OpenAI published "The Eternal Complement" on 1 October 2026, an essay in which Hemanth Asirvatham and Elliott Mokski argue the bottleneck on progress is not the idea but its execution. The more hypotheses a model generates, the costlier verifying them becomes.
Key takeaways
- Published 1 October 2026 on OpenAI's Intelligence Age platform
- The claim: machine intelligence and "institutional intelligence"?Institutional intelligence: An organisation's capacity to execute — coordinating people, hardware and procedures at a scale no single person can hold. — the capacity to execute — are complementary goods
- Sustaining Moore's Law?Moore's Law: A 1965 observation that transistor counts double at regular intervals. Not a law of physics but a trend sustained by spending. takes over 18 times as many researchers as in the early 1970s
- Cited in the essay: research productivity down 41% against a 23-fold rise in spending
- Two paths: a "deep civilisation" built on reasoning, a "wide" one built on experiment
Complements, not substitutes
Asirvatham and Mokski compress it into one image: machine genius designs monuments, institutions lay the stones. The James Webb Space Telescope illustrates it — roughly $10 billion, 300 institutions across 14 countries, build tolerances in tens of nanometres.
The data behind the claim
The strongest point comes from Nicholas Bloom, Charles Jones, John Van Reenen and Michael Webb in the American Economic Review in 2020: holding transistor density on its doubling path now takes 18 times the researchers of the early 1970s.
| Metric | Change |
|---|---|
| Researchers needed to sustain Moore's Law | over 18× more than in the early 1970s |
| Research productivity | −41% against a 23-fold rise in spending |
| Technician headcount vs scientists | growing twice as fast |
| Cost of a modern chip fab | five times what it was thirty years ago |
| James Webb Space Telescope | ≈ $10bn, 300 institutions, 14 countries |
Deep civilisation or wide
Then the fork. A "deep civilisation" maximises reasoning and minimises experiment, letting knowledge advance by thought as gaps in nature's description reveal themselves. A "wide civilisation" assumes the opposite and commits to experiment at scale.
Drug development is the example, where more AI-generated candidates mean a larger verification bottleneck. Asirvatham and Mokski back the second scenario, and the conclusion follows: execution overtakes ideation as the scarcest resource.
Why it matters
The position is awkward for both sides of the AI argument. It denies enthusiasts the promise that machine intelligence alone suffices, and sceptics the claim that models do not change the economy. If complementarity is real, a rising supply of ideas raises the value of labs, production lines and technicians rather than lowering it. That reframes the labour question from substitution to throughput, and suggests the valuation multiple sits outside the models themselves.
What's next?
- The essay announces no product — it is OpenAI's position on where the bottleneck sits, not a roadmap
- The claim is measurable: if research productivity?Research productivity: New ideas obtained per unit of research input — per researcher or per dollar. It falls when the same progress needs ever larger teams. stops falling as spending rises, the execution-as-bottleneck argument weakens
Sources
- TMTPost — OpenAI 新文章把超级智能拉回执行层，十亿个爱因斯坦也要有人去采矿
- American Economic Review — Are Ideas Getting Harder to Find?
