AGI Glossary Search
The working vocabulary of AGI — searchable and filterable.
38 terms
Agent
technical
An AI system that takes actions in an environment over multiple steps, often via tools, browsers, or operating systems.
AGI (Artificial General Intelligence)
core
A hypothesised machine system that can match or exceed competent human adults across the full range of cognitive tasks, without being retrained for each one.
AGI race
societal
The competitive dynamic between labs and nations pursuing increasingly capable AI systems.
AI literacy
societal
The set of skills needed to understand, work with, and critically evaluate AI systems in daily life and work.
Algorithmic accountability
societal
The principle that organisations deploying algorithms can be held responsible for the outcomes those algorithms produce.
Alignment
safety
The technical and governance problem of ensuring AI systems pursue goals that match human intent and values, including under distributional shift.
ASI (Artificial Superintelligence)
core
Hypothesised systems significantly exceeding the best human performance across virtually all domains. ASI sits beyond AGI on most road maps.
Catastrophic risk
safety
Risk of large-scale, irreversible harm, including from misuse, misalignment, or accident.
Closed-weight model
research
A model accessed only through an API; the weights themselves are not released.
Co-intelligence
societal
Working pattern in which a person and an AI system collaborate, with the human providing direction and judgment and the AI providing scale.
Cognitive automation
societal
Automation of tasks previously seen as requiring human cognition (analysis, writing, design).
Compute governance
safety
Governance of access to large-scale compute as a lever for AI policy, since frontier training runs require concentrated infrastructure.
Compute overhang
research
Situation where available compute substantially exceeds what current algorithms use efficiently, enabling rapid capability jumps when algorithms catch up.
Differential privacy
safety
A mathematical framework for limiting how much a system reveals about any single individual in a dataset.
Distillation
technical
Training a smaller model to mimic a larger one, transferring much of its capability into a more efficient form.
Emergent capability
research
A capability that appears at scale and was not present at smaller scales, sometimes without explicit training.
Evaluation (eval)
research
Structured tests of model capability or safety on benchmarks or scenarios designed to probe specific behaviours.
Existential risk (x-risk)
safety
Risk that the long-term potential of humanity is permanently curtailed, including extinction.
Foundation model
technical
A large model trained on broad data that can be adapted to many downstream tasks. GPT-class and Claude-class models are foundation models.
Frontier model
research
The most capable AI models at any given time, typically defined by training compute or benchmark performance.
Hallucination
technical
Confident generation of false or fabricated content by a language model.
Human-in-the-loop
safety
System design pattern in which human review or approval is required at decision points.
Inference compute
technical
The compute used to run a model at query time, as distinct from training compute.
Long context
technical
A model's ability to attend to very large input windows (hundreds of thousands or millions of tokens).
Mechanistic interpretability
safety
Research aimed at understanding the internal computations of neural networks at the level of circuits and features.
Model card
safety
Standardised documentation describing a model's intended use, training data, capabilities, and limitations.
Model evaluation suite
research
A standardised collection of benchmarks used to measure model capability or safety.
Narrow AI
core
An AI system designed for a single task or a small set of related tasks. Today's chatbots, translators, and recommendation engines are narrow AI even when powerful.
Open-weight model
research
A model whose trained parameters are released publicly, enabling local deployment and modification.
RAG (Retrieval-Augmented Generation)
technical
A pattern in which a model retrieves relevant documents at query time and uses them to ground its output.
Reasoning model
technical
A model trained or prompted to produce explicit intermediate reasoning steps, often with significant inference-time compute.
Red teaming
safety
Adversarial testing of an AI system to surface harmful or unintended behaviours before deployment.
RLHF (Reinforcement Learning from Human Feedback)
technical
A training method that fine-tunes a model using human preference judgments. Widely used to align large language models.
Scaling laws
research
Empirical relationships between model size, data, compute, and loss; used to predict capability gains from larger training runs.
Sovereign AI
societal
National strategies aimed at maintaining domestic AI capability, compute, and data infrastructure.
Synthetic data
research
Training data generated by other models or simulations rather than collected from humans.
Tool use
technical
The pattern of a model calling external tools (search, code execution, APIs) to extend its capabilities.
Transfer learning
technical
The ability of a model to apply what it learned on one task to a different but related task.