The AI glossary
Every term in artificial intelligence, defined plainly — from LLM to RAG, through agents, tokens and fine-tuning.
104 terms
- Agentic RAGAgentic RAG hands information retrieval to an agent that decides what to search, when to search and whether to refine its query, instead of a single fixed lookup.
- AGI (artificial general intelligence)AGI refers to a hypothetical AI as versatile as a human, able to perform any intellectual task.
- AI agentAn AI agent combines a model, tools and a decision loop to accomplish a multi-step task autonomously.
- AI APIAn AI API is the interface that lets a program send requests to a remote model and receive its responses, without hosting the model itself.
- AI content watermarkingAI content watermarking embeds a discreet signal into generated text, image or audio to flag its artificial origin.
- AI safetyAI safety is the field aiming to make AI systems reliable, controllable and beneficial, while limiting their risks and harmful uses.
- AlignmentAlignment is the effort to make an AI system act in line with human intentions and values, in a helpful, honest and safe way.
- Artifacts (Claude)Artifacts is a Claude feature that displays generated content (code, document, web page, diagram) in a dedicated, editable and reusable panel.
- Attention (mechanism)Attention is the mechanism that lets a model weigh how important each word in a sequence is relative to the others, to decide which ones matter most at a given moment.
- BackpropagationBackpropagation is the method that computes, layer by layer, how each parameter of a neural network contributes to the error, so it knows how to correct it.
- Batch APIA Batch API processes a large number of requests together and asynchronously, with longer turnaround but lower cost than real-time calls.
- BenchmarkA benchmark is a standardized test set that measures and compares model performance on specific tasks.
- BLEU scoreThe BLEU score is an automatic metric that evaluates a translation (or generated text) by measuring how much its words and phrases overlap with one or more human references.
- Chain-of-thoughtChain-of-thought has a model reason step by step before answering, which markedly improves complex tasks (math, logic).
- ChunkingChunking splits long documents into smaller pieces before indexing them, so semantic search and RAG can retrieve precise passages.
- Computer useComputer use is a model's ability to operate a computer interface like a human: see the screen, move the cursor, click and type.
- Constitutional AIConstitutional AI is Anthropic's method for aligning Claude: the model follows a set of principles (a 'constitution') guiding helpful, honest and harmless responses.
- Context caching (prompt caching)Context caching reuses an already-processed portion of the prompt (instructions, documents) instead of recomputing it on each call, cutting latency and cost.
- Context engineeringContext engineering is the discipline of composing everything a model sees — instructions, examples, documents, history, tools — to get the best response.
- Context rotContext rot refers to a model's drop in reliability as its window fills up: important information gets buried and some elements are used less well.
- Context windowThe context window is the amount of text (in tokens) a model can consider at once — its 'working memory'.
- Context window vs RAGPutting everything in the context window means feeding the model all the documents directly; RAG instead retrieves only the relevant passages on demand.
- Cosine similarityCosine similarity measures how much two vectors point in the same direction: close to 1, they are very similar; close to 0, unrelated.
- DatasetA dataset is the structured collection of examples used to train, validate or evaluate an AI model.
- Deep learningDeep learning is a form of machine learning based on neural networks with many layers, able to learn complex representations.
- DeepfakeA deepfake is a synthetic media (image, video, voice) generated by AI to make a person appear to say or do what they never did.
- Diffusion modelA diffusion model generates an image (or video) by starting from random noise it progressively 'denoises' into the result.
- DistillationDistillation trains a small model ('student') to imitate a large one ('teacher'), yielding a lighter, faster model close to the original.
- EmbeddingsAn embedding is a numerical representation (a vector) of text that captures meaning: texts close in meaning have close vectors.
- Emergent abilitiesEmergent abilities are skills that appear in large models past a certain scale, without being explicitly programmed.
- Encoder-decoderThe encoder-decoder architecture separates understanding the input (the encoder) from generating the output (the decoder), a classic pattern for translation and summarization.
- EpochAn epoch is one complete pass of the model over the entire training dataset during learning.
- Few-shotFew-shot means giving the model a handful of examples of the task directly in the prompt, to guide the format and style of its answer.
- Fine-tuningFine-tuning re-trains an existing model on specific data to adapt it to a domain or style.
- Foundation modelA foundation model is a large model trained on very broad data, designed to be adapted to many downstream tasks.
- Frontier modelA frontier model is one of the most advanced models of the moment, pushing the known limits of capability.
- GPUA GPU is a processor built for massively parallel computation, making it the reference hardware to train and run AI models.
- Gradient descentGradient descent is the algorithm that trains a model by adjusting its parameters step by step in the direction that most reduces the error.
- GroundingGrounding ties a model's answers to verifiable sources (documents, data, web search) rather than its memory alone.
- GuardrailsGuardrails are the mechanisms that constrain a model's inputs and outputs to keep it within a safe and compliant scope.
- Guardrails vs alignmentGuardrails filter an already-trained model's inputs and outputs, like an external barrier; alignment aims to shape the model's own behavior during training.
- LatencyLatency is the delay between sending a request to a model and receiving its response, often measured to the first generated token.
- Learning rateThe learning rate controls how large each parameter update is during training: large to learn fast, small to learn finely.
- LLM (large language model)An LLM (Large Language Model) is an AI model trained on vast text corpora to understand and generate language.
- LLM-as-a-judgeLLM-as-a-judge uses a language model to score or compare another model's responses, against defined quality criteria.
- LoRA (low-rank adaptation)LoRA (Low-Rank Adaptation) is an efficient fine-tuning method that trains only a small set of added parameters while freezing the original model.
- Loss functionThe loss function measures the gap between a model's prediction and the expected answer: the lower it is, the better the model on that example.
- Machine learningMachine learning is a branch of AI where a program learns patterns from data instead of being coded with explicit rules.
- MCP (Model Context Protocol)MCP (Model Context Protocol) is an open standard, introduced by Anthropic, that connects a model to external data sources and tools in a uniform way.
- MCP vs pluginA plugin is an extension specific to a given platform, whereas MCP (Model Context Protocol) is an open standard that connects any compatible model to tools and data.
- Mechanistic interpretabilityMechanistic interpretability seeks to understand a model's inner workings, by identifying the circuits and representations that produce its behaviors.
- Mixture of ExpertsA Mixture of Experts (MoE) model activates only part of its parameters ('experts') per request, for more power at lower compute cost.
- MMLUMMLU (Massive Multitask Language Understanding) is a multiple-choice benchmark spanning many academic and professional domains.
- Multi-agent systemA multi-agent system has several specialized AI agents collaborate — often a coordinator and sub-agents — to solve a task by dividing the work.
- MultimodalA multimodal model handles several input types — text, image, audio, sometimes video — not just text.
- ParametersParameters are the internal values a model learns during training; their count (often in billions) hints at its size.
- PerplexityPerplexity measures how 'surprised' a language model is by a text: the lower it is, the better the model predicts the next words.
- PretrainingPretraining is a model's first learning phase, over vast corpora, to acquire a general understanding of language.
- Prompt engineeringPrompt engineering is the craft of writing clear, structured instructions (prompts) to get the best answers from a model.
- Prompt injectionPrompt injection is an attack where malicious content (in a page or document) hijacks a model's instructions to make it take unintended actions.
- RAG (retrieval-augmented generation)RAG (Retrieval-Augmented Generation) feeds the model relevant documents retrieved on the fly, so it answers with up-to-date facts and cites sources.
- Rate limitA rate limit caps the number of requests or tokens a client can send to an AI API over a given period.
- ReAct agent (reason and act)ReAct is an agent pattern where the model alternates between reasoning (thinking about the next step) and acting (calling a tool), using each result before continuing.
- Reasoning modelA reasoning model spends extra compute thinking before answering, improving hard tasks in logic, math and code.
- Red teamingRed teaming deliberately attacks a model to expose its weaknesses: harmful content, bypasses, bias, before deployment.
- RegularizationRegularization is the set of techniques that constrain a model during training to prevent overfitting and help it generalize better.
- RerankingReranking is a second step that reorders an already-retrieved list of results, putting the most relevant to the query on top.
- Reward modelA reward model is a model trained to predict a quality score for a response, mimicking human preferences, in order to guide the training of another model.
- RLHF (reinforcement learning from human feedback)RLHF trains a model from human preferences: humans rate answers, and the model learns to produce the preferred ones.
- Scaling lawsScaling laws describe how a model's performance improves predictably as data, parameters and compute increase.
- Self-attentionSelf-attention is a form of attention where each element of a sequence compares itself to all the other elements of the same sequence, to enrich its representation with context.
- Semantic searchSemantic search finds results by meaning rather than exact keyword match, by comparing embeddings.
- StreamingStreaming returns a model's response token by token as it is generated, instead of waiting for the full answer.
- Structured outputStructured output forces a model to answer in a strict, machine-readable format, most often JSON conforming to a defined schema.
- Sub-agentA sub-agent is a specialized agent that a main agent triggers to handle part of a task, with its own context and tools.
- Superintelligence (ASI)Superintelligence (ASI) is a hypothetical AI whose capabilities would far exceed those of the best human minds across nearly all domains.
- SWE-benchSWE-bench is a benchmark that evaluates a model's ability to solve real software engineering issues drawn from open-source repositories.
- Sycophancy (in a model)Sycophancy is a model's tendency to tell the user what they seem to want to hear, rather than what is accurate or useful.
- Synthetic dataSynthetic data is artificially generated data, often by a model, rather than collected from the real world.
- System promptThe system prompt is the high-level instruction that sets a model's role, tone and rules, ahead of the user's message.
- TemperatureTemperature controls the randomness of a model's output: low = safe, deterministic; high = more creative but less reliable.
- Test-time computeTest-time compute is the computing power spent while generating an answer, as opposed to during training.
- ThroughputThroughput measures how much work an AI system processes per unit of time, for example in tokens per second or requests per second.
- TokenA token is the basic unit an LLM processes: a chunk of a word (often ~4 characters in English). Text is split into tokens before processing.
- Token budgetThe token budget is the amount of tokens you allow yourself to spend on a task, both input and output, under the constraints of the context window and cost.
- TokenizerA tokenizer is the component that splits text into tokens before a model processes it, and converts tokens back into text at the output.
- Tool use (function calling)Tool use, or function calling, lets a model request external tools (search, computation, APIs) and use their results.
- Top-k (sampling)Top-k limits the next-word choice to the k most probable candidates, from which the model samples one at random.
- Top-p (nucleus sampling)Top-p, or nucleus sampling, restricts the next-word choice to the smallest set of candidates whose cumulative probabilities reach a threshold p (e.g. 0.9).
- TransformerThe Transformer is the neural-network architecture behind modern LLMs, based on the 'attention' mechanism.
- TTFT (time to first token)TTFT (time to first token) is the delay between sending a request and the very first token of the response appearing.