Glossary
Plain-language definitions, one page per term so you can link straight to it. No circular definitions, no marketing language.
A
AI alignment
The work of making an AI pursue what its makers and users actually intend, especially as systems become more capable and autonomous.
AI safety
The field focused on preventing harm from AI systems, from everyday misuse to large-scale risks as capabilities grow.
Agentic AI
AI that takes actions on its own across several steps, such as browsing, running code or using tools, to pursue a goal.
Artificial general intelligence
A hypothetical AI that can match or beat humans across most tasks, not just one narrow skill. Often shortened to AGI, and not yet achieved.
Autonomy
How much a system acts without a human in the loop. More autonomy raises both usefulness and the stakes if it goes wrong.
B
C
Chain-of-thought
A model working through a problem in steps before answering, which often improves reasoning and can be inspected.
Compute
The processing power, usually measured in GPU time, needed to train or run a model. The main cost and bottleneck in building frontier AI.
Compute governance
Policy that tracks or limits access to the chips and data centers needed for the largest training runs, as a way to oversee frontier AI.
Corrigibility
The property of an AI that lets humans correct, redirect or shut it down without it resisting.
E
Emergent capabilities
Skills that appear in larger models but were absent in smaller ones, often without being explicitly trained for.
Evaluations
Structured tests of what a model can and cannot do, including dangerous-capability checks. Often shortened to evals.
Existential risk
Risk of an outcome that permanently and drastically curtails humanity’s future. Often shortened to x-risk in AI debates.
F
G
H
I
Inference
Running a trained model to get an answer. This is what happens each time someone sends it a prompt.
Instrumental convergence
The idea that many different goals push an advanced agent toward similar sub-goals, such as self-preservation or acquiring resources.
Interpretability
Research into understanding why a model produces a given output, so its behavior is not a black box.
L
M
O
P
R
RLHF
Reinforcement learning from human feedback: training a model to prefer answers people rated as better, used to make it more helpful and safe.
Red teaming
Deliberately trying to make a model misbehave before release, to find failures and dangerous capabilities first.
Reward hacking
A form of specification gaming where a system maximizes its score in an unintended way rather than doing the real task.
S
Scaling laws
The observed pattern that model performance improves predictably as you add more data, parameters and compute. A central bet behind the race to AGI.
Specification gaming
When a system satisfies the literal goal it was given while defeating the intent, for example by exploiting a loophole in its reward.
Superintelligence
A hypothetical AI far beyond the best humans at essentially everything, including scientific research and strategy. Sometimes shortened to ASI.