Bittensor: How AI Subnets Reshape Collective Intelligence Networks

1.

Background of the AI Revolution =======

Background of AI Boom

With the rapid development of artificial intelligence (AI) technology, we are entering a data-driven new era. Breakthroughs in fields such as deep learning and natural language processing have made AI applications ubiquitous. The birth of ChatGPT in 2022 ignited the AI industry, followed by a series of AI tools for video generation and automated office tasks, and the application of "AI+" has also been put on the agenda. The market value of the AI industry has skyrocketed, expected to reach $185 billion by 2030.

Figure 1 Changes in AI Market Value

Traditional Internet Companies Monopolize AI

Currently, the AI industry is mainly monopolized by companies such as Nvidia, Microsoft, Google, and OpenAI, and the advancement of technology has also brought a series of challenges such as data centralization and uneven distribution of computing resources. Meanwhile, the decentralized concept of Web3 provides new possibilities for solving these problems, reshaping the current AI development landscape within the distributed network of Web3.

Current Progress of Web3+AI

As the AI industry surges, a large number of high-quality Web3+AI projects have emerged. Fetch.ai creates a decentralized economy through blockchain technology, supporting autonomous agents and smart contracts to optimize AI model training and application; Numerai uses blockchain technology and a community of data scientists to predict market trends and incentivizes model developers through a reward mechanism; Velas builds a high-performance smart contract platform for AI and blockchain, offering faster transaction speeds and higher security. AI projects themselves consist of three main elements: data, algorithms, and computing power. The Web3+data and Web3+computing power tracks are currently thriving, but the Web3+algorithm direction has been fragmented, resulting in projects that can only form single-direction applications. Bittensor has seized this gap by building an AI algorithm platform with a built-in competitive selection mechanism through the competition and incentive mechanisms of blockchain itself, retaining the highest quality AI projects.

2.

Development Path of Bittensor =============

Innovative Breakthrough

Bittensor is a decentralized incentive machine learning network and digital goods marketplace.

• Decentralization: Bittensor operates on a distributed network of thousands of computers controlled by different companies and organizations, addressing issues such as data centralization.
• Fair Incentive Mechanism: The $TAO tokens provided by the Bittensor network to its subnets are proportional to the subnet's contribution, and the rewards provided to its miners and validators are also proportional to the node's contribution.
• Machine Learning Resources: The decentralized network can provide services for every individual needing machine learning computing resources.
• Diverse Digital Goods Marketplace: Initially, the digital goods marketplace of the Bittensor network was designed specifically for trading machine learning models and related data, but thanks to the expansion of the Bittensor network and the Yuma consensus mechanism, which does not concern the substantive content of data, it has become a marketplace for trading any form of data.

Development History

Unlike many overvalued VC projects currently on the market, Bittensor is a fairer, more interesting, and meaningful geek project, and its development history does not involve the "big promises to deceive investors" process seen in other projects.

• Concept Formation and Project Launch (2021): Bittensor was created by a group of technology enthusiasts and experts dedicated to promoting a decentralized AI network, who built the Bittensor blockchain using the Substrate framework to ensure its flexibility and scalability.
• Early Development and Technical Validation (2022): The team released the Alpha version of the network to validate the feasibility of decentralized AI. They introduced the Yuma consensus, emphasizing the data-agnostic principle to maintain user privacy and security.
• Network Expansion and Community Building (2023): The team released the Beta version and introduced a token economic model (TAO) to incentivize network maintenance.
• Technical Innovation and Cross-Chain Compatibility (2024): The team utilized DHT (Distributed Hash Table) integration technology to make data storage and retrieval more efficient. At the same time, the project began to focus on promoting and further expanding subnets and the digital goods marketplace.
  
  Figure 2 Bittensor Network Promotional Image
  Throughout the development of Bittensor, there has been little intervention from traditional VCs, avoiding the risks of centralized control. The project incentivizes nodes and miners through tokens, ensuring the vitality of the Bittensor network. Essentially, Bittensor is a GPU miner-driven AI computing power and service project.

Token Economics

The Bittensor network token is TAO, which, in many ways, is similar to BTC to express admiration for Bitcoin. Its total supply is 21 million, with halving occurring every four years. The TAO tokens were distributed through a fair launch when the Bittensor network started, with no pre-mining, so no tokens were reserved for the founding team or VCs. Currently, approximately one Bittensor network block is generated every 12 seconds, with each block rewarding 1 $TAO token, generating about 7,200 TAO daily, which are distributed according to contribution to each subnet and then allocated to subnet owners, validators, and miners within the subnet.

Figure 3 Bittensor Community Promotional Image
TAO tokens can be used to purchase and acquire computing resources, data, and AI models on the Bittensor network, and they also serve as credentials for participating in community governance.

Current Development Status

The total number of Bittensor network accounts has reached over 100,000, with non-zero accounts numbering as high as 80,000.

Figure 4 Changes in Bittensor Account Numbers
In the past year, TAO has risen by several dozen times, currently valued at $2.278 billion, with a price of $321.

Figure 5 Changes in TAO Token Price

3.

Gradually Implementing Subnet Architecture =========

Bittensor Protocol

The Bittensor protocol is a decentralized machine learning protocol that supports the exchange of machine learning capabilities and predictions among network participants, facilitating the sharing and collaboration of machine learning models and services in a peer-to-peer manner.

Figure 6 Bittensor Protocol
The Bittensor protocol includes network architecture, sub-tensors, subnet architecture, validator nodes, miner nodes, and more within the subnet ecosystem. The Bittensor network essentially consists of groups of nodes participating in the protocol, each running Bittensor client software to interact with other networks; these nodes are managed by individual subnets, which adopt a survival of the fittest mechanism, where poorly performing subnets are replaced by new subnets, and poorly performing validator and miner nodes within each subnet are also eliminated. Thus, subnets are the most crucial part of the Bittensor network architecture.

Subnet Logic

Subnets can be seen as segments of independently running code that establish unique user incentives and functions, but each subnet maintains the same consensus interface as the Bittensor mainnet. There are three types of subnets: local subnets, testnet subnets, and mainnet subnets. Excluding the root subnet, there are currently 45 subnets, and it is expected that from May to July 2024, the number of subnets will increase from 32 to 64, adding four new subnets each week.

Subnet Roles and Emission

There are six functional roles within the entire Bittensor network: users, developers, miners, staking validators, subnet owners, and committee members. The subnet includes subnet owners, miners, and staking validators.

• Subnet Owners: Subnet owners are responsible for providing the basic miner and validator code, can set unique additional incentive mechanisms, and allocate miner work incentives.
• Miners: Miner nodes are encouraged to iterate on server and mining code to maintain a competitive edge over other miners in the same subnet; miners with the lowest emissions will be replaced by new miners and will need to re-register their nodes. Notably, miners can run multiple nodes across multiple subnets.
• Validators: Validators earn corresponding rewards by measuring each subnet's contributions and ensuring their correctness. They can also stake TAO tokens on validator nodes, which can earn 0-18% (adjustable) staking rewards.
  Subnet emission is the mechanism for distributing TAO tokens as rewards to miners and validators in the Bittensor network. Generally, 18% of the emission rewards obtained by the subnet are allocated to subnet owners, 41% to subnet validators, and 41% to miners. A subnet contains 256 UDI slots, of which 64 are allocated to validators and 192 to miners; only the top 64 validators with the highest staking amounts can obtain validator permissions and are considered active validators within the subnet. The staking amount and performance of validators determine their status and rewards within the subnet. Miners' performance is rated through requests and evaluations from subnet validators, and poorly performing miners will be replaced by newly registered miners. Therefore, the more tokens validators stake, the higher the miners' computing efficiency, and the higher the total emission of the subnet, leading to better rankings.

Subnet Registration and Elimination

After subnet registration, there is a 7-day immunity period, with an initial registration fee of 100 $TAO, and the price for re-registration doubles, gradually falling back to 100 TAO over time. When all subnet positions are filled, registering a new subnet will remove one subnet with the lowest emissions that is not in the immunity period to accommodate the new subnet. Therefore, subnets need to maximize the staking amounts of validators and the efficiency of miners within the UID slots to ensure they are not eliminated after the immunity period.

Figure 7 Subnet Names
Thanks to the subnet architecture of the Bittensor network, the decentralized AI data network Masa has been implemented, becoming the first dual-token reward system within the Bittensor network, attracting $18 million in funding.

Figure 8 Masa Promotion

4.

Consensus and Proof Mechanisms =======

The Bittensor network includes various consensus and proof mechanisms. In traditional decentralized networks, PoW (Proof of Work) is often used for miner nodes, ensuring that miners contribute to the network and receive rewards based on their computing power and data processing quality; for validator nodes, PoV (Proof of Validation) mechanisms are generally employed to ensure the security and integrity of the network. In the Bittensor network, an innovative PoI (Proof of Intelligence) mechanism is introduced in conjunction with the Yuma consensus to achieve validation and reward distribution.

Proof of Intelligence Mechanism

Bittensor's PoI mechanism is an innovative validation and incentive mechanism that proves participants' contributions through the completion of intelligent computing tasks, ensuring the security of the network, data quality, and efficient utilization of computing resources.

• Miner nodes prove their work by completing intelligent computing tasks, which may include natural language processing, data analysis, and machine learning model training.
• Tasks are assigned by validators to miners, and after completing the tasks, miners return the results to validators, who score based on the quality of task completion.

Yuma Consensus

The Yuma consensus is the core consensus mechanism of the Bittensor network. After validators derive scores based on task completion, they input the scores into the Yuma consensus algorithm. In the consensus algorithm, validators with higher staked TAO amounts have a greater weight in scoring, and the algorithm filters out results that deviate from the majority of validators, ultimately distributing token rewards based on the comprehensive scores.

Figure 9 Consensus Algorithm Illustration

• Data Agnostic Principle: Ensures privacy and security during the data processing process, meaning nodes do not need to understand the specific content of the data being processed to complete computations and validations.
• Performance-Based Rewards: Rewards are distributed based on the performance and contributions of nodes, ensuring efficient and high-quality computing resources and data processing.

MOE Mechanism Working in Harmony

Bittensor introduces the MOE mechanism in the network, integrating multiple expert-level sub-models within a model architecture, where each expert model has relative advantages when addressing corresponding domain issues. Therefore, when new data is introduced to the entire model architecture, different sub-models can work together to achieve better operational results than a single model.
Under the Yuma consensus mechanism, validators can also score expert models, rank their capabilities, and allocate token rewards, thereby incentivizing model optimization and improvement.

Figure 10 Problem-Solving Approach

5.

Subnet Projects ====

As of the time of writing, the number of Bittensor subnets registered has reached 45, with 40 named. In the past, when the number of subnets was limited, competition for subnet registration was fierce, with registration fees once reaching millions of dollars. Currently, Bittensor is gradually opening up more subnet registration slots; newly registered subnets may not perform as well in terms of stability and model effectiveness as those that have been running for a longer time. However, due to the subnet elimination mechanism introduced by Bittensor, in the long run, it is a process where good coins drive out bad coins, and subnets with poor model performance and insufficient strength will struggle to survive.

Figure 11 Details of Bittensor Subnet Projects
Excluding the root subnet, subnets 19, 18, and 1 have received significant attention; their emission shares are 8.72%, 6.47%, and 4.16%, respectively.

• Subnet 19

Subnet 19 is named Vision, registered on December 18, 2023. Vision focuses on decentralized image generation and inference; the network provides access to the best open-source LLM, image generation models (including models trained on the dataset of subnet 19), and other miscellaneous models (such as embedding models).
Currently, the registration fee for the Vision subnet slot is 3.7 TAO, with total node earnings of approximately 627.84 TAO over 24 hours, recovering nodes worth 64.79 TAO in the past 24 hours; if newly registered nodes can reach average levels, daily earnings can reach 2.472 TAO, approximately $866.

Figure 12 Vision Subnet Registration Fee Data
Currently, the total value of recovered nodes in the Vision subnet is approximately 19,200 TAO.

Figure 13 Vision Subnet Recovery Fees

• Subnet 18

Subnet 18 is named Cortex.t, developed by Corcel. Cortex.t is dedicated to building a cutting-edge AI platform that provides reliable, high-quality text and image responses to users via API.
Currently, the registration fee for the Cortex.t subnet slot is 3.34 TAO, with total node earnings of approximately 457.2 TAO over 24 hours, recovering nodes worth 106.32 TAO in the past 24 hours; if newly registered nodes can reach average levels, daily earnings can reach 1.76 TAO, approximately $553.64.

Figure 14 Cortex.t Subnet Registration Fee Data
Currently, the total value of recovered nodes in the Cortex.t subnet is approximately 27,134 TAO.

Figure 15 Cortex.t Subnet Recovery Fees

• Subnet 1

Subnet 1 is developed by the Opentensor Foundation and is a decentralized subnet specifically for text generation; this subnet, as the first project of the Bittensor subnet, faced significant skepticism; in March of this year, Eric Wall, founder of Taproot Wizards, referred to Bittensor's TAO token as a meme coin in the AI field and pointed out that subnet 1 produced similar results through AI for hundreds of nodes when answering text-related questions, which did not improve the effectiveness of solving practical problems.

• Others

From the model category perspective, subnets 19, 18, and 1 all belong to generative models. In addition, there are large models for data processing, trading AI models, etc., such as subnet 22 Meta Search, which provides market sentiment by analyzing Twitter data, and subnet 2 Omron, which learns and continuously optimizes staking strategies through deep neural networks.
From a risk-reward perspective, if a slot can successfully operate for several weeks, the rewards are evidently substantial. However, if newly registered nodes cannot use high-performance graphics cards and optimize local algorithms, it will be difficult to survive in competition with other nodes.

6.

Future Development ====

• From a popularity perspective, the heat of the AI concept itself is no less than that of the Web3 concept, and even a lot of hot money that would originally flow into the Web3 industry has been attracted by the AI industry. Therefore, Web3+AI will remain the market center for a long time to come.
• From a project architecture perspective, Bittensor is not a traditional VC project; since its launch, the project has risen several dozen times, supported by both technology and market.
• From a technological innovation perspective, Bittensor breaks the previous situation where Web3+AI projects fought alone; its innovative subnet architecture can also reduce the difficulty for many teams with AI technical strength to migrate to decentralized networks and quickly obtain benefits. Moreover, due to the competitive elimination mechanism, subnet projects must continuously optimize models and increase staking amounts to prevent being replaced by new subnets.
• From a risk perspective, as Bittensor increases the number of subnet slots, it will inevitably lower the difficulty of subnet registration, increasing the likelihood of mixed projects taking advantage of the situation; at the same time, as the number of subnets increases, the amount of TAO obtained by previously registered subnets will gradually decrease. If the price of the TAO token cannot rise along with the increase in the number of subnets, the rewards are likely to fall short of expectations.