β€œWe’ll just add ChatGPT”

A founder says it in a product meeting. A developer hears it in a brief. A client writes it in a requirements doc. β€œWe’ll just use AI” β€” said with confidence, with a plan already forming in their head.

The trouble is, those three people may mean three different things. One means β€œwe’ll add a chatbot.” Another means β€œwe’ll run a recommendation algorithm.” A third means β€œwe’ll wire up a GPT-4 API call.” All three are technically AI. None of them are the same thing.

Before you build, it helps to have a map.

The outer ring: AI as an umbrella

Artificial intelligence is the widest term. It covers every attempt to get a machine to do something that, if a human did it, we’d call intelligent β€” recognizing a face, translating a sentence, recommending a film, detecting fraud, playing chess.

The word is old. It was coined in the 1950s, and it has gone through several cycles of hype and disappointment since. This matters: when someone invokes AI skeptically (β€œwe’ve heard this before”), they’re reacting to those older cycles. They’re not wrong to remember them. The question is what’s different now β€” and that story belongs to a separate post.

Inside the ring: machine learning

Most of what people call AI today is really machine learning β€” a subset of AI in which the machine learns patterns from data rather than following rules a human wrote. You don’t tell it β€œif the email contains the word β€˜prize’, mark it spam.” You show it ten thousand emails labeled spam or not-spam, and it figures out the rules itself.

This is a meaningful shift: the work moves from writing rules to curating examples. It made AI practical across a much wider range of problems.

Deeper still: deep learning and neural networks

Inside machine learning sits deep learning β€” a family of techniques that use layered mathematical structures loosely inspired by the brain (called neural networks). Deep learning is what made image recognition, speech recognition, and machine translation actually work well. Before deep learning, those tasks were partly handled by hand-engineered rules. After it, the machine learned the patterns directly from raw data.

You don’t need to know how the layers work. What matters is that deep learning is a technique inside machine learning β€” not a separate thing.

Today’s wave: large language models

A large language model, or LLM, is a deep-learning model trained specifically on text. It reads enormous amounts of written language β€” books, web pages, code, conversations β€” and learns to predict what comes next in any piece of text. That single job, done at sufficient scale, turns out to produce something surprisingly capable: a model that can write, summarize, translate, reason through problems, and generate code.

The β€œlarge” matters. The same architecture trained on far less data produces far weaker results. Scale unlocked the capability. (The architecture that made scale tractable is called the transformer β€” a technique introduced in a 2017 research paper that reorganized how models process sequences of text. Worth knowing the name; you don’t need to know the mechanism.)

Generative AI is the marketing term for this wave β€” models that generate new content (text, images, audio, video) rather than just classifying or predicting. It’s a real category, but it’s also partly a rebranding. Generative models are large deep-learning models. The distinction is in what they output, not in a separate technology.

The distinction that matters most: model β‰  product β‰  system

This is where the map pays off.

ChatGPT is not a model. It is a product β€” a chatbot built by OpenAI, wrapped around their GPT models, with a user interface, a memory layer, a moderation layer, and a pricing structure. When you use ChatGPT, you are using a product that sits on top of an LLM.

GPT-4 (and its successors) is the model underneath. You can access that model directly through an API β€” without the chatbot interface, without the memory defaults, with different behavior. Same underlying capability, different product shell.

An agent is something else again: a model given tools, memory, and planning so it can take actions in the world β€” searching the web, reading files, calling APIs, running code. An agent uses an LLM as its reasoning engine, but it’s a system wrapped around the model. This distinction β€” model, product, system β€” is the one worth internalizing. The post From LLMs to agents walks that journey in full.

Generative versus predictive

One more split that trips people up: not all ML is generative.

Predictive (or analytical) ML produces a score, a forecast, or a classification. Will this customer churn? What is the probability this transaction is fraud? Which of these three categories does this image fall into? The output is a number or a label, not new content.

Generative ML produces new content: a sentence, an image, an audio clip, a code snippet. The model invents something, within the patterns it learned, rather than assigning a label.

Both live under the machine learning umbrella. A fraud-detection model at a bank is ML. A tool that drafts a summary of your meeting is also ML. They feel completely different because they are doing completely different things β€” but they share the same lineage.

Where it gets fuzzy β€” and that’s honest

These categories are real, but the lines blur.

An LLM is, technically, a kind of machine learning model β€” so β€œmachine learning” and β€œLLM” overlap; LLMs are a subset. β€œGenerative AI” is partly a marketing umbrella that groups LLMs with image generators and audio models under one banner. The term isn’t wrong, but it’s not a precise technical boundary either.

β€œAI” itself is used so loosely that a spam filter, a recommendation engine, and a model that writes poetry all fit inside it. This is fine as long as you know what you mean when you say it.

The map isn’t perfectly crisp. That’s the honest version. The point isn’t to memorize the taxonomy β€” it’s to have enough separation to ask the right question: which kind of AI, doing what, in what context?

What to take away

  • AI is the wide umbrella β€” machines doing intelligent-seeming things.
  • Machine learning is how most of it works today β€” learned patterns, not hand-written rules.
  • Deep learning / neural networks are the technique inside most modern ML.
  • LLMs are deep-learning models trained on text at scale.
  • Generative AI is today’s wave: models that produce new content. The term is real but partly marketing.
  • ChatGPT is a product. GPT-4 is a model. An agent is a system wrapping a model with tools and planning.
  • Predictive ML gives you scores and classifications. Generative ML gives you new content. Both are ML.

If you want to understand why LLMs are the current wave β€” and why this moment feels different from past AI cycles β€” the next stop is Why this time is different.