synthesis · syntheses/What's Scarce When Execution Is Cheap.md
What's Scarce When Execution Is Cheap
Filed 2026-06-02. Three sources in this wiki — written in three different decades and disciplines — independently describe the same structure: when the cost of producing the default, conventional output collapses, value migrates to the scarce, high-information, high-judgment contribution that (for now) only humans reliably supply.
Answer
Norbert Wiener (information theory, ~1950), Dan Shipper (economics of expertise, 2026), and Ethan Mollick (organizational change, 2025) reach one conclusion from three starting points. The shared move: identify the thing AI makes abundant (and therefore cheap and low-value), then locate value in its scarce complement.
| Dimension | Norbert Wiener (1950) | Dan Shipper (2026) | Ethan Mollick (2025) |
|---|---|---|---|
| Frame | information theory & aesthetics | economics of expertise | organizational change |
| What's abundant / cheap | the predictable, clichéd message (Clichés Carry Less Information) | the commoditized "residue of competence" → Slop | individual execution — code, research, content |
| Why that's low-value | predictable = low information | everyone shares the same model trained on the same corpus → sameness | execution was never the point; it was a proxy for a goal |
| What becomes scarce / valuable | originality, the improbable, the idiosyncratic | difference — what only a human expert sees (AI Commoditizes Expertise) | deciding what to do (The Bottleneck Shift); organizational innovation |
| Who supplies it | the artist | the human expert | the experienced worker + Leadership, Lab, and Crowd |
The unifying thesis
All three are versions of: value = improbability. Wiener gives the information-theoretic floor — a message's value rises with how unexpected it is, so the statistically-likely default (what an LLM emits unless pushed) is by construction the low-value one. Shipper gives the market consequence — once that default is free and everywhere, buyers will only pay for difference. Mollick gives the organizational consequence — once execution is free, the binding constraint becomes the human judgment about which execution is worth doing. "Humans know what needs to be done right now" (Shipper), "the bottleneck is figuring out what research to do" (Mollick), and "clichés contain less information than… great artists" (Wiener) are the same sentence in three dialects.
Where they diverge (and why it matters)
-
Wiener's quiet time-bomb. The same excerpt says "today's LLMs corroborate" his theory — but LLMs can also be steered toward the improbable. If models learn to manufacture difference on demand, the human monopoly on high-information output is contingent, not permanent. This is exactly the AGI objection Shipper sidesteps and the open thread in AI and Employment — Replacement vs Expansion.
-
Is the scarce value actually being captured? Mollick's data says no — individual gains aren't reaching organizations (Individual vs Organizational AI Gains), partly because workers hide their best AI work (Secret Cyborgs). Shipper's Every says yes — but only because it did the organizational innovation Mollick prescribes. So the thesis may be true in principle yet unrealized in practice for most firms.
-
A second scarcity. All three treat human judgment as the bottleneck — but Token Shortage says compute is also scarce. If both judgment and compute bind, the constraint is whichever runs out first, and "just add more agents" fails on the compute side regardless of judgment.
-
A cognitive pillar (added 2026-11-08). Willingham explains why the scarce human contribution resists copying: judgment is bound to domain knowledge built through practice, not a portable skill. That supports the thesis. It also adds a risk none of the three pillars considers: if AI absorbs the practice that builds knowledge, the supply of judgment may shrink just as demand for it grows. See Method Reflection (T3) for the same question applied to this wiki.
-
The scarce thing may be checking, not deciding (added 2026-11-08). Anthropic — Investigating Vault Particles with Claude is the first case here where people did expert work without the expertise, by leaning on the model. It mostly worked. What they could not buy was the ability to tell when the model was wrong: "in areas we knew well, we caught mistakes quickly, but they cost us time in areas we didn't." If execution is cheap and even deciding can be assisted, the residual human function narrows toward verification — and verification is the one thing that can't be delegated to the system being verified (Verification Requires Domain Knowledge). That is a sharper, more defensible statement of the thesis than "judgment," and a more demanding one: it requires people to keep building knowledge they no longer need in order to produce.
The sharpest open question: if "value = improbability" and models get better at improbability, does the human premium erode — or does each capability gain just relocate the scarce judgment one level up (from doing → deciding → deciding what's worth deciding)? The three sources assume the latter without proving it.
Connections
- Placement and Judgment Over Raw Execution — companion synthesis (read as a pair): this one names judgment as the scarce complement to cheap execution; the companion adds placement and shows both expressing themselves as three moves that recur across the wiki's AI-and-work and engineering-craft domains.
- Clichés Carry Less Information — Wiener's information-theoretic floor for the whole thesis.
- AI Commoditizes Expertise / Slop — Shipper's commoditization → sameness → demand-for-difference.
- The Bottleneck Shift — Mollick's "what to do" scarcity (and the two-author convergence this synthesis extends to three).
- The Infinite Backlog — the felt experience of the judgment bottleneck: infinite work, finite deciding.
- The Human Sandwich — the operational form: the human supplies the framing and the judgment AI can't.
- Individual vs Organizational AI Gains / Secret Cyborgs — why the value migration may be real but uncaptured.
- Token Shortage — the competing scarcity the thesis underweights. AI-Accelerated Science now makes it concrete: AI cost caps research acceleration near 16×.
- AI-Accelerated Science / AI Use Tracks AI Capability — empirical test (2026): in science, AI absorbs analysis and writing while ideation ("what to study") stays underused and ranked most important. That supports the thesis. But 72% of surveyed scientists expect AI to match top researchers at ideation by 2030, so it is an open question whether the scarcity is durable (see the contradiction on The Bottleneck Shift).
Sources
- The Human Use of Human Beings — Wiener (excerpt) — the information-theoretic pillar.
- The AI Daily Brief — Why Agents Still Need Humans — the economic pillar (Dan Shipper / Dan Shipper — After Automation).
- Making AI Work — Leadership, Lab, and Crowd — the organizational pillar (Ethan Mollick).
- Willingham — Critical Thinking, Why Is It So Hard to Teach — the cognitive pillar (added 2026-11-08).
- Anthropic — Investigating Vault Particles with Claude — a case that narrows the scarce contribution from "judgment" toward "verification" (added 2026-11-08).