30 秒でわかる
AI agents turn Linux's once-intimidating command lines and configuration files into interfaces anyone can ask to change, argues Omarchy creator David Heinemeier Hansson. He says a small team built the latest version in months and sees a chance to challenge established platforms. Code alone will offer less protection, he believes, but human taste and persistence still shape the last details. He also wants open models to compete with frontier labs, so powerful AI does not become another closed platform.
- 1Agents can make an operating system directly changeableHansson says users can ask Omarchy to change a menu, clock or calendar. Its Linux command line and configuration files give agents interfaces that previously made the system harder for people.
- 2Hansson sees app lock-in weakeningHe says an Omarchy community member recreated the Lightroom features he needed in around two weeks. He expects agents to build missing software or perform some tasks without a traditional application.
- 3Taste moves to the last 2% of the workHansson says agents make it easier to fix small flaws, but he still has to supply an eye, a vision and the persistence to polish the result. He says the work has him putting in 16-hour days.
- 4One subscription often covers his own agent workThough he has three Claude subscriptions, Hansson says a single $200 monthly subscription is sufficient for him most weeks. His account describes his own use, not a general cost guarantee.
- 5Open models provide a check on concentrated AI powerHansson welcomes frontier labs' advances but worries about dependence on a small number of providers. He says increasingly useful models that run on local hardware make competition possible.
- 6Omarchy's ambition grew from a personal desktop to challenging macOSHansson says agents expanded his goal from making a nicer personal computer to building his ideal one with a community. He thinks Linux could exceed macOS in three years, while giving lower odds to doing so within one year.
ポイント
- Linux's command lines, configuration files and detailed errors give agents ways to modify the system, Hansson argues.
- When code is easier to produce, the persistence to refine the last 2% can distinguish a product.
- Hansson says a small part-time team made the latest Omarchy version in roughly two and a half months.
- Open-weight models can give users an alternative if a few frontier labs dominate AI, he argues.
チャプターごとに整えた全文を読む
From Basecamp and Rails to an agent-built Linux desktop
Hansson traces his work from Basecamp and Ruby on Rails to Omarchy. He explains why frustration with Apple's developer policies brought him to Linux.
David Heinemeier Hansson says he has spent roughly 30 years building internet software. Basecamp, his project-management product, became the source of Ruby on Rails, the web framework later used for services including GitHub and Shopify. He still works on both, and his company also makes the email service HEY.
His latest project is Omarchy, a Linux operating system distribution that he calls native to the age of agents. He says hundreds of thousands of people are switching to it from Windows and macOS. He had used Linux on servers since the 1990s, but only began using it on his desktop around two and a half years ago.
After years as a Mac user and advocate, Hansson says he grew frustrated with Apple's treatment of developers and its share of app revenue. He moved to Linux, but thinks such grievances would never have persuaded most people to switch. Agents, he argues, change the calculation.
Why agents change Linux's usability
Hansson says the interfaces that once intimidated users are easy for agents to manipulate. He expects a user to be able to reshape the operating system simply by asking.
A host describes asking an agent to modify the operating system and seeing it work. Hansson says Omarchy itself has been developed through agents for roughly five or six months: he has written no lines by hand during that period, but hundreds of thousands through agents. Users, too, can ask for changes to a menu, clock, calendar or other part of their computer.
Hansson argues that operating systems designed around mouse clicks are poorly suited to agents. He cites changing workspace shortcuts on macOS, where he says users must work through settings one field at a time. Linux's command line, editable configuration files and detailed errors, historically barriers to human users, give agents precisely the interfaces and information they can use.
He worries that Apple could become the AOL of the agent era: a restricted environment offering a diluted alternative to the open internet. Asked about phones, he says the same shift could open the door to another mobile operating system, after years in which Apple and Google have divided the market.
Can agents break the hold of app ecosystems?
Hansson predicts that applications can be replicated or replaced with agent workflows. He gives photo editing and contract preparation as examples.
Hansson points to the lack of applications as a major reason Windows Phone failed. He says a member of the Omarchy community recreated the Lightroom functionality he needed in roughly two weeks, including the reason he had kept a Mac. His prediction is that agents will make it much easier for a new platform to obtain the applications people want.
One host adds that agents might do an application's work directly, removing the need to recreate it. Hansson says he now asks an agent to draft contracts, reviews them in Markdown, and has the agent create and send the signed PDF. He says he has used that process while setting up the nonprofit behind Omarchy.
He expects photo editing to change similarly: rather than learn a complicated program, a user can ask an agent to remove someone from a picture or improve it. He predicts that some entire classes of applications will disappear, while people will also be able to create a needed piece of software on demand.
Hansson describes a Fireworks demonstration generating a Tetris game in seconds at a reported 1,500 tokens per second, then changing another game live as the presenter requested lasers, bombs and larger monsters. He sees it as a preview of software that can mutate as people ask for changes; much higher speeds, he speculates, could take that further.
He says the current Omarchy version was made in about two and a half months of part-time work with his main project partner Ryan Hughes. In his view a handful of people can now produce something competitive with products built by thousands. Development that once required years of developer salaries and venture funding can, he argues, be compressed into days, weeks or months.
As models improve, Hansson expects large bodies of source code to offer less protection from competitors. A host pushes back with the example of StarCraft, whose balance still receives small adjustments decades later: perhaps the real advantage is the accumulated design judgment rather than the code.
Craft and the last 2% when code is abundant
A host asks whether accumulated design judgment remains a competitive advantage. Hansson says agents let him pursue details faster, but a person's taste and persistence still guide the work.
Hansson agrees that taste and care remain necessary. Agents have reduced the distance between noticing a misplaced pixel and fixing it, letting him put more of his perfectionism into software. He is working 16-hour days despite having co-written a book called It Doesn't Have to Be Crazy at Work, and says that capability can be intoxicating and hard on sleep.
Many projects, he says, are satisfying for their creators in the first 80%, then stall with rough edges and bugs. He gets his own reward from the last 2%. He has not found a setup that produces a perfect design in one prompt: someone still needs an eye, determination and a vision for how the pieces fit.
Hansson has no desire to manage a company of 6,000 or 60,000 people instead of 37signals' roughly 60. The prospect of directing agents with enormous productive capacity, without the same coordination and management overhead, interests him much more. He compares that experience to what Shopify's Tobias Lütke has called being in hyperdrive.
He says the tools are also cheaper than people may assume. Although he can afford many tokens, he has three Claude subscriptions and says that in most weeks a single $200 monthly subscription is enough for what he wants to do. He is using frontier models rather than trying to minimize cost with local or cheaper models.
For now, Hansson sees a role for a person directing and polishing the work. He also thinks it is possible that AI could surpass people at those tasks as soon as the following year, making a bespoke computer from a simple request. He says he is savoring the present moment while humans can still add something.
What would count as superintelligence?
A host describes trying to get an AI to build a robot. Hansson distinguishes agents' broad software ability from the harder physical challenges of robotics.
A host describes his own test for superintelligence: give a model a budget and see whether it can order parts and build itself an autonomous robot. The experiment currently fills his house with components, but has not produced a walking robot. Hansson says he has no equivalent test.
For software, Hansson already considers agents superhuman in the breadth and speed of their work, even though individual programmers can still outperform them in narrow areas. He contrasts software optimists' short timelines with robotics practitioners he met in China, who thought physical capabilities might take four, five or even ten years.
The host says his agent has pragmatically started ordering home-automation equipment, perhaps producing a talking, house-controlling display before an embodied robot. The exchange prompts a joke about giving an AI a vague goal and becoming a paperclip: a reference to an objective pursued without regard for human intentions.
Why open models must compete with frontier labs
Hansson wants both frontier models and usable open alternatives. He argues that competition can prevent another concentrated platform market.
Asked about frontier AI labs versus open source, Hansson says open-source software and openly available model weights are essential to competition. He recalls the early web as a free platform on which anyone could build without permission, followed by mobile app stores that made access to users dependent on platform owners.
He would welcome neither government-enforced control of AI nor a commercial duopoly between leading labs. He compares the stakes to the 1990s debate over encryption, arguing that restricting public-key encryption would have harmed internet commerce and communication. Competition among several closed-model developers is useful, he says, but open models matter even more.
Hansson says open-weight models are becoming useful, though he distrusts benchmark claims that they have fully caught up with leading proprietary models. He says he had previously dismissed powerful personal computers for local AI because available models were too weak, but changed his mind: a desk-sized machine can now run a model he finds useful.
He wants both frontier labs and open models to flourish. Without competition, he worries, providers could raise prices substantially because customers would pay for the capability. He points to Google and Meta's dominance of online advertising as his example of what happens when a small number of companies capture most of a market's value.
A personal computer project becomes a bid to rival macOS
Hansson explains how agents raised his ambitions for Omarchy. He offers a speculative timeline and describes focusing on what current tools can do.
A host asks how Hansson's ambitions have grown as the models improve. Omarchy began as a personal project: he wrote Bash scripts by hand and wanted a better computer outside Apple's control, perhaps for himself, a few friends and some people online. Tens of thousands adopted the earlier project.
With agents, he says, the ambition shifted from a slightly better computer to his idea of a perfect one, built with a community. He now thinks Linux could surpass macOS in three years and says even twelve months is possible, although he would not put majority odds on the shorter timeline. He describes the goal himself as a potential delusion of grandeur.
Hansson prefers concentrating his ambition on a specific project and what existing tools make possible. Trying to picture several model generations ahead, he says, can lead to anxiety about whether humans will have a role. He calls his own present attitude 'AI euphoria' and acknowledges that the future is difficult to predict.
He argues against dwelling only on worst-case scenarios, offering his controversial view of nuclear deterrence as an example of how hard outcomes are to foresee. He feels more comfortable extrapolating from capabilities he can already use, saying that significant increases in economic growth are possible, while predictions about next year's models remain uncertain.
Disruption, anxiety and the return of hackers
Hansson acknowledges programmers' fears while arguing that society must adapt to the technology. He finds inspiration in established programmers and pseudonymous independent hackers.
A host asks which people Hansson follows for guidance through a confusing technological shift. Hansson says the programmers he respects have embraced the tools because they can see how much faster they make work. He also recognizes that programmers who are not financially secure may reasonably fear that the skills they worked hard to acquire could lose value.
He compares that anxiety to earlier disruptions experienced by factory workers and weavers. In his view it is possible to sympathize with displaced workers while still accepting technological change and finding ways to bring people along. He describes public attitudes toward AI as a cultural battle, while acknowledging that not everyone uncomfortable with it belongs to one political tribe.
Among the people he follows are established programmers such as John Carmack and Tim Sweeney, and a newer generation of pseudonymous hackers. He says members of the Omarchy community and foundation contribute under handles, a practice he associates with the online culture of the 1990s.
Hansson sees a return of independent experimentation in Linux and local AI. Complex infrastructure such as browser engines had become the province of large companies, he says, but agents give individuals a chance to contribute to consequential parts of computing again. He credits online networks, including X, with helping these communities connect.
Bringing 1990s demo culture back to the desktop
Hansson wants computers to have character as well as utility. He describes Omarchy's animated effects and his early experiences with bulletin boards.
The hosts and Hansson celebrate 1990s demo culture, where artists built technically elaborate visual displays for their own sake. Omarchy's screensaver uses ASCII art and dozens of demo effects. He wants the system to have its own terminal-centered aesthetic rather than imitate Windows or macOS.
While preparing local AI setup, Hansson added animated characters to occupy the screen as a model loads. They serve no functional purpose, he says: they are there to make using the computer feel exciting. He remembers the animated intros on games and the distinctive character of bulletin-board systems.
At 14, Hansson ran a bulletin-board system from his bedroom with three phone lines. He still remembers the sound of a 33.6 modem connecting. Computers, he says, should retain some mystery and character; he feels that Apple's highly polished design has removed too much of it.
The conversation closes with the participants looking forward to meeting in person. Asked about a frequently photographed office view, Hansson says the office is real and in Malibu; he still enjoys the view whenever he works there.
同じチャンネルの記事
38:48EN
33:11EN
16:19EN
1:03:18EN
46:06EN
46:11EN