This is part of our series on a year of Bittensor experience, leading up to our first anniversary at the 13th of July. After describing honest, rogue, calculative, blockchain-savvy, overfitting, devops, (over)analyzing, clueless, slow and lazy miners, we turn to the (un)lucky miner. One more miner type to go before we discuss incentive landscapes. Please let us know what you think in our Discord channel!
You don’t choose to be a lucky miner, you just sometimes happen to be one. A miner can get lucky by being in a top position due to failures of others, or by staying in a top position due to external events. In similar ways you can be unlucky.
A well known group of such winner/losers are the miners who were earning incentive during the firewalling of the blockchain during the July 2024 hack. (Sidenote: we were explicitly unlucky miners in that respect, having just missed one crucial round of model evaluation before all weights were effectively frozen. There also was good luck turning bad, as the chain reverted to stake-weighted emission after 5000 blocks, while miners still had to pay for their infra.) Another interesting example is how incentive is given to UIDs – not hotkeys – and how a UID can get rewarded after its hotkey is deregistered, benefiting the newly registered hotkey. This is exacerbated by the implementation of “commit reveal” which makes it so that a result posted in some epoch, is rewarded a few epochs later.
Although this may sound like “cost of doing business,” there may be strategies here that harm subnets. Especially in subnets with low registration fees, and winner-takes-all incentive logic, lazy miners may be tempted to keep registering UIDs just to catch the occasional random reward.
One way to combat this in the incentive structure, is to decay weights, rather than setting them to zero. A once-winning miner is then protected for a certain time, because his weight will not reach zero immediately, or not at all. The tiniest weight of 1/65535 will do. With such protection, only new miners that have never won before are susceptible to this attack, and new miners that only register to get lucky, will never succeed.
After the next post, about the idle miner, we will discuss incentive landscapes in light of all miner behaviors we have identified.
0 Comments