Foundation model history

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.

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.

Limitations and evidence. Broad training and reuse do not imply uniform ability across every downstream task. Evaluation must stay attached to a model version and adaptation setup, and model-level limitations can recur across several applications that share the same foundation.

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.

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.

Why it matters: Foundation model is best understood as part of a system, not as a label in isolation. The practical questions are what inputs it accepts, what outputs it produces, which evidence supports its capabilities, what constraints surround deployment, and how those details change over time. The Compendium separates the documented facts above from interpretation and records an as-of date so readers can re-check the primary source.

Reading note: this is original explanatory writing. It links to the cited source instead of reproducing protected text. Product availability, model names, pricing, and policies can change; follow the source for the provider's current position.

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.

Why it matters: Foundation model is best understood as part of a system, not as a label in isolation. The practical questions are what inputs it accepts, what outputs it produces, which evidence supports its capabilities, what constraints surround deployment, and how those details change over time. The Atlas separates the documented facts above from interpretation and records an as-of date so readers can re-check the primary source.

Reading note: this is original explanatory writing. It links to the cited source instead of reproducing protected text. Product availability, model names, pricing, and policies can change; follow the source for the provider's current position.