The AI Mastery Dictionary
130 terms in plain language — and the six pairs everyone confuses
Most AI glossaries are written by people who already understand AI, for people who already understand AI. This one is not. This lesson covers the vocabulary that actually unlocks the rest — what the words mean, which pairs get confused (expensively), which definitions are still being argued about, and the two habits worth more than any single term.
1 · The words that unlock everything else
A handful of terms do most of the work. An LLM is a system trained on huge amounts of text that predicts and generates language — a reasoning and prediction engine, not a search box and not a mind. The context window is how much it can hold in mind at once: think of it as the size of a desk, where once it is full, older papers fall off. Tokens are the chunks it reads and bills in, roughly three-quarters of a word each, which is why long documents get truncated and why you pay per token.
Hallucination is when the system states something false with complete confidence — a student who would rather invent an answer than admit they do not know. Grounding is the fix: giving it trusted external information so answers are rooted in reality. And retrieval-augmented generation is grounding done systematically — fetching relevant chunks of your own documents first, then answering from them. An open-book exam instead of relying on memory.
Get these five and most AI conversations stop being intimidating. Everything else is detail hung on this frame.
Temperature is the randomness dial — low for factual work, high for creative exploration. Most people never touch it and should.
2 · The six pairs everyone confuses
These are the highest-value clarifications in the whole dictionary, because each one is a mistake people make repeatedly and several are expensive. AI versus machine learning: AI is the goal, machine learning is the technique used to reach it today. Authentication versus authorisation: proving who you are, versus what you may do once inside — showing ID at the door versus whether your ticket admits you to the VIP section.
The anon key versus the service-role key is the dangerous one. Your database's public key is safe in browser code, but only because row-level security stands behind it; the service-role key bypasses that security entirely. Putting the second where the first belongs is the most common critical security mistake in AI-built apps, and it hands full access to anyone who views your page source.
Prompt injection targets your application, usually through a poisoned document your AI reads; jailbreaking targets the model's own safety rules. Data residency is where data physically sits; sovereignty is whose laws govern it — you can satisfy one and still fail the other. And encryption can be reversed with a key, while hashing cannot be reversed at all, which is precisely why passwords are hashed.
If a service can email you your existing password, they hashed nothing. Walk away.
3 · What is still being argued about
An honest dictionary flags what has not settled. Agentic AI broadly means a system that plans and takes multi-step actions rather than only answering — but every vendor defines it differently, so treat any single definition as provisional. Frontier model is used loosely, mostly in policy contexts: all frontier models are foundation models, but not the reverse.
Answer engine optimisation — structuring content so AI engines cite it — is real and worth doing, but its metrics and best practices are still forming. And llms.txt, a proposed file summarising your site for AI models, is not an official standard; studies of real traffic show almost no bots request it. Low risk, low proven benefit.
Anyone teaching these four as settled facts is overselling. Knowing which parts of a field are contested is itself a form of expertise.
When a term is contested, ask who is defining it and what they sell. That usually explains the definition.
4 · The vocabulary of building
The moment you build something real, a second vocabulary arrives. Environment variables keep secrets out of your code — PINs in a locked drawer rather than written on the wall. A webhook is an automatic message a service sends your app when an event happens: a doorbell that rings your app, rather than your app constantly checking the door.
Row-level security is the single most important setting in a Supabase project: it enforces at the database itself which rows each user may see or change — every file in the cabinet checking your ID before it opens. Without it, anyone holding your public key can read anything. And it defaults to deny-all when enabled, which means enabling it and writing the policies is the actual job.
Then the operational words: DNS propagation is the delay while an address change spreads worldwide, and it is normal rather than broken. Staging is the rehearsal copy; production is what real users touch. Test keys simulate payments; live keys move money — never ship the first, never experiment with the second.
Anything named secret, private or service-role stays server-side. Never in frontend code, never in a repo, never pasted into a chat.
5 · The words that decide what AI costs you
Spending accumulates in three ways, and confusing them is why budgets surprise people: token cost per word processed by a cloud model, compute cost for running hardware, and storage cost for keeping data. Total cost of ownership is all of it over time — not the sticker price but hardware, energy, maintenance and staff.
The two levers that actually reduce the bill are vocabulary too. Precomputation means doing expensive work once, ahead of time, so answers are ready and cheap forever after — it turns a recurring cost into a one-time one. Local models mean running on hardware you already own: no per-use bill, complete privacy, limited by your machine. Quantisation is what makes that practical, compressing a model to run on ordinary hardware the way a smaller JPEG keeps the picture.
And for any Nigerian business: the current data protection law is the Nigeria Data Protection Act 2023, not the older 2019 regulation. The Commission has confirmed the old one no longer applies. Compliance documents still citing it are out of date.
"Sufficient" is the standard when choosing a model, not "best". The savings live entirely in that one word.