What we're reading this week

After 15 years of blogging, Demirbas argues that writing is how you think and that a record of thinking is only its by-product. LLM-written docs now look like evidence of thought without containing any. The five rules at the end are familiar.

Highsmith has written with AI since Parkinson's changed how he processes language. He tests each passage by asking "is this judgment I earned, or fluency I accepted?" and argues the output looks the same whether the writer's capacity is growing or shrinking. The craft-guide half is skimmable.

Hullman argues that an AI-verified answer is not yet knowledge, because knowledge needs people who can answer for a claim and challenge it, and those capacities fade when nobody uses them. It applies philosophy of science to the question of what humans should keep doing.

Tornhill argues most of programming was always comprehension (5% coding and 70% understanding in a 2015 study), so experience now matters most for spotting problems in generated code. His example is an agent whose tests passed on badly built code.

A founder with no Swift or Kotlin built native apps with Codex and runs a 25-person agency on role-based agents. The write-up separates what AI supplied from what he supplied, and is candid about costs ($3-5 per 30-minute voice session), broad agent permissions and his worry about dependence. Better models also made most of his skill library unnecessary.

Eberhardt rebuilt a real app with the skills framework. The agent carried out engineering rituals faithfully (verbose user stories, redundant mock tests) without the judgment that gives them value. He also credits what works.

A solo developer updates his April workflow with evidence from production: an Android port built mostly by Claude, function signatures written ahead of time to steer style, and QA counted in estimates. He admits he hasn't found the right level of agent oversight.

Kelly asks what becomes scarce when every person has an AI genie and answers that difference and collaboration become the value. He traces the shareable unit of knowledge from the book to the page to the app to a bundle of specs and skills.

A structural look at selling the work instead of the tool. Early revenue proves demand but says little about delivery economics, automation gains get competed away under fixed-fee pricing, and accountability and human judgment may be what lets a firm hold its price.

Mollick runs Claude Fable 5.2 through filmmaking, hardware hacking, a business launch, literary scholarship and empirical research. The gaps he finds are in taste and self-assessment: a joyless course site, a 182-page paper where nothing was cut, and a multi-agent world-building project that grew an 87,500-word rule book and needed outside "consultant" agents to untangle it. What is left for the human is noticing when work is too long, repetitive or no fun.

An economic argument, backed by a model and an SSRN paper, that a wage cannot be both the price of an hour and a person's claim on output. In the model, wages fall to 17% of income if machines improve a hundredfold. The author changed his mind on taxing machines after running the numbers.

Linear's CEO argues that when building is nearly free, leadership's capacity to understand and steer becomes the bottleneck. His fix is to state intent up front (bets, priorities, what is out of scope) so people and agents can check their work against it.

Rare hard data from one company. About 80% of PRs include agent code and 43% ship with no human reviewer. The median engineer ships only about 10% more PRs while tenured engineers ship about 70% more, and new hires are told to skip AI for their first month because heavy early use hurt long-term performance.