How Does TikTok Actually Work (It’s Scary…)

How Does TikTok Actually Work? (It’s Scary…) 🔗 **Source:** [How Does TikTok Actually Work? (It’s Scary…)](https://www.youtube.com/watch?v=L0Ez2Aw-krQ) > [!note]+ Document Notes > tags: [TikTok, Recommender Systems, Digital Ethics] > author: [Cal Newport] > source: [YouTube] > reference: [https://www.youtube.com/watch?v=L0Ez2Aw-krQ] > date created: {{date:YYYY-MM-DD}} {{time:HH:mm}} > > # 🏷️ Title: Why TikTok Knows You Better Than You Know Yourself >  ## 🧭 Table of Contents - [📝 Summary](#-summary) - [📖 Main Content](#-main-content) - [🗒️ Notes](#️-notes) - [📚 References](#-references) - [🗣️ Paraphrase](#️-paraphrase) - [📰 Newsletter](#-newsletter) - [✅ Action Items](#-action-items) - [🧪 Hypotheses](#-hypotheses) - [📊 Methodology](#-methodology) - [❓ Questions](#-questions) - [🔮 Future Research](#-future-research) - [💡 Implications](#-implications) - [🧠 Learning Styles Assessment](#-learning-styles-assessment) - [📊 Accessibility Assessment](#-accessibility-assessment) - [🌍 DEIB Assessment](#-deib-assessment) - [✅ Logical Validity](#-logical-validity) - [🛡️ DARVO Analysis](#-darvo-analysis) - [🎯 Conclusion](#-conclusion) ### 🔗 References - Include links cited in or relevant to the article. ## 📝 Summary - The episode explains TikTok’s recommendation architecture, not a single “algorithm.” [Video](https://www.youtube.com/watch?v=L0Ez2Aw-krQ) - It likely uses two-tower models, fast feedback, and near-real-time updates. [Video](https://www.youtube.com/watch?v=L0Ez2Aw-krQ) - Stakeholders include users, ByteDance, Oracle, regulators, and media. [Video](https://www.youtube.com/watch?v=L0Ez2Aw-krQ) - Risks include amplification of harmful impulses and weak human values in curation. [Video](https://www.youtube.com/watch?v=L0Ez2Aw-krQ) - Bottom line: changing owners won’t fix value-blind optimization. Human curation still matters. [Video](https://www.youtube.com/watch?v=L0Ez2Aw-krQ) ## 📖 Main Content - TikTok’s feed uses a distributed recommender system, not a tweakable editor. [Video](https://www.youtube.com/watch?v=L0Ez2Aw-krQ) - Two towers encode items and users into vectors within a shared space. [Video](https://www.youtube.com/watch?v=L0Ez2Aw-krQ) - Candidate retrieval uses similarity; ranking applies business rules and heuristics. [Video](https://www.youtube.com/watch?v=L0Ez2Aw-krQ) - Short videos yield dense feedback, enabling rapid personalization. [Video](https://www.youtube.com/watch?v=L0Ez2Aw-krQ) - Real-time training updates user embeddings as sessions unfold. [Video](https://www.youtube.com/watch?v=L0Ez2Aw-krQ) - Trending signals mix with personal fit to extend exploration. [Video](https://www.youtube.com/watch?v=L0Ez2Aw-krQ) - Ethical concern: systems optimize engagement without moral judgment. [Video](https://www.youtube.com/watch?v=L0Ez2Aw-krQ) - Policy moves can restrict data access, yet won’t resolve value blindness. [Video](https://www.youtube.com/watch?v=L0Ez2Aw-krQ) - Quantum computing digression: interesting, but not a near-term LLM panacea. [Video](https://www.youtube.com/watch?v=L0Ez2Aw-krQ) ### 🗒️ Notes #### ⏱️ [00:00:05](https://www.youtube.com/watch?v=L0Ez2Aw-krQ&t=5) Oracle, data security, and US oversight - Oracle will guard US data and monitor changes. Policy goals drive this. [Video](https://www.youtube.com/watch?v=L0Ez2Aw-krQ) #### ⏱️ [00:02:00](https://www.youtube.com/watch?v=L0Ez2Aw-krQ&t=120) “Algorithm” hype vs reality - Media treats “the algorithm” as a unitary gadget. It is a system. [Video](https://www.youtube.com/watch?v=L0Ez2Aw-krQ) #### ⏱️ [00:07:00](https://www.youtube.com/watch?v=L0Ez2Aw-krQ&t=420) From algorithm to architecture - Think “recommender architecture,” not a magic editor. [Video](https://www.youtube.com/watch?v=L0Ez2Aw-krQ) #### ⏱️ [00:12:00](https://www.youtube.com/watch?v=L0Ez2Aw-krQ&t=720) Two-tower mental model - One tower encodes videos; one tower encodes users. Shared properties emerge. [Video](https://www.youtube.com/watch?v=L0Ez2Aw-krQ) #### ⏱️ [00:18:00](https://www.youtube.com/watch?v=L0Ez2Aw-krQ&t=1080) Retrieval and ranking - Similarity search narrows candidates; heuristics decide the next clip. [Video](https://www.youtube.com/watch?v=L0Ez2Aw-krQ) #### ⏱️ [00:21:00](https://www.youtube.com/watch?v=L0Ez2Aw-krQ&t=1260) Why short video wins - Short clips create strong, frequent labels. Cold start shrinks fast. [Video](https://www.youtube.com/watch?v=L0Ez2Aw-krQ) #### ⏱️ [00:22:00](https://www.youtube.com/watch?v=L0Ez2Aw-krQ&t=1320) Near real-time updates - User embeddings retrain continuously across a distributed cloud. [Video](https://www.youtube.com/watch?v=L0Ez2Aw-krQ) #### ⏱️ [00:23:40](https://www.youtube.com/watch?v=L0Ez2Aw-krQ&t=1420) Mix of personal and trending - Popularity signals inject novelty alongside fit. Feedback loops grow. [Video](https://www.youtube.com/watch?v=L0Ez2Aw-krQ) #### ⏱️ [00:27:00](https://www.youtube.com/watch?v=L0Ez2Aw-krQ&t=1620) Ethics and content power - Mass content affects norms. Systems map our darker drives too. [Video](https://www.youtube.com/watch?v=L0Ez2Aw-krQ) #### ⏱️ [00:31:00](https://www.youtube.com/watch?v=L0Ez2Aw-krQ&t=1860) Governance limits - US control may reduce security risk. It won’t add ethics. [Video](https://www.youtube.com/watch?v=L0Ez2Aw-krQ) #### ⏱️ [00:55:00](https://www.youtube.com/watch?v=L0Ez2Aw-krQ&t=3300) Quantum detour - Quantum helps niche problems; not LLM salvation now. [Video](https://www.youtube.com/watch?v=L0Ez2Aw-krQ) ## 📚 References - Cal Newport — *Why TikTok Knows You Better Than You Know Yourself* (video). [YouTube](https://www.youtube.com/watch?v=L0Ez2Aw-krQ) - The speaker references ByteDance research papers. Links were not provided in the transcript, so they are not cited here. ## 🗣️ Paraphrase - TikTok’s feed relies on a scalable, value-agnostic recommender architecture that matches users to videos. - The design learns quickly from short videos and updates user representations almost in real time. - Human oversight cannot easily embed ethics into this optimization loop. - Therefore, governance must address incentives, guardrails, and non-algorithmic curation. ## 📰 Newsletter - TikTok’s architecture shows why value-blind optimization spreads fast. [Video](https://www.youtube.com/watch?v=L0Ez2Aw-krQ) - WordPress folks can learn from its feedback loops and guardrails. - Try human-curated sources for news in community channels. - Resources: [Learn WordPress](https://learn.wordpress.org/), [Make/Community](https://make.wordpress.org/community/), [Developer Docs](https://developer.wordpress.org/), [Openverse](https://wordpress.org/openverse/). ## ✅ Action Items - **DevRel** - Publish a primer: “How recommender systems work for publishers.” - Run office hours on ethical analytics configurations. - **Developer advocacy: plans, talks, demos, and measurable outcomes.** - Demo a two-tower toy model using public data. - Track engagement uplift without dark patterns. - **Education: workshops, curricula, office hours, mentorship programs.** - Build short courses on metrics literacy and consentful analytics. - Mentor contributors on telemetry minimization. - **Developer experience: docs, SDKs, examples, templates, quickstarts.** - Ship examples for event logging with privacy budgets. - Add reproducible dashboards for content moderation metrics. - **Developer marketing: campaigns, content, funnels, and attribution.** - Promote “human-in-the-loop” editorial workflows. - Measure conversions with privacy-preserving attribution. - **Community management: moderation, code of conduct ops, surveys.** - Launch rapid-response moderation playbooks. - Add escalation trees and incident retrospectives. - **Internal/external/cross-project: align goals, SLAs, and feedback loops.** - Create SLAs for takedown requests and appeal handling. - Share quarterly ethics metrics with stakeholders. - **OSPO** - **Policy & Governance: contribution policies, decision rights, charters.** - Draft an AI curation policy and decision matrix. - Define when algorithmic ranking is allowed on .org properties. - **License Compliance: policies, scanning, review, and training cadence.** - Add model and dataset license checks to CI. - **Security Management: vuln intake, embargo policy, signing, attestations.** - Include model supply-chain attestations in SBOMs. - **Software Inventory: SBOM strategy, ownership, freshness SLOs.** - Track models, datasets, and feature flags as inventory items. - **Developer Support: training, tooling, legal guardrails, consultation hours.** - Offer consults on bias, safety, and privacy-by-design. - **Community Engagement: upstream funding, MoUs, operator agreements.** - Fund moderation tools used by multiple OSS communities. - **Strategic Alignment: KPIs, OKRs, executive reporting, initiative roadmaps.** - Set KPIs for harmful amplification reduction. - **Community** - **Maintainers: roadmap clinics, governance templates, release hygiene.** - Add “ethics checks” to release criteria. - **Contributors: onboarding paths, “first issue” rotations, pairing sessions.** - Pair newcomers with moderation and Docs squads. - **Sponsors: transparent budgets, milestones, deliverables, impact reports.** - Publish quarterly transparency reports on funding impact. - **Ecosystem: cross-project incident drills and coordinated disclosures.** - Run misinformation tabletop exercises with partner projects. - **WordPress** - **Core: priorities, issue triage, performance budgets, release criteria.** - Gate any ranking features behind opt-in and transparency. - **Core Performance: profiling plans, metrics dashboards, perf bounties.** - Track perf without invasive user profiling. - **Core AI: responsible AI guardrails, evaluation sets, pilot projects.** - Develop evaluation sets for editorial safety. - **Core Program: coordination rituals, cross-team dependencies, retrospectives.** - Add bias and safety reviews to release retros. - **Mobile: app backlog, telemetry review, offline UX improvements.** - Keep telemetry minimal and documented in-app. - **Playground: browser labs, reproducible bugs, education modules.** - Teach ranking pitfalls with safe sandboxes. - **Community: meetup kits, WordCamp playbooks, speaker mentorship.** - Offer talks on human curation best practices. - **Training: lesson plans, learning paths, certification pilots.** - Create “Ethical Analytics” learning path. - **Marketing: campaigns, narratives, social editorial calendars.** - Promote human editors over opaque feeds. - **Support: response SLAs, knowledge base gaps, escalation paths.** - Add policies for content takedowns and appeals. - **Design: UX audits, design tokens, research backlogs.** - Label why items appear in feeds (“Because you followed X”). - **Documentation: DevHub/HelpHub gaps, examples-first patterns, style guide.** - Document ranking transparency patterns. - **Photos: content drives, moderation guidelines, taxonomy hygiene.** - Strengthen content provenance practices. - **TV: submission guides, caption quality, content indexing.** - Enforce captions and age labeling. - **Polyglots: translation sprints, glossary alignment, locale coverage.** - Translate safety and moderation guides first. - **Accessibility: WCAG reviews, tooling, issues heatmap, audits cadence.** - Audit AI surfaces for motion, captions, and focus order. - **Meta: .org roadmap, API reliability, contributor tools.** - Log anti-abuse events with privacy budgets. - **CLI: command coverage, UX docs, plugin scaffolds, smoke tests.** - Scaffold moderation hooks and audit trails. - **Hosting: cross-host checklists, defaults, telemetry for reliability.** - Provide default rate limits and abuse protection. - **Tide: scoring transparency, false-positive review loops, plugin outreach.** - Encourage “why” messages for scores. - **Openverse: ingestion pipelines, moderation policy, education tie-ins.** - Add provenance signals and user labeling. - **Themes: review throughput, lints, mentoring, test matrices.** - Lint for invasive tracking patterns. - **Plugins: review throughput, security liaisons, provenance checks.** - Require privacy notes for any ranking features. - **The WPCC** - **Programs: fellowship tracks, microgrants, maintainer stipends.** - Fund safety tooling and documentation fellowships. - **Projects: governance clinics, documentation sprints, release teams.** - Host “ethical curation” clinics for plugin authors. - **Funding experiments: matched grants, milestone payouts, transparent ledgers.** - Pilot matched grants for moderation tech. - **FAIR** - **Technical independence: local metadata, provenance, and signature checks.** - Sign content feeds; publish verification tooling. - **Policies and covenants: funding transparency, sunset clauses, anti-capture.** - Codify anti-capture terms for funding ranking features. - **Productized governance artifacts: charters, operator agreements, scorecards.** - Offer an “Algorithmic Transparency” charter template. ## 🧪 Hypotheses - Transparent “why am I seeing this” reduces harmful engagement by 15%. - Human editorial interleaving lowers misinformation spread versus pure ranking. - Real-time updates increase fixation on extreme content under weak guardrails. - Short-form feedback loops accelerate preference formation among teens. - Visible opt-outs reduce time-on-feed without harming satisfaction. ### 📊 Methodology - Run A/B tests comparing pure ranking versus editor-mixed feeds. - Measure misinformation shares, dwell time, and report rates. - Survey teens on preference shifts after guided interventions. - Use pre-registered metrics and external auditing. ### ❓ Questions - Where should policy set hard limits on opaque ranking? - Which disclosures actually help users calibrate trust? - How can OSS projects model healthier defaults? ### 🔮 Future Research - Study long-term effects of “explanations” on behavior. - Evaluate small-team editorial councils in community feeds. - Test “cooldown” UX for high-arousal topics. ### 💡 Implications - Practice: ship transparent, opt-in ranking with human oversight. - Policy: require audit logs and user explanations. - Security: monitor coordinated manipulation using privacy-safe signals. - Education: teach metrics literacy across contributor onboarding. - Funding: prioritize moderation and provenance infrastructure. ### 🧠 Learning Styles Assessment - **Reading level estimate:** Flesch-Kincaid Grade ≈ 9–11 (brief, plain style). - **Visual learners** - Present: conceptual two-tower diagrams suggested. - Improve: flowcharts for retrieval and ranking steps. - **Aural learners** - Present: podcast episode format. - Improve: 3-minute audio summary per section. - **Verbal learners** - Present: definitions and plain terms. - Improve: side-by-side glossary and examples. - **Kinesthetic learners** - Present: none. - Improve: hands-on sandbox with safe datasets. - **Logical learners** - Present: stepwise architecture reasoning. - Improve: causal diagrams and confusion matrices. - **Social learners** - Present: discussion prompts for meetups. - Improve: cohort critiques of feed UX. - **Solitary learners** - Present: reflection prompts on media diet. - Improve: self-check quizzes. - **Naturalistic learners** - Present: none. - Improve: sustainability analogies for data footprints. ### 📊 Accessibility Assessment - **Screen reader semantics** - Present: headings and lists help navigation. - Improve: consistent alt text and language tags. - **Keyboard navigation** - Present: clear focus order recommended. - Improve: remove traps; ensure tab reachability. - **Contrast and readability** - Present: enforce 4.5:1 minimum contrast. - Improve: verify line height and spacing. - **Text resizing and reflow** - Present: responsive content suggested. - Improve: support 200% zoom without clipping. - **Media and captions** - Present: captions for videos. - Improve: downloadable transcripts and audio descriptions. - **Forms and interactions** - Present: labeled controls advised. - Improve: inline validation and forgiving inputs. - **Standards reference** - Map gaps to WCAG 2.2 AA criteria. ### 🌍 DEIB Assessment - **Cultural breadth** - Present: global platform context. - Improve: region-specific case studies and translations. - **Inclusive language** - Present: neutral, defined terms. - Improve: add plain-language summaries. - **Neurodiversity support** - Present: chunked sections. - Improve: icon+text labels and predictable layouts. - **Representation** - Present: varied stakeholder framing. - Improve: more diverse examples and imagery alt text. - **Psychological safety** - Present: respectful tone. - Improve: publish community guidelines and escalation paths. ### ✅ Logical Validity - **Strong reasoning** - Inductive link from architecture to outcomes is clear. [Video](https://www.youtube.com/watch?v=L0Ez2Aw-krQ) - Causal claim ties feedback density to personalization. [Video](https://www.youtube.com/watch?v=L0Ez2Aw-krQ) - **Fallacies** - Risk of hasty generalization across platforms. Mitigate with comparative data. - **Tone of voice** - Analytical, cautionary, and practical. ### 🛡️ DARVO Analysis - **Deny** - No explicit denial; system behavior framed as agnostic. - **Attack** - No personal attacks; critique targets mechanisms, not people. - **Reverse victim/offender** - None observed; emphasis stays on structural incentives. ## 🎯 Conclusion - TikTok succeeds through efficient, value-blind optimization. [Video](https://www.youtube.com/watch?v=L0Ez2Aw-krQ) - Ownership changes won’t add ethics. Human curation and transparency will. - Next steps: adopt opt-in ranking, publish explanations, and strengthen governance across WordPress and The WPCC programs.