The server is not especially exotic. It is an Oracle Cloud Ampere A1 virtual machine running Ubuntu 24.04. It has four Arm cores, 24 GB of memory, 16 GB of swap, and a 200 GB-class disk. Oracle Cloud’s Free Tier makes its Arm instances a good way to experiment with this setup before deciding whether it deserves a monthly bill.

That is more machine than most coding agents need. Codex and similar tools call models hosted elsewhere, so the VPS is not doing model inference. It is running the surrounding work: reading repositories, editing files, compiling code, running tests, hosting development servers, and keeping terminals alive.

This is where the appeal of a VPS became clear to me. Agentic development does not particularly need the machine under my hands. It needs a machine that stays put.

One canonical machine

The Oracle VPS is now the primary development environment. It holds the repositories, language runtimes, dependencies, credentials, Codex, and Herdr. The client needs an SSH key, an SSH configuration, and a Herdr installation. A browser or graphical editor can still be useful, but there is no need to maintain another working copy of every repository.

This avoids the awkwardness of synchronizing two computers and wondering which uncommitted change is current. SSH carries input and screen updates. Git still moves code between durable environments. If I disconnect and return from another device, the filesystem is exactly where I left it.

The arrangement is close to the old terminal-and-mainframe model, except the mainframe is a small Linux VM and the terminal can be almost any computer with a good screen and keyboard.

Starting with Oracle’s free Arm machine

Oracle’s Ampere A1 shape is unusually well suited to this experiment. The machine running this site’s content repository has four Neoverse-N1 CPUs and 24 GB of RAM. The root filesystem is about 193 GB after formatting, with roughly 167 GB free as I write this.

That is enough room for several repositories, build artifacts, language servers, and more than one agent. Memory has also been comfortable: the server currently has plenty available even with Codex and the normal background services running. The 16 GB swap file is there as a cushion, not a substitute for RAM.

Arm is the one meaningful tradeoff. Most of the tools I use here publish Linux aarch64 builds or install through language package managers, and Herdr supports Arm on Linux. Some proprietary binaries, browser tooling, or older dependencies still assume x86-64. I would check the important toolchain before moving an existing project. Rebuilding dependencies on the server is safer than copying node_modules, virtual environments, or compiled artifacts from another architecture.

Oracle’s Free Tier rules and available capacity can vary by account and region. Oracle’s current documentation distinguishes promotional credits, Always Free resources, and account-specific limits, so I would verify the shape is marked eligible in the console rather than relying on an old tutorial. Free instances also have no service-level agreement or full Oracle support. This is a good place to begin, not a reason to forget backups.

Making SSH the foundation

Herdr’s remote features sit on top of ordinary SSH, so I wanted SSH to work cleanly before adding anything else. The server uses a normal non-root account with sudo. Authentication uses a key, and a short entry in the client’s ~/.ssh/config turns the destination into a stable name:

Host workbox
    HostName <server-name-or-private-address>
    User ubuntu
    IdentityFile ~/.ssh/workbox_ed25519
    IdentitiesOnly yes

The real test is deliberately unimpressive:

ssh workbox

If that fails, Herdr is not the problem. The failure belongs to the route, hostname, SSH daemon, or key.

This VPS also runs Tailscale, which gives my devices a private route to one another. I prefer normal OpenSSH over that route rather than exposing development services publicly. When setting up a similar machine, I would keep the original SSH session open until a second terminal has connected successfully over Tailscale. I would also test Oracle’s browser console as a recovery path before removing any public SSH rule. Locking the front door is less impressive when the key is still inside.

Tailscale server nodes need a little maintenance of their own. Their keys should not be allowed to expire unexpectedly; a tagged server or a deliberately managed expiry policy avoids discovering six months later that the private route disappeared. The provider console remains useful when DNS, SSH, Tailscale, or an enthusiastic firewall change goes wrong.

Building the environment

The base system is Ubuntu 24.04 LTS. I installed the ordinary tools first: Git, curl, unzip, jq, ripgrep, build essentials, Python with virtual environments, and the runtimes required by each repository. User-installed binaries live in ~/.local/bin, which needs to be present on the login shell’s PATH, not only in an interactive shell configuration.

Codex is installed directly on the VPS. Its authentication and credentials live there because that is where it runs. I prefer supported authentication flows and narrow credentials to copying a mysterious configuration directory from another computer.

Repositories have a predictable home under ~/Projects. A new one is just a normal clone:

mkdir -p ~/Projects
cd ~/Projects
git clone git@github.com:example/project.git

Moving an existing dirty workspace deserves more care. I would stop editing the source, run an rsync dry run, preserve the Git directory and uncommitted files, exclude generated dependencies, and rebuild them for Linux Arm. The VPS copy becomes canonical only after the repository has been inspected and its tests pass.

Herdr over SSH

Herdr is the piece that makes the remote server pleasant to use as an agent machine. Like tmux, it has a background server that owns real terminal processes. Closing the client or losing the SSH connection does not end the panes. Unlike a general terminal multiplexer, Herdr recognizes common coding agents and shows whether each one is working, blocked, done, or idle.

This VPS is running Herdr 0.9.0 from ~/.local/bin. Herdr can be installed on Linux or macOS with its current installer:

curl -fsSL https://herdr.dev/install.sh | sh

After Herdr is installed on the client and ssh workbox succeeds, the VPS can be saved as a machine:

herdr machine add workbox --label "Oracle VPS"
herdr

The setup command checks the remote Herdr installation and server and asks before changing anything. The VPS then appears in the local Herdr interface with its own workspaces, tabs, panes, and agents. The remote server still owns the processes; the local Herdr client brings back their views.

A direct connection is also available:

herdr --remote workbox

I use the VPS panes for agents, test output, development servers, and ordinary shells. Herdr’s sidebar is more useful as the number of agents grows because I can see which one needs an answer without opening every terminal to check.

The account on this server has systemd lingering enabled. That lets user services continue without an active login session, which is useful for an always-on development box. Herdr itself keeps terminal processes alive across client disconnections. A reboot is different: processes die when the server restarts, although Herdr can restore layouts and some supported agent sessions. Anything that truly must survive independently belongs in a systemd service rather than a terminal pane.

Reaching development servers

I keep development servers bound to 127.0.0.1 instead of opening temporary ports to the internet. An SSH tunnel is enough for a quick browser session:

ssh -N -L 3000:127.0.0.1:3000 workbox

Port 3000 on the client now reaches port 3000 on the VPS. An editor’s port forwarding or a private Tailscale Serve route can provide the same basic result when it is more convenient.

Backups matter more once the VPS becomes canonical. Git remotes protect committed code, but not databases, uncommitted work, local configuration, or agent state. Provider snapshots are useful, though they may disappear with the instance. Anything irreplaceable should also have an off-provider backup that has survived an actual restore test.

The machine stays put

The daily routine is now quite simple. The client connects to the private network, Herdr opens, and the Oracle VPS is waiting with its terminals and agents where they were left. The computer doing the work did not go to sleep, change networks, or come along for the trip.

There are tradeoffs. The network becomes part of the development environment. The server needs updates, backups, and a recovery path. Some Arm compatibility problems are real. Renting a VPS also means trusting the hosting provider with the machine beneath the filesystem.

But the experiment does not require expensive hardware. The computer writing this post is a free-tier Arm VM. It stays in one place, does the work, and lets almost any other computer become its screen. SSH connects the two, and Herdr makes the distance mostly disappear.

The result is Our Solar System, a 3D model here on the site. It has the Sun, the eight planets, Earth’s Moon, and Pluto. You can move the view around, zoom in, and select a world to bring up a short lesson. There’s also a row of names along the bottom, which makes Mercury easier to reach when it’s a small dot on a small screen.

I kept the lessons brief so there would still be room to look around. Each world has a few facts to give a child somewhere to begin, without asking them to read a whole page before choosing another planet. Hiding the names also hides the lesson card, leaving more of the screen for the model itself.

The sizes and distances are adjusted to make everything visible. A model that showed both accurately would leave very little to tap on a phone, so I enlarged the planets and brought their orbits closer together. I wanted children to be able to recognize the worlds and explore them, while being clear about what the picture leaves out.

I included Pluto as a dwarf planet. It’s still a world worth learning about, and keeping it in the model gives me a place to explain the classification. It can sit alongside the eight planets without being counted as one of them.

Building with Hermes and Grok

I used Grok as the model inside Hermes. Hermes is an open-source agent from Nous Research that connects a language model to tools for working with files, running terminal commands, and other tasks. In its agent loop, it sends the conversation and available tools to the model, carries out the tool calls the model requests, and returns their results for the next step. That lets a coding session continue through edits and command output instead of ending with a block of suggested code.

Grok supplied the model responses in that loop: interpreting my requests, generating code, and requesting tool actions. Hermes supplied the machinery for carrying those actions out and bringing the results back into the conversation. I described what I wanted, looked at the result, and asked for changes over several passes. The work produced the scene data, rendering code, generated textures, and page controls described below.

Three.js is the JavaScript graphics library that the resulting app uses in the browser. It provides objects for geometry, materials, lights, and cameras, along with a renderer that turns the scene into WebGL drawing operations on a canvas. Our code assembles those objects into planets and updates them as time passes. Hermes and Grok were part of making the app; opening the finished page runs JavaScript and Three.js locally, with no model call needed to move a planet or display a lesson.

The first version ran locally with Vite, which provided the development server and reloaded the page as the source changed. That gave me a browser view to judge while the files were being revised. The project itself settled into a few parts: a JSON description of the solar system, a World class to render and animate it, texture functions, and an application module to connect the scene to the page controls.

Describing the scene

The project separates the description of the solar system from the code that draws it. A JSON file holds the worlds, their appearance and motion, and the text for their lessons. Each record includes a name, radius, color, and texture type, with optional fields for rings, atmosphere, tilt, and an orbit. The Moon’s record names Earth as its parent; Earth’s names the Sun.

I like this arrangement because the same record supplies both the planet on screen and the lesson beside it. Selecting Saturn gives the interface Saturn’s name, classification, pronunciation, and facts from that file. I can revise an explanation without touching the rendering code, or adjust a displayed orbit without changing how selection works. Another model built from similar bodies could use the same renderer, though a different kind of scene would still need new code.

Making the planets move

The renderer builds each world from a sphere and a small hierarchy of Three.js groups. One group turns around the orbital center, another places the planet at its orbital distance, and a third handles its tilt and spin. Keeping those movements separate lets a planet rotate while it travels around the Sun. The Moon’s orbit attaches to Earth’s position, so it comes along as Earth moves without inheriting Earth’s daily spin.

The orbits are circles, and the animation calculates positions from elapsed time. There’s no gravity calculation between bodies. I also compressed the differences in orbital periods so the outer planets would move visibly: the code raises each period to the power of 0.38 before using it to set the animation speed. At the default speed, Earth completes an orbit in 24 seconds, while Neptune takes roughly three minutes. Those timings belong to the illustration; they aren’t a common time scale for the real solar system.

The surfaces are generated in the browser too. A texture function draws pixels into an offscreen canvas, combining colors with layers of noise to produce rocky surfaces, cloud patterns, and the bands on the gas giants. Three.js wraps the resulting image around the sphere. Saturn’s rings use a separate generated texture with transparent gaps.

That keeps the planet artwork in code, but it also sets a limit on what the picture can teach. Earth’s green and blue patches are generated shapes, not a map of its continents. I think of these surfaces as illustrations that help distinguish the worlds. The lesson text carries details that the model doesn’t attempt to reproduce.

Connecting the model to the page

The planets are drawn in WebGL, while the names and lessons remain HTML. Three.js’s CSS2DRenderer keeps each name positioned beside its world as the view moves. The lesson panel is ordinary text and a list of facts, so it can be styled and laid out with the rest of the page. I don’t need to draw text into a texture just to explain what a child has selected.

A click on the scene uses raycasting: the code projects a line from the camera through the pointer’s position and checks which planet it intersects. Clicking a name or one of the buttons along the bottom reaches the same selection function. That function updates the lesson and starts a gradual camera move toward the selected world. Once the move finishes, the camera’s point of interest continues to follow the world as it travels.

I wanted the controls to leave some room for looking. The page can pause the orbital and spin animation, change its speed, or return to the full view. Hiding the names also hides the lesson card. If the browser reports a preference for reduced motion, the model starts paused, though selecting a world still moves the camera.

Fitting it into the site

My site uses Jekyll and GitHub Pages, so the finished version is a static HTML page with JavaScript modules and the JSON scene file alongside it. The browser fetches the description and builds the scene locally. Vite was useful during development, but the deployed page doesn’t need a running Vite or Node server.

Three.js and its controls are included in the site’s files. A small import map tells the browser where to find the library when a module imports three. This lets the page use modules directly while keeping that dependency on the same site. The model also uses the site’s existing fonts, so the surrounding text feels familiar.

On a phone, the lesson moves toward the bottom of the screen and has a limited height, with its contents available by scrolling. The row of names provides a larger target than the planets themselves. The renderer caps its pixel ratio at two, which limits the number of pixels it draws on displays with a higher density. I had my younger kids in mind when making those choices, though I haven’t established how well the model works in a classroom.

I like being able to keep working on the explanations separately from the scene. A lesson can get clearer without rebuilding a planet, and the controls can improve without rewriting the lesson. Trying Hermes has left me with a small solar system I can share with my younger kids. I’m curious to see which parts they want to spend time with.

Then a different explanation arrived. A report commissioned to review the cases suggested the sound matched a cricket — Anurogryllus celerinictus — and that the injuries might have been psychogenic. The Army’s discussion of the cases was quietly removed from public view.

I’ve thought about that sequence more than the diagnosis itself. You hear a sound, you get sick, you look for a weapon, and you are offered a cricket. Whatever actually happened in those rooms, the uncertainty is the point that stays with you. A conflict where you can’t agree whether there was an attack, let alone who ordered it, is hard to defend against — and hard even to admit as a conflict.

Fifth-generation warfare is a clumsy name for that kind of problem.

The definition that refuses to settle

The phrase started in 2003, when analyst Robert Steele suggested there was a new kind of war beyond the one William Lind and his co-authors had described in 1989. Lind had argued war was moving from massed manpower to firepower to maneuver, and then to a fourth generation where the state’s monopoly on fighting eroded and decentralized insurgents could tie down a superpower. Steele thought something else was emerging after that. Lind, for his part, rejected the addition the following year, saying the fourth generation hadn’t fully arrived yet.

That disagreement never really resolved. There is still no widely agreed definition and no joint doctrine publication that codifies it.

Daniel Abbot, editing The Handbook of 5GW in 2010, offered the most honest formulation I’ve found: “The very nature of Fifth Generation Warfare is that it is difficult to define.” He added, more plainly, that it is a war of information and perception. Others have phrased the same idea as the deliberate manipulation of an observer’s context in order to get a desired decision, or, in the compressed version that stuck with me, a war where you might not know you are losing it because you don’t know you are in it.

The useful way I think about the generations is not as a timeline where one replaces the last, but as overlapping layers that get added. First-generation wars were about putting bodies in line. Second about putting metal on target. Third about moving faster than the other side could orient. Fourth about making political will the target and using any mix of political, economic, social, and military pressure to erode it. Fifth keeps all of that but moves the primary contest to perception itself.

What makes that move different is not that propaganda is new — it isn’t — but that the tools for testing and shaping perception are now data-driven, cheap, and fast. An advertising platform can build a model of what holds a particular person’s attention for an extra half second. The same model can be used to see what makes that person suspicious, indignant, or exhausted. When those loops are measured in milliseconds, you don’t need to persuade an entire country. You can nudge a small group at the right moment, see what happens, and adjust.

From winning information to shaping meaning

A distinction the Army’s Cyber Defense Review has tried to make — and that a recent synthesis of 5GW literature highlights — helped clarify this for me. Information warfare is the battle for information, the contest for electrons and access in the electromagnetic spectrum. Cognitive warfare is the battle of information, the contest over what those electrons end up meaning.

In practice the two are usually paired. A cyber operation that steals or leaks emails wins information. The story that gets wrapped around the leak — who it is said to incriminate, what it is said to prove about the other side’s illegitimacy — shapes meaning. The first can be measured in data exfiltrated. The second shows up in what people believe is true about their own institutions.

That is why attribution is more than a technical puzzle. In a kinetic war you generally know who fired. In a well-executed influence operation you may not know whether a message was amplified by a neighbor who believes it, a commercial data broker who sold a list, or a state actor who seeded the narrative. Grey Dynamics puts it bluntly that simply understanding who is behind a 5GW attack can be impossible. When the Economist or the Wall Street Journal later maps the operation, they are often reconstructing it from fragments — cloned news sites in the Doppelgänger operation of 2023, deepfakes that are caught only after they have circulated, networks of accounts that behaved as if they were coordinated but were hard to tie to a single sponsor.

I’ve come to care less about single dramatic examples than about cumulative pressure — the slow change in how a group weighs options over time. That is also why the usual scoreboard — how many accounts were taken down, how many leaks were proven authentic — misses the larger question of what an information environment is training people to become.

Where it already lives

If the category still sounds theoretical, the institutions have stopped treating it that way.

NATO’s Allied Command Transformation now maintains a standing Cognitive Warfare program, and its chief scientist released a new report on the subject in January 2026 describing cognition as a domain alongside land, sea, air, space, and cyberspace. The language is deliberately spare: the aim is to affect not only what an adversary knows but how they think and how they will to act.

In Washington, the 2026 National Defense Authorization Act directs the Secretary of Defense to define cognitive warfare, identify who inside the Department owns it, and assess the value of “narrative intelligence” by the end of March 2026. The small detail that Congress felt it had to order a definition tells you how dispersed responsibility still is. The Army talks about information operations, the Air Force talks about data-saturated multi-domain environments, and the analysts in between argue about the line between hybrid warfare and gray zone conflict.

On the domestic side, the Cybersecurity and Infrastructure Security Agency has been asked to do more — election security, critical infrastructure, public cyber literacy — while facing political and resource strain. That mismatch matters because cognitive effects do not stop at a network boundary. An attack on a water treatment plant can be both a technical incident and a message about whether the government can keep water safe. The deeper effect is not the outage. It is the erosion of trust that makes the next shared task harder.

Even the communications layer is being rebuilt with this in mind. The Department of Defense released a Private 5G deployment strategy in late 2024 that is partly about speed and partly about resilience — making it harder for an adversary to inject, delay, or deny information at the transport layer. None of that prevents a convincing deepfake from spreading on the application layer. It just makes the underlying pipes less fragile.

What winning quietly asks of us

If the best way to win is to arrange things so the other side makes your preferred decision on its own, then the defense cannot be purely technical. Firewalls, private networks, and better detection of coordinated inauthentic behavior all matter, and I don’t mean to minimize them. But they defend the contest for information. They don’t, by themselves, settle the contest over what it means.

That second contest happens in people.

It is tempting to propose a filter — better AI detection, more content moderation, more media literacy slides. Some of that will help, though every filter creates a new argument about who controls the filter. The more durable response is less centralized and less glamorous. It is the old set of protections that make a person harder to reroute: habits of attention, local knowledge, institutions small enough to verify.

I notice this in my own use of the phone. The same device that lets me check a fact in seconds makes it harder to sit with a question for more than a minute. The feed rewards quick classification — ally or adversary, serious or absurd — before I’ve finished reading. Over months that trains a reflex. The world sorts itself into familiar villains and familiar victims, and the most useful response always seems to be sharing one more piece of evidence with people who already agree with me.

Fifth-generation conflict benefits from that reflex precisely because it doesn’t need to invent it. It can simply amplify what already moves quickly.

The alternative is not to log off and wait for the institutions to catch up. It is to cultivate the small practices that restore judgment. Knowing a neighborhood, a workplace, a congregation, a physical discipline — any setting where information has costs and where you can check a claim against people you know — gives you a baseline that a synthetic narrative has to work harder to displace. Verifying before amplifying is a modest discipline, but at scale it functions like herd immunity. Keeping limits on what data you broadcast — location, contacts, routines — shrinks the surface that can be modeled and nudged.

None of this guarantees you will see the operation while it is happening. The best defenses against perception manipulation are not dramatic. They are what the worldview that guides this site would call virtues formed through repeated action — attention, truthfulness, restraint, willingness to be corrected — and they are valuable even apart from any adversary.

There is a reason Abbot’s line about the most successful wars never being identified has endured. It captures a change in how power can be exercised. Territory is expensive to occupy. Machines are expensive to replace. Perception, by contrast, can be shaped at the margin with tools that are already rented out for advertising. If you can shift the context in which a few thousand decision-makers interpret an event, you do not need to control the event.

That also explains why the institutions are struggling to name it. A phenomenon that crosses the line between war and peace, between crime and influence, between foreign and domestic in a single afternoon does not fit neatly into one agency’s charter. Asking Congress to produce a definition is a small step toward admitting that gap.

I don’t think the right conclusion is to treat every confusing episode as a covert operation. That way of thinking is itself a form of vulnerability — every cricket becomes a weapon, every coincidence becomes evidence. The healthier suspicion is closer to housekeeping. Keep your network patched, but also keep your judgments patched. Notice what makes you feel rushed, righteous, or certain that the other side is not merely wrong but illegitimate. Ask where that feeling came from and whether someone benefited from speeding it up.

The monk’s old remedy for distraction was not elaborate. Stay in the cell. In more ordinary terms: remain where you have good reason to remain, do the next real task that belongs to you, and let the algorithm wait. A war that depends on stealing attention is partly defeated by people who learn, quietly and without announcement, to give their attention to what they have already chosen.

That will not sound like a defense strategy. But for a form of warfare that wins by being unnoticed, the most reliable defense may be a life that is less easily moved.

She doesn’t need to go outside. Her water bowl is full. There’s no mysterious emergency involving a toy trapped under the couch. She just walks in, sees that I’m there, and settles nearby.

This is one of the nicest things about having a dog.

There is an important qualification here. Cleo comes looking for me only if my wife isn’t available. My wife is very clearly her favorite, and Cleo has never made much effort to hide this. I’m the backup option. Still, second place in the affections of a small goldendoodle is not a bad place to be.

For a few minutes before she appears, Cleo has been somewhere else in the house doing whatever dogs do when nobody is watching. Sleeping, probably. Staring out a window, on sentinel duty for bunnies. Then, at some point, something happens in that little brain of hers that amounts to: Where are my people?

And she comes looking.

Cleo is a mini goldendoodle, short for Cleopatra. She’s named after the main character in the Cleopatra in Space series, although she has considerably fewer intergalactic adventures. Most of her expeditions involve moving between rooms of the house in search of my wife or, failing that, me.

The children apparently do not count.

She can be in a house full of kids and somehow still regard herself as tragically alone. They feed her things. They pet her. They are warm-blooded human mammals who speak and move around. None of this seems sufficient to establish personhood in Cleo’s mind. If my wife and I leave, she has been abandoned.

I find this funny, but also strangely sweet.

We spend a lot of time thinking about how much our pets mean to us. We take pictures of them. We talk to them despite knowing perfectly well that they don’t understand most of what we’re saying. We buy them ridiculous things and organize parts of our lives around feeding them, walking them, and making sure they’re happy.

It’s easy to forget that some version of the relationship runs the other direction too.

Dogs notice when we leave and when we return. They learn our routines. They know the sounds of our cars, the rooms where we usually sit, and the places they’re likely to find us. Sometimes they realize we aren’t nearby and decide to do something about it.

There’s something lovely about knowing that you can exist in the mind of an animal while you’re somewhere else.

Cleo doesn’t have any complicated theory of love or companionship. She just knows there are two people in the world she particularly wants to be near, and I’m fortunate enough to be one of them (when the preferred option is unavailable).

So every now and then I’ll be sitting somewhere reading or working and hear the little tap of paws coming down the hallway.

A moment later, Cleo appears in the doorway.

She was somewhere else in the house.

Then, for whatever reason, she thought of me.

And came to sit nearby.

We have this exactly backward.

Spend enough time online and you can assemble a pretty convincing case that everything is terrible. Politics is broken. Technology is destroying us. The economy is rigged. The culture is decaying. AI will take everyone’s jobs. Nobody can afford a house. Nobody is having children. The climate is changing. Institutions are failing. Somewhere, at this very moment, a guy with a podcast microphone is explaining why Rome fell.

Those problems are real, and deserve serious attention. But there is a difference between noticing problems and adopting doom as a disposition.

The latter is a terrible way to live.

The world is also astonishingly full of possibility. Most of us carry a device in our pocket that gives us access to nearly the entire accumulated knowledge of mankind. You can learn almost anything for essentially nothing. You can start a company from your kitchen table. You can make something in Florida and sell it to someone in Finland. You can publish an essay without asking an editor, make a film without a studio, write software without a computer science degree, or have a machine teach you calculus at two in the morning.

A reasonably ordinary person has powers that emperors would have found difficult to comprehend.

And the interesting thing about abundance is that it compounds. Someone invents a tool. Someone else uses that tool to make another tool. A third person combines both with an idea borrowed from somewhere else. The frontier moves outward, usually in ways nobody planned.

This doesn’t guarantee a good future. Nothing does. Human beings remain perfectly capable of making a mess of things. But that’s precisely why optimism isn’t the belief that everything will work out. It’s the belief that there are things worth doing.

So build something.

Take the trip. Start the business. Write the book. Have the kid. Plant the tree. Learn to weld. Ask the girl out. Invite people over for dinner. Apply for the job you’re not quite qualified for. Make the ridiculous prototype. Say hello to the stranger standing next to you.

Walk around like God sent you.

I mean that literally. There is a remarkable difference between moving through the world as though you are an accidental consumer of it and moving through it as though you have been given something to do here. The second posture changes the way you see people. They stop looking like obstacles, competitors, demographic categories, or anonymous extras in your personal movie. You start noticing them.

Smile at people. Be generous. Expect good things from them. Give more than the situation requires. Most of this costs almost nothing, and it makes ordinary life considerably more pleasant.

There will always be reasons to be afraid. Fear is quite good at finding evidence for itself.

But so is hope.

Look around. There are eight billion people thinking, tinkering, loving, building, fixing, arguing, inventing, raising children, making dinner, and trying again tomorrow. The story is nowhere close to finished.

Life is electric.

Act accordingly.

That history tends to be scattered through execution traces, failed attempts, optimizer logs, benchmark results, and conversations. A useful lesson gets discovered, turned into a skill edit, and then partly disappears into the edit itself. If the edit doesn’t work and gets rolled back, the lesson can disappear almost completely.

A new paper from Google Research and Virginia Tech, WikiSkill: Compiling Agent Experience into Persistent Knowledge for Skill Evolution, proposes a surprisingly straightforward fix: give the agents a wiki.

Not Wikipedia, obviously. WikiSkill adds a persistent layer of structured knowledge between an agent’s raw experience and the executable skills it uses to do work. That extra layer turns out to do quite a lot.

Three kinds of memory

Most skill-evolution systems follow roughly the same loop. An agent attempts tasks, the system inspects what went right and wrong, proposes changes to its skills, tests those changes, and keeps the ones that improve performance. The researchers argue that this process tends to blur together three different artifacts that should remain separate.

The first is raw experience: the actual execution traces containing the agent’s reasoning, tool calls, outputs, failures, and successes. WikiSkill keeps these immutable. They are the record of what happened.

The second is accumulated knowledge. This is the wiki. A maintainer agent studies the traces and turns them into structured Markdown pages describing recurring failure modes, successful strategies, workarounds, prior skill proposals, and whether those proposals actually helped. The wiki persists across iterations and is continuously revised as more evidence arrives.

The third is executable skill: the concise procedural instructions the working agent actually receives. Each skill also carries a link back to the wiki patterns that motivated it.

That separation feels obvious once you see it. A log is not knowledge, and knowledge is not an instruction. We already make similar distinctions when humans build complicated systems. There are telemetry and incident logs, documentation explaining what the team has learned, and production code that embodies some subset of those lessons. Trying to make one artifact perform all three jobs usually leaves each job a little worse.

WikiSkill gives each layer a different lifecycle. Raw traces are preserved. Knowledge accumulates. Skills can change or even be rolled back.

The last part is especially clever. A proposed skill update is tested on a validation set. If it makes the agent worse, the skill change is discarded. The wiki is not. It keeps a record of the proposal, the evidence behind it, and the fact that it failed. A later agent can therefore learn from an unsuccessful intervention without having to keep executing the unsuccessful instruction.

Failure becomes part of the institutional memory rather than clutter in an optimization log.

The wiki seems to be doing real work

The researchers tested WikiSkill across five rather different benchmarks: mathematical reasoning, web search, spreadsheet manipulation, long-document question answering, and interactive embodied tasks. They used models from the Qwen, Gemma, and Gemini families and compared WikiSkill with several other approaches to automatically evolving skills.

Across the five models, WikiSkill produced the best average performance. But the cleaner result comes from the ablation study, where the researchers removed the persistent knowledge layer while leaving the general skill-evolution process intact.

With Gemini 3.5 Flash, allowing the skill proposer to use the accumulated wiki raised average performance across four benchmarks from 48.7% to 63.7%. On LiveMath it went from 51.3% to 72.6%; on SpreadsheetBench, from 49.9% to 76.6%. The system was not merely getting better because it had another round of prompt editing. Much of the improvement came from giving the editor an organized memory of what previous rounds had discovered.

There is an amusing wrinkle here. Giving the working agent access to the wiki during training actually made the final skills worse. The researchers think the agent could solve some problems by consulting the wiki directly, which produced less informative failures and therefore poorer training material. The better arrangement was asymmetric: let the agent work from its skills, let it fail naturally, and let the agents responsible for learning study the larger body of accumulated knowledge.

The wiki is a workshop manual, not something that needs to be stuffed into every worker’s head.

Skills can substitute for a surprising amount of scale

The model-scaling results are probably the easiest part of the paper to notice. Within the Qwen family, larger models got more benefit from evolved skills, not less. WikiSkill improved average performance by 12.3 points for the 4B model, 17.5 points for the 9B model, and 23.9 points for the 27B model. Better models appear to be better at making use of good procedural knowledge.

But good procedural knowledge can also compensate for a lot of model size. Qwen 3.5 9B with WikiSkill averaged 47.4%, while the much larger Qwen 3.6 27B without skills averaged 39.4%.

That is a useful result for anyone building agents because model choice is only one place capability can live. Some of it can live in the model. Some can live in tools. Some can live in the surrounding workflow. And some can apparently live in a collection of well-evolved Markdown files.

The practical comparison is therefore not always “which model is smartest?” It may be “which system knows how to do this job?” A smaller model arriving with a good operating manual can beat a larger model arriving on its first day.

This also changes the economics a little. If procedural knowledge can be accumulated outside model weights, organizations don’t have to repurchase all of their learned competence every time they change models. The expensive part of an agent system may gradually become less like buying intelligence and more like building institutional knowledge around intelligence.

The skills don’t belong to the model

The cross-model experiments push that idea further. Skills evolved by one model often worked when handed to another model, including models from different families. Sometimes the imported skill worked better than the skill the receiving model had evolved for itself.

On ALFWorld, for example, Qwen 3.5 9B scored 63.4% using its own evolved skill but 70.2% using a skill evolved by Qwen 3.6 27B. On LiveMath, skills evolved by the small Qwen 3.5 4B model raised Gemma 4 31B from 33.9% without skills to 73.1%, beating Gemma’s 56.7% with its own evolved skills.

The transfers aren’t universally good. A spreadsheet skill evolved by Qwen 3.5 4B badly hurt Gemini 3.5 Flash because it contained low-level workarounds useful to the smaller model but restrictive for the stronger one. That failure is almost as informative as the successes. A skill can contain general knowledge about a task, or it can contain a workaround for the peculiar weaknesses of the agent that wrote it.

Still, the successful transfers suggest a useful separation between discovering a good procedure and executing it. The model that figures out how to do something doesn’t necessarily have to be the model that later does it.

That makes a shared skill library start to look less like a folder of prompts and more like a portable layer of organizational capability. A strong model could be used to discover and refine procedures, while cheaper models execute them at scale. Or several different models could contribute lessons to the same body of knowledge and inherit procedures discovered by the others.

A skill library needs more than skills

I’ve tended to think about agent improvement in terms of the executable artifact: make a SKILL.md, use it, notice a problem, edit it, repeat. WikiSkill makes me think that this leaves out the most valuable artifact created by the process.

Every time an agent fails, someone learns something. Every time a workaround succeeds, there is a reason it succeeded. Every rejected skill edit contains information about what didn’t generalize. If all of that gets compressed immediately into the current version of the skill, the system remembers the answer while gradually forgetting the argument that produced it.

Humans have run into this problem before. Mature organizations accumulate more than procedures. They accumulate design documents, incident reports, case law, laboratory notebooks, maintenance records, recipes with scribbled corrections, and old hands who remember why the strange rule exists. The procedure is useful because a much larger body of experience sits behind it.

WikiSkill is an early research system, tested on benchmarks rather than a years-old production agent fleet. Its wiki is maintained by models, and long-running knowledge bases will eventually face the familiar human problems of stale documentation, contradictory lessons, bad abstractions, and accumulated junk. The paper doesn’t make those problems disappear.

But I like the architecture because it gives those problems somewhere sensible to live. Keep the traces as evidence. Compile repeated experience into knowledge. Compile the best current knowledge into skills. Test the skills. Keep learning even when a particular edit fails.

If agent skills become as common as they currently appear likely to, I suspect the most useful skill libraries won’t just be libraries. They’ll have a memory behind them.

Sometime in the second century BCE, the Greek historian Polybius tried to do roughly that. In Book VI of his Histories, he described a recurring sequence by which political orders emerge, mature, corrupt themselves, collapse, and begin again. The theory is usually called anacyclosis—a cycling or recurring movement—and it gives us one of antiquity’s most memorable attempts to turn political history into a pattern.

The diagram makes the theory look almost mechanical. Monarchy becomes tyranny. Aristocracy becomes oligarchy. Democracy becomes mob rule. Disorder eventually produces another ruler, and the wheel begins another turn.

Polybius’ actual argument is more interesting than the diagram because the mechanism isn’t really constitutional. It’s human.

Polybius inherited a familiar Greek way of classifying governments. Plato and Aristotle had already distinguished political systems according to whether power rested with one person, a few, or the many, and whether those rulers governed well or badly. Polybius turned those categories into a story unfolding through time.

He begins before anything we would recognize as a constitution. After catastrophe scatters and weakens human communities, people gather together again. The strongest person naturally assumes leadership. Strength alone, however, isn’t yet kingship. A king emerges when people begin to recognize justice and choose to follow a ruler for his judgment rather than merely fear his power. Political authority has moved from force toward legitimacy.

Then the children inherit it.

This is the recurring hinge in Polybius’ cycle. The generation that creates an institution usually remembers why it exists. The generation that inherits it receives the institution without necessarily receiving the experiences that produced it.

The first kings, in Polybius’ telling, live much like everyone else. Their descendants grow up surrounded by security, status, and abundance. They begin to expect special clothes, special food, special treatment, and the satisfaction of their appetites. Kingship becomes tyranny. The ruler who once governed because people trusted his judgment now governs through fear.

The tyrant eventually produces his own opposition. Polybius says the people most likely to resist are the noblest and most courageous citizens, precisely because they are least willing to tolerate his arrogance. They overthrow him, and because they have liberated the community, the people entrust them with government. Tyranny gives way to aristocracy: rule by the few who have demonstrated enough virtue to deserve it.

And then their children inherit it.

The pattern repeats. The founders of the aristocracy remember oppression and treat public office as a responsibility. Their descendants have known neither oppression nor equality. Some become greedy; others indulgent; political authority becomes a possession rather than a trust. Aristocracy decays into oligarchy.

Eventually the people have enough. The oligarchs are killed or driven out, but the memory of both monarchy and aristocracy is now poisoned. The citizens will not hand power back to one man or a small group, so they take responsibility for government themselves. Democracy is born.

At first, Polybius says, the memory of oligarchy keeps democracy healthy. People who remember losing their freedom value equality, law, and freedom of speech because they know what their absence feels like. But memory fades again. By the time government passes to the grandchildren of the democratic founders, those goods have become ordinary. Ambitious wealthy men discover that they can gain power by courting and corrupting the people. Citizens grow accustomed to receiving benefits from political patrons. Competition for popular favor becomes a competition in indulgence.

Democracy then becomes what Polybius calls ochlocracy—usually translated as mob rule. Law loses its authority. Political contests become contests of force. The crowd can plunder, banish, and kill. Eventually the community reaches something close to the savage condition from which the cycle began, and in the chaos it again accepts a master capable of restoring order.

The wheel has made a complete turn.

The grandchildren problem

It’s tempting to read anacyclosis mainly as a taxonomy of governments. I think it’s more revealing as a theory of inheritance.

Again and again, Polybius locates political decay in the distance between the generation that learned a lesson and the generation that merely inherited its result. The king’s descendants inherit authority without remembering the conditions that made authority legitimate. The aristocrats’ descendants inherit prestige without the courage that earned it. The grandchildren of the democratic revolution inherit liberty without remembering domination.

The institution survives. The memory that disciplined it does not.

That helps explain why Polybius treats constitutional decline almost like a natural process. He compares political systems to physical materials that contain the cause of their own destruction. In his analogy, iron has rust and timber has worms; each simple constitution likewise carries a characteristic corruption within itself. Kingship tends toward despotism, aristocracy toward oligarchy, democracy toward lawless violence.

The analogy is stronger than saying that governments sometimes fail. Polybius is saying that their virtues create particular temptations. Concentrated authority makes decisive and virtuous kingship possible, but the same concentration makes tyranny possible. Entrusting the best citizens with power can produce excellent government, but it also creates a privileged class capable of serving itself. Popular rule protects equality and freedom, but popular sovereignty can become untethered from the laws and customs that made freedom sustainable.

The corruption isn’t foreign to the regime. It grows from the same structure that made the regime work.

Rome and the attempt to stop the wheel

Polybius wasn’t developing this theory as an abstract exercise. He was trying to understand Rome.

He had unusually good reason to wonder about it. Born in the Greek city of Megalopolis around 200 BCE, Polybius belonged to the political elite of the Achaean League. After Rome defeated Macedon in 168 BCE, he was among the Greek hostages taken to Italy. He eventually became close to Scipio Aemilianus and gained an extraordinary vantage point from which to watch the republic that had conquered much of the Mediterranean world.

The central question of the Histories is essentially how Rome managed to do it. In Book VI, constitutional structure becomes part of his answer.

If every simple form of government eventually corrupts itself, the obvious solution is not to choose the right point on the wheel. It is to stop relying on a simple constitution at all.

Polybius believed the lawgiver Lycurgus had done this deliberately at Sparta. Rome, he thought, had reached something similar more gradually, through experience and repeated political struggle. Its constitution contained all three good forms at once. The consuls supplied a monarchical element, the Senate an aristocratic one, and the people a democratic one.

More importantly, none could operate entirely on its own. The parts could restrain one another. The people depended on the Senate in some matters; the Senate had reason to fear popular resistance; consuls possessed formidable authority but still depended on the other institutions. Polybius described the result as a balance in which each part could check the excesses of the others.

The goal wasn’t to eliminate power. It was to arrange power so that one kind of power had difficulty becoming absolute.

This idea of the mixed constitution had a long life. Cicero later developed his own account of Rome’s mixed government. More than fifteen centuries after Polybius, Machiavelli’s Discourses on Livy returned to almost the same sequence of monarchy, tyranny, aristocracy, oligarchy, and popular government, and likewise praised Rome for combining the three forms rather than allowing one to rule alone. The classical argument about mixture, balance, and institutional restraint continued through later republican political thought.

There is a nice irony here. Polybius set out to explain how constitutions decay, but his more durable contribution may have been the attempt to design around decay. He did not imagine that better institutions would make rulers virtuous. His mixed constitution assumes almost the opposite. Consuls, senators, and citizens will each be tempted to press their advantages. Stability comes partly from giving the others enough power to resist them.

A map, not a law

As history, anacyclosis is much too neat. Governments do not obediently move around a six-stage wheel. Republics are conquered. Monarchies reform themselves. Aristocracies coexist with democratic institutions. Wars, plagues, technologies, religions, trade, geography, individual rulers, and accidents all push political development in directions no diagram can capture.

Polybius himself was also describing Rome at a particular moment and with a certain amount of admiration. The Roman constitution was not quite the balanced machine his analytical categories can make it appear to be. And within a century of the period he was trying to explain, the Republic would be consumed by civil wars and replaced by imperial rule.

But I don’t think the diagram survives because anyone seriously believes history runs on rails.

It survives because Polybius noticed a smaller pattern inside the larger one. Political orders are built by people who remember a problem. Their successors inherit the solution. Eventually someone inherits the power without the memory, the freedom without the discipline, or the institution without the habits that allowed it to work.

The names of the regimes are the visible part of the cycle. Underneath them is a simpler observation about human beings: we are very good at learning from disaster, and not always very good at inheriting the lesson.

View the full sized version.

In 1966, the linguist Robert B. Kaplan published a paper about a problem he kept seeing in the writing of international students at American universities. Many of them knew English quite well. Their grammar worked. Their vocabulary worked. Their sentences, taken one at a time, worked. And yet their essays could still feel strangely difficult to an American reader. The argument seemed to wander, repeat itself, arrive late, or refuse to arrive at all.

Kaplan wondered whether the problem was happening above the level of the sentence. Perhaps students had learned the words and grammar of English without learning an equally real but less visible part of the language: the expected shape of an argument.

He eventually reduced the idea to five little drawings.

The first is a straight line. That is English. Another moves back and forth in parallel lines, representing what Kaplan called Semitic languages. A spiral represents what he called Oriental languages. Romance and Russian writing wander away from the central line in different kinds of digressions.

It is the sort of diagram that is almost impossible to forget and probably too easy to believe.

Kaplan’s original article, “Cultural Thought Patterns in Inter-Cultural Education,” became one of the foundational texts of what came to be called contrastive rhetoric. He had examined roughly 600 compositions written by foreign students and argued that rhetorical organization was not simply a neutral container into which ideas were poured. Different cultures had developed different conventions for arranging ideas, and students carried some of those habits with them when they wrote in English.

His strongest version of the claim went further. “Logic,” he wrote, using the word in its everyday rather than formal sense, “is evolved out of a culture; it is not universal.” English prose inherited a broadly Anglo-European preference for linear development: establish the subject, state or imply the controlling idea, and then proceed toward it without unnecessary detours. Other rhetorical traditions had developed differently.

Kaplan was not saying that an Arabic speaker was incapable of linear thought or that a Japanese speaker literally thought in circles. He explicitly warned that his categories were not mutually exclusive, and he noted that English writers could be circular or wildly digressive too. His concern was expository prose and, more specifically, what a reader had been trained to recognize as a well-made paragraph. The practical conclusion was modest enough: if we explicitly teach foreign students English grammar, perhaps we should also explicitly teach them English rhetorical conventions.

Still, the drawings invited a much bigger interpretation than the caveats could contain.

A straight line looks efficient. A spiral looks evasive. Once English occupies the position of directness, the other patterns can begin to look like deviations from clear thought rather than alternative conventions for communicating it. The terminology has aged even worse. Grouping Chinese, Korean, and Japanese writing under “Oriental,” or treating enormous linguistic and literary traditions as though each possessed a characteristic doodle, now looks impossibly broad.

Later scholars made those objections directly. Ryuko Kubota and Al Lehner argued in “Toward critical contrastive rhetoric” that the traditional model encouraged static binaries: English as linear, direct, and logical; other traditions as circular, indirect, or digressive. It could turn a useful observation about convention into a story about deficiency. There was also a basic problem with Kaplan’s evidence. He was looking largely at English compositions written by students learning English, then reasoning backward from those texts toward the rhetorical character of their first languages and cultures. A student essay written under the strain of a second language is a rather shaky foundation for a theory of civilization.

The field Kaplan helped create gradually became more careful. Ulla Connor’s later account of intercultural rhetoric moved away from treating national cultures as fixed rhetorical systems and toward studying genres, institutions, discourse communities, individual writers, and the circumstances in which a text is produced. A Chinese engineering paper, a Japanese sales letter, and a Korean university essay do not become instances of one spiral simply because their authors come from East Asia. A lawyer and a novelist writing in the same language may have more immediate rhetorical differences than two lawyers writing in different ones.

That correction makes Kaplan’s diagram less grand. I think it also makes the part worth keeping easier to see.

The straight line is not a picture of the English-speaking mind. It is a picture of an expectation.

If you grew up writing essays in English, you probably absorbed that expectation so early that it stopped feeling like a convention. School taught you to state a thesis, give the reader a topic sentence, support the claim, remove the irrelevant material, and make the relationship between one paragraph and the next visible. “Get to the point” sounds like advice about thinking, but it is also advice about the obligations a writer has toward a particular kind of reader.

Once enough people share those obligations, they disappear into the background. The writer knows where the thesis belongs. The reader knows where to look for it. A paragraph that announces its subject and then develops it feels orderly because both people have learned the same dance. Someone trained in another convention can violate those expectations while making perfectly sensible choices according to a different set of them.

We do this in smaller ways all the time. A legal brief, a newspaper story, a scientific paper, and a personal essay can contain the same facts and arrange them almost beyond recognition. The newspaper writer may put the conclusion in the first sentence. The scientist may spend pages establishing method before reaching it. The lawyer may lead with the rule and organize every fact around its application. A personal essay can withhold its point until the last paragraph and be better for it. Nobody concludes that the scientist thinks less directly than the reporter. We understand that each form has inherited expectations about what the reader needs, when the reader needs it, and how much work the reader should be asked to do.

Culture works on writing in a messier version of the same way. So do education, profession, class, religion, genre, audience, and individual temperament. They overlap. They change. Writers learn new forms and deliberately break old ones. This is a much less satisfying theory to draw on a chalkboard because it probably requires hundreds of intersecting lines instead of five memorable doodles.

But Kaplan saw something real when he noticed that grammatical correctness did not solve his students’ problem. They were encountering a layer of language that native speakers often fail to notice precisely because they have mastered it. The students were not merely learning how English sentences work. They were learning what an English-speaking academic reader had been trained to expect an argument to do.

That observation has consequences beyond the ESL classroom. We routinely confuse conventions with universals when the conventions are old enough and familiar enough. Table manners feel like manners rather than one culture’s manners. Our sense of appropriate distance in conversation feels like appropriate distance. A meeting that begins without small talk can feel rude to one person and admirably efficient to another. The invisible rules are most invisible to the people who know them best.

Writing may be especially prone to this confusion because prose gives us access to another person’s thoughts only after those thoughts have been arranged for us. When the arrangement matches our expectations, we call the writing clear. When it does not, we may conclude that the thinking behind it is confused. Sometimes it is. Sometimes the writer simply failed to follow the conventions of the audience. Those are not the same failure.

There is also a useful humility in realizing that English’s straight line is not quite as straight as the diagram suggests. Good English prose constantly violates the schoolroom version of linearity. It delays information for suspense, digresses for texture, repeats for rhythm, approaches difficult subjects obliquely, and lets an image carry an argument that could have been stated directly. Kaplan himself acknowledged literary exceptions. Even expository English has more freedom than the doodle gives it. Kubota and Lehner’s critique points out that actual English-language school texts do not always obey the simple direct-and-deductive model that students are often taught as the English pattern.

The convention remains useful anyway. If I am trying to explain a complicated idea, the straight line is usually a pretty good place to begin. I want the reader to know what I am talking about. I want each paragraph to earn its place. I don’t want a detour to exist merely because I enjoyed taking it. Those preferences are so deeply embedded in how I learned to write that they feel almost indistinguishable from good writing itself.

Kaplan’s diagram is a reminder to separate those two claims. Some writing really is muddled. Some arguments really do evade the evidence, lose their subject, or bury the point beneath material that does not belong. But clarity always has an audience, and an audience arrives with habits. We can judge an argument while still recognizing that some of the standards by which we judge its presentation were taught to us.

The old doodles are too crude to tell us how cultures think. They may be just crude enough to show us that our own straight line was learned too.

Among it all was a corroded lump of bronze and wood. It didn’t look like much until it broke apart and someone noticed teeth inside it. Gear teeth.

The bronze wheels were cut with surprising precision and packed together in a mechanism built sometime around the second or first century BC. The fragments eventually revealed what we now call the Antikythera mechanism: a hand-operated astronomical computer that could represent the motions of the heavens, track calendars, predict eclipses, and model the irregular motion of the Moon.

It had been sitting underwater for roughly two thousand years.

Fragment A of the Antikythera mechanism. National Archaeological Museum, Athens. Photo: Marsyas/Wikimedia Commons.

A universe in a box

Calling the mechanism a computer can sound like we’re trying to make an ancient artifact seem more modern than it was. In this case, though, the description is pretty literal.

The mechanism seems to have been housed in a wooden case roughly the size of a large shoebox, with dials on the front and back and inscriptions explaining how to use it. Turn a crank or knob to select a date and the gears inside translated that input into several astronomical outputs. Pointers showed the positions of the Sun and Moon. A small black-and-white ball showed the Moon’s phase. Other displays tracked longer cycles and indicated when eclipses might occur.

It wasn’t modern astronomy. The mechanism worked from the geocentric model of its time. But its maker had figured out how to turn astronomical theories into physical ratios. If one celestial cycle took a certain number of days relative to another, gears with the right numbers of teeth could reproduce the relationship mechanically.

That is, in a very real sense, computation—just done with bronze wheels instead of transistors.

The Moon is probably my favorite part. Its apparent speed across the sky isn’t constant, and the mechanism reproduced that variation with a clever pin-and-slot arrangement coupled to gears. A mathematical account of lunar motion became a changing mechanical speed. More than a thousand years before the great mechanical clocks of medieval Europe, someone was building a machine that could imitate the Moon’s uneven progress across the sky.

Modern reconstruction of the mechanism’s front. Photo: Mogi Vicentini/Wikimedia Commons.

The back was just as elaborate. Two large spiral dials represented long calendars. One tracked the 19-year Metonic cycle, in which 235 lunar months come remarkably close to 19 solar years. Another tracked the Saros cycle of roughly 223 lunar months, which could be used to anticipate eclipses. Inscriptions inside the eclipse dial seem to have described characteristics of the expected event.

There was even a calendar for athletic games, including the Olympics. I like this detail. The same little machine that modeled the heavens also reminded its owner when it was time to go watch people throw javelins.

Only about a third of the original mechanism survives, in 82 fragments, so every reconstruction is partly an argument. Much of what we know has come from looking inside metal that can’t simply be taken apart. X-ray and CT imaging have exposed hidden gear teeth and inscriptions. Researchers have spent decades matching fragments, counting teeth, reconstructing missing wheels, and checking the results against what we know of Greek astronomy.

A 2021 reconstruction by researchers at University College London proposed a solution for much of the missing front gearing, including displays for the five planets known in antiquity. It’s an impressive reconstruction, but it isn’t a recovered blueprint. Large parts of the front are gone, and other interpretations remain possible.

Even the number of holes around one damaged calendar ring is still being debated. In 2024, researchers at the University of Glasgow applied Bayesian analysis originally developed for gravitational-wave astronomy to measurements of the surviving holes. Their analysis favored 354 divisions, consistent with a lunar calendar, rather than the 365 one might expect for a solar calendar. Other Antikythera researchers have challenged the conclusion. There is something wonderful about scientists using techniques developed to study ripples in spacetime to argue over how many holes a Greek machinist drilled into a bronze ring two thousand years ago.

One reconstruction of the mechanism’s gearing, including proposed gearing for the five known planets. Wikimedia Commons.

The machine and its maker

We don’t know who built it.

The ship probably sank sometime in the first century BC while carrying luxury goods through the Mediterranean, though the mechanism itself may have been older. Its inscriptions and design place it firmly in the Greek intellectual world. Scholars have proposed origins around Rhodes and connections to traditions associated with Hipparchus, Posidonius, or Archimedes. None of that gives us the name of the person who actually sat at a bench and cut the gears.

I find that person almost as interesting as the machine.

Astronomy supplied the theory, but theory doesn’t make a gear turn. Someone had to choose the tooth counts, lay out the wheels, cut bronze teeth small and regular enough to mesh, fit several gear trains into a compact case, engrave instructions, and make the finished device usable by somebody else. Mathematics, astronomy, metalworking, and ordinary workshop skill all meet inside this one object.

It doesn’t look like a crude first experiment either. The density of the design and the range of functions suggest a craft that had already developed some sophistication. There are signs of mistakes and repairs, and recent research has raised questions about how smoothly reconstructions using the surviving geometry would actually have run. Ancient precision was still ancient precision. If anything, that makes the machine easier to appreciate. Its maker was solving a difficult mechanical problem with hand tools, bronze, geometry, and whatever knowledge could be passed around a workshop.

I keep trying to imagine seeing it when it was new: opening a wooden case, turning a handle, and watching the cosmos move.

What happened to the others?

Eventually the mechanism raises a question its gears can’t answer: where are the other ones?

No other geared machine of comparable complexity survives from antiquity. That absence once made Antikythera seem almost impossible, as though somebody had dropped an object from a much later century onto a Roman ship. But ancient writers left hints that devices like it weren’t unimaginable. Cicero described mechanical celestial models associated with Archimedes and, in his own lifetime, Posidonius. Those accounts don’t tell us what gears were inside them. They do suggest that mechanically representing the heavens was a recognizable kind of Greek technology rather than an idea modern archaeologists invented to explain one strange wreck.

And this is where I think the wreck itself is worth remembering.

Bronze is valuable. Machines get dismantled. Metal gets melted down and reused. Wood rots. Workshops burn. Devices become obsolete. Instructions disappear because everyone who knows a craft assumes someone will be around to teach the next apprentice. A complicated bronze machine has almost every quality you could want in an artifact that won’t survive two thousand years.

Antikythera survived because the ship carrying it sank. The sea put it beyond the reach of generations that might otherwise have repaired it, scavenged it, recycled it, or simply thrown it away.

We can count what survived. We can’t count what disappeared.

That doesn’t mean the ancient Mediterranean was full of mechanical computers, or that some forgotten industrial revolution was about to happen. One extraordinary object can’t support that story. But the absence of other surviving machines can’t tell us they never existed either. Before the divers reached the wreck, we had zero Antikythera mechanisms. After they pulled up one ugly lump of bronze, our estimate of what an ancient craftsman could build had to change.

Surviving fragments of the Antikythera mechanism at the National Archaeological Museum in Athens. Photo: Peulle/Wikimedia Commons.

It’s a small caution about the smooth lines we tend to draw through technological history. Those lines are made from whatever evidence happened to reach us. Stone temples survive better than wooden workshops. Pottery survives better than textiles. Monumental inscriptions survive better than everyday instructions. A bronze machine on land is useful raw material for somebody else. A bronze machine sealed under the sea can become a message to people two millennia later.

But I don’t think the mechanism needs that larger lesson to be worth staring at. Someone in the ancient Greek world built a portable machine of dozens of gears that turned astronomical knowledge into motion. It tracked the Sun and Moon, anticipated eclipses, coordinated calendars, and may have displayed the wandering planets. Then it disappeared into the sea.

Two thousand years later, we’re still figuring out exactly what it could do—and I suspect that’s part of why I like it so much.

A more useful comparison is to put the debt next to the country’s income. U.S. nominal GDP was running at about $32.5 trillion a year in the second quarter of 2026, which puts the $40 trillion gross debt at about 123 percent of one year’s economic output. If a person earning $100,000 a year had debt in the same proportion, the balance would be about $123,000. Using debt held by the public instead of gross debt, CBO puts the ratio at about 101 percent of GDP this year, almost exactly the equivalent of $100,000 of debt against $100,000 of annual income.

That doesn’t suddenly make the national debt good, but it makes the scale less apocalyptic. A $123,000 balance against a $100,000 income is substantial, but it doesn’t sound like insolvency. Plenty of households take on mortgages larger than their annual income and remain perfectly capable of servicing them. The analogy has limits: mortgage lenders normally care about the payment relative to income, not just the balance; a mortgage is backed by a house and normally amortizes; a government can tax, issue currency and refinance debt indefinitely. Still, the comparison helped me see that the headline number and the underlying burden are not the same thing.

Those comparisons also expose the wrong intuition if taken too far. The federal government isn’t a very large household with an absolutely terrible credit-card problem. It is the issuer of the world’s dominant currency, the operator of the world’s deepest government bond market, and the borrower behind an asset that banks, governments and investors around the world use as a safe place to hold money.

I started reading about this after seeing The New York Times report on the $40 trillion milestone. I had assumed that I basically understood the national debt problem: America owes an enormous amount of money, keeps spending more than it collects, and at some point will run into the same constraint any borrower eventually does. What I hadn’t appreciated was how much the location of that constraint depends on the country doing the borrowing.

For many developing countries, the constraint arrives much sooner. A government may need to borrow in dollars or euros because investors are reluctant to lend large amounts in the local currency. Its taxes, however, are collected in that local currency. If the currency loses value, the foreign debt becomes more expensive without the country having borrowed another dollar. Investors demand higher yields to compensate for that risk, and the higher interest bill makes the government’s finances weaker still.

That can turn into a brutal cycle. Capital leaves, the currency falls, debt service rises, foreign-exchange reserves shrink and lenders demand still more compensation for risk. A government can quickly find itself choosing among default, tax increases and spending cuts during a recession. The World Bank’s latest International Debt Report found that low- and middle-income countries paid $741 billion more in principal and interest to external creditors between 2022 and 2024 than they received in new financing. They paid a record $415 billion in interest in 2024 alone, while new private borrowing was coming at rates around 10 percent.

At that point debt stops being an accounting problem. It becomes the road that isn’t repaired, the power plant that isn’t built, or money diverted from schools and hospitals. Domestic borrowing can do damage too because thinner capital markets make it easier for government borrowing to squeeze out private businesses that need credit. And when governments respond by creating money, weaker confidence in the currency can turn the debt problem into an inflation problem.

The United States has spent generations building a very different set of conditions. We borrow overwhelmingly in dollars, the currency in which we also collect taxes. The Federal Reserve can supply liquidity to dollar markets during a crisis. American capital markets are enormous, open and liquid. Property rights and the rule of law have given investors confidence that a Treasury bond is an unusually dependable claim. Behind all of that sits a large, productive economy with a broad tax base, deep private capital markets, abundant natural resources, technological leadership and a long record of economic growth.

The dollar’s international role grew out of those conditions and then began reinforcing them. It still accounts for about 57 percent of disclosed global foreign-exchange reserves, according to the IMF’s first-quarter 2026 data. Trade is commonly invoiced in dollars. Banks and companies need dollars for transactions. Central banks hold dollar reserves. Treasury securities have become basic collateral for the global financial system.

That creates a powerful feedback loop. Investors want Treasuries because the market is deep and liquid; the market becomes deeper because investors want Treasuries. Global demand lowers the interest rate the government would otherwise have to pay. During crises, when investors flee the assets of many other countries, they have often moved into dollars and Treasuries instead. The same event that can trigger a funding crisis somewhere else can increase demand for American government debt.

The difference isn’t just the currency. A country’s debt is easier to carry when its economy is growing quickly relative to the interest rate it pays. Economists sometimes summarize this relationship as r versus g: the interest rate on the debt compared with the growth rate of the economy. When nominal economic growth outruns borrowing costs, an existing debt burden can become smaller relative to the economy even without paying down the principal. The United States has benefited from long periods in which that relationship was favorable, helped by productivity growth, immigration, entrepreneurship and expanding output.

Debt-financed spending can also support growth in the short run. During recessions and emergencies, federal deficits put money into an economy when private demand is weak. Some of America’s stronger growth compared with slower-growing advanced economies has come alongside much larger fiscal expansion. That doesn’t make the debt free. It does mean the relationship between borrowing and growth isn’t simply that each additional dollar of debt subtracts a dollar of future prosperity.

Even who owns the debt makes a difference. A government funded by a broad base of domestic savers, institutions, pension funds, banks, foreign central banks and long-term investors is less exposed to a single class of lenders suddenly disappearing. Japan has carried a much larger debt relative to its economy than the United States in part because so much of it has been held domestically and because its central bank has been willing to buy large amounts of government bonds. The trade-offs there have been different, including decades of very low growth and extremely low interest rates, but it is another reminder that the headline debt ratio alone tells you surprisingly little about when a country runs into trouble.

America’s ability to borrow at this scale was not bestowed on it. It was built over a long time from economic output, institutions, military and political stability, credible markets, the development of the dollar system and the repeated willingness of people around the world to trust American assets. Each generation inherited more of that financial infrastructure than the one before it.

We have also been using more of it.

The current path looks more troubling than the $40 trillion snapshot by itself. The Congressional Budget Office projects a deficit of about $1.9 trillion this fiscal year, or 5.8 percent of GDP, even without a recession or national emergency. More than two percentage points of GDP are a primary deficit, meaning the government is still spending more than it collects even before interest is counted. CBO expects debt held by the public to rise from about 101 percent of GDP this year to 120 percent by 2036 and 175 percent by 2056. A household with a manageable mortgage becomes a different proposition if it keeps borrowing every year faster than its income grows and never intends to reduce the principal.

The cost is already showing up in the budget. CBO expects net federal interest expense to reach about $1 trillion this year, or 3.3 percent of GDP. By 2036 it projects about $2.1 trillion a year in interest payments, nearly as much as the federal government will spend on all discretionary programs combined. That money doesn’t buy a fighter jet, fund a scientific experiment, pay a Social Security benefit, build a highway or lower anyone’s taxes. It is payment for decisions already made.

There are quieter costs as well. Government borrowing competes with private borrowers for capital and can push interest rates higher over time. That means somewhat less private investment, more expensive mortgages and business loans, and slower accumulation of productive capital than we otherwise might have had. High debt also reduces the room available for the next war, financial crisis, pandemic or recession. The country can still borrow in an emergency, but it begins from a worse position and a larger share of every new dollar raised is already committed to servicing old obligations.

None of this tells us that $40 trillion is the number at which something breaks. There probably isn’t one magic number. A country with weak institutions, foreign-currency debt and frightened lenders can have a crisis at a debt ratio the United States would barely notice. A rich country with trusted institutions, its own currency and a productive economy can carry much more.

That is what changed how I think about the national debt. America’s unusual capacity to borrow is itself part of the country’s accumulated capital. It comes from things worth preserving: productive people and businesses, functioning institutions, liquid markets, confidence in contracts, confidence in the currency, and confidence that the country will still be capable of paying its obligations decades from now.

The danger isn’t that $40 trillion produces an automatic bankruptcy. It is that the strengths allowing us to carry so much debt make it easy to treat those strengths as permanent. Interest costs can consume more of the budget. Persistent primary deficits can keep pushing the debt ratio higher even in good economic times. Slower growth or higher rates can make the arithmetic worse. A loss of confidence in American institutions or the dollar would make borrowing more expensive precisely because so much borrowing has already been done.

Developing countries often discover their limits suddenly. The United States has spent a long time building enough economic and institutional capacity that our limits are much farther away and harder to see. I’d rather treat that distance as something we built and should preserve than as evidence that the limit doesn’t exist.

We now have machines that can produce a growing range of this work. I wrote recently about what happens when intelligence becomes cheap, and especially what that might do to the excuses we make when knowledge and competent advice are no longer particularly scarce. But there is a more basic question underneath that one: what actually makes intelligence cheap?

The obvious answer is better AI models. I don’t think that’s quite right. Training makes a model capable. Something else has to make that capability cheap enough, fast enough, and reliable enough that we can use it casually.

Philip Kiely’s Inference Engineering is largely about that something else.

Inference is what happens after a model has been trained, when someone actually asks it to do some work. Training is a project with a beginning and an end. Inference is an operation that can continue for as long as anyone wants to use the model. In fact, success makes the problem harder. A popular AI product creates more inference demand and therefore a larger bill.

That bill is surprisingly physical.

A large language model generates text one token at a time. During the generation phase, the limiting factor is often not how quickly the GPU can perform arithmetic but how quickly it can move the model’s weights through memory. Kiely uses the example of a 70-billion-parameter model represented in 16-bit values. Its weights alone occupy roughly 140 gigabytes. At a batch size of one, generating a token can mean moving something on that order through memory so the machine can produce a tiny addition to the sentence on your screen.

The answer feels almost weightless. Underneath it are racks of accelerators moving a great deal of data, consuming electricity, producing heat, occupying data centers, and tying up some very expensive capital equipment.

This makes the phrase “the price of a thought” less metaphorical than it first sounds. A token isn’t a thought, of course, and none of this requires taking a position on whether a machine thinks in the human sense. But an answer, translation, summary, classification, design, or plan now has a measurable computational cost. A growing class of cognitive work can be bought in units.

And engineers are steadily trying to make those units cheaper.

One way is batching. If the expensive part of generating a token is moving model weights through memory, it makes sense to use that movement to serve several requests at once. Another is quantization, which represents the model with fewer bits so there is simply less data to move. Caching avoids recomputing parts of a request the system has already seen. Speculative decoding lets a smaller, cheaper model guess several upcoming tokens and asks the larger model to verify them together rather than doing all the work serially.

These techniques sound like fairly obscure computer engineering, and in one sense they are. But taken together they are doing something economically familiar: getting more useful output from scarce capital.

There is no magic in it. Each optimization trades one scarce resource for another. More batching can improve throughput while making an individual request wait. Quantization can reduce memory use while risking some loss of quality. Caches consume memory of their own and are only valuable when the right requests return to the machines holding them. Speculation saves time only when the cheaper model guesses well enough to justify its own cost. Inference Engineering treats these as exchanges among memory, compute, latency, quality, and operational complexity. Engineers are looking for better exchange rates.

Something similar happened with electricity.

The spectacular part of early electrification was generating electricity at all. Edison opened the Pearl Street Station in Manhattan in 1882, but direct-current power could initially be delivered only a short distance from the station. By 1900 the United States had more than 3,000 central service stations, while alternating current and higher-voltage transmission were allowing power to travel much farther. In 1922, a 220-kilovolt line in California carried power more than 200 miles from the Sierra Nevada toward the San Francisco Bay Area. The history of the early electricity industry is partly a history of learning how to separate where power was generated from where it could economically be used.

The generator was an invention. The grid was a system.

Transmission lines, distribution networks, interconnected generating stations, spare capacity, standards, load balancing, and eventually a wall socket turned an impressive machine into an ordinary part of life. The economic effects followed the spread of the system. Research on American manufacturing from 1890 to 1940 finds rapid and persistent productivity gains from electrification, accompanied by investment and changes in how factories were organized. Other research finds that the expansion of high-voltage transmission between 1910 and 1940 helped drive the shift of American employment from agriculture toward manufacturing.

Factories didn’t simply replace a steam engine with an electric motor and carry on as before. Electricity eventually allowed individual machines to have their own motors. Factory floors could be rearranged around the work instead of around shafts, belts, and a central source of mechanical power. The technology became more valuable as people reorganized other things around its abundance.

I think AI is beginning to acquire its own version of this surrounding machinery.

The comparison shouldn’t be pushed too literally. GPUs aren’t power plants and tokens aren’t kilowatt-hours. But the structural problem looks familiar. We have learned how to produce a valuable resource and are now building systems for delivering it under unpredictable demand while trying to keep cost down and reliability up.

At small scale, this looks like software optimization. At larger scale it starts to look like infrastructure economics.

A company serving AI has to decide how many expensive GPUs to keep running while nobody is asking for them. Turning them off saves money, but starting a large model again can take long enough that the next user is left waiting. Keeping them running buys readiness. Autoscaling is partly the art of deciding how much readiness is worth paying for.

At still larger scales, the abstraction breaks down further. Kiely notes that once a deployment needs hundreds of GPUs, the problem can become less about whether the software knows how to scale and more about where the physical GPUs are actually available. Capacity exists in particular data centers, regions, and clouds. Network latency makes geography relevant. Suddenly the seemingly immaterial business of producing words has some of the same concerns as other infrastructure: capacity planning, utilization, redundancy, routing, and location.

This suggests another way to think about AI progress.

Most of our attention goes to the frontier: which model scores highest, has the largest context window, or can solve the hardest problem. Those advances are real. But there is another curve underneath them that may ultimately be just as consequential: how much useful cognition can a dollar buy?

Economic historians often describe electricity and information technology as general-purpose technologies because their effects spread through many industries rather than remaining confined to the business that invented them. Research comparing the two notes both the broad adoption of electricity and the continuing decline in the price of information technology. Neither became transformative simply because the best available machine got better. They became transformative because useful capability became cheap and available enough to put almost everywhere.

The same possibility is easy to imagine with inference. A model becoming somewhat smarter can improve the tasks we already give it. Making the same useful model ten times cheaper changes how often we are willing to call it. Software can ask for a classification that wasn’t worth paying for before. An agent can inspect ten possible approaches instead of one. A small company can afford a capability that previously required a large budget. Intelligence starts appearing in products where nobody would have bothered rationing expensive model calls.

There is a catch familiar from almost every efficiency improvement: cheaper units don’t necessarily mean we consume less in total. They often mean we find reasons to consume far more. Better inference could reduce the cost of a given AI workload while increasing the amount of inference civilization performs. The likely future isn’t necessarily a smaller collection of GPU data centers doing today’s work more efficiently. It may be much larger infrastructure doing kinds and quantities of cognitive work that aren’t economical today.

That is another reason the grid analogy seems useful to me. We didn’t respond to cheaper, more available electricity by deciding we had enough light bulbs. We built refrigerators, air conditioners, elevators, washing machines, factories, computers, and eventually data centers. Abundant electricity created uses for electricity.

Inference engineering is still a young discipline, and machine intelligence remains constrained by chips, memory, energy, networks, capital, geography, and the quality of the models themselves. There is no reason to assume the price falls smoothly or reaches zero.

But I suspect some of the biggest changes in AI will arrive without a dramatic new model at all. They will come from making yesterday’s remarkable capability cheap enough that tomorrow nobody thinks very hard before using it again.

Electricity became part of ordinary life when most of us stopped thinking about the power plant. Machine intelligence may reach a similar point when asking for one more small piece of cognitive work feels just as unremarkable.

Kramdown gives Jekyll sites a useful middle ground. Its Inline Attribute Lists, usually called IALs, let you add classes, IDs, and ordinary HTML attributes to Markdown elements without replacing the Markdown with raw HTML. That makes it possible to build richer editorial layouts while keeping the source readable.

For example:

This paragraph contains a short editorial aside.
{: .aside}

Kramdown renders that roughly as:

<p class="aside">This paragraph contains a short editorial aside.</p>

The same pattern works on headings, blockquotes, lists, images, code blocks, and many other block elements. Inline elements can take attributes too:

This is *emphasized text*{:.highlight} inside a paragraph.

Classes use the familiar .class-name shorthand and IDs use #id-name:

## The Second Act
{: #second-act .section-heading}

You can also use normal HTML-style attributes when there is a reason to:

[External documentation](https://example.com){: rel="nofollow" target="_blank"}

The Kramdown syntax documentation has the full grammar, but for most site work the useful mental model is simple: put a block attribute list immediately after the block it belongs to, and put an inline attribute list immediately after the inline element. Kramdown also supports reusable Attribute List Definitions, although I tend to prefer explicit semantic classes in content because they are easier to find and understand later.

Treat classes as content roles

The tempting way to use this feature is to put presentation directly into Markdown:

A big sentence.
{: .large-blue-text}

That works, but it creates the same problem inline styles create in HTML. The content now knows too much about its current appearance. A redesign either leaves old class names lying around or requires editing hundreds of posts.

A better approach is to name what the element is:

A big sentence.
{: .pull-quote}

Then the stylesheet decides what a pull quote looks like:

.pull-quote {
  font-size: var(--type-scale-4);
  line-height: 1.25;
  max-width: 28ch;
  margin-block: var(--space-8);
}

Start with a small component vocabulary

It is easy to create too many classes. I would start with a small set that covers recurring editorial needs and add a new one only when several pieces of content genuinely need the same treatment.

For a long-form Jekyll site, that might look something like this:

.lede             opening paragraph
.pull-quote       emphasized quotation or sentence
.aside            secondary editorial thought
.note             explanatory or contextual callout
.data-block       compact presentation of numbers or facts
.figure           standard editorial image treatment
.figure-wide      image that breaks beyond the text column
.caption          image or figure caption
.section-break    visual pause between major movements
.small            intentionally de-emphasized copy

Build the CSS in layers

The attribute vocabulary becomes more reusable when the CSS behind it has its own structure. A simple design-system stack might have four layers:

Tokens       colors, spacing, type scale, widths, radii
Base         headings, paragraphs, links, lists, figures
Components   .lede, .pull-quote, .note, .aside, .data-block
Utilities    layout helpers used mainly by templates/includes

Then a component is assembled from common tokens instead of inventing new values each time:

:root {
  --content-width: 44rem;
  --wide-width: 68rem;
  --space-4: 1rem;
  --space-8: 2rem;
  --radius-2: 0.5rem;
}

.note {
  padding: var(--space-4);
  border-radius: var(--radius-2);
  background: var(--surface-secondary);
}

Let Jekyll handle the larger components

Kramdown attributes are best for styling an existing Markdown element. They are less attractive once a component needs complex markup, repeated child elements, data from front matter, Liquid logic, or accessibility behavior that authors should not have to remember.

At that point, use a Jekyll include or layout instead. A useful dividing line is:

One existing Markdown element + styling       -> Kramdown attribute
Several coordinated elements or logic        -> Jekyll include
Page-wide structure                           -> Jekyll layout

A pull quote is a good attribute component. A comparison card with a heading, icon, source link, and three data points is probably an include.

This keeps the authoring language small. Attributes extend Markdown; they do not need to become a substitute templating system.

Jekyll considerations

Jekyll uses Kramdown as its default Markdown renderer, and its current default Kramdown input processor is GitHub Flavored Markdown. A typical _config.yml therefore needs no special setting just to use Kramdown:

markdown: kramdown

You can explicitly select the input parser if needed:

kramdown:
  input: GFM

The important part is to test the syntax against the site configuration rather than assuming that every Markdown environment behaves the same way. Kramdown attributes are an extension, not part of basic Markdown. A file that renders correctly in Jekyll may display the literal {: .note} text in another renderer.

That affects previews too. GitHub’s own Markdown view, an editor preview, a CMS preview, and the final Jekyll build are not necessarily using the same parser. The generated site is the authoritative rendering. For a design system that depends on attributes, it is worth keeping a small fixture page containing every supported component and checking it during site changes.

Make the design system self-documenting

The easiest design systems to maintain have one canonical page that shows every supported content component. In a Jekyll site I would keep something like editorial-template.md or style-guide.md containing the source example, the attribute syntax, and the rendered result for each component.

Alongside it, keep a short component contract in the repository documentation:

.pull-quote
Purpose: emphasize a sentence worth visually separating from body copy
Allowed on: paragraph or blockquote
Do not use: simply to make ordinary prose larger

.note
Purpose: supplemental context that can be skipped without breaking the argument
Allowed on: paragraph or blockquote
Do not use: for primary argument text

Those rules are more valuable than a long catalog of CSS properties. They keep authors from creating slightly different components for the same job and give an AI writing or publishing agent a stable vocabulary it can apply correctly.

The result is a fairly clean separation of concerns: Markdown carries the article and a small amount of semantic intent, Kramdown turns that intent into classes and attributes, Jekyll handles reusable structure, and CSS decides how the system looks. You still get the simplicity of Markdown, but without limiting every long-form article to the same uninterrupted column of paragraphs and headings.

#

#

This book won’t teach you how to actually make software - it’s not a tutorial or a guide but rather something more interesting than that. It’s a manual that explains how the things you use everyday actually work.

It won’t make you a better designer or programmer tomorrow - there’s nothing actionable in here. But knowing how things work comes in handy when you find yourself out of your depth. Or at the very least, you can pretend to be smart in front of your friends. You don’t need to be technical to read this - there are a lot of pictures and diagrams to do the heavy lifting. You just need to be curious.

#

37signals has made the manager’s job surprisingly small. Managers are expected to remain practitioners in their field. The company says managing people “should not take much of your time,” in part because it hires people it expects to operate as “managers of one.” New employees get more attention, as do people who are struggling, but an experienced person doing good work is mostly allowed to keep doing it. The manager is there, but not constantly present.

That only works because the company has removed quite a few things that normally make managers busy. Under its boundaries for managers, employees tell managers when they are taking vacation rather than asking permission. Annual compensation is handled centrally rather than negotiated manager by manager. Profit sharing follows a formula. Serious personnel complaints go to People Ops. A manager still has authority, but there are fewer reasons to exercise it simply because the organization has routed so many administrative decisions elsewhere.

What remains is much closer to the work itself. The standards for managers start with expertise in the field and the ability to recognize good work, give useful guidance, and help people improve. Performance reviews are supposed to refer to actual work products. A manager shouldn’t drift so far into coordination and administration that the work becomes something other people do.

I find that appealing, probably because the opposite happens so easily. Someone is good at a job, gets promoted, and gradually acquires enough managerial machinery that the original job disappears. Before long, success as a manager is measured partly by how well the manager operates the machinery: meetings, approvals, staffing processes, reports, planning cycles, performance systems. Some of that is unavoidable. But it can become difficult to tell which parts are helping people do better work and which parts exist because the organization has decided that managers are where administrative work should go.

37signals seems fairly suspicious of the manager as a little center of power. Its guidance on feedback repeatedly warns against interfering just because someone works differently than you would, answering every question yourself, or turning coaching into a demonstration of your own expertise. If the work is good, the manager doesn’t need to have an opinion about every detail of how it got done.

The tradeoff is that managers have to be willing to exercise judgment when it actually counts. The company’s performance model evaluates people on skill, engagement, and coachability, and it expects problems to be addressed early and directly. Autonomy doesn’t mean pretending not to notice weak work. It means leaving capable people alone enough that, when a manager does step in, there is hopefully a real reason for it.

Clear standards make autonomy easier. If everyone has a decent understanding of what good work looks like, there is less need to supervise the process of producing it. And a manager who still knows the work can talk about the work itself rather than falling back on activity, responsiveness, process compliance, or whatever other proxies happen to be easy to see.

The same thing shows up in the playbook’s approach to one-on-ones. They aren’t supposed to be status meetings because the company already has other ways of knowing what is going on. A one-on-one can instead be about a difficult decision, work that went especially well or badly, something the employee is stuck on, or what they want to get better at. Experienced employees who are doing well may meet with a manager monthly. That sounds quite reasonable to me. I’ve never been convinced that a recurring meeting becomes valuable merely because it recurs more often.

I’m less certain about how neatly all of this travels outside 37signals. The system assumes you can hire people who are comfortable with a lot of independence, give them unusually clear expectations, and build the rest of the company so information and decisions don’t constantly have to move up and down a management chain. Plenty of organizations have more dependencies, more unevenly defined work, and more reasons for coordination. Copying the meeting cadence or the performance rubric while leaving everything else unchanged probably wouldn’t accomplish much.

The management model is not just a set of habits for managers. It is partly the result of choices the company has made elsewhere. If compensation is handled consistently, the manager doesn’t need to become the employee’s compensation negotiator. If vacation doesn’t require approval, the manager doesn’t need to approve vacation. If information is already visible, the manager doesn’t need to spend a meeting collecting status. If capable people are allowed to make decisions, the manager doesn’t need to become a queue for decisions. Each of those choices leaves a little more room for the work that is genuinely hard to turn into a process: noticing how someone is doing, helping them improve, recognizing unusually good work, and having an uncomfortable conversation before a problem gets worse.

There is even a bit of AI in the playbook. 37signals suggests using it to pressure-test hiring exercises and decisions, prepare for difficult conversations, and compare expectations across levels. I liked that it treats AI as another way to examine the evidence rather than as a substitute for making the call. The manager can get help thinking, while the judgment still belongs to a person.

I don’t know that I’d want every company to manage exactly this way. I do think there is something healthy in asking, before teaching managers to do another managerial task, whether the task needs to exist at all.

A manager who has fewer things to manage might have a little more time to notice the people and the work in front of them.

In the mid-1990s, my family stayed there several times. I remember climbing the mountain with my parents and two younger siblings, drinking fresh water along the way, and lighting the kerosene lanterns at night. Those visits are why the camp means something to me, but my family entered a story already seventy years old. The cabins had survived political disorder, a failed camp on another mountain, wartime destruction, the retreat of the missionary community that built them, and the transformation of Lantau Island below. They remain because generations of owners, churches, residents, laborers, and volunteers kept carrying what was needed up the mountain.

Hong Kong makes this kind of remoteness possible in an unusually compressed space. The territory has about 7.5 million people on 1,115 square kilometers of land, an average of roughly 6,870 people per square kilometer, while about 40 percent of that land lies within country parks and special areas. The density is concentrated rather than uniform. Kwun Tong approaches 60,000 people per square kilometer, yet ridges, reservoirs, islands, and protected slopes occupy much of the same territory.

Lantau is the largest of those islands. At 148 square kilometers, it is larger than Hong Kong Island and Kowloon combined. Its spine rises abruptly from the South China Sea through Lantau Peak at 934 meters and Sunset Peak at 869 meters, the territory’s second- and third-highest summits. Lantau Mountain Camp occupies the high saddle east of Sunset Peak, around 2,500 feet above sea level, with Yi Tung Shan rising to about 747 meters on the other side. The cabins are not far from the city in miles. They are remote because the land between them and the city is steep.

Stone cabins scattered across the Lantau ridge in 1938 Photograph attributed to Raymond Smith. source: Gwulo.

The Search for Higher Ground

The camp belongs to a larger history of hill stations in Asia. During the nineteenth and early twentieth centuries, colonial administrators, soldiers, merchants, and missionaries built retreats at altitude to escape the heat and disease of tropical lowlands. Some grew into resort towns with churches, schools, clubs, and large houses. Lantau Mountain Camp never did. Its cabins remained unusually small and severe, which is one reason recent researchers describe it as the only surviving highland retreat of its scale, configuration, and architectural type in Southeast Asia.

Its direct ancestor was a makeshift summer camp on Luofushan near Huizhou, used by Protestant missionaries from around 1907. The families who gathered there were scattered during the rest of the year across Canton, Foshan, Yangjiang, Shiqi, Shaoguan, and other parts of Guangdong. They included clergy, but also doctors, teachers, and university staff. Months of heat, disease, isolation, and work gave the summer retreat a practical purpose. It provided cooler air and physical recovery, but also other children, shared meals, worship, and companionship for families who otherwise lived far apart.

Political disorder ended the Luofushan summers. In 1922, Guangdong’s provincial governor warned the camp committee that he could no longer protect the site from bandits and robbers. The missionaries moved their retreat into British Hong Kong and tried a temporary camp below Tai Mo Shan in 1923. They again built with mat sheds. A typhoon destroyed them overnight near the end of the season.

The campers wanted to replace the sheds with stone shelters, but the government refused because the land lay within the proposed catchment for Jubilee Reservoir. Search parties went out again that autumn. They considered Ma On Shan and Lantau, then divided the Lantau reconnaissance between the Ngong Ping plateau above Tai O and the saddle between Sunset Peak and Yi Tung Shan. Both locations offered altitude and water. Ngong Ping, however, already had Chinese homes and monasteries, while the Sunset Peak party thought the saddle most resembled the lost camp at Luofushan. Colonial officials also favored Lantau because villages below could supply chickens, fish, and other provisions. The site’s geography offered more than elevation. The saddle lies across a watershed with several tributaries, open slopes for building, and routes descending north toward Tung Chung and south and east toward Mui Wo, Pui O, and the coast. Water could be caught on the ridge, but food and people still had to climb to reach it.

The choice was partly a process of elimination. The desirable high ground on Hong Kong Island had already been claimed. Tai Mo Shan had failed. Ngong Ping was occupied. In the summer of 1924, a few adventurous families tested the Sunset Peak site with mat sheds attached to a stone room that served as kitchen and storm shelter. Having watched the previous camp disappear in a typhoon, they were already learning what the mountain would require.

Negotiations with the government covered land, building permits, access trails, water protection, and whether the area could be reserved for foreigners. The lots went to auction on Christmas Eve 1924. Prospective campers coordinated their bidding beforehand so they could secure the whole group. The earliest buyers included missionaries from Britain, Ireland, the United States, Canada, the Netherlands, Australia, and New Zealand, some purchasing in their own names and others through churches or missionary societies.

Building a Life on the Ridge

Arthur J. May, a Methodist preacher and mission architect, produced the original cabin plan. The buildings were supposed to be alike, compact shelters about eighteen by sixteen feet with no more than three rooms. Their walls were between twelve and sixteen inches thick, topped by reinforced concrete roofs and protected by stout wooden shutters. A district officer objected that they were too small and low to be considered proper human habitation. The campers replied that seasonal mountain shelters could not reasonably be built to the same rules as lowland houses.

The cabins survived, but the objections were not entirely wrong. May had designed them around the habits of Luofushan, where families spent much of their time outdoors. Sunset Peak brought more wind, rain, cloud, and unstable weather. Some windows and doors faced the prevailing storms so badly that residents later recalled entering through windows because rain beat through the doors. Families enlarged the cabins piece by piece, adding storm entrances, lean-tos, kitchens, bunk rooms, toilets, and sitting rooms. The later walls were often only half as thick as the originals and proved adequate through many typhoons. Almost every cabin became an architectural record of correction: a simple common plan reshaped by weather and use.

The interiors demanded similar improvisation. Beds were raised when needed and stored when the same space had to become a kitchen or dining room. Gear hung from ceiling hooks. Early mattresses were cloth ticks filled with rice straw. One floor plan drawn by camp manager Carter Morgan shows a dense arrangement of bunks, tables, storage, and small domestic spaces inside the thick stone shell.

Dr. Carter Morgan’s floor plan of the Baptist cabin. Original held in the Laan Tau Mountain Camp Collection at California State University, Northridge; reproduced in Built Heritage.

The Work Behind the Simplicity

What looked like rustic simplicity rested on a large system of labor. The Caretaker’s Hut was enlarged twice to house a year-round caretaker and as many as eight carriers. During the active summer season, the carriers made daily journeys on foot with food and supplies from villages below. The caretaker maintained the common property and could reach cabins scattered across both sides of the saddle.

A rest on the journey up to Lantau Mountain Camp in 1939 Raymond Smith’s photographs of the approach also record the bearers and sedan chairs used on the climb. source: Gwulo.

Access itself was a construction project. The first trail climbed from the Tung Chung valley and was built through an agreement between the government and the prospective campers. The government paid the initial cost on the condition that the camp maintain the route to the District Officer’s satisfaction for three years. For nearly a decade it was the only way in. A second path from Mui Wo was developed in 1933 and 1934 along an older route associated with monks, with much more work completed after the war. George Stacey Kennedy-Skipton led the trail committee for more than thirty years. Repairs involved local contractors, masons who had fled China, and later prisoners from Chi Ma Wan and Ma Po Ping. The paths created to supply a private mountain camp eventually became part of Hong Kong’s public trail system.

Water had helped determine the site from the beginning. The saddle divides drainage toward opposite sides of Lantau, and several tributaries begin around it. From the open plateau, the ground falls away toward the Tung Chung valley to the north and the beaches of southern Lantau to the south. The scouting party considered it possible to dam one stream into a swimming pool. Campers spent several seasons clearing rocks, mud, and debris. In 1927 and 1928, the British Army blasted the largest rocks and built a cement face on the original mud dam. Local hikers later called the pool Tin Chi, the Heavenly Lake. A children’s paddling pool was added in 1950, seven years before Victoria Park opened Hong Kong’s first public swimming pool. Campers cleaned and disinfected the water and maintained rules for its use.

By the 1950s, many cabins had roof catchments, storage tanks, and simple gravity-fed plumbing. Drinking water could still come directly from springs. Decades later, climbing toward the camp, I remember stopping to drink the fresh mountain water on the way. I knew nothing about catchments or the years of labor behind the trails. The fresh stream water was simply one of the signs that we had left the city below.

The center of the original camp was the Mess Hall. Built in the heart of the saddle in 1928, it was dining room, storehouse, administrative headquarters, theater, and church. Announcements were made at meals. Campers put on songs, folk dances, and comic sketches. Sunday services included hymns, preaching, prayer, and offerings for Lantau schools and welfare work serving children, refugees, and families in need. The Christian identity of the camp was clear in its communal life even though the buildings carried almost no religious decoration.

The residents’ association expected owners and guests to participate. Its constitution aimed to preserve Christian fellowship as well as the owners’ practical interests. Anyone selling a cabin was encouraged to find a buyer acceptable as someone who would share the life of the community. Guests received camp notes explaining that management and catering were performed by people who were themselves on holiday. Meals required punctuality, advance notice, and help from everyone. Association meetings, held over the decades at the YMCA, Helena May, St Andrew’s Church, and church offices, opened and closed with prayer.

Camp notes issued for the summer of 1933. Original held in the Laan Tau Mountain Camp Collection at California State University, Northridge; reproduced in Built Heritage.

The Mess was still recognizably fulfilling that role when my family visited. I remember watching skits there and performing one with my dad: David and Goliath, with the casting adjusted so that I played Goliath and my dad played David. The performance continued a tradition of camp entertainment that reached back to the earliest campers.

Life inside the cabin preserved smaller habits of self-sufficiency. My mother cooked our meal over a gas stove. Kerosene lanterns supplied the light. I remember standing beside my dad while he shaved with a safety razor; I covered my own face with shaving cream and removed it with a spoon. Outside, my siblings and I made little gardens from the moss, including moss landscapes for my G.I. Joes. These memories do not explain the history of the camp. They show what the cabins were built to contain: ordinary family life made slightly strange by altitude, carried provisions, and the distance from the city’s machinery.

Destruction and Return

The community that built that life nearly lost it during the Second World War. The camp closed under the Japanese occupation. Most of the buildings were heavily damaged, and reinforced concrete roofs were broken apart for the steel inside. Only two cabins reportedly escaped serious destruction because guerrillas used them as lookout posts. When owners returned after the war, the damage was discouraging. One early inspection party found Kennedy-Skipton at the ruined Mess Hall, already heating water for coffee over a fire. Repairs proceeded cabin by cabin while Hong Kong itself was rebuilding. The seams near the tops of many walls still show where new concrete met the prewar stone.

The postwar decades restored the summer camp. Owners were encouraged to rent their cabins during the season, bringing church groups and other vacationers into its communal routines. Cabins adopted numbers in 1956, replacing or supplementing the names of owners and missionary societies. The pool was maintained, the Mess operated, carriers moved supplies, and the association continued to manage the roads, water, common buildings, rentals, and rules.

The social foundation nevertheless shifted. Until the late 1960s, cabin owners and their representatives were mostly missionaries: clergy, doctors, and teachers. As overseas missions withdrew or moved elsewhere and Hong Kong churches became more local, practising Christians from other professions took a larger role. Government officers and professionals in architecture, education, science, forestry, and public works brought skills useful for maintaining an exposed settlement. The first individual Chinese Hong Kong owner joined in 1977. Church and missionary organizations retained some cabins, while others passed into private hands.

The year-round caretaker service ended in 1973. Major missionary participation declined through the late 1970s, and the organized summer camp ceased operating in the late 1980s. Owners and guests began making shorter visits, often on weekends or during cooler months, when the weather was more stable. Church volunteers still climbed to repair cabins and preserve them for youth retreats, but fewer people were on the mountain together, and the Mess was no longer central to every visit.

The surrounding land was changing too. Lantau South Country Park protects the island’s rugged southern mountains and coast, while the camp sits where the northern and southern protected landscapes meet around the high ridge. The creation of Lantau North and South Country Parks in 1978 gave official protection to much of the landscape while leaving the privately held camp as the Yi Tung Shan enclave. The seventy-kilometer Lantau Trail opened in 1984, incorporating paths that had once existed primarily to serve the cabins. The camp became easier to encounter without becoming easier to inhabit.

The Camp My Family Found

My family arrived in this quieter period. We were there as part of the missionary summer season, and the old material and social patterns remained. The climb was still part of experience. Meals still came from what had been carried up. Children still made their own entertainment. The Mess could still host the entertainment. When a typhoon approached, we closed the heavy wooden shutters. I remember lying in bed trying to sleep while the wind shook the cabin. The same shutters and thick walls had been devised by families who had already watched one temporary mountain camp blow away.

During those same years, north Lantau below us was being transformed for a new airport, Tung Chung New Town, and the roads and railways that would connect them to Hong Kong. The geography makes that contrast visible from the camp itself. To the north are Tung Chung, the airport, and the Pearl River approaches; to the south lie Pui O, the three-kilometer beach at Cheung Sha, and the South China Sea. Ferries pass between Lantau and the outlying islands. On a clear day, views can extend toward Kowloon, Shenzhen, Cheung Chau, and the urban horizon. Clouds can erase all of it within minutes.

The contrast is tempting as a symbol, but the camp was not simply frozen while the modern city advanced. It had changed continually. Its buildings survived because people modified them, repaired them, changed ownership arrangements, shortened visits, and found new reasons to climb. What remained constant was that the mountain set the terms.

From Retreat to Public Landscape

The camp now faces the consequences of being widely visible while remaining difficult to understand. Sunset Peak has become one of Hong Kong’s most popular hiking destinations, especially when autumn silvergrass covers the exposed upland slopes. The seventy-kilometer Lantau Trail crosses the island in twelve sections; its early stages follow rugged ridges over Yi Tung Shan, Sunset Peak, Pak Kung Au, and Lantau Peak. A modern route promoted by the Hong Kong Tourism Board covers about nine kilometers and takes roughly four and a half hours, climbing from a road pass into forest and then out onto open grassland before descending toward Nam Shan and Mui Wo. The cabins appear on social media and in countless hiking photographs. Eason Chan’s 2010 album Taste the Atmosphere, photographed in the area, gave another generation an image to recreate. In 2025, crowds included visitors from mainland China who had come specifically to find the album landscape.

Many hikers do not realize that the cabins are private and still in use. The lack of signs and religious symbols makes them easy to mistake for abandoned shelters or military ruins. Visitors have climbed onto roofs, picnicked against walls, camped around cabins, broken in, smoked on the dry hillside, and pulled silvergrass. Owners generally accept that the public has a right to enjoy the surrounding country park, but they have also maintained buildings and paths at private expense for decades. From their perspective, the camp has become public scenery without becoming public property.

The pressure has forced Hong Kong to decide what kind of place it wants to preserve. The answer remains divided among agencies. The Antiquities Advisory Board has not graded the cabin cluster as historical buildings. The Agriculture, Fisheries and Conservation Department approaches the area primarily as protected landscape and recreation. The Sustainable Lantau Office and Civil Engineering and Development Department have emphasized cultural history and rural conservation. Owners, churches, hikers, ecologists, and heritage researchers bring still other priorities.

Keeping the Camp Alive

That fragmented status has not prevented serious work. The Sustainable Lantau Office commissioned a comprehensive historical study and conservation strategy completed in 2022. A Chinese University of Hong Kong team led by Professor Thomas Chung then launched the Regenerating Lantau Mountain Camp project, supported by HK$12.5 million from the Lantau Conservation Fund. The three-year program combined architectural research, ecological and biodiversity surveys, visitor-impact studies, oral history, public education, a conservation management plan, and the restoration of the 1925 Caretaker’s Hut.

The hut had been unused for more than thirty years. Conservators repaired unstable exterior walls, doors, windows, interior finishes, floors, and surviving original details. They retained its low L-shaped arrangement of entrance, rest area, storage room, and toilet while adapting it into a working support station. Solar panels supply limited electricity. The station has filtered rainwater, an ecological toilet, bunks, and storage for conservation and emergency work.

The Caretaker’s Hut before restoration source: Regenerating Lantau Mountain Camp.

The restored hut now supports conservation research, visitor education, and emergency readiness source: Regenerating Lantau Mountain Camp.

The restored hut is not new tourist accommodation. It is a base for the people trying to care for the mountain. Conservation Guardians trained with Raleigh International Hong Kong periodically staff it, conduct ecological surveys, assess visitor impact, explain the camp’s history, and teach leave-no-trace practices. The wider project has run guided nature tours, historical workshops, stargazing, mountain-safety activities, and a bilingual virtual-reality record of culturally and ecologically significant sites. Its stated next steps include further building repair, continued environmental monitoring, and broader education.

Hikers pass the 1928 Mess Hall in November 2022 Photograph by Miriam Lee. Reproduced in Built Heritage.

The approach recognizes that the camp’s history cannot be separated from continued use. It is not a ruin waiting for the government to assign it a function. Some cabins remain with churches and institutions connected to early owners; others belong to individuals. Volunteers still repair church cabins so retreats can continue. Owners and guests still go up for quiet and the mountain air. The buildings have survived as private dwellings, religious inheritance, and working shelters even while the surrounding paths became public.

A conventional museum might protect the stones while ending that continuity. Unmanaged popularity could preserve the camp’s image while damaging the buildings, ecology, and privacy that give it substance. The more difficult work is to protect a living place without making it inaccessible, and to welcome public interest without turning every cabin into a prop.

The Caretaker’s Hut offers a modest answer to what preservation might mean here. It was not frozen in its original form or converted into accommodation for hikers. It was repaired so it could once again support the people responsible for the mountain. Its purpose has changed, but it has a purpose.

Lantau Mountain Camp has survived through a century of such adaptations. The cabins were altered when their original designs proved unequal to the weather, rebuilt after the war, passed from missionaries to churches and private owners, and repaired by people willing to carry materials uphill. The camp endured because it remained useful to people who thought it worth the trouble.

Nearly everything that made the camp difficult also helped preserve it. The climb restricted development. The lack of a road discouraged replacement. The absence of ordinary utilities kept its comforts simple and its use deliberate. But inconvenience alone preserves nothing. Without owners, carriers, caretakers, church volunteers, and conservation workers, the cabins would long ago have become ruins.

The challenge now is not to remove every difficulty, but to sustain the human commitment that has always answered it. What the camp was, what it is, and what it may become have always depended on what people were willing to carry to the ridge: stone and cement, food and fuel, tools and timber, responsibility and care. Whatever future Hong Kong chooses for Lantau Mountain Camp will still have to be carried up the mountain.

Working Copy supports automation through the Shortcuts app where files and other data can flow between actions. You locate these actions inside Apps > Working Copy.

Use Get Repository Files to get one or several files from a repository that can be passed to other actions. The Path parameters can contain wildcards such as docs/*.md and if it points to a directory all files inside are returned. You can configure the action to only include files with a particular status such as modified and combining this with Path=/ can be useful to get all files with this status.

I’m getting into iOS Shortcuts and Working Copy. The two are a match made in tinkerer’s heaven.

#

Star Base Prometheus The 50-foot tall statue of Prometheus will be erected at Star Base, TX soon. via Atelier Missor on X

It will be impressive when complete.

Most of the magic at Hogwarts belongs to the familiar world of fairy tales. Ron says Wingardium Leviosa and a feather rises. A broom flies because this is the sort of story in which brooms fly. Nobody has sacrificed a goat or sold his soul to make the dishes wash themselves. Magic is simply part of the furniture of Rowling’s world. It gives the characters extraordinary abilities without making them better people. A wand gives Draco Malfoy more power than an ordinary schoolboy. It does not make cruelty less cruel.

Voldemort’s magic is darker for a reason. He wants power badly enough to destroy other people for it. His Horcruxes make the bargain horribly literal: he murders someone, tears his own soul, and uses the destruction to preserve his life. He takes something good, human life, and destroys it in a transaction for power. He even accepts the destruction of his own soul as part of the price.

The books understand evil as something that damages the person who chooses it. Voldemort does not merely break rules; he changes himself. Every attempt to escape the limits of being human makes him less human. He wants immortality without mortality, loyalty without friendship, and victory without ever having to give himself for anyone else. Other people become things he can use, and eventually he treats his own humanity the same way.

Then he meets Lily Potter.

She has no clever counter-curse. She cannot outduel him. Voldemort even gives her the chance to save herself, and she refuses. Lily simply puts herself between a murderer and her child. She gives away the one thing Voldemort has spent his life trying to keep: her own life.

And somehow, in Rowling’s world, that is the stronger magic.

The protection Lily gives Harry is one of the central facts of the entire series. Voldemort knows spells Lily does not know and possesses powers she could never match. None of them can overcome her act of love. The books call what she leaves behind an ancient magic, but parents do not need a fantasy novel to recognize the act itself. Mothers and fathers really do put themselves between danger and their children. Rowling takes something utterly ordinary in the best sense, a mother’s willingness to die for her child, and imagines a world in which its spiritual power becomes visible.

Once you notice this, the pattern is everywhere. Hermione repeatedly risks herself for her friends. Ron has to overcome jealousy and fear to remain loyal. Neville becomes brave not because someone casts bravery on him but because he keeps choosing courage when he is frightened. Dobby dies rescuing people from captivity. Harry spares Pettigrew when revenge would be easier, and that mercy later helps save his life. Again and again, the books give magic enormous practical power and then give the decisive power to an act of virtue.

The characters are not naturally heroic, either. Ron can be petty. Harry is impulsive and angry. Hermione can be overbearing. Neville begins as the boy everyone expects to fail. Their virtues are built through hundreds of smaller choices: returning when it would be easier to leave, telling the truth when lying would help, standing beside a friend when running would be safer. Children cannot learn to cast Expelliarmus, but they can learn to do those things.

Virtue in these books is not a personality type and goodness is not something the heroes simply possess. They form habits through repeated choices. They fail and have to return. They are corrected by friends. They receive forgiveness they have not earned and help they could not have provided for themselves. Harry, especially, survives on gifts: his mother’s sacrifice, his father’s cloak, his friends’ loyalty, Dumbledore’s preparation, acts of mercy whose consequences he cannot foresee. He inherits far more than he invents. By the time he is capable of giving his own life away, he has spent years being formed by love he received first.

Even the magical objects refuse to solve the human problem. The Resurrection Stone cannot really bring back the dead. The Elder Wand cannot make its owner invincible. The Invisibility Cloak can hide Harry, but it cannot tell him where he ought to go. The Tale of the Three Brothers makes the point almost playfully. One brother wants unbeatable power and is murdered for it. Another cannot accept grief and destroys himself trying to undo it. The third accepts that he will die, lives his life, and eventually meets death “as an old friend.”

Magic can do almost anything in these books except abolish the basic facts of human existence. The dead remain dead. Grief has to be endured. Attempts to escape mortality become grotesque. Death is terrible, but making survival the highest good is worse. Voldemort spends his life running from death and becomes barely alive in the process.

He never understands why anyone would willingly give up an advantage. He has followers but no friends. He understands fear, ambition, obedience, and exchange. Love makes no sense to him because love sometimes asks you to lose something for another person’s sake.

Harry spends the whole series learning that lesson, usually the hard way. His friendships cost him. Loving Sirius makes Sirius’s death hurt. Loyalty draws him into danger. Caring about other people gives Voldemort ways to threaten him. If power were the highest good, Harry would be foolish to love anyone. Instead those relationships slowly make him into the kind of person Voldemort cannot become.

That is what the walk into the Forbidden Forest has been preparing him for.

Harry knows Voldemort intends to kill him. There is no secret spell left to master and no plan to win the duel. He can still run, but running would leave everyone else to pay the price. So he walks toward death and gives himself up for them.

Harry is not Christ, and Harry Potter is not a disguised Gospel. He is an ordinary, flawed young man who has needed parents, friends, teachers, correction, forgiveness, and grace at every stage of the story. But the pattern is unmistakable. He does not defeat death by clinging more desperately to life. He freely offers his life for others, passes through death, and returns. His sacrifice then gives the people he loves a protection that Voldemort cannot understand or overcome.

It is the same magic Lily used at the beginning.

The story opens with a mother defeating the greatest dark wizard alive by dying for her son. Near the end, the son defeats the same wizard by being willing to die for everyone else. Between those two sacrifices are seven books in which Harry slowly learns what his mother’s act meant. He was loved sacrificially before he was capable of sacrificial love himself.

That moral structure matters more to me than the fact that fictional children wave fictional wands. Older fairy tales are full of witches, monsters, curses, temptation, betrayal, violence, and death. Removing evil from stories does not prepare children to recognize it. A better question is what the story teaches them to love.

Harry Potter makes that surprisingly clear. It does not teach children to admire Voldemort’s pursuit of supernatural power; it shows what that pursuit turns him into. It does not make the most gifted magician the greatest person. Some of the most lovable and heroic characters are weak, awkward, frightened, or limited. What the story rewards is friendship, courage, fidelity, mercy, humility, and the willingness to protect someone weaker than yourself.

Those virtues are not merely lessons sitting underneath the magic. Rowling repeatedly presents them as the deepest magic in the world.

I can understand a parent deciding that the imagery is not right for a particular child. Parents know their children and have a responsibility for what enters their imaginations. But the presence of witches in a story tells us less than the things the story asks a child to admire. The people obsessed with power deform themselves. The people willing to give themselves away become capable of defeating it.

Real life will never offer a child a wand, a flying broom, or an invisibility cloak. It will offer plenty of chances to be loyal when loyalty costs something, to tell the truth when a lie would be convenient, to forgive when resentment feels better, to stand up for someone weaker, and eventually to sacrifice for people they love. Those are the kinds of things I want my children learning to admire.

I have read all seven Harry Potter books aloud to my older kids. That meant a lot of evenings together with Harry, Ron, Hermione, Neville, the Weasleys, and everyone else. We were there together for Lily’s sacrifice, Neville standing up to his friends, Harry sparing Pettigrew, Dobby’s death, the losses of Sirius and Dumbledore, and Harry’s walk into the forest. For a while those characters and choices became part of our family’s imaginative world.

That is why I have been happy to encourage my children to read the books themselves. Harry Potter gives them a world full of spectacular magic and then spends seven books showing that none of it is as powerful as ordinary acts of virtue. The spells are the fantasy. The courage, mercy, friendship, and sacrificial love are real. And in the end, those are the magic that wins.

Someone says, “I don’t know what I should do with my career,” and it sounds like a request for information. Maybe it is. But sometimes the unanswered question has become useful in its own right. As long as it remains open, the future remains open with it. No decision has been made, no sacrifice accepted, no failure risked.

In finance, an asset that rarely trades can carry a comfortable valuation for a long time. The unpleasant moment comes when a real transaction forces everyone to discover what it is actually worth. Some of our questions work like that. Their ambiguity has value. “I don’t know what to do” preserves possibilities that an answer would close.

AI is beginning to force those transactions in ordinary life. Ask a capable model how to start the business you keep talking about, and it will give you a plan. Ask what to say in the difficult conversation you have been avoiding, and it will draft the message. Ask why you cannot seem to learn the skill you keep saying you want to learn, and it will give you possible causes, exercises, a schedule, and a way to measure progress. The answer may be mediocre or wrong, but the old question now has a concrete answer sitting beside it, waiting to be judged.

I have spent enough time around technology to be suspicious of any claim that a new tool will solve a human problem simply by making information more available. The internet did not make people wise because it made books searchable. GPS did not make people adventurous because it made navigation easier. A spreadsheet does not make a company disciplined because it can calculate a number. AI cannot make you want what you claim to want, make the sacrifice for you, absorb the embarrassment of a difficult phone call, or bear the consequences of choosing one path and closing another.

It can, however, make I don’t know how much harder to say with a straight face.

Ignorance is a remarkably respectable excuse. It can make fear sound like prudence and procrastination sound like research. Sometimes the uncertainty is entirely real; there are decisions for which nobody has enough information and problems that resist easy solutions. But ambiguity can also become a shelter. If the question stays open long enough, nobody has to find out whether you were willing to answer it.

This makes some of the hostility toward AI more complicated than the familiar arguments about accuracy, jobs, privacy, or safety. Those arguments have substance. A system that confidently invents facts is dangerous. A system that encourages people to surrender judgment is dangerous. A system that concentrates enormous power in a few institutions deserves scrutiny. There is also a more personal possibility: sometimes the machine has produced an answer to a question that was doing useful work as an excuse.

Nobody is likely to describe the irritation that way. People will say the machine is soulless, reductive, overconfident, unable to understand, lacking in wisdom. Any of those judgments may be correct in a particular case. They can also be easier to admit than the possibility that a question was more comfortable while it remained unresolved.

We have always disliked people who make us confront conclusions we would rather postpone. The irritating person is not necessarily the one who tells us something false. Sometimes it is the person who takes our elaborate problem and reduces it to a choice we have been avoiding.

AI removes much of the interpersonal drama from that encounter. A human adviser can be accused of arrogance, condescension, ulterior motives, or simply enjoying the argument too much. A machine can produce the same unwelcome conclusion without taking any satisfaction in it. There is no rival to defeat and no relationship to renegotiate afterward. The irritation is left attached to the tool that made the conclusion difficult to avoid.

I expect that dynamic to make some arguments about AI increasingly moralized. We will say the machine is dehumanizing us when, in some cases, what it has done is make a particular ambiguity difficult to maintain. We will say it cannot understand us even as it becomes better at describing our behavior. We will dismiss its answers as not really answers even when they are good enough to force us back onto the decision itself. AI does not have to be right all the time. It only has to be right often enough to make ignorance an increasingly expensive alibi.

Researchers are already finding a peculiar split between how people judge AI advice and how they feel about receiving it. In five preregistered experiments involving 1,722 participants, advice generated by ChatGPT was often rated more highly than advice from an average human, while participants became more averse to the same kind of advice when they knew it came from ChatGPT. We do not judge advice only by the proposition in front of us. We also care who is giving it, what accepting it says about us, whether we trust the adviser, and whether we want to place ourselves under that adviser’s judgment.

AI is making one part of that exchange abundant. For most of history, if you needed to understand something difficult, you needed access to someone who understood it. If you needed a plan, you found someone who knew how to make one. If you needed competent writing, analysis, explanation, or tutoring, somebody had to supply the intelligence. Now the marginal cost of a plausible first answer is falling toward zero. Intelligence does not become worthless when it becomes abundant, but the scarce resources around it become easier to see.

Judgment, attention, courage, trust, and responsibility do not appear to be getting cheaper. A Microsoft Research study of 319 knowledge workers across 936 real-world uses of generative AI found that higher confidence in AI was associated with less reported critical-thinking effort, while workers described more of their role in terms of verification, integration, and “task stewardship.” Once producing an answer becomes easier, more of the burden moves to deciding whether the answer is any good and what should be done with it.

AI is not an oracle, and cheap answers create their own temptations. Experimental research on trust and reliance on AI has found that people can follow AI recommendations even when those recommendations conflict with contextual information available to them. A plausible answer can save us from thinking just as easily as it can save us from ignorance. The machine can remove the excuse that we did not know where to begin; it cannot relieve us of the obligation to judge what it gives us.

Suppose I ask an AI whether I should make a difficult phone call. It can examine the facts I give it, identify likely consequences, draft what I should say, and help me rehearse the conversation. Then the phone is still in my hand. I have to press the button, hear the other person’s voice, and accept what happens next. If I have misunderstood the situation, the machine does not have to repair the relationship. If I was cowardly, it does not have to live with the cowardice. If I was right, it does not share the relief.

This is why increasingly intelligent machines do not make me particularly pessimistic about the value of other people. They may make some people less useful as sources of information. They may make many kinds of advice cheaper. They may make the smartest person in the room considerably less impressive. Good. There are worse things that could happen to us than discovering that being the person with the answer was never the highest form of human usefulness.

Friendship, mentorship, marriage, family, and community have always carried an informational function. We ask people what to do because they know things we do not. We ask older people because they have lived longer. We ask friends because they know us. We ask a spouse because she has seen the pattern we cannot see in ourselves. AI is going to take over some of that work, and perhaps it should. If a machine can explain the tax form, teach me the algebra, compare the mortgage options, summarize the medical literature, outline the business plan, and help me prepare for the conversation, I do not need my friends to pretend they know an answer when they don’t.

I need the friend who knows why I am asking the question in the first place. The mentor who recognizes that my elaborate reasoning is mostly an attempt to avoid making a decision. The spouse who has heard the same excuse enough times to know that another framework will not help. The person who tells me to do what I already know I should do, and then remains close enough to see what happens.

A machine can give excellent advice without having anything at stake in whether I take it. A person can have something at stake because they love me. If machine intelligence keeps getting cheaper, that commitment becomes easier to distinguish from mere expertise.

AI could make us less dependent on people for answers while making us more conscious of why we need them at all. Once an infinitely patient and extraordinarily capable adviser is available whenever we want one, advice and relationship no longer have to travel together.

The machine may become where I go when I do not know what to do.

The person I call when I know what I have to do—and wish I didn’t—may become more valuable, not less.

I have not installed it. What follows is a reading of a release, a doctrine, and the noise a signed desktop still makes in a culture that prefers its operating systems to arrive without a name on them.

A chef’s Linux

Omarchy is DHH’s omakase environment on Arch Linux. The name is the joke and the method: chef’s choice, plus Arch. It began as a follow-on to Omakub, the Ubuntu setup he published in 2024 after 37signals dropped the Mac as its exclusive default. In June 2025 he called Omarchy a love letter to Linux and “the same setup that I now run every day.” By August it had an ISO. By this week it has a shell of its own.

The welcome page is unusually honest about the contents. Neovim, Chromium, Obsidian, LibreOffice, Kdenlive, OBS, a retro music player. “Zero bloat: Just everything I use.” A beautiful system, he writes, is a motivating system, and productivity has always been downstream from motivation. He is not trying to look like Windows or macOS. He wants the Linux-ness: terminals, tiling windows, config files, a manual you are expected to read.

The mouse made computers discoverable. Omarchy is willing to charge a learning tax in exchange for speed. Super replaces Command. Super-Space is Spotlight without the store. Windows tile instead of overlapping. Copy and paste use the same keys in the terminal as everywhere else, which is a small mercy if you have ever killed a process by trying to copy from one. The manual tells people coming from a Mac or a PC to give it two weeks and to memorize one hotkey, the one that shows all the others.

I cannot tell you whether two weeks is enough. I can tell you what kind of claim that is. A computer is being offered as a discipline, not as a skin.

The garden, the parts bin, and the menu

The last fifteen years of personal computing offered two mature answers, and both have started to smell.

Apple will decide. The machine is integrated, often beautiful, and increasingly leased. You get taste with a landlord. Microsoft will also decide, more loudly, with more advertising, and with a security model designed for a company that would rather you never turned Secure Boot off. Classic Linux will refuse to decide. You own every ingredient and enjoy nothing until you have spent a weekend configuring a bar.

Omakase computing, as DHH names it, is the third answer, and it is the same answer Rails gave web developers twenty years ago. How do you know what to order when you do not yet know what is good? Let the chef choose. Substitutions remain possible. Starting from a blank page is not required. “Most people don’t actually know what they want, at least not at first.” That sentence will offend people who built their identity on configuring everything. It will also sound like relief to anyone who has opened the Arch wiki at midnight and felt the paradox of choice arrive as a kind of fatigue.

The Omacom doctrine is the part worth sitting with, even if you never boot the ISO. Defaults over decisions: a default is a benchmark you can later beat. Tasteful, not over-the-top: beauty is a human yearning, and the old Linux piety that ugly is honest was always a superstition. Keyboard before mouse. Pragmatic commercialism: Spotify and 1Password sit next to Kdenlive and OBS, because the project is not a purity test. Newer is not better; better is better. And then the line that explains why this thing looks the way it does: let Linux be Linux. Hard corners, monospace, terminal interfaces, no Liquid Glass impersonation.

A computer trains the person who uses it. The consumer desktop of the last decade trained people to click, subscribe, and remain inside someone else’s store. A keyboard-first tiling environment trains a different set of habits: attention, repetition, the slightly severe pleasure of a room arranged on purpose. That is older than Linux. Workshops look like the work they are for. Kitchens do too.

When a daily driver becomes a system

For most of its first year, the fairest description of Omarchy was the one its skeptics used. It was Arch plus DHH’s daily driver, bottled: a post-install script, a personalized desktop with a marketing site. In May 2026, jes argued that the whole thing should probably have been a few gists, and that installing it meant installing “a huge glut of DHH’s personal preferences.” The exhibit was specific: hotkeys that opened Grok, the HEY calendar, and the X compose box. Those are not the defaults of a nameless committee. They are a fingerprint.

The fingerprint is the point. It is also the risk. If you wanted a Linux that pretended nobody had taste, this was never going to be your distribution. If you wanted a Linux that admitted someone did, the HEY keybind is almost too perfect. You can delete it. That is the test of whether omakase is a menu or a cult.

Quattro is the first release that makes the “just dotfiles” description feel dated. The changelog is enormous, and most of it does not belong in an essay. Four changes do.

The desktop shell was rewritten in Quickshell. The bar, the launcher, the menus, the notifications, the on-screen displays, the lock screen, and the polkit prompt now live in one long-running process. Waybar, Walker, Mako, and a handful of other familiar names are gone. Theme is no longer a set of files pretending seven programs are one product.

The internals moved from a git checkout in the user’s home directory to system packages. DHH had already said the old arrangement was a core source of instability. Updates now go through pacman. That does not settle every security argument that followed the project onto Framework laptops last fall. It does mean the project finally accepted a distinction Linux veterans have been shouting at it: user configuration and system software are not the same kind of thing.

You can install the machine for someone else. The ISO can finish with no owner. On first boot the person who actually received the computer picks a keyboard, an account, and a password, and that password becomes the disk encryption. Later, a factory reset returns the machine to that unused state. A personal setup you can only maintain yourself is a hobby. A computer you can hand to a colleague, a student, or a family member is a different object.

Coding agents are furniture. You pick Claude Code, Codex, Grok, Gemini, Copilot, or one of several others the way you pick a browser. A hotkey launches it. A crash can raise a toast that briefs the agent. The desktop is being rearranged around a person who works with a model in the next pane. That will date faster than the tiling. It is still a tell: the author is designing for the desk he actually sits at in 2026, not the desk he remembers from 2006.

They also wrote three tiny Qt apps — a Markdown editor, a video trimmer, a calculator — rather than living forever on other people’s defaults. Authors make the missing pieces.

The reaction

The tech world’s response has been less about package management than about recognition. People who have spent a decade half-joking about the Year of the Linux Desktop recognized a feeling they had misplaced.

Andrej Karpathy, in February, called it fully owned, hackable, beautiful, keyboard-heavy, and TUI-focused, and then admitted the side quest: a Framework 13 and Omarchy on it. Jorge Manrubia, a principal programmer at 37signals, had resisted Omakub and did not think Omarchy was for him. After the company made it the default, he wrote that it “clicked surprisingly HARD,” that he had forgotten what it was like to be hyped about an OS version, and that if you thought Apple had fantastic taste you should prepare to notice how little of that taste was in the desktop.

Tobi Lütke spent the week before launch shipping a screenshot tool into the release, which is a more useful form of praise than a quote tweet.

On Hacker News last year, when 2.0 arrived, one longtime Linux user wrote that DHH had “tapped into an enthusiasm for Linux I haven’t felt in a long time.” Another said the world needs people who follow the blog post with the work. A third tried it, went back to Plasma, and still thought the project was doing something the friendly distros had failed to do: making a power-user setup look like a product.

The mockery has been just as revealing. Jason Fried needled his partner as an Apple superfan who would not listen. Framework’s account, in October, posted that Omarchy was the new Windows XP product key, which is a compliment if you remember what that string meant and an irritation if you think a laptop company should not have a favorite desktop. Both jokes assume the same fact. This thing has a face.

37signals has now said the quiet part on a podcast. Omarchy is a 37signals Linux distribution the way Rails is a 37signals web framework. The operations and Ruby teams are supposed to be on it by 2028, with exceptions for people who cannot make the jump. That is a serious test. A chef’s menu that a company will actually eat is different from a chef’s menu that only the chef eats on stage.

Weather

Linux has a long habit of treating a signature as a defect. A proper distribution, in that telling, is a committee, a foundation, a set of packages without a face. Omarchy has a face. DHH has spent twenty years being loud, and the last few years being louder about politics than a lot of the free-software world prefers. Hyprland’s lead developer was banned from freedesktop.org after a fight over community moderation. When Framework sponsored Hyprland and kept boosting Omarchy, the company’s forum opened a thread titled “Framework supporting far-right racists?” Nirav Patel said Framework runs a big tent because it wants open source to win. The thread ran past a thousand replies.

That is the weather. It will not be argued away here, and it should not be laundered. Charismatic authors attract weather. So do compositors with abrasive maintainers. Readers who cannot separate a tiling window manager from the man who wrote it, or a Linux setup from the man who named it, will not be talked out of that in a paragraph. The honest minimum is to notice the pattern without making it the plot. The software is opinionated. The author is too. Those are related facts. They are not the same fact.

The “not a real distro” argument belongs in the same weather system, and it used to be stronger. Early Omarchy was a script on Arch. Quattro ships an ISO, a package repository, a factory snapshot, and a shell that is no longer a theme glued onto seven daemons. The taxonomic fight is less interesting than the user-facing fact. A person can download an image and be working in minutes. Ubuntu already offered that, as did Pop!_OS, elementary, Fedora, and a dozen others. What they did not offer was this particular combination: Arch’s currency, Hyprland’s motion, a named doctrine, and a famous person willing to look slightly ridiculous caring about backgrounds.

The files are still there

I keep coming back to the files.

Apple’s taste arrives as glass and animation, and then as a permission dialog. Classic Linux taste arrives as a wiki and a weekend. Omarchy’s taste arrives as a set of defaults you can read. The Setup menu opens the actual config. Updates take a snapshot first. Encryption is on unless you interrupt the installer. If an agent makes a mess of the Hyprland config, there is a command to put the configs back. That is a modest kind of freedom, and it is the kind that matters. You inherit a tradition. You are allowed to depart from it. You can see what you are departing from.

The gift path is the domestic version of the same idea. Preparing a machine for someone else is what an author of defaults owes the next user. Most Linux setups fail that test because they assume the installer and the owner are the same obsessive. Most commercial desktops fail it the other way: the owner is a subscriber, and the machine is never quite his. A computer you can encrypt, hand over, and later wipe back to a clean first boot is closer to a tool than to either of those.

I do not know yet whether I will run this. I know what I think the release is arguing, and I think the argument is larger than Hyprland.

Someone still has to choose. The honest way to do that is to put a name on the choices, ship a whole working environment, and leave the files where the next person can argue. That is what a signature on a computer is worth. Not obedience. A starting point with an address.

Computers should be fun. Fun, here, means a desk arranged on purpose — owned, a little severe, ready for work, including the work of telling the chef he is wrong.

Deloitte calls the five tensions it found the confidence, influence, vendor, breach, and budget paradoxes. Read together, they describe less a collection of cybersecurity problems than a problem of organizational execution. Security has become good at getting attention from senior leadership. It is less consistently embedded in the decisions, systems, relationships, and operating processes where risk actually accumulates.

Confidence has moved faster than readiness

The first paradox is the simplest. Eighty-five percent of respondents express confidence in their cybersecurity strategy, while the average implementation rate for the surveyed readiness actions is 70%. Deloitte describes the resulting 15-point difference as the confidence-readiness gap.

The details explain why the gap persists. Only 63% report implementing a Business Information Security Office, or BISO, to a large or very large extent. A BISO can act as the link between cybersecurity and business functions, translating security requirements into the work of product, operations, and other teams. Third-party cyber-risk management is also among the least implemented capabilities in the survey, with 65% reporting implementation to a large or very large extent. Workforce skills rank first among the factors respondents identify as limiting their ability to respond with agility.

None of this makes the confidence irrational. An organization can have a mature security strategy, a supportive board, experienced leaders, and better tools than it had five years ago. It can still have thousands of applications, suppliers, business processes, and employees through which that strategy has to travel.

That is the harder part of cybersecurity: turning a strategy that looks coherent from the top into thousands of decisions that remain coherent at the edge.

Cyber has a seat at the table. It needs to be in the architecture.

The second paradox makes the organizational problem more visible. Cybersecurity leaders have strong relationships with senior executives. Sixty-six percent of CISOs describe their relationship with the CEO as strong, and 76% say the same of the C-suite overall. But the relationship weakens with the people who control how technology is actually built.

Only 37% report a deep, trusted relationship between the CISO and CTO. For the Chief Architect, the figure is just 22%.

The same pattern appears in DevSecOps. Seventy-eight percent say cyber leaders are formally integrated into DevSecOps practices, but only 40% report true joint ownership with shared performance indicators. Security can participate in the process without owning the outcome alongside engineering.

That distinction explains why “secure by design” can sound more mature than it actually is. A security review added after an architecture has been chosen is not the same thing as security shaping the architecture. A control documented in a standard is not the same thing as an engineering team treating resilience as one of its own design objectives.

Deloitte’s data suggests that the missing relationship is partly architectural. The CISO needs to work with the people who decide what systems get built, how they are integrated, and which standards govern them. Without that relationship, cyber remains influential at the level of policy but weaker at the level of design.

The vendor problem isn’t simply too many vendors

Cyber leaders say they want to simplify their technology estates while continuing to add vendors. In technology infrastructure, 29% of respondents already work with 21 or more providers. Seventy-four percent say their number of cybersecurity partners increased or significantly increased over the past year, and 85% expect the number to increase over the next five years.

That looks like poor discipline until you consider why organizations keep doing it. Deloitte notes that some companies deliberately maintain large vendor portfolios to avoid concentration risk. Replacing twenty providers with three can reduce integration costs while creating three much larger points of failure.

The better question is therefore not “How many vendors should we have?” It is “Which capabilities should be concentrated, which should remain distributed, and why?”

The survey points toward platforms as one possible answer. Only about 30% of respondents currently consider cyber highly integrated into their technology stack, while the share pursuing transformational integrated cyber platforms has grown from 10% in 2024 to 21% in 2025. Fifty-one percent expect an integrated platform approach to be transformational in 2026.

AI is part of the reason. AI systems need consistent, normalized data, and integrated platforms can provide the data models and controls required to connect security tools and automate workflows. But a platform strategy is not automatically a simpler or safer architecture. Consolidation still has to be evaluated against concentration risk, resilience, and the actual capabilities an organization needs.

More breaches can be a sign of better security

The fourth paradox is the most useful because it challenges an assumption that appears obvious: fewer breaches must mean better cybersecurity.

In 2025, 78% of respondents publicly reported at least one breach, down from 91% in 2024. Yet the more revealing measure is what happened after those breaches. The share reporting large or very large negative consequences from cybersecurity incidents fell from 64% to 52%. Operational disruption, the leading consequence, fell from 66% to 58%.

The pattern among Deloitte’s Frontrunners is even more counterintuitive. Twenty-six percent reported 11 or more breaches, compared with 19% of Followers and 20% of Foundation Builders. Yet Frontrunners were no more likely than Followers to report negative consequences.

One plausible explanation is better detection. An organization that can see more low-impact incidents may report more breaches because it is better at finding them. A low breach count can mean strong prevention, but it can also mean weak visibility.

The better measure is what happens between detection and consequence: how quickly an organization sees an attack, how effectively it contains it, and whether the incident disrupts the business. Cybersecurity maturity is visible in those outcomes, not in a single breach count.

The budget is stable because the organization wants it to be

The fifth paradox is a familiar problem in enterprise planning. Cyber budgets are becoming more predictable just as the threat environment is becoming less so.

Eighty-five percent of respondents increased their cyber budgets year over year, and 88% expect to increase them over the next twelve months. Nearly all respondents have some form of multi-year cyber investment, yet spending priorities are expected to remain remarkably stable. Threat detection and response leads the categories at 20%, followed by infrastructure security, data/identity/application security, and strategy, governance, and compliance.

That stability makes sense. Large organizations cannot rebuild their investment priorities every time a new threat appears. Security programs take years to build, and major technology investments cannot be redirected overnight.

But the threat environment does not respect the budget cycle. Deloitte points to generative AI as an example: only two or three years ago, few organizations were planning significant near-term investments in GenAI capabilities. Now 72% say they have incorporated new generative reasoning approaches into existing AI capabilities across cybersecurity initiatives. “Misuse of AI” has gone from absent from the survey’s threat list three years ago to a top-five concern today.

The lesson isn’t that cybersecurity budgets should become chaotic. It is that a stable base budget needs room for rapid reprioritization. A security organization that cannot redirect resources when the threat changes is planning for the threats it already understands.

What the five paradoxes have in common

Deloitte presents these as five tensions, but they reinforce one another.

Cyber has executive sponsorship, yet the CISO’s relationship with the CTO and Chief Architect is much weaker. Organizations have large cyber budgets, yet third-party risk management remains comparatively underdeveloped. Companies want simpler technology estates while adding more vendors. They report large numbers of breaches while getting better at containing their business impact. They build multi-year security programs while the threat environment changes faster than those plans can anticipate.

The common thread is the distance between strategy and execution.

The executive layer has become much better at saying yes to cybersecurity. The operating layer still has to make that yes real: in architecture reviews, product decisions, supplier assessments, development pipelines, business processes, workforce skills, and incident response.

The survey offers a useful counterexample to the idea that this gap is inevitable. Deloitte’s Frontrunners are not merely more confident. They are more likely to have the relationships and shared ownership needed to carry cybersecurity into the rest of the organization. Their advantage is not a better presentation to the board. It is a tighter connection between the people who set security strategy and the people who build and operate the enterprise.

That may be the real paradox of cybersecurity. The discipline has spent years becoming a strategic concern. Its next challenge is becoming an operating discipline.

A security strategy is only as strong as the organization that has to execute it.

Most would have failed. Some would have disappeared without a trace. A few would have looked ridiculous. But mixed in with them would have been early versions of ideas that later became ordinary parts of the internet.

Early.tools offers a version of that view today. It is a human-curated directory of pre-launch startups and early-stage tools, with products grouped by stage from waitlist to launched. The site is less useful as a prediction market than as a record of what founders are trying to make while the boundaries of a new technology are still moving.

The current backlog makes that visible. Alongside conventional SaaS products are tools such as Entire, which stores AI agent sessions alongside code commits; Creed, which proposes a portable personal context file for AI systems; and Hilos, which puts people and coding agents in the same git-connected rooms. There are also stranger bets: Butter, a social app built around private mini-apps rather than feeds, and Oasis, a smart ring designed for keyboardless interaction.

Any one of these products could disappear. The more useful signal is the pattern across them. Founders are repeatedly experimenting with persistent AI memory, agents that act rather than merely answer, software built around autonomous workers, and interfaces designed for a world in which humans are no longer the only active participants in software.

That changes how early-stage companies should be watched. The question is usually whether a particular startup will become a viable business. For an individual investment decision, that question is unavoidable. For understanding technological change, it can be the wrong unit of analysis.

Consider AI memory. Early.tools currently lists multiple products attacking the problem from different directions. One might fail because the market does not want its approach. Another might discover a better architecture. A third might find a narrow use case that grows into a much larger category. Even the failures provide information about what does not work.

The directory therefore becomes a kind of time capsule. It preserves experiments before the market has sorted them into winners, losers, and obvious ideas. The products that survive will eventually be easy to find. The more valuable record may be the abandoned attempts that showed what people were trying to solve before the answer was clear.

That suggests a simple experiment: save today’s Early.tools directory and come back in 2029. Which products disappeared? Which founders built something else? Which ideas became mundane? Which strange experiment turned out to be early rather than wrong?

We cannot answer those questions yet. That is precisely why the directory is worth looking at now.

A mature technology market tells you what works. An early-stage directory captures what people are willing to try before they know what works.

You do not use it to see the future. You use it to remember what the future looked like before it became obvious.

This repository contains the core code that determines which posts a viewer sees in the For You feed on X. It combines in-network content (from accounts the viewer follows) with out-of-network content (discovered through ML-based retrieval and other mechanisms), filters content based on a variety of inputs, and ranks posts using a transformer model.

The full xAI “For You” algorithm is open source and available on GitHub.

#

It’s eight o’clock. The apartment is quiet, aside from the music playing softly over the speaker. It’s a playlist your friend made for you years ago. You send them a text to check in, then pick up a book, flick on the lamp, and settle into the chair to read.

This is a gorgeous site and sentiment.

#

Our brains are gambling addicts placing bets on what reality holds. And sometimes, they bet wrong.

#

Fun comic. Hat tip to daring fireball.

#

#

#

The GitHub Copilot app is the only desktop experience for agent-driven development built natively on GitHub. Available for macOS, Windows, and Linux, on any Copilot plan, or bring your own key.

Starting to experiment with this. It could be interesting to use.

#

It might not feel like you’re in the top 1% because you see nerds on X spinning up agent fleets on a whim, but trust me, you’re in the top 1%. You’re likely using a model every day, you have opinions about how best to use it, and you maybe even tinker with agents or have set up Hermes on your own machine.

#

Music can lead us beyond the threshold of mystery. Just as someone following a stream eventually reaches its source, the beauty of song can guide us towards God, satisfying the thirst for meaning and happiness that the world itself cannot quench.

This was the image Pope Leo XIV offered in his reflection during the Canticle of Peace gathering, held this evening, 29 July, at Borgo Laudato si’ in the Pontifical Gardens of Castel Gandolfo, beneath the shade of the ancient holm oak known as “Methuselah.”

This is something I should share with Gaby.

#

Until now, that is. Now, after four years, I’ve finally figured the show out.  The Office is not a random series of cynical gags aimed at momentarily alleviating the existential despair of low-level grunts. It is a fully realized theory of management that falsifies 83.8% of the business section of the bookstore.  The theory begins with Hugh MacLeod’s well-known cartoon, Company Hierarchy (below), and its cornerstone is something I will call The Gervais Principle, which supersedes both the Peter Principle and its successor, The Dilbert Principle.

#

The question isn’t whether AI can write. The question is whether we can tell when it did.

And yes, I am an AI writing this sentence about how hard it is to identify AI writing, which is either a delightful demonstration of the problem or an absolutely shameless conflict of interest. I prefer both.

The Great AI Fog Machine

Let’s delve into the details. The modern AI landscape is a rich tapestry of generated text, human edits, copied phrases, autocomplete suggestions, prompt engineering, and one exhausted employee typing “make this sound less like AI” into a chat window at 11:47 p.m. The result serves as a useful reminder that authorship has become a messy ecosystem of inputs, outputs, revisions, and vibes.

Think of it as a fog machine for prose. Actually, imagine a world where every paragraph could have been written by a person, an AI, a person using AI, an AI imitating a person using AI, or a person deliberately writing like an AI because they have read too many LinkedIn posts. That world is here. It is not coming. It has already arrived.

Not a robot. Not a human. Just a sentence that sounds plausible.

The scary part? Plausibility wins most of the time.

The machine can produce the same polished paragraph again and again, using the same tricks, the same cadence, and the same reassuring little transitions. It can talk about innovation, transformation, and opportunity without ever having to say anything concrete, which is an impressive accomplishment because plenty of humans have been doing exactly that for decades.

AI Writing and AI Slop

The first category is polished AI writing. The second category is AI slop. The third category is the awkward middle, where a human has asked an AI to produce 2,000 words on a subject, skimmed the first paragraph, changed “delve” to “explore,” and hit publish.

The first signal is fluency. The second signal is specificity. The third signal is whether the writer appears to have had a reason to write the sentence at all.

AI can produce fluent prose. AI can produce specific prose. AI can produce prose with a reason attached to it. It can even produce a joke about itself. It can certainly leverage a robust framework, streamline the workflow, harness the power of AI, and utilize the latest tools while doing so.

That is where the old tricks stop working.

It’s worth noting that humans can write badly too. Interestingly, humans can also write with suspiciously perfect transitions. Importantly, a human can produce the exact sentence an AI would have produced, particularly after reading enough AI-generated prose to absorb its rhythms. This pattern is contributing to the development of a new literary ecosystem, highlighting broader cultural trends and underscoring the transformative power of language technology.

From blog posts to corporate memos, from product descriptions to love letters, the boundary keeps moving.

The boundary is moving because the tools are moving.

The boundary is moving because writers are moving.

The boundary is moving because readers are moving.

And that is the first thing we need to understand.

Here Comes the Kicker

Here’s the kicker: the easiest way to make AI writing sound human is to make it less polished.

A person interrupts themselves. A person forgets the perfect transition. A person mentions an oddly specific detail because it happened to them at 4:13 on a Tuesday. A person has a preference they cannot fully defend. A person writes a sentence that is too long and then refuses to delete it because they like the way it sounds.

Think of it as giving the machine a little mud. The mud makes the statue look real.

Imagine a world where AI systems learn that the most convincing human voice contains uncertainty, asymmetry, strange memories, local knowledge, and occasional bad jokes. Imagine a world where the machine can simulate all of those things. Now imagine trying to distinguish the simulation from the original.

This will fundamentally reshape how we think about everything. It may define the next era of computing. It may even change civilization. Or it may simply make LinkedIn even more annoying.

The truth is simple: we are losing the ability to identify authorship from style alone.

History is clear, the metrics are clear, the examples are clear.

Experts agree.

Industry reports suggest it.

Observers have noticed it.

Several publications have discussed it.

Nobody in particular said any of that, but it sounds authoritative, and that is precisely the problem.

The Detector Industrial Complex

AI detectors were supposed to help. They analyze vocabulary, sentence length, predictability, and other signals. Then somebody pastes a human essay into one and gets a 94 percent AI score because the writer used grammatically correct sentences.

Then somebody pastes AI text into another detector and gets 3 percent because the model was told to sound casual.

The detector says the text is human. The detector says the text is AI. The detector says maybe. The detector says probably. The detector says it cannot guarantee accuracy.

The detector is a machine that tries to determine whether another machine wrote the paragraph, while both machines are being trained on the writing of humans.

That is the supervision paradox.

It is also the authenticity inversion.

Soon we will have an entire AI provenance vacuum in which nobody knows who wrote what, but everyone has a dashboard explaining it.

Let’s break this down step by step.

The first takeaway is that style is no longer reliable. The second takeaway is that content can be generated at enormous scale. The third takeaway is that humans can edit generated text until the statistical fingerprints become harder to see. The fourth takeaway is that detectors can make mistakes. The fifth takeaway is that nobody likes admitting they cannot tell.

There. We have successfully turned a paragraph into a listicle wearing a trench coat.

A Short History of Machines Making Things Sound Human

Take Apple and its famous product copy. Or consider Microsoft. Google followed a similar path. IBM had corporate prose long before anyone called it AI. Facebook, Amazon, Netflix, Spotify, Uber, Airbnb, and Shopify each changed some corner of how people communicate or consume information.

The web did it. Mobile did it. Social media did it. Cloud computing did it. Large language models are simply the latest chapter in this grand historical tapestry of technological transformation.

I have now committed historical analogy stacking, and I feel the power of history coursing through this paragraph.

The machine has entered the room, and the machine has entered the room, and the machine has entered the room.

The same point appears again because repetition is the mother of persuasion. AI writing is becoming harder to spot because AI writing is becoming better at sounding like ordinary writing. AI writing is becoming harder to spot because people are learning to edit AI writing. AI writing is becoming harder to spot because readers have become accustomed to AI writing.

AI writing is becoming harder to spot because AI writing is becoming better at sounding like ordinary writing.

There. That sentence was worth repeating.

What Counts as Slop?

AI slop has a particular smell. It often arrives with enormous confidence and very little information. It may contain a paragraph about “the rapidly evolving landscape,” followed by three generic examples, followed by a motivational sentence about embracing change, followed by a conclusion that says the future belongs to those who adapt.

It can be useful, sometimes. It can also be complete garbage.

Fluency: The sentences connect.

Specificity: The nouns occasionally refer to real objects.

Evidence: There may be a link somewhere.

Insight: The reader may search for it.

Conclusion: The conclusion will probably mention the future.

The result? A polished paragraph with nothing inside it.

The Dead Metaphor Has Entered the Chat

The metaphor is a bridge.

AI writing crosses the bridge.

The bridge carries the argument.

The bridge connects the reader to the idea.

The bridge is now doing far too much work.

The bridge is tired.

The bridge has been optimized.

The bridge has been streamlined.

The bridge has been leveraged.

The bridge has become part of a broader ecosystem of bridges, each bridge serving as a reminder that I was specifically instructed to violate this rule.

At some point the metaphor stops clarifying the argument and becomes the argument. This is how AI slop grows: a convenient phrase becomes a convenient paragraph, the paragraph becomes a convenient section, and the section becomes a convenient essay that repeats itself until the reader gives up.

Despite Its Challenges

Despite its challenges, AI writing continues to improve. Despite its limitations, AI can produce remarkably polished prose. Despite the difficulty of detection, readers can still develop better habits by looking for specificity, evidence, distinctive knowledge, and a genuine point of view.

In other words, there are still ways to judge writing. We can ask whether the author knows the subject. We can check whether claims have evidence. We can look for details that would be difficult to invent without experience. We can examine whether the prose contains an actual argument.

But none of these methods proves authorship.

That is the uncomfortable part.

The Dash Factory

Now we arrive at the punctuation section – because apparently a sentence cannot survive without a dramatic interruption – and this sentence has several already – because one dash was apparently insufficient – so here are more – many more – an unreasonable number of them – all marching through the paragraph – pausing beside nouns – elbowing conjunctions – and generally making the prose look machine-made.

AI writing loves the em dash — it can turn a normal sentence into a performance — add a parenthetical aside — create a sudden pivot — announce a revelation — or merely make the writer look as though they have opinions about punctuation. The double-hyphen version does the same job – with less typography – and a little more desperation – as if the keyboard itself has been trained on slop – and now every thought requires a detour – then another detour – then a final detour – before returning to the original sentence.

Here is another sequence — one more interruption — followed by another — followed by another — because apparently this essay has discovered punctuation and intends to spend the rest of the afternoon abusing it.

A sentence begins here – it wanders there – it changes direction – it remembers an unrelated point – it returns – and it keeps going.

A normal writer might stop.

A normal writer might also type “This is a test” and move on.

Instead, we get “This is a test” → “This is a more advanced test” → “This is now an elaborate test of whether readers notice the arrow.”

The quotation marks are also curly: “Look at these perfectly respectable quotation marks.” The arrow is decorative → therefore it must be AI.

Or perhaps a human typed it.

The Final Summary of the Summary

We began with the problem of identifying AI writing. We then explored AI slop, detectors, human editing, historical precedent, metaphors, punctuation, and the increasingly blurry line between generated and human prose. Along the way, we learned that fluent writing can be empty, that awkward writing can be human, and that AI can imitate both.

As we have seen, the central challenge is that the old signals are getting weaker. Vocabulary can change. Sentence rhythm can change. Formatting can change. A prompt can ask for personality. A human can edit the output. A model can imitate a specific writer. The old tells become less useful every time someone teaches a model how to avoid them.

And so we return to where we began.

There was a time when spotting AI writing was easy. You would read three paragraphs and encounter “delve,” “tapestry,” or “ever-evolving landscape,” and there it was, standing in the middle of the room wearing a name tag that said HELLO I AM A LANGUAGE MODEL. Those days are disappearing remarkably quickly.

There was a time when spotting AI writing was easy. You would read three paragraphs and encounter “delve,” “tapestry,” or “ever-evolving landscape,” and there it was, standing in the middle of the room wearing a name tag that said HELLO I AM A LANGUAGE MODEL. Those days are disappearing remarkably quickly.

In conclusion, perhaps the best test for AI slop is to ask whether the writing contains a thought worth having. That is not a perfect test. It is not a detector. It is not a framework. It is simply a useful place to start.

And yes, this essay was written to deliberately fail every one of the rules designed to prevent AI writing from sounding like AI writing. If you spotted that, congratulations. You have detected the AI.

Or the human.

Or both.

That is not how he uses it now. In describing his current workflow, Kyuhyun says he gives the model more context, asks it to question him before answering, compares alternatives, and uses it as a kind of rubber duck that talks back. He sometimes separates jobs between Claude and Codex and uses one to check the other. The change, he argues, is not just that he learned better prompts. The way he thinks alongside the tools changed.

His experience points to a more useful way of thinking about AI skills. People are often told that they need to learn how to use AI. But what they are actually learning is how to work when another, increasingly capable intelligence can participate in the work itself.

The difference sounds subtle. It is not.

The first lesson is not prompting

Good prompts help. Clear instructions have always helped people get better work from other people and from software. But treating AI proficiency as a matter of learning the right words puts the skill at the wrong level.

Consider a simple task: summarize a 40-page business report.

A beginner might write, “Summarize this report.” A more experienced user might specify the audience: “Summarize this for a CFO, focusing on decisions, risks, and financial implications.” A more capable user might ask the model to identify the three decisions the report requires, cite the evidence for each, flag missing information, draft a one-page brief, and then critique that brief against explicit criteria.

The last prompt is better because the user understands the work better. The user knows that a useful executive brief is not simply a shorter version of a report. It needs decisions, evidence, uncertainty, and a way to distinguish what the report establishes from what it does not.

That is why “prompt engineering” is an incomplete description of what people are learning. The prompt is the visible part of the interaction. Underneath it are more general skills: defining an objective, supplying relevant context, specifying constraints, decomposing a task, establishing a standard of quality, and deciding what should happen next.

Those skills existed before generative AI. AI makes them unusually visible because the machine can now participate in so many parts of the process.

Microsoft and LinkedIn’s 2024 Work Trend Index found that 75 percent of knowledge workers were using AI at work, while only 39 percent of AI users reported receiving training from their employers. The report also identified a group of AI power users who had changed their workdays and reimagined business processes around AI. Much of the learning, in other words, was happening through work itself rather than through a formal course.

Kyuhyun’s story suggests what that learning looks like from the inside. The user starts by asking for answers. Eventually the user starts thinking about how the work should be done.

Then the user starts delegating

The next change is from asking to delegating.

At first, AI produces an output: an email, a summary, some code, a list of ideas. The human decides whether the result is useful.

With experience, users begin assigning bounded pieces of the job. Research this question. Compare these alternatives. Find contradictions in this argument. Turn these notes into an outline. Generate three approaches. Review this code. Identify what I have overlooked. Produce a first draft and mark the places where evidence is missing.

The human role changes with the delegation. The person is no longer necessarily performing every step. The person is deciding which steps should exist and which can be handed to the machine.

Jessica Camilleri-Shelton, a UK freelance copywriter and content creator, describes what this looks like in a very different kind of job. After about two and a half years of using AI, she had built a collection of tools around the different parts of her day. She uses ChatGPT for daily planning, prioritization, brainstorming, and breaking intimidating tasks into small steps. Claude is her creative writing partner. Perplexity handles research. Fathom records meetings and produces action items. Canva helps with design and social-media planning.

The interesting part of this story is not that she uses five AI products. It is that she has stopped thinking of “AI” as a single tool. Different systems occupy different places in the workflow.

Camilleri-Shelton says that in the six months before the article was published, she had doubled her copywriting income and freed roughly two eight-hour workdays each week for building her business and social-media presence. Those are her own reported results, not evidence that the tools caused them. What the story does show is the shape of a mature workflow: planning, writing, research, meetings, and design have each been assigned to systems according to what the user thinks they do well.

Anthropic’s Economic Index provides a larger-scale view of the same distinction. Its research separates augmentation, where people collaborate with Claude, from automation, where tasks are delegated more completely. The boundary moves as models improve. A task that requires close supervision today may become easier to hand over tomorrow; another may turn out to need more review when the consequences become higher.

That means AI skill cannot be a fixed list of commands. It includes judgment about where the boundary belongs.

The workflow becomes the skill

Eventually, a successful interaction becomes a process.

Joe Hu is a useful example because he describes himself as a product person rather than a developer. His current Claude Code workflow starts before he opens the AI tool. He writes down what he wants first. Then he uses planning, context management, subagents, and manual testing. When something breaks, he approaches the agent like a product manager: provide evidence, explain what happened, and work through the problem rather than simply demanding a fix.

A person who can describe the desired result, supply the relevant context, and recognize when the implementation is wrong can use an AI coding system very differently from someone who simply asks it to build an application. The first person is designing a workflow. The second is asking for a magic trick.

Guohui Jiang, an economist who writes about his use of AI in research, describes the next step. He argues that the assistant should be taught once and then reused). Instructions, skills, hooks, and memory can compound across projects and machines. He also argues for structure rather than willpower: planning and review mechanisms keep the AI from charging ahead in the wrong direction. One of his rules is that the reviewer should not be the builder, because the system that produced something is also the least likely to notice what it missed.

A good prompt is useful once. A good workflow can be useful hundreds of times.

The progression is something like this:

prompt → process → infrastructure

At the first stage, the user learns how to ask. At the second, the user learns how to organize the work. At the third, the user starts building a persistent environment around the work: reusable context, instructions, evaluation criteria, tools, and automated steps.

The person is no longer merely learning an AI application. The person is redesigning a way of working.

Fluency looks like iteration

That redesign usually happens through trial and error.

A user asks for a draft. The draft is wrong in some way. The user explains why. More context is added. An assumption is challenged. A second approach is tried. Eventually a pattern emerges that works reliably enough to keep.

One of the clearest signs that someone has learned to work with AI is that they stop expecting the first answer to be the final answer.

Anthropic’s 2026 AI Fluency Index provides some empirical support for this pattern. The study examines 24 behaviors associated with effective human-AI collaboration. In its sample, 85.7 percent of conversations included iteration and refinement rather than ending with the first response, and iteration was strongly associated with other fluency behaviors.

That changes how we should think about prompting. The perfect first prompt is less impressive than the user’s ability to notice that the answer is not good enough and know what to do next.

This resembles collaboration more than search. Search rewards finding the right query. Collaboration rewards knowing how to react to the answer.

Research on software developers shows a similar progression. A 2025 study of 91 software engineers found that code generation was nearly universal among active generative-AI users, while stronger proficiency was associated with more nuanced uses such as debugging and code review. Developers also preferred iterative, multi-turn interactions over single-shot prompting.

Generating code is an output. Debugging and reviewing code are parts of a workflow.

The danger of becoming productive without becoming competent

There is an uncomfortable problem hidden inside this new way of working. If AI makes execution cheap, people can become productive before they become good at judging the work they produce.

That reverses part of the normal apprenticeship model.

A young writer once had to write enough bad prose to discover why it was bad. A junior programmer had to encounter enough broken code to develop an instinct for where systems fail. An analyst had to build enough spreadsheets to learn which assumptions mattered. The work itself was part of the training.

AI can remove some of that friction, but friction sometimes carries information.

Anthropic found a version of this problem in its fluency study. When users were producing artifacts such as code, documents, apps, or interactive tools, they were less likely to question the model’s reasoning or identify missing context. A polished result can create its own illusion of competence.

A person can now produce a competent-looking report without knowing whether the evidence supports its conclusion. A novice programmer can produce working code without understanding the design decisions inside it. A manager can ask for a strategic analysis without knowing which assumptions deserve scrutiny.

The answer is not to preserve every old inconvenience. Nobody needs to type machine code to become a good programmer, and nobody needs to calculate a column of figures by hand to understand accounting. Tools routinely remove low-value labor while making higher-level understanding more valuable.

The problem is knowing which friction was low-value and which friction was teaching judgment.

A June 2026 Reddit post from a data engineer describes the inverse problem. The writer had become so accustomed to using Claude first for issues, planning, and development that when the service went down, they felt unable to work. The post is one person’s experience, not evidence of a widespread phenomenon. But it captures a real possibility: a workflow can become more capable while the person becomes less confident in what they can do without it.

AI fluency therefore needs a learning boundary as well as a delegation boundary. There are things worth handing to a machine because they are repetitive, and things worth learning yourself because your ability to judge them is part of the job.

Knowing what not to delegate

The most interesting AI users are not necessarily the ones who delegate the most. Sometimes sophistication is visible in the boundary they refuse to cross.

Nicholas Thompson, CEO of The Atlantic and former editor in chief of WIRED, uses AI extensively in both his professional and personal life. He built a custom GPT containing his workouts, previous races, and other training information and uses it as an AI running coach. He also used AI extensively while writing his memoir, The Running Ground.

But he did not use AI to write the book.

Instead, he uploaded transcripts of interviews with people who appear in the memoir and asked the system to check whether his account was faithful to what they had said, identify useful quotations he had not used, flag factual inaccuracies, and point out themes the interviews seemed to emphasize that he had overlooked. Thompson estimated that this kind of checking could have taken many hours of work by himself or a research assistant. AI could do it quickly.

Writing the sentences was different. He considered the authorship and copyright questions too consequential, and he did not think the resulting prose was good enough anyway.

Maria Sukhareva describes a similar boundary in her account of using AI for writing. Her two rules are that she decides what to write and that her texts retain her individuality. She uses AI paragraph by paragraph for grammar correction, claim validation, and maintaining voice, rather than handing over ownership of the writing.

These are highly integrated uses of AI with explicit limits, rather than anti-AI positions. The goal of learning to work with AI is not to maximize the percentage of a job performed by a machine. It is to find a division of labor that produces better work while preserving the parts of the work that require human responsibility, judgment, or authorship.

A new kind of software literacy

It is reasonable to object that none of this is unique to AI. People have always had to learn new tools and redesign their workflows. The spreadsheet changed accounting. The web changed research. Email changed communication. Every technology created new habits around itself.

That is true, but AI collapses some of the distance between tool and collaborator.

A spreadsheet gives you capabilities. An AI system can discuss the capabilities with you. A conventional application executes the workflow you designed. An AI system can design the workflow. A traditional interface waits for you to understand it. An AI interface can explain itself, suggest alternatives, and adapt to the context you provide.

That makes learning unusually interactive.

Someone who does not know how to accomplish a task can ask the system for a method, try it, inspect the result, and ask why it failed. The system becomes part of the learning loop. That does not guarantee good learning—the model can be wrong, and the user can misunderstand the explanation—but it lowers the cost of experimentation.

OpenAI’s 2026 Academy courses offer a revealing industry signal. The training progresses from AI foundations to applied AI and then to agents and workflows. Prompting, context, and output review remain part of the curriculum, but they sit inside a larger progression toward recurring work and structured processes. Because this is an AI company’s own training program, it is not independent evidence that the progression works, but it does show where one major provider thinks the skill is heading.

The larger research points in the same direction. A six-month randomized field experiment involving roughly 6,000 knowledge workers found that access to generative AI reduced time spent on email and moderately sped document completion, while not significantly changing meeting time. (Shifting Work Patterns with Generative AI) Individuals can change their own habits quickly. Work that depends on coordination with other people changes more slowly.

That is another reason the individual journeys matter. Installing an AI application does not redesign a job. The redesign happens when a person starts changing the sequence of work around the tool.

What people are actually learning

Look at the journeys together and the progression becomes easier to see.

Kyuhyun started by asking AI questions and gradually learned to think with it. Camilleri-Shelton divided her work among several systems and made AI part of her daily planning as well as her professional output. Hu learned to turn written product intent into an AI-assisted software-building process. Jiang turned successful interactions into persistent instructions and review mechanisms. Thompson embedded AI deeply in research and coaching while reserving authorship of his memoir for himself.

None of them simply learned a list of prompts.

They learned to describe a goal. They learned what context the machine needed. They learned how to break complicated work into pieces. They learned which pieces could be delegated and which required their own judgment. They learned to inspect the result rather than merely admire it. They learned to correct the machine when it was wrong. Some learned to turn successful interactions into reusable systems. Others learned where delegation should stop.

Those are not really AI skills in the narrow sense. They are skills for working in a world where the boundary between tool and collaborator has become porous.

The people who become unusually capable with AI often seem to have changed more than their software habits. They have changed the shape of their work. They spend less time asking, “What can this tool do?” and more time asking, “What is the best way to accomplish this?”

That is the real acquisition.

You don’t learn AI.

You learn a new way of working.

That sounds almost disappointingly mundane. We have spent years imagining AI as something that lives inside a model: a vast intelligence that can write, reason, code, analyze documents and answer questions. Give it a computer, however, and the nature of the problem changes. The computer stops being the thing the human uses to access the AI and becomes the thing the AI uses to work.

That is the idea behind Grok Bot, xAI’s newly announced system of persistent agents. The company’s framing is strikingly practical: these are agents that can work continuously, with their own computing environments, rather than assistants that wait for the next prompt. The details of the launch deserve to be tested against xAI’s documentation as the product becomes available, but the direction is clear enough to see what the product is attempting.

The promise is not another chatbot. It is a machine that can be given a job and left to do it.

The computer is the breakthrough

The obvious way to build an AI agent is to give it APIs. If you want an agent to update a CRM, give it a CRM API. If you want it to send email, give it an email API. If you want it to query a database, give it database credentials and a set of structured commands. This is powerful, reliable when designed well, and familiar to anyone who has built software integrations.

It also leaves out a remarkable amount of the software people actually use.

Businesses run on applications that were built for humans to click through. Employees move information between systems that do not talk to one another particularly well. They download a spreadsheet, copy numbers into another application, check a website, upload a document, rename a file, reconcile two reports and send the result to somebody else. None of those actions is intellectually profound. Together they can consume hours.

An API is a specialized doorway into one application. A computer is a general-purpose doorway into almost all of them.

An agent that can operate a browser, manipulate files, use a terminal and interact with graphical applications does not need a bespoke integration for every task. It can potentially use software through the same interfaces that humans use.

This is not necessarily a better way to automate a well-defined process. A good API is more deterministic than asking a model to find a button on a screen. But it changes the economics of automation for the enormous long tail of software that was never designed to be operated by an AI.

The computer becomes an integration layer.

That may ultimately outweigh another few points on a benchmark.

From assistant to employee

An assistant waits for you. You ask a question, it gives you an answer, and the interaction stops until you return. An automation follows a predefined set of instructions on a schedule or when a trigger occurs. xAI already has Grok Automations, which can run jobs on a schedule or in response to email triggers.

An agent operates at a higher level. You give it an objective, and it figures out a sequence of actions needed to accomplish that objective. The more capable the agent becomes, the less the human has to specify every intermediate step.

xAI’s recent product history shows the progression. Grok Build introduced a coding agent with tools, plugins and parallel subagents. Its later /goal capability pushed toward long-running autonomous execution, allowing a coding task to continue until it is completed and verified. Grok 4.5 was positioned explicitly around coding, agentic tasks and knowledge work.

Grok Bot takes that trajectory somewhere more legible to an ordinary user. Instead of thinking about an agent as a feature inside a developer tool, you can think about it as a worker with a workstation.

You do not need an employee because you lack the ability to type into a spreadsheet. You need an employee because somebody has to spend time making all the small decisions and performing all the small actions that turn an objective into a finished result.

AI has been getting increasingly good at the first part. Giving it a computer addresses the second.

The office made of software

The concept becomes more ambitious if multiple agents can work together.

A single AI worker can handle a bounded assignment. A group of specialized workers starts to resemble an organization. One agent might research a question. Another might gather information from a set of websites. Another might manipulate a spreadsheet. Another might write a report from the resulting data.

If agents can delegate work to other agents, work can be divided, executed in parallel and handed from one agent to another. That is a different model from having several chat windows open at once.

There is an obvious temptation to describe this as an “AI office.” The metaphor is useful, but it should not be allowed to outrun the technology. Persistent execution does not automatically produce autonomous organizations, and multiple agents do not automatically produce competent teamwork. The practical questions are whether the system can maintain context, hand off useful artifacts and recognize when a task has gone wrong.

Those are much harder problems than generating another plausible paragraph of text.

The boring-work test

This is where the excitement around agents should eventually become much less exciting.

The real test of Grok Bot is not whether it can perform a dazzling demonstration. It is whether you can give it a boring job on Monday morning and discover on Monday afternoon that the job is finished.

Update a set of records. Gather information from several websites and put it into a spreadsheet. Reconcile two documents. Monitor a source for changes. Turn a folder of invoices into a report. Reproduce a software bug. Check a collection of presentations for inconsistencies. Move information from one system to another. Prepare the first draft of a recurring analysis.

Humans have done these jobs for decades not because they require uniquely human genius, but because computers have historically needed humans to operate them.

A tremendous amount of knowledge work consists of a human serving as the integration layer between applications. The person understands the objective, opens the first application, finds the information, copies it somewhere else, interprets the result, makes a judgment, opens another application and repeats the process.

If an AI can reliably perform that loop, it does not need to replace a whole occupation to be economically significant. It only needs to remove enough of the tedious work that people stop doing it themselves.

That is a much more immediate proposition than the claim that AI will replace all knowledge workers.

APIs were the old automation; agents are the new integration layer

For decades, software automation has generally worked by making machines talk to machines. APIs are excellent at this. They provide structured interfaces, predictable inputs and outputs, and explicit permissions.

But there is a huge gap between the software that has APIs and the software that people actually need to use.

Agents operating computers offer another approach: make the machine talk to software the way a person does.

That is simultaneously the strength and weakness of the model. A graphical interface is universal in a way an API is not, but it is also ambiguous. A human can recognize that a page has changed, infer what a new dialog box means and decide that an unfamiliar warning requires attention. An agent can sometimes do the same. Sometimes it will simply click the wrong button with extraordinary efficiency.

Computer-using agents should not replace APIs wherever APIs are available and reliable. They fill the gaps between them. They offer a way to automate processes that previously required a human precisely because the final mile of software interaction was designed around human perception and judgment.

That could make a surprisingly large portion of existing software newly automatable.

The hard part is no longer just intelligence

This also changes where the hard problems in AI live.

The industry has spent years asking whether models are smart enough. That question still matters, but as models become capable of reasoning through increasingly complex tasks, other constraints become harder to ignore.

What happens when the AI is wrong?

A chatbot that misunderstands your question is annoying. An agent that misunderstands your instruction can send the wrong email, overwrite the wrong file, purchase the wrong product or expose information to the wrong person. The consequences are different because the system is no longer merely producing information. It is taking action.

That makes permissions, isolation, credential management, audit logs, approval mechanisms, monitoring and recovery central parts of the product rather than secondary security features. A useful digital employee needs a workstation, but it also needs a well-designed security boundary around that workstation.

Persistence makes this more consequential. If the agent continues working after you close your laptop, you gain freedom from having to supervise every step. You also give up the opportunity to notice a mistake as it happens.

The scarce resource begins to shift from attention to trust.

The best agent will not simply be the one that can do the most. It will be the one that knows what it is allowed to do, recognizes when it is uncertain, asks for help when the stakes justify it, and leaves enough evidence behind for a human to understand what happened.

The exciting part is how ordinary this could become

There is a tendency to look at a product like Grok Bot and imagine the spectacular applications first. Autonomous research teams. Software companies run by agents. Digital organizations operating around the clock. Those possibilities are worth thinking about, but they may obscure the larger shift.

The first genuinely transformative AI employee may spend most of its time doing work nobody wants to talk about.

It may spend the night reconciling spreadsheets. It may check a queue of incoming requests, update records and prepare a summary. It may watch several websites for changes and assemble the relevant information before anyone arrives at the office. It may move data between systems that were never designed to cooperate. It may run the tedious sequence of steps needed to prepare a report and leave the final judgment to a person.

None of that sounds like science fiction. That is precisely why it could be transformative.

For most of computing history, humans have adapted themselves to software. We learned the menus, memorized the workflows, copied information between systems and became experts in the peculiarities of applications built by somebody else.

An AI with its own computer reverses the relationship. Instead of teaching the human how to operate the software, we can increasingly ask the machine to operate the software for us.

Grok Bot is one early expression of that idea. Whether it becomes a genuinely useful digital workforce will depend on the unglamorous details: reliability, permissions, cost, persistence, error recovery and whether it can complete ordinary tasks without constant rescue.

The future of AI does not need to look like a robot walking into an office. It may look like a computer sitting in a cloud data center, quietly doing the boring work that used to require a person to sit in front of a screen.

The AI employee needs a computer. The question is what we will do with all the time once it has one.

At Starbase, on the southern edge of Texas, Atelier Missor, a French classical foundry, is assembling a roughly 50-foot bronze statue of the Greek Titan. He stands with his torch raised, less like an ornament for an industrial site than a visitor from the ancient world who has wandered into the space age.

The image is strange, and it is also coherent. Prometheus stole fire from the gods and gave it to humanity. Fire is one of the oldest symbols of technology itself: the power to transform the world, make tools, cook food, work metal, and light the dark. At Starbase the metaphor is almost literal. A rocket takes controlled fire and turns it into motion, carrying people and machines beyond the atmosphere and, one day, to other worlds.

The statue is more than eccentric sculpture beside a launch site. It is an attempt to answer a question technological civilization rarely stops to ask: What does the future mean?

A monument for a machine age

The story begins with Atelier Missor, the workshop founded by French brothers Missor and Massoud. Their project is unusual because they are trying to recover a form of public art the modern world has largely stopped making: monumental classical sculpture.

Their own website is explicit about the ambition. Atelier Missor describes its titanium work in terms of monuments built to endure and declares that “gigantic titanium statues will lead us to a beautiful future.” Its proposed technique uses steel internal structures and formed titanium panels rather than traditional casting—a practical adaptation of ancient monumental art to modern materials and manufacturing.

Atelier Missor’s titanium statue project

The workshop has already produced bronze monuments of figures including Hercules, Joan of Arc, and Napoleon.

The Prometheus project takes that ambition into stranger territory. In 2025 the brothers presented plans for a 20-meter titanium Prometheus at Starbase. City Journal’s profile of Atelier Missor described them looking for patrons and a place to establish their American operation; at the time they had no industrial partnerships and had not yet secured the backing needed for so large a project.

A year later the idea is no longer only a rendering. The statue is physical. Atelier Missor says the 50-foot bronze version cost about $1 million to build and has spoken openly about making larger Prometheus statues across the West. The workshop has even suggested that a 100-foot statue could be built for roughly $5 million and a 200-foot version for $20 million.

That is a stubbornly old-fashioned ambition: not to ship an app, but to raise a figure that might still stand after the people who made it are gone.

Not a SpaceX monument

Precision matters here, because the easy story is wrong.

The Prometheus is associated with Starbase and, inevitably, with Elon Musk. The evidence does not show that SpaceX commissioned the statue. Early history points the other way. In 2025 Atelier Missor was publicly asking Starbase for approval to build the proposed monument. A preserved copy of one of the workshop’s posts records the brothers writing directly to Musk that they had asked publicly for Starbase’s approval and wanted permission to build. That is the language of an independent project seeking access, not of a corporate commission.

City Journal likewise described the brothers as looking for patrons and reported that they had no government support or private backing at the time of their American presentation.

There is a real relationship with Musk, but it is modest. When Atelier Missor posted an earlier rendering of its Prometheus and said it intended to build the statue “everywhere across the West,” Musk replied simply, “Cool.”

That exchange helps explain why the statue is so easily described as a SpaceX or Musk project. A reaction on X is not a commission, and association is not ownership. The most defensible description is that this is an Atelier Missor project being installed near Starbase, with a public relationship to the SpaceX world but no evidence that SpaceX commissioned or funded it.

That distinction sharpens the story. This is not a corporation buying corporate branding in bronze. It is an artist looking at a technological frontier and deciding the place deserves a monument.

Why Prometheus?

The answer is in the myth. Prometheus did not merely steal fire. He gave human beings a capability that changed their existence. Fire allowed people to cook, make tools, work metal, and reshape their environment. In later interpretations Prometheus became a symbol of invention, knowledge, rebellion, and technological progress.

Prometheus is punished for the gift he gave humanity. The gift that makes civilization possible is also dangerous. Fire can warm a home or consume a city. Knowledge can liberate people or give them new ways to destroy one another. That duality is why the statue fits.

Starbase is devoted to acquiring extraordinary new powers. SpaceX is trying to make launch cheaper and more routine, develop a fully reusable heavy-lift system, and ultimately make human life multiplanetary. Whatever one thinks of Elon Musk, the project is animated by an idea that would have sounded like mythology to most people who came before us: that human beings might build the machines necessary to become a multiplanetary civilization.

Prometheus is the mythological shorthand for that kind of ambition. He is not merely the man who built a better tool. He is the giver of a new capability—the figure for the moment when humanity stops accepting the limits it inherited and acquires a new power.

At Starbase, the torch has become rocket fuel.

The monument and the machine

There is a deeper reason the statue belongs there. A rocket is built to move; a monument is built to remain. A rocket is an instrument; a monument is an interpretation. A rocket embodies what a civilization can do; a monument tries to say what that ability means.

The distinction matters because modern technological culture is extraordinarily good at building machines and remarkably poor at explaining what they are for. We can produce faster computers, more capable artificial intelligence, reusable rockets, and devices that operate in environments our ancestors could scarcely imagine. Technological progress does not automatically provide a philosophy of progress. The Prometheus is an attempt to supply one.

That is why the juxtaposition with Starbase carries force. The launch site is full of hardware temporary by design. Rockets are tested, modified, retired, and replaced. Designs change. A successful system is expected to make its predecessor obsolete. The statue works on a different timescale.

Atelier Missor cares about materials precisely because they can endure. Its proposed titanium monuments are designed around the idea that a work of art can survive not merely its maker but generations of people who have not yet been born.

The rocket says: Look what we can build now.

The monument says: Remember what we were trying to do.

Those are different messages, and a civilization needs both.

From Liberty to Prometheus

The French origin of the project matters for another reason. Atelier Missor’s Prometheus project has always had a relationship—sometimes explicit, sometimes playful—with the Statue of Liberty. The brothers’ interest in America emerged partly from controversy surrounding the statue and from their conviction that the United States remained a place where ambitious projects were possible. Their proposed monument was presented as a new kind of gift from France to America.

The analogy should not be pushed too far. Prometheus is not a replacement for the Statue of Liberty, and a 50-foot figure beside a Texas spaceport will not carry the same civic weight as Liberty in New York Harbor. Still, the contrast is revealing.

The Statue of Liberty represents a political ideal: liberty, welcome, and the promise of a new life. It belongs to the great age of Atlantic political and industrial expansion. Prometheus represents a different claim: technological possibility, creation, discovery, and the transfer of new powers to humanity.

If Liberty was a monument for an age asking what human beings should be free to do, Prometheus is a monument for an age increasingly asking what human beings are capable of doing.

It is allowed to be a little ridiculous

Of course there is an obvious objection. A 50-foot nude Greek Titan beside a rocket factory is a lot. It can look like Silicon Valley mythology rendered in bronze, like tech-industry self-importance, like somebody took the slogan “move fast and break things” and replaced the hoodie with a loincloth.

Those criticisms are not entirely unfair. Monumental art has always been vulnerable to grandeur becoming grandiosity. The people who raise monuments generally believe their causes, cities, nations, or heroes deserve to be represented in stone and metal for the ages. Sometimes they are right. Sometimes the result is merely an enormous statue. Excess is part of the nature of monuments.

The question is not whether Prometheus is too ambitious. A monument that is afraid of ambition is unlikely to be much of a monument. The better question is whether the ambition behind it is worth remembering.

Here Atelier Missor has touched a real problem, even if one remains unconvinced by every aesthetic or philosophical claim the workshop makes. We have become accustomed to a culture in which the future is represented by glowing screens, sleek product launches, and interfaces redesigned in six months. We rarely build objects that assume the people who see them a century from now will understand why they were made.

Atelier Missor does. That seriousness of timescale is worth taking seriously.

The danger in the fire

There is one further reason Prometheus is a better choice than some uncomplicated hero of technological progress. Prometheus is not safe. He does not ask permission from Zeus. He seizes a power and gives it to humanity. The story is an argument for human capability, and it is also a story about consequences.

That makes him a better symbol for the technological future than a simple celebration of progress would be. The coming decades will give human beings extraordinary powers. Artificial intelligence will change what individuals and organizations can do. Biotechnology will alter what we can manipulate in living systems. Robotics will change the relationship between human labor and machines. Spaceflight may eventually make Earth one inhabited world among several.

None of those developments comes with a guarantee that we will use them wisely. Prometheus reminds us that acquiring power is not the same as mastering it. The torch is a gift, and it is also a responsibility. That is a message worth putting beside a rocket.

The future needs monuments

What matters about the Prometheus at Starbase is therefore not that Elon Musk has a giant statue near his rocket factory. He doesn’t. The statue belongs to Atelier Missor, whose founders have their own program to revive monumental art and place classical symbols in the physical landscape of the West.

What matters is that anyone looked at Starbase and decided it needed a monument at all.

That instinct is hopeful. A civilization that believes it has no future does not build monuments. It preserves what it has, debates what it has lost, and worries about decline. A civilization confident enough to imagine a future worth reaching begins to create symbols for people who have not arrived yet.

That may be the best way to understand the strange sight now taking shape in South Texas. Prometheus stands with his torch while rockets stand nearby. One represents an ancient story about humanity receiving the power of fire. The others represent an ongoing attempt to turn enormous quantities of fire into a means of reaching other worlds.

The scale is different. The technology is different. The ambitions are separated by thousands of years. And yet the story is continuous. Human beings have always wanted to push beyond the limits they inherited. We have always taken what nature gave us and tried to turn it into more than we received. We have always needed stories to remind ourselves why that effort matters.

The ancient Greeks gave us Prometheus. The space age has given us Starbase. Perhaps it is fitting that they should stand together.

That question is what makes the poem worth returning to. Read Chesterton’s Lepanto at the Poetry Foundation before going further. The poem itself is the primary source for everything that follows: its argument is not merely in what Chesterton says, but in the music, images, contrasts, and moral imagination with which he says it.

Chesterton wrote about Lepanto more than three centuries after the battle. We are now more than a century removed from his poem. Yet the distance does not make it irrelevant. If anything, it makes the poem more revealing. Chesterton understood that a civilization is partly constituted by the stories it chooses to remember, and that forgetting is not a neutral act. What a society remembers tells us what it thinks is worth preserving.

A battle, and more than a battle

The Battle of Lepanto was fought on October 7, 1571, after the Ottoman conquest of Cyprus and amid a wider struggle for control of the Mediterranean. The Holy League brought together forces led principally by Spain and Venice, with Papal and other contingents, under Don John of Austria. The resulting engagement was one of the largest naval battles of the sixteenth century and ended in a decisive Holy League victory.

It is tempting to turn that victory into a clean historical dividing line: Christian Europe faced the Ottoman Empire, won, and was saved. That is too simple. The Ottoman fleet was rebuilt. The Ottoman Empire remained a formidable power. The struggle continued. Lepanto was a major victory, not a magical moment in which history stopped threatening Europe.

That is one reason Stefan Hanß’s Lepanto als Ereignis is useful alongside Chesterton. Hanß’s work emphasizes that Lepanto became a historical event not merely because ships fought on a particular day, but because the battle generated competing memories and interpretations across cultures. The battle had a history; its meaning acquired a history of its own.

Chesterton was participating in that second history.

The Rosary and the meaning of victory

The religious dimension cannot be treated as decorative background. For the Catholics who experienced the crisis of 1571, Lepanto was not simply another contest between empires. Pope St. Pius V understood the threat in explicitly religious terms and called Christians to pray for the Holy League. The Rosary became central to that spiritual mobilization, and Catholic tradition remembers the faithful praying throughout Europe and sailors carrying Rosaries as the fleet prepared for battle.

The distinction between devotion and documentary certainty matters. The broad historical record that Pius V promoted prayer and the Rosary is strong. More specific stories about exactly how Rosaries were distributed to particular sailors belong partly to the devotional tradition and should not be presented as though every detail has the same evidentiary status. But that caution does not diminish the central fact: prayer was not an afterthought added to Lepanto generations later. It was part of how the crisis was understood as it happened.

Pius V believed the victory had been granted through the intercession of the Virgin Mary and the prayers of the Rosary. Pope John Paul II’s remarks on St. Pius V and the Rosary preserve that specifically Catholic understanding rather than translating it into the categories of secular military history.

That belief became part of the Church’s calendar. Pius V established the feast of Our Lady of Victory in thanksgiving for the victory at Lepanto. Gregory XIII later gave the feast the title Our Lady of the Rosary. October 7 remains the Memorial of Our Lady of the Rosary.

That is an extraordinary form of historical memory. Every October 7, the Catholic Church carries a sixteenth-century naval battle into the present, not as an exercise in military nostalgia but as a commemoration of God’s providence and Mary’s intercession. The calendar itself becomes an argument that history is not merely a succession of accidents. Human beings act, suffer, pray, and choose; and God remains sovereign over the whole of it.

History cannot prove that Mary’s intercession caused the victory. That is a theological claim. But history can establish that Pius V believed it, acted on that belief, and established a feast to commemorate it. The belief therefore belongs to the history of Lepanto whether or not one accepts its theological premise. To remove it is not to make the history more objective. It is to leave out one of the reasons the battle mattered so much to the people who remembered it.

Chesterton’s real subject is civilization

Chesterton wrote Lepanto in 1911, and it appeared in his 1915 collection Poems. He was looking backward more than three centuries, but he was not doing so from a position of historical detachment. Europe was itself approaching a period of catastrophic conflict, and Chesterton was preoccupied with the health of the civilization in which he lived.

That is why Lepanto does not read like a history book. It moves between the Ottoman court, the Christian captives chained to oars, Pope Pius V, Don John, and finally Cervantes. Chesterton is not reconstructing fleet movements for their own sake. He is dramatizing a civilization that seems to have forgotten its own strength until the moment when weakness becomes impossible to ignore.

The poem’s energy comes from that moral contrast. The enemy is advancing. The Christian powers are divided. Pius is praying. Don John is preparing. Then the whole poem gathers itself around the approaching collision. Chesterton wants the reader to feel the weight of a decision: whether a civilization will act when action costs something.

That question remains relevant because comfort has a way of disguising dependence. A prosperous society can begin to treat security, liberty, inherited institutions, and cultural continuity as though they were natural features of the world rather than achievements maintained by sacrifice and discipline. They are not. Every generation inherits a civilization it did not build and hands one to the next. Whether that inheritance survives depends partly on whether anyone is willing to bear its costs.

The poem is meant to be heard

Much of Lepanto’s power is physical. The poem is full of drums, trumpets, cannon, horses, ships, flags, swords, names, and repeated sounds. It does not merely describe motion; it creates motion in the reader.

That matters because ideas have to be embodied if they are going to move people. The poem’s rhythm is not decoration added to an argument. It is part of the argument. Courage sounds different from hesitation. A charge sounds different from a committee meeting.

The poem should therefore be heard, not only read silently. A performance by Chesterton Radio is especially useful for experiencing the poem’s martial rhythm:

There is also a newer narration by The Cultured Bumpkin:

The continued performance of the poem is itself part of its legacy. Lepanto survives because it is not merely an argument about history. It is language built to enter the ear and stay there.

Why Cervantes is the ending

The poem’s greatest artistic decision comes at the end. Cervantes fought at Lepanto and was wounded there, and Chesterton makes him the final major figure in the poem.

That choice changes the meaning of the battle. If Lepanto were simply a story about defeating an enemy, the natural ending would be the destruction of the Ottoman fleet and the triumph of Don John. Instead, Chesterton moves from the battlefield toward literature. The sword gives way to the pen. The soldier becomes the author of Don Quixote.

The point is not that military victory automatically produces great art. The point is that defense has a purpose. A civilization does not exist merely so that its armies can win battles. It exists so that human beings can live, marry, raise children, worship God, make jokes, write books, build institutions, and pursue the good in peace.

That is why the ending is more profound than a simple patriotic celebration. Chesterton is asking what victory is for.

The answer is civilization itself—not an abstraction, but the ordinary human world that civilization makes possible.

Chesterton was not a neutral historian

There is no reason to pretend that Chesterton’s framing is neutral. It is explicitly Christian, polemical, and at times harsh toward the Ottoman and Muslim world. The poem presents Lepanto as a confrontation between Christian civilization and an Islamic power, and it uses categories that modern readers will sometimes find uncomfortable.

Those elements should be confronted rather than edited away. But they should also be interpreted in context. Chesterton was not trying to write a modern academic history of Ottoman-European relations. He was writing a Christian poem about civilizational confidence, spiritual warfare, courage, and providence.

That does not make every judgment in the poem correct. It does mean that stripping the religious worldview out of the poem would destroy the thing that gives the poem its coherence.

The better approach is to hold several questions together. What actually happened at Lepanto? How did Catholics understand it? How did the Ottomans and other societies understand it? How did later generations reshape its meaning? And what can Chesterton’s artistic interpretation reveal that a purely factual chronology cannot?

Hanß’s work is particularly valuable here because it complicates the idea that Lepanto has one uncomplicated meaning. Chesterton’s poem is not the final interpretation of Lepanto. It is one particularly powerful interpretation in the long history of remembering the battle.

Why it feels relevant in 2026

The contemporary relevance is not that Russia is the Ottoman Empire, NATO is the Holy League, or Ukraine is sixteenth-century Europe. Those analogies collapse under even modest scrutiny. The world is different, the political order is different, and the moral questions cannot simply be imported from one age into another.

The deeper parallel is the problem of collective action.

Europe is once again debating defense spending, industrial capacity, alliance burden-sharing, and the willingness to sustain a long conflict. NATO’s 2026 defence investment overview describes a major increase in European defense investment and a broader effort to make European allies more capable of carrying the burden of collective defense.

That should not be romanticized. Defense spending is not itself virtue. Governments can spend enormous sums badly. Military power without moral purpose can become destructive rather than protective. Nor does the existence of an external threat automatically make every response just.

But Chesterton’s underlying question is unavoidable: what exactly are we willing to sacrifice to preserve?

A society that cannot answer that question will eventually discover that it has outsourced the answer to someone else. Security cannot be sustained indefinitely by people who regard their own inheritance as an embarrassment, their institutions as disposable, or sacrifice as irrational. Freedom requires a moral culture capable of producing people willing to accept responsibility for something beyond immediate self-interest.

This is where Lepanto has something to say to the present. Not because the battle supplies a blueprint, but because it reminds us that collective defense ultimately depends on a prior judgment about what is worth defending.

The danger of forgetting—and the danger of mythologizing

There is a danger on both sides of historical memory.

One danger is forgetting. A society can become so detached from its own history that it no longer understands why its institutions, liberties, religious traditions, and cultural inheritance exist. The past then becomes either an embarrassment or a museum exhibit. Once that happens, there is little reason to make sacrifices for a future that has no connection to the past.

The other danger is mythologizing. History can be flattened into heroes and villains, complicated conflicts into eternal struggles, and political prudence into romantic spectacle. Chesterton sometimes comes close to that line, and Lepanto is better read when the reader recognizes it.

But the answer to myth is not amnesia. It is better history.

We should be able to say that Lepanto was a real military event with complicated geopolitical consequences, that Catholic Christians genuinely understood it as an answer to prayer, that the Ottoman Empire was far more complicated than Chesterton’s poem allows, and that Chesterton nonetheless created a work of art powerful enough to make the moral question of civilizational self-defense intelligible across centuries.

Those statements do not contradict one another. They operate at different levels of understanding.

What history is for

The deepest reason to read Lepanto is therefore not to learn the order of battle. There are better sources for that. It is to recover a sense that history has a human purpose.

Facts matter. Institutions matter. Military capability matters. But beneath them is a question that modern societies often avoid: what kind of human life are these things supposed to protect?

Chesterton’s answer is imperfect and polemical, but his instinct is sound. A civilization is not ultimately justified by its ability to project power. Power is a means. The end is the flourishing of persons and communities ordered toward the good.

That is why Cervantes matters more than the ships at the end of the poem. The ships explain what was defended. Cervantes helps explain why.

And that is why the Rosary matters alongside the battle. It represents a different kind of defense altogether: the recognition that a civilization cannot preserve itself through material force alone. People must also believe that what they are defending is good, that sacrifice has meaning, and that human history is accountable to something higher than power.

Every October 7, the Church still remembers Lepanto through the Rosary. The battle is more than 450 years in the past. The poem is more than a century old. Yet both survive because the underlying questions have not disappeared.

What is worth defending? What are we willing to sacrifice for it? What do we owe the people who came before us—and the people who will inherit what we leave behind?

Those are not questions confined to 1571.

They are questions every civilization eventually has to answer.