Reshaping AI development: How Bittensor promotes the sustainable development of open source and decentralized AI

Author: IOSG Ventures

Preface

AI development has produced huge breakthroughs in recent years with the progress of data, computing power, and algorithm research, especially the emergence of OpenAI GPT-4, which represents the arrival of the basic LLM large model, which has promoted the improvement of productivity and society. Efficiency changes.

However, the disadvantages of these large closed-source models represented by GPT-4 have also been revealed, that is, centralized models usually have restrictions on third party integrations, which makes AI agents based on centralized models less scalable and interoperable. Get hit.

Therefore, open source large models such as Llama and other series of models are sought after by more and more researchers. However, open source does not mean transparency, and it also faces many challenges.

The main dilemma is that open source AI development does not provide any economic incentives for most contributors. Even if there are some competition rewards, these are usually one-time, and subsequent improvement and development work still needs to generate electricity for love. Unless it reaches a certain scale and has a large community of followers, there will be more possibilities to achieve income and more contributors to continue to improve.

Therefore, the Bittensor AI project attempts to use web3 token mining to make open source AI development more sustainable, verifiable and efficient. Through Yuma Consensus, we introduce and align resources with research side Miner, verifier Validator, and AI project side Subnet Creator, making the entire AI research more transparent and decentralized, and anyone can participate in AI contributions. And win the rewards you deserve.

The performance of tokens in the secondary market also confirms people's expectations. The price has increased from more than 50 US dollars in September 2023 to more than 500 US dollars in December 2024, an increase of 10 times!

Recently, Bittensor investor and founder of Digital Currency Group established Yuma, an accelerator that specializes in incubating subnet projects within the bittensor ecosystem, and served as CEO, showing his confidence and potential in the Bittensor project.

Source: Coindesk

Of course, the success of any project cannot be without doubts. Since the birth of the Bittensor project, there have been a lot of FUDs. In this article, we have summarized many questions that have not yet been fully answered, and tried to understand Bittensor’s future positioning and possibilities in the decentralized AI track through research and analysis.

What is Bittensor?

Bittensor was founded in 2021 by a team from Toronto, Canada, Jacob Robert Steeves, Ala Shaabana, and Garrett Oetken.

Bittensor is a decentralized AI infrastructure used by AI developers to build and deploy machine learning models or other AI-related developments on it. Many Web3 AI projects, regardless of whether they already have their own blockchain, can access Bittensor's blockchain "subtensor" and become a subnet of it.

What is Subnet?

Subnets form the core of the Bittensor ecosystem. Each subnet of Bittensor is an independent incentive-based competitive market. Anyone can create a subnetwork, customize what tasks the subnetwork is to perform, and design an incentive mechanism (in the machine learning analogy, the incentive mechanism can be understood as a target loss function that can guide model training toward the desired results) . You only need to pay the registration fee (priced in TAO) to create a subnet, and you will get a netuid for the subnet. Note that a subnet creator does not need to assume the running tasks within the subnet, but transfers the rights to operate the subnet tasks to others.

The task of operating this subnet provides another way for others to participate, by joining an existing subnet. If you are joining an existing subnet, you can participate in two ways: as a subnet miner or as a subnet validator. In addition to paying the registration fee denominated in TAO (if he is a validator, he also needs to pledge TAO), he only needs to provide a computer with sufficient computing resources, and register the computer and his wallet into a subnet subnet. At the same time, run the subnet miner module miner module or subnet verifier module validator module provided by the subnet builder on this computer (both modules are Python codes in the Bittensor API).

How does the competitive market for subnets work?

Here's how subnet competition works: Let's say you decide to become a subet miner. The subnet validator will assign some tasks to you to complete. Other miners in the subnet will also receive tasks of the same type of work. After all subnet miners complete their tasks, they submit the results to the subnet validator.

Subnet validators then individually evaluate and rank the quality of tasks submitted by subnet miners. As a subnet miner, you will be rewarded (priced in TAO) based on the quality of your work. Likewise, other subnet miners will receive corresponding rewards based on their performance. At the same time, subnet validators will also be rewarded for ensuring that high-quality subnet miners receive better rewards, thus promoting the continued improvement of the overall quality of the subnet . All these competition processes are completed automatically based on the code incentive mechanism written by the subnet creator.

Source: Steps on how Subnet Creator defines Incentive Mechanism

The incentive mechanism is ultimately a judgment on the performance of subnet miners. When incentives are well calibrated, a virtuous cycle can be formed where subnet miners compete to continuously improve on the required tasks.

Conversely, poorly designed incentives can lead to exploits and shortcuts that adversely affect the overall quality of the subnet and disincentivize fair miners.

The specific work of each subnet miner depends on the original purpose of the subnet creator, which can be more changeable or more specific. For example, the task of the miner in subnet 1 is to respond to the text prompt sent by the subnet validator and give the best prompt completion result, or the task of the miner in subnet 47 is to provide storage.

Each sub-network also has its own unique research and commercialization direction, such as trying to overcome technical difficulties in a certain AI field, such as decentralized AI training, verifiable reasoning, or providing some infrastructure and resources required by AI. , such as the GPU trading market or data annotation service, or a subnet that helps users identify AIGC deepfake technology, such as Subnet 34 - BitMind.

Bittensor currently has more than 55 subnets, and this number is still increasing!

Source: IOSG Ventures

The role of Subtensor blockchain

Obviously, the blockchain and the project token TAO played a big role in this series of competitions.

First, the Subtensor blockchain records the key activities of all subnets in its ledger. More importantly, the Subtensor blockchain is responsible for determining reward distribution to subnet miners and subnet validators. An algorithm called Yuma Consensus (YC) runs continuously on the blockchain Subtensor. Each subnet validator ranks the work quality of all subnet miners, and all rankings of each subnet validator will be sent together and provided as collective input to the YC algorithm. Generally, the rankings of validators in different subnets will arrive at Subtensor at different times, but the YC algorithm on Subtensor will wait until all rankings arrive, usually every 12 seconds, and the YC algorithm will calculate the reward based on the ranking input of all validators. These rewards (denominated in TAO) will be deposited into the wallets of subnet miners and subnet validators. The Subtensor blockchain will run the YC algorithm independently and continuously for each subnet.

The YC consensus algorithm mainly considers two factors. The first is a weight vector maintained by each subnet validator. Each element of the vector represents the weight assigned to the subnet miners. The weight represents the subnet miners according to the subnet. The historical execution of validators. Each subnet validator ranks all subnet miners by this weight vector. The second factor is the amount of stake per validator and miner. The on-chain Yuma consensus uses this weight vector and the number of pledges to calculate rewards and distribute them among subnet validators and subnet miners.

The Bittensor API will play the role of transmitting and connecting the opinions of the verifiers on the subnet and the Yuma consensus on the Subtensor blockchain. In addition, validators in the same subnet will only connect to miners in the same subnet, and validators and miners in different subnets will not communicate and connect with each other.

Source: Bittenso

Game Theory of Validator

To participate as a subnet validator or subnet miner, you must first register and stake. Registration means registering the key in the desired subnet and obtaining the UID slot in the subnet. The UID slot represents the verification rights of the subnet. Note that subnet verifiers can have multiple UID slots at the same time and verify for multiple subnets, but there is no need to increase the pledge amount. By staking the amount of TAO once, you can select multiple UID slots and verify for multiple subnets. (Similar to the concept of restaking).

Therefore, in order to obtain the most rewards, verification stakers will tend to choose to provide verification services for all subnets. However, not all staking validators have the right to actually provide staking services. Only the top 64 validators in a subnet ranked according to the number of pledges are considered to have the real verification permission of the subnet. This reduces the risk of validators doing evil, as the amount of stake becomes a high threshold and increases the cost of doing evil (you must have at least 1000 TAO to set a weight in the subnet). In order to increase the number of pledges, each validator will try to establish a good reputation and performance record, attract more TAO entrusted pledges to increase the number of pledges, and become the top 64 validators of the subnet.

Once the subnet validators and subnet miners (no staking is required to run the miners) have registered the keys to the subnet, mining can begin.

Unique token incentivized economy

All TAO token rewards are newly minted, similar to Bitcoin. Bittensor’s $TAO has the same token economics and issuance curve as Bitcoin. TAO supply: The total supply is capped at 21 million, which is halved every 4 years.

Bittensor starts with a fair launch, no pre-mined TAO tokens or ICO. Currently, the network generates 7,200 TAOs per day, 1 TAO is generated per block, and a block is generated approximately every 12 seconds. The total token supply is capped at 21 million, following a programmatic issuance plan similar to Bitcoin.

However, Bittensor introduces a unique mechanism where the issuance rate is halved once half of the total supply is distributed. This halving occurs approximately every 4 years and continues at each halving point of the remaining tokens until all 21 million TAO tokens are in circulation.

Although TAO adopts the issuance curve and philosophy of Bitcoin, due to its recycling mechanism, this curve is actively dynamic and not completely fixed like Bitcoin.

Recycling mechanism:

The daily token issuance for the current cycle is 7,200 TAO (the same as Bitcoin’s first cycle from January 2009 to November 2012).

However, a certain number of dynamic TAOs are recycled through key (re)registration every day.

To become a miner or validator, you must register a key with the network and meet other GPU and computing power requirements. Registration requires recycling TAO, that is, paying a certain amount of TAO to put it back into the network.

Each key (re)registration removes the TAO from the circulating supply and puts it back into the protocol’s issuance pool, where it can be mined again in the future.

This mechanism delays the planned 4-year halving time because the recycled TAO is dynamic. The recycled TAO may increase significantly when more keys are (re)registered, TAO recycling costs increase, or other subnets are released.

Not only that, registration is not only available for new joiners, but also for those who have been deregistered due to the following reasons:

• For miners, their models and reasoning are not competitive enough among other miners;

• For validators, they fail to consistently set the correct weights, maintain issuance, or do not have enough TAO (self-delegation + other delegators’ shares) in the key.

These factors themselves will also exacerbate the growth in registration demand.

The number of recycled TAOs = the total number of registered (or re-registered) keys in each subnet * average registration (or re-registered) cost)

Therefore, the first halving, originally planned to take place 4 years after launch, may be postponed to 5 or 6 years, or even longer. It all depends on the balance between TAO’s issuance and recycling.

The Bittensor network went online on January 3, 2021, and according to token recycling data from taostats, the planned halving date is expected to be delayed to November 2025.

Source: https://taostats.io/tokenomics

What is dTAO?

dTAO is an innovative incentive mechanism proposed by the Opentensor/Bittensor network, aiming to solve the problem of low resource allocation efficiency in decentralized networks. Unlike the traditional method of manually voting by validators to determine resource allocation, dTAO introduces a market-based dynamic adjustment mechanism that directly links resource allocation to subnet network performance, thereby optimizing the fairness and efficiency of reward distribution.

core mechanism

1. Market-based dynamic resource allocation
   The allocation mechanism of dynamic TAO is based on the market performance of subnet tokens. Each subnet in the network has an independent token, and its relative price determines the distribution ratio of TAO issuance among the subnets. As market information changes, this allocation ratio will be dynamically adjusted to ensure that resources flow to efficient and potential subnets.

2. Embedded liquidity pool design
   Each subnet is configured with a liquidity pool composed of TAO and subnet tokens (subnet/TAO token pair). Users can mortgage TAO into the liquidity pool in exchange for subnet tokens. This design incentivizes users to invest in high-performing subnets and indirectly supports the overall development of the network.

3. Fair token distribution mechanism
   Subnet tokens are gradually distributed through the "Fair Launch" model to ensure that the team needs to gradually obtain token shares through long-term contribution and construction. This mechanism avoids the risk of tokens being sold off quickly, while encouraging the team to focus on technology improvement and ecological construction.

4. Balancing the roles of users and validators
   The resource allocation of dynamic TAO is not only determined by the market, but also jointly affected by validators and users. Verifiers need to rigorously evaluate the team's technical capabilities, market potential, and actual performance, just like venture capitalists (VCs). By staking TAO and participating in market transactions, users further promote the formation of the subnet's market value.

Economic model analysis

1. Current financial support
   Data shows that subnets in the current network receive an average of approximately $47,000 in rewards per day, corresponding to an average annual support of approximately $17 million. This funding scale is much higher than the median value of seed rounds (approximately US$3 million) and Series A financing (approximately US$14 million) for traditional AI startups, providing a strong boost to the rapid development of the subnet.

2. future potential
   The current annual budget of Bittensor is expected to reach US$1.3 billion, which is comparable to centralized AI research institutions such as OpenAI and Anthropic. With the launch of dynamic TAO, the new issuance of TAO in the future will mainly flow to the liquidity pool of subnet tokens, further promoting the circulation of capital and value within the ecosystem.

3. long term incentives
   The design of dTAO greatly incentivizes the team to continuously improve its technology and applications by linking issuance volume to market performance. This mechanism also inhibits the short-term behavior of quickly cashing out through over-the-counter transactions (OTC), laying the foundation for the long-term sustainable development of the network.

Impact and Significance

1. Resource allocation optimization

dTAO dynamically adjusts resource allocation through the market to ensure that subnetworks with high utility and high growth potential receive more resources. This mechanism not only improves the overall efficiency of the network, but also promotes competition and innovation.

2. Decentralized AI ecological construction

Bittensor is not only a decentralized AI network, but also an incubation platform for AI networks through dynamic TAO. Competition and collaboration between subnetworks further promote the development of the decentralized AI ecosystem.

3. Incentives for ecological participants

Dynamic TAO balances the interests of users, verifiers and teams, and ensures that all participants can contribute to the growth of the network through an economic incentive mechanism.

4. Increased role of validator

Validators need to play a more important role in the network. They strictly evaluate the value and potential of subnets in a manner similar to venture capital to ensure the scientificity and rationality of network resource allocation.

The launch of dTAO marks a major advancement in the decentralized network resource allocation mechanism. Through market-oriented dynamic adjustments, embedded liquidity pool design and fair issuance model, dTAO achieves efficient and fair resource allocation. In addition, as an AI network incubation platform, it not only empowers the development of subnetworks, but also provides a new development path for the future of decentralized AI networks.

Agents application on Bittensor

Many people say that Bittensor is an AI coin represented by VC dignitaries, and it has lagged behind the application era where a hundred flowers are blooming in the developer framework ecosystem of major agents. With the recent rise of the AI ​​Agents craze, and the total market value of AI Agent-related tokens exceeding US$10 billion, projects represented by the Virtuals ecosystem, in particular, account for US$5 billion in total Marketcap market capitalization (including various practical categories). Investment and investment research analysis agents, such as $AIXBT, $VADER, $SEKOIA, etc.), Bittensor seems to be being left behind in the eyes of many people.

In reality, however, Bittensor still has a lot of “Alpha”. What many people don't realize is that Virtuals/ai16z's success in the field of consumer AI agents and the Bittensor subnet's efforts in decentralized AI infrastructure are complementary to each other.

As the TVL (Total Locked Volume) and influence of Agents expand, powerful training and inference infrastructure becomes more important.

Currently, Virtuals and Bittensor have many ecological cooperations.

Many consumer-facing virtuals protocol agents are powered by the Bittensor subnet, leveraging TAO's computing power and data ecosystem to create new possibilities, such as

$TAOCAT

• TAOCAT is an artificial intelligence agent built by Masa in the Virtuals ecosystem. Its main role is to serve as a staunch defender of TAO, actively participate in discussions on X, and speak out for TAO's influence.
• TAOCAT utilizes the real-time data infrastructure of subnet 42 Masa and the advanced LLM provided by Bittensor subnet 19 to compete for TAO token distribution in the Agent Arena on Bittensor subnet 59, creating a new paradigm of tokenized AI value capture. Any user interaction on X will become training data for TAO Cat.

Other projects supported by the Bittensor subnet include:

• $AION: The first Agent that can predict prediction results and participate in prediction market betting. The copy-trading copy trading function will also be launched soon.
• $SERAPH: The first project focused on verification infrastructure designed to certify the wave of AI Agents that are about to overwhelm our digital world.

The cooperation between Virtuals and Bittensor proves that huge practical value can be created based on Bittensor infrastructure. With the official launch of AgenTAO (SN62), this will become an important milestone for the automated software engineering Agent on Bittensor. All Bittensor subnets will gradually be developed by Agents on Bittensor. In the future, we will see more application-side AI Agents emerge from the Bittensor ecosystem!

Source: taogod

Conclusion

The future of Bittesor is exciting, and many research and investment institutions specifically surrounding the Bittensor ecosystem have begun to appear, similar to the Ethereum network. Including the DCG founder’s orders, podcasts, blogs, and OSS Capital, which focuses on Bittensor investment. It is dedicated to studying the Bittensor ecology and is a subnet research organization. A network of people similar to the Paypal gang’s Bittensor has been formed. , Contango, Canonical, Delphi Labs and DCG recently held a gathering, and many experts in the Crytpo x AI community began to rely on Bittensor and support it. Therefore, it is not unreasonable for Bittensor to surpass Virtuals on Kaito’s mindshare some time ago. Passing Virtuals is actually not unreasonable.

Source: BitMind Bittensor Subet 34

Next year in April 2025, Austin, Texax Bittensor will hold a 300+ person The Endgame Summit conference and hackathon, specifically focusing on introducing more subnets, validators and miners to the bittensor ecosystem and broadening its territory.

Endgame Summit

Of course, whether it is a centralized AI project or a decentralized AI project, the final standard will return to the product. Currently,

The Bittensor ecosystem has emerged and is flourishing.

Source: Outpost AI Research

Recently, the founder of Bittensor summarized the main achievements of various subnets of Bittensor in the past year on his personal X:

Source

Therefore, let us continue to look forward to Bittensor and see what products and use cases will come out of Bittensor in the future, and let us become the first choice for people to find solutions to specific AI problems!