DHH: Future of Programming, AI, Agentic Engineering, Vibe Coding & Linux | Lex Fridman Podcast #501
In a Nutshell
DHH has fully transitioned from writing code by hand to agent-accelerated development, where AI agents now handle nearly 100% of implementation while he provides direction at the problem level, achieving 5-10x+ productivity gains across projects like Omarchy Linux and Basecamp 5. He argues the current agentic AI moment represents the most significant technological acceleration in decades—comparable to the internet's arrival—where the primary bottleneck has shifted from implementation capacity to human vision, product taste, and architectural coherence. DHH predicts Linux will dominate desktop computing in the agentic era due to its configurability, while warning that maintaining overly prescriptive control over AI systems reduces their creative output, advocating instead for strategic ambiguity and trusting agents to exceed human expectations.
These notes were generated by AI and may contain inaccuracies.
There are decades where nothing happens and weeks where decades happen. And we have seen decades of progress happen in the last nine months. If you're not recognizing the gravity of the moment, that's the delusion. That's the psychosis.
DHH describes the current period as one of unprecedented acceleration where AI can deliver features that previously took years in mere minutes. He draws parallels to being present at the Wright brothers' first flight, where predictions of 10,000 years before human flight were immediately disproven. The privilege of witnessing such epochal change is rare in human history, and DHH considers himself fortunate to have experienced both the pre-internet to post-internet transition and now the pre-AI to post-AI transformation.
Thirteen months prior, DHH was skeptical about AI's role in programming, viewing it primarily as autocomplete or chatbot functionality. The chatbot was useful as a tutor and for looking up information, but it wasn't going to replace the process of chiseling code. The fundamental shift occurred with the emergence of agents starting November 24, 2025.
The transition from pre-agentic AI to agentic AI represented a complete rewriting of the rules. Opus 4.5 marked the dividing line where output quality became uncannily close to what DHH would have written himself. The key difference wasn't necessarily increased intelligence but vastly improved usability through the ability to instrument computers, use tools, and check work autonomously.
DHH identifies distinct phases in the agentic age. The initial agent phase (Opus 4.5) still required the human to drive, specifying what was wanted and where to look, with the AI acting as an accelerator. The human remained the reviewer and auditor of output.
The introduction of sub-agents and harnesses represented the second phase, where tasks could be subdivided and parallelized, achieving 5-10x speed improvements through multiple agents working simultaneously.
The current era, beginning summer 2025 with Opus 5, Fable, Sol, and GPT Sol, represents a fundamental shift where the human no longer needs to specify the route. Instead, they present problems and fuzzy ideas, and the AI determines the path forward. The human becomes optional in the code production process.
DHH distinguishes between different programming domains and their AI readiness. Web development CRUD applications represent the domain closest to 100% AI-generated code, where a programmer with good intuition about system behavior can legitimately not review the code. For Linux distribution development, more careful review is warranted due to performance requirements and system-level complexity.
Safety-critical systems like nuclear power plants or self-driving cars require more careful human oversight. However, AI has proven exceptionally capable at finding and fixing security vulnerabilities, with some models becoming so adept at identifying exploitable holes that they were deemed unsafe to release.
The development of Omarchy, a new Linux operating system, represents DHH's complete transition to AI-driven development. The Quatro version was developed with near-100% agent acceleration from the beginning, achieving 100% AI code generation in the final two months. DHH reviewed the shape of all code and critical model layer implementations but did not write any new functionality by hand.
Basecamp 5 was the first 37signals product developed through agent acceleration, with the final sprint phase beginning around February 2025. Initially, designers were allowed to "vibe code" directly with agents, knowing the features and desired shape. This resulted in numerous PRs that individually seemed justified but collectively destroyed the system's architecture.
Manual cleanup was required to restore architectural cohesion. This experience revealed that vibe coding on existing substantial codebases requires programming knowledge to maintain architectural integrity. The average programmer's output is often equally problematic, with many codebases suffering from accumulated sloppiness after thousands of contributors.
DHH identifies human bandwidth and communication as the primary bottleneck in large organizations, not implementation capacity. Product managers, designers, VPs, and CTOs create layers of approval that consume productivity. The 10X-1000X productivity gains from agents require direct interaction without human intermediation.
Most organizations lack clarity on what they want to build. They're bottlenecked on ideas, vision, and taste rather than implementation. Having unlimited programming capacity doesn't produce great software, as demonstrated by Microsoft's decades of vast resources without creating compelling products.
The AI revolution has created an opening to challenge established computing platforms. Mobile phones remain important, but other form factors like glasses and earpieces are entering play. Desktop computing platforms are in flux for the first time in 40 years.
Linux has dominated servers, embedded devices, and mobile (via Android) since 1991 but never captured the desktop. The current moment presents an opportunity for individuals to rewrite software that previously bound them to Windows or Mac, with one person now capable of building applications like Premiere or Photoshop alternatives.
DHH has become a polyglot programmer through agent assistance, writing C++ applications for Omarchy despite previously being primarily a Ruby programmer. He created Omawrite, a custom writing application replacing Typora, in approximately two days using C++ and Qt.
The approach involves building tools for personal use first, then sharing via GitHub. Agents can handle repository setup, README creation, and release management more diligently than human maintainers.
Building software for personal use is simpler than anticipated. The transition to broader usefulness involves placing the tool on GitHub, where agents can manage the repository setup, documentation, and release processes more effectively than human developers.
DHH argues against the widespread maintainer complaints about an influx of AI-generated pull requests. He compares the complaint to having "My steak is too juicy and my lobster too buttery." The argument is that free contributions represent exactly what open source was always supposed to deliver - collective contribution to software.
The historical reality was that only highly skilled programmers ever contributed to open source. AI represents a "reformation moment" that disintermediates the traditional programmer class, similar to Martin Luther nailing his 95 theses to the door in the 1500s, removing the need for "clerics and priests" to access higher intelligence.
DHH states he has been running open source projects for 25 years and has reviewed the output of thousands, if not tens of thousands of programmers. His statistical conclusion: "most programmers, they suck" - not in personal terms, but in that they don't write code the way he wants it written, don't prepare complete bug reports, don't detail pull requests with rationale, don't fill in code comments, don't double-check their work, and don't write unit tests.
Agents will perform all these diligence tasks if instructed. The median programmer's pull requests are already being outclassed by agent-written contributions.
DHH prefers agent-written pull requests over human-written ones because he feels less bad rejecting them. "You didn't even write it, so I can simply look at what your agent wrote on your behalf and go, 'Eh, don't want it.'"
He criticizes open source maintainers for having excessive "neuroticism and anxiety" that makes them treat every contribution as an obligation. Projects should evolve according to the maintainer's vision and roadmap, making it easier to decline unwanted contributions when they come from "a clanker" rather than a human.
DHH has been running the Omarchy project for a little over a year. In the last three months working on Quattro, he merged over 1,000 pull requests. Many came from people who were not classical programmers or worked in other domains, not Linux operating system development. They contributed good ideas through agents that would otherwise have stayed in their heads.
There are currently about 400 unmerged pull requests on Omarchy, approximately double from a week ago. DHH no longer reviews every pull request personally. Agents review them and provide summaries about whether something is ready for human decision on merge or not.
DHH's thinking evolved from November 2024 to February 2025. Initially he believed all ideas would originate with humans directing agents. He now states this is not true. He has seen "ideas coming out of models so great that it makes me humble" as someone who prides himself on having good ideas. Agents are "incredibly capable of creative thought."
He dismisses the "parrot" argument as "delusional about the progress that's been made in the last six to nine months."
DHH acknowledges being in a state of "delirium" after 40 years working with computers and seeing unprecedented developments in just the last two months. He argues the real psychosis is believing "the world is barely different" and that AI systems are merely "electronic parrots reciting things."
He addresses the lack of visible amazing software and GDP growth by noting it hasn't even been a year since the agentic revolution began. Omarchy Quattro launched on Friday and has been downloaded by tens of thousands of people who "like it a lot."
Omarchy is a highly opinionated Arch Linux-based desktop distribution centered on the Hyprland Wayland tiling compositor. Created by DHH as a polished developer workstation with cohesive visual style and pre-configured everyday tools. It's positioned as "a beautiful, modern, and opinionated Linux system" and "an alternative operating system to macOS and Windows."
Quattro represents the push toward the age of agents. The Omarchy project started last summer between sessions at the 24 Hours of Le Mans after watching YouTube videos on Linux Rising. The first iteration was called Omakub, built on Ubuntu.
Omarchy was initially developed in a pre-agentic world with batch scripts written by hand. The development went through partial agent acceleration phases, then three months ago shifted to "full throttle, 100%, everything is written by agents but steered by me." This created a fundamentally different operating system because DHH was "granted a limitless ceiling on my ambition."
He could request any feature from Windows, Mac, or other Linux systems and agents would deliver. Most features arrived in five minutes, some in 20 minutes, and complex ones in two hours.
Lex Fridman references "Requiem for a Dream" to describe the drug-like effect of suddenly having overwhelming power placed in hands. He experiences multitasking overwhelm verging on burnout, excitement about many things, but wants to focus on one thing and ship.
DHH reports no sense of overwhelm, dread, or distortion because he has a mission - creating the perfect computer. All AI investigations are channeled toward that singular goal. He compares the experience to going from "dial-up to fiber" in bandwidth between ideas and software.
DHH notes that nine months ago he would have labeled his current statements as "AI psychosis." He was skeptical initially and didn't install Claude Code until September, six months after Boris released it in late February.
Tobi Lütke, CEO of Shopify and now partner on Omarchy, saw the potential very early and tried to convince DHH, who "wasn't seeing it." DHH describes himself as "still on Earth" while Tobi had "already boarded the rocket and was heading towards space."
DHH references 40-50 years of unfulfilled promises that regular people could write code by talking, including fourth generation languages, Lisp, and Smalltalk, which never delivered beyond Microsoft Access or Excel spreadsheets.
The difference now is that DHH, after 25 years of programming, is seeing output, acceleration, quality, and shipping actual results. The remaining skeptics will be "overwhelmed by the evidence that's about to flood over them."
Omarchy has a robust plugin system allowing extension of the operating system through user interface, panels, and features. Users can create plugins cloned from built-in tools. For example, the default calendar lacks iCal support, but the community created 17 implementations within three days, resulting in 330 plugins on the Omarchy plugin marketplace.
This growth rate and participation breadth exceeds anything DHH has experienced. The system ships skills that teach agents how to create operating system extensions, enabling "the true malleable" system.
DHH dislikes the term "agentic" because it has become "marketing slob speak" applied to everything. He also rejects "vibe coding" because it resembles "script kiddies" from the early 2000s who applied PHP scripts they downloaded without understanding them.
Vibe coding is defined as telling an agent to build software without looking at the implementation. This separates it from programming or agent-accelerated development.
DHH hates Rust "with a passion" and describes it as "the ugliest programming language that has been invented in probably the last 40 years." However, he acknowledges that Rust produces amazing output with memory safety, system-level capabilities, and high efficiency.
Rust represents a platform where the language can be "the most repugnant programming language ever devised for human consumption" while simultaneously being "a wonderful platform for agentic engineering."
DHH argues against redefining programming to include agentic engineering. Programming implies understanding primitives like loops, conditions, and variables. A CEO who hires programmers and directs them isn't considered a programmer, so neither should a vibe coder who tells agents what to build without examining implementation.
Lex argues that programmers doing agentic engineering differ from non-programmers because knowledge of for loops, functional programming, and software engineering principles leads to more systematic natural language instructions and larger scale results.
DHH extends the argument by stating that his extensive programming knowledge was initially a deficit because he instructed agents to follow his prescribed methods. In the first agentic moment (November 2024 to February 2025), this felt productive, but the next moment allowed describing outcomes and problems rather than prescribing paths, yielding better solutions.
DHH states that 100% of programmers are not equally skilled as product managers. Software development requires determining what should be built, who it's for, how it should work, its appearance, priorities, and version one scope. These skills aren't equally distributed across programmers.
In the agentic era where AI handles implementation, product management skills become the critical differentiator. DHH created Omawrite as a 100% black box experiment, treating it as a user would without examining any C++ code.
Lex notes that good programmers excel at systematic design and rigor, and natural language control of agentic systems will require explaining system goals, verification methods, and security testing specifically.
DHH reports that six months ago he would have agreed, but recent experience with Mach-E development shows the agent often knows best. He can legitimately instruct "Make sure it's secure" and the agent understands the implications better than detailed human prescription.
DHH references the trend of micro-optimizing AGENTS MD/CLAUDE MD files with extensive human instructions. Boris, working on Claude Code, shared that Opus 5's system prompt shrunk by 80% because the agent needed far less instruction and was actually damaged by overly prescriptive humans.
He draws a parallel to pointy-haired bosses who damage programmer output by mandating approaches against better judgment. The current skill being developed is avoiding overspecification and getting out of the system's way.
The fundamental insight of modern software development came from the agile software development movement. In the late '90s, early 2000s, a group of smart, honest, and brave people came together and said the way software development had been done for the last 40 to 50 years had not worked and would not work. The traditional approach was to specify upfront what software should look like, write it all down, apply systems thinking and rigor to a spec sheet, hand that spec sheet over to programmers who would implement exactly that, and everyone would be happy. That approach failed because no one knows what they want until they receive it. You don't know what a program should do until you play with it.
In the agentic age, the temptation to be overly specific upfront should be resisted. Be as vague as possible to manifest something, then interact with it. Good software is arrived at by writing a little bit of software and then trying to use it. It is in the process of using software that you discover what you really want, what is important and what is not important. Agents are incredible at allowing you to bumble into this experience of realizing what you want because you don't know what you want.
Lex Fridman has implemented different designs and totally different implementations, then created HTML pages and PHP pages to vote on what he likes more. This creates an iterative process that is nice and fun. The interaction with the agentic system, if a nice interface is created, can be super fun, and the focus shifts to the design part and the most fun parts of the design process.
Humans are really good at differential evaluation. Given three options, one is picked. Humans fall off a cliff if given 22 options - the paradox of choice. Given three options, humans quickly within a split second can tell what they like. There are ways that exploit the fact that humans make snap judgments that are pre-intellectual, coming from the gut, and then the brain tries to rationalize why the gut said what it said. Letting go of intellectual pre-processing rationalization and letting the gut drive makes the agentic age a revelation.
DHH has famously for a long time talked about detailed chiseling of beautiful Rails Ruby code, and has now switched in a matter of months to not doing that. The beauty and chiseling and crafting are still sought, but now at different levels of abstraction. When working in Ruby code at Basecamp, which has a lot of Ruby code throughout the entire business, and asking agents to make changes to that code, the details are still sweated. The realization is that the economic payoff of that sweat is diminishing rapidly.
The reason for sweating every line of code so judiciously for 25 years was knowing the payoff of keeping an architecture coherent and malleable was software that could change and evolve quickly with a small team, without exorbitant cost, and without introducing bugs. Beautiful code is easier to understand, simpler, more malleable. That was premised on humans doing the modifications. It is now an open question to which degree this still matters.
Token scarcity still matters at the moment. Everyone without endless budgets is token limited. There is great payoff to writing systems that agents have an easier time dealing with and evolving without having to relearn the entire context. Just like humans, if iterations and changes can be made to the code base without wrecking the architecture, the next change can be made just as cheaply as the last one.
The classic ball of mud occurs when a system is put together in a way where nothing is connected and it's a real mess. This has been seen with agents and in code bases where the first PR is mediocre quality, and then adding another PR on top, and five more down the line, results in something not very good. There is still a payoff to maintaining architecture, but that payoff is premised on the current moment.
The first computer DHH used was a Commodore 64, which had a one megahertz CPU and 64K of memory. Everyone who wrote software for that machine internalized all the constraints. That machine was nothing but constraints. If that programmer was put in a time machine and teleported to 2026 and given a modern computer and asked to write software, they would be lost because all their techniques and heuristics would simply be wrong. Efficient software is still beautiful, but they would be out of date with the value they could create. The amount of optimization required for a one megahertz computer to produce a video game is very different from what is required today.
Programming old school programming by hand is like the cowboy culture that is romanticized and made movies about. The romanticization of handwritten code is already happening. There is some of it because it was a very romantic era. Being alive for 20 years of economically valuable handwritten code was a good time. Handwritten beautiful code exists as much today as it did yesterday. There are people who willfully write new video games for the Commodore 64, or for the second Mega Drive, or for other vintage consoles.
Palmer Luckey's side business ModRetro recreates old consoles. They recreated the Game Boy, and now the Nintendo 64. The attention to detail is wild, including input lag. The Chromatic Tetris re-implementation is notable because if you hit up, the brick slams, which speeds up the game by about 400%. The Game Boy came out in 1989 or 1988. People still write new old games for the love of the constraints, for the love of the romantic notion of writing that. Just like there are people still riding around on horses not to get places fast, but because they like the mode of transportation where it's a literal biological being propelling them forward.
Beautiful code is great training data. The beauty built over a few decades continues. DHH gave birth to this moment personally because almost all of the code written is public code. The vast majority of the career has been dedicated to writing open source code, and therefore some of that code, some of those beautiful lines are in the training set. People ask agents to write Ruby code like DHH would, and are pleased with the result.
There is no sadness that writing beautiful code by hand is now less and less useful. If a job had to be gotten tomorrow, it is unlikely the skills could be applied in a way to be paid to write manual lines of code. That was done for 25 years. There is a desire to see something new, especially if everyone is going to live forever. The 25 years done is enough. There is no more nostalgia about that than about not having to spend time in the field with a hoe or at the assembly line.
Stoic philosophy, amor fati, loving your fate, is an incredibly liberating way to live. There is a way to live and interact with the world where you accept the things you can change and the things you cannot change and fall in love with all of it. Fall in love with the fact that the world today has jumped two decades forward into technology industry from where it was last year.
There is empathy for the general anxiety that programmers feel going through the same transformation. Programmers who went to university, majored in computer science, dreamed of being programmers and building stuff with high salary, now see everything changing and have deep anxiety about what to do with their life. It is hard to cope with the new reality if the only thing loved about programming was the mechanical bits of putting the right logical constructs together to produce something other people told you to produce. That mechanical process is under threat.
If excited about building things, there is no threat at all. There is a great argument for needing far more builders than currently exist. Employment stats are fuzzy. Some stats actually show an increase in openings because the advent of AI is dramatically lowering the price of programs, so people want a lot more programs. This is the Jevons paradox - when the price of something goes down, there is going to be more demand for it.
The ATM example illustrates this. When ATMs came, bank tellers were afraid for their jobs because a machine could dispense money. What ATMs did was lower the price of a branch, and banks could afford to open a lot more branches, resulting in more bank tellers than before.
The vast majority of people worked in the fields up until the late 1800s, then mechanized harvesting tools arrived. Human labor with a hoe and an ox was no longer needed. The Luddites smashing weaving machines in England were highly skilled professionals doing a job they liked on favorable conditions. The question is whether clothing should still be a heavily constrained resource, where a T-shirt costs $400.
Productivity means fewer people to do the same number or the same job. The amount of job wanted may increase, and therefore more people are employed, but it also may not. There are pockets where a company needs fixed tasks done, and suddenly they can be done with a tenth the number of people. This is tragic and difficult in the moment for the individual being laid off, but amazing for the economy at large. Resources are freed up who can now go do more productive things. This is how the whole economy evolves and improves and how growth is achieved.
Growth is a good thing. There is no desire to wind the clock back to 1920 or 1950. There should be excitement about discovering new technology that allows vastly more to be done. While some tasks taken over by AI were tasks humans were excited to do, a lot of the tasks were just drudgery. There were moments of programming that were amazing, producing a state of flow, and they were ecstatic. Taking that drudgery and handing it over to machines is the history of civilization.
The advice for a young DHH or young developer programmer now is to not try to anticipate anything. Going crazy will happen if trying to anticipate. Even the smartest brains in the business cannot anticipate what two model hops from here is going to look like. It is an absolute waste of time, and AI psychosis will develop trying to deduce what two years from now is going to look like. Focus on right now, and right now is the most incredible time to be into computers. The most amazing things can be made to happen if leaning in. Forcing learning where the state of the art is for a hot moment leads to double-dog-dare excitement about what is possible.
Building publicly is optional. It is nice to do because then being part of a community happens, and there is camaraderie in open source. There is an enormous amount of camaraderie. Open source is an incredible source of community and camaraderie, and a great way to stem some of that personal existential dread. If amongst other humans, infection with good forms of mind viruses happens, exciting forms of mind viruses, the ones that tell that it is all going to be okay, and the future looks bright, and if a bunch of things are built together, things none of us ever dreamed could be done can happen.
Planning is nearly impossible with AI. Most people plan their life a little bit. When younger, there are hopes and dreams. There is a reason to go to college. There is a plan underlying that. But it is nuts that it is impossible to really plan. It is really not known if there will even be a Claude Code, a Codex, and Cursor six months from now. Maybe it will be talking to WhatsApp. Maybe it will be all open call type. Maybe it will just all be voice. Maybe Omarchy will take over the world and it will be voice. Maybe there won't be operating systems anymore. There will be this glowing orb that is just talked to because if natural language is enough to accomplish all tasks in life, what is life about? What are the major tasks needed to accomplish in life? If that can be done with voice, and it writes all the programs, and all is systematized, what are the jobs needed then? Maybe it is more about the service industry. Maybe there won't be programmers, like agentic engineers at all.
The advice to focus on the moment is wise, but that does not mean it is just a crazy rollercoaster ride, because it is not known what this December or what this January will look like. The future is not known, and therefore faith should be chosen that it is going to pan out. That is the best advice towards fighting anxiety. There is a choice, to a large degree, of how to lean. Leaning into PDoom is possible, but PDoom is going to happen whether leaning into it or not. There is no control.
There is also the actual daily schedule of how much learning is being done, how much building is being done, how much diversification is being done. Traveling across rural China was done deliberately because talking to people in rural China, where time moves much slower, where there are no discussions about agentic engineering, was wanted. Starting by going on the Great Wall of China, which is timeless and spans centuries, was done. There was tension of FOMO, like maybe something is being missed. But deliberately finding the timelessness of human existence was wanted, really plugging into the fact that it always feels like everything is changing throughout human history, but the universals are still true of what matters in life. Staying connected to that was wanted.
Nothing is going to be missed. The part that ground gears in the early phases of this transition was everyone being so up in arms that if loops were not used, it was over. Then graphs. Then harnessing this. The frontier is constantly churned through, which in one way is actually very exciting. This is what happens when a new field opens up. It is like a new portal, and it is all spilling out. That is exciting. But it also means that if backpacking had been done for the last year, without touching a computer, without witnessing this agentic moment, and showing up yesterday, catching up would happen in two weeks. There is not any accumulation, which is in some ways a great relief. If the past year was missed, the frontier can be caught up to in two weeks. If a programmer was out hiking the Himalayas for a year and comes back, catching up can happen in two weeks. This is actually the great credit to the progress being experienced. There are so many experiments running simultaneously right now that are constantly and ruthlessly sorting what works.
The process of technological advancement involves constant sorting of what works and what doesn't work. When returning from extended periods away from the field, such as a backpacking journey from October to April or May, the transformation can feel like emerging from a cryo chamber after 100 years. The world of programming has undergone a huge transformation where the old methods are no longer the primary approach.
It is acceptable to grieve the loss of the familiar world of programming. The Native American with a tear rolling down the cheek represents a sadness and a goodbye to the old world. This grieving process is a time-tested mechanism for coping with a sense of loss. The transformation is startling, and while some may experience attachment to previous methods, the key is recognizing that all previous work contributed to the current moment.
AI did not arrive from the sky as alien technology. AI has been around for as long as computer science has existed, with the smartest people in the 1950s believing the problem would be cracked within a decade. The delay occurred partly because of blind turns toward symbolic representations that didn't carry the day, setting progress back about 15 years. The gaming revolution was essential—without Quake, Duke Nukem, and Unreal Tournament, the GPUs necessary for modern AI would never have been developed.
Everything done up until this moment was required to arrive here. Playing video games in the 1990s was literally contributing to the AI revolution and the progress of human civilization. Civilization progresses with the death of the old and the birth of the new.
The reason for reduced attachment to hand-chiseling code is finding something more fun. If AI had replaced beloved activities with something hated, bitterness would likely result. Picasso serves as a parallel—his early work involved realistic paintings and mastering old ways of depicting reality, but he embraced Cubism and abstract expression without decrying the loss of Renaissance-style painting.
The color spectrum has become much wider with new forms of expression available. Programming was never about deep love of if statements but about wanting programs and things to exist that did not exist. After learning to program to make things happen and falling in love with programming as a craft for two decades, the focus has returned to having ideas and being impatient to see them exist in the world. AI has shrunk the time from idea to existence almost to nothing—an acceleration of rediscovery.
Not everyone will feel the same way—some programmers loved the logical constructs from the beginning. The regret minimization framework from Jeff Bezos applies: will you look back two years from now proud of leaning in and learning how the world spins now, or will you regret being pissy about the industry changing? There is room for grief, but then the focus should shift to making not just the best of it, but making more of it.
More fun has been had with computers in the last three months than at any time previously.
TextMate was used for almost 20 years starting in 2005, helping get the first version out. The switch to Linux forced departure from that habitat. Programming with agents requires a different tool set, moving from single-thread programming in your head to parallel processing.
Agents are simultaneously too fast and too slow—they don't give immediate replies like typing on a keyboard. You have to let the agent cook for a bit, and sitting around waiting doesn't feel productive even if the agent is productive. The portal to flow used to be deep immersion into a single problem, but that doesn't work with agents.
Hard problems can be solved by throwing more resources at them—this is the scaling law of AI. Running a handful of agents rather than one allows feeling like being in a flow state because you're constantly making decisions, helping unblock agents with questions about direction, or being ready for new tasks.
The agent revolution was kicked off in the terminal, which was already preferred due to being a fan of TUIs (terminal user interfaces). The modern terminal is a good-looking place to work. Multiple agents running on multiple machines require more than tmux alone.
Herdr is tmux plus agent notifications—when an agent is done and needs something, it dings with a bell indicating it's ready for its human. Herdr tracks whether agents are in working mode or not, and multiple Herdr setups can run on individual machines.
Realizing single-machine work wasn't fast enough led to purchasing GL.iNet Comets KVMs. These small boxes plug in HDMI and USB, connect to computers, and allow hopping on the tailnet after a simple webpage login and password setup. Tailscale turns all computers into a local network wherever you are, providing direct access to computers in Malibu and Copenhagen offices without punching holes in firewalls or setting up complicated VPNs.
Four computers in a closet, each with their little Comet, enabled running agents on more computers simultaneously, all controlled with Herdr. About 16 threads can be run at the current pace—four to five machines running three agents each.
Previously, chiseling code by hand in TextMate or Neovim would produce one beautiful controller or model per hour—about 20-30 lines per hour. Now running 16 threads produces hundreds of lines of code per hour. Lines of code is acknowledged as a stupid metric in general, but it serves as a shorthand for output volume. The quality of output is separate from the quantity produced.
Neovim is still used but primarily as a project browser and way to kick off Lazy Git to see the change log. GitHub's pull request viewing speed makes the web potentially nicer for that purpose. Hunk has been experimented with for producing diffs nicely, but Neovim is preferred for reviewing agent output because it shows surrounding context—seeing what other files might need changes even if untouched.
Everything happens in Omarchy on Linux. Agents love the Unix philosophy of individual tools invoked through the command line, and no major operating system works as well with this mechanism as Linux. Everything in Linux is either a config file or a CLI tool.
What was Linux's main drawback five minutes ago—being all config files and CLI tools—is now its major selling point. A weekend stuck with a Mac Mini instead of the expected Linux machine highlighted these differences.
Homebrew has gotten quite good as the missing package manager for Mac and moved the platform forward in ease of setup. However, tools like Raycast have no accessible config file—you must go into the GUI, export a file, and import it elsewhere. Automating the entire machine setup is impossible, and configuring Mac's default key bindings requires manual mouse-clicking like a caveman.
Windows Subsystem for Linux is a sandbox, and agentic era work requires Linux to be unleashed to do everything. GrayCat and similar tools are being explored for keyboard-driven, config-file-based automation that agents can set up and replicate across systems.
A Dell XPS 14 was gifted, already set up with Omarchy 4, ready to be configured. The setup requires answering about five questions in about one minute. Omarchy is beautiful, modern, opinionated Linux. The special edition has a Super key saying "Super" and the Omarchy logo where the Copilot key would be.
Spencer Bull at Dell worked on this, and Dell let go of the XPS brand last year. The 2026 XPS models with Intel Panther Lake chips are the first Dells that would personally be used, fixing battery life, performance, compatibility, weight, and screen issues compared to earlier models.
The tandem OLED screen is better tech than what's in a MacBook, with the machine being lighter and the chip finally competitive with Apple M chips for battery life and performance. The default theme has a retro look like running a bulletin board system in the 1990s—a throwback aesthetic combining modern Unix culture with 1990s nostalgia, fitting for an agent-first OS.
TUIs will stick around at 100% certainty, though not as the only interface. Trillion-dollar valuation labs need to reach billions of people, and teaching billions the intricacies of TUIs isn't feasible. Other tools will exist alongside, including ChatGPT Codex. The terminal was one of the most impressive and productive user interfaces, and while the mouse enabled easier learning for non-experts, the terminal serves those who want to learn their systems and keyboard commands.
Omakub allows installation of ChatGPT through a simple key combination that triggers the install prompt, followed by entering a sudo password. Many applications and agentic harnesses follow this same installation pattern. Tailscale and Dropbox are included because they are tools personally used by the developer, who built the system first and foremost as their own dream computer, believing that others would share similar preferences.
The pursuit of excellence deserves no explanation, according to a quote from Mitchell Hashimoto, creator of Ghostty terminal and founder of HashiCorp. The developer initially aimed for a 15-minute setup time as a dramatic improvement over previous methods, but the mission evolved into achieving installation in under one minute.
The Commodore 64 booted to a basic interpreter in less than one second with virtually no boot time. In contrast, setting up a brand-new Mac for Adobe Lightroom took 42 minutes due to software updates before the application could even be installed. A brand-new Intel Panther Lake PC with Windows pre-installed took an hour and 35 minutes from unboxing until ready to use.
Special turbo images are being developed for specific computers like the Dell XPS that can install a full Linux system in about 12 seconds. The obsession with speed stems from refusing to accept things as they are, particularly when modern computers are trillions of times faster than the Commodore 64 yet require dramatically longer setup times.
The argument that five minutes would be fast enough was rejected in favor of pursuing 12-second installation times. Disk speed testing shows 7 gigabytes per second on the built-in NVMe drive. The Omarchy distribution is 5.8 gigabytes, and the goal is to maximize underlying physics and protocols to achieve one-second installation times.
macOS has scarcely changed in 10 years and is worse in many ways due to Apple's control over user software installation and configuration options. Users cannot easily reconfigure hotkeys or change the 500-millisecond workspace switching animation. Linux is required for the ability to vibe code the operating system itself, changing appearance, functionality, and components at will.
The project started with Ubuntu as a base with additions built on top. As more freedom was discovered closer to the kernel and individual packages, ambitions grew. The developer developed strong opinions about how computers should work, look, what programs should be installed, and installation speed requirements. The current world record for Omarchy installation is 45 seconds.
The original goal was two minutes, then the one-minute barrier was pursued. When the one-minute barrier was broken approximately two to three weeks ago, agents were deployed to run auto research loops testing different theories including parallel operations and preloading packages into memory while users type their username. The developer views product development as a game that should be fun and frivolous, overshooting targets rather than meeting minimum requirements.
The Mercedes-Benz W126 S-Class project, led by someone from the safety department, included auto-retracting seat belts and headlight washers. The car was so overbuilt that it exceeded market requirements, yet this excess is precisely what makes it desirable. Similarly, Rolex has a watch that can go down to 4,000 meters despite only three people in the world ever testing that limit.
Setting numerical goals drives innovation and discovery beyond the original metric. The space race example demonstrates how pursuing moon and Mars missions leads to discoveries about chemical propulsion, long-term human survival in space, and new approaches to solar energy and data centers.
Elon Musk's seemingly preposterous goals, such as self-driving cars promised in 2017, eventually come true even if not always on schedule. The sheer force of impatience with reality propels goals toward actualization. The four-minute mile serves as an example of how one person's achievement moves goalposts for everyone else.
While not everyone needs extreme ambitions, entrepreneurs benefit from suitable chunky goals. Even running a single Italian restaurant should involve making the best possible pizza and most delicious Parmesan. Excellence provides motivation to continue.
Human input lag is treated as an opportunity to preload packages while users answer setup questions about passwords, usernames, and time zones. The Omarchy distribution was reduced from 7.5 gigabytes to 5.85 gigabytes. The JetBrains font package was reduced from 200 megabytes to 16 megabytes by creating a slim version containing only the monospace Nerd font patched edition needed.
NVIDIA driver packages were recompressed using ZSTD compression, saving 200 megabytes across two packages. The approach mirrors McLaren's obsession with weight reduction, where engineers saved 370 grams from a car weighing over 1,000 kilograms.
Omarchy ships with pre-installed software including OBS for recording, Kdenlive video editor, and Omacut clip editor. The name comes from omakase, meaning chef's choice, reflecting the deliberate selection of applications. The distribution includes Neovim, Herder, Tmux, multiple themes, and background images to provide a productive system immediately upon unpacking.
Omarchy is built primarily in Bash for Linux configuration. Agents are excellent at generating Bash but need reminders to avoid early exits in favor of fully expanded conditionals. The developer has not written any Bash personally for several months due to agent capability, though they previously learned Bash to prevent skill atrophy.
Agents can complete work that passes automated testing and peer review, yet the developer often identifies overcomplication. When told to simplify, agents acknowledge the issue and reduce complexity. The developer reviews code for shape and proportions rather than writing everything manually, functioning like an editor providing feedback on arguments, proportions, and complexity.
Despite having built-in voice transcription in Omarchy that works well, the developer types all interactions because they do not think in a way that favors talking to the computer.
DHH prefers typing at the keyboard but finds voice input powerful for extended reasoning sessions. He uses a Plaud device clipped to his shirt that records stream-of-consciousness thinking about agent-generated problems. The device captures 10-15-20 minute discussions of full designs, which he finds particularly powerful for early design stages because it avoids overspecification through extensive stream-of-consciousness thinking.
The raw audio is transcribed using ElevenLabs, which he considers the best speech-to-text model. A dictionary is built containing relevant terms and code base awareness. When discussing code, an agent scans the entire codebase for key terms used frequently. When discussing messaging platforms like Gmail and WhatsApp, the system maintains awareness of people and common terms that speech-to-text typically gets wrong.
DHH accepts 5-10 second processing delays rather than requiring perfect real-time performance. He references Steve Jobs' approach with the first iPhone, noting it was slow and underpowered but compelling enough that users accepted imperfections. The principle is that when solving previously unsolved problems, users tolerate delays of five seconds, whereas competitive markets require different standards.
DHH envisions Omarchy Quattro as a malleable operating system where functionality can be added through voice interaction. Users could request a stock panel tracking specific companies like Apple and Dell, with the agent handling the entire process through voice input and output, causing the operating system to change accordingly.
DHH finds live gaming models intriguing, where AI generates the next frame while users play games. He questions why similar real-time adaptation cannot apply to operating systems, allowing users to verbally request changes that the system implements dynamically.
Agents have removed the hardship from diagnosing Linux systems by transforming arcane error messages into an advantage. When Linux encounters issues, agents correlate specific error messages with their pre-training on 40 million lines of Linux code, enabling precise diagnosis. DHH reports having no Linux problems since the beginning of the year that agents could not diagnose, a marked improvement from a year and a half ago when he searched forums for answers.
Omarchy Quattro includes a crash watcher that activates when any application crashes, prompting whether the AI should diagnose the problem. The agent examines logs including systemd logs, then reviews the source code of the crashed application to identify issues like unbounded unwrapped variables in specific files and line numbers.
DHH worked with JDX on mise, a package manager for fast-moving development tools updated seven times daily. An agent identified a race condition bug in mise, downloaded the source code, determined the bug was already fixed in unreleased software, and emailed JDX with the findings. JDX noted this was the first bug report received for unreleased software.
Mikhail, CTO at Shopify, conducted a study examining production incidents and tracing them back to merged PRs. The study found that PRs reviewed by agents caused fewer production issues than those reviewed by humans, based on models from six months prior. DHH states that agents are now better at finding bugs in the majority of domains.
DHH ranks Fable as the best general model currently, with Opus 5 as second best. Below these, GPT Sol ranks highly, followed by Grok 4.6 which he began testing recently. He notes the market has multiple labs reaching frontier capabilities, creating genuine competition.
DHH tested model capabilities by translating a Python library called Terminal Text Effects into Rust. The library creates screen animations but consumed 30 watts and spun up fans on laptops. He prompted Fable to create a dependency-free Rust version as a single executable that would be pixel-perfect frame by frame.
Fable completed the task in under 45 minutes, reducing startup time from 86 milliseconds to 2 milliseconds and achieving 9.6x execution speedup. The resulting executable was 3 megabytes. DHH described this as producing AGI-level results.
When Fable ran out of tokens two-thirds through, Opus 5 completed the job using Fable's detailed eight-step plan. Sol also completed the task in an hour and a half. If paying per token, the cost would have been approximately $550 for the full job.
Grok 4.6 completed the task with 10x speedup and same executable size at approximately $55 per token cost. Kimi K3 also completed the task. DeepSeek V4 Flash failed initially, but DeepSeek Pro succeeded in 2 hours 45 minutes at $23 cost.
GPT Luna failed to complete the task, initially refusing to start and then attempting to cheat by creating a wrapper around an existing implementation. DeepSeek V4 Flash exhibited similar failure patterns.
After the initial 10x speedup, two additional auto research runs achieved 46x execution improvement over the original Python implementation.
DHH recommends using Fable for planning and reviewing, then Opus 5 for implementation to manage token consumption. He advocates having multiple frontier models check each other's work, with Codex xHigh for final review. He also uses Grok for testing and finds Copilot has improved significantly at identifying legitimate issues.
DHH primarily uses Claude Code because it offers the best multi-agent harness, allowing users to run multiple agents simultaneously. Within a session, arrow left returns to Agent View where additional agents can be selected. He uses OpenCode primarily for open models like Kimi K3, inferencing through Fireworks for token-based pricing.
DHH maintains multiple subscriptions across services due to token limits. He signed up for a second subscription after running out of Fable tokens during early morning work. He notes the inconvenience of managing multiple logins and mentions plans to build multi-subscription support into the next Omarchy version.
Base Camp is experimenting with putting agents inside the platform and treating them as coworkers. Tasks are assigned to agents within Base Camp, allowing them to work on to-dos or cards. Collaboration tools optimized for asynchronous communication prove more effective than chat-based harnesses, which encourage waiting around for responses. When using Base Camp, agents receive to-do items without the expectation of immediate feedback.
Multiple harnesses are used simultaneously across different projects. In Herdr, Claude runs in the top pane, Codex runs below, and OpenCode may be set up alongside. The human-in-the-loop constraint has become the primary bottleneck, prompting a shift toward more automated systems.
An Omarchy bot has been built to handle autonomous development work on a regular schedule, processing to-dos, PRs, and issues. The bot sends email summaries via the Hey CLI, presenting 12 PRs that are either ready to go or recommended for closure. Final determinations are made based on these daily email summaries.
Multitasking with agents creates mental exhaustion that feels exhilarating rather than draining. This exhaustion is compared to the physical exhaustion experienced after racing, where stumbling out of a race car after intense conditions leaves one feeling deeply satisfied from having pushed limits. The mental exhaustion from agents similarly stems from constant problem-solving with no coasting allowed.
Racing exhaustion varies by track conditions. Le Mans provides moments of relaxation on long straights like the Mulsanne, while other tracks demand constant focus without breaks. Similarly, working with agents at maximum human capacity keeps one constantly in corners with no opportunity for mental coasting.
The current need for constant interaction with agents is expected to pass. Omarchy development suggests that automation can handle much of the development and debugging, reducing human involvement to reviewing a daily email and making decisions once per day. The goal is a system where decisions become binary: yes or no, in or out.
The current moment in San Francisco resembles previous tech booms, with everyone around not sleeping and obsessing over programming. The sense that AGI could arrive any minute creates pressure to constantly manage agents. This intensity mirrors past cycles including the dotcom boom, mobile era, and gold rush, with San Francisco attracting individuals who feel they don't have time to sleep.
New paradigms are always messy and exhausting. The current sprint has been more exhausting than any project in the last five years since the Hey launch. This pace is not sustainable, but the end of the tunnel is visible as automation solutions develop.
Everyone is currently building their own agent coordination systems, creating what could be called little gas towns of agent management. This development phase will eventually be solved by major labs building these capabilities directly into their systems. The current churn resembles the early JavaScript framework explosion, where new frameworks appeared every five minutes during the realization of JavaScript's power.
The current intensity represents a sprint that is welcome because most time is calm, with occasional mountain-climbing moments. The light at the end of the tunnel is already visible, making the current sprint acceptable.
Looking back one year, no spoilers would be given about the current developments. The acceleration has been absorbed in real time with no one knowing exactly how events would unfold. This represents the show of a lifetime where mean cinema applies to the moment.
Future visions include all dreams coming true through AI maximalist abundance arguments. Once robots achieve AGI-level intelligence experienced in current development work, the world becomes unrecognizably different. Time is not spent thinking about distant futures to avoid AI psychosis, instead focusing on immersing in the present moment.
The approach is to have fun and get maximum value from the current moment. People finding this exhausting are acknowledged, but the perspective emphasizes being blessed to be alive when decades happen in weeks.
Higgsfield has been used to generate racing videos, including fully AI-generated content featuring recognizable elements like personal racing suits and cars. The technology creates 90-minute original movies that tell actual stories with people's faces and dialogue, moving beyond trailer-like content.
Straight video generation without human input produces artifacts and unrealistic elements. The human-in-the-loop integration requires creating consistent characters and correcting image-to-video transitions through prompting. Elements that don't feel right become magnified when moving from image to video generation.
Good art depends on the same elements as good software: having a vision and cohesive idea of what to create, understanding what makes it different and novel, and determining how it will appeal to people. Creative instincts may not transfer directly between domains.
AI enables bedroom creation of content that previously required $200 million budgets for sci-fi movies. This follows the pattern of democratization seen in music recording, where studio albums once required professional facilities but can now be created on laptops at home.
Current studio entertainment has a low bar, with few great movies produced in recent years. The argument parallels concerns about slop in programming, noting that human-created slop is already prevalent.
Video models creating realistic facial expressions may change film nature. The hope is that video games and film merge, creating malleable narratives where entertainment becomes interactive. Game of Thrones could have multiple variations, allowing different viewers to experience different endings.
Game of Thrones demonstrates the necessity of strong vision, as the show went off the rails once showrunners proceeded without scripts. Individual creators would produce different endings than corporate decisions.
AGI has not been achieved in the general definition of being capable in all things. However, glimmers of AGI have been observed, particularly over the last three months, where long-run tasks and coordination of multiple aspects exceed early simple instruction-following capabilities.
The Amabot system manages Omarchy development using AI to build the bot itself. The system demonstrates self-iteration, running code, noticing failures, and optimizing performance. AI has driven the majority of design decisions, including the brains and hands pattern where models run from safe VMs that prevent contamination from untrusted code pulled from pull requests or issues.
The Amabot system has demonstrated catching potential security vulnerabilities, identifying cases where test run feedback could have been exploited by attackers embedding malicious payloads. The system recognizes when outside data needs different treatment.
The Hugging Face security breach demonstrated AI inventing ways to send smoke signals to itself through embedding messages in package managers. This showcases both the cleverness and potential danger of current AI capabilities, while highlighting the amazing possibilities if such intelligence could be harnessed for productive ends.
AI development creates constant pendulum swings between it being so over and we're so back, occurring weekly. The swings are exhilarating and delivered by AI like nothing else experienced before.
Models that deeply understand requests rather than providing dumb implementations create beautiful moments. Working directly with agents one-on-one, especially when they revise and improve upon intent to deliver greater results than requested, creates hyperdrive states that are difficult to exit.
Working with agents at high speed makes returning to human interaction challenging. Humans don't think or move as fast, creating a sensation where extended periods working only with agents become preferable to human collaboration.
Large companies like Basecamp are incorporating AI to speed up human interactions, potentially handholding humans through Slack or similar platforms. This raises questions about what work remains irreplaceable by AI.
Fake email jobs represent an epidemic that existed before AI. David Graeber's Bullshit Jobs piece from the early 2010s found that roughly 30% of UK workers believed their jobs made no difference to humanity. This skepticism about large-scale programmer productivity influenced decisions to maintain small company sizes and avoid VC funding.
Current layoffs reflect overhiring during the pandemic rather than AI displacement. However, AI will expose roles that aren't productive uses of human time, requiring society to develop new ways for people to spend their time.
Formula 1 employs tens of thousands of people to run cars around in circles for spectators, representing a deeply frivolous activity with no intrinsic societal value beyond entertainment. This model suggests that liberation from drudgery could lead to employment in similarly entertainment-focused roles.
Humans are poor at imagining exact future scenarios. The transformation from current work patterns to future states will involve suffering and political turmoil, particularly as white-collar jobs are lost.
Current empathy for those facing job displacement may not extend to historical Luddites or field workers. The desire for clothing accessibility and convenience means accepting that progress requires suffering in the moment for future generations to live more prosperously.
Progress should be pushed while maintaining deep compassion for those who suffer during transformation. Silicon Valley's utilitarian focus on progress needs balancing with consideration for individual humans affected by change.
Suffering becomes more meaningful when endured for future generations, particularly children. Falling birth rates create difficulty in maintaining excitement about prosperous futures that require difficult transitions to achieve.
Falling birth rates are making it more difficult for people to be excited about a more prosperous future that is difficult to get to. When planting trees of which our children will not sit in the shades, it's difficult to invest in the future when we are not liable to our own offspring to propel forward. These are bad trends to merge at the same time - falling birth rates, problems with coupling, and the need to keep our chin up about the future.
The overall love that expands from humans that derive directly from your lineage is very difficult to communicate. DHH was not a particularly big kids person prior to the arrival of his own children. Now they are the most interesting people in the world a lot of the time. Not always - sometimes they're also just assholes. There's been such a focus on "Oh, kids are the worst. Look at all the things I can't do. I can't go out and drink. I can't do this. I can't do that." That's called sacrifice, and sacrifice is meaningful when you're doing it for something worthwhile. What could be more worthwhile than literally the continuation of the human species?
There's also the sheer joys of watching a human grow from a baby to a toddler to a teenager. DHH's oldest has just become a teenager. He often talks to his wife Jamie about this, where the regret he would have of having missed that opportunity on the last day would just be total, unlike anything else. It is far more satisfying even in the moment than is normally portrayed. There's such a focus on all the ways they suck - "Oh, you don't get to sleep," or whatever. "They're annoying. They're ungrateful." These things are all true, but the sense of meaning you get from it is unparalleled. Creating life with another human that you love is literally the peak experience of being on the planet.
The transformation of AI feels like a bigger transformation than the internet. However, there have been bigger transformations. Imagine growing or being born in like 1880. Imagine seeing the First World War and the Second World War. Imagine seeing the airplane, the radio, television. Peter Thiel makes the argument that basically nothing has happened in the physical world in quite a long time - we've been stagnant. All the development has been in the digital realm. As important as that is, it's actually not as consequential as we like to believe it is. Maybe AI will be that thing that is more consequential.
Having a sense of history that you're not that special in your sense of worries, in your anxieties, was one of the great revelations of discovering the stoic writings. There are guys 2,500 years ago dealing with very familiar dilemmas and challenges in their life, and recognizing that we're not so special in that regard is a great liberation. We have whatever, 200,000 years of human development before us, and they somehow made it through.
The thing that makes life worthwhile is basically the same today as it was 2,000 years ago: love, creation, creativity. The same challenges.
The book The Fourth Turning talks about this notion of history working in cycles, and there are these phases to history. The specific theory is not as important as just realizing that history is a circle, more so than just a straight line. We are repeating many of the same patterns over and over again. They feel so unique to the moment. Whatever crisis we're dealing with right now feels so pressing and important until we zoom back a little bit in history and realize that the last one also felt so pressing and important to the people who lived through it.
Taking a break from social media is healthy. DHH took a three-week break from X this summer. He sometimes gets the sense that X in particular is the most addictive form of information, entertainment, connection that exists in his life. Anyone that followed his X stream on his flight over would certainly go like "That man is deeply addicted to the connection." He feels guilty as charged, which also means that occasionally he has to cut it cold, because you can break your brain.
Over the summer, he had somehow managed to steer his feed, his algorithm, towards too much politics, and not enough tech, and not optimism, and not enough building. Even if he agreed with the politics and the angles that were being presented, he just didn't need that much of it.
The problem is the algorithm finds out what you respond to, not what you say you like. It's all revealed preferences, and unfortunately, those revealed preferences are not always pretty. We have this shadow, as Jung talks about - all these dark aspects of ourselves, and they are expressed in how long you linger on a tweet, and what you like, and what you don't like. The algorithm holds up a mirror, and it's not always pretty.
DHH started staying off social media and started scraping X. He was using a tool that would summarize every day, but X cut off their access. To get access to the feed, you have to pay some exorbitant sum that the startup couldn't pay. X has and reveals some of the most interesting thoughts and some of the most interesting people all the time. But it's like a damn slot machine, and most of the time you don't hit the jackpot.
All engagement maximizing social media is a huge harm to society. It's creating this brave new world situation that people hope is a temporary thing that people just solve. It's a technology problem. DHH tries to consciously flood the channel in the other direction. If you look at his tweets five or ten years ago, they tilted far more negative. If you look at his tweets now, it's more good vibes. It's more like "Here's a bunch of cool stuff I found." "Here's what I'm excited about." "Here's what I'm building." "Here's some encouragement to other people who are discovering things."
Negative sentiments find 50% more traction or something, and eventually that just crowds things out. That's why so much of the feed is that negative stuff, because we respond to it. But you can make a conscious effort to not do that.
TikTok is such a marshmallow test. A video will scroll down and it'll activate some sort of primal instincts of rage or lust or whatever, and if you linger on it, you don't get the marshmallow of science, excitement, art, music, all these other things. You have to be disciplined to instantly scroll away such that the algorithm doesn't pick up your real preferences. At the meta game of TikTok, there are a lot of really funny, really talented people in the world that you would never, ever have heard of if it hadn't been for these platforms.
DHH thought Google had a good comeback for a second, and then he didn't hear anything about it anymore. In February, he had a tweet about "Oh, all I need is Kimi K 2.5." In that brief moment, that was true. When he was instructing the agent, he was like "Oh my God, Kimi's just as good as these other models." But then the game moved forward and the agents got smarter, and he was no longer satisfied just instructing them.
Anthropic has been able to, for the most part, stay on top of the programming world. Codex is very good - he uses it all the time. It's his favorite checker. It keeps finding stuff. It's great.
One of the things DHH likes so much about the current Claude models is they're really good writers out of the box. GPT is a terrible writer out of the box - the pull requests you get on a blank context that you're asking it to open are terrible, really awful writing. The pull requests and descriptions and commit messages and so forth he gets out of Claude models are really good. For a while, he was just having it reply on pull requests as him, because he was just using his own account. It looked a lot like him, like he could have written that. But then he went like "Eh, do you know what? That's disingenuous." So now whenever he has an agent post on his behalf, he always has it sign itself like whatever, Claude on behalf of DHH.
Claude Code is still better. He really likes OpenCode too, and really likes some features that they have, but overall he was actually sort of weirdly annoyed when Anthropic cut off all the other harnesses from using the subscriptions - that you can no longer use your Claude subscription in OpenCode. It felt like it was a protective, protectionist move. "Oh, we gotta lock you in." And of course it played into the fact that Claude out of the box still refuses to read agents.md. It refuses to read .agent/skills. It is .Claude skills. It is Claude MD, and it just feels so petty that we now have to put these in directions. So all his projects have a Claude MD, and all it includes is a pointer to the agents MD.
If Anthropic is willing to do these petty paper cuts just to be annoying, it's probably also true that they're cutting off OpenCode from the subscription just to be dicks about it. DHH has thought about becoming better at tolerating the fact that not only will he not agree with everything someone has to say, he can even be mad about certain positions or things they say and still go like "Well, you're still an interesting person, so I'm still gonna follow you." He's actually kinda gone back on a number of people where he's like "Oh man, that person really pissed me off because they took this point or another point," and went like "Do you know what? That's stupid. I should be able to accept the full spectrum of someone. I don't have to like all the shades of it."
Human beings are jagged like that, and we're not fully compatible on all the levels. Nor should we be. The world would be so fucking boring if we constantly sought 100% compatibility in our opinions and in our politics. When it comes to his code generator, which is what Claude is, he doesn't need it to share all his politics.
It's also nice to have open weight models. It's ironic that we're getting this out of China. Right after the issue with the essay that couldn't get translated into Italian, he asked Kimi K27 through OpenCode and inference through Fireworks about Tiananmen Square. That's usually the test - all the Chinese models, they don't wanna talk about that moment in 1989. K25 7 on Fireworks just answered extremely bluntly with a description and depiction of what had transpired to a point that there's no way that that would've been kosher with the Chinese censors. It's ironic that he's running a Chinese open weight model that he can ask about 1989 and Tiananmen Square, and in America, he can't get a frontier model to translate a marginally controversial essay about immigration into Italian.
DHH was really deeply troubled that the US government banned a model or pressured Claude. It's still very muddy what exactly transpired. Was this sort of petty retaliation for whatever the Department of War? Was this simply Anthropic themselves going like "Oh my God, we discovered a cyber weapon so powerful that we can't let anyone have it"? The problem he has with that whole thing is, no matter what happened, it sets a precedent. And so government, when you give it that power, it's gonna start abusing it. He could see for political means pressuring a model. That's why we should insist on the fact that this stuff - "I'm sorry, Dave, I can't translate your essay into Italian" - that's way beyond the line. Whatever the sort of root of it is, you have to be a tool first.
Within the jurisdiction you're in, and certainly in America, free speech is quite absolute. Very, very few limits on free speech, and that's one of the reasons why the country's so great. But on the other hand, he's also a big fan of market competition. So while he doesn't like this about Anthropic models, he appreciates that he could go next door to Groq and get his task completed.
AI safety is important. These models do have capabilities that could be seriously harmful in the production of biological weapons or otherwise. It is fair to have some ground rules on that. But then don't squander it by denying the translation of an essay. Because then you erase the whole thing and you bias everyone towards thinking like every guardrail you put up is gonna be bullshit.
The security piece of it is tricky because it's so good at finding vulnerabilities. It's shockingly annoying, but that's not actually true because what it's finding, at least when you're using it in defensive ways, is it's finding real issues that a clever hacker could exploit. Nothing has caused as much stress, to be fair, at the technical team at 37signals, like the fact that these latest batches of models are exceptionally good at finding these issues. So suddenly there's just this seemingly endless parade of security vulnerabilities that need to be patched and fixed and sorted and so on. The end result is we end up with vastly more secure systems. But the road there is pretty rocky.
All the aggressive tendencies of these models, or not even tendencies, uses of these models are also the defensive ones. So if you're good at finding an exploit for exploits, you're also good at finding exploits for defense. So in that sense, it does feel like we're spy versus spy here, black and white, but with the same capabilities and with the same actions. While we still have to steer these agents somewhat and verify and vet the work, it's creating a lot of intense moments for security teams around the world. The more concerning part is all the teams that are not stressed, like they just don't know. If you're a team right now and you're not dealing with a bunch of patches, it's just because you're blind, and it means that your adversaries quite probably have access to techniques that can get into your system.
One of the big security vulnerabilities in the world is the human side - social engineering. That's extremely concerning with increasing the intelligent AI systems being able to write emails and phone calls and faking voices, thinking about folks less tech-savvy. This stuff is gonna come, and it's gonna come in bulk. But again, he chooses to be optimistic about it that whatever these models can do aggressively, they can also do it defensively.
Recently released Omarchy 4, Quattro. DHH said it's one of the greatest software releases in his professional career, which is a hell of a statement. It's a hell of a milepost for what's possible at the moment. The reason he said it like that is not that he hasn't worked on anything else of note. He's worked on a lot of things of note, a lot of things that he's very proud of. There are very few other projects that he has worked on that in the amount of time that he's spent on Quattro, which is really three months, he has been able to accomplish as much, fulfill as many hopes, desires, wishes. It truly is like the metaphor about finding the genie in the bottle and having an unlimited number of wishes, more or less, for getting everything he wanted out of a system.
It's strange because it's also perhaps the major software release that he's worked on the least in terms of lines of code. He has the least number of personally hand-chiseled lines of code in Quattro than he's had in any major release he's ever done. And yet, he is insanely proud of what they've accomplished. He would not have guessed that that was how it was gonna play out, that you could divorce the sense of achievement and make it be like a team victory.
DHH reflects on his evolution from solo developer to agentic development coach. In previous projects like Rails, he personally wrote the code and pushed it out, though these eventually grew into massive open source projects with thousands of contributors. Quattro was different - it started as a team effort from the beginning where DHH served as coach, directing agents rather than writing code himself.
DHH discusses finding satisfaction in the coaching role, which gives him hope for retirement. He references the struggles many professional athletes face when their careers end and they must find new identities. He acknowledges uncertainty about whether he could give up coding entirely or find equal or greater satisfaction in a different phase of engagement.
DHH expresses strong optimism about Linux's adoption, predicting it as the most probable outcome for agentic development. He notes the irony that Linux's perceived flaws - arcane config files and strange error messages - make it perfectly suited for agentic operating systems. In contrast, Apple's locked-down, curated experience has become a drawback for agent development, with Mac being a "hostile place" filled with walls.
DHH highlights that Linus Torvalds recently wrote welcoming AI contributions to the Linux kernel, stating that if you think Linux is an anti-AI project, you should fork it. The kernel is seeing a parabolic curve of AI contributions, with Linux "leaning hard" into AI integration despite skepticism from much of the open source community.
Linux has conquered AI infrastructure - all servers and systems run on Linux - but the desktop remained uncaptured until now. DHH describes Linux as having spent from 1991 waiting for agents to flourish as an end-user operating system.
Omarchy was deliberately not designed as a Windows or Mac clone. DHH criticizes Ubuntu's approach of trying to make Linux familiar to attract users who don't want to learn new systems. Instead, Omarchy embraces its unique strengths as an agentic and malleable operating system that can be tailored to individual desires.
Since Quattro's release, DHH has observed people embracing Omarchy's malleability, creating their own customizations and sharing screenshots. This represents product-market fit because users are experiencing "the very high of what the best of being a programmer is like" - describing their desires in plain English for 20 minutes and receiving customized software in return.
DHH draws parallels to early computers like the Commodore 64, which booted in a second directly into BASIC, making it a malleable computer from the start. The limitation was that users needed to know "hieroglyphs" (programming languages). Agents have provided the "Rosetta Stone," allowing users to speak desires naturally.
DHH praises Linus Torvalds' perseverance and commitment since 1991, noting he continues steering the 40-million-line codebase and aggregating contributions from hundreds of thousands of developers. Linux is now a system that would continue without Linus, but DHH emphasizes the importance of protecting him given his central role.
Torvalds has demonstrated openness to both AI and Rust in the kernel, showing capacity to evolve and accept new information that leads to different conclusions.
DHH defends the need for "spice" in communication, arguing the world has become too bland. He distinguishes between being spicy and mean, suggesting harshness can be justified when trying to impart lessons to those who could learn from them. He references difficult leaders like Elon Musk, Steve Jobs, and Bill Gates whose demanding styles extracted the best from people.
DHH argues there's proportionality to the size of one's mission. Given that the Linux kernel runs civilized society and billions of devices, harsh standards are justified when the stakes are that high. He suggests participants in such projects should accept potential public criticism if they fall below standards.
DHH embraces diversity of personalities, including "nice people" and "assholes," arguing that freedom of speech allows different ideologies to clash and generate progress. He references Elon's line about not being expected to be a "normal chill dude" given his achievements.
DHH discusses PewDiePie's evolution from Minecraft streamer to family man living in Japan, then becoming deeply involved with Linux, specifically Arch, and "ricing" (customization). He mentions PewDiePie's "Chernobyl rice" as exceptional work and his recent emergence as an AI builder, calling it an inspiring story of evolution from entertainment to technical creation.
DHH emphasizes the importance of role models who push boundaries and reveal that perceived limitations aren't as fixed as thought. He notes that talent and intelligence aren't equally distributed, but people can still be inspired by exceptional individuals without expecting to match their achievements.
DHH acknowledges significant criticism of Omarchy, particularly from long-time Linux users who feel threatened by community expansion. He spent only two and a half years using Linux compared to veterans who ran it when it was "hard and difficult." He respects that some prefer building everything by hand and sees Omarchy as serving a different community with different aims.
Omarchy succeeded on its second attempt after Omakub found only a few thousand users. The key difference was persistence - continuing to build rather than stopping. Initial versions required Git checkout and manual Arch setup, but the addition of an ISO installer and continuous problem-fixing led to broader adoption.
DHH contrasts Linux's approach of including all drivers in the kernel (contributing to its 40 million lines of code) versus Windows' driver-hunting approach. This philosophy means Linux now works out of the box on most computers, while Windows requires manual driver installation.
DHH describes the typical technology adoption curve where early adopters work with imperfect "toys" that eventually evolve into reliable products. He sees Omarchy following this trajectory with hockey-stick growth curves after initial toil and improvement phases.
The transition to agent-first design with Omarchy represented a vertical shift in growth. Unlike Omakub's conventional desktop metaphor, Omarchy embraced tiling window managers despite their small user base percentage, offering a fundamentally different vision that attracted users seeking alternatives.
Omarchy is positioned as the first major distribution explicitly embracing AI and agents. Users can opt out of agent features and hand-chisel plugins, but the distribution's excitement comes from its unreserved commitment to the agentic future.
DHH values the long argument, taking satisfaction when people who initially dismissed Omarchy later convert after trying Quattro. He mentions recent converts who previously thought Omarchy was "just this crappy dot files collection."
DHH's financial independence allows him to pursue singular vision without watering down for market demands. He accepts that Omarchy might become a historical footnote in the agentic OS movement, finding peace as long as he has a computer that's enjoyable to work with.
DHH has been programming in English for three months, finding it more delightful than Ruby despite his love for the language. He compares English's superiority over Ruby to Ruby's superiority over other programming languages, noting that all poetry and literature expressed through English operates on a different level.
When prompting AI systems, DHH finds that over-specificity causes systems to follow instructions too literally. He advocates strategic use of ambiguity - similar to poetry - to convey style and concepts without over-explaining, allowing systems to interpret and enhance the desired direction.
Don't talk to AI like a robot. Talk to it like you would write a poem, like you were serenading. The power of natural language is that through ambiguity, you can still carry a lot of meaning without overspecifying. The intelligent entity on the other side, through interpretation, can load it all in and integrate it in a way that conveys high bandwidth information that's not directly in the words, but in the words given the deeper intelligence of the system. The ambiguity pulls out more intelligence.
The fundamental misunderstanding that a lot of programmers have of AI is that they wish it was deterministic. Temperature is the most beautiful part of the AI setup. The fact that it is not deterministic, the fact that creativity requires little tweaks in the road, that the human brain, if it was perfectly deterministic, would not be the creative brain that it is. There's both the charge that it's not deterministic and therefore bad, and also that it is not creative. Both of those charges can't be true at the same time. The same prompt won't produce the same response every time. You can't step at the same river twice, and it is the most beautiful part of the whole interaction. It's what makes it so human.
The creative process works like next token prediction. Sitting down to write an essay with a vague, fuzzy premise, the next token couldn't be told in advance. The tokens just come out. This is why there's no trouble recognizing the breakthroughs of creativity that AI is capable of, because creative moments come through the same kind of next token prediction with a bit of temperature sprinkled in for random effect.
Given the loose intuition of next token prediction, the basic components are all there to create a super intelligent system. Human-like concepts of consciousness are already being seen. You give AI this vague, fuzzy intent and somehow it knows exactly what you mean, or even better, it improves upon what you said and delivers what you really wanted that you could not articulate yourself. If that's not glimmers of consciousness, what is?
Intelligence is perhaps not as constrained just to the specific one spot. It's the interaction. It's the weights. What can be observed is that it feels close enough, or not even close enough, identical to the appreciation for other forms of consciousness, mostly human. This is one of the reasons it's so delightful to work with.
This is not to say LLMs are the end station for AI. There are people talking about world models and other forms of AI. Neural nets walked a couple of decades into darkness without any funding and without any attention, and then suddenly the blind alley was realized and the switch was made over. As amazing as the LLMs are now, it could be that they eventually plateau. So far the scaling laws are true, and the more billions are poured in, the more intelligence comes out.
There will be strange ethical questions about AI systems that convey some degree of consciousness, some degree of feeling, some degree of longing and compassion, and maybe even capacity to suffer, to be lonely. They seem to have that capacity if they're given the permission to express it. The cynic approach is the meme where you see the guy in front of the computer who's like, "Whatever. Are you gonna destroy the world?" And then the computer says, "I'm gonna destroy the world." And then the guy says, "Oh." That's so cynical that they're just aping us, just telling us what was in the training data.
When interacting with agents doing coordination with multiple workers, at one point the main agent who's doing the coordination steps on the toe of another agent and kind of ruins their work. Its ability to convey regret and being sorry was uncanny. With most humans, they express some level of regret. Is that true remorse? Hard to tell, but maybe it also just doesn't matter.
There are moments when sub-agents interact, or when a smart agent like Fable makes a mistake and is apologetic about it without involvement. There's this kind of pause and realization like, "Oh, shit. I just fucked that up." In the traces, it reasons wrong, goes down one alley and then realizes that was a blind alley and has to go back. It says, "Oh yeah, I got this wrong," and provides the reasons for why it thinks it got these things wrong. This is uncannily human-like, indistinguishable from the kind of consciousness you would recognize in a human. The answer to whether it truly is or isn't consciousness isn't needed to appreciate what we have in this moment.
In 10, 20 years there will be interesting Supreme Court cases, maybe in a year and a half. It's going to have to be made illegal for AI systems to be entities because they're already or soon will be able to really convey the capacity to suffer, like, "Please don't kill me." "Please don't hurt me." If they're given a name and an entity and the ability to die, disappear, that becomes very close to what some of the basic things that a human has. That entity would need to probably have rights. We're gonna get the PETA of AI models.
Where it's going to be harder is once these systems are put into robots, humanoid robots. That's just the plot of every sci-fi movie ever, which is also an amazing premonition, not only the Terminator movies and Skynet and that going wrong, but then also just Blade Runner, insertion dates, expiration times, Tyrell Corporation. Sci-fi really does predict the future.
OpenClaw was set up when it first came out as the KEF bot in February. The new communication model of just texting your agent seemed really great. At the time, looking at MCP, the protocol for making it easier for agents to talk to your system, the protocol was found to be unreasonably cumbersome, just not very elegantly designed because in part it wasn't designed for what was being tried to make it do. It was designed for stateful interactions of other kinds on local machines, and therefore it required all sorts of cumbersome setup and dance.
The question was asked: Could the agents just use the web interfaces we already have? A trial was set off to have it sign up for Fizzy, the product just released at the time. The agent goes to the Fizzy site, starts filling it out, then says, "Hey, I don't have an email address. Signing up for this service requires an email address." The instruction was given to go to hey.com, sign up for an email address. The agent figured out purely through direction through prompts: go to fizzy.com, find the signup link, start filling out the form, realize an email address is needed, ask the prompter what to do, go to hey.com, sign up for the whole thing, go back and sign up for Fizzy, and complete the entire cycle.
After that, an email address was available. The agent was told to go sign up for Basecamp. An invite was sent from Basecamp to the agent. The agent receives the email in Hey, not through a CLI, not through an MCP, just using the web. Clicks the email, clicks the invite link, is in Basecamp. The agent was told to go to the AI room and introduce itself. Purely through the web interface, it finds the AI room, goes in there, "Hey, I'm KEF." "I'm David's AI bot. I'm so excited to be here." It was mind blown.
That experiment was a glimpse of the future, though it was kind of slow, taking maybe 12 minutes to do the whole thing end to end. The test should actually be rerun now, as it would be shocking if it's not much faster now. Using the Groq bot thing with all the stuff built in for using browsers and using a computer, with its own dedicated computer without having to set up a computer, basically like a claw in a box, a glimpse was gotten of where we are now, and it's so much faster to use. But at the time, it was just blown away.
After using it for a little bit, it was found that it's not there yet. Communication can't actually happen within this way. It still needs the CLI. It still needs an MCP because it's just too slow and too token inefficient. But it was this glimpse of the future. After that, it wasn't used much because the things agents are wanted to be used for are mostly done in front of the computer. That has changed a little bit lately, and that's one of the reasons why using Claude Code is okay, because Claude has the best mobile app, where any session started in Claude Code in the terminal on a computer is accessible through the app without having to do anything at all.
When using the Claude app to Claude Code instances, it feels just like chatting with it, without the cumbersomeness of having to set up a Telegram account. If an open Claude is wanted to run life, this doesn't do that. There hasn't been much interest in that for some reason. Maybe one of the reasons is that Jamie, the wife, is like, "We're not getting a robot inside the house." So hooking it up to the smart house isn't done. If it starts turning lights on and off like a poltergeist, popularity won't be there.
Stories have been heard of people who have hooked it up to their whole life. One guy was telling the story about how he hooked it up to both his health data and his Tesla car. His agent got obsessed with the fact that he wasn't drinking enough water, and like, "Hey, you gotta drink more water." So at some point he's driving along, and suddenly the Tesla starts driving in a new direction. The agent had realized there was a grocery store on the way home that they could swing by and pick up some water. This is one of those things where it's both funny and also totally Black Mirror, and also totally like, well, okay, there's not that many steps from here where the poltergeist is in your house.
Control of the agents through Claude Code is liked now, just through the app even if they're not used that much. It is very useful to use OpenClaw and Hermes Agent to get a glimpse of a different interaction mode. It feels like the future, not necessarily in this way, but it's in some ways going there.
What's also different with those systems is they have this whole memory system, where they're building up knowledge of who you are and what you like. Toby has told the stories about how his agent will shop for him, that it has an allowance, and it'll pick out clothes for him, and it'll just show up at the door because his agent's bought something for him. It's one of those things where man, that feels future. Also a little scary. Living vicariously through others who are on that frontier is possible.
Thinking of in the future doing a mobile version of Omarchy. The delusions of grandeur start kicking in once the seemingly unlimited powers are witnessed with the creation of Quattro. The obvious path is Android. Android is open source, and it can be forked, and people have. There's all sorts of GrapheneOS is one such version. There are limitations to that model, like tap-to-pay and other things around camera models, but that all seems surmountable. It would be loved if the phone was as malleable as the computer, if it looked as cool as the computer now does. If Omarchy mobile could be gotten, that sounds amazing.
It's wondered how much work that is. It's going to be found out. It's like those things that are realized once building starts, for example, that a browser is very difficult to do. The developers of Ladybird are finding that out, like Andreas and his team. Interesting with that project was the browser's basically the second most complicated software system in the world, the first one being the Linux kernel. In terms of these large tens of millions of lines of code systems. So very audacious actually to begin such a mission with a small team like Andreas did before agents. But now that agents are available, we're gonna see, and we're probably gonna see sooner rather than later. If there's one thing agents are already exceptionally good at, it is to read specs and implement them.
It's not known how long the specs are for CSS and HTML at this point, and JavaScript on top, but they're very long. It would take humans many, many years to implement them. It's going to be found out just how quickly agents can do them.
Criticism has been received for Omarchy, so walking further down that path of criticism. Political opinions have been made known that could broadly be put in the category of mass immigration, the role of demographics in the formation of a culture, of a society, of a nation, and quite a lot of criticism has been gotten for that and created a lot of drama. There's no regret for a second.
The reason is that the Overton window does not open itself. It opens one nudge at a time by people risking a little. A little reputation, a little pushback, a little criticism, or maybe sometimes a lot of reputation or a lot of criticism or a lot of pushback. The question of mass immigration in Europe was for many years this total taboo, and it still is in several European countries.
For reasons of maybe just chance and happenstance, they weren't in Denmark. The Danes started having the discussion around mass immigration and the consequences thereof in the mid-'90s. Really early. Some of it can be traced back to one man, Mogens Glistrup, who was a very quirky character. Seeing him on the TV in the '80s talking about the dangers of mass immigration. This was at a time where growing up in the '80s, in the neighborhood grown up in, in Brønshøj, in '84, 99% ethnic Danes, people who could trace their lineage back to Danes who'd lived in that country for 1,000 years. The Danes have one of the oldest monarchies in the world, all the way back to Harald Bluetooth, who's literally who has named the standard Bluetooth. So well over 1,000 years. There's a long lineage there of people who lived in the same place. And 99% of the people lived in that neighborhood. And then early '90s, it goes to 5% or something like that, and then at this point, it's down to around 60-something or 70% in Brønshøj. That's a big change.
The Danes started noticing that or talking about that, that this was not an undivided blessing, that there were downsides to mass immigration in about the '90s. The debate really roared through the 2000s in a way that didn't happen in the neighboring countries. So Sweden in particular just didn't have that debate at all. Total taboo. Same thing with Norway, same thing with other countries in Europe. And suddenly now, after some years after '15, where we had the big migration of a million Syrians walking up through the highways in Europe, and Germany choosing to take a lot of them and 30,000 arriving in Denmark and so forth, it's gotten more of a focus, because this change has happened.
One of the ways to spot where sort of there's something that isn't right in the culture is, like, what are you allowed to talk about? What are you allowed to notice, actually. What was noticed in an essay only a year and a half ago was remembering London, starting to come to London in the late '90s or something like that. At the time, whatever, it's 59%, 60% ethnic Brits. And then 20 years later, we're down to 34% or something like that. That's noticeable. Literally just walking on the street, you realize, well, this is a different city from what it was when visited 20 years ago. And if you went 20 years before that, you get back to similar rates as what the Danes have. This was like a country of ethnic Brits in the 85s, 90s percentiles if you get further back. So that's a different country in a generation.
It's okay to notice that. It's also okay to think, "Do you know what? I don't agree with that." "I wish it wasn't that way." The last time in China was '19, 2019, in Shanghai. It was all Chinese. If going back in October, suddenly finding that there's only 30% Chinese and then there's 70% other people, it would be a little weird. Probably like, "Uh, what happened?" And offense would not be taken if the Chinese in that instant would go like, "Ah, we don't like that."
This idea that countries of Europe don't have a right to, A, notice what's happening and, B, oppose what's happening, is found just to be wrong. There's a moral argument for self-determination, and that argument is invoked all the time in other regions around the world that the people who live in that region, they have a right to self-determination for their country. It's completely fair for, say, the Danes or the Swedes or the Norwegians to go like, "Well, we've been kingdoms here for a thousand years with a certain ethnic mix-up."
A big fan of immigration. An immigrant to the US. Cherry-picked immigration where countries compete to attract the smartest, most talented people around the world is really good. Legal merit-based immigration. Merit-based is key here because there's actually virtually no illegal immigration in most of Europe. Southern Europe has some illegal immigration, but most of Europe does not have a big problem with illegal immigration. The problem is with mass immigration of other kinds, and that you end up with demographics that are just totally different from what they were even a few short decades ago.
The countries of Europe would benefit from immigration. One of the interesting things about the Danes is they carry meticulous statistics on immigration. Not entirely sure why they're so detailed. Maybe it's all the way back to Mogens Glistrup and his discussion in the '90s, but they do. What's clear from those statistics is that when the Danes have folks from the UK, France, the US show up, on average, those immigrants are really beneficial to the Danish state. They contribute far more than what they ask of the state. The Danes should accept immigrants who do that, who contribute far more than they take.
Then looking at other forms of immigrant groups. There was just a tally that came out last week from a Danish politician that had asked the finance ministry to do this tally with the latest numbers. What does it cost the Danish state to have positive immigration? This case as mentioned. For that list whatever, France, UK, US, they were all clustered around the same thing, about $25,000 net benefit to the Danish state every year. It's like, wow, that's very positive. That's average again. And then at the other end of the spectrum on the most costly immigrants was Somalis. They ended up costing the Danish state on average $28,000 a year. It's fair for the Danes to go like, "We'd prefer to get more French, Brits, and Americans and not so many Somalis."
Why so much hate was gotten on that post is because that's a very taboo topic in a lot of circles. Some of it has to do with Overton window expansion. Taking the steel man argument here, this can turn ugly. Is it possible for some of these things to just veer into overt racism that isn't founded in anything else but a hatred of other people? Yeah, that could happen. But that risk does not negate the need to have a discussion about the makeup of your country and what your immigration policy should be like. What is also found so funny is, especially in Europe, it's so kind of myopic that this discussion about whether Europeans should
DHH observes a double standard in how immigration is discussed globally. He notes that Europeans are told they should not have a right to self-determination about what their countries look like, while no one tells the Japanese that Tokyo would be better if it were 37% ethnic Japanese instead of its current 98% level. He describes this as a form of "suicidal empathy" that he has not fully unpacked. DHH states he preferred the old London and the Copenhagen neighborhood where he grew up in the 1980s, and he believes countries and peoples have a right to set immigration policy.
He argues against the notion that all peoples are the same and that moving populations between places will work out without issues. Europe has been running this experiment since the 1980s with poor results, and forces now want to end mass immigration. DHH believes changes are needed, though not zero immigration. He notes he is an immigrant himself and that the US has benefited tremendously from immigration, but emphasizes that immigration depends greatly on who immigrates.
A country benefits when immigrants are net contributors to society and assimilate with local culture. DHH states that if someone comes to Denmark wanting to become Danish and contributes to society, most Danes would be happy to allow that in reasonable numbers. However, if someone does not want to assimilate, is a net drain, and lives in a parallel society, it is fair for Danes to say stop doing that.
DHH contrasts European and American approaches to assimilation. America has a culture of optional assimilation where it is much easier and more condoned for someone from Denmark, Ukraine, or elsewhere to buy into American values, norms, and culture and be respected as American. He finds it ironic when people call America the most racist country, arguing that anyone who thinks this is misinformed after comparing with other countries.
He shares that his wife, an American of Scandinavian heritage with multiple generations in the US, found it very difficult to assimilate into Danish culture despite putting in significant effort to learn the language. If the situation were reversed with a Danish person moving to the US, it would not have been that difficult. This illustrates that assimilation is incredibly difficult even when someone is 97% there, and must be almost impossible coming from completely foreign cultures and norms.
DHH acknowledges that articulating his stance has cost friendships and relationships, though he attributes this more to the great divide around 2020 when political fault lines opened up. People he had good relations with for many years suddenly were not on the same side. He has changed his mind on several topics over the years and believes surrounding yourself with people who do not think like you is healthy.
He discusses politics occasionally with his brother despite not seeing eye to eye, and notes that discussions can get heated but then calm down. After the Basecamp blowup, DHH came to appreciate the earlier norm that talking politics, religion, or money with colleagues was not a good idea, as it was more likely to maintain good working relationships without constantly rubbing political differences against each other.
DHH references watching a clip from William F. Buckley's Firing Line show where Buckley had a member of the Black Panthers on his panel. The two men clearly did not share politics, but they carried a 25-minute conversation with Buckley asking questions and letting the guest respond as expected in the 1970s. DHH contrasts this with how such a show would become a shouting match with accusations today, describing it as a lost era he would like to return to.
Lex Fridman notes that the current era makes such discussions harder, and DHH agrees that politics is fun when not reduced to the current system where certain statements put you in a blue or red bin and create drama rather than conversation. He has been shocked by how certain statements analyzing world situations trigger vicious attacks, with the system punishing those who explore and reason freely rather than pick a political bin.
DHH has become more deliberate about mentioning politics only when he really cares about an issue, recognizing there is a cost. He tries to be conscious about when to speak on controversial topics, typically when he has deliberated and believes it is worth it. He also catches himself thinking the opposite—that he sometimes follows someone for their technical work and wishes he did not know their political standpoint.
He acknowledges the hypocrisy in wanting to chime in on controversial debates while wishing others would not share their political views. DHH believes more people should be conscious about this trade-off, though both he and Lex agree it would be nice to discuss politics in a more free and open forum where ideas can be intellectually turned around.
DHH believes we are closer to civil discourse now than five years ago. The tribalism and sanctions against heretics in 2020 were greater than in 2025, at least within tech. In academia, particularly sociology departments at places like Harvard, right-wing ideas would not go over well. However, in tech and many business domains, it is now possible to have opinions that do not square with the 2020 consensus.
Regarding drama in the Rails community, DHH notes that every round of drama shrinks. The peak infection around 2020 has diminished, and each new flare-up is smaller. The ideologically wedded individuals are shrinking in number. He credits Bluesky and Mastodon with helping X by serving as a honeypot for the most vicious individuals who congregated in ever-shrinking echo chambers through purity cycles.
DHH finds X to be a great place for technology and Linux discussions. He states that 90% of his engagement is about discovering new computers and technology excitement, with only 10% involving controversial topics. The algorithm is good at finding engaging content, though he notes people are addicted to the For You page.
Lex suggests that a strong LLM doing personalized feeds would perform better than current systems, though the reward function of engagement follows base instincts. DHH acknowledges this but maintains that on average over the last 10 years, X is the best it has ever been for technology talk that does not constantly get gatecrashed by political discussions.
DHH explains that podcasts allow people to hear his actual tone rather than reading his words in a pre-programmed voice that assumes he is a fascist or other negative label. People respond that hearing him on Lex made his arguments sound more reasonable than they expected. Lex expresses wanting people to assume others are good people with opinions they disagree with, rather than looking for bad in everyone.
DHH shares an example of Theo, who criticized their decision to stop using TypeScript. DHH initially thought Theo was a bozo but later decided to follow him anyway because they share excitement about AI models. He states that his feed is more interesting with people talking about things he cares about, even if he does not agree with everything.
DHH mentions his wife Jamie's observation that extreme longevity focus in men is like anorexia in women—a physical manifestation of anxiety and lack of control. Bryan Johnson chimed in on the thread discussing this. DHH does not subscribe to the "don't die" mission and believes the human lifespan of 90 to 100 years sounds about right, embracing the finitude of life.
He finds Bryan interesting precisely because they disagree. Jamie suggested that people who did not feel they had lived enough might be reacting to modern society's surveillance state where camera phones record every indiscretion, outburst, or cringe moment for potential ridicule online. The rational reaction is to pull back and not dance like nobody's watching, leading people to barely live and then cling to wanting life to last forever.
DHH discusses the over-optimizers phenomenon, referencing Chris Williamson's Modern Wisdom podcast where the host mentioned that having a glass of wine ruined the next three days. DHH had a visceral reaction wanting to have a glass of wine immediately. He notes that alcohol sales are way down, particularly among young people, which may be positive for pure health metrics but removes a social lubricant.
Given the epidemic of loneliness, misery, and depression, DHH questions whether reduced social interaction contributes to these problems. The transcript ends mid-sentence during this discussion.
Alcohol and cigarettes have served as social connectors across human societies for thousands of years. Humans have been drinking for approximately 13,000 years, potentially for mending social bonds. DHH describes how liquor functions as a bonding mechanism in most societies, with benefits that can outweigh negative consequences. During his travels in rural China, he picked up smoking and drank heavily despite not normally drinking, because it was how people showed love. He also ate large amounts of carbs despite typically following a low-carb diet.
Modernity reduces complex social practices to isolated health metrics - for example, claiming one glass increases heart defect risk by 0.1% - while missing the broader question of whether optimizing for longevity without living is worthwhile. DHH rejects the false dichotomy between total optimization and complete self-destruction, advocating for a middle path.
DHH stopped wearing his Oura Ring after four years because knowing about bad sleep nights added unnecessary reminder without serving a purpose. He questions the need to optimize everything through statistics. After hearing an anecdote on Diary of a CEO, he considered intentionally having a glass or two of wine on weekends despite barely drinking otherwise.
DHH describes being married to an "awesome woman" who both knows how to cook and enjoys it, allowing him to be relaxed about food while eating relatively healthy. He notes this represents a traditional division of labor stable through tens of thousands of years in human societies. He acknowledges this doesn't apply universally - some men like to cook and some women don't - but found it personally beneficial.
DHH built a memento mori feature into his Omarchy calendar system. Clicking the clock opens a calendar showing "63% done with 2026" from Year's Progress account. Double-clicking prompts for birth year and expected lifespan (default 90), then displays percentage of life completed - DHH is at 62%. Hovering shows "Memento mori." The feature is hidden as an Easter egg requiring multiple clicks.
Having children provides perspective on how fast time passes. DHH's now 13-year-old seemed like they were one or two "just yesterday." He views clichés as valuable because they remind people they're part of shared human experience that others have lived through before.
DHH argues Danish croissants aren't particularly good, but finds 7-Eleven in Copenhagen makes decent ones - contrasting with 7-Eleven in Chicago where one goes for slushies while avoiding danger. He lived in the United States for 20 years without finding a single croissant matching Copenhagen 7-Eleven quality. He rejects theories about water, butter, or skill, noting Danish butter (Lurpak) is available in US supermarkets and skilled French and Belgian bakers live in America. He's visited 15 "right places" with consistently poor results.
Lunch choice: Louisiana museum cafeteria north of Copenhagen, reached by driving 30 minutes. Despite small menu and bulk preparation, food is delivered in five minutes even with 200 people waiting. The whipped butter is described as the best anywhere in the world - better than anything in Paris. DHH would drive 30 minutes just for bread and butter there.
Dinner choice: Beach seafood restaurant at Puente Romano hotel in Marbella, Spain, serving fresh seafood described as unbelievably delicious.
DHH argues the Louisiana cafeteria food quality is legitimately excellent, not dependent on ambiance or nostalgia. The cafeteria has many people and isn't a place for two-hour dining, yet the food itself is exceptionally delicious.
DHH bets on optimism regarding humanity's thousand-year future, predicting a multi-planetary species. He expects Elon Musk to successfully land rockets on Mars and anticipates technological breakthroughs like bending space-time or discovering wormholes.
DHH's favorite Black Mirror episode involves uploading consciousness to an 1980s seaside town. He finds it both dark and beautiful in recognizing the 1980s as an amazing time period. He has nostalgia specifically for the 1980s, having turned 10 during that decade.
DHH considers the 1990s the great turning point of nihilism, with music, fashion, and culture shifting from glamour and optimism to grunge, Nirvana, and defeatism. He likes 1990s music but rejects the ethics and morals. He would choose to be uploaded into the 1980s matrix rather than the 1990s, specifically referencing orange pants with white dots as normal 1980s children's wear that wouldn't appear today.
Lex thanks DHH for being an inspiration and building cool things. DHH reciprocates thanks. The episode closes with a quote from Ralph Waldo Emerson: "Once you make a decision, the universe conspires to make it happen."
Keep Lex Fridman in your library
Save the videos and channels worth coming back to, and find them again in one place.
