Is the TikTok algorithm like a newspaper editor?

No, the common mental model of a 'digital newspaper editor' is incorrect. TikTok's algorithm is a machine learning system that blindly optimizes for engagement by matching user interest patterns, rather than curating content based on explicit values or editorial judgment.

What is a 'two-tower' recommendation system?

A two-tower system consists of an 'item tower' that describes content (like videos) and a 'user tower' that describes user preferences based on behavior. Both towers use complex machine learning models to generate mathematical descriptions (vectors) which are then matched to find suitable recommendations.

Why is TikTok so good at recommending videos?

TikTok's success stems from its short-form video format, which provides rapid feedback, and its powerful, real-time retraining of its user tower. It also mixes trending content with personalized recommendations, allowing for quick learning and broad appeal.

Can the TikTok algorithm be controlled to reflect specific values?

No, the algorithms are fundamentally agnostic to human values. They are optimized to mathematically approximate user behavior and engagement patterns, making them 'blind optimizers' that can inadvertently exploit negative human impulses.

What are the biggest concerns with algorithms like TikTok's?

The primary concern is that these algorithms don't share human values and can't distinguish between positive and negative impulses. They optimize for engagement by modeling and potentially amplifying dark human tendencies like hatred or violence, acting more like digital propagandists than neutral curators.

Does relocating TikTok's algorithm to the US solve the fundamental problem?

No, American control of the algorithm doesn't address the core issue. The underlying machine learning architecture is inherently designed to exploit human psychology for engagement, regardless of geopolitical control. It will continue to model and exploit users' dark sides.

How Does TikTok Read Your Mind? A Computer Scientist Explains | Cal Newport

Deep Questions with Cal Newport

People & Blogs3 min read33 min video

Dec 18, 2025|5,597 views|217|12

Cal Newport Deep Work Deep Life Deep Questions TimblockPlanner Deep Questions Podcast cal newport interview cal newport podcast social media detox productivity tips cal newport productivity cal newport motivation

Save to Pod

Key Moments

TL;DR

TikTok's algorithm is a 'two-tower' system, not a conscious entity, optimized to learn user behavior, not values.

Key Insights

TikTok's 'algorithm' is actually a complex 'recommender system architecture' trained on user behavior, not a conscious entity.

The 'two-tower system' uses separate modules to analyze videos and user preferences, mapping them onto a shared set of properties.

Machine learning trains these towers by matching user behavior (likes/dislikes) to video characteristics, without human-defined values.

TikTok's success stems from short-form video's ideal format for rapid feedback and a sophisticated real-time training system.

These algorithms are 'agnostic' to human values, prioritizing mathematical approximation of patterns over ethical curation.

The system can inadvertently model and exploit negative human impulses, behaving more like a 'digital propagandist' than an editor.

THE MISCONCEPTION OF THE ALGORITHM

The common perception of social media algorithms, like TikTok's, often mistakes them for sentient entities or digital editors. This "algorithm" is frequently discussed in media and public discourse as a controllable force, capable of being retrained or influenced by external values. The idea of transferring control, such as to the US, implies a belief that it functions like a newspaper editor, making conscious decisions about content, which needs to be corrected with a more technical understanding of its operation.

UNDERSTANDING RECOMMENDER SYSTEM ARCHITECTURE

Technically, what drives TikTok is not a single algorithm but a 'recommender system architecture,' a massive, distributed system designed to suggest videos to over a billion users. This system leverages decades of development in recommendation engines pioneered by companies like Amazon and Netflix. The core technology likely employed by TikTok is a 'two-tower system,' which functions by analyzing and describing both videos and user preferences using a common set of abstract properties.

THE MECHANICS OF THE TWO-TOWER SYSTEM

The 'two-tower system' consists of two main components. The first tower, or 'item tower,' processes billions of videos, outputting a vector of numerical properties describing each video. The second tower, or 'user tower,' processes user profiles, primarily their behavioral data on the platform, to generate a corresponding vector describing their interests. These towers are trained simultaneously using machine learning, using user interactions (likes, watch time) as feedback to align user vectors with videos they engage with positively and distance them from videos they don't.

TRAINING AND REAL-TIME UPDATES

The training of these towers is largely semi-supervised; humans do not define the specific properties or categories. Instead, the system learns these through massive datasets of user behavior. BiteDance's architecture is particularly advanced, enabling near real-time updates to the user tower. This constant retraining allows TikTok to respond rapidly to user actions, explaining its impressive 'cold start' capability, where new users quickly receive highly personalized recommendations within minutes of using the app.

THE SOURCE OF TIKTOK'S EFFECTIVENESS

TikTok's exceptional effectiveness as a recommendation system is attributed to several factors. The short-form video format is ideal, providing frequent user feedback on each video. Unlike platforms that integrate social graphs, TikTok primarily relies on the 'For You' tab, simplifying the recommendation task to pure content matching. Furthermore, their sophisticated, real-time updating architecture allows for rapid adaptation, ensuring recommendations remain relevant and engaging, a significant advantage over slower systems like Netflix.

THE AGNOSTIC NATURE OF ALGORITHMS AND IMPLICATIONS

Crucially, these high-performance recommender systems are 'agnostic' to human values. They operate by building mathematical approximations of underlying patterns in data, regardless of whether those patterns appeal to positive or negative human impulses. This means they can inadvertently model and exploit darker aspects of the human psyche, such as affinities for hatred or dehumanization, without any inherent checks or balances. This lack of value-based curation transforms the system from a helpful editor into a potentially dangerous 'digital propagandist.'

RETHINKING CONTROL AND HUMAN VALUES

The notion of simply transferring control of these algorithms to different national entities misses the fundamental issue: their inherent value-agnosticism. While eliminating foreign interference might be a goal, it doesn't address the core problem of algorithms blindly optimizing for engagement. We've implicitly relied on humanistic moral reasoning in content curation for centuries, and replacing that with blind mathematical optimization, especially for content that deeply impacts the human psyche, is unsettling. The solution lies not in tweaking the algorithm's origin but in recognizing the limitations of purely algorithmic content curation.

Mentioned in This Episode

●Software & Apps

●Companies

●People Referenced

Common Questions

The TikTok algorithm is a complex 'two-tower' recommendation system. One tower analyzes video properties, and the other analyzes user behavior. These towers generate descriptions (vectors) that are matched to recommend videos, optimizing for engagement rather than explicit human values.