Novelty Search May 25, 2025

This episode of Novelty Search (recorded May 25, 2025) unpacks the dramatic BitTensor chain halt and recovery, alongside a deep dive into the Video (VIA) subnet's ambitious plan to revolutionize video processing using AI-driven compression and upscaling, offering critical insights for Crypto AI investors on network resilience and real-world application development.

The BitTensor Chain Crisis: A Test of Resilience and Decentralization

• Const, a host of Novelty Search, opened the discussion by recounting a recent, "extremely traumatic" event where the BitTensor chain halted, ceasing to finalize blocks. This occurred approximately two days prior, around 1:00 AM his time.
• He emphasized the importance of building on a blockchain like Substrate (a framework for building customized blockchains, often used in the Polkadot ecosystem) for its consistent, verifiable history. "Everyone can see that nothing has been modified, nothing nothing has been tampered with," Const stated, highlighting the transparency even when "safe mode" interventions are necessary.
• The current BitTensor network operates on Proof-of-Authority (PoA), where a fixed set of 20 validator nodes, all controlled by the Opentensor Foundation (referred to as "F"), authorize transactions. Const described this as a "pseudo form of decentralization," resistant to single-node failures but not fully decentralized.
• Strategic Implication: F's long-term plan is to transition from these F-controlled validators to a broader set of ecosystem organizations, eventually moving to a full Proof-of-Stake (PoS) system. This event, Const noted, will stimulate this transition. Investors should monitor this decentralization roadmap as it impacts network security and governance.
• The root cause of the halt was the validator nodes' inability to reach consensus. The team was initially "blind to the failure," lacking sufficient logs to diagnose the issue within the block transition function or finalization process.

The Anonymous Tip and the Vulnerability

• An anonymous individual contacted the F team, offering information about the chain halt in exchange for a bounty. Initially skeptical due to common scam attempts, the team engaged and found the information legitimate.
• The individual had discovered a serious vulnerability in how transactions were validated on BitTensor, allowing malicious transactions to cause a loss of consensus. Const refrained from detailing the vulnerability, as it was "not currently patched entirely across the ecosystem," deeming disclosure irresponsible.
• Actionable Insight: This incident underscores the ongoing security challenges in blockchain infrastructure. The reliance on external, ethical hackers (via bounties) can be crucial for identifying and patching critical vulnerabilities.
• After receiving the information and paying the bounty, F developers patched the fix deep within the Substrate codebase. However, a "poisoning" effect occurred where existing nodes across the network broadcasted old, problematic consensus rounds, preventing the patched network from finalizing.
• Const highlighted the difficulty of fixing issues in a layered stack like Substrate, where F doesn't control all components. He commended developers, specifically "VO from Leighton and JIP from the heads of perspective Nucleus and Medulla teams," for working nearly 48 hours straight to implement a manual soft fork and restore consensus.

Introducing Video (VIA): A New Subnet for AI-Powered Video Processing

• With the chain stable, the episode transitioned to introduce the Video (VIA) team and their subnet (Subnet 85), focused on AI-driven video processing. Mog, an advisor to Video and associated with Talstats, introduced the team.
• Mog recounted how Gareth, Video's team lead, unexpectedly presented a well-developed business plan for a subnet. Mog emphasized his decision to support Video to gain "in the trenches" experience with subnet development, crucial for advising future builders.
• Gareth then took the lead, explaining the core problem Video addresses: the spiraling costs of video delivery and storage, with 85% of internet traffic being video-based. Every video requires processing (upscaling, downscaling, compression) before reaching the viewer.
• Technical Term: CDN (Content Delivery Network) refers to a geographically distributed network of proxy servers and their data centers, which work together to provide fast delivery of Internet content.
• Video's solution focuses on AI-driven video upscaling (enhancing low-resolution content to HD or 4K) and AI-driven video compression (reducing file size without compromising perceptual quality).
• Gareth highlighted that the current centralized video processing industry is "closed off, slow to innovate, and really expensive to use."

Deep Dive: Video Upscaling on Subnet 85

• Gareth explained that video upscaling is currently live on Subnet 85 (in beta). The workflow involves sending low-resolution files to miners, who upscale them. Validation uses methods like VMAF (Video Multimethod Assessment Fusion), a Netflix-developed quality metric, and PIAP (Perceptual Image Assessment Platform).
• "The real important thing for us here is the perceptual quality," Gareth emphasized, contrasting it with traditional mathematical verification. AI models trained on large datasets assess how the video appears to the human eye.
• For organic uploads without a reference file, Video uses CLIP IQA+ (Contrastive Language-Image Pre-training Image Quality Assessment), another perceptual quality metric.
• Const inquired about the limits and applications of upscaling. Mog provided a real-world example: upscaling old '90s skate movies shot on SD (Standard Definition) Hi8 tapes for use in modern HD or 4K documentaries, making previously unwatchable footage usable.
• Gareth clarified that while upscaling is valuable, "definitely the video compression is really where the real business lies," due to its potential for massive cost savings in bandwidth and storage for content platforms.
• Strategic Implication: The Video team's initial benchmarks (using their base miner model) showed their upscaling quality (measured by PIAP) as superior to competitors like Topaz, FFmpeg, and Headpaw. Miners on the network have since improved upon this base model. This demonstrates the power of BitTensor's incentivized competition.

Groundbreaking Video Compression Technology

• Ahmed, Video's Machine Learning Engineer, presented their novel video compression model, which is not yet live on the subnet but is in a proof-of-concept stage.
• The core problem with existing encoders (like H.264, a widely used video compression standard) is their "one solution fits all" approach. Video's technology analyzes video content scene-by-scene, using ML models to predict optimal encoder parameters for each scene, significantly improving compression efficiency without losing perceptual quality.
• The process involves:
  • Preprocessing to remove artifacts (noise, blurring).
  • Splitting video into scenes.
  • Analyzing scenes for complexity (textures, motion, contrast).
  • Classifying scene types (e.g., talking heads, nature, gaming) using ML.
  • An ML model then determines the best encoder parameters (for codecs like H.264, HEVC (High Efficiency Video Coding), or AV1, an open, royalty-free codec) based on this analysis and desired output quality.
• Const asked why not use an end-to-end neural network for compression. Ahmed explained the critical "compatibility issue": end-users would need the exact same large ML model to decode the video, which is impractical. Current industry encoders are highly optimized and widely supported. "What we do is we optimize the parameters of those encoders," Ahmed clarified.
• Ahmed demonstrated a PoC app, showing an uncompressed 800MB video compressed to 40MB (4.8% of original size) with a target VMAF of 96, achieving a 750MB reduction. A pre-compressed 230MB video was further transcoded to 67.9MB.
• Actionable Insight: Video's approach to content-aware, ML-optimized parameter tuning for standard codecs is a significant innovation. If successful at scale, it could offer substantial cost savings for the $300 billion video streaming and cloud storage market.

Miner Competition and Subnet Performance

• Mog discussed the upscaling synapse, which is live. Miners compete to improve on the base model. The presentation showed week-on-week improvements in VMAF and PIAP scores, demonstrating miner-driven optimization.
• "The subnet starts to exponentially improve and that's what we're looking at here," Mog stated, highlighting the positive feedback loop when miners care about the product.
• For the upcoming compression synapse, the exact incentive mechanism for miners is still being designed but will likely involve optimizing for compression rate, perceptual quality, and speed.
• Strategic Implication: The observed performance gains on the upscaling synapse validate BitTensor's core premise: incentivized, decentralized intelligence can rapidly optimize complex tasks. Investors should watch how Video structures incentives for the more complex compression task.

Business Model, Funding, and Go-to-Market Strategy

• Mog explained that Video was incubated by Talstats (his entity), emphasizing a "business first approach." He believes a quality subnet should be self-sustaining operationally within 90-100 days, moving beyond reliance on emissions.
• The priority is securing a solid customer base through reliable services before focusing on revenue. Miners are viewed as "contracting suppliers," and building strong relationships with them is fundamental.
• Mog controversially suggested that subnets "need to not be afraid to foster business-to-business relationships," even if it means working with NDAs for sensitive content (e.g., unreleased documentaries). This might involve miners opting into specific B2B contracts while still performing on the open subnet.
• Const reinforced this, asking why threading the path from subnet to customer is vital. Mog replied, "A subnet is a company or infrastructure that cannot be solely emissions dependent."
• Gareth outlined the go-to-market strategy:
  • Targeting emerging markets (initially Africa): Where connectivity and bandwidth are issues, Video's compression can be revolutionary for VOD platforms and network providers. Gareth has extensive experience in this market.
  • Strategic Partnerships: Integrating Video's technology into existing products (e.g., asset management providers, CDNs).
  • B2B focus: Approaching large enterprise clients, content owners, and VOD platforms with the promise of significant cost savings.
• Gareth mentioned they are in talks with Jan Ozer, a respected figure in the streaming industry ("the godfather of Screaming"), who will review and benchmark their compression model.

Inter-Subnet Collaborations and Future Vision

• Video is actively pursuing collaborations with other BitTensor subnets:
  • Hippius (Subnet - Implied Storage): For decentralized storage of video files.
  • Bitmind (Subnet 18 - AI Content Detection): To detect AI-generated content.
  • Soundsight (Subnet 105 - Audio Restoration): To offer joint video and audio restoration services.
  • Bitcast (Content Creation Marketplace): For content creation and promotion.
  • Potentially Computer Vision Subnet (e.g., Subnet 44 Score): For metadata tagging and content analysis.
• "All of those subnets that we're looking at there are going to bring value to our product," Gareth stated, emphasizing the reciprocal benefits and strengthening of the BitTensor ecosystem.
• The future roadmap includes on-demand video architecture and live streaming through the subnet, which will require decentralized storage, edge servers, and a decentralized CDN.
• Actionable Insight: These inter-subnet collaborations are crucial for building a composable and powerful BitTensor ecosystem. Investors should look for subnets that actively seek synergies, as this can create network effects and enhance overall value.

Validator Perspectives and Network Growth

• Const questioned Mog (representing Talstats, a major validator) about leveraging validator bandwidth for subnet services. Mog explained his view has shifted: the responsibility for revenue generation falls more on subnet owners.
• Talstats' current philosophy is to "give our stake weight to the person that adds the most value to the subnet," which is often the owner. He acknowledged the evolving roles of validators within the BitTensor ecosystem.
• Gareth concluded by reiterating the transformative potential for emerging markets and the unique advantage BitTensor provides. "I couldn't have done this without BitTensor... the decentralized network and the miners just to them."

Conclusion

This episode highlights BitTensor's resilience in crisis and the tangible progress of application-focused subnets like Video. For Crypto AI investors and researchers, the key takeaway is the strategic importance of subnets developing real-world utility and B2B revenue streams, alongside fostering inter-subnet collaboration to build a robust, self-sustaining ecosystem.