Comprehensive analysis of the decentralized AI network Bittensor

Original title: "Demystify Bittensor: How's the Decentralized AI Network?"

Author: Ming Ruan, Wenshuang Guo, Animoca Brands Research

Compiled by: Scof, ChainCatcher

brief overview

• The development of artificial intelligence has reached a critical tipping point, and its profound impact on the world is not only inevitable, but also expected to grow exponentially in scope and depth. The market size is expected to exceed US$1 trillion in the next five years, and projects that can accelerate the growth of AI will create huge value.
• Powered by TAO tokens, Bittensor is a decentralized artificial intelligence network designed to allow projects to crowdsource AI-related “digital goods” as an alternative to decentralized AI training.
• Bittensor's functionality relies on two core mechanisms. The first is the innovative Yuma consensus, which enables validators to agree on reward distribution among miners. The second is the continuous issuance of TAO tokens, with 1 token issued every 12 seconds. The issuance rate is halved every four years, creating a predictable but decreasing supply of TAO.
• The basic building block of Bittensor is a subnet, which consists of three key players: 1) subnet owners, who set the goals of the subnet; 2) miners, who provide computing power and compete for a larger share of the rewards; 3) Validators, evaluate the performance of miners and receive rewards for their evaluation. The total reward allocated to a subnet is determined by the "root subnet" or subnet 0.
• Currently, there are more than 50 subnets operating on Bittensor, covering a variety of AI-related needs including infrastructure, data sources, model training, fine-tuning, etc. Some subnetworks surpass centralized alternatives in generating specific dimensions and demonstrate early success.
• Like many newly formed ecosystems, Bittensor’s economic model still has flaws. Its top-down design of funding allocation has yet to fully align the interests of all parties. Additionally, the price of the TAO token that supports network funding remains vulnerable to the upcoming halving event. We propose a method to solve these problems and improve the interest alignment and sustainability of the Bittensor ecosystem.

The need for decentralized AI

The field of artificial intelligence is advancing at an unprecedented pace, but it is not without challenges. Currently, centralized data training models dominate the field, mainly controlled by technology giants such as OpenAI, Google, and X (formerly Twitter).

Although centralized AI training has made remarkable achievements in recent years, it also has certain limitations. First, there are some problems in the data training process, such as unauthorized use of private information, data censorship leading to distortion of training results, and lack of traceability of data sources. In terms of algorithms, centralized models rely heavily on data quality and are often difficult to evaluate in real time for iterative improvements.

Decentralized AI training presents an alternative, but faces significant challenges, especially insufficient resources. Currently, the cost of training large models exceeds $100 million, making it nearly impossible for community-driven projects to compete. Decentralization efforts rely on voluntary contributions of computing power, data, and talent, but these resources are insufficient to support a project of similar scale. Therefore, the potential of decentralized AI remains limited and cannot fully compete with centralized AI in scale and influence.

Source: Statista

Bittensor Overview

Bittensor is a decentralized network designed to form an intelligent marketplace that enables high-quality AI models to be developed in a decentralized manner. By leveraging the incentive mechanism and rewarding participants for providing computing resources, expertise and innovative contributions, Bittensor has established an open source AI capability ecosystem, in which the native currency TAO is both a reward token and a credential for accessing the network.

Bittensor’s core components, including its Yuma consensus, subnet, and TAO token, were initially launched in November 2021 with the release of version “Satoshi Nakamoto” and built as a parachain for Polkadot. Subsequently migrating to a Layer 1 chain built on Polkadot Substrate in 2023, TAO's issuance plan remains unchanged.

The founder and operating body of Bittensor, the Opentensor Foundation, was co-founded by former Google engineer Jacob Steeves and machine learning scholar Ala Shaabana. The foundation currently has about 30 employees, almost all of whom are engaged in engineering functions. It lacks B2B market expansion, business Functions such as development, partnerships, or developer relations.

Fundamentals: How does Bittensor work?

Bittensor has developed an innovative network based on a dynamic incentive consensus framework that allows participants to contribute resources needed to produce machine intelligence. Each subnet operates as a model for a specific task, has its own independent performance evaluation criteria, and is passed Bittensor’s overall Yuma consensus distributes incentives.

Let's use an analogy to illustrate how subnets work. A subnet can be compared to a publisher of a magazine that organizes monthly writing competitions. Each month, an editor posts a topic for writers to compete for a $10,000 prize pool. The standard is "the work that best embodies the spirit of web3". Writers submit their articles to editors for review. All editors will evaluate all submitted works, and the editors' scores will determine the final ranking. The highest-ranking articles will be published and receive the largest share of rewards, while lower-ranking articles may also receive smaller rewards. All submitted articles and their ratings are shared with participating writers and editors for feedback and learning. Through such an incentive structure, writers will continue to participate and contribute, and the standards between writers and editors will gradually converge, allowing the magazine to publish high-quality articles that best "embody the spirit of web3."

In this analogy, magazine publishers represent subnets, writers represent miners, and editors represent validators. The evaluation process of articles summarized by editors is the Yuma consensus mechanism. In a real subnet, miners will receive TAO tokens instead of USD, which are distributed by the root subnet (subnet 0); validators will also be incentivized so that their standards are close to the aggregated score, thus earning more Lots of rewards.

Under this framework, subnet owners train and acquire intelligent capabilities from miners through verifiers to build AI modules with specific functions. In addition to subnets, Bittensor also has other layers that support the functionality of the entire network:

1. Application layer : External applications send requests to the subnet to get smart responses.
2. Execution layer : consists of a set of subnets, each of which trains and utilizes miners to achieve its goal of developing intelligence and other related capabilities.
3. Funding layer : The root subnet (subnet 0) is responsible for allocating TAO issuance to the subnet to fund its activities.
4. Blockchain layer : issues TAO and records transactions on the subnet chain.

a. Application layer

Users can interact with Bittensor through various applications, which are connected to or as a subnet. Users make requests for services, such as language translation or data analysis, and the application routes the request to the subnet through the validator API. The best miner answer is selected by validator consensus and returned to the user.

b. Execution layer

This layer consists of a set of subnets, all of which use Yuma consensus to train and utilize miners. Without going into depth about individual subnets, we will discuss the Yuma consensus and subnet participation here.

Yuma consensus

Yuma consensus is an algorithm that establishes consensus among validators and is used to determine reward distribution. The simplified way to think about it is that each validator assigns 1 point of its total score to the miners. Each miner's consensus score will then be the weighted average of all validator scores. This score determines the share of rewards miners can receive.

subnet

Task-specific subnets use a framework designed for their goals, called a "validation stack." Validators execute the stack, guiding miners toward value-creating tasks and ensuring that subnet goals are achieved. The goal of miners is to outperform their peers and win a larger share of the rewards.

To become a subnet owner, you need to pay a registration fee before you can connect to the subnet. The registration fee will fluctuate based on demand and is currently approximately 3,000 TAO. Fees will be refunded when the subnet is deregistered.

Becoming a validator requires staking TAO. The minimum stake threshold is determined by the validator ranked 64th in stake. Validators with larger stakes have higher weight in each subnet’s service requests and Yuma consensus. Additionally, a validator can serve in multiple subnets.

To join the subnet, either as a validator or a miner, you need to register for a slot called a "neuron". Registration fees vary by subnet and vary based on demand and supply, but are typically less than 1 TAO. Once the registration is cancelled, it will be withdrawn and will not be refunded. In addition to the entry fee, miners are required to custom configure their software and hardware to serve the purpose of the subnet. Therefore, miners are usually bound to a subnet.

As of December 2024, there are a total of 250 validators operated by 173 user accounts and 11,856 miners operated by 2,709 user accounts.

Incentive distribution

Within each subnet, the allocated TAO issuance is determined by the root subnet (root in the funding layer) and is allocated to validators and miners according to a predetermined ratio: 41% is allocated to miners, 41% is allocated to validators, and 18% is allocated To the subnet owner. Among miners, incentives are determined based on the "trust value" assigned by the validator. Among validators, incentives are distributed based on the validator’s “trust score” and stake amount. For validators who have delegated TAO mortgages, the validators will share the rewards with other stakers and distribute them after deducting the "share" they retain.

c. Funding layer

The root network, also known as subnet 0, is the funding layer in the Bittensor stack. Within the root subnet, each validator runs the verification stack across all subnets, assessing the quality, accuracy, and response time of each subnet's own benchmark return results to determine its score. These scores are then converted via Yuma consensus into the subnet’s share of the TAO issuance.

The root subnet has two unique characteristics: first, miners are replaced by the subnet as the subject of evaluation; second, the number of validators in the root subnet is fixed at 64. In the design, new applicants must stake a TAO that exceeds the minimum stake of existing validators to join the root subnet. However, this alternative mechanism is currently suspended, so a fixed set of 64 validators controls the root subnet, although the amount of collateral varies.

d. Blockchain layer

Subtensor is the blockchain in Bittensor and is responsible for issuing TAO. Validators in the subnet submit weight scores, transaction data, and performance metrics to the Subtensor blockchain. The validator nodes of the Subtensor blockchain are operated by the Opentensor Foundation using Proof-of-Authority and are responsible for verifying transactions, updating the Subtensor ledger, and managing reward distribution. Claims of decentralization around this blockchain layer are controversial, as the Opentensor Foundation team can pause the blockchain when needed.

Evaluation subnet

As of December 2024, there are 56 active subnets. These subnets cover multiple aspects of AI development, such as training data pipelines, computing power, training platforms, general AI models, and application-specific AI tools. Issuance distribution among subnets is uneven, with the top 10 subnets accounting for approximately 50% of total issuance.

In the remainder of this section, we'll dive into three interesting subnets to illustrate how subnet owners use them.

Subnet 18: Cortex.t

Cortex.t is a subnet developed by Corcel under the DSIS framework with the goal of generating dynamic synthetic data via GPT4o and GPT4 for model testing and unbiased artificial intelligence evaluation. It creates high-quality prompt-response pairs and archives them as synthetic question and answer data on wandb.ai, while leveraging techniques such as prompt evolution and data augmentation to optimize the output.

In the Cortex.t subnet, miners process hints requiring GPT4o and GPT4 output, and their accuracy, speed, and efficiency are evaluated by validators. These validators use the same model to form a synthetic database while managing the API server to send prompts. In addition, validators can sell bandwidth as a service under the DSIS framework for production-level applications.

Subnet 37: Model fine-tuning

The model fine-tuning subnet is designed to leverage decentralized capabilities to train advanced specialized models, such as chatbots or inference systems. Model fine-tuning is often time-consuming, computationally intensive, and requires specific skills. By applying the subnet structure, miners can contribute their skills and resources to improve the model and be rewarded accordingly.

Miner tasks are organized as competitions to fine-tune models. The base model, constraints, and goals are announced for each competition. Miners use the basic model as a starting point, run fine-tuning offline, submit the fine-tuned model to Hugging Face (an artificial intelligence community website), and submit model metadata to the Bittensor chain.

The validator retrieves metadata to identify the model and uses synthetic QA data from subnet 18. Verifiers then measure miner performance by evaluating the miner model's multiple-choice accuracy on the SYNTHETIC_MMLU task. Miners are ranked based on the validators’ scores and receive corresponding rewards.

Subnet 5: Open Kaito

Subnet 5 focuses on developing text embedding models to enhance applications such as semantic search and natural language understanding. Miners train these models using large-scale text datasets to ensure their efficient performance in generating embeddings. Validators drive continuous improvement by evaluating model performance against state-of-the-art benchmarks.

These models will be accessible via a public validator API to facilitate integration into a variety of applications, with the goal of exceeding existing performance standards.

The Open Kaito team recently claimed that subnet 5 miners have surpassed the OpenAI embedding model in external retention benchmarks. This is a breakthrough based on the lower loss and higher Top-1 accuracy achieved previously using large-scale synthetic data sets. This success is significant because it shows that decentralized AI model training can in some ways outperform centralized models.

Source:wandb

From these subnet examples, we can see early signs of Bittensor’s success as a decentralized AI ecosystem. At the individual subnetwork level, decentralized model training and fine-tuning is feasible. This is also supported by Prime Intelligence’s recent success in training the first decentralized 10 billion parameter model INTELLECT-1 . At the ecosystem level, sub-networks can cooperate with each other to form a value chain. Therefore, not only the model training part, but even the entire artificial intelligence value chain may be decentralized through Bittensor.

After evaluating the technical feasibility of decentralized AI on Bittensor, we will further explore the economics of its ecosystem to determine whether its design can support a sustainable AI training value chain.

TAO token economic model

Bittensor's incentive token TAO is built on its blockchain Subtensor. Currently, a new block is generated every 12 seconds, and 1 TAO is minted per block and distributed to network participants as rewards. The issuance rate of TAO will be halved for the first time after the first batch of 10,500,000 tokens are minted. Subsequently, whenever it reaches half of the issuance of the previous halving cycle (for example, the second halving is 5,250,000 tokens), the issuance rate will be The rate will continue to be halved. This halving mechanism sets the total supply cap of TAO at 21 million. As of December 2024, approximately 37.8% of the total supply, or 7.93 million TAO, has been minted.

The time of each halving event is mainly determined by the block generation rate and is also affected by token recycling. Every time 1 TAO is recycled (usually from blockchain handling fees or subnet participant registration fees), it prevents the minting of 1 new TAO, thereby delaying the halving time of the block generation cycle. According to the latest estimates, the next halving date is November 29, 2025 .

In the previous article, we have mentioned that the main purpose of TAO is as a reward token to incentivize network participants. Here are other uses for TAO:

• Staking : TAO holders can stake TAO directly to their own verifiers, or entrust them to verifiers and share the rewards of verifier activities. The minimum pledge amount is 0.1 TAO.
• Network Admission : Miners, validators, and subnet owners need to pay a TAO registration fee to join the Bittensor network.
• Governance : TAO holders can influence network governance decisions, such as protocol upgrades and issuance adjustments, through the "Senate", whose members include active subnet validators.
• Transaction fees : TAO is the fuel fee token for the Subtensor blockchain.

Is Bittensor sustainable?

Bittensor is positioned as a smart marketplace that allows users seeking artificial intelligence capabilities to attract validators and miners to cooperate by establishing subnets. However, its mechanism deviates from traditional market structures: subnet owners are not directly compensated for the services of miners or validators, nor are these contributors’ income tied to their output or workload.

In reality, Bittensor is more of a top-down funding system than a market. Subnet activity is incentivized through grants allocated by the "root" network, but the criteria for determining grant amounts do not always align with the actual value or workload of the subnet. Furthermore, the total grant pool is always fixed regardless of how the number of subnetworks in the ecosystem changes.

To better understand this practice, consider an analogy to a city. In this city, in addition to the magazine publishers we discussed earlier, there are also business entities such as book publishers, music studios, and acting schools. These businesses, whose sole source of revenue is city government appropriations, continue to receive funding as long as they are approved by the Appropriations Committee, regardless of their actual value creation. For new companies, they need to purchase land from the government to qualify. As new businesses are added, the total grant pool remains fixed regardless of how the number of businesses changes.

This non-traditional system distorts market dynamics and leads to inefficient allocation of resources. Grant allocators (the root network) lack effective means to measure subnet contributions, and subnets have no incentive to share profits from their AI capabilities with the broader ecosystem. Miners tend to use minimal effort in exchange for maximum TAO distribution, rather than increasing capabilities or taking on more workload.

Although a proposal for Dynamic TAO was put forward in January 2024 to address some of the inefficiencies, the proposal has not yet been implemented and its effectiveness remains uncertain.

TAO's incentive model is highly dependent on TAO's strong price, and most network participants' main income is TAO, which leads to continued selling pressure. To balance this pressure, staking has become the main mechanism, but the amount of token recovery brought by blockchain fees and registration fees is still limited.

There are two forms of pledge:

1. Validator Staking : Participants stake TAO to support network security and receive rewards, accounting for approximately 75% of all issued TAO. Validators currently distribute 2,952 TAO per day, with an annualized return of 16%. However, after the first halving, this allocation will be reduced to 1,476 TAO per day, and staking will become less attractive, thereby weakening the impact of staking on the balance of token supply and demand.
2. Subnet registration pledge : The subnet registration fee is approximately 3,000 TAO, and the addition of new subnets significantly affects the supply of TAO. But this also raises a problem: Since the total issuance of TAO is fixed, an increase in the number of subnets will dilute the rewards of all subnets, making it difficult for existing subnets to maintain operations, and may cause some subnets to exit the network.

Bittensor’s current economic model is not sustainable. Its top-down funding structure failed to efficiently allocate resources across subnets. More importantly, the demand for TAO is insufficient to support its post-halving value, which exacerbates the vulnerability of the network and threatens its long-term viability.

our proposal

We propose a two-part strategy to enhance Bittensor’s sustainability:

1. Incentive subnet contribution : Allow subnet owners to provide additional TAO rewards to their subnets, which will be added to the total incentive pool allocated to the subnet by the root network and distributed to participants through the existing consensus mechanism. This move will incentivize those subnet owners who derive significant value from the subnet to contribute funds to the subnet reward pool, thereby ensuring active participation of miners and validators and allowing subnet owners to become the purchasing power of TAO tokens, effectively support its price.
2. Prioritize allocations : The root network should prioritize allocations for new and high-potential subnets, while gradually reducing support for older subnets. This move will generate lower value subnets through natural elimination, ensuring that new subnets are fully funded without being diluted by the total number of subnets. Additionally, this approach will reduce the burden on root network validators, allowing them to focus on the growth of new subnets, more in line with a top-down funding model.

By implementing these strategies, Bittensor can create ongoing demand for the TAO token, helping it maintain its value rather than relying solely on staking. At the same time, these measures promote ecosystem growth by introducing natural selection mechanisms and pool resources to incubate new sub-networks.

Conclusion

Artificial intelligence undoubtedly represents the future of technological advancement. This is reflected in the high valuations of leading companies in the artificial intelligence value chain and its potential for widespread application in various areas of society. While centralized AI development has driven progress, it has also exposed the drawbacks of reliance on centralized data, model development, and profit concentration.

Bittensor provides a powerful alternative to decentralized artificial intelligence. Supported by a top-down funding model and a solid TAO token price, multiple subnets have demonstrated the potential to drive the development of artificial intelligence capabilities. Overall, Bittensor can form a comprehensive platform covering the entire value chain of the artificial intelligence ecosystem.

However, like other emerging ecosystems, Bittensor faces challenges, particularly with regard to the sustainability of its token economic model and the effectiveness of its reward distribution system after the first halving. To address these issues, we propose adjusting the reward model to prioritize support for new and high-potential subnets, making it work like a venture capital investment. This will enable existing subnet owners to fund their participants and benefit from decentralized AI contributions.

With this adjustment, we believe Bittensor’s incentive model can achieve sustainability, thereby turning our attention to the truly critical question: When will decentralized artificial intelligence create high-value real-world applications?