SN17 - 404GEN :: Decentralized AI generated 3D assets

This episode of Novelty Search offers a deep dive into 404GEN (Subnet 17 on Bittensor), revealing how decentralized AI is revolutionizing 3D asset generation for gaming, film, and beyond, and what this means for Crypto AI investors and researchers.

The Genesis of 404GEN and the Bittensor Vision

• Ben, founder of the company behind Subnet 17 (404GEN), introduces the team, which has grown from four to about ten members, combining expertise from research, 3D, gaming, film VFX (Visual Effects), and metaverses.
• The company, older than Bittensor itself, initially focused on Web2 solutions for AAA game developers (a term for high-budget, high-profile game developers) using image processing and computer vision.
• Ben explains his attraction to Bittensor: "When I saw the the white paper for Bittensor, what I realized then was that it was an opportunity to do more like democratization of of content creation." This shift aimed to empower smaller indie developers and focus on foundational tech.
• The discussion touches on the unique challenge of researching Bittensor subnets, where miner anonymity is common. The focus, as Constant (the host) and Ben agree, is on investigating the "algorithm, the incentive mechanism that's actually the technology," rather than individual miner actions.

The Multi-Billion Dollar Problem in 3D Content Creation

• Ben highlights the escalating demands for virtual world complexity, leading to soaring development time and costs.
  • Grand Theft Auto 5 took ~5 years and $265 million.
  • The leaked budget for GTA 6 is a staggering $2 billion.
• This isn't just about main characters; every object in an immersive world requires manual modeling and texturing.
• The gaming industry is valued at $300 billion, but 404GEN's scope extends to film, VFX, digital twins (virtual replicas of physical objects or systems), metaverses, and XR (Extended Reality), which encompasses augmented, virtual, and mixed reality.
• Strategic Implication: The immense cost and time bottleneck in 3D content creation presents a massive market opportunity for AI-driven solutions like 404GEN, promising significant efficiency gains and cost reductions.

Why Bittensor is the Ideal Platform for 3D AI Innovation

• The 3D AI landscape is currently fragmented with no dominant model or representation. Key contenders include:
  • Meshes: The industry standard, polygon-based representations (think triangles or quads forming a 3D shape's surface).
  • NeRFs (Neural Radiance Fields): Originating from reconstruction, these models excel at creating high-fidelity 3D scenes from 2D images by learning a continuous volumetric scene function.
  • Gaussian Splats: A newer technique representing scenes as a collection of 3D Gaussians (ellipsoid-like shapes with color and opacity). They offer benefits of both meshes (editability, speed) and NeRFs (realism, AI-friendliness).
• Ben notes, "3D AI right now there is no clear winner... that's really good for Bittensor... when there's no dominant solution because it just means that you're you're really incentivized to find the best um the best solution."
• The rapid evolution of open-source models allows Bittensor miners to quickly test and deploy innovations, a speed centralized competitors struggle to match.
• Investor Insight: The lack of a dominant 3D AI solution means the field is ripe for disruption. Subnets like 404GEN on Bittensor are positioned to rapidly iterate and potentially define future standards, offering high growth potential.

Deep Dive into 3D Representations: Meshes vs. Gaussian Splats

• Constant prompts Ben for a clearer distinction between meshes and Gaussian splats.
• Meshes: Composed of polygons (faces connected by vertices) forming a shell structure. Textures are applied via color data on vertices or projected onto faces. Representing fine details like fur on a cat requires an immense number of polygons, making it computationally heavy.
• Gaussian Splats: Described by Ben as "a point cloud on steroids." Each point has a unique scale, rotation (in X, Y, Z), color, and opacity. These "little ellipsoids" overlay to create the illusion of solid geometry without explicit surfaces.
  • They are lightweight, can be generated quickly, and leverage AI techniques from NeRFs.
  • Ben states, "on the deployment side, I think our team is really bullish that this is going to fundamentally change a lot of games and virtual worlds."
• While Gaussian splats offer advantages in fidelity, memory, and speed, they are not yet the industry standard. 404GEN has developed open-source tools to convert splats to meshes for compatibility with existing pipelines like Unity and Unreal Engine.
• Researcher Note: The shift towards representations like Gaussian splats, which are more amenable to AI processing and offer rendering efficiencies, is a key trend. Research into hybrid models and efficient conversion tools is critical.

Crafting Incentives and Validation in a Subjective Domain

• Ben addresses the challenge of evaluating 3D model quality, which has subjective elements. The example given is a "sturdy yellow drill cordless," where two different AI generations could both be valid yet distinct.
• 404GEN's validation mechanism has evolved:
  1. Initial Focus: Prompt adherence (does the model match the text request?).
  2. Adding Quality Metrics: Incorporating LPIPS (Learned Perceptual Image Patch Similarity) for texture artifact detection and structural similarity metrics.
  3. User-Based Constraints: Factoring in generation time (crucial for UGC - User Generated Content) and new input methods like image-to-3D (coming to testnet).
  4. Dynamic Evaluation with ELO: Implementing an ELO ranking system (a method for calculating the relative skill levels of players in zero-sum games) where models are battled head-to-head.
     • VLMs (Visual Language Models) are used to judge these battles. Ben notes VLMs are "very good at a relative ranking for two models head-to-head" but not for absolute scoring.
     • The system will use "fast validation" for quick user feedback and "slow validation" for more computationally intensive ELO battles.
• Actionable Insight: The sophisticated, multi-faceted validation system combining objective metrics, VLM-based comparisons, and user constraints is crucial for driving quality. Investors should look for subnets with robust and evolving incentive mechanisms that resist exploits and genuinely reward value.

Showcasing 404GEN: Real-World Applications and Impact

• Open-Source Dataset: Even non-winning models from the subnet contribute to what Ben claims is "the world's largest data set for AI research" in 3D models, fostering further innovation.
• Gaming:
  • A gameplay trailer in Unreal Engine (a popular game development engine) showcased cars and other assets generated by 404GEN, complete with physics.
  • "Splat Fest": An in-house game featuring ~30,000 Gaussian splats, demonstrating the diversity and realism achievable.
• Art and Architecture:
  • An AI-generated piece exhibited at the prestigious Venice Architecture Biennale, created by two non-3D artists using 404GEN assets in Unreal Engine. Ben, trained as an architect, emphasizes the significance: "This is exactly where I think this tech should be deployed... for people who are not trained in this industry."
  • A similar exhibit at the World Expo in Osaka, exposing Bittensor and Tao to millions.
• Indie Game Development (Unity):
  • A game created in Unity (another major game engine) by a developer new to the engine, using 404GEN assets for all 3D models and Claude (an AI assistant) for coding game mechanics.
  • 404GEN is a Unity Verified Solution: This involved extensive tech due diligence by Unity, a major Web2 validation. Users can make direct API calls to the 404GEN mainnet from within Unity.
• Strategic Implication: These diverse applications, especially the Unity partnership, demonstrate strong product-market fit and the potential for widespread adoption, moving beyond crypto-native users.

Roadmap, Future Vision, and the AI-Native Media Landscape

• Gateway API: Recently published and open-sourced, designed for scalability, better decentralization, and API key management. Ben suggests other subnets could benefit from this work.
• Upcoming Features: 2D-to-3D generation (for style consistency), enhanced mesh conversion (Blender, Unity), MCP (Multi-Chain Protocol/Platform) integrations, and "vibe coding" of games (using AI to assist in game logic creation).
• Vision: Ben envisions an "AI native like fusion of gaming and media," comparing the current state of 3D AI to early television broadcasts that merely replicated radio.
  • Future possibilities: Infinite, real-time generated worlds; personalized game experiences.
  • This future involves a Bittensor ecosystem: "LLMs, you know, inference for diffusion models, storage, compute, that's where you have a a Bittensor ecosystem."
• Investor & Researcher Focus: The development of robust APIs and a focus on integrating with existing Web2 platforms (like Unity) are key drivers for adoption. The vision of an interconnected Bittensor ecosystem powering next-gen media is a long-term bullish signal.

Q&A Highlights: Challenges and Opportunities

• Competition from Centralized AI (e.g., OpenAI): Ben acknowledges the possibility of large models generating entire 3D worlds. He sees potential for hybrid solutions where 404GEN could provide anchoring 3D context to video generation models, mitigating issues like context loss in pure video approaches.
• Scalability and Throughput: While initial tests focused on throughput (generating millions of models for their dataset), the current emphasis is on quality, allowing miners ~30-45 seconds per model. The new Gateway API is designed to handle future traffic increases.
• Web2 Competitors: 404GEN sees itself at the forefront, with current quality often surpassing closed-source, VC-funded alternatives that price assets around $10-$25.
• Copyright: A significant concern, especially with Unity. 404GEN employs prompt filtering and a "digital handshake" with miners to encourage the use of commercially viable, legally sound models.
• Gateway API Technicals (Max, Tech Lead): The Gateway API uses OpenRAFT (a consensus algorithm implementation) and a DHT (Distributed Hash Table) to involve all validators in handling organic traffic, aiming to solve common scaling and consensus issues in subnets.
• Monetization: The current Northstar metric is adoption, following Midjourney's strategy of building a strong user base before aggressive monetization. Future revenue could come from programmatic shares in successful projects using 404GEN assets or a curated asset store.
• Miner Behavior: Ben theorizes two types of miners: innovators pushing state-of-the-art AI, and those brute-forcing hyperparameters. The ELO system is designed to be robust against simple optimization, pushing for genuine quality improvements.

Conclusion: The Future of 3D Content is Decentralized and AI-Driven

• 404GEN's progress on Bittensor demonstrates AI's transformative potential in 3D asset creation, validated by real-world adoption and Web2 partnerships. Investors and researchers should monitor the evolution of its incentive mechanisms, API development, and its role in fostering an interconnected AI-native media ecosystem.

Actionable Takeaway: The development of sophisticated validation like ELO systems and robust infrastructure like the Gateway API are critical for decentralized AI networks. Tracking 404GEN's adoption metrics, particularly within platforms like Unity, will be key indicators of its market penetration and long-term value.