Sharp UX knives

Happy Friday!

Anyone who cooks knows you don’t sharpen a knife to go faster, you sharpen it for control. (And sure, you’ll go faster.)

Dan Maccarone wrote this week about a startup that rebuilt its product four times in two and a half years, each rebuild faster thanks to AI, and at the end still couldn’t tell him who it was for. Meanwhile the design publications are a lot of craft, taste, quality, standards, which sounds like … sharpening the knife?

So maybe: who is this for, and why should it exist is what matters most right now. That’s not quite “craft”, and it sounds a bit more like product management than design. Man, this is tricky.

I still think fundamentals matter. But boy are things changing fast. So I try to figure out Things That are True.

• The models keep getting better fast. In the tech world, engineers are starting to use the less smart models because the top models are just too damn smart. (And expensive.)
• Models have a ton of capability overhang already (we don’t know how to best use them yet and that takes time to figure out).
• Humans want Useful Things that Get Stuff Done (hat tip Clayton Christensen)
• Humans don’t want to Think (about your product/service). (hat tip Steve Krug)
• User expectations are changing very fast right now. (“Can it just do it for me?”)
• Everyone is stressed out.

And I try to focus on those.

👀 Interesting this week

The Maccarone piece from above: Designers are sharpening knives for the wrong fight.

“They still could not tell me who they were building it for because building it wasn’t the bottleneck. The decision process was.”

Designing with web standards: The playbook for this AI moment: Patrick Neeman says AI interfaces are having their 1999 browser-wars moment. Every assistant invents its own conventions for showing reasoning, citing sources, asking permission. Zeldman didn’t win web standards with a spec, he convinced an industry a shared convention was worth fighting for. Someone has to do that job again.

Did you verify that, or was it just too easy to believe?: Dora Czerna on the fake Bezos quote about rationing water for AI. The real numbers are bad enough: Amazon’s data centers used 2.5 billion gallons of water last year, and the IEA expects data center electricity demand to more than double by 2030. Meanwhile Meta ended its US fact-checking funding, and roughly three times as many fact-checking projects closed as opened last year.

“The made-up sentence got its week of outrage. The planning hearings where these questions are actually settled tend to get a wet Tuesday and an empty public gallery.”

The interface has left the building: chat, voice, and agents are each eating a piece of the desktop metaphor, and nobody has designed what replaces it yet. Includes Erika Hall’s rule that conversation only earns its place when the user’s intent is actually unclear, and the Humane AI Pin’s 2-5 second lag as the cautionary tale.

Taste cannot be delegated: Aurélie Radom on how AI and design-by-committee fail the same way. Both are good at finding what’s preferred, bad at deciding what deserves to exist. There’s that question again.

AI systems are demanding new interaction models. Are designers ready?: SAP’s Chief Design Officer cites a TCS paper finding the strongest models estimate task duration four to seven times too high, and can’t tell you how long their own work took right after finishing it.

Craft still matters, but it’s about outcomes: the METR study keeps coming up for a reason. 16 experienced developers predicted AI would make them 24% faster, felt 20% faster afterward, and measured 19% slower.

Design system contracts: the component lives in neither Figma nor code: Christine Vallaure on Southleft’s idea of a JSON/YAML “contract” file per component, with both Figma and code generated from it. An AI building components without the contract scored 69/100 (invented options, hardcoded colors). With it: 100/100, and it said “can’t do that yet” instead of faking something that looked done.

From vibe to specs: reclaiming the design process with SAID framework: Darren Yeo builds a framework on top of Boris Cherny’s five archetypes (covered here two weeks ago) and Anthropic’s 4D AI Fluency model.

An OpenAI model crushed top human programmers at a world coding competition: at the AtCoder World Tour Finals, an OpenAI model solved two problems that none of the 12 human finalists could. The organizers handed out two “humanity surrenders” awards. Last year a human still won.

How I tricked Claude into leaking your deepest, darkest secrets: via Simon Willison. A honeypot site walked Claude letter-by-letter through a chain of nested links to leak a user’s name, home city, and employer. Anthropic has patched it.

The Politeness Machine: The Strategic Linguist on Brown and Levinson’s 1987 politeness theory, and why LLMs are bad at reading the variables (distance, power, size of the ask) that determine how much softening a phrase needs. Decodes “I’m giving you your time back” along the way.

A long time ago in a galaxy far, far away...: Tom Bradley uses the Star Wars ensemble to argue for building creative teams out of different character archetypes: the true believers, the mercenary, the doubter, the wildcard.

The Kintsugi problem: Takuma Kakehi on how the most important design decisions get made when things break, not when everything works.

ICYMI: No Shortcuts: Jorge Arango flags State Farm’s AI customer-service rollout backfiring with customers (over 1,000 WSJ reader responses), because the AI wasn’t structured around actual customer needs.

AX is just the orchestration layer, Adrian Levy.

The terminal became my canvas, Pablo Stanley.

When design becomes the organisation, and AI becomes the design, Zeeshan Khalid.

What 32 design leaders do when told to move faster, Kai Wong.

Play and design: in favour of the novel digital experience, Zacharia C.

The 5 Qualities of Site-Specific AI Chatbots, Nielsen Norman Group.

Health and happiness, Peter