You cannot really hide the amount of compute required to train an LLM. Do we have actual clues that NASA is training their own frontier model?