AI Glossary
Plain definitions of the terms that keep coming up, each one linked to the explainer that covers it properly.
A
- Agent harnessscaffoldingThe code around a model that defines its tools, enforces limits, manages context and handles failure.
- AI agentA model given tools and a goal, running in a loop and choosing its own next action until it judges the job done.
- Attentionself-attentionThe mechanism letting each token weigh how much every other token matters to it.
C
- CalibrationWhether a model’s stated confidence matches how often it is actually right.
- Chain of thoughtCoTWorking through intermediate steps before answering, which measurably improves accuracy on multi-step problems.
- ChunkingSplitting documents into pieces before embedding them. It is usually the real determinant of retrieval quality.
- Context windowThe total amount of text a model can consider at once: its entire working memory.
E
F
G
H
- HallucinationA model stating something false with confidence: a predictable consequence of how models are trained and graded.
- HBMhigh-bandwidth memory, HBM3E, HBM4High-bandwidth memory: stacks of memory chips placed right beside an AI chip to feed it data very quickly.
- Hybrid searchCombining semantic vector search with traditional keyword search, because each fails where the other works.
I
L
M
N
O
P
- ParameterweightA single learned number inside a model. Collectively, what the model knows.
- Post-trainingalignmentThe smaller training phase that shapes a model’s behaviour rather than its knowledge.
- PretrainingThe first and largest training phase, where a model learns from raw text by predicting what comes next.
- Prompt cachingReusing the processed form of a repeated prompt prefix so you are not charged full price for it again.
- Prompt injectionText placed where a model will read it, written to redirect the model’s behaviour.
Q
R
- RAGretrieval-augmented generationSearching a corpus for relevant material and pasting it into the prompt so the model can answer from it.
- Reasoning modelA model trained to work through a problem internally before committing to an answer.
- RerankingRetrieving generously, then reordering the candidates with a slower, more accurate model.
- RLHFreinforcement learning from human feedbackTraining a model toward responses humans rate highly, rather than toward text that merely looks likely.
S
- SAE levelsSAE J3016, levels of driving automationThe 0-5 framework defining how much of the driving task a system performs, and who is responsible when it fails.
- Sensor fusionCombining cameras, lidar and radar so no single sensor’s blind spot becomes the system’s blind spot.
- Sim-to-realsim2real, simulation-to-reality transferTraining a robot in a computer simulation and then transferring what it learned to a real robot in the real world.
- System promptThe standing instruction placed before a conversation that sets a model’s role, constraints and format.
T
- Task horizonHow long a task, measured in human working time, a model can complete at a given success rate.
- Teleoperationremote operation, teleopA person controlling a robot remotely, often to record demonstrations the robot can later learn from.
- Test-time computeinference-time computeSpending more computation at the moment a question is asked, rather than only during training.
- TokenThe unit of text a model actually reads and writes, usually a word fragment, not a word.
- TokenisationSplitting text into tokens using a vocabulary learned from a training corpus.
- Tool usefunction callingLetting a model request an action (a search, a calculation, a database query) that your code then executes.
- TOPStera operations per secondTrillions of operations per second, a headline measure of how much arithmetic an AI chip can do.
- TransformerThe neural network architecture underneath essentially every modern AI model, introduced in 2017.
V
W
No terms match that filter.