390 just what are these words broadly mean? 00:25:20.242 --> 00:25:21.200 ROMAIN: Any volunteers? 00:25:26.920 --> 00:25:29.500 GARY GENSLER: I'll volunteer, then. 00:25:29.500 --> 00:25:33.650 So natural language processing is just 00:25:33.650 --> 00:25:39.550 simply taking human language, natural language, 00:25:39.550 --> 00:25:43.630 and processing it down to computer language, 00:25:43.630 --> 00:25:47.170 all the way down to machine readable code, 00:25:47.170 --> 00:25:50.050 or going the other direction. 00:25:50.050 --> 00:25:53.490 So you can almost think of it as input and output 00:25:53.490 --> 00:25:59.080 to the computer, or we have many, many languages 00:25:59.080 --> 00:26:05.170 represented on this call right here with 99 participants. 00:26:05.170 --> 00:26:06.910 You can think about it as translating 00:26:06.910 --> 00:26:09.640 French to German and German to French, 00:26:09.640 --> 00:26:13.060 but instead it's natural language, what 00:26:13.060 --> 00:26:16.420 we do, down to the computer. 00:26:16.420 --> 00:26:19.680 And this is really important in terms of user interface 00:26:19.680 --> 00:26:25.050 and user experiences, and we'll get to that a little bit more. 00:26:25.050 --> 00:26:27.300 And then we're going to talk a lot about which sectors 00:26:27.300 --> 00:26:30.240 in financial services are being most affected 00:26:30.240 --> 00:26:31.220 at this point in time. 00:26:31.220 --> 00:26:35.100 So I won't call on the class right now because I want 00:26:35.100 --> 00:26:38.260 to keep moving to go forward. 00:26:38.260 --> 00:26:41.490 So we're in to talk about the financial world and fintech, 00:26:41.490 --> 00:26:43.480 and then-- oops, I didn't change this. 00:26:43.480 --> 00:26:46.680 This slide will-- I'll shift, because that's 00:26:46.680 --> 00:26:48.840 from the other day. 00:26:48.840 --> 00:26:51.360 So we looked at this slide the other day, 00:26:51.360 --> 00:26:53.100 and it just helps us-- 00:26:53.100 --> 00:26:54.960 what is AI machine learning? 00:26:54.960 --> 00:26:58.590 Extracting useful patterns from the data, 00:26:58.590 --> 00:27:00.870 using neural networks that we talked about, 00:27:00.870 --> 00:27:04.040 optimizing to lower error rates. 00:27:04.040 --> 00:27:09.930 Optimizing so that you actually say with 99% or 99 1/2% 00:27:09.930 --> 00:27:13.800 of the time, this is a traffic light, this is a stop sign. 00:27:13.800 --> 00:27:15.810 Actually optimizing. 00:27:15.810 --> 00:27:18.750 There's lots of programs that you can use. 00:27:18.750 --> 00:27:20.940 Google has TensorFlow, and-- 00:27:20.940 --> 00:27:23.070 I don't know. 00:27:23.070 --> 00:27:25.260 Have many of you ever used TensorFlow? 00:27:25.260 --> 00:27:27.210 I don't know all of your backgrounds, 00:27:27.210 --> 00:27:32.197 but any show blue hands? 00:27:32.197 --> 00:27:34.030 I'm not going to call on you to describe it. 00:27:34.030 --> 00:27:36.280 I'm just kind of curious if there's many people that 00:27:36.280 --> 00:27:40.240 have used TensorFlow or not. 00:27:40.240 --> 00:27:42.556 ROMAIN: How about we go with Devin? 00:27:42.556 --> 00:27:44.380 GARY GENSLER: So Devin's actually used it. 00:27:44.380 --> 00:27:45.880 I wasn't going to pick on them up, 00:27:45.880 --> 00:27:48.993 but if you wanted to say anything about it, Devin. 00:27:48.993 --> 00:27:50.410 STUDENT: Yeah, I can very briefly. 00:27:50.410 --> 00:27:54.310 So it essentially gives like a plug-and-play method 00:27:54.310 --> 00:27:58.858 to do machine learning and build neural nets in Python. 00:27:58.858 --> 00:28:01.150 You don't necessarily have to have a full understanding 00:28:01.150 --> 00:28:03.070 of how under the hood works. 00:28:03.070 --> 00:28:05.200 You can just add bits as you want, take bits away 00:28:05.200 --> 00:28:08.133 as you want, and it speeds up the whole process. 00:28:08.133 --> 00:28:10.300 GARY GENSLER: And so what's important about that is, 00:28:10.300 --> 00:28:15.040 just as somebody that came of age in the 1970s 00:28:15.040 --> 00:28:18.250 and '80s didn't have to learn how to computer code all 00:28:18.250 --> 00:28:20.800 the way down to machine readable code, 00:28:20.800 --> 00:28:25.000 they could learn how to use-- 00:28:25.000 --> 00:28:30.170 by the 1990s, C++, or C And C++, and later, Python. 00:28:30.170 --> 00:28:34.640 And many people in this class know how to use Python. 00:28:34.640 --> 00:28:36.740 In the machine learning area, there's 00:28:36.740 --> 00:28:41.360 been plug and play programs like TensorFlow, 00:28:41.360 --> 00:28:43.250 where you don't need to actually know 00:28:43.250 --> 00:28:47.290 how to build the data and things like that. 00:28:47.290 --> 00:28:50.080 Most importantly, and this is if you're 00:28:50.080 --> 00:28:53.380 thinking about being a data analyst 00:28:53.380 --> 00:28:55.930 or actually building a business around it, 00:28:55.930 --> 00:29:00.730 it's the data and the questions you train on the data. 00:29:00.730 --> 00:29:03.790 And most studies have shown, as of 2019, 00:29:03.790 --> 00:29:08.560 that 90%, 95% of the cost of data analytics 00:29:08.560 --> 00:29:11.930 is what some people might call cleaning up the data, 00:29:11.930 --> 00:29:13.660 making sure the data's well labeled. 00:29:13.660 --> 00:29:16.480 We talked about structured versus unstructured data 00:29:16.480 --> 00:29:17.440 earlier. 00:29:17.440 --> 00:29:21.070 Really important about that labeling, the cost. 00:29:21.070 --> 00:29:26.430 If you have 1,000 people working in a machine learning shop, 00:29:26.430 --> 00:29:30.420 or 500 people or five, it is quite likely 00:29:30.420 --> 00:29:36.480 that a big bulk of their time is standardizing the data, so 00:29:36.480 --> 00:29:39.840 to speak cleaning up the data, making sure that the fields are 00:29:39.840 --> 00:29:46.610 filled, and ensuring that you can then train on this data, 00:29:46.610 --> 00:29:51.507 meaning it's labeled, or enough of it's labeled. 00:29:51.507 --> 00:29:53.840 And then thinking about what the questions you're really 00:29:53.840 --> 00:29:57.230 trying to achieve, what you're trying to extract. 00:29:57.230 --> 00:29:58.490 Why is it happening now? 00:29:58.490 --> 00:30:00.980 This is off of Lex Freeman's slide again, but why? 00:30:00.980 --> 00:30:04.520 Because the hardware, the tools, the analytics-- a lot 00:30:04.520 --> 00:30:08.960 has shifted in the last five or eight years. 00:30:08.960 --> 00:30:12.710 To give you a sense of what it's being used for, all of these, 00:30:12.710 --> 00:30:13.430 we know. 00:30:13.430 --> 00:30:14.540 We know already. 00:30:14.540 --> 00:30:18.090 It's dramatically changing our lives. 00:30:18.090 --> 00:30:20.940 When we're sitting at home, sheltering at a home, 00:30:20.940 --> 00:30:22.680 and you're thinking about the next movie, 00:30:22.680 --> 00:30:25.050 and you're on Netflix, Netflix is telling us 00:30:25.050 --> 00:30:27.900 what they think the next thing we should watch. 00:30:27.900 --> 00:30:31.170 That's training off of not just the knowledge 00:30:31.170 --> 00:30:33.750 of what each of us has been watching, 00:30:33.750 --> 00:30:37.080 but it's about what others are watching. 00:30:37.080 --> 00:30:39.270 It's the Postal Service that no longer 00:30:39.270 --> 00:30:43.260 has to have a human reading the text, 00:30:43.260 --> 00:30:46.020 our scrawl on the envelope. 00:30:46.020 --> 00:30:50.070 It's Facebook with the facial recognition programs and so 00:30:50.070 --> 00:30:55.380 forth, and autonomous vehicles that 00:30:55.380 --> 00:30:57.570 are now being tested on the roads 00:30:57.570 --> 00:31:00.240 but are very likely part of our future. 00:31:00.240 --> 00:31:04.800 Now, will they be rolled out in a dramatic way in five years, 00:31:04.800 --> 00:31:06.360 or will it be 15 years? 00:31:06.360 --> 00:31:08.100 But I would feel comfortable that we 00:31:08.100 --> 00:31:11.670 will have autonomous vehicles on the road 00:31:11.670 --> 00:31:16.080 sometime at least by the 2030, but maybe others 00:31:16.080 --> 00:31:17.610 would be more optimistic. 00:31:17.610 --> 00:31:21.540 So it's changing a lot in many, many fields. 00:31:21.540 --> 00:31:26.370 The question is now, how is it shifting this field of finance? 00:31:26.370 --> 00:31:32.295 Why do I put it at the center of what we're doing here? 00:31:35.500 --> 00:31:37.065 So we talked about this. 00:31:37.065 --> 00:31:39.570 This was just my little attempt, and so now you 00:31:39.570 --> 00:31:43.080 have a slide that does what we chatted about before. 00:31:45.690 --> 00:31:49.895 One important thing also is happening is, in finance-- 00:31:49.895 --> 00:31:51.270 [CLEARING HIS THROAT] excuse me-- 00:31:51.270 --> 00:31:55.680 people are grabbing alternative data, using alternative data. 00:31:55.680 --> 00:31:59.490 I said earlier that the most important questions 00:31:59.490 --> 00:32:02.340 are what's the good data? 00:32:02.340 --> 00:32:06.870 So then we think about, in finance, what type of data 00:32:06.870 --> 00:32:09.480 do we want to grab? 00:32:09.480 --> 00:32:12.810 Data analytics in finance goes back centuries. 00:32:12.810 --> 00:32:14.730 That is not new. 00:32:14.730 --> 00:32:17.040 The Medicis, when they had to figure out 00:32:17.040 --> 00:32:21.300 to whom to land in Renaissance era 00:32:21.300 --> 00:32:24.840 had to figure out who is a good credit or not. 00:32:24.840 --> 00:32:28.110 And two data scientists from Stanford started a company 00:32:28.110 --> 00:32:29.790 called Fair Isaac-- 00:32:29.790 --> 00:32:32.940 those were their last names, Fair and Isaac-- 00:32:32.940 --> 00:32:36.030 and that became the FICO company, the Fair Isaac 00:32:36.030 --> 00:32:37.670 Company. 00:32:37.670 --> 00:32:41.530 So the data analytics in finance and the consumer side 00:32:41.530 --> 00:32:45.540 has certainly been around since the 1950s and 1960s, 00:32:45.540 --> 00:32:47.470 but where we are now is to say what 00:32:47.470 --> 00:32:50.380 is the additional types of data that we might take, 00:32:50.380 --> 00:32:54.970 and not just banking and checking and so forth? 00:32:54.970 --> 00:32:59.540 But Alibaba can look at a company, look at a company 00:32:59.540 --> 00:33:03.440 very closely, and do a full cash flow underwriting. 00:33:03.440 --> 00:33:06.740 Alibaba, because they have AliPay, 00:33:06.740 --> 00:33:10.550 can see what that small business is spending 00:33:10.550 --> 00:33:12.500 and what that small business is receiving. 00:33:12.500 --> 00:33:15.440 Amazon Prime can't quite do it as much, 00:33:15.440 --> 00:33:18.570 but they can do a bit of it as well. 00:33:18.570 --> 00:33:21.150 And even Toast, which is a fintech company 00:33:21.150 --> 00:33:25.440 in the restaurant business until this Corona shut down, 00:33:25.440 --> 00:33:27.600 Toast could see a lot about what's 00:33:27.600 --> 00:33:31.590 happening restaurant by restaurant in their cash flows. 00:33:31.590 --> 00:33:37.600 They had the revenue side more than the expenditure side. 00:33:37.600 --> 00:33:41.700 Alibaba, much better data sets than Toast, 00:33:41.700 --> 00:33:47.840 but I wouldn't put at rest a Toast started earlier 00:33:47.840 --> 00:33:53.120 in the last year to do credit extension to the restaurant 00:33:53.120 --> 00:33:54.470 business. 00:33:54.470 --> 00:33:56.990 Now, do you need deep learning and machine learning 00:33:56.990 --> 00:33:57.680 to do that? 00:33:57.680 --> 00:33:58.940 Not necessarily. 00:33:58.940 --> 00:34:01.790 You can still use plain old regression 00:34:01.790 --> 00:34:07.160 and linear statistics, linear regression analysis, 00:34:07.160 --> 00:34:12.975 but machine learning and deep learning help you go further. 00:34:12.975 --> 00:34:14.600 And then there's, of course, everything 00:34:14.600 --> 00:34:18.350 about our usage, our browser history, our email receipts, 00:34:18.350 --> 00:34:19.429 and so forth. 00:34:19.429 --> 00:34:22.310 If we look at China, they've stood up 00:34:22.310 --> 00:34:28.670 a broader social credit system, and in that system, 00:34:28.670 --> 00:34:32.989 they can tap into data about users 00:34:32.989 --> 00:34:34.370 in many different platforms. 00:34:34.370 --> 00:34:36.460 Romain, do I see-- are you waving at me? 00:34:36.460 --> 00:34:38.406 Is there a question? 00:34:38.406 --> 00:34:39.239 ROMAIN: No, I'm not. 00:34:39.239 --> 00:34:40.440 Sorry for that. 00:34:40.440 --> 00:34:41.732 GARY GENSLER: That's all right. 00:34:43.510 --> 00:34:47.080 So natural language processing, I mentioned. 00:34:47.080 --> 00:34:50.139 I just want to say a few more words about it. 00:34:50.139 --> 00:34:55.120 Think of it as computers input and output interpretation. 00:34:55.120 --> 00:34:58.430 This sort of going from German to French, 00:34:58.430 --> 00:35:03.760 or going from computer language to human language and back 00:35:03.760 --> 00:35:04.780 again. 00:35:04.780 --> 00:35:08.710 That's this important back and forth, 00:35:08.710 --> 00:35:10.960 and so it's natural language understanding, 00:35:10.960 --> 00:35:14.500 meaning a computer understands something, and also 00:35:14.500 --> 00:35:17.570 natural language generation. 00:35:17.570 --> 00:35:20.260 So it can be audio, image, video, 00:35:20.260 --> 00:35:24.590 any form of communication-- even a gesture. 00:35:24.590 --> 00:35:27.480 This hand wave can be interpreted-- 00:35:27.480 --> 00:35:32.220 if not now in 2020, within a few years will be interpreted. 00:35:32.220 --> 00:35:37.210 A movement of your face will be interpreted as well. 00:35:39.840 --> 00:35:44.340 And so how it's being used is really quite interesting, 00:35:44.340 --> 00:35:48.740 but we all know about chat bots and voice assistance already. 00:35:48.740 --> 00:35:52.250 That's shifting our worlds. 00:35:52.250 --> 00:35:52.790 Yeah? 00:35:52.790 --> 00:35:54.650 ROMAIN: We have a question from Nadia. 00:35:54.650 --> 00:35:56.600 GARY GENSLER: Nadia, please. 00:35:56.600 --> 00:35:59.030 STUDENT: I have one question related to the chat bots. 00:35:59.030 --> 00:36:00.470 What kind of factors do you think 00:36:00.470 --> 00:36:02.300 will encourage people to use chat bots? 00:36:02.300 --> 00:36:04.910 Because now, I do think people prefer 00:36:04.910 --> 00:36:09.280 to talk to a person rather than chat bots. 00:36:09.280 --> 00:36:12.700 GARY GENSLER: Well, I think, Nadia, 00:36:12.700 --> 00:36:15.280 we might still prefer to talk to a person, 00:36:15.280 --> 00:36:20.950 but there's a certain efficiency in-- 00:36:20.950 --> 00:36:23.860 and I'm just going to stay in finance for a minute. 00:36:23.860 --> 00:36:29.750 But there's a certain efficiency that financial service firms 00:36:29.750 --> 00:36:34.010 find that they can use chat bots instead of putting 00:36:34.010 --> 00:36:35.250 a human on the phone. 00:36:35.250 --> 00:36:38.930 So even when you and I call up to a Bank of America, 00:36:38.930 --> 00:36:43.800 and we want to check in on something on our credit cards, 00:36:43.800 --> 00:36:47.720 we're put through a various series of push one 00:36:47.720 --> 00:36:50.150 if you want this, push two if you want that, 00:36:50.150 --> 00:36:51.870 and we're pushing buttons. 00:36:51.870 --> 00:36:55.610 That's not high technology, by the way, 00:36:55.610 --> 00:36:58.820 but that's an efficiency that Bank of America 00:36:58.820 --> 00:37:02.060 has interposed into the system instead of having 00:37:02.060 --> 00:37:04.640 a call center of humans. 00:37:04.640 --> 00:37:08.750 And so if they can move from a cost center of humans 00:37:08.750 --> 00:37:14.050 to an automated call center of chat bots, 00:37:14.050 --> 00:37:19.150 they can provide services at a lower cost and to more people. 00:37:19.150 --> 00:37:22.930 Now, you and I might still want a human on the other side, 00:37:22.930 --> 00:37:28.710 but business is interposing an automation, 00:37:28.710 --> 00:37:36.050 and that automation means, often, quicker response time. 00:37:36.050 --> 00:37:39.890 So many of us, you go into a website today, 00:37:39.890 --> 00:37:44.870 and there's a little bot window that comes up. 00:37:44.870 --> 00:37:49.940 And the first thing that comes up on so many websites-- 00:37:49.940 --> 00:37:52.460 and this is true if it's a financial site. 00:37:52.460 --> 00:37:57.290 It's true if it's a commercial website where 00:37:57.290 --> 00:37:58.850 you're buying something online. 00:37:58.850 --> 00:38:00.920 It's probably true of dating websites, 00:38:00.920 --> 00:38:03.410 that there's some little bot window that comes up and says, 00:38:03.410 --> 00:38:05.060 can we help you. 00:38:05.060 --> 00:38:09.620 That's not a human, but it does give us greater service. 00:38:09.620 --> 00:38:12.920 It gives us an immediate recognition somebody's 00:38:12.920 --> 00:38:15.280 answering a question. 00:38:15.280 --> 00:38:17.530 Nadia, do I sense you'd prefer not to have 00:38:17.530 --> 00:38:18.970 the chat bots interposed? 00:38:26.410 --> 00:38:28.000 Is Nadia still there? 00:38:28.000 --> 00:38:29.760 STUDENT: Oh, yeah. 00:38:29.760 --> 00:38:31.810 Yeah, because I do think sometimes, 00:38:31.810 --> 00:38:34.420 chat bots, we ask a question, but their answer 00:38:34.420 --> 00:38:38.320 is not really related to our questions. 00:38:38.320 --> 00:38:40.720 GARY GENSLER: So you would prefer a human 00:38:40.720 --> 00:38:44.950 because you think the human will interpret your question 00:38:44.950 --> 00:38:46.920 and be able to answer it better? 00:38:46.920 --> 00:38:48.840 STUDENT: Yes. 00:38:48.840 --> 00:38:52.290 GARY GENSLER: So if the chat bot could answer as well 00:38:52.290 --> 00:38:53.700 as Kelly could answer-- 00:38:53.700 --> 00:38:54.840 again, I'm sorry, Kelly. 00:38:54.840 --> 00:38:56.280 You're on my screen. 00:38:56.280 --> 00:39:01.380 But if the chat bot could answer as well as Kelly or Camillo 00:39:01.380 --> 00:39:03.480 or others in this class could answer, 00:39:03.480 --> 00:39:05.590 you'd be all right with that? 00:39:05.590 --> 00:39:08.213 STUDENT: Yeah, it's faster. 00:39:08.213 --> 00:39:09.630 GARY GENSLER: See, if it's faster, 00:39:09.630 --> 00:39:11.320 and it can answer as well. 00:39:11.320 --> 00:39:15.084 So those are some of the commercial challenges. 00:39:15.084 --> 00:39:18.230 ROMAIN: I think Ivy would like to contribute as well. 00:39:18.230 --> 00:39:20.340 GARY GENSLER: Sure. 00:39:20.340 --> 00:39:24.660 STUDENT: Yeah, I just wanted to offer a little bit of some 00:39:24.660 --> 00:39:26.490 of the consumer studies that we did 00:39:26.490 --> 00:39:28.610 when I was working for a startup where 00:39:28.610 --> 00:39:30.180 we were building chat bot. 00:39:30.180 --> 00:39:34.230 And interesting enough, I think most people are actually 00:39:34.230 --> 00:39:34.740 pretty-- 00:39:34.740 --> 00:39:37.410 and we did a pretty large survey, 00:39:37.410 --> 00:39:40.490 and most people were pretty open to the idea 00:39:40.490 --> 00:39:42.240 of working with a chat bot because I think 00:39:42.240 --> 00:39:44.310 that's become so pervasive. 00:39:44.310 --> 00:39:47.040 But then there's this idea, they want 00:39:47.040 --> 00:39:50.490 to know that it is a chat bot, that the company is very 00:39:50.490 --> 00:39:52.890 transparent about that, because people change 00:39:52.890 --> 00:39:55.110 their behavior when they are speaking to the chat 00:39:55.110 --> 00:39:58.110 bot or some kind of virtual assistant, 00:39:58.110 --> 00:40:00.340 as long as they know, just to build that trust. 00:40:00.340 --> 00:40:02.370 And also, I guess we try to be more 00:40:02.370 --> 00:40:05.190 explicit in our wording, both verbally 00:40:05.190 --> 00:40:07.500 as well as written texts. 00:40:07.500 --> 00:40:09.660 And then secondly, I think because-- 00:40:09.660 --> 00:40:12.900 I guess I wanted to pose this as a question, too. 00:40:12.900 --> 00:40:16.300 When I think about chat bots and things like that, 00:40:16.300 --> 00:40:20.540 the technology is not necessarily there, 00:40:20.540 --> 00:40:22.920 or it takes a long time. 00:40:22.920 --> 00:40:28.110 And so I'm just curious what your thoughts are in terms of-- 00:40:28.110 --> 00:40:35.360 for me, I see it as like AI gets us 80% of the way there, 00:40:35.360 --> 00:40:38.210 but we need the human touch 20% of the way there. 00:40:38.210 --> 00:40:41.280 And so I actually see a lot of companies either having 00:40:41.280 --> 00:40:43.030 that human at the end of the-- 00:40:43.030 --> 00:40:46.260 you ultimately still need a human at the end of the day. 00:40:46.260 --> 00:40:48.825 So I mean, I just wanted to explore that a little. 00:40:48.825 --> 00:40:51.950 GARY GENSLER: I think what Nadia and Ivy are raising, 00:40:51.950 --> 00:40:53.900 and I'm sure we're all grappling with this. 00:40:53.900 --> 00:40:57.480 We are living in a very exciting time, 00:40:57.480 --> 00:41:01.460 and I'm not talking about this corona crisis. 00:41:01.460 --> 00:41:04.580 That's a different type of challenge. 00:41:04.580 --> 00:41:06.320 But we're living in an exciting time 00:41:06.320 --> 00:41:10.190 where we can automate a lot of things that humans have done. 00:41:10.190 --> 00:41:12.230 We've automated so many things that humans 00:41:12.230 --> 00:41:14.420 have done for centuries, but we're now 00:41:14.420 --> 00:41:19.070 automating this interface, through chat 00:41:19.070 --> 00:41:24.310 bots and conversational interfaces, voice assistance. 00:41:24.310 --> 00:41:28.550 I would dare say that, of the 90-plus people in this class 00:41:28.550 --> 00:41:32.570 now, that most of us, if not all of us, at some point in time, 00:41:32.570 --> 00:41:34.540 have used Siri. 00:41:34.540 --> 00:41:36.460 I mean, if we're driving along a highway, 00:41:36.460 --> 00:41:38.800 and we're supposed to be hands free, 00:41:38.800 --> 00:41:45.210 we might talk and start up an app or something legally, 00:41:45.210 --> 00:41:46.240 legally. 00:41:46.240 --> 00:41:51.270 And there's a lot of automation that's going on. 00:41:51.270 --> 00:41:56.280 How many of us have called to arrange a reservation 00:41:56.280 --> 00:41:58.140 at a restaurant, and we're not quite sure 00:41:58.140 --> 00:42:04.050 if we're talking to a human or a conversational agent? 00:42:04.050 --> 00:42:05.940 But I think that what Ivy's saying 00:42:05.940 --> 00:42:12.690 is that there might always need to be a human somewhere there. 00:42:12.690 --> 00:42:15.630 I don't know, Ivy, if that's correct. 00:42:15.630 --> 00:42:18.390 That's where we are in 2020. 00:42:18.390 --> 00:42:20.310 Let's think about autonomous vehicles. 00:42:20.310 --> 00:42:22.760 Right now, we're not comfortable enough. 00:42:22.760 --> 00:42:25.080 The manufacturers aren't comfortable enough. 00:42:25.080 --> 00:42:27.840 The computer scientist aren't comfortable enough. 00:42:27.840 --> 00:42:29.550 The regulators aren't comfortable enough. 00:42:29.550 --> 00:42:31.080 The public's not comfortable enough 00:42:31.080 --> 00:42:35.910 to have autonomous vehicles on the road with no humans 00:42:35.910 --> 00:42:41.280 whatsoever, but that's not really necessarily 00:42:41.280 --> 00:42:43.710 where we'll be in 2030. 00:42:43.710 --> 00:42:45.090 Or take radiology. 00:42:45.090 --> 00:42:48.930 Right now, at least in advanced economies 00:42:48.930 --> 00:42:51.840 like in Europe and the US and elsewhere, 00:42:51.840 --> 00:42:53.580 in advanced economies, we say we still 00:42:53.580 --> 00:42:57.100 want a doctor's eyes on a radiologist report. 00:42:57.100 --> 00:43:01.920 So the mammogram might be read by some artificial intelligence 00:43:01.920 --> 00:43:07.380 machine learning trained data, but we still have a human. 00:43:07.380 --> 00:43:11.370 But is that really the tradeoff we'll make in a few years? 00:43:11.370 --> 00:43:13.140 And is it the right tradeoff to be 00:43:13.140 --> 00:43:16.140 made in less developed countries, where 00:43:16.140 --> 00:43:19.170 they don't have the resources to have the doctors? 00:43:19.170 --> 00:43:22.440 And now, we even look in the middle of this crisis, 00:43:22.440 --> 00:43:24.870 the corona crisis, if-- 00:43:24.870 --> 00:43:27.300 this is sort of God willing. 00:43:27.300 --> 00:43:31.380 If the Baidus of China and the Googles of the US 00:43:31.380 --> 00:43:35.280 and others sharp analytic AI shops 00:43:35.280 --> 00:43:39.540 come up with a way to extract patterns and develop 00:43:39.540 --> 00:43:43.770 some recognition as to who's most vulnerable, 00:43:43.770 --> 00:43:46.140 are we going to rely on that, or are we 00:43:46.140 --> 00:43:49.380 going to say a human has to also interpret it and be involved 00:43:49.380 --> 00:43:50.820 in it? 00:43:50.820 --> 00:43:52.410 And I don't know. 00:43:52.410 --> 00:43:54.270 So I think we're at an exciting time 00:43:54.270 --> 00:43:57.150 where we're automating more and more. 00:43:57.150 --> 00:43:59.250 I do agree with you, Ivy, there's always 00:43:59.250 --> 00:44:00.540 going to be a role for humans. 00:44:00.540 --> 00:44:06.360 I'm not terribly worried that we'll all be put out of a job. 00:44:06.360 --> 00:44:10.130 200 years ago, our ancestors, all of our ancestors, 00:44:10.130 --> 00:44:12.155 were, by and large, working on farms. 00:44:15.420 --> 00:44:16.950 That's the economies. 00:44:16.950 --> 00:44:20.340 And we have found other things to fill those roles 00:44:20.340 --> 00:44:21.930 and those needs. 00:44:21.930 --> 00:44:28.370 I think we'll still have the humans, but not in every task. 00:44:32.190 --> 00:44:34.200 So let me go through a little bit-- 00:44:34.200 --> 00:44:36.000 so the Financial Stability Board, this is 00:44:36.000 --> 00:44:38.520 their definitions I'm going to pass on, 00:44:38.520 --> 00:44:40.890 but this was in that paper that you all 00:44:40.890 --> 00:44:43.950 read about what big data and machine learning was. 00:44:43.950 --> 00:44:46.560 When I show this page to computer scientists, 00:44:46.560 --> 00:44:50.970 when I show it to colleagues of mine at MIT 00:44:50.970 --> 00:44:54.375 from the College of Computing, they look at it, 00:44:54.375 --> 00:44:56.250 and they say, jeez, that's funny that a bunch 00:44:56.250 --> 00:44:59.010 of financial treasury secretaries 00:44:59.010 --> 00:45:01.680 and central bankers and their staffs 00:45:01.680 --> 00:45:03.930 define big data machine learning this way. 00:45:03.930 --> 00:45:06.090 So I partly put it up because this 00:45:06.090 --> 00:45:10.170 is kind of what the regulators define as to what it is. 00:45:10.170 --> 00:45:12.810 Machine learning may be defined as a method of designing 00:45:12.810 --> 00:45:16.200 sequence of actions to solve a problem, known as algorithms, 00:45:16.200 --> 00:45:18.090 and so forth. 00:45:18.090 --> 00:45:22.440 Computer scientists would name it a little bit differently. 00:45:22.440 --> 00:45:24.230 So I said to you the other day that I 00:45:24.230 --> 00:45:28.220 think of financial technology as history is building 00:45:28.220 --> 00:45:32.570 on these things, but machine learning and deep learning 00:45:32.570 --> 00:45:34.480 is at this top level. 00:45:34.480 --> 00:45:37.620 In the customer interface, it's the chat box 00:45:37.620 --> 00:45:41.960 we were just talking about, and on the risk management side, 00:45:41.960 --> 00:45:45.420 it's extracting patterns to make better risk decisions. 00:45:45.420 --> 00:45:48.770 So it's in these two broad fields, 00:45:48.770 --> 00:45:54.040 I sort of think of it is the customer interface 00:45:54.040 --> 00:45:57.970 and then lowering risk and extracting patterns. 00:45:57.970 --> 00:45:59.800 And sometimes, it's not just lowering risk. 00:45:59.800 --> 00:46:02.020 It's enhancing returns. 00:46:02.020 --> 00:46:04.600 And so I was going to go through and chat 00:46:04.600 --> 00:46:09.700 about each of these eight areas, and not 00:46:09.700 --> 00:46:12.190 all of the areas we're going to talk about are as robust. 00:46:12.190 --> 00:46:15.390 Asset management, right now-- 00:46:15.390 --> 00:46:18.480 asset management from hedge funds all the way 00:46:18.480 --> 00:46:21.990 to the BlackRock and Fidelity are exploring the use 00:46:21.990 --> 00:46:24.530 of machine learning and AI. 00:46:24.530 --> 00:46:27.710 By and large, most high frequency trading shops, 00:46:27.710 --> 00:46:31.520 most hedge funds today, most asset managers today, 00:46:31.520 --> 00:46:35.740 are not using much machine learning and AI. 00:46:35.740 --> 00:46:40.150 I view that as an opportunity in the 2020s. 00:46:40.150 --> 00:46:45.640 I view that as a real possibility of a shift, a very 00:46:45.640 --> 00:46:46.630 significant shift. 00:46:46.630 --> 00:46:49.180 But where is it being used so far? 00:46:49.180 --> 00:46:51.730 So BlackRock has been announcing and saying 00:46:51.730 --> 00:46:53.320 that they're already using BlackRock 00:46:53.320 --> 00:46:55.960 as one of the world's, if not the world's, largest 00:46:55.960 --> 00:46:57.610 asset manager. 00:46:57.610 --> 00:47:00.820 Before this corona crisis, probably $6 or $7 trillion 00:47:00.820 --> 00:47:01.940 of assets. 00:47:01.940 --> 00:47:05.470 It's, of course, a lower number now. 00:47:05.470 --> 00:47:06.970 And BlackRock and others have been 00:47:06.970 --> 00:47:11.470 saying we're using machine learning to actually listen 00:47:11.470 --> 00:47:16.930 to all of the audio files, all of the audio files 00:47:16.930 --> 00:47:18.910 of the major companies when they announce 00:47:18.910 --> 00:47:22.080 their quarterly earnings. 00:47:22.080 --> 00:47:29.210 And they're also putting in news articles, digital news 00:47:29.210 --> 00:47:32.330 or articles about those announcements, 00:47:32.330 --> 00:47:37.790 and also feeding in some of the actual financial statements 00:47:37.790 --> 00:47:39.920 that are released. 00:47:39.920 --> 00:47:44.480 And that takes a little bit of natural language processing. 00:47:44.480 --> 00:47:47.780 You need some form of taking the audio files, 00:47:47.780 --> 00:47:52.580 taking the written files, and interpreting that. 00:47:52.580 --> 00:47:55.490 But with that data they're looking for sentiment. 00:47:55.490 --> 00:47:59.510 They're trying to interpret the sentiments 00:47:59.510 --> 00:48:01.550 and see if the stock price are moving 00:48:01.550 --> 00:48:05.700 based on all of that data. 00:48:05.700 --> 00:48:08.790 Now, that's BlackRock with $6 or $7 trillion of assets. 00:48:08.790 --> 00:48:11.650 They're deeply resourced, fidelity, and so forth. 00:48:11.650 --> 00:48:13.530 But if you go down the value chain, 00:48:13.530 --> 00:48:15.890 if you go down to smaller asset managers, 00:48:15.890 --> 00:48:21.920 they're not doing a lot, I would say, yet. 00:48:21.920 --> 00:48:24.530 But there are hedge funds that are specifically saying, 00:48:24.530 --> 00:48:26.720 we are data analytic hedge funds. 00:48:26.720 --> 00:48:28.130 We want to move a little further. 00:48:28.130 --> 00:48:31.130 We want to try to use this machine learning because it's 00:48:31.130 --> 00:48:34.225 a better way to extract correlations, 00:48:34.225 --> 00:48:35.600 and that's what it's looking for. 00:48:35.600 --> 00:48:37.580 It's pattern recognition. 00:48:37.580 --> 00:48:40.790 I think that you're going to see more and more high 00:48:40.790 --> 00:48:44.510 frequency trading shops and hedge funds exploring this. 00:48:44.510 --> 00:48:47.030 But one conversation I had in the last couple of months 00:48:47.030 --> 00:48:48.800 with the high frequency trading shop-- 00:48:48.800 --> 00:48:51.170 and it was a shop that had about 100 employees. 00:48:51.170 --> 00:48:54.100 It was not big, but it was big enough. 00:48:54.100 --> 00:48:57.770 It was certainly making money in those days. 00:48:57.770 --> 00:49:00.620 I don't know how it's doing now. 00:49:00.620 --> 00:49:04.430 But they said, look, we feel pretty good about what we do, 00:49:04.430 --> 00:49:09.500 and when we look for algorithms, when we look-- 00:49:09.500 --> 00:49:14.270 our algorithmic trading doesn't need all of that expenditure 00:49:14.270 --> 00:49:17.300 and all that resource intensive thing of machine 00:49:17.300 --> 00:49:21.180 learning and cleaning up the data and finding the patterns. 00:49:21.180 --> 00:49:26.930 And in fact, we think that is not flexible enough yet for us. 00:49:26.930 --> 00:49:30.950 We're looking for short term opportunities, 00:49:30.950 --> 00:49:35.110 and we think that regression analysis and our classic linear 00:49:35.110 --> 00:49:39.000 and correlation analyses are enough at this moment. 00:49:39.000 --> 00:49:41.480 What's going to come in 2023 is a different thing, 00:49:41.480 --> 00:49:45.290 but what they're doing now, [INAUDIBLE] not really needed. 00:49:45.290 --> 00:49:48.260 Questions about asset management just before I 00:49:48.260 --> 00:49:49.930 go to a couple other fields? 00:49:52.792 --> 00:49:55.250 ROMAIN: Now is the time to raise your hand if you have any. 00:49:57.760 --> 00:49:59.433 I don't see any, Gary. 00:49:59.433 --> 00:50:01.002 GARY GENSLER: So cost-- 00:50:01.002 --> 00:50:03.880 STUDENT: Sorry, Gary, I raised it in the last second, 00:50:03.880 --> 00:50:05.030 so Romain didn't see it. 00:50:05.030 --> 00:50:07.210 GARY GENSLER: [INAUDIBLE]. 00:50:07.210 --> 00:50:08.530 STUDENT: My question is-- 00:50:08.530 --> 00:50:11.320 so I fully understand their claims 00:50:11.320 --> 00:50:13.960 that this is not the claims of the asset manager, 00:50:13.960 --> 00:50:18.840 you were talking of high speed trading. 00:50:18.840 --> 00:50:22.320 I don't see why machine learning or any algorithm 00:50:22.320 --> 00:50:25.922 besides linear regression doesn't have enough flexibility 00:50:25.922 --> 00:50:27.630 to actually replicate what they're doing, 00:50:27.630 --> 00:50:31.860 but with a bit more of accuracy or computing 00:50:31.860 --> 00:50:35.250 performance, et cetera, because in the end, 00:50:35.250 --> 00:50:37.770 my understanding is that you need the same-- 00:50:37.770 --> 00:50:39.300 except when you go to deep learning. 00:50:39.300 --> 00:50:41.448 But you need the same data sources 00:50:41.448 --> 00:50:43.740 that you currently have, and the only thing that you're 00:50:43.740 --> 00:50:49.640 going to do is, instead of just having linear models predicting 00:50:49.640 --> 00:50:51.990 the different traits that you want to do, 00:50:51.990 --> 00:50:55.260 you have other models that can find 00:50:55.260 --> 00:50:58.330 alternative basically trends and et cetera. 00:50:58.330 --> 00:51:00.850 So I don't see the limitation there of using it. 00:51:00.850 --> 00:51:03.790 GARY GENSLER: So I think you raise a really good point. 00:51:03.790 --> 00:51:06.110 We're in a period of transition. 00:51:06.110 --> 00:51:09.570 I personally don't think that machine learning 00:51:09.570 --> 00:51:13.440 and deep learning is the answer to all these pattern 00:51:13.440 --> 00:51:16.350 recognition challenges in finance. 00:51:16.350 --> 00:51:23.280 But you will find I'm more to the sort of center maximalist 00:51:23.280 --> 00:51:25.480 than center minimalist, and those of you 00:51:25.480 --> 00:51:28.470 that know me, that when we talk about blockchain technology, 00:51:28.470 --> 00:51:31.740 you'll find I'm more to the center minimalist side. 00:51:31.740 --> 00:51:35.010 But what I mean by that is I think that the pattern 00:51:35.010 --> 00:51:37.900 recognition out of deep learning, machine learning-- 00:51:37.900 --> 00:51:41.340 and by pattern recognition I mean the ability to extract, 00:51:41.340 --> 00:51:47.810 with remarkable ability, correlations and then create 00:51:47.810 --> 00:51:51.650 certain decision sets based on those correlations-- 00:51:51.650 --> 00:51:56.660 is better than classic linear algebra, classic regression 00:51:56.660 --> 00:51:58.370 analysis. 00:51:58.370 --> 00:52:03.380 But it comes with a cost, and that's the tradeoff in 2020. 00:52:03.380 --> 00:52:06.190 Maybe in a handful of years, it'll be less cost, 00:52:06.190 --> 00:52:10.460 but I'm going to use an example, which 00:52:10.460 --> 00:52:14.060 is not about asset management, but it's about lending. 00:52:14.060 --> 00:52:18.740 I had this conversation just a few weeks ago with the CEO 00:52:18.740 --> 00:52:21.000 of a major peer to peer lending company, 00:52:21.000 --> 00:52:22.940 and as this is being recorded, I'm 00:52:22.940 --> 00:52:27.530 just going to maybe not say his name. 00:52:27.530 --> 00:52:31.820 And I said, do you use machine learning and deep 00:52:31.820 --> 00:52:36.410 learning to do your credit decisions, 00:52:36.410 --> 00:52:38.150 all your credit decisions? 00:52:38.150 --> 00:52:41.720 And he said, yes, we use a lot of alternative data. 00:52:41.720 --> 00:52:44.150 We run it into the decision sets, 00:52:44.150 --> 00:52:46.650 and we see what patterns emerge. 00:52:46.650 --> 00:52:49.040 And I said, so you extend credit then based on that? 00:52:49.040 --> 00:52:51.290 He said, well, not exactly. 00:52:51.290 --> 00:52:55.280 He said, what we do is, we look for patterns, 00:52:55.280 --> 00:52:57.920 and then when we find them, we then just 00:52:57.920 --> 00:53:04.300 use classic algebra and linear flagging when we actually 00:53:04.300 --> 00:53:05.770 extend the credit. 00:53:05.770 --> 00:53:09.927 So we use it to look for patterns, 00:53:09.927 --> 00:53:12.385 but then we sort of use the traditional way, and I ask why. 00:53:12.385 --> 00:53:13.843 He said, well, there's two reasons. 00:53:13.843 --> 00:53:17.800 It's less costly than running the whole thing all the time, 00:53:17.800 --> 00:53:21.940 and two, they can explain it better to regulators, 00:53:21.940 --> 00:53:24.850 and they can explain it better to the public. 00:53:24.850 --> 00:53:26.890 And at least in consumer finance, 00:53:26.890 --> 00:53:30.070 there were laws passed about 50 years ago-- in the US, 00:53:30.070 --> 00:53:32.620 it's called the Fair Credit Reporting Act. 00:53:32.620 --> 00:53:34.240 These laws were passed and said if you 00:53:34.240 --> 00:53:36.610 deny somebody credit you have to be able to explain 00:53:36.610 --> 00:53:39.430 why you're denying them credit. 00:53:39.430 --> 00:53:41.950 And of course, there are other laws in many other countries 00:53:41.950 --> 00:53:43.630 similar to that, and there are also 00:53:43.630 --> 00:53:46.450 laws about avoiding biases. 00:53:46.450 --> 00:53:49.720 In our country, we call it the Equal Credit Opportunity 00:53:49.720 --> 00:53:52.030 Act or ECOA, and in other countries, 00:53:52.030 --> 00:53:54.580 similar things about avoiding biases 00:53:54.580 --> 00:54:00.440 for gender and race and background and the like. 00:54:00.440 --> 00:54:04.420 So I'm just saying that, whether it's asset management, 00:54:04.420 --> 00:54:06.280 whether it's consumer credit, there's 00:54:06.280 --> 00:54:11.260 some tradeoffs of these new data analytic tools. 00:54:11.260 --> 00:54:17.580 And those tradeoffs, I think, in the next handful of years, 00:54:17.580 --> 00:54:22.800 will keep tipping in the way towards using deep learning. 00:54:22.800 --> 00:54:23.910 But they don't come-- 00:54:23.910 --> 00:54:28.300 they're not cost free is what I'm saying. 00:54:28.300 --> 00:54:30.040 ROMAIN: José has his hand up. 00:54:30.040 --> 00:54:31.390 GARY GENSLER: Please. 00:54:31.390 --> 00:54:34.150 STUDENT: So you talked a bit about how 00:54:34.150 --> 00:54:37.390 this is impacting the high frequency trading shops. 00:54:37.390 --> 00:54:39.730 Is there something similar happening 00:54:39.730 --> 00:54:43.450 on more like value investing long term or in the hedge 00:54:43.450 --> 00:54:44.240 funds? 00:54:44.240 --> 00:54:48.370 So I heard some of them are using satellite image 00:54:48.370 --> 00:54:50.188 to predict the number of cars-- 00:54:50.188 --> 00:54:51.730 to see the number of cars [INAUDIBLE] 00:54:51.730 --> 00:54:54.200 and predict the sales for that year, things like that. 00:54:54.200 --> 00:54:58.600 But I don't know, do you see a lot of headroom in this area? 00:54:58.600 --> 00:55:02.680 GARY GENSLER: I think there is headroom where I'm saying, 00:55:02.680 --> 00:55:06.100 I've mentioned two areas, but I think you're right. 00:55:06.100 --> 00:55:07.150 There's a third area. 00:55:07.150 --> 00:55:10.450 The two areas are sentiment analysis, just 00:55:10.450 --> 00:55:15.460 sentiment analysis around either an individual company 00:55:15.460 --> 00:55:17.800 or the overall markets and seeing 00:55:17.800 --> 00:55:21.730 what the sentiment, the mood, the sense of the crowd 00:55:21.730 --> 00:55:29.300 is off of words, images, or the like. 00:55:29.300 --> 00:55:33.410 And then also the high frequency traders 00:55:33.410 --> 00:55:35.950 and so forth just looking for the patterns in the short term 00:55:35.950 --> 00:55:37.060 trading. 00:55:37.060 --> 00:55:40.360 You're talking about more on the broader value orientation, 00:55:40.360 --> 00:55:42.880 and I absolutely share your view that we'll 00:55:42.880 --> 00:55:44.980 see more of that develop. 00:55:44.980 --> 00:55:48.670 I haven't heard a lot of it, but you're right about sector 00:55:48.670 --> 00:55:51.450 after sector that you might be able to analyze. 00:55:51.450 --> 00:55:55.050 But again, it has to have some ability 00:55:55.050 --> 00:56:01.000 to extract a pattern better than classic linear regressions 00:56:01.000 --> 00:56:02.200 in analysis. 00:56:02.200 --> 00:56:04.200 So let me just try to hit a couple more of these 00:56:04.200 --> 00:56:05.825 because we're going to run out of time, 00:56:05.825 --> 00:56:08.010 but we are going to talk Monday more about these. 00:56:08.010 --> 00:56:11.040 We talked about call centers, chat bots, robo-advising, 00:56:11.040 --> 00:56:13.710 and so forth, so I think you've got that. 00:56:13.710 --> 00:56:16.770 A lot of that is not just efficiency, 00:56:16.770 --> 00:56:22.150 but it's also inclusion. 00:56:22.150 --> 00:56:24.940 You can cover many, many more people 00:56:24.940 --> 00:56:27.760 by automating some of these tasks. 00:56:27.760 --> 00:56:30.940 It's just reality, it's the tradeoff 00:56:30.940 --> 00:56:38.620 of efficiency and inclusion, that you cover far more people. 00:56:38.620 --> 00:56:40.660 You can also be more targeted. 00:56:40.660 --> 00:56:44.280 You can be more targeted with advice and so forth 00:56:44.280 --> 00:56:45.760 as these are automated. 00:56:45.760 --> 00:56:48.500 It comes with the tradeoff that Ivy and Nadia 00:56:48.500 --> 00:56:52.120 were talking about earlier. 00:56:52.120 --> 00:56:54.880 Credit and insurance-- this is the concept 00:56:54.880 --> 00:56:57.550 of basically how you allocate or extend 00:56:57.550 --> 00:57:03.550 or price, either alone or price insurance. 00:57:03.550 --> 00:57:06.960 To date, the insurance companies and insurance underwriting 00:57:06.960 --> 00:57:09.510 are starting to grapple with this, starting to move on, 00:57:09.510 --> 00:57:11.610 and we'll talk about some fintech companies 00:57:11.610 --> 00:57:13.830 in this space. 00:57:13.830 --> 00:57:17.760 I think that insurance companies have been a little bit slower 00:57:17.760 --> 00:57:20.250 to do it than, let's say, the credit card 00:57:20.250 --> 00:57:23.340 companies, but the allocation-- 00:57:23.340 --> 00:57:26.940 I think this will be dramatically shifting. 00:57:26.940 --> 00:57:29.480 I think if you look at what's going on, again, 00:57:29.480 --> 00:57:32.970 in consumer credit and small business credit in China, 00:57:32.970 --> 00:57:35.430 through WeChat Pay and AliPay, they're 00:57:35.430 --> 00:57:40.640 much further along than we are, frankly, here in the US. 00:57:40.640 --> 00:57:47.190 We're still largely reliant on a 30 or 40-year-old architecture 00:57:47.190 --> 00:57:50.790 around the Fair Isaac Company, the FICO scores 00:57:50.790 --> 00:57:54.510 that are used in about 30 countries around the globe. 00:57:54.510 --> 00:57:58.500 These are still quite limited, but FICO itself, 00:57:58.500 --> 00:58:01.590 FICO itself rolls out new versions of FICO 00:58:01.590 --> 00:58:02.760 every few years. 00:58:02.760 --> 00:58:07.650 I think they're rolling out FICO 10.0 this summer, 00:58:07.650 --> 00:58:10.270 or were before the corona crisis. 00:58:10.270 --> 00:58:13.290 I think, if you look at the end of the 2020s, 00:58:13.290 --> 00:58:16.650 either FICO will not exist at all, 00:58:16.650 --> 00:58:21.600 or FICO 14.0 or 15.0 will look a lot more 00:58:21.600 --> 00:58:26.250 like a machine learning, deep learning type of model. 00:58:26.250 --> 00:58:29.280 What it is right now is pretty rudimentary 00:58:29.280 --> 00:58:34.490 compared to what it could be in 5 or 10 years. 00:58:34.490 --> 00:58:36.890 This is an area that's being used a lot-- 00:58:36.890 --> 00:58:38.850 fraud detection and prevention. 00:58:38.850 --> 00:58:40.940 The credit card companies, if you 00:58:40.940 --> 00:58:46.640 look at Cap 1 and Bank of America, Discover, American 00:58:46.640 --> 00:58:50.690 Express, they're deeply now using machine learning tools 00:58:50.690 --> 00:58:52.520 for fraud detection. 00:58:52.520 --> 00:58:57.230 Many of us probably remember just a year or two ago, 00:58:57.230 --> 00:58:59.360 you would still call up your credit card company, 00:58:59.360 --> 00:59:01.610 and you would say I'm traveling to France, 00:59:01.610 --> 00:59:04.430 I'm traveling to Italy, wherever I'm traveling. 00:59:04.430 --> 00:59:06.350 I want to put a flag on there that I 00:59:06.350 --> 00:59:09.320 might be using my cell phone-- 00:59:09.320 --> 00:59:13.470 using my credit card in one of these countries. 00:59:13.470 --> 00:59:15.560 Well, most companies now don't ask you 00:59:15.560 --> 00:59:17.270 to put a travel alert on it. 00:59:17.270 --> 00:59:18.800 Now, part of that is because they 00:59:18.800 --> 00:59:21.230 know where we're traveling because we're walking around 00:59:21.230 --> 00:59:24.870 with these location devices. 00:59:24.870 --> 00:59:28.285 The banks no longer need us to call them to tell them 00:59:28.285 --> 00:59:29.910 we're going to be in Paris because they 00:59:29.910 --> 00:59:31.130 know we're in Paris-- 00:59:31.130 --> 00:59:36.300 this location tracking company device. 00:59:36.300 --> 00:59:40.410 But in addition to that location device, 00:59:40.410 --> 00:59:42.360 they are also using machine learning 00:59:42.360 --> 00:59:45.570 to do fraud detection and prevention in the credit card 00:59:45.570 --> 00:59:47.910 space. 00:59:47.910 --> 00:59:50.390 Similarly, they're using it to track 00:59:50.390 --> 00:59:53.000 and try to comply with laws called anti money laundering. 00:59:53.000 --> 00:59:56.520 These two areas, fraud detection and any money laundering, 00:59:56.520 --> 00:59:59.630 which I might call compliance broadly, 00:59:59.630 --> 01:00:03.860 are two of the areas most developed right now in 2020. 01:00:03.860 --> 01:00:06.380 That doesn't mean they'll be the most developed later. 01:00:06.380 --> 01:00:09.560 I think a lot more will happen in the underwriting space. 01:00:09.560 --> 01:00:11.945 A lot more will happen in the asset management space. 01:00:15.350 --> 01:00:17.780 Robotic process automation-- I want 01:00:17.780 --> 01:00:19.490 to just pause for a minute. 01:00:19.490 --> 01:00:23.060 Does anybody have a sense of what these three words together 01:00:23.060 --> 01:00:26.290 mean-- robotic process automation? 01:00:26.290 --> 01:00:29.905 Romain, you get to see if there's any blue hands up. 01:00:29.905 --> 01:00:32.001 ROMAIN: Andrea. 01:00:32.001 --> 01:00:33.830 STUDENT: Hi. 01:00:33.830 --> 01:00:36.600 So robotic process automation is very simple. 01:00:36.600 --> 01:00:39.280 You have a lot of manual processes 01:00:39.280 --> 01:00:41.080 or manual work done in, for example, 01:00:41.080 --> 01:00:43.020 back office of the banks. 01:00:43.020 --> 01:00:46.030 And the idea here is, instead of people 01:00:46.030 --> 01:00:48.550 doing that, low skilled work or workforce, 01:00:48.550 --> 01:00:53.540 you can actually teach robots or the algorithms with the PC 01:00:53.540 --> 01:00:55.330 to do it instead of you. 01:00:55.330 --> 01:00:56.890 So for example, it can be anything 01:00:56.890 --> 01:00:59.050 as simple as just going through the forms 01:00:59.050 --> 01:01:03.490 and copying or overwriting and rewriting 01:01:03.490 --> 01:01:09.257 some of the words or parts of the forms to some other place. 01:01:09.257 --> 01:01:10.090 GARY GENSLER: Right. 01:01:10.090 --> 01:01:14.320 So robotic process automation can be as simple 01:01:14.320 --> 01:01:19.330 as you're giving your permission to a startup 01:01:19.330 --> 01:01:26.320 company, a fintech company, to access your bank account. 01:01:26.320 --> 01:01:31.000 And one of us gives a fintech company-- maybe Credit Karma. 01:01:31.000 --> 01:01:33.400 We give Credit Karma the right to go in and look 01:01:33.400 --> 01:01:35.020 at a bank account of ours. 01:01:35.020 --> 01:01:39.040 Credit Karma might not have permission from Bank of America 01:01:39.040 --> 01:01:40.490 go in, but they have my-- 01:01:40.490 --> 01:01:41.880 I'm permissioned them. 01:01:41.880 --> 01:01:46.777 I've given them my password and my user ID, and they go in, 01:01:46.777 --> 01:01:47.860 and they want to automate. 01:01:47.860 --> 01:01:49.360 Credit Karma wants to automate. 01:01:49.360 --> 01:01:52.150 They don't want to have a human actually have to type that all 01:01:52.150 --> 01:01:52.680 in. 01:01:52.680 --> 01:01:56.880 It can be as simple as just automating inserting the user 01:01:56.880 --> 01:01:59.830 name and the password and the like, 01:01:59.830 --> 01:02:03.250 but then you go further than it can navigate the web page. 01:02:03.250 --> 01:02:05.830 It can navigate and click the right buttons 01:02:05.830 --> 01:02:08.270 and get the right data and so forth. 01:02:08.270 --> 01:02:10.090 So robotic process automation can 01:02:10.090 --> 01:02:16.170 be helping a startup company say Gary Gensler, 01:02:16.170 --> 01:02:18.780 we want you as a client of Credit Karma. 01:02:18.780 --> 01:02:20.640 We, Credit Karma, will figure out 01:02:20.640 --> 01:02:23.310 how to interface with Bank of America 01:02:23.310 --> 01:02:26.670 and with Chase and the others. 01:02:26.670 --> 01:02:30.870 But also, the banks are using robotic process automation 01:02:30.870 --> 01:02:36.030 to automate so much of their both back office and their data 01:02:36.030 --> 01:02:37.170 entry. 01:02:37.170 --> 01:02:40.470 And one form, just to say, is many 01:02:40.470 --> 01:02:43.320 of us now feel very comfortable to deposit 01:02:43.320 --> 01:02:45.810 a check from our cell phone. 01:02:45.810 --> 01:02:49.560 So you can take your cell phone, take a picture of a check, 01:02:49.560 --> 01:02:51.750 and somehow, that gives an instruction, 01:02:51.750 --> 01:02:56.550 a digital instruction, to move money. 01:02:56.550 --> 01:02:59.580 Well, part of that's natural language processing, 01:02:59.580 --> 01:03:03.880 that the cell phone could take a picture, read all that data, 01:03:03.880 --> 01:03:07.920 put it into computer language, and actually move a digital. 01:03:07.920 --> 01:03:10.650 Part of that is robotic process automation. 01:03:10.650 --> 01:03:11.910 Romain, was there a question? 01:03:11.910 --> 01:03:14.298 I think I saw some flashing chat rooms. 01:03:14.298 --> 01:03:15.090 ROMAIN: Yes, sorry. 01:03:15.090 --> 01:03:16.820 You have 15 minutes left. 01:03:16.820 --> 01:03:17.890 GARY GENSLER: Oh, OK. 01:03:17.890 --> 01:03:20.710 Then was there a question or no? 01:03:20.710 --> 01:03:22.340 ROMAIN: No question at this point. 01:03:22.340 --> 01:03:23.632 GARY GENSLER: So now, trading-- 01:03:23.632 --> 01:03:27.280 trading is an area I spent a lot of time with Goldman Sachs. 01:03:27.280 --> 01:03:30.640 And in that trading of the day, we 01:03:30.640 --> 01:03:32.680 were automating everything we could automate, 01:03:32.680 --> 01:03:35.820 and this is in the 1990s. 01:03:35.820 --> 01:03:38.010 And ever since, anything is-- 01:03:38.010 --> 01:03:41.220 a trading floor in 2020 looks very different than a trading 01:03:41.220 --> 01:03:45.630 floor in the 1990s in terms of the day to day trading, 01:03:45.630 --> 01:03:49.290 and this is trading at the center of the markets, 01:03:49.290 --> 01:03:55.860 the platforms themselves, and of course, the high frequency 01:03:55.860 --> 01:03:58.020 traders on the other end of the market 01:03:58.020 --> 01:04:00.390 is basically just like asset management. 01:04:00.390 --> 01:04:02.010 What patterns can you see? 01:04:02.010 --> 01:04:04.200 Now, this is less about value investing. 01:04:04.200 --> 01:04:09.300 This is the patterns right in the nanoseconds 01:04:09.300 --> 01:04:11.640 and milliseconds and so forth. 01:04:11.640 --> 01:04:14.580 I'll make one note in terms of trading, which is not 01:04:14.580 --> 01:04:16.830 related to machine learning, but just related 01:04:16.830 --> 01:04:20.730 to the corona crisis that we are all living in. 01:04:20.730 --> 01:04:23.970 I have an overall belief that this coronavirus 01:04:23.970 --> 01:04:28.680 crisis will accentuate trends that we've already seen. 01:04:28.680 --> 01:04:32.700 In industry after industry, if we're locked down 01:04:32.700 --> 01:04:37.320 for two, three, or four months, or God forbid, for 18 or 24 01:04:37.320 --> 01:04:38.610 months-- 01:04:38.610 --> 01:04:41.130 if we're locked down for that long, 01:04:41.130 --> 01:04:44.820 we're going to find new ways to engage 01:04:44.820 --> 01:04:47.460 with each other in economic activity 01:04:47.460 --> 01:04:49.620 and social activity and the like. 01:04:49.620 --> 01:04:52.890 And we've already had some trends, deep trends-- 01:04:52.890 --> 01:04:55.770 we talked about them Monday-- 01:04:55.770 --> 01:05:00.270 that we're unlikely to be using many paper money and coinage 01:05:00.270 --> 01:05:01.880 money. 01:05:01.880 --> 01:05:07.250 Three months from now, nearly 70% or 80% the world 01:05:07.250 --> 01:05:11.390 will have forgotten how to use paper money and coinage. 01:05:11.390 --> 01:05:14.810 In fact, it will even be viewed as a disease delivery device. 01:05:14.810 --> 01:05:15.820 It might be dirty. 01:05:15.820 --> 01:05:17.570 It might be something we don't want to use 01:05:17.570 --> 01:05:21.900 because it could be a problem. 01:05:21.900 --> 01:05:23.810 Well, let me talk about trading for a second. 01:05:23.810 --> 01:05:28.340 The New York Stock Exchange and the world's largest stock 01:05:28.340 --> 01:05:30.620 exchanges are now electronic. 01:05:30.620 --> 01:05:34.050 They could have done that two years ago. 01:05:34.050 --> 01:05:35.880 When the Intercontinental Exchange, 01:05:35.880 --> 01:05:38.580 which is a big public company, bought the New York Stock 01:05:38.580 --> 01:05:43.110 Exchange a handful of years ago, five or so years ago, 01:05:43.110 --> 01:05:45.360 Jeff Sprecher, the entrepreneur who 01:05:45.360 --> 01:05:49.800 started the Intercontinental Exchange in 1998 or '99, 01:05:49.800 --> 01:05:53.280 he's always been an entrepreneur, an innovator, 01:05:53.280 --> 01:05:55.098 to do electronic trading. 01:05:55.098 --> 01:05:57.390 They could have taken the New York Stock Exchange fully 01:05:57.390 --> 01:05:59.040 electronic, but guess what? 01:05:59.040 --> 01:06:02.310 That's what happened in the last three weeks. 01:06:02.310 --> 01:06:05.750 So after we get out of this lockdown period, 01:06:05.750 --> 01:06:11.370 will we bring back the floor of these London and New 01:06:11.370 --> 01:06:16.200 York and Shanghai and Mumbai and so forth? 01:06:16.200 --> 01:06:18.420 Will we bring back the floors? 01:06:18.420 --> 01:06:21.711 I think quite possibly not. 01:06:21.711 --> 01:06:29.420 Not sure, but there's a lot that's shifting on, I think. 01:06:29.420 --> 01:06:31.460 Natural language processing-- we talked 01:06:31.460 --> 01:06:35.360 about in customer service, process automation, 01:06:35.360 --> 01:06:36.770 and sentiment analysis. 01:06:36.770 --> 01:06:39.860 These are sort of the slices that I think about 01:06:39.860 --> 01:06:42.690 in these fields. 01:06:42.690 --> 01:06:47.570 So I was curious how many people have ever used Siri? 01:06:47.570 --> 01:06:50.210 Probably almost every hand would go up. 01:06:50.210 --> 01:06:53.225 But how many people have ever used Erica? 01:06:57.610 --> 01:07:00.572 ROMAIN: So perhaps we have Shaheryar who'd 01:07:00.572 --> 01:07:01.780 like to share his experience. 01:07:01.780 --> 01:07:05.880 Sorry for your name-- mispronouncing it. 01:07:05.880 --> 01:07:08.650 STUDENT: Yeah, so it's essentially like Siri, 01:07:08.650 --> 01:07:11.390 but you actually-- it's a product of Bank of America, 01:07:11.390 --> 01:07:14.740 and you can use it to check your spending habits. 01:07:14.740 --> 01:07:17.290 You can also use it to, if you need things with regards 01:07:17.290 --> 01:07:20.150 to check depositing, or if you want to know something, 01:07:20.150 --> 01:07:20.990 it can do that. 01:07:20.990 --> 01:07:25.626 But currently, I believe it's not as refined as Siri, 01:07:25.626 --> 01:07:27.626 and I still think there is a lot room over there 01:07:27.626 --> 01:07:30.650 for improvement. 01:07:30.650 --> 01:07:32.780 GARY GENSLER: Why do you think it is that Erica-- 01:07:32.780 --> 01:07:35.510 and JP Morgan has one as well. 01:07:35.510 --> 01:07:38.450 I can't remember what her name is. 01:07:38.450 --> 01:07:44.960 They all do seem to be mostly female voices, 01:07:44.960 --> 01:07:46.060 if I'm not mistaken. 01:07:46.060 --> 01:07:50.980 But why is it that Erica and the like, 01:07:50.980 --> 01:07:55.040 a virtual assistant, as they're called in finance, 01:07:55.040 --> 01:07:59.550 aren't as developed as the Siris and Alexas, do you think? 01:07:59.550 --> 01:08:01.885 STUDENT: Sorry, can you repeat the question? 01:08:01.885 --> 01:08:03.260 GARY GENSLER: Anybody can answer. 01:08:03.260 --> 01:08:07.430 Why is it that the finance virtual assistants like Erica 01:08:07.430 --> 01:08:13.370 are not yet as fully developed as the Home and other ones 01:08:13.370 --> 01:08:16.050 like Siri and Alexa? 01:08:16.050 --> 01:08:17.240 STUDENT: I believe some of-- 01:08:17.240 --> 01:08:19.157 I think it's got something to do with the fact 01:08:19.157 --> 01:08:21.810 that the number of users for financial assistance 01:08:21.810 --> 01:08:25.370 is way lower as compared to Alexa, Siri, or Alexa 01:08:25.370 --> 01:08:26.510 or whatever. 01:08:26.510 --> 01:08:30.920 So I believe that is something which may explain it, 01:08:30.920 --> 01:08:35.508 the disparity between these two kinds of assistants. 01:08:35.508 --> 01:08:36.300 GARY GENSLER: Yeah. 01:08:36.300 --> 01:08:39.160 Any other thoughts on that, one why-- 01:08:39.160 --> 01:08:43.250 ROMAIN: Nikhil has a different take, and that we have Laira. 01:08:43.250 --> 01:08:45.899 STUDENT: Probably along similar lines, banks so far, 01:08:45.899 --> 01:08:49.609 the interactions probably have been in person or over phones, 01:08:49.609 --> 01:08:52.550 and they weren't used to processing data and processing 01:08:52.550 --> 01:08:53.310 requests. 01:08:53.310 --> 01:08:55.520 I think they have a smaller data set 01:08:55.520 --> 01:08:59.960 to go through and understand what problem customer 01:08:59.960 --> 01:09:01.000 questions are. 01:09:01.000 --> 01:09:04.040 And that's probably a limiting factor versus, say, Siri, 01:09:04.040 --> 01:09:07.819 they have much more data on everyday users. 01:09:07.819 --> 01:09:10.583 I think that's probably the biggest differentiator. 01:09:10.583 --> 01:09:13.000 GARY GENSLER: Yeah, and was there another comment, Romain, 01:09:13.000 --> 01:09:13.790 you said? 01:09:13.790 --> 01:09:14.990 Sure. 01:09:14.990 --> 01:09:16.779 ROMAIN: So I think Laira had her hand up, 01:09:16.779 --> 01:09:18.330 but I think she withdrew it. 01:09:18.330 --> 01:09:20.660 So perhaps we can hear from Brian. 01:09:20.660 --> 01:09:21.819 GARY GENSLER: OK. 01:09:21.819 --> 01:09:23.960 STUDENT: So in addition to the data, 01:09:23.960 --> 01:09:26.439 I think there's also a human capital element. 01:09:26.439 --> 01:09:29.410 It's possible that Apple has better human capital 01:09:29.410 --> 01:09:32.840 capabilities than do these financial institutions, 01:09:32.840 --> 01:09:34.510 so it's harnessing that data as well. 01:09:34.510 --> 01:09:35.420 GARY GENSLER: Yeah. 01:09:35.420 --> 01:09:37.359 So what we've just talked about was data, 01:09:37.359 --> 01:09:40.779 human capital, experience-- 01:09:40.779 --> 01:09:41.470 all true. 01:09:41.470 --> 01:09:43.660 Also, Erica and the financial firms 01:09:43.660 --> 01:09:48.470 only started more recently, and so forth. 01:09:48.470 --> 01:09:51.970 But the voice recognition programs and then 01:09:51.970 --> 01:09:58.660 taking that data that an Apple has or their competitors 01:09:58.660 --> 01:10:00.660 in big tech around the globe-- 01:10:00.660 --> 01:10:03.430 because it's not just here in the US-- 01:10:03.430 --> 01:10:06.400 is really remarkable now. 01:10:06.400 --> 01:10:08.710 Even to the extent that I don't know 01:10:08.710 --> 01:10:13.030 how many people use earbuds, but if you look closely 01:10:13.030 --> 01:10:17.380 at the user agreement on the airbuds that you use, it says-- 01:10:17.380 --> 01:10:22.060 and if it's an Apple, if I'm mistaken, 01:10:22.060 --> 01:10:24.280 the Apple lawyers will chase me. 01:10:24.280 --> 01:10:26.695 But the user agreement says that they 01:10:26.695 --> 01:10:32.350 can listen to that to help you to make sure 01:10:32.350 --> 01:10:34.930 that there's not a drop between your earbuds 01:10:34.930 --> 01:10:37.240 and your cell phone. 01:10:37.240 --> 01:10:45.050 They are picking up vast amounts of data, vast amounts of data 01:10:45.050 --> 01:10:47.240 from our text messaging. 01:10:47.240 --> 01:10:49.340 If you look at something like Google, 01:10:49.340 --> 01:10:53.060 they're picking it up from Google Chrome, Google Maps, 01:10:53.060 --> 01:10:55.180 Gmail. 01:10:55.180 --> 01:10:57.310 Multiple places that they can pick up, 01:10:57.310 --> 01:11:00.640 and we talked about this conceptually, big tech 01:11:00.640 --> 01:11:04.120 versus big finance versus startups in this triangle 01:11:04.120 --> 01:11:06.910 of competitive landscape. 01:11:06.910 --> 01:11:10.090 And why I wanted to sort of close on Bank of America Erica 01:11:10.090 --> 01:11:12.460 and this discussion of Erica versus Siri 01:11:12.460 --> 01:11:18.020 and Alexa is big tech using Google, just as an example, 01:11:18.020 --> 01:11:22.270 has this remarkable network that they're layering activities. 01:11:22.270 --> 01:11:25.430 Remember, we said data network activities-- 01:11:25.430 --> 01:11:29.040 that's the Bank of International Settlement way to put it, 01:11:29.040 --> 01:11:30.520 and what a perfect example to show. 01:11:30.520 --> 01:11:32.680 Google has Gmail. 01:11:32.680 --> 01:11:34.660 It has Maps. 01:11:34.660 --> 01:11:36.100 It has Google Chrome. 01:11:36.100 --> 01:11:39.110 It has Android, the operating system. 01:11:39.110 --> 01:11:42.070 So all of these different ways to build their network, 01:11:42.070 --> 01:11:44.980 and they layer activities on top of it, 01:11:44.980 --> 01:11:47.890 and then vast amounts of data come in 01:11:47.890 --> 01:11:52.430 and the human capital that was mentioned at the end there. 01:11:52.430 --> 01:11:56.350 And they have more experience to move it forward. 01:11:56.350 --> 01:11:58.400 Apple, similarly. 01:11:58.400 --> 01:12:04.290 Baidu and Alibaba in China and so forth, similarly. 01:12:04.290 --> 01:12:07.800 If I were a CEO of a big incumbent, 01:12:07.800 --> 01:12:12.774 yes, I would be very focused on the fintech startups, but I 01:12:12.774 --> 01:12:13.710 tell you-- 01:12:13.710 --> 01:12:16.230 I'd be looking at big tech in a way 01:12:16.230 --> 01:12:20.010 that their advantages are really significant, very 01:12:20.010 --> 01:12:22.260 significant advantages. 01:12:22.260 --> 01:12:25.225 Romain, I see some hands up maybe. 01:12:25.225 --> 01:12:28.310 ROMAIN: We now have Laira who has her hand up. 01:12:28.310 --> 01:12:32.060 STUDENT: Yeah, I'm just curious to know what do you think? 01:12:32.060 --> 01:12:37.580 So currently, we know that tech has kind of-- or AI has kind of 01:12:37.580 --> 01:12:39.890 emerged to the financial and payment space 01:12:39.890 --> 01:12:41.820 in the form of virtual assistants, 01:12:41.820 --> 01:12:45.170 but what do you anticipate the next step would be in terms 01:12:45.170 --> 01:12:48.370 of this integration of technology and AI 01:12:48.370 --> 01:12:50.090 into the payment space? 01:12:50.090 --> 01:12:54.610 What next after the virtual assistants? 01:12:54.610 --> 01:12:57.440 GARY GENSLER: We're going to have a whole class on payments 01:12:57.440 --> 01:12:59.960 specifically, but I think that what's 01:12:59.960 --> 01:13:04.060 happening in the payment space is 01:13:04.060 --> 01:13:07.690 we've seen specialized payment service providers. 01:13:07.690 --> 01:13:09.640 Of course, we've seen a lot of the competition 01:13:09.640 --> 01:13:12.400 starting with PayPal in 1998. 01:13:12.400 --> 01:13:18.460 This is not a new space for disruption. 01:13:18.460 --> 01:13:20.890 But what we've seen more recently 01:13:20.890 --> 01:13:24.640 is, in the retail space, whether it's companies I've mentioned 01:13:24.640 --> 01:13:27.490 earlier, like Toast, that got into one vertical, 01:13:27.490 --> 01:13:31.010 one slice within payments, which was restaurants-- 01:13:31.010 --> 01:13:38.280 they can provide a better product for that slice. 01:13:38.280 --> 01:13:42.780 And then can collect back to AI enough robust data 01:13:42.780 --> 01:13:44.550 within that slice-- 01:13:44.550 --> 01:13:47.640 this is using Toast as an example-- 01:13:47.640 --> 01:13:54.420 that they can provide better software, better hardware, 01:13:54.420 --> 01:14:00.090 and also less risk loans. 01:14:00.090 --> 01:14:03.030 Basically, as Toast started to provide lending 01:14:03.030 --> 01:14:08.430 to restaurants within that space, built upon the payment. 01:14:08.430 --> 01:14:13.740 So it's the marriage between the user experience providing 01:14:13.740 --> 01:14:14.550 the users-- 01:14:14.550 --> 01:14:17.170 in that case, the restaurants in the payment space-- 01:14:17.170 --> 01:14:20.940 but providing the users in that space something 01:14:20.940 --> 01:14:23.780 that the generalized platform-- 01:14:23.780 --> 01:14:26.340 a bank payment system is generalized. 01:14:26.340 --> 01:14:29.460 It's multi sector. 01:14:29.460 --> 01:14:31.635 It's a general product, and Toast 01:14:31.635 --> 01:14:33.510 was able to say, no, we can provide something 01:14:33.510 --> 01:14:35.233 that just restaurants-- 01:14:35.233 --> 01:14:36.900 there might be something a little unique 01:14:36.900 --> 01:14:38.760 about the restaurant business that we can 01:14:38.760 --> 01:14:40.740 provide software and hardware. 01:14:40.740 --> 01:14:42.750 In their case, it was tablets. 01:14:42.750 --> 01:14:45.330 They were providing tablets for the servers 01:14:45.330 --> 01:14:48.000 to walk around and take the orders. 01:14:48.000 --> 01:14:51.540 They could integrate the menu right into the payment app. 01:14:51.540 --> 01:14:54.630 So there was something a little bit unique about that. 01:14:54.630 --> 01:14:57.450 But then based on that, they get a bunch of data, 01:14:57.450 --> 01:15:01.020 and that data helps them with, I would say, 01:15:01.020 --> 01:15:04.770 underwriting decisions based on-- 01:15:04.770 --> 01:15:07.040 it doesn't have to be machine learning. 01:15:07.040 --> 01:15:11.010 But it's enhanced data analytics because of machine learning. 01:15:11.010 --> 01:15:12.020 So I hope that helps. 01:15:12.020 --> 01:15:15.670 I think on the conversational agents 01:15:15.670 --> 01:15:19.610 and the virtual assistants, what we're seeing in the payment 01:15:19.610 --> 01:15:21.230 space, because that was your question, 01:15:21.230 --> 01:15:26.960 is we're moving from card authorized payments 01:15:26.960 --> 01:15:30.260 to mobile app QR codes. 01:15:30.260 --> 01:15:35.540 Then the QR codes is not based upon virtual assistants, 01:15:35.540 --> 01:15:37.820 but it's an interesting question whether we'll 01:15:37.820 --> 01:15:42.830 get to some voice authenticated payments. 01:15:42.830 --> 01:15:48.080 There are a lot of uses of voice authentication already. 01:15:48.080 --> 01:15:50.930 Vanguard and many other asset managers, 01:15:50.930 --> 01:15:52.760 where you can have your brokerage accounts, 01:15:52.760 --> 01:15:55.220 you can call in and get voice authenticated 01:15:55.220 --> 01:15:57.640 before you can do a trade. 01:15:57.640 --> 01:16:00.280 And that voice authentication is just 01:16:00.280 --> 01:16:04.900 like other forms of authentication, 01:16:04.900 --> 01:16:09.840 but it's not perfect as of this moment.