What we're reading this week

A practitioner retrospective on how delegating cognitive work to agents bred dependency and intellectual passivity (a "productivity ouroboros"), argued from lived examples rather than future-of-work musing. The inversion in the title is the part to keep.

Frames compression as the actual work of understanding: if you can't simplify it, you don't understand it well enough. Then argues against handing that effort to a model, because what you lose is the check rather than the chore.

A retrospective on writing to think, naming concrete mechanisms: arguing a point as a forcing function, and changing your mind mid-draft as the signal that the writing did real work.

Fuses commons economics with distributed cognition: individually rational AI adoption can deplete the shared expertise a field needs to renew itself, and oversight depends on the very expertise AI use erodes. The loop generates no warning, because the warning is the thing being spent.

Jill Lepore traces identity verification across millennia and arrives at reCAPTCHA, which siphoned invisible unpaid cognitive labor from millions of people to train machine vision. A long history of tools quietly extracting attention.

A warts-and-all account of a non-technical consultant rebuilding her service business around three Claude skill files (a pipeline operator, a proposal builder, a "think like Grace" voice guide) plus a homemade Gmail replacement. Evidence that the skill-file discipline is not an engineer-only practice.

A credible practitioner (Honeycomb CTO) visibly changing her mind, with structural takes on reliability and verification as the new bottleneck ("pets to cattle" for code). Her question converts an identity problem into a work item, and her harder claim is that code review is overrated next to deciding what to build.

A clear structural thesis with a real objections section rather than thin trend-riding: implementation is cheap, understanding is the scarce bottleneck, so value funnels to the few whose judgment can be trusted. The sharpest turn is the inversion in the pull quote, where the 10x is measured on the person generating load and not on the people absorbing it.

Maps LLM-driven workflow upheaval onto two rival systems-design paradigms (control and decomposition versus complexity and emergence), grounded in resilience literature (Ashby, Le Coze, Box) with concrete diagnostics for telling which one you are actually operating under.

When agents generate code faster than anyone can read it, verification capacity is what runs out, and quality moves into constraints and back-pressure (mutation tests, complexity gates, CI blocks) rather than per-diff review. The useful distinction is back-pressure: a sensor that reports is a measurement, a sensor that blocks is a control.

Ben Thompson maps today's AI-infrastructure financing (Nvidia's residual-guaranteed vehicle, hyperscaler debt, Google's equity raise) onto the 1873 railroad-bond collapse, arguing where value concentrates when compute commoditizes.

Derives cultural sameness from squared-error loss collapsing to the conditional mean, shows it at four scales with real datasets, and offers a concrete exploration-versus-exploitation fix. The practical read: under-specified input does not produce a random plausible answer, it produces the average one, and every specific you supply narrows the conditional.

A practitioner interview on what stays scarce when AI generates interfaces in seconds: taste, empathy, and point of view. It complicates the consensus along the way, with agents as a new class of user, pushback on the junior-designer-crisis framing, and the "vibe designing" coinage.

Structural cultural analysis (Graeber, Debord, Amabile) arguing that AI breaks the faith which made abstract knowledge work bearable, with the loss falling hardest on "experience-first" workers who value the messy middle.

A substantive deep dive on a "build and own" engineering culture (engineers as CEOs of their projects), a fast onboarding-to-ownership loop, and the structural shift where ultra-low latency has stopped being the moat and models and reliability have taken over. Useful as evidence that ownership scope is the variable, at a firm with no agentic story at all.

A structural reframe of the leader and report relationship (which side of the table you sit on), plus a graded ladder of interventions that protect autonomy. The ladder is the durable half; it gives you something to do other than escalate.

A practitioner retrospective naming "95% done" as the most expensive project state, with concrete landing artifacts to get out of it: a scope-freeze checklist, a pre-mortem, and a single prioritized cut list.

A Principal Engineer's framework for compressing new-to-good, backed by concrete first-hand cases. Knowledge splits into Facts, Processes, and Concepts; only Concepts travel between jobs, which is why an experienced hire ramps fast. The completion signal in the pull quote is observable in a 1:1 without instrumenting anything.

A business-practical, counter-consensus case against the claim that AI makes juniors redundant. It frames the junior and senior boundary as jagged and treats the team as a system, tracing first- and second-order effects on how capability actually grows.

The feature author documents dynamic Claude Code "Workflows" with four reusable patterns (classify-and-act, fan-out-and-synthesize, adversarial verification, loop-until-done) and names the long-context failure modes each one exists to avoid. Every shape is a way of spending several short contexts instead of one long one.