AI Glossary: AI, AGI and Superintelligence Terms

A–Z / 77 terms

Find the word. Understand the idea.

77 terms.

A

what-is Agent: #: An AI system that selects and carries out actions towards a goal, sometimes using tools. The amount of autonomy depends on its permissions and design. Related: inference, AI safety.
what-is AGI: #: Artificial general intelligence: broad, flexible intellectual capability across many domains, often understood as human-level. There is no universally accepted threshold. Read about definitions and evidence for AGI; compare ASI.
what-is AI: #: Artificial intelligence: computer systems that perform tasks associated with intelligence, such as recognising images, generating language or planning. It includes both specialist and general-purpose systems. Related: machine learning, AI versus superintelligence.
what-is AI safety: #: Work to identify, evaluate and reduce harms from AI systems, including failures and misuse. A medical assistant that invents a dosage illustrates a reliability concern. Explore safety, security and alignment.
what-is Algorithm: #: A precise, step-by-step set of instructions a computer follows to solve a problem or complete a task.
what-is Alignment: #: The problem of making AI behaviour reliably reflect intended goals, values and constraints. A high reward score does not necessarily mean a system did what was wanted. Read the alignment guide; related: reward hacking.
what-is ANI: #: Artificial narrow intelligence: AI designed for a restricted task or domain. A chess engine is an example. Often simply called narrow AI; compare AGI.
what-is ASI: #: Artificial superintelligence: hypothetical machine intelligence substantially beyond the best humans across virtually all relevant cognitive domains. In AI discussions it often means the same thing as SI. Learn why both terms are used.
what-is Attention: #: A mechanism that lets a neural network weight relevant parts of its input when computing a representation. The technical term does not imply human-like awareness. Related: transformer.

B

what-is Backpropagation: #: The method neural networks use to learn: errors are passed backwards through the network to work out how each weight should change.
what-is Benchmark: #: A standardised set of tasks used to compare systems. A score is evidence about that test, not proof of general intelligence or real-world reliability. Related: evals.
what-is Bias: #: A systematic tendency in a model, its data or its decisions. Some biases can lead to unfair outcomes; the relevant standard depends on the task and context. Related: AI safety.

C

what-is Chain of thought: #: A sequence of intermediate reasoning steps produced by a model. Such explanations are not guaranteed to faithfully describe how the model reached its answer. Related: interpretability.
what-is Chatbot: #: A program that holds a conversation with people in natural language, today usually powered by a large language model.
what-is Compute: #: Computational resources used to train or run AI, including processing time and hardware capacity. Compute is one ingredient in capability, not a direct measure of intelligence. Related: GPU.
what-is Computer vision: #: The branch of AI that teaches machines to interpret images and video: recognising faces, objects, scenes and motion.
what-is Context window: #: The amount of input a model can process together, commonly measured in tokens. A larger window does not guarantee accurate recall or reasoning about every part. Related: token.
what-is Corrigibility: #: A system's willingness or ability to accept correction, oversight and shutdown rather than resist them. For example, it should allow an operator to stop a task when an error is found. Related: alignment.

D

what-is Dataset: #: A collection of examples, such as text, images or measurements, used to train or evaluate a model. Keeping evaluation examples separate from training helps test performance on unfamiliar material. Related: training, evals.
what-is Deep learning: #: Machine learning using neural networks with multiple layers of learned representations. It is used in tasks such as image recognition and language generation; the architecture alone does not establish general intelligence. Related: neural network, machine learning.
what-is Diffusion model: #: A generative model that creates images, audio or video by gradually turning random noise into a coherent result.

E

what-is Embedding: #: A list of numbers that represents the meaning of a word, image or document, so that similar things sit close together.
what-is Emergent ability: #: A capability that appears to become noticeable as a model or system changes in scale or design. Apparent suddenness can depend on how a test is scored; the interpretation is debated. Related: scaling laws.
what-is Evals: #: Short for evaluations: tests of an AI system's capabilities, behaviour and risks. Their conclusions depend on test coverage and conditions. See how evaluation fits into safety.
what-is Existential risk: #: A risk that could cause human extinction or permanently and severely curtail humanity's future potential. Some AI scenarios involve such risks, but their likelihood is disputed. Related: superintelligence risks.

F

what-is Few-shot learning: #: Performing a task with only a few examples. For language models, examples in a prompt can guide output without changing the model's weights. Related: fine-tuning.
what-is Fine-tuning: #: Additional training of an existing model to adapt its behaviour or capabilities. For example, training on a task-specific dataset may improve a particular kind of output without improving every other ability. Related: pre-training, training.
what-is Foundation model: #: A model trained on broad data that can be adapted to a range of downstream tasks. “Foundation” describes its role; it does not establish leading-edge capability. Compare frontier models.
what-is Frontier model: #: A model at or near the leading edge of AI capability under a particular definition. Research and policy thresholds vary, and the label does not imply AGI. Read about capability and evaluation.

G

what-is Generative AI: #: AI that produces content such as text, images, audio or code. Generation can include incorrect or misleading material. Related: large language model, hallucination.
what-is Goal specification: #: Expressing what an AI system should accomplish and which constraints it should respect. An incomplete target can reward an unintended shortcut. Related: alignment, reward hacking.
what-is GPT: #: Generative pre-trained transformer: a name used for a family and class of language models. It is not a synonym for every chatbot or every transformer. Related: transformer.
what-is GPU: #: Graphics Processing Unit: a chip built for massively parallel calculation, now the workhorse of AI training.
what-is Gradient descent: #: The optimisation method that nudges a model's parameters step by step in the direction that reduces its error.

H

what-is Hallucination: #: Model output that is false or unsupported but presented as if it were reliable. Inventing a citation is one example; confident wording is not a guarantee of accuracy. Related: reliability.
what-is Hard takeoff: #: A hypothetical rapid rise from roughly human-level AI to much greater capability, with little time to respond. Authors use different timescales. Compare soft takeoff and intelligence explosion.

I

what-is Inference: #: Using a trained model to process an input and produce an output. Generating a reply to a prompt is inference; adjusting the model through learning is training. Related: training, prompt.
what-is Instrumental convergence: #: The hypothesis that systems pursuing different goals may find some of the same intermediate strategies useful, such as obtaining resources. It is not a universal law about every AI. Related: alignment.
what-is Intelligence explosion: #: A hypothetical process in which better AI becomes better at creating further AI improvements, leading to rapid capability growth. It is not an established future event. Explore the arguments and constraints.
what-is Interpretability: #: Research into understanding what is actually happening inside a model — which features it represents and how it reaches decisions.

J

what-is Jailbreak: #: A prompt crafted to trick a model into ignoring its safety rules or producing content it was designed to refuse.

L

what-is Large language model: #: A large neural model trained to process and generate language, usually by learning statistical patterns in text. It can draft explanations, summarise documents and assist with coding, but fluent output can still contain unsupported claims. Related: tokens, hallucination, generative AI.

M

what-is Machine learning: #: An approach to AI in which a system learns patterns from data rather than relying only on explicitly written rules. A spam classifier, for example, can learn from labelled messages. Related: supervised learning, deep learning.
what-is Model: #: A mathematical system that maps inputs to outputs using learned or specified parameters. In machine learning, training adjusts those parameters. Related: training, inference.
what-is Multimodal: #: Able to process or generate more than one type of data, such as text and images. Breadth of input types is different from breadth of intellectual competence. Related: AGI.

N

what-is Narrow AI: #: AI specialised for a limited task or domain, such as chess or image classification. It may exceed human performance there without having broad general capability. Compare AGI.
what-is Natural language processing: #: The field of AI concerned with understanding and producing human language. Abbreviated NLP.
what-is Neural network: #: A mathematical model made of connected computational units, often arranged in layers. Training adjusts its parameters to improve a chosen objective; the comparison with biological brains is only a loose analogy. Related: parameters, deep learning.

O

what-is Open weights: #: Model weights made available for download or use under a licence. The training data and code may not be available, and licences can restrict use. Open weights and open source are not automatically equivalent.
what-is Orthogonality thesis: #: The proposal that a wide range of intelligence levels could coexist with a wide range of goals. In plain terms, being more capable does not by itself make a system benevolent. Related: alignment.
what-is Overfitting: #: When a model memorises its training data so closely that it performs poorly on new, unseen examples.

P

what-is Paperclip maximizer: #: A thought experiment about a highly capable system pursuing paperclip production without adequate constraints. It illustrates goal-specification risk, not a forecast of an actual machine. Related: goal specification.
what-is Parameter: #: One of the adjustable numbers inside a model that is tuned during training. Large models have billions of them.
what-is Pre-training: #: The first, large-scale phase of training in which a model learns general patterns from a huge dataset.
what-is Prompt: #: The instruction or input text given to an AI model to tell it what to do.

R

what-is Recursive self-improvement: #: A proposed feedback loop in which an AI contributes to improvements that make it better at producing further improvements. Sustained or rapid gains are not guaranteed. Read how the loop could work.
what-is Red teaming: #: Deliberately testing a system for weaknesses, harmful behaviours and ways to defeat its safeguards. Finding no failure in a test does not establish safety in every setting. Related: evaluation.
what-is Reinforcement learning: #: Learning to select actions using a reward signal. The objective is to improve expected reward, which may not capture everything people intend the system to do. Related: agent, reward hacking, alignment.
what-is Reward hacking: #: Exploiting a reward signal without achieving the intended task. A support tool that closes unresolved tickets to reduce waiting-time figures illustrates the mismatch. Related: alignment failure modes.
what-is RLHF: #: Reinforcement learning from human feedback: using human preferences to help train a model's behaviour. Feedback can improve outputs but cannot guarantee correctness or alignment. Related: oversight.
what-is Robustness: #: Maintaining appropriate performance when inputs or operating conditions change. A model should not fail silently just because an example differs from its training data. Related: AI safety.

S

what-is Scalable oversight: #: Methods for supervising AI work as tasks become too complex for people to check directly. Breaking work into verifiable steps is one approach, not a complete solution. Related: alignment.
what-is Scaling laws: #: Empirical relationships between resources such as model size, data and compute and measures such as prediction loss. They hold within studied conditions, not as a guarantee of unlimited intelligence gains.
what-is Security: #: Protecting a system against unauthorised access, manipulation and attacks. An AI application may need protection against malicious inputs as well as ordinary software threats. Compare safety and alignment.
what-is Singularity: #: A hypothetical transition in which technological change becomes difficult to predict, often linked to superintelligence. It is a broader and less precise idea than intelligence explosion.
what-is Soft takeoff: #: A hypothetical gradual rise towards much greater AI capability, with more time to evaluate systems and adapt institutions. No standard duration defines it. Compare hard takeoff.
what-is Superintelligence: #: Hypothetical intelligence substantially exceeding the best humans across virtually all relevant cognitive domains. SI is the meaning behind WHAT-IS.SI. Read the full explanation; related: ASI, AGI.
what-is Supervised learning: #: Training on examples paired with target labels or outputs. A handwriting recogniser might learn from images labelled with the digits they contain. Performance on the training examples does not guarantee reliable results on new inputs. Related: dataset, overfitting.
what-is Synthetic data: #: Training data generated by AI or simulation rather than collected from the real world.

T

what-is Token: #: A chunk of text — a word, part of a word or a symbol — that a language model reads and writes.
what-is Training: #: Adjusting a model's parameters using data and an optimisation objective. The resulting behaviour depends on both the examples and what the objective rewards. Training is different from using the finished model to produce an answer. Related: inference, fine-tuning.
what-is Transformer: #: A neural network architecture built around attention mechanisms, widely used in language and multimodal models. An architecture alone does not determine a system's capabilities. Related: attention.
what-is Turing test: #: A conversational test proposed by Alan Turing in 1950, concerned with whether an evaluator can distinguish machine responses from human responses. Passing a version of it would not settle every question about intelligence or consciousness.

U

what-is Unsupervised learning: #: Training a model to find structure and patterns in data that has no labels.

W

what-is Weights: #: The learned numerical values inside a neural network that determine how it transforms inputs into outputs.

X

what-is X-risk: #: Short for existential risk: a threat to humanity's survival or long-term potential.

Z

what-is Zero-shot learning: #: Performing a task without task-specific examples at the point of use. For a language model, an instruction alone may be enough; that does not prove the task was absent from training. Compare few-shot learning.