AI Key Terms Glossary

A quick-reference glossary of core terms that recur across almost any discussion of modern AI systems.

Core Vocabulary

These terms show up constantly in research papers, product documentation, and everyday conversations about AI. Knowing them precisely makes it much easier to follow more advanced material without getting lost in jargon.

TermDefinition
ModelThe overall system, including its structure and learned settings, that takes an input (such as text) and produces an output (such as a prediction or generated text).
ParametersThe internal numeric values inside a model, adjusted during training, that determine how it transforms inputs into outputs. Large models are often described by how many parameters they have, sometimes in the billions.
TrainingThe process of repeatedly adjusting a model's parameters using example data so its outputs get closer to a desired result over time.
InferenceThe process of running an already-trained model on new input to produce an output, this is what happens each time a user actually uses the model.
PromptThe text or instructions given to a model as input to elicit a particular kind of response.
Fine-tuningAdditional, more focused training applied to an already-trained model using a smaller, specialized dataset, so it performs better on a narrower task or domain.
EpochOne complete pass through the entire training dataset during the training process.
Inference latencyThe amount of time a model takes to produce a response after receiving an input.
Context windowThe maximum amount of text a model can take into account at once when generating a response, usually measured in tokens.

Digging a Bit Deeper

A model and its parameters are closely linked but not the same thing: the model is the overall structure and behavior, while the parameters are the specific numbers learned during training that make one trained model different from another with the same structure.

Training, fine-tuning, and inference are three distinct phases in a model's lifecycle, and it helps to keep them clearly separated.

  • Training happens first, often over days or weeks, using large general-purpose datasets to build a broadly capable base model.
  • Fine-tuning happens afterward, usually much faster, to adapt that base model toward a specific task, style, or domain.
  • Inference happens continuously after deployment, every single time a user sends a prompt and receives a response.

Context Window and Latency in Practice

A context window acts like a limited working memory: once a conversation or document exceeds it, earlier content can no longer be directly considered by the model when forming a response, even if it was part of the same session.

Note: These terms interact in practice: a longer context window generally means more text for the model to process per request, which can increase inference latency, so product design often has to balance the two.

Exercise: AI Terminology

What does the term "algorithm" refer to in AI?