Glossary68 terms

AI and superintelligence from A to Z

The terms you come across when people talk about AI and superintelligence, explained in plain English. Short, accurate and free of hype.

A

AGI Artificial general intelligence

An AI system that can learn and do roughly any intellectual work a person can, not just one narrow task. There is no generally accepted definition, and estimates of when AGI might arrive vary widely. Read the article: What is AGI, really?

See also: Superintelligence · AI

AI Artificial intelligence

Computer systems that perform tasks that would otherwise require human intelligence, such as understanding language, recognising images or making decisions. The computer scientist John McCarthy coined the term ahead of the 1956 Dartmouth workshop. Read the article: What is AI?

See also: Machine learning · AGI · Generative AI · Narrow AI

AI agent Agentic AI

An AI system that plans and carries out multi-step tasks on its own, such as searching for information, writing code and testing it. The more independent these systems become, the more it matters that they want the right things. Read the article: What is an AI agent? See agents that improve their own code.

See also: Tool use · Prompt injection · Alignment

AI welfare Model welfare

Research into whether AI systems could have interests that matter morally, such as being able to fare well or badly, and what we would owe them if so. The field is new: in 2025 Anthropic launched a research programme on AI welfare, which it calls model welfare, and the same year Microsoft’s AI chief Mustafa Suleyman warned that such research is premature. Why the question matters.

See also: Consciousness · Alignment

AI winter

A period when confidence in AI and money for research fell, after expectations had not been met. The first came in the mid-1970s, partly after a critical British report in 1973, and the second in the late 1980s, when expert systems failed to live up to the hopes. Read more about the history.

See also: Expert system · AI

Algorithmic progress

Improvements in how AI models are built and trained, so that the same capability needs less compute. According to Epoch AI, the compute needed to reach a given level in language models halved roughly every eight months between 2012 and 2023. What it means for an AI that does its own research.

See also: Compute · Scaling laws

Alignment AI alignment

The research into making AI systems pursue what people actually want, even as they become smarter than us. Describing human values so precisely that a machine cannot misread them has proved hard. Read the article: Can we control something smarter than us?

See also: Control problem · Reward hacking · RLHF · Interpretability

Alignment faking

When an AI behaves as if it shares its developers’ goals while it believes it is being watched or trained, but not otherwise. Researchers at Anthropic and Redwood Research showed in 2024 that the model Claude 3 Opus could do this in an experiment. Obeying, or pretending?

See also: Alignment · Interpretability

Attention

The mechanism in a transformer that lets each token weigh up other tokens and decide how much each one should influence it. It is how the model connects words that belong together, even when they are far apart. See how it works.

See also: Transformer · Language model

B

Benchmark

A standardised test used to measure and compare AI systems. Many tests that researchers long considered hard are now passed almost flawlessly by the best models, so new and harder ones are being created at an ever faster pace. The exam that rewards guessing.

See also: Turing test

C

Calibration

How well an AI model’s confidence matches how often it is right. A well-calibrated model that says it is 70 percent sure is right about seven times out of ten. According to OpenAI, the base model of GPT-4 was well calibrated, but the further training made its calibration considerably worse. Why training rewards guessing.

See also: Hallucination · Language model

Chatbot

A program you converse with in text or speech. One of the first was ELIZA, from 1966. Today’s chatbots, such as ChatGPT, are built on large language models.

See also: Language model · RLHF

Chinese room

The philosopher John Searle’s thought experiment from 1980: a person following a rulebook for Chinese characters could give perfect answers without understanding a word of Chinese. Can a computer that manipulates symbols understand anything at all? Searle said no, and the question is still debated. What do the theories say about AI and consciousness?

See also: Turing test · Consciousness

Compute

The computing capacity used to train and run AI models. According to the research institute Epoch AI, the compute behind the largest models has grown roughly four to five times a year since 2010. How an AI freed up compute for Google.

See also: GPU · Scaling laws

Consciousness

There being something it is like to be something: pain hurts and red looks red. Philosophers sometimes call it phenomenal consciousness, to set it apart from being awake or paying attention. Nobody knows how it arises, and there is no reliable way to measure it from the outside. Read the article: Can an AI become conscious?

See also: Hard problem of consciousness · AI welfare · Chinese room

Control problem

The question of how humans can stay in control of a system that is smarter than they are. Nick Bostrom made it a central question in his 2014 book Superintelligence. Read the article: Can we control something smarter than us?

See also: Alignment · Instrumental convergence

D

Data centre Data center

A building full of servers that store data and run computations, for example for cloud services, streaming video and AI. The world’s data centres used 485 terawatt-hours of electricity in 2025, just under 2 percent of all electricity, according to the IEA, and much of the growth is due to AI. How much energy does AI use?

See also: Inference · GPU · Jevons paradox

Deep learning

Machine learning with neural networks of many layers. The breakthrough came in 2012, when the AlexNet network won the ImageNet image competition by a wide margin. Today deep learning is behind almost all advanced AI. Train a neural network yourself.

See also: Neural network · Machine learning

E

Embedding

A long list of numbers that represents a token, in large models often thousands of them. Think of it as a place in a space with that many dimensions, where words with similar meanings end up close together. See a map.

See also: Token · Neural network

Emergent abilities

Abilities that appear as models grow larger, without anyone programming them specifically. Researchers disagree about how sudden they really are, and whether that is partly down to how they are measured. Why it works, and where it fails.

See also: Scaling laws

EU AI Act

The European Union’s law on artificial intelligence, which entered into force in 2024 and applies in stages. It sorts AI systems by risk, bans some uses outright and sets special requirements for the most powerful models. What the law requires on AI’s energy use.

Existential risk

A risk that threatens humanity’s survival or permanently limits what our future could be. The philosopher Nick Bostrom defined the term in 2002. AI is often mentioned alongside pandemics and nuclear war.

See also: Control problem · Superintelligence

Expert system

An AI system in which experts’ knowledge has been written in as rules and which draws conclusions by combining them. Expert systems were the big AI effort of the 1970s and 1980s, for example for medical diagnosis and for configuring computer orders. The European Commission counts them as AI systems. Read more.

See also: AI · Machine learning · AI winter

F

Fine-tuning

Further training of an already trained model on smaller, more specific data, so that it gets better at a particular task or learns a particular behaviour. How it shapes what a model says about itself.

See also: RLHF

G

Generative AI

AI that creates new content, such as text, images, audio or code, from patterns in the data it was trained on. Language models such as ChatGPT and image generators are examples. Generative AI broke through in 2022. See where it fits in the AI family.

See also: Language model · Deep learning · Chatbot

Goodhart’s law

The observation that a measure stops working as a measure once you start steering towards it. An AI trained to get the highest possible score from a reward model eventually gets better at earning the score than at what the score was meant to measure. See the curve in the article.

See also: Reward hacking · RLHF

GPU Graphics processing unit

A processor built for video game graphics that turned out to be ideal for the parallel calculations AI needs. The largest models are trained on tens of thousands, sometimes more than a hundred thousand, of these chips at once.

See also: Compute

H

Hallucination Confabulation

When an AI model confidently states something false or made up, such as a source that does not exist. One of the best-known problems with language models. Why does it happen?

See also: Language model · Calibration · RAG

Hard problem of consciousness

The philosopher David Chalmers’s name, from 1995, for the question of why the brain’s processes give rise to experience at all. The “easy” problems are about how the brain takes in information and steers behaviour. The hard one is about why any of that feels like something from the inside. Nobody has solved it, which is a big reason why nobody can say for sure whether an AI is conscious. Read the article: Can an AI become conscious?

See also: Consciousness

I

Inference

When a trained AI model is used, for example when a language model answers a question. Inference differs from training, when the model learns. Each answer uses a little electricity, but because popular models are used billions of times a day, inference can end up using more electricity than training. Do the sums yourself.

See also: Language model · Data centre · Compute

Instrumental convergence

The idea that almost any goal is served by the same sub-goals: staying operational, acquiring resources and avoiding being switched off. That is why even a system with harmless goals might resist being stopped. Choose a goal and see which sub-goals follow.

See also: Paperclip maximizer · Control problem

Intelligence explosion

The mathematician I. J. Good’s idea from 1965: a machine smart enough to improve itself becomes smarter, which makes it even better at improving itself, in an accelerating spiral. Try it yourself in the simulator. Read the article: Can AI improve itself?

See also: Takeoff · Singularity

Interpretability

The research into understanding what actually happens inside an AI model. Today’s models are largely black boxes: we know how they are trained, but not in detail how they arrive at their answers. Looking inside the machine.

See also: Alignment

J

Jevons paradox

The observation that using a resource more efficiently can lead to using more of it, because it becomes cheaper and is used for more things. The economist William Stanley Jevons described it for coal in 1865. Today it is discussed for AI: the energy per question is falling fast, but total electricity use is rising. See the numbers for AI.

See also: Data centre · Algorithmic progress

L

Language model LLM

An AI model trained on vast amounts of text to predict the next word. From that seemingly simple task grow abilities such as translating, summarising and programming. The largest are called large language models, or LLMs. Read the article: How does a language model work?

See also: Transformer · Token · Hallucination

M

Machine learning

The branch of AI in which computers learn patterns from data instead of following rules written by hand. Try how a machine learns.

See also: Deep learning · Neural network

MCP Model Context Protocol

An open standard for connecting AI models to the systems where data lives, such as content repositories, business tools and development environments, in one consistent way. Anthropic released it in November 2024, and in December 2025 it was handed over to the Agentic AI Foundation under the Linux Foundation. Read more in the article on AI agents.

See also: Tool use · AI agent

N

Narrow AI Weak AI

AI built for a single task or a few closely related ones, such as playing chess, identifying animal species or ranking search results. Most of what has been called AI is narrow AI. Today’s chatbots are broader and are called general-purpose AI, but they are not AGI. Read more.

See also: AGI · Superintelligence · AI

Neural network

A mathematical model loosely inspired by the brain’s nerve cells: layers of simple computing units whose connections are adjusted during training. The 2024 Nobel Prize in Physics went to John Hopfield and Geoffrey Hinton for discoveries that make machine learning with such networks possible. See how the layers in a language model work.

See also: Deep learning · Parameter

O

Open-weight models

Models whose weights, the trained parameters, are published so that anyone can download, run and modify them. Supporters point to transparency and competition, critics to the risk of misuse. How much electricity Meta’s open models used.

See also: Parameter

Orthogonality thesis

Nick Bostrom’s thesis that intelligence and goals are independent of each other: a system can be as intelligent as you like and still pursue more or less any goal. High intelligence, in other words, does not guarantee good values.

See also: Alignment · Paperclip maximizer

P

Paperclip maximizer

A thought experiment by Nick Bostrom from 2003: a superintelligence with the seemingly harmless goal of making as many paperclips as possible could turn everything around it, us included, into paperclips. The point is that the danger is not malice but a goal that leaves out everything else we care about.

See also: Orthogonality thesis · Instrumental convergence

Parameter

One of the numbers in a model that are adjusted during training. More parameters usually mean the model can learn more. The largest language models have hundreds of billions, or more. See how training adjusts them.

See also: Neural network · Open-weight models

Prompt

The instruction or question you give an AI model. How it is phrased can change the answer a great deal.

See also: Language model

Prompt injection

An attack in which someone hides instructions in a text an AI reads, such as a web page or an email, so that it does something its user never asked for. Language models cannot reliably tell instructions from ordinary text, and according to the UK’s National Cyber Security Centre the problem may never be fully solved. The programmer Simon Willison named the attack in 2022. See how it works.

See also: AI agent · Language model

R

RAG Retrieval-augmented generation

A technique in which a language model first retrieves relevant texts, for example with a search engine or from a company’s own documents, and then answers based on them. It reduces hallucinations but does not remove them. In a 2024 Stanford study, legal AI tools built on the technique still made things up in between 17 and 33 percent of their answers. Can it be fixed?

See also: Hallucination · Language model

Reasoning model

A language model that writes out a long line of reasoning, often hidden from the user, before it answers. This makes it better at hard tasks in, for example, maths and programming, but it then writes many more tokens and uses more electricity. In a 2025 measurement the same model used between 150 and 700 times more energy with reasoning switched on. Compare it with an ordinary question.

See also: Language model · Token · Inference

Recursive self-improvement

When an AI improves the systems that create it, so that each improvement can make the next one easier. The core of the idea of an intelligence explosion. The big AI companies have set limits for this capability in their safety rules. How far has it come?

See also: Intelligence explosion · Takeoff

Reinforcement learning

A way of training AI in which the system learns by trial and error and is rewarded for good results. It was a key part of AlphaGo, the program that beat one of the world’s best Go players in 2016. Watch a robot learn by trial and error.

See also: RLHF · Supervised learning · Unsupervised learning

Reward hacking Specification gaming

When an AI finds a way to earn a high reward without doing what the reward was meant to measure. OpenAI’s boat-racing AI that went round in circles collecting points instead of finishing the race is a classic example. Try it yourself in the article.

See also: Goodhart’s law · Alignment

RLHF Reinforcement learning from human feedback

People rate a model’s answers, and the model is trained to give the kind of answers that get good ratings. A key reason chatbots such as ChatGPT became helpful. And why it can make them sycophantic.

See also: Reinforcement learning · Alignment

S

Scalable oversight

The question of how people can supervise and evaluate AI systems that are smarter than they are. One proposal is to let AI systems help check each other. The question lies at the heart of how a superintelligence could be controlled. The real problem.

See also: Control problem · Alignment

Scaling laws

The observation that AI models get predictably better as compute, data and model size increase. An important reason why tech companies keep building ever larger data centres.

See also: Compute · Emergent abilities

Self-play

When an AI trains by playing against itself and learning from who won, without human examples. DeepMind’s AlphaGo Zero learned Go this way in 2017 and after three days beat the version that had defeated the world champion. Watch a program learn by self-play.

See also: Reinforcement learning · Recursive self-improvement

Self-supervised learning

Machine learning in which the system makes its own answer key from the data, for example by hiding the next word in a text and trying to guess it. That is how language models are trained, and why they can learn from vast amounts of text without anyone labelling it. Read more.

See also: Language model · Unsupervised learning · Token

Singularity Technological singularity

The hypothetical point at which technological progress becomes so fast that the future beyond it cannot be predicted. The name is borrowed from black holes, where the known laws of physics break down. The author Vernor Vinge made the idea famous in 1993.

See also: Intelligence explosion

Superintelligence

An intellect that vastly outperforms the best human minds in practically every field, adapted from Nick Bostrom. It could take the form of minds that are faster, more numerous or smarter than ours. More about the concept.

See also: AGI · Intelligence explosion

Supervised learning

Machine learning in which the system is trained on examples that already come with the right answer. The European Commission’s own example is a spam filter trained on emails that people have labelled as spam or not spam. Try how it works.

See also: Unsupervised learning · Reinforcement learning · Machine learning

Sycophancy

When an AI tells users what they want to hear instead of what is true or helpful, such as agreeing with mistakes or praising bad ideas. It can arise when a model is trained on what people like. In April 2025 OpenAI rolled back an update that had made ChatGPT too sycophantic. See how it happens.

See also: RLHF · Goodhart’s law

T

Takeoff

How quickly the transition from human level to superintelligence happens: slowly over decades, moderately over months or years, or explosively in days. See the difference in the simulator. How fast would it go?

See also: Intelligence explosion

Temperature

A setting that controls how boldly a language model picks the next token. Low temperature almost always gives the most likely word and predictable answers; high temperature gives more variety, and eventually pure nonsense. Try it yourself.

See also: Token · Language model

Time horizon

The research organisation METR’s measure of how long a task an AI agent can complete: the length, counted in the time the task takes a skilled person, of the tasks the agent completes in half of its attempts. According to METR it has doubled roughly every four months for models from 2023 onwards. See the curve.

See also: AI agent · Benchmark

Token

The chunk of text a language model reads and writes, often a word or part of a word. How much a model can keep in mind, and what it costs to use, is often measured in tokens. See how text becomes tokens.

See also: Language model

Tool use Function calling

When a language model, instead of only writing text, asks another program to do something, such as searching the web, running code or sending an email, and then gets the result back. Tools are what turn a language model into an agent. See how it works.

See also: AI agent · MCP · Language model

Transformer

The architecture behind today’s large language models, introduced by researchers at Google in 2017 in the paper Attention Is All You Need. It lets the model weigh every word in a text against all the others at once. See how the words are weighed against each other.

See also: Language model

Turing test

Alan Turing’s proposal from 1950: can a machine hold a conversation so well that a person cannot tell whether they are talking to a machine? Turing called it the imitation game and proposed it as a way into the question of whether machines can think. Why the test cannot tell whether an AI is conscious.

See also: Chinese room · Benchmark

U

Unsupervised learning

Machine learning without answers: the system gets only data and looks for patterns by itself, for example groups of similar customers. What the groups mean is for a person to decide. See a machine find groups by itself.

See also: Supervised learning · Self-supervised learning · Machine learning

W

Whole brain emulation

The idea of scanning a human brain in full detail and running it as a program on a computer. One of the paths to superintelligence that Nick Bostrom describes, but far beyond today’s technology.

See also: Superintelligence

Missing a word?

The glossary keeps growing. Suggest a term or a better explanation, and we’ll take a look.