A large, general purpose model that many different applications are built on top of. One expensive model is trained once, then thousands of products use it for different purposes. This is the economic structure of the current industry: a small number of organisations can afford to train these models, and a very large number build on them.
Technical view
The term was coined in 2021 to capture both the generality of these models and the fact that downstream systems inherit their properties. Inherited weaknesses, biases or failure modes propagate to everything built on top, which is why the concentration of the field around a handful of base models is itself a matter of concern for some researchers and regulators.
Related
Part of AI Terminology Index, section: Extending and Applying Models.