Generative Multimodal Models are In-Context Learners