100 AI & Prompt Engineering
Terms, Explained

Crafting inputs specifically designed to cause a model to behave incorrectly or unsafely.

An AI system that autonomously plans and executes multi-step tasks using tools and environment feedback.

A multi-step pipeline where an AI agent plans, acts, observes, and iterates to complete a complex task.

The research field focused on ensuring AI systems pursue goals and values intended by their designers.

The interdisciplinary field studying how to develop AI systems that are safe, reliable, and beneficial.

An Application Programming Interface that lets developers call AI model capabilities programmatically.

The core transformer operation that weighs the relevance of each token to every other token in a sequence.

An AI system capable of pursuing goals and executing long-horizon tasks with minimal human oversight.

A model trained only on next-token prediction over a large corpus, before any instruction tuning.

A standardised test suite used to measure and compare AI model capabilities across tasks.

Systematic errors in model outputs that unfairly favour or disadvantage certain groups or perspectives.

A prompting technique that asks the model to reason step-by-step before giving a final answer.

A model fine-tuned to handle multi-turn conversational exchanges between a user and an assistant.

Anthropic's family of large language models, known for safety, long context, and instruction-following.

An AI model whose weights and training details are proprietary and not publicly released.

A language model specifically trained or fine-tuned to understand and generate programming code.

A model that continues text from a given prefix, used before chat-style interfaces became standard.

Anthropic's technique for training helpful, harmless AI using a set of written principles as the training signal.

Automated or human review of AI inputs and outputs to prevent harmful, illegal, or policy-violating content.

The maximum number of tokens a model can handle in a single forward pass.

The maximum number of tokens a model can process in a single input-output interaction.

A structured collection of data examples used to train, validate, or evaluate a machine learning model.

A hint or nudge embedded in the prompt that steers the model toward a desired answer style.

A training method that aligns models to human preferences without requiring a separate reward model.

A dense numerical vector that represents a token, sentence, or document in a continuous semantic space.

Capabilities that appear suddenly in large models without being explicitly trained for.

The systematic process of measuring AI model quality, safety, and alignment against defined criteria.

The property of an AI system treating individuals and groups equitably and without unjust discrimination.

Providing a small number of input-output examples in the prompt to teach the model a task pattern.

A pretrained model whose weights have been updated on a specific dataset for a targeted task.

Continuing training of a pretrained model on a smaller, task-specific dataset to specialise its behaviour.

A large model trained on broad data that can be adapted to many downstream tasks.

A model feature that allows it to request the execution of developer-defined functions with structured arguments.

Google DeepMind's family of multimodal large language models, available in Ultra, Pro, and Nano tiers.

A technique where the model first generates relevant facts, then uses them to answer the question.

OpenAI's large multimodal model that set the benchmark standard for general-purpose AI capability.

Connecting model outputs to verifiable external sources to reduce hallucination and improve accuracy.

Programmatic constraints that prevent an AI application from producing or acting on harmful outputs.

When a model confidently generates false or fabricated information not supported by its training data or context.

The process of running a trained model to generate predictions or responses from new inputs.

Directly telling the model what to do using clear imperative commands.

Supervised fine-tuning on diverse instruction-response pairs to improve a model's ability to follow commands.

A model fine-tuned on instruction-response pairs to follow natural-language directives reliably.

Refining a prompt through multiple rounds of feedback until the output meets requirements.

A prompt designed to bypass a model's safety guidelines and elicit restricted content.

A neural network with billions of parameters trained on text to understand and generate language.

The time delay between sending a request to a model and receiving its first token in response.

Meta's open-weight large language model family, the dominant open-source alternative to GPT and Claude.

The raw, unnormalised score a model assigns to each vocabulary token before converting to probabilities.

A parameter-efficient fine-tuning method that updates only small low-rank matrices injected into model layers.

Mechanisms that allow AI agents to store, retrieve, and act on information across interactions.

A prompt that instructs the model to generate or improve other prompts.

A French AI company producing efficient open-weight models that punch above their size class.

A model that can process and generate across multiple data types such as text, images, and audio.

An instruction that tells the model what to avoid or exclude in its output.

A computational model loosely inspired by the brain, made of interconnected layers of weighted nodes.

A model whose weights are publicly available, enabling local deployment, inspection, and fine-tuning.

Specifying the exact structure, length, or schema the model should use when responding.

A family of methods that fine-tune large models by updating only a small fraction of their parameters.

The process by which an AI agent determines a sequence of actions to achieve a goal.

A mechanism that injects token position information into transformer inputs, since attention is order-agnostic.

The initial phase of training a model on massive text data to learn general language representations.

The input text sent to an AI model to elicit a specific response or behaviour.

Passing the output of one model call as the input to a subsequent call to complete complex tasks.

Reducing prompt length while preserving the information needed for accurate responses.

An attack where malicious text in external data hijacks the model's instruction-following behaviour.

A reusable prompt structure with placeholders that are filled in at runtime.

A memory-efficient fine-tuning method combining quantisation with LoRA adapters.

Augmenting model responses by retrieving relevant documents from an external knowledge base at inference time.

Interleaving reasoning traces and actions so a model can use tools and reflect on their results.

The capacity of an AI model to derive conclusions from premises through logical or analogical inference.

A model trained to perform extended internal reasoning before producing a response.

Systematically testing an AI system by attempting to elicit harmful or unintended behaviours before deployment.

A training technique that uses human preference ratings to align model outputs with human values.

The practice of developing and deploying AI systems ethically, transparently, and with accountability.

Shorthand for Reinforcement Learning from Human Feedback — the alignment training paradigm.

Assigning the model a persona or expert identity to shape the style and depth of its responses.

Empirical relationships describing how model performance improves predictably with more compute, data, or parameters.

An attention operation where a sequence attends to itself, allowing each token to gather context from all others.

Sampling multiple reasoning paths and selecting the most common answer to improve reliability.

A function that converts a vector of real numbers into a probability distribution summing to 1.

Sending model output tokens to the client incrementally as they are generated rather than all at once.

Constraining model generation to produce valid, machine-parseable formats like JSON or XML.

Training a model on input-output pairs where the correct output is provided as a label.

Training data generated by AI models rather than collected from human-created sources.

A privileged instruction block sent before the conversation that sets the model's persona and rules.

A sampling parameter that controls the randomness and creativity of model outputs.

The number of tokens or requests an inference system can process per unit of time.

The basic unit of text a language model processes, roughly corresponding to a word or word fragment.

The process of splitting text into tokens using a vocabulary and encoding algorithm like BPE.

The ability of a model to invoke external tools — APIs, code executors, search — to augment its capabilities.

A sampling strategy that limits token selection to the smallest set covering a cumulative probability.

The corpus of examples a model learns from during its training process.

Reusing a model trained on one task as the starting point for a related task.

The neural network architecture that underpins all modern large language models, based on self-attention.

A framework that explores multiple reasoning branches in parallel and selects the most promising path.

Training a model on raw data without human-provided labels, discovering structure autonomously.

A database optimised for storing and querying high-dimensional embedding vectors via similarity search.

A model capable of jointly reasoning over both images and text.

Asking a model to perform a task it has never seen demonstrated, relying purely on instructions.