Foundation models are large neural networks trained on massive mixed data. They can be adapted to many tasks without retraining from scratch. Examples include GPT-4, Claude and image models like Stable Diffusion.
Traditional machine learning required custom models for each problem. Foundation models instead capture broad patterns in text, images or audio, then specialise through fine-tuning or prompting.
These models enable rapid prototyping but raise questions around data provenance, scale and safety. Companies must weigh the benefits of generality against compute costs and potential biases.