Reference
AGI glossary
Short, neutral definitions of the words that carry most of the weight in AGI debates.
- Agent
- An AI system that takes actions toward a goal, often using tools, over several steps, rather than only answering a single prompt.
- AGI (artificial general intelligence)
- A loosely defined goal: an AI system that can perform a broad range of tasks at or above human level, rather than one narrow task. There is no single agreed definition or test. See What people mean by AGI.
- Alignment
- Research and practice aimed at making AI systems pursue the goals and values their designers and users intend.
- Benchmark
- A fixed set of test tasks with a scoring method, used to compare systems. Useful for comparison, but a high score shows performance on that test, not general ability. See How to read an AI benchmark claim.
- Benchmark saturation
- When top systems score near the maximum, so the benchmark can no longer tell them apart.
- Contamination (data leakage)
- When test questions, or close copies, appear in a model's training data, inflating its score.
- Emergent ability
- A capability that appears in larger models but not smaller ones. Researchers debate how much apparent emergence depends on how performance is measured.
- Evaluation (eval)
- Any structured test of a model's behaviour, including capability, safety, and reliability tests. Broader than a single benchmark.
- Frontier model
- An informal term for the most capable AI models available at a given time.
- Hallucination
- When a model states something false or fabricated with apparent confidence.
- Large language model (LLM)
- A neural network trained on very large amounts of text to predict and generate language. Most current chat assistants are built on LLMs.
- Narrow AI
- A system built for a specific task or domain, such as recognising speech or playing chess.
- Red teaming
- Deliberately trying to make a system fail or behave harmfully, to find problems before release.
- Superintelligence
- A hypothetical system that greatly exceeds the best human performance across almost all domains. Usually treated as distinct from, and beyond, AGI.
- System card / model card
- A document published with a model describing its intended uses, evaluations, limitations, and risks.