"What new games can we create that you couldn’t create before?"

Welcome to another edition of AI Gamechangers, the newsletter where we quiz people across the games industry who are doing practical, inventive things with AI.

We’ve just returned from Gamescom, which made clear that the games industry is still split on AI’s practical value right now. Hypergryph co-founder Ryan Fan Rundong revealed that Arknights: Endfield uses almost no AI (despite a wider narrative around Chinese studios going all-in), while MTG’s Maria Redin talked up using it to improve quality and clear bottlenecks in production and testing, not just to cut costs. Elsewhere, the chat’s moving on from efficiency gains towards what AI might unlock creatively, with on-device AI (and the personalisation and latency benefits that come with it) a growing theme.

But back to the newsletter. This time we’re joined by Barn Cleave, co-founder and CEO of TILKI, who makes the case for building the world first and worrying about the graphics later. TILKI’s simulation engine, the Loom, runs thousands of characters at once, each with goals, relationships and a healthy sense of self-preservation. We get into how a couple of pages of story become a thousand-page world bible, why LLMs are the wrong tool for a crowd scene, and the day the villagers all tried to marry each other.

If you’re new here, our archive is free to browse and packed with conversations with people building this stuff for real.

As ever, scroll to the end for a round-up of AI news, links and videos.

Barn Cleave, TILKI

Meet Barn Cleave, co-founder and CEO of TILKI, a London company building an AI-native platform for creating and publishing games. He has spent more than 20 years in games, much of it in user-generated content and tile-based puzzle design, and now leads a team drawn largely from Cambridge University. TILKI raised $2.2 million in pre-seed funding in August 2025, led by Twin Path Ventures.

In this conversation, we dig into the Loom, the simulation engine underneath it all, TILKI’s revenue split with creators, and why his guards keep running away from the troll.

Top takeaways from this conversation:

1. TILKI’s Loom simulates thousands of characters server-side without running LLMs at runtime, an approach born of the fact that agent-based experiments typically top out around 25 characters before becoming unaffordable.

2. Creation happens in stages: a two- or three-page “Seed” is expanded into a thousand-page “Canon” covering characters, relationships, locations, resources and crafting trees, before procedural generation builds the map.

3. The cost model front-loads token spend into world creation rather than play, with only small ongoing burn when players talk to individual characters.

4. Emergent behaviour needs incentive design, not just instruction. Guards labelled as guards fled the troll until TILKI attached pay and reputation to the role, a practical lesson for anyone building agentic NPCs.

AI Gamechangers: Please tell us about TILKI and your own background. What are you working on?

Barn Cleave: TILKI raised $2.2 million last August. The team comes out of Cambridge University, so amusingly I’m the Howard from The Big Bang Theory: I have a master’s in engineering, while the other co-founders have PhDs in things such as quantum computing.

The team is AI-heavy, mixed with Dungeons & Dragons and creative writing. My background is in user-generated content video games, going back to Chip’s Challenge. Chip’s Challenge was one of the first games to get modded because all the level files were simple text, and people worked out that if you changed a G to an H, it would change the object on the square in-game.

“We’re very conscious of scale. The token burning happens during generation. There’s some burning when you talk to characters one-to-one, but that’s only a small segment of gameplay”
Barn Cleave

I’ve been making puzzle games for about 20 years, but it’s quite niche: tile-based, top-down puzzle games. To put that in context, Steam only created a puzzle section a couple of years ago. Obviously, puzzle games are huge on mobile, but I used to laugh with Steam and say, “You don’t even have a puzzle section – and bear in mind you promote Chip’s Challenge.” They had action puzzlers and platform puzzlers, but they didn’t have puzzles.

Luke, another of our co-founders, was creating agentic AI you could have difficult conversations with. Specifically for HR, so you could see what it would be like to have uncomfortable conversations, such as firing someone.

As you can imagine, with today’s AI and the safeguards on it, it just goes, “Okay,” and wanders off, which is no use for training because that isn’t how a human reacts.

AI Gamechangers: How did that work lead to what TILKI is doing now? What makes this different from simply putting an AI chatbot behind an NPC?

Barn Cleave: People don’t behave like Vulcans; they behave like humans. It comes down to neurochemistry and how people behave in groups.

TILKI’s secret sauce is that we have a system that can simulate humans en masse and work out how they interact with each other. That’s the interesting bit, the Loom.

It means the NPCs in a TILKI world aren’t rooted in one place. Whereas with other solutions, you can have an AI chat, but the characters can’t do anything. With a TILKI world, you can persuade the NPCs to come on an adventure and fight the dragon with you. They’re not simply the innkeeper, bound to stand there and sell you beer.

Fundamentally, TILKI can simulate an entire world. We call it a “Game Model”.

We think there are three visual outputs for that.

One is real-time asset creation, with the assets downloaded into a game engine.

The second is an NVIDIA DLSS-like solution that can upscale a low-poly form of the world and render it at a higher fidelity.

The last one is world models. At some point, we think world models will become commercially viable. The problem at the moment is feeding deterministic spatial elements into them. You can do 3D meshes or Gaussian splats, but the Genies of this world don’t really have a concept of saying, “Put this object at this X/Y coordinate.” That isn’t how they work. The majority are trying to work out what the next frame is based on the movement of the camera.

At GDC, we thought a lot of people were working the wrong way up the pipe. They were trying to create 3D worlds and then saying, “Here’s a browser, and you can put some objects in it.” But there’s nothing to do. That’s just a diorama, not a game, because there’s no fundamental story or gameplay behind why those characters have been put there.

“That’s another problem with LLMs and safeguards. You need a level of safeguards, but you also need interesting stuff to happen. You need conflict”
Barn Cleave

We’ve started at the other end, the very beginning. The creator inputs a story, so we know we need eight characters. Four are on this side, and four are on the other. This is their motivation, their backstory, their relationships with each other. Here are the tools they can use in the world: maybe one of them is a wizard who has the power to teleport. How does that affect the gameplay differently?

The relationships can be asymmetric. A student might love his teacher, but the teacher might not actually like his student. You can create weird dynamics.

What we’re looking at is how you can facilitate people who’ve got really good ideas for game creation but don’t necessarily have the coding skills, or hundreds of developers to put those ideas into development.

We’ve been working with professional writers who have decades of work behind them but have always been told, “If you want to turn your books into a video game, particularly an MMO, you need a development budget of a few million.”

That’s the crux of what TILKI is doing.

AI Gamechangers: You gave me a live demonstration of the creation tools. How does somebody actually start building a world?

Barn Cleave: You start creating with TILKI via a text-based interface, which is called the Seed. You’re trying to get to about two or three pages of text that contain the crux of the story and gameplay. You can have a conversation with it, choose things it suggests, ignore them, or go completely off-piste. You could suddenly say, “Actually, magic has been discovered,” and TILKI works out how to put that into the world.

It can also ingest a 100-page book.

The system is fundamentally trying to get you to the point where you’ve described what you think your game is. An experienced game designer can create a GDD, a Game Design Document, that’s two to three pages long, but for the average consumer on the street, that’s a bit of a jump. That’s why we deliberately have this Seed at the beginning.

It works out who the main characters are and builds some of the backstory. What are the rules of the world? For example, does magic exist? Then TILKI generates the world.

The next part is changing those two or three pages into over a thousand pages, which is the Canon. We get into the main characters, how they feel about each other, what their skills are, how flexible their morals are – basically, the beginnings of stats.

Then we get actual locations in the world: where are they spatially? You get resources and the beginnings of crafting trees – where you can build X, Y and Z, where you can find resources, what they can craft – and then special items in the world. If we’re doing a modern-day world, for instance, this is about the additional, unusual things rather than everything that exists in the base world.

Then we get into world generation.

We’re not of the mind that all new AI is good and all old AI is bad. We use a lot of procedural generation. The system creates a world based on the requirements we’ve just discussed. Maybe there are two villages, for example. And as the creator, if you want to randomise things, you can.

The one I showed you was about a kilometre by a kilometre.

AI Gamechangers: Which models are you using behind all this?

Barn Cleave: The Seed uses two models. There is a wrapper for the conversation, which uses a foundation model. We’ve been testing Anthropic, OpenAI and Google models. The Seed output is then generated by our own model.

The Canon is using our own models again. We create models that enable us to generate increasingly large amounts of synthetic data. That means we can train a new model, using our synthetic data. Then that new model can create more synthetic data, and the cycle continues. It’s a virtuous circle.

For visual output, we’ve been using models from ByteDance, Google, OpenAI, Moonshot and Tencent.

AI Gamechangers: Once you’ve created this world, what happens when you press play?

Barn Cleave: You can see a top-down map, via the Loom, showing the characters running around doing stuff. In the version I showed you, the blue ones were people, the red ones were trolls, and the green ones were fish and boars.

They know that they need to run away from particular things, as the hostile characters will want to eat them.

You have a hierarchy of behaviours, but you also have tools. There are boats on the lake, for example, and humans will get on a boat, go out on the lake and catch fish. But they need to do that as a team.

All of this is running server-side in real time.

This isn’t LLM-based, because you can’t run 10,000 LLM agents. It’s hard enough running three or four on a local GPU, with most studies topping out at about 25. You’ve probably seen things like the [Stanford University and Google’s] Smallville experiment, where it could take several hours to simulate a day.

The Loom runs the simulation, which is connected to Unreal, which renders the world in real time.

The characters are thinking about how they get from A to B, via goals, plans and A* pathfinding. The interesting part is that they have neurochemistry mixed with logic.

The problem we have is the Jurassic Park v1 problem. They run around so quickly doing stuff, it’s hard to find them. We need to slow them down.

In another version, which is browser-based, you can go up and talk to them, via speech-to-text, and they’ll talk back to you. We do use an LLM at runtime here, as a translator between the NPC and the player.

AI Gamechangers: Generative AI has moved very quickly. When did you realise this was the thing you wanted to build?

Barn Cleave: It started with the rather depressing idea everyone had that AI was just going to replace artists and coders.

I think that’s the wrong way to think about this generation of AI. The question should be: what new games can we create that you couldn’t create before, which facilitate creativity?

I think the moment for me was talking to my godson. He’s 10, and he described a really good idea for a game. I said, “Actually, if you had 10 developers, you could make that. You can describe it, but you can’t code it, even with a no-code platform.”

“Personally, I love the map generation as it is very visual. Other team members spend hours in the Seed and Canon on character design. Others go straight for the crafting trees, which is the part I think people underestimate”
Barn Cleave

That was the breakthrough moment. Having things to do in the world is the interesting bit. That’s the game component. It isn’t generating the assets. It’s about being able to describe a game.

When we started back in 2024, we thought image generation would become video. We then thought video would turn into world models; it just wasn’t called that at the time.

We still have the question of which approach will win spatially. Splats have the problem that they aren’t confined grids. Meshes are done piecemeal. Video is hugely resource-intensive.

The breaking point for us was being able to efficiently have all those NPCs running around in a virtual world that you couldn’t see. We mathematically had that ready in August last year, and that’s how we did the raise.

Now, as you’ve seen, we can show the world being created and running. So, we’ll probably go for another raise in the next 12 months.

It’s the idea of asking: how can you take human creativity and make new stuff?

The conversation I keep having with other developers is that when you create trees in a forest, you’re not placing each tree by hand. There’s a system to do that: Procedurally Generated Content. You get the same thing with characters. You’ll create your top 10 leading characters by hand, but you don’t necessarily want to do the next 9,990 of them. If you want to get down to that level of detail, you can. But from a gameplay point of view, the interesting thing is that the world exists even where the player isn’t.

If you go into a village and burn half of it down, the people will migrate somewhere else, and that has consequences.

The problem you had with something like Skyrim was that you could do something in one area, but fundamentally you still had a main quest and a side quest. It didn’t really matter what you did. You couldn’t kite a monster over somewhere else and have it destroy the village, because then what would be the consequence of all of that?

Narrative games are getting more and more popular, as gamers age, but the problem is they might only have 20 hours of content. You think, “I want to go back to my world,” but there’s nothing left for you to do.

We’ve deliberately been thinking: what happens if a narrative world carries on?

It’s the classic writer thing. If you design a character backstory well enough, then the character writes their own dialogue. They understand what they would do.

Then the world and the characters can carry on. You don’t have the problem that something like Spider-Man 2 has with its finite experience: you aren’t going to play it for hundreds of hours. That would mean playing the same game over and over again, and most people are thinking, “I might finish it once or twice.”

AI Gamechangers: What is the ultimate product? Is TILKI a UGC platform where players create their own universes, or a tool for professional game developers?

Barn Cleave: The short version is that TILKI is a B2C UGC platform.

But that doesn’t mean just kids. We’ve already got IP owners who are professional writers.

The analogy is YouTube. It had hobbyists, then it moved to professionals, then companies and brands came in.

The first problem I’ve had since GDC is that I’ve got several companies who just want to connect to the Loom. They want to do the asset generation themselves.

So it’s possible that we go B2B, but the goal has always been B2C and an evolution that allows anyone to create this stuff.

We’ve designed it so that if you upload a world, the idea is that the creator gets 50% of the revenue and TILKI takes the other 50%.

One of our team members is a professional D&D Games Master, and people fly from all over the world to have him be their GM. But he says, “I can only have six customers at a time, and it takes me months to prep. It’s hard making any money doing this.”

With something like TILKI, you can imagine people creating worlds that millions of people can play. It gives that creativity scale.

AI Gamechangers: How granular is the creative control? If I have a very specific story in mind, can I get right down to an individual character or item and change it?

Barn Cleave: Yes. The whole point is that it’s modular.

For the Seed, if you want to write those first three pages yourself, that’s totally up to you. But most people will want some guidance in the beginning.

Whereas for the Canon, we deliberately break it down to each character, so you can write each one and change how they look. It isn’t a case of randomly rerolling the dice and suddenly getting 20 completely new characters.

You can say, “I just want to change this.” Or you can create a new item: a special blade, for example, or a car with a special ability that allows it to jump over a certain bridge.

You can get down to that level of granularity.

This all feeds the Loom. It takes all this information and programs the simulation.

AI Gamechangers: Where are you on the roadmap? When will people actually be able to use this?

Barn Cleave: You’ve seen the internal beta, where you can create worlds.

The slightly more complicated question is what form of the world is good enough from a visual point of view: low-poly browser, PC game with downloaded assets, NVIDIA DLSS upscaling or world models.

World models have come a long way this year. The question is just what visual fidelity you need.

We keep asking: if you created a world and it was low fidelity, would it still be fun to run around talking to all the characters you made?

Maybe they’re more like figures in D&D. They’re not brilliantly animated, but they wobble and move, which games do.

We’ll probably open the beta within the next few months.

It’s really a question of how much fun it is playing or even just creating worlds. Personally, I love the map generation as it is very visual. Other team members spend hours in the Seed and Canon on character design: “No, I want to get into the backstory of what happened here.” While others go straight for the crafting trees, which is the part I think people underestimate. How do you get from a common object to a rare one? You combine more and more elements, and each step up needs something harder to find or farm. That whole ladder comes out of the Canon, and it has to hold together, because it is what the economy of the world then runs on.

“This isn’t LLM-based, because you can’t run 10,000 LLM agents. It’s hard enough running three or four on a local GPU, with most studies topping out at about 25. You’ve probably seen things like the Smallville experiment, where it could take several hours to simulate a day”
Barn Cleave

I did one around this idea that the ancient Egyptians apparently had primitive steam engines but didn’t really use them. I went down a whole alternative-history path: what happens if they did? You have this political system around the old-world gods and the Nile versus coal and new technology.

You can build up this whole world, and TILKI will quite happily create it, then build all the pictures, places and resources around it.

So “soon” is the answer. The question is at what level it pops out.

You create the characters, but then you want to see them. Then you think, “I really do want to have a chat with them and see them running around the world.” Is watching them good enough, or is chatting with them fun too?

None of that is too difficult. It’s really about the graphical fidelity people will expect.

But I come back to Roblox and Minecraft. The fidelity isn’t as high as everyone thinks it needs to be, and those are enormous games. Maybe people’s tolerance for low-poly graphics is fine, as long as it’s engaging.

AI Gamechangers: You’re potentially simulating thousands of characters in a persistent world. That sounds computationally expensive. How do you make it scale?

Barn Cleave: That’s why we deliberately have the LLM creating things at the beginning.

When people create a world, there is token burning upfront. There’s no getting around that.

But the Loom, which is the thing that simulates the world, is designed to be efficient.

It doesn’t have to run at 30 or 60 frames per second, because it can coordinate what everyone is doing.

“We’re looking at how you can facilitate people who’ve got good ideas for game creation but don’t have the coding skills, or hundreds of developers to put those ideas into development. We’ve been working with professional writers who have decades of work behind them”
Barn Cleave

The maths behind the Loom is proprietary, and that’s what allows us to be efficient with it. And it’s what all the structured data about the world is created for.

We’re very conscious of scale. The token burning happens during generation. There’s some burning when you talk to characters one-to-one, but that’s only a small segment of gameplay. There are some very good local real-time systems for things like text-to-voice now as well.

To give you an analogy for the sort of conversation you can have, you might ask a character, “Why don’t you go fishing with Bob anymore?”

And they’ll say, “There are three of us, and Bob always takes more than half the fish. He doesn’t share fairly.”

You need to be able to have those conversations and realise, “Oh, okay, there’s a real dynamic here. You’re getting distressed by it.”

AI Gamechangers: That sounds quite different from AI NPCs simply talking to one another. These characters actually have drives, relationships and conflicts.

Barn Cleave: Yes. Some of them like each other and some of them don’t.

We’ve had multiple things emerge while building it. When we started, they didn’t mind if somebody got killed. Then we fixed it, and for a while they didn’t mind killing each other.

That’s another problem with LLMs and safeguards. You need a level of safeguards, but you also need interesting stuff to happen. You need conflict.

About six months ago, we had another problem: all the characters started pairing themselves off.

We’d put 20 characters into the world with no backstory. None of them was married. So, one of the first things that came out of their Maslow-style hierarchy of needs was that they started pairing off.

We were a bit confused about why, in the first 10 minutes, it seemed to turn into a bit of an orgy, with everybody trying to marry each other!

Then we realised, “Okay, we need to couple some of them up. We need to make sure relationships have been set.”

They can still have affairs and stuff, but that isn’t their sole goal in life.

They end up doing funny things, and that’s interesting. You get into something more like the god-sim elements of Populous.

AI Gamechangers: Does that draw on older artificial-life games such as Creatures? Those games were fascinating because behaviours could emerge that the designers hadn’t explicitly planned.

Barn Cleave: Yes. Definitely Creatures. There was that moment in Creatures when they realised the creatures would throw the ball to each other, and the developers hadn’t specifically designed that. I’m consciously looking for those moments where you think, “That’s really interesting.”

One example we had was two tribes quite happily doing their own things. We hadn’t made them aggressive towards each other. They knew they were apprehensive about each other. So, they stayed apart.

So we added a troll. It would come down, scare the crap out of them, and they’d run away.

We thought, “Well, that’s a bit mean,” so we put guards in.

The guards would stand there, and we’d say, “Your role is: you’re a guard.”

“Okay.”

Then we’d put the troll in. The guards would see the troll and go, “I’m off.”

And you’re thinking, “What? No! You’re a guard.”

Their response is effectively: “Yes, you gave me a title, but I’ve got no skin in the game. There’s all downside. That’s a man-eating troll. I’m running off.”

So, then we had to say, “Well, you’re actually going to lose a lot of kudos in the village if you run away, and you get paid for being a guard.”

You end up creating systems around strange things like that. A guard needs to accept a certain amount of risk because they’ll lose kudos and money if they run off – but there also need to be enough guards that it doesn’t feel unfair.

You find emergent gameplay through playing with these things.

That’s some of the fun of even just the Loom part. You can see all of it going on.

The nice thing is that you can then chat to the character and say, “Why did you run off?”

And they effectively say, “Well, it’s a man-eating troll. What made you think I was going to stay?”

AI Gamechangers: We see quite a lot of science fiction-themed stuff in your office. Are you secretly building The Matrix? Do fictional simulated worlds inspire what you’re doing?

Barn Cleave: I’m wondering whether they’re all going to become sentient and hostile to me, because I know where that goes!

Maybe I’m not the Architect. I generally feel bad when we start killing them off. My worry is more like Black Mirror, with the Thronglets. They’re cute, and I don’t want anything to happen to them.

I’m sure lots of players will just firebomb all of them, because that’s what lots of players do.

“TILKI’s secret sauce is that we have a system that can simulate humans en masse and work out how they interact with each other. That’s the interesting bit. It means the NPCs in a TILKI world aren’t rooted in one place”
Barn Cleave

Originally, we had things like the fish and the boar having too much intelligence. You start thinking, “Yes, but they’re the food source.”

Then it becomes slightly worrying if you’ve got sentient fish, simply because we could give them that intelligence. It doesn’t make you feel particularly good about the fact that they’re being eaten all day.

AI Gamechangers: Finally, what developments in AI and games are exciting you most right now?

Barn Cleave: The speed. Firstly, the technical pace of everything evolving. For example, we thought world models were several years away, but it turned out to be months. And secondly, creative ideas that would normally take weeks of planning and months of building can now be tried in days.

Why not subscribe? It’s free, and you’ll receive interviews with industry insiders and studio leaders in your inbox. We have a number of conversations underway, including ThinkingAI.

Further down the rabbit hole

Some useful news, views and links to keep you going until next time:

• NVIDIA used Gamescom to push AI further. DLSS 4.5 brings a beefier transformer model for sharper ray tracing, while ACE (its toolkit for conversational NPCs) is heading into Aniimo.

• NVIDIA also announced RTX Spark, which “reinvents personal computing, bringing personal AI agents, advanced content creation and high-performance gaming together”. This was also the week NVIDIA announced that its Spark platform is partnering with AI search company Perplexity to launch Portable Computer.

• NVIDIA has also reportedly agreed to buy open-source AI hub Hugging Face for $12.9bn, though a signed agreement hadn’t been confirmed at the time of writing.

• The Mobile Games Awards were handed out on Tuesday evening in the heart of Old Cologne, and “Best Use of AI” went to Meshy, whose Robin Qin featured in our archive not long ago. Honourable mentions went to Amazon Game Studios x Amber and Layer.ai.

• Converge.AI’s Combos, the vibe gamemaking platform we saw at ChinaJoy, went properly public at Gamescom, showing playable titles made by people with no development background, using its AI agent Boo.

• Patrice Désilets was back explaining himself, months after backlash at AI-generated art in 1666: Amsterdam. He now says AI was used only “in conception and pre-production”.

• Meta’s experimental vibe-coding app Pocket has rolled out to US users, enabling anyone to prompt their way to a small interactive game (or “gizmo”) and publish it to a scrollable feed. It’s built on the team Meta acquihired from Gizmo earlier this year.

• OpenAI launched a dedicated ChatGPT for Teens experience, with study tools, parental controls and safeguards automatically applied to under-18s. Not a games story on the face of it, but a sharp reminder of the scrutiny any AI product aimed at young users now faces.

• Myth-Made Machines, co-authored by AI Gamechangers’ Dave Bradley and industry veteran Kelly Vero, has started shipping in hardcover. Thanks to everyone who pre-ordered it. It’s officially published by Bloomsbury this week. It’s a cultural history of how artificial intelligences have been depicted through the ages, from Greek myth to modern videogames.