Kilometre eight. Zone 2. The effort drops low enough that the body goes quiet and the Default Mode Network opens — the circuit that connects dots you can't force at your desk.

This morning Karen Hao finished an argument for me.

Her book is Empire of AI. 250 interviews with OpenAI insiders. And her argument is not the one most people are having. Most people debate whether AI is good or bad. Dangerous or overhyped. That debate is a decoy.

Hao's point: AI is not the subject of the story. A small group of individuals is — and they chose AI as their vector. Same pattern every time a civilisation has concentrated power. Extract resources. Consolidate capital. Control the narrative. Bury the invisible labour. Swap the tool — printing press, railway, oil, search engine, LLM — and the micro-dynamic is identical.

That's why she calls it an empire. The word is precise.

We are in the middle of the biggest mental-model transplant in a generation, and most leadership teams are treating it as a procurement decision. Which licence. Which vendor. Which training. As if the question were how do we use this thing, when the real question is whose worldview are we installing when we use it.

A few kilometres later the DMN pulled in another voice. Nikhyl Singhal on Lenny's. Half of PMs in trouble. 30,000 out, 8,000 back in, all AI-first. The information mover is dead. The builder is reborn.

Singhal says find your moments of joy. Good advice for individuals. Not enough for organisations. A builder inside an empire is still inside an empire. A PM who ships faster inside an extractive worldview helps the extraction go faster.

Donella Meadows put mental models near the top of the leverage hierarchy. Not parameters. Not goals. Mental models. That is where organisations bend or break.

So the uncomfortable question. Inside your company right now, whose mental model is this AI running? Who decided? When? Did anyone else get a vote?

Stop arguing about the tool. Start fighting over the mental model.

Last week we looked at Limits to Growth — the constraint that tightens silently around every reinforcing loop. This week: what organisations actually do when they feel it tightening.

They reach for the fastest thing that makes the number move.

The trap is that the symptomatic fix works. Not permanently — but long enough to kill the urgency for the fundamental one. Every time the quick fix buys a quarter, the deeper investment loses a sponsor. The capability you needed to build atrophies precisely because the workaround performed.

In SaaS, the constraint was product stagnation masked by acquisition. The fix: more spend — discounts, switching promos, aggressive bundling. Revenue recovers. But every euro spent pulling forward demand is a euro not invested in product depth. The fix became the operating model.

In manufacturing, the constraint was quality erosion under volume pressure. The fix: outsource maintenance, bring in contractors, add another shift. But each workaround drains institutional knowledge. What started as a resource gap became a capability gap.

In healthcare, the constraint was diagnostic capacity falling behind volume. The fix: triage harder, shorten consultations, defer complex cases. Throughput improves. But diagnostic accuracy drops. The institution's diagnostic capability erodes — while its volume metrics look better than ever.

Senge called this the addiction structure — and the analogy is precise. Relief reduces the motivation to address the root cause. The root cause deteriorates. The need for relief grows.

What makes this archetype hard to see from inside is that every decision, taken individually, looks rational. Nobody chose to erode capability. They chose to hit a target. The erosion is a side effect distributed across time, invisible in any single review.

The constraint you found in Limits to Growth is the same variable here. The question is whether you treat it — or build past it.

Kilometre four. Amol Avasare, Anthropic's Head of Growth, tells Lenny that activation — not acquisition, not retention — is the single highest-leverage problem in AI growth.

I nearly stopped running. Not because it was new. Because it was confirmation.

I've spent months building a causal loop diagram of how SaaS revenue systems work. Thirty variables. Fifty-five links. Fourteen loops. Then someone at the fastest-growing AI company in history says out loud the thing the model predicts.

In the system I mapped, there's a loop called R9 — Frictionless Activation. Product quality deepens activation, which lifts conversion, which funds investment, which improves the product. It connects what you build to what you earn.

But Anthropic adds a twist. They're deliberately adding friction to onboarding — friction that helps users understand why the product is for them. Every time, it improved conversion. The variable isn't "friction." It's activation depth. Shallow activation converts fast and churns faster. Deep activation sticks.

$1B to $19B in fourteen months. They've stopped using linear charts internally — too exponential to be useful. That's what happens when reinforcing loops fire before balancing loops activate.

Anthropic indexes 70/30 toward big bets. Most growth teams do the opposite. When model capabilities grow exponentially, micro-optimizations capture a shrinking share of a growing pie. You can't incrementally optimise your way out of a structural problem.

And then the line that confirmed the map. Anthropic is "very comfortable forgoing metric impact to prioritize safety, brand, quality, and user experience." That's not a values statement. That's a constraint function. In my model, constraints prevent vicious loops from activating. The system is the strategy.

Pflaeging argues that every organisation has three structures. The formal hierarchy — which produces compliance. The informal network — which produces influence. And the value creation structure — which produces reputation.

His point is sharp: we over-invest in the formal structure and ignore the other two. Value flows centre-out through networks of teams, not top-down through reporting lines.

But what about the shadows of the other two?

Jo Freeman wrote "The Tyranny of Structurelessness" in 1970. When you remove formal hierarchy, you don't remove power. You make it invisible. Whoever controls information, whoever belongs to the in-group, whoever knows the unwritten rules — those people govern without mandate and without accountability.

Then there's Foucault's panopticon. It's a principle, not a building. You don't need a boss watching you if everyone is watching everyone. In informal networks, the disciplinary gaze doesn't come from above. It comes from all sides.

And James Barker's concertive control. In 1993, he studied a company that moved from hierarchical control to self-managing teams. The teams developed normative rules that controlled behaviour more powerfully than the old system ever did. Peers confronted each other, punished deviance, rewarded conformity. The iron cage didn't open. It tightened.

Each structure carries its own shadow. The formal hierarchy carries rigidity and bureaucracy. The informal network carries invisible power and social exclusion. The value creation structure carries peer surveillance and production at any cost.

The question isn't which structure to privilege. It's how to make the dynamics of power visible across all three.

Six kilometres into this morning's run, John Cutler said something that stopped me mid-stride: "It's more likely that your company will change you than you will change your company."

That's a hard sentence if you've spent any part of your career trying to make organizations work better.

Companies are gold rush towns, not lasting institutions. We design for permanence when we should be designing for passage. Most of the structures we fight to change were never built to last — they were built to capture a moment.

Leaders should be game designers, not chess players. The job isn't to move pieces. It's to shape the rules and conditions so that better moves become possible for everyone. This connects to Ashby's Law — a system's ability to respond has to match the complexity it faces. You get that by distributing control, not centralising it.

What senior leaders need isn't a dashboard — it's a flexible operating system that gives them real signal. Most planning tools give leaders the illusion of control while removing them from the texture of the work.

Cutler's most useful framework is a matrix of explicit versus implicit crossed with intentional versus not intentional. Some things in organizations are visible and deliberate — strategy docs, OKRs. Some are invisible but still deliberate — the unwritten norms leaders actively shape. Some are visible but accidental — bureaucratic leftovers nobody questions. And then there's the fourth quadrant: the invisible, unintentional patterns that quietly run the show. That's where the org actually lives.

You probably won't change the company. But you can shape the local environment — your team, your rituals, your standards. And when people look back, those are the moments they remember. Not the reorg. The Tuesday morning where someone felt safe enough to say the thing everyone was thinking.

That's the beautiful mess.

Jon Smart traces how modern management practices emerged from military doctrine, industrial engineering, and Enlightenment philosophy. But I wanted a wider lens.

For roughly 95% of human history, we organised through "reverse dominance." The group functioned as a collective alpha. Anyone who tried to command others was actively suppressed. That's the anthropological record.

The shift didn't start with factories. It started when states discovered that certain crops — wheat, barley, rice — were easy to store, transport, and tax. Quantifiable labour is controllable labour.

From there, the architecture layered up. Egypt gave us nested supervisory hierarchies. Babylon gave us performance incentives. China built the world's first meritocratic civil service. And then came the Prussians.

After Napoleon humiliated them at Jena in 1806, the Prussian military rebuilt itself around two competing philosophies. Befehlstaktik — lead by detailed orders, no discretion. Auftragstaktik — lead by intent, leave the how to the people closest to the work.

When American railroads needed organisational models, they hired former military officers. Those officers chose Befehlstaktik. Not because it was better. Because it was familiar.

Taylor formalised it. Fayol structured it. Weber justified it. And the 20th century exported it globally through business schools as Cold War soft power.

The entire command and control model was designed for manual labour — turning bodies into predictable, mechanical output. It was never designed for knowledge work, where productivity depends on judgement, creativity, and intrinsic motivation.

We didn't start with hierarchy. We built it. And we can choose to build something different.

Last week I described how four systemic archetypes spiral into each other. This week we zoom into the first. The one that starts it all.

A reinforcing loop drives performance. Growth feeds growth. Customers bring customers. For a while the curve looks exponential. Dashboards glow. People get promoted.

Then something shifts. Not suddenly — gradually. The same effort yields less. The curve bends. Because a balancing loop, one that was always there, started to bite. That balancing loop is the constraint.

In SaaS, the constraint is often invisible for years. Acquisition keeps masking it. Customers join, extract what they can, quietly leave. The net number looks fine — until it doesn't.

In manufacturing, a factory scales output by adding shifts. Throughput rises. Then quality incidents climb, maintenance gets deferred, experienced workers burn out. The reinforcing loop runs. The constraint tightens around it.

In healthcare, a hospital expands patient volume without expanding diagnostic capacity. Admissions go up. Outcomes go down. Growth created the pressure that degrades the service.

Here's what makes this dangerous. The early signal is success. The constraint doesn't show up in the metrics you watch. It shows up in the ones you don't. Retention you didn't track. Engagement that plateaued two quarters ago. A team that stopped raising concerns because every town hall celebrated the top line.

The instinct when growth slows is always the same: push harder. More budget. More headcount. More campaigns. For a while that works — which is exactly why the pattern persists. Senge called it "the harder you push, the harder the system pushes back."

The ones that break the pattern ask an uncomfortable question early: what is the constraint we're not measuring? Before the curve bends.

The budget was signed in December. Targets locked. Everyone aligned. Thirty days later there was already a gap you could see from orbit.

That's not an execution failure. What failed was the premise underneath the number.

Every growth system runs two forces at the same time. One pushes forward. One pushes back. Companies treat them very differently.

The forward force gets the corner office. New logos, green dashboards, confetti in Slack. The other force gets nothing. No team, no channel. It shows up six months later as a number that should've been higher.

When retention sags, what does the org reach for? Discounts. SPIFFs. Switching promos. These work — for a quarter. Then they erode whatever value was holding the product together, and kill the urgency to invest in what would actually fix it: stickiness, product depth, value customers can feel. This is Shifting the Burden — the fundamental solution atrophies while the symptomatic one gets a bigger budget every cycle.

And those conversion incentives? They attracted price-sensitive customers who churn faster. That aggressive upselling? It accelerated the very churn you were trying to stop. Each fix produced a delayed consequence that worsened the original problem. This is Fixes that Fail.

Underneath, the quiet killer. Under revenue pressure, the instinct is to defer product investment. But when the product hasn't evolved in years and pricing has been the heavy lifter, that deferral hardens the structural limit every quarter. This is Growth and Underinvestment — the archetype that makes all the others permanent.

Four patterns. One spiral. Every intervention that ignores the spiral makes it spin faster.

Before you intervene — find both forces. What's accelerating? What's resisting? Acceleration is a tactic. Diagnosis is strategy.

That's the work we do with noise.os. Less confetti. More stethoscopes.

I run most mornings. Not fast, not far. Enough to think.

There's something about the rhythm — the breath, the pace, the distance behind and ahead — that makes listening easier. Six kilometres into the run, a podcast becomes conversation. A problem becomes structure. An idea becomes clarity.

These are the notes from those runs. Not summaries. Not reviews. Just the moments where something someone said made me stop mid-stride, and the thinking that followed.

I read the way most people read — too much, too fast, too many tabs open.

But some things slow me down. A paragraph I read three times. A margin note that turns into a page of my own thinking. A citation that leads me backwards through a genealogy of ideas.

This is Marginalia — not reviews, not summaries. Just the notes in the margins, and where they took me. The thinking beneath the thinking. The threads no one else could pull.