Weight-copying is a flaw in Bittensor’s Yuma Consensus that allow subnet validators to exploit how TAO emissions get distributed.
For those unfamiliar with the issue, here is the TLDR:
In Bittensor, subnet validators need to give weights to miners according to the quality of their work. Validators then get paid according to how close their weights are from the average of weights. What dishonest validators do is to wait for others to publish their weights, compute the average, and publish that as their own weights. It allows them to get a perfect score without doing any of the intended work. They are weight-copying: copying the weights of others. More details on Learn Bittensor here.
In this article, I will look at WC (short for weight-copying) and its impact on Bittensor from different angles: the good, the bad, and the ugly.
The biggest risks are the unknowns. They are the black swans that destroy economies and protocols. The risks humans see coming are usually not that catastrophic, and often pass faster than we thought.
Fortunately for everyone in the Bittensor ecosystem, weight-copying is a well understood and widely recognized flaw. A lot of people are thinking about it, and many solutions are already being worked on.
Commit-reveal 3.0 is about to get released, and liquid alpha is another tool to help fight weight-copying. Dynamic TAO will change the validation game, and very likely have an impact on WC as well.
In short: don’t fret, Bittensor is still in its early days and weight-copying will not be the end of it.
Weight-copying is a serious problem in Bittensor.
First, it reduces the TAO incentives earned by validators that are actually doing subnet validation. Instead, these incentives go to validators that are not contributing to the network. This decreases the profitability of working validators, and lowers the incentives to actually improve the network by validating miners’ outputs.
Second, and maybe even more importantly, it undermines the evaluation work done by working validators. The goal of validators is to rate (weight) miners’ work. If working validators are always rated lower than weight-copiers, the weights actually given by the working validators can be skewed much more easily. If one working validator has different weights than others, it will skew the consensus much more if there are weight-copiers, than if everyone is working to set their own weights.
In other words, in addition to disincentivizing working validators, weight-copiers also weaken the consensus found by validators within the subnets.
The beauty of open and decentralized networks that actually work, is that there is no room for opinions or politics. In Bitcoin, a transaction is either valid or invalid. It isn’t honest or dishonest, good or bad.
When asked how to build resistant networks, Snowden answered by mentioning the “adversarial approach” of the Bitcoin Whitepaper (see min 16). The whole point of these open networks is not having to hope that people will not behave malicious ways or coerce them not to, but making it impossible and unprofitable for them to do so.
If, like Bitcoin, we want Bittensor to stand the test of time, it needs to be built with an adversarial mindset. It is the only thing that makes malicious behaviors technically and economically unsustainable in the long-run.
Bittensor is all about incentives. And right now, the incentives are flawed towards weight-copying. But introducing a cancel culture where some people have the say about what’s right or wrong is not how we can strengthen Bittensor and make it thrive.
Some compare Bittensor to Bitcoin, and I think they are right. But if we want to stand true to that comparison, the network need to stay neutral and resistant to bad actors.