Agent. A system in a loop with a goal: it acts, looks at the result and acts again to close the gap (Chapter 12). Chapter 15 draws a stricter line: an adaptive agent can also change its means when its actions stop working.
AI nurse. This book’s name for the machine it assembles, step by step, to answer the 2 a.m. message. An explanatory model built from abilities shown in different systems, not a real product. “The nurse” on its own always means the human.
Attention. The part of a language model that lets each word weigh every earlier word when predicting the next one (Chapter 8).
Automation bias. People’s tendency to accept a machine’s output with too little checking, especially when it is usually right (Chapter 16).
Backpropagation. The method that works out, in one backward pass, how to nudge every weight of a network (Chapter 5).
Base model. A network trained only to predict the next word. It knows a great deal but continues documents rather than helping (Chapter 8).
Borrowed meaning. Meaning a model inherits from the people who wrote its training text, without having checked it against the world itself (Chapter 14).
Calibration. Confidence that matches reality: of all the answers given with 90% confidence, about 90% are right. It must be measured on the right cases and can change with the task or model version (Chapter 16).
Compression. Replacing many examples with a shorter description that generates them. It is closely connected to generalization (Chapter 6).
Contamination. Test questions, or near copies, that a model has already seen in training, so that a high score shows memory rather than skill (Chapter 6).
Context window. Everything a language model can see at once: the question, its own reasoning, any documents. It is long but finite (Chapter 11).
Curiosity. Agency aimed at one’s own ignorance: the goal is to know more, and the actions are questions (Chapter 15).
Delegated goal. A goal given to a system by others, such as its builders and users, rather than one it has for its own sake (Chapter 15).
Expert system. A program built from hand-written if-then rules supplied by experts, such as MYCIN (Chapter 4).
Generalization. Doing well on cases a learner has never seen, under conditions like those it will really face (Chapter 6).
Goodhart’s law. When a measure becomes a target, it stops being a good measure. In AI this shows up as reward hacking and shortcut learning (Chapters 7, 10, 12).
Gradient descent. Learning by repeatedly nudging every adjustable number a little in the direction that reduces the error (Chapter 5).
Grounding. Tying some symbols directly to the world through perception and action (Chapter 14).
Hallucination. A model producing plausible but false or unsupported information, often expressed confidently (Chapter 14).
Harness. The ordinary software around a model that carries out its tool calls and feeds the results back (Chapter 12).
Imitation (supervised fine-tuning). Further training on example conversations, which teaches a base model the role of a helpful assistant (Chapter 9).
Inductive bias. The assumptions a learner brings about what the right rule is likely to look like (Chapter 6).
Irreducibility. A process that cannot be predicted faster than by running it step by step (Chapter 3).
Loss. A single number that measures how wrong a network is on its examples (Chapter 5).
Metacognition. Thinking about one’s own thinking: sensing how likely one is to be right, and acting on it (Chapter 16).
Neural network. A flexible machine built in layers of units, whose behavior is set by many adjustable numbers called weights (Chapter 5).
Overfitting. Doing much better on the training examples than on new ones (Chapter 6).
Reinforcement learning. Learning from the consequences of one’s own attempts: try, get a score, make better-scoring behavior more likely (Chapter 10).
Retrieval-augmented generation. Keeping a model current by searching a maintained document collection and placing the relevant passages in its context window, without changing its weights (Chapter 14).
Reward model. A second network trained to predict which of two answers people would prefer, used as the score in reinforcement learning (Chapter 10).
Shortcut learning. Learning whatever easiest feature goes with the label, such as a hospital’s markers or a surgeon’s ink, instead of the intended one (Chapter 7).
Spectrogram. A picture of sound: time runs across, pitch runs up, and brightness shows loudness. It lets the methods that learned to see learn to hear (Chapter 7).
Speech recognition. Turning spoken words into written ones. Modern systems learn it from many hours of recordings with transcripts (Chapter 7).
Token. A word or piece of a word, the unit a language model reads and writes (Chapter 8).
Turing machine. The simplest model of computation: symbols on a tape, a state, and a fixed table of local rules (Chapter 1).
Undecidable. A question that no program can answer correctly for every input, such as whether a program will ever stop (Chapter 3).
Universal machine. One fixed machine that can follow any rulebook written on its tape (Chapter 2).
Value of information. How much the answer to a question would change what you do (Chapter 15).
Weights. The adjustable numbers inside a network. They are set during training and usually fixed during use; feedback stored in notes can still change later behavior (Chapters 5, 12).