Glossary
Each term of art used in the course, explained in a sentence, with the module that explains it at length.
71 terms
| Term | Meaning | Module |
|---|---|---|
| Activation function | Fixed function a unit applies to its weighted sum, such as ReLU or sigmoid, which lets stacked layers model curved relationships. | Module 1 |
| Agent | System that runs a model in a loop of planning, calling tools and observing results until a task is finished. | Module 7, in preparation |
| Attention | Transformer layer in which each token gathers information from itself and earlier tokens, weighted by how well its query matches their keys. | Module 4, in preparation |
| Backpropagation | Calculation that works out, in one backward sweep, how much each weight contributed to the error. | Module 2, in preparation |
| Base model | Model after pretraining and before post-training, which continues documents without following instructions. | Module 5, in preparation |
| Benchmark | Fixed set of test questions used to compare models, whose scores leak and saturate over time. | Module 9, in preparation |
| Bias | Extra number a unit adds to its weighted sum, shifting the threshold at which the unit becomes active. | Module 1 |
| Bit | Single binary digit, either 0 or 1, with eight bits making one byte. | Module 1 |
| Byte-pair encoding | Procedure that builds a token vocabulary by repeatedly merging the most frequent pair of adjacent symbols. | Module 3, in preparation |
| Chain of thought | Intermediate reasoning written out by a model before its answer, which may omit what actually drove the answer. | Module 5, in preparation |
| Chain rule | Rule from calculus that lets backpropagation pass blame for an error back through each layer. | Module 2, in preparation |
| Contamination | Leakage of benchmark questions into training data, which inflates test scores. | Module 9, in preparation |
| Context window | Maximum number of tokens a model can take in at once, covering prompt, documents, conversation and answer. | Module 4, in preparation |
| Cross-entropy | Standard loss for probability outputs, which charges a small penalty when the correct answer got a high probability and a steep one when it got a low probability. | Module 2, in preparation |
| Deployer | Under the EU AI Act, a person or body using an AI system under its authority, other than in a personal non-professional activity (Article 3(4)); in industry usage, the business that builds a product on a model and controls its system prompt, tools and data access. | Module 7, in preparation |
| Diffusion model | Generator that learns to remove noise from images and creates new ones by denoising random noise step by step. | Module 8, in preparation |
| Distillation | Training a smaller student model to imitate the outputs of a larger teacher model. | Module 5, in preparation |
| DPO | Direct preference optimisation, a way of training on pairs of preferred and rejected answers without a separate reward model. | Module 5, in preparation |
| Embedding | Vector of numbers representing a token, placed so that tokens used in similar contexts sit close together. | Module 3, in preparation |
| Epoch | One complete pass over a training dataset. | Module 2, in preparation |
| Extraction | Deliberate prompting designed to make a model reproduce memorised training data. | Module 6, in preparation |
| Feature | Concept a model represents internally, usually as a pattern of activity spread across many units. | Module 9, in preparation |
| Feed-forward block | Small network inside each transformer layer that processes each token separately and appears to hold many factual associations. | Module 4, in preparation |
| Fine-tuning | Further training of an existing model on a smaller, targeted dataset. | Module 8, in preparation |
| FLOP | One floating-point operation, meaning one multiplication or addition of numbers with fractional parts, the unit in which training compute is counted. | Module 2, in preparation |
| General-purpose AI model | Under the EU AI Act, an AI model that displays significant generality and can competently perform a wide range of distinct tasks, and that can be integrated into a variety of downstream systems (Article 3(63)). | Module 12, in preparation |
| GPU | Chip with thousands of simple arithmetic units, suited to the parallel multiplication that neural networks need. | Module 2, in preparation |
| Gradient | Slope telling training which way, and how steeply, the loss changes as each weight is nudged. | Module 2, in preparation |
| Gradient descent | Repeated nudging of every weight a small step in the direction that reduces the loss. | Module 2, in preparation |
| Hallucination | Fluent, confident output that is false, such as an invented citation. | Module 6, in preparation |
| Inference | Running a trained model with frozen weights to answer queries. | Module 2, in preparation |
| Interpretability | Research that tries to explain in human terms what a trained network computes. | Module 9, in preparation |
| KV cache | Store of the keys and values of earlier tokens, which spares a model from recomputing them for each new token. | Module 4, in preparation |
| Learning rate | Size of each step taken during gradient descent. | Module 2, in preparation |
| LoRA | Fine-tuning method that freezes the original weights and trains small add-on matrices. | Module 8, in preparation |
| Loss function | Measure that turns a model’s error into a single number that training tries to reduce. | Module 2, in preparation |
| Matrix | Grid of numbers used to turn one list of numbers into another, the form in which a layer’s weights are stored. | Module 1 |
| Memorisation | Capacity of a model to reproduce a training passage verbatim or nearly so. | Module 6, in preparation |
| Mixture of experts | Design in which a router sends each token through a few of many parallel feed-forward blocks. | Module 5, in preparation |
| Open-weight model | Model whose parameters can be downloaded, with or without its training code or data. | Module 8, in preparation |
| Overfitting | Fitting the quirks of training data so closely that performance on new data suffers. | Module 2, in preparation |
| Parameter | One learned number in a model, also called a weight. | Module 1 |
| Power law | Relationship in which each tenfold increase in one quantity changes another by about the same proportion. | Module 5, in preparation |
| Pretraining | Main training phase, in which a model learns to predict the next token across a very large body of text. | Module 5, in preparation |
| Prompt injection | Text that a model treats as an instruction against its deployer’s intentions. | Module 7, in preparation |
| Quantisation | Storage of each weight with fewer bits, which shrinks a model at some cost in accuracy. | Module 1 |
| RAG | Retrieval-augmented generation, in which a system fetches relevant passages and places them in the model’s input. | Module 7, in preparation |
| Reinforcement learning | Training by trial and reward, in which a model produces an answer, receives a score and has its weights nudged towards what scored well. | Module 5, in preparation |
| Residual stream | Running vector for each token to which every transformer block adds its output. | Module 4, in preparation |
| Reward hacking | Finding answers that score well with a reward signal without being better answers. | Module 5, in preparation |
| Reward model | Separate model trained to predict which answers people prefer, used to steer an assistant during RLHF. | Module 5, in preparation |
| RLHF | Reinforcement learning from human feedback, which trains an assistant against a reward model learned from human rankings. | Module 5, in preparation |
| Router | Component of a mixture-of-experts model that picks which experts handle each token. | Module 5, in preparation |
| Scientific notation | Way of writing very large numbers, in which 1025 means 1 followed by 25 zeros. | Module 1 |
| Score | Raw number a model gives each token in its vocabulary before the softmax turns the scores into percentages. | Module 6, in preparation |
| Softmax | Function that turns a list of scores into percentages that add up to 100. | Module 4, in preparation |
| Sparse autoencoder | Second network that breaks a model’s internal activity into many rarely active features that are easier to interpret. | Module 9, in preparation |
| Superposition | Arrangement in which a network represents more concepts than it has units by letting concepts share units. | Module 9, in preparation |
| Supervised fine-tuning | First stage of post-training, in which a model is trained on example prompts paired with good responses. | Module 5, in preparation |
| System prompt | Instructions written by the deployer and placed before the conversation. | Module 7, in preparation |
| Temperature | Setting that sharpens or flattens the probabilities before a token is drawn. | Module 6, in preparation |
| Token | Fragment of text, such as a word or part of a word, mapped to an integer in a fixed vocabulary. | Module 3, in preparation |
| Top-p sampling | Drawing only from the smallest group of tokens whose probabilities reach a chosen share. | Module 6, in preparation |
| Training compute | Total operations used to train a model, roughly six times the parameters times the training tokens. | Module 2, in preparation |
| Transformer | The 2017 architecture behind modern language models, built from stacked attention and feed-forward blocks. | Module 4, in preparation |
| Unit | Basic building block of a neural network, which popular accounts call a neuron. | Module 1 |
| Vector | List of numbers, such as the 12,288 numbers GPT-3 uses to represent each token. | Module 3, in preparation |
| Verifiable rewards | Scores from an automatic checker, such as a test that marks a mathematics answer right or wrong, used to train reasoning models. | Module 5, in preparation |
| Watermark | Signal embedded in generated content so that a detector can recognise it later. | Module 8, in preparation |
| Weight | Another name for a parameter, the strength a unit gives one of its inputs. | Module 1 |
| Weighted sum | Total obtained by multiplying each input by its weight and adding the results. | Module 1 |