A working reference for AI terms you will see in product briefings, model cards, and technical conversations. Each entry starts plain, then explains what is actually happening under the hood.
AI designed to take a sequence of steps toward a goal, rather than only answer a single prompt.
An agent loops through a model, instructions, tools, and a working state. The model proposes the next action; software executes it and returns the result for another pass.
A defined way for software to ask another service for data or an action.
An API is usually an HTTP request with an authenticated key and structured input. Model APIs send text, files, or tool definitions and return structured model output.
A hypothetical AI that can learn and perform a broad range of intellectual tasks at a human-like level.
There is no accepted technical test for AGI. Current systems are specialized statistical models, even when they look broadly capable in conversation.
A standardized test used to compare models on a defined task.
A benchmark fixes a dataset, scoring rule, and often a prompt format. It measures performance on that setup, not every real-world use case.
The amount of text, code, or other input a model can consider in one request.
Inputs are converted to tokens and placed in the model's attention computation. A larger window raises compute and does not guarantee the model uses every detail equally well.
Related explainer: Context windows, without the marketing fog
A model that creates images, audio, or video by gradually turning noise into a coherent result.
Training teaches the system to predict and remove noise. At generation time it repeats that denoising process, steered by a prompt or reference image.
A numeric representation of text, images, or other data that captures useful similarity.
An encoder maps an item to a vector: a long list of numbers. Nearby vectors tend to represent related meanings, which makes similarity search practical.
Further training a base model on selected examples to make it better at a narrow job or style.
Training updates some or all of the model's weights using new examples. It can improve consistency, but it does not reliably add fresh factual knowledge.
A broadly trained model that can be adapted to many downstream tasks.
It is pretrained on a large, varied corpus to learn reusable patterns. Prompting, fine-tuning, or tool use then specializes it for a product task.
A model output that sounds plausible but is inaccurate, unsupported, or invented.
A language model predicts likely next tokens; it does not check claims against reality by default. Retrieval, tools, and verification workflows reduce this risk but do not erase it.
The process of running a trained model to produce an answer, prediction, or generated file.
The model's fixed weights transform input tokens through many numerical operations. For language generation, this runs repeatedly, one token at a time.
An AI model trained to understand and generate language by predicting likely sequences of tokens.
An LLM is typically a transformer neural network trained on massive text collections. It turns words into tokens, computes relationships with attention, and predicts a probability distribution for the next token.
A model that can work with more than one kind of input or output, such as text, images, audio, or video.
Separate encoders or shared representations convert each modality into compatible signals. The model learns connections between them during training.
Designing instructions and examples so a model reliably produces useful output.
Prompts set the context the model uses to predict its next tokens. Clear constraints, examples, and structured outputs narrow the range of likely responses.
A method that gives a model relevant source material at answer time.
Software searches a document index, usually with embeddings and keyword matching, then places selected passages in the model's context window before generation.
A small unit of text a language model reads and generates; it may be a word, part of a word, or punctuation.
A tokenizer converts text into integer IDs from a fixed vocabulary. Model pricing, context limits, and output speed are usually measured in these units.
The neural-network architecture behind most modern large language models.
Its attention layers let each token weigh the relevance of other tokens in the input. Stacked layers build increasingly abstract representations before predicting output.
A database optimized for finding items whose embeddings are mathematically similar.
It stores vectors in an index that approximates nearest-neighbor search, trading a little exactness for speed at large scale.