OpenAI-hosted essay argues AI’s value may lie in carrying out ideas

OpenAI published an essay on October 1 arguing that AI’s effect on scientific and economic progress may depend on its ability to carry out ideas, as well as produce them. In ‘The eternal complement,’ Hemanth Asirvatham and Elliott Mokski describe two possible futures: one in which AI helps people achieve more with existing resources, and another in which it generates more ideas than laboratories and institutions can put into practice.

The authors say the essay, the first in a series on the next economy hosted by OpenAI, expresses their own views rather than those of OpenAI or their colleagues. Their account is an argument about what could happen as AI improves. It does not establish that AI has already raised economy-wide research productivity or determine which future is more likely.

Why the authors call execution a complement to intelligence

In economics, complementary inputs become more valuable when used together. The essay applies that idea to intelligence and execution: a useful hypothesis still needs evidence, instruments, funding and people able to act on it. Better ideas can make a telescope more valuable, while a better telescope can make the questions scientists ask more valuable. The authors call the coordination needed to turn ideas into results ‘institutional intelligence.’

That coordination includes laws, funding mechanisms, supply chains and other institutions. The essay’s point is that even exceptional reasoning cannot, by itself, build an instrument or complete an experiment. It presents routine organizational work as part of the production of knowledge, rather than merely an administrative cost around it.

What earlier research says about the cost of finding ideas

The essay cites a 2020 American Economic Review paper by Nicholas Bloom, Charles Jones, John Van Reenen and Michael Webb on research effort and productivity. The paper’s recorded abstract says sustaining Moore’s law required more than 18 times as many researchers as it did in the early 1970s. Its authors also report increasing research effort and falling research productivity across several industries, products and firms.

Asirvatham and Mokski add a broader comparison in their account of that research: they say economy-wide effective research effort rose 23-fold from the 1930s while measured research productivity fell by a factor of 41. Those figures are presented in the essay as historical context for its argument that progress can require more supporting work. The earlier paper does not test the essay’s predictions about AI.

How AI could change the balance

The essay says AI can already write code, search unfamiliar literature and turn a sketch into a working prototype. The authors argue that such assistance could let more individuals pursue projects that previously required an organization. They also propose that more capable AI might eventually originate many research agendas of its own, increasing the number of ideas seeking scarce experimental time and resources.

In what the authors call a ‘civilization of depth,’ better reasoning and simulations would help researchers use existing evidence more effectively and select only the most useful new experiments. In that scenario, intelligence would reduce the resources needed for each useful result. The essay offers this as a possible direction, not as an observed outcome.

Their contrasting ‘civilization of width’ would see new ideas and research opportunities multiply faster than physical experiments and institutions could handle them. Materials, energy, construction and coordination would remain constraints. The essay uses medicine as an example: it argues that computer simulations cannot replace large human trials needed to establish whether new medicines are safe and effective.

The authors explicitly say they do not know whether the resources and institutions needed to use new insights will grow, shrink or stay stable. Their depth and width scenarios frame the open question behind the essay: whether stronger AI will make practical work easier faster than it creates more work to be done. Neither the essay nor the cited historical research resolves that question.