Sunny

← All weeks

What I’ve been reading (W38, 2026)

Starbucks spent years automating its stores and then reversed it · The clever part of Apple's folding iPhone Duo isn't the hinge, it's the shape

What stays scarce once AI makes everything cheap, the A4-shaped iPhone Duo, Zuckerberg on safety without coordination, the product decisions you refuse, and why agents can’t clean up their own slop.

  1. Starbucks spent years automating its stores and then reversed it: handwritten notes on cups, ceramic mugs, more baristas per store, because small details and hospitality drive satisfaction. Alex Imas (a behavioural economist) uses this to ask what stays scarce once AI can make most things cheaply, and his answer is the human in the product. When a sector gets automated its prices fall, people get richer in real terms, and they shift spending to things where who made it is part of the value; that shift, not the price drop, explains over 75% of how economies have restructured historically. Farming went from 40% of US jobs to 2% while people ate more than ever; the same move takes workers into care, teaching, hospitality, craft, and anything where provenance matters. People pay about double for an identical good when they know others are excluded from it, and AI involvement halves that premium because anything AI touched feels infinitely reproducible aleximas.substack.com/p/what-will-be-scarce
  2. The clever part of Apple’s folding iPhone Duo isn’t the hinge, it’s the shape. Both screens are roughly 1.4:1, the proportion of A4 paper, so the unfolded inner screen has the same shape as the outer one instead of the awkward square every Android foldable ends up with, and square is why nobody bought them. Apple has clearly built the best possible book-shaped phone; whether a book-shaped phone is a good idea is still open, and nobody at Apple called the Duo the future of iPhone the way they did with the iPhone X. It also has no Face ID, and Apple is pushing “unlock it with your Watch” instead daringfireball.net/2026/09/thoughts_and_observations_on_apples_surprise_and_shine_event
  3. Zuckerberg on why labs shouldn’t agree to slow down: the market already forces safety. Nobody uses an agent that doesn’t do what they ask, and labs are liable when models cause harm, so any lab that doesn’t focus on alignment will fall behind. The one concrete promise is that the significant majority of Meta’s compute goes to serving people rather than racing towards recursive self-improvement. Everything is something a lab can do alone: safety through balance of power between labs, not coordination among them. x.com/finkd/status/2099997096896274533
  4. The most important product decision is what you don’t build. A document hub or a notifications centre (two things this author has refused to build) starts as a list of PDFs or a bell icon with a red dot, and every stakeholder adds one more thing until you’ve rebuilt Google Drive and have to maintain it forever. What brings people back is laying out the second-order costs of running it over years, not the build cost, then handling each use case with something simple that already exists. Deleting is harder than refusing, because nobody gets promoted for removing things. Now that agents make building nearly free, the smarter use is pointing them at the pruning. liamnugent.me/posts/what-you-dont-build
  5. LLMs write code that is almost always correct, which is not the same as good. Correctness is easy to train for because tests give a clear pass or fail; sloppiness needs human taste, which is why asking a model to rate its own code is basically a random number generator earendil.com/posts/measuring-code-sloppiness