Describe a Game. 404-Gen (SN17) Builds the 3D World, Tests It, and Critiques It

The 404—GEN (SN17) team demonstrated during Novelty Search E086 that its Titan model now generates game-ready 3D assets in 30 seconds against the several minutes competitors like Meshy require.

Titan is shipping inside Colony by Parallel, a game in early playtesting where players describe weapons and armor to an NPC (Non-Playable Character) and receive custom-generated equipment moments later. 

Beyond Titan, the presentation also covered the subnet’s transition toward code-generating miners and the first end-to-end world model experiment run internally.

What Titan Does Inside Colony

The player interaction inside Colony makes real-time asset generation feel natural, hiding a complex orchestration pipeline underneath the surface.

1. Players describe a weapon or armor to an in-game NPC: Conversational prompting fits the fantasy setting without breaking immersion for the user during gameplay sessions.

2. Titan generates the raw mesh with high geometric detail: Cat helmets, custom swords, and unique shields all emerge as detailed 3D models within seconds without requiring pre-built asset libraries.

3. Atlas remeshes the output down to mobile-friendly polygon counts: Colony runs on mobile, so heavy raw meshes get compressed while retaining visual quality through projected texture maps.

4. Pivot points and centers of gravity get set correctly: Weapons fit into the player’s hand, helmets fit onto the head, all without human intervention across the pipeline.

5. The whole loop completes while the player walks across the map: Generation happens fast enough that the finished asset appears in inventory before the next scene loads.

The end result feels like the game is genuinely generating unique content for each player rather than serving pre-authored variations from a static asset bank.

The Titan Versus Meshy Head-to-Head

Comparing Titan against Meshy (the closed-source incumbent) reveals the specific advantages that made commercial deployment possible for game studios.

Beyond the head-to-head comparison, miners rebuilt every dependency blocking commercial deployment and produced cleaner geometry than Trellis 2 delivered raw.

They also added multi-view input for complex assets and normal maps for lightweight mobile rendering.

What Comes Next: Code Generation and World Models

The current competition on SN17 shifted focus toward two different frontiers pushing the subnet beyond raw mesh generation entirely.

1. Miners now submit 3js code that generates assets: The output becomes structured, editable, lightweight, and immediately deployable inside web-based games without additional processing.

2. Web-based games can now use SN17 through simple agent prompting: Its public repo lets any developer point an agent at the code with instructions to build a specified type of game from scratch.

3. The first end-to-end world model demo already runs internally: A prompted game world produced with SN17 assets, a world model trained on that game, and an agent trained to play the world through the model.

4. Applications extend beyond gaming into robotics and simulation: World models trained through visual diversity carry direct implications for physical AI training pipelines beyond entertainment.

The subnet effectively covers three forms of 3D representation now (Gaussian splats, meshes, 3js code), giving it the broadest visual foundation of any AI research team working on world modeling today.

A Real Commercial Game Beats Every Demo Reel

Mobile gaming reshaped the industry through new business models, new player types, and new forms of interaction that older console developers refused to embrace.

AI-native gaming appears to be following the same trajectory, and 404—GEN is closer to the frontier than any other subnet operating on Bittensor today.

Players describing custom weapons in natural language and receiving them inside the game engine seconds later is exactly the interaction pattern no traditional game studio can match.

Whether the incumbents adapt or get displaced depends entirely on how quickly they integrate the tooling SN17 is now making commercially available.