The Transformer is a deep learning architecture introduced by Vaswani et al. in 2017 in the seminal paper Attention Is All You Need. It replaces recurrence and convolution with a pure self-attention mechanism, enabling massive parallelism and long-range dependency modeling that earlier RNNs and LSTMs struggled to capture. The canonical model stacks an encoder and a decoder, each composed of identical layers built from two sublayers — multi-head self-attention and a position-wise feed-forward network — wrapped by residual connections and layer normalization. Because attention is permutation-invariant, positional information is injected through sinusoidal positional encodings or learned embeddings so the model can reason about token order. Scaled dot-product attention computes softmax(QK^T / √d_k) V, where the scaling factor stabilizes gradients for large key dimensions. Multi-head attention projects queries, keys, and values into several subspaces in parallel, letting the model attend to different relational patterns simultaneously — syntactic, semantic, coreferential, positional — before concatenating and projecting the heads back. The decoder mirrors the encoder but adds a masked self-attention layer that prevents tokens from attending to future positions during autoregressive generation, plus an encoder-decoder cross-attention that conditions output on the source sequence. Modern descendants specialize this blueprint: GPT collapses to a decoder-only stack and trains with next-token prediction; BERT uses an encoder-only stack with masked-language-model and next-sentence-prediction objectives; T5 keeps the full encoder-decoder and frames every problem as text-to-text; Vision Transformers (ViT) split images into patch tokens; Whisper, AudioLM, and MusicGen apply the same machinery to spectrograms; AlphaFold leverages it for protein structure. Scaling laws by Kaplan, Hoffmann, and others showed that loss falls predictably as parameters, data, and compute grow together, motivating the leap from BERT-base at 110M to GPT-3 at 175B and beyond. Efficiency innovations — FlashAttention's IO-aware kernels, rotary positional embeddings (RoPE), grouped-query and multi-query attention, sparse and sliding-window attention, mixture-of-experts layers, KV caching, speculative decoding — have together pushed context windows past one million tokens while keeping latency tractable. Today the Transformer is the foundational substrate of modern artificial intelligence — language, vision, audio, code, robotics, multimodal reasoning, and agentic systems all converge on the same essential building block: attention, residuals, normalization, and feed-forward projection, stacked deep and trained at scale.