Authors: Elena Burger, Anish Acharya, and Jonathan Lai, a16z
Compiled by: Deep Tide TechFlow
Deep Tide Editor’s Note: When disruptive technology enters an industry, its established players are often the first to feel the impact. But in a recent discussion, the a16z team argues that AI won’t destroy the existing landscape of gaming. Instead, it will accelerate the evolution of three key forces at once: independent developers, toolmakers, and world-building studios.
The article explores two dimensions—“new games vs. existing games” and “reducing costs vs. creating new possibilities”—to map out four quadrants of AI-powered game development. Through real-world examples including Blender modeling, code decompilation (Halo), and an individual bringing an old game back to life, it shows how a new generation of game origin stories is emerging as creative barriers fall.
To hear Elena Burger, Anish A, and Jon Lai share more on this topic, listen to the companion podcast for this article here.
Three archetypes in the games industry
If you want to understand an industry, the best approach is to study its most celebrated archetypes.
In the games industry, the first archetype is the indie developer. This category includes Eric Barone (creator of (Stardew Valley)) and Jonathan Blow (creator of (Braid), (The Witness), and (Order of the Sinking Star)).
There are also small teams like Tarn and Zach Adams, the co-creators of the epic world simulation game (Dwarf Fortress).
People love these developers and their games because, despite facing tremendous obstacles, they’ve achieved commercial success outside the sheltered confines of traditional studios. They’re the gaming world’s “nerds get the girl” underdog stories. The fact that these games succeed is part of their appeal—and, of course, they’re also genuinely fun to play.
The second archetype is the visionary toolmaker, represented by people like Tim Sweeney and John Carmack. Through hands-on programming experiments close to the hardware, they continually push the limits of its performance, expanding what others can create. Their work gives rise to engines, editors, and simulation systems, which eventually evolve into development environments that others can use to build things.
The final archetype is the world-building studio, represented by teams like Rockstar Games ((Red Dead Redemption 2) and the upcoming (GTA VI)) and Larian Studios ((Baldur’s Gate 3)).
They spend years building richly detailed, highly realistic game worlds filled with little secrets, often with more depth than any one player could discover.
For example, the wildlife in (Red Dead Redemption) includes possums that play dead and bears that bluff-charge, while (GTA VI) will have six in-game radio stations, including one hosted by a fictional character voiced by Lana Del Rey and her husband. People love these games because they feel incredibly real and alive.
These archetypes point to different values: indie developers can devote themselves to refining a distinctive vision and become famous when it resonates; visionary toolmakers push the boundaries of technology; and world-building studios create games of remarkable depth.
Usually, when a dauntingly disruptive technology enters an industry, these established archetypes are among the first to feel its impact. But AI may be different. In fact, AI’s most compelling promise in games is that it can accelerate all three archetypes at once.
The same technology can help an individual bring a wild idea to life, help engineers analyze a game’s underlying structure, and give a large studio with hundreds of specialists more opportunities to try out new ideas during production.
This was also a recurring theme in a recent conversation about AI and games between the three of us—Anish, Jon, and me. Each of us has several games we’ve wanted to make but put off for years, for different reasons: perhaps we lack a particular technical skill, or it’s hard to keep a large project going over the long term while working a full-time job unrelated to game development.
Lowering these barriers means that we—and developers far more professional than us—can try out more ideas and raise the ceiling on the ambition of our projects.
Four quadrants of AI-powered games
These opportunities fall along two dimensions: whether AI is helping with an existing game or creating a new one, and whether its main contribution is lowering costs or enabling new possibilities.
For aspiring indie developers, the opportunity is to bridge the gap between having a distinctive idea and having all the skills needed to express it fully. Even if someone excels at one aspect of game design, such as designing the core gameplay loop, they may still be unable to make the game they want if they lack skills in other key areas.
On our podcast, Jon described how agents can work across disciplines: helping write design documents, build assets in Blender, assemble scenes in an engine, inspect screenshots, and try to fix problems. So a solo developer can use tools in the “Produce” quadrant to quickly build a playable prototype and discover sooner what’s genuinely fun about an idea.
Of course, there’s still a lot of work between finding a spark and finishing a polished game. Anish’s experimental record-store project, inspired by (Habbo Hotel) and mentioned on the podcast, grew out of a complaint from a DJ friend: online record stores all looked like cookie-cutter e-commerce sites.
Anish wanted to create something closer to the experience of browsing record stores in Toronto in the 1990s, with loud music, mixtapes, rave flyers, and chance conversations.
Anish gave the model visual references from (Habbo Hotel) and began recreating that experience. Since production assistance was used to build a completely new game from end to end, the record-store project sits in the “Invent” quadrant of the chart.
Eventually, Anish made his record store and the surrounding streets moddable, so other players could build their own shops on the street. His son added a pet adoption center (which, Jon pointed out, may be a nod to the popular Roblox game (Adopt Me)).
The animals in the record store can be adopted and follow their owners around—behavior Anish hadn’t included in the original project. He says the coding agent added this feature automatically and made the existing record-store codebase backward-compatible with the new mechanics. This is “Extend”: players bring new ideas into a world they already know.
Jon’s eight-year-old son offers a smaller, more contemporary example. He wanted to play (Ghost of Tsushima), but Jon thought the game was too violent. So he started asking ChatGPT about its mechanics and boss-fight design, then used Codex to help make a (Minecraft) mod inspired by it.
Many of these emergent model capabilities, such as operating Blender or using 3D libraries like Three.js and Spark.js, have really only taken shape over the past few months.
When OpenAI released Astra, people were immediately struck by its sophisticated computer-use capabilities: it could carry out long-horizon tasks in graphics software and game engines like Blender, Unity, and Unreal.
Anthropic’s Opus models, meanwhile, have been especially impressive at end-to-end game design and implementation. A few months ago, you might have only been able to ask a model to generate some usable assets or a game design document.
Now, you can practically just type, “Make this game more fun,” and the game really does get more fun. Simon Willison wrote a great post exploring these developments in more detail.
Toolmakers and the decompilation wave
The steady improvement of model capabilities gives us a useful entry point for exploring the archetype of the “visionary toolmaker” and its role in the age of AI. AI is getting better and better at putting games together, and just as remarkable is how good it is at taking them apart.
This process is called decompilation: reverse-engineering a compiled program to recover source code that people can inspect and modify. Doing this also calls for creative engineering practices.
In Maurice Heumann’s account of an AI decompilation project, he directed a group of agents to decompile a first-person shooter. At first, the agents generated code that was readable, but didn’t behave correctly when run.
This led his team to introduce a validation process: recompiling the reconstructed functions and comparing them with the original binary, while accounting for references whose memory addresses might change. It’s a vivid illustration of a new generation of technical pioneers: as project overseers, they design constraints that allow agents to solve problems reliably. This is a classic example of the “Recover” quadrant.
On the podcast, Jon noted that, by his count, there are already at least four decompiled versions of (Halo) running in browsers. We’re fully prepared for the gaming world’s new benchmark to shift from “Can it run (Doom)?” to “Can it decompile (Halo)?”
Big studios’ attitudes and new opportunities for studios
For big studios, it would be easy to reject AI outright, as many players have done in recent months. It would be just as easy to feel anxious about the growing wave of decompilation and resort to legal action. That’s why it’s so intriguing to see major studios show genuine openness and excitement.
In October this year, Tim Sweeney outlined an organizational restructuring focused on AI-assisted development, intended to help employees create more and better work and extend those advantages to teams using Unreal Engine. Meanwhile, Xbox announced the formation of Xbox Labs, focused on rebuilding games for new hardware, developing browser ports, and creating mashup remasters.
Xbox formed a dedicated team to explore these possibilities and share tools. The significance of this move is that established giants see opportunities both to expand production of new titles and to breathe new life into their existing game libraries.
For world-building studios, AI offers a twofold opportunity: it can dramatically shorten development cycles and help them rethink what’s worth attempting in the first place. As Jon pointed out on the podcast, investing years and hundreds of millions of dollars in a game can push producers toward familiar, established IP.
AI opens up new possibilities by shortening the time from idea to prototype. By compressing development cycles, studios can prototype more game ideas faster, taking greater risks on original IP or experimental genres—projects that would have struggled to get the green light under traditional development timelines.
Creation itself is a kind of play
The three of us are now making games ourselves. At the end of the podcast, we talked about where we’d like to take this creative freedom. Anish has been working to resurrect (Solar Realms Elite), a multiplayer game he played in the early days of BBSes, whose source code has long since been lost.
Fortunately, a design document and an executable survived. He’s kept at it, restarting the project with each new generation of models. Now he has a version running on his computer that he describes as a highly faithful recreation.
Jon wants to revisit (Civilization II) and give the citizens in the game their own agendas, turning the rule of an empire into a process of constantly persuading different factions to cooperate. Elena, meanwhile, is considering reviving some old-school dating sims she made in Unity before the age of AI.
The most obvious criticism of this current wave of AI-generated games is that they might be fun for their creators and a small group of friends, but we’re still a long way from truly vast and coherent AI-generated games.
Critics might say these early AI games are fun more because of the novelty of “making a game yourself” than because of their entertainment value. But is that really a bad thing? From building with LEGO bricks and doodling with crayons to creating worlds in (Minecraft), making things has long been a form of play.
This means we’re witnessing one of the things that has always fascinated us most about games: the emergence of a grassroots ecosystem of how-to guides, tacit knowledge, and creative exchange. The history of games is full of stories about people discovering something brilliant by tinkering with someone else’s work. The most famous example in modding may be (Counter-Strike), which began as a 1999 mod for (Half-Life).
We can take this history seriously without expecting every (Minecraft) mod or vibe-coded game to become the next (Counter-Strike).
Ultimately, there will be many more games. Someone who gets into development by rebuilding an old game might develop a deep interest in the underlying systems—after all, it’s hard to really get started with decompilation without learning what a game binary is. Someone who builds a game world for friends might keep expanding it for years and eventually attract a huge player community.
These are classic plotlines and origin stories for industry archetypes we already know and admire. In an industry so fascinated by “obsessed people making incredible things,” we should be thrilled to see more people getting the chance to get started.