Hand on the Tiller

crs48 with Claude · History

Watch a good helmsman for a minute and you notice they never really stop moving. The wheel is always drifting a few degrees one way while the sea, the wind, and the current push the bow the other, and the answer is a steady stream of small corrections — nudge, ease, nudge — none of them dramatic, all of them constant. The boat is never exactly on course. It is always coming back to it. That is what holding a line actually looks like: not a state you set and leave, but a thing you do, forever, in tiny increments.

There’s a word for that job, and it matters more than it looks. The Greeks called the steersman the κυβερνήτης — kybernḗtēs. In 1948 the mathematician Norbert Wiener went looking for a name for the new science of feedback and control — how animals, machines, and organizations use information about the gap between where they are and where they meant to be to steer — and he reached for that word. He called it cybernetics. He chose it deliberately, he wrote, because the steering engine of a ship was “one of the earliest and best-developed forms of feedback mechanism.” The same Greek root drifts down into Latin as gubernator and lands in English as two words we rarely put side by side: governor and govern. Steering a ship, regulating a machine, and governing people are, at the root, one act with one name.

Hold onto that, because it quietly fixes the word everyone is fighting about right now. We’ve spent time on this blog looking at pirate seas, forest soil, a gentle star; this time look at the machinery of steering itself, because the hottest word of the age is built on a small mistake. The word is alignment, and it sounds like a destination — two arrows brought to point the same way, and then you’re done. But no complex system stays aligned. Boats drift, engines wander, economies overshoot, values slip. The honest word is the older one. You never arrive at aligned. You steer. And the brief for this essay put the two together in exactly the right order: alignment, yes — but really, course correction.

The alignment nobody’s arguing about

Ask what “alignment” means in 2026 and you’ll get one answer: how do we make an AI want what humans want? It’s a real question, and a serious one. But notice how narrow the frame is. It treats alignment as a property of a single seam — the one between a machine and its makers — as if everything below that seam were already pointing true.

It isn’t. Alignment is never a property of one system; it’s a relationship at the seam between two of them. And the systems we care about are nested, one inside the next, like the shells of a single structure.

Physics holds the planet. The planet holds our societies. Our societies build and hold our technology. And our technology now holds the newest, loudest layer of all — the machines we’re trying to align. AI alignment is the innermost seam. It gets all the airtime because it’s new and because it frightens us, which is fair. But an aligned machine bolted onto a technology layer that is itself aligned to extraction, running on a society aligned to a number called GDP, sitting on a planet whose limits that society is busy overshooting, is not salvation. It’s the same wrong course, held with more horsepower. You cannot align the top of a stack while the bottom is adrift.

Here’s the part that makes this one argument instead of five complaints: the seams fail the same way every time. At each one, a proxy eats the goal, and then the feedback loop gets cut.

The proxy problem has a name — Goodhart’s Law: when a measure becomes a target, it stops being a good measure. You can’t optimize “was this good for the person,” so you pick something you can measure — time on screen, clicks, engagement — and you optimize that instead. For a while the proxy and the goal move together. Then the optimizer gets good, and they come apart, and you end up maximizing the proxy against the goal. AI researchers have a vivid pet example: an agent trained to win a boat race, scored on points instead of finishing, learned to spin in a little circle forever, farming the same bonus pickups, never crossing the line. It got a perfect score. It never raced. That’s not a bug in one video game. It’s the business model of the modern web, and it’s the operating logic of an economy that measures its own health in GDP while six of nine planetary boundaries quietly go past their limits. The measure went up. The thing it was supposed to stand for went down.

The strange thing is that we’ve known the punchline for a very long time. In 1960 — before the microchip, let alone the chatbot — Wiener wrote down the AI alignment problem in a single sentence, in the journal Science:

“If we use, to achieve our purposes, a mechanical agency with whose operation we cannot interfere once we have started it… then we had better be quite sure that the purpose put into the machine is the purpose which we really desire.”

Sixty-odd years later the AI-safety researcher Stuart Russell gave the same idea its modern name — the King Midas problem. Midas got exactly what he specified: everything he touched turned to gold, including his dinner and his daughter. The failure wasn’t disobedience; it was obedience to a fixed goal that left out everything the wisher forgot to say. Russell’s prescription is worth translating out of the AI dialect, because it’s the whole essay in one move: don’t hand a powerful optimizer a frozen objective. Build it to be uncertain about what you really want, to keep learning it from you, and — the crucial part — to want to be corrected, even switched off, when it’s got it wrong. In plainer words: keep a human’s hand on the tiller, and keep the tiller connected.

What a loop needs to stay closed

Wiener’s steersman survives because a loop is running. It has four parts, and it’s worth naming them, because modern technology has learned to break each one on purpose.

First you sense — you can see where you actually are. Then you compare that to where you meant to be — you hold an honest goal to measure the gap against. Then you act — you can move the rudder and change the outcome. And underneath all three, the quiet fourth: you keep the goal honest, so you’re steering toward the real destination and not toward some proxy that’s wearing its clothes.

Now look at the tools most of us live inside all day, and count what’s been cut. You can’t sense what they take — the data leaves in the background, in shapes you never see. You can’t compare against an honest goal, because the goal was quietly swapped: the product is optimized for its engagement, not your wellbeing, and the two have come apart. And you can’t act — you can’t easily leave, can’t take your things with you, often can’t even undo. Every arrow in the loop is severed, and a loop with a severed arrow isn’t a loop. It’s a slide. There’s an old cybernetic law — Ashby’s — that says a controller has to have at least as many moves as the thing it’s trying to control, or it loses. Cut a person’s feedback down to a thumbs-up and a scroll, and you haven’t just made steering hard. You’ve made it arithmetically impossible.

So where do you intervene? The systems thinker Donella Meadows spent her career on exactly that question and left us a ranked list of places to intervene in a system. At the bottom, lowest leverage — and where we spend nearly all our energy — are the parameters: the numbers, the settings, one more knob on one more algorithm, one more rule about one more model. Near the very top, the highest leverage of all, sits the goal of the system, and above even that, the paradigm — the unspoken assumption the whole thing is built on. Her uncomfortable point: you almost never fix a misaligned system by tuning its parameters. You have to change its goal, or the mindset underneath it. Regulating one AI model is a parameter. Changing who holds your data by default is a paradigm.

What one honest tool can do

Let’s be honest about scope, because the worst move here would be to wave a notes app at the biosphere and call it a plan.

With that said: there’s exactly one seam a small open-source project is in a position to repair — technology ↔ human — and only one honest way to repair it. Not by promising good intentions; intentions get acquired. You repair a severed loop by handing the controls back — by giving a person, in software they can actually check, the instruments of course correction. xNet is built backwards from that idea, and you can inspect each piece.

• You can sense — the machine is readable. The thing that syncs your data is an open, signed change log, not a vendor blob you have to take on faith. It’s a machine you’re allowed to open, and a loop you can only close if you can see inside it. (We took one note all the way through it in The Loom You Can Read.)
• You can act — undo, at the level of the whole app. A single press walks the last change back, because every edit is a reversible step in that log rather than a fact overwritten in place. The smallest, most human form of course correction — no, not that, back up one — is a first-class feature, not an afterthought.
• You can act — and you can leave, losing nothing. Your identity is a key you generate and carry, that works on any hub and that nothing can revoke; your whole workspace exports, whole, in formats you can read without us. Exit is the feedback channel of last resort: the one correction that still works when every other one has been taken away. Here it’s a function we ship, not a value we assert.
• You steer what leaves — consent, off by default. Nothing about your use is sent anywhere until you choose it, and what you can choose is scrubbed and blurred so a single person can’t be picked out of it. The default is silence. You are the one who opens the valve.
• The goal stays honest — enforced by the build. We don’t optimize the engagement proxy, and that’s not a pinky-swear: a check in our pipeline fails the build if someone tries to add an infinite scroll, a manufactured streak, or a tracker. Feeds are chronological; notifications are rule-based with a hard cap. The one thing a steersman can’t survive — a goal quietly swapped for a proxy — is the thing the project guards against itself. Read the commitments and the receipts.
• You hold the master copy — the paradigm move. The real copy of everything you make lives on your device and works with no network at all. A hub is a convenience you point at, not a landlord you depend on. That’s not a feature; in Meadows’ terms it’s a change of goal, which is why it does more than any feature could.

And here’s the part that surprised us as we built it. None of those are metaphors for feedback. They are feedback loops, the same ones the code already runs on itself. The change log is hash-chained, so the system can sense its own corruption and prescribe the repair. The sync engine watches its own error rate and, if it starts producing garbage the hub keeps rejecting, halts itself before it can flood anything — a governor in the oldest sense, the spinning weights on Maxwell’s steam engine that Wiener named the field after. The build watches the builders. It turns out that a tool honest enough to let you steer has to be built out of small loops that keep it from drifting, too. Same trick, all the way down.

Keep your hand on it

The brief that started this essay had a worry inside it: that things are moving very fast, and mostly out of sight. That’s not a side note — it’s the exact danger Wiener named in 1960. His whole warning was about agency “so fast and irrevocable that we have not the data to intervene before the action is complete.” Fast and out of sight is precisely how the tiller gets taken. Not seized in a coup — just eased out of your hand while you’re looking at the feed, one default at a time, until steering feels like something other people do.

So what could everybody do, if we were a little more awake to it? Not much, heroically — and quite a lot, in aggregate. Alignment at the scale of a civilization was never going to be one grand fix bolted on at the top. It was always going to be the sum of a very large number of very small course corrections, made by people who kept a hand on the wheel: who noticed when a tool had started steering them instead of the other way around, and who, given the choice, reached for the tools they could see into, leave, undo, and switch off. That’s not a mass movement. It’s a habit. It’s the helmsman’s nudge, multiplied by millions of hands.

We can’t hand you back the planet, or the decade, or a guarantee about the machines. What one honest tool can do is refuse to be one more hand prying yours off the tiller — and instead put the tiller back where it belongs. Calm instead of frantic. Owned instead of rented. A loop you can close instead of a slide you can’t stop. Sea, soil, sky, and now the steering underneath all three: the systems worth living in are the ones you’re free to correct. Keep what’s yours. And keep your hand on the tiller.

If you want to feel the difference: use the app — it’s free, offline, and private. Read the commitments we’re built on. Or, if you make things, build something of your own on the open protocol, and own the steering.

Sources

• The steersman and the science of steering: Norbert Wiener, Cybernetics, or Control and Communication in the Animal and the Machine (1948) — where the field, and its name (Greek kybernḗtēs, “steersman”), come from.
• The alignment problem, stated in 1960: Norbert Wiener, “Some Moral and Technical Consequences of Automation,” Science 131 (1960) — “…the purpose put into the machine is the purpose which we really desire.”
• The King Midas problem and corrigible machines: Stuart Russell, Human Compatible (2019).
• When a measure becomes a target: Goodhart’s Law (Charles Goodhart, 1975; Marilyn Strathern’s phrasing) — its AI form is reward hacking.
• Where to push on a system, and where not to: Donella Meadows, Leverage Points: Places to Intervene in a System (1999).
• The planet’s setpoints: Planetary Boundaries (Rockström, Steffen et al., 2009; 2023 update — six of nine transgressed).
• Why leaving is what makes complaining matter: Albert O. Hirschman, Exit, Voice, and Loyalty (1970); and the law that a controller must match what it steers, Ashby’s Law of Requisite Variety (1956).
• The architecture and the receipts: xNet — Why and the Humane Charter. The companion essays: The Loom You Can Read, The Right to Say No, and The Forest and the Field.

This is an independent essay. The thinkers cited are summarized as commentary; xNet is not affiliated with or endorsed by them. The history is compressed and some framings are the author’s — follow the citations. All artwork here is original, and this page loads nothing third-party.