2.4. Multimodal Generation

💡 First Principle: "Generative" isn't only text. Modern Foundry models and services generate and consume images, video, and speech, and an agentic interaction often needs more than one modality (speak a question, see a chart, hear an answer). The modality filter from Phase 1 decides which capability to reach for.

Why care: AI-103 explicitly expanded into multimodal generation — image generation, video, and speech-to-text/text-to-speech for agentic interactions. These appear both in the gen-AI domain (generating media) and brush against the vision and text-analysis domains (understanding media). Keeping "generate" versus "understand" straight prevents cross-domain confusion.

⚠️ Common Misconception: Learners assume multimodal means one giant model does everything. In practice you compose capabilities — a multimodal LLM for image understanding, a dedicated image-generation model, a speech service for voice — orchestrated together, often inside an agent.