Your Best Friend, the Algorithm
You have that one friend who always knows where to eat. You say "I'm hungry," and they say "Let's get ramen" — and it's perfect. Not because they memorized your favorite restaurants, but because they know your patterns. You like warm food, you like noodles, you tried Thai last week and loved it, and it's cold outside.
They didn't run a calculation. They just... know you.
Netflix does the exact same thing. Except instead of one friend who knows you, it's an algorithm that knows 260 million people — and it finds the ones who are just like you.
The "People Like You" Trick
Here's the core idea behind recommendation systems, stripped down to its simplest form:
Step 1: Look at everything you've watched and rated.
Step 2: Find other people who watched and rated the same things similarly.
Step 3: Look at what those people watched that you haven't seen yet.
Step 4: Recommend those shows to you.
That's it. If you and someone in Brazil both gave 5 stars to the same 20 shows, and they also loved a documentary you've never heard of — Netflix bets you'll love it too.
This approach is called collaborative filtering, and it's the backbone of almost every recommendation system you use — Netflix, Spotify, Amazon, YouTube, TikTok.
But How Does It Actually Know?
Let's make it concrete. Imagine a giant spreadsheet. Every row is a person. Every column is a movie. Each cell is a rating from 1 to 5 (or empty if they haven't seen it).
| | The Matrix | Inception | Titanic | Avengers | |---|---|---|---|---| | You | 5 | 5 | 2 | ? | | Person A | 5 | 5 | 1 | 5 | | Person B | 2 | 1 | 5 | 2 | | Person C | 4 | 5 | 2 | 4 |
Look at the pattern. You rated The Matrix and Inception highly but didn't love Titanic. Person A and Person C have almost the same pattern. Person B is the opposite.
So what would you rate Avengers? The algorithm looks at Person A (rated it 5) and Person C (rated it 4) — the people most similar to you. It predicts you'll probably give it a 4 or 5. And it recommends it.
That's recommendation AI. It's not reading your mind. It's reading patterns.
The "You Might Also Like" Problem
There's a second approach that works differently. Instead of finding similar people, it finds similar items.
If you watched and loved Inception, the system looks at Inception's properties: it's a sci-fi thriller, directed by Christopher Nolan, starring Leonardo DiCaprio, with a complex plot and stunning visuals.
Then it finds other movies with similar properties: Interstellar, The Prestige, Shutter Island, Tenet.
This is called content-based filtering — and it's why Netflix shows you a row called "Because you watched Inception."
The Real Magic: Combining Both
Modern recommendation systems don't use just one approach. They blend both:
- Collaborative filtering: People similar to you liked this
- Content-based filtering: This item is similar to things you liked
- Your behavior: What time you watch, how long you watch before quitting, what you search for, what you hover over but don't click
All of these signals feed into a neural network (remember those from Article #1?) that learns your preferences at a level of detail that's almost scary.
That show Netflix recommended at 11 PM on a rainy Tuesday? It probably noticed that you watch lighter comedies late at night and dramas on weekends. It knew it was raining because that affects viewing patterns. And it knew you just finished a heavy drama, so you might want something lighter.
The Cold Start Problem
Here's an interesting challenge: what happens when you're brand new to Netflix? You haven't watched anything. There are no patterns to find.
This is called the cold start problem, and every recommendation system faces it. Solutions include:
- Asking you to pick genres you like when you sign up
- Showing you the most universally popular content first
- Using your age, location, and device as rough signals
- Learning very fast from your first few clicks
Within just 5-10 interactions, the algorithm already has enough data to start personalizing. That's how fast pattern matching works.
Try It Yourself
Think about the last 5 songs Spotify recommended that you actually liked. Now think about what they have in common:
- Same genre? Same tempo? Same era?
- Did you listen to something similar recently?
- Do your friends listen to the same artists?
You just reverse-engineered a recommendation system. The algorithm asked the same questions — it just did it with math across millions of users simultaneously.
The Big Takeaway
Recommendation systems aren't mind-reading. They're pattern-matching at massive scale. They find people who behave like you, find items that look like things you've enjoyed, and make educated guesses.
The same technique that suggests your next Netflix binge also powers Amazon's "customers also bought," Spotify's Discover Weekly, YouTube's "Up Next," and TikTok's entire For You page.
What's Next
In the next article, we'll tackle the technology behind ChatGPT and autocomplete — how a computer learns to finish your sentences, write essays, and hold conversations. It starts with a surprisingly simple question: "Given all the words so far, what word comes next?"
This is part of the AI from Scratch series — making AI and machine learning understandable for everyone, no PhD required. Follow along on Medium or at netcausal.ai/blog.