This is part of our series on a year of Bittensor experience, leading up to our anniversary at the 13th of July. After discussing miner archetypes we now discuss incentive landscapes, starting with the Dirac type. Please let us know what you think in our Discord channel!
Subnets with a winner-takes-all mechanism reward one top miner, and nobody else. Although the concept seems a good fit for a competitive environment, such subnets have specific issues. As a second place earns a miner nothing, it takes a courageous miner to try and compete with a good number one. The runner up invests time and money to develop code, possibly train models, rent compute, without the guarantee of beating the number one. Money spent may run into the $10k-100k range in practice, in line with the possible reward; it’s all-or-nothing.
Possibly, the number one has a few improvements prepared, that it never had to deploy before. It will decide to use them just after the runner up thinks it can take the top position. If the top miner manages to do this trick a few times, the runner up will eventually stop trying, and the top miner can stay on top, even without having to put all their cards on the table.
This is not good for the subnet: competition is not optimal, and, yet again, the top miner pursues his own interest, that is at odds with the stated goal of the subnet. As we once put it in a Novelty Search presentation, a subnet is only as good as its second best miner. You need to reward a number two.
The next article discusses the flat incentive landscape.
0 Comments