AI Tech Fundamentals for Non‑technical Decision Makers Part 1 of 5
Written for people who have to make business decisions involving AI-enabled systems. No mathematics, no code, and no assumption that you have read anything else first.
What exactly is a model, and what kinds are there?
The object
A very large array of numbers, plus a modest amount of code to run them. That's the whole thing.
Almost everyone pictures something more. It is not a program with logic you could read through. It is not a service. It is not a database of answers. Three properties follow that explain a great deal of what comes later:
| Property | What it means |
|---|---|
| It is frozen | The numbers do not change while you use it. Nothing you do alters the model. |
| It is identical for everyone | Your copy is not your model. There is no personalization inside the numbers — every customer is served from the same ones. |
| It starts fresh every time | It does not know that you spoke to it a minute ago, or that you exist. |
So everything that feels animate is built around it. Memory, personalization, knowing today’s date, looking things up — none of that is in the model. All of it is machinery someone assembled on the outside, and later on we'll explore how much difference that machinery makes.
Capability
Mechanically it does one thing: it takes some text and works out what is likely to come next, then does that again. Part 2 will take that mechanism apart.
Functionally, that one operation buys a short and surprisingly enumerable list. This is the useful version:
| It can | Which looks like |
|---|---|
| Continue | carry on from text it was given |
| Recall | produce facts absorbed during training |
| Transform | translate, summarize, reformat, extract, rewrite |
| Generate | produce new material from a description |
| Follow instructions | do what was asked |
| Converse | hold a coherent exchange across several turns — when the earlier turns are handed back to it each time |
| Judge | classify, score, rank, choose between options |
| Work in steps | reason toward an answer through intermediate stages |
Equally short, and every item is something people assume it does:
| It cannot | Consequence |
|---|---|
| Remember | Nothing carries from one request to the next. |
| Look anything up | No search, no database, no files — unless something outside it fetches them and puts them in front of it. |
| Do anything | No code runs, no email sends, no record updates. |
| Check its own work | Nothing verifies the answer before it arrives. |
| Learn from you | Nothing you say changes the numbers. |
Every one of those absences is a product somebody had to build. That is the reason two companies using the identical model can ship things that feel nothing alike. We'll explore this in a future section.
Types of models
This terminology causes more confusion in meetings than it should: there is no single way to categorize models. People say “what kind of model is it” to mean any of many different things, and rarely say which.
The plain description comes first in each row, and the word people actually say is in brackets after it. There will be a glossary at the end.
| Axis | The question it answers | Common words you will hear |
|---|---|---|
| What it is for | What job was it built to do? | text generation (large language model, LLM) · turning text into numbers for search (embedding model) · image generation (diffusion model) · speech · classification |
| How finished it is | How far through the making did it get? | raw, straight out of training (base) · taught to follow instructions (instruct, chat) · one that works before it answers (reasoning) |
| How big | How many numbers, and what does it take to run? | small enough for a laptop (small language model, SLM) · one good chip · a rack · undisclosed |
| Who can have it | Are the numbers published? | open weights · available only as a service (closed, proprietary, API-only) |
| How it is arranged | Is all of it used on every request? | all of it, every time (dense) · split into parts, a few woken at a time (mixture of experts, MoE) |
These axes are independent, and that is the point. One model can be small and openly published and able to work before answering — all at once. So when somebody says “what kind of model is that,” the useful reflex is to ask which of those do you mean? It is a small habit that prevents a lot of talking past each other.
Each axis gets filled in where it belongs. Part 2 covers how the numbers got there in the first place; Part 3 covers what it costs to let a model work before it answers, and the vocabulary of size and packing — dense, quantized, distilled — in more detail than is useful here.
A fork worth knowing now
Almost everything written about AI describes the first kind of model and implies it covers the second. It does not, and assuming otherwise is the most common wrong idea non-technical people carry.
Produced one piece at a time, left to right, each new piece chosen in light of everything already written. Committed as it goes — what has been written does not change.
Nothing is predicted and nothing is next. The whole image exists from the first moment as random noise, and is refined everywhere at once, over and over, until something is there. No part is finished before any other.
They are different machines with different economics, different controls, and completely different ways of going wrong. Part 2 takes the first apart; Part 4 will give you the tools to explain why the second produces six fingers and cannot spell.
The habit worth forming now is small: when someone says “AI” or "model," feel free to ask which kind they mean.
Reference
These are the words that arrive in vendor decks, launch posts and meetings. Most of them name a position on one of the axes above. A few name nothing at all, which is worth knowing before you nod at one.
| Term | What it means |
|---|---|
| Large language model LLM | A model whose job is generating text. Usually what somebody means by “a model” with no other qualifier. It tells you the first axis and nothing else — not the size, not whether the weights are published, not how it is arranged inside. |
| Foundation model | A model trained broadly, then adapted to many different jobs, rather than built for one. The term was coined at Stanford in 2021.1 What makes it useful is that it names a business model as much as a technical one: one expensive general model, many cheap specializations on top. |
| Frontier model | Not a category but a position — whichever models are currently the most capable. Membership changes with every release, and last year's frontier model is this year's mid-tier one, at a tenth of the price. Treat it as a claim about the calendar, not about the product. Where it has to mean something legally, it gets pinned to a measurable threshold instead — the EU AI Act uses one. |
| General-purpose AI model GPAI | The EU AI Act's legal term for roughly what people mean by a foundation model. It matters if anything you buy touches the EU, because obligations attach to the term. |
| Multimodal model | A real term, used for two different things. A model that can read more than text — images, audio, a PDF as a picture — is multimodal on the way in, and that is now ordinary for major text models. A model that can produce more than text is multimodal on the way out, and that is much rarer. In a vendor deck it nearly always means the first. |
| Vision-language model VLM | The precise name for the common case: text out, but text or images in. |
| Diffusion model | An image generator. Named after how it works — starting from noise and removing it — which is a different mechanism from text generation, and the subject of the next section. |
| Embedding model | Turns text into numbers so that similar meanings sit close together. It generates nothing. You will meet it as the thing that makes search over your own documents possible. |
| Term | What it means |
|---|---|
| Base model | Straight out of the first stage of training. It continues text rather than answering questions, and is nearly always the wrong thing to buy. Offered mostly to people who intend to finish it themselves. |
| Instruction-tuned instruct, chat | Taught to treat your text as a request rather than as something to continue. This is what almost every product uses. |
| Reasoning model | One that produces work and iterates before it answers. Better on problems with a right answer, slower and more expensive on everything. Part 3 covers what that costs and Part 5 when it is worth it. |
| Fine-tuned | Trained further on somebody's own examples. Says nothing about how much, on what, or whether it helped. |
| Term | What it means |
|---|---|
| Open weights | The whole file with the numbers is downloadable. You can run it on your own machines, keep your data off somebody's API, and modify it. You also inherit the job of running it. |
| Open source | Usually a misuse of the term. Almost always it means open weights: the numbers are published, but the training data and the code that produced them are not. That is not what open source means anywhere else in software, and the gap matters if you need to know what a model was trained on. |
| Closed proprietary, API-only | Available as a service. You send text, you get text, you pay per token, and the models themselves never leave the provider. |
Most of these words answer one axis and stay silent on the other four. “It's a frontier multimodal LLM” tells you the vendor thinks it is current, that it reads images, and that it makes text. It tells you nothing about the size, the price, whether the weights are published, or whether it works for you.
Sources
What's coming next
Each answers one question, and they are meant in order — though nothing stops you jumping to the one you need today.