Podcast: Understanding Unity’s Vector (with Felix The)

On this week’s episode of the podcast, I am joined by Felix The, SVP of Product and Engineering – Advertising at Unity, to discuss the inner workings of Unity’s Vector product and the strategic integration of engine-level data into the advertising ecosystem. We explore how Unity is rebuilding its machine learning infrastructure to provide more granular predictions and better performance for mobile gaming advertisers. Among other things, we discuss:

• How the integration of real-time game engine signals can improve user acquisition performance for mobile game advertisers
• Why the shift toward massive unified models represents a fundamental departure from the traditional fragmented approach to machine learning
• Whether the use of runtime data provides a decisive competitive advantage over traditional software development kit signals for predictive modeling
• What the transition from manual creative production to generative exploration means for the long-term sustainability of performance marketing budgets
• If the ability to test core gameplay loops through playables before full development can significantly reduce traditional soft launch risk
• How personalized creative units tailored to micro-cohorts will solve the persistent challenge of declining engagement in broad audience targeting

Thanks to the sponsors of this week’s episode of the Mobile Dev Memo podcast:

• ⁠⁠⁠⁠INCRMNTAL⁠⁠⁠. True attribution measures incrementality, always on.
• Xsolla⁠. With the Xsolla Web Shop, you can create a direct storefront, cut fees down to as low as 5%, and keep players engaged with bundles, rewards, and analytics.
• ⁠Branch⁠. Branch is an AI-powered MMP, connecting every paid, owned, and organic touchpoint so growth teams can see exactly where to put their dollars to bring users in the door and keep them coming back
• Voyantis. Voyantis uses predictive AI to transform customer value into high-impact signals that boost ROAS across Google, Meta, and more.

Interested in sponsoring the Mobile Dev Memo podcast? Contact Mobile Dev Memo advertising.

The Mobile Dev Memo podcast is available on:

• YouTube
• Apple Podcasts
• Spotify

Transcript

Eric Seufert: Welcome to the Mobile Dev Memo podcast. I am your host, Eric Seufert, and I am joined today by Felix The. Felix, welcome to the podcast.

Felix The: Thank you, Eric. Thank you for having me.

ES: I am very glad to have you here to discuss all things Vector and Unity Ads. I have been looking forward to this conversation. Before we dig into the meat of the conversation, can you please introduce yourself to the audience?

FT: I have been at Unity for nine years, so I have seen the different stages of Unity, the ups and the downs. I have a background in both optimization sciences and game development from my undergrad. Unity is the perfect combination of the two. I am passionate about video games, our creators, and finding ways for them to be able to monetize their creation and get discovered. Unity plays an important role in the ecosystem, and I am happy to be a part of it.

ES: For some context, we have known each other for quite a while. My wife works at Unity; she has been there for over a decade. But we met in person for the first time over seven years ago because we were doing the class in San Francisco when we were about to have babies. It was the CPR class taught by the firemen. We both lived in the same area then.

FT: Our kids are one week apart, I think.

ES: Also, full disclosure—and this is public—I am an advisor to Unity. I am advising on many of the AI initiatives that Unity is undertaking, and those are diverse and wide-ranging. Today, we are only going to be talking about the ad network initiatives. I always like to start these conversations with a big, broad, conceptual question. Since we are talking about Unity today and Unity Ads, I will start with Vector. What is Vector? How does it work, and what is its purpose?

FT: Vector is Unity’s AI-powered growth and user acquisition platform. It is the fastest-growing part of our business and represents a full rebuild of our machine learning and data infrastructure that powers our advertising platform. At a high level, Vector connects gameplay behavior, monetization signals, and campaign performance into one continuous learning model.

What really sets it apart is the architecture. Vector is specifically designed to ingest and interpret very large volumes of gameplay-related data. The model we use is also very centered around games and application discovery. It makes a natural fit with Unity and our engine business because we understand how games are built and how players actually behave inside them. By pairing the gameplay insight with our ad network at scale, Vector can learn faster and make better predictions, which translates into stronger results for developers and advertisers.

Since launch, Vector has grown rapidly. Revenue in January was up 72 percent year-over-year. It has been growing sequentially quarter-over-quarter, and we believe we are still in the early adoption curve of growth. Now, our focus is to work on the core engine driving performance across the growth business by making sure all the data elements we are partnering with developers to use can power Vector in ads and beyond. We invest in Vector because we understand games not just from the perspective of machine learning and AI, but also the way games are built and played.

ES: This has been an ongoing initiative for some time. Was there a start date or a moment when this initiative took shape?

FT: The most significant moment was somewhere between Q2 and Q3 of 2023 because that is when the industry actually saw the impact. In R&D, you do not really have a single release. In AI and machine learning for ads, there are elements of the model you need to update sequentially, and there is the data architecture you need to update. That work was ongoing, but we saw the majority of the benefit happen around Q2 and Q3 of last year, so that is when we went public with the message about the impact of Vector.

ES: That is a good point about the R&D process. I want to hover there for a minute, particularly with respect to ads. I was on a crusade for a long time against the use of the phrase AI in this context because my argument was that this is ML, not AI. But I gave in because it is more exciting to say AI. Nonetheless, that has become the big topic of discussion across the entire landscape. Every time a company mentions AI in earnings, you get LinkedIn posts about how the new algorithm is changing advertising wholesale. But the reality is that these systems have been in place for years, and they make sequential updates to them. This is a research project. Talk to me about that process of constant improvement and iteration in systems like Vector.

FT: There are two paths. One is that you can tune based on scientific advancement; there is always a better model or a better way to develop certain protocols, parameters, or adjacencies. For example, do you want to tune the model based on a certain outcome or based on understanding multiple outcomes in a multi-head approach? Those are modeling techniques that improve over time.

The second path for iteration is that you do not have to limit yourself to model development alone. Based on the availability of data, how that data is represented, and the nature of the data sources, it opens up avenues of outcomes for modeling. Think about the model on two tracks. The modeling team can improve the model independently, but you also want the team to understand what is possible based on data. Based on what is possible with data, the techniques you employ and the approach itself will become fundamentally different. Those are the two things that tend to be iterative in nature.

ES: Within that data sphere, you have opportunities with data augmentation, purposefully noising the data to add diversity or to train the model to pick up on what is noise and what is signal. On the modeling side, you could go down to changing the optimizer with a different hyperparameter. You are making adjustments; some are big and broad, like a multi-head architecture, or they might be changing the momentum in an Adam implementation. This is a regular, ongoing process where researchers are focused on this iteratively. It is not just that Version 2 is out and that was the first time we deployed a dramatic change.

FT: That is right. The Vector we released for 2023 is not the Vector that people experience today. We did not say it is Vector 2, but because we have researchers and scientists actively working behind the scenes, every time there is a better way to do something through modeling or data, we release it and let the customers experience the benefit.

ES: Talk to me about using the engine data for ads targeting. How does that data contribute to better outcomes, and what might that data look like?

FT: Unity sits at a unique intersection of creation and growth. Unlike most companies that are squarely in the growth area, we have a lot of developers in mobile who make their games with Unity—more than 70 percent of the top 1,000. Our engine has that coverage. There is a portion of those developers not using services around user acquisition or ad monetization. What the engine data allows us to do, with the right developer framework and consent, is use gameplay data to benefit them when they are ready to engage in those services. This is also data they get in return for using products like diagnostics.

This means Unity potentially has access to a footprint of data that you typically do not get in many other sources. The second part is that our runtime data, from a data quality perspective, is the gold standard. When the game boots, that is the Unity runtime; it is the first thing that gets instantiated. Our runtime data does not have to worry about quality issues like SDKs getting initialized late or missing certain things. We also do not have to worry about joining data from disparate third-party sources where it only takes one source to not do their job for the whole thing to lose value. Instead of pulling from a patchwork of sources or manually exporting datasets that have been stitched together, we get access to the cleanest quality data straight from the runtime and the boot time when applications get loaded. That is the power of the Unity runtime: coverage and quality of the data signal.

ES: A point you made to me that I had not really considered is that having the time series is really important, and knowing the fidelity of the time series is also vital. When you are dealing with SDK data, you do not actually know if event A happened before event A minus one; you might be getting it in the wrong order. But with the runtime, you know with certainty that these events happened in a specific sequence. When you are doing sequential modeling, it is imperative that you believe the sequence is real.

FT: Causality is important. If causality is broken, then advanced modeling techniques like understanding sequences of events become moot. Also, a lack of causality can lead you to think correlation is causation, which is dangerous. The idea of having clean data with the right sequence of events represented correctly without joins is powerful.

ES: The only way to truly know the sequence is correct is to have access to the runtime. With an SDK, you are always dealing with transmission issues that could cause it to be out of order. Developers also have to know what to send you for you to help them, and that is not always clear. No one may know; it may take that experimentation process to unearth what is actually valuable. This was the process that took shape for many developers when SKAdNetwork was first launched. We had to be thoughtful about conversion values. Do we know what is valuable? Many developers had no idea. Once you went through the process of trying to adjudicate the value of those events, it became clear that a lot of developers had never really thought about it. They were sending things they assumed were valuable, but in fact, they were not. Encoding a single event was not really that meaningful; you really needed to encode an event onto a sequence. When the sequence happened, that was the event that unlocked usefulness.

FT: To build on what you said, it is not just the manual work by developers—people can interpret things differently. The lack of standardization across events makes it really hard for AI products to use. Standardization and consistency without manual intervention that can cause error is one of the benefits of runtime data.

ES: And the volume. If I am a game developer and I have one game, I have one bank of data and I do not see that much. Conversion events are rare. People buying stuff in a mobile game is rare. When I wrote Freemium Economics, I talked about the five percent rule as a rough heuristic where you expect five percent of your users in a freemium product to monetize. But the reality with mobile gaming now is that given some categories, that number could be sub-one percent. You are talking about a very tiny minority of users actually doing the thing that you care about, which is making a purchase or contributing revenue. That has changed because ads are a more material portion of mobile gaming revenue than when I wrote that, but you still have this broad stratification of revenue profiles. If you think about LTV, it is much more stratified now than it was in 2014, and that matters because that long tail is where all the value is.

FT: You are right about that. Sure, ads become more relevant and there is more diversification, but modeling is also about that long tail. It is about the signal versus noise and the power of differentiation. Sometimes the power of differentiation is less about understanding that CPM varies, and more about a whale purchasing ten thousand dollars in a game. That signal differentiation coming from transactions still matters a lot in our ecosystem.

ES: Anybody that has ever worked on classification with extreme class imbalance knows that if you have one percent of users making IAPs, and your classifier just says everyone is not going to make an IAP, it is correct 99 percent of the time. It looks really good if you say your classifier has a 99 percent correctness rate, except it is wrong for the group of users we care about.

FT: Correct.

ES: Why now? Unity is what, 15 years old? Unity Ads was the Applifier acquisition in 2012. Unity has had an ads product for quite some time. What was the impetus for Vector?

FT: Prior to Vector, the team did work to improve our system. Typically, when you already have a working deep learning model in the past, it is always a big decision when you have a viable business running but you are wondering if this is the system you want to replace. Essentially, you are flying a plane while you are building a new one at the same time. You need to find a way to chart a course for a smooth cutover.

It is not simple, and any R&D rebuild comes with a real risk. We got really comfortable in taking that risk when we saw a few things line up: new modeling techniques, new information available for us, and new machine learning frameworks that we used in Vector that we did not use in the past. We also had next-gen inference tech that simply did not exist when the original system was built. Combining that with our belief that this is the pivotal moment where we can build the best user acquisition solution for games, it became clear that rebuilding was the right course of action. We are pretty pleased with the outcome.

ES: Talk to me about some of those new techniques. I had Meta’s VP of AI ad products on, and I was asking him pointed questions. I am not going to ask you for any secret sauce, but what has been an unlock?

FT: Generally speaking, I believe in big models, not small ones. By big models, I mean you could create models that are an army of smaller models trying to do different things, or you could combine them so they can actually understand what each other needs to understand. Auto-correlation, covariance, the same training pipeline, and understanding the causality of decisions made by another model—all these are factors. I feel like the future is about big models.

The second part is that even the data platform and inference world has changed quite a lot. Something that we could not imagine in the past could be something that we imagine now. On the data side, having a full sequence of data about a player will unlock new modeling techniques like sequential modeling. I believe the future of understanding a player is not just about understanding them at a certain point, but actually being able to predict what is next. That is powerful because you can think about the applications of what you could do across multiple uses.

Third, big models should not scare us. In the past, a big model without the utility of data was pointless, but we surpassed that. Inference techniques and new research allow a company like us to do this in practice because we can manage the cost of a big model by optimizing inference. I believe the future of this space is about big models, how to optimize your cost, and using new data sources to make your model even bigger.

ES: What does the mid-term roadmap look like? How do you expect Vector to evolve?

FT: We are very keen on starting to utilize the runtime data. A lot of the Vector performance that we have celebrated to date is without that. We are pretty confident, based on early signals, that this is something we can utilize, and that will unlock new modeling techniques.

Outside of advertising, you can see how AI has shaped the world of development and how people interact with products. We do not think the current classic way for people to interact with a product will matter as much in the long run. What matters is interacting with a product where you can express the intent of what you want to do, whether it is a marketing campaign, the money you want to make from ads, or IAP. A lot of this hinges upon our ability to embrace AI beyond just the predictive ML we offer at the core. We believe the new workflow for Unity Ads will be centered around embracing AI that can influence your campaign performance. The third piece is what I mentioned earlier about big models.

ES: What is nice about the mobile gaming context is that you probably have a clearer picture of what you want. If I am a D2C advertiser, I could be setting a number of objectives at different points in the funnel. For mobile games, you want the user to install, enjoy the game, and be engaged. You are probably less interested in just someone going to the App Store page, which you have no visibility into anyway because once they click out, you do not know what happens. You do not have any visibility there; that is just how the App Store and Google Play work.

FT: Yes. In mobile gaming development and general gaming development, the thing that will make a lot of impact is minimizing your pivot and your change at the 11th hour. The issue is that you cannot always avoid it because you cannot model fun easily. The game industry is a hit-driven business; you never know until you try. But there is a creative way to minimize the risk. One way some developers are starting to minimize risk is to not follow the traditional soft launch, pivot, spend, tweak, and global launch model. They try to minimize development costs by porting an existing title, re-skinning it, or changing game mechanics slightly to create a vertical slice, but not the full game.

Then, they launch that as a playable unit where they do not measure for installs—there is nothing to install—but they measure engagement. That tells you a lot about marketability and the viability of the game concept. When you see something is wrong, that is the signal they use to pivot. At that point, the cost of pivoting is very low. They have completely flipped the traditional model of expensive 11th-hour changes. I think that is a really healthy shift for the industry. If more people adopt this, we will see less R&D cost spent down a rabbit hole without a viability of success.

ES: That concept just accelerates everything because you are getting actual real data much faster. Then the calculus comes down to how well we can proxy the data we get from this to commercial data. That was always the issue with soft launches: how well can we proxy day-seven retention or acquisition costs to when we are spending five million a month? But this allows you to collect actual real data much faster.

FT: When you soft launch, the game is complete. That is the issue. If the soft launch is not looking very good, you can make some alterations, like balancing the economy if IAPs are too expensive or free gems are too cheap. But what if the core loop itself is not fun? At that point, the game is done. It is hard to undo ‘done.’ The idea of this iterative playable testing is that the game is not done, and you get the feedback from the people that matter most.

ES: The MVP for a soft launch candidate is 90 percent of the way to the finished product. You have to build the systems, the core engine, the economy—you have a lot of things in place. It is a fully baked product, and you are just doing tweaks. If you could just do a vertical slice with a playable, the baked-in decisions are reduced.

FT: Yes, that is right. Instead of stipulating about what the game will do out in the wild, you get that data.

ES: We hear a lot about AI in game development and all the ways developers will be able to avail themselves of these tools. How does Unity support developers as AI penetrates further into the process?

FT: Unity, at the inception of our core value and product design, is open. We have a vibrant plugin ecosystem. Many of the tools are made by third parties. Unity is used as the assembly point where you check that all those tools are used in the right context to build the game you want to ship. Then you ship it with Unity. In fact, we ship with the game. In Unity’s case, when you click play in the editor, that is your game running.

We never shy away from opening ourselves up to third parties. We have Unity AI, which is a first-party AI we built. Are we going to open up to third parties? Yes. We have launched the Muse AI gateway with a couple of partners. We are also planning to find a way to open this up with more people that use other workflows through our server. We never shy away from third parties because Unity is the destination point where people bring their ideas to life.

One reason why we invest in a first-party AI product is that we believe we understand games. A lot of models out there are very large, like LLMs. Sometimes using a large language model that is generally smart for a remedial task like creating an NPC character is overkill. It is expensive, and they are not specialized or tuned for the object of making an NPC. Creating an NPC has dependencies: you need the mesh, the texture, the concept design, and animation rigging. There is a sequence of operations. We believe Unity can excel in our own agent because we understand the Unity context. We can make a model that is very cost-effective and very smart at making video games. We do not care about them being smart in doing anything else, but we understand developers. Developers do not want to spend too much burning tokens to create games. They want to make games by consuming tokens as an expression of productivity. We want to invest in our first-party AI to solve that problem.

ES: I think we are going to see a recognition at some point that the way a lot of people have been interfacing with these development tools does not scale. I see people brag about building a tool with code and then, when they need to add a new feature, they basically just rebuild the entire thing with that feature in it. They are one-shotting everything. The reality of video games is an enterprise-grade entertainment product. You need to think about not just the visual fidelity, but the mechanics, the services, the monetization, and the economy design. All these interactions require thought.

FT: You got it. It is like one-shotting a game. The idea of game development with AI should be ‘human in the loop.’ You have context of what you have done before and you build based on that so you do not regenerate. You want to change a small bit without touching anything else. That requires a deep understanding of how humans and AI interact.

ES: These token costs accumulate into real costs. I think people don’t recognize that the view of GenAI as just continually re-prompting or one-shotting things is a mistake. All those core design principles developed in software engineering over decades are important, and they are going to be more important in the GenAI case because you could have all these agents burning tokens with no sort of discernable or scalable output.

FT: I agree with that.

ES: What are you most optimistic about with mobile gaming?

FT: At its core, gaming has never been about having the best graphics or the most complex design; it is about fun. Mobile continues to be the best place for fun to show away. Smartphones are powerful enough to display rich graphics and experiences, and they are portable enough that anyone can access content at a massive scale. That is exactly where Unity sits.

Unity is mindful that games are not just about graphical fidelity but all the ongoing services and the design of the game itself that make it a good experience. Mobile will also benefit from this AI revolution. Mobile has the most device penetration, the most volume, a working discovery dynamic, and flexible monetization models. It also has support for graphics at any level of complexity. We sit to enable this mobile propagation to exist and thrive. With content generation with AI becoming more and more accessible, more devices will get entertainment content. I have high hopes for how mobile will go in this AI era.

ES: Felix, this was great and very informative. Thank you very much for taking the time. I really appreciate that we got the chance to have this conversation.

FT: Same here, Eric. We should do this more often.