Looking under the hood of multimodal AI