Why This Glossary Exists
AI moves fast. New terms appear in product announcements, research papers, and developer forums almost every week. If you have ever read a blog post about AI and hit a wall of acronyms - RAG, LoRA, RLHF, embeddings, fine-tuning - you know how quickly the jargon can get in the way of actually understanding what is going on.
This glossary is our attempt to fix that. We define terms in plain language first, then give you the technical context if you want to go deeper. You do not need a machine learning background to use it. We wrote it for product people, marketers, founders, and developers who are working with AI tools without necessarily having studied the underlying research.
We update the glossary regularly as new concepts enter common usage. If you think a term is missing or a definition could be clearer, let us know via the contact page.
Definitions written for humans, not researchers.
New terms added as the field evolves.
214+ terms across every letter.
Useful for beginners and experienced practitioners.
10 terms
Showing all terms starting with B
A standardised test or dataset used to evaluate and compare the performance of AI models across specific tasks.
Bidirectional Encoder Representations from Transformers - a Google language model that reads text in both directions, excelling at understanding context.
Systematic errors in model outputs caused by skewed training data or flawed design, leading to unfair or inaccurate results.
An AI system whose internal decision-making process is opaque or not interpretable, even to its developers.
The algorithm used to train neural networks by calculating gradients of the loss function and propagating them backward through the network layers.
A training technique that normalises the inputs of each layer to speed up training and improve the stability of deep neural networks.
The number of training samples processed together in one forward and backward pass during model training, affecting speed and convergence.
A decoding algorithm used in text generation that maintains the top-k most probable sequences at each step to find a near-optimal output.
A probabilistic approach for hyperparameter tuning that uses past evaluation results to intelligently choose the next set of parameters to try.
A tokenisation algorithm that iteratively merges the most frequent character pairs, used in GPT and other LLMs to build their vocabulary.
Commonly Misunderstood AI Terms
These are the terms that come up in almost every AI conversation but are often used loosely or incorrectly. Worth knowing what they actually mean.
A type of AI trained on massive amounts of text to predict and generate language. Models like GPT-4, Claude, and Gemini are all LLMs. The "large" refers to the number of parameters - the internal settings the model learned during training.
When an AI generates information that sounds confident and plausible but is factually incorrect. It is not lying - the model genuinely does not know what it does not know. This is one of the most important limitations to understand when using AI-generated content.
The practice of writing inputs to AI models in ways that produce better outputs. A well-structured prompt gives the model context, a clear task, constraints, and sometimes examples. It is less about magic words and more about being precise about what you want.
A technique where an AI retrieves relevant documents or data before generating a response, rather than relying solely on what it learned during training. RAG is how tools like Perplexity AI give you up-to-date answers with citations.
Taking a pre-trained model and training it further on a specific dataset to specialise its behaviour. A fine-tuned model for legal documents will behave differently from the same base model fine-tuned on medical records, even though they started from the same foundation.
A way of representing text, images, or other data as numbers in a high-dimensional space, so that similar things end up close together mathematically. Embeddings power semantic search, recommendation systems, and many RAG implementations.