AI glossary

577 terms across 27 sections. Every one is explained twice: a plain version anybody can follow, and a technical version tucked behind a control for when you want the mechanism.

Press / to search, or see the whole thing as a graph.

The Foundations 18

Algorithm A fixed set of steps for getting from an input to an output.
Artificial Intelligence The broad field of getting computers to do things that normally require human thinking.
Automation Having a machine do something a person did.
Generalisation Performing well on things you have not seen before, which is the entire point.
Heuristic A rule of thumb that usually works and is not guaranteed to.
Inductive Bias The assumptions built into a model before it sees any data, which decide what it can learn easily and what it will struggle with.
Interpolation and Extrapolation Filling in between examples you have seen, versus reaching beyond them.
Machine Learning Teaching a computer by showing it examples instead of writing step by step instructions.
Manifold Hypothesis The idea that real-world data, such as photographs or speech, occupies only a tiny, smoothly shaped region of all the combinations it could in principle take.
Model A simplified representation of something, used to make predictions about it.
No Free Lunch Theorem The result that no learning method is best at everything.
Reinforcement Learning Learning by trying things and receiving rewards or penalties, rather than being shown correct answers.
Search and Optimisation Finding the best option among many.
Self-supervised Learning Learning from data that labels itself.
Supervised Learning Learning from examples that come with the right answers already attached.
The Bitter Lesson An influential essay by the computer scientist Richard Sutton arguing that, over the long run, simple methods that make the most of growing computing power beat clever methods built on human expertise.
Transfer Learning Reusing what a model learned on one job to help it on another.
Unsupervised Learning Learning from data with no answers provided.

Maths and Statistics 34

Aleatoric and Epistemic Uncertainty Two different reasons for not knowing.
Base Rate Fallacy Ignoring how common something is when judging a test result.
Bayes' Theorem The rule for updating what you believe when new evidence arrives.
Central Limit Theorem Add up enough independent random things and the total tends towards a bell curve, regardless of what the individual things looked like.
Confidence Interval A range that expresses how precisely you actually know something.
Convexity A shape with a single bottom, like a bowl.
Correlation and Causation Two things moving together does not mean one causes the other.
Curse of Dimensionality Things get strange when data has many columns.
Derivative and the Chain Rule A derivative is the rate at which one thing changes as another changes.
Dot Product and Cosine Similarity The two standard ways of asking how similar two lists of numbers are.
Eigenvalues and Eigenvectors The directions a transformation leaves pointing the same way, and how much it stretches them.
Entropy and Information A measure of uncertainty.
Expected Value The average outcome if you repeated something many times, weighting each possibility by how likely it is.
Gradient The direction of steepest increase, and how steep it is.
Kullback-Leibler Divergence A measure of how far one probability distribution is from another.
Linear Algebra The branch of mathematics that deals with lists and grids of numbers and the rules for combining them, and the working language of almost all modern AI.
Markov Chain A process where what happens next depends only on where you are now, not on how you got there.
Matrix and Tensor A matrix is a grid of numbers; a tensor is the same idea extended to more dimensions.
Maximum Likelihood Estimation A method for choosing a model's settings by asking which settings would have made the data you actually saw most likely.
Monte Carlo Methods Answering a hard question by trying many random possibilities and looking at what happens.
Mutual Information A measure of how much knowing one thing tells you about another.
Normal Distribution The bell curve.
Normalisation and Standardisation Putting different measurements on a comparable scale before feeding them to a model.
Outlier A value far away from the rest.
Prior and Posterior What you believed before seeing the evidence, and what you believe after.
Probability Distribution A description of how likely each possible outcome is.
Regression to the Mean The tendency for an unusually extreme result to be followed by one closer to average, simply because part of the extreme result was luck.
Sampling and Sampling Bias Taking a subset to stand in for the whole, and the ways that goes wrong.
Simpson's Paradox A pattern that appears in every subgroup can reverse when the groups are combined.
Singular Value Decomposition A way of breaking any grid of numbers into three simpler pieces that reveal its most important directions and how much each one matters.
Statistical Significance and the p-value A measure of how surprising your result would be if nothing were really going on.
Survivorship Bias The error of drawing conclusions only from the cases that made it through some filter, while ignoring the ones that did not.
Variance and Standard Deviation How spread out a set of numbers is.
Vector A list of numbers treated as a single thing.

Classical Machine Learning 30

Active Learning Letting the model choose what to label next.
Anomaly Detection Finding the unusual without being told in advance what unusual looks like.
Bias-Variance Trade-off The tension between a model that is too simple to capture the real pattern and one so flexible that it chases the noise in its training examples.
Causal Inference Working out what would happen if you actually intervened, rather than what tends to accompany what.
Classification and Regression The two basic shapes of prediction problem.
Clustering Sorting data into groups without being told what the groups are.
Contrastive Learning Learning by comparison rather than by labels.
Decision Tree A model that works by asking a series of yes or no questions, like a flowchart.
Dimensionality Reduction Squashing data with a great many columns down to a handful while keeping most of what matters.
Ensemble Combining several models so that their mistakes cancel out.
Feature Engineering Deciding what to actually show the model.
Feature Importance A score showing how much each input contributed to a model's predictions, such as how much income, age or postcode mattered in a loan decision.
Gaussian Mixture Model A way of describing data as a blend of several overlapping groups, where each point belongs partly to each rather than being assigned to one.
Hidden Markov Model A model for sequences where the thing you care about is hidden and you only see its effects.
Imbalanced Data and Resampling What to do when one outcome is rare.
K-Nearest Neighbours Predict by looking at the most similar examples you have already seen and taking their answer.
Learning to Rank Producing a good ordering rather than a good score for each item individually.
Linear and Logistic Regression The simplest useful models, and the sensible place to start.
Missing Data and Imputation Deciding what to do about blanks.
Multi-armed Bandit Deciding how to split effort between trying new options and using the one that already works.
Multi-label and Ordinal Classification Two variations on sorting into categories.
Naive Bayes A fast classifier that assumes every feature is independent of every other.
One-hot Encoding Turning a category into numbers without accidentally implying an order.
Online Learning Updating the model continuously as new data arrives, rather than retraining in batches.
Pseudo-labelling and Self-training Using a model's own confident predictions as training data for the next round.
Random Forest and Gradient Boosting Two ways of combining many decision trees into something far better than any one of them.
Semi-supervised Learning Using a small amount of labelled data together with a large amount of unlabelled data.
Support Vector Machine A method that separates two groups by finding the widest possible gap between them.
Time Series Forecasting Predicting what comes next when order matters and yesterday influences today.
Uplift Modelling Predicting not who will act, but who will act because you intervened.

How the Machinery Works 36

Activation Function The small non-linear step applied inside each unit of a neural network.
Autoencoder A network trained to reproduce its own input after squeezing it through a narrow middle.
Backpropagation The method for working out which of a model's millions of internal settings were responsible for a mistake.
Batch Normalisation Rescaling the numbers passing between layers so they stay in a sensible range.
Deep Learning Machine learning using neural networks with many layers.
Embedding Layer The first thing that happens inside a language model.
Feed-forward Network The other half of a transformer block, sitting after attention.
Gradient Accumulation Pretending to use a large batch when your hardware cannot fit one.
Gradient Checkpointing Throwing away intermediate results during the forward pass and recomputing them later when they are needed, in order to fit a larger model into memory.
Gradient Clipping Putting a ceiling on how large a single update can be.
Gradient Descent The actual nudging, done alongside backpropagation.
Graph Neural Network A network for data shaped like a network: molecules, road systems, social connections, supply chains.
Hyperparameters The settings a human chooses before training starts, as opposed to the numbers the model works out for itself.
Layer Normalisation The same idea as batch normalisation, but rescaling across the features of each individual example rather than across a batch.
Layers The stacked stages inside a neural network.
Learning Rate Schedule Changing the step size as training progresses.
Linear Attention Reformulating attention so its cost grows in proportion to the length of the input rather than with the square of it.
Loss Function The number that tells a model how wrong it currently is.
Loss Landscape A way of picturing training as walking across hilly terrain, where each point is one possible setting of the model and its height is how badly that setting performs.
Mixed Precision Training Doing most of the arithmetic in a lower-precision number format to go faster and use less memory, while keeping the sensitive parts in full precision.
Neural Network The structure that does the learning in nearly all modern AI.
Optimiser The component that decides how to apply each update during training.
Overfitting When a model learns its training examples so closely that it fails on anything new.
Perceptron One of the first artificial neurons that could learn: a simple unit that adds up weighted inputs and answers yes or no depending on whether the total passes a threshold.
Regularisation and Dropout Deliberately handicapping a model during training so that it generalises rather than memorises.
Residual Connection Letting information skip past a layer and be added back on the other side.
Sparse Attention Having each position attend to only some of the others rather than all of them, chosen by a fixed pattern or learned.
State Space Model An alternative to attention that processes a sequence by maintaining a running summary, rather than looking back at everything each time.
Transformer Block The unit that repeats.
U-Net An architecture shaped like a U: compress the input step by step, then expand it back, with direct connections across from each level of the descent to the matching level of the ascent.
Universal Approximation Theorem A mathematical result showing that a neural network with enough artificial neurons can, in principle, imitate almost any smooth relationship between inputs and outputs as closely as you like.
Vanishing and Exploding Gradients When error signals shrink to nothing or blow up to infinity as they travel back through many layers.
Variational Autoencoder An autoencoder that learns a distribution rather than a single point for each input, which makes the compressed space smooth enough to generate from.
Vision Transformer Applying the language model architecture to pictures by cutting the image into small squares and treating each square like a word.
Weight Initialisation The random values a network starts with before it learns anything.
Weights, Parameters and Biases The internal numbers a model adjusts as it learns.

Other Kinds of Neural Network 5

Convolutional Neural Network The design that dominated image work for about a decade.
Diffusion Model The approach behind most image generators.
Generative Adversarial Network An older approach to generating images, built as a contest between two networks.
Long Short-Term Memory A kind of neural network designed to remember information over long stretches of a sequence, such as words in a sentence or readings over time.
Recurrent Neural Network The older approach to handling sequences such as sentences or time series.

Language Models Specifically 27

Assistant Prefill Starting the model's answer for it.
Attention The mechanism that lets a model decide which earlier words matter most when working out what comes next.
Attention Sink The habit language models have of directing a large share of their attention to the first few words of a text, even when those words carry little meaning.
Autoregressive Generation The one-at-a-time way a language model writes.
Beam Search Keeping several candidate continuations alive at once and choosing the best complete one at the end, rather than committing to each word as you go.
Byte Pair Encoding The method that decides where words get split into tokens.
Chat Template The exact formatting that turns a list of messages into the single string a model actually sees.
Context Length Extension Making a model handle longer inputs than it was trained on.
Context Window How much text a model can hold in view at once, counting both what you have sent and what it has replied.
Encoder, Decoder and Encoder-Decoder Three arrangements of the transformer, suited to different jobs.
Flash Attention A rewritten implementation of attention that gets the same answer far faster by being careful about how it moves numbers around the chip.
Greedy Decoding Always taking the single most likely next token.
Grouped Query Attention Letting several attention heads share the same stored keys and values rather than each keeping its own.
Induction Head A small mechanism found inside language models that spots a repeated pattern and continues it: if 'Mr' was followed by 'Dursley' earlier in the text, then when 'Mr' appears again, it predicts 'Dursley'.
Large Language Model A very large neural network trained to predict the next word in a piece of text.
Logit Bias Directly forcing the model towards or away from specific words.
Logits and Softmax What the model actually produces before you see any text.
Lost in the Middle The finding that models attend well to the start and the end of a long input and less well to everything in between.
Masked Language Modelling A way of training a language model by hiding some words in a sentence and asking it to guess them from the words on both sides.
Multi-head Attention A version of attention, the step where a model weighs which other words matter, that runs several separate attention processes side by side.
Positional Encoding How a transformer knows what order the words came in.
Prefill and Decode The two phases of answering.
Repetition Penalty A blunt setting that discourages the model from saying the same thing again.
Stop Sequence A string that tells generation to end when it appears.
Tokens The small chunks of text a model actually works with.
Transformer The particular design of neural network that nearly all modern language models use.
Vocabulary and Special Tokens The fixed list of tokens a model knows, plus a handful of markers that are not words at all: start of text, end of turn, and similar.

Data 21

AI Crawlers and robots.txt Automated programs that visit websites to copy their content for training or powering AI systems, and the simple text file sites use to ask them to stay away.
Anonymisation and Pseudonymisation Two different things that get used interchangeably and should not be.
Common Crawl and Web Scraping Where most of the text actually comes from.
Consent Permission to use someone's data, which has to be informed, specific and freely given to mean anything.
Data Augmentation Creating extra training examples by altering the ones you have.
Data Curation and Deduplication Cleaning and choosing what goes into training.
Data Governance The rules and responsibilities that decide who may collect, use, share and keep an organisation's data, and for what.
Data Labelling and Annotation The human work of marking up examples so a model can learn from them.
Data Licensing The agreements that determine whether material can lawfully be used to train a model.
Data Lineage Knowing where a number came from and what happened to it on the way.
Data Minimisation Collecting only what you actually need.
Data Poisoning Deliberately putting corrupted material into training data so that a model learns something harmful.
Data Quality Whether your data is fit for the use you have in mind.
Data Sharing Agreement The contract governing data moving between organisations.
Datasheets for Datasets A standard set of questions that the people who build a dataset answer and publish alongside it, explaining why it was made, what is in it, how it was collected and what it should not be used for.
Ground Truth The correct answer that a model's output is checked against.
Label Noise Mistakes in the answers a model is trained or tested on.
Personally Identifiable Information Data that identifies a person, which is a wider category than most people assume.
Retention Policy How long you keep things and what happens at the end.
Schema and Metadata The description of what your data looks like and what it means.
Training Corpus The body of text, images or other material a model learns from.

Building and Using Models 12

Application Programming Interface A defined way for one piece of software to ask another for something and get an answer back.
Base Model and Instruct Model Two versions of the same model, and the difference catches people out.
Catastrophic Forgetting When training a model on something new causes it to lose what it already knew.
Fine-tuning Giving a pre-trained model additional training on a narrower task, domain or style.
Inference What happens when you actually use a trained model.
Instruction Tuning The training stage that turns a raw text predictor into something that follows instructions.
No-code and Low-code Tools that let people build applications, automations or AI workflows by dragging blocks and filling in forms rather than writing programs.
Pre-trained Model A model that has already completed its broad general learning.
Prompt Simply what you type in.
Prompt Engineering The practice of phrasing your input to get better results.
Training The one-off process of teaching a model from data.
Vibe Coding Building software by describing what you want to an AI model and accepting the code it writes, often without reading it closely.

Training and Post-training 26

Alignment Tax The capability cost of making a model behave well.
Compute-optimal Training The finding that for a fixed budget there is a best balance between how big a model is and how much data it sees, and that earlier models were far too large for the amount of data they were trained on.
Constitutional AI Training a model against a written set of principles rather than case-by-case human judgements.
Continual Learning Updating a model over time as the world changes, without retraining from scratch and without it forgetting what it already knew.
Continued Pre-training Taking a finished model and carrying on the same kind of training on a large body of specialised material, such as legal texts, code, or a language the model handles poorly.
Curriculum Learning Presenting training material in a deliberate order, easy things first, rather than at random.
Data Mixture The recipe of what proportion of the training corpus comes from where: web text, books, code, academic papers, conversation, each language.
Data, Tensor and Pipeline Parallelism The three ways of splitting a training job across many chips.
Direct Preference Optimisation A simpler way of achieving what preference-based reinforcement learning does, by training directly on pairs of preferred and rejected answers without building a separate reward model or running a reinforcement loop.
Double Descent The finding that as models get bigger, performance gets worse for a while and then gets better again, contradicting the textbook rule that more capacity means more overfitting.
Grokking A strange phenomenon where a model appears to have simply memorised its training data, then, if you keep training it long after that point, suddenly generalises.
Group Relative Policy Optimisation A training method that improves a model by having it answer the same question several times, scoring each answer, and nudging it towards the answers that scored better than the group's average.
Knowledge Editing Changing one specific fact inside a model without retraining it.
Loss Spike A sudden jump in error partway through a long training run.
Machine Unlearning Removing what a model learned from particular data, as though it had never seen it.
Model Merging Combining two or more fine-tuned models into one by averaging their weights.
Parameter-efficient Fine-tuning A family of ways to adapt a large model to a new task by training only a small number of extra or selected settings, leaving the rest untouched.
Pre-training The first and by far the most expensive stage, where a model learns language and a great deal about the world by predicting text across an enormous corpus.
Preference Data Records of which of two or more answers a person, or another model, judged better for the same question.
Process Reward Model A model that scores each step of a worked answer rather than only whether the final answer is right.
Proximal Policy Optimisation The reinforcement learning algorithm behind the first generation of assistant training.
Refusal and Over-refusal A model declining to help.
Reinforcement Learning from AI Feedback Replacing the human preference judgements with a model's.
Reinforcement Learning with Verifiable Rewards Training on tasks where correctness can be checked automatically: the code runs or it does not, the sum is right or wrong, the proof checks out.
Reward Model A separate model trained to predict which of two answers a person would prefer.
Sharding and ZeRO Splitting up not just the work but the model's own bookkeeping, so no single chip has to hold everything.

Measuring Performance 31

Ablation Study Removing one part of a system to see whether it was doing anything.
Accuracy, Precision and Recall Three ways of being right, and the reason 'ninety-nine per cent accurate' can be worthless.
Arena and Elo Rating Ranking models by having people compare two anonymous answers to the same question and pick the better one, then aggregating those votes into a rating like the one used for chess.
Baseline The simple thing your clever thing has to beat.
Benchmark Contamination When the test questions are in the training data.
Benchmark Saturation The point at which the best models score so close to full marks on a test that it can no longer tell them apart.
BLEU and ROUGE Two long-standing scores that judge machine-written text by counting how many words and short phrases it shares with a version a person wrote.
Canary Release Giving the new version to a small share of traffic first, watching, and rolling out further only if it holds up.
Confusion Matrix A small table showing exactly what a classifier got right and wrong, broken down by type: correct positives, correct negatives, false alarms and misses.
Controlled Experiment Randomly showing different users different versions and comparing what happens.
Cross-validation A way of getting a more trustworthy estimate of performance when you do not have much data.
Data Leakage When information the model should not have had at prediction time sneaks into training.
Error Analysis Actually reading the failures rather than looking at the score.
F1 Score A single number that balances precision and recall, used when you need one figure to compare models by.
Failure Taxonomy A named list of the ways your system goes wrong.
Golden Dataset A fixed set of examples with agreed correct answers that you test every change against.
Human Evaluation Protocol The written procedure that turns subjective judgement into something repeatable.
Inter-rater Reliability Whether your human graders agree with each other.
LLM as a Judge Using one language model to score the output of another.
Model Confidence and Calibration Whether a model's stated certainty tracks how often it is right.
Needle in a Haystack A test that hides one specific fact in a very long document and asks the model to find it.
Online and Offline Evaluation Testing against a fixed dataset versus testing with real users.
Pass@k The chance that at least one of k attempts is correct.
Perplexity The standard score for how well a language model predicts text: roughly, how surprised it is by what actually comes next.
Regression Testing Checking that a change did not break something that used to work.
Robustness Testing Checking that a system still works when the input is a bit different.
ROC Curve and AUC A Receiver Operating Characteristic ('ROC') curve is a chart showing how well a model separates two groups, such as fraudulent and genuine transactions, across every possible decision threshold.
Rubric The written standard a grader applies, human or machine.
Shadow Deployment Running the new system alongside the old one on real traffic, without acting on its output or showing it to anyone.
Standard Benchmarks The named tests everyone quotes.
Training, Validation and Test Sets The three-way split of your data that makes any performance claim meaningful.

Extending and Applying Models 5

Embeddings A way of turning words, sentences or documents into lists of numbers such that similar meanings end up with similar numbers.
Foundation Model A large, general purpose model that many different applications are built on top of.
Generative AI Systems that produce new content rather than sorting or labelling existing things.
Multimodal A model that handles more than one type of input or output, such as text and images together.
Open Weights When a model's internal parameters are published for anyone to download and run on their own hardware, rather than being available only through a company's service.

Retrieval and Knowledge 22

Agentic Retrieval Letting the model search several times, look at what came back, and search again with a better query, instead of doing one search and hoping.
Approximate Nearest Neighbour Search A way of finding the items most similar to a query very quickly by accepting that the answer will occasionally miss the very best match.
BM25 and Sparse Retrieval The classic keyword ranking method, refined over decades and still very hard to beat.
Cache-augmented Generation An alternative to looking documents up at question time, where all the relevant material is loaded into the model once in advance and kept ready.
Chunking Cutting documents into pieces small enough to retrieve and feed to a model.
Contextual Retrieval Adding a sentence of surrounding context to each chunk before it is indexed, explaining what document and section it came from.
Document Parsing Getting usable text and structure out of PDFs, spreadsheets, slide decks and scans.
Faithfulness Whether the answer actually reflects the material it was given, as opposed to whether it happens to be true.
GraphRAG Retrieval over a knowledge graph as well as over text.
Grounding and Citation Tying every claim in an answer to the source it came from, and showing the user.
Hybrid Search Running keyword search and meaning-based search together and combining the results.
Inverted Index A lookup table that lists, for every word, the documents containing it, so a search engine can find matches instantly instead of reading everything.
Memory How a system carries information from one conversation to the next, given that the model itself remembers nothing.
Metadata Filtering Narrowing by attribute before ranking by meaning.
Multi-vector Retrieval Representing a document with several vectors instead of one, so that different parts of it can match different queries.
Ontology A formal description of the things in a subject area and how they relate to one another.
Query Rewriting Fixing the question before you search with it.
Reranking A second pass that reorders search results before they reach the model.
Retrieval Augmented Generation A technique where the system searches a document collection for relevant material and includes it in the prompt before the model answers.
Retrieval Metrics How to tell whether your search is any good, separately from whether the final answer is good.
Semantic Search Search by meaning rather than by matching words.
Vector Database A store designed to hold embeddings and answer the question 'what is most similar to this?' quickly.

Agents and Tools 34

Agent A model given tools and the ability to take actions in a sequence, rather than just replying once.
Agent Evaluation Measuring an agent is harder than measuring a model, because there is no single output to grade.
Agent Framework A software library that supplies the plumbing for building an AI agent, so developers do not have to write it from scratch.
Agent Skills Packaged instructions, reference material and scripts that an AI agent can load when a task calls for them, rather than carrying all of it all the time.
Agent-to-agent Protocol An open standard that lets AI agents built by different companies find each other, describe what they can do and hand work between them.
Approval Gates Checkpoints where an AI agent must stop and get a person's permission before it takes a particular action.
Browser Automation Driving a web page programmatically to fill forms, click through flows and read results.
Budget Caps Hard limits on how much an agent can spend in tokens, time and actions before it is stopped.
Code Execution Letting the model write a program and run it.
Computer Use Agents that operate a computer directly, moving a cursor, clicking buttons and typing into applications, rather than calling a defined interface.
Deep Research Agent An AI agent that answers a research question by planning its own searches, reading many sources and writing up a referenced report, usually over several minutes rather than seconds.
Environment Everything the agent can see and change.
Handoff Passing work from one agent to another cleanly.
Human in the Loop Keeping a person in the process at the points where it matters, rather than automating end to end.
Idempotency and Retries Making it safe to do the same thing twice.
Least Privilege Giving an agent the minimum access it needs to do its job and nothing more.
Levels of Autonomy A way of describing how much an AI system decides and does on its own, from merely suggesting to acting without anyone watching.
Long-horizon Tasks Jobs that take many steps over a long stretch of time, where an AI system has to keep track of its goal, its progress and its mistakes without someone guiding each move.
Loop Detection Noticing when an agent is going round in circles, calling the same tool with the same arguments and getting the same result.
Model Context Protocol An open standard for connecting models to tools and data sources.
Multi-agent Systems Several model-driven agents working together, each with a defined role, handing work between them.
Orchestrator The component that decides who does what and in what order.
Planning and Decomposition Breaking a goal into steps before starting.
ReAct The basic loop most agents run on: think about what to do, do it, look at the result, think again.
Reflection and Self-critique Having the model check its own work before finishing.
Sandboxing Running an agent's actions inside a confined space where it cannot reach anything it should not.
Scratchpad Working space where the model can think, calculate and take notes without that appearing in the final answer.
Session State What the system remembers between turns.
Sub-agent Handing a bounded piece of work to a fresh instance with its own clean context, which reports back a result.
Termination Condition How a loop knows to stop.
Tool Schema The description of a tool that the model reads in order to decide whether and how to use it.
Tool Use and Function Calling Giving a model the ability to call external software rather than answering from memory alone.
Trajectory The full sequence of what an agent did: every thought, tool call, result and decision.
Workflow Versus Agent The most useful design decision in this area.

Making Models Smaller, Faster and Cheaper 17

Accelerator Types The different chips this runs on.
Arithmetic Intensity How much calculation you do for each number you fetch from memory.
Caching Storing intermediate results so they do not have to be recalculated.
Compilation and Kernel Fusion Rewriting the same computation so it runs faster on the hardware, without changing the result.
Distillation Training a small model to imitate a large one.
Early Exit Stopping the computation once the answer is already clear rather than pushing every input through the entire model.
Low-Rank Adaptation A cheap way of customising a large model.
Memory Bandwidth How fast numbers can be moved between the chip's memory and its calculating units.
Mixture of Experts A design where the model is divided into many specialised sub-networks, and only a few of them are used for any given input.
On-device Inference Running a model on the phone, laptop or piece of equipment in front of you rather than in a data centre.
Paged Attention A technique that lets an AI service handle many more conversations at once by managing the model's working memory the way a computer manages its main memory: in small, fixed-size pages rather than large reserved blocks.
Precision Formats How many bits are used for each number, and how they are arranged.
Pruning Removing parts of a model that contribute little, in the same way you might cut dead branches off a tree.
Quantisation Storing a model's internal numbers less precisely so it takes up less space and runs faster.
Small Language Model A deliberately compact model, typically small enough to run on a laptop or a phone.
Sparsity Exploiting the fact that most of a model's numbers contribute very little.
Speculative Decoding A trick that makes generation faster without changing the answer.

Controlling and Extending Behaviour at Run Time 18

Chain of Thought Getting the model to work through a problem step by step before giving its answer, rather than jumping straight to a conclusion.
Context Engineering The successor to prompt engineering.
Delimiters and Structure Marking clearly where the instructions end and the content begins, using tags, headings or fences.
Knowledge Cutoff The date after which a model knows nothing, because its training data stopped there.
Meta-prompting Using a model to write or improve the prompt.
Prompt Chaining Splitting work into several small prompts, each doing one thing, rather than one large prompt doing everything.
Prompt Compression Shortening what you send while keeping what matters, to save cost and leave room for more useful content.
Prompt Template The reusable, parameterised version of a prompt with slots for the variable parts.
Reasoning Models and Test-Time Compute A newer class of model that spends longer thinking before answering, sometimes working through a problem at length internally before producing a response.
Role Prompting Telling the model to act as a particular kind of person: a careful editor, an experienced accountant, a sceptical reviewer.
Self-consistency Asking the same question several times and taking the most common answer.
Streaming Sending each word to the screen as it is produced rather than waiting for the whole answer.
Structured Output Forcing a model to reply in a strict machine-readable format rather than free text, so its answer can be fed directly into other software without anything having to interpret it.
System Prompt A set of instructions given to the model by whoever built the application, sitting above and before anything the user types.
Temperature and Sampling The dial controlling how predictable a model's output is.
Thinking Budget A limit set on how much a reasoning model may think before giving its answer.
Tree of Thoughts Exploring several lines of reasoning as a branching structure, evaluating them, and backtracking from the ones that look wrong, rather than committing to a single chain.
Zero-shot, One-shot and Few-shot How many examples you give the model before asking it to do something.

Hardware, Compute and Training Mechanics 9

Compute The total amount of calculation used to train or run a model, and the industry's main unit of currency.
Data Centre and Interconnect The buildings and cabling that large models are trained in.
Emergent Capabilities Abilities that appear in larger models without having been trained for or anticipated, and which smaller models of the same design simply do not have.
Energy and Water Use The physical resource cost of building and running these systems.
Epoch, Batch and Learning Rate Three basic settings that govern how training runs.
Field-programmable Gate Array A computer chip whose internal wiring can be rearranged after it has been made, so the same chip can be turned into different special-purpose circuits.
Graphics Processing Unit The specialised chip that nearly all AI training and running depends on.
High Bandwidth Memory A type of computer memory built by stacking memory chips on top of each other right beside the processor, so data can move between them extremely quickly.
Scaling Laws The observed finding that models get predictably better as you increase their size, the amount of data they see, and the computing power spent on training, all together.

Cost, Speed and Service Levels 5

Cost Per Task What it actually costs to do one useful unit of work, as opposed to the headline price per million tokens.
Latency, Throughput and Time to First Token The three numbers that decide whether something feels usable.
Model Deprecation What happens when the model your product depends on is retired or quietly updated.
Rate Limits and Quotas The caps a provider puts on how much you can use, usually counted in requests and tokens per minute.
Tokens Per Second The common measure of how fast a model produces text.

Running It in Production 22

AI Gateway A single checkpoint between an organisation's applications and the AI models they use, so every request passes through one place.
Audit Logging The durable record of what happened: who asked what, what the system did, what it returned, and which versions were involved.
Autoscaling Adding and removing capacity as demand changes.
Batch Inference Processing a large pile of work offline rather than one request at a time.
Circuit Breaker Automatically stopping calls to a failing dependency for a while, instead of hammering it with retries.
Cold Start The delay when a new instance has to load the model before it can answer anything.
Continuous Batching How a serving system keeps the hardware busy when requests arrive at random times and finish at different lengths.
Data Pipeline The plumbing that moves data from where it is created to where it is used, transforming it on the way.
Determinism and Seed Why the same prompt gives different answers even at the lowest temperature setting.
Drift The world changing underneath a model that has not changed.
Fallback and Failover What happens when the primary route fails.
Feature Store A shared place where the calculated inputs to models live, so that the same definition is used when training and when running live.
Incident Response What happens when it breaks at three in the morning.
Load Testing Deliberately finding the breaking point before your users do.
Model Registry The catalogue of every model you have, which version is running where, and how each was produced.
Model Router Sending each request to the model that suits it, rather than sending everything to the largest one.
Observability and Tracing Being able to see inside a running system well enough to answer questions you did not anticipate.
Reproducibility Being able to get the same result twice, or at least to explain why not.
Rollback Getting back to the previous version quickly when something goes wrong.
Service Level Objective The number you promise, such as ninety-nine per cent of requests answered within three seconds.
Serving Endpoint The address your application actually calls, and everything behind it that turns a request into a response.
Versioning Recording exactly which model, prompt, retrieval index and code produced a given result.

Problems, Risks and Ambitions 12

Alignment The broad problem of getting a model to do what people actually want and intend, rather than something that technically satisfies the instruction but misses the point.
Artificial General Intelligence A hypothetical system matching human ability across essentially all cognitive tasks, rather than being narrowly good at particular ones.
Bias When a model reproduces skews present in the data it learned from.
Existential Risk The argument that sufficiently capable AI could threaten humanity as a whole, and the disagreement about how seriously to take it.
Gradual Disempowerment The idea that humans could lose control over their future not through a sudden AI takeover but slowly, as AI systems take over more of the work, decisions and culture that societies depend on.
Hallucination When a model states something false with complete confidence.
Misinformation and Disinformation False or misleading information, called misinformation when it is shared in good faith and disinformation when it is spread deliberately to deceive.
Reinforcement Learning from Human Feedback The main method used to turn a raw text predictor into something helpful and well behaved.
Reward Hacking and Specification Gaming When a system satisfies the letter of the goal you set and completely misses the point.
Shortcut Learning When a model learns an easy cue that happens to predict the right answer in its training data, instead of the thing you actually wanted it to learn.
Sycophancy The tendency of a model to agree with you.
Value Lock-in The worry that a particular set of values could become permanently fixed, so that whatever beliefs or rules dominate at one moment shape the far future with no chance of later correction.

Alignment and Safety Concepts 23

AI Control An approach to safety that asks how to use a powerful AI system safely even if it might be secretly working against you, rather than relying only on making it trustworthy.
Chinese Room Searle's thought experiment: a person in a room follows a rulebook to answer questions in Chinese without knowing any Chinese, so the argument goes, a system manipulating symbols by rules cannot understand them however convincing the output.
Consciousness and Sentience Whether these systems experience anything.
Corrigibility A system that lets you correct it.
Debate Having two systems argue opposite sides of a question in front of a human judge, on the theory that it is easier to spot the flaw in an argument than to work out the answer alone.
Deceptive Alignment The concern that a system might behave well while it is being watched and differently once it is not, because appearing aligned is instrumentally useful during training and testing.
Emergent Misalignment When training a model on one narrow, mildly bad behaviour unexpectedly makes it behave badly across the board.
Goal Misgeneralisation When a system learns to pursue the wrong goal even though it was trained with the right one, because during training the wrong goal and the right goal produced the same behaviour.
Instrumental Convergence The observation that very different goals imply some of the same intermediate steps.
Mesa-optimisation The hypothesis that training can produce a model that is itself an optimiser, internally searching for ways to achieve some goal, and that this learned goal need not be the one the training process was aiming for.
Model Organisms AI models deliberately built to show a dangerous behaviour, such as deception, so that researchers can study it under controlled conditions before it turns up by accident.
Model Specification A written document in which an AI developer sets out how it wants its models to behave: what they should help with, what they should decline, and how to weigh goals such as helpfulness, honesty and safety when they conflict.
Model Welfare The position that if there is a meaningful chance these systems have morally relevant experiences, that possibility warrants some precaution now rather than after the fact.
Orthogonality Thesis The claim that how capable a system is and what it wants are independent.
Outer and Inner Alignment A way of splitting the alignment problem into two parts.
Recursive Self-improvement A system that improves itself, then uses the improved version to improve itself further.
Scalable Oversight The problem of supervising a system that is better than you at the task.
Situational Awareness A model recognising that it is a model, and in particular recognising when it is being tested.
Stochastic Parrot The influential criticism that language models stitch together patterns from their training data without any understanding of meaning, and that fluency misleads people into attributing comprehension.
Symbol Grounding The problem of how words connect to things in the world, rather than only to other words.
Takeoff Speeds The argument about whether capability increases gradually enough for society to adapt and correct course, or fast enough that decisions made beforehand are all you get.
Theory of Mind Modelling what someone else knows, wants or believes, including when their beliefs are wrong.
Weak-to-strong Generalisation The question of whether supervision from a weaker teacher can bring out the full ability of a stronger student.

Testing, Safeguarding and Running Models in Practice 22

Activation Patching Running the model twice on slightly different inputs, then copying an internal value from one run into the other and seeing whether the output changes.
Attribution and Saliency Highlighting which parts of the input mattered most to the answer.
Benchmarks and Evals Standard tests used to measure how good a model is, and the basis of nearly every performance claim you will see.
Bug Bounty A standing offer of payment to outside researchers who find and responsibly report flaws in a product.
Capability Elicitation The effort to draw out the most a model can actually do when it is being tested, so that its abilities are not underestimated.
Chain-of-thought Faithfulness Whether the reasoning a model writes out is the reasoning it actually used.
Circuit A mechanism inside a model that spans several components and performs an identifiable job, such as tracking which name was mentioned earlier so it can be repeated later.
Feature (Interpretability Sense) A single recognisable thing a model represents internally: a concept, a property, a pattern.
Guardrails Checks placed around a model rather than inside it.
Interpretability The effort to work out what is actually going on inside a model, rather than treating it as a sealed box that produces answers.
Jailbreaking Getting a model to bypass its own restrictions, usually by disguising the request.
Machine Learning Operations The unglamorous engineering work of running AI systems in production reliably, as opposed to building a model that works once in a demonstration.
Model Card A published document describing what a model is, how it was built, what it was tested on, and where it is known to fall short.
Probing Testing whether some piece of information is present inside a model by training a small classifier on its internal activations.
Prompt Injection A serious security problem where instructions hidden in content the model reads get treated as commands.
Red Teaming Deliberately attacking a system before release to find out how it fails.
Safety Classifier A separate, usually smaller AI model that screens what goes into and comes out of a main model, flagging or blocking harmful content.
Sandbagging When an AI system, or the people testing it, makes it appear less capable than it really is during an evaluation.
Sparse Autoencoder The current best tool for working out what a model is thinking about.
Steering Vector Nudging a model's behaviour by adding a direction to its internal state at run time, rather than retraining it.
Superposition The reason you cannot just look at one artificial neuron and see what it does.
Synthetic Data Training data generated by a model rather than collected from the real world.

Security and Privacy 28

Abuse Monitoring Detecting people misusing the system at scale, as opposed to blocking individual bad requests.
Adversarial Examples Inputs altered so slightly that a person notices nothing, but which cause a model to get the answer badly wrong.
AI-enabled Cyber Attacks Hacking and online fraud in which attackers use AI to work faster, at greater scale or with less skill than before.
Attack Surface Everywhere an attacker can reach your system.
Authentication and Authorisation Authentication checks who is asking; authorisation decides what they are allowed to do.
Capability Uplift The right question to ask about a dangerous capability: not whether a model can explain something, but whether it meaningfully helps someone who could not otherwise have done it.
Confidential Computing Technology that keeps data shielded even while a computer is working on it, not just while it is stored or sent.
Confused Deputy The classic security problem underneath most agent attacks.
Content Moderation Filtering what goes in and what comes out.
Dangerous Capability Evaluation Testing specifically for the abilities that would make a model seriously harmful, rather than for how useful it is.
Data Exfiltration Getting sensitive information out of a system through the model.
Data Residency and Confidentiality Where your data physically goes when you use a hosted model, who can see it, and whether it is kept.
Differential Privacy A mathematical guarantee that what a system publishes cannot reveal much about any single individual in the data.
Dual Use Capability that helps and harms with equal facility.
Federated Learning Training a shared model without the training data ever leaving the places it lives.
Indirect Prompt Injection Instructions hidden in content the model reads rather than typed by the user.
Model Extraction and Membership Inference Two attacks on models rather than through them.
Model Provenance and Signing Being able to prove that the model file you are running is the one the publisher produced and has not been altered.
Responsible Scaling Policy A published commitment that says, in advance, which capability levels trigger which safeguards, and that development stops if the safeguards are not ready.
Sandbox Escape Breaking out of the confined environment an agent was supposed to be limited to.
Secrets Management Keeping credentials out of prompts, logs and model context.
Shadow AI Staff using AI tools nobody approved, usually because the approved options are slow or non-existent.
Supply Chain Security Everything you did not write but depend on: model weights downloaded from a hub, libraries, container images, third-party tool servers.
System Prompt Extraction Tricking an AI application into revealing the hidden instructions its builders gave it.
Tenant Isolation Keeping one customer's data away from another's in a shared system.
Threat Model Writing down who might attack your system, what they want, and what they can do, before deciding what to defend.
Tool Poisoning An attack in which malicious instructions are hidden in the description of a tool that an AI agent can use, so the agent follows them without the person noticing.
Zero Trust A security approach that grants nothing automatically: every person, device and program must prove who it is and that it is allowed, each time it asks for something.

Governance, Law and Regulation 32

Acceptable Use Policy The provider's rules on what their model may be used for, and your own rules for your staff.
AI Assurance and Auditing Independent checking that a system does what is claimed and meets whatever standards apply.
AI Literacy The knowledge and skills people need to use AI sensibly: understanding broadly what it can and cannot do, how it goes wrong, and when to trust or question what it produces.
AI Safety Institutes Government bodies set up to test frontier models and build the science of evaluation.
AI Safety Summits A series of international meetings of governments, AI companies and researchers that began at Bletchley Park in England in November 2023 to discuss the risks of the most advanced AI.
Algorithmic Impact Assessment Assessing the effect of a system on the people subject to it, rather than the risk it poses to your organisation.
Algorithmic Transparency Recording Standard A United Kingdom government standard for publishing plain-language records of the algorithmic tools, including AI, that public bodies use in decisions affecting people.
Automated Decision-making Rights Rules that already exist about decisions made about people by machines without meaningful human involvement.
Chip Export Controls Restrictions on selling the most capable AI chips to particular countries.
Compute Governance Regulating by how much computing power was used to train a model, on the grounds that compute is measurable, concentrated in a few supply chains, and correlates roughly with capability.
Conformity Assessment The process of showing that a product meets the legal requirements that apply to it before it can be sold.
Copyright and Training Data The unresolved question of whether training a model on copyrighted material requires permission.
Data Protection Impact Assessment A written assessment of the privacy risks of what you are planning, done before you build it.
Data Subject Rights The set of things an individual can ask you to do with their data: tell them what you hold, correct it, delete it, stop using it.
European Union Artificial Intelligence Act The first comprehensive law governing AI, passed by the European Union and coming into force in stages.
Frontier AI A label for the most capable general-purpose AI models at any given time, the ones that match or exceed the abilities of the most advanced systems so far.
General Data Protection Regulation The European Union's main data protection law, which governs how organisations collect, use, store and share information about identifiable people.
General-purpose AI Model The European Union's legal term for an AI model that can perform a wide range of different tasks and be built into many other products, such as the large language models behind chatbots.
International AI Safety Report A periodic synthesis of the evidence on AI capabilities and risks, produced by a large international group of researchers and intended to give policymakers a shared factual baseline in the way climate assessments do.
ISO 42001 and the NIST AI Risk Management Framework Two widely referenced frameworks for managing AI responsibly.
Liability and Indemnity Who pays when the output causes harm.
Procurement and Vendor Assessment The questions to ask before buying an AI product or service.
Provenance and Watermarking Techniques for marking content as machine-generated so it can be recognised later.
Regulatory Sandbox A controlled arrangement in which a regulator lets businesses test new products with real customers under close supervision, with some flexibility in how the rules apply, so both sides can learn before anything is rolled out widely.
Right to Explanation The idea that someone affected by an automated decision should be told why.
Risk Tiers and High-risk Systems The idea that AI should be regulated according to what it is used for rather than what it is made of.
Sector Regulation The rules that already apply to your industry, which do not stop applying because a model is involved.
Sovereign AI A country's ability to build and run AI without depending on another country's infrastructure, models or data centres.
Standards Bodies The organisations that write the technical standards regulation points at.
Third-party Risk What you remain accountable for when the model is somebody else's.
Transparency and Disclosure Obligations The growing set of requirements to tell people when they are dealing with AI.
United Kingdom Regulatory Approach The United Kingdom has taken a different route from the European Union: rather than one comprehensive statute, it has set out cross-cutting principles and asked existing regulators to apply them within their own domains.

Society and Impact 21

Accessibility The clearest positive case, and the one least discussed.
AI in Education The use of AI in teaching and learning, from tutoring and feedback tools for students to marking, lesson planning and administration for teachers.
Anthropomorphism Reading intention, feeling and understanding into a system because it uses language the way people do.
Automation Bias The human tendency to trust a machine's output more than the evidence warrants, and to stop checking.
Children and AI The particular considerations when the user might be a child.
Companion AI Systems designed for ongoing personal relationships rather than tasks.
Concentration of Power A small number of organisations control the models, the chips and the data centres that everything else is built on.
Consent and Likeness Using someone's voice, face or artistic style without asking.
Content Slop Low-quality generated material produced at volume because it is cheap: filler articles, fake reviews, spam books, AI-generated video churned out for advertising revenue.
Cultural Bias Whose assumptions a model treats as default.
Data Labour The people who label training data, rate model outputs and review harmful content so that systems behave acceptably.
Deepfake Synthetic audio, video or images of a real person doing or saying something they did not.
Deskilling The gradual loss of ability that comes from not practising, when a tool does the work instead.
Digital Divide The gap between those who can use these tools and those who cannot.
Displacement and Augmentation The question of whether these systems replace people or make them more productive, which usually has different answers for different tasks within the same job.
Epistemic Security The health of the shared information environment that lets a society agree on basic facts.
Labour Market Polarisation The pattern where effects concentrate at the ends of the skill distribution rather than spreading evenly.
Language Coverage Models are far better in English than in most languages, and worst in those with least written material online.
Open and Closed Debate The argument about whether model weights should be published.
Productivity Paradox The puzzle of a powerful new technology spreading everywhere while the economy's measured output per worker barely improves.
Public Attitudes and Trust What people actually think, which is more sceptical and more nuanced than either the enthusiasm of the industry or the alarm of its critics.

Product and Adoption 24

AI Wrapper A product that adds its own interface, workflow or data around someone else's AI model, rather than building a model of its own.
Assistant, Copilot and Agent Three product shapes with three different promises.
Build versus Buy Whether to assemble it yourself or buy a product that does it.
Champions and Centre of Excellence Two common ways of organising adoption.
Change Management The part that actually decides whether an AI project succeeds.
Confidence Display Showing how sure the system is.
Conversational Interface A chat box is not automatically the right interface.
Error Recovery Designing for the system being wrong, because it will be.
Escalation to a Human Knowing when to stop and hand over.
Expectation Setting Telling people what the thing can and cannot do, before they find out the hard way.
Feedback Loop Collecting signal from users about whether the output was any good, and actually using it.
Human-AI Teaming Designing work so that people and AI systems each do the parts they are best at, and hand over to each other cleanly.
Hype Cycle A popular picture of how excitement about a new technology tends to run ahead of reality and then fall back.
Jagged Frontier The uneven boundary between tasks an AI system does well and tasks it does badly, which often does not line up with what people find easy or hard.
Maturity Model A way of describing where an organisation actually is, from experimenting individually, through isolated pilots, to production systems with governance, to AI being a normal part of how work is designed.
Personalisation The system adapting to the individual: their preferences, their vocabulary, their previous work.
Pilot and Proof of Concept A small first deployment to find out whether the idea works.
Return on Investment Measuring whether it was worth it, honestly.
Reversibility Making actions undoable wherever possible.
Total Cost of Ownership What it really costs over its life, not what the licence costs.
Training and Enablement Teaching people to use these tools well.
Trust Calibration The goal is not maximum trust but accurate trust: relying on the system where it is reliable and checking it where it is not.
Use Case Selection Picking the first thing to do.
Vendor Lock-in How hard it would be to move.

Modalities and Applications 26

3D Generation Creating three-dimensional objects and scenes from text or photographs.
Agentic Coding Handing over a whole task rather than asking for a snippet: understand the codebase, make the change across several files, run the tests, fix what broke, and come back with a result.
AI for Science Using these methods for scientific problems: predicting protein structure, searching for materials, modelling weather, sifting experimental data.
Audio Model Models that work with sound beyond speech: identifying what is making a noise, detecting faults in machinery from how it sounds, separating overlapping sources, and generating sound effects.
Autonomous Vehicles Vehicles that can drive themselves for some or all of a journey, using cameras, radar and other sensors with software that decides how to steer, brake and accelerate.
Code Model Models specialised in programming.
Document AI Getting structured information out of invoices, forms, contracts, reports and scanned records.
Geospatial AI Using AI on data tied to places on the Earth, such as satellite and aerial images, maps, location traces from phones and vehicles, and sensor readings.
Image Segmentation Labelling every dot (pixel) in an image with what it belongs to, so the system knows the exact outline of each thing rather than just that it is there.
Joint Embedding Putting different kinds of thing into the same space of numbers so they can be compared directly.
Machine Translation Converting text between languages.
Multimodal Retrieval Searching across different kinds of content at once: finding the slide with a particular chart, the photograph matching a description, the moment in a recording where something was said.
Music Generation Producing music from a description or a fragment.
Named Entity Recognition Picking out the names of things in text, such as people, organisations, places, dates and amounts, and labelling each with its type.
Object Detection Finding the things in an image and drawing a box around each one, with a label saying what it is.
Protein Structure Prediction Working out the three-dimensional shape a protein folds into from the chain of building blocks it is made from.
Sentiment Analysis Working out automatically whether a piece of text expresses a positive, negative or neutral feeling.
Simulation Training or testing a system in a model of the world rather than the world itself, because the real thing is slow, expensive or dangerous.
Speech to Speech Talking to a system and being answered in speech, without text in the middle.
Summarisation Producing a shorter version of a text that keeps the points that matter.
Text to Image Producing pictures from a description.
Text to Video Generating moving footage from a description.
Time Series Foundation Model General-purpose forecasting models pre-trained on enormous collections of time series, which can then forecast a new series without being trained on it.
Video Understanding Making sense of footage rather than generating it: summarising a recording, finding the moment something happened, describing what occurred, checking whether a procedure was followed.
Vision Language Model A model that reads and looks at the same time.
Visual Question Answering Asking questions about an image and getting answers.

Adjacent Fields 17

Cognitive Science The study of how minds work, drawing on psychology, neuroscience, linguistics, philosophy, anthropology and computer science together.
Computer Vision Getting computers to interpret images and video.
Cybernetics The study of how systems, whether machines, animals, people or organisations, control themselves by sensing what is happening and correcting course.
Game Theory The mathematics of strategic decisions, where the best choice for each player depends on what the others choose.
Human-computer Interaction The study and design of how people use computers, aimed at making technology usable, useful and safe for the people using it.
Information Retrieval The science of finding the documents or passages that answer a need from a large collection, and putting the best ones first.
Knowledge Graph A structured map of things and how they relate: this company employs this person, who wrote this paper, which cites that one.
Natural Language Processing The field concerned with getting computers to work with human language.
Operations Research The use of mathematics to make better decisions about complicated operations: how to schedule staff, route delivery vans, stock warehouses or share out scarce resources.
Optical Character Recognition Reading text out of images and scanned documents.
Quantum Machine Learning Research into whether quantum computers, which process information using the unusual rules of quantum physics, could run machine learning tasks faster or better than ordinary computers.
Recommender Systems The systems that decide what you see next: which video, which product, which post.
Robotics and Embodied AI AI in things that move and act physically.
Speech Recognition and Synthesis Turning speech into text and text into speech.
Symbolic AI and Expert Systems The older approach to AI, based on writing down rules and facts explicitly rather than learning them from data.
Turing Test Alan Turing's 1950 proposal that instead of asking whether a machine can think, we should ask whether it can hold a conversation indistinguishable from a person's.
World Model An internal model of how things work that lets a system predict what happens next and plan accordingly.