It has to be the case that OpenAI, Anthropic, and Google are doing internally what they accuse others of doing: using expensive advanced versions of their models to train lighter-weight more efficient models. Have they explicitly acknowledged that Astra was used to train Sol?