Foundation model
A foundation model is a large AI model trained on broad data that can be adapted to many downstream tasks. Foundation models exist for language, images, audio and multimodal use — the language ones are called LLMs.
A foundation model is the general-purpose engine underneath modern AI products. Trained once on huge amounts of data, it can then be pointed at translation, summarisation, search, image generation, code completion and dozens of other tasks. A large language model is the text version of this idea; the same pattern now runs on images, audio and video too.
The honest take is that foundation models are powerful but generic. Out of the box they know a lot about the world and almost nothing about your business — your products, your customers, your tone of voice, your data. Getting reliable behaviour on your own domain usually means combining a foundation model with retrieval-augmented generation, fine-tuning, or both, plus a proper evals harness.
The common mistake is to reach for fine-tuning first because it sounds the most "AI". In practice, sharper prompt engineering plus RAG is faster, cheaper and easier to change when the underlying model improves. Fine-tuning earns its place when behaviour needs to be deeply consistent or when you are optimising the cost of inference at scale. The rule of thumb: start generic, measure, and only specialise where the data tells you to.
