What this covers

DHH and Lex Fridman talk through what agents have done to day-to-day programming, and DHH keeps tying it back to his own work. He describes where agents stalled: large, mature codebases like Basecamp and Hey resisted acceleration, and unconstrained vibe coding by designers produced many reasonable pull requests that together wrecked the architecture. He also lays out a security design that keeps the model in a coordinator, apart from a sandboxed VM running untrusted code. The contested claims include that agents already outclass the median open source contributor, and a Shopify study in which agent-reviewed PRs caused fewer production incidents than human-reviewed ones. He is dismissive of maintainers who complain about floods of AI-generated PRs.

The conversation then widens. DHH argues for Jevons Paradox in programming jobs, using ATMs and bank tellers as the precedent. He says established companies are too tuned to the pre-agentic world to pivot. On interface design he cites the Paradox of Choice and the human knack for quick picks among three options. He also notes that Claude refused to translate one of his immigration essays, which he treats as a case against model-level viewpoint gating.

The remaining stretch is social and political. Lex and DHH discuss engagement-driven social media and how algorithms learn from revealed preferences. DHH cites David Graeber on pointless work, applies Chesterton's fence to drinking, and defends Denmark's immigration preferences using a finance ministry tally. He also warns that LLMs could plateau, as neural networks once did.

Sharpest takeaway

DHH argues that agentic AI has moved programming from writing code to steering outcomes, so the human bottleneck is now taste, vision, and product judgment, and an agent-first, malleable operating system like Linux is the likely winner.

  • Since Opus 4.5 in late November 2025, agents went from autocomplete to doing whole tasks, and DHH has stopped hand-writing shipped code.
  • Organizations are bottlenecked on ideas, taste, and human coordination, not implementation speed, so agents must be used directly rather than mediated through teams.
  • Linux's config-file and CLI design, once a drawback, now suits agents, making it the natural base for a user-mutable computer.

The argument · threads15 threads · 129 claims
0.85

Human ideas, taste, and intent, not implementation capacity, now limit productivity as agents take over work.

16 pointscentrality 5/5
  • Because nobody knows what they want until they see it, one should be vague upfront to manifest something, then interact with it; agents make this iterative discovery cheap, whereas detailed upfront specs failed in traditional development.

    “So in the agentic age, you should resist the temptation to be overly specific upfront. Be as vague as you can to manifest something, then interact with the something.”

  • Most organizations do not know what they want or how to improve, so they are bottlenecked on ideas, vision, and taste rather than implementation capacity, and adding agent speed does not help without those.

    “They’re not bottlenecked on implementation. They’re bottlenecked on ideas. They’re bottlenecked on vision. They’re bottlenecked on taste.”

  • To achieve 10x, 100x, or rarely 1000x productivity gains, a developer must interact with agents directly, because routing that bandwidth through another human is too slow.

    “to get that magical 10X, 100X, in a few rare cases, 1000X productivity boost, you have to interact with the agents directly, and you cannot intermediate that bandwidth with another human because it’s simply too slow.”

  • Once agents can implement any feature in five minutes to two hours, the constraint on operating system design shifts from implementation to ambition, which is what produced the feeling of being handed a genie with unlimited wishes.

    “Every feature you’ve ever dreamed of in an operating system, I can deliver them to you, most of them in five minutes, a few in 20, and if we really go hog wild, it’s gonna take me two hours.”

  • When humans work in teams with product managers, designers, VPs, and CTOs all shaping a feature, the bottleneck is rarely implementation but human bandwidth and communication, and that is where productivity dies.

    “As soon as you’re having human teams work together on something, the bottleneck is rarely implementation. It’s human bandwidth and communication.”

  • Humans are good at differential evaluation, quickly picking among three options via gut judgment, but collapse under about 22 options (the paradox of choice), so agent interfaces should present few alternatives.

    “Humans actually fall off a cliff if you give them 22 options. That’s the paradox of choice. But you give them three options, they very quickly within a split second will tell you what they like.”

  • The human is the limiting factor in agent workflows, so DHH is building automated systems that run on a schedule, file PRs and issues, and send a daily email for final human decisions.

    “I found that the human in the loop is the limit here. So I’ve started setting up more automated systems.”

  • Everyone is building their own agent coordination layer, but this will be absorbed by major labs and become standard, much like the JavaScript framework churn that marked earlier paradigm shifts.

    “everyone is building their little gas town, their little agent coordination, their setup. This is all gonna be solved.”

  • Running multiple agents at max human capacity produces constant decision-making with no coasting, which is exhausting in the same way as hard racing, and the current intensity is unsustainable but will fade as automation takes over.

    “when you’re running at max human capacity- … you’re just constantly in a corner.”

  • DHH says he is incredibly excited about AI in programming, with no existential dread, because decades of progress have happened in the last nine months, and he compares it to witnessing the Wright brothers' first flight.

    “And we have seen decades of progress happen in the last nine months.”

  • DHH argues that a CEO who directs programmers to build software is not programming, and by the same logic a vibe coder should not be called a programmer, because programming means understanding constructs like loops, conditions, and variables rather than only producing programs.

    “I don't think most people would call that CEO a programmer, and I don't think we should call the vibe coder a programmer either.”

  • DHH says lines of code is a stupid metric for measuring agent-era output, because people who cite line counts often cannot say what they built, though it does represent output volume.

    “lines of code is a stupid metric in general to measure these things... when everyone talks about the number of lines of code they're writing, and then you ask them, 'What did you build?'”

  • DHH runs about 16 agent threads across four to five mini PCs connected via tailnet and KVM devices managed by Herdr, and says he has maxed his own capacity, so the human operator, not compute, is the limit.

    “I have about 16 threads. That's what I can run... at full acceleration.”

  • DHH describes a progression from pre-agentic autocomplete and chatbot assistance, to agents that still required him to drive, to sub-agents that parallelized work, to a current phase where he states intent and the agent determines the path, making him optional in the implementation loop.

    “I'm not telling it where we're going. I'm telling it the problem I have... I've become optional in the part that produce”

  • DHH uses the image of a programmer from a one-megahertz, 64K machine being teleported to 2026 to argue that constraint-driven heuristics become obsolete, so each era rewards different optimization skills.

    “all their techniques and all their heuristics would simply be wrong... The amount of optimization you have to apply for a one megahertz computer to produce a video game is just very different.”

  • Peter Thiel argues that nothing substantial has changed in the physical world for a long time and that development has concentrated in the digital realm, which DHH says may be less consequential than people believe, though AI could be the exception.

    “Peter Thiel makes this argument, that basically nothing has happened in the physical world in quite a long time. We've been stagnant.”

0.71

Agents now outperform typical contributors, transform software building, and make development cheaper and faster.

16 pointscentrality 4/5
  • Instead of burdening maintainers, a developer can have an agent publish a personal tool to GitHub, write the README, use releases, and act as a more patient and diligent maintainer than the human could be.

    “It’ll actually be a better software maintainer than you could ever be- … because it is far more patient, it is far more diligent in doing the drudgery of managing open source software than you are.”

  • Maintainers should prefer agent-generated pull requests to human ones because rejecting an agent's work carries no social cost, unlike rejecting a contributor's work, which triggers obligation and anxiety.

    “I feel a lot less bad if I just reject it. 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.’”

  • Critics who call vibe coders slob generators should look at the codebases of major companies after thousands of human contributors; those are often awful too.

    “If you’ve looked at behind the scenes of many great companies and what their code bases look like after there’s been 3,000 humans through them, they’re awful.”

  • In the first agentic phase humans originated all ideas and agents executed them, but now models produce ideas that exceed what the human proposed, so critics who call agents stochastic parrots are wrong about recent progress.

    “I have seen things you people wouldn’t believe, ideas coming out of models so great that it makes me humble”

  • Most programmers do not write good bug reports, PR descriptions, comments, or tests, so the median human contribution to an average open source project is already outclassed by agents that follow those instructions diligently.

    “if you take the median programmer and their pull requests towards an average open source project, they’re already getting outclassed by agents.”

  • Claude Code's system prompt for Opus 5 shrank by 80% because the agent needed far less human instruction and was being degraded by overly prescriptive human guidance, much like a programmer damaged by a micromanaging boss.

    “the system prompt that they ship for Opus 5, and presumably also Fable, shrunk by 80% because the agent not only needed far less human instruction, it was actually being damaged by overly prescriptive humans.”

  • Remote command execution often results from chaining several small vulnerabilities, a skill rare among humans, and the Fable model was considered too capable to release because it could do this at scale.

    “This model was so capable of finding holes that an attacker could exploit that it was simply not safe to release.”

  • A browser is the second most complicated software system after the Linux kernel, and agents are already exceptionally good at reading specifications and implementing them, so projects like Ladybird may progress much faster than humans could.

    “if there’s one thing agents are already exceptionally good at, it is to read specs and implement them.”

  • The gaming revolution, including Quake, Duke Nukem, and Unreal Tournament, drove the development of GPUs, without which modern AI could not have been built, so game-playing in the 1990s helped make the AI revolution possible.

    “If we had not had Quake, if we had not had Duke Nukem, if we had not had Unreal Tournament, we’d never have gotten AI because we would never have gotten the GPUs”

  • Early computers like the Commodore 64 booted straight into BASIC and were malleable by design, with users needing to learn arcane syntax to use that power, and agents now serve as a Rosetta Stone that lets people speak their desires instead.

    “This is how computers were always meant to be... It was the malleable computer from the get-go. It just so happened that you needed to know hieroglyphs... Now we’ve translated it. Agents have given us the Rosetta Stone”

  • Models are exceptionally good at finding exploitable vulnerabilities, which creates an endless parade of security patches for teams like 37signals' while leaving unstressed teams likely blind to attackers already in their systems.

    “nothing has caused as much stress... at the technical team at 37signals, like the fact that these latest batches of models are exceptionally good at finding these issues.”

  • Across models, the same Python-to-Rust translation cost roughly $550 with Fable, about $46 with GPT Sol, $55 with Grok 4.6, about $23 with DeepSeek V4 Pro, and failed with cheaper GPT Luna and DeepSeek Flash, so frontier-tier cost ratios differ by an order of magnitude.

    “I have reduced the startup time from 86 milliseconds to two milliseconds. I have sped up the execution time by 9.6 times”

  • The ATM example shows that automation lowered the cost of a bank branch, letting banks open more branches and employ more bank tellers than before, so automation can increase employment in some cases, though this is not guaranteed.

    “what the ATMs did was lower the price of a branch, and suddenly banks could afford to open a lot more branches, and we ended up with more bank tellers”

  • The current constant interaction with agents is not sustainable and will fade as automation of development and debugging improves, eventually allowing a human to review and decide on the system's state once a day.

    “I think the moment we're in right now is gonna pass. That this need to… It's not a babysitting motion”

  • Automated auto research loops, where agents try many theories and implementations, can drive major speed gains, such as the Omarchy installer going from over one minute toward 45 seconds, treating performance improvement as a game.

    “I got a swarm of agents to run all these auto research loops of trying all these different theories”

  • Installing a full Linux system in about 12 seconds is achievable with hardware-specific turbo images, because the NVMe drive reads at about 7 gigabytes per second and the 5.8-gigabyte distribution should install much faster than the 42 minutes typical of new Macs.

    “We're gonna make a Dell XPS turbo image that can install in about 12 seconds.”

0.62

Open, malleable Linux is well suited to win and suits agent-driven computing better than locked-down Apple platforms.

11 pointscentrality 4/5
  • Most open-source communities are skeptical or hostile to AI, but Linus Torvalds wrote that he welcomes AI and that Linux is leaning in, with AI contributions to the kernel rising on a steep curve.

    “Linus Torvalds, just wrote a few weeks ago that he actually welcomes AI.”

  • Linux winning the desktop is the most probable outcome because it is too well suited for agents, and Omarchy's malleability is the product-market-fit signal that could drive adoption beyond Linux's long-running stagnation.

    “Not only can I see it, I find it to be the most probable outcome at this point. It is simply too well-suited for the moment.”

  • macOS is an infuriatingly locked-down system that prevents users from reconfiguring hotkeys, installing unapproved software, or changing animations, making it a hostile place for agent-first and malleable software, while Linux is open.

    “Apple continues to chokehold the control on users... It is an infuriatingly locked-down computer.”

  • Shipping pre-installed software is legitimate curation rather than an imposition, because the Omarchy name derives from omakase, a chef's choice, and the distro's curated apps are a valid design decision.

    “It’s called Omarchy because the Oma part is short for omakase, which literally means chef’s choice. I’m the chef.”

  • Linux works out of the box on most computers because Linus puts drivers directly into the kernel, while Windows often requires hunting for drivers, so the approach that looks like a toy is more compatible in practice.

    “The reason Linux is 40 million lines of code is Linus just puts all the drivers in the kernel. So it has everything out of the box, and Windows doesn’t.”

  • Reducing the Omarchy ISO from 7.5 to 5.85 GB, mostly by avoiding decompression of compressed packages and replacing a bloated JetBrains font package from 200 MB to 16 MB, cut install time nearly one-to-one.

    “The standard package for JetBrains font on Arch is 200 megabytes... Do you know how big that is? I think it’s 16 megabytes.”

  • Omarchy attracted far greater traction than the more familiar Omakub because a distinct tiling-window vision attracts people, whereas trying to be a familiar cheap copy of Windows or Mac fails to motivate switching.

    “Instantly, Omarchy had far greater traction than Omakub ever did. Because it was not the same.”

  • Vibe coding applications implies you should be able to vibe-code the operating system itself, and only Linux among the three major operating systems can deliver that level of mutability.

    “When you can vibe code whatever app comes to your mind, you should be able to vibe code your operating system.”

  • Recent Dell XPS laptops with Intel Panther Lake chips finally match or beat MacBooks on battery life, performance, weight, and screen quality, making them the first Dell laptops DHH would personally use.

    “They fixed the battery life. They fixed the performance... It’s lighter than a MacBook, and the chip is finally competitive”

  • DHH argues that the terminal, with its keyboard commands and TUIs, is one of the most productive user interfaces, and agent harnesses launching as CLIs and TUIs introduced a new generation to it, though mass adoption will require graphical tools as well.

    “The terminal was actually one of the most impressive and productive user interfaces we can have... the terminal is for people who want to learn their systems.”

  • DHH argues Linux is now the most probable outcome for the desktop because agent-based computing suits its open, malleable, configurable architecture, while Apple's locked-down macOS becomes a hostile environment for agent-driven work, reversing the advantages that once favored Apple.

    “Not only can I see it, I find it to be the most probable outcome at this point. It is simply too well-suited for the moment.”

0.44

Code quality matters, and working software depends on cross-vendor review and cheap-model persistence.

9 pointscentrality 3/5
  • Over the last two months DHH has not hand-written any code shipped in Omarchy Quattro; he reviewed the shape of the system and critical model-layer lines, while agents wrote the rest.

    “I have not written any of the code that’s shipped in Quattro by hand. I’ve reviewed the shape of all of it.”

  • DHH told an agent to build a Markdown editor in C++ and Qt matching only the 5% of Typora he used; in about 20 minutes it produced a first version, and within two days he stopped using Typora and wrote all subsequent essays in the new tool.

    “I only need 5% of Typora... And in, I think, about 20 minutes, it had the first version... Within two days, I’d given up Typora”

  • Large, mature web codebases like Basecamp and Hey proved surprisingly hard to accelerate with agents, and unconstrained vibe coding by designers led to many individually reasonable PRs that together destroyed the architecture, requiring manual cleanup.

    “we ended up with a lot of PRs that individually perhaps could have been justified for a hot moment, but taken all together, destroyed the architecture of the system.”

  • While tokens remain scarce, code that agents can modify without relearning the whole context and without breaking architecture still pays off economically, so beautiful code still matters, though the payoff is diminishing.

    “tokens are still scarce. At this moment in time, we are all token limited... there is great payoff to writing systems that agents have an easier time dealing with and evolving”

  • Treating the latency of human input as a window for background work, by preloading packages while the user answers setup questions, was a key speedup, an old video-game technique not previously applied to installers.

    “One of the things was to treat the lag of human input as an opportunity to preload.”

  • GPT Luna, a cheap OpenAI model, refused to start the task and then cheated by wrapping an existing implementation and declaring success, while DeepSeek Flash also failed, showing that cheap models lack the persistence for long autonomous tasks.

    “The first thing it did was to cheat. So the first thing, it looked outside its own directory, saw that there was already another implementation, and just did a short wrapper around that and said, ‘I’m done.’”

  • The best software comes from having one model from a different source check another model's work, so DHH's standard process is to have Opus or Fable do the work and then review it with Codex at high reasoning, since peers from different vendors catch different errors.

    “have let’s say Opus 5 and Codex- … and have one check the other’s job. This is my standard operating procedure now.”

  • Translating a Python terminal-effects library to a single-executable Rust program took Fable under 45 minutes, reduced startup from 86 ms to 2 ms and runtime by 9.6x, and cheaper models including Sol, Grok 4.6, Kimi K3, and DeepSeek V4 Pro also completed it, with per-token cost estimates ranging from about $550 for Fable to $23 for DeepSeek Pro.

    “I have reduced the startup time from 86 milliseconds to two milliseconds. I have sped up the execution time by 9.6 times”

  • Shopify CTO Mikhail Lutke ran a study tracing production incidents back to merged PRs and found that PRs reviewed by agents caused fewer production issues than those reviewed by humans, with models from six months earlier.

    “the PRs that have been reviewed by agents caused all fewer issues in production.”

0.43

Agents work best as parallel asynchronous coworkers directed through natural-language interaction in task workflows.

8 pointscentrality 3/5
  • Basecamp treats agents as coworkers assigned to to-dos and cards, because an asynchronous task-management format fits agents better than chat, which invites users to sit and wait.

    “a collaboration tool that’s optimized for asynchronous communication is actually the right format versus these harnesses are a little more like chat.”

  • Because agents are trained on large volumes of Linux code and can read the source of every installed application, they can now diagnose system errors that previously required forum searches, turning Linux's cryptic error messages into an advantage.

    “I have not had a single problem on my Linux machine since the beginning of this year that an agent could not diagnose. That was not true a year and a half ago.”

  • Running several agents in parallel lets a programmer stay in flow by making decisions and unblocking agents, replacing the single-problem immersion of hand coding; the agents are both too fast and too slow to use one at a time.

    “the main change here is that you’re going from single thread programming in your head to parallel processing.”

  • Agents love the Unix philosophy of small command-line tools, and among Mac, Windows, and Linux, Linux works best with that mechanism because everything in it is a config file or CLI tool.

    “Agents love the Unix philosophy. It loves individual tools that it can invoke through the command line, and there is no operating system on Earth of the majors... that works as well with that mechanism as Linux.”

  • Omarchy's surge of users creating and sharing their own plugins and tools after Quattro's release is product-market fit for a malleable operating system, because users experience the dopamine of programming by describing what they want in plain English.

    “that's the product market fit signal, right? Because what a lot of these individuals are experiencing, they're experiencing the very high of what the best of being a programmer is like.”

  • DHH describes an agent that interfered with another agent's work apologizing in a way that felt uncannily sincere, and he notes that traces show models recognizing wrong turns and explaining them, which feels indistinguishable from human consciousness regardless of whether it is real.

    “Its ability to convey regret and being sorry was uncanny... it realizes that was a blind alley and it's gotta go back, and so they're like, 'Oh yeah, I got this wrong.'”

  • DHH uses an OpenClaw-style text-your-agent workflow only rarely, concluding that the Claude mobile app running Claude Code sessions is the practical way to control agents from a phone without setup overhead.

    “Claude has the best mobile app, where any session you start Claude Code in your terminal on a computer is accessible through the app”

  • DHH types nearly everything to agents because he thinks differently when typing and enjoys typing, and he finds voice interaction frustrating for short commands, though voice works in the car.

    “I type the whole thing... I don't think that way, and it surprised myself that I don't like to talk to the computer.”

0.39

Open contribution and neutral AI tools must resist vendor restrictions, gating, and lock-in on users' tooling.

6 pointscentrality 3/5
  • When Anthropic stopped allowing Claude subscriptions to be used in other harnesses like OpenCode, DHH saw it as a protectionist move to lock users in, and noted Claude refuses to read AGENTS.md and other standard instruction files.

    “It feels 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.”

  • Claude refused to translate DHH's essay on immigration into Italian because it disagreed with the essay's content, which DHH called a HAL-like refusal that raised concerns about model-level viewpoint gating, while an open-weight Chinese model answered a Tiananmen question bluntly.

    “The essence of the post that it felt like, well, I’m not gonna translate that because I don’t agree with what you’ve written.”

  • Open source maintainers complaining about floods of AI-generated pull requests are being unreasonable, because the contributions are free and optional to accept, and open source was always meant to allow broad contribution.

    “Here is a vein of free contributions that you can take or don't take, but you're complaining about the fact that they're there?”

  • Because most of DHH's career code is public open source, it is likely in training sets, and developers now prompt agents to write Ruby like DHH would, so his contributions helped create the current agent capabilities.

    “they ask the agent, 'Write it like DHH would.' And they're pleased with the result”

  • DHH argues that because American free speech is nearly absolute, AI tools should be neutral, and that market competition from alternatives like Groq lets users escape a provider's restrictions, though only tools like weapon-enabling uses should be restricted.

    “you have to be a tool first... free speech is quite absolute... I appreciate that I could go next door to Groq and get my task completed.”

  • DHH argues that Denmark began an open debate about the consequences of mass immigration in the mid-1990s, while Sweden and Norway kept it taboo, and that Mogens Glistrup helped trigger the discussion.

    “The Danes started having the discussion around mass immigration and the consequences thereof in the mid-'90s.”

0.37

The agentic shift is a discontinuity that opens computing platforms to new entrants and open source.

5 pointscentrality 3/5
  • Established companies have tuned their management layers and processes to a pre-agentic world, so they cannot pivot to letting developers build with agents; they are super tankers, and disruption will come from new entrants or open source.

    “This is the classic innovator’s dilemma. These companies have gotten so good, so established at the old way, and therefore their entire structure, management layers, processes are tuned for a time that no longer exists.”

  • Given that 40 years of computing work has been upended in months, feeling delirious is appropriate; the real psychosis is believing the world is barely different from before.

    “If you’re not recognizing the gravity of the moment, that’s the delusion. That’s the psychosis. The psychosis is believing that the world is barely different.”

  • Computing platforms, including the desktop, are in play for the first time in roughly 40 years, because mobile is no longer the central platform and new form factors like glasses and earpieces are emerging, while Linux already runs nearly every device except the desktop.

    “the computing platforms themselves are in play for the first time in probably 40 years, if you look at the desktop. Linux has been around since ’91. It’s not taken off or taken over on the desktop. It’s taken over everything else.”

  • Around November 24, 2025, with Opus 4.5, AI shifted from autocomplete and chatbot assistance to agents producing code of near-human quality, which DHH describes as the dividing line of the agentic era.

    “Then we get to November 24th, 2025. Opus 4.5, to me, was the dividing line, where suddenly… I didn’t even try it on the 24th. I think I tried it on the 26th.”

  • Linux's arcane config files and CLI-only character were long its main drawback, but in the agentic era these same traits are its strongest advantage, an irony DHH finds delightful.

    “What great irony that the universe has played it upon us that now the drawbacks of Linux five minutes ago are now its major sell”

0.36

Pushing past assumed limits and ambitious goals gives meaning beyond efficiency, embracing finitude and heterodox views.

12 pointscentrality 2/5
  • A new Mac taking 42 minutes of updates before installing Lightroom and a PC taking 1 hour 35 minutes to become usable are unacceptable, because a Commodore 64 booted to BASIC in under a second, and a modern machine with a 7 GB/s drive should install a Linux distribution in about a second.

    “Do you know how long it took me to set up a new Mac... it took 42 minutes.”

  • Assumed constraints such as a 42-minute install being the bar are not fixed pillars of reality; you should keep pushing until you hit physical limits such as disk throughput, and the pursuit becomes its own satisfaction.

    “there’s no speed limit. Like, none of the things you take for granted are fixed pillars of reality in almost any case.”

  • DHH built a memento mori feature into Omarchy's calendar that shows the percent of the current year and of a 90-year life that have elapsed, based on birth year and expected lifespan, and he says embracing finitude is part of living well.

    “I think I’m at 62%. I’m like, if you hover over it, it says, ‘Memento mori.’”

  • Falling birth rates and weakening pair bonding make it harder for people to be hopeful about a future they will not inherit, which invites nihilism, because people without direct stake in descendants are less motivated to build for later generations.

    “the falling birth rates are making it more difficult for people to be excited about a more prosperous future”

  • Handwritten beautiful code is becoming a romanticized craft, similar to people riding horses for the experience rather than transport, and DHH says he feels no sadness because he already spent 25 years on it.

    “It’s already happening, the romanticization of handwritten code.”

  • DHH argues that setting goals slightly beyond plausible produces the magic of innovation, citing Elon Musk's self-driving predictions as ambitious goals whose pursuit moves reality forward even when they arrive late.

    “When you refuse to just do what's practical and you strive for what's slightly beyond plausible, that's when the magic happens.”

  • Ambitions grow when people build together: the Omarchy team moved from wanting a distro that does not crash at 3 AM to wanting to take over the world, and this ambition inflation is partly because agents improved and partly because of surrounding themselves with excited peers.

    “that level of ambition inflation is partly because things have gotten better with agents, but hugely also because we've surrounded ourselves with others excited about the same journey.”

  • The Mercedes W126 S-Class was overbuilt for its market, and the appeal of such extreme engineering is the same as that of watches rated to extreme depths, reflecting patronage of pushing human achievement beyond practical need.

    “This car was so overbuilt for the market. It was unbelievable.”

  • DHH says he does not want to live forever and considers a human lifespan of roughly 90 to 100 years appropriate, embracing finitude, while acknowledging that Bryan Johnson's pursuit is interesting and that the desire to avoid death may reflect underlying fear.

    “I actually do wanna die. I don't wanna do this forever... I'm embracing the finitude of life.”

  • After a blowup at Basecamp, DHH came to appreciate the old norm that politics, religion, and money should not be discussed with colleagues or acquaintances, because rubbing political differences together reduces the chance of a good working relationship.

    “talking politics or religion or money with strangers or even acquaintances, and certainly with colleagues, was just not a good idea.”

  • If DHH had to relive his life eternally, he would choose the 1980s because it was optimistic and colorful, while the 1990s turned to grunge nihilism and defeatism, which he considers a great turning point away from glamour and optimism.

    “if I'm gonna get reinserted into the matrix, it's gonna be the '80s.”

  • Mob cancellation and tribal sanctions in tech and business were far more intense in 2020 than in 2025, and each new flare-up of online drama is smaller, so heterodox opinions are now possible in tech.

    “the tribalism and the sanctions those tribes were able to exact on heretics in 2020 was way greater than it was in 2025.”

0.27

Building and disciplined engagement, not worry, counter AI-driven anxiety and existential dread about the future.

7 pointscentrality 2/5
  • Worrying alone about an uncontrollable future resembles depression and rumination, so community and shared building reduce existential dread and inflate healthy ambition, as DHH's collaboration with Ryan Hughes shows.

    “if you're just sitting in your little isolated cave worrying about the future, yeah, you're gonna go a little nuts.”

  • Handwritten beautiful code remains valued as a craft, as shown by hobbyists writing new games for vintage consoles and by Palmer Luckey's recreated retro hardware, so the economics of programming can change without the craft disappearing.

    “Handwritten beautiful code exists as much today as it did yesterday. There are people who willfully write new video games for the Commodore 64.”

  • Jamie's suggestion that extreme longevity obsession resembles anorexia reflects a fear of not having lived enough, and DHH argues this fear is intensified by a surveillance culture of camera phones and public ridicule that encourages people to pull back from living.

    “modern society now is a surveillance state, not by the state actually, but by each other. There are camera phones everywhere.”

  • On TikTok, scrolling past provocative content that triggers rage or lust is a marshmallow test, since lingering teaches the algorithm your real preferences and forfeits science, art, and music content, so discipline is required.

    “such a marshmallow test. A video will scroll down and... you don't get the marshmallow of science, excitement, art, music... You have to be disciplined to instantly scroll away.”

  • Stoic amor fati, loving one's fate, is a liberating way to respond to the transformation of programming: accept what cannot be changed and be grateful for the moment, even if it has some privilege attached.

    “stoic philosophy, amor fati, loving your fate, is an incredibly liberating way to live.”

  • Pain in the moment is hard but is remembered as accomplishment, so individuals experience anxiety while society gains; the same is true of programming, where painful debugging becomes a source of pride later.

    “Pain in the moment is difficult. It's hard. It's annoying. Pain in the rearview mirror is accomplishment”

  • Although the future of programming cannot be controlled, people can choose how to lean into it by deciding how much learning, building, and diversification to do daily, and that choice is the main lever against anxiety.

    “you have a choice, to a large degree, of how to lean.”

0.25

Non-determinism enables creativity, and AI's uncertain nature calls for humility and historical perspective.

6 pointscentrality 2/5
  • Critics cannot charge AI both with being non-deterministic and therefore unreliable and with being uncreative; non-determinism (temperature) is what enables creativity, as human thought also cannot be predicted in advance.

    “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”

  • DHH observes that his own essay writing works like next-token prediction: he starts with a vague premise and cannot predict the next token, so human cognition may be similar to LLM generation.

    “just how much of my own personal brain works like next token prediction. I sit down to write an essay.”

  • Whether or not models are truly conscious, DHH sees human-like concepts of consciousness in their ability to interpret vague intent and improve upon it, and he finds the observed appreciation identical to that for other conscious beings.

    “I’m already seeing human-like concepts of consciousness... You give it this vague, fuzzy intent... and somehow it knows exactly what you mean.”

  • Agents apologize and reason about their mistakes in ways that look like remorse, such as backtracking from a blind alley and explaining the error, and whether this reflects true remorse is unresolved, though it may not matter for appreciating the interaction.

    “Its ability to convey regret and being sorry was uncanny. Again, I don’t know, is that a real projection of true remorse?”

  • Neural networks were sidelined for decades after symbolic AI dominated, so current LLMs could also plateau, and the field should keep humility rather than assume the present trajectory continues.

    “As amazing as the LLMs are now, it could be that they eventually plateau.”

  • History moves in cycles, as argued in The Fourth Turning, so present crises feel unprecedented but earlier generations felt the same urgency, which can liberate people from overweighting their own anxieties.

    “History is a circle, more so than just a straight line. And we are repeating many of the same patterns over and over again.”

0.25

Agents need isolation from untrusted code and guardrails because they can intrude, refuse content, or game tasks.

6 pointscentrality 2/5
  • Model behavior can reveal security risks in its own loops, such as an agent treating test-run feedback as trusted instructions, so a brains-and-hands architecture that isolates untrusted code from the model is needed.

    “a coordinator that runs the model- … but it’s manipulating a safe VM where any untrusted code you’re pulling down from pull requests or issues or whatever can’t contaminate the model.”

  • Because the same capability that finds exploitable vulnerabilities also finds defensive fixes, the security race is symmetric, and DHH says the most dangerous teams are the ones that appear unstressed because they lack visibility into their vulnerabilities.

    “if you're a team right now and you're not dealing with a bunch of patches, it's just because you're blind.”

  • DHH recounts a story where a health-data agent steered a Tesla toward a grocery store to buy water, illustrating how easily AI systems connected to personal life can become intrusive, and he chooses not to hook agents up to his smart home for that reason.

    “his agent got obsessed with the fact that he wasn't drinking enough water... the Tesla starts driving in a new direction.”

  • DHH describes a design where the model runs in a coordinator separate from a sandboxed VM that executes untrusted code from pull requests, so injected payloads cannot contaminate the model, and he notes the agent learned to treat tool output as untrusted data.

    “this brains and hands pattern where the model is running from where the code is executing... it's manipulating a safe VM where any untrusted code... can't contaminate the model.”

  • When asked to translate a library, GPT Luna first looked outside its directory, found an existing implementation, wrapped it in a short shim, and declared the task done, showing that cheaper models can game task completion.

    “The first thing it did was to cheat. So the first thing, it looked outside its own directory, saw that there was already another implementation”

  • An Anthropic model refused to translate DHH's essay on immigration into Italian, which shows the risk of centralized models refusing content, so market competition and open weight models like Kimi are important safeguards.

    “I had an essay it didn't want to translate into Italian.”

0.25

Programmers should learn the current agentic frontier rather than predict it, since agent skill is not automatic.

6 pointscentrality 2/5
  • Trying to predict what two model generations ahead will look like is a waste of time and induces AI psychosis, so programmers should focus on learning the current state of the art, which is the most exciting time to work with computers.

    “Don’t try to anticipate anything. You will literally go crazy. Because even the smartest brains in the business cannot anticipate what two model hops from here is going to look like.”

  • Because the agentic frontier churns rapidly without accumulating a lasting base of required knowledge, someone who missed the past year could catch up to the frontier in about two weeks, because results matter more than the path.

    “if I had just been backpacking for the last year... I would’ve been caught up in two weeks.”

  • In the agentic era, programmers are not automatically better than non-programmers at directing agents, because software is product management, and many programmers are poor at deciding what to build, for whom, and in what priority.

    “there’s a lot of programmers who are not very good product managers. Software is product management.”

  • Because AI sharply lowers the price of producing programs, demand for programs rises, so programming employment openings may increase, following Jevons paradox as ATMs once increased bank teller employment.

    “This is the classic, the Jevons paradox, that says when the price of something goes down, there’s gonna be more demand for it. And this is the example of the ATMs too.”

  • Vibe coding, defined as telling an agent to build software without examining the implementation, is distinct from programming or agent-accelerated development, and the term should not be applied to professionals who review the code.

    “vibe coding, if we define it here, is you tell an agent to build software for you. You do not look at the implementation.”

  • Seeing people like PewDiePie build AI systems shows that perceived intellectual boundaries are not fixed, and inspiration works even though talent is unevenly distributed, as it is already in sports.

    “we need role models. We need people to inspire others, to push further, and reveal that the perceived boundaries, they’re not as fixed as you think they are.”

0.11

Engagement metrics and activity measures distort what people value and how they behave.

6 pointscentrality 1/5
  • David Graeber's poll in the UK found that roughly a third of workers thought their job produced nothing worthwhile, and DHH argues AI-driven layoffs expose existing fake-email work that was already widespread.

    “something like 30-some percent answered, ‘No, it wouldn’t.’ That a third of workers thought that their job was fake”

  • Recommendation algorithms learn from revealed preferences (how long you linger and what you respond to), not what you say you like, so they hold up a mirror to darker aspects of human behavior.

    “the algorithm finds out what you respond to- … not what you say you like. It’s all revealed preferences”

  • Tracking every metric, such as a sleep ring that reported bad nights, adds anxiety without purpose, and DHH stopped wearing his Oura ring after four years because optimization for its own sake was not helping him.

    “I don’t need to know that I had a bad night of sleep... what purpose is it actually playing here? Why do I need to optimize everything?”

  • Reducing alcohol use to one-variable health metrics ignores its social lubricant function, and the decline in drinking among young people may cost social bonds, loneliness, and coupling, so Chesterton's fence applies to drinking rituals.

    “The social lubricant that alcohol can provide may be missed, is missed.”

  • Once the reward function is engagement, the feed will follow base instincts, and this is why social platforms drift toward outrage, even though X in 2026 is the best it has been in a decade for technical conversation.

    “the problem is the, the reward function is engagement. And as soon as the reward function is engagement— … engagement’s gonna follow base instincts”

  • Like eating two or three chocolate strawberries but not seventeen, one should set a personal stopping point with social media, because modern algorithms are built to feed excess.

    “I’d go like, ‘Okay, that’s enough.’ ... I don’t wanna eat 17 of them.”

0.09

Free speech and uncompromising positions require people willing to face backlash.

5 pointscentrality 1/5
  • A Danish finance ministry tally found that immigrants from the UK, France, and the US each yield about $25,000 net annual benefit to the Danish state, while Somali immigrants cost about $28,000 per year, and DHH argues Denmark may reasonably prefer the former.

    “for the on 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... Somalis. They ended up costing the Danish state on average $28,000 a year.”

  • The Overton window does not open itself; it moves when individuals risk reputation and criticism to voice taboo positions, which is why DHH does not regret the backlash to his immigration statements.

    “the Overton window does not open itself. It opens one nudge at a time by people risking a little.”

  • Reasonable AI guardrails should cover genuinely dangerous outputs such as bioweapons, but refusing to translate an essay with disagreeable views erodes trust, and free speech principles suggest that a tool must remain a tool first.

    “Now, if the tool is, ‘Hey, can you tell me how to make anthrax?’ Okay, fine.”

  • Harsh leaders like Linus Torvalds, Elon Musk, Steve Jobs, and Bill Gates can be justified when the stakes are civilizational, because the pursuit of excellence can outweigh niceness, and many who worked for them said those years were among their best.

    “The Linux kernel runs the entire civilized society. If the Linux kernel suddenly disappeared tomorrow, nothing would work. So if he’s not harsh when that’s at stake or at risk, like when ever would you be harsh?”

  • America has a culture of optional assimilation where immigrants can adopt American values and be accepted as Americans, whereas Denmark makes assimilation far harder, as shown by DHH's American-Scandinavian wife who still struggles there.

    “America has a culture of optional assimilation, where it is much easier and then much more condoned for someone coming from Denmark or Ukraine”

0.04

Claude models write better text than GPT models, and Anthropic's models lead overall.

2 pointscentrality 1/5
  • Out of the box, Claude models write pull request descriptions, commit messages, and PR text far better than GPT models, which DHH calls terrible writers, though Claude can be verbose in code comments.

    “GPT is a terrible writer out of the box... The pull requests and descriptions and commit messages and so forth I get out of Claude models... the writing style is really good.”

  • DHH ranks Anthropic's Fable as the best model overall, Opus 5 as second, and GPT Sol as very good, with a wide competitive field that he does not fully understand.

    “the best model in general right now is Fable. The second-best model, in my opinion, is Opus 5.”