Skip to content
AdvancedUnderstand an asset

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.

Read next

Crypto quizzes

Answer a few questions and get your result instantly.

Start