Instacart is using deep learning to optimize how shoppers pick grocery orders. They developed a sequence prediction model that learns patterns in store layouts and product locations to recommend the most efficient picking path. This reduces the time it takes shoppers to fill orders, allowing Instacart to deliver groceries faster. The model went through several iterations to arrive at an architecture using shared embeddings and hidden layers that achieved a score of 0.7, representing a 10x improvement in training speed and 10% better accuracy over prior versions.