Latest / The Adults In The AI Room / Adam Sutherland on AI and Personalization in Media & Entertainment - Episode #4

Transcript

Adam Sutherland on AI and Personalization in Media & Entertainment - Episode #4 · The Adults In The AI Room

• 0:04
  SPEAKER_00: Listening to the money and the interpretation to the future of cognitive assistance in money in all aspects of personal and personal life.
• 0:22
  SPEAKER_00: Thanks so much for listening.
• 0:23
  SPEAKER_00: If this is your first time, we invite you to subscribe wherever you listen to podcasts.
• 0:28
  SPEAKER_00: And now here are your hosts, Seth Early and Chris Featherstone.
• 0:36
  SPEAKER_03: Welcome to today's podcast.
• 0:38
  SPEAKER_03: I'm Seth Early.
• 0:39
  SPEAKER_01: And I'm Chris Featherstone.
• 0:41
  SPEAKER_01: Thanks.
• 0:41
  SPEAKER_03: I'm really excited about our guest today.
• 0:43
  SPEAKER_03: He's a multidimensional renaissance man with a broad array of interests.
• 0:47
  SPEAKER_03: He's currently leveraging his extensive background in media and publishing as the worldwide business development lead for data science and analytics in media and entertainment at Amazon Web Services.
• 0:58
  SPEAKER_03: Please welcome Adam Sutherland.
• 1:00
  SPEAKER_03: Hello, thank you.
• 1:01
  SPEAKER_03: Nice to be here.
• 1:02
  SPEAKER_02: And thank you for that very glowing introduction.
• 1:04
  SPEAKER_01: I feel like there should be a lot of applause too.
• 1:06
  SPEAKER_01: Yeah.
• 1:08
  SPEAKER_01: Everybody's standing ovacious.
• 1:09
  SPEAKER_01: That's right.
• 1:10
  SPEAKER_01: And the soundtrack.
• 1:10
  SPEAKER_01: Yeah.
• 1:11
  SPEAKER_03: You should have roaring applause.
• 1:12
  SPEAKER_03: Yes.
• 1:13
  SPEAKER_03: And Chris, I know you know Adam from your Amazon days.
• 1:16
  SPEAKER_03: So do you want to give us a color background?
• 1:19
  SPEAKER_01: Yeah, I mean, I'll ask him to maybe put his earphones on mute as I talk about him because yeah, he's one of my favorite humans, not only at Amazon, but just in general.
• 1:28
  SPEAKER_01: And I think the key, you know, is it's rare when you meet um just a genuinely great person that is good in all these aspects and stuff.
• 1:36
  SPEAKER_01: And so yeah, I was excited that um Adam that you would even say yes to doing this with us because when our history, just in terms of working together, which I learned an amazing amount of stuff from you.
• 1:48
  SPEAKER_01: And I think you have, you know, an unbelievable perspective that I think the world needs to know.
• 1:53
  SPEAKER_01: And not that we're going to be the ones that expose everything that you know in your head, but just in terms of, you know, if if folks can understand you and uh and understand what you do, then I think they'll be better for it.
• 2:05
  SPEAKER_01: So thanks for the time.
• 2:08
  SPEAKER_02: Really appreciate it.
• 2:08
  SPEAKER_02: I've been blessed, slash, very lucky to to work with some really great people uh over the years, including yourself, Chris.
• 2:15
  SPEAKER_02: So thank you for that.
• 2:16
  SPEAKER_02: I appreciate it.
• 2:17
  SPEAKER_03: So I just a little bit of uh background here.
• 2:20
  SPEAKER_03: I understand you have an undergraduate in Asian studies.
• 2:24
  SPEAKER_03: What does that mean?
• 2:25
  SPEAKER_03: And uh uh do you speak?
• 2:28
  SPEAKER_03: Yeah, yeah, yeah.
• 2:30
  SPEAKER_03: So uh the yeah, so actually Japanese guy know is Taihenoshikata desu, anatola sushi o put in the guide, Josu Tesu.
• 2:37
  SPEAKER_03: And I'm sure I totally missed that.
• 2:39
  SPEAKER_03: It's something about you're really good at making sushi or something.
• 2:43
  SPEAKER_04: Yeah, okay.
• 2:45
  SPEAKER_03: I used to own a sushi bar many, many uh lives ago.
• 2:49
  SPEAKER_03: Oh, hilarious.
• 2:51
  SPEAKER_02: Yeah, so my undergrad was actually in um Japanese literature, uh, which um anyway, so with I did have a minor in business, which was you know, was the only way that my parents would agree to pay for college if I if I didn't just do Japanese literature.
• 3:08
  SPEAKER_02: The you know, the the typical hubris of an 18-year-old, I I I wanted to um study something that would never get boring.
• 3:14
  SPEAKER_02: And so I thought another culture, right?
• 3:16
  SPEAKER_02: Because there's a lot to another culture and uh eventually landed on literature because that as a window into another culture.
• 3:23
  SPEAKER_02: Of course, now with you know the the benefit of not being a uh uh 18-year-old, I realized that you know you can study anything.
• 3:32
  SPEAKER_02: And if you study it well enough, it'll never get boring.
• 3:34
  SPEAKER_02: But at that time, I yeah.
• 3:36
  SPEAKER_03: Just curious, did you ever do any martial arts?
• 3:38
  SPEAKER_03: And did you spend much time in Japan?
• 3:41
  SPEAKER_02: Yeah, so I I actually lived in I lived in Japan for uh five years total.
• 3:46
  SPEAKER_02: Um, and yeah, I did martial arts, I did uh Thai boxing and some and some other stuff, uh, starting when I lived in Chicago.
• 3:54
  SPEAKER_02: And then, you know, it's funny.
• 3:56
  SPEAKER_02: So there's a very similar kind of path here.
• 3:58
  SPEAKER_02: So studied Japanese in college, started martial arts, you know, in the US, and then moved to Japan and uh tried to speak Japanese with actual Japanese people and also don't joined a dojo with a few uh with a few world champs and uh humble, you know, humble pie across the board.
• 4:18
  SPEAKER_03: Thai boxing, you said?
• 4:20
  SPEAKER_03: Oh that's yeah, muay Thai boxing, yeah.
• 4:22
  SPEAKER_03: That is brutal, brutal elbows, knees, anything goes.
• 4:25
  SPEAKER_03: I did Shotokan and then I yes, I I have a black belt and shot a con, but I'm doing aikido because that's much easier for old people.
• 4:35
  SPEAKER_02: Yes, yeah, you know, I I had I had a I had a really good friend in that I with whom I lived in Japan, and yeah, Aikido is awesome.
• 4:44
  SPEAKER_02: And you know, once again, time of life, but also personality-wise, like it just didn't fit for me.
• 4:49
  SPEAKER_02: Like I just, if somebody was gonna cause some grief, I felt that once in a while maybe I should be the first one to to you know make sure that we uh stop the trouble.
• 4:58
  SPEAKER_02: Whereas I had a friend who was very much on the other end of the spectrum, very much pet, and he took to Aikido like a fish to water and got like his Don, and then he was teaching in Chicago.
• 5:08
  SPEAKER_02: So yeah, it's a beautiful, beautiful art sport, whatever you want to call it.
• 5:12
  SPEAKER_02: But it is do them both.
• 5:14
  SPEAKER_02: So the okay, yeah, good, exactly.
• 5:16
  SPEAKER_02: Passive and aggressive.
• 5:18
  SPEAKER_02: That's right.
• 5:18
  SPEAKER_02: Yeah.
• 5:19
  SPEAKER_01: So, Adam, you've got an amazing background to right, and I think the the culture and the depth.
• 5:25
  SPEAKER_01: Um, I'd love to get a take, and for those listening, because most of this that we're trying to, you know, to do is not just another podcast about AI, but more on the lines of kind of the human experience of how one goes from where they're at in their career to either embracing it, practicing it, you know, um, creating, you know, all this new innovation and how it hits them.
• 5:51
  SPEAKER_01: Love to get an idea from you and your kind of your history on, you know, through your career, how did it lead you to, let's say, what you're working on today and what's you know, what's been you know the impact.
• 6:02
  SPEAKER_02: Yeah.
• 6:03
  SPEAKER_02: So um the career has been kind of a combination of tech and content pretty much from the beginning, you know, after except for my first stint in Japan.
• 6:13
  SPEAKER_02: But even there, I was teaching English and then ended up kind of starting up a school and doing all things like that.
• 6:18
  SPEAKER_02: So um I I then worked, believe it or not, for Encyclopedia Britannica.
• 6:23
  SPEAKER_02: Uh we, you know, we had uh and learned an important lesson at my time at Britannica, which is very much kind of like the Panasonic and uh uh VHS um Betamax conversation.
• 6:38
  SPEAKER_02: Uh uh Encyclopedia Britannica actually was one of the we had the first CD-ROM, we had the first website, we had all the other other kinds of things.
• 6:45
  SPEAKER_02: We were actually very cutting edge from a technology perspective.
• 6:48
  SPEAKER_02: The business model did not allow us to take advantage of it.
• 6:52
  SPEAKER_02: You had a sales force that was basically making money on big ticket and uh commissions, and so that just it was in, you know, we couldn't get rid of it.
• 7:02
  SPEAKER_02: So anyway, the so doing kind of tech for uh in Secretary Britannica, and then actually moved from Britannica over to Discovery Channel as the head of uh Discovery Online.
• 7:16
  SPEAKER_02: Long story short, moved around in uh the like I said, that tech content space, moved into MA, where I would, you know, worked with Discovery and uh Travel Channel and others to look on uh to figure out, you know, we have content, we have a brand.
• 7:31
  SPEAKER_02: And then uh to take care to take advantage of that in new platforms, we have to build by borrow, right?
• 7:40
  SPEAKER_02: And so I was the buy-borrow guy.
• 7:43
  SPEAKER_02: So figure out who we partnered with, what platforms we might, you know, actually end up acquiring, or content that we would acquire.
• 7:50
  SPEAKER_02: Um uh before I kind of got really steeped in AIML, I was at uh National Geographic as the head of strategy, and then did a complete pivot over to a little startup in the area which did speech to automatic speech recognition, so ASR and machine translation, small operation that had mostly kind of a government uh customer set.
• 8:14
  SPEAKER_02: And they brought me on as the hired CEO to help them figure out what we could do in media.
• 8:20
  SPEAKER_02: And it turns out in in telco as well.
• 8:23
  SPEAKER_02: So uh the kind of a long-winded way to answer your question, I got into it from not from the technology, although you know, have love technology and it's always been about how does that technology solve a problem.
• 8:35
  SPEAKER_02: Specifically when it comes to AIML, it was that.
• 8:38
  SPEAKER_02: It was all right, we've got this cool technology.
• 8:41
  SPEAKER_02: We know that there's some early indications that media could use it for captioning and for, you know, um uh asset uh metadata and and call center.
• 8:52
  SPEAKER_02: And so got into there from like, okay, how do I turn that technology and make it into a business?
• 8:58
  SPEAKER_02: Uh and so that was how you know I got there.
• 9:01
  SPEAKER_02: What I find interesting, and there's a we do a thing uh at my current gig where we will uh spend some time talking to high school kids about how to get into the high the world of high tech.
• 9:13
  SPEAKER_02: And it it the to me, it's a few things.
• 9:16
  SPEAKER_02: One, you don't necessarily have to be an engineer.
• 9:18
  SPEAKER_02: I think that's a super important point, right?
• 9:21
  SPEAKER_02: The there there's there's room for engineers and there's room for people who you know work backwards from a business perspective.
• 9:28
  SPEAKER_02: Um, but at the end of the day, it's just remaining curious, right?
• 9:32
  SPEAKER_02: And and and not uh um not ever kind of saying, you know, I I'm I I'm afraid to learn something new, I'm afraid to go into or I already know that.
• 9:43
  SPEAKER_02: Or I already know that, yeah, exactly.
• 9:45
  SPEAKER_02: Or I know exactly how to do that.
• 9:46
  SPEAKER_02: And it's been I've done it a hundred ways, a hundred times this way.
• 9:49
  SPEAKER_02: So that's got to be the way to do it.
• 9:51
  SPEAKER_02: Um so yeah, so the that move from NetGeo to that the small startup was fantastic because I brought the industry experience and I knew nothing about AIML, like zero.
• 10:06
  SPEAKER_02: And next thing I know, I've got a team of 40 scientists sitting at universities around the world and all this, and it was great because we could just learn from each other uh, you know, for you know the first six to nine months.
• 10:17
  SPEAKER_02: So it was it was awesome.
• 10:19
  SPEAKER_01: We get into a lot of the de you know demystifying the we'll say the dark arts or the magic arts of AI, right?
• 10:26
  SPEAKER_01: And I think you know, the cool part of what you said is, and this is what we like to really glean out is that there is no one size fits all.
• 10:36
  SPEAKER_01: In fact, there is no, you know, there's so much semantics that go into it.
• 10:42
  SPEAKER_01: And what I hear you saying is I know this industry, I'm being asked to apply new ways and innovative ways of actually doing more with less if you want to go cliche, but doing more and creating more innovation and more time to value, you know, less time to value, but more more time um, you know, for these types of things.
• 11:02
  SPEAKER_01: And that's really all it is, is actually just asking those initial questions and then putting smart people around you to help answer them.
• 11:09
  SPEAKER_02: Yeah, uh from my perspective, absolutely, right?
• 11:11
  SPEAKER_02: So it but but I'm not the one building the models, right?
• 11:15
  SPEAKER_02: I'm not the one doing the data engineering.
• 11:17
  SPEAKER_02: Um it is it's a really interesting thing.
• 11:21
  SPEAKER_02: You know how that whenever whenever whenever something becomes so kind of important, it it almost kind of becomes transparent, right?
• 11:27
  SPEAKER_02: So electricity, of course, is the the old example.
• 11:29
  SPEAKER_02: I think eventually we'll get there, right, with with AIML.
• 11:33
  SPEAKER_02: You're gonna have people, of course, you're gonna have people who are building these things, but what the the actual availability of, you know, to be to put it in AI speak, the actual availability of an inference endpoint to solve the problem that you've got, that's I don't want to say ubiquitous, but it's gonna be pretty easy to find those things for most use cases.
• 11:55
  SPEAKER_02: Uh and the real heavy lifting is gonna be for specific workflows, workloads that uh that require problems and very specific business problems, right?
• 12:05
  SPEAKER_03: Yes, exactly.
• 12:06
  SPEAKER_03: Question.
• 12:06
  SPEAKER_03: So so what's a day in the life for Adam look like these days?
• 12:10
  SPEAKER_03: What do you what do you spend your time day to day?
• 12:12
  SPEAKER_03: Give us an example of how you interact with your stakeholders, with your team saying.
• 12:18
  SPEAKER_02: Yeah, so um the as once again, we're gonna keep with cliches.
• 12:23
  SPEAKER_02: Uh, I like to look at everything as a portfolio, right?
• 12:25
  SPEAKER_02: And so there's a certain amount of consumption and a certain amount of production.
• 12:29
  SPEAKER_02: Um, from my perspective, the production is uh I will be pulled into conversations with customers, um, you know, willingly, I make it sound like they're, you know, dragging me, kicking, and screaming, get pulled into conversations with customers to help with that uh that initial triage, right?
• 12:46
  SPEAKER_02: So either it's the the the interesting ones, but which is coming less and less, actually.
• 12:52
  SPEAKER_02: But when I first started in this role, the number of times I would get into a conversation with somebody who said, I need AI.
• 12:58
  SPEAKER_02: Right.
• 12:59
  SPEAKER_02: Right.
• 12:59
  SPEAKER_02: And that should be on the downside.
• 13:01
  SPEAKER_02: That should that should definitely be declining.
• 13:04
  SPEAKER_02: It is.
• 13:05
  SPEAKER_02: It is.
• 13:05
  SPEAKER_02: And so now we have conversations about, hey, I have this specific use case, and it's more around is there a managed service that can do it, or do I need to um have my own uh my own data scientists?
• 13:18
  SPEAKER_02: And so go ahead.
• 13:20
  SPEAKER_03: Well, what kinds of problems?
• 13:21
  SPEAKER_03: Like what do you mean you say you get pulled in?
• 13:23
  SPEAKER_03: They're not saying I want AI generically, but what are they trying to give some examples of types of things people are trying to do?
• 13:30
  SPEAKER_02: Yeah, so obviously it's very industry specific, so I'll focus on on media at this point.
• 13:35
  SPEAKER_02: There's probably the biggest one and the one that we are um most frequently dealing with is just this idea of understanding X.
• 13:48
  SPEAKER_02: And so and the X is generally content or customer generally, okay.
• 13:53
  SPEAKER_02: So when it comes to content, you have companies that have huge libraries or small libraries, they've got an archive.
• 14:00
  SPEAKER_02: And because of the way that the industry has historically worked, the metadata in most of those assets is where they are very buttoned up is on the metadata around stock footage or music rights or things like that.
• 14:16
  SPEAKER_02: Any place where you could get sued, right?
• 14:19
  SPEAKER_02: Or any place where mess messing up means that you're gonna um either owe somebody money or you're gonna miss out on money that you are owed.
• 14:27
  SPEAKER_02: Uh, what is less kind of buttoned up is what's actually in my content.
• 14:33
  SPEAKER_02: I'll give you an example.
• 14:35
  SPEAKER_02: I I worked at national, I was the chief strategy officer at National Geographic for uh four and a half years, and we spent a lot of money every year getting third-party footage that we knew we had somewhere in the archive.
• 14:52
  SPEAKER_02: But finding it, so the difference between knowing there's an elephant in that clip and going in and finding the actual clip that would work for the show, that effort was cost us more in terms of manpower, et cetera, than just buying a clip from a third party.
• 15:07
  SPEAKER_02: Um, and that's and it's not that infrequent.
• 15:10
  SPEAKER_02: So that's one of the big use cases is using AI to transcribe, to use image recognition, video recognition, um, to do maybe some NLP on top of that to really get a better idea of what's inside your content.
• 15:23
  SPEAKER_03: Which is not far from what organizations are doing with their dark data, you know, when they when they stuff that they don't know what is.
• 15:29
  SPEAKER_03: I have all this unstructured information, it's it's all over the place.
• 15:33
  SPEAKER_03: I don't really know what's there.
• 15:34
  SPEAKER_03: And so I'm trying to use AI to mine it, to text mine it, to yeah, what the value is and to make a clustering model so you can start grouping things into, yeah.
• 15:44
  SPEAKER_02: Yeah.
• 15:44
  SPEAKER_02: We don't so it the uh uh there is a really a really interesting use case with with uh a particular customer where they have spent a lot of time um taking movie clips and tagging them for kind of emotions, right?
• 16:03
  SPEAKER_02: So the the the the use case that I just discussed, where you take a clip, you run it through, you'll get a bunch of tags, right?
• 16:10
  SPEAKER_02: Car, house, animal, you know, you'll get Chris Featherstone if his face is in the library, like you get all that information.
• 16:18
  SPEAKER_02: It and that's the first pass.
• 16:20
  SPEAKER_02: The next pass is this idea of okay, well, well, if I see Cafe, Arc de Triomphe, this is a very obvious one, right?
• 16:27
  SPEAKER_02: It's a Cafe Arc de Triomphe and you know, something else.
• 16:29
  SPEAKER_02: Oh, we're in Paris.
• 16:30
  SPEAKER_02: Okay.
• 16:31
  SPEAKER_02: That next layer of kind of meta uh assumption on top of the tags that you've got is really interesting.
• 16:38
  SPEAKER_02: There's a company that we've got that is uh that we're working with that did a lot of that uh tagging themselves.
• 16:44
  SPEAKER_02: So they took a clip and they said, this is you know, triumphant, this is uh teamwork, these are these kind of really amorphous concepts.
• 16:52
  SPEAKER_02: But now that they've got all that tag data, we're working with them to build a model where you can do that off of the metadata that's created by these managed services.
• 17:02
  SPEAKER_02: So when you see smiling and more than three people and you know, this kind of stuff, you can assume it's teamwork, for example.
• 17:10
  SPEAKER_02: So there's a lot of really interesting stuff that you can just that basic tagging that I talked about is a huge improvement and and provides a lot of value.
• 17:19
  SPEAKER_02: Once you've got it, there's all kinds of cool stuff you can do on top of it.
• 17:23
  SPEAKER_03: Well, I have one question.
• 17:25
  SPEAKER_03: Uh, I want to know how come I can't find anything on my thousands of streaming apps and uh all the TV opportunities.
• 17:34
  SPEAKER_03: I go in and sometimes search isn't working, or I go for a category, there's no category, or you know, it's just yeah, it seems all over the place on these days, the wild west.
• 17:46
  SPEAKER_02: Yeah, it's the there's definitely a streaming fatigue uh in the it's a good question.
• 17:53
  SPEAKER_02: So, you know, the the this goes back to the knowing your customer, right?
• 17:56
  SPEAKER_02: So it's a perfect segue to knowing your customer.
• 17:58
  SPEAKER_02: Um, the other thing that we hear a lot from media companies is you know, there used to be a really clear delineation of work, right?
• 18:08
  SPEAKER_02: So I I'm you know, XYZ company, I make content, and then I hand it off to somebody else who distributes content, and then they, you know, and then they work with an advertising agency to monetize that content.
• 18:20
  SPEAKER_02: Those clear swim lanes are anything but clear these days.
• 18:24
  SPEAKER_02: So you've got companies, you know, born in the cloud, like Netflix, for example, who from the beginning they they had that whole value chain.
• 18:32
  SPEAKER_02: So they had to understand what the customer was doing and what they were, you know.
• 18:35
  SPEAKER_02: Um, and so of course, one of these one of the big requests is to personalize uh content uh based on what somebody has done in the past.
• 18:45
  SPEAKER_02: Uh lots of lots of complications there, right?
• 18:49
  SPEAKER_02: If people share accounts, I think we talked about this before.
• 18:51
  SPEAKER_02: If you share accounts, then it doesn't always work.
• 18:55
  SPEAKER_02: The other thing is that you it you've got to, you don't have to, but most of the folks that I talk to are very cognizant about not going too far down that route because you create like an echo chamber.
• 19:06
  SPEAKER_02: So there's this idea of manufactured serendipity where 80% of the recommendations will be recommended, and then 20% are totally random, for example, just to make sure that something is interjected uh pretty frequently.
• 19:21
  SPEAKER_02: Seth, I think you might be on mute.
• 19:23
  SPEAKER_03: Sorry, my my wife, oh thank you.
• 19:26
  SPEAKER_03: My wife is always using my profile when she's watching stuff.
• 19:30
  SPEAKER_03: And uh I heard of some guy who changed the icon, changed the title of the icon uh of his profile to update windows now.
• 19:41
  SPEAKER_03: Oh, that's hilarious.
• 19:42
  SPEAKER_03: No one touched it.
• 19:44
  SPEAKER_03: Yeah, yeah, yeah.
• 19:46
  SPEAKER_03: He kept his profile sacred, but that's funny.
• 19:50
  SPEAKER_02: My my home internet, my home Wi-Fi, I named it ransomware for the same reason.
• 19:56
  SPEAKER_02: Just, you know, if anybody is trying to hack in that they would they would avoid it.
• 20:00
  SPEAKER_01: I just called mine honeypot, right?
• 20:02
  SPEAKER_01: Honeypot, you saw.
• 20:06
  SPEAKER_02: So the the you know, I mean we talked about this before as well, I believe, but one of the best things that Amazon.com ever did, in my opinion, was that little checkbox for this is a gift.
• 20:18
  SPEAKER_02: Because the the power of Amazon was that collaborative filtering and that, you know, well, of course, the availability of all the things they've got.
• 20:25
  SPEAKER_02: But also I go in and I see stuff that they get me.
• 20:28
  SPEAKER_02: You know, I'll be honest.
• 20:29
  SPEAKER_02: I go in, I'm like, oh yeah, I actually do want that.
• 20:32
  SPEAKER_02: When I started to use it to buy presents for people, especially my kids, it messed everything up.
• 20:38
  SPEAKER_02: And so I got all of that, I started getting a jumbled, but and then once again, and once they entered that, you know, this is a gift, and I was able to say, okay, I'm buying this, but don't put it in my profile, especially essentially, it went back to being really, really uh uh efficient, efficient and targeted.
• 20:55
  SPEAKER_01: Adam, I'd love to get your take too, in terms of you know, you're you interact with customers all day, every day, especially with media and content, search and discovery, like monetization as well as moderation.
• 21:09
  SPEAKER_01: You know, those are big areas, right?
• 21:11
  SPEAKER_01: But what do you see like as a common thread as the biggest barrier to entry for a lot of these orgs and what they're you know trying to do?
• 21:18
  SPEAKER_02: So, you know, the the biggest maybe it's too big, right?
• 21:22
  SPEAKER_01: Too big a question, but you know, well get your take.
• 21:24
  SPEAKER_02: So yeah, so uh I'll tell you what we see a lot, and and and I'll I'm happy to dive down into those different use cases and talk a little bit more about them.
• 21:32
  SPEAKER_02: But what we see a lot is um where before the conversation was, hey, I want AI, right?
• 21:39
  SPEAKER_02: The now the conversation is I want um I want to use your services to help me uh monetize my content, right?
• 21:46
  SPEAKER_02: To use the example you're talking about.
• 21:48
  SPEAKER_02: And so we say, yeah, sure, you can use transcribe and you know recognition in our case, right?
• 21:54
  SPEAKER_02: You can use these tools to find the tags where you, you know, and figure out where you might want to put an ad, but you very quickly figure out that the people you're talking to don't control the value chain.
• 22:05
  SPEAKER_02: And and you know whether it's the value chain or it's the data sources or it's all these other things like the the you either have conversations with people who are way down in the weeds of the technology or you have conversations with people who are trying to solve business outcomes.
• 22:21
  SPEAKER_02: And either one you realize that they're not talking to each other.
• 22:25
  SPEAKER_02: So it's it it's working with customers to say, okay, yes, we can do this.
• 22:29
  SPEAKER_02: And in some cases you can do I don't want to kind of call it cosmetic because there's still value there.
• 22:34
  SPEAKER_02: You can do what they're asking and you can give them some of these benefits when you know in the back of your mind that the real value would be if they integrated this through their whole stack.
• 22:45
  SPEAKER_02: And so one of the things that we we try to do uh so to your point Chris when you once once you have better metadata you can there's search and discovery right you can use these tools to help with things like captioning and subtitling you know for uh accessibility and and global distribution you can do content moderation so you can find things that might be inappropriate from one market but you know appropriate for another and then content um monetization what we try to do is we try to tell folks that uh when we can to to do the first part to do the tagging right even if you're not ready to do all those use cases do the tagging because you know for the use case that you are ready for whether that is search and discovery or captioning what have you once you've got it you've got that foundation then you can do all the other stuff um but the biggest the biggest hurdle is generally finding talking to a customer who has the connects the dots across the different inputs to get where you want to be when you talk about tagging I mean someone once said to me just tag everything with everything and it was hard to explain to them that that's not really going to work right yeah yeah everything with everything.
• 24:06
  SPEAKER_02: So you know what should how what level of granularity that should they be thinking about especially with these types of assets and uh what standards should we think about and when did it start getting meeting uh bespoke or custom uh structures metadata structures yeah good very good question so most of the services you know from us and others will allow you to set uh what we call a confidence interval or confidence score and so the uh and we have this I have this conversation all the time the trade-off being if you're gonna do this for metadata okay like would you rather have false positives or would you rather miss a few things okay and uh it depends on use case so you have the ability all the outputs of these managed services are JSON files which maintain that that confidence score so you can have a customer for example who is comfortable with a 60% confidence score for the transcription if they're using it for captioning and subtitling because there's going to be a human in the loop you know down the the the workflow um and uh or and then you have at the other end of the spectrum where somebody wants 80 to 90 percent you know uh assurance that this is something that needs to be moderated or vice versa depending on what their workflow looks like so you can use that confidence score as a trigger.
• 25:29
  SPEAKER_02: Uh the to your question about when does something need to be bespoke, a generalized model is a generalized model, right?
• 25:37
  SPEAKER_02: And most of the services that we've got and and others of course have the ability to create some customization whether it's a vocabulary or a face library or things like that.
• 25:48
  SPEAKER_02: But the more specific the use case the more likely you're going to want to use a bespoke model.
• 25:55
  SPEAKER_02: I'll give you an example uh you know a lot of these a lot of the recognition services including recognition from from AWS they uh you can go through and you can find all the things I mentioned right you've got you know the general labels in terms of cat dog indoor outdoor face happy sad all this kinds of stuff all these kinds of stuff things talked to a customer recently who wanted to do um wanted to be able to say I'm gonna film this thing and people are going to be doing sporting events sporting things okay I want to know not only which sport they're doing or what sport movement they're doing I want to be able to know whether they're doing it well or not and so that's exactly the kind of scenario where I said hey you know out-of-box services generalized models not going to work for you exactly but but you can you know build your own and here's how we've done it in the past et cetera right I think I think one thing to to help everybody remember too Seth is that when we're talking about this, you know, Adam had brought up we got managed services and then we have custom bespoke models that people can run on their own.
• 27:04
  SPEAKER_01: And you know the notion between the differences an organization has put one together and put it out there and created all the security all the scalability all the high availability around it with some generalized models that anybody could use as their infrastructure whereas the inverse of that is build it all yourself.
• 27:23
  SPEAKER_01: Here's some maybe some TensorFlow or some PyTorch or something like that to get started.
• 27:29
  SPEAKER_02: However everything yeah exactly so well and that's you know it's important to know but anyway go ahead Adam no no absolutely that that's exactly right and and the other thing so the one thing is how specific is your use case right the other thing is like where what is this to your business right so is is what you is what you is the outcome that you're looking for is that a differentiator for you right so are you a I'm making this up but are you a translation house?
• 27:58
  SPEAKER_02: Okay if you're a translation house then a a a generalized translation model is not going to give you any advantage to the translation house down the street if you're both using the same generalized model.
• 28:10
  SPEAKER_02: So you know if it's a strategic differentiator for you then it's worth the investment in the data scientists competitive advantage right yeah exactly exactly competitive advantage yeah exactly so that that's another that's another conversation that we have all the time um yeah so uh and then the what was gonna say and and I did want to say that the the the use cases that we mentioned in AI you know AI is getting better all the time and it's really once again the conversation that I'm not having that I used to have is like why is this not perfect?
• 28:47
  SPEAKER_02: I thought that AI would could be perfect and most people understand now that that it is it is you know general um but the the there are scenarios and and captioning and content moderation are two of them where oftentimes you're still going to want a human in the loop right but you're just gonna use AI for that initial heavy lifting uh and it does two things it it reduces the amount of of just work that or media that not somebody needs to go through.
• 29:20
  SPEAKER_02: The second thing is it actually when it comes to captioning for example it can change the nature of the workforce you need.
• 29:27
  SPEAKER_02: So you you no longer need somebody who is a trained captioner.
• 29:31
  SPEAKER_02: By the way you know the the the demand is outstripping the supply for captioners at this point.
• 29:38
  SPEAKER_02: So you you can know you no longer need somebody who's a trained captioner you just need somebody who can do post editing.
• 29:43
  SPEAKER_02: And so it changes kind of the nature of what you have to do in terms of who you hire and and how you set up your business.
• 29:49
  SPEAKER_03: I wanted to ask a quick question about personalization of recommendation.
• 29:54
  SPEAKER_03: Recommendation engines are really important across lots of areas personalization is really important you know in many different contexts it's it's really you know people talk about information overload it's filter failure right the right information for me based on what I'm looking for based on my my characteristics based on my interests and so on how different and how much of a competitive advantage are you know are these recommendation ends engines becoming for the various uh media organizations and I would imagine they can be quite different and they're also going to look at different signals and you know do they do they have um a greater fitness to purpose for certain genres or certain types of users or certain segments or talk a little bit about um recommendation engines.
• 30:44
  SPEAKER_02: Sure so uh first off there's two there's two big categories at least the the the way that I look at it there's two big categories there's content for people and then there's people for content right so content for people is the what I you know the the kind of core we're talking about here you've got a lot of stuff and you show up and you you know log into your account and you want to see something there that is similar to what you watched before and you're you know it's gonna make you happy.
• 31:09
  SPEAKER_02: People for content is uh you know how do you grow your audience?
• 31:13
  SPEAKER_02: How do you have the right marketing messaging to the right cohort and things like that.
• 31:16
  SPEAKER_02: And and currently you you have you have models like algorithms that can do both but the underlying data that trains it is a little bit different.
• 31:25
  SPEAKER_02: And so to your point Seth about kind of you know how do these compare and the not an expert in personalization but this is the way that I see it the algorithms the methods the whatever those are those are out there.
• 31:41
  SPEAKER_02: There's not like a secret algorithm that somebody has built and you know so it really comes down to as it does with a lot of things in this space it comes down to data governance.
• 31:51
  SPEAKER_02: What are you collecting you know how well uh and consistent are you with the how you manage that data and and you feed all that into the different algorithms um and then you can tweak I mean some of the customers we work with will have three or four different personalization engines running at any one time.
• 32:10
  SPEAKER_03: And the hyper parameters of those engines the tweaking and the tuning of okay how are we going to say what level of injection of random recommendations will we have and then what we will provide those and then what is the algorithm going to base that those uh recommendations on what aspects of the characteristics of the content or the user or the behaviors right so yes Yeah exactly so and where where the the two things that I mentioned before knowing your customer and knowing your content where they where they do uh intersect is in this area of personalization.
• 32:43
  SPEAKER_02: And so you've got and I don't mean to make it sound like any of this is is you know uh easy peasy um but there's there are things that people know well and to your point like what's the data how do we you know train the model what is important what gets the right weights all that kind of stuff and and for the most part when when you're training machine learning models you don't you don't tell the model that you put the information in is then it comes back and it tells you what the weights are and what the relative importance is and things like that.
• 33:12
  SPEAKER_02: But two things one to your point there there I know of customers who have for example video rails where one rail is editorial editorially curated.
• 33:23
  SPEAKER_02: There's a human being that says hey here's the cool stuff going on.
• 33:26
  SPEAKER_02: The next one is just for you.
• 33:28
  SPEAKER_02: The one below that is based on another algorithm so that they and then they constantly are optimizing in terms of where those rails fall right so you show up and maybe they've decided that they're always going to have the human curation rail at the top but everything else is up for grabs based on what you interact with.
• 33:49
  SPEAKER_03: Human curation is based on kind of merchandising and what you want to put out there and how you understand your segment and it's kind of that and your brand that artisan curation right yeah exactly and and the and that kind of brand differentiation right here's what we you know here's what we think is is good for you.
• 34:05
  SPEAKER_02: I mean you know the New York Times bestseller list and you know the the fact that you go there and you're like well these are people who read a lot of books and they you know they can tell me what's good you know um what I find really interesting and this is not there are some out there doing it and by the way you know Netflix is at the cutting edge of all this stuff in case you in case you didn't know I'm pretty sure everybody knows but what what I find really interesting is you know we have this concept of a data lake and and one of the things about a data lake that is that is interesting is you can take a bunch of content a bunch of data I'm sorry and you collect the data and you kind of keep it in low cost storage because you might need it someday.
• 34:51
  SPEAKER_02: Okay so the a great example is you know people track where customers fall off in a video how long they consume a video do they watch the whole thing do they say whatever and that's been around for a long time.
• 35:05
  SPEAKER_02: Once you start to so if you but if you marry the content metadata with that information you can start to say by demographic well hold on a second my demographic drops off after generally after this this this and this right generally after a really violent scene generally after you know XYZ and so you can start to actually make editorial content decisions whether that's what you you know what you leave what you buy what you um create yourself all these other things based on the behaviors you're seeing within the content not just that on kind of like that whole level but you can or you can start to say you know not just you know binary but you can say I'm gonna keep making action films but I'm gonna dial down the I'm I use violence I'm gonna dial down the violence in these shows either across the board or you could have a specific version of that content but dial down violence for a specific cohort.
• 36:07
  SPEAKER_02: You can start doing some really cool things in terms of the level of personalization that you would be doing.
• 36:13
  SPEAKER_01: Hey I wanted to um for those of you for those on the podcast who may be jumping in a little bit late we're speaking with Adam Sutherland who's a leader in in AI and ML for media and content and everything you know great in the world that seems fun.
• 36:29
  SPEAKER_01: I wanted to also um you know like like like speech and rec and stuff is is speech and recognition is fun.
• 36:36
  SPEAKER_02: Natural language processing is fun but when you get into media and content it's super visual and it's it means a lot of things to a lot of people but I wanted to switch gears a bit and kind of get a take from you Adam on what emerging trends are you starting to see and what is you know what should folks maybe be keeping um a lookout from maybe from a consumer perspective and from a business perspective just in terms of content yeah so I'll tell you so I mean there's some stuff that's really interesting from and I don't want to call it blocking and tackling but it is to your point about the visual medium there are some techniques emerging and that you know with with ML what you'll find is it's not about whether or not there's a technique or an algorithm for it.
• 37:19
  SPEAKER_02: It's about whether or not you can start doing it affordably right um but there are uh some very interesting techniques coming up with uh hyper like with up resing and and you and the way it works is you'll take an old flick you know hopefully 32 frames per second or maybe even less and you'll the machine the algorithm will actually create seven or eight frames between each frame which are slight little derivations to get you from that one frame to this so now you've got an HD file right before you didn't um so up resing is very cool.
• 38:01
  SPEAKER_02: The the other is there's a there is a burgeoning um I god I don't remember the name of sunspot or something like that.
• 38:08
  SPEAKER_02: There's a burgeoning area of content creation just creation so there uh and I'll I'll find it and I'll send it to you guys afterwards uh to to include here but it is a short film that where the script was was completely AI generated.
• 38:28
  SPEAKER_02: So they took a bunch of like sci-fi scripts and they fed it into a model and then they had this thing and so and then they and they hired actors and it you know you'll recognize the actors when you when you see it like they hired actors and they did like a five minute movie where the entire script was um that was generated.
• 38:46
  SPEAKER_02: It's not good.
• 38:47
  SPEAKER_03: I mean it's not great.
• 38:49
  SPEAKER_02: Well one day it's gonna be better right it's yeah and what about synthetic uh actors yeah yeah well so the you know we were talking about personalization i i've heard one of the use cases i've heard i haven't actually seen it but i've heard that some networks have are starting to build models where they will not necessarily feed the scripts or or not necessarily create scripts but they'll take the scripts or the the screenplays or whatever it is run it through and then they've got a model that says this is going to be a hit or it's not going to be a hit.
• 39:23
  SPEAKER_02: Obviously not that binary it's not like yes or no but they're starting to incorporate the just like they're starting to use AI in job search right with resume reviews and things like that there's also this this kind of emerging use case where media where media companies will use AI to let them know directionally is the script a hit or is it not?
• 39:41
  SPEAKER_02: Yeah.
• 39:42
  SPEAKER_01: Well see I'm wondering I'm wondering then if they're gonna use as the baseline either Seinfeld or arrested development right as those baseline models because those are the hits yeah yeah exactly I was going to say or raising Arizona right yeah let's go back to I mean I was just thinking about upraising like cool head Luke right and what that's gonna do.
• 40:03
  SPEAKER_02: Yeah exactly well um yeah go ahead the the last thing I'll say is the on the commercial side from a business perspective two things one uh the ability to use uh AI in some capacity think of a sporting event right and if you're in the the stadium you will see those those kind of uh interactive banners or those banners across the thing and they switch and they flip with you know from one brand to another and if you're in the stadium you're you have no idea what the frequency is what's driving that what's driving it is how off how how much screen time that brand has had on the telecast.
• 40:43
  SPEAKER_02: So you can automate that right now there's a human washing there with a stopwatch and we're like okay that's been 30 seconds okay switch it you can automate those kinds of things the ultimate expression of that is you can start to take brands you can track how long a brand has been in a video and all that other stuff and get that implied value for product placement you can also start to superimpose other brands.
• 41:05
  SPEAKER_02: So maybe you made a show and that show was you know the lead sponsor of that show was a making was a car company A.
• 41:14
  SPEAKER_02: And now five years later you're going back into your archive you're going to make something new out of that show and you you know if if the rights allow you can now take car company A and you can swap that out with car company B and you can use AI to superimpose over every place that that showed up in the initial one with the new one.
• 41:34
  SPEAKER_01: Very very cutting edge not happening at scale now but it's one of the things that is driving some well I mean to your point I remember you know seeing like hey I'm gonna look at um you know like a TV show like Top Gear and within that TV show of Top Gear I don't know where I'm gonna put product placement yet.
• 41:52
  SPEAKER_01: However I go through and I analyze everything and see where the keywords are actually spoken and then optimize you know whenever that is spoken to boom.
• 42:01
  SPEAKER_02: Oh he said Mercedes boom Mercedes goes in or he said this car company and then goes in you know those types of things so more ad hoc print process but amazing amazing so you know I don't know if there's um you know anything more that you can you know inlight this with I mean shoot just going through uh you know an hour is just the tip of the iceberg but anything you want to predict for median entertainment anything kind of blue sky or uh uh interesting predictions or things that we'll see down the road you mentioned AI um uh synthesized scripts how about synthesized actors and what else what else could we start seeing what else is possible yeah well I think what's gonna be really interesting is the uh the the rights around those derivatives so you yeah so if you're just making something new okay then it's work for hire for example you're making a synthetic person or you're doing a you know an animation um uh what's going to be interesting is the ability and we see this already is the ability to take for example an actor's voice train a model to sound like that actor so now you have you know voiceovers and and this is a really specific use case that's come up is I don't know if you've ever watched I'm sure you've watched foreign films that are dubbed and and when I lived in Japan there was a Schwarzenegger film that was dubbed and I was like that doesn't sound like Schwarzenegger at all right well now you can actually train a voice that can speak another language in that actor's voice.
• 43:35
  SPEAKER_02: And the question becomes what are the rights around that who gets paid is it yeah um so there's gonna be a lot of these kind of derivative things that are gonna raise a lot of questions.
• 43:50
  SPEAKER_03: And it and it still seems like such a free for all like my my smart TV is trying to take over the the the interface it's trying to be the set top box I have to keep telling it no.
• 44:01
  SPEAKER_03: I want to go to Apple TV.
• 44:03
  SPEAKER_03: I want to go to, you know, I want to control this, not you.
• 44:05
  SPEAKER_03: So it's very interesting because it forgets that I've already subscribed to other things, you know.
• 44:10
  SPEAKER_03: Yeah.
• 44:10
  SPEAKER_02: So well, there's definitely going to be you're seeing it already in terms of just actual MA, but the the pendulum swing from everybody.
• 44:20
  SPEAKER_02: So, you know, once again, I have a little bit of a bias here given that my job is data science and analytics.
• 44:26
  SPEAKER_02: The battle for data is is the new battleground.
• 44:30
  SPEAKER_02: And so, you know, not only kind of who owns that data, but who owns the in the interface with you that takes advantage of that data.
• 44:40
  SPEAKER_02: So, you know, I just recently I signed up for a service and over the I I cut the cord for about six months, and then I signed up for an over-the-top service, which essentially is another is like a cable bundle, but it's over the top.
• 44:55
  SPEAKER_02: But but the the main difference is it's got that personalization aspect to it.
• 45:00
  SPEAKER_02: Whereas across all the services, across all the services that are bundled underneath it.
• 45:05
  SPEAKER_02: So that's going to be the new kind of paradigm, which is not just what do I know about you, but but who's who's in the lead on that customer experience thing?
• 45:15
  SPEAKER_02: Yeah.
• 45:15
  SPEAKER_02: Yeah.
• 45:16
  SPEAKER_02: And this, I I neglected to say this before, but the all of these things, uh, I you know, I in the media space, and this isn't a prediction, this is just kind of maybe a a proclamation, if you will.
• 45:29
  SPEAKER_02: Lots of content and good content.
• 45:31
  SPEAKER_02: That's cave, that's table stakes.
• 45:33
  SPEAKER_02: Like everybody expects that if I open up X service or Y service, that I'm gonna have good content and I'm gonna have a lot of content.
• 45:39
  SPEAKER_02: Even if it's a niche service, I'm gonna have a lot of good content in that niche, right?
• 45:44
  SPEAKER_02: The where where everybody is competing now is on that customer experience.
• 45:49
  SPEAKER_02: And it's not just personalization, it's uptime, it's all the other things that go into it, but the real driver for a good customer experience is data.
• 46:00
  SPEAKER_02: And just get searched to work.
• 46:06
  SPEAKER_02: Exactly.
• 46:07
  SPEAKER_01: Listen, Adam, um, it's always a pleasure, man.
• 46:10
  SPEAKER_01: Love hearing, you know, your insights and anecdotes and you know where things are going.
• 46:15
  SPEAKER_01: I always learn stuff.
• 46:16
  SPEAKER_01: And so I want to appreciate, you know, just tell you we appreciate the time and thanks for.
• 46:23
  SPEAKER_03: Thank you so much.
• 46:24
  SPEAKER_03: And I also want to just before we close, thank our sponsors, which is the which are the Marketing AI Institute and CMS Wire, two wonderful resources for digging in deeply to many of these topics.
• 46:37
  SPEAKER_03: So thank you to those folks.
• 46:39
  SPEAKER_03: Chris, Sharon, uh, Sharon in the background, uh, our producer and Adam, thank you so much.
• 46:45
  SPEAKER_03: Really appreciate it.
• 46:45
  SPEAKER_03: This has been a lot of fun.
• 46:51
  SPEAKER_00: Thanks for listening to the Early AI podcast with your hosts Seth Early and Chris Featherstone.
• 46:57
  SPEAKER_00: We hope you enjoyed the program and took away some ideas that you can start to implement in your organization right now.
• 47:03
  SPEAKER_00: If you're stuck and need help, check out www.early.com for information about early information science consulting services.
• 47:12
  SPEAKER_00: And of course, check the show notes for links to any resources we mentioned today.
• 47:16
  SPEAKER_00: That's all for now.
• 47:17
  SPEAKER_00: See you next time.