What is machine learning, really? Let's start with a thermostat that refused to behave
Last winter, my heating system developed a personality. Not a smart one. The thermostat kept switching the boiler on at 6 a.m., even on days nobody was home before noon. I could have read the manual, found the scheduling menu, and fixed it in ten minutes. Instead I did the unreasonable thing: I spent an evening logging temperatures, occupancy, and outdoor weather in a spreadsheet, then tried to predict the boiler's behaviour from those columns. Twelve rows in, something clicked. I wasn't programming the thermostat anymore. I was trying to make a machine learn a pattern from data that I couldn't fully articulate myself.
That is machine learning in one sentence: instead of writing rules, you feed examples to an algorithm and let it infer the rules. Classical software is a recipe. Machine learning is a student.
Key takeaways
- Machine learning means a program improves at a task by finding patterns in data, not by following instructions you wrote by hand.
- It sits inside artificial intelligence, and deep learning sits inside it. ChatGPT belongs to the deep learning layer, not to "AI" as a vague umbrella.
- Three families cover most of what you'll meet: supervised, unsupervised, and reinforcement learning.
- You need less maths than the internet claims for a first project, and more clean data than anyone warns you about.
- A classical model often beats a neural network on small tabular datasets. I learned this the expensive way.
- Your first useful model can be built in an afternoon with Python, pandas, and scikit-learn.
What is machine learning in simple words for beginners?
Machine learning is a way of getting computers to solve problems by showing them examples rather than explaining the solution. You want to filter spam? You don't write a rule for every dodgy phrase. You show the algorithm fifty thousand emails labelled spam or not spam, and it works out the boundaries itself. Those boundaries are what we call a model.
Here's the part that trips people up. A traditional program does exactly what you told it. A machine learning model does what the data told it. Which means if the data is biased, sloppy, or unrepresentative, you get a confident machine that is confidently wrong. I've shipped one of those. It predicted customer churn beautifully on my training set and then flagged almost every loyal client as a flight risk in production, because my sample was pulled from a period when we'd just raised prices. Nobody had cancelled yet. The model had learned my pricing mistake, not customer behaviour.
The one-line definition worth memorising
Machine learning is the practice of building systems whose performance on a task improves as they are exposed to more relevant data, without a human rewriting the logic each time. That's it. No jargon required.
Why patterns beat hand-written rules
Hand-written rules break the moment reality shifts. Think about recognising a cat in a photo. You could try to encode "pointy ears, whiskers, four legs" — and then watch it fail on a sleeping cat, a cat in a box, a cat photographed from above. A model trained on a few thousand labelled images doesn't need your rules. It learns features you would never think to write down, and it generalises to cats it has never seen.
The trade-off is real, though. You lose the ability to read the logic line by line. When a rule-based system denies a loan, you can point at the rule. When a model denies it, you point at a file of weights and hope your explanation holds up. That is why regulated industries still mix both approaches, and why I always keep a simple, interpretable model as a baseline before reaching for anything fancy.
Where machine learning actually sits in the AI world
People use "AI" for everything these days, which makes the word almost useless in conversation. Here's the hierarchy I draw on a whiteboard whenever a colleague asks, and it has never failed me:
- Artificial intelligence — the broad ambition of making machines do things that look intelligent.
- Machine learning — one way to get there, by learning from data.
- Deep learning — machine learning using neural networks with many layers.
- Large language models — deep learning systems trained on enormous amounts of text.
Each layer is a subset of the one above. So the honest answer to most "is this AI or ML?" arguments is: both, and the question was badly framed.
Is ChatGPT AI or machine learning?
ChatGPT is machine learning, and therefore also AI. More precisely, it is a large language model, a deep learning system trained to predict the next word in a sequence, and that seemingly humble objective produces the conversational behaviour you see. The correct chain is: AI contains machine learning, machine learning contains deep learning, deep learning contains large language models. Saying "ChatGPT is AI" isn't wrong, it's just imprecise, like calling a violin "music".
Types of machine learning, without the textbook fog
Most introductory material lists the categories and moves on. That's a waste, because the category determines what you can actually build. Here's how I'd explain them to someone joining my team.
| Family | What you give it | What you get back | A task it suits |
|---|---|---|---|
| Supervised | Labelled examples | Predictions for new inputs | Predicting whether a customer will cancel |
| Unsupervised | Unlabelled data only | Groups, structure, oddities | Segmenting users by behaviour |
| Reinforcement | A goal and a reward signal | A policy for acting | Teaching a robot to walk |
| Generative (deep learning) | Vast amounts of text, images, audio | New content in the same style | Drafting replies, generating images |
Supervised learning covers the majority of business use cases you'll encounter, which is why almost every course starts there. Unsupervised learning is where I've had the most pleasant surprises — a clustering pass on support tickets once revealed a billing bug our team had been quietly absorbing for months, purely because the algorithm noticed an unusually tight group of similar complaints.
When is deep learning actually necessary?
If your data fits in a spreadsheet, try a classical model first. Gradient boosting, logistic regression, random forests — these are fast, interpretable, and on tabular data they frequently match or beat neural networks. I once replaced a small neural network with a gradient boosting model on a tabular dataset and improved accuracy by roughly nine percentage points while cutting training time from hours to under a minute. Deep learning earns its keep on images, audio, language, and anything where the raw input is high-dimensional and unstructured. Reaching for a neural network on a 2,000-row CSV is a very common beginner mistake, and I made it for months.
How to learn machine learning as a beginner?
Skip the temptation to start with a three-hundred-hour maths course. You'll quit in week two. The route that worked for me, and for the two juniors I've mentored since, looks like this:
- Get comfortable with Python basics — variables, loops, functions, files. Two or three weeks of steady practice is plenty.
- Learn pandas and how to actually clean a messy dataset. This is where most of your time goes in real projects, and almost no beginner course prepares you for it.
- Build one supervised model end to end with scikit-learn: load data, split it, train, evaluate, look at the errors.
- Add the maths as it becomes necessary: linear algebra for how models represent data, statistics for evaluation and uncertainty, calculus only when you need to understand how training adjusts weights.
- Do three small projects on data you care about, not on the iris dataset for the fourth time.
Notice the order. Tools before theory, theory before depth. A colleague of mine took the opposite path — eight months of mathematics before writing a single line of model code — and admitted afterwards that the concepts only clicked once he had something running to attach them to.
Do you need heavy maths to start?
You need enough statistics to know why accuracy lies on imbalanced data, and enough linear algebra to understand that a model is often just a weighted combination of inputs. That's a few weeks of focused effort, not a degree. What you genuinely cannot skip is data hygiene: missing values, duplicates, leakage between training and test sets. Leakage is the silent killer. My thermostat experiment was a fine learning exercise, but if I'd had an occupancy column that quietly encoded the answer, my "successful" predictions would have meant nothing.
Where you already meet machine learning, every day
Recommendation feeds, fraud detection on your card, voice transcription, the spam folder, route predictions in your maps app, the autocomplete on your phone keyboard. None of these were programmed with explicit rules for your specific case. They learned from millions of prior examples and generalised to you. If you want a copy-and-run starting point, search for a scikit-learn tutorial repository on GitHub rather than a theory PDF — a working notebook you can modify teaches more in an afternoon than a fifty-page handout.
And here's the thought I'd leave you with. The hard part of machine learning was never the algorithm. It's deciding what question is worth asking of your data, and being honest about whether the answer holds up outside the notebook.
