Gaming Gets Smarter: Insights from King's Head of AI

Luka Crnkovic-Friis, Head of AI at King, talks to AI Business about how AI can accelerate gaming development. Luka outlines King's data-focused approach to development and how the gaming industry has shifted in the last 12 months amidst the AI wave.

Listen to the podcast below or read the edited transcript of the interview.

AI Business: How does AI fit into mobile gaming and how does it differ from traditional console and PC gaming?

Luka Crnkovic-Friis: Traditional console and PC gaming is a lot about creating an immersive environment, telling a story — essentially more related to traditional types of entertainment, like watching movies or reading a book.

Mobile gaming is quite different. Our players play in short bursts to de-stress, it's not so much about showing them some fantastic immersive environment, but more about making the game work for them and their needs.

Mobile gaming was a big revolution when it moved from being on social media, like Facebook, to handsets. The reason for that revolution was that it was more of an adjustment to the player's wants. What AI allows us to do is take the next step and personalize it more, adjusting a game to exactly what an individual player wants, rather than to a group of players.

AI Business: Where does AI fit into King's games? Walk us through some use cases.

Crnkovic-Friis: It's a very broad usage. We have about 300 use cases that we're working on ranging from purely commercial to gameplay related. The use cases touch on everything from level tweaking and level generation to personalizing offers to the player to essentially try to do better in reaching out with communication with visual arts and things like that.

That's half of the story. When it comes to gaming, the other half is not specific to us, or gaming in general - the productivity increases and new creative opportunities.

In terms of specific projects, we're running about 180 million predictions per day for our machine learning platform.

The challenge has been not so much in getting a good model that predicts player behavior but what to do with it. Even if it's, for instance, a recommender system that recommends power-ups in the game, even if you build one that players tend to use twice as much as they do with the previous methods, what to do with how they integrate with how these play with the other game mechanics and so on. There's a lot more complexity than just a machine learning model.

AI Business: You joined King in 2022 following the acquisition of Peltarion. Talk us through how you've seen AI change at King since you joined, considering the recent generative explosion.

Crnkovic-Friis: King has always been very data-driven. Ten years ago, when Candy Crush was launched, it was based around the data and A/B testing. There was a machine learning team in place that was doing a lot of great stuff. What the Peltarion acquisition meant for King was a boost in terms of the number of people and technologies available.

Six months in, we had the generative explosion which was unexpected. I've been working for 25 years, and generative and large language models are nothing new. But just the speed that has developed in the past six months has been remarkable. And that sort of makes you rethink things from scratch because completely new opportunities open up.

For instance, if we were talking about personalization, a typical approach would be to have a central system that tries to learn player behavior, and then adjust the game, according to the individual.

But what if we don't do that? Instead of doing a central system, what if each player has their own AI buddy who learns how they prefer to play, what they're playing, and so on. That AI in turn talks to an AI that’s at expert levels, and they communicate. … There's a ton of new possibilities opening up, in terms of game design, and so on.

AI Business: Generating assets in gaming isn’t new: 2016’s No Man’s Sky proved procedural generation can power gaming. Where do you see generative AI fitting in?

Crnkovic-Friis: The difference is in the algorithms behind it. The traditional type of pre-neural network generative models are mathematical models, where you have a known equation that you iterate on, and it can create textures, levels, or whatever you optimize it for.

The cool thing about AI is that it's really good at imitating the human style of doing things like writing and doing levels, so it offers up much higher quality results than something that's coded by a limited algorithm.

AI has applications just about everywhere, from generating assets in the game to complete generation of levels and even code. In our case, we've been working with level generation, in particular, level tweaking — helping level designers using our playtesting bot reinforcement learning base that can very quickly play through levels and determine if it is too difficult, etc. …

We have over 12,000 levels in Candy Crush, so have tons of data. Literally, billions of people have gone through levels. And that can be used to train a model that then comes up with suggestions for a new type of level. You can then condition the model to tailor levels to certain types of players.

It’s not just that we’re doing levels up and adding levels, we are constantly improving all levels across the game. If you played Candy Crush a few years ago and played level 50, if you played it again today, it won't be the same level. It will go through a lot of modifications and improvements; we constantly iterate on it. And the generative model can come up with a suggestion and then we have level designers who look through, critique it and determine improvements. This gives level designers an opportunity of producing many more levels than they could have otherwise.

AI Business: I love the process in which it takes to develop a game, but obviously you need thousands of man hours, thousands of developers to build it. How much of a boon is AI to help those overworked staff people create concepts, themes and overall assets in games?

Crnkovic-Friis: Generative AI will be a massive thing for game developers. All the internal tests that we've done show that designers, once they use the technology, become very positive towards it. There is initially fear about it replacing tasks and work and so on. Generally, when developers see this can help visualize things quicker, they can come up with drafting, and even come up with fairly good results. But you still must specify the creative vision. And that's not always something you can easily do just in words. For example, Midjourney can produce good-looking images, but it may not be exactly the image that you want.

This is still in the research and experimental stage. Also, the legal situation is very unclear, there's still work to do. But we see Adobe has AI standards now in their latest Photoshop release, so I'm guessing that will lead to very broad adoption.

AI Business: How cautious should developers be considering the legal issues around copyrighted materials? What kinds of conversations have you had about this at King?

Crnkovic-Friis: We have our AI labs team and creative teams on Candy Crush looking at that. The legal situation is not so much about the use of copyrighted material; we have plenty of assets of our own. It’s more about the copyrightability of AI output that’s unclear.

I see that it can only come down to one thing in the end: It will be allowed because it's such a massive productivity boost for so many. But still, it's very much a grey area right now. We are treading very cautiously in that particular part.

AI Business: Where does AI fit into mobile gaming over the next 12 months?

Crnkovic-Friis: There are two components: productivity increases and assistants. If you’re working in an organization and running projects, get a copilot to help you structure content. We’re doing a lot of that now and trying to get that going as fast as possible. And then there’s using AI to be a time saver but then there’s just general cognitive acceleration.

On the actual game side, one of the big areas is personalization. When you play Candy Crush or one of our other games, typically the core game experience is one part of it. But you have various events, competitions and minigames going on as well, and users have very different preferences. What AI allows us to do is tailor it to the user’s preferences.

Another part is automation. Right now, we're basing most of our stuff on A/B tests, and we're running hundreds of them at any single time. That gets very complex and it’s hard to coordinate decisions. AI can help make decisions and analyze results from tests.

And finally, content generation. It’s a bottleneck that we've always had, building high-quality levels, even if we have a lot of people working on it.

To keep up-to-date with the AI Business Podcast, subscribe on Apple and Spotify or wherever you get your podcasts.

Stay updated. Subscribe to the AI Business newsletter