What is a foundation model?

A foundation model is a large AI model trained on a broad sweep of data that can be adapted to many different tasks, rather than being built for just one. LLMs are the best-known example, but the idea extends to images, audio, and video too. The name captures the point: it's a foundation you build on.

How is a foundation model different from older AI?

Instead of training a separate model for translation, summarizing, coding, and question-answering, you train one massive general-purpose model and then adapt it to each specific job. That reuse is what makes modern AI economical. It's a real shift from older AI, where each application needed its own narrowly trained model built from scratch.

The defining traits are scale and generality: foundation models are trained on enormous, diverse datasets with huge numbers of parameters, which gives them broad, transferable capabilities that often generalize to tasks their creators never explicitly planned for.

How do you adapt one to a specific task?

You rarely retrain a foundation model from zero. Instead you steer the one you have, usually one of three ways:

• Prompting — describing the task in the input, with no training at all.
• Fine-tuning — further training the model on examples from your specific domain.
• Retrieval — feeding in relevant outside documents so the model can ground its answer in them.

Each option trades effort for control, and you can combine them. This is why a single model can suddenly do so many different things.

What are the trade-offs?

The catch is cost and control. Foundation models are expensive to build, which concentrates them in a handful of well-resourced labs, so most people use someone else's rather than training their own. They can also carry whatever biases or gaps exist in their training data, and because that data is so vast, those problems are hard to fully audit.

Understanding this is key to understanding why one model like an LLM can be so capable — and why its blind spots matter.