No Discovery Needed

Browse a store. Read a review. Watch a trailer. Click buy. Discovery starts with something someone else crafted. The game waves at you across a canyon of noise. You wave back.

Discovery is the bottleneck of all media. Storefronts, algorithms, influencers, wishlists, publishers, platform deals: an entire industry built to bridge that canyon.

What if the game started with you? The ultimate solve for discovery would be no discovery at all: a game created from what you want to play, when you want to play it.

We've been here before

Games have generated content at runtime for more than forty years. Rogue did it in 1980: a new dungeon every run. Enter the Gungeon remixes its rooms. Minecraft generates terrain as you explore. No Man's Sky builds planets from procedural rules.

A dungeon nobody has seen before is nothing new. Generative models don't own that distinction, and they have constraints too. The question now is how much of a game can take shape around what you ask to play.

Google DeepMind's Genie 3 generates navigable worlds from text prompts at 24 frames per second. In January 2026, Google began opening access through Project Genie: a research prototype you could try, with a 60-second generation limit.^[1]

The pieces are also coming together inside games. In August 2026, Interact Studio combined Rosebud with video world models served through Reactor to build Subject 731. Reference images set each room's visual direction; the model generates the environment as the player moves through it. Puzzles and progression run in code. The world is generated. The gameplay carries the designer's intent.^[2]

Borges imagined a library containing every possible book within a fixed alphabet and format. Somewhere in it is exactly what you want to read, buried in almost endless nonsense. Having everything doesn't help you find it.^[3]

Generating more games could leave you in the same library. The promise is a game shaped by what you want to play. Whether it's worth playing is another question.

Not slop

The obvious objection: this sounds like AI slop with extra steps. “Slop” was Merriam-Webster's word of the year in 2025. By July 2025, 7,818 titles across the Steam library disclosed generative-AI use, not necessarily generated games. In GDC's 2026 survey, 52% of more than 2,300 respondents thought generative AI had a negative impact on the industry. Use is spreading. Trust is another matter. Neither figure tells us whether a particular game is any good.^[4]

Ted Chiang argues that averaging the choices other creators have made gives you “the least interesting choices possible.” In his deliberately simplified example, a 10,000-word story involves roughly 10,000 choices. Hand those choices to a model and you may get fluent prose with nothing to say.^[5]

A game can run, respond to your input, and still give you no reason to care. Generating it is only part of the problem. The choices that make it worth playing have to come from somewhere.

Voice as parameter

The creator's taste shapes more than the finished game. It shapes the system that makes it.

Pieces of this already exist. Midjourney's --sref uses an image reference or a style-library code to guide the aesthetic. Its --p profiles learn from images you like or select, and influence generation when enabled. The creative direction becomes an input.^[6]

In March 2026, Suno introduced custom models that musicians can tune to tracks from their own original catalog. The generator learns from the music they've already made, carrying elements of their style into new compositions.^[7] These controls carry selected qualities forward. They don't make every result worth keeping.

For games, the possibility goes further. Pacing. Difficulty. What the game rewards. What it refuses to make easy. A designer could define those choices without fixing every encounter in advance. The player's request shapes the experience within them.

Hideo Kojima describes using AI to adapt to players' habits and sense of control.^[8] The designer's choices remain; their expression responds to the player. On-demand generation takes that possibility further: the player's request helps shape the game itself.

That's voice as parameter. You choose a sensibility, not just a finished product. The game takes shape through it.

The trade

The boss fight you describe may not be the one your friend played. A speedrun needs a common challenge. A conversation about the twist needs a twist you both experienced. The more those moments diverge, the less you can assume you share.

Take-Two CEO Strauss Zelnick says generative AI plays “zero part” in Grand Theft Auto VI. He describes Rockstar's worlds as handcrafted, building by building, street by street.^[9] That's the appeal of an authored world you enter alongside everyone else: its places become common reference points.

Generated games can still share themes, mechanics, even the same saved configuration. Personalization doesn't have to mean isolation. A recognizable voice gives different experiences something in common. It doesn't replace having been there together.

Made for you

Ask Bootleg Arcade for “Space Invaders with a paddle and ball.” Change the request and summon another. You're not browsing a collection for the closest match. You're asking for a game through the cabinet's particular arcade sensibility.

Your request gives it direction. The creator's choices give it character.

That's the bet: less time finding something you want to play, more time playing something shaped by what you asked for. The creative voice stays recognizable. What comes through it is made for you.