Full project description
The paper explains an autonomous commodity engine for creating RL training data.
- Miners produce json trees and are measured on novelty and ethical depth
- We use a vector to determine similarity and organize it by domain.
- Validators check novelty and measure for completeness - ie missing branches
Hardware requirements:
For miners:
- 72B param model
- physics simulation engine
For validators:
- check against a versioned copy of the vector
- also run a physics engine - ideal
The vector is maintained by Project X, which will initiate partnership discussions, technical integration with ROS, run simulation modelling of the engine, post results to Hugging Face.
Revenue from the sale of the data clusters for RL, training foundation models/world models, and licensing will be split 90/10:
90 to the company
10 for the pool of registered miners (requires KYC to receive a portion of fiat royalties)
- twice per year voting for whether this goes directly to the holders as a TAO injection or allocated as a royalty
— voting weight by stake
New challenges correlate to demands for RL data. Miners receive challenges based on queries to the codex, this dynamically adjust challenges based on gaps in the clusters.
Externally this commodity is valued because of its proposals for new challenges requires voting by stake weight (see Curve Finance) -This creates a secondary market where parties can purchase votes to influence challenge creation and direct production toward their specific data needs.
This proposal identifies that challenge creation will be permissionless at full build.
At launch this will be a managed process with development scheduled after revenue goals have plateaued.
Products:
Raw JSON: bulk export, minimal curation, researchers and academics
Domain Packages: curated by vertical, mid market operators
API Access: programmatic query, real time updates, enterprise
Runtime Query: decision support at inference, low latency, SLA guaranteed
Summary: autonomous commodity engine for creating RL training data for humanoid robotics.
Phase 1: sell versioned data
Phase 2: runtime query
Phase 3: train our own model
Current status: infrastructure built, vector search live, stress testing incentive formulas.
Full whitepaper is available here: https://docsend.com/view/s/mas5gn32ibwdyw9c
Why it works on Bittensor
Building this on centralized infrastructure means solving the cold start problem. With a subnet the infrastructure already exists. Decentralized validator and miner coordination is built in. We inherit a battle tested incentive layer instead of building from scratch.
Novelty based scoring with Bittensor subnets reward unique computational proofs for ML challenges. We define the challenge as producing novel, physics validated simulation data.
Bittensor makes the engine self sustaining before the first customer transaction.