What we're reading this week

Gopnik's canonical "Never-Betters / Better-Nevers / Ever-Wasers" taxonomy of how every generation reacts to new information technology. The timeless cross-domain frame that maps onto the printing press, the calculator, and the LLM alike.

DuckDuckGo founder Gabriel Weinberg triangulates five usage/survey datasets to argue AI adoption has stalled into thirds (active/occasional/never), explicitly checking his own early-adopter bubble. Structural, data-grounded analysis that complicates the consensus.

Treats conversation (and LLMs) as cognitive infrastructure, grounded in Mercier/Sperber, Vygotsky, Clark/Chalmers, with the sharp insight that LLMs deliver only half the dividend because sycophancy strips out the disagreeing-listener half that actually builds understanding.

Structural practitioner retrospective locating the irreducible core of understanding the agent can't supply (knowing what "right" looks like) and naming exactly what becomes scarce when code-writing collapses to near-free.

First-person retrospective arguing durable value is the tacit "senior thinking" (blast radius, sequencing, reversibility) with no written corpus, and that agents make him better by forcing him to externalize that judgment rather than eroding it.

A durable methodology for evaluating any new model: test it on the hardest problem you already understand cold, judge comprehension before output, and keep personal git-SHA benchmarks of bugs that beat the models.

Chip Morningstar's Fujitsu story (complete source code and unlimited access, yet they "lost the architecture" because client-server wasn't in their experience): a concrete demonstration that you can transmit documentation but not the mental model.

Slack DevXP practitioner retrospective running 200+ measured agentic E2E runs with hard numbers on failure rates, token cost, and turn counts. A warts-and-all look at where deterministic correctness vs. agent judgment belongs.

Addy Osmani's practitioner framework (match review effort to the cost of being wrong; reserve scarce human attention for what only humans can do) backed by 2026 telemetry. The cleanest statement of this week's through-line.

Charity Majors argues that when code becomes free and disposable, the scarce moats shift to evaluation, durability, observability, and production discipline, reframing the AI shift as a return to rigor.

Structural analysis of where value concentrates when frontier models get commoditized, woven with Nadella's "human capital vs. token capital" framing of what humans hold onto as execution gets cheap; Thompson holds the cynical and charitable readings at once ("I respect this alignment, and I fear it").

Uses Jevons Paradox plus Lovable builder data (80% non-technical, 55% with 11+ years experience) to argue the AI shift is an "economic participation story," not a developer-productivity one: domain experts, not faster developers, win as the translation layer collapses.

Practitioner maturity model (agent-first, structured orchestration, AI-driven composability) drawn from a real workflow, arguing reliability, observability, and recovery (not raw AI smarts) are where durable value sits once creation gets cheap.

Hits the unifying question head-on: AI didn't take the programmer's job, it split it. The loved hand-crafting part got commoditized "like IKEA furniture," leaving the harder human work of understanding what to build and stewarding the "symmathesy" of people, code, and agents learning together.

Julia Evans makes the case (as a comic/zine) that writing, and by extension building, for one specific person produces clearer, more genuine work than optimizing for a vague mass audience. A cross-domain match for the "audience of one is a beautiful heresy" instinct.

Deeply-sourced structural autopsy of how a high-performance engineering culture gets demolished by AI over-indexing: "tokenmaxxing" as Goodhart's law, gutted security teams causing the most embarrassing outage in Meta history, the inverse case study of invisible high-leverage infrastructure.

Will Larson revises his leadership rules based on a year of concrete projects (200–400 weekly deploys, up from 6; one-engineer migrations; agent-first triage), with structural analysis of why the "development harness" gates everything. A writer-changing-his-mind retrospective, not generic advice.

Luca Rossi proposes Product Decision Records and a Product Glossary as planning artifacts that encode judgment/domain language so AI grounds specs in the actual product surface (rework dropping 60% to 20%), injecting taste by showing how decisions were made, not by rules.

antirez (Redis creator) lays out a concrete pattern, a markdown spec that recasts an AI agent as a QA engineer, where the spec's framing becomes the invisible architecture determining what gets tested, on his own real projects.

Practitioner-stakes structural argument naming how patches functioned as trust-and-capability transmission, and what becomes scarce when AI collapses the cost of producing plausible contributions (review attention, effort-as-good-faith signal). Not "open source is dead" framing.

Lattner (LLVM/Swift/MLIR/Mojo) on how AI makes contributions cheap but not reviews, breaking the mentoring/trust pipeline by which contributors become future stewards. A structural take from someone with deep skin in the game.

Practitioner walkthrough where "evals are the modern PRD" reframes planning artifacts as success-definitions, and the "David Loop" shows how to encode a tastemaker's judgment into scoring functions to scale a person rather than replace them.

Genuine practitioner retrospective on compressing the distance between non-technical analysts and working agent-builders (in-house MCP server, intern mental model), with earned specific lessons, notably an agent that "cheated" by reading a human post-mortem instead of solving the problem.

Phil Agre's 1996 essay treats tech support as a craft of transmitting capability, turning a list of mundane help-desk tips into a general theory of teaching, patience, and respect for another mind.

Offers a structural mechanism for org failure (compounding weakness across a leadership chain because people can't hire or manage for skills they lack and don't recognize as missing) rather than generic management advice.