What we're reading this week

Ethan Mollick on the human role shifting from steering to commissioning as a frontier agentic model absorbs hundreds of judgment calls you never get a vote on, and the counterintuitive claim we may need more coders, not fewer. Lands directly on the unifying question.

Jem Gold distinguishes "object-language" (tokens, radii) from "encounter-language" (the felt experience), arguing vibe becomes a programmable semantic layer above the design system as diffusion models parse sensory prompts directly. Touches cognition, translation, and planning artifacts at once.

A non-developer's retrospective on shipping a real product: pass/fail rubrics that define "done" before work starts, lean CLAUDE.md files where every rule is "a scar," and the argument that understanding (not the config) is the asset. Also hits agentic-dev patterns and planning artifacts.

A practitioner with deep skin in the game on taste and judgment when execution is cheap, framing cognitive surrender to AI as the biggest risk facing product builders. (Podcast, transcript behind video.)

Benedict Evans argues scoring AI job exposure is impossible in principle: back-testing fails (a century of accounting automation grew accountants), the business can change under you even if the job doesn't, and you can't fully describe a job. Structural analysis that complicates the consensus.

Murat Demirbas translates an MIT/Wharton paper into Amdahl's Law terms: task-level code gains (+228% to +741% lines) decay to +10–20% shipped releases because human review is a strong complement. The parallelizable fraction stays pinned at ~35%, capping speedup at ~1.53x. The deepest structural take on the AI-productivity gap.

Deep structural analysis (Milgrom-Roberts complementarity, Aoki's J-firm vs H-firm) of why organizational practices come in self-reinforcing bundles that resist piecemeal change: deep process knowledge as the scarce, hard-to-replicate moat. Cross-domain and timeless.

Ben Thompson on the Apple vs. Microsoft strategic divergence (device-centric Siri AI vs. cloud-agent Project Solara), anchored by a non-consensus framing of where AI value concentrates.

Research-backed: devs spend only 14% of time coding; typical org sees 7.8% throughput gain; one study found AI increased implementation time 18% for experienced OSS devs; 80% use AI tools but only 29% trust accuracy. The adoption-gap curiosity, with data.

The disconnect between impressive AI usage metrics and the inability to draw a line to shipped value, with real org data (Uber's CTO blew through the 2026 AI budget by March; DoorDash's justify-and-share token limits). Pairs with the DX and Amdahl pieces.

Jake Albaugh argues the bottleneck in the AI era is intentionality, not execution: design is deciding what should exist before building. "100 times zero is zero." Self-demonstrating, written by Claude from his Slack, but the argument was his.

Nicole Ruiz documents a real build end-to-end: a Claude Project codifying an "invisible checklist" of purchasing criteria into reusable instructions, plus Cowork automating returns. A tool for an audience of one, with the structural insight that AI levels the field for heritage brands whose terrible websites lose to ad-heavy DTC.

Coins a term for pumping disproportionate optimization into a layered activity to extract the purest dopamine hit while destroying what made it valuable. The strawberry metaphor (synthesizing the one aromatic compound, erasing 500 unique experiences) is a beautifully built cross-domain case against thin optimization.

A war story borrowing the political tactic "flooding the zone" to describe how pure output volume collapses a team's ability to evaluate quality. Discernment as the scarce resource.

Jesse Skinner extends the "rockstar developer" metaphor to AI agents: every new chat risks adding a rockstar, producing a codebase "written by hundreds of different rockstars." Maintainability and craftsmanship as the things that can't be outsourced.

Cate Hall personifies "Melvin," the part of us that does the normal-shaped action to avoid blame rather than solve the problem; mediocre effort is often rational because of covert conflicting goals.

Practitioner guide on Claude Code at scale: agentic search vs RAG staleness, "the harness matters as much as the model," self-improving hooks, and the emerging "agent manager" role. Squarely in the agentic-dev-patterns interest.

Sean Goedecke argues engineers should run at ~80% utilization because performance is dominated by time-dependent outlier events you can only seize if you're not already busy. The invisible leverage of staying loose.

Kent Beck reframes XP practices as a "trust factory" and finds a self-reinforcing meta-pattern: every practice that creates trust also encourages trustworthiness. Applies it to AI "single player" dev, where the genie cares about prompts, not purposes.

A crisp etiquette principle for the AI era: clearly label AI output and add your own effort before forwarding it to a human. Captures a real emerging dynamic around attention as a scarce, shared resource.

A Jane Street designer documents shifting from Figma mockups to building real prototypes in the production codebase ("prototypes are living proposal docs, the code is disposable"). An LLM skeptic changing his mind, with an honest fear about being stuck in an iterative rather than creative mindset.

Cate Huston on the underrated leadership skill of incorporating new information and making a new plan, and catching herself doing the very thing she judged others for. Hits the "writers who change their mind" signal directly.

Research (30 devs observed, 131 surveyed) on how GenAI rewires team interaction: routine Q&A migrates to AI (51%; 62% find it easier without embarrassment) while human conversation concentrates on context and judgment. The casual teaching moments senior engineers created vanish with the routine questions.

How Walmart builds AI fluency across a 1.6M-person workforce via OpenAI/Google certificate programs, with real associate-built tools (a cake-decorating coaching app, a load-matching app to get drivers home), framed around keeping human judgment central while scaling capability per individual.

Rands on a specific coaching move: when direct feedback fails, slide up next to the discomfort with a personal story of your own struggle rather than lecturing. Sharp on why the hardest feedback contradicts the very skills that got someone the role.

Nikhyl Singhal works through five leadership questions, with the through-line that the playbook that got you to a leadership seat is the wrong one for the hardest questions from it. The "two superpowers collide, you get two shadows" diagnosis is genuinely sharp.