The acceleration of technological change is evident everywhere, and it’s benefiting games in unseen ways. In 2016, King had only 2,000 levels for Candy Crush Saga.
As the publisher King turns 20, there are now 15,000 levels for Candy Crush Saga, which has helped the company generate $20 billion in revenue and five billion downloads to date. The game had 238 million monthly active users in the second quarter.
One of the people responsible for making the tech happen at King is Steve Collins, CTO at the company. He has held that role since 2020, and he has more than 30 years of experience in entrepreneurship and leadership in the technology sector.
In his role at King, Collins leads the implementation of King’s technology strategy and the management of the shared technology team which is responsible for the centralized products, platforms and services that King has created to build and operate its live mobile games.
GamesBeat Next 2023
Join the GamesBeat community in San Francisco this October 23-24. You’ll hear from the brightest minds within the gaming industry on latest developments and their take on the future of gaming.
Before joining King, Collins was a partner at Frontline Ventures, a European VC Fund investing in technology start-ups, and prior to that was a founder and CTO of Havok, Kore and Swrve. Now he’s very focused on King’s cloud and machine learning strategies. (This is the sort of thing we’ll talk about at GamesBeat Next 2023 on October 23-24 in San Francisco; here’s a code for 25% off: deangbn2325)
Here’s an edited transcript of our interview.
GamesBeat: How long have you been at King now?
Steve Collins: I joined just before lockdown, so I started in 2020. My first act as CTO was to help us implement that transition. It was amazing to go through. A terrible time, obviously, but King was already a pretty digital-native organization. We adapted quickly to the new way of working.
In some ways it’s been interesting to be part of that transition. King now is very much in the hybrid mode. We allow each of our teams to figure out how they want to show up, and it’s been incredibly successful for us as an organization.
GamesBeat: Do you remember how big King was then? Has it grown significantly in the intervening years?
Collins: It probably has grown a certain amount, but we still–my team is the shared technology organization, where we invest heavily in a number of big initiatives, not the least of which is AI. We’re also going through a very large cloud transformation. We’re moving into the cloud as an organization, something we weren’t doing before I joined. We had migrated our data org to the cloud, but we’re in the process now of nearly completing that, which is quite incredible. We’ve gone from 100% of our games in our data centers to 100% in the cloud over a number of years. I’m looking forward to finishing that.
GamesBeat: Do you have your own engine? Is it a Unity shop?
Collins: All our live titles are powered by what we call the Fiction technology. It’s our own technology. In fact, we’ve even spun out game engines in the past. But Fiction is what we’re investing in now. We have an entire team on that. Imagine Unity, but designed specifically for mobile casual games. That’s its superpower. We’ve invested in creating what we think is the best mobile casual game engine available to us. It’s entirely grown from first principles. Originally it was developed to create some games from before Candy Crush, but now it’s very focused on Candy, Soda, Farm, and the rest of our live titles.
There’s a sort of common theme among those titles, which gives us the ability to focus on the best technology, the best engine development, the best renderer for that sort of game. It’s great for our artists to be able to take advantage of that.
GamesBeat: Do you ever revisit whether you want to have your own engine versus using publicly available technology?
Collins: My history before this is as a provider of middleware technology for console games. I’m very aware of the different arguments. For some of our newer games we’re exploring, games of a different nature, we’ve looked at Unity. We’ve been a partner with Unity for a number of years. That’s gone great for those types of games. But for our core live titles it makes no sense at all, because we have this great IP that we own, that we’ve invested in heavily with a team around it.
One of the things this particular technology gives us is incredible reach. We’re about making the world playful, and that means the whole world. It doesn’t matter what type of device you’re on. We support a huge range of iOS, Android, and other devices. We support Facebook and desktop and Kindle. The diversity of hardware platforms that we have support with a consistent player experience, that’s the technical challenge. Our games are long-lived. Candy is 10 years old and more. Farm and Soda are catching up on that. As an always-on service you have to bring the game along as the technology shifts.
We had some major shifts recently, like Apple deprecating OpenGL in favor of Metal. We had to do a full repurposing of the rendering engine to support that. No player had any issues with that. We were able to do that entirely transparently for hundreds of millions of players.
GamesBeat: Back in 2016 when I did a story, there were 2,000 Candy Crush levels. Now it’s 15,000. What accounts for that acceleration in the ability to launch new levels?
Collins: A couple of things underpin that. One is the voracious appetite of our players. That drives us. We have thousands of players who are at what we call “end of content.” They’ve played through all of our 15,000 levels. Which is utterly amazing and inspiring for us. That keeps the drive. We know that people come back every second week. They look forward to our new drops and new episodes that we hit every two weeks. They love that consistency over the last decade or more.
But I would say we’re getting better at it. We’ve learned over 10 years of delivering Candy levels. Our level designers are amazing. They’re taking advantage of more and more smart tooling. We’ve invested heavily in AI to support our level designers to be able to cope with that capacity and demand from our players. We’ve talked about that quite a bit. But that AI-driven level testing we do, and also AI-driven level design recommendation engines–as the designers work on levels, we can use the AI to essentially approximate what millions of players might behave like in a level and give feedback nearly live. “Hey, maybe you want to make this about 10% more difficult. Here are some things you might want to do.” But it’s ultimately in the hands of the designer, what to do in the end.
GamesBeat: Was there anything else that made you more productive and able to accelerate like that?
Collins: There’s always a few things. You learn as an organization. Our teams get better. Our organization design gets better. The tools we develop internally get better. We’ve made that workflow increasingly smooth. We’ve removed the blocks to level designers, artists, and engineers, just getting stuff done. We feel very strongly about looking at all your internal core loops as an organization. What can you do to remove the friction, remove the blocks from people being in flow? Trying to get to that flow state in all of our crafts is something we try to enable.
Some of the recent innovations in language models and things of that nature are very exciting for that. They bring even more tools to our disposal, specifically around content. The ability to consistently produce high-quality levels, and also look over our earlier levels–we’re constantly looking back and seeing if we can make the narrative more fun for players by tweaking and fine-tuning the level progression. That’s how we engage players. That’s how we make the world more playful, by providing that roller-coaster ride through Candy, Soda, and Farm levels.
GamesBeat: You acquired Peltarion last year and their AI and machine learning people. Can you give more details about that?
Collins: Running through a bit of the history, we started out looking at AI as a tool around 2016, when the first project kicked off to do some AI player development. Trying to create an AI that plays our games, particularly the levels, as if they were a certain type of player. We’ve invested in that for quite some time. Three or four years ago we realized it was a huge opportunity.
King is an amazingly data-driven organization and always has been. That’s in the DNA. We’re pretty well-known in the industry for our consistent use of A/B testing and data. When you run game services like this over many years, and when you have hundreds of millions of players, you rely so much on the telemetry and what the data tells you about how players enjoy your games. We reach out and talk to our players and get that qualitative feedback, but you can’t talk to all 238 million monthly active players, so you use data. That’s something we’ve built on top of.
A couple of years ago we made a decision to level up on the amount of AI. We saw all these areas in King where we could take advantage of AI, but we weren’t systematically doing it. We were doing it on more of a case-by-case basis. We felt it was a good opportunity to create a real AI strategy, which is what we’ve done. Part of that was looking outside to see other companies that are well-aligned in their skill sets with what we need to do. As I’m sure you know, competition for these sorts of people with those skills and capabilities is very high.
We were fortunate that we’d been working with Peltarion, which was a Swedish-based company a couple of blocks down the road from us in Stockholm, and we’d known them from the Swedish AI community at large. We had the opportunity to acquire them, which we did last year. They were a team of about 45. Their background was focused on bringing tools and capabilities to larger enterprises to help them in AI transformation. Identifying enterprises that weren’t heavy adopters of AI and helping them make that change. It was a good pattern match for the things we wanted to do here at King.
We now have Luka, the ex-CEO, as our head of AI and machine learning at King. He works directly with me. He’s leading the charge in transforming King into an AI-driven organization.
GamesBeat: The boom in generative AI came after that. How has that changed things?
Collins: If anything, it reinforced that it was a great time to bring that capability to bear. Having a non-organic lift in the resources we had available for AI. Now we have a team of more than 50 exclusively focused on AI capability and AI tooling development. There are more than 100 people in the organization who are working with AI throughout our game teams. We’ve significantly leveled up in our investment. What happened in 2022 and 2023 with ChatGPT and so on, the pace of change has accelerated enormously. We’re well-positioned given the team we have. These are people with decades of experience in AI. They were very familiar with a lot of the technology – attention mechanisms with transformers, the paper from Google in 2016, all that stuff. We’re able to take advantage of it.
We’re looking at it in two ways. Generative AI, particularly large language models, like any tech organization we think it’s going to have a real impact on how we work. The tools are generally available to everybody at King. We’ve invested in that. We’re in a learning mode at the moment. We’ve been running a number of pilot studies tuned to our use cases. Anecdotally, things like Copilot are great tools for accelerating coding. That’s something we’re looking at. We’ve run some studies internally and replicated some of that, which is exciting. You have to test for yourself as well.
There’s no question this is interesting. There are still some issues, things that need to be worked out around ownership and copyright and all those things. We’re very much in learning mode before we dive in. But I expect this is going to be a big transformation in how software engineering is done from here on.
GamesBeat: Things like being able to trace ownership of what’s fed into the model, that seems to be an issue for everyone.
Collins: It’s particularly true when you get into art and content generation. You see the conversations underway in that space around ownership and how these models are trained. That’s important, because people who create content should be rewarded when that content is used. We feel strongly about that. Having said that, when you look at generative in its broadest sense, with large language models in that category, then it’s much broader than that. You have to understand the provenance of the model, who owns it, who governs it, what statements can be made about safe and fair use of that model. What does that mean for the content you generate using that model? That’s exactly the conversation the industry and the world are having, trying to figure it out.
GamesBeat: As far as the impact in the organization, there was a report by Bain recently saying they expected that in games, a lot of the impact would fall on production. That’s a contrast to the early impression that the impact might be on pre-production, things like concept art. There’s some analysis being put toward what part of the organization changes the most. Have you been able to figure out some of that?
Collins: I don’t know if anyone has it figured out yet. It’s moving so quickly. It’s exciting and also challenging. The position we’re taking on it is to learn fast. Explore, experiment, learn as much as we can and keep up with the pace of change. There’s no question that a large number of roles–it’s not that they’re going to be impacted. They’ll be energized by this capability. These new tools are super.
Thinking of an example, looking at our data, how do we explore our data? How do we produce insights from that data? It’s always a challenge. You have a lot of data, but how much of that is truly driving your business? We think language models, or the multimodal models of the future, are going to be great at sifting through large volumes of information and distilling it down into something more consumable. That could be one of the routes that we use to get to know our players more, get to know more segments of players who maybe have a very specific experience of our games. We can get better at understanding what more and more of our players want and not constantly be thinking about the mean player, because there’s no such thing.
Our entire mantra is to use the data to understand your player more so you can make the game more engaging and make the world more playful. This is hopefully a tool we can bring to bear on that. We have a lot of data scientists in our organization clamoring to get access to this capability, because it’s going to make them more effective at what they do.
GamesBeat: There are some interesting uses that can be targeted at consumers, too. The Gran Turismo people created an AI racing agent that competes against some of the best human players in the world. It’s a competitive driver racing alongside them that can beat some of the best humans. Can that be more generally applied to almost any game, even yours? You could have a competition between players and AI.
Collins: The whole genre of games that benefit from the idea of competition–whether it’s multiplayer or a single-player game with non-player characters, going all the way back to day zero in gaming, we’ve had the concept. Stretching the analogy a bit, the ghosts in Pac-Man had rules that made them behave like an entity with its own motivations. That was compelling at the time.
Games like the racing games, this rubber-band capability has always been important, to give the impression that you’re racing against a driver as skilled as you. Now you don’t have to have that. Instead of just giving the illusion of skills to another driver, you can have a driver that’s been trained using reinforcement learning or whatever technique to really behave like a human player. That’s great, because that means when there’s no one around in your matchmaking sessions, you can have compelling races against non-player characters.
It doesn’t get rid of the fun of playing against your friends, though. That’s the one thing you learn. Just because Deep Blue solved chess doesn’t mean no one wants to play a human anymore. You don’t spend all your time just playing computer opponents. But the ability to enrich worlds with NPCs that are going to be more and more sophisticated is really interesting. I don’t know if you’ve come across a paper from Stanford called Smallville. There are companies starting to do this. There’s the idea of connecting a system like ChatGPT into the NPCs in something like a Sims game, giving them personalities and roles and letting it evolve forward to see what comes out of that. There’s something fun in there for that type of game.
GamesBeat: Maybe some of that isn’t necessarily what you’d expect to see in Candy Crush.
Collins: Well, we do the opposite. We actually create smart players, or dumb players, or competitive players, or non-competitive players. We try to simulate what a player would do in a situation so we can test levels. We’re doing it in a different way, but we create all these AI players. Increasingly our goal is to try and create players that have all the variations and all the diversity that our player base has. That’s really hard to do, and generally we may never get there. But the more we can do that, the more we can test levels in advance. We don’t have to react. We don’t have to change and tweak when things are live. We can be more purposeful in the narrative or emotional journey that we’re trying to create and bring players on with our levels.
GamesBeat: What are some near term things you expect to happen, either for King of the game industry as a whole?
Collins: I do think that some of these technologies are going to give rise to different types of gaming. I don’t necessarily think it will disrupt the entire game industry, but it’s going to allow us to explore different types of games that are more adaptive to players. I don’t know what that looks like, and there’s interesting tension between design and agency and, let’s say, a simulation that’s free to do what it wants to do. Developers like ourselves, we’re trying to create a story, a narrative, even if it’s expressed in different ways. We’re trying to intentionally bring players on an emotional roller-coaster. That’s what we do. We want to engage and let people have fun.
We’ll get better at doing that with tools that let us understand players more. We’re going to get better at doing that because we’ll have more sophisticated tools that allow us to test scenarios more effectively. Hopefully as an organization we’ll level up because we have these great cognitive agents assisting every craft at King be able to do more. That will be incredible. Beyond that, what it means for the future of the industry, it’s hard to say. If you asked me that 18 months ago I’d probably have had a completely different answer than what I give today. That’s the pace of change.
GamesBeat: One thing I wasn’t sure about was where the cost of the large-language models (LLMs) goes. People are going to get more ambitious and that can drive costs up. You would expect, though, with the normal advances in computing, that they would figure out how to build these in a less expensive way. I was starting to hear things like, every time a human sends a query back to AI in a game there’s a cost associated with that. Do that too much, talk too much with a smart AI character, it could run up a pretty big bill.
Collins: Absolutely. As the demand increases, the technology will advance. If you look at models today, they’re experiencing their own form of Moore’s Law. Consistently, for the past six or seven years, the models have doubled in parameter count. There’s no sign of that ending any time soon. There’s another interesting result from Google last year, or early list year, what’s called the Chinchilla paper. It postulated that, given all the data in the world, how much more data could we feed into these models and they would still get better? Or some variation of better. The result is we’re about an order of magnitude away from even getting to that capping-out point.
These models are going to get more sophisticated. Multimodality is interesting. We have Google’s Gemini coming later this year, where it’s trying to bring video and audio and text training into one place. No idea what that looks like, but it sounds very interesting. Increasingly, as you get into consumer-based applications where these things will be called at scale, something will have to be done about cost, because it would be too expensive today.
What’s really interesting to me is neural radiance fields, and the idea of using neural networks to approximate physical worlds. Encoding light fields and how light interacts in the real world essentially as a neural network. It’s learning how to render. Nvidia is doing this to a certain extent right now. There are some interesting extrapolations of that, where you can give it a description of a world and then it will create it for you, create an entire 3D universe for you to explore.
DLSS is a really small subset of that. You can imagine that extrapolated out. Rather than just being a rendering accelerator, this is an idea generator from first principles. Where that goes in the future–the idea of an automatic idea generator is quite something. But I think we’re a long way off from that yet.
GamesBeat’s creed when covering the game industry is “where passion meets business.” What does this mean? We want to tell you how the news matters to you — not just as a decision-maker at a game studio, but also as a fan of games. Whether you read our articles, listen to our podcasts, or watch our videos, GamesBeat will help you learn about the industry and enjoy engaging with it. Discover our Briefings.