@sil That's a strong argument, but what if we had a deterministically trained LLM with only free software inputs?