AI, ML, LLM — what the words really mean

A meeting. The engineering lead says “we should use ML for this, not LLMs.” The marketing lead says “isn’t that all AI?” The consultant says “yes, but deep learning is different.” Three people, three different referents, and all three statements are technically defensible.

The words AI, machine learning, deep learning, LLM, generative AI, NLP, and foundation model get used as if they were synonyms. They aren’t. They name specific things that nest inside each other like Russian dolls — each strictly contained within the larger category. Once the nesting is clear, those overlapping uses stop being noise and start carrying information about what the speaker actually means.

The nesting, from outside in

AI is the biggest box. Any computer system doing something that traditionally required human intelligence. This includes:

• A chess engine from the 1990s
• Your email’s spam filter
• The autocomplete in your search bar
• A self-driving car
• ChatGPT
• A handwriting recognition system

Some of this is rule-based (a human wrote down the rules). Some is learning-based (the system figured the rules out from examples). Both qualify as AI.

Machine learning (ML) is the subset of AI that learns from data instead of being explicitly programmed. The system is shown examples; it figures out a pattern; it applies the pattern to new cases.

ML examples:

• Show 100,000 emails labeled spam/not spam → spam filter
• Show historical sales data → forecasting model
• Show user behavior + which ads they clicked → ad ranker
• Show fraud and non-fraud transactions → fraud detection

ML has been quietly running businesses for decades. Most of what powers credit scoring, Netflix recommendations, Amazon product ranking, and fraud detection is ML. Not new. Not flashy. Industrial-strength infrastructure.

Deep learning is a flavor of ML that uses neural networks with many layers. The “deep” refers to the number of layers. Deep learning is especially good at messy, unstructured data — images, audio, text — where classical ML struggles.

Most modern AI breakthroughs are deep learning specifically. Image recognition that exceeds humans, real-time translation, voice transcription that actually works — deep learning underneath.

A neural network is the structure deep learning uses. Loosely inspired by neurons in the brain. A bunch of interconnected nodes that pass signals to each other. The “weights” of these connections are what get adjusted during training.

Generative AI is the subset of deep learning that generates new content rather than classifying or predicting from existing content. A spam filter classifies (is this spam?). A generative model creates (write me a paragraph; make me an image).

The shift from predictive AI to generative AI is what made AI suddenly visible inside businesses. The older AI ran in the background — spam filters, recommendation engines. Generative AI is what you talk to.

Large Language Model (LLM) is generative AI specifically for text. GPT, Claude, Gemini, Llama, Mistral — all LLMs. The first chapter walked through the mechanism — the next-word loop — that sits inside every one of them.

“Large” means: many parameters (the weights inside the model), trained on a lot of data. Smaller language models exist; they’re just less capable and cheaper to run.

A foundation model is a generic term for a large model trained on broad data that can be adapted to many tasks. Most LLMs are foundation models. Large image models can be too. The term means “this model is general-purpose enough to use as the base for many specific applications.”

Two more terms you’ll hear

Natural language processing (NLP) is the field of getting computers to work with human language. It’s a research area, not a kind of model. NLP predates LLMs by decades and used to involve hand-crafted rules and classical ML. LLMs are the current dominant approach in NLP.

Transformer is the specific architecture most modern LLMs (and many image models) are built on. It’s a particular way of structuring the neural network’s layers, and it underpins the current generation of models. The “GPT” in ChatGPT stands for “Generative Pre-trained Transformer.” Calling something “a transformer” is roughly equivalent to calling a car “a four-wheeler” — true, technical, not usually decision-relevant at the business level.

A picture of the nesting

Each box is a strict subset of the one outside it. Every LLM is generative AI is deep learning is ML is AI. Not every AI system is an LLM.

How these words get used in practice

The same terms carry different meanings in different mouths. The patterns are stable enough to be worth naming.

“We use AI,” in current marketing copy, almost always means generative AI or LLMs underneath. The word “AI” has been narrowed by usage — older AI categories (rule-based systems, classical ML) are still in production everywhere, but they rarely get marketed under the bare “AI” label anymore. In a more technical context, “AI” can still refer to a classical-ML system 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, a category to assign — and a classical ML model trained on the company’s specific data will be faster, cheaper, and more reliable than wiring up a chat model.

“Deep learning,” in casual conversation, often functions as a synonym for “modern AI” with a flag of technical seriousness. In a strict technical conversation, it specifically means neural networks with many layers.

“Generative AI,” from a consultant or a vendor, means the recent wave: LLMs, image generators, voice models. The capability that has been making headlines in recent years.

A heuristic for which tool fits

A rough field guide for matching a problem to a tool family:

Classical ML (often without LLMs) is the right tool for:

• Predicting a number (sales, demand, churn risk)
• Classifying into pre-defined categories with clear features (fraud / not fraud, lead quality A/B/C)
• Pattern-matching on structured data (transactions, sensor readings, click-streams)

These problems have years of data, clear labels, and stable inputs. Classical ML gives you faster, cheaper, more explainable answers than LLMs would.

LLMs (generative AI) is the right tool for:

• Anything where the output is text someone will read
• Anything where the input is messy unstructured language (emails, transcripts, documents)
• Anything where rules are impractical to write down (drafting, summarizing, extracting)

Both, combined, is increasingly common:

• An LLM reads a document and extracts structured information; a classical ML model uses that information to make a prediction.
• Classical ML routes a customer ticket to the right team; an LLM drafts the first response.

The dumb mistake is treating LLMs as the answer to every problem. They are not. The other dumb mistake is dismissing LLMs because “we already do ML.” Different tools, different jobs.

The terms, in one place

A reference card for the vocabulary as it is actually used.

• “AI” — in current usage, usually generative AI / LLMs. In stricter usage, anything from a chess engine to a spam filter.
• “ML” — can be classical (a model trained on a company’s own structured data for prediction or classification) or modern (LLMs). The two are very different things sharing a label.
• “Deep learning” — strictly, neural networks with many layers. Loosely, a flag of technical modernity; often used where “LLM” or “modern AI” would be more precise.
• “Neural network” — the underlying structure of deep learning. True of almost every modern model; not, on its own, very informative.
• “Generative AI” — the wave of models that produce content (text, image, audio, video) rather than only classifying it.
• “LLM” — generative AI specifically for text. ChatGPT, Claude, Gemini, Llama, Mistral.
• “Foundation model” — a large, general-purpose model trained on broad data and adapted to many downstream tasks. Most LLMs are foundation models.
• “NLP” — the research field of getting computers to work with human language. Predates LLMs by decades; LLMs are the current dominant approach.
• “Transformer” — the specific neural-network architecture that almost every modern LLM (and many image models) is built on. The “T” in GPT.

The point of holding these distinct isn’t vocabulary for its own sake. It’s that the differences between them are load-bearing. A claim about “AI” is a different claim depending on whether it means a classical ML model, an LLM, or a rule-based system — and the implications for what the system can do, how it fails, and what it costs are different in each case.