Glossary Wiki
Key terms in AI, explained simply.
A
- Agent
- An AI system that can take actions on its own, such as using tools, browsing, or running code, to complete a goal rather than just answering in text.
- AGI (Artificial general intelligence)
- A hypothetical AI that can learn and perform any intellectual task a human can, rather than being good at just one narrow job.
- AI safety
- The field working to keep AI systems from causing harm, from today's misuse to risks from far stronger future systems.
- AI slop
- Low-effort machine-generated content flooding the web, recognisable by its confident blandness.
- Air-gapped
- A machine kept fully off the network; the strictest way to guarantee an AI tool cannot send anything anywhere.
- Alignment
- The work of making sure an AI system actually does what people intend and holds to human values, especially as it becomes more capable.
- API (Application programming interface)
- The doorway one program uses to talk to another; AI providers sell model access through APIs priced per token.
- Artificial intelligence (AI)
- Software that performs tasks we usually associate with human intelligence, like understanding language, recognising images, or making decisions.
- Attention
- The mechanism inside a transformer that lets a model weigh which earlier words matter most when predicting the next one.
B
- 1-bit model
- A model whose weights are stored as single bits instead of 16, shrinking a 27B model to a few gigabytes with modest quality loss.
- Backpropagation
- The method that works out how much each weight contributed to an error, so gradient descent knows what to adjust.
- Base model
- A model fresh from pre-training, before instruction tuning; it continues text rather than answering questions.
- Benchmark
- A standard test or dataset used to measure and compare how well AI models perform on a task.
- Bias
- Systematic skew in a model's outputs, usually learned from patterns in its training data, that can make results unfair or inaccurate.
C
- Catastrophic forgetting
- When fine-tuning on new material erases skills a model already had.
- Chain of thought
- A prompting style, or model behaviour, where the AI works through a problem step by step instead of jumping straight to an answer.
- Checkpoint
- A saved snapshot of a model's weights during training, used to resume, compare or roll back.
- Chunking
- Splitting documents into pieces small enough for retrieval and context windows; the size chosen quietly shapes answer quality.
- Closed model
- A model reachable only through its maker's API, with weights, training data and inner workings kept private.
- Compute
- The raw processing power, usually measured in chips and hours, used to train or run an AI model.
- Context window
- The amount of text a model can consider at once, including your prompt and its own reply. Anything beyond it is forgotten.
D
- Data labeling
- The human work of tagging training examples, telling the model what counts as correct.
- Dataset
- The collection of text, images or other examples a model learns from; its quality and coverage shape what the model can do.
- Deepfake
- AI-generated audio, image or video that convincingly imitates a real person.
- Dense model
- A model that uses all its parameters for every token, unlike mixture-of-experts designs that activate only a few.
- Diffusion model
- A model that creates images, audio, or text by starting from random noise and gradually refining it into a clear result.
- Distillation
- Training a smaller model on the outputs of a stronger one so it inherits capability at a fraction of the cost.
E
- Embedding
- A list of numbers that represents the meaning of a word, sentence, or image so a computer can compare how similar two things are.
- Embodied AI
- AI that acts through a physical body, usually a robot, closing the loop between perception and action.
- Emergent capabilities
- Skills that appear once models pass a certain scale without being trained for directly, like arithmetic or translation.
- Epoch
- One full pass through the training dataset; models usually train for several.
- Evals
- Structured tests that measure what a model can actually do, from exam questions to full task simulations.
F
- Federated learning
- Training one shared model across many devices where the data stays put and only weight updates travel.
- Few-shot learning
- Getting a model to do a task by showing a handful of examples in the prompt, with no retraining.
- Fine-tuning
- Taking an already trained model and training it a bit more on a smaller, specific dataset so it is better at a particular task or style.
- Foundation model
- A large model trained on broad data that can be adapted to many different tasks, serving as a base others build on.
- Frontier model
- One of the most capable models available at a given moment, usually from a handful of large labs.
G
- GAN (Generative adversarial network)
- A setup where two networks compete, one creating fake data and one judging it, which pushes the creator to produce realistic results.
- GGUF
- The file format local runners use for models, packing weights and metadata into one portable, often quantized file.
- GPU (Graphics processing unit)
- A chip originally built for graphics that turned out to be ideal for the parallel maths behind training and running AI models.
- Gradient descent
- The core training algorithm: nudge each weight a little in the direction that reduces the model's error, millions of times.
- Ground truth
- The verified correct answer a model's output is judged against.
- Guardrails
- Rules and filters added around a model to stop it producing harmful, unsafe, or off-limits responses.
H
- Hallucination
- When an AI states something false or made up as if it were true, because it predicts plausible text rather than checking facts.
- Hyperparameters
- The settings chosen before training or generation, like learning rate or temperature, tuned by people rather than learned.
I
- In-context learning
- A model's ability to pick up a pattern from examples inside the prompt and apply it immediately, without weight updates.
- Inference
- The act of running a trained model to get an answer. This is what happens every time you send it a prompt.
- Instruct model
- A base model after instruction tuning, trained to follow requests and hold conversations; the kind most people actually use.
- Instruction tuning
- Fine-tuning a base model on example instructions and answers so it learns to follow requests instead of just continuing text.
- Interpretability
- The study of what is actually happening inside a model's weights and activations, and why it produced what it did.
J
- Jailbreak
- A prompt crafted to push a model past its safety rules and produce output it was trained to refuse.
K
- Knowledge cutoff
- The date a model's training data ends; anything after it the model cannot know without being told.
- KV cache
- Memory that stores what a model has already read in a conversation so it is not recomputed for every new token; it grows with context length.
L
- Large language model (LLM)
- A model trained on huge amounts of text that predicts and generates language, powering tools like chatbots and writing assistants.
- Latency
- The wait between sending a request and getting a response; for chat, the wait before the first word appears.
- Latent space
- The internal map where a model places concepts, where distance means similarity and directions can carry meaning.
- Local AI
- Running models on hardware you control instead of a cloud API: private, offline-capable and free of per-token bills.
- Logits
- The raw scores a model assigns every possible next token before they are turned into probabilities.
- LoRA (Low-rank adaptation)
- A fine-tuning method that trains a small add-on layer instead of the whole model, cutting memory and cost sharply.
- Loss function
- The score a model tries to minimise during training; it measures how far predictions are from the right answers.
M
- Machine learning
- A branch of AI where systems learn patterns from data and improve with experience, instead of being programmed with fixed rules.
- MCP (Model Context Protocol)
- An open standard for connecting an AI assistant to outside tools and data, so a connector written once works in any compliant app.
- Memory bandwidth
- How fast a chip moves data to its processors; for local text generation it matters more than raw compute.
- Mixture of experts (MoE)
- A model design that routes each input to a few specialised sub-networks instead of using the whole model, saving compute.
- Model card
- A model's spec sheet: what it was trained on, what it is for, its limits and its licence.
- Model extraction
- Rebuilding a rival model's capability by harvesting its outputs at scale, usually against the provider's terms.
- Model merging
- Combining the weights of two fine-tuned models into one, a common local-scene way to blend abilities without training.
- Model server
- A program, like LM Studio or Ollama, that loads a model and serves it to other apps through an API.
- Multi-agent system
- A setup where several AI agents work on one task, splitting roles like researcher, writer and reviewer.
- Multimodal
- An AI that can work with more than one kind of input or output, such as text, images, audio, and video together.
N
- Neural network
- A model loosely inspired by the brain, made of layers of connected units that adjust as they learn from data.
- NPU (Neural processing unit)
- A chip built specifically for AI maths, common in phones and new laptops, using far less power than a GPU.
O
- Offloading
- Splitting a model between fast GPU memory and slower system RAM when it does not fully fit, trading speed for size.
- On-device AI
- Running models directly on a phone or laptop, so data stays local and nothing depends on a connection.
- Open source
- Software whose source code is free to inspect, modify and share; for models the honest term is often open weights instead.
- Open weights
- A model whose trained parameters are released publicly so anyone can download, run, and adapt it, though the training data may not be shared.
- OpenAI-compatible API
- The de facto standard interface for chat models; tools that speak it can swap between providers and local servers freely.
- Overfitting
- When a model memorises its training data too closely and performs well there but poorly on new, unseen examples.
P
- Parameters
- The internal numbers a model adjusts during training. More parameters can mean more capacity, and counts like 70B refer to these.
- Phoning home
- When software contacts its maker's servers without the user asking; caught by watching its network traffic.
- Post-training
- Everything done to a model after pre-training, from instruction tuning to preference training, turning raw prediction into a product.
- Pre-training
- The first, broad training stage where a model learns general patterns from a very large dataset before any task-specific tuning.
- Prompt
- The instruction or question you give an AI model to tell it what you want.
- Prompt caching
- Reusing the processed form of a repeated prompt prefix so long system prompts stop costing full price on every call.
- Prompt engineering
- The craft of wording and structuring prompts to get more useful, accurate, or reliable results from a model.
- Prompt injection
- An attack that hides instructions inside content a model reads, tricking it into following the attacker instead of the user.
- Prompt steganography
- Hiding a signal inside normal-looking prompt text, as with invisible Unicode characters that flag a session without the user seeing.
Q
- Quantization
- Shrinking a model by storing its numbers at lower precision, so it uses less memory and runs on smaller hardware, with a small quality trade-off.
R
- RAG (Retrieval-augmented generation)
- A technique where a model looks up relevant documents and uses them to answer, so it can cite sources and stay current.
- Reasoning model
- A model trained to spend extra steps thinking through a problem before answering, which improves results on maths, coding, and logic.
- Reasoning tokens
- The hidden thinking a reasoning model produces before its visible answer; they cost time and budget like normal output.
- Red teaming
- Deliberately probing a model to find weaknesses, unsafe outputs, or ways it can be misused, so they can be fixed before release.
- Reinforcement learning
- Training a model through trial and error, rewarding good outcomes and penalising bad ones, so it learns a useful strategy.
- RLHF (Reinforcement learning from human feedback)
- A training method that uses human ratings of model answers to teach it to be more helpful, honest, and safe.
S
- Safetensors
- A file format for model weights that cannot hide executable code, made to be safe to download and load.
- Sampling
- How a model picks each next token from its probability list; settings like temperature make the choice safer or more varied.
- Sandbox
- An isolated environment where an agent's code and commands run with limited access, so mistakes cannot touch the wider system.
- Scaling laws
- The observed pattern that model performance improves predictably as you increase data, parameters, and compute together.
- Seed
- The starting number for a generator's randomness; fixing it makes the same prompt reproduce the same output.
- Self-hosting
- Running software, including AI models, on your own machines instead of renting the hosted version.
- Self-supervised learning
- Training where the data labels itself, as when a model learns by predicting the next word of ordinary text.
- Semantic search
- Search that matches meaning rather than exact words, so a query about cars finds documents about automobiles.
- Small language model (SLM)
- A compact model, roughly under 10 billion parameters, built to run fast on phones and laptops.
- Speculative decoding
- A speed trick where a small model drafts tokens and the big model checks them in bulk, with identical output.
- Speech recognition (ASR)
- Turning spoken audio into text; local models like Whisper made it free and private.
- State of the art (SOTA)
- The best published result on a benchmark at a given time; a moving target in AI.
- State-space model (SSM)
- A neural architecture that handles long sequences with far less memory than attention, used alongside it in hybrid models.
- Streaming
- Sending a model's answer word by word as it is generated instead of waiting for the whole reply.
- Superintelligence
- A hypothetical AI that far surpasses the best human minds across essentially all fields.
- Supervised learning
- Training on examples paired with correct answers, so the model learns to map one to the other.
- Synthetic data
- Data generated by a model or simulation, rather than collected from the real world, used to train or test other models.
- System prompt
- The hidden instructions an app gives a model before the conversation starts, setting its role, rules and tone.
T
- Telemetry
- Data software sends home about how it is used; some AI tools ship it switched on by default with no visible off switch.
- Temperature
- A setting that controls how random a model's output is. Low values make it focused and predictable, high values make it more varied.
- Test-time compute
- Extra thinking a model spends at answer time instead of training time; the idea behind reasoning models.
- Text-to-speech (TTS)
- Generating spoken audio from text; modern local systems produce natural voices on ordinary hardware.
- Time to first token
- How long a model takes to start answering; long prompts stretch it because the model must read everything first.
- Token
- The small chunk of text, often a word piece, that a model reads and generates. Models measure input and cost in tokens.
- Tokenization
- The step of splitting text into tokens so a model can process it.
- Tokens per second (tok/s)
- The standard speed measure for text generation; around 10 to 15 reads like natural typing pace.
- Tool calling
- A model's ability to trigger outside functions, like a search or a script, by emitting a structured request instead of plain text.
- Top-k sampling
- A sampling setting that limits the model's next-token choice to the k most likely options.
- Top-p (nucleus sampling)
- A sampling setting that limits choices to the smallest set of tokens whose probabilities add up to p, trimming unlikely words.
- Training
- The process of feeding data to a model so it adjusts its parameters and learns to perform a task.
- Transfer learning
- Reusing what a model learned on one task as the starting point for another, instead of training from zero.
- Transformer
- The neural network architecture behind most modern AI, which uses attention to handle sequences like text efficiently.
- Turing test
- Alan Turing's 1950 test of whether a machine's conversation can pass as human.
U
- Unified memory
- A design where CPU and GPU share one pool of RAM, letting Macs load models that would need a big graphics card elsewhere.
- Unsupervised learning
- Training on raw data with no labels, leaving the model to find structure on its own.
V
- Vector database
- A database built to store embeddings and quickly find the items most similar in meaning to a query. Often used with RAG.
- Vibe coding
- Building software by describing what you want to an AI and accepting its code with little review, prized for speed and blamed for bugs.
- Vision-language model (VLM)
- A model that reads images and text together, so it can describe screenshots or answer questions about a photo.
- VRAM (Video RAM)
- The memory on a graphics card; on most PCs it caps the size of model you can run.
W
- Watermarking
- Embedding a hidden, detectable pattern in AI output so machines can later tell it was generated.
- Weights
- Another word for a model's learned parameters, the values that determine how it turns an input into an output.
- World model
- An AI's internal picture of how an environment behaves, letting it predict outcomes before acting.
Z
- Zero-shot learning
- Asking a model to do a task with no examples at all, relying on what it absorbed in training.