foundation model

The term was coined in 2021 by Stanford’s Center for Research on Foundation Models (CRFM), in a report led by Rishi Bommasani with Percy Liang as senior author. They picked “foundation” to underscore the model’s “critically central yet incomplete character”: it is the base layer other systems build on, not a finished product. One pretrained checkpoint gets specialized for many jobs instead of training a fresh model per task.

Adaptation and why the name matters

Adaptation happens through fine-tuning, instruction tuning, prompt engineering, LoRA, or retrieval (RAG). Two properties from the CRFM report matter here. Scale produces emergent capabilities the base training never targeted. And because most apps build on the same few bases, defects homogenize: a flaw in the foundation model propagates to every product downstream.

Foundation model is broader than large language model. Every LLM is a foundation model, but so are image models (DALL-E, Stable Diffusion) and multimodal models. CRFM chose the wider term because the paradigm was never only about language. Regulators later wrote their own versions: the EU AI Act settled on the label “general-purpose AI model” for a model “trained with a large amount of data using self-supervision at scale” that can “competently perform a wide range of distinct tasks,” while the 2023 US executive order used a training-compute threshold (10^26 floating-point operations) to define dual-use foundation models.

For search, the load-bearing fact is that every AI answer engine (Google’s AI Overviews, ChatGPT, Perplexity, Gemini) sits on a foundation model. What that model learned during pretraining, and what it retrieves at query time, decides whether your brand appears in the answer.