Could you anticipate someone's musical preferences based on their taste in movies? Perhaps, but Qloo, an innovative predictive algorithm that maps consumer tastes, goes far beyond that. Their models can make relational taste predictions about your preferred snacks, favorite bars, book choices, dream travel destinations, and much more. In this episode of The Insights Factory, Ian sits down with Alex Elias, Founder and CEO of Qloo, to discuss their extensive consumer taste data set and their ability to map consumer preferences across categories.
Could you anticipate someone's musical preferences based on their taste in movies? Perhaps, but Qloo, an innovative predictive algorithm that maps consumer tastes, goes far beyond that. Their models can make relational taste predictions about your preferred snacks, favorite bars, book choices, dream travel destinations, and much more. In this episode of The Insights Factory, Ian sits down with Alex Elias, Founder and CEO of Qloo, to discuss their extensive consumer taste data set and their ability to map consumer preferences across categories. They also delve into Qloo’s geospatial precision and the anonymity embedded in their AI and ML models.
---------
Key Quotes:
“With the geospatial model, we like to do these ridiculous data polls where it's like, 57th Street and 11th Avenue, plus Tolstoy, plus, Jill Sander, the fashion brand, and have it extrapolate where that person would want to grab a bagel or something like that.”
“One example recently was with another brand of product, where the cross category correlates to film preferences and TV preferences. It showed a penchant for horror and fear, and campy horror in particular. And that ultimately informed some sort of creative that was about the fear of running out of a product.”
“I would encourage people to approach this whole world of AI and data with a degree of optimism and hope and just excitement, because I think there's so much about it that could serve us. And a lot of the kind of doom and gloom narratives around it are likely leading to a degree of trepidation and sort of antagonism that ultimately is counterproductive, I think.”
---------
Show Timestamps:
(01:06) Why Alex started Qloo
(03:29) Building a symmetrical panel
(08:05) Impacting the bottom line through consumer taste data
(10:45) Interacting with Qloo
(15:51) Anonymity and privacy
(18:08) Entity instead of identity
(19:57) Thoughts on AI
(25:39) What do customers use Qloo data for?
(32:21) Navigating the gap between the business problem and the data owner
(37:16) What technology excites Alex moving forward?
(42:49) Approaching AI with optimism
--------
Sponsor:
This podcast is proudly sponsored by Seek, the leader in cloud-based creation and delivery of industry-focused insights.
--------
Links:
Connect with Alex Elias on LinkedIn
Connect with Host, Ian Cook, on LinkedIn
Follow Seek on LinkedIn
Ian Cook: [00:00:00] Some days it feels like we're drowning in data, but data isn't knowledge. Data without context or the expertise to understand it is just something taking up space in your warehouse. That's where being insights driven comes in. Insights driven organizations focus on generating actionable insights rather than just collecting and munging data.
Insights drive better decisions. Welcome to the Insights Factor.
Thank you for listening. I'm your host Ian Cook, the CTO here at Seek. Joining me today is Alex Elias, the founder and CEO of Clue, a fascinating AI driven company that I have the privilege to interact with fairly regularly, and we are thrilled to have him on the podcast with us today. Thanks for joining us, Alex.
Alex Elias: Thank you, Ian. Uh, appreciate it. It's great to be here with you. Absolutely.
Ian Cook: So let's start at a very, very basic [00:01:00] question. Why did you have to start Clue?
Alex Elias: That is a great question. Uh, maybe not so basic, but I'll try and, I'll try and keep it basic. So the journey, we started this company over a decade ago, which is quite atypical in the tech world, uh, certainly in the AI world.
Uh, what we saw was a world where a lot of the type of data that was predictive of consumer taste was highly siloed. So you had, at the time, it was Pandora, you know, covering music. Uh, an even earlier predecessor that many people don't know about is Firefly, which was started in the 90s, acquired by Microsoft, was one of really the first collaborative filtering algorithms, which is essentially a fancy word for you may also like.
They specifically focused on music. Music has always been kind of the pioneering.
Ian Cook: Oh, I grew up with the Napster and the LimeWire, Firefly, I remember terribly well, oh yeah.
Alex Elias: Totally, totally. So, so basically [00:02:00] what we saw was there was a massive opportunity to build this kind of omnibus overarching. Uh, data set that could actually predict within and across some of the most important areas of consumer taste.
Um, so you, you think about music, the media nexus, uh, movies, TV, but then also the kind of geolocational, transactional nexus, so things like restaurants and, and hotels. Uh, we felt that this was most interesting as a consumer premise, so we wanted to kind of... disentangle the recommendation science from these large kind of B2C silos.
So, you know, most people were just getting recommendations as part and parcel to using Expedia for where to travel or, you know, Pandora for music. But we felt like it would be interesting to build this holistically and in a way that Served consumers. Uh, and we soon found that was a challenging business, a challenging concept.
Uh, so the going direct to the consumer part, [00:03:00] direct to the consumer, meaning me as
Ian Cook: somebody could sign up as back, back in the day, there was a point I could sign up for a clue myself. Absolutely. Absolutely.
Alex Elias: Through, okay. We had an io. So we were kind of right at the beginning of the mobile wave. So we had an iOS app that scaled to north of a million registrants.
Uh, it had a very laborious onboarding. It required that everyone described their tastes across, you know, eight different categories, because we were very interested in that symmetry, building a symmetrical panel. We were sort of obsessive over it. Could I
Ian Cook: dig in a little bit there? A symmetrical panel?
What would you, what, what, for the audience that may not be aware of this, what is that?
Alex Elias: Yeah, so basically, uh, the challenge, the challenging aspect of what we hoped to do is to be able to link these different worlds. So link the world of music to film, to literature, even to all the way to restaurants and fashion.
Uh, and, you know, it turns out generating, uh, kind of [00:04:00] recommendations within any of those categories is a lot, is a lot easier because there's kind of a content base. There's objective. characteristics that link things. So, a Wes Anderson movie, uh, is objectively linked in some way to another Wes Anderson movie.
Right. As a horror movie is to another horror movie, but jumping across categories requires kind of a nuanced, uh, data set, ultimately, of, of people's taste across, across categories. And we've, you know, we, we soon discovered there is a predictive nexus across these different domains. And in fact, one of the.
You know, and this is now getting very attenuated, but I think it's interesting. Uh, one of the, uh, one of the founding sort of impetus for us was, was the Netflix prize competition. So they had a collaborative filtering competition for who could improve the efficacy of their algorithm. At the time, it was DVD rating data.
Right. Uh, and what confounded, uh, in the end, uh, it was kind of a... failed [00:05:00] experiments. It was a million dollar prize. But, um, the, the, uh, the film Napoleon Dynamite was so enigmatic and weird that it accounted for this outsize, uh, sort of error rate for folks. And we actually revisited that once we had enough symmetry, enough of a, of a panel and found interesting, reliable pathways that were predictive through unexpected areas like fashion and music.
to, to this film recommendation. So that was, you know, we, we saw even as that consumer premise, you know, it was, uh, the business model was predicated on affiliate revenue. We had affiliate relationships with iTunes, with Amazon. Um, but we found over time that it's not a, it's not a great business model. It was a challenging consumer premise to scale.
Um, but we stuck with it because we always felt that the data asset itself needed to happen. Like someone needed to. Take great care to kind of build this, this, uh, taste data [00:06:00] set that's spanned across all these different categories and served people ultimately. Um, what we soon found is that developers were the more interesting audience for the tech we built.
Uh, and it was actually a very senior product manager at Twitter in 2015 that reached out. and asked if we could, she was a fan of, of Clue as a consumer, uh, and it got pretty sophisticated by that point. We had a high level of scale of data, uh, and she reached out and asked if we could externalize through APIs, uh, all that intelligence for kind of a forthcoming commerce product they were rolling out.
Uh, so that was our first, uh, and that's a whole nother hour long story, uh, because I'm sure anything working that all exploded. Real, it gets real
Ian Cook: interesting real fast. Indeed. I'm, I'm interested, you said you revisited the Netflix idea. So my recollection of the Netflix process, you know, being somebody who sort of came up through the AI, [00:07:00] ML world, was that it was very interesting.
The thing that actually won, the technically won, the The competition was this aggregated, um, model of models. So they were essentially bagging and boosting to try to get a better result. But the only point of doing it was to get a higher score in some objective measure. And what was interesting, and this is something that we hit on a lot in this podcast, is that's fine, but objectively it's not that useful to Netflix because there was just this tiny little increase in like, well, I can predict how you're going to rate it just a tiny bit more, but it's not.
An insight that's useful. Like there's nothing I've learned about people that helps me say like, all right, terrific. Now I know their ratings. I mean, ultimately, as we see now, when you log into Netflix, there's. For a while there was, I like it, no I don't, or I like it a lot, now I think it's just thumbs up, thumbs down, but did you run across that same kind of thing as like, you have a ton of this information and there's a, there's a, almost a hurdle to get to to say like, great, [00:08:00] but what am I learning or what is it that I'm showing you about it?
Your taste in information. Absolutely,
Alex Elias: yeah. I think, uh, there's definitely, and I think to your point with that experiment, it was a highly incremental game that, you know, didn't have bottom line business impact. Um, and candidly, you know, we work with Netflix now, so coming full circle. Oh, well, okay. Uh, which is an exciting, that's something we could speak to some extent, uh, a little more about.
But it's, uh, you know, I, I think what we found, Uh, and, and I think what you might be hinting at is there's also a lot of demand now for kind of, uh, baking in some explainability, um, and, and sort of, uh, trying to articulate some of the causal mechanisms by which these recommendations come about, particularly in the enterprise use cases.
So I think for, uh. For Netflix users, you pointed out, they've kind of deprecated a lot of the explicit rating mechanisms for many reasons. I think they found that, you know, and A, that [00:09:00] began in the DVD era, you know, now with streaming, all these OTTs have an incredible amount of implicit feedback. If I could just ask
Ian Cook: you to, when I hit an acronym, I like to make sure people hear it, OTT.
Alex Elias: Uh, over the top, uh, it's, it's a media term just for a streamer, essentially, it's kind of a fancy, unnecessary, uh, acronym. Um, but the, but yeah, the basic, you know, the implicit feedback mechanisms they now have of, you know, when someone's started a show, stopped, all the events that are tracked, it turns out are, A, far more, seamless from a solicitation standpoint.
Um, and then the explicit rating mechanisms, which again, were a legacy of the kind of, you know, evolution from Blockbuster into, into DVD mailers and so on. But, um, but you know, the sad thing is that something's actually lost in that. And I think that's something we've, You know, we felt, and this comes full circle to a company we acquired called TasteDive, uh, which puts [00:10:00] us firmly back at our roots.
Uh, TasteDive, many people know it as TasteKid. It was, it's kind of a cult recommendation platform. Maybe one of the only ones now that still exists from the kind of 2000, the late aughts. Uh, started by some former IBM folks, and it covers every category that Nearly every category that Clue covers, uh, and it is entirely explicit sentiment.
So it's, you know, people coming in and saying, I love this, I dislike this, and getting recommendations accordingly. Uh, it's, yeah, it's an interesting platform, totally ad free, and it kind of gives us a lot of those first party, you know, joins that we always, We always wanted, but, but yeah, to your point, um, I think there, yeah, there's a lot of, a lot of interesting evolutions in, in that recommendation science space.
Ian Cook: So just to be very specific about what it is, if I were to interact with Clue by API or anything else, it really is, and I think people will want to hear and know is you mentioned like Wes Anderson movies, [00:11:00] horror movies, you mentioned music. I think it works a bit like, if you like this, then you'll like this other thing, but at such a deeper level and across, what I find is fascinating, so many entities.
Like, all right, if I like a Wes Anderson movie, maybe I'm gonna like these types of movies. I happen to be a huge horror movie fan, so, uh, you know, you like Ty West, you're probably gonna like A24 movies. If you like Wes Anderson, maybe you'll like some, uh, Sofia Coppola movies, but you're going from, if you like this movie, Not only the types of movies, but now into types of music, then into types of brands of food.
And this is something that we at Seeker are working on a whole lot, is to be able to like, think through, you liked Flamin Hot Cheetos, do you like, uh, you know, Takis, the blue flame Takis, or whatever they are. And, you know, we think through that process, but this is what you actually get out of the data you have.
Alex Elias: Yeah, yeah, absolutely, yeah. So we've, we've kind of put the whole thing on steroids, kind of beyond what we even initially envisioned, and it sort of came at [00:12:00] the intersection of being able to acquire more and more data sources with a commitment to anonymity, which I think has, has helped us tremendously along the way.
Uh, and then you combine that with the sort of all the accelerants in the world of compute and AI generally, you know, And there's been these kind of leap, you know, these step functions that, you know, we, we hadn't really expected in terms of efficacy. You know, one of the most notable is really our geospatial model where, you know, we've gotten to a point in 2019, we began tinkering and we had a very big client, the largest billboard company in the world, JCDecaux, who sort of commissioned the model essentially.
And so we began by sort of chucking a tremendous amount of compute. And essentially geo hash the whole globe down to precision seven, which is roughly 150 meter, you know, grid globally. So when you say precision
Ian Cook: seven, for people who have not worked with geo data, that's if you get lat long, you have a decimal point and then you have a [00:13:00] certain number of numbers, you have a.
A number of significant digits after that decimal point. That's as, yeah, it should be clear, but the more numbers that happen after that, the closer you're getting to one single spot on the globe and at the level of seven, seven digits after the period you're talking 150 effectively square meters.
Alex Elias: Yeah, something around there.
I want to, I want to confirm the exact number of meters in that order of magnitude. And that's a beautiful description of, of that. Uh, and, and yeah, so it, it, it was one of those aha moments. When we actually, you know, we obviously spent a lot on the compute and, and the exploration of it and the end result was so phenomenally impressive, even with this kind of anonymized sort of data corpus that, you know, one of our, one of our biggest shareholders, I won't mention his name, but, um, we, we actually, in one of the initial meetings, he was expressing a lot of skepticism And we, we derived the latitude and longitude for his home [00:14:00] address in the suburbs, uh, outside of New York.
And, uh, and, and we, we put that in and it actually predicted the brand of pants he was wearing as a sort of, and that was this, this aha moment, but it's, it's surprised us.
Ian Cook: Did J. C. Decoe have the same kind of, did you, like when you presented it to them, was there a moment they just kind of lit up and were like, I, you know, you kind of had this.
We weren't entirely sure you could do it, and then you see,
Alex Elias: like, wow. Yeah, totally. There was, there was many, many moments along the way. Uh, I think fashion was one of the most efficacious initially. Um, but then starting to get into media, and we actually got into a point where there's these kind of, it's almost like a pro forma prediction of, you know, when, because we, we began indexing, uh, media entities even in a pre production state.
Um, so, you know, once we know that a film has certain attributes, is optioned based on a certain book, has certain director, actors. We can actually tie it into the, that [00:15:00] nexus. And so that, that's where things started getting really impressive was, you know, even being able to generate predictions for things that hadn't yet accumulated, um, taste data, which is, you know, firmly now we're in the world of this kind of the, the, the excitement around AI.
But, but yeah, there, there's been a lot of, a lot of those. You know, they've used it for very interesting things to try and convince advertisers to look at novel areas of cities that they wouldn't have otherwise considered, you know, particularly the luxury brands. Uh, there's been a lot of, a lot of interest there.
So don't
Ian Cook: just buy the huge billboard off of the big highway because that's the most people driving by you're trying to target, get much better. Like, these are people with a much higher likelihood to purchase. Totally. Show them that it
Alex Elias: exists. Yeah, exactly. Inventory that might not be considered in kind of tough parts of Belgium for a luxury automaker, for instance.
So you've
Ian Cook: mentioned AI. We'll get there in a minute. I want to get back to something you've said a couple of times that I find really interesting. You've mentioned anonymity [00:16:00] and privacy, and you also use the word entity quite a bit, as opposed to say individual. Is that, that feels purposeful. Can I ask about that?
Alex Elias: Yeah, yeah, absolutely. And very, very well, uh, well identified. Uh, so entity, uh, denotes, uh, uh, uh, basically an atomic unit of taste. And it, and it's confusing because it could exist at multiple layers of the hierarchy. So an entity could be, you know, a musician. It could be a song, an album. So there's, you know, there's this hierarchical.
So entities are also interrelated. And in various complex ways, um, but the reason I, uh, we emphasize entity is that. In, in, in our world, in, in the world of AI that we partake in it, it's sort of a, uh, alternative to identity. Um, and particularly as you start looking at multiple entities, um, being combined and passed through as a kind of proxy of, of taste or as a proxy of, uh, the [00:17:00] individual.
Um, it's, it's a, it's a powerful, it's a, it's a powerful lens because it allow, A, there's an elasticity of how you define an individual. So identity is kind of a fixed deterministic thing, uh, and, you know, it's also slightly nefarious. It's something that Uh, ultimately, you know, there, there's very big questions about consent and what sort of, and, and I think, uh, obviously the Wild West era is hopefully starting to, you know, to, to come to, to a close, but there was an era where You know, a lot of the financial services firms that we work with, uh, in 2016 17, you know, might have been less interested in, in a probability based approach because you could get so much identity based join.
So you could, if you know that, uh, person X has applied for an auto loan at a certain time, uh, that, that is a real, that, that's not inferred, that's like a hard piece of, that's a hard join. Right. And at that
Ian Cook: point, you're just disambiguating. I've got [00:18:00] nine. And. Ian Cooks, and now you subscribe to this fishing magazine, and I know you bought pants at this, like,
Alex Elias: we know it's you.
Exactly. And multi billion dollar businesses in that era were built on exactly, as you said, just identity disambiguating on some level or another. And Entity, on the other hand, is a proxy For taste, it's elastic. You can pass. So, the way our service and the way that this kind of generally would work is you can supply a wide range of tastes, of context, of factors.
So, it could be location plus an array of, you know, merchant IDs and restaurants and songs and albums, some of which may never actually exist in the corpus of an individual's, you know, own life experience. Like, we like to With the geospatial model, we like to do these ridiculous sort of, uh, you know, uh, [00:19:00] data polls where it's like, you know, 57th Street and 11th Avenue, plus Tolstoy, plus, you know, Jill Sander, the fashion brand, and have it.
And have it extrapolate, you know, what the, where that person would, uh, wanna grab a bagel or something like that. And there's obviously that, you know, that bit of data, that individual may never exist or never come to exist and that bit of data may never, but it's, but, but that's the, the beautiful thing about this kind of contextual approach.
Ian Cook: You know, I get to be an Ian Cook that lives here, but I also get to be a Run the Jewels fan in Pittsburgh at one point, but also a, uh, Sturgill Simpson fan that's traveled up to Northern Vermont. And, you know, both of these things are interesting to brands at different times, whereas, you know, if you link it to me too much, you start saying like, well, but I know exactly who you are and I'm just going to send you this because I think that's what people in Pittsburgh want.
Right. All this makes me want to do is, uh, you know, the Spotify end of the year. [00:20:00] Right. Wrap up. One of the things I wanted to do the sort of sort of instant I learned more about this data is like, I just want to start like looking over time. Like this past year, Pittsburgh's top snack was this, and they also listened to this much and they like, I just think it would be great to see sort of that kind of happen throughout the year.
We should
Alex Elias: collaborate on that because I think it is, it is some of the most interesting sort of output and you know, um, I think seek is kind of the, you know, the, the. Brilliant, the brilliant minds in that, in that application of data.
Ian Cook: It would be fun. We could put that kind of insight together and people can sign in and see the dashboard and just sort of, sort of get a sense of that.
Cause I just, it, what I find interesting is, you know, you found it to be a tough consumer model and, you know, people don't want to pay for anything. We've discovered that people love everything to be as free as possible, but we're also so interested in our own tastes. Spotify thing comes out at the end of the year and everybody's like, who did you listen to the most?
And totally, totally. So it'd
Alex Elias: be fascinating. I think that's kind of the highest. and best use of, you know, if we can reveal something about oneself to people, you know, in a [00:21:00] consent framework and in a, in a fun way, that's, that's beautiful. So your
Ian Cook: company has been using AI for a while. Obviously, you know, I have lived in a world where I started out doing statistics, eventually I was doing AI and I was kind of always just doing the same thing.
I got older and the name kind of got cooler, but you guys are really embedded in this world. So what is your view on the way things have moved recently? Obviously, we have this explosion with large GPT. But you've been working in AI well before that became a big deal. Had you been in the sort of natural language processing space?
Are you moving more towards it now that there's a, you know, a desire for this?
Alex Elias: Yeah, I mean, it's, it's something that's always been a useful tool in the toolkit. Uh, it's, it's something we've worked with at various layers of the stack. Um, so, you know, the language models originally from a entity recognition standpoint and being able to extract metadata and decompose meta, you know, [00:22:00] metadata from a rich set of, set of data sets.
Uh, so that's where a lot of the language stuff initially emerged. There's a lot of interesting recommendation based applicability of these, these LLMs even, uh, and, and otherwise where, you know, the, the, the word order, I mean, ultimately. And LLM is one of the great summarizers in the world, but there's unexpected kind of utility in the in the recommendation space.
So we have been looking at it for a long time. I do think we're in this new era of kind of commercial and consumer surface area and ergonomics. There's a whole new world on how people are expecting to Uh, to, to be able to interact with these AIs and that's exciting for us because I think we found that there's, uh, there's a lot of applicability for kind of being an oracle or source of truth of some sort into these LLMs that then, you know, create this kind of another world of of insight and I'll give one example, uh from Q4 of last [00:23:00] year, there was a A hackathon we did with a, with a, with a large media client and, you know, they were looking at various, at the time it was GPT 3's API combined with Clue's API.
And one of the, one of the winning concepts was this ability to input one taste, just essentially one, it's a one word prompt, uh, and it passes through Clue's. AI, and will generate, so if I put in a music artist I love, or the, you know, the, the, the dinner spot I love, or anything else, it'll recognize that, it'll then generalize it, and create a sort of rich tapestry of, of taste across all the different joins and categories that we cover, and then it passed that into GPT 3 at the time, and generated this, uh, essentially long form essay that insulted your taste.
And it was, uh. It was an extraordinary result. It was kind of the, you know, the leading concept and it just, it was, it was mind blowing to kind of see that compounding [00:24:00] factor of taking, you know, the great summarizer, uh, uh, and combining it with this kind of taste oracle and then seeing what that, what that compounds into.
And so the magic of just putting in one little kernel and having it insult you in such a detailed way was extraordinary. And it was, and it was sort of a lot of things I actually. Uh, expected to, you know, an insult comic to, to sort of, you know, heave on me. That is, that
Ian Cook: is genius. I had not heard that. So like, I put it in like one restaurant, like Alinea and I can, you know, and somebody else, like, I'll get back an essay about my, like, super over fancy tastes and I must think I'm elite.
And just cause I put in like a high end restaurant and then everything, that is, that's, that's really interesting. Uh, I was listening or Pardon me, I think I was reading a story. Somebody brought up that there is a bot right now on, I guess we're calling it X, not Twitter. Right. Um, that is meant to mimic Grimes.
[00:25:00] So, a former, a former relationship for Elon Musk. But the person's so convinced that It must have human intervention because it is just so good at being grimes that I thought that's a fantastic way. So if you can, like, what I find interesting is like the interesting uses of these things, right? So what's interesting about that is somebody didn't just go, all right, you know, I want to know, I like.
some song, what other song do I need? But like going from all this information to such a very specific like, all right, here's what's funny about your particular taste. So yeah, yeah,
Alex Elias: absolutely. And, and perhaps the, uh, problematic Turing test should be replaced with a Grimes test. There you go.
Ian Cook: I love it. But that does get to the question of like what people are using it for.
So you've got a bunch of customers. What do you find them using the outputs to do? I'm don't mean to dig into anything specific, but the kinds of insights that you are helping them surface, where does that lead for a company to be [00:26:00] like, Oh, I just never thought this about my customers, or maybe it's something different.
Alex Elias: Totally, yeah. So the, one of the, one of the main themes we've seen in kind of the most high volume utilization of, of Clue is, uh, is basically to enrich, uh, you know, the understanding of particular elements of their business or customers on a one to one level, but in a privacy centric way. So being able to, just to put it, contextualize it in terms of a few examples, financial services has been a rich space.
Principally because they have so much data to begin with. There's a lot of transaction arrays. So you imagine if. Uh, but at the same time, there's big gaps in their understanding. So they might, you know, uh, a particular credit card company, for instance, would understand, uh, the, the, the merchant IDs that someone transacted at.
Um, they would know basket sizes. They would know that you, uh, have a Spotify, uh, you, you know, so then you start getting into, they know you [00:27:00] have a Spotify subscription, but they have no idea. or bought a StubHub ticket. They have no idea what the SKU was, what tickets you bought, what music you listened to, what TV shows you streamed on your Netflix subscription.
They do see the charge. Um, so that's where Clue, in those contexts, we can take in totally anonymized fragments of data, uh, pertaining to areas they know a lot about, and then we could translate that into actionable tastes. in any other category. This is done within milliseconds, totally anonymously. Uh, and then they have the full rights to run with that output inference.
So, so
Ian Cook: they know you bought a StubHub ticket and what you're kind of bringing to this is like, well, combined with some other things. There's a good chance that that ticket was to the Ares tour for Taylor Swift, and this is a person who's also going to shop at X store and likes these types of foods.
Alex Elias: And might travel to this location and so on, and so if they're now deciding from, you know, [00:28:00] who to send a concert ticket pre sale offer to.
across a CRM with, you know, 45 million people. Uh, and they could now essentially prioritize and rank that dynamically using clue, clues output. Uh, and that applies across the board. Like we work with a smart TV, uh, universal guide situation where they have a gaming hub and they want to recommend games. Uh, but all they have access to is linear TV viewing data.
So that's Going across categories, being able to now prioritize, and so in a lot of those contexts. The measure of success is very much a practical one of driving lift, you know, were it for were it not for this extra lens or this extra enrichment, there would not be a 273 percent increase in, you know, in the click through and redemption.
So those are those are the areas that are. That are, you know, some of the most concrete, [00:29:00] tangible integrations. Um, there's a whole host of other situations that are more about understanding the consumer. Um, and, and I think that's where our worlds intersect to an extent where there's a lot of. A lot of brands out there who want to use data to better understand and, and resonate with and distribute to and so on all these decisions.
That's one of the things
Ian Cook: that intrigues us so much about Clue is we focus a lot on stores and what's being sold at a store, volume in and out, out of stocks, things like that. But when we also say this is selling well, like you see a product that's selling well, and you're just going off what we would consider velocity, it's sold this much over this many days at this many, this many dollars per unit.
But if we want to start understanding what are other things that would go well at that store, we need to understand, well, what are the affinities for that area? What are the types of things that people who also buy this also buy other items? So that we can start providing a much more holistic picture [00:30:00] again, without having to put, you know, Ian Cook's face up and go, here's what he bought today.
And, you know, he seems to have a real Snickers addiction, might need to worry about that with a doctor, but let's also get him some Pepsi or something like that. But like we, we are, and that's one of the things we're doing with our insights club, but it's, it's. So interesting to me that to be able to do that and to think about that as, as a almost embedded sense of like, now I have this store that goes from, well, here's a box with an input and an output to here's a whole representation of a population that is really fun to
Alex Elias: watch.
Totally, totally. And if I could return the compliment, I think what intrigues us about Seek is. You know, we've built this very sort of focused taste engine, uh, that has the potential to do many things and we're not always in a position to ascertain that and I think Seek has been incredible with this kind of last mile of taking a kernel or a category of insight and supplying it in conjunction with [00:31:00] something else that all of a sudden creates this.
Enormous accretive sort of situation and we've seen it, you know, again and again with, with, with that seek kind of that, that seek touch, I'll call it, but being able to take, you know, and there's been a few examples, I'll sort of mention them and, uh, in an anonymous fashion, but being able to take a CPG brand and prioritize by taste, which Walmarts, You know, nationwide would be most relevant and help drive unexpected distribution decisions or all the way to one example recently with another brand of product, uh, where the cross category correlates to film preferences and TV preferences.
You know, I, I love this example, showed kind of a, a, a penchant for horror and fear and campy horror in particular, and that ultimately informed some sort of creative. That was about the fear of running out of a product, you know, and uh, and that was something that, you [00:32:00] know, we, we wouldn't, uh, we've, when you just focus on correlating things, you don't always see how it can apply and unlock value.
And I think, uh, and there's sometimes these adjacent data sets and these adjacent capabilities that You know, that, that really unlock it. Um, so it's, yeah, it's an exciting intersection. Well,
Ian Cook: appreciate the kind words. One of the things that underpins what we're doing and what I'm going to ask you about next is there's often this gap between, say, the business problem and the, the data owner.
So the business people may not be as trained in the data and to understand it and use it. The people who were trained in the data may not have the same sense of the business problem, or at least the nuance. We have some great people on staff who help. Awesome. Bridge the gap, but especially with something fairly sophisticated, but really elegantly delivered tool like Clue may really appeal to some technical folks.
Also the affinity may really appeal to some business folks. How often do you have to sit in the middle and sort of get them to both understand what's happening here?
Alex Elias: Very often. And I'd say we're [00:33:00] not, uh, we haven't yet mastered that, that art and we've tried to, and it's a, it's a big human resourcing challenge and technical cause.
You know, we don't always want to pull our most adept engineering resources into. You know, in the client meetings, but at the same time, you know, so, so there is, there is a gap, uh, there, uh, I think that gap is rapidly being eradicated, so not only, uh, you know, clients are becoming more technically adept, you have CPG companies that have incredible data scientists that that headcount might have not even existed, you know, 18 months ago, um, I think there's more, uh, uh, Uh, more and more deafness at kind of consuming developer centric software, uh, data, applying it.
So, so I think we're in this kind of new era of, you know, taking technology, uh, and applying it throughout functions, including the CMO suite and others that, that is exciting, I think, for, for people to do what, what we do ultimately. But there, yeah, there definitely [00:34:00] is, uh, there's a huge, uh, bridge that, that we need to cross.
kind of, uh, to cross there. And, uh, especially in terms of unlocking value for the end customer. So how there's oftentimes business folks within an org that are, that are interested in the, in, in the AI, they might be curious about it. They might have some sense of some value proposition. Um, but then there is, you know, that last mile of actually applying it to where.
Uh, yeah. Contact me. Maybe the most onerous part of kind of ultimately, you know, um, being able to, to. Uh, drive customer success in our, in our world, at least. Um, and I'm sure you might see that as well, although you, you all are the experts at bridging that divide. So it might. Well, it's one of the
Ian Cook: things we, uh, you know, one of our motivating kind of beliefs here is that the, the [00:35:00] technology should not get in the way of the insights.
And if you just throw data at people, they're going to be drowning in it. They're going to look around and say like, well, this is. Great. And it's currently happened to a lot of companies. It's, I have this massive warehouse and then somebody told me I needed a lake house and now I have to have something totally different.
Is that the same thing I had? Is it different? What do I do with all these items? So we're hoping to get past that. And I think there's some, um, Similarity there and what Clue is trying to do is you have a junction of an amazing set of all of this data and you're providing the mapping to go with it. And there's, of course, all of the technical underpinnings of having built a knowledge graph.
And I imagine there's a lot of semantic work that goes into those things, but that's such a different customer that you go in and say, like, the reason for Clue is blah, as opposed to the business customer and being like, I can immediately go from, like you said, like. Being out of stock of, uh, toilet paper to, you know, campy horror movies, Friday the 13th, part 74.
These things mean something together and you should pay attention to that. [00:36:00] So I think it's a really interesting kind of place to be. You can serve both and do, do these things really well. Because these companies aren't going to want to spend all the money to own all of the data that you already have combined together.
Alex Elias: Totally. Yep. Absolutely. I think that's super well said. So as you look
Ian Cook: out, you're a technologist, you have been now for Clue for 10 years, what kinds of things are exciting you that you want, um, people to, to be aware of and that you're going to be trying to work more on with Clue and thinking about for the future?
Alex Elias: Yeah, absolutely. I mean, there's so much we're excited about. Uh, you know, one that's probably less interesting to people, but I'm personally really excited about is the ergonomics of everything. We've been making significant investments in the kind of ergonomics of interacting with, uh, With our data, our APIs, our AI generally, we want to get to a world where there's just little to no user error, you could throw any bit of context and clue, we'll have something interesting to say.
Um, two other things that we're really excited about. [00:37:00] One is, uh, creating, and this is part of our, the intersection of our worlds, is creating more of a non technical, uh, interface and way to sort of interact with it in novel ways. And also extract. A lot of clue heretofore has been very, very much, um, delineated and prompted, so you have to kind of specify the output, you have to specify the category, the filters, the parameters.
Um, we want to get to a world where there's more of a fundamental discovery element, even in the choice of domains, you know, so. Uh, to a point where you just come and say, you know, this is, uh, here, here's a brand, uh, here's Cartier, like, what is, what is interesting about the taste? What are, what are the areas even, you know, not specifically tell me the podcast that might be most correlated or the locations or anything else, but just like, give me, give me a proactive, And that's a very challenging problem, but I think we're making huge strides on [00:38:00] it.
The last thing I'm super excited about is explainability. We're investing a lot in causal kind of explainability, which is a complex thing as the models get more and more complex. And so, um, it's an, it's an interesting sort of tension between pursuing Efficacy, black box, sheer compute with, you know, explainability.
And one of the things that we've lost over time, so the, you know, when you look back at the collaborative filtering days, matrix factorization, there's a, there's an underlying degree of causal, you know, there's a causal nexus there that's explainable. Um, you, we're recommending this because you liked X, Y, and Z, or because people like, you know, and that, That has gotten somewhat lost as you get more into the realm of neural networks and, you know, these sorts of black box, you know, transformer models even.
But, you know, now we're sort of getting over the hill where, you know, some of these technologies actually are now More capable of sort of [00:39:00] reintroducing and assessing the weightings of various inputs and dynamically. So one of the big things we're working on and I think we're getting close in various areas is novel explainability paradigms where Without sacrificing performance, you know, that round trip of like, give me this result in 10 milliseconds.
Um, which is one of the things the language models have a lot of trouble with, by the way. Because the world we're in, we're a lot of integrations. We have to be able to report back within, especially in a presentational purpose. Exactly, yeah. So, can we achieve explainability? Because clients are more...
They're, they're at, you know, we've got no, it's so, it's sort of like the post modern relationship with AI, where it's like, alright, we're now super impressed with this, we now believe it's, you know, it's efficacious, we want to now know why, you know, and, and sort of hold it accountable in a way, and so, we want to get ahead of that, there's some regulatory Anticip anticipatory regulatory concern around that, I'd say as well.
I mean there's, [00:40:00] you know, the EU draft rags of, you know, the AI regulations and, uh, my background is . I did my jd, you know, at N Y U and focused a lot on internet privacy. So I was very intimately close to the G D P R, you know, the, it was at the time, you know, This is some research on the Consumer Privacy Bill of Rights that eventually evolved into, you know, some of the domestic paradigms like CCPA, but I think with AI, you're going to start to see more and more.
The draft EU regulations were surprisingly lenient in some ways. So they were focused on, you know, key areas. They didn't have the sweeping breadth that I think people feared. Um, so it was more about, you know, certain areas of concern, like consumer lending, hiring. Um, but I do think we're going to get to a point where consumers, clients, customers, and regulators all explainability.
And we want to get ahead of that. That's fantastic.
Ian Cook: You've said a couple of things that resonate really heavily with me. Well, everything you've said has resonated, but [00:41:00] I came from, obviously, like I said, statistics, causal inferences, one of the things I focused on most and understanding and running experiments and being able to say like, this is what drove this result.
And I, you know, sort of got bypassed a little bit by some of this neural network stuff as I was. Growing and building. And I, you know, got very much into, I can put this into production, but like you said, what's it really doing underneath? And now we're at the point where there's models with some nice explainability tools out there, but at a certain point, there's a research frontier that even the researchers are not entirely sure why it does what it does.
So it's, uh, it, it has a chance to get ahead of us pretty darn fast. The other thing I am absolutely going to steal the term ergonomics. So I think talk to any product person and the idea of any part of friction is a chance to lose people. Like you just like, you're using it, you just like hit a bump, you move people.
So the fact that you have to focus on that, like you said, while people might not find it interesting, but this is one of the things that you have to do, building software, building a product and think about this should be [00:42:00] totally transparent, but the effort to get into the point where it is transparent to the individual being like, Oh, I just used it as opposed to like, yeah, but I took away.
Everything that would make you think, uh, now I have to like, okay, like, like you said, with the early one, it's like, now I've got to fill out. All right, here are the top 20 movies I've seen in the past couple of years. Here are my favorite albums. Here's the.
Alex Elias: Here's the fashion designer, Alex Harris, right now.
And I have to, by the way, I should give credit to, uh, so please do run with that term. It's fantastic. Uh, I should give credit to our brilliant head of product, Kobe Santos, who's repeatedly used that term, and I sort of, I, I, I stole and or borrowed it from, from him. It's, it describes
Ian Cook: it really well. Right.
Terrific. Well, Alex, any last patterning thoughts for our
Alex Elias: You know, very excited about the world ahead. I would encourage people to, you know, approach, uh, approach this whole world of AI and data, uh, with a degree of optimism and sort of just excitement. Because I think there's so [00:43:00] much about it that could serve us and a lot of the kind of...
You know, the, the narratives, the sort of, you know, uh, the, the, the kind of, uh, doom and gloom narratives around it are likely leading to a degree of trepidation and sort of antagonism that ultimately is, is counterproductive, I think. And, you know, I think there's just, it's an exciting frontier, uh, and, you know, just, we feel lucky to be, be a part of it and excited about exploring the intersection of Clue and Seek as well, because I think there's, there's so much there.
Excellent. Well, thank
Ian Cook: you. I am also really excited about getting a chance to work more closely with you, but want to say, you know, I think Clue is an amazing company and look forward to seeing all your success in the future. So Alex Elias, founder and CEO of Clue, thank you again for joining us and taking
Alex Elias: your time.
It was a pleasure. Thank you, Ian. Appreciate it. This podcast
Ian Cook: is proudly sponsored by Seek, the leader in cloud [00:44:00] based creation and delivery of industry focused insights. Thank you for listening. If you liked it, please feel free to rate and subscribe wherever you get your podcasts.