The legal framework around distillation is interesting to ponder. Models-as-a-Service companies have the right to impose terms of service. But those same companies have argued that fair use, etc. permit them to use publicly available data for training. As more content everywhere becomes LLM generated, that legal argument would seem to collide with their desired terms. There's also the business angle here. Direct distillation is a very valuable service. Google's market position potentially enables them to offer this more liberally than others.