Defining Foundation Models: Boundaries, Mechanisms, and Stack Abstractions

Foundation models are large, broadly trained models that can be adapted to many downstream tasks through prompting, retrieval, fine-tuning, or additional components. [1] The term requires precise explanation because it identifies a specific information flow, training choice, runtime mechanism, or governance boundary, and treating it as a synonym for advanced AI makes claims impossible to test. [1] The definition relies on three practical commitments: an identifiable input, a characteristic transformation, and an outcome that can be evaluated against a stated objective. [1]

Modern AI stacks build abstractions on top of one another, where pretraining creates reusable capability and adaptation changes behavior. [1] A model's performance can be determined by surrounding data, interfaces, hardware, permissions, and people even when the underlying model is unchanged. [1] Therefore, a useful explanation separates the model’s learned behavior from the product that decides when, where, and with what authority that behavior is used. [1] The nearest misleading shortcut is a narrow model trained from scratch for one prediction target, which changes the causal story and resource requirements. [1]

Defining AI systems precisely separates the underlying statistical representations from the final deployment abstractions that users interact with. By establishing clear operational boundaries, practitioners can better evaluate when to rely on a broad foundation model adapted via retrieval or prompting, versus when a narrow model built for a single prediction target is appropriate. Treating "foundation model" as an umbrella term for all advanced AI obscures the distinct hardware and resource commitments required for broad pretraining compared to narrow predictive pipelines. What the sources don't address: How the specific computational costs and data volume requirements of broad pretraining compare to the costs of training these narrow prediction models from scratch.