Original author: @dvzhangtz
Original source: X
Today, Binance launched @TheBittensorHub (TAO). What did this project do to make it rank in the top 30 of all tokens before it even hits the big stage?
In a word, this project does not directly contribute data or computing power. Instead, it uses blockchain networks and incentive mechanisms to schedule and screen different algorithms, thereby forming an algorithm (model) market with free competition and knowledge sharing in the AI field.
Summary
The three elements of the AI era are algorithms (that is, models), computing power, and data. We have relatively clear solutions on how to combine computing power and data with blockchain for distribution, but how to decentralize algorithms and how to concentrate everyone's efforts on co-building algorithms and models has always been a question.
Bittensor solves this problem.
Bittensor divides AI into different tasks, and each task starts a subnet. Each miner provides his own model for the task he is good at, and the network rewards the miner who does the best in each task. And through its unique Yuma consensus algorithm, it ensures that all participants do not do evil.
In terms of token economics, its token TAO is a pure mining coin. There is no VC round, private placement round, ICO/IEO/IDO, or foundation reservation. The only way to obtain tokens is mining and buying them on the market. As a mining coin, miners will have a strong motivation to pull the price in the early stage. Now 89% of the tokens in the entire network are staked and locked.
Bittensor provides direct ToC products, chatbots, and APIs for developers, providing a good user experience.
Background Part 1
Problems in the AI Market
Current issues:
Today’s AI models are overly dependent on benchmarks and lack diverse evaluation methods, resulting in a winner-takes-all model.
At present, each model is relatively independent
Bittensor believes that this is not conducive to the progress of AI as a whole and the efficiency of algorithm innovation. Isolated models and AI services that only select winners mean that once someone wants to develop a new model, they may have to start from scratch; assuming that model A is proficient in Spanish and model B is proficient in writing code, when a user needs AI to explain the code with Spanish annotations, it is obvious that the two algorithms work together to output the best effect, but this is not possible in the current environment; in addition, since third-party application integration requires the permission of the AI model owner, limited functions also mean limited value, and the combined force of the entire AI field has not actually been released. Therefore, the big goal of the Bittensor project is to enable different AI algorithms and models to collaborate, learn and combine with each other, so as to form a more powerful model to better serve developers and users.
This has several benefits:
Build a collaborative network of different models, use this network to judge the quality of models provided by different people, and let them collaborate
For users who use AI, there are no middlemen to make a profit, and they can get AI services at a lower price.
Many models are developed by small teams themselves, so that small companies that are not giants can also enjoy such benefits and have ownership of AI
Background Part2
Thoughts from BTC
BTC provides a set of accounting services. This system itself does not provide computing power. It only defines the way in which computing power is verified, that is, its POW mechanism.
Under the incentive of the POW mechanism, miners around the world try to optimize their mining machines, find lower electricity costs, and do better mining (providing BTC accounting services).
However, we know that this POW is doing a meaningless calculation, but if it can allow miners around the world to continue to actively find ways to reduce the cost of mining TAO in this market, optimize the algorithm, improve the accuracy, and reduce electricity costs, so as to provide better model services, it would be great to think about it.
technology
Part 1 Roles and Functions in the Network
Miners: They can be understood as providers of various AI algorithms and models around the world. They host AI models and provide them to the Bittensor network. Different types of models form different subnets, such as models specializing in images or sounds.
Validator: An evaluator within the Bittensor network. Evaluates the quality and effectiveness of AI models, ranks AI models based on their performance on specific tasks, and helps consumers find the best solution.
Nominator: Delegating tokens to a specific validator to show support, or you can delegate to a different validator. It's a bit like in DeFi, you stake your tokens to Lido to get returns. User: The end user of the AI model provided by Bittensor. It can be an individual or a developer seeking to use AI models for applications. (6/19)
Technology Part 2
Connections between characters
Users need better AI models;
Miners provide their own AI models;
Validators are responsible for screening better AI models according to different uses;
Nominators choose to support different validators.
To put it simply, it is an open AI supply and demand chain: some people provide different models, some people evaluate different models, and some people use the results provided by the best model.
Technology Part 3
How to ensure that miners and validators do not do evil: Yuma Consensus
In an open network, without centralized supervision, service providers are naturally willing to do evil in order to "reduce costs and increase efficiency". Possible evils include:
Miners do not run the model service properly, and the result is random numbers
Miners only connect to validators who collude with them. They are both players and referees.
Honest nodes may not be fully connected due to network reasons, while a large number of dishonest nodes communicate with each other to achieve collusion
Bittensor provides a solution to this problem, which I think is the biggest innovation of the project. The following paragraph is the key and difficult part of this article.
Technology Part 4
In order to explain Bittensor's solution, we need to first talk about some basic knowledge in machine learning: training, loss function, gradient descent
If you want to teach a puppy how to pee on a mat (classic case, no bad directions) - (If you want to teach a machine a piece of knowledge)
Method 1: If the dog urinates on the mat, reward him with a piece of meat, if not, spank him
Method 2: If the dog urinates on the mat, reward it with a piece of meat, if not, spank it; and the farther away from the mat, the harder the spanking (calculate the loss function)
Method 3: Every time the dog takes a step, a judgment is made:
If it walks towards the mat, reward it with a piece of meat, if it does not walk towards the mat, spank it
(Each time training is performed, the loss function is calculated once)
Method 4: Make a judgment every time the dog takes a step:
If it walks towards the mat, reward it with a piece of meat, if it does not walk towards the mat, spank it;
And put a piece of meat in the direction of the mat to attract the dog to walk towards the mat.
(Each time training is performed, the loss function is calculated, and then the direction that can best reduce the loss function is carried out for gradient descent)
Technology Part 5
OK, now that we have a general understanding of these concepts, let’s go back to Bittensor.
Bittensor divides AI into different tasks, and each task starts a subnet. When a subnet is created, the creator will define what tasks the subnet needs to complete (translation, question answering, image generation, etc.), and define a measurement indicator (loss function, BLEU, etc., the loss function is used as an example below)
We have some miners and some validators who participate in a Bittensor subnet to provide some type of AI service.
How to measure a miner's generation results
Credibility score: judge whether each miner is trustworthy
Some miners may not be connected to some validators due to poor network conditions (the connection relationship between validators and miners is represented as C matrix in the white paper)
This value is weighted by the stake of each validator (S in the white paper) to get a credibility score (the more validators with high stakes a miner connects to, the higher this value is, which is T in the white paper)
Effect rating: judge the quality of each miner's output
For each miner’s result, do the following steps:
Each validator connected to the miner uses its own model to calculate the loss function for the miner's output and further derives the score (W in the white paper)
Different validators have different importance. The stake amount of each validator (S in the white paper) is used as the weight and multiplied by the result of the first step to obtain a value. This is the score of all validators for this result (R in the white paper)
Final score = effect score * credibility score (I in the white paper)
Directly distribute tokens based on the final score of each miner (St+1 in the white paper)
How to judge a validator
Set a loss function and continuously optimize the network. The optimization goal of this function is to make more high-scoring nodes in the network connected to each other.
Token Economics Part 1 https://taostats.io/tokenomics/
What does the TAO token incentivize?
For validators: The more accurate and consistent the screening and evaluation of AI models, the more rewards you will get. Obviously, to become a validator, you need to stake a certain amount of TAO tokens. For miners: Respond to user needs and provide your own models, and receive TAO tokens based on your contributions. For nominators: Delegate your own TAO to validators, similar to liquidity staking rewards. For users: Pay TAO tokens to start tasks, which is equivalent to consumption (11/19)
Token Economics Part 2
Release Mechanism
TAO has a supply of 21,000,000 (a nod to BTC) and also has a 4-year halving cycle, with the reward for each block halved every 10.5 million blocks. A total of 64 halving events will occur, with the most recent halving cycle occurring in August 25.
Currently, one TAO is sent to the network every 12 seconds. Roughly speaking, 7,200 TAOs are produced per day, with miners and validators taking half each. There are no common routines such as VC rounds, private rounds, ICO/IEO/IDO, foundation reservations, etc. It can be understood as a pure mining coin.
Token Economics Part 3
Let's analyze in detail how the release is allocated to each subnet:
The TAO released every 12 seconds will be distributed to 32 subnets according to a certain percentage (determined by the root subnet, which is subnet 0, based on the performance of each subnet). Then it will be distributed to each miner and validator in each subnet.
The validators of subnet 0 are the 64 validators with the largest stake in the entire network. If you want to know how they will distribute tokens, you can run the installation tutorial in the last part of this document and check the code. metagraph.weights.shape is a torch.Size([64, 33]), which contains the weights of the 64 validators for each subnet.
How to operate Part 1
Want to use its AI services
Directly through its chatbot: https://app.corcel.io/chat
Call the API directly (API key required): https://api.corcel.io/docs
The API key is here: https://app.corcel.io/dashboard/api-key…
The github of this calling platform: https://github.com/corcel-api
How to operate Part 2
View transaction history
Use a url in a similar format, followed by the address
https://x.taostats.io/account/5HKtJCP7tMwncxNBH4EfjwKnSZmey2G3PdCNKYSMyp6xFY1d…
https://x.taostats.io/account/5F4tQyWrhfGVcNhoqeiNsR6KjD4wMZ2kfhLj4oHYuyHbZAc3… (foundation address)
Operation method Part 3
As a nominator, you want to get staking rewards
Wallet - Download extention - Create new wallet - Create - Confirm - Generate - Retype - Password - Save address
Withdraw to address - click stake - you will need a validator address
Operation method Part 4
The current staking income is https://taostats.io/staking/, and the income is about 20% of apy. Nearly 90% (5.3M / 5.9M) of the tokens in the entire network are being staked and can be unstaked at any time.
Operation method Part 5 Go to the validator page (https://taostats.io/verified-validators/…), copy an address, click stake in the wallet, enter the password, and it will succeed.
reference
Official website for display properties: https://bittensor.com
Documentation (introduction logic): https://docs.bittensor.com
If you want to deeply understand its logic, you need to use two white papers
https://bittensor.com/whitepaper
https://bittensor.com/about
Official data website (on-chain data, pledge data, etc.): https://taostats.io
Application interface (ToC application): https://app.corcel.io/chat
Github (a lot of development information to read): https://github.com/opentensor
This article also refers to a large number of research reports, and I would like to thank them here!
Tao to the Moon！
