AI, ML, LLM: what the words really mean

AI, machine learning, deep learning, LLM, generative AI, NLP and foundation model get used as if they were synonyms. They name different things, and those things sit inside each other: every LLM is machine learning, and most machine learning is not an LLM. On a live decision the distinction decides what gets built, because it decides which kind of machine reads the text, which one produces the number, what each costs per request, and whether anyone can see why the answer came out that way.

The nesting, from outside in

AI

The widest term. Any computer system doing something that traditionally required human intelligence:

A chess engine
A spam filter
The autocomplete in a search bar
A self-driving car
ChatGPT

Some of this is rule-based: a person wrote the rules down. Some is learning-based: the system worked the rules out from examples. Both count as AI.

Machine learning (ML)

The part of AI that learns from data instead of being programmed rule by rule. The system is shown examples, it finds a pattern, it applies the pattern to new cases.

100,000 emails labelled spam / not spam → a spam filter
Years of sales history → a forecasting model
Past fraud and non-fraud transactions → fraud detection

ML has been running businesses quietly for decades. Credit scoring, Netflix recommendations, Amazon’s product ranking and fraud detection are mostly ML. It is old, unglamorous, and heavily relied on.

Deep learning

A kind of ML built on neural networks with many layers. “Deep” is the number of layers. A neural network is a mass of interconnected nodes passing signals to each other, loosely inspired by neurons in the brain. The strengths of those connections are the weights, and they are what training adjusts.

Structured data arrives in rows and columns: transactions, sensor readings, CRM fields. Unstructured data is everything else: an email, a call recording, a photograph, a scanned contract. Classical ML handles structured data well and struggles badly with the rest.

Deep learning is what made unstructured data workable: image recognition that beats people, real-time translation, voice transcription that actually works.

Generative AI

The part of deep learning that makes new content rather than sorting what already exists. A spam filter classifies: is this spam? A generative model produces: write me a paragraph, make me an image.

That shift is what made AI visible inside businesses. The older AI ran in the background. Generative AI is what you talk to, and it covers images, voice and video as well as text.

Large language models (LLMs)

Generative AI specifically for text. GPT, Claude, Gemini, Llama, Mistral are all LLMs. What an AI model actually is walks through the mechanism inside every one of them: the next-word loop.

“Large” means many parameters (the weights inside the model) trained on a lot of data. Smaller language models exist; they cost less to run and can do less.

A picture of the nesting

6 boxes. AI holds machine learning, which holds deep learning, which holds generative AI. Inside generative AI sit LLMs for text, and models for images, audio and video beside them.

The nesting is near enough to strict for any business decision. The exceptions are historical: simple text generators existed before neural networks did, so not everything that generates is deep learning. Read it inward: every LLM is machine learning, and most machine learning is not an LLM.

4 more words in the same conversation

Natural language processing (NLP) is the field of getting computers to work with human language. A research area, not a kind of model. It predates LLMs by a long way and used to run on hand-written rules and classical ML. LLMs are its current dominant approach.

Transformer is the architecture almost every modern LLM is built on, a particular way of arranging the neural network’s layers. The “GPT” in ChatGPT stands for “Generative Pre-trained Transformer.” Nearly every modern model is one, so the word says almost nothing about any particular model.

Foundation model is the generic term for a large model trained on broad data and adapted to many specific jobs. Most LLMs are foundation models; large image models can be too. People reach for it when they mean the general-purpose thing underneath, as opposed to the product built on top.

Reasoning model, also called a “thinking model” or a model in “extended thinking” mode, is not a new box in the nesting. It is an LLM trained to work through a problem before answering: it generates its own intermediate steps, checks them and revises, instead of producing the answer in one pass. That working is itself generated text, which is why the product pauses on a thinking indicator and why the answer takes longer and costs more. It pays off on problems that have to be got right in order, one step at a time, and wastes money on questions that never needed the working. The major labs all ship models that work this way, often as a setting on the same model.

How these words get used in practice

The same terms carry different meanings in different mouths, in patterns stable enough to name.

“We use AI,” in current marketing copy, almost always means generative AI or LLMs underneath. Usage has narrowed the word. Rule-based systems and classical ML are still in production everywhere and rarely get marketed as “AI” any more. In a more technical context, “AI” can still mean a classical ML model running quietly in the background.

“Use ML for this, not LLMs,” from an engineer, usually means: the problem has structured inputs and outputs, a number to predict or a category to assign, and a classical model trained on the company’s own data will be faster, cheaper and more reliable than wiring up a chat model.

“Deep learning,” in casual conversation, usually means “modern AI” with a flag of technical seriousness attached. In a strict technical conversation it means neural networks with many layers. “Generative AI,” from a consultant or a vendor, means the recent wave.

Which tool the lead problem needs

What classical ML does with leads. Train a model on the company’s own history: 3 years of enquiries and which of them turned into deals. It learns from the fields that were already tidy: company size, industry, source, deal size asked for. The output is a number per lead, plus a readout of which inputs pushed it up or down. It runs as a file on the company’s own server, answers in milliseconds, and costs nothing per lead once trained. What it cannot do is read the message: to this model, “we’re replacing our current provider in Q1 and need three sites covered” and “just browsing” are the same empty field.

What an LLM does with leads. It reads the message, where the intent actually lives. It pulls out what they want, by when, at what scale, whether they named a competitor, whether they mentioned a budget, and writes a 2-line summary a salesperson can act on. What it cannot do is know which of last year’s leads closed. Ask it to rank them and it gives a confident ordering based on enquiries in general, not on what has closed at this company. It is also a network call billed per request, answering in seconds rather than milliseconds, at a fraction of a cent a lead. At a few thousand leads a month that is a few dollars and nobody notices the seconds. At a few hundred thousand, both the bill and the wait behind it become real.

What the two do together. The LLM reads each message and turns it into fields: intent, urgency, scale, competitor named, budget signalled. The scoring model takes those fields alongside the tidy ones and produces the ranking, now informed by what the message said. The LLM drafts the first reply. 3 parts, each doing the work the others can’t.

All 3 positions in an argument like that can be defensible at once. A scoring model does beat a chat model at ranking. The part that reads the message is a different kind of machine from the part that scores it. And all of it is AI. The disagreement is rarely about the technology. It is about which word each person is using for which piece.

The general shape:

Classical ML fits problems shaped like the lead ranking: a number to predict or a category to assign, the company’s own history to learn from, and inputs that arrive as tidy fields.
LLMs fit problems shaped like reading the enquiry: language nobody has tidied going in, text a person will read coming out, and no way to write the rules down in advance.
Both together is increasingly what ships, in exactly the shape above: language model at the edges, scoring model in the middle. What AI is good at goes through that boundary in more detail.

Both mistakes are common. Reaching for an LLM where a scoring model would be faster, cheaper and easier to explain is one. Passing on LLMs because the company already runs ML is the other: the two share a label and very little else.

The terms, in one place

“AI”: in current usage, generative AI or LLMs. In stricter usage, anything from a chess engine to a spam filter.
“ML”: either classical (a model trained on a company’s own structured data to predict or classify) or modern (LLMs). 2 very different things sharing a label.
“Deep learning”: strictly, neural networks with many layers. Loosely, a flag of technical seriousness, usually where “LLM” would be more precise.
“Neural network”: the underlying structure of deep learning. True of almost every modern model, so it tells you very little on its own.
“Generative AI”: models that produce content (text, image, audio, video) rather than only sorting it.
“LLM”: generative AI for text.
“Reasoning model”: an LLM that generates its working before its answer. Slower and more expensive per answer, stronger on multi-step problems.
“Foundation model”: a large, general-purpose model adapted to many specific jobs.
“NLP”: the research field of computers working with human language. Older than LLMs; LLMs are its current dominant approach.
“Transformer”: the architecture under almost every modern LLM. The “T” in GPT.

None of this is vocabulary for its own sake. The lead problem prices the difference: milliseconds against seconds, nothing per lead against a fraction of a cent, a readout of which inputs moved the score against an answer you have to take on trust. A claim about “AI” is a different claim depending on which one is underneath.