This episode details Score's ambitious plan to revolutionize sports analytics using Bittensor, transforming raw game footage into high-value, annotated data for clubs, hedge funds, and fantasy platforms through a novel computer vision subnet.
Introducing Score: The Winning Layer on Bittensor
- Tim and Max introduce Score, a project building on the Bittensor network since mid-last year, positioning itself as a "winning layer." Bittensor is a decentralized network that incentivizes machine intelligence creation and distribution.
- Their vision is to fundamentally disrupt how participants achieve success ("win") across various domains, starting with sports.
- Max emphasizes the team's diverse expertise in data science, gaming, and sports, aligning with Bittensor's core incentive mechanisms focused on performance and "winning."
- A key strategy is finding "new meta" and democratizing access to advanced sports analytics, particularly for lower-tier football clubs lacking resources. Max notes the counterintuitive lack of deep data science adoption beyond the very top teams.
Market Opportunity: Disrupting Sports Data Verticals
- Score aims to build both the foundational data layer and vertically integrated businesses leveraging that data.
- They identify large total addressable markets (TAMs) in sports data, betting, fantasy sports, and professional sports scouting/management.
- Tim highlights a "flywheel" effect: providing data to verticals generates secondary data and insights that feed back into and improve the core Score layer, creating a positive feedback loop. "We sort of have this feedback mechanism where we're providing data to these stakeholders... [their] secondary data... is fed back into what we're doing which ultimately increases the intelligence of the foundation."
The Technical Challenge: Limitations of Existing Sports Data
- Tim explains that while football is data-heavy (passes, shots), deriving deep insights remains difficult, even at the Premier League level. There's often a disconnect between management/scouts and data teams.
- Current high-end data collection (like FIFA's) is extremely expensive, slow, and requires significant hardware (dozens of cameras) and human resources (large annotation teams).
- Score identified an opportunity to produce similar quality data significantly faster (claiming >1000x), cheaper, and with near-human accuracy by leveraging Bittensor's decentralized compute.
Score's Computer Vision Subnet: Leveraging Asymmetry
- The core technical innovation lies in exploiting an asymmetry: the computationally intensive task of object and keypoint detection on video frames versus a lightweight, effective validation mechanism.
- Miners on the Score subnet (currently Subnet 44) process video clips, performing frame-by-frame object detection (players, ball) using bounding boxes and identifying key points on the pitch.
- This asymmetry allows Score to scale the annotation process efficiently. Tim states, "That's really the asymmetry we've uncovered and that's what we're really pushing down on because that's where we see the opportunity to really scale this."
How Score Works: From Footage to Actionable Insights
- Validators provide footage data (currently sourced by the team, future plans for user uploads) to miners.
- Miners detect players, the ball, and pitch key points on each frame.
- Validators score miner submissions using CLIP and Homography (more below).
- The best submissions are compiled and stored, creating a raw annotated dataset. This raw data itself has market value (talks with data brokers mentioned).
- Strategic Insight: This raw data feeds into Score's "Player Scoring System" or "Football Value Function," designed to assess player impact on each game state – the core intellectual property being developed.
- This processed data powers various applications: fantasy sports (Six Dubs), AI agents for betting (DKing for a sports hedge fund), scouting tools (Scout, advised by Bryant McDermott), and coaching analytics (Coach).
Validation Upgrade: Moving Beyond VLMs to CLIP and Homography
- Initially, Score used OpenAI's multimodal vision model (VLM) for validation. This proved expensive, non-deterministic (probabilistic scoring), limited in scope (only 2/750 frames evaluated), and susceptible to gaming by miners.
- Key Technical Shift: They migrated validation to a combination of CLIP (Contrastive Language–Image Pre-training) for semantic validation and Homography for geometric validation.
- CLIP: Embeds image crops (bounding boxes) and text descriptions ("football player") to check semantic alignment. It's more accurate, faster, cheaper (CPU-runnable), and deterministic.
- Homography: Uses detected pitch key points (minimum 4 out of 32) to geometrically map the pitch, allowing checks on player movement plausibility and position. This is also deterministic.
- Investor Takeaway: This upgrade is crucial for scalability and reliability, enabling Score to process more footage accurately and cost-effectively, underpinning the value proposition.
Securing the Data: A Landmark Footage Partnership
- Max announces a major global footage partnership providing access to data from 283 leagues (approx. 400,000 matches per year).
- The deal includes the current season plus a rolling pre-season archive, ensuring miners always have fresh, relevant data.
- Strategic Significance: Max frames this as a significant "moat" for the subnet, potentially creating the largest annotated football footage dataset on the market and adding immense value to the miners' work.
The Football Value Function: Towards a Universal Player Score
- With the annotation problem addressed, Score is focusing on the "Football Value Function" – a system to score player impact in real-time.
- Max describes the ambition to create a new "ELO rating" (a method for calculating relative skill levels, famously used in chess) for football players, and potentially a universal method for scoring any moving object in video.
- They are collaborating with team member Peter Cotton (author of "Microprediction"), using Monte Carlo simulations (computational algorithms relying on repeated random sampling) of millions of games to understand player value and impact.
- Research Angle: This involves complex modeling to correlate positional data with game outcomes (e.g., goals), potentially assessing the economic impact of actions when combined with betting odds. Findings and papers are expected later in the year.
Expanding Miner Capabilities: Future Tasks and Data Points
- Score plans to increase the complexity of tasks for miners:
- Identifying player teams.
- Identifying and tracking specific players over time.
- Calculating object velocity.
- Calculating proximity to the ball.
- Later goals include "event spotting": miners programmatically detecting goals, fouls, free kicks, etc., adding richer metadata directly within the subnet. Tim notes this competes with established players like Hawk-Eye but believes Bittensor's incentive mechanism can outperform them.
Real-Time Ambitions: Unlocking Live Data Applications
- The speed of miner inference makes real-time data processing feasible.
- Score plans to introduce a streaming challenge type, moving beyond discrete 30-second clips.
- Market Impact: This opens doors for live betting insights, real-time tactical feedback for coaches during games, and enhanced broadcast overlays.
Beyond Football: Score-as-a-Service and New Sports
- Leveraging their robust video annotation capability, Score plans to offer "Score-as-a-Service."
- This involves validators potentially building front-ends for external users (Web2 AI companies, data brokers) to upload video content for annotation, possibly accepting TAO (Bittensor's token) as payment in the future.
- Expansion: Score announces Cricket as the next sport, partnering with former player Nick Compton to navigate the ecosystem. This signals broader ambitions beyond football.
Traction and Partnerships: Building Real-World Use Cases
- Max provides updates on existing partnerships:
- Working with the second-largest sports hedge fund (managing >$5Bn/year), providing an edge in soccer markets.
- Ongoing talks with top European football clubs and institutions (sales cycles are long, often tied to off-season).
- Building proprietary front-end applications (like a fantasy app targeting the 2026 World Cup) to showcase the subnet's capabilities and drive adoption. Max states, "We want to proudly say that this fantasy app is powered by Bittensor."
The Moneyball Analogy and Universal Scoring Potential
- Jacob (host) draws a parallel to "Moneyball," the data-driven revolution in baseball. Tim and Max agree, noting football's complexity (fewer scoring events) makes it a harder challenge, but one they believe technology and Bittensor's miners can solve.
- The discussion revisits the universal scoring concept, applicable beyond sports to areas like self-driving cars, retail analytics (shoplifting detection), and even satellite imagery analysis for hedge funds tracking commodities.
Technical Deep Dive: Ground Truth, Accuracy, and Speed
- Jacob probes the validation mechanism and ground truth. Tim clarifies that CLIP handles object detection validation semantically, while homography checks geometric plausibility. Absolute ground truth isn't always required for validation.
- However, they plan periodic benchmarking using datasets with known ground truths (like SoccerNet) to precisely measure miner accuracy against human annotation standards.
- Max provides a stark speed comparison: human annotators take ~20-30 seconds per frame, while top Score miners process a 30-second clip (many frames) in ~3 seconds. FIFA's real-time annotation involves large teams tracking single objects.
Bittensor Integration and Future Development
- Jacob raises the possibility of the score predictor model itself being built on Bittensor. Max expresses enthusiasm but notes the complexity of running multiple subnets.
- Key Bittensor Insight: Jacob reveals upcoming Bittensor functionality allowing "subnets to have subnets," enabling Score to potentially run multiple mechanisms (like annotation and prediction) under the SN44 umbrella without launching a new token, keeping value consolidated. Max responds, "Oh right that's that's huge. Then yeah then it's opening like a yeah a lot of different avenues for us."
Partner Perspectives: Value Over Crypto Skepticism
- Max emphasizes that partners (hedge funds, clubs) are primarily interested in Score solving their problems faster, cheaper, and better. The crypto/token aspect is secondary and often abstracted away.
- Max: "When you prove people that you are fixing a real problem faster, cheaper, and in a more resilient way, they tend to forget that they don't like crypto." The value proposition overrides concerns about the underlying technology's nature.
Subnet Collaboration: Working with Microcosmos
- Score is collaborating with Microcosmos (SN13), using their sentiment data to enhance prediction models, particularly around betting odds direction.
- They are also exploring using Microcosmos (potentially SN9) to fine-tune their own CLIP models, highlighting a belief in inter-subnet synergy within the Bittensor ecosystem.
The Endgame: Buying a Football Club
- Max outlines the long-term, ambitious goal:
- Prove Score's data leads to winning.
- Raise capital (~$20-30M), potentially leveraging finance partners.
- Acquire a lower-division (e.g., English League 2) football club.
- Implement Score's data science via the subnet for recruitment and tactics.
- Aim for promotion, showcasing the system's power.
- They are also producing physical merchandise (jerseys) for the "Bittensor FC" concept.
Reflective and Strategic Conclusion
- Score demonstrates a potent application of decentralized AI for high-value data annotation in sports, validated by significant partnerships and a clear path to monetization. Investors should track their progress in scaling data throughput, refining the Football Value Function, and expanding into new verticals like Cricket and Score-as-a-Service.
