This is part of our series on a year of Bittensor experience, leading up to our anniversary at the 13th of July. We discuss 9 miner archetypes before digging deeper into incentive landscapes. After discussing the honest and the rogue miner, the calculative miner is considered. Please let us know what you think in our Discord channel!
Sometimes the incentive calculation, that takes miner submission scores and eventually turns them into weights, is so convoluted, that seemingly strange strategies help mediocre submissions win.
For example, when the absolute performance is not taken, but “wins” of one miner over a group of other miners, the calculative miner may decide to post competing submissions designed to lose hard to its primary submission, boosting its score up to the top. This was a very effective strategy in some LLM training subnets.
Needless to say, this benefits the calculative miner, while harming the subnet in multiple ways. It makes a suboptimal result rise to the top position, and disincentives honest miners that are trying to pursue the stated goal of the subnet. Once the incentive mechanism contains strong inter-miner effects, one can expect issues like this to arise.
Just as in the discussion of the rogue miner, the lesson to be taken from this is that the validator code should contain very clear log messages, that the logs should be easy to find and analyze, that data is logged or collected in a way that allows structured analysis, for later scrutiny. The miners are often the ones debugging the validator, if not for their own immediate gain, then for lowering the gains of competing miners, by fixing exploitable bugs.
The next article will discuss the blockchain-savvy miner.
0 Comments