Right, so sort of like https://github.com/bigscience-workshop/petals but for the training phase. I suppose different training runs could be proposed via a RFC type of procedure. Then it’s not only the open source model maintainers that put the effort, but also supporters of the project can “donate” their hardware resources.