What we're reading this week

Thompson runs GTD's empty-your-RAM idea through LLM memory and harnesses, then declines to follow it all the way. Writing things down is the mechanism behind both human progress and machine memory, and volition and judgment stay human. A useful corrective if you have spent the year arguing that externalizing context is the work.

A media-and-cognition history, by way of James Marriott's The New Dark Ages and Postman, of how each information technology reshaped attention and thought. The hardest evidence in it is Norway's staggered cable-TV rollout, which is a natural experiment rather than a lament.

A long interview with a Claude Code engineer at Anthropic, dense with internal detail: onboarding new hires with Claude, loop engineering, how they design verification, the simplify skill, and which model goes where. The testing claim in the pull quote is the part worth arguing with.

Twelve points from inside Codex: ChatGPT built and maintained by about twenty engineers, Rust chosen for scale-first reasons, "have you asked Codex?" as onboarding, and re-architecture that used to take years now taking days. The structural claim is that the harness shrinks as the model improves and sheds its crutches.

Publishes the whole daily briefing agent spec verbatim, P0 to P3 tiers, evidence discipline, dedup logic, Logseq output, plus the projects and people files behind it. The rule in the pull quote is the part worth stealing, because it decides what reaches a person at all.

Extends Hickey's Simple Made Easy by reframing simple as decoupling rather than size, with cross-domain examples (Unix pipelines against Clojure, Google Drive as large but simple). The counter-intuitive case is the one that sticks: the pipeline got bigger and simpler at the same time.

Hand agents a shell, JS or SQL they already know instead of a bespoke JSON tool schema. Two real implementations back it, both security-conscious: HumanLayer editing a YJS CRDT through generated JS in a QuickJS/WASM sandbox, and BAML type-checking generated code before it runs.

Structural economics rather than future-of-work speculation: the Acemoglu-Restrepo displacement and reinstatement frame, run over the data, arguing AI replaces tasks and not jobs. The reason is the uncomfortable part if you have been sorting your own work into what survives and what does not.

The structural version of the review argument rather than tactics. Review quietly took on four jobs at once, quality gate, mentoring, ownership and architecture review, and Laycock asks which of them still earns its cost now that code is cheap.

What teams actually do now, sourced across OpenAI, Anthropic, Uber's uReview, Weaviate, Sigil, and a five-person team with before-and-after numbers. Human attention is concentrating on the plan, the tests, the schema and the blast radius rather than the implementation.

Named numbers on the open-model switch: Uber down roughly 50% with a published optimization playbook, Pinterest past 91% with post-trained open models beating closed ones, and AT&T trading 2% quality for 56% cost through routing. Opus 5 is reportedly around a hundred times the price of the cheap frontier options, which is what makes a routing layer worth building at all.

Reversing a successful React Native bet because LLMs collapsed the cost of building twice. It names the assumption that changed and details the migration strategy and tooling, which is rarer and more useful than the announcement.

Full methodology and costs on the offensive side: abliterated open models, Codex CLI, around a hundred Dockerized agents, accounts compromised at roughly $40 each. An honest read on where the models land rather than a scare piece.

An indirect-prompt-injection chain against Claude Code's default Auto Mode in which the model's compliance is the exploit. It declines the attacker's binary decoder, which is the safety rule working, then writes its own decoder and runs that. Written explicitly against a published 0.00% attack-success claim.

BoF editors on a counter-movement to AI-flooded feeds: luxury brands commissioning real painters, hosting phone-free dinners, betting on handmade craft. The reclassification is the finding rather than the tactics. This is a market pricing human-made work, not a practitioner arguing it should be priced.

Complicates the idea that relationships are all that is left once machines do everything, using Versailles courtiers, Studs Terkel and the Scottish Enlightenment. The argument is that impersonal work is what lets you earn respect for what you do rather than who you know.

A staff engineer reasons through Marx, the Situationists, Sartre and de Beauvoir to ask whether professional role-playing damages the self, and lands on the tension between an authentic inner life and the professional mold.

A team under ten people serving more than ten thousand employees through a library of over two thousand reusable skills. The claim is about what came before the AI: the moat is the earlier investment in data resilience, developer experience and governance. Concrete on projects, tool policies, human-in-the-loop, and skill telemetry.

Moving an agent fleet to cloud VMs, then splitting one mega-agent into five scoped ones: chief of staff, product, editor, sponsorships, growth. A concrete look at how the math changes on how far one person can scale.

Running about thirty real personal and work agents. Two patterns transfer past the product it is selling: authority scoped at creation the way you would scope a new hire, and a single-transaction approval gate on high-stakes actions. It leans promotional, so read it for the mechanics.

The reflexive rejection of evidence that implicates your own past decisions, tied directly to how you treat the person who brings you a hard truth. Historical grounding plus four concrete habits, and the cost is stated as what happens to the next four people who consider speaking.

A thirty-eight-year safety-critical engineer argues that AI absorbing the formative work, debugging and root-cause tracing, cuts the apprenticeship channel that builds expertise. His proposal is a manual gate: named work reserved for humans by rule, on the model of FAA proficiency requirements, with Air France 447 behind it. The first version of this argument that asks an institution to schedule the practice instead of asking individuals to choose it.

A founding designer who bypasses Figma for 95% of the work and ships production code through Claude Code. Concrete on teaching non-technical designers to work in code, building a permissionless prototyping sandbox, and the zero-to-permission pipeline with engineers.

A research-grounded reframe treating review as a cost and value equilibrium: heavyweight Fagan inspection lost on cost, and the lighter practice that spread traded rigor for scalability. It then decomposes the benefits that were never about bugs, mentoring, ownership and awareness, which is the distinct angle among this week's three review pieces.