Foundation models are large neural networks pretrained on broad, diverse data at scale, producing general-purpose representations that can be adapted to many downstream tasks through fine-tuning or prompting. They underpin most modern large language models, vision models and multimodal systems, and their training typically relies on vast web-scale corpora such as Common Crawl. Their broad capability comes at the cost of high training compute and emergent, sometimes unpredictable, behaviour.