Project Introduction

Bittensor is essentially a permissionless peer-to-peer (P2P) network that leverages a blockchain token economy to incentivize the creation and operation of AI products. For developers, the Bittensor network offers a method for a decentralized artificial intelligence application marketplace, and for end-users, it allows direct access to network resources at a lower cost. The vision of the Bittensor network is to harness the power of the digital market to drive the development of the most important digital commodity—artificial intelligence. Its goal is to build the most powerful AI network, enabling every ordinary person to benefit from it and take ownership, fostering a bottom-up rather than top-down development model.

Subnets

All AI applications on the Bittensor network occur on its subnetworks, or subnets. Each subnet has a dedicated use case. Currently, there are 32 subnets on Bittensor, as shown in the following figure.

Chart 1: Subnet ecosystem, source from X@0xai_dev

Here are some typical examples of subnets:

Subnet1 belongs to the category of text generation subnets, where validators send prompts to miners, and miners generate results based on these prompts. The miner with the best result will receive a reward.

Subnet5 is a subnet for generating images from text, where miners create images based on customer requirements. Validators rank the images based on their aesthetic appeal and how well they match the customer’s prompt words. Additionally, validators receive a minor penalty for images that are too stylistically similar, to encourage diversity in the image models hosted by the miners.

Subnet8 is a subnet that uses artificial intelligence to predict financial market trends, currently focusing mainly on predicting Bitcoin price trends. However, as the project develops, it will gradually expand to other financial markets or specific areas (such as sports betting). The latest data on this subnet shows that the average daily mining reward is $133,000, with an estimated annual miner reward of $32 million.

Chart 2: Data related to subnet8, source from www.taoshi.io

Roles in Blockchain

1) Miners: Can be understood as the providers of AI models or algorithms, hosting AI models and offering them to the Bittensor network. Different subnetworks in Bittensor have various models, such as text generation models, image generation models, etc.

2) Validators: Act as evaluators of the Bittensor network, aiming to assess and verify the results completed by miners to help customers obtain the best solutions. To become a validator, a user must be among the top 64 holders of TAO and register a UID on any of its subnetworks. (However, looking at the list of validators, it seems that most institutions from the project’s ecosystem are included, perhaps later other organizations or users will become validators.)

Chart 3: Validator list, source form www.taostats.io

3) Nominators: Nominators delegate their TAO tokens to validators to show their support and earn staking rewards. Validator information is open and decentralized, allowing nominators to research and choose suitable validators to stake their tokens based on publicly available information.

4) Users: The ultimate users of AI models in the Bittensor network.

In one sentence, the relationship between these four roles can be summarized as: Users/clients present their demands; miners generate task results using AI models on the corresponding subnetwork based on these demands; validators evaluate the results and select the best solution for the clients; nominators choose validators they support to stake tokens.

Technical Architecture

The Bittensor network is a decentralized peer-to-peer machine learning protocol. Within the network, machine intelligence is measured by other intelligent systems over the internet in a continuous and asynchronous peer-to-peer (P2P) manner. This system focuses not only on the model’s ability to complete specific tasks but also evaluates the model’s capability to produce information. The network uses a digital ledger to record the achievements of researchers (miners/developers) and provide rewards, enabling them to benefit from the work created by artificial intelligence. The network is divided into two parts: the AI layer that processes intelligence and the blockchain layer responsible for recording and rewarding.

The blockchain layer is a Layer 0 blockchain based on Polkadot Substrate, responsible for executing consensus mechanisms, ensuring node identity, and incentivizing network nodes. Located beneath the artificial intelligence layer, the two layers communicate through inter-process communication. To fairly distribute incentives among all participating nodes, the Bittensor network utilizes consensus and leverages stake-weighted trust (achieved through the participation of validators and nominators). The AI layer, besides inference and training, is also responsible for abstracting the Bittensor kernel and ensuring the compatibility of node neural networks with the inputs/outputs of other nodes in the network.

Chart 4: Blockchain and AI system, source from bittensor.com

Yuma Consensus

The Yuma Consensus is a decentralized peer-to-peer consensus algorithm designed to achieve fair distribution of computational resources across the node network. The Bittensor network is supported by the Yuma Consensus algorithm. It adopts a hybrid consensus mechanism that integrates Proof of Work (POW) and Proof of Stake (POS). Nodes in the network perform computational work, verify transactions, and create new blocks, which are also verified by other nodes. Contributors who pass verification receive token rewards. Compared to traditional consensus mechanisms, this hybrid mode combines the advantages of both consensus mechanisms. On one hand, it avoids the excessive energy consumption of the POW mechanism, addressing environmental concerns; on the other hand, it circumvents the centralization risks present in POS, ensuring the network’s security and decentralization.

Token Economy

Bittensor’s token, TAO, serves as the network’s reward token, access token, and governance token, allowing token holders to also stake their tokens. A TAO is produced every 12 seconds, equating to the issuance of 7,200 tokens daily. Newly minted tokens are distributed evenly among miners and validators. The total supply of TAO is set at 21 million, with the issuance rate halving once half of the supply is issued. This halving occurs every 12 seconds per block, equivalent to a halving every four years, with each half-mark of the remaining issuance triggering a new halving event until all 21 million TAO are in circulation. This can be seen as a homage to Bitcoin. As of the time of writing, the token’s circulating supply has reached over 6 million, with a market capitalization of 3.5 billion USD, ranking it 26th on CoinGecko.

The following image is a token distribution snapshot from taostats.io, indicating that TAO was a fair launch, with no pre-sales to VCs, etc. Currently, the circulating supply accounts for about 30% of the total supply.

Graph 5: Token Economy Model, source from source from www.taostats.io

Summary

In an article published last month by Vitalik Buterin, “What are the most fruitful intersections between cryptocurrency and artificial intelligence?”, it was mentioned, “Using crypto incentives to foster better artificial intelligence can be achieved without falling into the rabbit hole of fully cryptography-encrypted methods, and approaches like Bittensor are among these.” This highlights Vitalik’s endorsement of the Bittensor project. The project selects the best options from existing algorithm models through incentive mechanisms, meaning it does not produce algorithms but rather transports them, thereby promoting the development of the decentralized AI market. With the continued popularity of artificial intelligence, the project’s market value has nearly tripled since January of this year. Personally, I find the rich ecosystem of applications within its subnet, such as medical health diagnostics, 3D asset creation, audio generation, image generation, distributed model pre-training, etc., to be quite interesting and worth further exploration.