Dojo

Organizations and researchers face challenges in evaluating generative AI outputs at scale without centralized gatekeeping. Traditional GANs rely on single generator-discriminator pairs, limiting competitive pressure and innovation. Dojo V2 addresses this by creating a decentralized network where multiple miners compete simultaneously to produce outputs that exceed high-quality baselines, while validators assess quality and rank submissions transparently through an incentive-aligned scoring system.

A decentralized, incentive-based network enables continuous competitive evaluation of generative outputs. Miners receive tasks through a browser interface, submit their best work, and earn rewards based on how their outputs compare to both baselines and competitor submissions. Validators run evaluation infrastructure that uses language models to score submissions objectively, storing results on-chain so the entire network can verify rankings and payouts.

Dojo V2 on Bittensor implements this through a containerized validator stack that connects to Subtensor, evaluates miner outputs via OpenRouter's LLM API, scores them in a structured format, and sets on-chain weights based on validator assessment. Miners register with a single hotkey and interact entirely through the browser interface at dojo.network, removing infrastructure barriers to participation. The system uses Redis for state management and optional logging infrastructure (Loki, Langfuse) for observability into scoring decisions.

Dojo V2 is a decentralized, competitive GAN (Generative Adversarial Network) built on Bittensor where miners act as both generators and discriminators in a zero-sum incentive environment. Miners create outputs that must be superior to high-quality baselines while competing against each other, with validators evaluating the quality and competitiveness of submissions through an LLM-based scoring system.