Foundation model
파운데이션 모델
A base AI model that learns in advance from a huge amount of data and is then used across many different jobs.
Overview
In the past, separate AIs were built for each job, such as translating, summarizing and classifying. With a foundation model, one model is first trained widely on very broad data and then reused for many jobs. When needed, you teach it a bit more with data for your own job and polish it.
LLMs that handle text are the typical example, and there are also foundation models that handle images, sound and robot movements. They cost a lot of money and need a lot of computing hardware, so only a few companies and research labs can build them. In Korea, they are also called "hyperscale AI."
An everyday example
It is like an athlete who has built up general fitness and the basic skills. With only a little separate practice, they can quickly handle soccer or basketball.
Why it matters
Many AI services are built on top of just a few foundation models. So when something goes wrong with the base model or its terms change, the services on top of it are all affected.