Review Foundation Model

A foundation model is a broadly trained model designed to support adaptation or use across many downstream tasks rather than one narrowly defined application.

A developer starts with a model that already learned broad patterns, then uses prompting, retrieval, fine-tuning, tools, or application logic to make it useful for a particular job.

Many products can share the same underlying model while behaving differently because their data, instructions, tools, and controls differ. Risks and dependencies may also be inherited from the common foundation.

One foundation model supports a writing assistant, a document classifier, and a customer-service tool through different application layers and data connections.

Foundation does not mean neutral, complete, or suitable for every use. The model still reflects training choices, limitations, and conditions of use.

One broad base can support many different applications.

* [[large-language-model|Large Language Model]] * [[fine-tuning|Fine-Tuning]] * [[retrieval-augmented-generation|Retrieval-Augmented Generation]] * [[vendor-lock-in|Vendor Lock-In]]

* [Bommasani et al., Stanford CRFM, *On the Opportunities and Risks of Foundation Models*](https://crfm.stanford.edu/report.html) — originating report defining foundation models as broadly trained models adaptable to many downstream tasks. * [Bommasani et al., arXiv record for the foundation-model report](https://arxiv.org/abs/2108.07258) — versioned primary research record and authorship.

* Compare adaptation through [[fine-tuning|Fine-Tuning]] and [[retrieval-augmented-generation|Retrieval-Augmented Generation]].