Last updated: October 8, 2026.
| Document | Bittensor: A Peer-to-Peer Intelligence Market |
|---|---|
| Official-version author | Yuma Rao |
| Date | Publication date unstated; official version checked October 8, 2026. |
| Status and license | PDF: preprint under review, 15 pages; paper license unstated. SDK license: MIT. |
The paper’s central question
Bittensor: A Peer-to-Peer Intelligence Market asks whether machine-learning systems can value one another’s contributions and reward useful information through a digital ledger. The official paper credits Yuma Rao. The linked original PDF has 15 pages and describes itself as a preprint under review; it does not establish a 2022 publication date. This reference follows that official version. [1]
Publication history and source scope
An earlier arXiv entry with a similar title lists Yuma Rao, Jacob Steeves, Ala Shaabana, Daniel Attevelt and Matthew McAteer. It was submitted in March 2020 and withdrawn in November 2021 as incomplete and obsolete in content and authorship. That record should not be used to overwrite the authorship of the separately hosted official paper or to imply that its current version was itself withdrawn. [2]
Peer ranking, incentives and bonds
The historical model represents peers as learned functions. Each peer can use other peers’ outputs and submit weights indicating their contribution. A stake-weighted ranking distributes incentives, while a connectivity-based consensus term addresses self-rewarding coalitions. Bonds reward peers for identifying contributors that others later value. [1]
Conditional computation and knowledge extraction
The paper also proposes selecting a sparse set of peers to reduce communication and distilling their combined behavior into models that can work offline. These are model-design proposals, not evidence that every present subnet trains one shared neural network or exposes the same tensor format. Its collusion analysis depends on honest-stake, connectivity and convergence assumptions; it does not establish universally correct AI outputs. [1]
Later implementation: subnets and evaluation
Current Bittensor subnet documentation describes separate incentive markets. Miners produce the subnet’s commodity, validators assess miners, owners define mechanisms and parameters, and stakers back validators. Work is performed off-chain; the chain records participation, weights and rewards. The commodity may be inference, storage or predictions rather than the original paper’s particular peer-training model. [3]
Yuma Consensus is not block consensus
Yuma Consensus aggregates subnet evaluations using stake-weighted agreement, including clipping weights above a supported consensus level and calculating miner incentives and validator dividends. Bonds are reward-accounting quantities; an unfavorable reward calculation should not automatically be described as confiscation of a holder’s token balance. The internals guide also distinguishes legacy bond smoothing from the Yuma3/liquid-alpha path. [4]
Chain security and evaluation assumptions
As checked on October 8, 2026, the official chain guide describes Subtensor as Proof of Authority: authorized validators produce blocks through Aura and finalize them through GRANDPA. Its Nominated Proof of Stake transition is described as planned. Subnet scoring validators and block-producing authorities therefore have different roles. Holding or staking TAO does not by itself confer permission to finalize blocks. [5]
Dynamic TAO is a later, separate design
The separate Dynamic TAO paper proposes subnet-specific alpha tokens and market-based emission allocation through TAO/alpha pools. Its original presentation uses constant-product AMMs and subnet token prices. This should not be reduced to “more TAO staked always means more emissions,” or attributed to the earlier intelligence-market paper. The implementation has continued to evolve. [6]
Staking, pool exposure and supply
Current pool documentation describes subnet staking as exchanging TAO for alpha, with unstaking exchanging alpha back. Price impact, fees and changing pool prices affect the TAO recovered. It describes weighted pools, a generalization of constant-product pools. Root staking is a separate TAO-denominated case without that subnet swap. Rewards do not guarantee a positive return in TAO terms. [7]
Halving history and changing parameters
The emissions guide dates dTAO’s introduction to February 2025. TAO’s 21-million cap is accompanied by halvings based on total-issuance thresholds, not an unconditional four-year calendar. Recycling can affect when thresholds are reached. The guide reports the first halving in December 2025. Exact active emission settings should be checked against the chain rather than inferred from this historical paper or a documentation preview. [8]
What the paper does and does not establish
The paper offers a mechanism for pricing and rewarding informational contributions. Successful incentives still depend on the chosen task, evaluation quality, participants and economic assumptions. Payment or agreement does not independently prove factual correctness or customer demand. Registration costs likewise do not guarantee that only useful subnets survive. [1] [4]
License and reference boundaries
No explicit reuse license was identified in the reviewed paper. The Bittensor SDK repository’s MIT license is a separate software license. This page is a technical-paper reference; token prices, rankings, guaranteed staking yields and claims of being the first decentralized AI market are outside its verified scope. [9]
Social Media Sentiment
Not applicable to this technical-paper summary; project or token sentiment is not inferred from the paper’s claims.
Last updated: 2026-10
Related Terms
See Also
Sources
- Yuma Rao, Bittensor: A Peer-to-Peer Intelligence Market — Primary source; checked October 8, 2026.
- Earlier arXiv record: submission history and withdrawal notice — Primary source; checked October 8, 2026.
- Current network documentation: subnets, participants and TAO/alpha — Primary source; checked October 8, 2026.
- Yuma Consensus: stake-weighted scoring, bonds and assumptions — Primary source; checked October 8, 2026.
- Chain consensus: PoA authorities, Aura and GRANDPA — Primary source; checked October 8, 2026.
- Dynamic TAO paper: subnet tokens and market-based allocation proposal — Primary source; checked October 8, 2026.
- Staking and pools: swaps, price impact and root-network exception — Primary source; checked October 8, 2026.
- Emissions documentation: issuance thresholds and dTAO history — Primary source; checked October 8, 2026.
- Bittensor SDK repository license: MIT — Primary source; checked October 8, 2026.