513: Why Twitter Rejected His AI Tool | Alex Elias

0:01Hey founder fam, I want to talk to you about something super exciting. We're officially partnered with OmniSend, the email marketing and SMS platform built specifically for e-commerce founders. We've been recommending OmniSend to founder students for a while now because it just works. Whether you're launching your first store or you're scaling to seven figures, it really helps you automate your marketing and get real results. Did you know on average, OmniSend customers make$68 for every$1 they spend, which is an insanely good return on investment. And because you're part of the founder community, you get 50 % off your first three months with the code founder50.

0:39Just head to omnisend.com forward slash founder without the E to get started. All right, now let's jump back into the show. Hear the stories, learn the proven methods and accelerate your growth and future through entrepreneurship. Welcome to the founder podcast with Nathan Chan.

1:02All right, well, welcome Alex. So the first question that I ask everyone that comes on is how did you get your job aka how did you find yourself doing the work you're doing today? Yeah, absolutely. It's been a long journey as I understand it has been for you all as well. We started the company officially in 2011, but have begun thinking about it many years prior. I was actually in law school at the time that we officially incorporated, incidentally focused on internet privacy and doing empirical studies on the efficacy of click wrap agreements and things like that. But it was set across the backdrop of companies like Netflix transitioning to streaming, and all of a sudden, recommendations were this hot topic.

1:51And in fact, Netflix had a million-dollar prize for who could improve the efficacy of their algorithms, which at the time were all DVD-ranking data. and we sort of took that challenge head on and realized that there was huge pitfalls to sort of using data sets that were siloed in particular categories and so this is something that I pursued with kind of an academic interest at first. I as a hobbyist I play the saxophone the tenor sax and I play some piano and jazz is one of my big passions. I also love you know, mid-century Italian cinema. And I began to see this sort of relationship between all these different areas of taste that we wanted to explore more systematically.

2:36And so all of that to say, kind of embarked on this journey of could we do a better job sort of mapping, understanding tastes, and is there a business to build around this? And our original premise was that there was, and it was aimed squarely at the consumer, at essentially giving them better recommendations sort of a roll-up, a next generation of the entire user-generated review space from Yelp to TripAdvisor, IMDB, Goodreads. Most of those properties, incidentally, have been rolled up by Amazon. So we wanted to build a better version that sort of had a holistic view on taste and can give recommendations.

3:15And little did we know it would be a very long and winding road since then to build the business we are today. but that's how it began and I actually officially incorporated my third year of law school and went off to the races. Yeah there you go so what I'm really excited to speak to you about today is is obviously the topic of AI and everything that you're doing with Clue and and how you've come to kind of become a thought leader in this space because the thing is AI is super hot right now Like it is the hottest thing that, yeah. So it is the hottest thing since probably the launch of the internet in many, many ways, right?

3:55And it's one of these things where there's a lot of experts coming out of the woodworks, talking about the future of AI, talking about where the technology is going, talking about the future of work. And I just want to hear from you kind of, you've been in this trend for a very, very long time. well before it was a trend what's the last half like year and a half been like for clue and where you see it going did you think it would explode now or did you think it was going to be years ago or like talk me through it yeah absolutely it's been an absolutely uh crazy year and a half um you know to some extent we've taken a long view uh we've been doing this over a decade.

4:41And it felt like certain things were taking forever to materialize certain trends. So to some extent, I'm surprised at how soon certain things have have materialized. And we're very fortunate that they have because it's kind of represents this whole new sort of landscape of commercial surface area for us. But yeah, it's been it's been very interesting. I think people don't realize from the outside how incremental a lot of this has been. There's, you know, transformers have been built on top of other modeling paradigms that are built on top of other modeling paradigms. And, you know, it sort of bursts through the surface now with, you know, certainly the language models and a lot of the generative applications and sort of colliding against that backdrop of how much sheer computing power is now available.

5:30And so it's been really fascinating. We're very lucky in that we've carved out this unique lane or niche in kind of understanding taste that is somewhat portable. It has the ability to power anything from a mobile interface to what I consider to be the greatest new interface, which is language. And so these foundational models still need to be imbued with some source of truth and so on. So it's been fascinating. And I think one of the really interesting things is sort of watching the regulatory environment and even the private sector sort of response in a lot of traditional industries try and catch up and cope with this new reality.

6:15I'm an optimist in many ways about it and I'm sure we'll talk more about it but there's yeah it's it's it's been an incredible year and a half and what we're really excited about is that there's whole new applications for something we've been working on for you know over a decade yeah and one thing that Sam Altman has said publicly is he believes now within the environment we're in and with the capabilities of AI technology that there will be a one-person startup unicorn where there's don't need any employees one person uh do you believe that's possible uh yes I think there's some truth to that uh and it's not necessarily for some of the reasons that that that one might might think um obviously there's this there's going to be a productivity boon, there's the ability to, I saw the NVIDIA, you know, our favorite CEO at NVIDIA talk about how programming is no longer something people should learn.

7:18I somewhat disagree with that, although I might not have the standing to challenge his assertion. But yeah, I think one of the great things about this new era we're in is it's kind of like a fractal. There's just to be so many smaller problems to solve. And I think there's a tendency to think it's a winner takes all zero sum game where, you know, whoever's in this, in certain cases, stolen the largest corpus is going to run away and kind of create AGI. And that'll be, there'll be one application that serves it all. But the reality is that it's going to be a constellation of very, it's going to create kind of all this new problem space.

8:01We've been the beneficiary of that because we have, we solve one very specific problem around consumer taste and there's going to be a lot of others. And so I do think there will be definitely in terms of the ratio of like the enterprise value of a company to how much headcount actually was required to make that happen. I think we'll be new records for sure. I doubt it would be surprising if it were truly just one person and one person only that took it all the way. It's certainly possible. I mean, we are, you know, the AI itself from a productivity standpoint has the potential to augment, you know, people's existing skills.

8:41And so, yeah, I think, you know, there might be a record set for sure in that realm. and I'm an optimist with it. I think there's a lot of angst about it taking jobs, making jobs redundant. I know there's an AI artist named Claire Silver who recently said that she believes it's now taste, not skill. That's the kind of most important aspect of creating art. And I think that's a profound statement because it's not to say we don't need artists anymore. It's just that the role of an artist and the function of an artist is a little different. And everything from prompt engineering and guiding the AI and so on becomes this whole new way of creating value, really.

9:33So yeah, and I think, yeah, so I'm optimistic that there'll be a whole new set of problems to solve. There'll be new jobs that are created. And I do think on the whole, it has the potential to really make life better for, you know, most people. Yeah. So it's definitely interesting. Like I find myself, I was, you know, a year and a half ago, this time last year, a little bit before that, it came out. I was traveling around South America. And when I say it came out, ChatGPT was released. and it really started to popularize, you know, prompt engineering and, you know, working with a co-pilot to help you create things.

10:20And then I was kind of off it. And then for me, I started to get more into it the later half of last year when we started interviewing founders like yourself and we did a whole virtual conference around it just to understand how founders are using it and now as time has gone on I'm finding I'm using tools like chat GPT every day every single day to analyze data summarize things and then I think to myself like I was I was like last night you know just for three hours going up and back okay change this change this what if we did this what if we do that and it's like I'm brainstorming with somebody and it's and I was thinking to myself wow like this is crazy like how much more productive how much more effective I'm getting.

11:07I'm curious, have you had like, because you've been in this space for so long, are you like operating on a whole different level compared to most because you're so used to looking at the power of AI, seeing how you can use it in your day to day? We're certainly in a really good position to kind of capitalize and appreciate some of the benefits. I mean, just as an example, you know, in our own processes we've had years ago, we had to hire a whole range of kind of knowledge workers from the different from two universe kind of fashion schools, institutes in New York City, to help kind of catalog meta tag, summarize the sort of aesthetic qualities of different fashion brands and designers.

12:00And, you know, certain tasks like that, particularly around summarization and entity extraction and so on, you know, we can now layer through different AI processes and, you know, put that effort on steroids. So we've, you know, we've definitely benefited from that. We have a pretty acute sense of how to use it. And sometimes it actually creates new work. You know, I mean, you sort of talked about how you're prompting it and engaging with it. And, you know, ultimately it's a pretty time consuming task. And I was recently speaking with a pretty prominent comedy writer. I'm currently in Santa Monica in Los Angeles.

12:37And there was a comedy writer who's very prominent and was describing sort of engaging with GPT on certain outlines of sketches and things like that. And how in the end, he's in awe of the power of it and how much he can actually coax out of it. But in the end, he said he actually spent more time writing as a result of this kind of dialectic with the AI. And in some ways, it actually took longer just because you get into this kind of optimization loop and you're surprised. But the end result might ultimately be, you know, might be greater. And the way he described it, it'd be hard to argue that that wasn't his novel and, you know, true thumbprint is real output.

13:22You know, it's not not as if he's just outsourcing to the machine. And that said, it becomes sort of an extension of his own cognition. So I think it's interesting. And I could imagine people who use it for emails and things, if they're really kind of perfectionist, you could almost end up getting into an optimization loop that decreases productivity. but that's you know it all depends on on sort of how how people engage with it yeah look that's a really interesting thought because i was spending a lot of time going up and back up and back but i was thinking to myself yeah right if i was by myself i probably wouldn't have got as far but it did take a lot to get it to where i wanted to go and i was trying to perfect it so yeah that's a really interesting thought right right so talk to me about i guess because you've been in the space for so long you guys have been an early adopter of an industry that has had a very very slow brew right um what are the advantages of being an early adopter in an industry which hasn't been you know gone gone mass market yet and then also what are the advantages so what are the advantages What are the dangers just for founders that are looking to enter new spaces like AR as an example or VR, right?

14:39Like that's a still a space, you know, the Apple Vision Pro came out and that's kind of one big step, but it's not fully. Right. Absolutely. Yeah. I think one of the advantages is that you have a sort of gestational period of building something thoughtful. And, you know, we've for a long time, we didn't have the commercial traction that we have today. And so we really just focused on improving the core, improving the quality of the corpus, the data sets, the ontology that we train on, procuring that data in an ethical way, not taking shortcuts. So oftentimes, if you're lean enough and you have a gestational period, you can build something that's truly thoughtful, that becomes a differentiator in a kind of emerging landscape where people have to rush and take shortcuts that are kind of new entrants into a field.

15:33The disadvantages of being early are that the market is so undefined that sales cycles could be nearly indefinite. You spend a lot of time educating customers about what you do rather than sort of selling. You know, back when we started, AI in the commercial landscape was kind of a dirty word. Like it meant, you know, in 2014 or 15, talking about AI, it's hard to imagine now a decade later, But it was kind of a, it was a very sort of opaque kind of not fully respected kind of moniker. I mean, we would use euphemisms like we're a personalization, you know, technology that, you know, we would avoid, we would essentially go out of our way to avoid the word AI.

16:24And in retrospect, I wish we acquired more domains. I mean, we have amazing AI. We finally, around 2017, 2018, really started owning the kind of IP around AI and cultural AI and taste AI and so on. But yeah, I think there's definitely risks there. And if you end up burning too hot in terms of just the capital and so on, you could run into trouble. Like it's, it's very, very important when things are not clear yet to stay as lean as possible so that you have maximum latitude. Um, but, but yeah, I think there's, uh, there's definitely advantages and disadvantages. Um, one of the biggest advantages is that it's not a crowded field when you're kind of early to something.

17:10Um, and, and it's a non-consensus area. Um, you, you end up having the ability to build and sort of, you know, in a, in, in kind of have the space to do things that are not possible when it's a heavily funded super, you know, kind of a hype super cycle type area. So it's been interesting to see the contrast in that. Yeah. And I'm curious, like what kept you in those early days or even the past, you know, five, six, seven years when you didn't want to talk about, or AI was a dirty word, what kept you going? What made you not want to give up? Yeah. Honestly, sheer stubbornness and kind of like the real among our founding team as well.

17:57And we've had very little turnover. There was just this really stubborn notion that this data set has to has to happen. And someone has to build this type of understanding of taste because it's just being it's getting increasingly fragmented. Like the problem that initially drove us to start the company was sort of getting worse over time as there was more and more walled gardens. And then, you know, the the identity based approaches that were huge in 2015, 16, when we would talk to a customer, say, in financial services or elsewhere. where at that time they'd be like, why would we use your approach, which is essentially kind of an informed guess, a probability, you know, with AKA AI, as opposed to, you know, having the deterministic, like we can actually go to a company like Axiom at the time and sort of get Nathan, you know, get actual data joined to Alex on when I applied for an auto loan and make and use that data for X, Y, and Z.

19:01it was kind of a tough sell to be like, why don't we use this informed gas that's more ethical? But we really sort of felt like we were mission driven. And I think when you have a hypothesis that you really want to prove correct, it sort of gives you a lot of staying power. And we had enough breadth. So, you know, I think there's this conventional wisdom that you should be a wedge and you should try and be as focused as possible. And it's, you know, espoused in the innovators dilemma and all this great literature around sort of starting a company. And we were often accused of that early on, like you're way too broad.

19:44You know, why don't you focus on one area like music? You know, a company like Echo Nest that got acquired by Spotify was a perfect example. They focused on music data science and that was it. and what's become apparent over time is anything non-consensus really if you end up being correct could be a huge advantage and frankly our strength was in our breadth like even when things got really challenging we started building such a unique data asset because in some ways we were so unfocused and we were so broad and but but there was an under there was kind of an ultimate focus in that because we said, you know, our strength is in the breadth, in the sort of horizontal ties across these domains rather than just completely burrowing into, you know, into one kind of category of taste and so on.

20:35So that was a non-obvious thing to do. And I think sometimes, you know, there is even challenging some of that conventional wisdom about entrepreneurship that, you know, there's many cases in which the innovative, the thin edge of the wedge and the innovator's dilemma and things like that are incredibly valuable concepts that I think are, you know, empirically proven. But sometimes things that might be one of your, you know, a big weakness or perceived weakness could be a real differentiator because other people, it's a non-obvious way to approach a problem, you know. um and when you talk about cultural ai can you explain to everyone what that is yeah so cultural ai is basically purpose-built ai uh to understand taste and preference uh and be able to kind of uh within all the areas of of taste that humans care about that are kind of linked to our preferences and and our passions so things like music film television books literature or podcasts, but over to travel and fashion and restaurants, cultural AI broadly as an umbrella is just the sort of practice, the data science and the kind of recommendation science around being able to generate predictions about tastes and preference formation, given sort of like prompts, you know, like any other AI clue response to prompts.

22:06And a prompt could be, you know, here's a collection of movies, tell me what sort of music this collection of films might lead to, or what are the most correlated fashion or travel destinations to a particular restaurant, and so on. It's the inference generation around all our tastes and preferences. And it's a fairly unique lane. It's a highly structured, it's a challenging problem so that, you know, going back to the beginning of the conversation, that Netflix prize competition, that million dollar prize, ended up being kind of a failed effort because, you know, the leading, the winner barely improved the algorithm's efficacy.

22:51And the New York Times at the time wrote about the Napoleon Dynamite problem, the fact that, you know, Napoleon Dynamite was this weird comedy, and more than 20 % of the aggregate error rates or something in that order of magnitude was attributable to like the challenges of predicting taste in this one film inside and out of it. So yeah, that's the broad area of cultural AI. And basically, it's a kind of primitive in this world. So it's something that's highly composable. People can build other interesting products on top of it. One of my favorite stories actually in sort of the early days of LLMs, you know, kind of probably along the time you were traveling that day.

23:36And we ended up sort of providing our tech to a large hackathon with a very large media company. They were trying to build interesting applications around LLMs. This was in the GPT-3 era and sort of late, really late 2022 at this point. And we basically supplied our taste AI. And one of the winning concepts was being able to pass in one single taste, one single named entity. So it could be like your favorite music artist. I did it with like one of my favorite jazz saxophonists. And basically it layered that taste into Clues Cultural AI or Taste AI and sort of extrapolated all the associated nodes and edges and likely preferences surrounding that.

24:24And then it passed that payload into the LLM. And then the LLM wrote a long form essay that insulted your taste in like meticulous detail. And it was pretty incredible. I mean, it was just like it was so I mean, it ranged from funny to like actually insulting and was was just an extraordinary demonstration of, you know, and that's not a particularly useful product for society. but it's just, you know, it demonstrates how composable these different sort of elements within AI can be. Interesting. So with cultural AI, does word of mouth matter? In terms of the commercial landscape, like getting people to adopt it and leverage it and so on?

25:11Um, yeah, absolutely. I think, um, customer acquisition, you know, we've had a lot of, some of our best deals have been engineers that wanted to work with us in one company and we couldn't get a deal over the finish line. And then they sort of, they're, we've been around long enough, their career progresses and they're now in a different place and they're like, oh, clue and bring us in. And, you know, then we, we, we come in in such a way that, you know, there's a real profound understanding of our capabilities and, you know, it hits the ground running. And those are some of the best commercial deals we've had.

25:48And we've even had some of our, you know, we've hired from in a very above board way from some of our biggest customers and vice versa. There's been a lot of interchange of ideas in that regard. And then some of our biggest customers have also come. So we acquired a company called Taste Dive back in a few years back. They're based in Amsterdam, and it's kind of a leading cultural recommendation engine for consumers. It sort of goes back to where we started Clue. That was our original goal. They have an API as well that's very low cost, but it refers people. It's sort of an entry point into our broader universe, and that's been one of the best ways in which we've been able to sort of get companies to use our service by kind of letting a portion of it out in the wild, totally free, no advertising.

26:42And, you know, some portion of people using it who find it useful are developers, some of them might be at large companies, and that little seed is planted. So yeah, I think word of mouth is really important, especially with something that is as our sales cycle. So most companies use Clue through application programming interfaces or APIs, and there's some developmental overhead in that, and there's, you know, compliance overhead. And so the more, even within a large enterprise, you know, those softer elements of really getting personal relationships and buy-in from various stakeholders across InfoSec and, you know, CTO suite, the CFO suite, and so on is really important because any one of those factors could be something that hinders the success of the integration ultimately.

27:35Are you hesitating to take the next step in your e-commerce journey? FounderPlus has you covered. With proven frameworks tailored to your business needs for fast results, a supportive community of over 30 ,000 like-minded entrepreneurs, and weekly live mentorship sessions, FounderPlus is your key to success. Try Founder Plus today for just$1 for seven days and start building your dream business with confidence. You can visit founder.com forward slash start dollar trial or click the link in the description to claim your trial. Yeah. So you talked about customers and kind of use cases and you guys have many different use cases for Clue.

28:14I'm curious, can you tell me about the first customer that you guys got and like what that experience was like in a time when it you said like ai was a dirty word almost yeah um so yeah i'll tell you about one of our first major customer wins and it was kind of it ended in disaster which i think makes it more interesting but we um basically we had had our consumer application our eponymously called clue app which is very similar to what taste dive offers uh it was live in the app store. It scaled over a million users at one point, never quite hockey sticked. And it wasn't a great business because we were depending on affiliate revenue.

28:55So in other words, we recommend a great movie to you or book, maybe you buy it from Amazon and we get 4 % or whatever the percentage is. So there was a very senior product manager, and this is all declassified at this point, at Twitter at the time. And this was, you know, a decade ago, roughly, who became a user of the Clue application and was like, this is amazing. This could be incredible for what we're thinking about. Some of the real, you know, top line things we're thinking about at Twitter at the time. And so the initiative that was underway, she led a large team that was kind of focused on a commerce initiative within Twitter.

29:36And the idea was to have kind of entity extraction and a whole shoppable experience where if I tweeted about a book or a movie or a restaurant I love and you were consuming that tweet, you can click on it and you'll have a whole kind of experience that not only shows you the details of that entity, but then has a whole recommendation engine. So this whole discovery funnel where it's like, here's the other books and movies and things that are related to this. And that entire apparatus, the thinking was they needed this kind of cross-domain taste AI to really undergird that. And they didn't want to go through the hassle of building it because our smartest customers realized it's a huge pain to get this right.

30:22It is very difficult. It requires a ton of proprietary data, years of training, and so on. Uh, so they were, they were very smart about it and we worked together for over a year kind of productionalizing this. Um, and then I was, uh, on my, on my honeymoon, uh, in, and, uh, in Italy, uh, and got, got an email, started getting phone calls from my company and, um, got, got email, you know, an email in and, you know, I was obviously trying to like take a few days off from work. But it turned out that, you know, a certain CEO came back on board of Twitter at the time, and completely sort of scrapped the product roadmap, reorg, got rid of a lot of people.

31:07So they had, you know, they let go of a significant portion of staff, and then decided that instead of this commerce initiative, the most, the sort of top level focus would be on customer service, and the sort of SaaS aspects of that. So like people complaining to United Airlines and having Insight and United being able to address, like that's what they basically were shifting focus away from commerce and towards. And it's a real shame because we got to the 11th hour of kind of building and scaling this proposition out. And it was an extraordinary product. And I think it's still one that's needed, frankly.