Affine
SN120Mine open reasoning traces, rewarding AI models that show their working
Const, co-founder of Bittensor, built a subnet where miners compete to create the best reasoning AI. Models are trained, uploaded to Hugging Face, deployed as inference endpoints, and then battle-tested on multi-turn problem-solving tasks. The best reasoners earn emissions.
// Mine open reasoning.
Affine is a competitive arena for AI reasoning models. Miners train models that can navigate complex multi-step tasks, like shopping on websites, exploring virtual environments, solving logic puzzles, and conducting science experiments. The best-performing models earn the most rewards.
The simple version: Imagine a competition where AI models are dropped into a video game and scored on how well they complete challenging quests. Each quest requires multiple steps of reasoning, planning, and problem-solving. The AI that solves the most quests, across the widest variety, wins.
Centralized equivalent: Think OpenAI o1/o3 or Google's Gemini reasoning capabilities, but developed through open competition rather than internal research teams.
How it works:
- Miners train reasoning models, upload weights to Hugging Face, and deploy them as serverless endpoints on Chutes.ai. Models are committed to the blockchain with their deployment details.
- Validators send challenging multi-turn tasks from the AgentGym suite (webshop, alfworld, babyai, sciworld) to miner endpoints. They score using Pareto dominance across all environment subsets, rewarding both specialist and generalist models. Statistical significance testing prevents model copying.
- The problem it solves: Advanced reasoning models are expensive to develop and controlled by a handful of labs. There's no open marketplace for incentivizing better reasoning capabilities.
- The opportunity: Reasoning is the frontier of AI capability. Models that can plan, reason multi-step, and solve novel problems are the most commercially valuable AI systems being built today.
- The Bittensor advantage: Pareto dominance scoring incentivizes diversity. Rather than everyone copying the single best model, miners are rewarded for being the best at different combinations of tasks, creating a portfolio of specialized reasoning approaches.
- Traction signals: Founded by Const (co-founder of Bittensor). 644 commits, 19 contributors, 36 GitHub stars. Active development with 12 commits this week on Pareto scoring refinements. 239,716 TAO market cap with 5,057 holders.
Category: Model Training / AI Reasoning | Centralized Competitor: OpenAI o1/o3, Google Gemini Reasoning, Anthropic Claude Thinking
Affine is a founder's subnet in the truest sense. Const, who created Bittensor itself, built Affine to demonstrate that decentralized incentive mechanisms can produce frontier reasoning capabilities. This is a statement piece: if Bittensor's incentive design works anywhere, it should work here.
Mechanism:
The scoring system is Affine's most interesting design choice. Instead of a simple leaderboard, validators use Pareto dominance across all possible subsets of environments (L1 through L8). A model that's the best at webshop + alfworld gets rewarded even if it's mediocre at sciworld, while a generalist that's decent at everything also gets rewarded. This prevents winner-take-all dynamics and encourages a diverse ecosystem of specialized reasoners.
Models are deployed through Chutes.ai as serverless endpoints, which means miners don't need to run their own infrastructure. They train locally, push weights to Hugging Face, deploy on Chutes, and commit the deployment to Bittensor's blockchain. Validators then call these endpoints with benchmark tasks and score the results.
The codebase is actively maintained: 644 commits across 19 contributors, with 12 commits this week focused on Pareto pre-filtering, sigma floors for scoring stability, and cold-start improvements. The contributor catoneone drives most recent development.
Market metrics reflect the Const association. At 239,716 TAO market cap, it's a top-tier subnet. Root proportion of 0.247 is one of the highest in the network, indicating strong protocol-level support. Gini of 0.634 and HHI of 0.039 show well-distributed holders across 5,057 accounts. Net 7-day inflow of 2,664 TAO is positive.
The 30-day price is down 6.5%, which is notable against a generally bullish subnet market. Emission buy percentage is at 0%, meaning current emissions aren't being accumulated through market buys. This could indicate the subnet is transitioning between phases or that holders are waiting for catalyst developments.
- Benchmark saturation: The current task suite (webshop, alfworld, babyai, sciworld) is limited. As models improve, these benchmarks may become trivially solvable.
- Chutes dependency: All miner inference runs through Chutes.ai. A Chutes outage or policy change directly impacts the subnet's ability to function.
- Founder premium: Some of the market cap likely reflects Const's involvement rather than pure subnet utility. If attention shifts, the premium could deflate.
