Glossary of AI terms
67 terms you need to know — from the basics to advanced concepts
Artificial intelligence (AI)
The ability of machines to perform tasks that usually require human intelligence — speech recognition, decision-making, translation.
Machine learning (ML)
A subfield of AI where models learn from data without rules being explicitly programmed.
Neural network
A computational model inspired by the human brain — it processes and transforms data through layers of interconnected nodes.
Deep learning
Machine learning with many neural network layers, able to recognize complex patterns in data.
Model training
The process in which a model repeatedly goes through training data and adjusts its parameters.
Inference
Using an already trained model to generate answers or predictions.
Model parameters
A model's internal values set during training — GPT-4 has ~1.8 trillion parameters.
Benchmark
A standardized test for measuring and comparing the performance of AI models.
Open-source model
A model whose weights are publicly available — e.g. Llama, Mistral.
Proprietary model
A model whose weights are closed — e.g. GPT-4, Claude, Gemini.
Bias
A model's tendency to prefer certain answers based on patterns in its training data.
Dataset
A collection of data that a model is trained or tested on.
Label
A tag assigned to data so the model knows what's what. The foundation of supervised learning.
GPU
A specialized chip originally built for graphics, today the key hardware for training AI models.
Guardrails
Safety constraints built into a model that prevent harmful or inappropriate outputs.
Turing Test
A test devised by Alan Turing — a machine passes it if its answers can't be told apart from a human's.
LLM (Large Language Model)
A large language model trained on a huge amount of text, able to generate, translate and analyze text.
Token
The basic unit of text an LLM processes. Roughly 1 token = 0.75 words. The price of an API call is calculated in tokens.
Context window
The maximum amount of text a model sees at once. The bigger the context, the more information the model can process.
Temperature
A parameter that affects how creative the outputs are. Low (0–0.3) = precise answers. High (0.7–1) = more creative outputs.
Top-P (Nucleus Sampling)
A parameter that limits the choice of the next token to the most probable options. An alternative to temperature.
Hallucination
When a model confidently generates false or made-up information as if it were fact.
System prompt
An instruction at the start of a conversation that defines the model's behavior, role and constraints.
Multimodal model
A model that processes several types of data at once — text, images, audio, video.
GPT (Generative Pretrained Transformer)
The large language model architecture from OpenAI. The foundation of ChatGPT and other models.
AGI (Artificial General Intelligence)
A hypothetical AI able to learn and solve any task like a human — it doesn't exist yet.
ASI (Artificial Superintelligence)
A hypothetical AI that would surpass the best human abilities in every respect.
NLP (Natural Language Processing)
Natural language processing — the field of AI that enables machines to understand and generate human text.
MCP (Model Context Protocol)
An open standard for communication between AI models and external tools, databases and services.
Prompt
The input instruction or question you give an AI model.
Prompt Engineering
The discipline of designing and optimizing prompts to get the best outputs from AI models.
Zero-shot prompting
Giving a task without any examples. The model answers based on its training alone.
Few-shot prompting
Giving a task together with a few input/output examples. Significantly improves answer accuracy.
Chain-of-Thought (CoT)
A technique where the model thinks step by step before giving its final answer. Improves results on complex tasks.
Role prompting
Assigning the model a role ("You are a senior developer..."). Affects the tone, style and depth of answers.
Instruction prompt
A precise command telling the model what to do, how to do it and in which output format.
Meta-prompt
A prompt that generates other prompts. Used to automate prompt engineering.
Prompt injection
An attack where malicious content in the data overrides the system prompt's instructions.
Jailbreak
An attempt to bypass a model's safety constraints using specially crafted prompts.
RAG (Retrieval-Augmented Generation)
An architecture that retrieves relevant documents before generating an answer and adds them to the context. Solves the problem of the model's outdated knowledge.
Embedding
A numerical representation of text (a vector) that captures its semantic meaning. Similar texts have similar vectors.
Vector database
A database optimized for storing and searching embeddings. E.g. Pinecone, pgvector, Chroma.
Semantic search
Searching by the meaning of text, not by keywords. Uses embeddings.
Chunking
Splitting long documents into smaller pieces before storing them in a vector database. Your chunking strategy has a major impact on RAG quality.
Reranking
A second RAG step where the retrieved documents are sorted by relevance before being passed to the model.
Grounding
Anchoring a model's answers in specific data or documents. Reduces hallucinations.
AI agent
An AI system that independently plans and carries out steps toward a goal using tools and decision-making.
Agentic AI
An approach where AI autonomously decides on a sequence of actions without human intervention at every step.
Tool use
An agent's ability to call external tools — search, calculator, APIs, databases.
Multi-agent system
An architecture where several specialized agents work together on a complex task.
Orchestrator
An agent that manages other agents — it splits up tasks, collects results and coordinates the workflow.
Memory (agent memory)
A mechanism for storing context and the results of previous steps. Short-term (in the context) vs. long-term (in a database).
ReAct
A technique (Reasoning + Acting) where the agent alternates between thinking and acting in an iterative loop.
Human-in-the-loop
A design where a human approves or corrects the agent's decisions at critical moments.
Supervised Learning
Learning from data that comes with correct answers as a template. The model learns to map inputs to outputs.
Unsupervised Learning
Learning from raw, unlabeled data. The model finds patterns and structures on its own.
Reinforcement Learning
Learning by trial, reward and punishment. The agent learns optimal actions by interacting with its environment.
Computer Vision
The field of AI that enables machines to recognize and analyze images, videos and visual data.
Fine-tuning
Further training of an existing model on specific data for a particular use case.
RLHF (Reinforcement Learning from Human Feedback)
A technique where people rate the model's outputs and thereby teach it to give better answers.
LoRA (Low-Rank Adaptation)
An efficient fine-tuning method that changes only a small portion of the parameters. Cheaper and faster than full fine-tuning.
Synthetic data
Training data artificially generated with AI. Solves the shortage of real data.
Overfitting
When a model works great on training data but poorly on new data. It has memorized the data.
GEO (Generative Engine Optimization)
Optimizing content for visibility in AI models like ChatGPT, Perplexity, Gemini. The successor to traditional SEO.
AI citation
A link an AI model cites as a source when generating an answer. The main goal of a GEO strategy.
Share of Voice in AI
The share of answers in which an AI model mentions your brand or website for relevant queries.
Authoritative content
Content structured so that AI models treat it as a reliable source and cite it.