Bittensor (TAO) : A comprehensive presentation of a protocol combining AI and blockchain

What is Bittensor (TAO)?

Bittensor is a blockchain designed to revolutionize how intellectual and computational resources are exchanged. Its ambition is to create an open and decentralized peer-to-peer marketplace that fosters technological innovation through the exchange and valuation of artificial intelligence (AI) models.

Unlike traditional blockchains, which focus on monetary transactions or smart contracts, Bittensor positions itself as a marketplace dedicated to collaboration and innovation in AI.

Bittensor consists of a main network, the Subtensor, and is built around a system of sub-networks called Subnets. These Subnets create a competitive environment where market participants — miners and validators — tackle a wide array of computational and specialized tasks.

A key element of Bittensor is Yuma, a unique consensus mechanism based on the quality of contributions. Participants are rewarded in TAO tokens, proportional to the relevance and effectiveness of their responses. This competitive model drives participants to continually refine their computational models to achieve better results and earn more TAO rewards, thereby fostering technological innovation.

---

Bittensor’s Thesis

It’s easy to overlook one of the blockchain’s most transformative features: its ability to align the actions of a decentralized group of participants who don’t know each other, all working toward a shared goal, through financial incentive mechanisms.

For instance, Bitcoin uses mining as an incentive system to secure the network. As Bitcoin’s price increases, mining rewards become more attractive, drawing in new miners and strengthening the network’s security. This decentralized mechanism aligns miners' interests through financial rewards, without requiring them to know or interact with each other.

Bittensor builds on this powerful concept but extends it to a new domain: artificial intelligence. Rather than relying on a narrowly defined incentive system, such as securing transactions, Bittensor provides a framework that enables anyone to design custom incentive systems for a variety of computational tasks.

The network revolves around Subtensor, a main blockchain that hosts and coordinates multiple Subnets. Each Subnet is configured to address specific needs, such as text or image generation via AI, computational power lending, financial predictions, and more.

Bittensor provides an open framework where individual performance within the network is rewarded, and collective competition fosters the emergence of more effective and sophisticated solutions to real-world computational challenges. Thus, Bittensor not only creates a market for exchanging digital resources but also redefines how disparate entities can collaborate to tackle complex problems on an unprecedented scale.

---

Architecture and Technology

Subtensor

Subtensor acts as the backbone of the Bittensor network. As the main blockchain, it is responsible for two key functions:

• Task Coordination: It logs summaries of all tasks completed across Subnets and ensures transparency regarding contributions.
• Reward Distribution: Through the Yuma consensus mechanism, Subtensor automatically calculates and distributes TAO rewards to all participants based on the quality of their contributions.

Additionally, Subtensor supports the transfer of TAO tokens between addresses. Recently, it has also introduced Ethereum Virtual Machine (EVM) compatibility, enabling the execution of smart contracts directly on the Subtensor. This advancement paves the way for more complex applications on Bittensor, such as liquid staking, lending, and borrowing TAO.

Subtensor operates on a Proof of Authority model, meaning that only approved members can validate the network. Currently, all Subtensor nodes are owned and operated by the Opentensor Foundation, granting it full control over Subtensor. The network is supported by lightweight nodes and archive nodes that verify and store blockchain data but do not participate in transaction validation.

---

The Subnets

Subnets are autonomous sub-networks connected to Subtensor. Each Subnet is dedicated to a specific task and operates with a unique incentive system designed to drive contributors toward achieving its goals. Subnets consist of three primary actors:

• Subnet Owners: Responsible for creating and managing the Subnet, they define the tasks to be accomplished and the incentive system. They also provide the methodology for validating results and promoting the Subnet to attract contributors.
• Miners: These individuals execute assigned tasks using their computational models, which they can optimize over time to enhance their efficiency and maximize their rewards.
• Validators: Validators assign and distribute tasks to miners, evaluate the results, and submit their ratings to Subtensor. These ratings determine the rewards for each contributor.

---

Task Lifecycle

Let’s break down the lifecycle of a task within a Subnet, using Subnet 1 — specialized in conversational AI model development — as an example:

1. Task Creation: A validator within Subnet 1 creates a task. This task can originate from the validator itself or from an external service accessing the validator's API (e.g., an application like ChatGPT).
2. Task Assignment: The validator formats the task to meet Subnet standards and assigns it to multiple miners.
3. Task Execution: Each miner uses its computational model to respond to the task, submitting its output to the validator.
4. Evaluation: The validator evaluates and ranks all responses based on quality, selecting the best one to forward to the external service that made the initial request.
5. Reward Distribution: Validator ratings are sent to Subtensor, where the Yuma consensus determines and distributes TAO rewards to Subnet participants.

This mechanism is adaptable to any computational task requiring intelligent problem-solving.

Yuma Consensus

The Yuma consensus lies at the heart of Bittensor’s functionality, particularly in the distribution of rewards. Each validator assigns ratings to miners based on the quality of their work. These ratings are aggregated by Subtensor, which uses the Yuma consensus to calculate the rewards for all network participants.

Generally, the more highly a miner is rated by multiple validators, the more TAO they earn. Similarly, validators whose ratings align closely with the average ratings of all validators in the Subnet are also rewarded with higher TAO allocations.

Reward allocation within a Subnet follows this distribution:

• 18% for Subnet Owners
• 41% for Validators
• 41% for Miners

This system incentivizes miners to continuously improve their computational models, leading to better service for end users. With properly configured incentive structures, Subnet Owners can drive similar advancements in other computational tasks, illustrating the nearly limitless potential use cases for Bittensor.

Root Network

The Root Network, also known as Subnet 0, plays a critical role in distributing rewards among the various Subnets within Bittensor. While we have discussed how TAO rewards are distributed among participants within a single Subnet, the Root Network determines how rewards are allocated across all Subnets.

Every 12 seconds, Subtensor generates a new block and issues 1 TAO. However, this token is not evenly distributed among all Subnets. Instead, the Root Network is responsible for determining the allocation of these rewards. Subnet 0 is unique, as it comprises the 64 largest validators across all Subnets, selected based on the amount of TAO delegated to them.

In this structure, the miners of Subnet 0 are the other Subnets currently registered on the network. Just like in any other Subnet, the validators evaluate the performance of these miners and assign scores. These scores are sent to Subtensor, which uses them to decide how TAO is distributed to each Subnet. Ultimately, this allocation impacts all participants within those Subnets.

---

Life and Death of a Subnet

Bittensor’s design imposes a limit on the number of Subnets that can exist simultaneously. This cap is currently set at 64 but is gradually increasing to accommodate network growth, with a long-term goal of reaching 1,024 Subnets.

This limitation fosters competition among Subnets, ensuring only those that provide the most value to the network can remain active. The value of a Subnet is determined by the amount of TAO it receives, which is in turn decided by the votes of the Root Network validators.

If a Subnet consistently receives the fewest rewards, it becomes vulnerable to removal. When a new Subnet registers, the lowest-performing one is delisted. Newly registered Subnets enjoy a seven-day grace period during which they cannot be removed, allowing them time to attract miners and validators.

Registering a new Subnet requires the Subnet Owner to deposit TAO through a Dutch auction mechanism. The cost of registration decreases over time, with a minimum of 100 TAO. Once a Subnet is registered, the registration cost doubles for the next Subnet. Importantly, when a Subnet is delisted, the registration cost is fully refunded to its owner.

However, the fees for registering validators and miners are much smaller and are not refunded upon delisting. Instead, these fees are returned to the TAO pool, awaiting reallocation through the network’s reward mechanisms.

Bittensor EVM

Recently, the Opentensor Foundation announced the integration of the Ethereum Virtual Machine (EVM) into Subtensor. This feature enables the execution of smart contracts directly on the Bittensor network, paving the way for a variety of decentralized applications, particularly in decentralized finance (DeFi).

The inclusion of EVM is a strategic decision, as it significantly lowers the barrier to entry for developers already familiar with Ethereum’s ecosystem, which remains the most widely adopted blockchain for smart contract development. With EVM support, Bittensor can attract a broader developer community and foster the creation of advanced applications, such as liquid staking, TAO lending, and borrowing.

---

TAO: The Cryptocurrency of Bittensor

Role and tokenomics

TAO is much more than a simple token; it is the cornerstone of Bittensor's economy, playing a vital role in value distribution and coordination within the network. Its key uses include:

• Rewarding Network Participants: Subnet Owners, validators, and miners receive TAO for their contributions.
• Staking and Network Security: Users can stake TAO with validators to participate in the validation of miners' contributions.
• Transaction Fees: Transferring TAO and utilizing the EVM on Subtensor requires TAO.
• Subnet and Role Registration: Depositing TAO is necessary to register new Subnets or to enroll as miners or validators.

With a maximum supply capped at 21 million, TAO follows a model inspired by Bitcoin, with no premine, no private investors, and no team allocations. Tokens are minted at a linear rate of 1 TAO every 12 seconds, equivalent to 50 TAO every 10 minutes. Currently, approximately 8.1 million TAO are in circulation, with 75% staked via validators—a strong indicator of robust community engagement.

Inflation and halvening

The distribution of TAO is governed by the halvening mechanism, reminiscent of Bitcoin's halving events, which reduce TAO emissions by half. However, unlike Bitcoin, TAO halvings are based on the number of tokens issued rather than the number of blocks mined.

Two unique mechanisms influence this process:

• Recycled TAO: Tokens used for role registration (miners and validators) are reintroduced into circulation, delaying the next halvening.
• Validator Reserves: Root Network validators can choose to withhold a portion of the available rewards for future distribution. Currently, approximately 8% of TAO per block is kept in reserve.

These mechanisms ensure that as Bittensor’s activity grows, the next halvening is delayed. Over time, TAO emissions could balance out with recycled TAO, potentially making the next halvening the final one. This feature addresses a common concern in the Bitcoin community: what happens when the last Bitcoin is mined in 2140? Will transaction fees be sufficient to sustain miners? If Bittensor withstands the test of time and gains widespread adoption, this problem may never arise.

---

Dynamic TAO (dTAO)

Background

Bittensor's current reward system relies on the Root Network to allocate TAO emissions to each Subnet. These emissions are crucial, as they measure the value brought by each Subnet and, in turn, influence its long-term viability within the network.

This centralized model, however, gives the 64 Root Network validators significant influence, raising potential risks of collusion between leading validators and Subnet Owners. Such scenarios could result in the exclusion of valuable projects or an excessive allocation of TAO to projects offering little value.

In the long run, this system could create a significant mismatch between the actual value provided by Subnets and the TAO emissions they receive, eroding user trust and discouraging new Subnet creators from adopting Bittensor. Despite the good intentions of most validators, this scenario remains likely as long as reward allocation depends on a small group of actors.

---

How dTAO Works

The Dynamic TAO (dTAO) update addresses these challenges by delegating reward determination to market dynamics. Each Subnet introduces its own token, called dTAO, along with an associated liquidity pool. The value of dTAO tokens fluctuates independently of TAO.

When users wish to stake TAO on a particular Subnet, they must first convert their TAO into that Subnet's dTAO. Since dTAO prices vary, careful consideration is required. If the Subnet gains popularity, its dTAO appreciates, allowing users to exit with more TAO than initially staked (excluding staking rewards). Conversely, if the Subnet declines, users risk losing value as its dTAO depreciates.

---

Implications of dTAO

With Dynamic TAO, reward allocation will now be determined by the market value of each Subnet's dTAO. As Subnets gain popularity and attract more TAO, their dTAO prices rise, leading to increased TAO emissions.

Additionally, rewards will be distributed in dTAO rather than directly in TAO. Consequently, network participants (Subnet Owners, validators, or miners) must convert their rewards into TAO. This conversion may lower the dTAO price of their Subnet, reducing future rewards.

This mechanism ensures that Subnets receiving disproportionate emissions relative to their value will see a decrease in rewards over time. Conversely, Subnets offering significant value but receiving fewer rewards will attract investors seeking to acquire their dTAO, boosting staking participation and TAO emissions.

Ultimately, Dynamic TAO transforms Bittensor's economic governance, aligning reward allocations with each Subnet's real value. This dynamic market model could inspire similar applications in other blockchain projects.

---

The Limitations of Bittensor

As we’ve observed so far, Bittensor offers a unique and ambitious approach with use cases already in motion. However, the project is not without its limitations and points of centralization. While some of these issues are being addressed, they merit examination to fully understand the challenges facing the network. Here’s an overview of Bittensor’s key limitations and the efforts to overcome them.

Proof of Authority

Subtensor, the main blockchain supporting Subnets, operates on a Proof of Authority (PoA) consensus mechanism. This means that only verified nodes can participate in validating transactions. Currently, all Subtensor nodes are controlled by the Opentensor Foundation, which has full authority over the network, including the ability to censor transactions. This is a significant issue for a blockchain valued at over $5 billion.

In the short term, the Opentensor Foundation has no incentive to act against the network, as doing so could undermine user trust. However, for Bittensor to grow and establish itself as an independent and decentralized network, this authority must gradually be transferred to an open validation system based on Proof of Stake (PoS).

This challenge is well-known within the community and by the project team, which plans to transition to a PoS consensus system over time. However, as of now, no timeline has been set for the development or implementation of this solution.

Blockchain Size

Although transaction validation is restricted to nodes controlled by the Opentensor Foundation, anyone can contribute to decentralizing the chain by hosting a lightweight or archival node. These nodes don’t participate in consensus but instead record all blockchain data to ensure continuity in the event of a network failure.

According to the documentation, hosting a lightweight node requires only 128 GB of storage, as it stores a limited subset of transactions. This increases to 1.5 TB for archival nodes, which store all blockchain data since its inception. This suggests that by early 2025—approximately two years after its launch—the blockchain size could reach around 1 TB. By comparison, ten years after launch, Ethereum's blockchain is around 900 GB, and Bitcoin’s is about 600 GB after 15 years.

As Bittensor’s popularity grows and the number of Subnets continues to increase, the blockchain size is expected to grow exponentially. This is bad news for decentralization; if validation becomes open to all and requires storing tens of terabytes of data, only a small number of individuals would be capable of participating in network validation, rendering Bittensor's decentralization impractical.

Governance of Bittensor

Like Subtensor validation, Bittensor's governance is currently restricted to a select few entities. It consists of two distinct components:

• The Triumvirate: Made up of three Opentensor Foundation employees responsible for proposing network updates (akin to a governing body).
• The Senate: Votes on proposals from the Triumvirate. It comprises the top 12 validators with the most delegated TAO.

While it’s understandable that a young project might need centralized points of control during its development phase, centralization can also pose risks. It may enable a small number of entities to manipulate the rules to their advantage. The Opentensor Foundation does plan to eventually open governance to everyone.

Weight-Copying

Earlier, we discussed the Yuma consensus mechanism, which calculates rewards for all network participants. It relies heavily on the weighting of ratings provided by network validators: the better a miner is rated, the higher their rewards. Validators whose ratings align closely with the average across all validators in a Subnet also receive higher rewards.

However, a critical detail has yet to be addressed: miners are not rewarded immediately after completing their work. A delay exists between task completion and reward calculation, allowing validators time to submit their ratings to Subtensor. During this delay, some validators exploit the system by copying other validators’ ratings and aggregating them to calculate the optimal ratings to submit. This practice, known as weight-copying, enables them to maximize rewards without performing the rating work themselves.

A recent update aims to mitigate this issue. Validators can now submit a hash of their ratings to Subtensor instead of the ratings themselves. This hash is a unique digital fingerprint representing the data but does not reveal it to others. Once the delay ends, validators reveal their ratings, and if they match the previously submitted hash, the ratings are accepted by Subtensor. Otherwise, they are rejected.

While this update should reduce the impact of weight-copying on honest validators, it is still too early to determine its effectiveness fully. Advanced weight-copying techniques, such as analyzing past ratings to predict future ones, may still emerge. This remains an ongoing issue.

The Root Network

As previously discussed, TAO emissions are currently entirely controlled by the top 64 validators, giving them the power to determine the fate of all Subnets. This presents a major problem, comparable to the Proof of Authority model and Bittensor’s governance system. However, unlike those issues, a solution is already in the works: Dynamic TAO.

Dynamic TAO will regulate token emissions through market forces and is expected to launch in the early months of 2025.

---

Ecosystem and Subnets

Thanks to its unique design, Bittensor naturally boasts a large ecosystem composed of dozens of projects building their own Subnets, offering services based on Subnets, or facilitating TAO token delegation. Here, we will focus on a few Subnets and explore how they operate.

This overview aims to showcase the diversity of applications made possible by Bittensor’s technology. In theory, with a well-designed incentive system, any computational task can form the basis of a Subnet. Subnets generally fall into two categories:

• Those offering services with commercial purposes, such as when miners’ results are integrated into third-party applications;
• Those supporting scientific research in non-financial fields.

It’s worth noting that the list of active Subnets is constantly evolving. The Subnets described below may no longer be active if you are reading this article several months after publication.

Subnet 1: Apex

Apex ranks as the third-largest Subnet by TAO emissions (excluding the Root Network). Updated in early 2024, its primary objective is to test text generation models for continuous improvement. Validators submit various tasks to miners, such as information retrieval, summarizing web pages, or answering technical questions. Apex is also integrated with Chattensor, a ChatGPT-like service hosted on the Bittensor platform.

Subnet 4: Targon

Targon is Bittensor’s second-largest Subnet. Miners in Subnet 4 are tasked with analyzing responses from language models and providing sources to verify the accuracy of the answers. Miners are evaluated based on both the precision and speed of their responses, with bonuses awarded to the fastest performers. Targon has been integrated into Sybil.com, creating an AI-powered search engine.

Subnet 6: Infinite Games

Subnet 6 may not be as widely known, but its concept is intriguing. It serves as a prediction market focused on forecasting future events in politics, sports, and technology. It operates similarly to Polymarket, but here miners are the ones making predictions.

Miners are rewarded based on the accuracy of their forecasts, incentivizing them to continually refine their algorithms. These AI models excel at processing large datasets critical for predicting event outcomes. Subnet 6’s results have diverse applications, including politics, insurance, and sports betting.

Subnet 10: Sturdy

Sturdy’s mission is to offer optimal decentralized finance (DeFi) strategies. Miners within this Subnet propose capital allocation strategies designed to maximize yield. The best strategies are then featured on Sturdy’s platform, where anyone can deposit funds to be automatically invested.

Subnet 25: Protein Folding

Before diving into Subnet 25, let’s set the context. Proteins are crucial for our biological processes. They act as tiny keys that activate or inhibit certain functions within our bodies. During their formation, proteins fold into specific shapes that determine their function—an essential insight for developing new medicines.

However, identifying proteins with the right structure is a complex task. It requires selecting an initial sequence of amino acids and using computational tools to predict how the protein will fold. Significant advancements have been made in recent years, such as Google’s AlphaFold software.

Unfortunately, these tools still require immense computational power. Subnet 25 aims to contribute to protein research by delegating computational tasks to miners. These miners run simulations and are rewarded based on how closely they approximate a protein’s final shape. To date, Subnet 25 miners have successfully folded over 400,000 different proteins, demonstrating Bittensor’s potential to support scientific research funding.

Subnet 46: Neural AI

Subnet 46 specializes in transforming simple prompts into 3D digital models. Neural AI’s objective is to simplify the creation of virtual worlds through AI. Miners handle the generation process and are rewarded based on the quality of the models they produce.

Subnet 48: NextPlace

Subnet 48 also focuses on predictions, but exclusively in the real estate market. NextPlace operates as a marketplace where miners collect information on available properties in the U.S., such as location, number of rooms, and living area. They predict the sale prices of these properties and are rewarded based on the accuracy of their forecasts.

Subnet 51: Celium

Celium is a marketplace for renting GPU computational power, similar to platforms like Render or IO. The difference here is that the computational tasks are performed by miners on the Bittensor network.

---

Conclusion

With its innovative architecture based on Subnets, Bittensor promises many and varied applications combining artificial intelligence and decentralization. However, despite its potential, current uses of Subnets can seem repetitive, with incentive systems not always up to scratch. Added to this are major challenges, notably the centralization of transaction validation and governance. Overcoming these challenges will be necessary to guarantee credible decentralization and attract new players.

Bittensor has the potential to become an Amazon of blockchain. It remains to be seen whether it will succeed in fulfilling its ambitions without compromising its goal of decentralization.