Glossary

Forty terms you will meet in any AI article, defined without jargon.

43 terms

Algorithm: A fixed sequence of steps for solving a problem. A recipe is an algorithm. Most of computing is algorithms; only some of it learns.
Alignment: The work of getting a system to pursue what people actually want rather than a proxy that was easier to measure.
Artificial general intelligence: A hypothetical system that matches human ability across most cognitive tasks. Nothing built today is close, and forecasts disagree by decades.
Attention: A mechanism that lets a model weigh how much each word in a sentence matters to each other word. The core idea behind transformers.
Backpropagation: The method that assigns blame for an error backwards through a network, so every weight learns how it should change.
Bias (statistical): A systematic error that pushes predictions in one direction. Distinct from social bias, though the two often appear together.
Chatbot: Any program you converse with in text. Older ones followed scripts; current ones are language models.
Classification: Predicting which category something belongs to: spam or not, malaria or not, which of ten digits.
Computer vision: Getting useful information out of images and video — what is in the picture, where, and how it is moving.
Corpus: The body of text a language model is trained on.
Dataset: The collection of examples a model learns from. Its shortcomings become the model's shortcomings.
Deep learning: Machine learning with neural networks of many layers. "Deep" refers to the number of layers, nothing more profound.
Diffusion model: An image generator that starts from noise and removes it step by step until a picture matching the prompt emerges.
Embedding: A list of numbers representing a word, image or document, arranged so similar things sit close together.
Epoch: One complete pass through the training data. Training usually takes many.
Feature: One measurable input the model uses — age, pixel brightness, days since last purchase.
Fine-tuning: Taking a model somebody else trained expensively and adjusting it on your own smaller dataset.
Generative AI: Models that produce new text, images, audio or code rather than only labelling existing things.
GPU: A graphics processor. Its ability to do thousands of small multiplications at once is why modern AI became practical.
Ground truth: The correct answer, established by something other than the model, used to train and to score it.
Hallucination: A confident, fluent, false statement from a language model. It is generating plausible text, not looking anything up.
Hyperparameter: A setting you choose before training, such as learning rate or layer count, as opposed to a weight the model learns.
Inference: Using a trained model to make a prediction. Training happens once and costs a fortune; inference happens constantly.
Label: The known answer attached to a training example. Labelling is slow, human and expensive.
Large language model: A model trained on enormous quantities of text to predict the next piece of a sequence. Everything it does follows from that.
Loss function: The measure of how wrong a prediction is. Training is the search for weights that make it small.
Machine learning: Programs that improve at a task by processing examples, instead of following rules a person wrote out.
Model: The learned thing itself — the structure plus the numbers it settled on during training.
Multimodal: A model that handles more than one kind of input, such as reading an image and answering questions about it in text.
Natural language processing: The field concerned with getting computers to work with human language.
Neural network: Layers of simple units, each passing a weighted sum through a threshold. Loosely inspired by neurons; not a brain simulation.
Overfitting: When a model memorises its training examples and fails on new ones. The student who learned the past papers, not the subject.
Parameter: One learned number inside a model. Current large models have hundreds of billions.
Prompt: The instruction and context you give a language model. Small changes in wording change the output more than people expect.
Regression: Predicting a number rather than a category: a price, a temperature, a yield.
Reinforcement learning: Learning by trying actions and receiving rewards or penalties. How game-playing and robotic systems are usually trained.
Supervised learning: Learning from examples that already carry the right answers.
Token: The chunk of text a language model actually handles — roughly a short word or part of a word.
Training: The process of adjusting a model's parameters until its predictions on known examples stop improving.
Transformer: The architecture, introduced in 2017, behind nearly every current language and multimodal model.
Turing test: Turing's 1950 proposal: can a machine's text be told apart from a person's? A conversation starter more than a benchmark.
Unsupervised learning: Finding structure in data that carries no labels, such as grouping customers who behave alike.
Weight: The strength of one connection in a network. Training is the adjustment of weights, and nothing else.

Words that sound technical and are not

Some vocabulary in AI carries far less meaning than its tone suggests. Neural means "arranged in layers of weighted sums". Deep means "several layers". Learning means "numbers were adjusted to reduce an error". Attention means "a weighting over positions in the input".

None of them describes an inner life. Keeping the plain translation in mind makes a lot of breathless writing much easier to evaluate.