WEBVTT 1 00:00:00.000 --> 00:00:00.740 2 00:00:00.740 --> 00:00:03.160 SAM RANSBOTHAM: We hear a lot about how companies 3 00:00:03.160 --> 00:00:04.840 can use AI to work more efficiently 4 00:00:04.840 --> 00:00:06.450 and improve profitability. 5 00:00:06.450 --> 00:00:08.560 At Warner Music Group, the company 6 00:00:08.560 --> 00:00:10.790 achieves those goals by helping customers discover 7 00:00:10.790 --> 00:00:12.310 the music they like the most. 8 00:00:12.310 --> 00:00:14.390 Find out how AI can help bring music 9 00:00:14.390 --> 00:00:16.810 to your ears in today's episode. 10 00:00:16.810 --> 00:00:19.330 KOBI ABAYOMI: I'm Kobi Abayomi from Warner Music Group, 11 00:00:19.330 --> 00:00:22.290 and you're listening to Me, Myself, and AI. 12 00:00:22.290 --> 00:00:24.290 SAM RANSBOTHAM: Welcome to Me, Myself, and AI, 13 00:00:24.290 --> 00:00:27.270 the podcast on artificial intelligence in business. 14 00:00:27.270 --> 00:00:28.780 In each episode, we introduce you 15 00:00:28.780 --> 00:00:30.440 to someone innovating with AI. 16 00:00:30.440 --> 00:00:33.610 I'm Sam Ransbotham, professor of information systems in Boston 17 00:00:33.610 --> 00:00:36.660 College, and I'm also the guest editor for the AI and Business 18 00:00:36.660 --> 00:00:40.128 Strategy Big Ideas program at MIT Sloan Management Review. 19 00:00:40.128 --> 00:00:42.170 SHERVIN KHODABANDEH: And I'm Shervin Khodabandeh, 20 00:00:42.170 --> 00:00:45.260 senior partner and managing director at Boston Consulting 21 00:00:45.260 --> 00:00:45.760 Group. 22 00:00:45.760 --> 00:00:46.840 KOBI ABAYOMI: Nice to meet you. 23 00:00:46.840 --> 00:00:48.210 SHERVIN KHODABANDEH: Welcome to the show. 24 00:00:48.210 --> 00:00:49.290 KOBI ABAYOMI: Thank you. 25 00:00:49.290 --> 00:00:50.665 SAM RANSBOTHAM: Shervin and I are 26 00:00:50.665 --> 00:00:53.360 excited to be talking today with Kobi Abayomi, senior vice 27 00:00:53.360 --> 00:00:55.597 president for data science at Warner Music Group. 28 00:00:55.597 --> 00:00:57.430 Kobi, thanks for taking the time to join us. 29 00:00:57.430 --> 00:00:58.090 Welcome. 30 00:00:58.090 --> 00:00:59.040 KOBI ABAYOMI: Yes, thank you. 31 00:00:59.040 --> 00:00:59.660 Thank you for having me. 32 00:00:59.660 --> 00:01:01.270 SAM RANSBOTHAM: Let's get started. 33 00:01:01.270 --> 00:01:03.580 Kobi, you've got a new position at a new company 34 00:01:03.580 --> 00:01:05.120 for you, Warner Music. 35 00:01:05.120 --> 00:01:06.710 Can you tell us about your role? 36 00:01:06.710 --> 00:01:07.910 KOBI ABAYOMI: Sure. 37 00:01:07.910 --> 00:01:09.850 I started here about a year ago. 38 00:01:09.850 --> 00:01:11.860 I lead the data science effort of the company. 39 00:01:11.860 --> 00:01:12.910 It's a music company. 40 00:01:12.910 --> 00:01:14.570 And what is a music company? 41 00:01:14.570 --> 00:01:17.920 A music company is a support network for artists; 42 00:01:17.920 --> 00:01:21.630 it's a repository for licensing rights; 43 00:01:21.630 --> 00:01:24.440 it's a creator of music content. 44 00:01:24.440 --> 00:01:27.170 So [it's] a modern music company -- 45 00:01:27.170 --> 00:01:28.910 one of the three major companies. 46 00:01:28.910 --> 00:01:31.190 The others are Sony and Universal, 47 00:01:31.190 --> 00:01:35.490 and each of them probably undertakes a suite of activity 48 00:01:35.490 --> 00:01:39.110 that's relatively similar, which is find new talent, 49 00:01:39.110 --> 00:01:43.370 maintain and monetize music that's already been produced -- 50 00:01:43.370 --> 00:01:46.470 both from the recorded and publishing sides -- 51 00:01:46.470 --> 00:01:50.280 and then find other ways to monetize current and past 52 00:01:50.280 --> 00:01:53.360 artists, through merchandise, through licensing, 53 00:01:53.360 --> 00:01:57.390 through sync (that's when music is played in other media, 54 00:01:57.390 --> 00:02:02.130 commercials, movies, etc.), and the stuff that we all like 55 00:02:02.130 --> 00:02:05.080 and enjoy: going on tours, live music, things like that. 56 00:02:05.080 --> 00:02:07.170 And so the company has its fingers 57 00:02:07.170 --> 00:02:09.289 in all parts of the music landscape. 58 00:02:09.289 --> 00:02:11.403 SAM RANSBOTHAM: Your role is senior vice president 59 00:02:11.403 --> 00:02:13.820 for data science, and I don't think you mentioned the word 60 00:02:13.820 --> 00:02:14.540 data in there at all. 61 00:02:14.540 --> 00:02:15.180 KOBI ABAYOMI: Sure. 62 00:02:15.180 --> 00:02:16.525 SAM RANSBOTHAM: What role does data play? 63 00:02:16.525 --> 00:02:16.880 64 00:02:16.880 --> 00:02:18.020 KOBI ABAYOMI: Let me give you the other side 65 00:02:18.020 --> 00:02:18.840 of the page, then. 66 00:02:18.840 --> 00:02:20.130 Data plays a large role. 67 00:02:20.130 --> 00:02:22.598 This is a legacy media company, [of] which there are many. 68 00:02:22.598 --> 00:02:24.640 From a data science perspective, much of the work 69 00:02:24.640 --> 00:02:25.890 is infrastructural. 70 00:02:25.890 --> 00:02:29.950 You are in an organization which did things one way and now 71 00:02:29.950 --> 00:02:33.770 is faced with a more competitive landscape with tech 72 00:02:33.770 --> 00:02:37.540 companies that have now their feet on the media 73 00:02:37.540 --> 00:02:39.750 side of the consumer space. 74 00:02:39.750 --> 00:02:43.530 And so a lot of the work is around infrastructure, 75 00:02:43.530 --> 00:02:46.750 codification, and technology, really -- 76 00:02:46.750 --> 00:02:50.160 being able to ingest data, turn it into features, 77 00:02:50.160 --> 00:02:53.460 and turn it into insights that are meaningful so that 78 00:02:53.460 --> 00:02:57.100 the business can compete and operate now that it's beset 79 00:02:57.100 --> 00:02:59.620 with a volume of data that it wasn't before. 80 00:02:59.620 --> 00:03:01.360 Much of the way money comes in the door 81 00:03:01.360 --> 00:03:04.670 for a music company nowadays is through digital streaming. 82 00:03:04.670 --> 00:03:07.180 The distribution channel is digital, right? 83 00:03:07.180 --> 00:03:11.280 The Spotifys and Apple Musics of the world are the ones who 84 00:03:11.280 --> 00:03:15.720 deliver our product directly to my listening ears -- 85 00:03:15.720 --> 00:03:18.520 as well as your listening ears. 86 00:03:18.520 --> 00:03:23.930 And we get paid for the use of that product and inventory. 87 00:03:23.930 --> 00:03:28.310 And then what happens is, the data about the use of that 88 00:03:28.310 --> 00:03:32.720 product -- which originates and is at the highest resolution 89 00:03:32.720 --> 00:03:35.240 on the distribution side of the channel -- 90 00:03:35.240 --> 00:03:36.850 we get a version of it. 91 00:03:36.850 --> 00:03:42.615 So a lot of the things that have become pro forma in D2C 92 00:03:42.615 --> 00:03:44.990 [direct-to-consumer] businesses, which are recommendation 93 00:03:44.990 --> 00:03:47.490 systems, audience segmentation, things like that, 94 00:03:47.490 --> 00:03:50.000 are things now media companies -- not just this company; 95 00:03:50.000 --> 00:03:52.840 any sort of legacy media company that has a distribution channel 96 00:03:52.840 --> 00:03:54.970 that's changed from analog to digital -- 97 00:03:54.970 --> 00:03:58.680 are now having to re-factor and re-understand: 98 00:03:58.680 --> 00:04:00.010 How do we understand market? 99 00:04:00.010 --> 00:04:02.130 How do we understand audience groups? 100 00:04:02.130 --> 00:04:04.727 I'm old enough to remember and have participated in -- 101 00:04:04.727 --> 00:04:06.810 back when I was a college student -- focus groups. 102 00:04:06.810 --> 00:04:10.462 [Laughs.] And I remember the college I went to, I guess I 103 00:04:10.462 --> 00:04:11.920 was a good focus group participant, 104 00:04:11.920 --> 00:04:15.840 and they kept calling me back in for different campaigns -- 105 00:04:15.840 --> 00:04:18.265 you know, Budweiser or Hertz Rent-a-Car -- 106 00:04:18.265 --> 00:04:20.640 and I remember being asked a bunch of different questions 107 00:04:20.640 --> 00:04:21.310 in a group. 108 00:04:21.310 --> 00:04:24.320 Well, now, with the digital consumption of things, 109 00:04:24.320 --> 00:04:27.090 we can augment those older methodologies 110 00:04:27.090 --> 00:04:30.380 of what people actually like, who they actually are, 111 00:04:30.380 --> 00:04:32.910 what their affinities are, what their behaviors are. 112 00:04:32.910 --> 00:04:36.430 And so for this company in particular -- 113 00:04:36.430 --> 00:04:38.540 for legacy media companies in general -- 114 00:04:38.540 --> 00:04:42.970 a lot of the data science work is around building the capacity 115 00:04:42.970 --> 00:04:47.825 for them to be able to enjoy what's now available data-wise 116 00:04:47.825 --> 00:04:49.200 in understanding their consumers. 117 00:04:49.200 --> 00:04:51.017 SAM RANSBOTHAM: One of the challenges 118 00:04:51.017 --> 00:04:53.350 that Shervin and I often have is when we talk to people, 119 00:04:53.350 --> 00:04:55.978 they're very narrowly focused on their particular job. 120 00:04:55.978 --> 00:04:57.520 And so one of the things we're always 121 00:04:57.520 --> 00:04:59.750 kind of thinking about is, "How do we expand this 122 00:04:59.750 --> 00:05:01.917 so that everyone can relate?" because not everyone's 123 00:05:01.917 --> 00:05:03.010 in the music business. 124 00:05:03.010 --> 00:05:04.430 But you've done that for us. 125 00:05:04.430 --> 00:05:07.370 I mean, you've just really laid out the challenges of legacy 126 00:05:07.370 --> 00:05:10.250 companies and how that applies not just with media, 127 00:05:10.250 --> 00:05:12.880 but I mean, so much is digitally distributed. 128 00:05:12.880 --> 00:05:14.853 I think that's an interesting broadening 129 00:05:14.853 --> 00:05:17.145 that you've already done there to explain that process. 130 00:05:17.145 --> 00:05:17.560 131 00:05:17.560 --> 00:05:19.643 KOBI ABAYOMI: Oh, well, thank you for saying that. 132 00:05:19.643 --> 00:05:21.790 This is the speech that I give inside the company. 133 00:05:21.790 --> 00:05:24.210 So, everybody in the company doesn't have an understanding 134 00:05:24.210 --> 00:05:25.240 of "What is data science? 135 00:05:25.240 --> 00:05:26.198 Why are you guys here?" 136 00:05:26.198 --> 00:05:29.590 And I think anybody who's in data science 137 00:05:29.590 --> 00:05:32.060 is beset with that existential question. 138 00:05:32.060 --> 00:05:34.940 SHERVIN KHODABANDEH: You started by saying 139 00:05:34.940 --> 00:05:37.890 a lot of the challenges, given that it's a legacy company, 140 00:05:37.890 --> 00:05:40.690 are infrastructural challenges, and you nicely 141 00:05:40.690 --> 00:05:41.940 walked us through that. 142 00:05:41.940 --> 00:05:45.480 But then you also talked about the axiom of existence 143 00:05:45.480 --> 00:05:48.310 in a legacy company for a data science team, 144 00:05:48.310 --> 00:05:51.010 which is a little bit of a -- or maybe a lot of -- 145 00:05:51.010 --> 00:05:53.830 a cultural and organizational challenge as well. 146 00:05:53.830 --> 00:05:56.778 Comment a little bit about that side of the coin. 147 00:05:56.778 --> 00:05:57.570 KOBI ABAYOMI: Sure. 148 00:05:57.570 --> 00:06:00.210 I've been doing statistics long enough that I remember 149 00:06:00.210 --> 00:06:02.250 when academic departments were getting rid 150 00:06:02.250 --> 00:06:04.880 of their stats departments, where statistics became 151 00:06:04.880 --> 00:06:08.018 biostatistics, econometrics, psychometrics, and people 152 00:06:08.018 --> 00:06:09.310 [said], "We'll do it ourselves. 153 00:06:09.310 --> 00:06:10.030 We'll teach ourselves. 154 00:06:10.030 --> 00:06:11.572 We'll teach our own service classes." 155 00:06:11.572 --> 00:06:13.950 Princeton got rid of their stats department. 156 00:06:13.950 --> 00:06:15.700 I remember the University of Texas 157 00:06:15.700 --> 00:06:17.662 when they folded what they did in statistics 158 00:06:17.662 --> 00:06:19.620 into their industrial engineering [department]. 159 00:06:19.620 --> 00:06:22.250 This was all happening 20, 25 years ago. 160 00:06:22.250 --> 00:06:24.500 The generation of data caught up. 161 00:06:24.500 --> 00:06:26.480 Now we have devices that generate data, 162 00:06:26.480 --> 00:06:30.520 and a lot of the ideas and notions we had around 163 00:06:30.520 --> 00:06:34.040 large-scale computational statistics have now come 164 00:06:34.040 --> 00:06:37.177 to life (1) out of necessity, and (2) out of ability -- 165 00:06:37.177 --> 00:06:39.010 that we have the ability to do these things. 166 00:06:39.010 --> 00:06:40.468 You can do these things on laptops; 167 00:06:40.468 --> 00:06:42.740 you don't have to access the Condor mainframe 168 00:06:42.740 --> 00:06:46.355 and run jobs in parallel, things like that. 169 00:06:46.355 --> 00:06:48.230 To now answer that in the context of business 170 00:06:48.230 --> 00:06:51.190 and cultural sorts of things, business is people, right? 171 00:06:51.190 --> 00:06:53.840 We're lucky to organize people for similar tasks. 172 00:06:53.840 --> 00:06:55.340 I say this: I love cars. 173 00:06:55.340 --> 00:06:57.060 I remember one day having a thought: 174 00:06:57.060 --> 00:06:59.590 What an amazing thing a car is. 175 00:06:59.590 --> 00:07:04.470 Take Toyota, for instance -- this large concern which spans 176 00:07:04.470 --> 00:07:06.730 all these different activities: We make dashboards; 177 00:07:06.730 --> 00:07:09.500 we make dials; we make door handles. 178 00:07:09.500 --> 00:07:12.060 And essentially, even in a deeper way, 179 00:07:12.060 --> 00:07:14.750 we turn minerals and dust into this product -- 180 00:07:14.750 --> 00:07:17.760 regular product -- where I can have a Corolla, 181 00:07:17.760 --> 00:07:21.560 you can have a Corolla, and they're more or less the same. 182 00:07:21.560 --> 00:07:26.620 What an amazing thing to marshal so many people 183 00:07:26.620 --> 00:07:31.030 to get out this regular, complicated thing. 184 00:07:31.030 --> 00:07:32.800 In any business, you have that. 185 00:07:32.800 --> 00:07:35.710 You're appealing to the Jungian zeitgeist 186 00:07:35.710 --> 00:07:39.790 to produce something which is coherent and then reproducible. 187 00:07:39.790 --> 00:07:41.390 It's how business makes money. 188 00:07:41.390 --> 00:07:44.520 In the music business, one of the challenges 189 00:07:44.520 --> 00:07:50.310 is the way people understand music, 190 00:07:50.310 --> 00:07:51.820 and how people consume it. 191 00:07:51.820 --> 00:07:55.160 There's a narrative that making a hit, say, [is] 192 00:07:55.160 --> 00:07:57.190 just a chance thing that could happen, 193 00:07:57.190 --> 00:08:01.420 or idiosyncratic or only known by the guy who's 194 00:08:01.420 --> 00:08:05.070 got the cool Adidas sneakers on who's out at the club right 195 00:08:05.070 --> 00:08:05.570 now. 196 00:08:05.570 --> 00:08:09.710 And what I always say in having these conversations 197 00:08:09.710 --> 00:08:12.440 about culture: Are people predictable? 198 00:08:12.440 --> 00:08:13.990 Do people have patterns? 199 00:08:13.990 --> 00:08:17.170 And can we just pay closer attention to what we do? 200 00:08:17.170 --> 00:08:19.800 And that's all really data science is, right? 201 00:08:19.800 --> 00:08:23.130 Writing things down, pretending that there was an experiment. 202 00:08:23.130 --> 00:08:26.520 From what assumptions do our actions arise? 203 00:08:26.520 --> 00:08:29.000 And then, how do we notice the results 204 00:08:29.000 --> 00:08:32.320 of those things, mediated and engaged by those assumptions? 205 00:08:32.320 --> 00:08:34.927 These are the sorts of conversations 206 00:08:34.927 --> 00:08:36.510 that I have inside the company, that I 207 00:08:36.510 --> 00:08:38.539 have with people on my team. 208 00:08:38.539 --> 00:08:41.710 And I think one of the beautiful things about music, 209 00:08:41.710 --> 00:08:43.520 and why I find myself lucky to be doing 210 00:08:43.520 --> 00:08:45.930 data science at a music company, is 211 00:08:45.930 --> 00:08:47.510 that there's resonance there. 212 00:08:47.510 --> 00:08:51.790 And I think people who generally have a natural inclination 213 00:08:51.790 --> 00:08:55.510 to this version of the art form and its complexity 214 00:08:55.510 --> 00:08:57.860 hear things like that and get that argument. 215 00:08:57.860 --> 00:08:59.600 It makes sense. 216 00:08:59.600 --> 00:09:02.330 We're trying to figure out I've had this conversation 217 00:09:02.330 --> 00:09:05.040 with some of my colleagues at other data 218 00:09:05.040 --> 00:09:07.870 science groups, at other music or music distribution 219 00:09:07.870 --> 00:09:08.370 companies. 220 00:09:08.370 --> 00:09:10.120 And what, essentially, you're trying to do 221 00:09:10.120 --> 00:09:12.930 is to curate more listening. 222 00:09:12.930 --> 00:09:15.300 I think anybody in media these days -- 223 00:09:15.300 --> 00:09:17.530 I think when we look back at this epoch, 224 00:09:17.530 --> 00:09:21.660 the epitaph will be "This was the competitive area 225 00:09:21.660 --> 00:09:26.920 of attention" or "the era of few eyeballs and many billboards." 226 00:09:26.920 --> 00:09:30.590 We're all trying to extract a greater slice 227 00:09:30.590 --> 00:09:32.740 of people's attention budget. 228 00:09:32.740 --> 00:09:34.490 We're even more advantaged because there's 229 00:09:34.490 --> 00:09:38.560 a need to consume that sort of content. 230 00:09:38.560 --> 00:09:40.280 Your brain wants to hear something; 231 00:09:40.280 --> 00:09:42.275 your brain wants to hear something novel. 232 00:09:42.275 --> 00:09:43.650 And then the other part of it is, 233 00:09:43.650 --> 00:09:45.840 as people grow and audiences grow, 234 00:09:45.840 --> 00:09:49.190 what you're appealing to are people's past experiences 235 00:09:49.190 --> 00:09:53.460 and their experiential sort of understanding and incorporation 236 00:09:53.460 --> 00:09:55.660 of what they've heard at different times, 237 00:09:55.660 --> 00:09:58.480 and we can see this in consumption patterns in music. 238 00:09:58.480 --> 00:10:03.470 As the total addressable market of digital listeners increases, 239 00:10:03.470 --> 00:10:07.670 you see a tilt toward certain vintages of catalog listening. 240 00:10:07.670 --> 00:10:10.280 You get a population, and you give them the opportunity 241 00:10:10.280 --> 00:10:12.530 to listen to all sorts of stuff; turns out 242 00:10:12.530 --> 00:10:15.270 they want to listen to the stuff that they 243 00:10:15.270 --> 00:10:18.280 associate with good feelings and good times in their lives. 244 00:10:18.280 --> 00:10:21.400 And so what we want to do is to really understand 245 00:10:21.400 --> 00:10:25.290 that interaction between a person's affinities 246 00:10:25.290 --> 00:10:28.560 for certain sorts of sounds and experiences, 247 00:10:28.560 --> 00:10:30.580 and curate that, in a way. 248 00:10:30.580 --> 00:10:33.940 Some of that curation can happen from the artist as the locus, 249 00:10:33.940 --> 00:10:34.440 right? 250 00:10:34.440 --> 00:10:37.810 Like, John Mellencamp makes you feel a certain way, 251 00:10:37.810 --> 00:10:40.710 and you become a fan of that and very invested 252 00:10:40.710 --> 00:10:43.723 in that particular artist in a certain way. 253 00:10:43.723 --> 00:10:45.390 This sound makes you feel a certain way. 254 00:10:45.390 --> 00:10:49.460 Electronic music, deep house makes you feel a certain way. 255 00:10:49.460 --> 00:10:54.100 And our job is to understand that well enough to be able 256 00:10:54.100 --> 00:10:57.610 to dynamically have the content and have it curated 257 00:10:57.610 --> 00:11:01.780 and delivered so that we can capture and respond to what I 258 00:11:01.780 --> 00:11:06.560 consider a natural desire to consume our product -- music. 259 00:11:06.560 --> 00:11:07.060 260 00:11:07.060 --> 00:11:08.710 SHERVIN KHODABANDEH: Very well said. 261 00:11:08.710 --> 00:11:11.340 And you could do it with a clear conscience. 262 00:11:11.340 --> 00:11:12.115 263 00:11:12.115 --> 00:11:13.240 KOBI ABAYOMI: That's right. 264 00:11:13.240 --> 00:11:14.050 Most of the time! 265 00:11:14.050 --> 00:11:14.550 266 00:11:14.550 --> 00:11:18.710 SHERVIN KHODABANDEH: You gave us a very good glimpse 267 00:11:18.710 --> 00:11:23.490 of why it's purposeful and interesting to the organization 268 00:11:23.490 --> 00:11:27.000 and to the company to do these types of AI-driven use cases. 269 00:11:27.000 --> 00:11:28.740 What makes your job hard? 270 00:11:28.740 --> 00:11:30.218 What are some of the challenges? 271 00:11:30.218 --> 00:11:31.010 KOBI ABAYOMI: Sure. 272 00:11:31.010 --> 00:11:32.270 I'll say a couple of things. 273 00:11:32.270 --> 00:11:35.730 First, it's a difficult problem as a technical thing. 274 00:11:35.730 --> 00:11:38.320 We have different verticals on my team 275 00:11:38.320 --> 00:11:41.330 dedicated to the specific legs of a stool, 276 00:11:41.330 --> 00:11:44.600 or understanding audiences and doing that 277 00:11:44.600 --> 00:11:47.050 through a mediated view of them. 278 00:11:47.050 --> 00:11:49.270 Again, we're more like a B2B business 279 00:11:49.270 --> 00:11:51.620 than a direct-to-consumer, right? 280 00:11:51.620 --> 00:11:55.030 Spotify and Apple are mediating us. 281 00:11:55.030 --> 00:11:58.090 How can we best understand our audiences 282 00:11:58.090 --> 00:12:01.650 in a dynamic and high-resolution way with the information 283 00:12:01.650 --> 00:12:02.240 that we have? 284 00:12:02.240 --> 00:12:05.270 This is where data science can add a lot of value. 285 00:12:05.270 --> 00:12:09.360 Our content itself --understanding our content 286 00:12:09.360 --> 00:12:12.770 itself in a way that's useful for data science. 287 00:12:12.770 --> 00:12:17.500 Turning sound into information, and not just 288 00:12:17.500 --> 00:12:19.960 sound into information but the song itself. 289 00:12:19.960 --> 00:12:22.840 The song itself has a product from an artist. 290 00:12:22.840 --> 00:12:26.140 The song itself has a product from a collection of producers. 291 00:12:26.140 --> 00:12:29.900 The song itself has a product from a collection of musicians. 292 00:12:29.900 --> 00:12:32.070 Having the data science infrastructure 293 00:12:32.070 --> 00:12:37.400 that allows us to encode all of this and associate in a way 294 00:12:37.400 --> 00:12:41.610 that we can model it and offer meaningful inference 295 00:12:41.610 --> 00:12:42.740 to the business. 296 00:12:42.740 --> 00:12:47.770 As a scientific exercise, there are a lot of interesting things 297 00:12:47.770 --> 00:12:50.600 that are very meaty, and I enjoy that part of it. 298 00:12:50.600 --> 00:12:52.783 I say that in answer to "What part is difficult?" 299 00:12:52.783 --> 00:12:55.450 The other part of it is -- and I think this is not just speaking 300 00:12:55.450 --> 00:12:59.460 for what we're doing in particular in the group that I 301 00:12:59.460 --> 00:13:03.910 work with but just in general -- back to this need 302 00:13:03.910 --> 00:13:06.115 for justification of existence. 303 00:13:06.115 --> 00:13:08.490 Part of what I think any data science department is going 304 00:13:08.490 --> 00:13:11.160 to be beset with, especially the legacy company, 305 00:13:11.160 --> 00:13:13.940 is satisfying the need for information product 306 00:13:13.940 --> 00:13:16.950 to the business while at the same time 307 00:13:16.950 --> 00:13:21.560 setting the business up for what the future is going to be. 308 00:13:21.560 --> 00:13:24.690 If you set up a data science department that just answers 309 00:13:24.690 --> 00:13:27.130 what's being asked now, that's not a useful data science 310 00:13:27.130 --> 00:13:28.273 department. 311 00:13:28.273 --> 00:13:29.690 I'm trying to think of an example. 312 00:13:29.690 --> 00:13:32.120 [What] if Google had stopped at "We're just 313 00:13:32.120 --> 00:13:34.210 going to do PageRank and that's it, 314 00:13:34.210 --> 00:13:36.560 and we're going to sell ads through PageRank"? 315 00:13:36.560 --> 00:13:38.160 Somebody asked the question, "Well, 316 00:13:38.160 --> 00:13:39.590 what if we had a mapping product? 317 00:13:39.590 --> 00:13:41.020 Well, what if we had an email product? 318 00:13:41.020 --> 00:13:42.000 What if we had these things?" 319 00:13:42.000 --> 00:13:44.375 And I'm sure to the people who were there at the business 320 00:13:44.375 --> 00:13:47.080 at the time in the past, these were 321 00:13:47.080 --> 00:13:50.470 new and novel and something that required some convincing. 322 00:13:50.470 --> 00:13:55.110 We were in the same position in collaboration and conversation 323 00:13:55.110 --> 00:13:58.430 with the business to push the frontier of what 324 00:13:58.430 --> 00:14:02.577 can be done, to excite their imagination on what's possible. 325 00:14:02.577 --> 00:14:03.120 326 00:14:03.120 --> 00:14:05.370 SAM RANSBOTHAM: How are you figuring out those things? 327 00:14:05.370 --> 00:14:06.560 How do you figure out what's next? 328 00:14:06.560 --> 00:14:07.730 It seems easy enough to -- 329 00:14:07.730 --> 00:14:08.140 KOBI ABAYOMI: Yeah. 330 00:14:08.140 --> 00:14:09.770 SAM RANSBOTHAM: You hear the complaints to tell you what 331 00:14:09.770 --> 00:14:10.940 you should be doing right now. 332 00:14:10.940 --> 00:14:11.500 KOBI ABAYOMI: That's right. 333 00:14:11.500 --> 00:14:11.810 That's right. 334 00:14:11.810 --> 00:14:12.230 [Laughs.] 335 00:14:12.230 --> 00:14:13.700 SAM RANSBOTHAM: How are you figuring out what to do 336 00:14:13.700 --> 00:14:15.130 and where you should be going? 337 00:14:15.130 --> 00:14:17.670 KOBI ABAYOMI: [It's a] work in progress. 338 00:14:17.670 --> 00:14:21.150 Here's an example: This past year I had to have eye surgery, 339 00:14:21.150 --> 00:14:24.433 and in the beginning, the questions 340 00:14:24.433 --> 00:14:26.600 I might have I was like, "Oh, I can feel something's 341 00:14:26.600 --> 00:14:28.850 wrong with my eye; what needs to be done?" 342 00:14:28.850 --> 00:14:30.480 And so then you go to the specialist, 343 00:14:30.480 --> 00:14:32.750 and the specialist goes through a series of things 344 00:14:32.750 --> 00:14:34.757 to get to the point where, "Here, this 345 00:14:34.757 --> 00:14:36.840 is the outcome you want, but these are things that 346 00:14:36.840 --> 00:14:38.330 need to happen along the way." 347 00:14:38.330 --> 00:14:41.380 You have to be able to follow along 348 00:14:41.380 --> 00:14:44.380 that path of investigation with somebody 349 00:14:44.380 --> 00:14:48.790 who's an expert to be able to get to the overall coarser 350 00:14:48.790 --> 00:14:49.790 goal. 351 00:14:49.790 --> 00:14:53.070 When we, in conversation with our business partners, 352 00:14:53.070 --> 00:14:56.337 are able to frame what they want So, an analogy: 353 00:14:56.337 --> 00:14:58.170 They're the people who need the eye surgery, 354 00:14:58.170 --> 00:15:00.610 and at times, we're the people who 355 00:15:00.610 --> 00:15:04.390 can provide a scalpel or a scleral buckle 356 00:15:04.390 --> 00:15:06.680 or a bubble in their retina, or whatever 357 00:15:06.680 --> 00:15:10.490 it is that needs to address what they're asking for. 358 00:15:10.490 --> 00:15:12.490 When we're able to have conversations like that, 359 00:15:12.490 --> 00:15:15.620 it's very useful for us, because we're 360 00:15:15.620 --> 00:15:17.410 able to break up where they're trying 361 00:15:17.410 --> 00:15:19.910 to get to into the pieces that we know that we can construct 362 00:15:19.910 --> 00:15:20.987 to get there. 363 00:15:20.987 --> 00:15:22.570 Those are conversations that we really 364 00:15:22.570 --> 00:15:26.970 enjoy having, but then, more directly, I 365 00:15:26.970 --> 00:15:29.280 don't think any of us knows where this space can go 366 00:15:29.280 --> 00:15:33.700 and what the next frontier of music enjoyment is. 367 00:15:33.700 --> 00:15:37.080 I don't think people could have predicted things like Snapchat 368 00:15:37.080 --> 00:15:40.800 or the proliferation of short-form video in the way 369 00:15:40.800 --> 00:15:42.240 that it's very popular right now. 370 00:15:42.240 --> 00:15:46.990 I think we'll all be surprised at what happens in music 371 00:15:46.990 --> 00:15:48.510 in concert with data science. 372 00:15:48.510 --> 00:15:51.310 Some of the things that we're excited about are things that 373 00:15:51.310 --> 00:15:52.950 are happening in the virtual world -- 374 00:15:52.950 --> 00:15:55.210 Web3 sort of conversations. 375 00:15:55.210 --> 00:15:58.023 I had pooh-poohed virtual reality and the metaverse 376 00:15:58.023 --> 00:15:58.690 for a long time. 377 00:15:58.690 --> 00:16:02.160 I have a 12-year-old daughter; she put her VR glasses on me 378 00:16:02.160 --> 00:16:03.120 a couple of weeks ago. 379 00:16:03.120 --> 00:16:04.240 I didn't want to take them off! 380 00:16:04.240 --> 00:16:05.270 I was like, "Oh my God." 381 00:16:05.270 --> 00:16:06.679 And when I took them off -- 382 00:16:06.679 --> 00:16:09.096 SAM RANSBOTHAM: Was that before or after your eye surgery? 383 00:16:09.096 --> 00:16:10.070 384 00:16:10.070 --> 00:16:11.380 KOBI ABAYOMI: That was after. 385 00:16:11.380 --> 00:16:13.400 [Laughs.] And so what that just told me is, 386 00:16:13.400 --> 00:16:17.040 there's a lot of power in these new spaces and new ways that 387 00:16:17.040 --> 00:16:20.350 people are consuming things -- gaming, the metaverse -- 388 00:16:20.350 --> 00:16:24.740 and new ways that we can deliver experiential goods, 389 00:16:24.740 --> 00:16:28.590 if you will, around music that are edifying to people and grab 390 00:16:28.590 --> 00:16:30.090 another erg of the attention budget. 391 00:16:30.090 --> 00:16:33.880 SHERVIN KHODABANDEH: Kobi, how do you find your team? 392 00:16:33.880 --> 00:16:35.980 What makes a good data scientist? 393 00:16:35.980 --> 00:16:36.822 394 00:16:36.822 --> 00:16:37.530 KOBI ABAYOMI: Oh! 395 00:16:37.530 --> 00:16:37.840 That's great. 396 00:16:37.840 --> 00:16:39.140 You know, we're very lucky. 397 00:16:39.140 --> 00:16:41.223 I'll say I've worked in a couple different places, 398 00:16:41.223 --> 00:16:43.570 and I'm lucky to have an amazing team. 399 00:16:43.570 --> 00:16:46.430 And I want to give credit to this company that gave me 400 00:16:46.430 --> 00:16:49.050 the latitude to go by some of the principles 401 00:16:49.050 --> 00:16:51.970 that I had figured out along the way around who you look for 402 00:16:51.970 --> 00:16:55.110 and who you need as a good data scientist. 403 00:16:55.110 --> 00:16:57.670 I try to find a team that's complementary. 404 00:16:57.670 --> 00:17:00.248 Some people are good in computational stuff. 405 00:17:00.248 --> 00:17:02.040 Some people are good in theoretical things. 406 00:17:02.040 --> 00:17:04.300 Some people are good in operations research. 407 00:17:04.300 --> 00:17:06.280 Some people are good in Bayesian statistics. 408 00:17:06.280 --> 00:17:09.740 You try to get an orthogonal basis so that you can 409 00:17:09.740 --> 00:17:11.609 define the space pretty well. 410 00:17:11.609 --> 00:17:13.849 And then, just as far as people, I just 411 00:17:13.849 --> 00:17:15.660 look for people who are inquisitive. 412 00:17:15.660 --> 00:17:20.660 One of the nice things about hiring over the past year 413 00:17:20.660 --> 00:17:24.410 at Warner Music is, a lot of people who have a preternatural 414 00:17:24.410 --> 00:17:27.720 interest in music bubble up and [say], "Hey, wow -- 415 00:17:27.720 --> 00:17:29.980 I didn't know I could do music in data science," 416 00:17:29.980 --> 00:17:32.600 so they're excited about the subject matter. 417 00:17:32.600 --> 00:17:34.660 And those are people who've worked out 418 00:17:34.660 --> 00:17:36.380 and who've been fantastic and great. 419 00:17:36.380 --> 00:17:39.500 If you want to just level-set as far as ability, 420 00:17:39.500 --> 00:17:40.897 I give you a take-home problem. 421 00:17:40.897 --> 00:17:42.480 We have a couple different versions -- 422 00:17:42.480 --> 00:17:44.563 they're related to music and data science and what 423 00:17:44.563 --> 00:17:48.040 we're working on -- and just leave people to sit and think 424 00:17:48.040 --> 00:17:49.120 about them. 425 00:17:49.120 --> 00:17:51.290 And then, lastly, I look for people 426 00:17:51.290 --> 00:17:55.530 who we can enjoy and treat each other like people 427 00:17:55.530 --> 00:17:56.860 and as equals. 428 00:17:56.860 --> 00:18:00.140 For a team like this, where the work is hard and foundational, 429 00:18:00.140 --> 00:18:01.830 we need a lot of collaboration. 430 00:18:01.830 --> 00:18:04.140 We need a lot of communication. 431 00:18:04.140 --> 00:18:07.720 So, good personalities, where we act with one another almost as 432 00:18:07.720 --> 00:18:09.120 if we were family members. 433 00:18:09.120 --> 00:18:11.460 A lot of these things, you know, I learned -- 434 00:18:11.460 --> 00:18:12.543 I'll tell you the truth -- 435 00:18:12.543 --> 00:18:14.510 I learned as a professor finding grad students. 436 00:18:14.510 --> 00:18:16.140 It's a strong relationship. 437 00:18:16.140 --> 00:18:18.450 It's not quite parent-child, but it's 438 00:18:18.450 --> 00:18:20.930 a very strong relationship. 439 00:18:20.930 --> 00:18:23.590 SHERVIN KHODABANDEH: Kobi, we have a new segment in the show. 440 00:18:23.590 --> 00:18:27.500 We're going to ask you a few questions rapid-fire style. 441 00:18:27.500 --> 00:18:29.590 What's your proudest AI moment? 442 00:18:29.590 --> 00:18:30.338 443 00:18:30.338 --> 00:18:32.130 KOBI ABAYOMI: You know, we're turning sound 444 00:18:32.130 --> 00:18:35.340 into vectors of information that we can construct embeddings on. 445 00:18:35.340 --> 00:18:37.820 And then we look and see the topology 446 00:18:37.820 --> 00:18:40.470 of the space that we form after ingesting the sound. 447 00:18:40.470 --> 00:18:42.000 And I'll tell you, it works. 448 00:18:42.000 --> 00:18:43.580 For instance, I've been listening 449 00:18:43.580 --> 00:18:47.110 to a lot of what I'll call horn-driven, vocal, 450 00:18:47.110 --> 00:18:51.700 '70s rock lately: Redbone, Chicago, groups like that. 451 00:18:51.700 --> 00:18:54.470 And when you look at the model that we built, 452 00:18:54.470 --> 00:18:56.800 you'll get the other bands and sounds, 453 00:18:56.800 --> 00:18:58.383 even with all the variance and things 454 00:18:58.383 --> 00:19:00.300 that happened across different types of albums 455 00:19:00.300 --> 00:19:01.930 and different tracks, and you'll pick 456 00:19:01.930 --> 00:19:04.240 out the ones that you would say, like, "Oh yeah, yeah. 457 00:19:04.240 --> 00:19:04.740 Right. 458 00:19:04.740 --> 00:19:06.280 I would listen to those guys, too." 459 00:19:06.280 --> 00:19:08.650 I started listening to a lot more Hot Chocolate 460 00:19:08.650 --> 00:19:11.270 because of the model we use; they have this song, 461 00:19:11.270 --> 00:19:12.900 "Everyone's a Winner, Baby." 462 00:19:12.900 --> 00:19:16.310 That's something that came out of constructing this model. 463 00:19:16.310 --> 00:19:19.370 SHERVIN KHODABANDEH: What's your favorite activity that 464 00:19:19.370 --> 00:19:21.530 does not involve technology? 465 00:19:21.530 --> 00:19:22.907 KOBI ABAYOMI: I like to swim. 466 00:19:22.907 --> 00:19:24.240 I was a swimmer for a long time. 467 00:19:24.240 --> 00:19:25.140 I still like to get in the water. 468 00:19:25.140 --> 00:19:25.640 I love it. 469 00:19:25.640 --> 00:19:28.720 SAM RANSBOTHAM: What did you want to be when you were a kid? 470 00:19:28.720 --> 00:19:29.580 KOBI ABAYOMI: Oh! 471 00:19:29.580 --> 00:19:31.480 All sorts of things, but I'll say 472 00:19:31.480 --> 00:19:33.240 the one that was the most salient was, 473 00:19:33.240 --> 00:19:35.260 I wanted to do something with cars. 474 00:19:35.260 --> 00:19:39.880 I wanted to design cars, make cars, make engines run. 475 00:19:39.880 --> 00:19:42.120 [Laughs.] I loved cars. 476 00:19:42.120 --> 00:19:46.410 I loved them as just complicated devices. 477 00:19:46.410 --> 00:19:48.680 All these things need to work in concert, 478 00:19:48.680 --> 00:19:51.620 you know: chemistry, thermodynamics, physics. 479 00:19:51.620 --> 00:19:52.260 I loved cars. 480 00:19:52.260 --> 00:19:54.820 SHERVIN KHODABANDEH: What worries you about AI? 481 00:19:54.820 --> 00:19:58.870 KOBI ABAYOMI: People are worried about face recognition 482 00:19:58.870 --> 00:20:00.780 and things like that and how well that 483 00:20:00.780 --> 00:20:02.340 works across different groups. 484 00:20:02.340 --> 00:20:04.075 That's a bit disconcerting. 485 00:20:04.075 --> 00:20:06.200 When you think about the underlying problem and you 486 00:20:06.200 --> 00:20:09.270 think about what's likely being used underneath it, 487 00:20:09.270 --> 00:20:11.730 what's likely being used underneath it is [a] neural 488 00:20:11.730 --> 00:20:15.910 network classifier, which likely has a strong assumption 489 00:20:15.910 --> 00:20:18.570 for multivariate Gaussian distribution or whatever -- 490 00:20:18.570 --> 00:20:21.132 sort of pixilation of a face and things like that -- 491 00:20:21.132 --> 00:20:23.590 and you think about those couple of things and you're like, 492 00:20:23.590 --> 00:20:24.500 "Yeah, some things could go wrong." 493 00:20:24.500 --> 00:20:26.050 SAM RANSBOTHAM: What are you hoping 494 00:20:26.050 --> 00:20:28.300 we can get from artificial intelligence in the future? 495 00:20:28.300 --> 00:20:29.155 496 00:20:29.155 --> 00:20:30.530 KOBI ABAYOMI: So this is the part 497 00:20:30.530 --> 00:20:31.640 that I had a little canned speech for you. 498 00:20:31.640 --> 00:20:33.473 So when we think of artificial intelligence, 499 00:20:33.473 --> 00:20:36.120 we know as practitioners that it's not 500 00:20:36.120 --> 00:20:37.500 artificial intelligence, right? 501 00:20:37.500 --> 00:20:41.960 It's sort of augmenting intelligence, or auxiliary. 502 00:20:41.960 --> 00:20:44.760 It's intelligence that we're creating to perform 503 00:20:44.760 --> 00:20:46.890 different tasks for us. 504 00:20:46.890 --> 00:20:49.140 And I think that that's what we should expect from it: 505 00:20:49.140 --> 00:20:51.990 to release people to think about other things, right? 506 00:20:51.990 --> 00:20:54.453 Maybe one of the nice things if we do get self-driving 507 00:20:54.453 --> 00:20:55.870 down is that people have more time 508 00:20:55.870 --> 00:20:58.360 to read books and be introspective and read 509 00:20:58.360 --> 00:20:59.570 beautiful literature. 510 00:20:59.570 --> 00:21:01.970 And so, what I'm hoping is that we use it 511 00:21:01.970 --> 00:21:05.493 in ways that edify the human experience 512 00:21:05.493 --> 00:21:06.410 and not smash it down. 513 00:21:06.410 --> 00:21:07.910 SHERVIN KHODABANDEH: Very well said. 514 00:21:07.910 --> 00:21:08.410 515 00:21:08.410 --> 00:21:11.060 SAM RANSBOTHAM: Kobi, we really enjoyed talking to you. 516 00:21:11.060 --> 00:21:14.770 I think you're in an interesting space because as the production 517 00:21:14.770 --> 00:21:16.780 costs of music have dropped so much, 518 00:21:16.780 --> 00:21:19.920 it's created so much music, but you kept using the word curate 519 00:21:19.920 --> 00:21:22.142 a few times, and you definitely mentioned a few times 520 00:21:22.142 --> 00:21:24.600 it's an experience good -- that we don't know what music we 521 00:21:24.600 --> 00:21:26.058 like until we hear it. 522 00:21:26.058 --> 00:21:28.350 But once we've done that, we've already spent the time. 523 00:21:28.350 --> 00:21:31.960 And so this role of curation is very interesting, 524 00:21:31.960 --> 00:21:34.230 and it seems like a great application 525 00:21:34.230 --> 00:21:35.343 of your particular skills. 526 00:21:35.343 --> 00:21:37.510 Thank you for taking the time to talk with us today. 527 00:21:37.510 --> 00:21:39.220 SHERVIN KHODABANDEH: Thanks a lot for all your insights. 528 00:21:39.220 --> 00:21:39.390 529 00:21:39.390 --> 00:21:40.390 KOBI ABAYOMI: Thank you. 530 00:21:40.390 --> 00:21:42.340 SAM RANSBOTHAM: Thanks for listening. 531 00:21:42.340 --> 00:21:46.410 Next time, we'll talk with Ya Xu, head of data at LinkedIn. 532 00:21:46.410 --> 00:21:49.352 Please join us. 533 00:21:49.352 --> 00:21:50.810 ALLISON RYDER: Thanks for listening 534 00:21:50.810 --> 00:21:52.320 to Me, Myself, and AI. 535 00:21:52.320 --> 00:21:54.770 We believe, like you, that the conversation 536 00:21:54.770 --> 00:21:56.990 about AI implementation doesn't start and stop 537 00:21:56.990 --> 00:21:58.170 with this podcast. 538 00:21:58.170 --> 00:22:00.660 That's why we've created a group on LinkedIn, specifically 539 00:22:00.660 --> 00:22:01.780 for leaders like you. 540 00:22:01.780 --> 00:22:04.530 It's called AI for Leaders, and if you join us, 541 00:22:04.530 --> 00:22:06.560 you can chat with show creators and hosts, 542 00:22:06.560 --> 00:22:10.160 ask your own questions, share insights, and gain access 543 00:22:10.160 --> 00:22:12.660 to valuable resources about AI implementation 544 00:22:12.660 --> 00:22:14.750 from MIT SMR and BCG. 545 00:22:14.750 --> 00:22:19.880 You can access it by visiting mitsmr.com/AIforLeaders. 546 00:22:19.880 --> 00:22:22.590 We'll put that link in the show notes, 547 00:22:22.590 --> 00:22:25.030 and we hope to see you there. 548 00:22:25.030 --> 00:22:30.000