The AI
glossary.
Every AI word that matters, in plain English — each one with an analogy to make it stick, and a link to the full explanation. Skim it, bookmark it, or send it to anyone who's tired of nodding along to jargon they don't actually understand.
No hype, no equations. If a term shows up in an AI conversation, a pricing page, or a news headline, it's here — defined once in a sentence, then again as a picture you can hold in your head. Terms with a deeper page get a → deep dive link.
The core five
Model
Plain: a mathematical function that predicts the next chunk of text — the thing people mean when they say "an AI."
Picture it: an extremely well-read autocomplete. → deep dive
Parameters / Weights
Plain: the billions of tuned numbers that make up a model; each one is basically "how much this input matters." The model is its weights.
Picture it: a mixing board with billions of sliders — the slider positions are the weights. → see the animation
Token
Plain: the small chunk of text a model reads and writes — roughly ¾ of a word. Everything is measured in tokens.
Picture it: LEGO bricks of language; the model builds and reads only in these little pieces. → deep dive
Context window
Plain: how many tokens the model can consider at once — its short-term memory.
Picture it: the size of its desk; only what fits on the desk can be used right now. → deep dive
Prompt
Plain: the text you give the model to act on — your instructions plus any context.
Picture it: the brief you hand a contractor. A clear brief gets a better build.
How a model is made and used
Training
Plain: tuning a model's weights by showing it mountains of text until its guesses get good.
Picture it: adjusting all the mixing-board sliders, over and over, until the mix finally sounds right. → deep dive
Loss
Plain: a number for how wrong a guess was; training drives it down.
Picture it: a golf score — lower is better, and training is practicing to shoot lower. → deep dive
Pretraining vs Fine-tuning
Plain: pretraining builds the general brain from scratch (expensive, one-time); fine-tuning cheaply specializes an existing model.
Picture it: a full general education vs a weekend certification course. → deep dive
Inference
Plain: actually running a trained model to produce an answer, one token at a time. The ongoing cost of AI.
Picture it: the trained musician now performing — versus the years of practice that got them there. → deep dive
Temperature
Plain: how much randomness the model uses when choosing the next token.
Picture it: a creativity dial — low is safe and repeatable, high is wild and varied.
Hallucination
Plain: when a model states something false with total confidence.
Picture it: the student who never says "I don't know" and just confidently makes something up.
The chips it runs on
GPU
Plain: the parallel-math chip that's the AI workhorse — thousands of cores working at once.
Picture it: thousands of interns each doing simple arithmetic simultaneously (a CPU is one genius doing it in order). → deep dive
FPGA
Plain: a chip you can rewire after it's manufactured, for low-latency custom work.
Picture it: a workshop of reconfigurable machines you rebuild for each new job. → deep dive
ASIC
Plain: a chip permanently etched for exactly one job — fastest and most efficient at it, useless for anything else.
Picture it: a factory line built to make one product superbly. → deep dive
TPU / NPU
Plain: ASIC-style AI accelerators — the TPU is Google's neural-network chip; the NPU is the little AI chip in your phone or laptop.
Picture it: a purpose-built appliance rather than a general-purpose computer. → deep dive
How models are sold and controlled
Frontier model
Plain: the biggest, most capable models (GPT, Claude, Gemini), reached through an API and paid per token.
Picture it: renting time on the best brain money can buy. → deep dive
Local / open model
Plain: a model you run on your own hardware — private, with no per-token meter.
Picture it: owning the brain outright instead of renting it. → deep dive
Open weights / open source
Plain: a model whose trained weights are published for anyone to download, run, and fine-tune.
Picture it: the finished dish (and sometimes the recipe) handed out for free. → deep dive
Closed source
Plain: a proprietary model whose weights are secret and offered only through an API.
Picture it: a secret-recipe dish you can only taste at the restaurant. → deep dive
Following the money
Input / output tokens
Plain: input is what you send the model; output is what it writes back. You pay for both, and output usually costs more.
Picture it: a taxi that charges for the miles in and the miles out — with the ride home priced higher. → deep dive
Cost per task
Plain: what it actually costs to finish a job — not the price per token, but the total once you count how many tokens and retries it took.
Picture it: the full grocery bill vs the price per item; quantity is what gets you. → deep dive
Token efficiency
Plain: how much useful work a model does per token ("intelligence density"). A more efficient model gets there with fewer tokens.
Picture it: the person who answers in one sharp sentence vs the one who takes three rambling paragraphs. → deep dive
Jevons paradox
Plain: when something gets cheaper to use, people use far more of it — so cheaper AI raises total demand for compute, not less.
Picture it: wider, faster highways that somehow fill with even more traffic. → deep dive
Distillation
Plain: training a smaller or new model on the outputs of a bigger one, so it learns to imitate the bigger model.
Picture it: an apprentice getting good by watching a master work.
What models can be wired up to do
LLM (Large Language Model)
Plain: a model trained on huge amounts of text to predict tokens — the technology behind assistants like ChatGPT and Claude.
Picture it: the engine; the chat app is the car built around it.
Multimodal
Plain: a model that handles more than text — images, audio, or video too.
Picture it: giving the model senses beyond reading — now it can see and hear.
RAG / Retrieval
Plain: fetching the few relevant documents and feeding them to the model at question time so it answers from the right source.
Picture it: an open-book exam — it looks up the correct page before answering.
Agent / tool use
Plain: a model set up to take actions — call tools, browse the web, run code — not just chat.
Picture it: giving the brain a set of hands.
Want any of these turned into something your business actually uses — not just understands? That's the whole job. Let's talk →