Attention

Embeddings, projection, Q/K/V, dot product, softmax, weighted sum, positional encoding, matrix shapes, and the intuition behind attention — built up from first principles.