Multimodal Foundation Models: From Specialists to General-Purpose Assistants