Six groups, 43 words.
The basics
What AI, models and machine learning actually are.
Large language model (LLM)
An AI model trained on huge amounts of text to predict the next piece of text, which lets it answer questions, write and summarize. ChatGPT, Claude and Gemini are built on LLMs.
Source: Google for DevelopersFoundation model
A large general-purpose model trained on a broad mix of data, which other products and more specialized models are then built on.
Source: Google for DevelopersArtificial intelligence (AI)
Computer systems that do things we would normally say need human intelligence, such as understanding language, recognizing images, making predictions or making decisions.
Source: EU AI ActModel
The trained program at the heart of an AI product: a large set of learned numbers that turns an input (your question) into an output (its answer). ChatGPT is a product; GPT-5 is a model.
Source: Google for DevelopersNeural network
The kind of model behind modern AI: layers of simple maths units, loosely inspired by brain cells, connected by adjustable weights.
Source: Google for DevelopersMachine learning
The part of AI in which computers learn patterns from examples and data instead of being given step-by-step rules.
Source: Stanford HAIDeep learning
Machine learning using very large neural networks with many layers; the approach behind today's chatbots, image generators and speech recognition.
Source: Stanford HAINarrow AI
AI built for one kind of task, such as recognizing faces or transcribing speech, as opposed to general-purpose AI.
Source: Stanford HAIGenerative AI
AI that creates new content, such as text, images, audio, video or code, in the style of the data it was trained on.
Source: NIST AI 600-1Multimodal
A model that can handle more than one kind of input or output, such as text, images, audio and video, rather than text alone.
Source: IBM ThinkUsing AI
The words you meet when you type into a chatbot or hand it a job.
Context window
How much text (counted in tokens) a model can take in at once, including your question, any documents you give it and the conversation so far. Anything beyond it is not seen.
Source: Google for DevelopersHallucination
When an AI model states something false or made up as if it were true, such as an invented quote, fact or court case.
Source: Google for DevelopersPrompt
The text (or image, or file) you give an AI model to tell it what you want.
Source: Google for DevelopersChain of thought
Step-by-step working that a model writes out before giving its answer. Reasoning models do this by default.
Source: Google for DevelopersAI agent
An AI system that does not just answer but takes actions for you, such as browsing websites, running code or filling in forms, over several steps.
Source: Google for DevelopersToken
The small chunk of text (often part of a word) that a model reads and writes. Prices, limits and context windows are counted in tokens.
Source: OpenAI Help CenterInference
Running a trained model to get answers, as happens every time someone uses a chatbot. Contrasts with training, which builds the model.
Source: IBM ThinkRetrieval-augmented generation (RAG)
A way of making a chatbot look things up in a set of documents or a database before it answers, so its answer can draw on, and cite, that material.
Source: IBM Research blog (Kim Martineau)Reasoning model
A model trained to work through a problem step by step (a hidden or visible chain of thought) before answering. It is slower and costlier, but usually better at maths, code and logic.
Source: OpenAIHow it's built
Training, weights, chips and the data centers that run them.
Fine-tuning
Extra training given to an already-trained model to make it better at a particular job, style or subject.
Source: Google for DevelopersDistillation
Training a smaller, cheaper model to copy the behavior of a bigger one, so it runs faster at some loss of ability.
Source: Google for DevelopersParameters (weights)
The numbers inside a model that are adjusted during training and hold what it has learned. Big models have hundreds of billions or more.
Source: Google for DevelopersTraining data
The text, images, code and other material a model learns from during training.
Source: EU AI ActTraining run
One long job, often lasting weeks to months on thousands of chips, in which a model learns from its training data. Its size is usually measured in FLOP of compute.
Source: EU AI ActCompute
Computing power: the chips, and the amount of calculation they do, used to train and run AI. It is often counted in FLOP.
Source: EU AI ActFLOP (floating-point operation)
One basic arithmetic step on decimal numbers. The total FLOP used to train a model (e.g. 10^25) is a standard way to say how big a training run was.
Source: EU AI ActGPU
A graphics processing unit: a chip first built for video games that can do huge numbers of calculations in parallel, which makes it the workhorse for training and running AI.
Source: IBM ThinkData center
A building full of computers, storage and networking equipment. The largest AI data centers hold hundreds of thousands of chips and draw hundreds of megawatts of power.
Source: IBM ThinkOpen-weight model
A model whose trained parameters (weights) are published for anyone to download and run or modify. It is not always fully 'open source', since training data and code may stay private.
Source: NTIAFrontier model
One of the most capable AI models at a given time, usually from the biggest labs, and the main focus of safety rules and testing.
Source: GoogleTesting and safety
How AI is measured, attacked and checked.
Benchmark
A fixed test, such as a set of exam questions or coding tasks, used to score and compare AI models. Scores can mislead if models were trained on the answers or the test is too easy.
Source: IBM ThinkDeepfake
Fake but realistic video, audio or images of real people, places or events, made or altered with AI.
Source: EU AI ActAlignment
Work to make AI systems pursue the goals and values their makers and users intend, and not harmful or unintended ones.
Source: IBM ThinkRed-teaming
Deliberately attacking or trying to misuse an AI system before release, to find dangerous or embarrassing failures so they can be fixed.
Source: US Executive Order 14110Jailbreak
A prompt crafted to trick a model into ignoring its safety rules, for example to get instructions it is meant to refuse.
Source: NIST AI 600-1Prompt injection
An attack in which hidden or malicious instructions, typed in or planted in a web page, email or file the AI reads, make it do something its user did not intend.
Source: OWASP Top 10 for LLM Applications 2025System card (model card)
A report a lab publishes with a model, describing what it can do, how it was tested, safety results and known limits. The idea comes from 'model cards'.
Source: Mitchell et al.Rules
Terms that come from the laws now in force.
General-purpose AI model
The EU AI Act's legal term for a model that can be used for many different tasks, like the ones behind ChatGPT, Claude or Gemini. Makers of these models have duties under the Act.
Source: EU AI ActSystemic risk
In the EU AI Act, risks from the most powerful general-purpose models that could have large effects across society, which bring extra testing and reporting duties.
Source: EU AI ActDual-use foundation model
A US government term (2023) for a broad, very large model that could be used for good purposes or to cause serious harm, for example in cyberattacks or weapons.
Source: US Executive Order 14110AI Act
The European Union's law on artificial intelligence. It bans some uses outright, sets rules for 'high-risk' uses such as hiring and credit, and puts duties on makers of general-purpose models, with rules coming in stages from 2025 to 2028.
Source: European CommissionThe big ideas
What people mean by AGI and superintelligence.
Superintelligence watchAGI (artificial general intelligence)
AI that can do most of the thinking work people do, at least as well as people. There is no agreed test, so claims that it has arrived depend on whose definition is used.
Source: OpenAI CharterSuperintelligence
A hypothetical AI far smarter than the best humans in almost every field, not just a match for them.
Source: Nick BostromVideos on how AI works.
Talks, interviews and explainers, each with our short summary. They play in YouTube's own player.
All videos on how AI worksMIT 6.S191 (2025): Large Language Models (Google)
Peter Grabowski of Google, who leads an applied research team working on Gemini, gives a guest lecture on large language models as part of MIT's 2025 Introduction to Deep Learning course.
Anthropic · Mar 27, 2025Tracing the thoughts of a large language model
Anthropic introduces two interpretability research papers that map the internal circuits behind its Claude model's thinking. One example suggests the model picks rhyme words before writing the line that leads to them, which hints that it plans ahead even though it outputs one word at a time.
Made by the company itselfAlexander Amini · Mar 3, 2025MIT Introduction to Deep Learning (2025) | 6.S191
Alexander Amini gives the opening lecture of MIT's 2025 Introduction to Deep Learning course (6.S191), on the foundations of deep learning and neural networks.