Alembic raises $145M to prove causal AI can measure what brand marketing actually drives revenue

And so then our clients started asking us to be like, "Well, can you do our FPNA for finance? Can you do our supply chain? " And so that's really kind of where the business came from.

But the core of it is is that like when you have low information or low amounts of stuff, you want to be able to actually affect the metrics you care about. Take us through some of the case studies. Who have you worked with? What's the most concrete example of uh you you talked about buying a stadium.

Have you literally found whether or not Crypto. com Arena pencled out? Have you have you looked at these? You know, I grew up at the Staple Center. I I I miss Staples. Uh, it's been renamed, but but and I want to at least know the red stapler that likes I want to at least know that that they was worth it financially.

Is it was it worth it financially? Can you tell me that? Yeah. So, we did a check staple stock since they [laughter] Yes. Should Staples have stuck with the sponsorship? See, the problem is it doesn't count for execution risk. Yes. Yes. So, when we uh there's there's always unknowables, but but what can you know? No.

So, we did a great uh case study with Delta Airlines that we actually presented with their CMO at the the Nvidia conference at the GTC conference where they were sponsoring the Olympics. Yeah. And one of the things about these large sponsorships is there's two big aspects people don't talk about.

One is every time you've ever worked in the business side of the house, they go, "Hey, you need enough historical data to be able to do something. " Yeah. Right. The second is it takes a lot of time to get a response. Well, when they were doing this uh when they were doing the Olympics, they were doing the promotions.

One of the most interesting things we found out is we analyzed all this ad work that's coming out. You have only two weeks. You're holding up tens of millions of dollars on the P&L line while you're doing this, right? These are not cheap options.

And when we pulled the study, we actually found out that the best piece of content that functioned for them, the most profitable was not actually the content that was the ads. The ads did okay, right? 30 seconds, 60c spots.

But what ended up happening is they actually, if you watch the Olympics, they had the Delta medal presentation ceremony where like every time an American athlete would win a medal, they would take it, put it on, you know, their shoulders and that would be, you know, a really emotional moment.

Well, it may seem kind of obvious in hindsight, but when you have the Eiffel Tower in the background, really emotional moments, you sell a lot of tickets to Paris. But the key is if you can know that within a few days, you can either double down, right?

you can uh sponsor the next Olympics, you can do all these things and then you can actually act on it. Now the problem also with when you do big huge campaigns like that is you walk into the lounge for Delta Airlines and Team USA is the Wi-Fi passcode. It's not like one simple campaign, right?

You pivot an entire company to a messaging set, but we were able to tell them within a couple days down to the dollar a material amount of cash that they were able to pull back and then they could actually show that to you know executives. So tragic oldest oldest problem in advertising, right?

Of like I know my advertise like half of my advertising. That's a line. I'm 100% sure. Yep. That 50% of my brand marketing is working if I only knew which 50%. Exactly. Which has been true in brand marketing forever going back to the 1980s. Yeah.

I'm sure in entertainment especially because there's less like you the attribution is really challenging. We did a bunch of marketing and then people went to movie theaters and then some people streamed it and it's hard to actually track it all the way through.

What the the brilliance of what Tomas and Olympic have done is is that using well first of all and he should explain it because he's the brainiac here the new math that came out of contact tracing from co which actually has impacted many different businesses in terms of research and you know causality and attribution.

We got advertising learnings out of contact tracing. By the way, it was absolutely there's no that's a biggie. [laughter] Very American to to like you know entrepreneur says turn it into an advertising.

But when you think about it is is that either for that for co testing or for any of these attributions what they are able to do which was not possible even only three or four years ago which is to be able to ingest billions of rows of data. Yeah. From all sources. Yeah.

um and using machine learning and AI to actually digest that and make sense of that and actually then be able to see directional in in time you know it's they'll he'll explain neural networks and how this actually works which as I said is way above my pay grade here and it but I can tell you which is to your to your question here which is when we went to the top brand marketers the first reaction was no way.

Yeah. You know, this is like, you know, yeah, it's not possible. And this is use like telling us, you know, uh, you know, pot of gold at the end of the rainbow or, you know, Dumbo flies or, I don't know, whatever you [laughter] want to say. And they went not not possible.

And that's when these guys would come in and do these PC's. Yeah. Disney, Mars, Accenture, Delta. And you know, in the world, once is an accent, twice as a coincidence.

three times you go okay well this thing actually delivers as fanciful as the notion may be it's real yeah can you tell me a little bit of uh the history of how uh like product placement works in Hollywood because that feels like the classic example of difficult to measure but you've been in that car in this world by but it's but it's when you start to actually think about this there's a trillion dollars a year glo globally spent on brand marketing and brand marketing is everything from crypto on the Staple Center [laughter] to an interstitial to imagine in the Disney parks all the things that you're you know that they're able to offer to their brand partners to putting a logo on a car.

What do you what you know what's the value of having Oracle Yeah. on the race car. on a race car in this or or a patch on a base on a baseball player or a basketball player, you know, so is it has it always just been intuitive or has it been more relationship driven? No, it literally is intuitive. It is.

I mean, he can tell you there's there's a there's an old math mm. He'll he can again Tomas can really take you through explain it and that literally dates back to the 1980s. And one uh it's directional, not specific. And two, the lag Yeah. between the time it happens and you're able to gather the data on it was months.

Yeah. So it doesn't the value of it, you know, is really questionable. I I guess my question for you is uh it it's very clear that with the with the progress in AI there's the ability to find insights in data like clearly we see this across everything.

Uh my question is like there there's so little ground truth that how can your clients AB test your solution against something else like if I go to a you know a a coding agent and I ask it to generate some code I can run the code at the end and I can say well yeah the code worked right or I can read the report but if I go to you and and I say how much was this sponsorship worth and you tell me it's $4 million and I say okay like maybe it was four I I that that could have just been guess.

So, how do you justify the the results that you spit out since we can't run a control trial? So, validation is one of the number one things we get asked constantly. I'm sure when we give out uh papers, it's really funny.

We have one paper where it's like like a two-page brochure and then the next 15 pages is how we validate. Yeah. Right. Exactly. And you know, one of the things that people don't talk about is it's actually very possible to test these things. So, what you can actually do is you can do what we call backfit testing.

could set the machine backwards in time and then you can test against things that you know you can withhold information. You can do [clears throat] all sorts of things to actually see whether you know you feed the machine up to like you know what all the results are going to be this year.

So when you predict what something's going to be you can see how close you got. And so I think that um when we're testing things we use a variety of both synthetic data testing methods. Yeah. Because we're not you have to keep in mind we're not a transformer model or an LLM. Sure. It's an entirely new methodology.

And so when we grab this stuff we have both synthetic data we can build.

So there's systems like for the nerds like tiger bite that does causal synthetic data or there's real data like efmri data that we actually know that there's a cause and effect that we can pull from physical world or physical bodies that you can test the algorithms against as well and then from the business side of the house there's lots of ways to do it.

Now one of the things I like to correct on this is it's like we often say the mantra of our company is we're about being directionally correct not specifically wrong. Okay?

And so when you need to make decisions, a lot of time when we do these simulations, we're like, here's the top 20% of things you need to absolutely keep doing cuz they knock it out of the park. Here's the 20% of things that are just literally hurting you. Okay?

And when you have zero information, like you're in a pitch black room and you have no idea where the exits are, would you rather take a really really good educated guess about where the exit is or would you rather wander around the dark? Yeah.

And so I think that one of the things about business intelligence is when you have zero information, the value of the information you can get is that much more important. Yeah. And so we do confused about the the back testing thing.

Like if I like how would you go back and and assess the value of like Budweiser sponsoring Super Bowl 40 or something. It's like you can't you can't run the counterfactual. Well, counterfactual simulations are assumptions. So think about like this. U a lot of people talk about probabilistic graph, right?

graph analysis probabilistic graphs whether they're causal whether they're anything else is um to a certain extent I mean a little pendantic LLMs are almost a probabilistic graph right undirected till you query it a neural network and exactly and so when we do this type of analysis the thing about it is is that we actually have seen lots of instances of the same exact thing it's about the specificity so we just did a calculation literally yesterday that we looked at one trillion connections across six months for a Like that's the type of scale of analysis we're doing.

We haven't seen when you have it when you say you the counterfactual when you're saying what does it look like when you have a base state. We've seen the operating version of a company over years. And so we know what the base state looks like when there's no influence from that. The key is you have to have enough data.

So in the old world when we're doing all this analysis, we used to say you hear about the term overfitting, right? Everybody's worried that is the model just biased. Well, in the old world, you'd say, I want to reduce the number of features, reduce the dimensionality to prevent overfitting, right?

To prevent overprediction. Well, in the modern world with like computational statistics nowadays or AI as we call it, you want more features. You get more accurate the more data you have. That's counterintuitive for how people think about these type of things.

So my answer to you is we have to have immense amounts of data, right? And so we ingest a lot of it. And then that gives us enough vision, right, that we can see what a state would be and what it would not be. And then we also provide our customers with confidence and everything else.

So there are times where we really really know, right? And we go this we have a 100% bet on. We've seen enough examples of this. And sometimes we go, you're asking for a call on this and yeah, we've got a pretty good guess, but you should still keep it. The interesting thing is is, you know, um, he's going fishing.

Sure. To see where, you know, the best fish are. Sure. And interestingly, we did one of these uh PC's for a very very big branded company. Uh and they were looking for the positive impact of uh event driven Sure. um uh brand marketing they were doing. Yeah.

when they went out and sucked in all of this data to do this assessment for them. In addition to finding what were the sort of positive impacts of this, which were modest, what they actually caught in the net was unknown to them, a promotion being done literally in Canada by a, you know, little subsidiary. Interesting.

That was just a you know, like a a regional commercial. Yeah. Which somehow or another bled over into the states. So was being run in Canada by a subsidiary there. Bled over into the states and had a tremendous negative adverse impact on brand. Yeah. Those Canadians. [laughter] I mean, yeah.

The classic example is like social media manager intern of your Canadian, you know, offshoot is doing something that goes viral in America and everyone. This one was even funnier. bought out um a national the national sports national hockey league final spot. Okay. And so you that's going to be huge in Canada. Yeah.

So they but that broadcast across the entirety of North America. Okay. So every so like suddenly this little like thing there a drippy ugly hamburger and a hand and I'm sure maybe natively they thought it was funny. Okay. Interesting. Interesting.

[laughter] And so you're looking at like when when sales data is happening relative to when the campaign goes on and then you tease out from there. Yeah. Think about like uh you know one of the big inspiration for the company was Renaissance Technologies out of New York. Sure.

So the high frequency trading firms and everything they can do a pretty good job about knowing when things affect each other or not but you have to have an immense amount of incredibly high speeded data sets. Talk about Accenture. You're partnering with them.

Are they are they uh just an investor or are they also a goto market channel? Well they actually started as um one of these uh tests we did because they were curious customer. Yeah. And the interesting which is the best way when you think about it John.

So they started as a customer then said wait a minute we we have a multi multi-billion dollar business around marketing go to market it seems like a lot of people go to Accenture for these questions Bane and BCG and Mackenzie and so then it went to hey can we help take you to market which they've been fantastic at I can imagine and then that led to when this when Tomas's doing his most recent round them stepping is the biggest uh venture investor they've ever done.

Wow, that's amazing. So big companies, Fortune 500 have been hiring Accenture and other consulting firms and research firms to help them understand uh what is driving results positive and negative in their business for a long time.

Uh what's the timeline to productizing what you're doing to a degree that a a much smaller company, let's say a company with like a million dollar a year advertising budget can actually start to get value out of this? Oh, that's interesting.

Oh, so this is actually one of my favorite questions because it has to do with the law of long-term vision of the company. Um, one of the problems you get when you have mid-size or smallsiz firms, uh, and they're near and dear to my heart is that they literally have never done things.

So when you have a really large corporation, right, they've been in a podcast no matter or not, somebody's mentioned them, right? They have all of these data sets across everything. But when you're a smaller company, right, you say a million dollar year business or something like that, you haven't done everything.

So there's no actual priors, there's no data, we don't know how they react. But eventually when we see enough of the universe, right, just like you hear about world models, just like you hear about anything else, we'll actually know what the causal universe looks like.

what is the actual most likely outcome of when somebody does something. Yeah. So in the future in a in a say a couple years we'll actually be able to build synthetic data sets that you can send us any query and we could respond to you what the most likely outcome would be.

And where this gets really really important is I think that um especially when you're dealing with private businesses, the world of LLMs and everything nowadays, I think they're amazing, but they're quickly converging, right? There's not going to be that much difference for a client between like Chat GBT and Claude.

Mhm. The problem is that if you're using that for business intelligence and business decisions, what are you going to do when your competitor gets the exact same answer and strategy you do? That's a real problem.

And so we believe that we will take the best private data sets in the world, do stuff for just them, get the get our overall learnings in other places and then we can actually provide people with strategies that are unique to them. Yeah. Right. Augment the other sets of intelligence.

But yeah, there are certain brands that will get a better return on investment from being in the Super Bowl than others. Yeah. If you are a no-name brand and you just put up a 30 secondond spot in the Super Bowl, people are going to be like, I your website might crash, right? Like all sorts of things. Yeah. Yeah.

your web server might you might not be ready for it but also people might just be like I'm not ready to learn about that as opposed to I actually going to am in the market for beer right now and thanks for showing me those clouds that makes yeah but I mean to kind of like talk to your question like that is an absolute dream that I have right of being able to level the playing field across that but also provides these massive businesses they can spend $20 million to figure out what's really working and then do a lot more of that whereas a small firm is like kind of doing the vibes based analysis and started my career uh helping companies like advertise on uh YouTube channels and with podcasts and you know they might run a $200,000 campaign and there's some like direct attribution that they get either from a landing page or a code but then they're like wait our conversion rate is just going up on the site generally totally is that being driven by changes that we made at the site level changes that we made to the offer is it just overall lift from from podcast advertising or is it some other strategy entirely and so the the the more you can bring like real business intelligence to small companies, the more they they'll be able to actually compete against the the big guys.

And that makes me happy because everybody should have the level playing field and whoever has the best strategy and product should do well, right? Um I think that it's important to note as we kind of talk about this thing. We spent years years building the signal processor for this thing.

We had to figure out how to bring in all this unstructured and semi-structured data and be able to basically do data dog for unstructured data first before we could even try the causal thing. And so we have years of working on that.

And that ingestion pipeline that that skill there is what allows us to do what you're talking about. You can't just be like I'm going to slap a model on top of it, right? You actually have to be able to uh have a sensor that can actually understand every data feed. Mhm.

Have you found a company yet that Jeffrey can't get a intro to or directly connect you to the CEO? Like this guy. Oh, you want to meet this guy? Yeah, I'll give him a call right now. Here. Here. You ever want to not take a bet on something? That's one of the things I would not take a bet on. Right.

[laughter] Jeffrey will find them. Yes. Yes. Yes. Hunt you down. Well, well, [laughter] congratulations on the progress. Thank you so much for coming by the Ultra Dome and uh yeah, good luck with the next with the next phase putting the capital to work.

Uh it's going to be an exciting time and I'm excited to hear more of these case studies as they roll out. Yeah, I really appreciate it both of you. Yeah, get let us know when you're ready for your first podcast customer. We got Yeah. Yeah.

We got to figure analysis that we want to I mean, we we just do the we do the vibe space and we did a billboard. We did a billboard in New York in the house really well. We should just do it for fun. We should just do it for fun anyways. Like there's nothing more that I like than like looking at data sets. Yeah. Yeah.

We ran we ran a billboard campaign in uh in Manhattan. We ran two exactly two billboards. We have no idea whether or not it worked. It seemed to work because people shared it a lot on social media. For a million dollars, he could tell you that. [laughter] Well, we'll be a bargain that for you all.

We could do some drinks. It'll be [laughter] fun. Well, thanks so much for coming on the show.

Uh, congratulations on all the progress and we will go back to our regular scheduled programming and I will tell you about AIO customer relationship magic the AI native CRM that builds, scales and grows your company to the next level. Um, Buco Capital Bloke is also blackpilling on the timeline.

Uh, there it is a blood bath in the markets. Nvidia is down 4%. Microsoft Michael Bur rage quit. Michael Bur rage quit. We need to talk about Michael Bur. I'm sure there's something in the stack. Um, let's go through some quick updates as we run through the show. Uh, oh, this is cool. This is a white pill.

Google the mine. Sema 2, our most capable AI agent for virtual 3D worlds. Tyler, what's the deal with Sema 2? Yeah, this is really cool. Um, so this is a it's like a general model that that can basically play like any video game. Sure.

Um, which is different because like you've seen a lot of like even early OpenAI, there was like Dota 2. This is amazing for me because I don't have any time to play any video games anymore since I have kids.

And if with this agent, I could just tell it to go play the game and then I I could go have fun and then describe the fun and email it to me and I'll have Nick read the email and summarize that to me in a text message. That would be my experience of the video game.

No, seriously, what I actually want this for is I hate how modern video games it takes 20 minutes to set up. Have you Have you ever experienced this? It's like going through the tutorial. Remember when I Yeah. When I made you play Halo? When I made you play Halo?

It It took like 10 minutes for you to actually get into the game and then the actual game took you 5 minutes to play. And so, uh, like there are a lot of times when I hear a new game, I'm like, I only have 20 minutes in my weekend to play this game. I want to jump straight into the action.

I don't want any of the opening unskippable cutscenes. I don't want any of the of the tutorial, learn how to jump, learn how to crouch. I already know how to move. I know what the stick does. Don't tell me that. Uh this is gonna solve that for me hopefully. Yeah.

But this is actually I think this is like one of the most interesting papers this year. Yes. Um a big part of it is um it's general so it you can put in the new game and then it does like selfplay essentially. Sure. Um which is like really important because that's like it's teaching itself.

This is basically the first like agentic model that can like see a new completely new environment. Yeah. And then do selfplay. Yeah. Uh where it gets better. So you um I wonder how wild I I'd love to know how how diverse the inputs are. Like does it expect an Xbox controller's worth of inputs?

Does it expect a uh a keyboard's worth of inputs because that's more inputs that it would need to learn? That's fascinating. And then I wonder what happens if you start marrying it to generative world models in the future as