What is a foundation model?

In plain English

A foundation model is a big, general model trained once on a very broad pool of data, then reused as the starting point for many specific applications. The name captures the idea: it is a base layer other things are built on top of.

Before this approach, each task meant its own model, trained from scratch on its own carefully labelled dataset. Now a developer can take an existing foundation model and adapt it — by giving it instructions and examples in a prompt, by connecting it to a company’s documents, or by fine-tuning, which means continuing training on a smaller specialised dataset.

Large language models are the best-known kind, but foundation models also exist for images, audio, video, protein structures and multiple data types at once.

Why it matters

This is the economic engine of the current AI wave. Training a foundation model costs an enormous amount in computing power and data, so only a handful of organisations do it. Everyone else builds on top, far more cheaply. That makes AI widely accessible, but it also concentrates influence — and any flaw or bias in a popular base model propagates into every product built on it.

An example

A law firm does not train its own AI. It takes an existing general model and connects it to the firm’s library of past contracts, so staff can ask questions in plain English and get answers grounded in the firm’s own documents.

What to watch out for

“General-purpose” does not mean reliable everywhere. A model trained mostly on English web text may be weak in other languages or in specialist domains. Adapting a model also does not erase what it learned before, so inherited errors and biases can resurface. And relying on someone else’s model means your product depends on their pricing, availability and policy changes.