📄 Bittensor ($TAO)¶
- 👤 Author: [20210456 / Lee Sangbum]
- 📆 Presentation Date: [2025-07-16]
1. Overview¶
- Project Name: Bittensor($TAO)
- Category: Decentralized AI Infrastructure
- Key Technologies / Platforms: Decentralized Machine Learning, Proof of Intelligence (PoI) Consensus, Peer-to-Peer AI Network
- Official Links:
- Website
- Whitepaper
- Docs
- GitHub
- X
- Discord
📌 Summary¶
Bittensor is a decentralized machine learning network, an AI infrastructure based on blockchain technology. It is a system where people contribute their AI models to the network, and are rewarded with TAOs based on their contributions.
2. Background & Problem Statement¶
• Limitations of Centralized AI Approaches¶
Companies such as Google, OpenAI, and Meta collect vast amounts of data through web crawling, partnerships, and their own services. This data is stored in private data centers and is not accessible to the public. Training AI models requires immense computational power, typically provided by server farms equipped with thousands or even tens of thousands of GPUs or TPUs.
As a result, AI development has a very high entry barrier for most individuals and smaller companies. Users have little visibility into how their data is used, and since AI models are trained on datasets curated by these companies, there is a risk of bias in the resulting models.
As AI becomes more powerful and influences our entire lives, if a few companies monopolize it, they can use AI as a power. Monopoly of AI infrastructure can cause social problems such as imbalances in information access, data and privacy issues, threats of censorship and bias, and concentration of power.
• Waste of Power due to Meaningless Computation¶
Traditional Proof of Work (PoW), as used in Bitcoin and early Ethereum, consumes enormous amounts of energy to solve arbitrary cryptographic puzzles that have no real-world value beyond securing the network. This leads to massive environmental costs and wasted computational resources. Proof of Stake (PoS) attempts to address energy inefficiency but shifts power toward wealth concentration, as those with more tokens gain disproportionate control over block validation. Both systems ultimately lack a mechanism for producing useful outputs beyond consensus.
• Bittensor's Approach to AI through Blockchain¶
Bittensor is trying to address centralization and accessibility limitations in AI development. It towards diversity, transparency, and gives many people opportunities to participate to drive the development of AI.
To induce participation, Bittensor introduces Proof of Intelligence (PoI), a consensus mechanism designed to reward participants not for arbitrary work or token holdings, but for contributing valuable AI model outputs. Bittensor's philosophy makes Miner's operation, which consumes power, meaningful in reality.
Bittensor differentiates itself from traditional centralized AI systems by leveraging blockchain technology to decentralize and transparently record AI model operations, performance evaluations, and reward distributions in a trustless and automated manner.
3. How It Works¶
🔍 3.1 Introduction¶
• How to solve the problem?¶
Bittensor decentralizes AI model training and inference by creating an open, blockchain-based network where anyone can contribute models and receive rewards. This structure removes the dependency on centralized corporations and democratizes access to AI technology.
| Category | AI Companies | Bittensor |
|---|---|---|
| Data Collection | Exclusive, Private | Decentralized, Open Participation |
| Computing Resources | Centralized Server Farms | Globally Distributed Individual Hardware |
| AI Model Operation | Closed, Internal Only | Open and Decentralized on the Network |
| Reward Structure | Company-Centric Profit | Contributor-Centric (TAO Token Rewards) |
| Transparency | Low | High (Blockchain-Based) |
| Accessibility | High Entry Barrier | Open to Anyone (with Some Conditions) |
| AI Technology Ownership | Concentrated in a Few Large Corporations | Decentralized Network, Shared Ownership |
• What is the core idea behind the solution?¶
The core idea is a peer-to-peer machine learning network powered by blockchain. Participants run machine learning models as nodes, respond to requests from other users, and get rewarded in TAO tokens based on the quality and usefulness of their outputs.
🏗️ 3.2 Architecture¶
• High-Level Overview of the System Architecture¶
The following roles define the ways to participate in Bittensor:
Miners - Miners work to produce digital commodities by running AI models that generate responses to queries. - Their outputs are evaluated for quality and usefulness by validators. High-performing miners are rewarded in TAO tokens according to their contribution to the network.
Validators - Validators evaluate the quality of miners’ outputs and assign performance scores. - Validators play a key role in maintaining the integrity and fairness of the network by ensuring only valuable contributions are rewarded.
Subnet Creators - Subnet creators design and manage subnet-specific incentive mechanisms. They define the type of work miners should perform and how validators should evaluate that work. - They enable customization within the Bittensor ecosystem by creating specialized AI tasks and reward structures.
Stakers - Stakers are TAO token holders who support specific validators by staking their tokens. By staking, they contribute to validator credibility and may receive a portion of the validator’s rewards. - Staking allows TAO holders to participate in the network’s governance and reward distribution without running a miner or validator node themselves.
🎯 3.3 Core Components¶
Subtensor Blockchain
Subtensor Blockchain serves as the foundational ledger of the Bittensor network. It records miner performance scores, validator assessments, staking data, and reward distributions in an immutable and transparent way. By leveraging blockchain technology, Subtensor ensures that no single party can manipulate contributions or payouts. It functions as both the governance layer and the reward settlement system for all AI-related activities within the Bittensor ecosystem.
Scoring Mediator: Yuma Consensus
Yuma Consensus is the core mechanism within Bittensor that mediates and finalizes the weight scores assigned by validators to miners. While individual validators assess the quality of AI responses from miners, there needs to be a reliable way to aggregate these scores in a trustless and decentralized environment. Yuma Consensus collects scores from all validators, adjusts for inconsistencies, filters out malicious or unreliable inputs, and produces a unified, tamper-proof ranking that is recorded on the Subtensor blockchain. This ensures fairness, prevents Sybil attacks, and maintains the integrity of the reward distribution system across the network.
Subnet Mechanism
The Subnet Mechanism allows Bittensor to support diverse AI tasks by segmenting the network into independent subnets. Each subnet can define its own incentive structure, evaluation logic, and AI modality, such as chatbots, code generation, or financial prediction. This modular architecture enables specialized AI development while maintaining a unified token economy and consensus layer across the entire network.
Wallet System: Coldkey and Hotkey The Wallet System manages cryptographic identities within Bittensor through coldkeys and hotkeys. Coldkeys hold long-term ownership rights and staking records, while hotkeys handle active network participation such as mining and validation. This dual-key structure enhances security by separating long-term assets from operational activities.
🔁 3.4 Workflow Overview¶
[Source from tao.news]
Step 1: A user submits a query to the Bittensor network, requesting an AI-generated response.
Step 2: The query is forwarded to multiple miners (model nodes), which generate and return answers.
Step 3: Validators assess the quality of the miners' responses and assign performance scores.
Step 4: Subtensor blockchain records performance scores, calculates rewards, and distributes TAO tokens to miners and validators.
Step 5: Users receive the AI-generated answers, and the ecosystem updates its weight and reward records transparently.
4. Token Economy¶
• TAO Status¶
TAO's market cap is approximately $3.59 billion. Each token is approximately $380. The total estimated supply is 21 million TAO, and the current circulation is approximately 9.39 million TAO.
Bittensor’s TAO token operates on a continuous issuance model designed to reward network participants, with no fixed supply cap but a controlled inflation rate that decreases over time. Tokens are minted automatically based on the performance evaluations of miners and validators, recorded on the blockchain. While there is no formal burn mechanism, certain network activities like subnet creation may have a partial burn or token lockup effect. TAO allocated to early contributors and investors follows a vesting schedule of 18 to 36 months, whereas rewards earned by miners and validators are immediately liquid without any lock-up period.
• Use of TAO¶
| Stakeholder | How They Use / Earn TAO |
|---|---|
| Miner | Earns TAO by providing high-quality AI model responses |
| Validator | Earns TAO by evaluating miners' responses |
| Subnet Creator | Pays TAO to create and manage a subnet |
| Staker | Stakes TAO to validators and earns a portion of rewards |
Added by Jason¶
Alpha Token¶
TAO is the native token of the Bittensor network, while Alpha tokens are custom tokens created by individual subnets to power their own incentive and utility systems. TAO rewards are allocated to subnets, but each subnet uses its own Alpha token to coordinate internal economics.
| Attribute | Description |
|---|---|
| Definition | Native token of a specific Bittensor Subnet |
| Purpose | Used to reward and coordinate Miner and Validator behavior within the subnet |
| Issuer | Subnet creator (via Smart Subnet template) |
| Token Supply | Customizable (finite or infinite, inflationary or fixed) |
| Utility | Access to services (e.g., inference, APIs), staking, governance, etc. |
| Value Backing | Determined by market demand, not pegged to TAO |
| DEX Tradability | Swappable with TAO via Subtensor’s internal liquidity pool (AMM model) |
| Redemption to TAO | Optional — only possible if subnet explicitly supports redemption (rare) |
| Reward Source | Subnet-defined logic (e.g., inference quality, API usage, reputation) |
| Emission Control | Fully managed by the Subnet’s reward function and minting policy |
| Liquidity Strategy | Provided by subnet owner or community; determines how tradable the token is |
| Analogy | Like an ERC-20 token on Ethereum, with built-in on-chain emission logic |
Example: Subnet 34 BitMind Alpha Token¶
BitMind is a Bittensor subnet focused on deepfake detection and image classification. It incentivizes miners to build models that distinguish between real and fake content, with validators evaluating performance for TAO-based rewards. Think of it as a decentralized, always-on ImageNet challenge with real-time scoring and tokenized incentives.
📘 Bittensor Subnets Examples¶
| Subnet (UID) | Name | Focus Area | Description | Key Highlights |
|---|---|---|---|---|
| SN3 | MyShell TTS | Text-to-Speech (TTS) | Decentralized platform for building and deploying AI-generated human-like voices. | Used by content creators for fast, low-cost audio generation. Try it at BittAudio. |
| SN5 | OpenKaito | Decentralized Search Engine | Community-powered alternative to KaitoAI. Users contribute data sources, indexing logic, and ranking algorithms. | Can integrate with other subnets (e.g., SN13 for Reddit data). Transparent and composable. |
| SN34 | BitMind | Deepfake Detection | Detects and combats deepfakes by incentivizing models that distinguish between real and fake content. | Addresses real-world threat of synthetic media; critical for AI safety. |
| SN30 / SN41 | Bettensor / Sportstensor | Sports Prediction | Crowd-sourced AI-powered prediction markets for sports outcomes. Bettensor and Sportstensor set odds via collective forecasting. | Combines AI with human wisdom for fairer and more accurate betting systems. |
| SN16 | BitAds | Performance-based Advertising | Shifts from click-based ads to sales-based rewards. Marketers (miners) are paid only when actual sales occur. | Actively running campaigns; integrates with platforms like Shopify; evolving via real-world testing. |
🧠 Comparison Table: Bittensor vs Ethereum¶
| Concept | Ethereum | Bittensor |
|---|---|---|
| Base Chain | Ethereum Mainnet | Subtensor (Bittensor L1) |
| Native Token | ETH |
TAO |
| Block Reward Recipients | Validators (via Proof-of-Stake) | Subnets (via TAO stakers × weight) |
| Participation Mechanism | Stake ETH to validate blocks | Stake TAO + contribute as a model or evaluator |
| Reward Basis | Proportional to staked ETH | Proportional to stake × weight |
| Independent Tokens | ERC-20 tokens | Alpha tokens (Subnet-native tokens) |
| Smart Contract Deployment | By developers on EVM | By Subnet creators using the Smart Subnet Template |
| Utility Ecosystem | DApps (e.g., Uniswap, Aave) | AI services (e.g., model inference, evaluation, APIs) |
5. Project Status & Plan¶
Project Status:¶
Bittensor is live on mainnet. There are more than 118 activated subnets. It has an active ecosystem of miners, validators, and subnet creators operating globally.
Community & Developer Base:¶
There is an active developer community with consistent discussions on Discord, GitHub contributions, and subnet experimentation. More than 60 official repositories on GitHub are actively managed, and the number of weekly commitments by developers is maintained continuously.
Partnerships & Funding:¶
Bittensor is self-bootstrapped but has gained organic traction. No major corporate partnerships have been formally announced as of mid-2025, but community-driven subnet projects are growing.
Open Source:¶
Yes. Bittensor’s core components, including Subtensor blockchain code, are available on GitHub. Also, some subnets publish validator scoring logic or subnet management tools as open-source projects. This allows participants to transparently review how the subnet operates and offers opportunities to contribute or customize the system.
Token Listing & Market Activity:¶
TAO is listed on several decentralized and centralized exchanges, such as Binance, Coinbase, Kraken, and Uniswap. TAO has increased in value per unit by about 47% over the year.
Bittensor is not a simple concept or event project, but a working ecosystem.¶
6. User Experience & Hands-on Review¶
Participating in TAO Mining on Testnet¶
Attention : This project does not encourage readers to participate in mining or invest in TAO. This was done for the purpose of helping understand the Bittensor system. You are responsible for all investments and decisions.
The purpose of this experiment is to understand the decentralized AI network structure throughout the reward mechanism. It directly observes how miners provide AI models, validators evaluate the quality of responses, and token rewards are generated as a result.
Experiment Setup Summary
1️⃣ GPU Specification • NVIDIA RTX 3060 Ti (GDDR6 8GB VRAM)
2️⃣ LLM Used • Mistral-7B-Instruct-v0.1 (4-bit quantized)
3️⃣ System Environment
• OS: Ubuntu 24.04 LTS (WLS environment)
• Main Libraries & Tools: Python 3, CUDA Toolkit, Bittensor 4️⃣ Participating Testnet UID
• 171(correspond to mainnet UID 45 / code generation, bug fixing)
Mainnet UID 45 corresponds to Testnet UID 171. You can ask questions to utilize the service provided by Bittensor Mainnet 45. Visit gen45.ai.
This response was generated using the AI model of the miners of Mainnet UID 45. Let's participate in creating these responses with local LLM.
Step 0. Install CUDA Driver¶
Install the appropriate CUDA driver for your OS. In this experiment, WLS2 based on Windows 11 was used.
Check if CUDA driver is successfully installed
nvcc --version
nvidia-smi
Step 1. Prepare a Bittensor SDK¶
• Create Virtual Environment
sudo apt update && sudo apt install python3-venv -y
python3 -m venv ~/taominervenv
source ~/taominervenv/bin/activate
• Clone Bittensor
git clone https://github.com/opentensor/bittensor.git
cd bittensor
pip install -e .
• Create Bittensor Wallet
btcli wallet new --name tao1
Determine your wallet name and hotkey.
The number of words can be set freely from 12 to 24.
These words are called Mnemonic Phrase. User can recover all the keys including coldkey and hotkey by using phrases. Keep these safe from exposure to others.
The directory where your wallet is stored is
~/.bittensor/wallets/
• Check Wallet Coldkey Address and The Balance
btcli wallet overview --wallet.name tao1
Step 2. Test Your LLM in Local(Optional)¶
Ask a locally installed LLM to make sure it provides the correct answer. This process is necessary when tuning the model or prompting it to satisfy the answer conditions required by subnet.
Install essential packages
pip install --upgrade pip
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
pip install transformers accelerate bitsandbytes sentencepiece
pip install bittensor
Save the script to the directory as below
~/mistral_test
├── miner.py
└── mistral_runner.py
mistral_runner.py ```python= from transformers import AutoConfig, AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig import torch
Model identifier to load¶
model_id = "mistralai/Mistral-7B-Instruct-v0.1"
Configure 4-bit NF4 quantization¶
quant_config = BitsAndBytesConfig( load_in_4bit=True, # Enable 4-bit loading bnb_4bit_compute_dtype = torch.float16, # Compute in float16 bnb_4bit_use_double_quant = True, # Use double quantization bnb_4bit_quant_type = "nf4" # Use NF4 quantization type )
Load custom Mistral config, trusting remote code¶
config = AutoConfig.from_pretrained( model_id, trust_remote_code=True )
Load a slow (SentencePiece) tokenizer, trusting remote code¶
tokenizer = AutoTokenizer.from_pretrained( model_id, use_fast=False, trust_remote_code=True )
Load the model with quantization and custom config¶
model = AutoModelForCausalLM.from_pretrained( model_id, config=config, device_map="auto", # Automatically map layers to devices quantization_config=quant_config, torch_dtype=torch.float16, trust_remote_code=True )
Encodes the given prompt, generates an output response from the model¶
def generate_response(prompt: str) -> str: # 1) Tokenize and send inputs to the model's device inputs = tokenizer(prompt, return_tensors="pt").to(model.device) # 2) Generate up to 128 new tokens outputs = model.generate( **inputs, max_new_tokens=128, pad_token_id=tokenizer.eos_token_id ) # 3) Decode generated tokens into text return tokenizer.decode(outputs[0], skip_special_tokens=True)
miner. py
```python=
from mistral_runner import generate_response
def main():
print("Mistral Prompt Test ")
print("Input('exit' to quit): \n")
while True:
prompt = input(">>> ")
if prompt.lower() in ["exit"]:
print("Test Ended")
break
response = generate_response(prompt)
print(f"\n{response}\n")
if __name__ == "__main__":
main()
Execute Mistral
python miner.py
Step 3. Scripting for AI Model Application and Execution¶
• Clone SWE-Rizzo
git clone https://github.com/brokespace/code.git
cd code
• Install Requried Packages
pip install -e .
Save the following syntax in ~/code/.env
PROVISIONING_API_KEY = <your-openrouter-provisioning-key>
miner. py(original) ```python= async def forward_swe(self, synapse: LogicSynapse) -> LogicSynapse: return miner_process_swe(self, synaps)
Change forward_swe function as below.
miner. py(revised)
```python=
async def forward_swe(self, synapse: LogicSynapse) -> LogicSynapse:
prompt = synapse.messages[0]
max_toks = getattr(self.config.miner, 'max_tokens', 128)
synapse.completion = generate_response(prompt, max_new_tokens=max_toks)
return synapse
Save the script to the directory as below
~/brokespace
├── neurons
│ ├── miner.py
│ └── mistral_runner.py
Step 4. Start Mining on Testnet¶
Before mining, you need to stake a certain amount of TAO in the chain. In Testnet, you have to pay a test TAO that has no real value, which can be obtained by requesting it on the Bittensor official Discord channel.
• Register wallet hotkey into testnet netuid 171
btcli subnet register --wallet.name tao1 --wallet.hotkey miner1 --subtensor.network test --netuid 171
• Stake a certain amount of tTAO in the chain
btcli stake add --wallet.name tao1 --wallet.hotkey miner1 --amount 0.0050 --subtensor.network test --netuid 171
• Start mining on testnet 171
cd ~/code
python neurons/miner.py \
--netuid 171 \
--subtensor.network test \
--wallet.name tao1 \
--wallet.hotkey miner1 \
--neuron.device cuda \
--axon.port 8091 \
--miner.max_tokens 128 \
--logging.debug \
--miner.name mistral4b
When essential processes are completed, the "Miner running..." loop is repeated and mining begins.
Why?¶
• Low Activity in Subnets(Usually on Testnet)
Unlike Mainnet, Testnet has very few residents and does not have active query requests. Even if validators are present, without user queries being sent to the subnet, miners do not receive prompts to generate responses. It is important to select a subnet with active request traffic.
• Subnet-Specific Response Format Requirements
Each subnet defines its own rules regarding response format and structure. Miners must adhere strictly to these specifications; otherwise, their responses may be ignored or scored as invalid. This may include JSON formatting, token count limits, or specific instruction-response pairs. Careful review of subnet documentation and alignment of model outputs is necessary to avoid zero scores.
• Validator Weight Synchronization Delay Even with traffic and correct responses, there can be a delay before validator-assigned weights reflect on the blockchain. This is a normal part of Bittensor’s consensus cycle. Waiting through several update intervals may be necessary before miners see their rank change.
7. Why Blockchain¶
• Why does this problem specifically require blockchain to be solved well?
Bittensor aims to create an opened, fair, and incentive-driven marketplace for AI models. This requires a transparent and immutable system for recording model performance, distributing rewards, and managing ownership. Blockchain provides the necessary trust layer without relying on a central authority.
• What does blockchain enable that traditional solutions cannot?
Traditional AI platforms rely on centralized servers and opaque evaluation processes controlled by single organizations. Blockchain enables decentralized verification of model quality, transparent reward distribution using native tokens, and tamper-proof record-keeping, ensuring that no single party can manipulate or monopolize the ecosystem.
• Is decentralization critical in this use case? Why or why not?
Without decentralization, AI model development and rewards would remain under the control of a few large corporations. Bittensor’s blockchain ensures that anyone, from individual developers to small organizations, can contribute, be evaluated fairly, and receive proportional rewards.
8. Insights & Limitations¶
✅ Key Takeaways¶
Bittensor has presented a new concept that allows anyone to participate in AI development, breaking away from the model developed and owned by giant companies.
⚠ Limitations / Open Questions¶
• Emerging contenders or expanding gap?
Bittensor’s decentralized AI ecosystem offers an open marketplace where anyone can contribute AI models and earn rewards. However, those with abundant capital and technology stil have advantages. Some people point out that it does not completely reverse the reality that huge capital dominates the AI industry.
• Is it really transparent? Although Bittensor aims for a transparent AI infrastructure, it is not transparent unless miner opens which AI model he or she used to make the answers and how it was made. Validator only evaluates the response quality of Miner. The AI model architecture, parameters, and datasets used by Miner are not recorded on the blockchain. Therefore, biases in AI models and data ethics problems may still exist.
When decentralized AI networks give ethically inappropriate answers, the responsibility for them is unclear. There is also a lack of clear interpretation of legal issues such as intellectual property rights and user privacy.
• Difficult to participate in the system with Miner and Validator. Typical Query Senders may need to pay TAO tokens on some subnets, but in most cases, they can receive AI responses without the need for cost, technical knowledge, or high-spec hardware. However, running a miner or validator requires technical knowledge and significant GPU resources, still limiting broader participation.
9. Reflections & Discussion¶
💡 Personal Reflections¶
• Replacing The Operation to Obtain Rewards with Meaningful Activities in the Real World
Bittensor shows that blockchain mining doesn't have to be about wasting energy on meaningless tasks. By tying mining rewards to useful AI tasks, it replaces traditional proof-of-work with real-world value creation. This approach could inspire future blockchain systems to focus on socially beneficial computation.
• How has your understanding of AI/blockchain changed after studying this? Previously, blockchain and cryptocurrency were thought of as tools for decentralization from an economic point of view. While studying Bitensor, I learned that blockchain can also serve as an infrastructure for a distributed AI ecosystem.
❓ Discussion Questions¶
- Can this project realistically compete with large corporate AI platforms?
- What are potential risks of decentralizing AI model validation and training using blockchain?
- What distinguishes this project from other decentralized AI infrastructure initiatives? What are its unique strengths or advantages?
10. Insight from others¶
After each presentation in class, we will form small groups for each case for discussion. At that time, please discuss with your group the questions posed in Section 9, and write any key points or insights from your discussion group here.
11. References¶
- [Bittensor Introduction] (https://bittensor.com/intro)
- [Bittensor Docs] (https://docs.learnbittensor.org/)
- [TAO Stauts] (https://taostats.io/)
- [Github opentensor/bittensor] (https://github.com/opentensor/bittensor)
- [Github opentensor/text prompting] (https://github.com/opentensor/text-prompting)
- [Subnet 45 - Rizzo Network] (https://rizzo.network/subnet-45/)