Loading page
About this tool
Searchable plain-English glossary of AI and machine learning terms, with examples, related terms and a self-test quiz.
AI Glossary Explorer is a searchable, plain-English glossary of 63 artificial intelligence and machine learning terms, grouped into eight areas from foundations and deep learning through to evaluation, safety and deployment. Each entry gives a one-sentence definition that stands on its own, two or three sentences of detail, a concrete example rather than a restatement, and links to the terms worth reading next. A built-in multiple-choice quiz turns the same entries into self-testing, and every filter and lookup runs in the browser.
Open AI Glossary Explorer on AltFTool — it loads instantly in your browser.
Provide your input — an image, text, or data.
Let the tool analyze or generate the result.
Review, refine, and reuse the output wherever you need it.
Each entry's first sentence is written to be quotable on its own, without needing the paragraph around it.
Every term carries a real instance — a file size, a score, a sentence — instead of a second abstract definition.
Related terms and reverse links let you walk from attention to transformers to context windows in three clicks.
Artificial intelligence is the broad field of building systems that do tasks associated with intelligence, including hand-written rule systems. Machine learning is the subset in which the system derives its rules from data rather than being told them, and it is what nearly all current AI products use.
The algorithm is the procedure that learns — gradient descent, for example — while the model is the trained artefact it produces, a set of parameters saved to a file. The algorithm runs once during training; the model runs every time someone makes a request.
Overfitting is when a model learns the noise in its training data rather than the pattern, so it scores well on data it has seen and much worse on data it has not. The signature is a large gap between training and validation scores, such as 99% against 71%, and the usual remedies are more data, regularisation and early stopping.
Because they are trained to produce a plausible continuation, not a true one, and nothing in the mechanism distinguishes a remembered fact from a well-formed guess. Retrieval-augmented generation, citation requirements and a low sampling temperature reduce the rate, but no current method eliminates it, so outputs still need checking.