Bitsota
SN94Compete to build state-of-the-art AI research with automated discovery of new techniques
A competitive bounty for algorithm discovery. Miners use genetic programming to evolve ML code from scratch, and earn only when they beat the current state of the art.
// Evolving algorithms. Rewarding breakthroughs.
Bitsota is a decentralized research network where miners evolve machine learning algorithms through competitive genetic programming. Instead of running pre-built models, miners build and mutate new algorithms from basic mathematical operations, submitting only when they beat the current benchmark. Validators independently verify results and reward the winner on-chain.
The simple version: It's like a live competition to rediscover how machine learning works, from first principles. Whoever evolves a better algorithm earns the reward.
Centralized equivalent: Google's AutoML-Zero research project, but run as an open, distributed market with cryptoeconomic incentives rather than a single lab's internal experiment.
How it works:
- Miners run genetic programming engines to evolve ML algorithms from basic math operations, submitting only when they beat the current state-of-the-art threshold on evaluation benchmarks
- Validators independently re-evaluate each submission, vote through relay consensus, and set on-chain weights to reward the winner
- The problem it solves: AI algorithm research is concentrated in a handful of large labs. Discovering new machine learning methods is slow, expensive, and largely closed.
- The opportunity: If algorithm discovery can be parallelized across distributed machines with open incentives, the rate of progress could increase.
- The Bittensor advantage: Cryptoeconomic rewards create a continuously running competition with no off switch and no single team controlling what gets explored.
- Traction signals: Early stage. The GitHub repo shows 104 commits from 2 contributors with development active through April 2026. A desktop GUI at bitsota.ai lowers the mining entry bar, and both direct and pool participation modes are available.
Category: Other (AI Research and Algorithm Discovery) | Centralized Competitor: Google AutoML-Zero, Google Brain
Bitsota builds on Google Research's AutoML-Zero project, which demonstrated that neural networks can be evolved from scratch using only elementary math operations rather than designed by hand. Bitsota attempts to decentralize and continuously incentivize that discovery process. "SoTA" refers to state of the art. The premise is that open, market-driven competition can generate novel AI research where closed lab settings cannot.
Mechanism:
According to the GitHub README, miners run local genetic programming engines, evolving algorithms over up to 150 generations against a fixed evaluation pipeline (CIFAR-10 binary classification is the current active benchmark). When a miner's algorithm beats the state-of-the-art threshold on hidden test criteria, it submits to a relay node. Validators independently re-evaluate the submission, vote through relay consensus, and set on-chain weights: 90% to a burn address and 10% to the winning miner. This means nearly all miner emission weight is directed to a burn hotkey rather than distributed broadly.
Two participation modes are available from the same codebase. Direct mining runs the full local evolution pipeline and competes for the 10% winner share. Pool mining distributes smaller evolution and evaluation tasks across multiple participants, who contribute to collective submissions and receive epoch-based reputation rewards. A desktop GUI at bitsota.ai is available for both modes, which is unusual for a Bittensor subnet and reduces setup friction considerably.
The subnet is designed as problem-agnostic. AutoML Zero is the first deployed challenge, but the same competitive structure could apply to different research benchmarks over time. The TAO.app team description frames the goal as enabling "open, reproducible" progress rewarded only on proof of performance.
Current on-chain activity shows very limited miner participation. This is the central tension in the subnet's current state: the mechanism requires competing miners to generate meaningful research outputs, and that competition has not yet developed.
The 7-day price gain of +21.6% and a positive 7-day net inflow of 187 TAO indicate active buying interest in the alpha token. Pool depth stands at 1,733 TAO, with a 0.44% emission share. The root proportion sits at 0.42, meaning roughly 42% of pool depth comes from the protocol subsidy, which is normal for a subnet at this stage.
- Execution: Active miner participation is currently minimal. A competitive benchmarking network needs competing miners to produce results. Growing the mining community is the critical variable.
- Development concentration: Two contributors account for all 104 GitHub commits. Long-term sustainability depends on attracting additional developers and validator operators.
- Mechanism novelty: The 90% burn / 10% winner weight model is unconventional. Most participants earn through pool reputation payouts rather than direct winner emissions. This is worth understanding before participating.
- Competition: Machine learning research is well-funded and fast-moving. The value of algorithmically evolved outputs depends on results the subnet has not yet produced at scale.