Point the Machines at What’s Scarce

In Part 2, I argued that machine labor on demand should fill the gaps people don’t want to fill.

But labor isn’t the only reason things are scarce. Some things are limited by physics.

This piece is about Change #2: pointing machine labor on demand, and above all its intelligence, at what’s scarce, until the scarce things become abundant.

Stocks and flows

To see why some things stay scarce, it helps to think like a systems scientist.

The best guide I know is Donella Meadows, lead author of The Limits to Growth (1972) and Thinking in Systems, published in 2008, seven years after her death.

Her core idea is simple. Every system is made of stocks and flows.

A stock is anything you can count at a given moment: water in a bathtub, grain in a silo, ore in the ground, trees in a forest.

Flows are what fill a stock and what drain it: the faucet and the drain.

If more flows in than out, the stock grows. If more flows out than in, it shrinks, no matter how big it was to begin with.

That’s what scarcity really is. A thing is scarce when we drain a stock faster than we refill it, or when the faucet is too small for everyone who needs what’s in the tub.

And that’s what abundance really is: inflows that keep up with what we use, without draining the stocks we depend on.

Keep that picture in mind. Everything in this piece is about turning up the faucets and protecting the tubs.

Natural limits

In the world we live in today, everything eventually runs into the natural scarcity of resources.

There is a finite amount of resources on our planet, and a finite amount of space. As a result, there are natural limits to growth.

Which is why Change #2 is dedicated to funneling machine intelligence into reducing those scarcities.

Intelligence and labor on demand fundamentally change our economic system. As many have written, it forces us to identify the key bottlenecks to automation and focus our efforts there (to accelerate towards this future).

This could be through technological breakthroughs...

Through novel discoveries of new materials...

Or new sources for gathering existing materials...

And through raw intelligence and labor thrown at identifying every resource that becomes a bottleneck in the economy, and then solving for said bottleneck.

If we pool intelligence and labor resources to do whatever it takes to increase supply of said scarce resources, we can ultimately eliminate the bottleneck by producing, or mining, or redirecting, or synthesizing whatever key materials we’re missing to provide universal basic goods and services at scale.

Why things are scarce

It helps to be precise about why things are scarce, because the reasons differ, and so do the fixes.

Some things are scarce because we lack the labor to provide them: doctors, builders, caregivers. Machine labor on demand solves those directly.

Other things are scarce because of production bottlenecks: chips, power plants, housing, supply chains. Those need machine intelligence pointed at the bottleneck itself, which is the heart of Change #2.

And others yet are physically finite: land, minerals, fresh water in a given place. Those need better extraction, substitution, and recycling, and eventually mining materials from beyond Earth.

A town with some farmland and a town in the middle of the desert have very different odds of feeding themselves in a sustainable manner.

There is also scarcity by human nature: beachfront property, original art, the attention of one particular person.

No amount of labor fixes the demand for those kinds of things, which is why money, or something like it, will likely still be needed to coordinate demand around those.

The bottlenecks nobody talks about

The AI industry, particularly on podcasts like Dwarkesh Patel’s, often talks about scaling resources for the purpose of training better AIs and producing recursive self-improvement that leads to an intelligence explosion that leads to more and better AIs.

However, that’s where the chain of thought always seems to end.

The thinking on what resources are truly scarce and need to be scaled typically stops at things like NVIDIA chips and Dutch EUV machines (which make said chips).

There’s rarely enough talk about all the other important things we need to produce at scale, and what it would take to scale those things from today’s level of scarcity to an actual state of abundance.

That’s where machine intelligence should be applied most directly: not only to scale the creation of more machine intelligence, but to scale the production of everything else we need to provide for humanity to thrive.

This includes, but is not limited to:

Food and water.
Energy.
Housing.
Healthcare services, treatments, and surgeries.
Sanitation, waste removal, and climate control.
Precious metals for chips and machinery.
Transportation services and vehicles.
Communication infrastructure.
Entertainment.
Physical goods and hardware.

Seven of these hold the essentials every household needs covered: food, water (sanitation included), energy, housing, healthcare, transport and communication. The rest are what it takes to make them, and what makes life more than getting by.

The idea here is to identify every single physical bottleneck for each of these categories, and direct machine intelligence towards reducing those critical constraints in the most sustainable way possible, on a rotating basis as the bottlenecks change over time, for as long as human civilization lasts.

Why the “sustainable” way? Not for greenwashing or virtue-signaling; to literally sustain supply.

If you’re able to increase a steady state of supply for all the constraints in a system, you can keep the stocks and flows at a level that creates abundance without burning out.

So machine intelligence and labor should also be directed toward optimizing the flow of supply to maintain this steady state. The discoveries can speed up; the stocks they draw on should hold steady. Preparing for peak demand on the year’s peak day, but for everything.

How machine intelligence breaks scarcity

The first kind of scarcity, labor, is what machine labor on demand solves directly. Production bottlenecks and physical limits need something more: new knowledge. We have to discover better ways to make things, and better things to make.

That’s where machine intelligence may matter most. Not as a worker, but as a scientist. Here’s how I see it playing out across the fields that matter most.

Biology: the data-rich frontier

Biology got to AI first because it had the most data. Decades of genome sequencing and protein structure research gave models something to learn from, and the results came fast.

The 2024 Nobel Prize in Chemistry went to David Baker for computational protein design, and to Demis Hassabis and John Jumper of Google DeepMind for predicting protein structures.

For scarcity, this matters because biology is how we make food and medicine.

Designed proteins and enzymes can mean crops that need less fertilizer and water, proteins brewed by microbes instead of raised on farms, enzymes that break down plastic waste, and drugs discovered in years instead of decades.

Food and healthcare, two of the biggest items on the list, run through biology.

Materials science: where the flywheel takes off

Materials science started with less data than biology. That’s changing fast, and I think this is where the scarcity flywheel really takes off.

In 2023, Google DeepMind’s GNoME predicted 2.2 million new crystal structures, including about 380,000 it expects to be stable. For context, only about 20,000 of the crystals found by experiment are stable by the same test, and other computer methods had raised that to about 48,000.

In a companion paper, Berkeley Lab’s A-Lab, where AI directs robots to make new materials, produced 41 of 58 attempted compounds over 17 days of independent operation.

Some researchers have since questioned how novel some of those compounds were, which is a healthy sign that the field is checking its own work.

In 2024, Microsoft and the Pacific Northwest National Laboratory used AI to screen 32.6 million candidate materials for a new battery electrolyte, narrowing them to 18 promising candidates in about 80 hours of computing.

The winner uses roughly 70% less lithium than today’s lithium-ion batteries, and the team built a working prototype battery with it.

And the 2025 Nobel Prize in Chemistry went to Susumu Kitagawa, Richard Robson, and Omar Yaghi for metal-organic frameworks: porous materials that can pull water out of desert air and capture carbon dioxide.

Remember the town in the desert with poor odds of feeding itself? This is the kind of science that changes those odds.

Here’s why materials science is the flywheel. Better materials make better chips, batteries, solar panels, magnets, and robots. Better chips and robots make more and faster labs.

More labs find better materials. Each turn of the wheel speeds up the next one, and unlike almost any other field, materials touch every bottleneck on the list: energy, compute, housing, transportation, and the machines themselves.

Physics and energy: the master input

Every other scarcity eventually comes down to energy. Cheap, abundant energy lets you desalinate water, run microfactories, recycle metals, and power the machines that do everything else.

Machine intelligence is already being applied here: to control the superheated plasma inside experimental fusion reactors, to design better solar cells and batteries, and to balance supply and demand across power grids.

That last one matters directly for the microgrids I describe in Part 5.

Chemistry: making things with less

Much of modern industry runs on chemical reactions that take enormous amounts of energy, like making fertilizer, cement, steel, and plastics.

Better catalysts, many of which will likely be discovered with AI, can make the same things at lower temperatures, with fewer inputs, and with less waste. Chemistry also powers recycling: recovering metals from old electronics and batteries is, in effect, mining what we’ve already mined.

Earth science: finding what’s already there

Some scarcity is just a search problem. KoBold Metals uses AI to explore for minerals, and in 2024 announced a huge copper find at its Mingomba project in Zambia.

It also shows where the bottleneck moves next: the company estimated the mine could cost around $2 billion to build, with first output targeted for the early 2030s. Finding is getting faster. Building still takes time.

Computing: doing more with less

Finally, machine intelligence can make itself cheaper. More efficient algorithms mean more answers per watt and per chip, which is exactly what’s needed to keep AI from being something only a few labs can afford.

Efficiency is how intelligence becomes a local resource instead of a centralized one.

Put these together and you can see how each kind of scarcity gets attacked. Labor shortages get machine labor. Production bottlenecks get new materials and better chemistry.

Physical limits get substitution, exploration, recycling, and eventually space. Only land in a given place and scarcity by nature, like the beachfront and the original painting, stay scarce.

That’s how scarce things become abundant.

Abundance answers what we make. It doesn’t answer how we organize ourselves to build it.

Today, that job belongs to the corporation. In Part 4: Quests Should Replace Corporations, I’ll argue that it shouldn’t.