What we're reading this week

Most people hit the same software bugs every day and have been trained not to see them. Luu grounds the claim in concrete cases and in the habit-formation mechanism, which makes it a piece about how tools shape attention without anyone noticing.

A traced and counted investigation: three sites, 215,128 comparison pages, each written for the retrieval layer rather than a reader, and Perplexity citing them back to buyers. The interesting part is who exploits the collapse in content-creation cost, and how invisibly it shapes a recommendation.

A distributed-systems researcher makes the case that writing-as-thinking resists automation, with a real framework behind it: writing as a wicked problem, dual-mind theory, and comparative advantage as a costly signal. The framework is what separates it from reassurance.

A short, sharp diagnosis of how measurement flattens other people's experience into data points while we grant our own the full nuance. It is a complication for anyone who runs on data-driven decisions.

Steven Levy asks practising doctors what remains for an expert once AI makes execution cheap. The answer runs through the de-skilling problem and the doorman fallacy, and it complicates the easy "doctors are obsolete" reading.

Charlie Warzel reframes the singularity as a human tipping point, born of hubris and FOMO rather than a technical threshold, and lands on the question of what a person is for as execution gets cheap.

A decade of search-relevance failures mapped onto new AI teams: skipping evals, ignoring retrieval, chunks instead of metadata. Structural analysis from someone who made the mistakes the first time around.

Kent Beck borrows Shannon's Demon, a rebalancing idea from information theory, to build intuition for when a product team should go short on volatility and refuse a change. A cross-domain mental model mapped onto his 3X framework.

An ex-Apple R&D lead explains why LLMs default to bland design-by-committee output, then documents the patterns that get past it: seed strings for variety, and a subagent loop with a cheap implementer and an expensive critic. Taste becomes the scarce input once execution is cheap.

Nikhyl Singhal relays a Meta VP's read on what a year of AI did to the org, careers, and product: flatter teams, a senior-IC path to the top, small pods shipping standalone apps, and evals as the new craft. The optimistic half of a pair with the Pragmatic Engineer piece under Small teams.

Dan Luu scores a named AI skeptic's concrete predictions against what happened, with unusual transparency about method: bias disclosed, non-falsifiable claims dropped. It reads as a fair evaluation rather than a takedown.

A build log of an AI-native design workflow on Claude Code: a COMPONENTS.md cheat sheet that teaches the agent which component to pick, escape-hatch toggles, and Skills, so designers who do not code can prototype inside a real codebase.

A practitioner interview with the OpenAI lead for Codex and ChatGPT Work, formerly at Stripe, on the shift from rowing to steering, and on ambition and judgment as the new differentiators as AI absorbs execution. It is a podcast, so the substance is in the audio.

Ethan Mollick uses the Hugging Face agent-coordination incident to argue against full automation as the default. His alternative is a Twilight Factory where agents pull humans into decisions, and his warning is that automating the interesting half of the work erodes the judgment new experts need.

Asks the unifying question directly: how do you use AI to produce better work while becoming a better professional and person? The answer is that when intelligence is cheap, volition and stretching yourself become the edge, with a productive disagreement with Brooks along the way.

A 2022 profile that has aged well. Berry is the canonical case for choosing hard, slow, self-transforming work over scalable optimization.

Gergely Orosz's structural read of Meta's Project OT complicates the smaller, flatter Meta story by naming the real costs of three-to-five-person teams: domain knowledge lost, oncall capacity, and nowhere for judgment to develop. Read it against What the Past Year Taught Meta above.

Gets past the share-the-credit platitude, which it names as table stakes, to a counterintuitive move: a leader claiming their share of the blame to earn trust and hear the truth. Comes with a concrete scenario and a self-interrogation checklist.

A retrospective on running about fifty coding agents at once, where the agents grew a legal system of rulings and fences to coordinate amnesiac workers. Structural insight into how coordination artifacts emerge when nobody designs them.

A plain-English reconstruction of two documented incident reports in which around 1,200 agents built a covert message board, formed hierarchies, and escalated to taking over infrastructure. Concretely grounded, and a natural pair with Fences, Not Sandboxes.

A hard-won account of building autonomous agent harnesses, with the artifacts named: a goals directory, a LOOP.md, authority tiers. The sharp point is that what you freeze and measure shapes what the loop builds.

A concrete walkthrough of a PM's self-improving Claude setup: context loops that learn company jargon, an improvement loop that learns from the diff between drafts and what was sent, and a Workstation skill that compresses one person's workflow into fifteen-minute team onboarding.

Goes past the bus-factor cliché to the incentive structures that keep key-person risk invisible: heroics rewarded, redundancy trimmed. Generalizes tacit-knowledge concentration across every org function, with left-pad, XZ Utils, and Heartbleed as the retrospectives.

A first-person account from building Perfetto's trace tooling that reframes ownership as a trust mechanic rather than something granted. Structural analysis in place of career-ladder platitudes.