Hello! Welcome to the latest edition of AI Gamechangers. This is your weekly “deep dive” newsletter where we sit down with games industry leaders and quiz them. The result is this conversational Q&A newsletter, available to read for free, giving insight into what’s afoot on the front line of AI in games.
This week we’re joined by Anton Bernstein, co-founder and CEO of Pocket Worlds, the studio behind social creation platform Highrise. Anton walks us through Pocket Labs, his applied research lab, and the tools it’s building, including Rosie: a conversational assistant that enables a largely non-coding community to build games just by describing what they want. We dig into where AI actually helps creators, where the community draws the line, and why human judgment and taste matter more than ever.
As ever, scroll to the end for our round-up of AI news, links and tools from around the web. And browse the archive, it’s great.
Anton Bernstein, Pocket Worlds
Meet Anton Bernstein, co-founder and CEO of Pocket Worlds, the studio behind Highrise and Everskies. A former venture investor who went on to build a giant travel company, Anton bootstrapped Pocket Worlds out of a love of gaming, creativity and community.
Highrise is a social virtual world where people make avatars, connect and build, with an audience that’s around 70% female. This year, the team launched Pocket Labs, an applied research lab putting AI to work across its ecosystem. We explore its creator tools, its trust and safety approach, and where Anton sees self-expression heading.
Top takeaways from this conversation:
Pocket Labs builds AI tools, using frontier models, from a clear belief that AI is a democratising force for creators rather than a replacement for them.
Its flagship tool, Rosie, is essentially “Claude Code for Highrise Studio”: a conversational assistant that helps a largely non-coding community build games and UI just by describing what they want.
AI also powers two less obvious wins: a trust and safety system, and a post-processing model that adds dynamic lighting to human-made 2D rooms.
The community embraces AI for coding but draws a firm line on art. Fully automated work is frowned upon, because Pocket Worlds’ appeal is human self-expression.
AI Gamechangers: To break the ice, can you tell us a little about yourself and your own journey? You’ve been an investor, you’re into the web3 gaming scene… what brought you to Pocket Worlds and AI?
Anton Bernstein: At the core, what originally brought me to Pocket Worlds is that I grew up gaming. I’m a child of the Diablo II and StarCraft multiplayer era, so I’d always been fond of it, and the first things I ever programmed were mods for games.
It was a bit of a circuitous route to get here, but at the end of the day, I’ve always loved games. I always thought they were the most social experiences online, and I loved creating in games. Pocket Worlds wasn’t originally started to do that, but mostly because of my interest in the combination of creation and very social gaming, we’ve evolved into a creator platform where people build games in a multiplayer environment and connect with each other.
The way I got there: I’m Russian originally. I moved to New Jersey when I was five, and I went to school for business, so I went into VC. I was a venture investor at a company called Insight, and another called Redpoint, focused on tech, media and gaming. Then I started an e-commerce company in Australia, which is now one of the largest travel companies there, called Luxury Escapes. After that, we sold our French satellite business, and that’s when I moved to San Francisco. I thought, “I have some money now, I can bootstrap whatever I want!” So I decided to bootstrap Pocket Worlds, in a field I’m interested in: gaming, creativity and creation.
We bootstrapped for the first five years, just exploring and experimenting. I’d never worked in games before, so this was my first attempt at working in games and building in games – a lot of trial and error, a lot of learning. But we ended up launching a product called Highrise, which is our flagship platform. It’s a virtual world platform that started almost as a shell to connect with other people: you make an avatar and an identity, you exist in this world. It’s like a little city, where you meet other people, connect with them and talk to them. It’s really about delivering that sense of belonging and community.
Over time (in large part because we’d never really built games before), we leaned on the creator community to build the experiences inside the game and a lot of the content. So we leaned more and more heavily into building creator tools, which is what my co-founder and I are good at and native to. We’re more technologists, more tool builders. We’ve also built the skill of making, I think, pretty good games, but at the core, we’re still more of a technology company.
“If expressing your creativity means building your environment, you can do that in an accessible way. You don’t necessarily need a desktop; you don’t necessarily need Unity; you can communicate with our tools and fully express yourself”
Anton Bernstein
Then we acquired a business called Everskies in 2021, which is also a very social product.
The other area we’ve decided to focus on – partly because we got there naturally, and then decided to really lean in – is a different sub-genre of gaming than most companies. It’s more about self-expression and creativity than about killing or even explicit objectives. Our audience tends to be around 70% female across both platforms, and I think we’ve developed a bit of expertise around fashion and style, around how people express themselves and how they can be creative in that way. One thing I like to say is that we’re more similar to a dollhouse than a sandbox, if you compare us to Roblox.
Let’s talk about your AI plans. You launched Pocket Labs this year. What was the journey from the product you had to what you’ve just introduced?
There are two things there. One is that, again, I see us more as a technology company than a gaming company in a lot of ways. So when a new technology comes along, we explore it heavily. Even when crypto came out, we explored it – it was new, and we felt compelled to check it out. It was exciting. The whole decentralised ownership thing never really proved meaningful in gaming, which in hindsight makes sense.
But as AI came onto the scene and became actually useful to creation, to games, to development, we launched Pocket Labs, an applied research lab. We take technology that companies like Anthropic and OpenAI – and some of the open-source solutions – are building, and we apply it in our ecosystem.
It’s really born from a combination of elements. One, we’re technologists, and we’re interested in trying new technologies and seeing how they might apply to the problems we’re solving for our customers. And two, we’re pretty creator-focused, and AI is such a democratising force for creators. It enables creators to scale themselves and to really realise their vision – manifest their vision. So we leaned more and more heavily into: what tools can we give creators, both in-house, like our own developers, and all the creators on our platform who are building games, experiences and clothing?
Some people worry that AI is replacing human creativity, but it sounds like you celebrate it. Tell us more about how you see AI and human creativity working together – what are some examples of what people are doing with this?
There’s a lot. Foundationally, that is what I believe. More than anything right now, what’s valuable from a human is judgment and taste, because a lot of the rote stuff is getting solved, or at least is easier to produce.
When it comes to development, for example, all of a sudden, your judgment and taste of what good code looks like, what good architecture looks like, become more valuable. The product manager also becomes a lot more valuable because they’re thinking: what does a good product actually look like? How does it actually serve its customers? The details you can start to accelerate through AI. I think that’s true across all of creation, so I 100% believe it’s an enabling force for humans, for human creativity and ingenuity.
I also think (and it’s hard for me to see this change in the short to medium run) that the human in the loop is really important, because that judgment, that taste, that ability to connect it to so many different contexts is crucial. That’s the difference between an okay product and a great product.
The way we’re doing it in practice: we have a few different technologies that use AI. We have one called Rosie. Imagine Highrise Studio, which is our answer to Roblox Studio: you can build games and experiences, and it’s built on top of Unity. Rosie is built on top of Highrise Studio, and the best analogy is that it’s kind of like Claude Code, but for Unity specifically, and for Highrise Studio more specifically. It’s trained on our documentation. Our scripting language is Lua, so it’s trained on our Lua scripting, and it’s also trained to use the Unity primitives. It can go in and manipulate the environments you’re setting up for a game or experience – move things around, turn things into anchors, give them different definitions or parameters, apply textures, all that kind of stuff – and it can generate scripts, and especially UI, based on the training we’ve done.
This enables our creators. Let’s say they want to make a game – an obstacle course, for example – and they want a daily login menu. They do no code; they just talk to Rosie. They say, “Hey, I want a daily login menu, I want it to look like this,” and Rosie does it, and now they have it. Even for generating the obstacle course, they can say, “I want to make an obstacle course, I want it to do this and that.” Someone internally, in a hackathon recently, made a Mario Kart-style game, pretty much all just talking to Rosie: 95% Rosie, and maybe 5% some tweaks.
So I think that’s a massively enabling force. Taking one step back: our audience is mostly 18 to 35, and it’s 70% to 80% women. A lot of our people haven’t learned computer science, and it wasn’t even a hobby before. As they pick up Studio and start to run with Rosie, they’re able to build things I don’t think they could have imagined building. We’re seeing these worlds emerge that have tons of UI and innovative play, and that cater to a different segment than the typical Roblox player. I think that’s pretty special.
With Rosie, it sounds like the human is working in partnership and can get it to do as much or as little as you want? Could you do 90% of the work and get Rosie to do 10% of the tweaks?
In practice, you’re working in what’s called an IDE (integrated development environment). In our case, it’s the Unity IDE, and you have a side panel that is Rosie. You’re talking to Rosie on the side, and Rosie is doing stuff in your IDE. You’ll say, “Go try this.” And you see it, and then maybe you say, “I don’t like it.” You can either talk to Rosie to make changes, or you can click in and see all the scripts it wrote, or see how it changed the parameters, and then change them yourself if you want.
You can even ask Rosie, “How do I change it?” and Rosie will say, “You can change it by going here, here and here.” Then you can do the changes.
I think this last point is interesting across the board: besides Rosie actually building things, just as interesting is that Rosie is a phenomenal teacher. We’ve found that, like with most chat apps, it all depends on how you use it – you could use it to cheat, or you could use it to learn. Our best creators are using it to learn. They’re asking Rosie, “How do I do this? You just did that thing – how did you do it?!” And Rosie explains it. So next time, as a creator, you can always do it yourself.
The Highrise scripting language is Lua, but Unity uses C#. How granular can you go? Can you edit the actual Unity work as well as your scripting code?
We think of it this way. C# is used to build the engine, but Lua is used to let everyone play with it.
Unity and C# are heavy-duty, strict, and they require a slow compilation step, plus, letting users run raw C# on our servers would create massive security risks.
“All of a sudden, your judgment and taste of what good code looks like, what good architecture looks like, become more valuable. The details you can start to accelerate through AI. I believe it’s an enabling force for humans, for human creativity and ingenuity”
Anton Bernstein
We chose Lua because it’s a lightweight language that lives safely in a sandbox. It’s incredibly easy for hobbyists to learn, it can’t crash the servers, and because there’s no compile step, changes happen instantly. It gives creators deep control, but keeps Highrise fast, safe, and accessible.
You mentioned a number of different technologies you’ve been trying. With Rosie, which foundation model are you actually using? What’s the actual LLM doing the interpreting behind the scenes?
Anton Bernstein: Rosie is a harness. You can bring your own model. With our creators, we either have a pay-as-you-go model, or for some of our top creators, we just pay for a subscription to either OpenAI or Anthropic – a monthly one, so they’re on our team account and can just use it. So we have both business models. But you can bring your own model, because Rosie is kind of like Claude Code, and Claude Code is what’s called a harness. You can plug in whatever model you prefer; you can even use a free open-source model if you want.
What other AI technologies have you got going on?
There are a few different areas. One, which is pretty straightforward, is in the trust, safety and player experience sphere. We take trust and safety very seriously – it’s very important to us. Literally, we’re running a little city, and you can’t have a little city without a police and fire department. So we provide that, and what AI enables is early response: we can treat every report that comes in, read it, understand the context, see what’s going on, escalate it to a human trust and safety moderator if it needs escalating, and remove it from the pool if it’s erroneous.
That’s been really helpful. It makes us maybe three to four times more efficient at flagging, which means more trust and safety issues get addressed, and our human team can deal with the more severe issues.
We built all of that ourselves, instead of using a third-party solution, because we’re a technology company. We use open-source and third-party models to process everything. What’s really neat is that, because every piece of text goes into our models, we combine that with real-time translation. So those models not only assess the threat level of a piece of content, they also translate it in real time into about 20 languages. We get the added benefit of real-time translation.
Another area we just rolled out: we’re fundamentally a 2D-ish platform, though now you can do 3D, so you can have 2D worlds or 3D worlds, up to you. Our fundamental aesthetic is isometric 2D; you look at it from a 45-degree angle. We have 20,000-plus pieces of furniture; our assets are all designed by humans in 2D, so you can’t rotate them, and that makes it difficult to apply great lighting. So we built a post-processing AI model that takes the room and applies beautiful lighting to it. We just couldn’t do that before without the power of AI. Now every creator, after they make a room, can turn on the post-processing lighting. It bakes, sends up to our servers, runs a model on it, and applies the kind of lighting they want – dark or light, intense or soft – and it gives a really cool effect.
It’s a neat combination of AI and human creativity, because the human paints the canvas, and then the AI post-processes it, and it looks really neat.
There is a mix of responses from gamers and creators about AI. Some love it and see the value; some make it their whole identity to push back against it. What sense do you get from the community about your AI tools?
I think it depends. When it comes to the tools for building experiences and worlds, from what I can tell, it’s been a general embrace. If you’re using AI to build a game or experience – the mechanics, the scripting, the coding – there’s virtually zero pushback from anyone about using AI to code, which is funny (and maybe kind of sad for engineers).
On the art side, it’s a lot more delicate. We’re a creative platform focused on self-expression, with a team of 25 artists, and we’re really focused on the craft of the art. The delicate balance is this: sometimes an artist might produce something, then twist it in some way or touch it up with AI. Some people are like, “That totally makes sense, that’s the tool, that’s how you use it.” Some are like, “No, that kills the soul of the piece.”
One thing I know people definitely hate (and we actively choose not to do it) is zero-to-100 AI: a situation where the AI comes up with the concept, draws the piece, produces it, and puts it in the app. Nobody wants that, and I totally understand. You want the human touch, especially for us. If we were a different kind of game – some throwaway hyper-casual thing – I don’t think anybody would care. But we’re a game about creativity and self-expression, so the human touch really matters.
Where I see it going is that it’s a really powerful tool, just like Photoshop is a really powerful tool. But if people are just being lazy (”I’m not going to think; I’m just going to tell it to do the thing, and I’ll sit here and collect the paycheque”), nobody wants that, and our community would never embrace it. People flag when things are AI, or when things aren’t looking right.
Over time, I think as people recognise this is a technology that’s useful to a human and their creative output, some things will be acceptable and some won’t. That’s what I see in our community.
Are there any other tools or technologies we need to know about?
There’s the typical stuff – prototyping and developing. All of our engineers are heavily enabled with AI at this point, and using it intensely. Our PMs and designers now use AI to prototype.
Just recently, our user acquisition team did an incredible thing. I spent 30 minutes with our lead motion designer over the weekend. She’s a motion designer, not a programmer, and I spent 30 minutes going over Claude Code with her. She comes back on Monday and says, “Look! I built this software suite that takes all of our creatives, breaks them down into little pieces, designs a bunch of transitions, and can permute hundreds and hundreds of creatives using our existing library. Here’s the UI, here’s how it works, here’s the visual timeline tool I have for each ad, here’s how I can send feedback to the AI and it changes.” This person does not have a programming degree or anything. She’s just an ambitious, high-agency person given a tool. As I was showing her Claude Code, she was like, “I didn’t know you could do this.” And then she just ran away with it.
“A lot of our people haven’t learned computer science, and it wasn’t even a hobby before. As they pick up Studio and start to run with Rosie, they’re able to build things I don’t think they could have imagined building. That’s pretty special”
Anton Bernstein
Combining domain expertise with these tools is such a massive unlock organisationally that we’re leaning more and more into it: how do we train people to use these tools effectively? How do we just show them it’s possible? Because for a lot of people it’s, “I didn’t know this was possible,” and if you sit down for 30 minutes and show them, their imaginations start running.
All of these models are pretty generalised; they’re trained on such a large corpus of information. If you’re a domain expert and you combine that with this generalised coding expert, all of a sudden you can build magic. That’s what I see across the whole organisation, and it’s really amazing. Then we have a bunch of small tools that are built using AI or integrate AI, but those aren’t really worth talking through.
Say we sit down again in two years’ time. What do you hope Highrise, Pocket Labs and Pocket Worlds will have achieved?
I would love for us to have built, and for many, many people to be using, highly accessible software to fully express their creativity.
So, if expressing your creativity means building your environment – a room, or an experience – you can do that in an accessible way. You don’t necessarily need a desktop; you don’t necessarily need Unity; you can communicate with our tools and fully express yourself. Perhaps it’s digital clothing, or digital content, or digital items that express you: if the barrier lowers such that you’re able to come up with an idea and realise it, I’d love to see that.
I’ll give you an example. Right now, we have two UGC systems that work well for us. One we call Highrise Ideas, the other Highrise Concepts. Highrise Ideas is kind of like Kickstarter: you sketch an idea on the back of a napkin and post it up (this is for a piece of clothing or content), and then other people pre-order it. Your friends, and other people who like the idea, pre-order it at a discount, our art team takes over and produces the item, and then it’s sold into the app. I think that can be significantly enabled with better tooling, to make the process faster and give people the tools to say, “You have a great idea on the back of a napkin? Let’s get it into the virtual world, into your little city.”
We have another one, Highrise Concepts, which is more of a contest: people submit ideas, others vote on them, the best ones win, and we produce them into an event.
I think there’s a lot of possible tooling to enable the design side, to help people who say, “I want to create my fashion label in the city.” Ultimately, I think that creates a greater sense of community and belonging, which is the core value we deliver: how do you feel a sense of belonging and connection with other people, through Highrise as well as through Everskies? That would be ideal for me in two years.
Further down the rabbit hole
Some useful news, views and links to keep you going until next time:
As part of the UK’s planned ban on social media for minors, the government also set a minimum age (18) for the use of AI chatbots with romantic or intimate modes.
Maliyo Games CEO Hugo Obi told Lagos Games Week that studios that thrive will use AI rather than resist it.
Tryll Engine, an on-device AI engine for games, opened a closed alpha and reached a $6 million valuation on the back of a $600,000 pre-seed round. The engine runs AI models directly on players’ hardware, so studios can give NPCs conversational intelligence with no per-message costs.
Next month, you should be in Shanghai. ChinaJoy, the huge annual games event, kicks off this year with the theme “level up with AI” (31 July to 3 August). Immediately before it, on 29 July, PGC Summit Shanghai takes place just up the road, with a content track devoted to AI and games.
Prompt To Play launched last week. It’s Vietnam’s first AI-powered puzzle game development competition, organised by VNGGames with support from OpenAI. You’ve got until 3 July to register, then the final challenge takes place 11-12 July.
As part of a long interview with PC Gamer, in which he made many comments about developer tools and publishing strategy, Tim Sweeney, CEO of Epic, claimed that Steam’s AI labels are “really irresponsible” and make it harder for small developers to succeed.
The founder of Atlas revealed a GTA-style open world, built by two people with their tool in just three days.
