PrefaceThe Last Scarcity

1. Two costs falling together #

Epoch AI compared models at a fixed capability threshold. On the MMLU general-knowledge benchmark, GPT-4-0314 scored 86 in March 2023, at a weighted API price of about $37.50 per million tokens. By February 2025, Gemini 2.0 Flash crossed the same line with a score of 88, at about $0.18. In roughly twenty-three months, the price of reaching that fixed capability fell to about one two-hundredth of what it had been (Epoch AI, 2025).

Those limits matter. Epoch AI weights input and output tokens three to one, and it leaves reasoning models out of the per-token comparison because they generate large volumes of intermediate tokens. The series measures the API price of a fixed benchmark capability. It does not measure what it costs a company to finish a job. Connecting the model, checking the output, reworking it, and owning the result sit outside that $0.18.

Even so, the curve is changing the floor price of digital knowledge work. Wherever the inputs and outputs are digital, and quality can be checked in a digital environment, many steps once bought by the hour can go to a model first. First-draft code, research collation, document summaries, report structure, and first-pass translation all sit in that class. A model call does not drive the cost of the whole job to zero. It does drive down the repeatable execution inside it.

The same capability also shortens the path to knowledge that already exists. Material that used to require search, screening, a full reading, and then a summary can now be organized by AI first, with a person returning to the original sources to check. What got cheaper is the friction of finding and first-pass processing. Facts did not become true on their own, and expertise was not thereby understood, verified, and transferred.

When getting it done and looking it up both get cheaper, advantages built on skilled execution and on information gaps get thinner. Cheap model calls will not tell an organization which problems are worth solving, which standard to apply to the problem in front of it, or whether the material on hand is enough to act. The scarcity that remains I call the last scarcity.

2. "Last" does not mean the others have gone #

Execution has not gone to zero. AI cannot turn construction, nursing, or last-mile delivery into a model call. Audit, compliance, and licenses do not vanish because tokens got cheaper. Even in the digital world, people still pay to connect the system, check the output, rework it, and take responsibility. When this book says execution is approaching zero, it names a direction in digital knowledge work: the marginal price of execution is falling fast. It is not a mathematical zero, and it is not a promise that every industry arrives on the same day.

Compute and energy have not stopped being scarce either. Chips, power, data centers, and access can all become bottlenecks, and organizations do not face the same price or the same difficulty getting them. The difference is that compute and energy have prices and supply constraints. They can enter a budget, a purchase order, and a resource plan. Capital can ease some of those bottlenecks. It cannot answer whether they are worth buying, where they should go, or when spending should stop.

An organization can also hire talent, retain advisors, or buy capability through acquisition, and so borrow answers and experience from outside. Purchasing cannot confirm the conditions of use for the organization in front of you, and it cannot take the result off the person who is accountable. Models, compute, and outside experience can all become material for judgment. They do not become this organization's judgment on their own.

So "last" does not mean the last scarcity in the universe, and it does not deny the value of compute or energy. It names the competition this book is about. When execution and retrieving existing knowledge are no longer the main barriers, the harder gap to copy is whether an organization can form a reliable judgment on a problem it has not seen before.

3. Four words, one line #

This book will keep using four words. They name the same thing at different layers, and they are not interchangeable. The last scarcity is the name of the thing. Adaptive insight is the capability in a person. Judgment is the asset in an organization. A standard is judgment written down, in a form people and AI can reuse.

Adaptive insight is not mystical. It shows up as four moves in a row. First you find what standard the present judgment is missing. Then you borrow a standard, and its conditions of use, from people and organizations that have already verified it. Then you judge whether it can move onto the problem in front of you. Then you verify with a cheap experiment, and you revise from the result.

When one person completes the four moves, the result is personal adaptive insight. Personal judgment becomes organizational judgment only when the organization uses it in common. Judgment becomes a standard only when it is written down, so it can be checked and reused. Rewrite the DNA is not mainly about how one person gets sharper. It is about how an organization settles adaptive insight into shared judgment standards and shared context. Personal adaptive insight is hard to scale by copying people. An organization has to get that capability out of a few heads and into an asset the rest can use.

4. From the crack to rewriting the DNA #

Fourteen chapters in four parts, following the sequence of one organizational evolution.

Part I, "The Great Surplus" (Chapters 1–2), finds the crack. How far have the costs of execution and of retrieving existing knowledge actually fallen, and where does the claim apply? Why do most organizations get no stronger as the tools multiply? The job here is to see which premises of the old world are giving way.

Part II, "The Last Scarcity" (Chapters 3–5), takes apart the capability the tools did not deliver. How it runs in a person, how it becomes an asset in an organization, and why it is the limiting resource an AI-native organization has to run.

Part III, "Producing the Scarcity" (Chapters 6–9), is how an organization produces it. The work starts with the standards cycle, then Clarity Method and Interference Method, then the company-wide consensus cycle. The same judgment moves from one person to a team, and then to the organization.

Part IV, "Rewriting the DNA" (Chapters 10–14), puts that capability into five fields: people, time, product, relationships, and the organization itself. The last chapter runs a first replacement in ninety days, so the new way of judging actually enters daily work.

The book distinguishes fact from claim. External numbers carry sources. Company-reported and unaudited figures are marked for quality. Where evidence is missing on one side, the gap stays visible; a neat story is not a law. Each chapter ends by landing the judgment in something a reader can start on Monday. The book is the complete argument as it stands. Regenic.ai keeps revising in public. bioby.ai is the first field experiment.

You do not have to accept every later conclusion first. A better way to read is to treat each chapter as a judgment that can be argued with: look at the evidence, then the boundary, then test it on your own organization.

5. Start with two questions #

Before you go further, ask your organization two things:

1. If the cost of execution and of retrieving existing knowledge in digital work keeps falling, how much of what is most valuable in the company today will still be valuable tomorrow?
2. Faced with a problem we have never done, can we find what standard is missing, borrow one that has already been verified, judge whether it transfers, and correct it with a cheap experiment?

If the first question has no answer, the organization's advantage may still sit on execution and information gaps that are getting cheaper. If the second has no answer, the organization is still depending on a few people to improvise: when they leave, it pays for the same education again.

Neither question requires you to believe AI can do everything. They require the opposite: first separate what is getting cheaper from the responsibilities that still have to sit with people. Everything that follows can start from a rebuttal. Chapter 1 takes the most direct objection: will execution really approach zero?

Chapter Acceptance Self-Check (preface special contract) #

1. Claim restatable in one sentence ✓.
2. Framework ✓ (costs fall together; judgment does not; four moves of adaptive insight).
3. Case contract ✓: no cases in the preface; one verified data set (Epoch AI, 2025, ~200×); strongest counter (compute/energy) answered.
4. Four core words pinned once ✓ (name / adaptive insight / judgment / standard).
5. Reading map + chapter rules + two honesty clauses + prototype note ✓; closing two questions + doubt handoff to Chapter 1 ✓.
6. Fluency ✓: rewritten in English voice from the Chinese authority; current prose-standard.