Does this imply that some type of decentralized training mechanism is possible? Like an accumulation for ML models. I suspect in the limit You will just have even more massive models which will require even more demand on the hardware. I also wonder if new capabilities emerge from the merging that are not present in any one model.