AI, ML, Deep Learning, GenAI

Four buzzwords that are often used as if they meant the same thing. They do not: they fit inside one another like nesting dolls. Here is the family tree, layer by layer.

At a wedding kenduri, an uncle tells you his company is “doing AI”. Your cousin says she is learning “machine learning”. A news headline warns about “generative AI”, and a friend swears by “the LLM” on her phone. Are they all talking about the same thing? Partly. These terms are like the address on a letter: country, state, town, street. Each one is more specific than the last, and each sits inside the one before it.

The nesting dolls

The simplest way to hold the whole picture in your head is a set of circles, each inside the next. Artificial intelligence is the biggest. Inside it sits machine learning. Inside that, deep learning. Inside that, generative AI. And inside generative AI you find the large language models behind today’s chat assistants. Open the layers one at a time.

LLM

The key idea: every LLM is generative AI, nearly all generative AI today is deep learning, all deep learning is machine learning, and all machine learning is AI. But it does not work the other way round. Plenty of AI is not machine learning, and plenty of machine learning is not deep learning.

Layer by layer

Artificial intelligence: the whole field

Artificial intelligence is the broad project, named in 1955, of making machines do things that seem to need intelligence (see What Is AI, Really?). It is an umbrella for many different techniques: logic, search, planning, probability and learning. The computer opponent in an old video game that chases you round a maze is AI in this broad sense, even though it learns nothing.

Machine learning: learning from examples

Machine learning is the part of AI where, instead of writing the rules yourself, you give the computer lots of examples and let it work out the pattern. The term was popularised by Arthur Samuel at IBM, whose 1959 paper described a checkers (draughts) program that improved by playing against itself until it beat him. Everyday examples: credit-card fraud alerts, product recommendations on Shopee or Lazada, and estimated arrival times in ride-hailing apps. Much of this is done with fairly simple methods — not every learning system is a giant neural network. How the learning actually happens is the subject of How Machines Learn.

Deep learning: many-layered networks

Deep learning uses neural networks — systems loosely inspired by brain cells, made of many simple units connected in layers. “Deep” just means many layers. Its strength is that it can learn directly from raw, messy data. Older methods often needed a person to say which features mattered (“look at the word ‘free’ and the number of exclamation marks”); a deep network finds useful features by itself. Deep learning took off around 2012, when a network called AlexNet won an image-recognition contest by a wide margin, helped by fast graphics chips. It powers speech recognition, face unlock, photo search and medical-image analysis.

Generative AI: making new things

Generative AI is deep learning that produces new content rather than just labels. A classic model looks at a photo and says “cat”. A generative model takes the words “a cat wearing a songkok” and draws one. Image generators such as Midjourney, DALL·E and Stable Diffusion, voice-cloning tools, music generators and video tools all belong here. Many image generators use a technique called diffusion; text generators use a different design called the Transformer.

Large language models: the chat engine

A large language model (LLM) is a generative model for text, trained on a vast collection of writing to predict what word comes next. Scale that up enormously and add further training to follow instructions, and you get the assistants many people now use daily. The next pages explain how in detail: What Is a Large Language Model?

A small honesty note: the borders are not perfectly neat. LLMs are also used for non-generative jobs such as sorting customer emails, and many newer models handle images and audio as well as text. The nesting dolls are a map, not a law of nature.

Two schools: rules versus learning

For most of AI’s history there were two broad camps. Symbolic AI, sometimes nicknamed GOFAI (“good old-fashioned AI”), tried to write intelligence down by hand: facts, logic and if–then rules. The learning camp tried to let machines discover knowledge from data. Compare them side by side.

How it works: experts write rules. “If the patient has a fever and a rash and has been bitten by mosquitoes, consider dengue.”

Strengths: you can read exactly why it decided something; it needs no training data; it is precise where the rules are clear.

Weaknesses: the world is too messy to write down. Try writing rules to recognise your mother’s face in a crowd. The 1980s boom in expert systems foundered partly on this: the rule books became huge, brittle and expensive to maintain (see Expert Systems and the Second Winter).

Neighbours: robotics and data science

Two other fields are often mixed up with AI. They overlap with it, but neither sits neatly inside the dolls.

ROBRoboticsMachines that act in the physical world

Robotics is about building machines that sense and move: factory arms, warehouse robots, robot vacuum cleaners, surgical robots. Many robots use little AI — a car-painting arm repeats the same motion perfectly. Others use a lot: a robot vacuum that maps your living room, or a self-driving car that recognises motorcycles weaving through traffic. So robotics overlaps AI where robots must perceive and decide for themselves. And a chat assistant is AI with no body at all.

DSData scienceGetting insight from data

Data science is the practice of collecting, cleaning, analysing and presenting data to answer questions: which branches of a pharmacy chain are growing, which months are busiest for a clinic. Data scientists often use machine learning as one tool among many — alongside statistics, spreadsheets and charts. So data science overlaps strongly with machine learning, but a lot of it involves no AI at all.

STATStatisticsThe older cousin

Much of machine learning grew from statistics: both are about finding patterns in data and measuring uncertainty. Statisticians tend to care more about explaining why; machine-learning engineers tend to care more about predicting what. The boundary is friendly and blurry.

Key point

When someone says “AI”, ask: which layer do they mean? A rule-based chatbot, a fraud model, a face-recognition network and a large language model are all “AI”, but they behave, fail and need checking in very different ways.

Check your understanding

Quick check

Which statement is true?

Deep learning is one kind of machine learning. But there is AI without learning (hand-written rules) and machine learning without deep networks.

A 1980s system that diagnoses engine faults using 500 rules written by mechanics is…

It follows hand-written rules rather than learning from data, so it sits in the outer ring only.

What makes an AI system “generative”?

Generative models produce new text, images, audio, video or code rather than only labelling or predicting.

Where does a large language model sit in the nesting dolls?

LLMs are deep neural networks that generate text, so they sit inside every other layer.

// Pause and reflect

  1. Think of an app you use for work or banking. Which layer of the dolls do you think its “smart” features come from?
  2. When a company says its product “uses AI”, what one question would you now ask to find out what that really means?

Your notes are saved privately in this browser.

Try this

Draw four circles inside each other on a scrap of paper and label them AI, ML, DL and GenAI. Now place five things you use — a spam filter, face unlock, a chat assistant, a calculator, a video recommendation feed — in the smallest circle each belongs to (or outside them all). Explaining your drawing to someone else is the best test that it has stuck.

Key terms on this page

Sources & further reading