Foundation model
Also called general-purpose AI model
A large model trained on broad data once, then adapted for many different products instead of each product training its own.
Training a large model from scratch costs a fortune in computing power and data. So the industry works in two steps. First a lab trains one big model on a vast, mixed collection of text, code and images. Then that model is adapted, many times over, for particular jobs: a chatbot, a coding tool, a customer-service assistant.
Researchers at Stanford named these "foundation models" in 2021, and the name stuck. The idea cuts both ways: one strong model lifts everything built on it, and one flaw in it, such as a bias or a blind spot, shows up in all of those products at once.
For example
An airline's chat assistant, a law firm's drafting tool and a homework helper can all run on the same foundation model, each with its own instructions and documents on top.
Put it to use
Where you'll meet foundation model in ChatGPT, Gemini and Claude, how to use it there, and the mix-up most people make.
Daily Bytes Plus is in the Android app.