The Factory Hasn’t Been Redesigned Yet
The productivity everyone's chasing lives on the other side of a redesign nobody's started.
When factories got electric motors in the 1880s, they swapped out the steam engine and changed nothing else. The floor plan stayed the same, the belt-driven machines stayed the same, the workflow stayed the same. For 30 years, output barely moved. The returns came when companies tore out the floor and redesigned everything around the new technology.
Tommy Geoco channels this history through Carlota Perez’s framework for technological revolutions: “We have swapped the motor, but we have not yet redesigned the factory.” Most teams have installed AI but haven’t changed how the work flows around it.
The numbers make it concrete. Google tells Clive Thompson that its 100,000+ developers work 10 percent faster with AI. Geoco’s studio went from one video a month to eight, though at the cost of running, as he puts it, 50 to 100 cognitive cycles a day, each with the same emotional weight. Jason Lemkin replaced a $10K/year sponsor portal in days. The difference between 10% and 10x comes down to whether you reorganized the work or just plugged a new tool into what you already had.
Most of us are still plugging in. David Oks finishes the famous ATM parable and the ending changes everything. US bank teller employment held steady through the entire ATM era, still at 332,000 as late as 2010, then collapsed to 164,000 by 2022. ATMs didn’t do it. The iPhone did. ATMs automated teller tasks better. The iPhone eliminated the reason to visit a branch at all.
AI tools that generate Figma variants or fill out documentation are automating tasks within the existing workflow. The harder question is whether AI changes the workflow itself, making some of those tasks unnecessary. Staff engineer Sean Goedecke does the math on his own profession. His assessment is that no breakthrough is required. Just incremental improvement on what AI already does.
Madison Utendahl closed her award-winning Brooklyn agency. Ten people, all women, every award possible. She didn’t close it because it failed. She closed it because the model underneath it broke. Lower fees meant more clients to hit the same revenue. More clients meant more pitching, more context-switching, more burnout. Then clients started generating moodboards with Midjourney before sending the brief. The old factory couldn’t absorb the new motor.
And the people inside these factories don’t all have the same options. Brad Frost names the privilege in the “just don’t use it” position: the people who can afford to sit this out tend to have seniority or institutional protection. The designers entering the field don’t. Anthropic’s researchers found that hiring of young workers in AI-exposed roles has quietly slowed, not because AI replaced them, but because companies stopped posting the listings. Companies are starting to figure out what a redesigned factory might look like.
So what does the redesigned factory look like? Julien Bek at Sequoia draws a clean line between intelligence work and judgment work. Intelligence work is rule-based execution AI can already handle. Judgment work is experience, taste, strategic calls. His argument: stop selling the tool and start selling the outcome. Close the books instead of selling QuickBooks. And here’s what makes the model compound: every task the autopilot completes teaches it something the copilot never learns, because the copilot hands that knowledge back to the human. The moat for the next generation of products won’t be the interface or even the model. It’ll be the accumulating dataset of domain-specific decisions.
Thu Do set up Figma MCP + Claude Code and audited her entire design system in 10 minutes. That’s the design version of this: tokens used to be nice-to-have for consistency. Now they’re infrastructure for AI-to-code workflows. The bar shifted from human readability to machine readability. That’s what tearing out the floor looks like for a design team.
I’ve been telling the comforting version of this story for months: every major tool shift expanded the field, more designers exist now than 40 years ago, the pattern holds. Oks convinced me that’s only the first half. The second half is what happens when the work reorganizes and the jobs stop needing to exist in their current form. New roles and new kinds of design work will emerge from the transition. The people in the old factory don’t automatically end up in the new one. And we’ve barely started tearing up the floor.
What I’m Consuming
The Color Statistic That’s Been Wrong for 80 Years. Kevin Muldoon decided to actually count how many colors the human eye can distinguish, instead of repeating the “10 million” figure everyone cites from a 1939 estimate nobody verified. His answer: roughly 273,000 surface colors at lab-detection threshold, and about 9,256 that you’d notice in daily life. The original estimate assumed human color perception was shaped like a box. It’s shaped like a mollusk. Three independent sources across 87 years landed on the same corrected number. The right answer was in the literature all along; it just lost the popularity contest to a bigger, rounder number. (Kevin Muldoon / UX Collective)
Mastering Midjourney: How to Create Consistent, Beautiful Brand Imagery Without Complex Prompts. Claire Vo interviews Jamey Gannon, an AI creative director who specializes in brand imagery. The counterintuitive move in Gannon’s workflow: she relies on style references and image refs rather than elaborate prompts, which produce more consistent results with less effort. The real value is in how she packages and delivers the system to clients so they can keep generating on-brand assets without her. (Claire Vo / Lenny’s Newsletter)
‘A Language We Share’ Traces a Photographic Lineage Between Gordon Parks and Beverly Price. A new exhibition at Brooklyn’s Center for Art and Advocacy puts Gordon Parks and Beverly Price in direct conversation. Parks embedded himself in American life from the 1940s onward, using photography as advocacy. Price, who picked up a camera a decade after her release from incarceration, entered into a dialogue with Parks by documenting the same Anacostia neighborhoods he’d photographed decades earlier. Both focus on children, policing, and the forces that threaten communities. Runs through June 19. (Grace Ebert / Colossal)
How You Source Great Designers. Soleio lays out how he finds design talent for startups: follow what designers are making on X, build referral nodes through a small number of well-connected design leaders, and create opinionated job pages that work while you sleep. His best advice is the simplest—ask candidates which 2-3 companies they’d interview at if they could pick any, and why. That one question tells you whether your startup can compete for them. For those of you who aren’t hiring—these are great tips on being a better candidate. (Soleio)
Designing the Shift. Ileana Marcut built a system with Claude Code to collect anonymous reflections about how AI is changing people’s work and identity. Part research, part interactive art, I love this. In her post about making it, she reveals the most instructive part is where it broke: Claude Code assumed every contributor was a designer (like Marcut) and baked that bias through the entire processing pipeline, from prompts to labels to sentiment analysis. The system was built for anyone in product, design, or creative roles, but Claude Code assumed every contributor was a designer and threaded that assumption through every layer. A PM writing about replacing designers got classified as a designer experiencing identity threat. The gaps you don’t specify are the gaps AI fills with its own assumptions. (Ileana Marcut)
How to Code with AI Agents. Peter Steinberger, creator of OpenClaw, describes the arc of working with AI coding agents: from short prompts to overengineered slash-command systems to short prompts again, but with hard-won intuition underneath. His sharpest observation is that experienced engineers sometimes struggle more than beginners because their expertise becomes a burden. They try to force their approach on the agent instead of letting it find its own path. Steinberger designs his codebase for agents to navigate, not for himself to read, and commits straight to main with local tests. (Lex Fridman)