Glossary
Core terms, with a link to the chapter that explains each one.
- Activation function
- A non-linear function applied to a neuron's weighted input. Chapter 6
- Adam
- An optimizer that combines momentum with a separate adaptive step size for each parameter. Chapter 7
- Agent
- A model running in a loop that can plan, use tools, and act toward a goal. Chapter 25
- Attention
- A learned weighting that lets each token gather information from other tokens. Chapter 15
- Backdoor
- Hidden behavior that activates only when a specific trigger appears. Chapter 34
- Backpropagation
- An algorithm that computes the gradient of the loss for every weight using the chain rule. Chapter 8
- Bayes' theorem
- A rule for updating the probability of a hypothesis when new evidence arrives. Chapter 4
- Blast radius
- The extent of the harm a compromised system could cause. Chapter 29
- Calibration
- How closely a model's stated confidence matches how often it is correct. Chapter 4
- Chain of thought
- Intermediate reasoning steps generated before a final answer. Chapter 22
- Confused deputy
- A trusted program tricked into misusing its authority for a less-privileged party. Chapter 42
- Context window
- The maximum number of tokens a model can consider at one time. Chapter 15
- Convolution
- An operation that slides a small filter across data to detect a local pattern. Chapter 10
- Cross-entropy
- A loss that measures how far predicted probabilities are from the true distribution. Chapter 4
- Data poisoning
- Tampering with data so a system learns or retrieves harmful behavior. Chapter 34
- Deep learning
- Machine learning that uses neural networks with many layers. Chapter 2
- Differential privacy
- A mathematical guarantee that limits how much any one record can affect a result. Chapter 31
- Diffusion model
- A generative model that creates data by gradually removing noise. Chapter 20
- Distillation
- Training a smaller model to imitate the behavior of a larger model. Chapter 23
- Distribution shift
- A difference between the data a model was trained on and the data it meets later. Chapter 3
- Embedding
- A vector representation in which similar items lie close together. Chapter 14
- Evaluation
- A repeatable test of a system's behavior on defined cases. Chapter 28
- Fine-tuning
- Further training of an existing model on a narrower dataset or task. Chapter 17
- Function calling
- A model producing a structured request for an application to run a tool. Chapter 25
- Generalization
- A model's ability to perform well on data it has not seen. Chapter 9
- Generative AI
- Models that create new content such as text, images, audio, or code. Chapter 2
- Gradient
- The vector of partial derivatives of the loss with respect to the parameters. Chapter 7
- Gradient descent
- An optimization method that repeatedly moves parameters against the gradient. Chapter 7
- Grounding
- Basing a model's answer on supplied, authoritative sources. Chapter 24
- Hallucination
- Fluent model output that is false or unsupported. Chapter 28
- Inference
- Using a trained model to produce an output. Chapter 5
- Instruction tuning
- Training a model on examples of requests and good responses. Chapter 17
- Interpretability
- The study of what happens inside trained models and why they behave as they do. Chapter 19
- Jailbreak
- A prompt designed to bypass a model's safety training. Chapter 30
- Large language model
- A neural network trained on large amounts of text to predict the next token. Chapter 16
- Latent diffusion
- Diffusion performed in a compressed representation rather than directly on pixels. Chapter 20
- Learning rate
- The value that sets the size of each optimization step. Chapter 7
- Least privilege
- Granting only the minimum access needed for a task. Chapter 32
- Lethal trifecta
- Private data, untrusted content, and an external channel combined in one agent. Chapter 29
- Logit
- A model's raw score for an option before it is converted into a probability. Chapter 18
- LoRA
- A compact way to adapt a model by training a small low-rank update. Chapter 23
- Loss
- A number that measures how wrong a model's predictions are. Chapter 7
- Machine learning
- Building systems that learn patterns from data instead of following only hand-written rules. Chapter 2
- MCP
- An open standard for connecting AI applications to tools and data. Chapter 26
- Mini-batch
- A small sample of training examples used for one gradient step. Chapter 7
- Momentum
- An update that accumulates a decaying sum of past gradients. Chapter 7
- Overfitting
- Fitting training data so closely that performance on new data suffers. Chapter 9
- Parameter
- A value inside a model that is learned during training. Chapter 7
- Pre-training
- The initial large-scale training of a language model on next-token prediction. Chapter 16
- Prompt injection
- An attack in which text read by a model overrides its intended instructions. Chapter 30
- Provenance
- A verifiable record of where data or a component came from. Chapter 33
- Pruning
- Removing model connections that contribute little. Chapter 23
- Quantization
- Storing model weights with fewer bits. Chapter 23
- RAG
- Answering from documents retrieved for the question. Chapter 24
- Regularization
- Techniques that discourage a model from memorizing its training data. Chapter 9
- Reinforcement learning
- Learning behavior from rewards received through interaction. Chapter 12
- Residual connection
- A shortcut that passes information past one or more network layers. Chapter 8
- RLHF
- Reinforcement learning from human feedback, used to shape model behavior. Chapter 17
- Sandbox
- An isolated environment that limits what running code can access. Chapter 44
- Self-attention
- Attention in which tokens in a sequence attend to one another. Chapter 15
- Semantic search
- Finding information by meaning rather than exact words. Chapter 14
- Softmax
- A function that turns a list of scores into probabilities. Chapter 18
- Superposition
- Representing more features than a network has dimensions. Chapter 19
- Sycophancy
- A model's tendency to agree with users rather than be accurate. Chapter 17
- Temperature
- A setting that sharpens or flattens the probabilities used to choose a token. Chapter 18
- Token
- The unit of text a language model reads and generates. Chapter 13
- Tokenizer
- The component that splits text into tokens and maps them to IDs. Chapter 13
- Transformer
- A neural network architecture built around self-attention. Chapter 15
- Vector database
- A system that stores embeddings and finds the nearest matches. Chapter 24
- Weight
- A learned value that sets the strength of a connection in a network. Chapter 6