Skip to content
AI Stack Nest

Glossary

Everything to pick, install and run open models on your own computer.

Glossary

Plain-language meanings for the words you meet when you start with open models, such as parameters, quantization, context window and tokens. Read the list once from top to bottom, or search for a single term when you get stuck.

63 terms. Jump to:

AI bias
Systematic errors in model output that unfairly favour or disadvantage certain groups, often reflecting patterns in the training data. Bias can appear in how a model describes people, what advice it gives, or whose language and norms it treats as default.
AI safety
A research field focused on ensuring AI systems behave as intended and do not cause unintended harm as they become more capable. It covers both near-term practical risks and longer-term questions about how to control increasingly powerful systems.
API
A way for programs to send requests to a service, for example to ask a model a question from code.
Adversarial input
Input designed to cause a model to produce an incorrect or harmful output by exploiting weaknesses in how it processes text or other data. Adversarial examples are used both to attack systems and to test their robustness.
Agent
An AI system that can take a sequence of actions using external tools, rather than only generating a single text response. Agents decide what step to take next, take it, observe the result, and continue until a goal is reached.
Agentic loop
The repeated cycle in which an AI agent takes an action, observes the result, and decides what action to take next. The loop continues until the task is complete or the agent determines it cannot proceed without help.
Annotation
The process of labelling data so that a model can learn from human judgments. For example, marking sentences as positive or negative, or rating which of two responses is more helpful. Annotation quality directly affects model quality.
Attention mechanism
A part of transformer models that helps the model focus on the most relevant parts of an input when generating each word of the output. It computes how much each word should influence every other word in the sequence.
Benchmark
A standardised test used to measure and compare the performance of different AI models on specific tasks. Benchmark results give a rough sense of relative capability but do not always reflect real-world usefulness.
Chain of thought
A prompting approach that asks the model to work through a problem step by step before giving a final answer. Writing out intermediate reasoning steps often leads to more accurate results on tasks involving logic or multi-step calculation.
Checkpoint
A saved snapshot of a model's state during or after training. Checkpoints allow training to be resumed after an interruption and allow specific versions of a model to be shared or deployed.
Completion
The act of an AI model generating text that continues from, or responds to, the input it received. The term comes from early language model interfaces where the main use case was completing a partial sentence or document.
Constitutional AI
A training approach that uses a set of written principles to guide a model toward helpful and harmless behaviour, often without requiring human ratings for every individual example. The model is trained to critique and revise its own outputs against the principles.
Context window
The maximum amount of text, measured in tokens, that a model can take into account at one time.
Credit
A unit some AI tools use to meter usage, for example per character of generated speech.
Data poisoning
An attack where harmful or misleading data is introduced into a training dataset to manipulate a model's behaviour. Poisoned data can cause a model to produce incorrect outputs or behave in unintended ways on specific inputs.
Distillation
A process where a smaller model is trained to mimic the behaviour of a larger one. The goal is to produce a more efficient model that retains much of the larger model's performance, making it cheaper and faster to run.
Embedding
A way of representing words, sentences, images, or other data as lists of numbers so that similar items end up close together in mathematical space. Embeddings power semantic search and retrieval systems.
Explainability
The degree to which a model's reasoning or output can be understood and interpreted by humans. High explainability is especially important in regulated industries where decisions must be justified, such as finance and healthcare.
Few-shot prompting
Putting a few worked examples in the prompt so the model copies their pattern.
Fine-tuning
Additional training on a smaller, specific dataset that adapts a general model for a particular task, style, or domain. It is faster and cheaper than training from scratch and builds on the patterns already learned.
Foundation model
A large model trained on broad data that can be adapted to many different tasks. It serves as a base that other models or applications build on, rather than being designed for a single specific use.
Generative AI
AI systems that can produce new content, such as text, images, audio, or code, rather than only classifying or categorising existing content. The output is generated based on patterns learned during training, not retrieved from a database.
Grounding
Connecting a model's output to real, verifiable information rather than relying solely on what it learned during training. A grounded response is supported by specific evidence, such as a retrieved document or a live data lookup.
Hallucination
A confident but false or invented answer from a model.
Inference
Running a trained model to produce an answer, as opposed to training it.
Jailbreak
An attempt to get an AI model to ignore its safety guidelines by disguising a restricted request as something else. Developers work to identify and patch jailbreak techniques, but new ones are regularly discovered.
Large language model (LLM)
A model trained on large amounts of text to predict and generate language.
Latency
The time between sending a request to an AI model and receiving a response. Latency matters especially in real-time applications like chat interfaces or voice assistants, where long waits disrupt the experience.
Local model
An AI model that runs entirely on a user's own device rather than on a remote server. Running locally keeps data private and works without an internet connection, but requires enough memory and storage to load the model file.
Machine learning
A type of software that improves at a task by learning from examples rather than being programmed with explicit rules. The system finds patterns in data and uses them to make predictions or decisions on new inputs.
Model card
A document published alongside an AI model that describes its intended uses, training methodology, known limitations, and potential risks. Model cards are a transparency practice that helps users decide whether and how to use a model.
Model hosting
Providing access to an AI model through a service so that users can run it without managing the underlying hardware. Hosted models are accessed through an interface or API, and the provider handles the computing infrastructure.
Multimodal
Describes a model that can process and generate more than one type of data, such as both text and images. A multimodal model might accept an image as input and produce a written description, or accept text and produce an image.
Neural network
A mathematical system made of layers of connected units that transform input data into output predictions. The term is loosely inspired by the brain, though the similarity is mostly structural rather than functional.
Open-weight model
A model whose trained files are published for download. Its license may still limit how you can use it.
Output moderation
Automated systems that screen model outputs for harmful or policy-violating content before they reach the user. Moderation layers are separate from the main model and can be tuned independently.
Overfitting
When a model learns its training data so specifically that it fails to generalise to new, unseen data. The model performs well on examples it has already seen but poorly on anything outside that set.
Parameter
A numerical value inside a model that is adjusted during training and shapes how the model processes information. A model with more parameters can capture more complex patterns but requires more memory and computing power to run.
Pretraining
The initial large-scale training phase where a model learns from a broad dataset before any task-specific adjustment. Pretraining is the most expensive step and gives the model its general knowledge and language ability.
Prompt
The instructions and text you give an AI model.
Prompt engineering
The practice of designing and refining inputs to an AI model to get more useful, accurate, or specific outputs. It includes techniques like adding context, specifying roles, using examples, and setting constraints.
Prompt injection
Text that tries to override a model's instructions, often hidden in a document or web page.
Quantization
Storing a model's numbers with fewer bits to make it smaller and lighter to run, usually with some loss of quality.
Reinforcement learning from human feedback
A training technique where human raters evaluate model outputs and those ratings are used to steer the model toward more helpful and less harmful responses. Often abbreviated as RLHF, it is widely used to make chat models more useful.
Responsible AI
A set of practices and principles aimed at making AI systems fair, transparent, accountable, and safe. Responsible AI is not a single standard but a collection of values that different organisations interpret and apply in different ways.
Retrieval-augmented generation
A technique that connects a language model to an external set of documents at query time, so it can retrieve relevant information before generating an answer. Often abbreviated as RAG, it allows models to answer questions about specific or recent information without retraining.
Safety alignment
The process of training an AI model to follow guidelines, avoid harmful outputs, and behave in ways considered acceptable. Alignment shapes what the model will and will not do, separately from its raw capabilities.
Sampling
The process a model uses to select the next token when generating text. Various sampling strategies control the trade-off between diversity and coherence in the output, and are often adjusted alongside temperature.
Semantic search
A search method that finds results based on meaning rather than exact keyword matches. It uses embeddings to compare the meaning of a query with the meaning of stored documents, returning relevant results even when the words differ.
Structured output
A model response formatted in a predictable, machine-readable way, such as JSON or a table, rather than as free-form text. Structured output makes it easier for other software to parse and use the model's response automatically.
Synthetic data
Artificially generated data used to train or test AI models when real data is scarce, expensive, or sensitive. If the synthetic data does not accurately represent the real distribution, the model trained on it may not perform well in practice.
System prompt
Instructions given to an AI assistant before the user's conversation begins, used to set its behaviour, role, or tone. The end user usually does not see the system prompt, but the model treats it as authoritative guidance throughout the conversation.
Temperature
A setting that controls how random or predictable a model's output is. Higher temperature produces more varied and sometimes more creative responses; lower temperature produces more focused and consistent ones.
Throughput
The number of requests a model can process in a given time period. High throughput is important when many users are sending requests simultaneously and the system needs to handle the load efficiently.
Token
A small piece of text, often part of a word, that a language model reads or writes.
Tokenizer
The component of a language model system that converts raw text into the numerical units the model processes. The same word may be split into multiple tokens depending on the tokenizer, and the token count affects how much text fits in the context window.
Tool use
The ability of an AI model to call external functions or services, such as running code, searching the web, or reading files. The model selects the right tool, passes it arguments, and uses the result when generating its response.
Training
The process of showing a model large amounts of data so it can adjust its internal values and become better at a task. Training is usually the most expensive and time-consuming step in building an AI model.
Transformer
An architecture used in most modern language models that processes all words in an input at once by computing relationships between them. It replaced older sequential architectures and scales well with more data and computing power.
Vector database
A database that stores data as numerical vectors and finds similar items by measuring the distance between them. It is commonly used in RAG systems to retrieve the most relevant documents for a given query quickly.
Weights
The internal numerical values of a trained model that encode the patterns it has learned. Weights are set during training and determine how the model responds to any given input. The terms weights and parameters are often used interchangeably.
Zero-shot learning
Using a model to perform a task it was not explicitly trained on, based only on a description of the task. No examples are provided; the model relies on general patterns from its training to figure out what is expected.