The map of AI

Four nested circles

Artificial intelligence is the broad goal: systems that perform tasks we associate with intelligence — perceiving, reasoning, planning, using language. Early AI was mostly hand-written rules and search. Machine learning is the subset where the behavior is learned from data rather than programmed: you choose a family of functions and let an algorithm pick the member that fits examples. Deep learning is the subset of ML whose function families are neural networks with many layers, trained by gradient descent (4.1 From perceptron to MLP; universal approximation, 4.4 Backpropagation, fully derived). Generative AI is the subset of deep learning that learns to produce data — text, images, audio, code — by modeling the distribution of the data itself (7.6 Decoder-only: GPT and autoregressive generation, 8.9 Generative models II: Diffusion and flow matching).

An agent is not a fifth circle; it is a system built around a model. It perceives (reads inputs and tool results), reasons (the model), and acts (calls tools, writes files, browses) in a loop until a goal is met (10.1 What is an agent? Loops, autonomy and workflows vs agents). The model is the brain; the agent is the brain plus hands, memory and a loop.

Intuition. A model answers one question. An agent decides which questions to ask, in what order, and what to do with the answers.

Four ways a machine can learn

The big story of the last decade is the third row: large language models are self-supervised — the internet provides the labels for free, because the next token is always known — and then refined with supervised examples and reinforcement learning (8.1 The LLM lifecycle and model landscape, 8.3 Alignment: RLHF, PPO, DPO, GRPO and Constitutional AI).

A short history, 1943 → 2026

Two patterns run through the table. Progress alternates between ideas (backprop, attention) and scale (GPUs, data, compute). And almost every breakthrough was a fix for a specific failure of its predecessor: XOR motivated hidden layers, vanishing gradients motivated LSTMs and ResNets, the seq2seq bottleneck motivated attention, attention's quadratic cost motivates today's efficient architectures (7.10 Efficient inference: KV cache, GQA, FlashAttention, speculative decoding, quantization, 7.12 Beyond vanilla attention: state space models, long context and hybrids).

The vocabulary you need on day one

• Model. A function $f_\theta$ that maps inputs to outputs; $\theta$ (theta) is the collection of all its adjustable numbers.
• Parameters. The entries of $\theta$: the weights and biases. GPT-2 small has about 124 million; frontier models have hundreds of billions (7.4 Inside the block: residuals, LayerNorm, FFN and embeddings).
• Loss. A number $L(\theta)$ measuring how badly $f_\theta$ fits the training data; smaller is better (4.3 Loss functions catalog).
• Training. Adjusting $\theta$ to reduce $L(\theta)$, almost always by gradient descent (1.10 Optimization: from gradient descent to Adam). Inference is using the trained $f_\theta$ on new inputs. Training is expensive and done once; inference is cheap and done billions of times.
• Dataset splits. Train to fit, validation to choose settings, test to report an honest score exactly once (3.1 What is learning? ERM, generalization and the ML workflow, 3.5 Evaluating models: metrics, validation and the bias–variance trade-off).
• Overfitting. Fitting the training data's noise so that performance on new data drops; generalization is doing well on data you did not train on (3.6 Regularization and generalization).
• Hyperparameters. Settings you choose rather than learn: learning rate, number of layers, batch size.
• Epoch and batch. One pass over the training data; one small group of examples processed together.
• Embedding. A learned vector that represents an item — a word, an image, a user — so that geometry (distance, angle) reflects meaning (1.1 Vectors, dot products and norms, 6.3 Word embeddings).
• Token. The unit a language model reads and writes, usually a word piece; about three-quarters of an English word (6.1 Text to numbers: tokenization). The context window is how many tokens a model can consider at once.

How the phases build on each other

The math of 1.1 Vectors, dot products and norms through 1.10 Optimization: from gradient descent to Adam is the language. Classical ML (3.1 What is learning? ERM, generalization and the ML workflow) teaches you to think about generalization with models simple enough to analyze. Deep learning (4.1 From perceptron to MLP; universal approximation) replaces hand-made features with learned ones, using the same losses and optimizers. Vision (5.1 Convolution: the operation and its math) and sequences (6.4 Recurrent neural networks and BPTT) are the two great application domains whose problems produced the transformer (7.1 Attention Is All You Need: the big picture). Large language models (8.1 The LLM lifecycle and model landscape) are transformers at scale plus alignment; retrieval (9.1 Why RAG, and the canonical pipeline) and agents (10.1 What is an agent? Loops, autonomy and workflows vs agents) are what you build with them.

Where this shows up later

• Each paradigm gets its own treatment: supervised learning — 3.1 What is learning? ERM, generalization and the ML workflow; self-supervised pretraining — 7.6 Decoder-only: GPT and autoregressive generation; reinforcement learning — 4.11 Reinforcement learning essentials.
• The history becomes a reading list in 11.2 Landmark papers: the reading list and the Papers timeline page.
• The vocabulary is used, without definition, from 1.1 Vectors, dot products and norms onward.