Already everywhere, often invisible
Face unlock on a modern phone uses a depth-sensing camera and a neural network trained to compare the current face against a stored embedding, similar in spirit to the embeddings from the RAG course but tuned for facial features, deciding in milliseconds whether it's really you. Self-driving systems from companies like Waymo and Tesla fuse detection and segmentation across multiple cameras, tracking every car, pedestrian, cyclist, and lane marking in real time, dozens of times per second, since a wrong or late detection has real safety consequences.
Medical imaging is one of the highest-stakes uses: models trained on labeled X-rays, MRIs, and pathology slides can flag a suspicious mass or classify a skin lesion, often matching or exceeding average human specialist accuracy on narrow, well-defined tasks. These systems are typically deployed as a second opinion that flags cases for a radiologist to review, not as a fully autonomous diagnosis, because false negatives in medicine are unacceptable.