1.1 YouTube/Netflix Recommendations
Or "Why YouTube Knows You'll Love That Cat Video You Didn't Even Search For"
Have you ever found yourself lost in YouTube Shorts or scrolling through your Netflix feed, and time just flew by? The credit (or blame!) for this lies with one of the most powerful and widespread AI systems in the world—the recommendation system.
Let's break down how it works at the most basic level, without complex terms.
Analogy: The Super-Observant Movie Buff Friend
Imagine you have a friend who:
- Watches you closely 24/7.
- Remembers everything: what movies you watched, how many minutes you watched each one, at what time of day, whether you liked it, if you skipped boring parts, if you quit halfway through.
- Compares you to millions of other people who watched the same things.
Example: You watched the movie "The Avengers" and gave it 5 stars.
A normal friend would say: "Cool!"
The AI friend (the recommendation system) does the following:
Step 2: Searches its huge database for "other users who also liked 'The Avengers'".
Step 3: Analyzes: "And what else did these similar users watch? What other movies did they also rate 5 stars?"
Step 4: Finds a pattern: "Aha! 80% of the people who loved 'The Avengers' also highly rated 'Thor: Ragnarok' and 'Guardians of the Galaxy'."
Step 5: Draws a conclusion: "With a high probability (80%), you will like 'Thor' and 'Guardians'." And it serves them up in your recommendations.
This method is called collaborative filtering—that is, filtering (selection) based on the collaboration (similarity) of users. The AI doesn't understand the movie's content; it only understands numbers, patterns, and connections between people and content.
Content-Based Filtering: The Second Approach
But Netflix and YouTube have gotten smarter. Now they also look at the content itself. This is the second powerful method—content-based filtering. Let's say you just finished the series "Stranger Things."
The AI "breaks down" this series into tags (labels): science fiction, 1980s, teenagers, horror, nostalgia, Stephen King (in spirit).
Then it searches its library for other series and movies with the most similar sets of tags.
And it recommends "His Dark Materials" or "The Dark Side of the Moon" to you, even if they were watched by a completely different group of people.
What Makes Recommendations So "Sticky"?
The secret is that modern systems combine both approaches and add a ton of other data:
- Your behavior: You watch comedies more often on Friday nights, and documentaries on Sunday mornings. The algorithm will notice this and suggest different content at different times.
- Trends: If half the world suddenly starts watching the Korean series "Squid Game," the algorithm will subtly hint: "Hey, don't you want to see what everyone's talking about?"
- Platform goals: YouTube has a goal—to keep you on the platform as long as possible. So it will recommend not necessarily the "objectively best" content, but the content most likely to make you click "Next." Sometimes this leads to recommendations becoming increasingly radical or sensational (this is called a "filter bubble" or "echo chamber").
In simple terms: What does the algorithm do when you open YouTube/Netflix?
- "Creates a portrait": It instantly creates your digital portrait based on your entire history.
- "Looks for neighbors": It finds thousands of other users with maximally similar portraits.
- "Peeps at their players": It looks at what these "neighbors" are watching and liking right now.
- "Sorts by probability": For every video in its library, it calculates the percentage probability that you will watch it to the end and be satisfied.
- "Serves the menu": It shows you the top 20 videos with the highest percentages in a nice interface.
The bottom line: Recommendation systems are not magic, but high-level mathematics and data analysis. This is an AI that has turned our habits, clicks, and likes into a giant map of interconnections. It doesn't read your mind, but it has studied you and millions of other people so well that it anticipates your desires—and often does it better than we can articulate them ourselves.
This is the power of "Narrow AI"—it's genius at one specific task (recommendations), but it doesn't possess intelligence or consciousness.