Novelty Search May 29, 2025

This episode explores how BitTensor's Subnet 17 is revolutionizing 3D content creation through decentralized AI, innovative incentive mechanisms, and the emerging dominance of Gaussian Splats, offering a new frontier for Crypto AI investors and researchers.

Understanding BitTensor's Incentive-Driven AI Development

• Const, the host, opens by discussing the unique nature of research on BitTensor, a decentralized network that incentivizes the creation and operation of AI models. He highlights a conversation with a researcher initially confused about studying a BitTensor subnet without knowing the miners' exact operations.
• The core insight, as Const explains, is that the "technology" itself is the incentive mechanism and the algorithm designed by the subnet creators. This is analogous to studying a neural network training algorithm rather than individual neuron activations.
• Const: "The purpose is to investigate the algorithm, the incentive mechanism that's actually the technology... the mechanism that you guys have invented."
• This framing is crucial for understanding how decentralized AI networks like BitTensor foster innovation by setting "rules of the game" and incentive designs, rather than centrally controlling AI model generation.

Introducing Subnet 17: Democratizing 3D Content Creation

• Ben, founder of the company behind Subnet 17 (also known as 404.17), introduces his team: Max (Tech Lead) and Monica (Marketing and BD Lead).
• Subnet 17 aims to provide an overview of their work, its importance on BitTensor, their incentive design, validation learnings after a year on mainnet (the live, operational blockchain), and applications built by third parties using their APIs.
• Ben shares that his company, older than BitTensor itself, originated from a Web2 background with EU research funding, focusing on 3D, gaming, film VFX, and metaverses.
• The BitTensor whitepaper resonated with Ben, offering a path to "democratization of content creation," moving beyond large studios to empower smaller indie developers by working on the foundational layer of the tech stack.
• Strategic Implication: Subnet 17's journey from Web2 to BitTensor underscores the potential for decentralized networks to disrupt established content creation pipelines, offering new avenues for investment in foundational AI infrastructure.

The Problem: Escalating Costs and Complexity in Virtual World Creation

• Ben outlines the core problem Subnet 17 addresses: as compute and graphics capabilities grow, so do demands for realistic virtual worlds, leading to escalating time and costs for developers.
• It's not just main characters; every 3D object in immersive experiences is typically textured and modeled by hand.
• Example: Grand Theft Auto 5 took ~5 years and $265 million. GTA 6's leaked budget is $2 billion.
• The gaming industry is valued at $300 billion, but Ben emphasizes that Subnet 17's applications extend to film, VFX, digital twins (virtual replicas of physical objects or systems), metaverses, and XR (Extended Reality, encompassing augmented, virtual, and mixed reality).
• Investor Insight: The immense cost and time investment in traditional 3D content creation signals a significant market opportunity for AI-driven solutions that can offer scalability and efficiency.

Why BitTensor for 3D AI: A Fragmented Landscape Poised for Growth

• Ben explains that the 3D AI space currently lacks a clear winner or dominant representation method, making it fertile ground for BitTensor's competitive, incentive-driven model.
• The three main 3D representation "protagonists" are:
• Meshes: Polygon-based structures, the industry standard for entertainment.
• Nerfs (Neural Radiance Fields): Used for high-fidelity reconstruction of existing structures, originating from a reconstruction background.
• Gaussian Splats: A newer method combining elements of meshes and Nerfs.
• Ben draws a parallel to the rapid growth of 2D AI and video AI, predicting 3D AI is "about to explode."
• Ben: "I think it's a really good time to be mining on our subnet because essentially what we have is this combination of there's no dominant solution. There are rapidly growing use cases."
• The lack of a dominant solution incentivizes miners on BitTensor to innovate rapidly. New research papers can see their innovations tested and deployed on the mainnet within days or hours, a speed unmatched by centralized competitors.
• Actionable Insight for Researchers: The dynamic and fragmented 3D AI landscape, coupled with BitTensor's rapid iteration cycle, presents a rich environment for research into novel 3D representations and generation techniques.

Deep Dive into 3D Representations: Meshes, Nerfs, and Gaussian Splats

• Ben elaborates on the characteristics of each representation:
• Nerfs: Standard for high-fidelity reconstruction, offering realism.
• Gaussian Splats: Gaining momentum in AI research communities.
• Meshes: Most compatible with existing industry pipelines.
• Subnet 17's team focuses on Gaussian Splats because they offer the high-quality reconstruction of Nerfs but with a fraction of the memory usage and much faster speeds, crucial for gaming and entertainment.
• Gaussian Splats also overcome limitations of meshes regarding geometric complexity and rendering pipeline issues with different levels of detail.
• Technical Explanation: A mesh is a polygon-based representation (e.g., triangles or quads) forming a shell structure with faces and vertices, textured by color data on vertices or faces. Achieving detail like individual hairs on a cat requires many polygons, making it computationally heavy.
• Technical Explanation: Gaussian Splats are described by Ben as "a point cloud on steroids." Instead of just XYZ coordinates, each point in a Gaussian Splat has its own unique scale and rotation in all three axes, plus color and opacity. This allows for the representation of complex geometry and appearance through an overlay of these transformed points, creating a realistic look without explicit surfaces. They are lightweight, generated quickly, and align well with AI techniques used for Nerfs.
• Const notes that Gaussian Splats match better with neural networks on the AI side. Ben adds that on the deployment side, they are poised to fundamentally change games and virtual worlds by allowing more unique geometry and larger scenes due to being less computationally intensive than meshes.
• While Gaussian Splats don't yet fully integrate into all existing workflows, their benefits are driving the creation of new tools and plugins for major pipelines like Unity and Unreal Engine. Subnet 17 has even developed and open-sourced its own plugins where needed.
• Const inquires about the hardware aspect, suggesting meshes might have dedicated hardware accelerators. Ben confirms this, noting it's an entire ecosystem built around meshes, from hardware to digital sculpting tools.
• Further Explanation of Gaussian Splats: Const asks for more detail on how a point cloud describes a surface. Ben clarifies that a traditional point cloud (often from scanning or photogrammetry – a technique using multiple photos to triangulate points in space) gives an approximation of an object's shape. Historically, these points were connected by faces to create a mesh. Gaussian Splats, instead, "stretch" and rotate each point (non-uniformly) and assign color/opacity. The overlay of these "little ellipsoids" gives the feeling of a concrete geometry that doesn't explicitly exist as a surface, similar to how a radiance field works. This makes them lightweight and aesthetically versatile. They can be converted to meshes if a solid object is required.

Subnet 17's Incentive Design and Validation Mechanism

• Ben addresses the challenge of evaluating 3D model quality, which has subjective elements. There's often no single "ground truth."
• Example: Two AI-generated "sturdy yellow drill cordless" models are shown, both plausible, making it hard even for a human to definitively pick the better one.
• Subnet 17's validation framework involves:
• Baseline Metrics: Traditional 3D AI research metrics like cosine similarity, structural similarity (for artifact detection in structure), and LPIPS (for texture artifacts).
• Exploit Proofing: Ensuring consistency in quality scores between models, using VLMs (Visual Language Models) for additional quality metrics and classifiers. VLMs are AI models that can understand and generate text related to images.
• Dynamic Rankings: An ELO ranking system (a method for calculating the relative skill levels of players in zero-sum games) that rewards/punishes models based on VLM evaluations in head-to-head "duels."
• Human-in-the-Loop: Due to the visual nature, humans can intervene at multiple stages to check outputs.
• The subnet has shown tremendous progress from early, "fun but not game-ready" models to highly realistic outputs.
• Validation evolution:
• Initial focus: Prompt adherence.
• Added: Objective quality (agreement, consistency, artifact detection).
• Current consideration: User-based constraints (generation time, image-to-3D input, support for mesh formats).
• The next iteration of validation will balance:
• Slow Validation: Computationally intensive, ELO-based head-to-head battles.
• Fast Validation: Keeps miners responsive and quickly rewards users, aligning with industry expectations.
• Ben explains the ELO system using the drill example:
• The VLM found no structural/visual artifacts in either drill.
• However, the right drill had significant teal color (prompt asked for yellow) and an unrequested base.
• The left drill was deemed better. Ben notes VLMs excel at relative ranking in head-to-head comparisons, even if absolute scores are less reliable.
• Strategic Consideration: The development of sophisticated, multi-faceted validation mechanisms like Subnet 17's ELO system is critical for ensuring quality and fairness in decentralized AI marketplaces, offering a model for other AI subnets.

Applications and Impact: From Open-Source Datasets to Major Exhibitions

• Even models that don't "win" in the validation process have value. Subnet 17 adds them to an open-source dataset for AI research, which has become the world's largest, attracting academic interest.
• Winning models find use in:
• Games: A gameplay trailer is shown where all 3D assets (cars, rockets, background/foreground models) except VFX were generated by Subnet 17, demonstrating physics compatibility in Unreal Engine (a popular game development engine). The assets showcased creative, non-realistic designs like dog-shaped cars.
• Immersive Worlds: An in-house project created a world with ~30,000 Gaussian Splats, published as a playable game to showcase the technology's capability for diversity and realism. Ben emphasizes this makes the tech understandable beyond Web3 audiences.
• Architectural Art: An AI-generated piece using Subnet 17 assets was exhibited at the Venice Architecture Biennale, a prestigious event. Ben, trained as an architect, highlights the significance. The piece, created by two non-3D artists in a few weeks using Unreal Engine, demonstrates the democratization of high-quality 3D content creation.
• Ben: "This tech is really about there are people who are not trained in this industry... and they're able to manifest kind of their crazy visions."
• World Expo: A similar piece is displayed at the World Expo in Osaka, exposing BitTensor and Subnet 17 to an estimated 28 million visitors unfamiliar with crypto.
• Actionable Insight for Investors: The diverse, high-profile applications (gaming, art, expos) demonstrate strong product-market fit and the potential for mainstream adoption of AI-generated 3D content from decentralized networks.

Unity Partnership and Democratizing Game Development

• Another game is showcased, built in Unity (another major game engine, popular with indie developers) by someone with no prior Unity experience.
• All game assets were generated on Subnet 17 (404). Game mechanics (car movement, boosts) were coded using Claude (an AI language model).
• Const asks about the specifics: assets like track pieces, arrows, skyscrapers, tunnels, and light posts were individually prompted and then manually composed in Unity.
• Achieving consistent style with text-only prompts is challenging, involving trial and error and descriptive prompts (e.g., "cyberpunk style"). This led to user feedback requesting image-to-3D generation for better style consistency, a feature launching on testnet next week.
• This means users can input an image directly into Subnet 17 to guide 3D model generation.
• Const contrasts this with traditional development: a similar game might take a small indie team (10-15 people) several months (e.g., 6 months), assuming a clear vision from the start. AI significantly speeds up iteration and exploration of creative ideas.
• Ben emphasizes AI's role in making creative processes more efficient, allowing smaller teams to achieve more or individuals to explore more concepts. This resonates even with creative industries initially skeptical of AI.
• Key Milestone: Subnet 17 became a "Verified Solution" by Unity. Ben explains this is a significant Web2 validation, involving extensive tech due diligence by Unity on their tech stack, incentive design, licensing, and quality consistency.
• Ben: "That was a really big web 2 testament that then is now generating direct [offset/benefits]."
• Strategic Implication: The Unity partnership is a major validation for Subnet 17 and BitTensor, bridging the gap between decentralized AI and established Web2 platforms, potentially opening up significant user acquisition channels.

Beyond Gaming: The Broader Horizon for 3D AI

• Const prompts Ben about applications beyond video games.
• Ben reiterates that while the $300 billion gaming industry is a major focus, Subnet 17 is seeing use and interest in:
• Digital Twins
• Film and VFX
• Metaverses
• XR (Extended Reality)
• CAD (Computer-Aided Design)
• Architecture
• Investor & Researcher Takeaway: The foundational 3D generation capabilities of Subnet 17 have far-reaching implications across numerous industries, suggesting a much larger total addressable market than gaming alone.

Reflective and Strategic Conclusion

This episode reveals Subnet 17's pioneering work in democratizing 3D AI on BitTensor, leveraging novel representations like Gaussian Splats and robust incentive mechanisms. Crypto AI investors and researchers should monitor the evolution of decentralized 3D content generation and its expanding applications, particularly the impact of direct Web2 integrations like the Unity partnership.