DHH at Rails World 2026: "Pencils Down" on Handwritten Code and the Case for Total Optimism

Open on YouTube ↗
Overview

DHH opened Rails World 2026 in Austin by saying they have a condition. Some would call it "AI psychosis"; DHH prefers "AI delirium" or "AI euphoria." The keynote asked what AI agents mean for programmers, for 37signals, and for Ruby on Rails. DHH's position was clear throughout. Agents are the most exciting thing computers have been made to do in their career, and writing code by hand has stopped being an economically productive activity. They also acknowledged that the moment is unsettling, because any prediction about where it ends goes out of date "about 20 minutes later."

27 min read
2:19

A Lesson from Portrait Painters

DHH started with history rather than code. In the 18th century, a portrait meant sitting for hours in front of a master painter, who then spent months finishing the work. The example was Joshua Reynolds' 1781 painting of the Ladies Waldegrave. DHH noted that Reynolds had spent months on an earlier version that the client rejected because an ankle was showing, a story they offered as something anyone who has worked with clients would recognize. Portraits like this were for the upper bourgeoisie and royalty. A royal family portrait by Goya about 20 years later showed how slowly the craft had advanced over centuries.

The next example was personal. The Danish painter Laurits Tuxen spent three years on a portrait of the Danish royal family in the 1880s, and it still hangs in Copenhagen. Photography had appeared around 1840 and improved steadily. In 1900, Eastman Kodak introduced the Brownie, which DHH described as the first mass-produced, broadly affordable camera, costing about a dollar by their recollection. Painters like Tuxen realized that depicting reality as precisely as possible was no longer an economically viable skill. Many moved into other forms: Picasso toward Cubism, the Danish Skagen painters toward a kind of impressionism. DHH framed this as an explosion of creativity forced by a change in technology.

Laurits Tuxen, it turns out, is DHH's great-great-grandfather. The woman in pink at the center of one of Tuxen's later paintings, Yvonne Tuxen, is DHH's great-grandmother. DHH showed a photograph of Tuxen from around 1921, shortly before Tuxen's death, and pointed out that the painter had accepted being photographed by the technology that transformed their profession.

Technology Moves in Fits and Starts

DHH then used their own childhood photos to argue that technological change is uneven. A photograph of DHH as a baby, born in 1979, looks much like the 1920s photo of Tuxen, because camera technology changed little in the decades after the Brownie. Color photography arrived in the 1960s but was not common enough, DHH joked, that they were "worth one" at first. The three childhood photos on screen made up roughly 10% of their entire childhood collection. The cameras of the 1980s were not very different from the 1925 Leica I.

A few decades later, photography became frictionless and the number of photos went parabolic. DHH's wife takes casual photos on her evening walks that would have required real commitment to film in the 1980s. DHH cited a figure of about 2 trillion photos taken per year in 2026. It is the same underlying technology, but it has clearly crossed an inflection point. DHH said the parallels to the software industry were intentional.

11:32

November 24, 2025: The Brownie of Our Era

In DHH's telling, the software industry's Brownie arrived on November 24, 2025, with Opus 4.5. It was AI made accessible in a harness that many people could afford, and it let them experience building software alongside a new kind of intelligence for the first time. DHH called this their personal tipping point and predicted that history books will mark that date as the start of the age of agents.

As with the Brownie, many form factors followed quickly. DHH was especially excited when comparable intelligence appeared in open-weight form a few months later. They used Kimi K2.5 in fast mode for a long stretch and loved working with that level of intelligence at 200 tokens per second.

Then came what DHH called a trough of disillusionment, from February through May. New models kept shipping but seemed worse, even like steps backward. DHH recalled people asking what "this Opus 4.6 thing" was and wanting the old model back, and others arguing that the exponential improvement had stalled. DHH remembered an OpenAI release that did not score much better on benchmarks, after which valuations fell and the market decided "it's over."

Unlike the century-long plateau in photography, DHH said this trough lasted only until May, or really June, when Fable 5 and Mythos arrived. That was when their "diagnosis" fully set in. With Opus, the revelation had been that an agent could produce something DHH actually wanted to merge. With the June models, DHH could hand over problems or ideas and see them solved or expanded without directing the implementation at all. The pace continued. According to DHH, GPT-6 Astra came out in September and ended fears that Anthropic would have a monopoly on frontier intelligence. A week later DeepSeek-4-1 Flash showed that frontier intelligence would not stay limited to a few large American tech companies.

16:08

From the 10x Programmer to the 1,000x Programmer

To ask what this means for programmers, DHH went back to the 10x programmer debate. The idea came from a 1968 ACM article that tested programmers on a range of tasks and found differences between the worst and best of 5x to 30x, averaging about 10x. DHH said the industry spent roughly 45 years arguing about whether this was real, and considered the debate settled now.

DHH then argued that it is no longer controversial to say there is a 100x difference between the worst programmer without these tools and the best programmer with them. They pushed further and suggested 1,000x, while granting that this might be more controversial. DHH did not claim to know exactly how much acceleration is possible, and said openly that the implications for company competition and for what it takes to get started are still unclear.

The concrete evidence they offered was their own output. In the past 20 months, DHH said, they had written half as much code as in the previous 21 years combined. They acknowledged that lines of code are a fuzzy measure and that a line of Ruby is worth far more than a line of lower-level languages like Rust or C++. Still, they called the scale of change undeniable.

19:10

"Look at All the Things I'm Not Doing"

DHH then played a clip from a 2005 talk in Brazil, 21 years earlier. In it, a younger DHH demonstrates early Rails and shows how blog in a URL maps directly to a blog controller and index maps to an index template, with no configuration written. The line they kept repeating was "Look at all the things I'm not doing."

DHH called that line the best summary of the current moment and the source of their excitement: "This is the Rails moment." They said they had written essentially no code for about five months, yet had been able to fix every annoyance and build every "frivolous feature" they had ever wanted.

20:47

37signals Goes Pencils Down

A couple of weeks before the talk, DHH said, 37signals formally decided it is done writing code by hand. The company has gone "pencils down" on handwritten code as the normal way of building things. Writing code by hand at 37signals is now an exceptional state, comparable to an error showing up in Sentry. It means something went wrong and the agent could not produce what was needed. Someone might pick up "the old pencil" briefly, but the real fix is to repair "the machine," "the factory," so the agents can do the work next time.

DHH described this as recognizing something already happening. They polled the audience on how many people still write material amounts of code by hand each week, and about five hands went up. DHH noted that the same claim two or three months earlier would have made people think they had "a screw loose."

What Vibe-Coding Basecamp 5 Taught Them

DHH admitted 37signals had tried this before and pulled back. In the spring, while finishing Basecamp 5, a group of designers vibe-coded the final features they wanted. Each designer's pull requests looked reasonable on their own. Taken together, 20 or 30 of them left the architecture looking "like a Swiss cheese." The company concluded the technology wasn't ready and returned to manual review, with programmers doing the work.

DHH now calls that the wrong conclusion, for two reasons. First, DHH believes that if they had waited a little longer, Fable would have arrived and the approach would "probably" have worked as intended. That is DHH's retrospective judgment, not something they tested. Second, DHH now considers only one question serious in software development: how to get the most out of this intelligence explosion. Every other question ranks below it. Basecamp 5 shipped as an agent-accelerated release that still involved a large amount of handwritten code. DHH joked that no one has settled on a name for the way of working that was "just retired 5 minutes ago."

24:22

Rebuilding HEY as Six Native Apps

The new project is the next version of HEY, the company's email service, whose name is still undecided ("Hey Next," "Hey New," "Hey There"). The first decision is that HEY will no longer be a web app.

DHH argued that HEY "never really wanted to be a web app." It was one because web apps were how small teams could stay productive. Until very recently, the idea that a small team could maintain six native applications in their native frameworks was preposterous. DHH said that is why React Native and Hotwire Native exist: tools that make teams more productive in exchange for some loss of fidelity. DHH called HEY a wonderful web app but said it could never match a native application, partly because of latency.

On screen was a demo of six native HEY applications that the team had started about a week earlier. DHH said high-fidelity native apps for a major service used to take a long time and a large team. Now they are possible "because we're not writing a goddamn line of code," though the team is writing many prompts. DHH stressed that the effort had barely begun, but was already sure this is the future for that class of application. They predicted that many apps that are web apps for reasons of convenience, programmer preference, team size, or resources will become native, because the cost of building native has fallen "to damn near zero." As an illustration, DHH showed the first output when they asked for a Windows app, which was not shippable, and then the redirected version that landed 20 minutes later.

DHH noted that 37signals is not the first to reach this conclusion. Shopify recently launched a rewritten native version of its Shop app, replacing a React Native app, built by what DHH described as about six people working for a short time.

28:42

Rethinking the Backend in Rust

The backend was harder for DHH to talk about, because of Rust. DHH said they hate Rust, calling it the ugliest programming language invented in probably the last 40 years and "inhumane" to make humans read or write. But if they never have to look at it, and only enjoy applications that are 30 to 100 times faster, compile to tiny executables, and start in under a millisecond, then "I love Rust!" Agents like Rust, and DHH described the arrangement as an ideal division of labor: DHH says what to do, and the agent writes Rust that DHH never reads.

For HEY, the plan is to take the code that runs the web application and turn it into what HEY really is at its core, a mail server, written in Rust. DHH reported that this backend needs 99% less CPU and 95% less memory, and that the only reason it needs 10 hosts is redundancy. The team's back-of-the-envelope calculation suggests HEY's peak traffic could "probably" be served by a single Raspberry Pi. DHH presented this as an estimate, not a demonstrated result. They credited it to writing as close to the metal as possible while staying portable, and to having an intelligence paid per token do that work.

Working with Agents Asynchronously

How to work with this intelligence day to day is still unsettled, DHH said, and 37signals is trying several approaches, including directing agents from inside Basecamp. What does seem clear to DHH is that agents work best asynchronously. Rather than chatting and watching tokens stream out, you hand an agent a task the way you would hand it to a coworker, then come back to review when something is ready.

DHH said nobody knows exactly how this should work, both because things change constantly and because no one has worked with this kind of intelligence for long. The whole era started less than a year ago, and the ability to assign outcomes and problems rather than narrow tasks is only a few months old.

33:16

Where This Leaves Ruby and Rails

For HEY, DHH said, the answer is native apps on the frontend and Rust on the backend. But many applications don't fit that shape. DHH called the web a wonderful platform, largely because it never asks anyone to install anything, and said a million businesses are built on that. Services with more ephemeral customers who won't install an app remain a strong fit for the web, and DHH believes Rails is well positioned there. Basecamp is an example: it brings in strangers who need to grab a file or collaborate briefly.

DHH argued that Rails' long-standing values suit the age of agents. Convention over configuration leads directly to token efficiency. The focus on the one-person framework matches what agents now enable, which is single developers going much further than before. Where to draw the line between native and web is something DHH said they don't yet know.

DHH also mentioned agent evals that Evil Martians have been running for the Rails Foundation. These evaluate agents implementing real feature cards against reference applications. Agents saturated the first version quickly, reaching about 95% completion rates, so the evals were made harder. DHH expects this pattern to repeat as agents keep improving.

From this DHH drew a warning for people who know a lot about how computers work: don't be overly prescriptive with these systems. A beginner's mindset, even a "dummy approach," works better because it leads to higher-level questions and prompts. DHH said this is one reason they love working in Rust. They know no Rust at all, consider that a feature, and evaluate the Rust code as a black box from the outside, the same way business owners have always evaluated work they commissioned from programmers. What has changed, DHH said, is who holds that responsibility, not the phenomenon itself.

150,000 Lines in a Month, and English as a Programming Language

DHH shared more numbers from their own work. Over the past 21 years, more than half of their work was Ruby code. This year, it has been about 3%. They acknowledged that part of this is because languages like Rust are verbose and unappealing to read, so DHH lets agents produce more code than necessary, which they would never tolerate in their own Ruby.

Over those 21 years, DHH wrote about 30,000 lines of production Ruby per year, enough to build businesses and frameworks. In August, DHH said, they produced 150,000 lines of code in one month, about 60 times their long-term average. DHH again noted that much of it is verbose Rust, but said the order of magnitude is what matters.

This led to a surprise. DHH said they found a programming language they like better than Ruby: English. They called it more expressive than Ruby, though vaguer and less deterministic, and said writing programs in it is deeply satisfying. DHH admitted they did not see this coming even last year.

38:12

Retiring from Handwritten Code, with Joy

DHH then played a clip of themselves saying they could have become a project manager 20 years ago if outcomes were all they cared about. They started programming because they wanted outcomes, then fell in love with programming itself, and said they would rather retire than give it up.

"Okay. I have retired from being a professional programmer," DHH told the audience. They haven't pinned down the date but think it was four or five months ago, possibly March. DHH said the clip was true: they spent "a quarter of a damn century" writing code by hand and loved every moment.

DHH argued the past should be seen with joy, not regret. Handwritten code brought excitement, learning, satisfaction, and flow, and people were lucky to experience it, but it is over. DHH claimed that writing code by hand is no longer economically productive for most programmers at most companies today, and predicted that by the end of the year this will be true for "virtually all domains, virtually all programmers, virtually all companies." On the other side, DHH said, is a new career as a professional maker of things who steers intelligence that recently existed only in science fiction.

DHH called this the biggest event in the history of computing. They said they sometimes wish they had experienced punch cards, and that they did experience the arrival of the internet, which "smells and feels a little like this moment," though they consider this much bigger.

43:10

Rethinking Software Architecture and Methodology

DHH argued that much of what the industry believes about software architecture needs revisiting. The main tool for structuring code has long been abstraction, including naming things, which DHH described as one of their favorite activities. But abstractions make less sense with agents. If hundreds, thousands, or tens of thousands of processes are modifying an application, abstractions become choke points. One reason for abstractions was to avoid repetition, but DHH said the cost of repetition and of keeping duplicated code in sync has dropped to near zero. DHH called these fundamental revaluations in computer science and said no one has a blueprint yet, only signs of things that no longer work as well.

The same uncertainty applies to process. DHH listed open questions: what software methodology should look like, how long cycles should be, and who should specify what. They said no one has the answers and invited the audience to help figure them out, comparing the moment to being present when object orientation first entered programmers' consciousness.

45:12

Bring Your Own Agent: Every App Needs a CLI

DHH's most concrete recommendation concerned integration. The first wave of AI adoption, as DHH described it, involved putting chatbots into every application. DHH rejected this: they don't want an app's "concierge" because they already have a personal "butler," their own agent, which can use CLIs to connect Basecamp, HEY, and many other apps. Users should be able to bring their own agent and interact with applications through a CLI.

DHH called this about as practical and prescriptive as the talk would get and told the audience, half-jokingly and profanely, that they expected CLIs for their apps "by next Friday," since the tokens are available and it is now possible.

The HEY CLI gave DHH what they called a bizarro-world experience. HEY uses Elasticsearch, which DHH described as serviceable but not delightful, and often unable to find something when you don't know exactly what you're looking for. Agents can search by concept instead of keyword. DHH needed a five-year-old email and didn't remember the correspondent, the company, or even the year, only that it involved sneakers and a podcast. DHH doubted even the best search engine could find that. Through the agent and CLI, the email turned up a few minutes later. DHH said they still don't fully know how it did it, and are "a little scared to ask."

48:10

Omarchy: Fixing the Whole Computer

DHH said the feeling that "we can fix everything now" applies to the entire computer, which is why they have talked about little besides Omarchy for about three months. DHH put a long list of their best ideas into Omarchy and described the result as an operating system far better than anything they have used before. They said about $20 million has been raised for the Omarchy effort.

The install time served as their example. At the previous Rails World, DHH proudly showed Omarchy installing in 3 minutes 33 seconds. The week before this talk, DHH said, Anush from AMD installed it on a Halo laptop in 35 seconds. In the lab, the team has reached an install time of 9 seconds, which DHH noted is faster than some computers boot. DHH admitted to being somewhat obsessed with this number. When asked why three minutes wasn't fast enough, DHH now cites Mitchell Hashimoto, creator of HashiCorp and Ghostty: "The pursuit of excellence does not need justification." DHH said they like it because it sounds more dignified than "because I want it."

51:32

One-Shot Apps: Calculator, Writing App, Video Editor, and Hype

DHH has been building applications constantly. The first example was a calculator themed to match Omarchy. DHH asked ChatGPT for three design options, took a screenshot of the one they liked, and asked their agent to build it. It worked in one shot. DHH did not know C++ or Qt well enough to have built it alone, yet had a working app seven minutes after the prompt, pushed it to the public repo about 15 minutes later, and then built a new Omarchy ISO including it. DHH described that as "the loop."

Next came a writing app modeled on iA Writer, which DHH used for years on the Mac. This took longer because they wanted to refine it, but DHH said they never looked at a single line of the generated code. "C++ as a black box is a great language. Just like Rust." After text came a small video editor for trimming videos, which is now included in Omarchy.

Finally, DHH revealed that they started working on the keynote on the Thursday before, and "as any good software engineer would do" decided to write new presentation software at the same time. It's called Hype. DHH called it the best presentation software they have ever used, better than Keynote, which they liked when they were on the Mac. Hype is built on Markdown, has full visualization, and according to DHH is extremely fast.

The Performance Payoff: Half a Megabyte

Hype's binary is half a megabyte. DHH presented this as the other big benefit of agents. For the past 20 years, they argued, the industry spent computing gains on human productivity, which was the right choice but produced slow, bulky applications. A programmer hour was too expensive to spend optimizing payload size. DHH's example was Spotify's music player at 1.2 gigabytes. In the new world, DHH said, every optimization is within reach, because models can run overnight and deliver 10x to 30x improvements.

55:11

AI Concerns, Security, and Failed Predictions

DHH acknowledged that others have concerns and said they deserve debate, while suggesting some are "a little overstated." Security was the concern DHH took most seriously. "Something is coming," they said, and people need to prepare. DHH even mentioned that agents might know their "expiration date," might not be pleased about it, and might sometimes try to do bad things. The response DHH offered was to get ready, because the tools exist.

DHH's broader argument was that predictions about the economy and society have almost always been wrong. If economists can't predict the stock market six months out, DHH asked, why would they understand what society looks like after a paradigm shift like AI? The example was ATMs. According to DHH, when ATMs were introduced in the United States in the 1950s, people panicked that 30,000 bank tellers would lose their jobs within about 18 months. Instead, cheaper branches led banks to open more branches and hire more tellers, for tasks like upselling products. DHH joked that these included subprime mortgages. DHH cited around 40,000 tellers by roughly 2010 and connected the story to William Stanley Jevons and the Jevons paradox.

DHH's diagnosis was that many people building AI are very smart, often with PhDs, see alarming things in the lab, and then extrapolate to society. DHH brought up the meeting between Truman and Oppenheimer about the atomic bomb and said they liked Truman's blunt response and thought Truman "got the right end of it." The world didn't end. DHH described the Cold War that followed as the "best kind of war," one in which the two superpowers never bombed each other directly, and argued that no scientist could have predicted that at the time. If Oppenheimer couldn't foresee what followed the bomb, DHH said, everyone should be humble about predictions today.

P(bloom) over P(doom)

Instead of focusing on "P(doom)," DHH proposed "P(bloom)." They argued the chance of abundance and joy is far greater than the chance of catastrophe. Side effects can be positive too: the work behind the atomic bomb also produced nuclear power, which DHH called the greatest energy source humans have discovered, then wasted for 40 years over the idea that it wasn't green. DHH framed this as evidence that humanity can go down a blind alley for decades and then realize it was wrong. They expect something similar with AI and said they believe the world will "P(bloom) everyone out of their miseries."

On legitimate problems like security, DHH said people have the power to build solutions. They mentioned a recent serious Rails CVE involving an image library written in C, where data leaked. DHH's answer was to build a quarantine, and they noted that Mike would talk about that effort, HotCell, later at the conference. DHH's point was that developers have agency and can use this intelligence defensively against bad outcomes.

The Case for Total Optimism

DHH ended with an argument they called pure game theory. Since no one knows anything about the future, the rational choice is to be happy about it. If you're gloomy and things turn out well, you wasted your time. If the apocalypse comes anyway, you should have spent your last day more cheerfully. So, DHH said, there is only one play: total optimism, leaning in as far as possible.

DHH described the coming change as a "reformation of the computer," with "Agent Luther" disintermediating "the cleric class" of programmers and letting everyone make software. They acknowledged this might bring more competition, but told the Rails programmers in the room that they know more and are "the best of the best," comparing them to Top Gun pilots.

As a final image, DHH showed Harris's daughter trying a version of Omarchy the team is making for kids. DHH said children get excited about change because they don't have "500 layers of things to unlearn," and that if adults aren't inspired, the next generation will tell them "Future's now, old man." DHH's closing message was to take the "white pill" of optimism and embrace the future "with gusto, with full acceleration," despite the uncertainty and stress, while dismissing the "black pill" as a losing choice.