affinefoundation/affine repository overview

https://github.com/AffineFoundation/affine-cortex⁠

Affine Mine open reasoning.

Introduction Affine is an incentivized RL environment that pays miners who make incremental improvements on a set of tasks (such as program abduction or coding). The mechanism is sybil-proof, decoy-proof, copy-proof, and overfitting-proof.

How does Affine work?

Affine validators incentivize miners to submit models to Subnet 64 on Bittensor (a.k.a Chutes) where they are inference load balanced and publicly available. These models are evaluated on a set of RL environments, with validators looking for models that dominate the Pareto frontier—namely, models that outcompete all other models across all environments. The network uses a winners-take-all mechanism where miners are incentivized to copy, download, and improve the Pareto frontier model.

Why Affine?

Directed incentives for RL have never been achieved. The ability to direct intelligence and aggregate the work effort of a large, non-permissioned group of individuals on RL tasks will unlock rapid advancement in intelligence. We intend to commoditize reasoning (intelligence's highest form) and break the intelligence sound barrier.