What we're reading this week

L.M. Sacasas argues AI is an environment rather than a neutral instrument, complicating the "just use it wisely" consensus.

A novelist/translator reframes AI as the latest chapter in externalizing cognition; the real danger is humans being trained to behave like machines.

Charity Majors openly changes her mind: code is a disposable cache of understanding, while the durable product is shared, encoded knowledge. (Also touches Transmission of Capability.)

Practitioner heuristics for rejecting working code precisely when output outpaces understanding; the bottleneck is the human's consolidated grasp, not the model.

Anil Seth's rigorous cross-domain argument (neuroscience, philosophy, thermodynamics) that consciousness is rooted in biology, complicating the "machine consciousness is when-not-if" consensus.

A warts-and-all retrospective where the author offloaded the typing but kept ownership of the verification harness (oracle parser + property-based fuzzing); a sharp case study in what understanding stays load-bearing.

A genuine engineering retrospective (Bayer's PRINCE/Agentic RAG) naming concrete decisions: hybrid-search weighting, reranking, removing an LLM SQL-reviewer that false-flagged valid queries, splitting one agent into three reflection loops.

The case that learning hard things builds the judgment AI can't supply; effort and tradeoff-weighing become the differentiators, and high agency beats raw coding skill in hiring.

Sarah Guo builds a falsifiable "absorption frontier": anything measurable gets benchmarked and commoditized, so value slides to untrainable ground truth, permission, and accountability. The sharpest single frame of the week.

From our own desk: Britton reads Anthropic's 400,000-session analysis and finds non-software professionals succeed nearly as often as engineers (26% vs 30%), because knowing what to build matters more than knowing how. The receipt for this week's value-concentration through-line, with a three-question test for what's safe to vibe-code.

Gergely Orosz's reported synthesis of the Meta account-takeover outage (AI-generated, AI-reviewed code) plus data on how much ships without human review and how Anthropic/OpenAI/Uber changed their workflows.

A data-grounded companion to the review-bottleneck story: a Faros report (22K+ devs) showing starting work is now easy and shipping is hard: volume up, quality down, even for top teams.

Kent Beck revisits his own principle in light of near-zero code cost, reframing it from thrift to two pieces of price theory (optionality + NPV) that survive cheap generation.

Brandur builds an actual cost model for when LLM-collapsed build costs flip buy-vs-build: concrete math, not "SaaS is dead" hype.

An executive coach to the OpenAI research team argues emotional clarity and discernment are the scarce skills when knowledge and effort are nearly free.

A Mozilla Distinguished Engineer's retrospective on three agentic workflows, making the structural case that the harness matters more than the model, with genuine warts.

Opens on the value-concentration thesis (compete on quality/UX because vibe-coding makes building cheap) then backs it with a transferable craft pattern any instant-search builder could lift.

A first-person ML-engineer retrospective that explicitly updates her view, with concrete hardware, a model-by-model progression, and a sandboxed Docker+Pi+LM Studio agentic setup.

A practitioner retrospective on building a Claude Code skill that brings Cursor-style click-to-inspect into the terminal: actual prompts, v1 bugs, and an explicit token-ROI call to stop short of auto-running (the "build exactly the tool you want" angle).

Ted Gioia treats authenticity as a hard, slow, transformative quality and names it as what becomes scarce when AI makes output cheap: argument, not nostalgia.

A cross-domain structural analysis mapping Pasolini's "false permissiveness" onto GenAI and alignment-as-conformism, framing art as process vs. commodity.

A short reflection on why writing-as-thinking works: the value of voicing what feels too obvious to say. The "audience of one" / "write to remember how to think" instinct, not generic blogging advice.

From our own desk: how we rebuilt this site, one engineer plus Claude, replacing the CMS with MDX in version control and emitting clean JSON-LD so agents can read our structure, not our marketing prose. The small-team-big-output story behind the page you are on, and a working example of owning your content infrastructure instead of renting it.

A practitioner interview on running an AI-native org shipping 8x more code, how Claude routines reshape a manager's operation, and which roles AI transforms next.

Will Larson updates leadership rules for the agent era with concrete data (200-400 deploys/week, individuals owning 95% of migrations), arguing judgment and durable teams become the moat.

A retrospective on solo-maintaining a 10K-star open-source app part-time with AI agents: validation as the bottleneck and "releasing more control than you're comfortable with." (Intro free, depth paywalled.)

A Meta VP describes a concrete transformation: the PRD is now a paragraph + prototype + eval set, an agent replaces the org-chart "compression algorithm," AI captains are insiders only. Dense across four curiosities, warts and all.

A structural taxonomy (in/on/off the loop) that operationalizes "design the harness, don't micromanage the output," escalating to agents recommending improvements to their own harness. Pairs with Loop Engineering below.

Addy Osmani specifies the "design the loop instead of prompting" pattern with five named building blocks mapped onto actual Claude Code/Codex primitives plus a worked example.

Argues planning quietly tangles product and technical decisions; splitting design into product/technical/program steps and front-loading every decision is where the leverage lives.

Jason Cohen's structural teardown of why RICE's "confidence" term is incoherent and what to replace it with: how a planning formula skews decisions and which judgment the tool can't express.

James Stanier analyzes how the junior-to-senior pipeline actually worked (low-stakes failure, scar tissue) and what breaks when AI absorbs the tasks where that learning happened.

Names which parts of struggle actually drive growth vs. busywork, grounded in desirable-difficulties research: what's lost when AI removes the friction where learning used to happen.

A retrospective on copying-as-discipline (Virgil Abloh's 3% approach) as the mechanism by which taste and instinct develop; "side quests became the main quest" lands as real insight.

Rands on writing as a lifelong habit, the power of naming things, stories as a feedback tool, and the return of the coding manager in the AI era. (Summary paywalled; podcast free.)