How AI Could Bring Halo and Other Classic Games Back to Life
- AI agents could make unusual game ideas economically viable by reducing the time and cost of prototyping. Anish Acharya describes Astra operating development tools directly and correcting its own output, shortening “the time between idea and prototype.” He tentatively imagines 5–10-year development cycles becoming 2–5 years, allowing studios to fund more experiments and new IP.
- Human taste remains essential even when building a working game becomes easier. Acharya’s constraint is that “the last 20% of the game is 80% of the work”: sound, feel and detail turn functioning software into something worth playing. Jonathan Lai adds a different opportunity—“creation is consumption,” so making a game can itself be the entertainment.
- Halo’s browser revival illustrates how AI could renew classic games and accelerate remixing. Acharya says he has heard of four community projects making the original Halo playable through WebGPU. Mods already produced multibillion-dollar franchises; easier experimentation gives players a “place to remix from” rather than requiring an original concept from a blank canvas.
- Community and distribution may determine whether easier creation benefits existing IP owners. Acharya distinguishes extensions that strengthen a shared game from forks that might split its community. Lai says revenue shares and royalties remain unresolved; platforms with persistent communities across games offer another route to capturing value.
- Lower production costs could reopen genres whose audiences struggle to support conventional development budgets. The discussion ranges from demanding RTS games to Lai’s enthusiasm for train-conductor simulators. These audiences can be narrow yet committed: Lai says enthusiasts will pay “a lot of money” for experiences that match their interests.
- The larger ambition is gameplay that adapts to players rather than requiring every possibility to be scripted. Lai’s “never-ending game” would generate adventures around a player’s choices, while dating-game characters could become responsive agents. These remain possibilities; Acharya also points out that some proposed innovations already exist in games such as Crusader Kings III without frontier AI.
1. Agents shorten prototypes and expand the risk budget
Acharya identifies Astra’s ability to operate Blender, Unreal and Unity as the unlock: a hobbyist can draft a game design document, generate concept art, build and rig 3D models, add generated audio and export a playable prototype. It can also take screenshots and correct its own output. Earlier models explained what users should do; direct operation shortens “the time between idea and prototype.”
Lai separates that from generation through code: he remembers “Opus 4.7,” adding “maybe it was Opus 4.8,” becoming unexpectedly good at Three.js browser graphics. Its element-by-element construction would be “maddeningly meticulous for a human.” Code generation and competent tool use are distinct capabilities converging on game creation.
Using GTA 6 as the extreme, Lai says it is coming “next month” after development beginning around 2013–2014—a 12–13-year cycle. Acharya tentatively imagines moving 5–10-year cycles toward 2–5 years. He also suggests reducing a ten-year, hundreds-of-millions project to five years and tens of millions. Development would remain expensive, but permit more experiments.
Long commitments favor sequels and established shooter genres because producers cannot casually risk ten years and hundreds of millions; markets and creative tastes can change before launch. Lai describes another use of productivity: retain the decade and increase ambition. If GTA 6 delivers an intricate city, comparable effort with AI could build a universe.
2. Human taste stays scarce while creation becomes play
Lai’s pushback is whether AI can produce balance and genuine fun. Acharya says AI is “not going to replace human taste and creative judgment”: “the last 20% of the game is 80% of the work.” Getting characters moving and shooting establishes functionality; making a shotgun sound different from a pistol helps establish feel. Those judgments remain human territory “for the foreseeable future.”
Acharya thinks “the vast majority” of future games will have just one player: their creator. That is acceptable because “creation is consumption”; bringing assets and characters to life has a “Zen gardening aspect.” Lai connects this to Minecraft and Lego, where making and sharing things already constitutes play.
Acharya’s record-shop project began with a DJ friend complaining that online shops were merely song lists. Toronto record stores in the 1990s were social spaces, full of conversations and loud music. Using Opus 5.5 and Habbo Hotel reference art, he generated a shop where selecting music prompted a DJ to put on a record and characters to dance.
Extending that shop into Toronto’s Queen Street, Acharya let other players build storefronts through prompts. His son created a pet-adoption center whose animals actually followed players: the coding agent changed the engine to support an unplanned mechanic. This was “dynamic abilities,” beyond procedural content. When Lai mentions Roblox’s Adopt Me!, Acharya concedes his son might have drawn inspiration from it.
3. Remixing revives classics and seeds new franchises
Lai’s eight-year-old wanted Ghost of Tsushima, but was barred from its bloody samurai combat. He used ChatGPT to explore its mechanics and boss battles through online wikis, then Codex helped him build a Minecraft mod recreating the experience. It worked and circulated among friends. Now, Lai says, his son stops asking to buy certain games and decides to make Minecraft versions instead.
Lai’s historical argument is that mods already create enormous businesses: Counter-Strike began as a Half-Life mod; League of Legends drew inspiration from Dota, itself a Warcraft III mod. AI lets ordinary players combine familiar mechanics and immediately test whether they are fun. Existing games solve the blank-canvas problem by providing “a place to start and a place to remix from.”
Lai describes decompiling games to recover source code and assets, and combining two games’ assets into a third. Acharya says he has heard of four community projects making the original Halo—a 2001 Xbox game—playable through WebGPU in a browser. Established IP can reach people on devices relevant to them.
Lai’s personal revival is Solar Realms Elite, an asynchronous multiplayer game from the dial-in bulletin-board era. Its source code was lost, but a design document and executable survive. Earlier model generations could not recreate it; he now has what “feels like a faithful recreation” running on his computer. Preservation is a concrete use case alongside new creation.
4. Cheaper production gives niche genres another opening
Acharya anticipates a renaissance in RTS games, including the elusive great iPad RTS. Lai links the genre’s dwindling hardcore audience to punishing competitive demands: actions per minute matter, and other players can be overwhelmingly faster. Conventional market-size analysis discourages development; AI makes him excited about an “explosion in RTS games.”
Lai’s favorite niche example is a train-conductor simulator: keep a Tokyo train on schedule, stop accurately and collect passengers. Overshoot and the last carriage may miss the platform, docking your paycheck. The host asks whether this is fun; Lai insists it is. Such specific interests need not command mass audiences if enthusiasts will pay “a lot of money.”
Acharya connects RTS-style interfaces to agent orchestration. Lai argues ambitious visual interfaces may become sensible once their creation cost falls. Acharya says command-line, text-heavy orchestration is unintuitive for ordinary consumers, so there will “probably” be a more accessible visual interface. Whether it resembles StarCraft, Agent Empires or Factorio is less clear; the implementation remains open.
5. Persistent communities offer a path to value capture
Lai says original IP owners historically have not benefited much from some successful descendants, citing Dota and League of Legends; he is explicitly unsure how much Half-Life’s developers earned from Counter-Strike. Embracing AI remixing therefore leaves a commercial gap. Revenue shares and royalties are being discussed, but “there’s a lot to figure out there still.”
Acharya distinguishes extending Minecraft’s shared foundation, which benefits the owner, from forking a game, which offers “no benefit” and might split its community. He also questions whether Roblox’s thinner shared game makes it harder to bootstrap than Minecraft. Lai likens Minecraft mods to fan fiction built on shared culture, but says both its centralized approach and Roblox’s decentralized entry through different games can work.
Lai sees room for distribution built around AI creators rather than existing Steam, App Store or Roblox channels. He recommends checking out Spawn, explicitly saying “we’re not investors”: its embodied agent, Savvy, accepts bug reports, changes games during play and creates portals between them. Games are multiplayer by default, preserving community across experiences. Lai calls it early, but unusually ambitious.
Lai says developers he speaks with already treat AI as “table stakes” for design documents, concept art and prototyping. Acharya expects lower costs to improve studio economics and working conditions in an overworked industry dependent on publisher contracts. Lai calls productivity the first wave; the second is new mechanics and genres that could “massively expand the market for games.”
6. Adaptive worlds could turn scripted content into responsive play
Lai’s “holy grail” is an MMO that adapts to player preferences: someone who stays on the beach instead of entering the scripted forest could build a ship and receive an ocean quest line, complete with pirates. World models might enable it. He invokes Ender’s Game’s “Mind Game,” keeping the forecast conditional: “it’s possible we’ll see our own version of the Mind Game at some point.”
Lai wants to revive Civ II’s 16-bit Windows version, then imagines Civilization units as agents with independent agendas: the ruler sets strategy and influences subjects instead of micromanaging every unit. Acharya immediately counters, “This game exists,” pointing to Crusader Kings III’s personalities, schemes and subjects who refuse orders. Acharya says it achieves that complexity without frontier AI.
Dating simulators present another content bottleneck. Acharya says the Chinese-developed Love and Deepspace generates hundreds of millions in revenue, but each new romantic character requires extensive content and a rich backstory. He proposes agents with encoded personalities, reacting in real time and knowing about the player’s day. He calls the genre “wide open” and ripe for AI disruption.
Full transcript
Grand Theft Auto 6 is finally coming out next month. That game has been in production since 2013 or 2014. It’s a 12- or 13-year development cycle.
Games have long been one of the most complex products to create. But with AI helping build some of that specialized knowledge, you no longer need to be a professional.
There’s having an agent look at something and try to recreate a game from scratch. There’s also the meshing and merging of two games or two art styles. Multibillion-dollar game franchises have been created using mods. Counter-Strike, for example, was a Half-Life mod. League of Legends was inspired by Defense of the Ancients, or Dota, which was itself a Warcraft III mod.
So imagine what the next 12-year game will look like.
AI is now making it possible to build niche games, like esoteric genres that previously wouldn’t have been economically viable.
What’s a game that each of you has been thinking about making for years, and now you finally can?
1. How AI Can Now Build a Game
It seems like there’s been this incredible shift in the gaming industry, and also in what we’re seeing from solo developers and tinkerers. Tim Sweeney from Epic has come out and said a bunch of really interesting things about AI. The CEO of Xbox, Asha Sharma, has also come out and said they’re thinking a lot about the interplay between hardware and software on Xbox. It seems like there’s a tide shift. So, yeah.
Want to talk about that.
2. Decompiling and Reviving Dead Games
Yeah, maybe I’ll start off. Anish, I’d love to hear your thoughts because you’ve been building some games in your free time.
Yes, I have. Yes.
I feel like there was a big technology unlock about a month ago with Astra, because it was very good at directly using popular game development software. Astra is very good at using Blender. It could actually work directly within Unreal and Unity. In previous generations of models, they would just explain to the user, “You should go and do this in Blender,” or, “You should do this in Unreal.” But Astra could actually operate the software for you, and that was a huge unlock.
For the first time, a hobbyist who’s never built any games before could use Astra to draft the game design document, generate some concept art, move that concept art into Blender, build 3D models from it, animate them, rig them, move them over to Unreal, and then add some music and sound files, which can also be generated. Then you can export that as a playable, and you literally have a game that you one-shotted using Astra.
3. Mods: Where Billion-Dollar Franchises Come From
The other thing that’s been very valuable is that Astra can also QA itself. It can take screen caps of the game and learn from its own output: “I need to fix this bug,” or, “This doesn’t look right.” That’s why the initial prototypes people are building on X are surprisingly good. I think that was a huge technology unlock, and it has a number of ramifications that we’re starting to see in the games industry.
The first effect is that it’s dramatically shortening the time between idea and prototype. That’s something Tim Sweeney talks about in some of his Twitter posts: AI is now making it possible to build niche games, or esoteric genres, that previously wouldn’t have been economically viable. Because it takes less time to make a game, you can now spin up a prototype of a strange or weird game idea. It costs less money and takes less time.
I think that’s huge for the games industry, because one of the structural challenges in games is that it takes a really long time to make them. Grand Theft Auto 6 is finally coming out next month, and that game has been in production since 2013 or 2014. It’s a 12- or 13-year development cycle.
Crazy.
I’m sure it’s going to be an amazing game, but by any other software development standard, taking 10-plus years to make HR software or a CRM is kind of insane. Shortening the time from idea to prototype—let’s see if we can get a 5- to 10-year development cycle down to 2 to 5 years. It’ll still take some time, but I think it’ll enable game developers to take more risks and have more at-bats, run more experiments, and generally increase the richness and diversity of games that are created, which I think is really exciting.
Yeah. To build on some of those comments, it’s interesting because we’ve seen two approaches emerge. I actually think the first thing that really became applicable to gaming was the emergence of Three.js manipulation. I think it was Opus 4.7, maybe it was Opus 4.8, but it had this property it wasn’t designed to have—at least Anthropic said it didn’t—which is that it was awesome with Three.js, which is in-browser 3D. The way it does it is by laying things out almost on an element-by-element basis, in a way that would be maddeningly meticulous for a human.
So coding is one path to game generation. It felt like what Astra did really well, to your point, Anish, was tool use: it could use Blender, and Blender in the cloud, really effectively. These are two megatrends within models, one kind of emerging and one programmed, that are converging to drive a lot of the acceleration in game development.
You know, it’s funny you say it takes 12 years to make HR software. I think the amount of time it takes is why there are so many sad stories in gaming. Think of all the game stories that are left untold. How many games were canceled three years into a six-year cycle or five years into a six-year cycle? The market has moved on, player tastes have changed. You go through financial cycles and creative cycles, where ideas that were creatively interesting three years ago now feel flat. It’s so hard to work on something for 13 years and see it all the way through, and then have it be great. It’s a miracle that it works at all.
4. What Shorter Dev Cycles Unlock
Yeah, no, for sure. Another aspect of these long development cycles that people don’t talk about, but I think has a very real impact on the industry, is that they have a suffocating impact on innovation. If you’re a producer at an AAA game studio and you know it’s going to take 10 years and hundreds of millions of dollars to ship a game, you’re going to be very conservative in what you greenlight.
Interesting. Right.
You’re going to greenlight sequels or new IP. I think we’re on the 20th Call of Duty installment. That’s Grand Theft Auto 6, right?
Yes.
It’s not a new IP. You’re going to skew toward genres that have established demand, which is why we see so many first-person and third-person shooters over experimental genres.
When you lower the bar for shipping a game—it’s going to take less time, five years instead of 10, and tens of millions instead of hundreds of millions—it’s still going to be expensive.
Yeah.
I think it just means that people can take more creative risks.
Totally.
You can actually ship a new IP. You can bet on some game developers who have never shipped anything before, greenlight them, and give them a budget to make something. I think that’s a really exciting story for the game industry, and it’s going to unlock a lot of structural innovation.
Right. Another way to explore this new capability is to increase ambition. Imagine what the next 12-year game will look like: GTA 7. You were talking today about how intricate the city is in GTA 6.
Yeah. They have custom radio stations that they designed because they were probably just like, “Well, we have a year, and this team isn’t doing anything. Let’s just record a bunch of radio stations and stuff like that.”
It’s incredible, right? But imagine if they then spent 10 years with the current tools. Forget about a city or even a world. You could build a universe with that level of detail. Why not?
Yeah.
Yeah, yeah, yeah. No, I think that’s absolutely right. Expand the scope and ambition of games; take more creative risks. I think AI can literally create an abundance era in games that wasn’t previously possible, because we can make games that have never been seen before. You can take more creative risks. Another element we haven’t talked about that I think is equally important is that AI is also changing who can make games.
Games have long been one of the most complex products to create. It’s a true multidisciplinary craft: you need to know how to code, art, game design, music, and sound. The execution required to make a great game is very high, much more so than snapping a great photo or making a video.
Those are hard, but I think there’s a level of difference there. You can have AI help fill the roles that require specialized knowledge you may not have. I think you no longer need to be a professional or have a full team of people with different specialties. AI can plug some of the gaps in your skills and knowledge, and that’s democratizing game creation in a way that wasn’t previously possible.
Someone with just an idea and some reference art can work with Astra to iterate their way to a game prototype. Of course, it’s a prototype, not a fully playable game, and many people won’t finish it. They’ll make the first level or scene and then drop out. But some percentage of them will make it through and decide they love game development and want to turn it into a job. I think we’ll see an explosion in the number of game creators, and that’s really exciting.
5. Can AI Make a Game Fun?
You guys must talk to founders who are either creating tools for these would-be AI-enabled game developers or working on AI-developed games themselves. The hardest thing to me would seem to be game balance and making something that’s actually fun and engaging. It seems like that’s an art, and something that requires human intuition, know-how, and a different kind of creativity. What are you seeing in terms of what AI can do?
I think you hit it spot on. The AI model isn’t going to replace human taste and creative judgment. That’s still really critical. There’s a saying in the games industry that the last 20% of a game is 80% of the work. That’s where all the magic comes together to make something fun. The first 80% is just putting together a game that works: you can start it, there’s a character, it can shoot things, and there are monsters to shoot.
The last 20% is so critical. How does the world feel? Going back to your point about custom radio stations in Grand Theft Auto 6, when you fire a gun, what does it sound like? The shotgun should sound different from the pistol. When you use a laser gun, there should be recoil, and so on. All the little details are what make a magical experience. I think that’s going to be the province of humans for the foreseeable future. How does something feel? What do players want? How should this gun feel? Human taste and judgment will still be critical in those situations.
Yeah. And what’s going to make me want to spend 1,000 hours playing Grand Theft Auto or Baldur’s Gate 3? These incredible games have a level of detail, but also a sense of consequence and depth to the world. Maybe that’s something LLMs can help extend.
Yeah, exactly.
What do you think?
I agree. I don’t know that it makes creating a really fun game trivial, because that’s really hard. It’s an intersection of many things: some designed, some serendipitous, some cultural, some social. Many things go into making a fun game. I do think that in this new era, sometimes creation is consumption. The vast majority of games may be played by just one person: the creator. And that’s okay, because there’s programming or creating games for the output, for the artifact, and then there’s doing those things as an activity.
It’s actually very tranquil to make and work on a game, especially because it has an art element. You see the assets come to life, and the characters say things you didn’t program them to say. I really do think there’s a kind of Zen gardening aspect to game creation that will become increasingly popular.
I think that’s such a great point. We’ve seen that creation is play sometimes. We’ve seen it with Minecraft, one of the most successful and longest-running games of all time, and even before video games with things like Lego, where kids create things for fun and share them with each other.
That’s definitely the case with AI today as well. The act of creating something, bringing something to life, is itself fun. AI is turbocharging what people can create, so you don’t need to know everything about code, art, and 3D to make something you can be proud of.
Yes.
And have fun making it.
6. Building a Habbo Hotel Record Shop
100%. Anish, you built this Habbo Hotel-themed record shop. What was that process like?
It started with a conversation with a friend who’s a DJ. He was saying, “All the online record stores look the same. It’s just a list of songs.” We were talking about how, when we used to go record shopping in Toronto in the ’90s, going to the record shop was like going to the barber shop. You’d go there, it would be filled with people, and you couldn’t leave without having five conversations. Someone at the back would be playing tracks really loud. There were mixtapes for sale, and raves were being promoted. The whole thing was a real social environment.
So he said, “Hey, why can’t online record shopping be like that?” I thought, “That’s cool. Let me try to create that.”
This was with Opus 5.5. I gave it a shot with just a couple of Habbo Hotel reference art components, and it far exceeded my expectations. It didn’t just generate the room; it also somehow grabbed all kinds of ’90s music. It had curated R&B, dancehall, hip-hop, and soul. You could pick a track, look through the tracks, and put one on. Then the DJ at the back would grab the record and play it. People were dancing a little. It was very intricate, so it was a total inspiration.
From there, I said, “What if I extended this and turned it into Queen Street in Toronto in the 1990s?” So I turned it into a whole street. The last touch, which I think was the most science-fictional thing that would never have been possible, was making it so every other storefront was something another player could build out just by prompting. The interesting thing is that they could customize not just the art, but the capabilities. My son went in and created a pet adoption center. You could actually adopt the pets, and they’d follow you around.
Of course, I didn’t design this animal-following capability. The coding agent updated my game engine to create that mechanic. In the past, procedural level generation was really inspirational, but that’s just dynamic content. Now we’re talking about dynamic abilities and whole different mechanics, right?
7. His Son Rebuilt Ghost of Tsushima in Minecraft
Yes. Has your son played Adopt Me! on Roblox? It’s one of the most popular Roblox games.
No, something like that.
Is it similar?
Oh, maybe he has, and that’s how he came up with it. I thought, “This is so creative.”
For sure. I think that’s a great example. My son has also recently gotten into game development using AI. It started because he wanted to play a game called Ghost of Tsushima.
It’s a fantastic game, but you basically play a samurai in feudal Japan. As you can imagine, it’s very bloody and kind of violent.
I told him, “Actually, no, you’re 8. You’re not playing this bloody samurai game.” But he really wanted to play it. He started talking to ChatGPT and asking about it. Somehow, he discovered he could build a Minecraft mod of Ghost of Tsushima.
No way.
He used GPT to deconstruct the game, because there were all these wikis online breaking down how the mechanics and boss battles work. Essentially, Codex helped him build a Minecraft mod, which he installed. It’s Ghost of Tsushima as a Minecraft mod.
And it works.
It works really well. He’s been sharing it with his friends, and now he doesn’t ask us to buy new games anymore. He just says, “I want to play this game. I’m going to make it as a Minecraft mod.”
That’s a lot of inspiration. What I also love about that is that it illustrates the fact that, as we think about how games get better, realism isn’t the only, or even necessarily the most interesting, axis for improvement. The fact that you can create a concept from two existing concepts that never would have existed is extraordinary.
That’s right.
That’s right. I think this is a good transition to talk about decompilation. That feels like a new trend, and there seem to be different strains of it: having an agent look at something and try to recreate a game from scratch, or meshing and merging two games or art styles. I’d love to disentangle all the different things going on right now.
I hope there are two trends getting turbocharged by AI. The first is that the games industry has long been driven by modding. Modding means taking a game you already love and building an add-on or extension that changes the gameplay somehow. Then it becomes its own game, its own genre. Historically, some of the biggest games in the industry have come from mods.
Counter-Strike, for example, a really big esport, was a Half-Life mod. League of Legends, one of the most successful MOBAs of all time, was inspired by Defense of the Ancients, Dota, which was itself a Warcraft III mod. My favorite example is Fortnite and PUBG, which were inspired by a PlayerUnknown’s Battle Royale mod for Arma. Multibillion-dollar game franchises have been created using mods. That’s the history: some of the best and biggest games in the world have been created this way.
8. The Renaissance of Niche Genres
AI is making it really easy to mod games right now. Just like my son did, an ordinary player who loves an IP or a game can point Astra at it and say, “Make me Ghost of Tsushima, but add box crafting. Add the relationship system from Animal Crossing.” Then you can make it and see if it’s fun. If it is, other people start playing it, it becomes a genre in itself, and it can blow up in a big way.
Part of this goes back to the creation loop. It’s hard for someone to create something new using a blank canvas.
Yeah.
If you put a text prompt in front of someone and say, “Make me a game,” you wouldn’t know where to start. Working within the framework of an existing game someone else has made, and saying, “Add this thing; remix it with this other thing,” is powerful because it gives you a place to start and a place to remix from.
Yeah.
Yeah, it’s interesting. You mentioned modding. I think the two other areas I’ve seen are, first, decompiling so that you have the source code and the assets.
Yep.
I think this gets really interesting for beloved but old games. Take Knights of the Old Republic, for example. BioWare is out of business, and they’re never going to make another Knights of the Old Republic game. Extending it or forking it is really cool. The third, maybe most futuristic thing I’ve seen—
There’s this guy Blendy who’s been doing a lot of amazing things on X, BLNDI, where he takes the assets from two existing games and combines them into a third game.
You know, which is a totally outer-space concept, but it works. Some of the games are really fun.
Yeah. On the retro games front, I think what sparked the Xbox Labs announcement—when Asha talked about wanting to support AI and new frontiers of development, specifically—was that so many community projects were popping up that took Halo, the original 2001 Xbox game—
Classic.
—and decompiled and recompiled it for modern platforms. I heard about four community projects that made Halo a playable WebGPU game, so you can basically play Halo on the web now. It’s incredible, and it goes back to what you were saying earlier about Three.js. 3D on the web has never been more powerful or accessible.
I think Xbox has likely seen things like this happen and is saying, “We used to have whole teams that would take a Windows game and port it over to the iOS ecosystem.”
Or take old games that are defunct and try to remaster them. Now AI can help a lot with those efforts. You’re taking an established IP and exposing it to a new audience, like kids today who have grown up and never actually had a chance to play Knights of the Old Republic.
Yeah, but now they can.
That’s right. And potentially on a console, PC, or device that’s relevant to them.
Yep.
Yeah, I think it’s really cool. I also think there are genres of games with really high production costs, where historically you simply don’t get the ROI. People love RTS games.
Everybody has been obsessed with RTS. There have been so many efforts to make a great iPad RTS, and we’ve never really gotten one. Now I think we’re going to see a complete renaissance of genres like RTS.
Such a— It’s what I’m personally really excited about. I think Tim Sweeney talked about this too. I love this really niche genre of games called job simulators. It’s a strange concept: you basically play a day in the life of a particular job. It’s typically not a glamorous job. You’re not the president or anything. The one I played most recently is this train conductor simulator.
Oh my gosh, amazing.
You can basically be a conductor of subways and trains in various cities around the world. I’ve been a JR, like a Tokyo subway conductor. It’s exactly as it sounds: it’s not glamorous at all. You show up for work, you have a timetable to make, you have to drive the train and stop at every stop along the way to pick up passengers. You have to do it perfectly. If you’re late stopping, the last car doesn’t make it in, so not everyone can get on the train, and your paycheck is docked for that. It’s such a niche genre. I can’t imagine that many people want to be train conductors.
Yeah.
But if they are, now you can make games like this, and people like me who really love it—
Yeah.
—will pay a lot of money to play these games. Is it fun?
I find it really fun.
You know, it’s funny: if you look at old computing magazines and the rankings of top software, number one is usually VisiCalc and number two is something like Microsoft Flight Simulator, those early flight sim games. I actually think simulation as a genre really has a hold on people. People love sims.
That’s right.
Yeah, it’s just an evergreen kind of game.
Maybe there’s a particular type of job that resonates with different people, but maybe you never had a chance to discover it because the simulator genre just wasn’t that big.
That’s right.
But now you can role-play a day in the life of almost any job in the world that you can envision. Wouldn’t that be really cool?
Did you guys ever play Untitled Goose Game?
Oh, I know about this game. Yeah.
It’s a goose simulation.
That’s right. You pretend to be a goose and go around.
So maybe I am into the sim genre. Mischievous goose.
It was really big during the—
That’s right. Goose. I don’t know what that says about me.
9. Managing AI Agents Like an RTS Game
Chaos. Actually, going back and merging the conversation about RTS games, one thing I’ve seen is agent orchestration schemes that look like classic RTS games. There’s an Age of Empires skin over your agent orchestration.
Yes, yes. There’s a Minecraft one too. All the different structures are different, like, code—
10. When Creating Is the Game
Exactly. What do you guys think of those? Is that a toy kind of rendering? Is it just an extension of the fact that it feels like a game to manage your agents that way, or is it the start of a new genre of simulator? Being the general manager of a line of business at any company—Credit Karma, Google, or wherever—is a sort of sim. You’re working with all these teams, you have limited control and incomplete information, and then you have market forces.
I don’t know that it’s that different. I actually think that now you can trivially create these really ambitious interfaces. In that world, you might do that and it might actually be optimal, whereas in the old world the cost was so high it was insane. It made no sense.
Yeah, that’s right.
Sure.
I think one thing that GUIs, graphical user interfaces inspired by games and used for agent management, make me think of is that the average consumer finds agent orchestration non-intuitive. Working from a command line to manage your agents, using a primarily text-based interface, is non-intuitive. I agree there will probably be a more mass-market visual interface—
—that an everyday consumer can use to manage a collection of agents.
And whether it looks like StarCraft or Agent Empires or Factorio, I’m less sure of the exact implementation.
11. The Business of AI Games
But it does feel like we’re early in figuring out the user interface for primarily agent-work players.
Yeah.
Yeah. So, John, one thing I’m really curious to hear from you on: Tim and Asha have come out strongly supporting the community around AI-developed games, which is awesome. But do you think their companies and ecosystems are set up to capture the value from this trend, or are they in an awkward position today, knowing they need to change?
Yeah, I think it’s a good question. If you take modding as an example, historically the original owners of IP haven’t benefited that much from the mods created on top of their IP, right?
Mm-hmm.
The original creators of Dota didn’t really benefit that much from League of Legends when it was created. I’m not sure the original Half-Life developers earned that much from Counter-Strike. There’s a gap in the business models that needs to be figured out. If you’re really embracing AI as a remixing, modding, decompilation, and recompilation engine to empower your player base, how do you benefit commercially as a company?
I think folks are still early in figuring that out. I’ve heard people talk about trying to figure out whether there’s a revenue share or royalty that can go back to the company, so you’re empowering creators. There’s still a lot to figure out there.
I mean, the strongest form of that, in my mind, is Minecraft. Minecraft really benefits from the mods ecosystem. It’s interesting because we’ve used the words “remix” and “extend” interchangeably, but I think additions that extend the core game and share a core foundation of community, plus a base game, really benefit the IP owner. Whereas when you’re forking the game, there’s no benefit. In fact, you might even be splitting the community.
That’s right. That’s right.
Yeah. It goes back to something you and I have talked about separately: how do you create new forms of distribution for the next generation of AI games, or games created using AI? Today, many of these games still go through traditional distribution channels. They’re posted on Steam or the App Store, and if you’re making something in Roblox, it’s still within the Roblox ecosystem. There’s an opportunity for a company to create a distribution platform that natively harnesses and empowers all these AI game creators.
Yeah.
And that doesn’t really exist today, because all the games are going through traditional channels, right?
Well, you should check out Spawn, actually. We’re not investors, but I’m a huge fan of what they’re doing. It’s Jacob and the team, and they’ve created a shared community and game generation platform. You work with an embodied agent named Savvy. She’s in all the games with you. She can take bug reports on the platform, but she can also modify the game while you’re in it.
Like a copilot.
Yeah, that’s right. She can create portals for you, so you can step into other games.
Yep.
And you can bring other players with you. Every game is multiplayer by default, with the same community.
Mm-hmm.
It’s early, but it’s the most impressive, ambitious effort I’ve seen at doing something like this, and I’m sure there’ll be more.
Yeah. No, I think that’s exciting. I’m a fan of Jacob, and I’ve been following the company for a while. I think the broader opportunity is: how do you create a community that persists between games? You’re not just bringing a couple of people to play one game and a couple of people to play another, with no connection between them. That’s a huge opportunity. Minecraft and Roblox are great examples: once you’re inside the Roblox ecosystem, if you get bored of one game or you’ve beaten it, there are hundreds of thousands, if not millions, more experiences you can participate in.
Can you contrast the two, John? One thing I’ve noticed is that all the games in Roblox are extremely different, so the core shared game, if you even call it that, feels pretty thin. Whereas in Minecraft, the core shared game is very rich. Do you think you need a really rich game? It feels like pure portal-based worlds without a shared core gaming mechanic are harder to get off the ground. Or is that not true?
I think there’s a benefit to a Minecraft model where it’s more centralized, where everyone starts off playing the same game. To your point, you fork it, mod it, and make it your own when you create Minecraft mods. The benefit is that everyone has a shared experience and is familiar with the IP, and you’re departing from that. Minecraft mods are almost like fan fiction: everyone watched the same Game of Thrones series, so they know who Daenerys and Jon Snow are, what the Watch is, and so on. Minecraft has a lot of shared culture around the Enderman and the dragons.
That feels very different from a more decentralized approach, which is how I’d characterize Roblox. People can enter the Roblox ecosystem through any game. Like I mentioned earlier, there’s Jailbreak, Brookhaven, and a bunch of other games. You have a very different experience, and then you have to learn where to go from there and how to be a creator. Both approaches are totally viable. It’s a philosophical difference in how you want to extend the game and harness the community, and both can work.
12. What Excites and Scares Studios
Yeah, both can work for sure. Going back to our discussion about studios and publishers, what do you think they’re most excited by, and what are they most worried about right now?
I think getting their costs down has got to be super exciting, and being able to create more IP with less financial risk seems hugely exciting. For a lot of these studios, they live and die by winning the right to develop a game for a publisher. If the game doesn’t do well, they may not get that opportunity again. You have hundreds of people who are overworked like crazy. It’s got to be one of the worst qualities of life just being in that industry. People do it anyway because they love it, but it’s tough. I think the quality of life for the average person in the industry, and honestly the quality of the business, should go way up.
Yep. I think what might worry them is that Xbox—Microsoft, I guess, owns Minecraft—but they wish they had an asset like that, closer to a network and a social product than a pure entertainment product.
Yeah.
I’m sure Fortnite, despite being the most beloved game in the world five years ago, isn’t discussed as much now. There’s this holy grail, what you call a “never-ending game” or a “forever game,” but it’s so hard to get there. Maybe with these technologies, they’ll have a better shot.
Yeah.
In my conversations with game developers, I’d say they’re all using AI already. AI has become table stakes as a tool to be more productive. People use AI to generate game design documents, generate concept art, quickly iterate on an idea, and go to prototype. As I mentioned earlier, it shortens the distance and time between an initial idea and a prototype. Everyone is universally excited about that. I think of that as wave one of AI adoption among game studios.
Mm-hmm.
I think the next wave, which is what Anish is even more excited about, is how to use AI to create new types of gameplay: new game mechanics and new game genres that don’t exist today.
Yeah.
That could massively expand the market for games. You’re not just making GTA 6 faster and cheaper; you’re making a new type of game that couldn’t have existed before. The article I wrote earlier on the concept of the never-ending game is something I’m personally passionate about. I’d love to see it happen.
The holy grail for an MMO is a game that infinitely adapts itself to what players want. If you land on a beach in an MMO and you’re supposed to go to the forest because that’s how the game starts, maybe you really love the ocean and decide to stay on the beach, build a ship, and sail off into the ocean. The MMO today couldn’t handle that, because it’s scripted linearly to expect you to go into the forest, where something happens that sets you up on a quest.
In an adaptive MMO, one built using world models, for example, you could generate an entire quest line for someone who really likes the ocean. Now there are pirates attacking you, and there’s a nautical warfare element.
Sometimes these discussions go into science fiction. In *Ender’s Game*, they have the concept of the Mind Game. The game adapts to what Ender wants to do in any given moment. I think it’s possible we’ll see our own version of the Mind Game at some point. That would be a game you can’t stop playing because it adapts to whatever you want to do at any given moment.
I mean, perhaps we’re playing it at this moment.
Yeah, exactly.
The Matrix.
Yeah. Or *Ender’s Game* itself, right? I don’t even think—
The story is that he’s in the game. Yeah.
13. The Games They'd Finally Build
Yeah. What games have you guys put the most hours into?
For me, I’ve played *The Witcher 3*. I’ve completed it twice, half-completed it a bunch of times, and I absolutely love it. The other game I adore is *Dune 2000*.
Oh.
That was my favorite RTS growing up. It’s so good. Somebody has already ported it to play on Mac, and it’s as good as it ever was.
That’s awesome.
For RTS fans, if you haven’t played it, I highly recommend *Dune 2000*.
Yeah. I’ve probably put thousands of hours into *League of Legends*. Some of my best friends and friendships have been made in that game, so it has to be my most-played game. Close behind that is *StarCraft*. I’ve played all the versions, back to *Brood War*, *StarCraft II*, and so on.
It’s a classic example of a game where the audience for hardcore RTS games has dwindled over time because it’s such a hard genre to play. Your APM matters. You can’t play competitively without feeling bad about yourself because everyone else is so much faster and better at the game. There aren’t many developers trying to make new, modern RTS games because the audience size doesn’t support it if you look at pure market size and analysis. But with AI, I’m excited to see an explosion in RTS games. I’ll play as many of them as I can.
Nice. I’m a big *Crusader Kings III* fan, which isn’t really RTS. I describe it to people who don’t know it as *The Sims* mixed with *Age of Empires*. You get personality management—
Relationships, politics.
Yeah, exactly. But you also get this grand overworld where you manage your dynasty and your realm. It’s so deep.
Wow.
Anyway, to close us out, what’s a game each of you has been thinking about making for years that you can finally make now?
I can go quickly. I grew up running a bulletin board system, which tells you how old I am. Before the internet, people would dial in. BBSs were famous for having asynchronous multiplayer games because only one person could phone in at a time. You’d play your turn, phone up, connect, play your turns, log off, and then the next day play your next turns and see everyone else’s turns at the same time.
There was a whole genre of amazing asynchronous multiplayer games, including two written by one of Google’s earliest employees, when he was still at Stanford. His name is Amit Patel. He made a game called *Baron Realms Elite* and another called *Solar Realms Elite*.
Unfortunately, he lost the source code for *Solar Realms Elite*, which is the better of the two, but he has a game design document and an executable. For every generation of models, I’ve tried asking, “Can you recreate the game?” They haven’t been able to do it, but now I have what feels like a faithful recreation running on my computer.
Incredible. That’s a great story.
Yeah. Something along a similar vein for me: I’m a big *Civilization* player, and the original *Civ II* generated some of my most endearing childhood memories. It’s the 16-bit Windows version of *Civ* that you need a floppy disk to play. It’s not playable today anymore; it’s a 16-bit executable. I’d love to recreate that game on modern hardware.
I’ve also long thought about what *Civ* looks like with agents, because the fantasy of the game is that you’re the ruler of an empire. But that fantasy breaks down over the course of *Civ* when you’re micromanaging literally every unit in your empire. At some point, you could imagine *Civ* where every unit was an agent with its own agenda and goals. As the ruler of the empire, you’d set strategy and try to influence your subjects to grow in the direction you want. You can’t just have them do exactly what you want; you have to coax and influence them.
I have amazing news for you: this is *Crusader Kings III*. The game exists, and it doesn’t use frontier AI at all.
Is that true? It’s the same game?
Yeah, pretty much. You manage schemes, different people’s personalities, and their opinions of you. Sometimes people won’t do what you want because their opinion of you is too low. There are different ways to conquer a realm or territory. You can murder people ahead of you in the succession. It’s super deep. It feels closest to the never-ending, endless game you were talking about: what it might feel like when it’s super deep and realistic, with way more permutations of choices.
You should definitely check it out. During the pandemic, I got COVID at the end of 2020, started teaching myself C, and started making a Japanese-style dating sim with Unity.
Popular.
Yeah. It was right around Unity’s IPO, so I was curious about Unity and how it worked and how easy it was. It was basically branching dialogue trees, so it wasn’t that advanced in terms of programming. I didn’t really get that far. There was some basic management of health and thirst.
I actually think the dating sim genre is ripe for AI.
Yeah. There’s a genre called otome games, which are basically dating sims for women, where you go on dates with men of different personalities and archetypes. There’s a game called *Love and Deepspace*, made by a Chinese developer, that’s blowing up. It’s bringing in hundreds of millions in revenue.
The challenge is that it takes a lot of time and work to create the content, because the expansion packs for the games are basically characters. There’s a new guy you can date, and that’s an update for the game. They have to build out his rich backstory: he likes cats, walks on the beach, but is terrified of heights, and so on. You have to discover all of this.
There’s no reason they can’t be agents in the future. These are people with their own personalities. Their world is encoded in their database, and they react to you in real time as you have conversations with them. They know about your day. I think the dating sim genre is wide open and ripe for disruption by AI.
Yeah. Amazing. Fun note to end on.