Generative AI · Established · Intermediate

Foundation Model

A large model pretrained on broad data that serves as a reusable base for many downstream applications.

What Foundation Model is

The term, coined at Stanford in 2021, captures a shift: instead of training a model per task, organisations adapt one general model many times.

How it works

Foundation models are pretrained with self-supervision at scale, then adapted by prompting, retrieval, fine-tuning or adapters. Providers expose them through APIs or release weights for self-hosting.

Why it matters

It reframes AI strategy around adaptation and evaluation rather than model building, and concentrates capability in the few organisations that can afford pretraining.

Common uses

→Base for chat assistants
→Vision and speech backbones
→Domain adaptation in medicine, law and finance

Strengths

✓Amortises enormous pretraining cost
✓Strong zero-shot ability

Watch for

✓Concentration of power and single points of failure
✓Opaque training data
✓Vendor dependence

Continue exploring

Large Language Model Transfer Learning Open-Weight Model Scaling Laws