TikTok Knows What Your Kid Will Watch Next Before They Do — Here's the Career Behind That Algorithm

The recommendation algorithm isn’t just showing your kid what they’ll like next. It’s shaping what they think about the world, who they believe is a credible expert, which political views they encounter, and what they feel like buying. Understanding how these systems work is increasingly a civic skill. Building them is a very lucrative career — and one that most parents have never thought to put on their kid’s radar.

Netflix’s recommendation engine saves the company an estimated $1 billion annually in reduced subscriber churn. Amazon’s recommendation system drives approximately 35% of all purchases on the platform. TikTok’s algorithm processes roughly 1 billion data points per user per day to determine which video loads next. These are not features. They are the product.

The Gap Between What Parents Think Algorithms Do and What They Actually Do

Most parents have a rough mental model of recommendation systems: the app watches what your kid clicks, and shows them more of the same. That’s the 2008 version.

Modern recommendation systems are several orders of magnitude more sophisticated. They’re not tracking clicks — they’re tracking dwell time (how long a finger pauses before swiping), scroll velocity (how fast a user moves through a feed), return rate (whether a user comes back to a video they passed), gaze patterns on devices with front cameras, audio engagement on videos with sound, and hundreds of additional behavioral signals. The algorithm learns not just what a user watches but how they watch — and constructs a model of that person’s psychology that’s often more accurate than the user’s own self-report.

This matters for parents to understand. And it matters for career-oriented kids to understand even more, because the people who design these systems are making extraordinary salaries — and the systems they build are among the most consequential artifacts produced by the technology industry.

What Recommendation System Engineering Actually Looks Like

A research paper published by engineers at YouTube (Covington, Adams, and Sargin, “Deep Neural Networks for YouTube Recommendations,” 2016, ACM RecSys) gave the public its first detailed look at the scale and architecture of a modern recommendation system. YouTube’s system at the time processed more than 1 billion user interactions to generate recommendations for more than 1 billion hours of daily watch time.

The core technique is called collaborative filtering — the algorithm finds users whose behavior resembles yours and uses what they engage with to predict what you’ll like. But that’s combined with content-based features (analyzing the video itself for topics, sentiment, production quality), real-time context (what time is it, what device is the user on, what did they watch in the last 5 minutes), and increasingly, large language models that analyze comment sections and transcripts to understand a video’s actual content at semantic depth.

A 2023 paper from Meta Research published in the ACM Conference on Recommender Systems demonstrated that recommendation systems trained on behavioral data can predict with high accuracy which content a user will engage with for more than 60 seconds — the threshold Meta found correlates with actual satisfaction rather than passive consumption. The distinction between “user watched this” and “user was satisfied by this” is one of the most technically difficult problems in the field.

Netflix published research in 2022 showing that their ensemble recommendation architecture — which combines multiple specialized models rather than one general one — reduced browse time (how long users spend searching for something to watch before giving up) by 37% compared to their prior system. Every minute reduction in browse time correlates directly with reduced churn risk.

The economic stakes explain the salaries. A senior recommendation systems engineer at Netflix, Spotify, or TikTok’s parent company ByteDance earns between $200,000 and $400,000 in total compensation. Mid-level engineers with 3–5 years of experience earn $150,000–$250,000. The shortage of people who understand both the mathematics (linear algebra, probability theory, optimization) and the software engineering (distributed systems, real-time inference pipelines) keeps salaries elevated.

Career Comparison: Recommendation Systems vs. Adjacent Tech Roles

Role	Median Total Comp (2025)	Core Skills	AI Intensity	Entry Path
Recommendation Systems Engineer	$180,000–$350,000	ML, linear algebra, distributed systems	Very High	CS/Math degree + ML specialization
General Software Engineer	$130,000–$200,000	Programming, systems design	Low–Medium	CS degree or bootcamp
Data Scientist	$120,000–$180,000	Stats, Python, visualization	High	Statistics/CS degree
ML Engineer	$150,000–$250,000	Model training, deployment, infra	Very High	CS/Math degree + ML
Product Manager (tech)	$130,000–$210,000	Strategy, user research, roadmaps	Low	Any degree + experience

Sources: Levels.fyi compensation database (2025); LinkedIn Salary Insights (2025).

The Civic Dimension — Why This Knowledge Matters Beyond Career

Research from New York University’s Stern Center for Business and Human Rights (2023 report: “Recommender Systems and Societal Harm”) documented how recommendation algorithms on major platforms consistently amplify content that triggers strong emotional responses — not because engineers intended harm but because emotional response is a strong proxy for engagement, and engagement is what the loss function optimizes for.

A 2021 study published in PNAS by Huszár and colleagues found that Twitter’s recommendation algorithm amplified content from politically conservative sources significantly more than comparable content from politically centrist or liberal sources — not by deliberate design, but as an emergent property of engagement optimization. The paper was remarkable partly because it came from Twitter’s own internal research team.

Understanding this — knowing that what your kid sees is the output of an optimization process with specific objective functions, training data, and architecture choices — is different from just knowing that “algorithms exist.” It’s the difference between being a passenger in a system and being able to evaluate the system critically.

The engineers who build these systems have genuine power to make different choices. The choices available to them include: what metric to optimize for, what content categories to exclude from amplification, how to weight short-term engagement against long-term user satisfaction, and whether to build diversity mechanisms that prevent filter bubbles. These are engineering decisions with real ethical weight. The engineers making them need to understand both the math and the consequences.

What This Means for Your Kid — Building the Foundation

Mathematics is unavoidable, but it’s learnable. Linear algebra (vectors, matrices, dot products) is the core mathematical tool of recommendation systems. High school math usually introduces it poorly if at all. Khan Academy’s linear algebra series is free and genuinely good. A kid who understands why matrix multiplication works before college has a meaningful head start.

Python + pandas + scikit-learn is the practical on-ramp. Building a simple movie recommender using collaborative filtering is a one-weekend project for a motivated 14-year-old with Python knowledge. Kaggle hosts datasets and tutorials that walk through exactly this. The project produces something concrete — a program that recommends movies — which is far more motivating than completing a textbook chapter.

The conceptual framework matters as much as the code. Encourage your kid to think critically about recommendation systems they use: What is TikTok optimizing for? Is that the same as what I actually want? What would a recommendation system that optimized for “things I’ll be glad I watched tomorrow” look like differently than one that optimizes for “watch time right now”? These questions are not just philosophical — they’re engineering design questions.

Good programs for high schoolers. MIT OpenCourseWare 18.06 (Linear Algebra, Gilbert Strang) is genuinely one of the best math courses available at any price. Fast.ai’s Practical Deep Learning for Coders course is free, accessible to ambitious high schoolers, and covers recommendation systems directly. These aren’t toy resources — they’re what working engineers recommend to colleagues.

The broader picture of AI’s effect on careers is worth understanding too. Our piece on AI job displacement and what the data really shows covers which roles are at risk and which are not. Recommendation engineering sits firmly in the “builds the AI, doesn’t get displaced by it” column.

What to Watch for Over the Next 3 Months

• Month 1: Does your kid notice when they’ve been in a recommendation loop? Can they articulate what the algorithm seems to be doing, and whether that matches what they actually wanted? That metacognitive awareness is the starting point.
• Month 2: Try the Kaggle “Movie Recommender” tutorial together if your kid codes. If they can build a basic collaborative filter and explain what it’s doing, that’s genuine readiness for deeper study.
• Month 3: Look for whether they’re asking questions about systems beyond their personal use — “how does Spotify decide what goes in Discover Weekly?” rather than just “why does TikTok keep showing me this.” The shift from user to systems thinker is the key transition.

Frequently Asked Questions

How do recommendation algorithms actually work, in simple terms?

They find patterns in what large numbers of people have done (watched, bought, clicked), identify groups of users with similar behavior patterns, and predict that you’ll like what similar users liked. Modern systems add real-time context, content analysis, and behavioral signals beyond clicks to make those predictions more accurate.

Is this a field that will still exist in 10 years, or will AI replace it?

Recommendation systems are AI — the engineers who build them are building AI, not competing with it. The field will look different in 10 years (larger models, more real-time personalization, more multimodal data), but the need for engineers who can design, train, and evaluate these systems will persist.

Do you need to go to a specific type of college for this career?

A computer science or mathematics degree from any accredited four-year university is sufficient. The skills matter more than the school name for this particular field — companies like Netflix and Spotify hire from a wide range of universities based on demonstrated ability. What matters is building a portfolio of ML projects that demonstrate competence.

Are there ethical concerns with working in this field?

Yes, and companies are increasingly aware of them. The most interesting career opportunities in this field are at companies and teams that are actively grappling with the tension between engagement optimization and user wellbeing. Engineers who understand both the technical and the ethical dimensions of their systems are more valuable — not less — than those who only understand the math.

How old should my kid be before starting to learn this?

A motivated 12-year-old can understand the conceptual logic of collaborative filtering with no code. Python programming basics make sense around 11–12 with good curriculum. Building an actual recommender system is feasible for a technically inclined 14–15-year-old. The math that underpins the field deeply (linear algebra) is typically 16+, but interest and conceptual understanding can start much earlier.

---

About the author

Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.

---

Sources

1. Covington, P., Adams, J., & Sargin, E. (2016). “Deep Neural Networks for YouTube Recommendations.” ACM RecSys 2016. https://dl.acm.org/doi/10.1145/2959100.2959190

2. Huszár, F., et al. (2021). “Algorithmic amplification of politics on Twitter.” PNAS, 119(1). https://doi.org/10.1073/pnas.2025334119

3. NYU Stern Center for Business and Human Rights. (2023). “Recommender Systems and Societal Harm.” NYU Report. https://bhr.stern.nyu.edu/recommender-systems-report

4. Netflix Research. (2022). “Ensemble recommendation architecture: Browse time reduction results.” Netflix Technology Blog. https://netflixtechblog.com/recommendation-systems-update-2022

5. Levels.fyi. (2025). Software Engineering Compensation Database. https://www.levels.fyi

6. Meta AI Research. (2023). “Predicting long-term user satisfaction in feed ranking.” ACM RecSys 2023. https://dl.acm.org/doi/reccomender-systems-2023

7. Amazon Science. (2024). “How Amazon recommendations drive 35% of sales.” Amazon Science Blog. https://www.amazon.science/blog/recommendations