From Hacker News to TikTok - How Algorithms Learned to Hook Us

What It Covers

A developer named Corey gets trapped in Instagram's AI cat video loop, prompting a deep investigation into how social media ranking systems evolved from Hacker News's simple gravity-based upvote algorithm through Facebook's engagement optimization and YouTube's collaborative filtering to TikTok's real-time 30-minute behavioral modeling.

Key Questions Answered

• •Hacker News Gravity Algorithm: The original 1005 Hacker News ranking system, written in Arc Lisp, works like Flappy Bird: upvotes push content up, time decay pulls it down. No machine learning, no personalization. This "gravity" model creates shared community front pages where everyone sees the same consensus content, making it structurally incapable of personalized rabbit holes.
• •Facebook's MSI Trap: In 2017, Facebook introduced Meaningful Social Interactions as its core engagement metric, optimizing for comments, shares, and reactions. This inadvertently replicated "sort by controversial" at 2 billion-user scale. Internal Haugen documents, now court evidence in 1,700+ lawsuits, show researchers identified the anger-amplification effect but leadership declined fixes that would reduce growth numbers.
• •YouTube's Collaborative Filtering Model: YouTube's Covington-Adams-Sargon paper reveals the system takes a user's last 50 watched videos, maps them into vector space, finds users with similar coordinates, then recommends what those users watched next. This batch-processing model updates nightly, meaning preference shifts take weeks to register, producing broader, slower-moving recommendation drift.
• •TikTok's 30-Minute Real-Time Model: TikTok's leaked ByteDance engineering document confirms the system uses only the last 30 minutes of viewing behavior as its primary signal, updating via a streaming Kafka-Flink pipeline after every swipe or pause. The Wall Street Journal's bot test found 93 of 224 videos served to a "sad" profile within 36 minutes were about depression or self-harm.
• •Algorithm Reset as Practical Intervention: Meta provides a documented reset function under Instagram's content preferences labeled "reset suggested content," which clears recommendation history across Explore, Reels, and Feed simultaneously. For users over-indexed into a narrow content category through concentrated interaction, this is the fastest available mechanism to force the algorithm to rebuild its behavioral model from scratch.