A new essay on OpenAI’s website makes an unglamorous bet about superintelligence: the machines may spend nearly all their time on logistics, coordination and construction, and very little on breakthroughs. The piece, titled “The eternal complement,” is the first in a series on what its authors call the next economy. An authors’ note says it reflects their views, not those of OpenAI or their colleagues, and the page we read names no individual author or publication date.
The argument starts from a mismatch. Humans have mapped the early universe, yet no person has traveled past the Moon. The authors offer two explanations. Either deep truths can be reached by thinking from one planet, or our ideas have outrun our ability to carry them out.
They lean on the James Webb Space Telescope as evidence for the second reading. Galileo needed two lenses and a tube. Webb cost ten billion dollars and carries eighteen mirror segments built to a fifty nanometer tolerance. Fourteen countries contributed to the build, with three hundred separate organizations involved. In the authors’ telling, seeing farther took a bigger bureaucracy, not just brighter minds.
The essay cites economists Nick Bloom and coauthors for the productivity picture. Compared with the start of the 1970s, keeping Moore’s law on pace now takes a research workforce more than eighteen times larger. Across the whole economy, effective research effort has grown twenty-three-fold since the 1930s. Over that same span, productivity per unit of research fell forty-one-fold. A modern chip fab costs five times what one did three decades ago.
From there the authors sketch two possible futures. In a “civilization of depth,” a superintelligence reasons so well that it needs only a handful of carefully chosen experiments, runs most of its tests in simulation, and keeps a small physical footprint. In a “civilization of width,” every answer raises more questions that only real-world testing can settle, so the need for labs, factories and energy outgrows any gain in cleverness.
Drug development is their example of width. Computer simulations of biology cannot yet stand in for trials on large numbers of people, so more candidate drugs from a machine means a heavier testing queue. A superintelligence could plan a hundred years of experiments before dinner, the authors write, and then wait the full hundred years for nature and machinery to respond.
If width wins, the essay’s conclusion is blunt. The authors say the real edge of a superintelligence would be its readiness to serve as the bureaucracy itself, and that “almost all machine intelligence would be deployed not to do the brilliant, but to do the boring.” Their example is a Dyson sphere, a hypothetical star-harvesting structure, which they describe as mostly a matter of building and shipping things across space, with very little breakthrough thinking involved.
The essay does not leave humans without a role. It offers three reasons: people have an obligation to keep seeking understanding, they may hold a comparative advantage in creative work over routine coordination, and human variety has value of its own. In the near term, the authors expect AI to make execution cheaper, shifting the scarce skill toward taste, meaning the judgment about what is worth building.
Treat the piece as framing, not evidence. It contains no experiments on AI systems, and its own closing line concedes that the outcome depends on whether brilliant minds require ever more physical resources or can get by on fewer. The statistics it cites describe human research, not machines.
The framing still has a market angle. Labs usually sell models on peak intelligence, with benchmark wins and discovery stories. This essay, from inside one of them, suggests the durable value may sit in reliable agents that handle coordination at scale, which is a different product from a better genius. Builders choosing where to invest should weigh orchestration, tooling and real-world integration at least as heavily as the next jump in raw model quality.
Reported from the essay “The eternal complement,” published on OpenAI’s website; the page carries no publication date.
