Every conversation about AI these days comes with a pile of jargon. Tokens. Weights. Inference. Nodding along only gets you so far.
This A to Z guide breaks down the core terms and technologies behind the AI revolution in plain language, from the companies building frontier models to the techniques that let AI run on your laptop. No computer science degree required.
Bookmark it for the next time a term makes you go, "Wait, what?"
Anthropic (an-THROP-ik)
Anthropic is the AI company behind Claude. It was founded in 2021 by seven former OpenAI employees, including siblings Dario and Daniela Amodei. The founders left OpenAI over disagreements about how safely AI should be developed. It's now one of the leading AI labs (alongside OpenAI, Google DeepMind, and others), with safety research and coding performance as its core differentiators.
Bias (BY-əss)
In AI, bias refers to the way a model's outputs can skew unfairly and favour certain groups, viewpoints, or patterns. Bias happens because of imbalances in the data it was trained on. (TL;DR: garbage in, garbage out.) Since models learn from human-created text and images, they can absorb and amplify human prejudices, which is why reducing bias is a major focus of AI research.
ChatGPT (CHAT-jee-pee-TEE)
ChatGPT is OpenAI's flagship chatbot and the product that kicked off the modern AI boom when it launched in November 2022. It reached 100 million users faster than any consumer app before it, and for many people "ChatGPT" has become shorthand for AI chatbots in general.
Diffusion (dih-FYOO-zhən)
Diffusion is the technique behind most AI image generators like Midjourney and Stable Diffusion. During training, images are progressively blurred into random static, and the model learns how to undo that damage step by step. Later, when you type a prompt, the model starts from pure static and gradually "cleans it up" until the prompted image emerges.
Ethics (ETH-iks)
AI ethics is the field concerned with the moral questions AI raises: fairness, privacy, job displacement, misinformation, and who's accountable when AI causes harm. It overlaps with safety research but focuses more on AI's impact on society than on controlling the technology itself.
Foundation Model (fown-DAY-shən MOD-əl)
A foundation model is a large, general-purpose AI model trained on massive amounts of data that can then be adapted to many different tasks — writing, coding, translation, analysis — rather than being built for just one job. GPT-4, Claude, and Gemini are all foundation models.
BONUS: Frontier Model (frun-TEER MOD-əl)
A frontier model is one of the most advanced, cutting-edge AI systems at any given moment from one of the leading AI companies. These models push the boundaries of what AI can do. The term is used by governments and labs alike to describe systems powerful enough to warrant extra scrutiny. GPT-4, Claude, and Gemini all qualify.
Guard Rails (GARD-raylz)
Guard rails are the rules and restrictions built into an AI system to keep it from producing harmful, illegal, or off-limits content. They're why a chatbot will refuse to help with certain requests. They're exactly what "jailbreaks" try to get around.
Human-in-the-loop (HYOO-mən-in-thə-LOOP)
Human-in-the-loop describes any AI system where a person stays involved in the process — reviewing outputs, approving decisions, or correcting mistakes rather than letting the AI act fully on its own. It's a common safeguard in high-stakes settings like medicine, hiring, and law.
Inference (IN-fer-əns)
Inference is what happens every time you actually use an AI model. It takes your prompt and generates a response, one word (or piece of a word) at a time, with each new word predicted based on everything that came before it. It's the counterpart to training: training is the one-time, enormously expensive process where the model learns, while inference is the model putting that learning to work, billions of times a day. It's also where the ongoing costs of AI live. Every response requires serious computing power, which is why AI companies are racing to make inference faster and cheaper.
Jailbreak (JAYL-brayk)
A jailbreak is a prompt designed to trick an AI into ignoring its guard rails that gets it to produce content it was built to refuse. Jailbreaks range from clever role-play scenarios to elaborate technical exploits, and AI companies are in a constant cat-and-mouse game patching them. The scary thing is that it’s not just people that try and jailbreak AI models.
Knowledge Graph (NOL-ij graf)
A knowledge graph is a way of storing information as a web of connected facts — "Paris is the capital of France," "France is in Europe" — rather than as plain text. Google uses one to power the info boxes in search results, and they're increasingly combined with AI models to make answers more accurate.
LLM (el-el-EM)
LLM stands for "large language model", it’s the technology behind AI chatbots like ChatGPT and Claude. LLMs are trained on enormous amounts of text and work by predicting the next word in a sequence, which turns out to be enough to produce essays, code, conversation, and more.
Meta (MEH-tuh)
Meta is the parent company of Facebook, Instagram, and WhatsApp, and a major AI player in its own right. It's best known in AI for its Llama models, which it releases with open weights — meaning developers can download and run them for free, unlike the closed models from OpenAI and Anthropic.
Nvidia (ən-VID-ee-uh)
Nvidia is the chipmaker whose graphics processors (GPUs) turned out to be ideal for training and running AI models. Its chips power nearly every major AI lab, and the AI boom briefly made it the most valuable company in the world.
Open Weights (OH-pən wayts)
An open-weights model is one whose underlying parameters (the "weights") are published for anyone to download, modify, and run themselves. It's often called "open source AI," though purists note that true open source would also include the training data and code, which usually stay private.
Prompt Engineering (PROMPT en-jin-EER-ing)
Prompt engineering is the craft of writing instructions that get the best results from an AI model by adding context, examples, and structure rather than just asking a bare question. The skill is evolving with agentic AI where systems can run in loops, checking their own work and retrying until they hit a goal. The emphasis is shifting from writing the perfect instruction to clearly defining the goal and letting the AI figure out the steps.
Quantization (kwon-tih-ZAY-shən)
Quantization is a technique for shrinking AI models by storing their numbers with less precision, kind of the same way an algorithm can compress a photo. The model gets smaller and faster with only a modest drop in quality, which is what makes it possible to run capable AI on a laptop or phone.
Robotics (roh-BOT-iks)
Traditional robotics is the field of building machines that sense and act in the physical world. Robotics + AI = Physical AI, as companies race to give robots the kind of general intelligence that chatbots have shown with language with humanoid robots as the most visible (and hyped) frontier.
Singularity (sing-gyuh-LAIR-ih-tee)
The singularity is a hypothetical future point where AI becomes capable of improving itself faster than humans can keep up, triggering runaway technological change. It's a staple of both serious debate and science fiction, with predictions ranging from "within decades" to "never."
Tokens (TOH-kənz)
Tokens are the chunks of text an AI model actually reads and writes. They usually represent a word or piece of a word ("understanding" might be split into "under" + "standing"). AI usage and pricing are measured in tokens, which is why you'll see limits described as "context windows" of so many thousand tokens.
Uncanny Valley (un-KAN-ee VAL-ee)
The uncanny valley is the unsettling feeling caused by something that looks almost (but not quite) human. Coined in robotics in the 1970s, the term now applies to AI-generated faces, voices, and videos that are realistic enough to be eerie rather than convincing.
Vision (VIZH-ən)
In AI, vision refers to a model's ability to interpret images and video and recognize objects, read documents, and describe photos. Modern chatbots are increasingly "multimodal," combining vision with language so you can show them a picture and ask questions about it.
Weights (WAYTS)
Weights are the millions or billions of numbers inside an AI model that encode everything it has learned. Training a model is essentially the process of adjusting these numbers, and the final set of weights is the model. That’s why publishing them is the AI world's version of open source: anyone can download, run, and build on the model for free (see Open Weights).
XAI (EKS-ay-eye)
XAI stands for "explainable AI", the effort to make AI systems whose decisions humans can understand and audit, rather than mysterious black boxes. Confusingly, xAI (lowercase x) is also the name of Elon Musk's AI company, maker of the Grok chatbot, so context matters.
YOLO (YOH-loh)
YOLO ("You Only Look Once") is a famous object-detection model that can identify and locate multiple objects in an image in a single pass, fast enough to work on live video. It's widely used in self-driving research, security systems, and anywhere machines need to see in real time.
Zero-shot Learning (ZEER-oh-shot LUR-ning)
Zero-shot learning is when an AI handles a task it was never explicitly trained or given examples for, like translating into a language pair it hasn't seen demonstrated. It's one of the surprising abilities that emerged as language models got larger.
