At the core of modern AI lies the principle of learning representations from data. The shift from handcrafted features to learned embeddings transformed natural language processing, computer vision, and speech recognition. Word embeddings, then contextualized representations from models such as BERT and its successors, captured semantic relationships that had eluded earlier symbolic approaches. This representational power, combined with the scalability of transformer architectures, gave rise to large language models whose capabilities surprised even their creators. The emergence of abilities like in-context learning and chain-of-thought reasoning in sufficiently large models prompted a reevaluation of what statistical pattern recognition can achieve. It became clear that scale alone, when paired with diverse training data and simple objectives like next-token prediction, can induce behaviors that appear to reflect rudimentary forms of reasoning, abstraction, and even theory of mind.