This is part of our series on a year of Bittensor experience, leading up to our anniversary at the 13th of July. After discussing the Dirac landscape and at its opposite, the flat incentive landscape, we now look at what happens if you take the bad features of each of those and try to combine them. Please let us know what you think in our Discord channel!

Note that some subnets pose as winner-takes-all – and therefore look Dirac-ish – but effectively have a skewed or flat landscape: if the 100% winner changes per epoch, and is selected based on a noisy scoring metric, miners have every incentive to hoard UIDs to increase their chances. This invites the monopolist behavior as described in the flat incentive landscape.

What makes this combination especially terrifying, is that miner deregistration becomes a feature of the subnet game, as all but one miner get zero weight per epoch. It makes (de)registration and price forming a central part of the subnet, by adding a layer of complexity to the subnet with no objective benefit.

One way the scoring metric can be needlessly noisy, is when it depends on inter-miner effects. These are often overlooked – see the calculative miner for an example. Another example is a case where miners were required to submit unique solutions, that is, unique compared to other submissions – while they could not see the solutions of other miners before the submission deadline. Considering that generating possible winning solutions in that subnet was not that difficult, it is not hard to imagine how that played out: the miner controlling the largest amount of UIDs had the highest chance to win, and used its winnings to hoard ever more UIDs, keeping the registration fee high as well.

This example underscores that it is important that the challenge given to miners is hard. When the challenge becomes too easy, because miners have optimized their approaches beyond expectations of subnet owners, this cannot be solved realistically in the scoring metric or incentive landscape alone. The challenge must become challenging again to fix the subnet. This too might not be trivial, especially in cases where the underlying problem is actually fully solved by the miners. The risk of such an evolution of a subnet, is that miners will eventually be solving artificial challenges, that have ever less to do with the stated goal of the subnet or any real problem.


0 Comments

Leave a Reply

Avatar placeholder

Your email address will not be published. Required fields are marked *