What is a Foundation Model?

Before foundation models, most AI systems were trained narrowly for one task on task-specific data. The foundation model approach inverted that: pretrain one very large model on a broad corpus (text, and increasingly images, audio, or code) so it learns general patterns of language, reasoning, and world knowledge, then reuse that single model as the base for many different applications. GPT, Claude, Gemini, and Llama are all foundation models in this sense — each one general-purpose model serving as the foundation for chatbots, coding assistants, agents, and countless other products built on top. The term "foundation" captures both the strength and the risk: because so many downstream systems build on the same base, any bias, blind spot, or capability gap in the foundation model propagates broadly across everything built on it, and improvements to the base model lift every application built on it at once. Foundation models are typically adapted to specific needs via prompting (no weight changes), fine-tuning (updating weights on task-specific data), or retrieval augmentation (grounding responses in external documents) rather than being retrained from scratch for each use case.