Where ML Already Shows Up in Your Day
You used ML before breakfast
Machine learning can feel like something that lives in research labs and big tech offices. It does, but it also lives in your pocket. On a normal day you probably use it dozens of times without noticing once.
In this lesson we walk through one ordinary day. For each moment, we will ask the questions from the last lesson: what is the machine predicting, what does it look at, and is it supervised or unsupervised? By the end, you will start seeing ML everywhere, and you will know how each one works at a basic level.
Morning
7:00 am. You pick up your phone and it unlocks when it sees your face. The model is answering one question: is this the owner, yes or no? That is classification. It learned what your face looks like from the photos it took when you set it up. The input is a camera image, so it uses deep learning.
7:05 am. You start typing "good morn" in WhatsApp and the keyboard suggests "morning". The model predicts the most likely next word from the words before it. It learned from huge amounts of text where the next word was already known, so it is supervised. Every word someone typed next acted as a label.
7:15 am. You open YouTube or Instagram and the feed is full of things you tend to like. For each possible post or video, a model predicts how likely you are to watch it, like it or skip it. The posts with the highest chance go to the top. It learned from what millions of people did in the past, which is recorded as labels: watched or skipped.
8:30 am. You check Google Maps before leaving and it says 34 minutes to the office. That estimate comes from a model that looks at distance, current traffic, the time of day and how long this trip took other people in the past. The answer is a number of minutes, so it is regression.
Afternoon
12:30 pm. You order lunch and the app says it will arrive in 32 minutes. Another regression model: it looks at how busy the restaurant is, distance, rain, the time of day, and how long similar orders took before. The "restaurants you might like" row on the home screen comes from a recommendation model, the same idea as your YouTube feed.
1:00 pm. You pay the delivery person by UPI. In the second between tapping "pay" and seeing the green tick, your bank's systems check whether the payment looks like fraud. Some of that is classification, trained on past payments marked as fraud or genuine. Some of it is anomaly detection from lesson 1.3: does this payment look very different from how you usually pay?
3:00 pm. Your email has quietly moved three "you won a lottery" messages into the spam folder. Classic classification: spam or not spam, learned from messages people marked in the past. Every time you press "report spam", you are adding a label.
Evening
6:00 pm. You look at earphones on a shopping app, and underneath it says "customers also bought: a phone cover, a charger". That comes from spotting which products tend to be bought together across millions of orders. Nobody labelled those pairs. The pattern came out of the data, which makes it unsupervised.
8:00 pm. The cricket broadcast shows a win predictor: "India 64%, Australia 36%". A model looks at runs needed, balls left and wickets in hand, and compares with how thousands of past matches in a similar situation turned out.
10:00 pm. Too tired to type, you send a voice note and your phone turns it into text. Speech recognition is deep learning: audio goes in, words come out. It learned from a huge collection of recordings where people had already written down what was said.
That is more than ten uses of machine learning in one ordinary day, and we skipped plenty. Autocorrect, photo gallery search, "is this a scam call" warnings, music recommendations.
Beyond your phone
ML is also doing serious work in places you might not see directly:
Key Points
- Healthcare: models read eye scans to catch diabetic eye disease early, which helps where there are not enough eye specialists to check everyone.
- Farming: a farmer photographs a sick leaf and an app suggests which disease it probably is. That is classification from an image.
- Banking: when you apply for a loan, a model estimates how likely you are to repay, based on people with similar histories.
- Weather: forecasts use models trained on decades of past weather data.
- Factories: sensors on machines feed anomaly detection models that warn when a machine is about to break down.
Every example has the same shape
Face unlock, traffic estimates, spam filters and cricket predictors look like completely different products. Underneath, they all work the same way:
Some inputs go in. A model, which learned from past examples, turns them into one prediction. The app then does something with that prediction: shows it to you, sorts your feed, or blocks a payment. Once you see that shape, you can look at any "smart" feature and work out roughly how it is built.
When ML is the wrong tool
After a day like that, it is tempting to think ML should be used for everything. It should not.
Think about calculating GST on a bill, working out your electricity bill from units used, or sorting names in alphabetical order. The rules for these are exact and known. A few lines of normal code get them right every single time. A model would only learn to get them "mostly right", which is worse.
ML is worth it when two things are true. The rules are too messy for a person to write down, like spam, faces or traffic. And being right most of the time is still very useful, even if the model is sometimes wrong.
Tip:A simple test: could you write the exact rules on a sheet of paper? If yes, write normal code. If the rules keep changing, have too many exceptions, or you cannot even describe them, that is where ML helps.
ML makes mistakes, and sometimes unfair ones
Every example in this lesson is sometimes wrong. The lunch arrives in 45 minutes instead of 32. A real email lands in spam. That is normal. A model learns patterns, and patterns have exceptions.
There is a more serious kind of mistake too. A model learns from past data, so it also learns whatever unfairness was in that past data. If a bank's old loan decisions were unfair to some group of people, a model trained on those decisions can quietly repeat that unfairness. This is a real problem, and it is part of why people who build models have to check them carefully, not just trust them.
Where people get confused
Key Points
- "Anything smart in an app is AI magic." Most of it is a model making one specific prediction, learned from past data. Once you know that, it stops being magic.
- "These apps use one big AI." Usually each feature has its own model. The ETA model and the "restaurants you might like" model in a food app are different models doing different jobs.
- "If a model is wrong sometimes, it is broken." Every model is wrong sometimes. The question is whether it is right often enough to be useful.
Quick check
Google Maps says your trip will take 34 minutes. Regression or classification?
Regression. The prediction is a number (minutes), not a category.
A clothing brand wants to split its customers into groups for different ads, and it has no labels for customer types. Which kind of learning?
Unsupervised, specifically clustering. There is no answer column, so the model groups similar customers and people decide what each group means.
Should you train an ML model to calculate the GST on a bill?
No. The GST rules are exact and known, so normal code gets it right every time. ML is for problems where the rules are too messy to write down.
Recap
Key Points
- You use ML dozens of times a day: face unlock, keyboard suggestions, feeds, maps, delivery times, fraud checks, spam filters, voice typing.
- Every one has the same shape: inputs go in, a model trained on past examples makes one prediction, and the app acts on it.
- Use ML when the rules are too messy to write and being mostly right is still useful. Use normal code when the rules are exact.
- Models make mistakes, and they can repeat unfairness from the data they learned from.
Next up: the ML lifecycle. Every one of those features started as a question someone asked. How do you get from a question to a model running inside an app?
Theia
Can you explain this concept?
Practise Where ML Already Shows Up in Your Day with realistic interview questions. Choose spoken or typed answers, get feedback on your reasoning, then return to the material that needs work.
Sign in to practise this conceptThe session stays focused on this lesson. Your selected interview level sets the bar.