Foundations & Models
Foundation Model
Pretrained once, very broadly and at great expense, then adapted for many purposes — that is the division of labor foundation models have established. A few providers bear the training costs; everyone else builds on top. This saves enormous effort and creates dependency at the same time: the strengths, weaknesses and worldview of the base model migrate into every product built on it.
An example
A hospital needs three things: summarizing discharge letters, coding findings, answering inquiries. Before 2020 that would have been three projects, each with its own data collection, its own training, its own team. Today it is one purchased base model and three different sets of instructions for it. The effort has moved from model development to embedding: Which data may the model see, who checks the result, what happens in case of objection?
A closer look
The term comes from a 2021 Stanford report and refers to models that are trained broadly once and then reused many times. Economically, this shifts the costs: training is a one-off large expense, while adapting to a specific case is comparatively cheap.
Why it matters
Foundation models have inverted the cost structure of AI projects. The expensive part is no longer the model but the integration and the control. For procurement this means: the question “Which model?” is rarely the most important and almost always reversible — models can be swapped. The question “How do we check the output, and where does it depend on our data?” is the expensive one, and the one you keep when you change providers. Anyone who reverses the order buys dependency instead of capability.
Commonly misunderstood
Broadly trained does not mean equally good everywhere. In specialist domains with little public text — law, medicine, internal processes — the base capability is thin and has to be supplemented.