“Google's research, even before the Gemini program, focused on jointly modeling words in the context of vision, language, and video, not just language alone.”