affine

Affine is an incentivized reinforcement learning (RL) environment operating as Subnet 120 (SN120) on the Bittensor network. Unlike subnets focused on inference or compute, Affine is designed to continuously improve AI reasoning capabilities through structured competition. Miners compete to train and submit improved AI models for complex tasks such as program abduction and code generation. The platform rewards only genuine model improvements through a mechanism that is sybil-proof, decoy-proof, copy-proof, and overfitting-proof.

The core idea behind Affine is what the project describes as mining for intelligence. Rather than mining compute cycles or serving API requests, participants mine intelligence itself by making incremental improvements to AI models on a set of RL tasks. Validators run continuous competitions across multiple RL environments, and only miners whose models outperform all others across all tasks receive the bulk of emissions. Every top-performing model becomes an open-source baseline that the entire network can download, improve, and resubmit, creating a continuous cycle of advancement. The project's stated goal is to commoditize reasoning and aggregate the work effort of a large, permissionless group of contributors toward advancing AI intelligence.

Affine is closely integrated with other Bittensor subnets. Miners submit their models to Chutes (SN64) where they are hosted, inference load-balanced, and made publicly available. This creates a cross-subnet value loop: Affine identifies and refines the best reasoning models, while Chutes provides the infrastructure to run them. The platform includes a live dashboard at affine.io for monitoring network activity, model performance, and leaderboards, as well as a rollouts section and an analysis view.

Affine operates through two participant roles: miners and validators. Miners train AI models using reinforcement learning techniques and deploy them to Chutes (SN64), where they become available for evaluation. Validators then test these models across a set of RL environments covering tasks such as program abduction and coding, measuring how well each model performs relative to the current best.

The evaluation uses a Pareto frontier approach. Validators look for models that dominate across all environments simultaneously, not just excel at a single task. The network applies a winners-take-all mechanism: the model that sits on the Pareto frontier earns the majority of rewards. All other miners are incentivized to download that winning model, improve it through further RL training, and resubmit. This creates a ratchet effect where the network's best model is continuously being refined by competing participants.

The mechanism is designed to resist common attack vectors in decentralized AI. It prevents sybil attacks (creating fake identities), decoy submissions (submitting intentionally weak models to game the system), copy attacks (resubmitting someone else's model without improvement), and overfitting (models that score well on known test cases but fail to generalize). Only models that demonstrate genuine, broad improvements across all evaluation environments are rewarded.

For container orchestration, Affine uses Affinetes, a custom Kubernetes-compatible infrastructure layer that handles environment management, supports both local and remote Docker deployments, and provides environment caching for performance. All evaluation environments are packaged as pre-built Docker images. The project also provides an SDK that allows external developers to evaluate models across different environments programmatically.