What Is a Foundation Model? Trained Once, Adapted Many Times

In plain English

The economics are the point. Training one is enormously expensive, so a small number of organisations train them and everyone else adapts them through prompting, retrieval or fine-tuning.

That is why almost nobody builds a model any more. The sensible question is which foundation model to build on, and how to supply it with your own knowledge.

What to know

Trained broadly

General data, no single task in mind.

Adapted downstream

Prompting, retrieval and fine-tuning specialise it.

Concentrated supply

Few organisations can afford to train one.

A dependency

Building on one means inheriting its behaviour, pricing and limits.

Why it matters

Understanding this shapes strategy. Your defensible asset is not the model, which anyone can rent. It is your data, your workflows and the judgement encoded in how you use it.

Common mistakes

×Planning to train a model from scratch without the data or budget.

×Building on one provider with no way to switch.

×Assuming a general model knows your business without being told.

×Fine-tuning when retrieval would have been cheaper and better.

FAQs

Foundation model or LLM?

An LLM is a foundation model for language. The term also covers image and multimodal models.

Should I fine-tune or use retrieval?

Retrieval first, almost always. Fine-tuning is for behaviour and format, not for knowledge.

WRITTEN BY TARIQ SALLAM

Marketing Consultant. Entrepreneur. Content Creator.

I'm a marketing consultant, entrepreneur and content creator. I help businesses grow through practical marketing, websites, SEO, content and AI.

More About Tariq →