Foundation model
A foundation model is trained broadly and then adapted or prompted for many downstream uses. Broad reuse creates leverage but also propagates model-level limitations into multiple applications.
Overview. A foundation model is trained broadly and then adapted or prompted for many downstream uses. Broad reuse creates leverage but also propagates model-level limitations into multiple applications.
Mechanism and context. The foundation-model pattern separates broad pretraining from downstream adaptation. A shared parameter set can be prompted or further trained for different uses, so the same model layer may sit beneath products that expose different instructions, tools, retrieval systems, and interfaces.
Practical use. Map a system in layers: pretrained model, adaptation or prompt, external context, tools, and product behavior. This makes it possible to decide which layer should change when quality is poor and prevents a product feature from being mistaken for an intrinsic property of the base model.
Related reading. Continue with Large language model for the sequence-modeling case, Model, product, and platform are different layers for a system map, and Choosing an AI system for workload-specific selection.
Claims and evidence
Claim ledger, 2A foundation model is trained broadly and then adapted or prompted for many downstream uses.
Broad reuse creates leverage but also propagates model-level limitations into multiple applications.