"People will use machine learning for personalised experiences beyond NPCs"
Brandon Da Silva of ArenaX Labs talks deploying AI agents in multiplayer games, web3 competitive gaming and more.
Welcome to AI Gamechangers, the weekly newsletter bringing you conversations with games industry leaders about the practical role of AI.
We’re amassing a fantastic archive of Q&As with people from all sectors of the games world, from development to user acquisition, from monetisation to QA testing. Check out the archive to read them all for free.
In today’s edition, we meet Brandon Da Silva, co-founder, CEO and CTO of ArenaX Labs, which is doing great things with bots for multiplayer games - but that description barely scratches the surface because with years of experience in machine learning, Brandon and the team are investigating player behaviour cloning, web3 competitions, and more.
We have some fascinating Q&As planned for the months ahead, so please remember to subscribe. Next week, we’ll meet ethical recruiter and Women in AI member Anna Moss to discuss the job market and the impact of AI on recruitment.
Scroll to the end for the latest AI games news, including the Nintendo Switch 2, OpenAI losses, TikTok and Nvidia, BioShock developer Ken Levine, and more.
Brandon Da Silva, ArenaX Labs
Meet Brandon Da Silva, the co-founder of ArenaX Labs, serving as its CEO and CTO. ArenaX Labs is building technologies at the intersection of multiplayer games, web3 and AI.
Its main products are ARC, an AI-powered SDK that helps with monetisation, multiplayer bots and more; AI Arena, a game where your AI characters fight against those trained by other players; and SAI, a machine learning competition designed to challenge scientists and researchers.
We spoke with Brandon in late 2024 for this piece, but just today, the team announced a partnership with Eliza Labs to “accelerate AI agent innovation towards AGI”. Details are in the news section below.
Top takeaways from this conversation:
ArenaX has developed machine learning and behavioural cloning to create personalised AI agents that can mimic a player's individual playstyle in their fighting game, AI Arena.
The firm’s ARC Agents technology provides a solution to the player liquidity challenge in multiplayer games by enabling studios to deploy AI agents when needed that realistically emulate human players.
ArenaX is integrating web3 technologies like tokenised AI models into their games, enabling competitive ownership and trading (though they are balancing this with traditional gaming experiences).
Beyond game development, ArenaX is also leveraging game environments as a platform for advancing machine learning research through their SAI initiative, which hosts coding-based competitions for researchers.
AI Gamechangers: Please tell us a little bit about yourself and your business.
Brandon Da Silva: It might sound strange how I jumped into gaming. Prior to this, I used to be a quant, building algorithmic trading strategies for the market. I would use machine learning very heavily. I did that for a number of years before I eventually made the leap and started a company about machine learning and gaming.
While building these algorithmic trading strategies with machine learning, oftentimes I would test out my algorithms on games in my spare time. I'd make these algorithms, and I would see if they worked on games first before bringing them to the markets!
“The core loop of the experience: continuously improving your AI until you're happy with it. You can enter it into ranked battle, where you compete against the AIs that other people have trained.”
Brandon Da Silva
Eventually, I started building my own game environments. Because I thought, “My AIs are able to do all this stuff; let me build my own with these unique challenges.” I would make agents for games. I did that for a number of years.
Then, in late 2020, I had a wild idea of building personalised agents for a game called AI Arena. It's a fighting game, and the idea was that each person could train their own personalised agent to play like them. It watches how you play, and it picks up on your play style. It's something that we call “behavioural cloning”, and it's basically like a digital clone of yourself that can play the game for you autonomously. That was the idea that kickstarted this whole thing.
We now do a lot of other things, but that was how I went from quant to building AIs for games!
What’s the user experience of playing the game? You’ve mentioned behavioural AI, but what kind of game is it?
The underlying genre in which we're applying this machine learning is a platform fighting game – something akin to Super Smash Bros.
The idea is to teach an AI to play like you, which involves a few steps. The first step we call “data collection”. This is where you are actually playing it – you get your controller and your keyboard, and you play it, and you show it your moves.
The next step, once we collect all that data, is that the user can tell the AI how it wants it to learn from that data. I collected this data, now what do I want my agent to do with this data; how do I want it to learn? For example, you can say, “I want you to only focus on the relative distance to the opponent in this training session.” Maybe in another training session, you're like, “I only want you to focus on the active projectiles on the screen.” Maybe you want to learn how to shield them.
And then, after you learn something, we allow you to visualise what the AI is “thinking”. We call it the “AI inspector” – you can dive in and see what it's going to do in every single scenario. Really, it's a probability distribution. Maybe in one scenario, there's a 70% probability I'm going to punch, a 10% probability to do a special move, and so on. You can see exactly what it's doing. Maybe you don't want it to do this in this scenario, so let's hop back into data collection and collect more data.
That's the core loop of the experience: continuously improving your AI until you're happy with it. Once in a while, if you want to, you can test it out. Send it into “simulation mode”, where you can simulate how it actually performs against bots.
Finally, if you're feeling up for it, you can enter it into ranked battle, where you compete against the AIs that other people have trained.
AI is such a popular term at the moment, covering many things and meaning different things to different people. What do AI and machine learning mean to you, and how are you using them here? In layman’s terms, what is your model doing?
This is funny because when I used to build quant strategies, the higher-ups running strategies would ask, “What the heck is AI? What is machine learning? What is the difference?!”
My definition of AI is that it is some intelligence that’s embedded into a machine that can perform something autonomously.
The question then arises, “How does it acquire this intelligence?” There are two dominant methodologies. One of them is called “expert systems” or “heuristics” – the old-school way, where it's essentially a bunch of statements, a rules-based approach. Back in the day, if you’re programming Deep Blue to play chess, for example, you get a bunch of experts into a room, and they tell you the agent should do this or that in a specific scenario. You basically make a long rules-based system. That takes a lot of time. It requires experts, and it doesn't really scale well.
The other approach to acquiring the intelligence is machine learning. That’s what we focus on. It’s basically learning through data. There are no experts. (Well, perhaps there could be to shape certain things!) But, to a large extent, you just need to know the structure of the environment – then collect the data, and learn from that data.
“We didn't actually want to license out our tech. But people just kept asking us. We thought our game was this niche thing that nerds would be interested in! But it turns out that it solves a really big problem, which is player liquidity for online matchmaking.”
Brandon Da Silva
We focus on machine learning, and then specifically, we focus on two subcategories. We map from states to actions, and we learn this in two ways: one is reinforcement learning, which is learning through a reward mechanism, and the other is through imitation learning, which is what I described with AI Arena, where we're trying to copy what the expert does. This is the behaviour cloning, but the end result is the same.
What is the game state, and what is the action I'm going to take, given this game state? This differs very much from what you see now in the news. It’s all LLMs and generative AI there, people making assets – we're completely different. We focus on in-game actions.
Was it easy to raise investment for this business? And how did you come together and form your leadership team?
We raised in the summer of 2021, after we built an initial prototype for AI Arena. That was an interesting process – I've never raised money before, and we had a few practice runs, some VCs, and eventually we nailed it down, and it wasn't super difficult initially. We did a follow-up round, which was harder to raise, given the market. It was a deep bear market at the time! But we still were able to raise, and in two rounds, we raised $11 million in total.
The leadership team is myself (I’m the CEO and CTO). [There are] two co-founders; one is Wei [Xie, COO], and the other one is Dylan [Pereira, CDO]. Dylan leads the art department. He's actually my cousin, so that was an easy connect! Wei used to work with me at the investment firm; he was my boss, the guy who hired me in 2016. I've always had an excellent working relationship with Wei; he's a close friend. Both of my co-founders are brilliant. It was a no-brainer to put us together because we had the necessary skill sets.
And where are you in your roadmap right now? How complete is the game technology you just described?
We've been testing and iterating on AI Arena for years. We operate on the web3 side of things, too, and we have launched on web3 , which was a big milestone for us.
We've been running the game on testnet for a long time. In September, we did our first test on mainnet to see how the reward distribution would work. Remember, ours is a super-competitive PvP game, so the reward mechanism is important; it’s important to reward players so they continue to improve the agents to make the AI better and better. With any game, there are balance issues, and we're continually trying to improve the balance of the game. But overall, players are happy. We're really focused on the competitive tier and intentionally designed it like that – we designed AI Arena to be a proof of concept for the tech.
As well as AI Arena, you also have more general tech for AI agents. Can you tell us a little more about the wider scope of your business?
Absolutely. We have a product that we call ARC Agents.
We didn't actually want to license out our tech to anybody, because we were thinking, “This is sick! We want to keep it for ourselves!” But people just kept asking us to license it. We were surprised. We thought our game was this niche thing that nerds would be interested in! But it turns out that it solves a really big problem, which is player liquidity, specifically for online matchmaking.
“If you're getting matched up against a super-intelligent AI and getting smoked every time, it's not a good experience. Being able to properly transfer a human skill onto an agent is important. Humans come with a variety of different skill sets.”
Brandon Da Silva
Now they can actually train agents that act like humans, which is very important – if you’ve played a lot of online games, you know when you’re facing a bot, right? It doesn't perform well. A lot of new games, especially indie games, as they're starting up, they have really long wait times. The queues to join – it’s just not a good experience for new people. That's really tough for [the developers] because they need it to be a good experience in order for it to take off. It's a chicken and egg problem.
Many studios were coming to us to help solve this player liquidity issue. We were surprised, but eventually, we had so many that we said, “Okay, you know what? We're going to make this into a product.” Since then, we've been expanding the product offering for ARC, and we have a lot of really exciting stuff we're working on.
How about the issue of skill level? Matchmaking is an art and a science. How do you ensure that the AI bots that are deployed will be a good fit for the game?
If I take a step back to when we were building AI Arena, people would ask me why we don't just use reinforcement learning to make super-intelligent AIs. And I was very adamant about the fact that I wanted it to be a representation of the player! I wanted it to be a player in digital form, essentially.
It turns out this is great for solving the [skill level] thing. What you can do now is make an Elo ladder of all these models.
Maybe you're way better at playing the game than me, but you know what? My agent is still valuable because there are a lot of players that play like me, and it can match up against them. So with ARC, we're structuring it in a way where we have competitions, where we're incentivising people to contribute agents to each slot in the ladder.
If you're getting matched up against a super-intelligent AI and getting smoked every time, you know it's not a good experience. I think being able to properly transfer a human skill onto an agent is super important. Humans come with a variety of different skill sets, and we want to be able to hit all of them.
You’re working with AI, multiplayer gaming and web3. Are your players embracing these cutting-edge gaming technologies?
For better or for worse, we mostly started out getting exposed to the web3 crowd first. They obviously had no problem with the technology – we put that at the forefront.
We were inspired by a company called Numerai. They basically run a stock market prediction competition, where people will submit their predictions. If you have a better model, you submit better predictions, and you get rewarded accordingly. I remember in 2020, I was looking at this, thinking, “That's so interesting. What if each person can own their own AI model? And it competes against each other, so you can see which AI models are better?” That's why I started thinking about the whole web3 angle: because you can tokenise it and have ownership of it, and if it's valuable and does really well in the competition, you can get compensated for it.
That's where the whole thing started. I was enamoured with this idea of owning this AI. At that time, I hadn't heard anybody even think about something like this. I was looking to invest in something like that, but nobody did it, so I definitely wanted to build it!
How do you scale up from there? There are a lot of people playing multiplayer games who aren't interested in web3 and may be a bit hesitant about AI as well. What's the education journey that you need to take them on?
If I'm being fully transparent, it's kind of steep! The web2 crowd – and they could be fully justified in this! – a lot of times don't want to have anything to do with web3. I'm not here to say whether that's right or wrong. Everyone has their own preferences. But there are certain people that will be impossible to convince, no matter what. There's no getting around that, no matter how much education you do.
What we're finding interesting, though, is that we have activations with web2, and there are certain people that, if we start to give out rewards in fiat currency, for example, so they’re earning, will be hungry and want to earn more. So we can say, “Hey, you can come over to this web3 side, and you can earn substantially more!” Their competitive nature makes them say, “Oh, yeah! And I'm better than these web3 gamers anyway!” So that's how we found the most success in transitioning people from web2 over to web3.
Some people on the web2 side just like playing the game. They just want custom skins; they want to personalise their character, and that's all they care about, which is totally cool. But certainly, there’s a subset of very competitive people that we're moving over.
I do want to note that we're intentionally keeping the web3 side hyper-competitive. There are only around 400 slots in terms of the NFTs that we give out to compete on the web3 side. We're keeping it small because we essentially want it to be the best of the best. We want it to be a spectacle.
AI is still the hottest topic in town. What’s your take on where the games market is going? If we do this interview again in five years’ time, what impact will AI have had on the multiplayer game space?
What I want to say is, “Everyone's using ARC!”
I think the world's going to be very different. I don't know if this will play out, but my personal belief is that ARC or something similar is going to be populating a lot of lobbies and multiplayer games. People will be using machine learning for a lot more personalised experiences beyond just NPCs.
“We're using games as exciting problems to solve for machine learning researchers. We're abstracting away some of the boring stuff that they're doing research on and re-framing it as a game so it's more exciting and interactive.”
Brandon Da Silva
The low-hanging fruit now is OpenAI; everyone's just using API calls to get some intelligent agents, right? I think it's going to extend beyond that. It’s less interesting to me how everyone talks about generative models speeding up the game development process. “Indie devs will be able to make AAA titles [with AI],” and that type of stuff. I used to make generative models, before it was cool! But now, I'm specifically focused on things like imitation learning and reinforcement learning.
One of the exciting things we're working on is actually understanding the nuances of how each individual player plays, so that we can personalise the experience to them. Imagine if you know how they move, probabilistically; you can put things along their way. You'd know where they're likely to go, so you can put things there to either make it easier or harder. And you can imagine how much personalisation you can get out of this once we have really good predictive models! I think that will be a big part of games going forward.
Is there anything else you're working on or that you’ve seen that you think we should know about?
There's one more thing. We’ve talked mostly about core integration into games, using machine learning for games. There's something else that we're doing! It's called SAI. It's a machine learning research competition.
We're using games for machine learning – it’s the flip side. We're using games as exciting problems to solve for machine learning researchers, and we're essentially abstracting away some of the boring stuff that they're doing research on and re-framing it as a game so that it's just more exciting and interactive.
It’s focused on engineers and researchers. We basically provide the Python library. You install it, and you have access to all of our environments, and you train these agents with whatever algorithm you want, and then you upload it onto our platform, and it instantly starts competing against the other agents.
Further down the rabbit hole
The latest must-read news stories, in a handy list:
Speculation about the Nintendo Switch 2 is rife, fuelled by a patent that was applied for in 2023 but just published this week (“Systems and Methods for Machine Learned Image Conversion”). It shows how Nintendo could use AI-upscaling to make games look better.
ChatGPT creator OpenAI looks set to incur losses of $5 billion in 2024, compared to estimated revenue of $3.7 billion, according to reports. Boss Sam Altman posted this week that the recently introduced Pro tier is currently losing money.
TikTok’s parent company, ByteDance, told suppliers it planned to spend $7 billion to access Nvidia AI chips outside of China this year, potentially making it one of Nvidia’s biggest global customers.
Speaking of Nvidia, its Taiwan manufacturing partner, Hon Hai (which is also the world’s largest maker of Apple iPhones), saw growth in its revenue and company value because of demand for AI servers.
Hollywood’s Creative Artists Agency is collaborating with YouTube on a programme to help actors and other talent find and take down AI-generated deepfakes on the video platform.
Aethir has announced a $40 million initiative with partners including Beam. The core business model is to acquire unused GPUs from companies with excess capacity and make them available to web3 gaming and AI startups, creating a decentralised marketplace for computing resources.
Legendary BioShock developer Ken Levine had pragmatic things to say about generative AI in an interview at the end of last year. In a conversation with GamesIndustry.biz, Levine spoke about how AI is great for analytics and organising the bug database but bad for creating stories and persistent worlds, so he won’t be using it for that.
Tickets are still available (just) for this month’s big London games industry conference, PG Connects (20-21 January). One of the track rooms features the topics AI Advances (sponsored by Layer AI) and Practical AI (sponsored by Filuta AI), plus there are three other stages and 2,500+ professionals to network with.
ArenaX Labs, creators of ARC, and Eliza Labs, announced a partnership this week focused on AI agent development. The agreement will combine ARC's behavioural learning capabilities with Eliza's agent development framework, focusing on games, simulations, and robotics applications. It’ll roll out in Q1 this year, with a long-term goal of accelerating innovation towards Artificial General Intelligence.