What is Bittensor (TAO)?
- Author
- CRYPTO PORT Editorial
- Published
- Updated
- Reading time
- 6 min
In short
Bittensor is a network that pays contributors of machine learning models. Within each 'subnet' — a specific task domain — miners supply model outputs and validators score them, and the native asset TAO is distributed according to those scores. Its issuance schedule is modelled on Bitcoin's.
Key points
- Unlike most chains, it evaluates the output of AI models and pays according to that evaluation
- Independent 'subnets' cover distinct tasks such as text generation, images and prediction
- Participants play two roles: miners who serve models and validators who score them
- Built on Substrate, with a reward schedule that halves in stages
Definition
A decentralised network in which machine learning outputs are peer-evaluated and the native asset TAO is distributed based on those evaluations.
Bittensor targets the fact that building and serving AI models is largely locked inside a few large companies. Training requires compute and data, and the results are reachable only through those companies' APIs. Bittensor tries to build a market in which model providers are paid directly.
Subnets are the central construct. Each subnet defines a task — text generation or embedding production, for example — and miners return outputs for it. Validators query multiple miners, compare and score their outputs, and record those scores on chain.
Rewards are aggregated through a mechanism called Yuma Consensus. Validator scores are not simply averaged; they are weighted by how closely they agree with other validators, which dampens the influence of outliers. Acting as a validator requires staking TAO.
In practice, external applications consume the inference APIs that subnets expose. Which subnets exist and how much emission each receives is decided by on-chain mechanisms, so subnets effectively compete with one another for rewards.
Watch out for
- · Output quality is judged by validator scoring, so the scoring criteria themselves can be gamed
- · Subnets are created and retired frequently, and hardware built for one can become useless quickly
- · Mining requires GPUs and other compute, which carry ongoing power and equipment costs
Frequently asked questions
Does holding TAO give access to AI models?
Holding it is not itself an access right. TAO is used for reward distribution, validator staking and subnet registration; model access goes through whatever interface each subnet provides.