Context
Games are one of the hardest software products to build, because a single title needs code, art, game design, music and sound to come together. In this episode, a16z general partners Anish Acharya and Jonathan Lai talk about what happens as AI models start to write game code and operate creative tools directly. Both have been building games themselves with AI, and Jonathan's son has too, so much of the discussion comes from hands-on experiments rather than theory. For PMs, games are a useful stress test: they show what happens to a product category when the cost of making things collapses, who gets to create, and where value moves when users can remix and extend what you built.
The Big Idea
When AI shrinks the cost and time from idea to playable prototype, the bottleneck moves from production to taste, and the biggest new products come from users remixing and extending existing worlds rather than starting from a blank prompt.
The guests point to a long history of mods turning into billion-dollar franchises, and argue that AI makes modding and remixing available to almost anyone. The open question is who captures the value, because historically the original owners of a game rarely did.
Key Insights
1. Models now operate tools directly
Jonathan describes a model the hosts call Astra as a "huge unlock" because it could directly use Blender, Unreal and Unity, not just explain how to use them.
- Before: earlier models would tell a user "you should do this in Blender" and leave the work to them.
- Now: a hobbyist who has never built a game can have the model draft the game design document, generate concept art, build and rig 3D models in Blender, move them into Unreal, add generated music and sound, and export a playable build.
- Self-QA: the model can take screen captures of the game it made, spot that something is broken or looks wrong, and fix it. Jonathan says this is why early prototypes shared on X were "surprisingly good."
- A second path: Anish notes a separate trend, Anthropic's Opus models (he thinks 4.7 or 4.8) turned out to be very good at Three.js, the in-browser 3D library, laying out scenes element by element in a way that would be "maddeningly meticulous for a human." He sees coding and tool use as two trends converging.
The PM lesson is that "the model advises" and "the model does the work inside the user's existing tools" are different products, and the second one changes who can finish a project.
2. Cheaper builds unlock riskier bets
Long, expensive development cycles make studios conservative, and that is the structural problem AI attacks first.
- The cost of slowness: Jonathan cites Grand Theft Auto VI, in production since about 2013 or 2014, a 12 to 13 year cycle. By any other software standard, taking that long to ship HR or CRM software would be "kind of insane."
- Why it kills innovation: a producer who knows a game will take 10 years and hundreds of millions of dollars will green light sequels over new IP and stick to proven genres like shooters. He points to the 20th Call of Duty installment and GTA VI as examples of sequels winning.
- What changes: he hopes cycles drop from 5 to 10 years toward 2 to 5, and budgets from hundreds of millions to tens of millions. That is still expensive, but it lets studios fund new IP and first-time developers.
- The cancellation problem: Anish adds that many games die three or five years into a six-year cycle because the market and player tastes moved on.
- Niche becomes viable: genres that never had the return, such as real-time strategy (RTS) games on iPad or job simulators, become worth making. Jonathan, a fan of a train conductor simulator, says people like him "will pay a lot of money" for these niche games.
The flip side, raised by Anish, is that cheaper production can also go toward more ambition, a GTA-level studio with these tools could build "a universe" instead of a city.
3. The last 20% stays human
AI makes the first working version easy, but it does not make a game fun. Jonathan quotes an industry saying that the last 20% of a game is 80% of the work.
- The first 80%: a game that works. You can start it, there is a character, it can shoot things, there are monsters.
- The last 20%: how the world feels. A shotgun should sound different from a pistol, a laser gun needs its own recoil, and all the small details add up to something magical.
- Who owns it: both guests expect human taste and judgment to own this layer "for the foreseeable future." Anish adds that fun comes from many things at once: design, serendipity, culture and social dynamics.
For PMs building with AI, this is the shift in where the team's time goes: less on assembling the working product, more on the feel and judgment calls that make people want to spend a thousand hours with it.
4. Remixing beats the blank canvas
Jonathan argues that a text box saying "make me a game" is a poor starting point, because most people would not know where to begin. Starting inside an existing game gives users a place to start and something to remix from.
- Mods built the industry: Counter-Strike began as a Half-Life mod. League of Legends was inspired by Defense of the Ancients (Dota), itself a Warcraft III mod. Fortnite and PUBG trace back to the PlayerUnknown's Battle Royale mod of Arma.
- AI lowers the bar to mod: a player can point a model at a game and say "make me Ghost of Tsushima, except add voxel-based crafting" or add the relationship system from Animal Crossing, then see if it is fun.
- The real example: Jonathan's eight-year-old son was not allowed to play Ghost of Tsushima because it is violent. He used ChatGPT to break down the game's mechanics and boss battles from online wikis, then used Codex to build a Minecraft mod version of it. It works, he shares it with friends, and he no longer asks for new games, he makes them as Minecraft mods.
Anish takes a lesson from that story: realism is not the only axis of improvement. Combining two existing concepts into something that never existed can be just as valuable.
5. Users can now add capabilities
Anish built a record-shop game inspired by Habbo Hotel art, after a DJ friend complained that online record stores are just lists of songs. He wanted to recreate the social feel of 1990s Toronto record shops.
- What the model did: with a few pieces of reference art, Opus 5.5 built the room, pulled in 1990s R&B, dancehall, hip hop and soul, and had a DJ at the back play the tracks you picked.
- The extension: he expanded it into Queen Street in 1990s Toronto, and made every other storefront something another player could build out just by prompting.
- The surprise: his son built a pet adoption center where pets actually follow you around. Anish never designed that mechanic. The coding agent updated his game engine to create it.
His framing: procedural generation gave games dynamic content, but this gives them "dynamic abilities." For PMs, this means user customization can now reach the product's behavior, not just its appearance, which raises new design questions about what users are allowed to change.
6. Extending helps owners, forking splits
Historically, the companies that owned a game gained little from the mods built on top of it. Jonathan doubts the original Dota creators gained much from League of Legends, or Half-Life's developers from Counter-Strike.
- The gap: business models for AI remixing are not figured out yet. He has heard talk of revenue shares or royalties flowing back to the IP owner, but says much remains open.
- Anish's distinction: "remix" and "extend" are not the same. Additions that extend a core game and share its community and base game benefit the owner, and Minecraft is the strongest case. Forking a game gives the owner no benefit and may even split the community.
- Distribution gap: Jonathan notes AI-made games still ship through Steam, the App Store or inside Roblox, and sees room for a distribution platform built natively for AI game creators.
This is the core PM question for any AI product that lets users build on it: design for extension that keeps users in a shared base, or watch value leak out through forks.
7. Creation itself is the product
Anish argues that in this era "sometimes creation is consumption." Many games may only ever be played by the person who made them, and he says that is fine.
- Why it works: making a game is enjoyable on its own, especially seeing assets come to life and characters say things you did not program. He calls it a "Zen gardening" aspect of creation that he expects to grow.
- Precedent: Jonathan points to Minecraft and LEGO, where people create for fun and share with each other.
- The funnel: Jonathan expects most AI-assisted creators to build one level and drop out, but some share will fall in love with game development and turn it into a job, leading to "an explosion" in the number of game creators.
For PMs, this means success metrics for creation tools should not assume an audience. A project with one user can still be a successful session.
Mental Models & Frameworks
The 80/20 rule of polish
The games industry saying that the last 20% of a game is 80% of the work splits a product into two layers. The first layer is functional: it runs and the core loop works. The second is feel: sound, recoil, small details, depth and consequence. Use it to decide where AI should do the work (the functional layer) and where scarce human judgment should go (the feel layer). It also warns against judging an AI prototype as nearly done just because it works.
Two waves of AI adoption
Jonathan describes how game studios are adopting AI in two waves.
- Wave one, productivity: AI is already "table stakes." Developers use it to generate design docs and concept art, iterate quickly and shorten the time from idea to prototype. Everyone is excited about this.
- Wave two, new product types: using AI to create new mechanics and new genres that could not exist before, which could expand the whole market. The point is not making GTA VI faster and cheaper, but making a kind of game that was impossible.
PMs can use this to place their own roadmap: wave one makes existing work cheaper, wave two is where new categories and bigger markets appear.
Three strains of remixing
The guests break AI-powered remixing into three distinct activities.
- Modding: adding an extension to a game you love that changes the gameplay. Some mods become genres of their own.
- Decompiling: recovering source code and assets from old games so they can be extended, forked or ported. Anish's example is Knights of the Old Republic, whose studio will not make another. Jonathan notes several community projects made the original 2001 Halo playable as a WebGPU game in the browser.
- Merging: combining assets from two existing games into a third. Anish cites a creator on X who does this, and says some of the results are genuinely fun.
Centralized versus decentralized platforms
Anish contrasts two ways to build a platform for user creations.
- Minecraft (centralized): everyone starts in the same rich base game, then mods it. Jonathan compares mods to fan fiction where everyone has watched the same series, so shared culture (Endermen, dragons) carries across.
- Roblox (decentralized): each game is very different and the shared core is thin. People can enter from any game, then have to learn the rest of the ecosystem.
Anish wonders whether portal-based worlds without a shared core mechanic are harder to get off the ground. Jonathan says both approaches are viable and the choice is philosophical: how you want the product extended and how you harness the community.
The never-ending adaptive game
Jonathan, who has written about this idea, calls a game that adapts infinitely to what players want the "Holy Grail" for massively multiplayer games. Today a game that starts on a beach and expects you to go to the forest cannot handle a player who stays, builds a ship and sails off. An adaptive game, perhaps built on world models, could generate a whole quest line for that player, with pirates and naval warfare. He compares it to the "mind game" in Ender's Game, a game that you cannot stop playing because it bends to whatever you want at that moment. Anish adds that studios want a "forever game" that works more like a network and social product than a pure entertainment product.
Trade-offs & Nuance
Lower cost versus higher ambition
Cheaper production can be spent two ways. Jonathan focuses on spending it on risk: more new IP, more experiments, more first-time developers. Anish points out it can also be spent on scope, so a 12-year game could build a universe rather than a city. Both come from the same capability, and a team has to choose which one its savings go toward.
Democratized creation versus finished products
AI lets someone with only an idea and reference art build a prototype. Jonathan is clear that this is a prototype, not a fully playable game, and many people will stop after the first level. The opportunity is real but the completion rate is low, so a creation product should not expect every user to finish.
Game-style interfaces for agents
The host asks whether game skins over agent orchestration, such as an Age of Empires look over coding agents, are toys or a new genre.
- Anish: running a business line at a company like Credit Karma or Google already is a kind of sim, with limited control, incomplete information and market forces. Now that ambitious interfaces are cheap to build, a game-like interface might actually be the best one.
- Jonathan: the average consumer finds managing agents from a command line unintuitive, so a mass-market visual interface for managing many agents will probably appear. He is less sure it will look like StarCraft or Age of Empires, and says the industry is early in designing interfaces for agent workflows.
Common Mistakes
Mistake: Prompting from a blank canvas
Giving users an empty text box and asking them to "make a game" leaves most people unsure where to begin. Jonathan's better approach is to start inside something that already exists and let users add to it or remix it with something else, which gives them a starting point.
Mistake: Treating realism as the goal
It is easy to assume better games means more realistic graphics. Anish argues that realism is not the only, or even the most interesting, axis for improvement. Combining two existing concepts, like a samurai story inside Minecraft, can create something more valuable than more detail.
Mistake: Letting forks split community
Treating every user remix as equally good for the product ignores that forks can take users away from the base. Anish notes that extensions sharing a common game and community benefit the IP owner, while forks give no benefit and may split the community.
Practical Application
Map functional versus feel work
- Do: list the parts of your AI product's experience and tag each as functional (it works) or feel (it delights). Assign AI-generated work mainly to the functional list.
- Then: put your team's design and review time on the feel list, the equivalent of making a shotgun sound different from a pistol.
- Why it works: the games industry finds the last 20% of polish takes 80% of the work, and the guests expect human taste to own that layer for the foreseeable future.
Replace blank prompts with templates
Audit every place your product asks a user to start from an empty text box. For each one, offer an existing example, template or prior project that the user can modify or combine with something else. Jonathan's argument is that people struggle to create from nothing but are good at saying "this, but add that."
Decide your extension policy early
If users can build on your product with AI, write down which changes should extend a shared base (keeping users in one community, as Minecraft does) and which would fork it. Then decide how creators and the platform share value, since the guests say revenue-share and royalty models for AI remixing are still unsolved.
Revisit shelved niche ideas
Pull out ideas your team rejected because the market looked too small for the build cost. The guests argue that when AI cuts the time and cost to build, niches like RTS games or train conductor simulators become worth making, so re-run the math on each shelved idea with the new build cost.
Questions to Consider
- Which ideas did your team reject only because the build cost was too high for the market size, and would they pass if AI cut that cost sharply?
- If users could use AI to add new capabilities to your product, not just change its look, which changes would you allow and which would break it?
- Does your platform keep user-made extensions tied to a shared base and community, or does it make forking away easier than building on top?
- For the AI features you ship, which part of the work is "making it work" and which is the human-judgment layer of feel and detail that makes users stay?
- Are you measuring a creation tool by audience reach when many users may create mainly for the enjoyment of making something?
Bottom Line
AI is collapsing the cost of building, which makes creative risk, niche products and user remixing viable at scale. What stays scarce is the human taste behind the last 20% of polish, and the business model that lets the platform owner benefit when users extend rather than fork what it built.
Case Studies Mentioned
Mods that became franchises
Several of the biggest games began as player-made mods of other games. Counter-Strike started as a Half-Life mod. League of Legends was inspired by Dota, a Warcraft III mod. Fortnite and PUBG trace back to the PlayerUnknown's Battle Royale mod of Arma. The lesson is that users remixing a product can create bigger businesses than the original, but Jonathan notes the original owners rarely captured much of that value.
Spawn and its embodied agent
Anish highlights Spawn (a16z is not an investor), a shared community plus game generation platform. Players work with an embodied agent named Savvy who is inside every game with them. She takes bug reports, modifies the game while you are playing, and creates portals to step into other games, bringing other players along. Every game is multiplayer by default with the same community. Anish calls it early but the most ambitious effort he has seen at a community that persists across AI-made games.
Recreating a lost BBS game
Anish grew up running a bulletin board system, where async multiplayer games let each player dial in, take turns and log off. Amit Patel, one of Google's earliest employees, wrote Solar Realms Elite while at Stanford but lost the source code, leaving only a design document and the executable. Anish tried every generation of models to recreate it without success, and now has what feels like a faithful recreation running on his computer. It shows how capability jumps can suddenly make a previously failed task work, so it is worth re-testing old failures with each model generation.
Tools & Products
| Tool / Product | What it does | Why it was mentioned |
|---|---|---|
| Blender | 3D modeling, rigging and animation software | The guests say newer models can now operate it directly to build game assets. |
| Unreal and Unity | Game engines for building and exporting games | A model working directly inside these engines let hobbyists export playable games. |
| Three.js | A library for 3D graphics in the browser | Anish says Anthropic's Opus models turned out to be very good at it, giving a coding path to game generation. |
| Codex | An AI coding agent | Jonathan's son used it to build a Minecraft mod version of Ghost of Tsushima. |
| Spawn | A shared community and AI game generation platform | Anish's example of a distribution platform built natively for AI-made games. |
| Minecraft | A sandbox game with a large mod ecosystem | The guests' strongest example of an IP owner benefiting from user extensions. |
| Roblox | A platform of user-made games | The contrast case: a decentralized platform with a thin shared core. |
People to Follow
Anish Acharya
A general partner at a16z who builds games with AI in his spare time, including a 1990s Toronto record-shop game where other players can add new storefronts and mechanics by prompting. He brings a platform and business view, including the distinction between extending and forking a product.
Jonathan Lai
A general partner at a16z focused on games, who has written about the "never-ending game" that adapts to its players. He brings the studio and industry view, from development cycles and green-light economics to the two waves of AI adoption.
Experiments to Try
Experiment: Remix start versus blank start
- Hypothesis: users who start from an existing example they can modify will finish more projects than users who start from an empty prompt.
- Change: for new users of a creation feature, show half a gallery of editable starting projects and the other half an empty prompt box.
- Test: run it as an A/B test over a few weeks of new signups.
- Measure: share of users who reach a finished or shared project, and how many iterations they make before stopping.
Experiment: Let users add behavior
- Hypothesis: allowing users to add new capabilities by prompting, not just change appearance, will increase sharing and return visits.
- Change: for one bounded area of the product, let an AI agent add new behavior a user asks for, similar to the pet-adoption mechanic a coding agent added to Anish's game.
- Test: compare engagement for users with this option against users limited to cosmetic changes.
- Measure: creations shared with others, return visits, and how often added behavior breaks something that needs fixing.
Notable Quotes
"the last 20% of the game is 80% of the work and that's where all the magic comes together to make something fun." (Jonathan Lai)
"it's hard for someone to create something new using a blank canvas." (Jonathan Lai)
"realism is not the only or even necessarily the most interesting axis for improvement." (Anish Acharya)
"in the past, it was really inspirational to have procedural level generation, but that's just dynamic content. Now we're talking about dynamic abilities." (Anish Acharya)
"I really do think there's a sort of Zen gardening aspect to game creation, which will become increasingly popular." (Anish Acharya)
