Ben James, founder and CEO of Atlas and 404-GEN, has spent years helping AAA gaming studios like Square Enix and Parallel accelerate 3D asset production with AI. But after discovering Bittensor in early 2024, he saw something bigger — a chance to build an open, decentralized system for 3D model generation that could reshape how digital worlds are created.
In an exclusive conversation featured on Taopill, a newsletter powered by Unsupervised Capital, James discussed the vision behind 404-GEN, setting the stage for how decentralized AI can transform creative industries.
For James, 404-GEN isn’t just another AI lab — it’s a new kind of creative network. Running on Bittensor’s Subnet 17, 404-GEN leverages decentralization to incentivize global contributors to train, refine, and improve generative 3D models collaboratively. Its mission is bold yet focused: to democratize 3D content creation and unlock a new wave of AI-native virtual world-building.
A New Model for Incentivized Creation
When 404-GEN first launched, its incentive system rewarded miners — the AI models contributing to the network — based on familiar performance benchmarks such as generation speed, prompt similarity, and artifact detection. But as the subnet matured, these metrics became too easy to surpass.
The team pivoted toward a competitive ranking system inspired by chess’ Elo model, where miners now “duel” one another using identical prompts. Each duel is judged by a Visual Language Model (VLM) that determines which output best matches quality and creativity standards. Results are logged transparently on a public dashboard, with explanations from the VLM itself.
This design adds a layer of dynamic competition to 404-GEN — one that continuously pushes miners to improve and innovate. According to James, this is the first Elo-based evaluation framework used on any subnet across Bittensor, introducing an evolving meritocracy for generative AI.
Matching and Beating State-of-the-Art
The results have been impressive. The assets produced by 404-GEN rival, and in some cases surpass, outputs from leading closed-source systems. Today, the models can generate high-quality 3D assets suitable for production environments, particularly for secondary or supporting objects in AAA games.
While a few assets still require minor post-processing like rigging or modular adjustments, the quality gap is rapidly closing. 404-GEN’s models are already being used by developers to prototype detailed “hero” assets — the kind of centerpiece objects that define a game’s visual identity.
From Gaming to Global Creation Infrastructure
Gaming remains 404-GEN’s strongest proof of concept, with developers integrating the subnet’s models directly into their workflows through plug-ins and API access. The team’s integration with Unity, one of the world’s most widely used game engines, marks a major milestone.
Each month, hundreds of developers download 404-GEN’s Unity plug-in — the first decentralized 3D generation system officially verified by Unity. This allows creators to generate assets “in-engine,” directly within Unity’s interface, making it seamless for more than 1.2 million active Unity users to build with decentralized AI.
James sees this as more than convenience — it’s the foundation for adoption. By embedding 404-GEN into tools creators already use, the subnet reduces friction and makes decentralized AI creation feel native, not experimental.
Beyond gaming, 404-GEN’s reach is expanding into adjacent creative fields such as film, visual effects, and extended reality (XR).
The team is also exploring Gaussian Splat-based world models, which could power digital twins and simulation environments for industrial, educational, and training applications.
The Unexpected Frontiers of 404-GEN
What’s surprised James most is how far creators have taken 404-GEN beyond its original scope. At the Super AI Conference in Singapore, he met a VR developer who had built an entire game using 404-GEN assets after discovering the tool through the Unity store — without ever realizing it was running on Bittensor.
This kind of discovery highlights an important truth about the subnet’s mission: for many creators, the underlying blockchain architecture matters less than the quality of the assets it produces. The network’s real success lies in making decentralized AI invisible — a background engine that simply works.
404-GEN’s assets have even appeared in mainstream venues, from the Venice Architecture Biennale to the 2025 Osaka World Expo, and users are experimenting with physically 3D printing generated models or building design businesses based on its outputs.
The Growing 3D Dataset — and What Comes Next
Over time, 404-GEN’s miners have collectively generated more than 28 million high-quality 3D models, forming one of the largest decentralized 3D datasets in existence. This repository is already being shared with research groups across the U.S., Europe, and Asia to advance open-source model development.
James believes this dataset will become a core economic and research resource — one that can be monetized through curated asset packs, bulk licensing, and marketplace integrations, while also fueling a virtuous cycle of feedback and improvement within the subnet itself.
Toward a Self-Sustaining Subnet Economy
Economically, 404-GEN takes a pragmatic, market-oriented approach. Early adopters help identify the most valuable features, while the team iterates based on real usage data and industry demand. The subnet’s growing revenues are directed back into its native token — ensuring that network success translates into tangible value for the ecosystem and its participants.
In James’ view, this closed-loop model is essential for building a sustainable AI economy on Bittensor — one that rewards open collaboration, technical excellence, and real-world adoption in equal measure.
A Decentralized Future for Creative Worlds
404-GEN captures what a mature Bittensor subnet looks like in practice: a blend of deep technical innovation, clear market focus, and community-driven growth. By leveraging on decentralized infrastructure with creative industries, it points toward a new future — one where world-class AI models are not confined to corporate labs but built, refined, and owned collectively.
For Ben James, the vision is simple yet transformative: empower creators everywhere to build richer, smarter, and more open virtual worlds — one 3D asset at a time.
Be the first to comment