#501 – DHH: Future of Programming, AI, Agentic Engineering, Vibe Coding & Linux

David Heinemeyer Hansson, DHH, creator of Rails, CTO of 37signals, race-car driver, and now behind Umachi Linux: a year ago you were wary of AI coding; now you’re lit up. Thrilled, terrified, or both?

I’m insanely excited. The dread is intellectual; emotionally, this is pure joy. We lived through pre- and post-internet, and now pre- and post-AI—what a ridiculous privilege. My view didn’t reverse; the technology did. Autocomplete helped, chatbots tutored, but late 2025 brought agents that used tools, checked themselves, and produced code astonishingly close to what I wanted.

Then they learned to split work among subagents. I can give them a fuzzy problem instead of a route. It’s early GPS versus today: eventually, for my work, I can trust the system to find and recover from mistakes.

That depends on the domain. In a database-backed web app, AI can write nearly all the implementation; kernel-level, performance-sensitive, or safety-critical systems still need deeper inspection.

Security is the funny counterexample: models are terrifyingly good at chaining small flaws into real exploits, which makes them exceptional defenders. On Umachi Quattro, I wrote none of the shipped code by hand; I checked important model-layer work and mostly let agents handle UI and supporting pieces.

Basecamp was harder. We let designers loose early, got individually plausible changes, and collectively made an architectural mess. A substantial existing system still needs taste and a keeper of coherence—though plenty of human codebases are already slob factories.

Why haven’t Photoshop, Premiere, and the big apps suddenly accelerated? A small group with agents ought to rebuild the parts people actually need.

Implementation is seldom the big-company bottleneck. Meetings, approvals, communication, and not knowing what to make next consume the energy; agents manufacture bad ideas faster, too. These firms are supertankers built for a vanished era, while one determined person can make their essential five percent.

When I moved to Linux, I missed my simple writing app. I asked an agent for a small C++ and Qt replacement, had a usable first pass in twenty minutes, and within two days replaced it with Amarite, which I’ve used for every essay since.

That’s the beautiful one-user-mind version of software. But turning a personal tool public can intimidate people.

Just ask the agent to publish it: repository, documentation, releases, tedious maintenance. Open source is getting a reformation: more people can bring ideas forward, while maintainers retain the right to say no. High standards matter, but agents can be more diligent than the median contributor. On Umachi, I’ve merged more than a thousand pull requests in months while agents filter duplicate and broken submissions.

Do good ideas still begin with people, or can models genuinely surprise you?

I used to think humans supplied every idea. I don’t anymore. Model proposals have humbled me, someone who prides himself on good ideas; calling them parrots is badly behind the capability curve. I’m delirious, sure, but the delusion is acting as if nothing profound changed. Quattro is evidence: a beautiful, opinionated modern Linux alternative built for this era.

I started Umachi watching Le Mans and too many Linux videos. Once agents took over implementation, ambition lost its ceiling: point at a feature and say, “I want that,” and usually get a first version in minutes, not months.

That power can feel drug-like. I’m excited and overwhelmed, tempted to chase ten things rather than finish one.

I have a mission: make the perfect computer. The bandwidth from an idea to software went from dial-up to fiber, but it’s directed. Linux is ideal because agents thrive on command-line tools and editable configuration. An arcane error or config file is wonderfully legible to a model trained on the source.

I’m more voice-heavy: I record ten or twenty minutes of messy design thinking, then models transcribe it with project-aware terminology into a useful prompt. It preserves uncertainty instead of obeying a brittle spec.

I type because I love typing, but I may use voice too narrowly. Old agile wisdom still holds: don’t pretend you know the whole product. Manifest something, use it, notice what’s wrong, and let taste choose among alternatives. I care about beautiful Ruby and architecture, but hand-perfecting every line is becoming cowboy culture: romantic and fun, not the only sensible way to travel.

That can be sad for young programmers who studied the craft and wonder what happens to their jobs.

If all you loved was arranging prescribed mechanical logic, that work is threatened. If you love making things, judgment, priorities, design, and solving problems for people, we need more builders than ever. Don’t forecast two model jumps ahead; lean into what’s possible today. Find other makers, because isolation turns uncertainty into rumination.

I feel a real goodbye to the old world. But the old craft helped create this one: decades of code became training material.

Exactly. We gave birth to it. I became a programmer because I wanted ideas to exist, then fell in love with craft; agents return me to that impatience with a wider canvas. There’s room to grieve briefly, then learn the new physics and make more of the world.

The same question is arriving in film, art, and games: a person with vision may make the sci-fi film that once needed millions.

Great art, like good software, needs cohesive vision, but the barriers are collapsing. I’m optimistic while recognizing disruption can be brutal: fewer people for some tasks means real pain. Yet machines have always taken drudgery, and people invent new purposes.

Even amid acceleration, family, food, history, and stepping away from feeds may be remarkably old.

Love, creation, children, friendship—those aren’t obsolete because models got smarter. Take breaks, don’t let algorithms feed you seventeen chocolate-covered strawberries, and remember people survived enormous transformations before us.

You’d still bet on humanity a thousand years out?

I can barely predict next year, mate, but I’ll take optimism: Mars, stranger physics, whatever comes. And if I’m uploaded into some future simulation, put me in the 1980s—bright colors, optimism, maybe not the orange pants. Thanks for having me.