What subnet?

A subnet is a competition market. Anyone can create a subnet, or participate in an existing subnet. You create a subnet by paying the registration cost (in TAO) and you will receive a subnet netuid. There are two ways you can participate in an existing subnet: Either as a subnet miner or as a subnet validator. You bring a computer that has sufficient computing resources, register that computer, along with your wallet, into a subnet. On this computer you then run either the subnet miner module or the subnet validator module (Python code) provided by the subnet owner.

Incentive Mechanism

In a Bittensor subnet:

• The task-performing entities are called subnet miners.

• Entities that create the tasks, score the output of the subnet miners and reward them, are called subnet validators.

Incentive mechanisms are a fundamental part of Bittensor. They drive the behaviour of the subnet miners and govern the consensus amongst the subnet validators. Each subnet has its own incentive mechanism. Subnet developers should design incentive mechanisms carefully so that these mechanisms comprehensively and explicitly promote all the desired behaviors and penalize undesired behaviors.

Subnet owner responsibilities​

A subnet owner is responsible for:

• Defining the specific digital task to be performed by the subnet miners.

• Implementing an incentive mechanism that aligns miners with the desired task outcomes.

Design with Yuma Consensus as a black box​

Though a subnet incentive mechanism works in conjunction with the Yuma Consensus in the Bittensor network, you must design your subnet incentive mechanism by treating Yuma Consensus as a black box.

Make it easy for participation​

To attract high-performing subnet miners and subnet validators, make sure that you publish sufficient documentation on your subnet.

Components of incentive mechanism​

A subnet incentive mechanism must contain the definition and implementation of the following behaviors:

Subnet protocol​

See 1 and 3 in the above diagram. A subnet protocol, which is unique to the subnet, must define how a subnet validator will query the subnet miners, and how a subnet miner should respond to the query.

For example, a subnet validator might send a query containing the task description to the subnet miners. The subnet miners will perform the task and then respond to the subnet validators with the results of the task the miners performed. Note, however, that query-response is only one of the ways of subnet miner-and-subnet validator interaction. An alternative example is when the subnet validators and subnet miners use additional shared resources such as databases, and these resources can be used to evaluate miner performance.

Subnet task​

See 2 in the above diagram. The task is one of the key components of any incentive mechanism as it defines what miners will perform as work. The task should be chosen so that miners are maximally effective at the intended use case for the subnet. In other words, the task should mimic an intended user interaction with a subnet. Examples of tasks are responding to natural language prompts and storing encrypted files.

Subnet reward model​

See 4 and 5 in the above diagram. Just as the task describes what miners should do, the reward model dictates how it should be done. Similarly, just as tasks should mimic user interactions, reward models should mimic user preferences or desired outcomes.

As with any machine learning model, a subnet has an objective function that it is continuously optimizing. The reward model defines the quality of all miner behvaiour in the subnet (both intended and unintended).

Operationally, it is the mathematical object that converts miner responses into numerical scores. A reward model can in fact contain as many different reward mechanisms as are necesary to align miners with the intended task.

Miners will continuously compete to achieve the highest reward possible. If the reward is capped at an upper limit, miners may not be motivated to improve further. Hence care should be taken to enable continuous improvement of the miner, rather than stagnation.

Discourage exploits​

The incentive mechanism is ultimately the judge of subnet miner performance. When the incentive mechanism is well calibrated, it can result in a virtuous cycle in which the subnet miners continuously improve at the desired task due to competition.

On the contrary, a poorly designed incentive mechanism can result in exploits and shortcuts, which can detrimentally impact the overall quality of the subnet and discourage fair miners.