Bittensor

A March 2020 paper by Yuma Rao, Jacob Steeves, Ala Shaabana and Daniel Hoogland put the idea into the public record. ArXiv now labels that paper withdrawn. It can identify the authors and the starting question, but it cannot be treated as a current whitepaper or proof that today's code implements every original claim.

The project's own chain history begins Kusanagi in January 2021 and says it stopped in May after consensus problems. Nakamoto relaunched the network in November 2021. Finney, forked on 2023-03-20, addressed performance constraints and became the setting for the subnet system. This lineage matters because Bittensor is an evolving incentive experiment, not a design that arrived fully formed in 2020.

On-chain, Subtensor tracks accounts, stake, registered subnets, validator weights and emissions. Off-chain, miners and services answer requests, train or serve models, rent compute, store data or make predictions. Validators run their own evaluation processes and submit weights. The chain settles those claims; it does not execute one shared model or merge every model into a global brain.

This separation lets subnet creators define different commodities and scoring mechanisms. It also moves much of the product truth outside consensus: a valid block can faithfully record a score without proving that the score captures what an end user calls useful. Understanding Bittensor therefore requires examining each subnet's service and evaluator, rather than treating TAO as a certificate for an undifferentiated AI network.

Yuma Consensus takes the weights validators assign to miners, finds a stake-weighted median and clips outlying support before rewards are calculated. The documentation calls the object being measured subjective utility and places security on a majority of relevant stake being honest. Yuma is distinct from Aura and GRANDPA, which order and finalize blocks.

That mechanism can reward evaluators who find useful output, but the rule can be learned. In 2024 the OpenTensor Foundation described weight copying: validators could observe public consensus weights, reproduce them without doing the underlying evaluation and still seek rewards. Commit-reveal hid weights during the vulnerable interval, mitigating the tactic without turning subjective evaluation into an objective ground truth.

The central test is therefore behavioral. If validators spend stake and effort on independent evaluation, the market may surface useful specialists. If copying, collusion or benchmark gaming is cheaper, the same accounting system can price conformity to the incentive mechanism instead.

The first dTAO block, 4,920,351 in February 2025, gave each non-root subnet its own alpha token and TAO/alpha pool. Staking TAO to a subnet is an AMM exchange into alpha, and unstaking exchanges back. A staker therefore accepts price impact, fees, MEV exposure, alpha volatility and the chosen validator's performance; different subnet alphas are not interchangeable claims on the same product.

The allocation rule did not stop changing with Taoflow. At finalized block 9,000,241 on 5 September 2026, Finney ran version 454. Its deployed code matched the official release. It started from capped moving-average alpha prices, multiplied each share by one minus the proportion of miner incentives withheld from owner hotkeys, and then applied a rank-based Hill emission gate before renormalizing. The old net-flow function remained unused in that path. This makes token demand and the handling of insider-directed rewards part of how subnets compete for TAO. Participant alpha still splits roughly 18% to owners, 41% to miners and 41% to validators and stakers. TAO and each alpha approach separate 21 million limits, and recycled TAO can be issued again; one ecosystem-wide cap or an exact four-year clock would misdescribe that system.

Bittensor's current block-production design is permissioned PoA: approved authorities run Aura and GRANDPA, while NPoS is a future plan. The documented mainnet release route requires signatures from both keys of the deployment multisig and then two of three Triumvirate signatures through the sudo proxy. Within a subnet, the owner can shape incentive code and multiple operating parameters and receives the 18% owner share; root-level controls reserve other changes. When the 128-subnet limit is full, a low EMA alpha-price subnet outside immunity can be removed with its UID state.

The July 2024 incident made operational power visible. A malicious PyPI package published as bittensor 6.12.2 stole decrypted keys; the foundation said Subtensor itself was not compromised. At 19:41 UTC on July 2, OTF placed authority validators behind a firewall and put the network into safe mode, stopping transactions while the threat was investigated. The response limited further movement, yet it also demonstrated that network flow could be interrupted through a concentrated operational path.