Week 20 · 2025-05-12 → 2025-05-18 · 14 newsletters

Taste, Tools, and the Quiet Inbox

ai-interface-and-taste · ai-in-the-vertical · engineering-and-company-memory · capital-and-craft

Mid-May 2025, fourteen emails across seven days. The inbox was thin, the way 2025 inboxes still tended to be before the daily-briefing era pulled in everything. No dominant news story. What there was, when you cluster it, was a small but coherent argument about taste, tools, and where AI is actually landing in real workflows. A few sharp design pieces, two solid engineering memoirs, one capital-strategy essay, and the usual scatter of grace notes.

AI Interface and Taste: The Conversational Box and Its Discontents

Two pieces this week wrestled with the same problem from opposite ends. Julie Zhuo at The Looking Glass ran the long version in "Conversational Interfaces: the Good, the Ugly & the Billion-Dollar Opportunity," arguing that the chat box, the very thing that broke consumer app records for ChatGPT, is now where AI design innovation is stuck. Her frame: the conversational text box was instantly usable by the entire world, but that universality is also its ceiling. Three years in, the medium that ushered in the new era of UX is the same medium pinning the next one down.

Gabby Lord at OMGLord ran the consumer-side companion in "Is ChatGPT your therapist?", noting that her designer friends are not worried about AI taking their jobs so much as quietly using it as a 2 AM crisis line. The behavior has seeped past the discourse. The interface that Zhuo finds creatively exhausted is, in practice, doing emotional work the design conversation has not caught up to.

Carly Ayres at Good Graf brought it home with "We're so back, why everyone is suddenly so thirsty for designers," a read on Jony Ive at Stripe Sessions and Airbnb's Summer 2025 skeuomorphic refresh. The tension she names, between taste and traction, is the same tension Zhuo names from inside the chat box. Ive's line about the cable management ("somebody gave a sh*t about me") landed because, as one reply she quotes put it, "we're at peak software without a soul." Designers got reminded of why they got into this, and the market is, as Ayres notes, "less sure."

The take: the three pieces, read together, draw the same shape. The chat box is the most-used and least-loved interface in software right now, and the people building things are starting to notice the gap. The thirst for taste is the thirst for the next interface, not nostalgia for the last one.

AI in the Vertical: Historians and Doctors

Two pieces this week showed AI landing in specific professional workflows, far from the chat-box debate. Mark Humphries at Generative History introduced Archive Studio, the open-source follow-up to Transcription Pearl. The headline number: Gemini 2.5 Pro now gets better than 95% WER on first-iteration transcription of handwritten historical documents, to the point that the correction step his team built last fall may no longer be worth the time. The framing is honest: many of the features are experimental, and the goal is as much to discover what the models cannot do as what they can.

Justin Mares at The Next ran the more provocative version with "AI is already better than doctors." His issue covered Kettle & Fire opening the country's largest bone broth plant after fifteen months of construction, but the throughline is the same one Humphries is on: AI is quietly outperforming the human baseline inside specific verticals where the workflow is well-defined and the cost of error is bounded.

The take: the loud AI conversation in mid-May 2025 was about chatbots and agent demos. The quiet one was about Gemini 2.5 Pro doing 95% WER on cursive and AI clinical reasoning beating GP intake. The verticals are where the real benchmarks moved this week.

Engineering and Company Memory: Rust, Meetings, and Mark's Memo

Three pieces formed an accidental cluster on how technical organizations carry their own history. Jacob Voytko at Client/Server led with "Ubuntu Is Betting Big on sudo's Rust rewrite," the news that Ubuntu 25.10 will ship sudo-rs by default in October. Canonical's stated goal of "oxidising" Ubuntu, replacing the most core programs with memory-safe alternatives, starts at exactly the place where a C-era vulnerability is worst: a tool whose role is to grant root. Voytko's pull of the 2021 Qualys CVE (sudoedit -s '' as a ten-year-lurking root escalation) is the right horror story to anchor the case.

Voytko's second piece, "Malicious compliance by booking an available meeting room," went the company-history route: Larry Page's 2011 meeting memo at Google (decision-maker required, ten-person cap, fifty-minute hours), and the engineer culture that flowered around gaming it. It is the kind of internal-folklore essay the engineering Substacks do better than anyone, and the implicit argument is that organizational rules outlive the conditions that produced them in ways their authors rarely anticipate.

Internal Tech Emails ran the perfect companion: Mark Zuckerberg's April 2022 memo to Chris Cox and Tom Alison, declaring that the Facebook app's "cultural relevance is decreasing quickly" and naming the friend graph as the structural problem. Three years later, reading it as a primary document, the diagnosis is sharper than most of the public commentary at the time. Zuckerberg saw it: friending was out of vogue, the graph was stale, the request was heavyweight, and the follow-graph and pure-algorithmic models had eaten the cultural relevance Facebook used to own.

The take: the three pieces are doing the same thing in different registers. Voytko on sudo is software-level institutional memory; Voytko on Page is process-level; Zuckerberg on FB is strategy-level. The pattern is that the load-bearing artifacts of a technical organization (a C codebase, a meeting culture, a graph structure) are exactly the ones that get hardest to revisit, and the cost of not revisiting them compounds.

Capital and Craft: PE Envy and the Statue's Hair

Two pieces sat on their own but worth surfacing. Matt Brown's Notes ran "Why VC and software have PE envy," the cleanest framing I have read of the structural difference: PE buys the practice and runs the playbook directly, VC funds software that codifies the playbook for any practice to implement. Both start from the same two premises (the average company in a vertical is underrun, and you have the playbook), and both have started reaching for each other's tools. The dental-empire framing is the right cartoon to carry the argument.

Rob Thomas at The Mentor ran the craft-side companion in "Details," opening with the braiding on top of the Statue of Liberty's hair. Bartholdi finished the statue in 1886, a decade before the first airplane, with no expectation that anyone would ever see the top of her head. The argument is the obvious one (taste is disciplined attention to nuance, details set you apart), but the framing is sticky enough to do the work.

The take: Brown is arguing that capital strategy is converging, and Thomas is arguing that the differentiator in a converging world is still the unseen braid. Read together, they are the same essay from two angles.

Grace Notes

Guillermo Flor at Don't Age, launching a new newsletter on the health-longevity-sport-tech intersection, wrote his origin story around two months on Whoop: HRV metrics dropping on bad-sleep weeks, dinners moving earlier, and an honest admission about getting addicted to the metrics paranoia. His other publication, VC Hiring, ran the weekly list of 32 a16z portfolio companies hiring at >$10M raised and 36 sub-ten-person teams backed by top-tier funds. Lingthusiasm ran a long interview with Hanna-Máret Outakoski on Sámi language reclamation, joiks written down as early as the 1500s, and the role of "language showers" in teaching children to read. JJ Chou ran a Taiwanese-language piece on high-income households that legally pay no income tax (307 out of 6.629 million filing households, per his stats).

Three Takeaways from the Week

The chat box is the most-used and least-loved interface in software right now. Zhuo named it from the design side, Lord named it from the consumer side, Ayres named it from the taste side, and all three pieces are pointing at the same gap: the medium that delivered the AI moment is not the medium that will carry the next one. Watch the verticals (Humphries, Mares) for what the next interface actually looks like in practice.

The quietest AI story of the week was Gemini 2.5 Pro getting better than 95% WER on handwritten historical documents, to the point that Mark Humphries' correction tooling from last fall may already be obsolete. That is not a chatbot story or an agent story. It is a "the model is now competitive with a domain expert on a specific bounded task" story, and those are the ones that keep landing while the front-page conversation chases benchmarks.

If you only revisit three pieces from the week, I would suggest Julie Zhuo on conversational interfaces for the cleanest read on where AI design is stuck, Carly Ayres on the thirst for designers for the taste argument that sits underneath it, and Mark Humphries introducing Archive Studio for the quiet vertical-AI proof point that did not make any front pages. The inbox was thin this week. The signal that was there was sharper than the volume suggests.