The AGI glossary

Every term you will meet on this site, defined once, properly. Where a term is contested, the entry says so rather than picking a side.

Written by Sarah Wood

Editor · Founder, Creative Sauce AI, which helps businesses put AI to work

Last reviewed 19/09/26

How to use it

Terms are defined here once and linked from everywhere else on the site, so no module has to stop and explain the vocabulary a second time.

Where a term is genuinely contested - AGI itself, emergent ability, agent - the entry says so rather than quietly picking the definition that makes the sentence easier to write.

Terms

A

AgentA system given a goal rather than a single instruction, which then takes multiple steps - often using tools - to pursue it. The word is doing very little quality control: a two-step script and a system running unattended for hours are both called agents.

AGIArtificial general intelligence. Contested - see Module 1, which sets out five published definitions that require different evidence.

AlignmentThe problem of getting a system to pursue what was intended rather than what was literally specified or what it inferred from training. Distinct from capability: a more capable misaligned system is worse, not better.

ASIArtificial superintelligence. A hypothetical system substantially beyond human capability across essentially everything. A separate claim from AGI and a much stronger one.

B

BenchmarkA fixed set of problems with known answers, used to score systems. See Module 2 for the six ways the score and the headline come apart.

C

Capability overhangThe gap between what a system can already do and what anyone has worked out how to make it do. Relevant because capability can appear to jump without the underlying system changing.

Chain of thoughtProducing intermediate reasoning steps before an answer. Improves performance on multi-step problems. The written steps are not necessarily a faithful account of how the answer was actually reached.

ComputeProcessing used to train or run a system, usually measured in floating-point operations. One of the few quantities in this field that can be stated precisely, which is part of why it features so heavily in forecasts.

Construct validityWhether a test measures the thing it claims to. The central question about every benchmark and the one least often asked.

ContaminationTest questions appearing in training data, so a score measures recall rather than capability.

D

Distribution shiftThe gap between the conditions a system was measured in and the conditions it is used in. The usual explanation for a system that benchmarks well and disappoints in practice.

E

Emergent abilityA capability that appears at larger scale without having been trained for specifically. Genuinely contested: some apparent emergence is an artefact of using a scoring metric that is all-or-nothing.

Evaluation (eval)A structured test of a system’s behaviour. Broader than a benchmark: includes safety testing, red-teaming and task-based assessment.

F

Fine-tuningFurther training of an existing model on a narrower dataset to specialise it.

Foundation modelA large model trained broadly and then adapted to many downstream uses, rather than built for one.

Frontier modelAmong the most capable systems in existence at a given moment. A moving description, not a category.

G

GeneralityThe range of tasks over which a system is useful. One of the two axes in the levels framework, and the property the word AGI is named for.

Goodhart’s lawWhen a measure becomes a target it stops being a good measure. The reason benchmark scores decay in usefulness as they gain importance.

H

HallucinationProducing fluent, confident, false content. Not lying: the system has no separate representation of what is true to depart from.

I

InferenceRunning a trained model to get an output, as opposed to training it. Most running costs are inference costs.

InterpretabilityResearch into what is actually happening inside a model. Relevant to AGI because most capability claims are currently behavioural - we observe what a system does, not how.

L

Long-horizon taskWork requiring many dependent steps over an extended period. Currently one of the sharpest limitations, because small per-step error rates compound.

M

MultimodalWorking across more than one kind of input or output - text, images, audio, video.

O

Open weightsA model whose trained parameters are published, so anyone can run or adapt it. Not the same as open source: the training data and code usually are not included.

P

ParametersThe learned numerical values inside a model. A rough and increasingly unreliable proxy for capability.

Post-trainingEverything done after pre-training to make a model useful and well-behaved. A large share of the difference between a raw model and a product.

Pre-trainingThe initial large-scale training phase on broad data, before any specialisation.

R

Reasoning modelA model trained to spend additional computation working through a problem before answering. The extra computation is the mechanism, and it costs time and money per answer.

Reinforcement learningTraining by reward signal rather than by labelled examples. Underlies much post-training work.

ReliabilityHow consistently a system performs across repeated attempts and varied conditions. Distinct from capability, and the gap between the two is the practical story of the last few years.

S

SaturationA benchmark approaching its ceiling, so it no longer distinguishes between systems.

Scaling lawsObserved regular relationships between compute, data, model size and performance. Empirical regularities, not physical laws, and the debate is over how far they extend.

Situational awarenessA system’s representation of its own circumstances - that it is being evaluated, for instance. Raises the question of whether test behaviour predicts deployed behaviour.

System cardA document published alongside a model setting out evaluations, limitations and risks. Usually the most informative thing released on launch day.

T

Tool useA system calling external software - search, code execution, a database - rather than answering from its own parameters alone.

Training cutoffThe date after which a model has seen no data. Relevant to contamination checks and to what a system can know.

TransferApplying what was learned in one setting to a genuinely different one. Central to the learning definition of AGI in Module 1.

TransformerThe neural network architecture, introduced in 2017, underlying most current systems.

Last reviewed 19/09/26

Enrol early

Get the modules as the course develops

We will email you when each module publishes. You will also get AGI, Made Clear, keeping you up to date on AGI and ASI developments in a field that moves faster than anyone can follow.

Double opt-in. One-click unsubscribe. We never share your address.

Looking for the news and the arguments?

AGI Made Clear covers what is happening, the evidence behind it and the claims that are not supported. This site teaches the subject. That one tracks it.

Visit AGI Made Clear →