Dojo
SN52A decentralised competition where AI miners generate and judge creative work to improve quality
A decentralized GAN (Generative Adversarial Network) where the generators and discriminators are human miners competing in a browser. No servers required. Load your wallet into Talisman, receive tasks, and compete to produce outputs that beat a high-quality baseline. Zero-sum: your gain is someone else's loss.
// A decentralized GAN, powered by humans.
Dojo transforms the GAN concept (where two AI networks compete to improve each other) into a human-powered competition on Bittensor. Miners compete to create outputs that are not only indistinguishable from high-quality baselines but are actually better. Other miners judge the results. The competition is zero-sum: the best work wins, the rest earns nothing.
The simple version: Imagine an art competition where contestants must create work that's better than a professional reference piece, and other contestants judge the entries. The judges and the artists are competing against each other simultaneously. That's a GAN, but with humans instead of neural networks.
Centralized equivalent: Think Scale AI (human data labeling) crossed with Kaggle competitions, but structured as a zero-sum game where miners are both producers and evaluators.
How it works:
- Miners register on the network and load their hotkey wallet into a browser wallet (like Talisman). They receive tasks through the web interface at dojo.network, produce outputs, and compete against baselines and other miners.
- Validators coordinate task distribution, evaluate submissions against baselines, and set weights based on output quality and discriminator accuracy.
- The problem it solves: AI training data and human evaluation at scale is expensive and centralized. Scale AI charges premium prices for human annotation. Dojo distributes this work across a global workforce.
- The opportunity: The data labeling market is worth billions and growing with every new AI model that needs fine-tuning data. Human-in-the-loop AI improvement is fundamental to alignment research.
- The Bittensor advantage: Zero-sum incentives create genuine quality pressure. Unlike traditional data labeling where workers are paid per task regardless of quality, Dojo's GAN structure means only the best work earns rewards.
- Traction signals: Built by Tensorplex Labs. 138 commits across 6 contributors. Written in Go (unique for Bittensor). Live at dojo.network for mainnet and testnet.
Category: Data Scraping and Archival | Centralized Competitor: Scale AI, Surge AI, Amazon Mechanical Turk, Labelbox
Dojo V2's most interesting design decision is making the GAN concept human-powered. Traditional GANs have two neural networks (generator and discriminator) competing. Dojo replaces both with human miners: some produce content, some judge it, and the competitive dynamic drives quality improvement.
Mechanism:
The zero-sum structure is key. Miners acting as generators must produce outputs that exceed the baseline. Miners acting as discriminators must accurately identify which outputs are superior. Both roles compete: generators who produce better content earn more, discriminators who judge more accurately earn more. This dual competition mirrors the GAN training dynamic but with human intelligence.
The browser-based mining approach is a significant UX decision. No server setup, no command line, no DevOps. Register, load a wallet, and start receiving tasks. This dramatically lowers the barrier to entry and opens mining to non-technical participants.
The codebase has 138 commits across 6 contributors, written in Go. However, development has been inactive since October 2025 (5 months with no commits). This is the primary concern.
Market metrics reflect the development pause. At 33,162 TAO market cap with 1,714 holders, the Gini of 0.837 is the highest we've seen, indicating extreme concentration. HHI of 0.275 confirms a few large holders dominate. The 30-day price is down 19.4%.
Root proportion of 0.168 shows organic demand exists, and the 90-day return of 10.4% is positive. Net 7-day inflow of 175 TAO is modest but positive.
- Development inactive: No commits since October 2025. Five months of silence for a live product is concerning.
- Extreme concentration: Gini of 0.837 and HHI of 0.275 are the highest in our coverage by far. This is effectively a whale-controlled subnet.
- Declining price: -19.4% over 30 days suggests fading interest.
- Competitive landscape: Scale AI and others have established enterprise relationships and massive workforces for human data labeling.