Where AI already runs your day: study notes

 

0.06 Where AI already runs your day: study notes

Learn AI by Story

AI is already working behind most of the apps you use, and almost all of it falls into four families: vision, language, recommendations and robotics.

1. The idea in one sentence

Everyday AI is not one big brain; it is many narrow tools, each trained for one job, that sense something (an image, a sentence, a click), find a pattern, and act on it.

2. The story from the video

Meera wakes up convinced she has not used any AI today. Then she counts: her phone unlocked with her face, her map picked the fastest route, her keyboard finished her sentences, and tonight a video app will choose her next show. By lunch AI had helped her dozens of times.

3. Key definitions

Term	Precise definition	Plain English
Narrow AI	A system built and trained to do one specific task or a small set of tasks	One job, done well
Computer vision	AI that extracts information from images and video	Software that sees
Natural language processing (NLP)	AI that reads, interprets or generates human language	Software that reads and writes
Speech recognition	Converting spoken audio into text	Turning voice into words
Recommender system	A model that predicts which items a user will like, from the past behaviour of that user and similar users	The thing that picks your next video
Collaborative filtering	A recommendation method that uses the overlap between users' histories	People like you liked this
Robotics	Building machines that sense, decide and act in the physical world	AI with a body
Sensor	A device that measures the world (camera, microphone, GPS, lidar)	The robot's senses
Actuator	A part that produces movement (motor, gripper)	The robot's muscles
Model	The learned pattern-matching program produced by training	What the AI actually is
Inference	Running a trained model on new input	Using what it learned
Personalization	Adapting what you see to your own behaviour	A feed that feels made for you
Feedback loop	When a system's output changes the data it later learns from	You click it, so you see more of it

4. Step-by-step: how everyday AI works

Almost every AI feature follows one loop: sense, think, act.

1. Sense. Something captures raw input: a camera frame, a microphone clip, the text you typed, or a log of your clicks.
2. Turn it into numbers. Computers only handle numbers, so pixels, sound waves and words are converted to numerical form (for example, an image becomes a grid of brightness values).
3. Think. A trained model compares the input with patterns it learned from many past examples, and produces a prediction: this is Meera's face, this word likely comes next, this video will probably be liked.
4. Act. The prediction triggers something visible: the phone unlocks, a suggestion appears, a robot arm moves.
5. Learn from feedback. Many systems record whether the prediction worked (did you tap the suggestion?) and use that to improve later.

Why this is practical: training happened once, in advance, on a huge amount of data. When you use the feature, only the fast step (inference) runs, so the answer feels instant.

5. The four families in depth

Vision

• Face unlock compares a live scan of your face with a stored numerical description of it. Good systems also check that it is a real face and not a photo.
• Photo apps group pictures by person, place or object by recognizing what is in each image.
• Medical imaging models can flag areas on X-rays or scans that deserve a closer look. They assist doctors; they do not replace the doctor's judgment.
• Cars use cameras (and often other sensors) to detect lanes, signs, people and other vehicles.
• Typical tasks: classification (what is in the picture), detection (where are the objects), segmentation (which pixels belong to which object), and text reading (OCR).

Language

• Autocomplete and spell-check predict the next word or a fix from what you have typed.
• Translation maps a sentence in one language to another.
• Voice assistants chain several models: speech recognition to turn audio into text, language understanding to work out what you want, an action, and speech synthesis to answer aloud.
• Chatbots built on large language models generate fluent text by predicting what comes next, one piece at a time.
• Typical tasks: classification (spam or not), extraction (pull dates from an email), summarization, translation and question answering.

Recommendations

• The system logs what you watch, skip, like or buy.
• It compares your history with many other people's histories.
• It predicts a score for every item you have not seen and shows the highest ones.
• Used by video, music and shopping apps, news feeds and ad systems.
• Two classic methods: collaborative filtering (people similar to you liked X) and content-based filtering (X is similar to things you liked). Real systems blend both with many other signals.

Robotics

• A robot needs sensors (cameras, distance sensors), a decision model, and actuators (motors).
• Warehouse robots move shelves or sort parcels; robot vacuums map a room and plan a path; factory arms weld, paint and assemble.
• Often only part of a robot is learned AI. The rest is ordinary engineering: control software, safety limits and mechanics.
• Robotics is harder than software-only AI because the physical world is messy, slow and unforgiving of mistakes.

6. What the video skipped

• Other families. Beyond these four there is forecasting (weather, demand, fraud scores), search ranking, generative AI that creates text, images or audio, and decision-making AI used in games and logistics.
• Most products mix families. A voice assistant uses language models, a map app uses search plus traffic forecasting, and a self-driving car uses vision, prediction and planning together.
• Learned versus rule-based. Some features people call AI are just fixed rules. Learned systems improve with data; rule systems do not (see roadmap 0.04, rules versus learning).
• Trade-offs. Personalization raises convenience but can create filter bubbles. Face recognition raises privacy and fairness questions. Systems can be wrong in ways that are hard to predict, so important decisions need human oversight.
• Where the computing runs. Some AI runs on your phone (face unlock, keyboard), some in data centres (large chatbots, recommendations). On-device AI is faster and more private; cloud AI can be much larger.
• History. Many of these tools matured after 2012, when deep learning, big datasets and GPUs came together (see 0.03 and 0.09).
• Connects to: 0.01 (narrow vs general AI), 0.02 (AI vs ML vs DL), 0.05 (kinds of learning), 0.08 (myths).

7. Worked example: how a recommendation is made (illustrative data)

Suppose a small app has three users and four videos. A tick means the user liked the video.

User	Cooking	Cricket	Coding	Music
Meera	tick	tick	tick	not seen
Arun	tick	tick	tick	tick
Diya	tick	no	no	no

1. Meera's liked list (cooking, cricket, coding) matches Arun's on three items, but Diya's on only one.
2. So Arun is the most similar user to Meera.
3. Arun liked Music, and Meera has not seen it yet.
4. The system gives Music a high score for Meera and shows it near the top of her feed.

Real systems do this with millions of users and thousands of signals, but the logic is the same: find similar people, borrow their tastes, predict yours.

8. Comparison of the four families

Family	Input	Output	Everyday example	Main risk
Vision	Images, video	Labels, boxes, decisions	Face unlock	Misidentification, privacy
Language	Text, speech	Text, intent, translation	Autocomplete, chatbot	Confident wrong answers
Recommendations	Click and watch history	A ranked list	Next video	Filter bubbles
Robotics	Sensor readings	Physical movement	Robot vacuum	Safety in the real world

9. Common mistakes and myths

• "AI is one thing." It is a collection of separate narrow tools.
• "If it feels smart, it understands." A recommender has no idea what a video is about in a human sense; it finds statistical patterns in behaviour.
• "AI only means chatbots." Most AI in daily life is invisible: spam filters, fraud checks, route planning.
• "Robots are always humanoid." Most robots are single-purpose machines such as arms and vacuums.
• "Personalized means private." Personalization needs data about you; check the privacy settings of the apps you use.

10. Real-life uses

• Banks flag unusual card payments within seconds.
• Email services filter spam and suggest replies.
• Streaming apps rank thousands of titles for each person.
• Farms use drone cameras to spot crop disease.
• Hospitals use imaging models as a second pair of eyes.
• Delivery companies plan routes with traffic forecasts.

11. Self-check questions

1. Name the four AI families from the video. Vision, language, recommendations, robotics.
2. What are the three steps of the basic AI loop? Sense, think, act.
3. Why does a trained model answer instantly? Training was done in advance; only inference runs when you use it.
4. What is collaborative filtering? Recommending items that similar users liked.
5. Which family does autocomplete belong to? Language.
6. In the worked example, why was Arun chosen for Meera? His likes overlapped with hers on three videos, more than Diya's.
7. What is the difference between a sensor and an actuator? A sensor measures the world; an actuator moves something.
8. Why might a recommender create a filter bubble? It shows more of what you already clicked, so you see less variety.
9. Give one example where AI assists a human rather than replacing them. Medical scan flagging for a doctor to review.
10. Is face unlock narrow or general AI? Narrow; it only does face matching.

12. Key takeaways and go deeper

• Everyday AI = narrow tools running the sense, think, act loop.
• The four big families are vision, language, recommendations and robotics, and most products mix them.
• AI helps most when a human can check important results.
• Go deeper: read about how image classification works, how recommender systems use matrix factorization, and what a sensor-actuator loop looks like in a simple robot vacuum.
• Next roadmap ID: 0.07, AGI and superintelligence: what people argue about.

Learn AI by Story: one concept, one story, every day.