Deep Questions with Cal Newport

Ep. 372: Decoding TikTok’s Algorithm

September 29, 2025

Key Takeaways Copied to clipboard!

  • The TikTok recommendation system is likely a highly optimized, two-tower machine learning architecture, not a controllable 'digital newspaper editor,' which allows it to relentlessly model and exploit human patterns, including dark impulses, without inherent ethical guardrails. 
  • The unique success of TikTok's algorithm stems from its platform format—short, sequential video delivery providing rapid feedback—which is ideal for training these recommendation systems faster and more effectively than other social platforms. 
  • Transferring control of the TikTok algorithm to US entities will not fundamentally solve the platform's dark impulses because the problematic behavior is baked into the blind mathematics of machine learning-based content curation, not just foreign influence. 
  • The pivot from hype about AI scaling to quantum computing as the next breakthrough reveals that some commentators are driven by a psychological need for massive technological disruption rather than deep technical understanding. 
  • Quantum computing is a highly specialized field requiring deep knowledge of quantum mechanics and is not a near-term solution for breaking the scaling barriers in generative AI. 
  • Lifestyle-centric planning, which involves systematically improving all aspects of one's life rather than pursuing one radical change, is an effective strategy for cultivating a deep life that resists the pull of algorithmic curation like TikTok. 
  • Parental curation of values through active involvement in children's lives is crucial to counteract the value-free, attention-maximizing curation provided by algorithms on platforms like TikTok, especially for children approaching high school age. 
  • Algorithmic curation systems, by definition, do not share human values, and relying on them for mass information distribution is inherently a recipe for cultural disaster, leading to bizarre, dangerous, or shallow content trends. 

Segments

TikTok Deal and Algorithm Fear
Copied to clipboard!
(00:00:03)
  • Key Takeaway: The Oracle deal aims to secure US control over TikTok’s recommendation technology to mitigate foreign influence fears.
  • Summary: Oracle will oversee the security of American user data and monitor updates to TikTok’s recommendation technology. A copy of the core algorithm will be licensed to an American investor group holding an 80% share. This deal addresses the fear that a foreign government could gain influence via the powerful recommendation engine.
Algorithm Mental Model Correction
Copied to clipboard!
(00:04:04)
  • Key Takeaway: The common mental model of a social media algorithm as a ‘digital newspaper editor’ is incorrect for modern systems.
  • Summary: The public often imagines the algorithm as a conscious editor making value-based decisions about content placement. In reality, these systems are complex, automated recommendation architectures. Transferring control based on the editor model may be misguided if the underlying technology operates fundamentally differently.
Two-Tower Recommendation Architecture
Copied to clipboard!
(00:06:58)
  • Key Takeaway: TikTok likely uses a two-tower system where item and user descriptions are mapped into a shared property vector space.
  • Summary: The system is a distributed recommender architecture, not a single algorithm, likely employing a two-tower approach: one tower describes videos (items) and the other describes user interests (users) using the same set of latent properties. These towers are trained semi-supervised using user behavior data (likes/dislikes) to ensure user vectors closely match liked item vectors.
TikTok’s Performance Advantage
Copied to clipboard!
(00:19:18)
  • Key Takeaway: TikTok’s superior performance is due to its short-form video format and a highly advanced, real-time distributed system architecture.
  • Summary: Short-form video is the ideal format because it maximizes user feedback per session, allowing the system to learn preferences rapidly. ByteDance built an impressive distributed system capable of retraining the user tower almost in real-time based on immediate user interaction data. This architecture enables the platform’s famous ‘cold start’ capability, learning new users’ interests within minutes.
Machine Learning Blindness to Values
Copied to clipboard!
(00:25:22)
  • Key Takeaway: Machine learning recommendation systems are agnostic to human values, blindly approximating underlying patterns, which inevitably includes modeling negative human impulses.
  • Summary: These systems are optimized only to win a training game by creating mathematical approximations of observed patterns, regardless of whether those patterns reflect positive or negative human traits. This process means the system will model affinities for hatred or violence just as easily as it models positive interests, creating a ‘digital propagandist’ rather than a curated editor.
Value-Centric Approach to Tech Use
Copied to clipboard!
(00:30:51)
  • Key Takeaway: The fundamental problem with social media is baked into the mathematics of machine learning curation, making simple fixes like changing ownership ineffective.
  • Summary: Cal Newport advocates for a value-centric, pragmatic approach to technology use: if a product reduces valued aspects of life without commensurate upside, it should be avoided. Relying on algorithms for civic functions like news delivery removes essential humanistic moral guardrails that have historically governed mass content production.
Ethical Sorting vs. Value Assessment
Copied to clipboard!
(00:44:46)
  • Key Takeaway: An ethical obligation to stop using social media based on who else uses it is an untenable ’ethical sorting’ approach; focus instead on whether the technology harms personal or civic values.
  • Summary: The speaker rejects the idea of constantly policing associations or consumer choices based on external ethical rankings, finding it leads to tribalism. A functionalist assessment—does this technology make my life better or worse?—is the more practical framework for engagement.
News Consumption via Algorithms
Copied to clipboard!
(00:49:41)
  • Key Takeaway: Using TikTok for news is a terrible method because algorithmic curation lacks the necessary human ethical guardrails present in traditional media.
  • Summary: Algorithmic curation is the ‘worst possible curation’ because it lacks human empathy, truth correlation, or normative standards. Staying informed requires human curation, even if biased, as humans inherently integrate cultural and ethical guardrails that blind machine learning systems ignore.
Quantum Computing and AI Shift
Copied to clipboard!
(00:53:48)
  • Key Takeaway: The current hype linking quantum computing to an imminent AI superintelligence phase shift is likely a reaction to the scaling limits encountered by current LLMs.
  • Summary: Understanding technology’s core operation is crucial for ethical assessment, which is why this question is relevant to the algorithm discussion. Current large language models have hit scaling walls, leading to speculation that quantum computing will provide the necessary breakthrough for superintelligence. This speculation attempts to revive hype after scaling plateaus were reached.
Quantum Computing vs. AI Hype
Copied to clipboard!
(00:54:55)
  • Key Takeaway: The sudden pivot to quantum computing as the key to superintelligence reveals the psychological motivation behind previous AI hype.
  • Summary: The speaker notes that many people are asking about quantum computing unlocking superintelligence after scaling limits were hit in current language models. Quantum computers are fundamentally different from classical computers, relying on quantum mechanics for specific problems like factoring large primes or certain simulations. There is no obvious connection between current quantum capabilities and breaking the scaling barriers required for generative AI.
Case Study: Composer’s Deep Life
Copied to clipboard!
(01:02:29)
  • Key Takeaway: Lifestyle-centric planning, which systematically improves all facets of life, succeeds where radical, single-focus career changes often fail.
  • Summary: Kieran initially pursued a radical change to become a full-time composer, which led to burnout and financial decline by ignoring other lifestyle factors like security. True success came from lifestyle-centric planning, where he maintained computer consulting to fund and support his composition ambitions, allowing him to work deeply in both areas. Constructing a deep life outside of digital distractions makes it easier to resist algorithmic curation.
Parenting and Digital Curation
Copied to clipboard!
(01:08:58)
  • Key Takeaway: Parents must actively serve as the primary filter for their children’s values, preventing exposure to the value-agnostic curation of social media algorithms.
  • Summary: The speaker prioritizes spending time with his children, coaching sports and serving on school boards, because this time is essential for implanting normative values. Unrestricted access to platforms like TikTok exposes children to algorithmically curated values that lack a humanistic core, which can lead them down negative paths. Parents must be the interpreter of the world, deciding what is valuable, rather than outsourcing that role to recommendation systems.
TikTok Article Reactions
Copied to clipboard!
(01:18:04)
  • Key Takeaway: TikTok’s algorithmic curation amplifies fringe content, leading to negative societal outcomes ranging from inadequate child data protection to the viral spread of doomsday predictions and dangerous stunts.
  • Summary: The first article revealed Canadian privacy officials found TikTok’s child data protection inadequate, highlighting the danger of letting children under 13 use the app. Another article showed how the algorithm amplified a niche ‘Rapture Talk’ trend, giving wide exposure to an unsubstantiated prediction of the world ending. A third report detailed teens facing criminal charges after a friend died during a stunt promoted by the algorithm, which favors attention-grabbing, dangerous content.