Score (SN44) Is Expanding From Computer Vision to Open World Models

Score, the team behind Bittensor Subnet 44, has outlined how the project evolved from a sports-prediction product at the early stage into an open computer-vision network, with its next target being an open world model.

The team describes the journey as a series of pivots, with each stage expanding the problem Score was trying to solve.

Stage 1: Sports Prediction

Score started with Score Predict, a sports-prediction product the team now openly considers a failure.

The product was not good enough, but that failure pushed the team toward a broader question: instead of predicting sports outcomes, could Score understand what was happening inside sports footage?

That became the first major pivot, from prediction to perception.

Stage 2: Building Computer Vision Through Sports

Sports became Score’s testing ground for computer vision.

The team developed systems that turn broadcast and match footage into structured information, including player identity, positioning, movement, and events.

The work expanded beyond football into cricket and, soon, volleyball, while also moving into commercial products for professional and youth sports.

The choice of sports was deliberate. Long video sequences, changing camera angles, occlusion, movement, and spatial relationships make sports footage a difficult computer-vision problem.

Score’s conclusion was that the technology being developed for sports could have applications far beyond sports.

SCORE – From football to the entire world

Stage 3: From Sports Analytics to Vision AI

Score’s focus then shifted from building sports products to solving broader problems in vision AI.

The team identified three challenges: training strong models efficiently, preventing open competitive systems from being gamed, and supporting both public and private models.

This became the foundation for Bittensor Subnet 44.

Score designed the subnet around two tracks: efficient open-source vision models that anyone can use, and specialized private models built for commercial applications.

The incentive system went through repeated iterations as the team tested how miners and validators could exploit different evaluation and reward mechanisms. The subnet only moved forward after reaching a system it considered sufficiently robust.

Stage 4: Score Studio

The next step is turning the subnet into a usable product.

Score Studio is designed as the front end for Subnet 44, allowing developers, researchers, and businesses to bring their own problems and data (or generate new data) and use the subnet to train, evaluate, and deploy vision models.

The platform covers data generation and labeling, model training, evaluation, workflow composition, and deployment. Models can remain private or be released openly, while vision-language models (VLMs) are also planned.

Score Studio is scheduled to launch on September 16.

This is the project’s immediate priority: turn the infrastructure Score has built into a product that creates real usage and revenue.

Stage 5: Building a World Model

The longer-term ambition goes beyond computer vision.

Score wants to use Subnet 44 to build an open-source world model that can learn not only what is happening in an environment, but how that environment behaves.

Traditional vision models can identify objects, people, and events. Vision-language models add the ability to describe and reason about what is being observed.

A world model would go further by learning relationships between objects, actions, and outcomes, allowing it to model what could happen next.

The team has already teased early work toward this direction.

John Wick scenes are one of the hardest to track. we ran our subnet vision models on one of them.

Shot at night, barely any light. The camera whipping around, cars sliding sideways, people cutting through the gaps, all of it smeared with motion blur. this is the footage most… pic.twitter.com/eAEa0eGHLQ

— Score (@webuildscore) September 1, 2026

The Flywheel

Score’s proposed model creates a feedback loop between the subnet, the world model, and Studio.

Subnet contributors improve the world model → the world model generates environments and training data → builders use that data to create specialized models → those models generate demand through Studio → revenue and usage feed back into the network.

Score believes this could address some of the biggest problems facing vision AI, including scarce training data, expensive annotation, rare events, and environments that are difficult or dangerous to capture in the real world.

From Sports to Open Vision AI

Score’s trajectory is therefore less about using sports to build toward a much larger vision.

The project started with sports prediction, moved into sports computer vision, expanded into an open competitive vision network, and is now building a product layer around that infrastructure.

The eventual goal is an open world model trained through Bittensor and made accessible through Score Studio.