New subnet design lets miners compete on training data instead of weights

const_reborn · x · 2026-07-24

A new subnet design makes miners compete on training data, not model weights

The referenced Cascade #sn91 approach changes the usual training-subnet format: miners do not submit model weights.
Instead, every miner trains the same 4M-parameter model from scratch, with the same seeds and the same compute budget.
The only thing miners provide is a data generator, so the competition becomes about the quality of the data rather than the learned weights.
The post frames this as a novel mechanism-design idea for how decentralized training could be organized.

More from coding & agent