Foundation model
A foundation model is a large AI model trained on broad data at scale that can be adapted to many different tasks, such as writing, coding, image analysis or search. The term was coined by Stanford researchers in 2021. Large language models are the best-known type of foundation model.
Foundation models are pre-trained once, at great expense, on large and varied datasets, then adapted through prompting, fine-tuning or retrieval for specific uses. One model can power a chatbot, a coding assistant and a document-analysis tool. Bommasani et al. (2021) introduced the term to highlight both this reuse and the risk that flaws in one model spread to every product built on it.
Why it matters: most AI software is a layer on top of foundation models from a small number of providers. Knowing which model a product uses helps you judge its capabilities, its data-processing chain and how dependent it is on another company's pricing and policy decisions.
Example: a startup builds an AI meeting assistant by calling a foundation model through an API and adding its own recording, transcription workflow and interface.
Limits: a foundation model's general training does not make it an expert in your domain, and products built on the same model can still differ widely in quality because of their prompts, data and design.
Related terms
Learn it step by step
- Lesson 1What is artificial intelligence (AI)?
- Lesson 2What is machine learning?
- Lesson 3What is deep learning?
- Lesson 6What is a large language model (LLM)?
- Lesson 11What are parameters in an LLM?
- Lesson 16What is fine-tuning?
- Lesson 21What is computer vision?
- Lesson 23What is AGI and how is it different from AI?
- Lesson 24What is open-source AI?
- Lesson 27Is AI bad for the environment?
Sources
- On the Opportunities and Risks of Foundation Models. Bommasani et al., arXiv, August 2021
Topic: AI fundamentals
Last updated