Hello everyone, I'm Tony Bai.
In this AI-dominated, uncertainty-filled year of 2026, the entire software industry seems to be shrouded in a collective anxiety. We discuss every day: when AI can write a week's worth of our code in one minute, how much value do we "human programmers" have left?
Just as everyone was pessimistically predicting that "programmers are about to be devalued," a Silicon Valley titan known for his sharp tongue and minimalist philosophy swam against the tide and threw out a shocking thesis:
"We may have already witnessed the peak of 'average programmer' salaries. But for those top-tier, truly knowledgeable developers, AI is making them more valuable than ever before."
The man who said this is David Heinemeier Hansson (DHH)—the father of the Ruby on Rails framework, and the co-founder and CTO of 37signals (Basecamp & HEY).
Just a few months ago, DHH was one of the staunchest critics of AI programming. He had publicly mocked Copilot as an annoying intern that interrupted his train of thought and generated garbage code.
But in a recent in-depth interview, he staged a dramatic "self-reversal." He not only admitted that he had "completely surrendered," but also described his current workflow as "Agent First on Everything."
What exactly happened behind this shocking 180-degree turn? In this information-packed conversation, DHH not only复盘ed in detail the "aha moment" that led to his "awakening," but also offered profound, even brutal, insights into programmer value in the AI era, team collaboration, and the future of "software craftsmanship."
From "Repugnant" to "Irresistible": DHH's Path to "Awakening"
DHH candidly admitted that during the era of Copilot and early Cursor's "Autocomplete," his loathing for such tools peaked.
"I was incredibly angry. It was always trying to guess what I wanted to write before I had even thought it through. 'Did you want to write this?' 'Did you want to write that?' Shut up! Let me finish my own sentence!"
He even pessimistically believed that the entire industry was heading toward a thoughtless, stupid future driven by the "Tab key," joking that he might have to go to Denmark to grow potatoes.
The turning point came in the winter of 2025. Two key variables completely changed the rules of the game:
Qualitative leap in models: The release of Anthropic's Claude Opus 4.5 model. DHH found that the quality of code generated by this model consistently and steadily shocked him for the first time. The code it produced was, in many cases, something he would be willing to merge himself.
Revolution in interaction paradigms: The emergence of Agent Harnesses represented by Open Code and Claude Code. AI was no longer that annoying "code completion machine," but had become a "digital colleague" with its own terminal, capable of independently using tools (Bash, network).
DHH described that when these two variables combined, he experienced the "second enlightenment" of his career—the last time was in the early 2000s when he first discovered the elegance of the Ruby language.
"I'm no longer the person typing at a keyboard; I feel like I've put on a super mecha suit. I suddenly grew 12 hands and can operate 7 screens simultaneously. My capability as a programmer has been amplified exponentially."
We May Have Already Passed "Peak Programmer Salary"
When asked whether AI will replace programmers, DHH didn't shy away from presenting a harsh view:
We have likely already witnessed the peak of the golden age of "programmers (as an ordinary profession)."
He believes that in the past, programmers were able to command high salaries because they were the "bottleneck resource" for producing software. A product manager with a brilliant idea had to wait in line for expensive programmers to spend weeks implementing it.
But now, the bottleneck is shifting rapidly.
"When a product manager can generate usable code themselves using AI, everything changes. In any company where software development is viewed as a 'cost center' (which is precisely the vast majority of software development scenarios in the world), the pressure for salary cuts and layoffs will be inevitable."
But does this mean all programmers will be eliminated?
Quite the opposite. DHH believes that AI is triggering a dramatic "value polarization."
The collapse of the middle layer: The value of ordinary programmers who only know how to "translate requirements into code" is being infinitely diluted. Because AI does this faster and cheaper.
The soaring value of top talent: Senior engineers with extremely high "Taste," "Aesthetics," and "Architectural Judgment" are having their value amplified 10x or even 100x by AI.
Because they are the ultimate gatekeepers who can judge "whether what AI generates is right or wrong, beautiful or ugly." They have evolved from "manual laborers" to "art directors."
When AI Can Write All Code, What's Left for Us?
In this conversation, DHH repeatedly emphasized one phrase: Aesthetics is truth.
He believes that whether in mathematics, physics, or software engineering, an elegant solution is often the correct one.
"Jobs cared about the wiring inside the Mac's case because he intuitively knew that only those who care about the layout of printed circuit boards will obsess over every pixel of the user interface."
In the AI era, this pursuit of "beauty" is not only not outdated, but has become more important than ever.
Because when you have infinite "computing power (AI)," the only scarce resource is "Taste."
DHH believes that the core competitiveness of future top software engineers will no longer be "knowing how many sorting algorithms," but:
Product sense: A deep understanding of "what we should do and what we shouldn't do."
System design capability: Abstracting vague business requirements into clear, elegant architecture.
Extremely high aesthetic standards: Being able to guide AI to generate code that not only works but is also pleasing to the eye and easy to maintain.
Code implementation is becoming cheap; code "taste" is becoming priceless.
A Day in the Life of a Master: How I Command My AI "Legion"
DHH shared his current "Agent-First" workflow in detail, which is textbook-level:
He uses tmux to create a three-panel layout in his terminal:
On the left is the Neovim editor.
On the top right is Open Code running Google Gemini.
On the bottom right is Claude Code running Claude Opus.
"Almost all my work starts with one of the Agents. I give it a vague instruction, then watch it generate a first draft. Then I throw the draft to another Agent to critique and refactor. I let the two of them 'argue' back and forth. Finally, I jump into Neovim and act as the final 'referee.'"
He shared a case that shocked even himself:
37signals' Linux distribution, Omarchy, had a backlog of 250 unhandled PRs. He spent 90 minutes having Claude help him review 100 of them.
10% were merged directly.
20% Claude thought the idea was right but the implementation was terrible, so it rewrote a version directly.
The majority of the rest were either judged as "not needed" or identified by Claude as "poor implementation with no good approach," and were closed directly.
"This would have been at least a week's worth of work before. More importantly, half of the PRs involved fields I don't understand, where Claude is a smarter, better reviewer than I am."
Explosion of Ambition: The Cost of Exploring an Intuition Has Been Reduced a Thousandfold
In the interview, DHH mentioned an extremely inspiring concept: AI is making "Ambition" cheap.
He gave an example where he had an Agent create a complete, executable plan for a long-shelved requirement (implementing Windows dual-boot for Omarchy) in just a few days. In the past, he wouldn't have even had the willingness to spend 4 hours researching it. Because it was "important but not urgent" and "very troublesome."
"The cost of exploring an intuition has been reduced a thousandfold. We can now take on projects we wouldn't have dared to even think about in the past."
He shared a real case from inside 37signals: an engineer named Jeremy used AI to launch a crazy project called "P1 Optimization." He wanted to optimize the fastest 1% of requests in the system to make them even faster.
In the world of traditional performance optimization, this was simply a waste of time.
But Jeremy, in just a few days, by having the Agent疯狂 analyze and refactor, submitted 12 PRs, and managed to compress the latency of that 1% of requests from 4ms to under 0.5ms, achieving a 10x performance improvement.
When the cost of exploration approaches zero, those edge optimizations that were previously considered "futile efforts" will collectively converge into an overwhelming product advantage.
Conclusion: A Renaissance of "Craftsmanship"
At the end of the interview, DHH expressed his extreme optimism about the future.
He believes that AI hasn't made programming boring; on the contrary, it has allowed him to rediscover the greatest joy since discovering Ruby in the early 2000s.
DHH's "awakening" is not just a tech titan embracing new tools. It's more like a manifesto:
In the AI era, the "Craft" of software engineering hasn't died; it has simply ascended from the micro-level of "carving code" to the macro-level of "shaping taste" and "mastering systems."
AI is ruthlessly eliminating those coders who only know how to "turn screws," but at the same time, it is handing an unprecedentedly powerful weapon to those true "artisans" who love to create and possess extremely high aesthetics and taste.
Are you ready to pick it up?
Source link: https://www.youtube.com/watch?v=JiWgKRgdgpI