[goldenshale]

This is a great idea. Being able to dynamically scale up model sizes as datasets and use cases expand without needing to retrain from scratch could enable a Cambrian explosion of interesting stuff building on top of a Llama type model trained in this way.