Foundation model
What is Foundation model?
A large model trained on broad data that many different applications are built on top of, instead of one model per task. Today's large language models are foundation models. You adapt one to your needs, rather than training your own from scratch. The most capable ones are often called frontier models. Stanford's researchers coined the term in 2021.
Why it matters
A foundation model is a large model trained on broad data that many different applications are built on top of, instead of training one model per task. Today’s large language models are foundation models. This matters because it changed who can use advanced AI. Instead of gathering data and training a model from scratch, which few organisations can afford, a company adapts an existing foundation model to its needs. That lowered the barrier enormously. The trade is dependence. You build on a model someone else trained, and its data, limits and biases come baked in.
In practice
A startup builds a legal-document assistant by adapting a general foundation model with its own case data, rather than training a language model from nothing, which would cost more than the whole company is worth. This is the normal pattern now. You take a broad base model and specialise it, through fine-tuning or by feeding it your documents at query time. That pattern has defined how modern AI gets built.
