深入了解 AppLovin 的 $100B 广告引擎

When we first started, it was, “I want to become a billion-dollar company.” When we got to a billion, it was, “Okay, let’s become a $10 billion company.” When we went public, we wanted to become a $100 billion company. We dropped 92% in the first 18 months of being public. That dream seemed very, very far away, but post-Axon 2, we recovered, and now we’re a $100 billion company.

Once we’re able to make predictions more accurate, advertisers see better returns and our business grows. I don’t want our engineers to sit next to AI. I want our engineers to sit on top of AI. Our team was elevated by AI, so we were able to manage the same size of a team while we’re solving more challenging problems. I would attribute the success of Axon largely to what we decided not to do, not actually to what we did.

How are you going to become a trillion-dollar company? Adam and Geo, welcome to Sourcery.

Thanks for having us.

Thank you for having us.

Well, thank you for having me at AppLovin’s HQ. We’re out here in Palo Alto. This is Palo Alto, right?

Yep.

Okay, cool. So, Geo, you are the CTO. Adam, you’re the CEO. But I wanted to do something a little bit different. I listened to a bunch of the podcasts you had before, and it’s very apparent you guys have an amazing underdog story.

There was one that I caught, and it was towards the end of the interview when you started to reveal the guts of the machine, the engine that started to propel AppLovin to become one of the most efficient, well-performing companies out there. You guys have great margins. You’re very profitable, all this awesome stuff. But it started when someone started to ask you hard questions. And, Geo, you were that person.

So you joined the company in November 2022. What was it like when you first joined AppLovin? What were your first impressions?

1. Geo Finds Room To Grow

My first feeling when I joined AppLovin was definitely a mixture of excitement and a sense of responsibility. In the beginning, it was like a breeze of fresh air. I came from big tech companies. The process there was much heavier, so I joined AppLovin and saw that the team here was very lean.

Both Adam and the CTO, Basil, at that time, were deeply involved in the day-to-day decisions and execution. It was easy to work directly with them. They were super approachable. That feeling was refreshing.

But on the other side, soon I realized the company was small. A lot of things—the models, the infrastructure, people’s mindsets, and the way they handled iterations—had a lot of room to grow.

More importantly, I felt the biggest problem at that time was that the junior people in this company were not elevated. They were not given the opportunity to understand the business, so they were given a lot of low-level tasks without the ability to have their own ideas and raise questions.

But the good thing is that at that time, I didn’t take this negatively. I wasn’t frustrated or disappointed. I saw it as an opportunity, and that’s why I joined AppLovin. I was glad to see that the whole team at that time had a very open mind. When I brought a lot of feedback and new ideas to the team, people were very willing to work together. So we took that as an opportunity and started to make changes and improve.

2. Hiring The Axon Two Architect

Going back to Geo’s hiring and the interview process, the old CTO and I both interviewed him. We were, at that point, somewhere around a $5.5 billion company when we started talking to Geo. We had a model that powered our advertising business, the Axon 1 model, and it was written using outdated machine-learning techniques.

We were looking for a researcher to come in and help us architect the Axon 2 model, which we released, I don’t know, probably 5 or 6 months after Geo joined, and he was the architect behind it. We liked Geo a lot in the interview, and both of us gave him the hard pitch on joining.

He was excited to join, and I went to do my November 2022 earnings call. He was supposed to join, I think, a couple of days later. The stock went down 30% that day, and I’m like, “Damn, we might not get Geo in.”

I don’t know why he joined us, but he’s one of the weird people who says, “It went down 30%, so I’m going to get more equity. Now I have more upside in this thing. I should join. I should even join quicker.” We were lucky that he ended up joining us at that moment.

It was smart. It’s arbitrage. You know you’re going to make it go up.

I wasn’t thinking that way, but I was in big tech. If I just wanted to stay on the winning team, I would not have chosen to leave. I wanted to find a place where I saw opportunities and could actually make a difference.

So what were some of the big questions that you started to ask?

I guess I didn’t start asking questions. I knew what my position was when I joined, right? I saw all those gaps, and there wasn’t a great model at the time. But before training the model, you need to have the right infrastructure for it.

My first month here, I just started writing code. It was very easy. I was working directly with Basil. We did not have a lot of meetings, which, by the way, in the beginning, I wasn’t very used to. I asked Basil to set up weekly one-on-one meetings with me, and he didn’t understand what that was because it’s not a thing here at AppLovin.

But soon we canceled that meeting because we communicated with each other mostly through code, and that was much more efficient.

3. The Lean Company Culture

Taking it back to when you start a business, teams are always small, and you try to build something that has product-market fit and scale up a company. When we started, the company was probably 10 people at the beginning, and then maybe for the first 3 or 4 years, a maximum of 50 people. So we always ran really lean.

In an advertising company, if you can’t make money, then you probably shouldn’t be in business. It doesn’t make any sense, because you’re trying to deliver value to advertisers. You get paid for the value you deliver, and, in theory, that should generate you something more than what your costs are.

We were profitable really, really early. We had difficulty raising, too, so we had to be profitable really early. But the leanness of the culture came from the beginning.

The types of people that we hired back then—Basil, the CTO, was the first lead engineer and then was promoted to CTO 4 or 5 years later—but his mindset was, “I’m just going to do it.” He would just write code to solve every single problem.

Up until Gio joined, Basil probably wrote, I don’t know, 60% of the code of the company himself. The mentality was always that individual contributors can help us succeed in this very, very competitive field. So we never built this framework of one-on-ones, meetings, and process.

I think that was what appealed to Gio when he came in. One of the challenges with that style, though, is that you don’t have a team constructed to actually go through challenges and solve problems as a group. He can talk about how, over the last 3 years of his tenure, he’s really developed that muscle.

But my own view, just being an individual contributor myself, was always: hire a bunch of great people and let them solve problems. If they need help, they can come to us, and we’ll give them help, but otherwise, it’s sort of just off to the races. That was what Gio stepped into.

How big was the engineering team when you joined?

It was probably the same size as now.

100 people.

100 people, right? Yeah.

Yeah.

Over the last few years.

How big was the organization?

The company—back then, we had gaming businesses, too, that we’ve since sold off. Not including those, and not including Adjust, which is a software-as-a-service company we bought and never integrated, I’d say the company has been around 400 people for a while.

About 2 or 3 years ago, we probably cut it down from 600 to 400, and we haven’t really changed all that much since. People come and go, but that’s what you’ve got.

The other interesting thing for Gio when he joined was that he came in, and we created the Slack room when it was Basil, Gio, and myself. We talked through the problems, and it was just real-time feedback.

I don’t know if he and Basil even saw the sun for those 5 or 6 months, because they were just coding and racing against each other. But the dynamic was, “Okay, well, this is a new path.”

We were effectively rebuilding the engine of the company. We didn’t really care about what existed. Everything else that had been built—the data platform, the software tools for developers called Mediation Service, all these things that we had as assets—were separate, but this core engine, this algorithm, they were completely rebuilding.

They basically went in to—I don’t know where they worked out of. I know where Basil worked out of. I don’t know where Gio worked out of. But they just went in and started coding against each other and racing to get to a direction that they thought was the right place to go.

It was cool for me to see. I think it was different for Gio to see because he was used to much bigger processes around what engineers could and couldn’t do at companies.

One of the first things I told Basil when we first started developing this company was, “Technology moves really fast, and you’ve got to be humble about what you have and know that if technology goes faster than what you’ve developed, you’ve got to throw away what you have and rearchitect it, rebuild it, and it’s probably going to happen every couple of years or even faster in the future.” That was always one of the core things that we believed in engineering: Who cares about what we built? Let’s always make sure we’re working on something that’s current, innovative, or cutting-edge.

So you have a lean team. It’s about 100 people on the engineering side. Everybody’s, quote-unquote, an individual contributor.

Yeah.

What was the process like for architecting Axon 2.0?

4. The Axon Two Build

When Axon 2.0 was being architected, the team at that time was very different from today, right? So I’m going to explain what happened at that time. When I do interviews, a lot of the time people are very curious about what I did to change AppLovin. But very rarely do people ask how AppLovin changed me.

Aw.

Actually, when I joined AppLovin, that was a change for me as well, because the moment I saw how Adam and Basil worked, it gave me a new definition of what it means to be hands-on. I used to be hands-on. I loved writing code, but joining AppLovin showed me a totally different level.

Initially, when I joined AppLovin, I actually planned to join 2 months later than my original date. Basil and Adam were trying to convince me to join first, and I said, “I already have a planned trip to Europe for a month for Christmas. I wouldn’t have much time to work during that month anyway. How about I join after that?” But Basil insisted. He said, “You can just join for a week. Do a week of work, and then go on your vacation.” I said, “Okay, fine.”

But that week fundamentally changed me, because after that week I decided not to cancel my vacation, but to work through my vacation. I still remember what happened at that time. It was the moment that ChatGPT-3.0 had just come out.

Mm-hmm.

But it was definitely not enough to write code for you at that time, so we still had to write code in the traditional way. I was on my flight to Italy, and it was a 12-hour flight, and I had to continue coding. At that time, people still knew how to code without an agent. I don’t think you can still do that without the internet today.

I was on the flight, and there was no internet. I was coding the Axon 2.08. At that time, I was building the training infrastructure, and at a certain point I needed to check the documentation. I needed to Google some information about the tool I was using, but I was on the flight and had no access to the internet. I had to force myself. Luckily, in my IDE I had access to the source code of all the libraries I was using, so I had to go deep into the source code to try to understand how things worked.

That was a pretty interesting experience. At that time, I was asking myself, “A man is given an infinite amount of time,” because I was literally trapped on the flight. The time was infinite, but I had no access to external help. I just kept coding, and when I was in Italy during that vacation, I kept coding.

My in-law is a lovely Italian lady, and she couldn’t understand why anyone had to work that hard during their vacation. She kept asking me, “Hey, Giovanni, are you okay? Giovanni, are you okay? Are you losing your job?” I said, “No, no, no. I’m just very passionate about what I’m doing.”

Surprisingly, initially I thought I would actually start working at AppLovin in 2023, but by the time I came back from that vacation, we were already ready with the infrastructure part of Axon 2.0. The moment I came back, we started training the first generation of the model.

Oh, wow. So that was a big reset for you?

Yeah, it was a reset. I feel like that kind of philosophy was already in me, and I definitely loved that almost feeling. But prior to joining AppLovin, I didn’t have a chance to experience that kind of working mode, and that month made me feel, “This is the place. I’d love to work here.”

So big tech is very different.

It was very different, yeah.

Damn. And then, in terms of rearchitecting the actual algorithm and building out this algorithm, because it was a huge, major inflection point for AppLovin, what was that like for you?

5. How Axon Two Works

Right. So let me go a little bit deeper about what Axon 2.0 actually is. The problem that a recommendation system is trying to solve is that you have many, many users and many, many items, and the combination of each user and each item is enormous. The recommendation system has to study the relationship of each combination.

Before, the last-generation model tried to model this by clustering those combinations into different clusters and building this tree-based model that was basically a huge set, with hundreds of thousands of if-else branches. Each cluster of user-item combinations sat at the end of one of these if-else branches.

But obviously, as you can see, the real world is so complicated, and the relationship changes dynamically as time changes, as the weather changes, as there are promotion events, holidays—all of these things are changing. These tree-based models are not able to accurately model the dynamic relationship between users and items.

The Axon 2.0 approach uses a concept that nowadays people are very familiar with, thanks to language models. It’s called semantic embedding. Basically, we use learnable embedding tables to encode those IDs—item IDs and user IDs—that otherwise would have no meaning. We encode these IDs into semantically meaningful embeddings, and they carry statistical information. Then we pass that information through a deep neural network, which allows us to study the complicated interactions between users and items.

The result of this kind of model is that it’s much more powerful at recognizing patterns, and it’s also powerful at extrapolating these patterns to unseen data and unseen user-item pairs. Once we built this, our model was able to make more accurate predictions. It was also cheaper in terms of infrastructure. Once we were able to make more accurate predictions, advertisers saw better returns, and our business grew.

Why was it cheaper?

This just has to do with some details of how the model is being inferenced.

Mm.

The neural network is made up of a lot of standard GEMM operations—basically, general matrix multiplication—which is what GPUs are famous for. GPUs are highly optimized for this kind of operation.

But the selection tree, the old-generation selection trees, has a structure that is very hard to optimize on a GPU machine. So the new-generation model is very, very powerful and efficient with modern GPU architectures.

So you’re not token maxing?

We’re not? Oh, token masking is a concept in large language models. It’s a different term.

I token max every time I hear Gio talk. I have to put in an LLM to translate what Gio said for me.

So Adam, can you translate what he just said?

I’m not smart enough, unfortunately.

But I’m really curious. As you’re talking about the deeper infrastructure and the deep learning that occurs within the model and the tree and that kind of thing, what are the main categories of data that you’re trying to build patterns off of?

Basically, the most important type of data is the user’s interaction with the ads, because every ad we show to a user is an important feedback data point for us. The user either reacts to it or doesn’t react to it. If you collect that information, it builds a very powerful prediction machine for the user’s next action.

It’s also important to understand our ads in this context. We’re not just a static native or banner ad. The ads inside games are much more similar to what television ads are, but specifically for games. You have somewhere around 60 seconds that people sit on an ad, and they’re not just sitting there doing nothing; they’re playing mini-games that we serve them.

So they’re engaging in something that is representative of what they can go download, and there’s tons of interaction data that comes from that. If you think about the average time spent on most websites that are really successful, a user spends, call it, 40 minutes a day on a really successful social property. If a user sees 10 ads on our platform a day, that’s 10 minutes of engagement with a website that’s effectively like a mini-game-serving website. Tons of data is able to be extracted from those engagements.

What other data do you have behind those ads? Because you serve them to effectively, like, a billion people.

Yeah.

I see this on your website. They sit there for 60 seconds, and they get served this ad, so what is the typical click-through...

How do you determine that?

Well, you have the engagement, so you have how they’re playing and what they’re engaging with. Then you follow them through a conversion funnel, right? We serve the ad, the user engages with the ad, and the user goes to install or doesn’t install. Sometimes people will just say, “I don’t want to engage with this ad at all. I’m not going to do anything on my device other than skip the ad.” So you know that person really doesn’t like that ad.

On the other extreme, someone might follow the ad all the way through to the app store and download an app, so now you know, okay, this is pretty warm. They’re interested in it. But just a download for us means nothing. We’re trying to deliver engagement and revenue to the advertiser. So to get really hot, the user then has to engage very regularly and generate revenue, whether ad revenue or in-app purchasing revenue for the customer.

The benefit of doing what we do is that we can serve the beginning of the conversion funnel and then close the loop all the way at the end. The advertiser gives us data to help close that loop, and then Gio’s system can take all of this data, which across our business is a ton, and put it all into this model to determine, okay, now I’ve got one thing that this person’s really hot on. What other users are like this person? What should I start predicting for that person?

What I thought was interesting about learning more about AppLovin is that the business model is predicated on performance, but it’s advertising to make people more money. So can you explain—was it always like that? Were you always making advertising so someone could build their business and create this reinforcement system to make more money, put more money in and get more money out, build their business, that kind of thing?

Yeah, totally. We want an advertiser to be able to spend money on our platform, measure it, and know that they end up getting more revenue and profit from the money that they spent. If you can give them a user and they make a profit on it, then they’ll scale to pretty large numbers. You don’t have to sell that through.

The reason that was originally our goal as a company is that I realized I don’t have the ability to sell all that well, and I don’t have the patience to really build up a traditionally large sales force. This is what most advertising companies do: They just scale out the go-to-market team when they have anything that works. What we wanted was a system where the company on the other side just knows it works, so they want to put money in, and we’re delivering so much value that they’re really desperate to work with us, not the other way around.

We’re not really selling anything other than the fact that they’re getting the user, they’re getting profit, and they’re happy. What makes it cool to do that is that the platform ends up becoming much more catered to businesses that you almost have never heard of before. We get companies that scale on our platform that are small businesses to most people but have very large P&Ls, and they’re able to do that because of a solution like ours. That always made us feel really good. It also fit the DNA of the company well.

When did you feel comfortable enough to switch not just from gaming ads, but also extend into e-commerce?

Yeah.

6. Expanding Beyond Gaming

It was about a year after we scaled Axon 2.0. I think, at that time, we realized the model architecture we had would work beyond just gaming, right? Of course, there were new things we had to build: the new business, the new pipeline, and the new product flows. But we found that the core technology should be general enough to support different verticals.

Yeah, I think when Gio built and architected Axon 2.0, he knew that this was going to function for any vertical. But we always got pushback from anyone who was covering the business—whether an analyst, an investor, or someone scrutinizing the business—saying, “Well, you only have gaming data, so how could it work?”

One thing that people lose sight of is that it’s a billion people on the other side, not just a billion people who are only playing games and doing nothing else. So if you understand the person, then you can actually deliver value to them outside of just a game-to-game scenario. You can only do that with a very sophisticated model.

The other piece of that is that you’re getting advertisers sharing data with you. This is what powers a lot of what Facebook was able to build over the years. Originally, when they went public, there was this notion that you couldn’t monetize a social network. How could you monetize this blank canvas? Well, they have a lot of people, and you start understanding the behaviors of those people through what people are doing on advertiser properties. You pixel websites, you get data.

And so we took the same approach, started pixeling websites, started getting data, and understood the user more than just their game behavior. The user does a lot more, and the people who are playing mobile casual games turn out to be a lot of heads of households. It skews slightly female, but it is this sort of 30- to 50-year-old age group that’s the heavy mobile casual gamer.

We had this really good user. They’re engaged with a really immersive advertisement, and we started collecting the data that the model could interpret to make these recommendations, and it just started working. But I think Gio had a sense that this was very doable because he understood that the architecture isn’t just looking at one vertical. It’s not a verticalized model.

Yeah. I think another important thing is having a great engineering team. The value of having a great engineering team is that a lot of times you don’t have to guess these kinds of things. At the time, we did analysis and discussions. There are reasons to support why this is going to work. There are reasons that tell us why this might not work.

So our solution was that we quickly built the prototype, and it only took us a few months to launch the first model with some testing advertisers. The result was overwhelmingly good, and that gave us the confidence that this was going to work, and we started building.

What’s your process for testing?

For testing the e-commerce?

Yeah.

So, in the beginning, it was building, not really testing, right? The building story was also fascinating. I think, at that time, the CTO, Basil, and some other senior leaders in the company were attending a Google conference in Vegas.

That morning, we were having breakfast and discussing how we would build the first data flow engine for this new product. We were waiting for our menu to come, and there was no paper, so we found a napkin and started discussing and drawing the design on the napkin. You probably did... Do you know this story?

No.

Did I never tell you this story?

No.

Okay, you should talk more and hang out more with Basil. Then we were having this discussion, and by the time the food arrived, we already had a pretty good design. Then we all went back to our hotel room.

Before we arrived in San Francisco, we actually had the first prototype for that data engine. Then it took us a couple of months to build up the data, the model, and the infrastructure, and we recruited a few advertisers. I think it was the business team who found the advertisers for us, and we put them online.

Initially, the model had very limited data, so it was very premature. But even with a very premature model, we were able to see people making a purchase. The ROI, the ROAS was decent, and we were like, “Okay, this is going to work.”

What’s cool about this story, which I’d never heard before, is that one of the cultural philosophies we had was almost no product management and engineers who were competent enough about business problems to write their own architecture and product to solve business problems that maybe even a business team didn’t know existed.

Here you’re not hearing about a product org that’s talking to clients, writing a spec, handing engineers a spec, and having them write the code to spec. None of that happened. You have engineers who didn’t even talk to a business person sitting down together at breakfast, on a napkin, writing out a solution to a problem that we never even discussed—and then creating something that could become a big business.

I think the team still is constructed that way. I don’t know how many product managers are in Gio’s org, but it’s probably fewer than a handful, and the rest of it is engineers who are competent enough to know what they’re going to develop.

So are you then working directly with the advertisers on tweaking it or making sure that it’s working, rather than going to another team to ask how they’re handling their relationship with that person to know whether it’s going well or not?

The communication with advertisers is still through the business team, but our engineers understand what a business needs. Right? Our engineers, when we have to talk to advertisers, sometimes you have to schedule a call or a meeting. I don’t think that’s a great way to scale our engineering team. We leverage the business team, but engineers understand what is important for the product.

If you’re delivering something like an audience to advertisers, the result is sort of hand-wavy. There’s no measurement of what actually occurred, and to give engineers a path, you’ve got to give them pretty clear metrics on whether it’s working or not. If you’re living in a revenue-based world or a return-on-ad-spend-based world, there’s not much of a debate with the advertisers. Either they’re going to put more money in because the return on ad spend is good, or the return on ad spend isn’t good.

If someone is selling lipstick and the lipstick costs $10, and it costs $5 for them, they’re going to have a $5 spread. They can’t spend more than $5 on selling the lipstick; otherwise, they’re in the red. It’s really easy to put that into a formula and say, “Engineering, here’s where we’re at. We’re either good or we’re bad, and we need you to be here for this to be scaled.”

Did you always have that intuition, or was that also built up?

I think that’s definitely built up. Before joining AppLovin, in my previous job, I was already carrying both the engineering part and the product part of the work. As I shared in the beginning of the podcast, when I joined AppLovin, I saw that the engineers were kind of disconnected from the business. So one part of my job was to try to help our engineers understand the full context.

Instead of assigning tasks to them and telling them exactly what they have to do, I show them how their work is impacting the business. Because our engineers are very talented, once they’re provided with that kind of context, they have a much better sense of what to build.

What do you think the biggest mistakes are that people make when they hire new talent and try to onboard them into these teams?

The biggest mistake?

Yeah.

I wouldn’t be able to comment in general, but I think on our team, a few mistakes could be made. First of all, not setting the right context. I feel people usually assume that junior people need hand-holding, and therefore they only provide very granular tasks for them, but I don’t think this is a great way of helping them grow. So I think that’s one type of mistake.

The other mistake—this is something that I recently learned—is exactly the opposite. Sometimes, when a junior team member does a good job, we tend to give them too much encouragement without continuing to challenge them. Sometimes their ego will grow, they’ll lose their humility, and they’ll stop questioning themselves. That also becomes a problem for their early growth.

Mm.

Well, let me cover the prior question too, and then we’ll get to that one in a second.

Yeah.

I think the big mistake I’ve seen is people really hand-holding new talent and bringing them along slowly. One thing that’s a little unique is that whether someone is an intern or a full-time employee who’s just started, maybe even right out of school, they can push code within the first week of being here.

At the end of the day, we’re an advertising company. If we went to a really high-IQ graduate out of one of the top universities and said, “You’re going to build ads,” it’s probably not easy to get them excited about this opportunity versus all the other opportunities they have in front of them, especially in the field of AI. Here they come in at the intersection of math and engineering, and they’re building models or data and plumbing to support the infrastructure that we need to run the models that we have.

They come in and just get to deploy code, so they get hands-on experience. There’s a lot of controls in place to make sure their code doesn’t blow up the whole system, but there’s no process in front of them where it’s, “Go through all of this training, review this, this, and this, and talk to these 15 people who will bring you along over 6 months, and then now you’ve got to be a little bit hands-on. A year and a half later, okay, now you can push your first test.”

Here, within a week, they can push a test. They start getting results. That moment is a pretty substantial dopamine hit. They get something that they developed or thought to develop quickly into production, reaching a ton of users, and they get a real-world feedback loop around it.

This isn’t just engineering; it’s on the business side, too. We don’t have the traditional training processes to bring people along. There’s no training manual. There’s no “come in and someone’s going to tell you what to do.” We want people to come in and be very curious and learn for themselves.

At least my own experience in learning is that when I was told what to learn, I never retained anything, but when you’re hands-on, you can retain. And now, in the world of AI, it’s even more powerful. If someone can come in and say, “Look, there’s no training manual, but I’m going to build my own training manual,” and prompt AI and start asking the questions they need, and if they ever hit a roadblock in learning, they go to someone on the team, not only are they able to use their own curiosity to educate themselves and bring themselves along faster than others, they’re able to say, “Look, I’m willing to raise my hand and speak up in an organization I’m new in.” By enabling that, we’ve found we’re able to bring people along much faster than we otherwise could.

Sorry for interrupting there.

No. I mean, that was a great answer. You’ve been really hardcore about making sure that your team is AI-active, AI-enabled, and AI-first, whatever you want to call it. How do you evaluate people on that? How do you make sure that they’re continuously upgrading themselves and upgrading their skills?

7. Engineers Sit On Top Of AI

I think AI works differently in the business team compared to the engineering team. I can speak about the engineering team. AI has been evolving very rapidly in the last few years, but there’s still a boundary between what AI is good at and what AI is not yet good at.

For example, even today, AI is extremely good at answering questions, writing code, and probably also doing research and helping you shape your opinions. But AI is still not good at long-horizon planning. A lot of work that we have to do today requires long-horizon planning, so we still need a human to be involved. Maybe one day AI will be better at that.

The way I see the relationship between AI and our employees, our engineers, is that I don’t want our engineers to sit next to AI. I want our engineers to sit on top of AI. As AI improves, the boundary between AI and humans is also moving, right? Previously, AI was only useful for correcting syntax errors while coding, and then AI became better. AI can complete an entire phrase, an entire paragraph of code.

Nowadays, you can just give AI an idea, and AI can write an entire PR. As AI improves, humans just sit on top of it and focus on what AI is not good at. Because of this, I think almost every engineer’s output and productivity have been dramatically multiplied in the last few years. As AI is advancing, I believe the output efficiency of our engineers will just keep growing.

I asked Max Levchin this question because he’s a hands-on CTO and CEO type. He’s super technical. I asked him, “Do you still think people need to learn how to code?”

Mm.

I think at this moment, yes.

Why?

Yes. But I’m not sure that my answer will stay the same in the future, right?

Okay.

At this moment, I think AI still very often makes bad coding decisions. If our engineers aren’t capable of coding themselves, they won’t be able to judge and correct those mistakes.

Then you'd blow up.

Yeah. That's probably what a lot of companies see, right? We don't have that problem at this moment. Our code is still pretty clean. Internally, we also have this concept: what is the system you want to prototype—

Mm.

—and what is the core system that you have to keep absolutely clean, right? When people just want to set up a dashboard or create a temporary project for a presentation, they do whatever they want. It doesn't have to be robust or maintainable. But for all the core infrastructure, the core system, this is what I told the team: You can use AI however you want, but at the end of the day, you as a human are held accountable for every decision that your AI made for you.

So it is absolutely not acceptable when I come to an engineer and say, “Hey, why did you implement this system this way?” If the engineer tells me, “I don't know. AI did this,” I tell them that this is not acceptable.

I think the other thing is that AI allows everyone to write code, obviously, and commoditizes a lot of that. As a company, you need to create alpha from somewhere, and your smart engineers can't just go to AI and say, “Give it to me out of the box.” If that were the case, then everyone would have technology that looks identical to each other.

We say this in ad creative, too: If AI starts writing all the ad creative, then all of it is eventually going to look the same, and the user won't respond to the ad. So how do you create alpha? You have to have really smart people who understand systems and understand what they're asking the AI to output for them. They need to create variance on that—not just use a simple prompt to get an output that anyone could deliver, but create something that understands the product, runs its own product management, and directs the AI in a specific way to create that alpha.

I think that's going to be a skill that's required if companies want to stay ahead of each other. Otherwise, everyone's going to land in the same place on a lot of these functions.

Can I ask a glaringly obvious question? I don't know if you've maybe been asked this many times, or maybe you guys do this internally, but I was listening to a podcast about this. It was an advertising agency trying to educate people on how to make the best AppLovin ads. What makes the best ad? They were talking about making 20 to 50 videos.

Yeah.

100 videos. I had no idea you had to make this many videos.

8. Making Ads That Convert

Because our ads have a lot of time with the user, and half the ads the user sees are between levels, think of it as a commercial break they can skip quickly. But half the ads they opt into watching because they want to watch them to get some sort of reward in the game.

Let's say you played a puzzle game and lost a life. By watching the ad, you can get another life, or you could just pay a dollar. But here it gives you something of monetary value, so you're going to watch the full 60 seconds. You've got a lot of time with the user.

Then the second part is that usually the first ad isn't going to convert the user. They're going to see multiple ads in a conversion funnel to get to the point of transaction. If it's 10 ads to convert, you had 10 minutes with the user. If you're giving them 10 minutes of content, it better be interesting; otherwise, you're actually going to turn them off.

There's no really good formula for this. I've tested tons of ads in my life. I've been in advertising for 21 years now, and I don't know that there's a formula for creating a great ad because none of us can really predict what a consumer is going to respond to all that well.

It requires, one, a lot of shots on goal. People have to test a lot of things. If you create 30 ads a week, probably 1 of those might be interesting, and you don't want to just change very small things around. You want to create concepts that are differentiated to find what a user is actually going to respond well to when it comes to that specific brand.

What works for that specific brand won't translate to 10 other brands as well. Each brand has to figure out what a user is going to respond to in our framework, where they have all of this time, and then they have to start creating multiples of that because they get so many shots with the user to engage them before the final conversion happens.

There's not a great answer to this. People want to come up with best practices: “Here's a list of 10 things to go do.” Simplify it. But at the end of the day, this is one of those variables that can create alpha in an advertising system. Advertisers that invest in creative and are good at it, whether they're using AI tools to develop it or using people to develop it, can create lifts in an advertising system where everything else is automated.

Has there been any sort of resolve on whether UGC is taking over traditional ads? I know with the proliferation of clipping, it's really cool seeing it on social media. I put out content every day, and I put out content before where there was this prebiotic-probiotic bar. I bought it, ate it, and said, “Oh my gosh, this is like Ozempic.” I tweeted that. They made that tweet an ad, and it became a 4X ROI ad for them. They made so much money on it.

Yeah.

I got no money. But there are different kinds of things that have made ads more effective than others in this proliferation of UGC, and what looks like clipping and what looks like just a normal social feed has kind of risen. Have you seen anything with that?

Some advertisers can make that work for us because there's a wide range of things that work, but I'd say social is really high ADHD, so everyone wants to get to the point really quickly.

Mm-hmm.

Typically, a social ad will say, “You've got 3 seconds to capture attention. That's it. Otherwise, you've lost the user.” In our world, where the user is sitting and engaging with the ad for such a long, extended timeframe, the user is just frankly different. The person who's playing Solitaire or Mahjong is not the power user scrolling on Instagram and TikTok all day long. You've got a different person. Probably the power user is older, and their mind is not working in the same way as someone who's grown up in today's culture.

So you have smarter people on your—

I wouldn't call them smarter—

Oh my God.

—different, differently constructed minds. They're just wired differently.

This is part of the education process. In gaming, we've been in the space for 12 years. The game customers brought playables to market—the mini-game preview in the ad—about 10 years ago. The ad has stayed pretty consistent since then because it's just a really good ad. You get the mini-game, and then you get the download behind it.

We got into this consumer vertical, and more specifically e-commerce, roughly 18 months ago. When you're so early in a market, you have to educate the market on what works. Where alpha comes into play here is with some of the customers who say, “I'm going to invest in this platform early on and really learn how the users behave, who the users are, what they're willing to buy, and what kind of ads to build.” They create alpha.

They get to create much larger campaigns with successful metrics on us than those who come in and just go, “Look, I'm going to run the same UGC ad I do on social over here.” We don't have 3 seconds. They had a lot longer to actually engage the user and have that user remember their brand, but they chose not to take that.

It'll take time for us to make sure agencies and advertisers understand these concepts, but it does create a world where, early on in a platform, advertisers who are paying attention get alpha.

How has AI changed the making of ads?

It's early. It's not like you can get a good clip, a short ad, out of some of the large language models. The challenge is to get them to put together a 30- to 60-second video advertisement for a brand without messing anything up, because anything wrong in that video, the brand isn't going to approve it to run.

It can't mess anything up, it has to be engaging, and it has to work for a pretty long, extended timeframe. That's not easy yet. So it's not at a place where you're just going to get a massive amount of advertisements flowing in that are diverse and creatively inspired and going to create lifts in campaigns.

But tools are available now so that humans can create ads much faster and at a much lower cost. We've seen much more ad content coming into the system. It's just not at a point where the average marketer can type into a box and say, “Here's a great ad,” and off we go.

Yeah. From our product, we saw a lot of AI-assisted content on the platform from advertisers. I think it's important to still have the human in the loop to guardrail the process, because advertising is about trust. If you create content that breaks the brand image, it will break the trust, and it's harder to convince users to buy.

There are 2 different types in terms of AI-generated content. There are 2 different approaches you can try.

One is making this tool entirely automatic, right? If AppLovin is building such a product, this is what we aim for, because we're not serving one single advertiser; we're serving all the advertisers. We have to create a tool that can fully automate everything. The other choice is to create a tool that allows advertisers to interact with and collaborate with the AI tool to create their own content. Right now, we see that the second type of tool has been pretty widely adopted, but I think there's still more challenges for the first approach, and we're still working on that.

Have you guys thought about buying any of these kinds of companies? There's a company called Higgsfield. How do you scope out R&D on this kind of new market that's opening up?

It is much easier to get content that is short, so they're optimized for social and UGC. These things, if you want to create a 10-second clip and put it on TikTok or Instagram, you can do out of the box, and Higgsfield is a tool for influencers. That's not the advertiser on the other side of our product.

If you get a brand as big as Wayfair and you hand them something out of the box that looks not ideal for their brand, they lose trust in you as a platform, and you certainly can't run that on their behalf. They have to get something that is compelling enough for their brand to be as good as what a human builds, maintain brand trust, and have extended time in it. There's nothing in the market that we've seen that solves that problem. Generally, when we see these problems that are hard to solve and specific more to us than just the general market, our team has to develop, too.

Okay. So we covered a lot internally. To take it a little bit back into macro, a lot of people are trying to rebuild right now for the AI era. I've listened to many of your interviews, Adam, where you're talking about how you build for the next 3, 4, 5 years ahead of you. Whatever you built right now, you built back in 2022 or 2023, so I'm curious: Does that still hold? How do you rebuild in such an era where things move so fast?

9. Rebuilding For The AI Era

I don't really want to rebuild. In the last 3 years, we built a lot of things that made us feel so proud of. AI is powerful, but I don't think that's a reason for us to rebuild. What AI really helps us do is build on top of what we have today. With the empowerment of AI, we were able to achieve much more.

I've been thinking about this growth trajectory, this growth journey of our company. As we showed previously, the size of the engineering team was basically the same size as it is today as it was 3 years ago. Typically, when a company goes through such growth in its business, the scale of its business is much larger, and the type of problems it's solving is also much more sophisticated. This kind of growth has to be paired with the growth of headcount and team size.

Luckily, I feel our growth was perfectly in sync with the advancement of AI. As the problems we are solving became more complicated and our team was elevated by AI, we were able to manage the same size of team while solving more challenging problems. As we're speaking now, we built this new-generation model 3 years ago. Now, with the assistance of AI, the team was able to tackle and work on a new-generation model that would be even more powerful. So, back to your question, I don't think we need to rebuild anything, but we are definitely building on top of what we have today.

On the business side, we keep the team really lean as well, and you wouldn't be able to scale out the products we want to scale out with the team sizes that we targeted in the past. But if we can get to a place where eventually everyone on the business team is doing something and then asks the question, “Could this be automated and done across the whole advertiser base?”—and they actually understood how to use AI, or they go to the central folks and a team that can create tools, dashboards, and automation for them, who can take that process and automate it—we'll get to a place a couple of years from now where hopefully most every process is automated.

The business interpersonal skills matter a lot. You're not going to be able to replace sitting down with the customer at a QBR and understanding their business. Because our AI, as an enterprise, can understand our business, it can't understand the businesses of advertisers. We always say our job is to be their growth consultant. It's to understand what their needs are and then deliver value that meets their needs on our platform.

There's a lot of interpersonal relationship skills there that AI can't replace. What we want to get to is a place where the team is scalable, most of the processes are automated, and those people are really sitting down with the customer, making a difference, and building relationships.

So I guess most people are trying to rebuild right now because they're acting a little bit more reactively versus proactively, which is what I was trying to get at. In terms of others that are trying to rebuild right now, what do you think the biggest mistakes they'll make are?

This is a very interesting question, right? I think everyone here would agree that AI is helping us improve the productivity of every engineer and every business person tremendously. But it's also interesting to see that not every company or organization has been able to multiply its output as a result of the empowerment of AI.

I think a big mistake people are making is that, because of AI, building things becomes easier. People are making a lot of bad decisions and wasteful, bad-taste ideas that actually offset the empowerment of AI. You build things faster, but you're making a lot of bad-taste ideas, and in the end, you might not see the end result. This is also why I don't think we just have to rebuild everything simply because of AI. We need to understand what problem we need to solve. AI really helps us solve the problem, but it doesn't change the problem we have to solve.

I think the fact is probably 99% of people who use AI today just use it as a search engine. You're left with, “Okay, as an organization or running an organization, that's really not what you want,” because it could do so much more. But how do you get your organization in a position where your people actually know how to use these tools better?

You've got AI-native companies that were started in this era. That's easy. They're hiring for it, and they're structured as such. We were able to stay lean all the way through and maintain a high bar on talent, so most of our people should technically be AI-native and figure it out. If they're not, they're probably not going to survive here because they'll fall behind.

Then you've got other companies that aren't really using it appropriately today because their people aren't AI-native, and they're not built as an AI-native company. They've got people who, in some ways, may even push back on the usage because they want to justify their jobs. So how do you turn those organizations into AI-native? It's really, really hard.

You can't just cut people out and say, “Okay, the rest are going to figure it out,” because how do you know who to cut and who not to cut? You can't just go, “Here's a budget. Go token-max and figure it out,” because people start building a bunch of slop. Then you have business people building dashboards and engineers building dashboards, and everyone's paying Anthropic and OpenAI for the same stuff. It's like, “What are we doing here?”

That doesn't get you there, and I think this is going to be a challenge for a lot of companies. I don't think there's a great answer to it because technically, you almost have to replace nearly everyone and rebuild the culture to be able to deal with the world as it is today. But that's not really possible for most companies, so there's no clear answer for how those companies can get there.

Bummer.

Yeah. I think a lot of people think the biggest differentiation between a great company and a not-so-great company is the ability to build.

Mm-hmm.

But I actually think it's different. It's the taste. It's the taste to know what to build and when not to build.

I'm curious, Gio, what are the tools that you use today, and how have they changed?

I'm a big open-source fan.

Okay.

For the IDE, I use VS Code.

Mm.

In terms of coding, I use Claude for coding.

Mm.

But for text tasks, I like to use ChatGPT these days. I used to be a Gemini person, but recently I switched to ChatGPT. ChatGPT often challenges me, often disagrees with me, but it provides a very, very good perspective. But these things change. A year ago, exactly, I felt exactly the same contrast, and I switched from GPT to Gemini. So who knows what's going to happen in the next few months?

Yeah. I was big on ChatGPT for tech stuff because, for all these episodes, I'll take the transcript. I'll be like, “Okay, give me a summary. Let's work with that. Let's make a newsletter out of it.” Then I switched to Claude, and I spent a whole day yelling at Claude last week. It was so exhausting, and I hope it doesn't retaliate.

It's okay. Yeah, Claude is very logical, but I still feel it's not perfect for—

It started using UK English verbiage and numbering systems. It was writing dates in—

Yeah.

It made no sense. It was really aggravating.

It actually invents its own slogans and code words.

Yeah. Okay, so I'm really curious. As it relates to the LLMs and chatbots, that is new ground.

Mm-hmm.

You guys primarily, if not 100%, advertise on gaming apps, right? So do you extend into LLMs and chatbots? How do you see the landscape shaping up?

Yeah, you mean in terms of advertising?

Yeah.

Yeah, I think it's still early. If you think about my statement a couple of minutes ago, if 99% of the use case is search, then ads in chatbots are probably going to look and feel a lot like search.

What is a search ad? There are 2 types of advertising. One is bottom-of-funnel, which typically is search, and a user goes to a search engine and says, “I want this type of pair of shoes,” and the result will give them those shoes and some other shoes. Sometimes users will click other things, but usually it'll just close the conversion funnel. The consumer knows what they want to buy, the search engine places them where they need to go, and the transaction is complete.

There's tons of value in that because you don't want that consumer to land on your competition, but the user already knows what they're going to get. Then you have top-of-funnel advertising, where the consumer doesn't actually know that they want to buy something. They see something and discover that they want to buy it. The best example, and the largest example, is Instagram.

You go to Instagram today, and a lot of people will say, “I do my window shopping and figure out what I want to buy on Instagram,” while the ads are top-of-funnel. That consumer didn't open Instagram knowing that they were going to buy that dress, but they saw a dress, and they went and transacted.

We're trying to show consumers things that they don't know they want to get, whether it's games that they have never seen before or something in shopping that they didn't understand they wanted to buy. We want to start the funnel and then close the loop. Our solution isn't the highly relevant bottom-of-funnel advertising.

Mm.

If that becomes a big space outside of the frontier labs and some of the main destinations in large language models, we can look, but there's a really good search advertising business out there that probably extends into it.

For us, what is compelling as we think about what's next in our business is: What are the other spaces that create this form of discovery where you don't actually know what the user wants? An example of that that we've been working on and still doing some research and development on to see if we can extend our platform to is connected TV—television.

The user doesn't know what they want. If you can start them down the flow and convert them, that's super high value. There's the open web and video placements inside the open web. If you're on a recipe site and you see a video, maybe the obvious ad would be something related to the recipe, but the less obvious ad might be something completely unrelated and might create more value.

That's another space that we look at over time, again, on this 3-, 5-, 10-year time horizon, where we can really extend our product offering. We certainly will track what's going on in the usage of these chatbots because there's just explosive usage there, but it's probably not the most well-suited space for our advertising model.

It was interesting. I was listening to this podcast you did with Patrick O'Shaughnessy back in 2022, and you were explaining how delicate but dynamic and competitive the gaming ecosystem is, where you are partners with Apple. You have a lot of gaming apps on the Apple App Store, and then you're also partners with Google on ads. Are there any other environments where you think that could be replicated?

Unclear. I mean, you're just in such a duopoly state in mobile, right? Anyone who has a business today to the consumer—and even if ads are B2B, it's B2B2C—you're going to sit probably on top of 1 of those 2 platforms, most likely both.

If you think about connected TV as an example, it's completely fragmented across operating systems. The highest-used operating system is not dominated by 1, and so you don't have that same framework. The open web, by definition, is open. So probably where we end up venturing out to is going to be more open.

If we ever got into advertising with some of the chatbots or large language models, you'd say, “Would one of these frontier labs become an operating system?” It doesn't look like they're going that way, so even there you're going to have fragmentation and a proliferation of a lot of different sites.

10. The Trillion Dollar Roadmap

Okay, I have a hard question. How are you going to become a trillion-dollar company?

I mean, as an advertising business, again, I said this earlier in the pod, but you've got to make money. Otherwise, you don't have a really good advertising platform. Fortunately, we make a lot of cash, and we like to think, just in traditional finance terms, you're worth your cash flow after SBC and what you expect that to be many years out into the future, and the terminal value of it.

If you think about where we are today, we probably, on an EBITDA basis, are over a $7 billion run rate, and we generate somewhere around 75% cash off the EBITDA dollar. That's sort of after SBC. If you took us 3 years ago, that number was much, much smaller. It was probably 1/20th of where we've gotten to in 3 years.

To take it from this scale and be worth a trillion dollars, we've got to believe that we can get to $30 billion-plus in cash flow a year. If we could get to $30 billion-plus in cash flow a year, it depends on how quickly you can do that in the future, but investors would probably give you a pretty good multiple on that cash flow to get you there.

When we think about the things that we're working on, everything that we work on has to have a very large economic opportunity. Typically, in investor terms, you think about total addressable market. We've done really well in the games category. We've built the number 1 place for mobile game developers to go market themselves.

But the mobile gaming category is not big enough in terms of user acquisition dollars to support the type of cash flow we need to be valued at a trillion dollars, so hence you go into the consumer side. You really widen the types of dollars that you can go get and the companies that you can service, but on its own, probably in the shorter-term time horizons—and again, when I talk about shorter-term, this is 5 years-plus—it can't get us there.

Then you have to start thinking about what else beyond that. One of the things that is very compelling as you look at the companies that have been B2B2C or B2C and are worth over a trillion dollars is that they started with a product, generated a ton of value off that product, and started extending to adjacent categories, and executed exceptionally well there to get there.

I think it's Gio and my job as partners in the business to ask, “What are those other things that we're also going to do on top of just executing on our core for the next 5 to 10 years?”

Yeah. On the engineering side, we also have to be prepared for that kind of growth. When we are building systems—for example, the data infrastructure—we just have to project that, in order to support a company at a trillion-dollar value, what scale do we have to factor into the design?

When we're designing, let's say, a product, we have to factor in whether this design will one day be able to support expansion into different verticals.

So you're not going to build a hyperscaler?

Probably not. We don't have much CapEx. One of the things I also learned early on is you have to be very confident about where you're going as an executive, and you also have to set goals.

As you keep adding zeros, it gets harder. When we first started, it was, “I want to become a billion-dollar company,” and I conveyed it to the team. We had a plan for it. When we got to a billion, it was, “Okay, let's become a $10 billion company.”

When we went public, we wanted to become a $100 billion company. We dropped 92% in the first 18 months of being public. That dream seemed very, very far away, but after Axon 2, we recovered, and now we're a $100 billion company.

Again, it gets harder when you add zeros, but on the other side of it, as you scale a business, you matter more. Your technology is more compelling. The data that you have access to is much larger in scale. You reach network effects. A lot of things get easier, which is why, even though it feels very, very daunting, if you have a plan around it, or people have confidence that you can direct it that way, you can eventually get there.

That's awesome. To go back a little bit, your stock dropped over 90%. There are companies that are doing that right now. The market's super volatile. SaaS is getting hit really badly, but you guys played offense. You dug your heels into the ground—I don't know if that's the right term—and then you just went after it.

Do you think this is a different environment for that circumstance? Do you think those companies—what does it take to pull out of—

It—

—a 90% dive?

I think it depends on the business model. The challenge you have when the stock price goes down is your stock-based compensation as a percentage of total goes way up.

Let's say you're a company that typically issues 2% or 3% of its equity every single year to the team. There's a dollar value with that. If the stock goes down 90%, are you going to issue 30% dilution every year? You'd never recover from that.

We always thought about that in fixed-dollar terms. We want to pay our people exceptionally well, but we don't want to overpay our people, and we keep a lean team, so it's always manageable.

On the other side of that, we generate a ton of cash. When the stock tanked, we didn't have to say, “Look, we've got to go convince investors to buy the stock.” We actually didn't do any investor relations for well over a year at the bottom, because what were we going to convey to investors? “We're rebuilding our technology. Our stock is dirt cheap.”

Nobody buys something that's dirt cheap. They want to see a vision. They want to see the long-term opportunity and growth prospects. Investors, just like the private markets, will always chase something that they expect is going to grow really quickly, but they very rarely tend to buy something where they think it's just cheap, because things trade wherever they deserve to trade.

We were able to flip it and say, “We're just going to buy our own shares. We're going to become our best investor. We're going to give our team nice equity compensation, but not overdo it, because we can't afford excessive dilution. On the other side of that, we're going to deploy every dollar we make and more, and just start buying back our own shares and be our best investor.”

That was only possible because we had conviction in what we were building and conviction in our future growth opportunities. The disconnect with public-market investors was that they felt like we had no growth opportunities. We, on the inside, fundamentally believed we could extract a ton of growth out of the business with the right technologies.

I think the challenge is that not every company is constructed the same way. Some don't have the cash flow to do that. Some don't have the capacity to actually grow. They're trading where they trade because those growth prospects aren't there.

Especially when you talk about enterprise SaaS, you have very robust platforms, but they're not algorithmic. So how do you actually accelerate revenue without go-to-market teams? If you have to hire more go-to-market people, your stock is low, and your burn goes up, it's just a downward spiral. It is company-specific, and some of these companies will have a tough time recovering.

What do you think's going to happen to them?

A lot of times, when companies reach that point where things are tough and their revenue multiple and potential cash flow multiple end up pretty low, they start getting accumulated by private equity or effectively plucked off the public markets and taken private, because the only way to recover is a complete restructure.

If you do big layoffs, you fire a lot of people, cash-flow it, and lever the business up, and that fits the private equity model. It doesn't fit the public markets all that well anymore.

Once you can't convey that you've got pricing power in a big market on the other side, you tend to lose all your multiple in the public markets, and then you're not a good public company.

As we think about the next 12 months—I don't know if you guys think about things in 12 months. Some people think in the next week. I don't know. It's crazy right now. What are you most excited about?

You?

What I'm most excited about is—I think what kept me excited in the last 3 years has been pretty constant. The team is constantly making changes to the models. We are improving our infrastructure.

Also, as I previously shared, as the business is growing, the number of people stays the same, which actually means every team member is growing together with the company. By the way, that is not typical, no matter what company you are looking at.

All the hottest companies we're seeing are growing, but their headcount grows at the same time, which means individuals in those companies do not have the same growth speed as the company itself. I think AppLovin is different.

Every individual in the company has outsized growth. Seeing the company grow, seeing myself grow, and seeing the people on the team learn more and become more mature—that just keeps me excited.

I don't really know the feeling. I don't get excited.

What do you feel? Do you feel anything?

I'll give you an example. When we went public, a lot of people celebrated and thought, “Okay, we went public. It's a great milestone.” For me, it was, “People just bought our stock at a $30 billion market cap. Holy shit, we've got to make those people money.”

I never think of the world in excitement terms. I also don't get down on the other side of the spectrum. Everything is a grind. We're trying to build something that hopefully lasts years after we're both gone, and we're thinking about time horizons very far into the future.

If you get excited, you over-index on something short-term that probably isn't relevant long-term. On the other hand, if you get down, it's also really hard. I've just learned to manage both of those emotions and really try to think, “Okay, let's make sure that we have the right people here and the right path, and then everything else will fall into place.”

Are there any certain parts about what's accelerating in AI, or at least turbulence in the market, or anything macro that is concerning you or keeping you up at night?

I had a 92 sleep score last night, so I sleep pretty well.

In a competitive field like this, when we first went out looking for seed funding with a $4 million valuation, we were told, “Big companies are going to put us out of business. We're not even worth a million over four.”

When you're competing in a field like this, you can't lose sleep over things around you. What you have to do is be very aware of everything around you, and you have to have some aspect of being visionary—seeing around corners that others can't—and trying to build, again, for the long term, just the right way.

Because we've always, in some ways, been disadvantaged and had to become advantaged in the face of those disadvantages, I don't know that anyone at the company loses sleep over what's around us.

It's more that the world of technology is moving really, really quickly. If you live in a world of tech debt and you're unwilling to move quickly with it, then you're potentially going to lose, and that would get us nervous.

But we're constructed for the way the world is today. We got a little lucky that some of the things we believed ended up fitting perfectly today. We couldn't do what we're doing today with the team that we have if we were not now in partnership with AI, enabled by AI to go faster.

A lot of the stuff that's happening in the world is exciting for us because it fits exactly what we're constructed to do. But the reason I don't get overly excited about it is because you still have to execute. At the end of the day, the burden of execution is still present. You have to have the right people. You have to have that hunger. You still have to be able to push forward hard.

Okay. I have to ask you both this now, based off of that. What happens when we hit—and if we hit—the singularity?

What do you mean by singularity?

Yeah, I don't know what that means. Some people say we already hit it.

Everything is autonomous. I don't know.

So some people say we already hit it.

The machine takes over.

I think that's a bigger question than AppLovin. Sometimes I have to believe that singularity moment will never come. This is not an opinion from fact; it's an opinion from my standing. As a human, I have to defend the value of humans, right?

Overall, I think we're seeing how AI is advancing. I feel pretty optimistic as a company. I think it's thanks to our structure. Our structure allows us to benefit from the improvement of AI much more than our competitors. So overall, if AI just improves at the current speed, I feel it's a positive thing for us.

You could probably find 100 very smart people right now on Twitter saying we're at the singularity now. It's a weird one. I think a lot of us in tech live in the Twitterverse. The world is much bigger. We get questions like, “Well, isn't the chatbot going to tell everyone what to buy and just complete the transaction, and no one will ever do e-commerce off of an ad again?” What that loses sight of is that people actually like to shop, and people want to look at different products.

People want to click buttons on a shopping site, and people want to hit buy and then track their purchase and get the box and open the box and whatever. At the end of the day, we're all still human beings. Different people have different understandings of how these technologies can work, but probably most of the real world is not going to understand how these technologies work at any level like we all talk about in the Twitterverse. The reality is, whether we get to that point or not, it's going to be a different impact for different people, and we're still going to have to understand that humans need human interaction. Humans need human solutions as well, at the end of the day.

I have 2 more questions left. One, maybe that was your answer, but what is your hottest take right now?

I think, in this era, AI allows everyone to build almost everything they want. Taste is a very important thing, and I think it has been underestimated.

Yeah. I was just listening to David's interview with Micky Malka, and Micky Malka was talking about how Silicon Valley lost the ability to make beautiful things.

Yeah. When I say taste, I don't just mean aesthetics, right? It's also engineering taste. From personal experience, I would attribute the success of AppLovin largely to what we decided not to do, not actually what we did. I saw these mistakes that all the small companies make when they're trying to challenge giants: What they try to do is copy exactly what the giants are doing. But the problem is, you're just such a smaller team. You have much fewer resources. You have no chance to win if you're trying to do exactly the same thing as your giant competitors.

I think this is where taste comes into play. We know our difference. We know where our strength is, so we deliberately decide what to focus on and what not to work on.

Where do you get your inspiration from? Do you get it from outside sources, from abstract things?

I think first principles.

Okay.

Just first principles. I guess you just have to ask the questions of what, how, and why. A lot of times, people overly focus on what. People say, “There's AI. Let's do some AI in the company. Let's see how AI can help our business.” Then they will start hiring a big team working on AI without asking how. There's AI—how do you apply AI in your product? And then why? Does your product actually need that, right?

This applies to AI. It also applies to previous generations of technology. When we were building Axon 2.0, the team was really small. I had a lot of experience in the industry. I knew how a recommendation system looks in many big companies. I think it was impossible. It would have been a big mistake if we tried to hire a huge team and replicate every single piece of that. So it is important to filter through all the noise and figure out what the one most important thing is, and have everything you can in your team focus on it.

My hot take, going back to marketing, is that there's a lot of animosity toward AI in the world right now, and people, again, in the Twitterverse talk about AI as if it's so amazing. If 99% of people are just using it as a search engine, and then they're hearing a lot of smart people saying it's going to take jobs away and there's going to be all this loss here and there, why would you be a proponent of AI? To me, it comes down to marketing. The way to flip people is to always talk about the positive of something and let them really see it, even when they don't understand it.

Tied back to our business, we serve ads. At the end of the day, ads could be seen as annoying. Most people don't notice the ad. If we just went out and said, “We serve ads,” that wouldn't get anyone excited. But you have to understand that an advertisement is creating a transaction that otherwise wouldn't have happened, so it's creating economics. It's helping expand the economy, it's helping grow businesses, and it's helping those businesses hire people.

When you have a really good ads framework, you actually are creating a positive catalyst on the economy. Once we learn to speak about it in those terms and really cherish the moments where we know some developer or some shop is able to hire people because they're able to get growth that's profitable on our platform, that makes everyone here feel good. To tie it back to the AI narrative, until people start understanding what positive use cases the average person can have that are incremental to search, they're not going to be proponents of AI, and there's no reason to expect them to be.

I like to ask this question from more of a personal performance standpoint. For you, Adam, Michael Barton said you were the most locked-in person that he's ever met, but I'm curious, from maybe both of your standpoints, how did that come to be, and is there anybody that you've been inspired by that keeps you motivated? How did you become that locked in?

I don't know if I'm locked in. I always optimize my life to remove all distractions and focus on the one thing that I was good at, which was building a business in the category that I'm in. For me, early on, it was finding my passion. I don't know why advertising became the passion, but it was sort of an extension of something I used to do: derivatives trading, finance, math, algorithms, placement, prediction. All those things came together in advertising.

So I found my passion, and then I tried not to do things around it that would create distraction. I'm not out there super active on social media. I'm not out there making a whole bunch of other investments. I'm not out there speaking at a lot of different conferences. I'm basically 100% wired in to try to do the best job that I can with this, and if I can do that well, then I feel like I'll have succeeded. And so that's my answer.

I came to AppLovin as a pretty junior, middle-level team member. All those things about being an executive are pretty new to me, like attending conferences and giving speeches. Honestly, I haven't really thought a lot about that. But I think an important part of my success is that, I don't know how, but in my early life, I developed this desire not to disappoint. I don't like to disappoint people.

Whenever people around me—my leaders, Adam—have some expectations of me, I just want to do anything I can to make it work.

Why do you think most people don't do that?

I don't know. Maybe it's a type of idealism I've had since I was young. It's a slight personality thing.

Does that become an obstacle sometimes?

It could be. Sometimes I feel like it becomes an obstacle because I have a hard time pausing and enjoying the achievements we have. I don't know if Adam feels the same thing. There was a lot of excitement in the past because of the success of AppLovin, but I find it hard to press the pause button and just enjoy that moment.

We do have a trillion-dollar milestone ahead of us.

Yeah, maybe.

I think then you have the next one. This is actually a common trait I've found in highly competent, successful people: There's no moment to really enjoy because there's no sort of goal that you get to. Once you get to one, you're onto the next one. When you're raised in households where the A is not good enough, you have to get the A+, a lot of times that builds this in people. The downside of it is, it's hard to reach fulfillment or happiness, or some of the emotions that a lot of others would have.

But if you're always achieving for more, you're pushing yourself to keep going, and no moment is good enough.

Are there any other traits you look for in a player?

I think the type of people I really like on my team usually have two characteristics. One is intelligence. The second is having a low ego. I think the intelligence part is easy to understand. They have to be smart, learn fast, have higher standards for themselves, and get things done. This is the characteristic a lot of companies are looking for.

But I think the second part, having a low ego, is often overlooked. When a person has a low ego, they tend to listen, empower their team, and help others. More importantly, they take feedback, and then they proactively question themselves. This is a very important part of growth.

Yeah, and it usually helps with the third thing, which I look for and always ask about: how they react to adversity. Nothing is easy in the world, and some people grew up with easier upbringings than other people. Usually, people who know how to respond well to adversity do really well here.

Yeah. This is kind of a danger for high performers because when you are a high performer, a lot of people around you observe you as a high performer. They trust you. They don't question you, right? So it becomes extremely important for those high performers to be able to question themselves.

I've seen some people do extremely well at this, but some others might forget to be the ones who challenge themselves.

I know I said I already asked a difficult question, and I almost left without asking this one, but Adam, you built out a gaming advertising company. You're in the gaming space. You know who else was in the gaming space? We had Torsten Reil, who then went on to found Helsing. We had Palmer Luckey, who went on to found Anduril. What comes after AppLovin for you?

That's a tough question. I'm probably—hopefully, if I'm good enough and earn people's trust around me—going to be the CEO for a long time. I don't think about what comes after this at this point, as a public company CEO running a business this large.

I think it changed for me along the way, because when we were smaller, I did used to think about building again, and I do really love the early stage. But not a lot of people in the world have had the privilege to be able to work with people and together build something from zero to $100 billion-plus. Where we are, we've just got so much opportunity in front of us, I can't really sit there thinking about what's after this.

Wow. That's a great place to end. Well, Gio, Adam, this was such an interesting, different, and thoughtful conversation. Thank you so much for giving us a deep dive inside the engineering side of AppLovin and the growth that you guys have achieved. I really appreciate it.

Yeah. Thanks again for having us.

Thank you for having us.