Tech internship postings dropped 30% since 2023, according to Handshake.[1] Meanwhile, applications surged 7%. For those who do land roles, the outlook is grim: 37% of companies terminated new Gen Z hires within their first year in 2024.[1:1]
Gergely Orosz, author of The Pragmatic Engineer newsletter, says the job market has not been this tough for new grads and junior engineers in the last 20 years, ever since the dot-com bust.[2] On Indeed, there are 35% fewer software developer job listings today than five years ago—a 3.5x drop from the mid-2022 peak.[3]
The bootcamp promise is collapsing. In 2021, Kim enrolled in Codesmith expecting a six-figure engineering job. By graduation in 2023, only 37% of graduates landed full-time roles within six months—down from 83% two years prior.[4] "Coding bootcamps were already declining," said Allison Baum Gates, early General Assembly employee. "But AI has been the final blow."
The diagnosis is simple: we built factories and now we need workshops.
Karri Saarinen, Linear's CEO, said it bluntly at Stripe Sessions 2024:
"We focused on building larger teams. We then made them run like factories churning out new things at a consistent pace. [...] Software today borderline works. But we are supposed to be professionals."
The factory model made sense when scale was the bottleneck. More hands meant more features. More features meant more users. More users meant more revenue. But AI changes the equation. Scale is no longer the constraint—AI handles scale. A single developer with Claude ships what took a team of ten.
"Personal taste is a moat" is becoming consensus.[5] Forbes called it "Silicon Valley's new competitive moat."[6] Paul Graham laid the philosophical foundation 20 years ago: "When you're a doer, taste becomes a practical matter—knowing what looks good isn't enough, you have to know what can be made to look good."[7] Karri Saarinen (Linear): "Building quality requires the skill and taste to recognize it."[8] Brian Chesky built Airbnb on the principle that "design is not a department—it's a way of thinking about the world."[9]
But here's what no one explains: how do you build taste?
Graham says "practical knowledge." Chesky says "learn from people with better taste than you." Saarinen says "hire for it and instill the belief." They're all describing the same thing: apprenticeship. Taste isn't a Udemy course. It's osmosis—sitting next to someone better, watching a thousand decisions get made, absorbing tradeoffs until they become intuition.
When any idea can be vibe-coded, the future of tech jobs is mastery, not hustle. How do we achieve it?
Mastery is the new bottleneck: the judgment to know what to build, the taste to know when it's right, the understanding to fix what breaks. AI amplifies craftspeople; it doesn't create them. Industrialized food didn't eliminate restaurant chefs—it made the good ones more valuable. The same pattern is coming for software. The craftspeople who understand why, who have taste, who can teach—they'll be the ones left standing.
Here's the chain: to solve complex problems, you need mastery. To have mastery, you need masters. To have masters, you need trained juniors. Factories don't train—they optimize for output.
The pipeline isn't slowing. It's breaking. Without internships, juniors can't get experience. Without junior roles, they can't build careers. Without junior hires today, companies won't have seniors in five years. The senior engineers everyone wants to hire didn't appear fully formed. They were juniors once—sitting next to someone better, absorbing tradeoffs, watching decisions get made. Mastery was built layer by layer: a thousand bugs debugged, a hundred bad decisions witnessed, years of osmosis before independent judgment.
Cut the pipeline and the cycle breaks. Today's "senior-only" teams are borrowing from a generation trained elsewhere. Who trains the next one?
Every culture with extreme craftsmanship discovered the answer: apprenticeship in small workshops where masters work alongside learners. The French call them ateliers. The Japanese, shokunin workshops. The Italians, bottega. Production is pedagogy.
Tech has startups and corporations. Neither fits the pattern.
The missing archetype
Every civilization that mastered a craft developed organizations to transmit it:
| Culture | Archetype | Key Trait |
|---|---|---|
| France | Atelier / Maison | Luxury, taste, heritage |
| Germany | Mittelstand / Meister | Precision, longevity |
| Japan | Shokunin (職人) | Lifetime devotion |
| Italy | Bottega | Artisanal workshop |
| Korea | Jang-in (匠人) | State-preserved mastery |
| India | Guru-Shishya | Sacred teacher-disciple bond |
| Tech | ??? | Scale, speed, disruption |
Tech has no archetype for small excellence. Startups optimize for growth. Corporations optimize for scale. Neither creates conditions for craft transmission.
Paul Graham wrote in "Made in USA" (2003): "The Japanese are obsessed with making things well."[10] But his prescription—work fast, ship fast, iterate—was for an era when basic patterns were still being discovered.
Twenty years later, those patterns are understood. The challenge has shifted from "make it work" to "make it excellent." Excellence doesn't reward speed—it rewards the meticulous craftsperson. (See Software Meal Tier System: Michelin Star vs Fast Food.)
The shokunin mindset has organizational prerequisites. You can't develop it in a factory.
The historical blueprint
Ateliers aren't new. The pattern repeats across centuries and cultures: production as pedagogy.
Hermès trains 450 apprentices over four years—one artisan, one bag, twenty hours of work. Recruitment selects florists, jewellers, butchers. "Open to all backgrounds—what matters: manual aptitude, personality, motivation." Teachability over credentials.
Escoffier's brigade industrializes formation without industrializing the result. Commis III → II → I. Each rank designed to graduate to the next. A two-star restaurant runs on fifteen people. No permanent junior caste.
Florence's bottegas trained Michelangelo at thirteen, Leonardo under Verrocchio, Raphael under Perugino. The bottega wasn't a school separate from production—production was the school. Sixty workshops juggled commissions from church, guild, and patron. Variety shaped mastery.
What this looks like in tech
The pattern exists in software—but barely.
37signals: the archetypal atelier
The closest thing to a true Tech Atelier:
- Small by design: ~80 people, never raised VC, profitable since 1999
- Multi-product: Basecamp (v1-v4), HEY, Writebook, Campfire
- Transparent tech: Rails monolith, open-source philosophy, public learnings
- Apprenticeship: Juniors pair with DHH and Jason from day one, graduate to leading features
- Long-term: No exit planned, optimizing for decades
- Teams of two: one designer + one programmer, no feature longer than six weeks
What doesn't work
- FAANG "mentorship": 1 hour/week with an overloaded senior. No osmosis, just check-ins.
- Bootcamp → agency pipeline: Ticket factory. Ship features, learn nothing about systems.
- Startup "learn on the job": Chaos without structure. Sink or swim isn't apprenticeship.
The difference: ateliers design for transmission. Factories hope it happens accidentally.
Defining the Tech Atelier
A Tech Atelier is a small, taste-driven software company where:
- Excellence over growth: craft quality is the metric because it creates value, not user counts or revenue multiples
- Apprenticeship by proximity: masters work alongside learners—knowledge transfers through osmosis, not documentation
- Long-term orientation: building for decades, not exits—decisions optimize for sustainability
- Multi-product portfolio: varied work proves consistent craft and trains across different constraints
- Transparent Technologies: you can only learn from code you can read—a codebase juniors can understand, debug, and eventually teach
How does this compare to what most people experience?
| Dimension | Startup | FAANG | Atelier |
|---|---|---|---|
| Size | 10-100+ | 1,000-100,000+ | 2-25 |
| Growth goal | 10x/year | Market dominance | Sustainable excellence |
| Junior trajectory | "Move fast, learn later" | Onboarding → specialized silo | Apprentice → Journeyman → Master |
| Knowledge transfer | Docs (if lucky) | Internal wiki / Slack | Master-apprentice proximity |
| Exit horizon | 5-7 years (IPO/acquisition) | Infinite (public) | Decades (no exit) |
| Code legibility | Tech debt | Microservices labyrinth | Monolith juniors can read |
The only clear example: 37signals. (See above)
Atelier-adjacent: Ghost transmits via open-source. Linear has craft but has VC pressure. Levels.io has output but no transmission.[11] None are true ateliers—but working in any would develop more craft than years at FAANG.
A Tech Atelier is not: a design agency (projects, not products), a startup (growth over craft), a startup studio (ships products but no elevation of craft), a consultancy (time, not products), or a solo operation (no transmission).
This relates well to Digital-Mittelstand with a small difference: refinement. Mittelstand = industrial framing. Atelier = luxury framing. An Atelier is a Mittelstand with taste.
How do you tell the difference between a tech atelier and a startup studio? Look at the output. Is it a chef-d'œuvre in the making?
The chef-d'œuvre
A Middle Age compagnon building cathedrals proved mastery through a masterpiece. What's the modern tech equivalent?
Not a weekend MVP. Not a single product. Historically, no atelier was mono-product—the bottega juggled commissions, bespoke tailors made suits and coats and shirts. In the AI era, this matters more: a single product can be cloned in weeks. A portfolio proves consistent craft across constraints.
The portfolio IS the chef-d'œuvre. A body of work showing durable taste, repeatable craft, and the ability to teach. 37signals is the clearest example: Basecamp, HEY, new products announced. Teams of two, no feature longer than six weeks. Same philosophy, multiple "commissions" over time.
Why ateliers need juniors
The objection: AI writes boilerplate, seniors review. Small teams ship faster without training overhead. Why would an atelier need juniors?
The objection misunderstands what makes ateliers work.
Craft requires humans with taste at every level. A Michelin kitchen doesn't run on celebrity chefs alone. The commis knows when the sauce is right. The line cook spots when plating is off. That judgment, that taste, is what AI cannot replicate. (See Software Meal Tier System.)
Masters who never train lose their edge. Teaching forces articulation. Explaining why clarifies your own thinking. Hermès artisans become Maîtres by training the next generation—not by hoarding technique.
A pipeline for mastery means company longevity. The atelier that trains juniors survives across generations. Without that pipeline, you're a one-generation operation, hiring ever-more-expensive seniors until the pool runs dry.
Without juniors, you can't test if your craft is transmissible. A master who never teaches doesn't know if their knowledge is tacit (locked in their head) or codified (transferable). If your senior can't explain why you chose Postgres over MongoDB to a junior, they don't understand it—they just followed a pattern.
Ateliers are a place where juniors become the seniors worth working with that build a future worth living in.
The deeper stakes
This might read as a training problem. But the stakes are higher. The atelier model is about what kind of work we want to do—and whether we'll still understand what we're building.
Bernard Stiegler warned about proletarianization: losing savoir-faire when knowledge gets externalized to machines. AI will rob us of theoretical knowledge too. The bootcamp grad deploys to Vercel but can't explain how HTTP works. They're operators, not craftspeople.
"The hyper-industrial societies that have risen from the ruins of the industrial democracies constitute the third stage of completed proletarianization: in the nineteenth century, we saw the loss of savoir-faire, and in the twentieth the loss of savoir-vivre. In the twenty-first century, we are witnessing the dawn of the age of the loss of savoirs théoriques, of theoretical knowledge..."
— Bernard Stiegler, Nanjing Lectures (2020)[12]
The atelier resists proletarianization. Master-to-apprentice transmission preserves the why, not just the how.
Ivan Illich called this conviviality: when a technology enhances autonomy rather than creating dependence. The test: if the company disappeared, could the apprentice build independently?
- FAANG internal tooling: Non-convivial. Proprietary stack, useless outside.
- Basecamp: Convivial. Learn Rails, take it anywhere.
The atelier creates masters. The factory creates operators.
The path forward
AI will keep compressing the space between "can ship" and "has taste." The gap will widen between those who understand systems and those who consume abstractions.
The answer is not more bootcamps. It's more ateliers.
If you're a senior engineer at a small company:
- Treat the next hire as an apprentice, not a "resource"
- Pair daily. Explain the why. Make them read the whole codebase.
- Design your stack so juniors can understand it—that's the Transparent Technologies test
If you're building a product:
- Consider multi-product from the start
- One SaaS is a bet. Three is a portfolio proving craft.
- Optimize for decades, not exits
If you're a junior:
- Seek ateliers, not résumé brands
- One year at a place like 37signals teaches more than five at Google
- Look for: small teams, open codebases, founders who still code
The archetype exists. The philosophy exists. What's missing is recognition—and the willingness to build this way.
No juniors, no masters. We need tech ateliers.
Stack Overflow Blog, "AI vs Gen Z" (2025). https://stackoverflow.blog/2025/09/10/ai-vs-gen-z/ ↩︎ ↩︎
Gergely Orosz, Shift Mag interview (2025). https://shiftmag.dev/the-pragmatic-engineers-career-advice-for-tough-times-4357/ ↩︎
Gergely Orosz, "Software Engineer Jobs at Five Year Low" (2025). https://blog.pragmaticengineer.com/software-engineer-jobs-five-year-low/ ↩︎
Reuters, "Bootcamp Bust: How AI is Upending Software Development" (2025). https://www.reuters.com/lifestyle/bootcamp-bust-how-ai-is-upending-software-development-industry-2025-08-09/ ↩︎
Cong Wang, "Personal Taste is the Moat" (2026). https://wangcong.org/2026-01-13-personal-taste-is-the-moat.html ↩︎
Josipa Majic, Forbes, "Why Taste is Silicon Valley's New Competitive Moat" (2025). https://www.forbes.com/sites/josipamajic/2025/10/17/the-democratization-of-market-research-why-taste-is-silicon-valleys-new-competitive-moat/ ↩︎
Paul Graham, "Taste for Makers" (2002). https://paulgraham.com/taste.html ↩︎
Karri Saarinen, "Why is quality so rare?" (2024). https://linear.app/now/why-is-quality-so-rare ↩︎
Brian Chesky, Lenny's Podcast interview. ↩︎
Paul Graham, "Made in USA" (2003). http://www.paulgraham.com/usa.html ↩︎
Pieter Levels. https://levels.io/ ↩︎
Bernard Stiegler, Nanjing Lectures (2020). https://www.academia.edu/44537529/Nanjing_Lectures_Bernard_Stiegler_2020_ ↩︎