Plain-language definitions of the AI terms people meet most often, as the words were used in mid-2026, ordered A to Z.
- Start with what a model is. Tokens are what it reads, attention and the transformer are how it reads them, and the weights are the model itself.
- Grounding is the main defence against invented detail. RAG searches first and has the model answer from what it found, instead of from memory.
- A benchmark is a rough guide, not proof. Scores can be gamed; an eval on your own cases tells you whether something works for you.
- Running models locally comes down to memory. VRAM is the limit, and quantization and mixture-of-experts are how larger models fit.
- Agents widen the attack surface. Prompt injection hides instructions in content a model reads, and red-teaming is how you find the gaps first.
Two to three sentences each, written for a person rather than a search engine — though the search engines are welcome. Terms are defined as used in mid-2026; the page carries a service date and will be re-serviced as the language moves.
A
Agent
A model given tools, memory, and a goal, allowed to take multiple steps on its own. The word covers everything from a script that retries twice to software that plans a workday.
Agentic workflow
A pipeline where the model decides the next step — searching, calling tools, revising — instead of a human sequencing every action.
Alignment
The training and tuning work that makes a model behave according to human intent and values: helpful, honest, refusing harm. The finishing pass after raw capability.
API
The programmatic doorway to a model: send a request with your prompt and settings, receive the model output. How software talks to AI without a chat window.
Attention
The transformer mechanism that lets every token weigh its relevance to every other token. The reason context matters at all — and the load-bearing idea of modern AI.
B
Batch processing
Running many requests together, typically slower but cheaper. The economical lane for jobs that do not need an instant answer.
Benchmark
A standardized test set used to compare models. Useful directionally; notoriously gameable; never a substitute for testing on your own cases.
Treat a benchmark score as a hint about a model, then run an eval on your own cases before you rely on it.
C
Chain of thought
A model working through intermediate reasoning before answering. On newer reasoning-trained models this happens internally — asking for it explicitly is often redundant.
Checkpoint
A saved snapshot of a model at a point in training. What you actually download when you download a model.
Context engineering
The successor skill to prompt engineering: assembling everything a model sees — instructions, documents, examples, tools, memory — as one designed working set.
Context window
The maximum amount of text, in tokens, a model can consider at once. Has grown from a few thousand tokens in early models to far larger windows; check the current figure for the model you use. Ordering within it now matters more than fitting.
D
Deepfake
Synthetic audio, image, or video convincingly depicting a real person doing something they did not do. The reason provenance tools exist.
Diffusion model
The architecture behind most image and video generation: it learns to turn noise into pictures step by step, steered by your description.
Distillation
Training a smaller model to imitate a larger one — cheaper and faster to run, at some cost in capability. How much of the on-device world gets made.
E
Embedding
A list of numbers representing meaning. Texts with similar meaning get nearby numbers, which is what makes semantic search and RAG possible.
Eval
A repeatable test of model output quality against your own cases and rubric. The difference between it-seems-fine and knowing.
F
Fine-tuning
Continuing a model training on your own examples so it adopts a style, format, or domain. Heavier than prompting; lighter than pretraining.
Frontier model
The most capable model class available at a given moment, usually cloud-first. The moving line everything else is measured against.
Function calling
A model responding with a structured request to run a tool — search, database, calculator — instead of prose. The joint that makes agents possible.
G
GPU
The parallel processor that trains and often runs models. Its memory, VRAM, is usually the binding constraint for running models locally.
Grounding
Tying a model answer to supplied or retrieved sources instead of its trained memory. The main defense against fabricated detail.
Guardrails
The layer of rules, filters, and checks around a model that constrain what it will produce, separate from the model’s own training.
H
Hallucination
Confident, fluent output that is factually wrong or invented. Not lying — the model has no concept of truth, only likelihood. Verify accordingly.
Human-in-the-loop
A workflow with a mandatory human checkpoint before output ships. What separates using these tools from adding to the flood.
I
Inference
Running a trained model to get output — as opposed to training it. What you pay for per token; where the datacenter electricity goes after launch.
J
Jailbreak
A prompt crafted to trick a model past its safety training. An arms race this site documents rather than participates in.
L
Latency
Time until the model starts and finishes responding. Network adds a few hundred milliseconds to cloud calls; on-device eliminates the round trip.
Local LLM / on-device AI
A model running entirely on your own hardware: private by construction, free per token, capped by your memory and the model size.
LoRA
Low-Rank Adaptation: small trainable add-on weights that customize a model cheaply — the standard way styles and subjects are added to open image models.
M
Mixture of Experts (MoE)
An architecture that activates only relevant portions of a large model per request — big-model capability at smaller running cost. Why capable local models became feasible.
Model collapse
Degradation that occurs when models train on machine-made copies of machine output. The Quarantine wing, stated as an equation: what trains on copies, dies of them.
Multimodal
A model that takes or produces more than text: images, audio, video, documents. Standard equipment on frontier systems now.
N
NPU
Neural processing unit — the accelerator in modern phones and laptops built specifically to run models efficiently on-device.
O
Open weights
A model whose trained parameters are downloadable. Licenses vary from fully permissive to research-only; the license, not the adjective open, is what matters.
Orchestrator
The controlling layer — human or model — that breaks a goal into tasks, assigns them, and reviews results. How large AI work is actually managed.
P
Pretraining
The long, expensive first phase: the model learns language and world patterns from a vast corpus. The expensive part, done once before any fine-tuning.
Prompt injection
Malicious instructions hidden inside content a model reads — a webpage, an email — attempting to hijack its behavior. The classic attack on agents.
Provenance / C2PA
Cryptographic metadata attached to media recording what made it and how it was edited. The emerging infrastructure of is-this-real.
Q
Quantization
Compressing model weights to lower precision so they fit in less memory and run faster, at a small quality cost. The magic that puts big models on laptops.
R
RAG
Retrieval-Augmented Generation: search first, then have the model answer from what was found. The standard pattern for grounding answers in your documents.
Rate limit
The provider-imposed cap on how fast you can call an API. The reason production systems queue, retry, and batch.
Reasoning model
A model trained to deliberate internally before answering — slower, costlier, stronger on hard problems. Brief it with goals and budgets, not step-by-step scripts.
Red-teaming
Systematically attacking your own model or app — jailbreaks, injections, edge cases — before someone else does.
RLHF
Reinforcement Learning from Human Feedback: humans rate outputs, and the model is tuned toward what rated well. A core alignment technique; manners, made load-bearing.
S
Scaling laws
The empirical finding that capability rises predictably with more data, parameters, and compute. The economic engine of the entire buildout.
Schema / structured output
Asking a model for machine-readable output in an exact shape, usually JSON, and validating it. In 2026 the output contract is part of the prompt.
Semantic search
Search by meaning, via embeddings, rather than exact words. Finds renal failure when you searched kidney problems.
Slop
Low-effort machine-generated content published at volume for clicks and ad revenue. A problem this site covers — distinct from AI itself, which is the tool.
Streaming
Receiving a model answer token-by-token as it generates, rather than waiting for completion. Why chat interfaces feel alive.
Synthetic data
Machine-generated training data. Useful deliberately and in moderation; catastrophic as an accidental diet — see model collapse.
System prompt
The standing instructions a model receives before any user message: role, rules, boundaries. Where durable behavior lives.
T
Temperature
The randomness dial. Low means consistent and repetitive; high means varied and risky. Check the vendor’s advice before changing it: on Gemini 3.x Google recommends leaving it at the default (sourced here).
Token
The unit models actually read — roughly three-quarters of an English word. Context windows, prices, and speeds are all measured in these.
Training run
The weeks-to-months campaign of pretraining a model across thousands of accelerators. The single most expensive object in modern software.
Transformer
The 2017 attention-based architecture underlying essentially all modern language models and much else. The frame every floor is bolted to.
V
Vector database
A database built to store embeddings and find nearest-meaning matches fast. The retrieval half of RAG at scale.
VRAM
The GPU’s own memory. The hard wall that decides which models your hardware can hold; quantization exists to negotiate with it.
W
Watermarking
Embedding a detectable signature in generated output to mark it as synthetic. Technically fragile, politically inevitable, actively evolving. Claude models in use today, including Opus 5 and Opus 5.5, already watermark their text to comply with the EU AI Act; Anthropic’s help page lists Fable 5.1, Mythos 5.1, Opus 5.5 and Opus 5, with older models due by 2 December 2026. Only organisations granted access can check for the mark for now, and a detected mark is a signal, not proof. More on Claude’s text watermark.Claude Help Center, How Claude marks AI-generated content, read at source 23 Sep 2026: “Generated text will carry embedded watermarks”; “Watermark detection is currently in private preview, available to eligible organizations as required under EU law (such as regulators, law enforcement, media, fact-checkers, independent researchers, educational organizations, and EU civil society groups)”; “A detected mark provides a signal that content was processed by Claude, but is not fully conclusive.” Its model table, read the same day, ticks text watermarks for Fable 5.1, Mythos 5.1, Opus 5.5 and Opus 5, and says “Anthropic is adding watermarks to outputs from models released before August 2, 2026, with all covered by December 2, 2026.”Anthropic, Introducing Claude Opus 5.5, 22 Sep 2026, read at source 23 Sep 2026: “As with Fable 5.1, Opus 5.5 comes with our watermarking measures to comply with the EU AI Act”.Anthropic, How Claude’s text watermark works, 14 Aug 2026, read at source 22 Sep 2026: “Future Claude models will generate text that contains a watermark. This is a way of determining the likelihood that Claude was involved in writing the text, and we, along with several other major AI providers, are implementing this change to comply with the EU AI Act.”First-hand: until 23 Sep 2026 this entry said Anthropic’s post “describes future models, not current ones”. That was wrong when written: Anthropic, Claude Fable and Mythos 5.1, September 2026, read at source 22 Sep 2026: “This required us to add a watermark” … “to the outputs of models released after August 2, 2026.”
Weights
The billions of learned numbers that ARE the model. Everything else — apps, chat windows, APIs — is furniture around them.
Z
Zero-shot / few-shot
Asking with no examples versus a handful. Examples remain one of the highest-leverage prompt moves per token spent.
Part of the Stay Human record. Practice the vocabulary in the syllabus; watch it get poured into being at The Assembly.