Foundation Models

Foundation models are large-scale, pretrained neural networks designed to serve as a base for many downstream tasks. Trained on unlabeled web-scale data, they underpin applications in NLP, vision, and even multi-modal AI.

Their strength lies in transferability: a single model (e.g., CLIP, GPT, DINOv2) can be fine-tuned or prompted to tackle everything from translation to segmentation.

Still, foundation models raise crucial questions: How general is too general? What biases are inherited? How can we ensure efficiency, fairness, and interpretability?

Foundation models blur the boundary between model and infrastructure—their capabilities define what's possible downstream.