Artificial intelligence
The AI jargon decoder: from tokens to neolabs
Jump to a section
The basics
Where every other term hangs off.
Artificial intelligence (AI)
The umbrella term for computers doing tasks that normally need human intelligence: understanding language, recognising images, predicting, planning, creating.
Machine learning (ML)
A branch of AI where computers learn patterns from data instead of having every rule programmed by hand.
Deep learning
Machine learning using large neural networks with many layers.
Neural network
A mathematical system loosely inspired by interconnected neurons. Modern models hold billions of adjustable values.
Generative AI (GenAI)
AI that creates new things — text, images, audio, video, software, designs — rather than only analysing what already exists.
Foundation model
A model trained broadly enough that it can later be adapted to many different tasks.
Large language model (LLM)
A model trained on huge quantities of language so it can predict, generate and manipulate text.
Small language model (SLM)
A smaller, cheaper language model. Useful when cost, speed, privacy or running on a device matters.
Multimodal model
A model that works across more than one kind of information: text, images, audio, video.
Vision-language model (VLM)
A model built to understand images and language together.
Reasoning model
A model tuned to spend more computing effort working through a hard problem before answering.
World model
A model that represents how an environment works, so it can predict what happens after an action.
Frontier model
Shorthand for the most capable models available or in development at a given moment.
Narrow AI
AI built for particular tasks rather than general intelligence.
AGI (artificial general intelligence)
The still-debated idea of AI able to handle a very broad range of intellectual tasks at or beyond human level.
ASI (artificial superintelligence)
A hypothetical AI well beyond human capability across most intellectual domains.
How a model actually works
The plumbing behind the chat box.
Token
The small chunk of information a language model reads and writes. Not quite a word: it might be a short word, part of a long one, or a punctuation mark. This is what "one million token context window" and "cost per million tokens" are counting.
Tokenisation / tokeniser
The process, and the software, that breaks text into tokens before the model sees it.
Parameter
A number inside a neural network that the model adjusts during training. "70 billion parameters" means an enormous number of learned numerical settings.
Weights
The learned numerical values inside a model. In everyday conversation, weights and parameters are used almost interchangeably.
Model architecture
The underlying design of a model: the blueprint for how its components interact.
Transformer
The architecture behind the modern LLM boom. It can spot relationships across a whole sequence rather than reading one word at a time.
Attention
The mechanism that lets a transformer work out which parts of the input matter to which other parts.
Self-attention
Attention applied within the same input sequence. In plain terms: which words should the model weigh when interpreting this word?
Embedding
A numerical representation of meaning. Embeddings turn words, paragraphs or images into lists of numbers so computers can compare similarity — useful for search, clustering, recommendations and classification.
Vector
That list of numbers, representing something as a position in mathematical space.
Vector database
A database built to store and search embeddings. It finds things that are conceptually similar rather than matching exact words.
Latent space
The abstract mathematical space in which a model represents learned concepts and relationships.
Context: what the AI can see
The single biggest source of disappointing AI output.
Context window
How much information a model can consider at once: your prompt, the conversation so far, documents, instructions and tool outputs. Bigger window, more material in play at the same time.
Context
Everything available to the model while it produces the current response.
Long context
Models able to process very large amounts of text or other information in one interaction.
Context engineering
Deciding what information an AI should receive, when it receives it and how it is structured. Prompt engineering is about the instruction; context engineering is about the model's whole information environment.
Talking to an AI
The part everyone thinks is the whole job.
Prompt
The instruction or information you give a model.
Prompt engineering
Designing prompts so they produce good output consistently.
System prompt
High-level instructions that define how an assistant should behave.
User prompt
The actual request made by the person using it.
Zero-shot prompting
Asking for a task with no example.
One-shot prompting
Giving one example.
Few-shot prompting
Giving several examples so the model can infer the pattern you want.
Chain-of-thought
Intermediate reasoning steps used while solving a problem. The idea grew into reasoning models, where extra computation goes into harder problems.
Prompt template
A reusable prompt structure with slots that get filled in automatically.
When an AI generates an answer
Speed, cost and the words behind your invoice.
Inference
Running a trained model to produce an output. Training creates the model; inference is using it. Every chatbot answer is inference happening somewhere.
Inference-time compute (test-time compute)
Letting a model spend more computation while solving a problem, rather than just making the model bigger. Central to reasoning models.
Temperature
A setting for how predictable or varied output is. Lower is more deterministic; higher introduces more variation.
Top-p
Another control over how widely the model considers possible next tokens.
Latency
How long you wait for the AI to respond.
TTFT (time to first token)
How long before the first part of a response appears.
Tokens per second
How quickly output is produced once generation starts.
Throughput
How much AI work a system can get through in a given period.
Batch processing
Sending many AI jobs through together instead of one at a time.
Rate limit
A cap on how many requests or tokens an application can send in a set period.
Training a model
How the thing got to be the way it is.
Training data
The information used to teach a model.
Dataset
An organised collection of data used to train, test or evaluate AI.
Corpus
A large body of text used for language research or model training.
Pretraining
The initial large-scale phase where a model learns patterns from enormous quantities of data.
Post-training
Everything done after that to make the model more useful, specialised, safe or capable.
Fine-tuning
Extra training that adapts an existing model to a particular task, domain or style.
SFT (supervised fine-tuning)
Fine-tuning on examples of desired inputs and outputs.
RL (reinforcement learning)
Training where a system learns behaviour from rewards or feedback.
RLHF (reinforcement learning from human feedback)
Training a model using human judgments about which outputs are better. An important technique behind conversational AI.
RLAIF (reinforcement learning from AI feedback)
The same idea, with another AI supplying some of the feedback.
Self-supervised learning
Training where the learning signal comes from the data itself rather than manual labelling.
Supervised learning
Training on examples that include known correct answers.
Unsupervised learning
Finding patterns in data with no supplied labels.
Transfer learning
Taking what was learned for one purpose and applying it to another.
Synthetic data
Artificially generated training data, often produced by other AI models.
Data annotation / labelling
Adding useful labels or information to training data, by people or by software.
Distillation
Training a smaller model to reproduce useful capabilities of a larger one: teacher abilities into a cheaper student.
Quantisation
Reducing the numerical precision used to represent a model so it needs less memory and compute.
Pruning
Removing the parts of a network that contribute little.
Sparsity
Designing models so only part of the network is active for a given task.
Mixture of experts (MoE)
A model made of specialised components, with only some "experts" activated per input.
Getting AI to use your information
The business end of most real projects.
RAG (retrieval-augmented generation)
Instead of relying only on training, the system searches another source — your documents, for instance — and hands the relevant material to the model before it answers. Search first, then ask the AI to answer using what it found.
Retrieval
Finding relevant information from a database, document set, search engine or other source.
Semantic search
Searching by meaning rather than exact keywords.
Reranking
Taking retrieved results and using another system to decide which are genuinely most relevant.
Grounding
Tying a response to trusted external information rather than the model's internal guesswork.
Knowledge base
An organised collection of information an AI system can search or reference.
When AI gets it wrong
Vocabulary for the risk conversation.
Hallucination
When a model confidently produces information that is unsupported, wrong or invented. It is not lying: the answer is statistically plausible but not grounded in reality.
Grounded answer
An answer backed by supplied or retrieved evidence.
Model drift
When behaviour or performance changes over time, often because the world around the system changed.
Bias
Systematic patterns in output that favour or disadvantage particular outcomes, perspectives or groups.
Calibration
How well a model's confidence matches how often it is actually right.
Measuring whether AI is any good
Ask for these before you sign anything.
Evaluation (eval)
A structured test that measures how well an AI system performs.
Benchmark
A standard test used to compare models.
Benchmark saturation
When models get so good at a benchmark that it stops telling them apart.
Human evaluation
People judging model output instead of relying only on automated scores.
A/B testing
Testing two versions of a model, prompt or product against each other.
Regression
When a newer version gets worse at something the old one handled fine.
AI agents
The vocabulary of the next two years.
AI agent
A system that does not just answer a question but takes actions towards a goal: searching the web, reading documents, updating software, sending messages, using tools.
Agentic AI
The broader idea of AI that can plan and take multiple actions with some autonomy.
Agentic workflow
A process where models, tools and software work through several steps to finish a larger task.
Tool use
Letting a model interact with external tools: calculators, databases, browsers, calendars, business applications.
Function calling
A structured way for a model to ask software to run a particular function.
Orchestration
The software layer coordinating models, agents, tools, data and workflows.
Multi-agent system
Several agents taking different roles or collaborating on a task.
Memory
Information stored so an application can use it next time instead of starting from scratch.
Human-in-the-loop
A setup where people stay involved in reviewing, approving or steering AI decisions.
Copilot
AI designed to work alongside a person rather than run on its own.
MCP
The acronym you will hear all year.
MCP (Model Context Protocol)
An open standard that makes it easier for AI applications to connect to external systems: data sources, tools and workflows. Think of it as a common connection standard — instead of a bespoke integration every time, systems talk through a shared protocol.
Building AI products
How software teams actually ship this.
API (application programming interface)
A way for one piece of software to talk to another. Most companies do not train giant models; they reach them through APIs.
SDK (software development kit)
A set of tools that makes building on a platform or API easier.
Endpoint
The particular API address a service is reached through.
Model provider
A company that develops models or provides access to them.
Inference provider
A company whose infrastructure runs models and serves their output.
Model router
Software that decides which model handles a request. A simple query goes to a cheap, fast model; a hard one goes somewhere more powerful.
AI wrapper
Used dismissively for software that mostly puts an interface or workflow around someone else's model. Since almost every product is built on layers of other technology, it is not always a useful distinction.
AI-native
A company or product designed around AI from the start rather than having AI bolted on later.
Horizontal AI
AI built to work across many industries or functions.
Vertical AI
AI built for one industry or specialised workflow: legal, medical, construction.
Open versus closed
Where procurement conversations get stuck.
Open source AI
A broad and sometimes contested label for AI whose software, and possibly other components, are publicly available.
Open-weight model
A more precise term: the trained weights are public. That does not mean the training data, code or process is.
Proprietary / closed model
A model controlled by an organisation that does not release its full weights or underlying technology.
Model licence
The legal terms covering how a model may be used, modified or distributed.
The hardware underneath
AI feels like software. Underneath is a heavy industry.
Compute
The processing power needed to train or run AI. "More compute" means access to more of it.
GPU (graphics processing unit)
A chip popularised for graphics that turned out to be very good at the parallel mathematics modern AI needs. Hence its place at the centre of the boom.
Accelerator
A specialised processor for running particular workloads efficiently.
TPU (tensor processing unit)
Google's specialised processors for machine-learning workloads.
NPU (neural processing unit)
A processor for accelerating AI operations, increasingly built into phones and laptops.
FLOPS
Floating-point operations per second, a measure of computational performance. Big training runs get discussed in total FLOPs.
GPU cluster
Large numbers of GPUs connected to train or serve models.
Training cluster
Infrastructure used mainly to train models.
Inference cluster
Infrastructure used to run trained models for users.
Data centre
The physical facilities holding servers, networking, power and cooling.
Model serving
The infrastructure and software that make a trained model available to people and applications.
Cloud terminology
Including the two words nobody could define last year.
Cloud computing
Renting computing resources over the internet instead of owning all the hardware.
Hyperscaler
An enormous cloud provider running infrastructure at global scale. The giant general-purpose platforms.
Neocloud
A newer kind of cloud company focused on AI infrastructure, particularly GPU-heavy workloads, rather than offering every cloud service. Specialised cloud infrastructure dedicated to AI.
Sovereign AI
The idea that a country or region should have more control over the infrastructure, data and models behind its AI capability.
AI factory
Large-scale infrastructure that turns data and compute into models, intelligence or output.
Safety and responsible AI
The governance vocabulary.
Alignment
The problem of getting a system's behaviour to match human intentions, rules or values.
Guardrails
Technical or policy controls that constrain what a system can do.
Safety layer
Systems around a model that detect or block undesirable output or actions.
Moderation
Identifying content that may breach a platform's rules or safety requirements.
Red teaming
Actively hunting for weaknesses, vulnerabilities or dangerous behaviour before someone else finds them.
Jailbreak
An attempt to talk a system into ignoring its intended restrictions.
Prompt injection
Instructions hidden inside content the system processes, trying to override its real instructions. It matters a great deal once agents browse websites, read email and use tools.
Data poisoning
Manipulating training or reference data to change a model's behaviour.
Interpretability
Research into why models behave the way they do internally.
Explainability
Methods for making AI decisions or output understandable to people.
Responsible AI
Developing and deploying AI with regard for safety, fairness, accountability, privacy, reliability and social consequences.
Physical AI and robotics
When the model has to move something.
Physical AI
AI that perceives, reasons about and acts in the physical world. Robotics and autonomous vehicles are the obvious cases.
Embodied AI
AI inside, or controlling, a physical system that interacts with its environment.
Robotics foundation model
A broadly trained model meant to support many robotic tasks rather than one programmed behaviour.
Simulation
A virtual environment where AI or robots can learn and be tested without every experiment happening for real.
Digital twin
A digital representation of a real object, facility or system that can be monitored, modelled or simulated.
And finally, what on earth is a neolab?
Yes, it is becoming a real industry term.
Neolab
A new generation of very well-funded, research-first AI startup built around a small elite technical team, chasing breakthroughs in model architecture, reasoning, robotics or scientific AI, often before there is a product or revenue. Neolab: trying to invent the next breakthrough. Neocloud: providing the computing infrastructure AI runs on.
The 30-second cheat sheet
If you remember nothing else, remember these.
