The Branches We Skipped

For ten thousand years, finite attention has decided which branches of the tech tree humanity could climb. The bottleneck has shifted before. This time it's shifting at the level that creates every other shift. Cognition itself.

The aeolipile sat on a shelf

In Roman Alexandria, around 60 AD, a mathematician named Heron built the first steam engine. A hollow sphere with two angled nozzles, mounted on a pipe. Fill the pipe with water, heat it, and the escaping steam spins the sphere. It worked. Heron described it in a manual on pneumatics and used it to demonstrate the principle to students. Nobody ran anything on it. Slave labor was cheap, and there was no economic gradient that demanded mechanical power. The aeolipile sat on a shelf for 1,652 years. Newcomen's atmospheric engine in 1712 is the point where the branch finally caught.

In Song dynasty China, around 1040, a commoner named Bi Sheng invented movable type. Clay characters, reusable, fully functional. Four hundred years before Gutenberg. It did not catch. Written Chinese has thousands of characters and the cost-benefit collapsed against woodblock printing, which was already good enough. The branch died in the cavity it was born in. The same idea, fired into a different cavity four centuries later, in alphabetic Latin Europe, conquered the world in fifty years. Same branch, two reaches, only the second had pump.

In 1901, Greek sponge divers pulled a corroded lump of bronze out of a shipwreck off the island of Antikythera. Inside was a geared analog computer, built around 150 BC, designed to predict astronomical positions. The precision and complexity would not be matched again for roughly 1,500 years, until the astronomical clocks of the medieval European monasteries. Whoever made it took the knowledge to the grave. The Hellenistic Mediterranean had no economic or scientific use loud enough to keep the gain on it. The branch went dark with one death.

Three branches. Three reaches. Three long dormancies. None of these required new physics. None required resources humanity did not have. They required a cavity that would let the gain flow to them, and the cavities they landed in had other dominant signals. The branches stayed subthreshold for centuries. Sometimes for most of recorded history.

A weird thing about lasers

The pattern I just described has a name in physics, and it is worth knowing the name because the name tells you why the pattern is scale-invariant. It shows up in a human lifetime, in a civilization, and in a box of light.

Pump energy into a laser cavity and something unintuitive happens. There are many modes the light could oscillate in: different frequencies, different spatial patterns, different polarizations. The cavity supports all of them mathematically. In principle, the energy could spread across all of them.

It doesn't. One mode wins.

One mode's gain is slightly higher than the others, and the nonlinear dynamics amplify that small edge until the winning mode has absorbed almost all of the pump and suppressed everything else. Physicists call this mode competition. The losers aren't broken or forbidden. They are subthreshold: dynamically real, dynamically starved. They exist in the cavity as possibilities. They just never get to oscillate.

You don't have to be a physicist to feel what this is like. You've felt it on a radio. Spin the dial across the AM band and you'll hit one strong station that drowns everything else in static. Other stations are broadcasting. Their signals are reaching the antenna. The receiver simply can't pick them up, because one signal is so much louder than the rest that it captures all the gain. Cut the strong station's power and suddenly the dial fills with stations that were always there.

A radio receiver is a cavity. So is a laser. So is most of what runs on a finite flow. So was Heron's Alexandria. So was Bi Sheng's China. So was whoever made the Antikythera mechanism.

Hold that picture. Mode competition is not just a thing lasers do. It is what happens any time a finite flow is allocated by a winner-take-most rule. Capital chasing returns. Search traffic chasing top-ranked links. A firm's payroll chasing whichever bottleneck currently determines output. Attention chasing whichever task currently eats the most hours.

Which is the cavity I actually want to talk about.

An investor's week as a cavity

Picture a first-year associate in private equity or credit. The job description says reasonable things: source deals, run diligence, build models, write memos, present to the IC. In practice, the week doesn't look like the job description. It looks like whichever task ate the most hours.

Traditionally that task is model-building. Pulling CIMs apart, tying out financials, rebuilding a three-statement model from someone else's template, chasing footnotes across PDFs, formatting the output so the MD can skim it in ninety seconds. The work is real and it matters. But it is coordination with spreadsheets, not judgment. And it captures the cavity. Sixty, seventy percent of the week flows into it.

The rest of the job, the part that would actually train the associate into an investor, sits in the subthreshold. Reading industry reports deeply enough to form a view. Calling operators who actually run the businesses. Building unusual analyses that nobody asked for but everyone would want if they saw them. Spinning up real-time data pipelines off direct API calls. Orchestrating AI agents to test a thesis a hundred ways in parallel.

None of this is impossible in a pre-AI week. It is just subthreshold. The dominant mode eats the pump. The rest oscillates at zero amplitude.

The radio in Section 2 and the week above are the same picture, drawn twice. A finite flow. A dominant mode. Subthreshold space. The dominant signal is loud not because it deserves to be, but because it is the one the receiver locks onto when the gain has nowhere else to go.

When the cavity changes

Now watch what happens when AI absorbs modeling.

Not "augments." Not "speeds up." Absorbs. Today's frontier models do most of this and not yet all. They can pull a CIM apart, tie out financials, draft a three-statement model from raw filings, source the comps, format the output. The pieces they can't do yet are the pieces that get cheaper every six months. Within a few model generations, the dominant signal in the cavity loses the gain it used to capture by default. Its underlying value didn't change. Its cost did, and cost is what determined which mode the pump went to.

What happens next is the part most takes about AI miss.

The associate doesn't get "free time." She gets a different job, built from the same person. The pump that used to flow into modeling flows somewhere else, and where it flows is determined by which mode now has the highest gain. For her, that is probably agent orchestration. The 100-agent debate she always wanted to run on a thesis but couldn't find the hours for. The real-time data pipeline she sketched on a napkin and shelved. The novel analysis nobody assigned her but everyone would read. Mode #8 becomes mode #1. Same cavity, completely different output.

On Visual 3, drag the slider up and watch which mode emerges as AI eats Modeling. The loudest mode in the rebalanced cavity is not the one that used to be loudest. It is the one whose underlying gain had been waiting all along.

This is mode competition at one human scale. It is not a story about productivity. It is a story about which mode the pump reaches, and what the cavity rewards once the old dominant signal stops capturing the gain.

The same shape, one scale up

The same physics runs at the firm scale. A firm is also a cavity with modes competing for a finite flow: payroll, management bandwidth, executive attention. The dominant mode for most firms is coordination. The literature calls this the Coasean boundary. A firm exists up to the point where the cost of coordinating one more person inside the firm exceeds the cost of contracting for it outside. Coordination is what captures the pump.

When AI absorbs coordination overhead, the bookkeeping, the scheduling, the compliance, the glue, the firm's cavity changes geometry. The dominant mode loses its edge. Business shapes that were subthreshold start to oscillate. A one-person telehealth clinic. A five-person SaaS with eight figures of revenue. A specialist agency in Manila serving clients in Manhattan. None of these were viable as real businesses ten years ago, not because they were bad ideas but because the coordination overhead would have eaten them.

I wrote about this scale at length in The New Stable Orbits. The argument there is the argument here, told at the firm scale instead of the week scale. Same mechanism. Different cavity. The set of viable shapes is determined by which mode currently captures the pump, and AI is changing which mode captures the pump in a structural, persistent way.

Two scales now. One human week. One firm. Both cavities. Both with dominant modes, both with subthreshold space, both rebalancing when the pump finds new geometry. The mechanism is scale-invariant.

Which raises the question of how far up the scaling goes.

All the way out: the tech tree

Pull back one more time. Past the week. Past the firm. All the way out.

Humanity has been exploring a space of possible knowledge and capability for about ten thousand years. At every branch point, every moment where a generation could have worked on one kind of problem or another, built one kind of tool or another, asked one kind of question or another, we picked. We had to. Attention is finite at civilizational scale the same way it is finite inside a week, and the branches that paid off immediately are the branches that got worked on. Everything else waited.

The path we actually took feels, in retrospect, like the only one. Agriculture, metallurgy, the printing press, steam, electricity, computation, the internet. A single glowing line through an enormous space. It looks inevitable because it is the only line we can see.

On the left of the trunk: things we abandoned. Heat engines that could have powered countless small workshops, foreclosed in favor of centralized power. Tesla's wireless transmission of electricity at Wardenclyffe in 1901, defunded by JP Morgan and dormant for a century before short-range wireless charging recovered a fragment of it. Electric cars that outsold gasoline cars in 1900, killed by cheap oil and the Model T, briefly revived as the GM EV1 in the 1990s and then literally crushed in the desert when GM recalled and destroyed every car in 2003, before Tesla, the company, reactivated the branch in 2008. Movable type that died once in China and got reborn four centuries later in Europe. Index funds, mocked when Bogle started Vanguard in 1975, eventually eclipsing the active managers who had laughed at him. Each one a branch we reached and walked away from, because the cavity around it rewarded something else.

On the right of the trunk: things AI is reaching now. Materials we did not need (chemistry too vast to search by hand). Proteins we could not fold (biology below the resolution of every previous instrument). Credit at the long tail (people and businesses traditional underwriting could not see). Sensing at scale (instruments collecting more signal than human teams could analyze). Languages we could not speak (translation between any two of seven thousand human languages, at conversational quality, in real time). These are not predictions. These are branches already lighting up, in real time, in real labs and real markets, because something finally changed the cost of reaching them.

This has happened before. The agricultural revolution moved most of humanity off subsistence food production, and an enormous share of attention that had been locked into one mode for millennia activated other modes: manufacturing, governance, professional inquiry, art as a livelihood. The industrial revolution moved labor off manual production and let services, professions, and modern science scale into populations of millions. Each of these was a bottleneck-shift at the species scale. Each opened branches that had been latent for thousands of years.

What is different about AI is not that the bottleneck is moving. The bottleneck has moved before. What is different is which bottleneck. Agriculture freed bodies from working the soil. Industrialization freed muscles from pulling the lever. AI is the first shift that moves the bottleneck on cognition itself, and cognition is upstream of every other bottleneck.

Someone had to think of the plow before anyone could use it. Someone had to think of the steam engine, the assembly line, the integrated circuit, the internet. Every prior bottleneck-shift was created by a cognitive act. AI is the first technology that changes the cost of those cognitive acts at scale. It does not just open one new branch of the tree. It changes the rate at which any future branch can be reached.

Physicists have a name for a system that has access, in principle, to states it will never actually visit. They call it broken ergodicity. A glass is the canonical example: a liquid that got stuck in one pocket of its possibility space on its way to finding its crystal ground state, and now sits there forever, exploring a tiny region of configurations while the rest of the state space remains mathematically real and dynamically unreachable. The glass isn't missing the other states. It is just stuck.

Every generation worked the branch it was born on. The dormant branches accumulated. A vast shadow tree of things we could have built, questions we could have asked, disciplines we could have founded. Some of that shadow tree we can name. Materials we couldn't synthesize. Proteins we couldn't fold. Climate systems we couldn't model. Languages we couldn't speak. And some of it we cannot name at all. Categories of question that required a kind of cognitive leverage no individual mind could supply. Whole regions of possibility space that nobody has mapped because nobody has had the capacity to even start.

The horizontal band below the tree, in the visual above, is that unnamed region. It does not get smaller as reach grows. AI lights up branches on the right that humanity could not climb alone. It does not light up the territory beyond. That territory is the deepest claim of the essay, and the most important one. The reach is changing. The frontier of what we can name is moving. But the unmapped is still there, and it is still bigger.

What we are living through is not the invention of new knowledge. It is the first moment in the history of the species when the barriers between branches are low enough that small groups of humans, with enough leverage, can reach into regions of the tree that used to require entire civilizations to even approach. The shadow tree is still shadow. But the reach just got much, much longer.

The fractal

The week. The firm. The species. Three stories. One shape. At every scale: a finite flow of attention, a dominant mode that captures it, and a space of dormant possibilities the flow never reaches. The nouns change. The mechanism does not.

The pattern is scale-invariant because the constraint is scale-invariant. Attention is finite inside one mind and finite inside one species, and in between it is finite inside every structure humans have ever built. The dormant modes are always there. The unreachable pockets are always real. The only question is whether the dynamics ever change enough to let them activate.

AI changes the dynamics at every scale at once. Not a productivity tool on one layer. A phase change across all of them simultaneously: the week, the firm, the tree.

What we never had the hands to climb

The honest version of this argument is that the branches we are about to reach are not automatically good. They are just reachable. Some will be abundant. Some will be strange. Some will be hard to integrate into the lives of the people who suddenly have access to them. The claim is not that exploring the dormant field is utopian. The claim is that the exploration itself is the event, and the event is bigger than the productivity conversation we are currently having about AI.

Most of what gets written about AI right now assumes the cavity stays the same and AI just moves work around inside it. Faster modeling. Faster writing. Faster customer service. That framing is small because it treats AI as a productivity multiplier on the dominant mode. The dominant mode is not the point. The dominant mode is the thing AI is absorbing. The point is what the pump reaches once the dominant mode stops capturing it.

Spin the dial back to the radio. The associate doesn't become faster at modeling, she becomes an investor who runs real-time pipelines off direct API calls and orchestrates agent swarms to attack a thesis from a hundred angles. The firm doesn't become a leaner version of itself, it becomes a shape of firm that wasn't in the old distribution at all. And the civilization doesn't move faster down the current branch of the tech tree, it starts reaching into branches nobody has been on.

The strategic question isn't how do I use AI to do my current job faster. It is which of my subthreshold modes is about to become reachable, and what am I going to do when it does.

The constraint on human progress was never the space of possibilities. It was the rate at which we could reach into it. The dial just got turned down on the only signal we could hear.