AI Terminology Cheat Sheet

AI conversations are full of jargon. If you’ve ever read an article about AI and felt lost by the third paragraph, this post is for you. Here’s a plain-language glossary of the terms you’ll encounter most. Bookmark it and come back whenever you need a refresher.

The basics

Artificial Intelligence (AI) — Any system that performs tasks normally requiring human intelligence. This is the broadest term and covers everything from a simple spam filter to ChatGPT.

Machine Learning (ML) — A subset of AI where systems learn from data instead of following pre-written rules. You feed it examples, it finds patterns, and it applies those patterns to new data.

Deep Learning — A subset of machine learning that uses neural networks with many layers. It’s behind most recent AI breakthroughs, including image recognition, language models, and speech processing.

Neural Network — A computing system loosely inspired by the structure of the brain. It consists of layers of interconnected nodes that process data. Each layer transforms the data and passes it to the next.

Algorithm — A set of instructions or rules that a computer follows to solve a problem. In AI, algorithms define how a model learns from data and makes predictions.

Language models and text

Large Language Model (LLM) — An AI model trained on massive amounts of text data that can generate, summarize, translate, and analyze text. GPT-4, Claude, Gemini, and Llama are all LLMs.

Token — The basic unit that an LLM processes—roughly three-quarters of a word on average. The word “cybersecurity” might be split into “the two tokens cyber” and “security”.Context Window — The maximum amount of text an LLM can process at once, including both input and output. Measured in tokens. A larger context window means the model can handle longer conversations or documents.

Prompt — The input you give to an AI model. Your question, instruction, or the text you want it to work with.

Prompt Engineering — The practice of writing better prompts to get better outputs from AI. Involves being specific, providing context, giving examples, and iterating.

Hallucination — When an AI generates information that is factually incorrect or entirely fabricated while presenting it as fact. A byproduct of how language models work, not a bug that can be fully fixed.

Temperature — A setting that controls randomness in the model’s output. Low temperature produces more predictable, conservative responses. High temperature produces more varied and creative (but potentially less accurate) responses.

Fine-tuning — Taking a pre-trained model and training it further on a specific dataset to improve its performance for a particular task or domain.

RLHF (Reinforcement Learning from Human Feedback) — A training technique where human reviewers rate the model’s outputs, and the model is adjusted to produce more of the kind of responses that humans rated highly.

Training and data

Training Data — The dataset used to teach a model. For LLMs, this typically consists of a massive collection of text from the internet, books, and other sources.

Pre-training — The initial phase of training where the model learns general patterns from a large dataset. This is the computationally expensive part that gives the model its broad capabilities.

Inference — When a trained model processes new input and generates an output. Every time you send a prompt to ChatGPT and get a response, that’s inference.

Parameters — The adjustable values inside a model that determine how it processes data. More parameters generally mean a more capable model. GPT-4 is estimated to have over a trillion parameters.

Overfitting — When a model learns the training data too well and performs poorly on new data it hasn’t seen before. It memorized the examples instead of learning the underlying patterns.

Bias — Systematic errors in a model’s outputs caused by biases in the training data. If the training data contains stereotypes or underrepresents certain groups, the model’s outputs will reflect that.

Model types and architectures

Transformer — The neural network architecture behind modern LLMs. Introduced in 2017, it processes text in parallel rather than sequentially, enabling much faster ttrainingon large datasets. GPT stands for “Generative Pre-trained Transformer.”

Generative AI — AI that creates new content — text, images, audio, video, code. ChatGPT generating a response, DALL-E generating an image, and Suno generating music are all generative AI.

Diffusion Model — The architecture behind most modern image generators (Midjourney, DALL-E, Stable Diffusion). It works by learning to remove noise from images, which lets it to generate new images from random noise.

GAN (Generative Adversarial Network) — An older generative AI architecture where two networks compete: one generates content, the other tries to detect fakes. This competition improves the generator over time. Used in early deepfake technology.

Multimodal — A model that can process and generate multiple types of data — text, images, audio, and video. GPT-4 is multimodal because it can process both text and images.

Open Source vs. Closed Source — Open source models (like Meta’s Llama) release their weights publicly so anyone can use and modify them. Closed source models (like GPT-4) keep their weights proprietary and are only accessible through the provider’s API or interface.

Practical terms

API (Application Programming Interface) — A way for software to interact with an AI model programmatically. Developers use APIs to integrate AI capabilities into their own applications.

RAG (Retrieval-Augmented Generation) — A technique where the AI retrieves relevant information from an external database or document set before generating a response. This reduces hallucinations by grounding the response in actual data.

Embedding — A way of representing text (or other data) as a list of numbers that captures meaning. Similar concepts have similar embeddings. Used in search, recommendation, and RAG systems.

Agent — An AI system that can take actions beyond just generating text — browsing the web, running code, calling APIs, managing files. Agents use tools to accomplish multi-step tasks.

Chatbot — An AI application designed for conversational interaction. ChatGPT, Claude, and Gemini are chatbots built on LLMs.

Guardrails — Rules and filters built into AI systems to prevent harmful, biased, or inappropriate outputs. These are part of the safety layer that sits on top of the base model.

AI and cybersecurity terms

Deepfake — AI-generated synthetic media — fake audio, video, or images designed to look or sound like a real person.

Adversarial Attack — An attack designed to fool an AI model by feeding it carefully crafted inputs. For example, adding subtle noise to an image that makes an AI misclassify it.

Data Poisoning — Deliberately corrupting an AI model’s training data to make it produce incorrect or harmful outputs.

Model Extraction — An attack where an adversary queries a model repeatedly to reverse-engineer its behavior and create a copy.

AI Red Teaming — Testing AI systems for vulnerabilities, biases, and safety issues by simulating adversarial conditions. Similar to traditional security red teaming but focused on AI-specific risks.

Bottom line

You don’t need to memorize all of these at once. Bookmark this page and refer back to it as you encounter these terms in articles, conversations, and your own work. The vocabulary will become natural as you spend more time with the technology.