“Google's Gemini Omni model is trained on multiple data types like text, audio, and video, based on the hypothesis that a model improves at one modality by observing others.”