Why social media is different from past moral panics about technology

Executive overview

Critics of social media are often dismissed with the "moral panic" argument: we overreacted to comic books and television, so we're probably overreacting now. But social media content differs in kind, not just degree, from earlier media.

The ultra-processed food analogy explains why: TikTok-style content is engineered to hijack desire mechanisms the same way Doritos are — and the right response to both is elimination, not moderation.

The moral panic objection and why it fails for social media

• The standard pushback: every new medium (comics, TV) triggered moral panic, and we were fine
• Processed food analogy shows this logic has limits — there's a meaningful line between processed and ultra-processed
• Pasta and Kugel = processed; Doritos and Oreos = ultra-processed. We treat them differently
• Social media content = ultra-processed: stock user-generated material algorithmically reconstituted into hyper-palatable form
• Streaming TV, NYT games = processed media; fine in moderation, not fine as the whole diet
• Ultra-processed content is not culturally rooted, not hard to eliminate, and causes demonstrable psychological harm
• Conclusion: internally consistent case that social media is genuinely exceptional, not just the next step in a normal trajectory

How TikTok's algorithm actually works

• Common mental model — a scheduler matching index cards of user interests to content — is wrong
• Reality: a weighted multi-dimensional scatter plot; each video is a vector of numbers, not semantic labels
• Watch time (percentage watched) is the primary input signal; no rich user profile is maintained
• Algorithm loop: probabilistically sample videos, weight toward clusters with high prior watch time, occasionally explore randomly
• "Multi-armed bandit" stochastic optimisation — simple logic, run at scale, produces eerily accurate results fast
• Consequence: no nuanced control knobs exist; you can filter extreme content or turn the system on/off, but little in between
• The algorithm is not intelligent — it is simple, fast, and self-reinforcing

Building analytical thinking capacity

• Immediate fix: eliminate context-switching during hard thinking; even brief phone checks cut effective IQ non-trivially
• Start with the 20-10 rule: 20 pages of reading per day, 10 minutes of solitary reflection
• Reading builds unique neural circuits — humans have no innate reading circuitry, so training it is genuine cognitive development
• Reflection practice means following a single thought to its source; must be practiced separately from reading
• Progress incrementally: 20/10 → 30/15 → 50+/20-30; at the upper range, cognitive capacity exceeds most people's
• Walk during reflection time — movement aids thinking (the Peripatetic school was not accidental)

Remote work and time blocking

• Without physical separation or visible supervision, motivation circuits weaken; structure must be self-imposed
• Time block plan every remote day: defined start, task blocks, breaks, email windows, shutdown
• Keep working hours tight and intense; when done, be done — no blending
• Structure matters more remotely than in-office, not less

Secondary skills and avoiding pseudo-productivity

• Secondary skills (typing speed, tool customisation) are often used to avoid the stress of direct performance evaluation
• "Busy and innovative" can mask avoidance of the question: how good is my actual output?
• Fix: obsess first over producing unambiguous, measurable value — promotions, client revenue, shipped work
• Secondary skills become genuinely useful only once there is a concrete bottleneck they would remove

AI agents and workflow

• Agentic AI is currently most useful in one domain: software development (Cursor, Windsurf, Roo)
• Core loop: LLM plans steps → agent executes → errors caught → repeat per step; enables more complex builds than single-prompt coding
• Main practical problem: inference latency creates idle time; filling it with social media causes context collapse
• Better fill: stay within project cognitive context (e.g. work on the project to-do list) rather than switching fully out
• The "waiting for compile" problem is decades old; agentic AI adds a new instance of a familiar constraint

Books read in May 2025

1. Building — Mark Ellison; memoir on craftsmanship by a high-end Manhattan carpenter
2. Thoreau's Acts — Caleb Smith; academic survey of attention and focus in 19th-century American writing
3. Nine Innings — Daniel Okrent; classic baseball book structured around a single Brewers–Orioles game
4. The Let Them Theory — Mel Robbins; long-running #1 bestseller
5. Against a Machine — Paul Kingsnorth (advance copy, fall release); anti-machine critique rooted in place, folklore, and tradition rather than classical left theory