Alembic and the Future of AI in Marketing - Ep. 263

0:10Hello, and welcome to the NVIDIA AI Podcast. I'm your host, Noah Kravitz. A quick note before we welcome today's guest. The AI Podcast has a new home on the web at ai-podcast.nvidia.com. You can find all of our episodes there, as well as links to listen to the show on your choice of podcast platforms. If you like what you hear, please take a moment to follow, subscribe, or even leave us a review. And if we're missing your favorite platform on that page, or you just want to tell us something, drop a line at ai-podcast.nvidia.com. Thanks for listening, and let's get right to it. My guest today is working at the leading edge of marketing intelligence.

0:47He's got a fascinating backstory, his company does, and today they're using data-backed strategy and AI to help brands transform their marketing. Thomas Puig, founder and CEO of Alembic, is here to discuss it all, and I've got just enough of a background in marketing myself that hopefully I can, you know, carry my end of the conversation. We'll see. Tomas, welcome, and thank you so much for joining the NVIDIA AI Podcast. Thank you for having me. Pleasure to be here. So I kind of hinted at it, but it's always better coming from the guest than me. We try to do these intros, but interesting story behind Alembic, and I'm sure I only know the tip of the iceberg there.

1:25So can you tell us what Alembic is and the story behind founding the company? Yeah, so we've been around a little while. We're really an applied science company, and so it took us many years to build technology. It began with three people originally, and yeah, still with us today. Myself and my background started at Ames Research Center, originally like when I was a kid, basically, winding into quantitative economics and then decided, actually, I prefer music, the arts and marketing. I went that route for quite a while until I ended up backwards. Right. The other founder was a guy named John Adams.

1:57John Adams, very storied infrastructure engineer in the Valley. He was the 13th employee at Twitter and took the company from the time it was a Mac Mini with a bad Ethernet cable under his desk all the way through the IPO. I think for many years, he was the longest serving person not on the board of directors. Wow. Okay. And then Seth Little, who is a world-renowned creative director and designer who has rebuilt brands for Lego and even done work for Apple, stuff like that over the period of time. The three of us got together, and there were a number of reasons why we really chose this field. But the most important is that we have felt that marketing and anything around the creative and even the arts and trying to promote it had been and no one had been able to be a storyteller for years.

2:40Everything becomes so obsessed with like trying to get that last little click, that last little engagement, that last little performance that we just about lost the plot. And at the same time, I had a few deep beliefs about where things were going with technology that married into that. Yeah. And one of the things that, and this kind of brings it to there, is the company really came out of a lot of the mathematics that were born during the pandemic. Okay. How so? A lot of people think mRNA is the only big tech to come out of the pandemic. The other thing to come out of it was a lot of incredible math.

3:11It was really one of the first times that we used huge scale compute and modeling to actually be able to analyze something causally in real time, right? As like an emergency was happening. Right. We'd done weather and stuff before, right? But nothing like that. And so we're like, well, nuclear, weather, disease, drug discovery, everything else is using these type of super compute modeling and deep learning. Why is everybody who does creative, the arts, marketing, everything else stuck with math in the 1970s? So we're a bit more unique in a company also as well that we actually run our own private cluster in our own cage.

3:46Like we physically own our hardware. Oh, wow. OK, you guys, I should have said the beginning, you're San Francisco based? San Francisco based. Hardware is North Virginia. OK. I'm assuming this is a very nerdy audience. And so I'm going to be like, we L2 patch straight into AWS East 1 out of Equinix. Yeah, go, go. And so when we kind of found the company, we sat there and we're like, well, what would it take? Well, the problem is, is that what it would take to actually do this was an entire deep offshoot of a branch of mathematics. So one of the things we deeply believe, and we talk about a lot, is that the profit at companies follows the flow of information.

4:19Okay. So if you want to follow the money, everybody says follow the money, but really follow the information. For all the information, you will find the money. You'll find the money, right. Right. And what's so important about that is I believe all the alpha, all the profit that will exist in corporations in the next while will all come from private data sets. We are seeing major models and LL models. In fact, there was a paper just released where they showed that models, when training on the same similar public data sets, end up more than 90 % the same all the time, by the way. Okay. okay we are seeing a convergence there and so and jensen spoke to this actually more recently too where we will buy it like electricity right right but that means that it will also be converged and commoditized the way i put it though is that these models are converging right and so as they converge they will be the difference between buying say bp and shell gasoline sure right no one has the same private data now a private data could be a songwriter writes his own song that is a private piece of thing that he wrote that is his, you know, what it is.

5:22Could also be a giant corporation that has a huge corpus of data that nothing else can see. And so I believe that within the next period of time, the thing that will generate the most data will actually be lived human experiences will generate the most brand net new data, right? I'm not saying that there's not a ton of data that's like derivative or whatever, net new data. And so this could be, let's take Disney as an example fake example yeah espn disney plus hulu the parks magic bands all the toys everything else sure marvel a lot star wars a lot a lot monstrous right like you've got a century of scripted ip so all of this means that that being able to take that and feed that and learn from it you know understand how it's structurally connected and then act upon it with agents and models or what the world's going to be like right and that's where the profits can come from right and so So everybody, the more you can act on that with your data, the more it's going to inform what you can and perhaps should do and how you do it and everything.

6:21Yeah. And so one of, you know, the first company kind of really do this in the other days was I'm a huge fan of Renaissance Technologies. They were the first high frequency trading firm kind of out of New York. Okay. I'm a mathematician. Right. And they really, you saw this in the quant space first. Then you saw Palantir try to do it in, they were really trying to find a needle in a haystack, right? Like one actor out of a group. Right. But what really happens is you have this massive data. And so, you know, there's a few reasons I kind of see this already. You know, first, look at open source models, DeepSeq, et cetera.

6:52They're getting close to matching the premium APIs in terms of performance and everything else. Second, companies, these larger firms, you are seeing larger firms be a much bigger part of this than the mobile revolution in the early days. While you have a couple of small companies like Kerser and everything that have a small number of employees that are pushing it through or just being the large LLM stuff. there's serious amounts of capital that are backed by the large backers. Yeah. The private data has much higher prominence in the sub and they have huge corpuses to apply this to. And the third is, is that anything that generates derivative data, say you do something and there's an action taken, it's like compounding interest.

7:28Right, it's a flywheel. Which is a cycle flywheel, just like you're saying. And so it feeds upon itself. And so if we think about like the 2010s, right? Right. 2010s, the advantage went to the team to capture and operationalized new data or audience firehoses. The rise of Facebook, the rise of these things. In 2020s, where we're seeing right now is who could stack the most GPUs. Hardware mattered so much. Right. And like you literally see teams being separated, the haves and have nots, by the amount of horsepower they have to apply to these things. Some of that's being offset by innovation. Right.

8:01Deep seek, et cetera. But still, to a large extent, to serve that, you got to have the power and chips. So I think 2030s, it's going to be who turns their private data and its exhaust into a deep learning asset. Right. Is then applied against. How do you work with your customers? How do you help brands, you know, start to tap into this data, wherever they're at in that process, and do things with it? Gain insights is what we always talk about. But, you know, how do you help them? So think about it like this. I'd say that two things to know about Alembic. One, every single piece of data that ever comes into our system is anonymous aggregate data.

8:38We allow no PII ever. There's literally never been an incident in our system. Just like when you think about where we really based off the biomedical math, right? You don't have PII there either for like a third phase trial, right? You're not tracking. Great. The second thing to know is that we ingest enormous amounts of it. Yeah. We've brought in 100 billion rows of data in three days for clients. And this can be like every transaction from 17 ,000 stores, stuff like that. The problem our clients have and anybody has is, and this is one thing that I find very interesting about why you haven't seen these whole like query the database tools take off as much as you've seen the generative tools take off.

9:16Because the problem actually is, you ever read Hitchhiker's Guide to the Galaxy? Long time ago, yeah. So you get the very end of the book, right? They're like, what's the mean of life, the interest, and everything? Yeah. And they go, well, the problem is we don't know the question. Right. When you have that much data, you don't even know the question to ask. Yeah. To be able to analyze it. Or worse yet, you think you know the question and you're so far off. Then I'm down a rabbit hole. Oh, and then you make a wrong assumption, right? And it becomes a huge pain. And so what we have to do first is we have to ingest an enormous amount of data and we have to signal process it.

9:46We actually use, we announced this at GTC, actually. The way we do signal processing and probably one of the coolest pieces of tech we have is we actually use spiking neural networks to do it. That gives us a lot of superpowers. And the two problems we had to solve for it were this. That's why, why go build? You know, spiking neural networks typically have been done on neuromorphic hardware, which usually people think of wetware like half biological computers. Right. We wrote a simulator for the wetware as a kernel in NVIDIA, just like you wrote a simulator for quantum. Yep. We wrote the simulator for Neuromorphic on the hardware, and that's how we run an SNN.

10:20It's actually fully custom to the NVIDIA chip. Obviously, the answer must be good, but how well does it perform? Really well, because the two problems that I had to solve were this. One, how do you compare apples and oranges? Right, yeah. How is a Nielsen rating or a TV or a Spotify view the same as a phone visit or walking into a store, right? Different modalities, different mediums, right? They all count as marketing, but exercises, touch points, something. But yeah. Marketing has the worst job because their little job is everything. Right. Right. Second reason it's the worst job is everybody on earth thinks they're a marketer.

10:55The second thing we had to solve was marketing. You always hear about this. Oh, yeah, we'll do a campaign for a couple of weeks. So you have no time history and no repeatability. But how do you look for outliers with no time history? So those two problems had to be solved. And SNNs were the solution for that. We turn everything into spike signals, spike encodings. That's always the Apple's news. For people who don't know, one sentence definition of what an SNN is? A spiking neural network is meant to be a digital twin of the human brain to where it can actually process signals like neuron spikes, just like your brain would.

11:27And what it's doing is it's seeing signal and then you're firing neurons off a propensity. There's a lot more explanation to that. The reason that they're really well-loved is they're incredibly fast and they are online, as we call them. So that means that they are always that and they're evolving. so you don't actually train them in advance. So they're an evolving network. So they really originally were really well-loved in people thought about that for IoT and they're fast. We use them to be able to find a way to detect outliers. So a human, a baby can pattern match from the moment it's born almost and teach it.

12:04So when you're looking for outliers in data, normally what we would do is we use prediction, grab all the previous data, see where the highest highs would be predicted Should be lowest lows and anything outside of that outlier. There's other ways to do it too, but for this audience, we'll be overly redacted. With what we do is we're actually looking at patterns and actually seeing those changes from that neuromorph, like the neural perspective in SNNs. And so we can do outlier detections with no time history. No time history, right. The reason this is important is say we actually, we presented at GTC a case study with the CMO of Delta and they sponsored the Olympics.

12:42And there's a great recording of this. Actually, we did a session. And the problem is the Olympics happens for two weeks every two to four years. You know how you sponsor it. If you do daily data, that's 14 time steps. What do you do with that? So doing that signal processing, first, we had to do that. The second thing is we actually had to figure out how to do the connections, connect the chain of events. So say you watch the Olympics, then you Google for a Delta flight, you click on a Google flight tab, and then you buy it. How do you connect those things in time? So that's where we ended up building all of our causal mathematics.

13:16So we use causal inference and transfer information mathematics to be able to do that. So the first thing is we have to see the signal at all. We build each time series with its own mini neural network. That neural network thing gets chained together with causal links, and then we can build chain reactions. Then you can mine the chain reactions for intelligence. Right. Okay. So let's stay with that example then of Delta and the Olympics. A minute ago, you were talking about, you know, with a client being able to ingest their data. And so I'm wondering, can you map like these technologies you're talking about in the spike neural networks and you started talking about causal AI, which was the other other thing I wanted to ask you about?

13:51I need to hit it. What can you do with the data and how do these chain reactions and, you know, how does it apply? What does the client see kind of it? I would say at the end, but, you know, you know, it's funny at the end. I actually believe that all dashboards should never exist. Dashboards only exist to derive intelligence from them, right? Nobody wakes in the morning and goes, you know what I want in life? Another dashboard. Yeah. So what you do is you like look at the dashboard and then you write an email. So the way our intelligence comes out is it actually comes out kind of like, uh, actually the model for it was like the president of the U S gets an intelligence briefing every day.

14:25Yeah. Yeah. Provide them as literal intelligence brief. Great. Yep. Oddly enough, we use L11s at the company, but not in the same way other people do. So there's a game called Mad Libs. I'll play with my kids. It's like you have a paragraph and then you have a blank and you fill in the blank and make a funny sentence. That's how we use LLMs. They never make a decision about data ever, but we do love them for user experience. So they'll write like kind of by the report, but then the causal, the chains will actually just put in the actual data sets. And that means that it kind of inoculates them against the hallucination issue because everything that's an actual real important data point that's not like an is or an and or connector is handled by a deeper method, a deep learning method.

15:04And then we use the rest for user experience and communication. Oh, interesting. Those are derived from the chains, but let's take Delta as an example and talk about the Olympics. Sponsoring the Olympics on a national level is expensive. We're talking eight figures at least. I think NBC reported hundreds of millions, maybe even a billion dollars in revenue off of the Olympics alone. I forget what it was. I'd have to a look. It was massive. And when you're doing that, there's two types of things that occur. One is you buy a whole bunch of 30 and 60 second ad spots that you do. The other is, is that you see this in sports all the time.

15:36You have things that are named after companies. So if you watch the Olympics, you saw that the medal presentation ceremonies were Delta. Okay. You know what sells a whole lot of tickets to Paris? Watching the Eiffel Tower in the background in a really emotional moment with a player as you put a gold medal on them. That was actually more effective than the ads themselves in some circumstances. Yeah, no, I can see it. How do you connect that, right? Somebody's booking a flight, they're doing some planning. There's a delay between those things. It's not like you do it while you're watching the TV.

16:07So you have to calculate the known, we call it optimal lag, right? Between the time series, right? How do you calculate those things? Right. And so what you do is you go, now we think about this like common sense. Of course, I saw that and then someone's going to buy something. It's much harder to mathematically prove it. I can't even imagine. You know, we work with a finance organization where we actually look at what causes ETF, electronically traded fund flows. Conversions can be anything. You know what I mean? You could be selling something or you could be trying to change the volume of something or you could actually care about the most open source contributors you get.

16:42What we do is when we look at these chain reactions, we have nodes and edges as we call them, right? Nodes are kind of the points like connect the dots. The edge is the connection between them. We calculate every connection that could possibly exist as we build stuff. So then we can dynamically search the intelligence afterwards. And so we can literally morph our reports based off what the user wants to see. So if you're like, I want to know all the things that are about selling a ticket. And the next day you know, I want to know all the things that are about game frequent pliers. All you do is change the focus.

17:12It redoes the search in real time instantaneously. And so that type of stuff, what I find really interesting is a lot of people are working AI and deep learning right now, the reason why we have such a sticky in these huge, big enterprise customers is, and we're not a cheap system, is that they derive real value out of it. And also, we understand that we have to meet them where they are. How do you mean meet them where they are? Well, I find that a lot of people are talking about AI and form factor. Should it be hardware? Should it be software? How should speak? Everything else. But I never hear anybody says, is the thing that outputs actually useful?

17:46You know, I believe that actually we're having a paradigm shift in the entire thing. For the longest time, there were a ton of studies that said 90 % of the world was consumers, 9 % was curators, and 1 % was creators. And this actually held true for pretty much as long as people could remember. But now, suddenly with AI, the curators are actually creators. So the 1 % is now 10%. So you've 10x the footprint of people who are wanting to build things, and the strategic monstrous shift that occurs with that is large. the way you put that it reminds me of i mean it reminds me of the 2000s and the web first you know coming out and anybody could publish if you wanted to you know what reminds me of the most is actually when you could start recording your own record in your living room right it used to be you had to have these ginormous studios we see them in films right and everything else and then you have people like beckon everything being like i wrote a hit record on my whole track yeah and the democratization of that i don't think anyone would say music is worse off because of it.

18:48The music industry, the monetization may have suffered, but the actual quality of the music and the stuff coming out, I wouldn't say that you've had a, I'd say we have access to more independent music than we have ever had. Oh, for sure. No, the difference between how my kids get music and how I did, you know, couldn't be more different. So like kind of like when we talk about that, like suddenly let's take code in this example. We use a lot of AI to write code now and this is everybody else. But code used to be or is in a lot of places. like a bespoke thing, right? Like it would be the most beautiful furniture ever and somebody's hand crafting this beautiful object.

19:24Or before the assembly line existed for Ford, right? Any motor vehicle would be like handcrafted by every single little individual thing. But now it's no longer bespoke. Now you have this sense of like, you actually have mass crafted, mass generated objects. In the beginning, I'm sure mass generated cars did not equal the bespoke cars. Of course. Nowadays, I would say probably most of the mass generated cars are probably better than the supposed cars from a safety perspective. Code is doing the same thing. And so now you have two fronts. You have everything able to be like process information.

19:57And the second thing is the curatorial class, right? Those with taste are now creators by default. The playlist is a Veblen good something. Right? Now, I love tools that give people agency to control themselves. I am deeply worried long term about access because that can be hard but i think about the number of musicians who never would have had anyone to hear them at all if a poor track didn't exist yeah i think about the number of writers or filmmakers or brilliant people who are never going to be heard at all if they did not have the democratization of tools when we bring this back around to marketing attention everybody's attention is really what there is is there's only it's a supply side limited there's only so much attention in the world there's only like for so long every company in the entire world is competing for the same slice of attention at the end of the day.

20:47The only commodity that can never be increased is a person's heartbeats. How many till the day you die? Then you can't sacrifice. So when you're doing that, marketing and everything else, and we're trying to do that, we're trying to understand the universe and get down to it. Now, we've got a little like theoretical here, but I think it's an important paradigm shift that we're discussing because you can't separate at the end of the day marketing from the rest of the businesses it governs. So, so bring it back down sort of to, you know, business us and bolts level. How is AI, I mean, as much as you can and, you know, be specific because it's so many ways.

21:22But what are some of the biggest ways you see, Nate, that's a better way to ask it, that AI is already changing how brands relate to their, their audiences and, you know, what's happening kind of just down the road. I think this idea of, and it's funny, as I was prepping to talk to you, I was thinking about, we've done a lot of episodes on health and medicine lately. And this idea of precision medicine, right? And being able to offer, you know, each individual, the individualized care, preventative care, you know, technology enabling us to provide that level of personalization. And so there's a similar thing in some ways, as I understand it, with marketing.

22:02and I mean, to really probably not do it justice, but I break it down in my head to cookies following me around. Yeah. And that being kind of a, you know, to me seems like a brute force, maybe outdated at this point way of doing it. But what's AI doing for all this and what's it going to be doing for brands and engagement and personalization? You take a little bit of an anti stance here. Human beings love building tribes and community and the most visceral experiences that they have. let's take sports people will often say sports is the new church that's not my phrase it's just what everybody you know you want that community whether that is in that format or it's in the sports format or something else that means that you need everybody to experience the same thing and be able to respond to it and if you think about the fondest experiences you can remember almost almost always were there with somebody else right you remember doing something with somebody or how it affected somebody or how it did something.

23:02And so while I believe that personalization is interesting, I actually believe that being able to really build these experiences that people have and to be able to bring people closer together, those are the marketers that are going to win. Mathematically, I talk about this a lot where you're like, what does an LLM mostly do it predicts the next best token inherently that next best token is the one you would expect that's what's supposed to do that's the right that's kind of in a weird way like kind of like that's kind of like it's going to give you the median right it's going to give you the answer you should have but what great artist what great marketer wakes up in the morning goes you know what i want to be the median and i don't know i can really feel it's gotten better But I can still feel the aesthetic of AI imagery and everything that comes out.

23:52Because it is still trying to, it is almost going to have its own aesthetic. And so I think that what will happen is that the uniqueness of experiences and everything else and the premium of that will rise. And that we will have this kind of push against where it's not like all about personalization forever. Everybody's going to realize you still are going to want to experience that with other stuff. Interesting, yeah. And so for maybe 10 % of your life, you're going to want hyper-personalized, like I want to talk one-on-one with it, right? Or maybe it's an ephemeral thing. You're using, I heard a great founder talk recently about how he uses ChatGPT and he writes a story with his kid at bedtime and he just doesn't keep it.

24:31It's audio. Okay. They're just playing with it. That's ephemeral. But you don't go to the water cooler the next day or get online and talk on Discord with your buddy about that. Yeah. And so I feel like often human nature is lost within this. some marketers, our job is to create those experiences. And then when it's a touch point for preference that you've already created, then try to enhance it with some personalization. But I feel like people have lost the plot a bit. Yeah. Interesting. I don't disagree. I just, it's interesting to think about that in the context of, as I said, all the personalization that jumps to mind when thinking about these types of things or the potential personalization.

25:10Well, I always think about personalization is once you've engaged with the brand, How does the brand find the optimal path for you? Yeah. Deeply important. Yeah. If I'm already an Amazon customer, I need the happy path. If I am an Apple customer, what iPhone do I buy? If I have some friends who love Alice and Olivia dresses, they've got a ton of different dresses. Which one should they get, right? They're going to have preferences within that. But I'm talking about like, how do you even get them to like the designer? Right. You have that shared moment. Profit and the margin comes from brand. but what is the value of brand?