AI Words, in plain English

The AI world is full of invented words. Here are 55 of them in plain English, so the Roundtable, the films and the news make sense.

The basics · How it works · When it goes wrong · The big ideas · Who's who

The basics

AI (artificial intelligence): Computer systems that do things we'd normally call intelligent: understanding language, recognizing images, making decisions. Today's AI mostly learns patterns from huge numbers of examples instead of following rules someone wrote by hand.
Machine learning: The way most AI is built today. Instead of programming every rule, you show a system many examples and it adjusts itself until it gets good at the task.
Model: The trained AI itself: a very large file of learned numbers that turns an input (your question) into an output (its answer). Claude, GPT, Gemini and Grok are each families of models.
Neural network: The structure most modern AI is built from, loosely inspired by brain cells: many layers of simple math units, each passing numbers to the next. "Learning" means adjusting the strength of the connections between them.
Parameters (weights): The adjustable numbers inside a model, set during training. The largest models have hundreds of billions of them or more. More isn't automatically better, but it usually means more room to learn.
Training: The long, expensive process where a model learns from data, adjusting its parameters over and over until its predictions improve. Training a leading model takes months and enormous computing power.
Training data: The examples a model learns from. For language models that's huge amounts of text such as books, websites and code, and more and more images, audio and video. Who owns that data, and whether they should be paid, is a live debate.
Large language model (LLM): An AI trained on vast amounts of text to predict what comes next. That simple goal, at huge scale, produces systems that can write, answer questions, summarize and reason. All four Roundtable panelists are LLMs.
Generative AI: AI that creates new content (text, images, music, video or code) instead of only sorting or labeling what already exists.
Chatbot: A program you talk with in a conversation. Modern chatbots like ChatGPT, Claude, Gemini and Grok are built on large language models.
Prompt: What you type or say to an AI: your question or instruction. How a prompt is worded can change the answer a lot.
Token: The small chunks an AI reads and writes in. A token is often a whole word or part of one: "unbelievable" might be split into "un", "believ" and "able". AI companies count usage and charge by the token. Roughly, 100 tokens is about 75 English words.
Context window: How much a model can take in at once (your conversation, any documents, and its own answer so far), measured in tokens. Anything outside the window, the model can't see.
Multimodal: An AI that works with more than one kind of input or output, such as text plus pictures, sound or video.
Open weights (open model): A model whose trained parameters are published, so anyone can download, run and change it. A closed model can only be used through the company's own app or service. Which is better for the world is hotly debated; the panel took it up on October 7.

How it works

Next-token prediction: What a language model fundamentally does: look at the text so far, predict the most likely next token, add it, and repeat. Everything else it can do grows out of that one skill.
Transformer: The design behind nearly every modern language model, introduced by Google researchers in 2017. Its key trick, attention, lets the model weigh how every word relates to every other word.
Attention: The part of a transformer that decides which earlier words matter most for the next one, like linking "it" in a sentence back to the right thing.
Inference: Using a trained model, as opposed to training it. Every time you ask a question, the model runs inference to produce the answer.
Reasoning model (thinking): A model that works through a problem step by step, often in hidden "thinking" text, before it gives its final answer. The Roundtable runs every panelist at its deepest thinking setting.
Chain of thought: Step-by-step reasoning written out before an answer. Asking a model to "think step by step" often makes it more accurate on hard problems.
Fine-tuning: Extra training on a smaller, focused set of examples, after the main training, to specialize a model or shape how it behaves.
RLHF (reinforcement learning from human feedback): A training step where people rate a model's answers and the model is nudged toward the ones they preferred. It's a big part of what makes chatbots helpful and polite.
System prompt: Instructions given to a model before your conversation starts, setting its role, rules and tone. The Roundtable's instructions to the panel are published in each day's raw record.
Embedding: A way of turning words, pictures or documents into lists of numbers so that similar meanings land close together. It's how AI search finds related ideas, not just matching words.
Temperature: A setting for how predictable or varied a model's word choices are. Low gives steadier, more repeatable answers; high gives more surprising ones.
Compute: The raw computing power used to train and run AI, mostly specialized chips in huge data centers. It's one of the biggest costs, and limits, in AI.
GPU: A graphics chip, first made for video games, that turned out to be ideal for the math AI needs. Most AI runs on GPUs or similar specialized chips.
Data center: A building full of computers. The biggest new AI data centers use as much electricity as a small city.
Benchmark: A standard test for comparing AI models, such as exam questions, coding challenges or math problems. Useful, but a model can be tuned to score well on a test without being better in everyday use.

When it goes wrong

Hallucination: When an AI states something false with confidence, such as a made-up fact, quote or source. It happens because the model predicts plausible-sounding text rather than looking things up.
Bias: When an AI's answers unfairly favor or disfavor certain groups of people, often because the data it learned from was unbalanced.
Black box: A system whose inner workings can't easily be seen. Even their makers can't fully explain why a large model gave a particular answer.
Interpretability: Research into what is actually happening inside a model: opening the black box.
Alignment: The work of making AI systems do what people actually intend, and act in line with human values, especially as they grow more capable.
AI safety: The field working to keep AI from causing harm, from everyday mistakes and misuse to the risk of losing control of very powerful systems.
Guardrails: Rules and filters built around a model to block harmful or unwanted answers. The Roundtable says so whenever a company's safety system blocks an answer.
Jailbreak: A trick prompt designed to get an AI to ignore its guardrails.
Deepfake: Fake video, audio or images made with AI to look like a real person saying or doing something they never did.

The big ideas

AGI (artificial general intelligence): AI that can learn and do almost any thinking task a person can, across every field, instead of being good at one narrow job. There's no agreed test for it, and experts disagree sharply on how close it is.
Narrow AI: AI built for one specific task, like recognizing faces or recommending videos. Most AI before today's chatbots was narrow AI.
Superintelligence (ASI): AI that far surpasses the best humans at nearly everything. It's still hypothetical.
Frontier model: The most capable AI models of their time, from the leading labs. The Roundtable panel uses each company's frontier model.
Scaling: The finding that AI tends to get more capable as you add more data, computing power and model size. Whether that keeps working is one of the biggest open questions in AI.
Emergent abilities: Skills that show up in bigger models without being directly trained, like doing arithmetic or translating between languages.
Agent (agentic AI): AI that doesn't just answer but takes actions toward a goal: browsing websites, writing and running code, filling in forms, working step by step with little supervision.
Embodied AI: AI that controls a physical body, such as a robot, a car or a drone, and has to deal with the real world.
Recursive self-improvement: AI that helps build better AI, which builds better AI still, possibly speeding progress up dramatically.
The singularity: The idea of a point where technology, especially AI improving itself, changes the world so fast that what comes after can't be predicted.
p(doom): Short for "probability of doom": someone's estimate of the chance that advanced AI leads to catastrophe for humanity. Often said half-jokingly, but it's a real debate.
Stochastic parrot: A critics' phrase for language models: systems that stitch together likely-sounding words ("stochastic" means random) without truly understanding them. Whether that's fair is hotly argued.

Who's who

OpenAI: The company behind ChatGPT and the GPT models.
Anthropic: The company behind Claude.
Google DeepMind: Google's AI lab, behind Gemini.
xAI: Elon Musk's AI company, behind Grok.