Part of a series “old econ papers”
Paul Romer’s Endogenous Technological Change: once somebody discovers a useful algorithm, thousands of people can use the knowledge at the same time. This nonrivalry is what lets Romer’s model produce sustained growth.
Ideas are expensive to create and cheap to copy
Romer combines:
- nonrival ideas, which create increasing returns for society;
- some excludability, such as patents or secrecy, so innovators can capture part of the return;
- deliberate R&D, so technological progress is an economic choice instead of magic arriving from outside the model.
The problem
If more researchers create proportionally more growth, why didn’t the huge increase in scientists during the 20th century generate a huge permanent acceleration? 2020 paper Are Ideas Getting Harder to Find? finds the same pattern in modern data: much more research effort is needed to maintain similar rates of progress in several domains.
What if the producers of ideas become nonrival too?
Human scientists are not nonrival — I can’t copy a good biologist 100,000 times and run all copies in parallel. If an AI system becomes a useful research worker, then we get:
That is a much stranger economy. Imagine a model that is good at a research loop, once trained, creating another instance may mostly require compute, so effective research labor could scale much faster than human population. This is one reason Trammell and Korinek’s work on transformative AI growth gets much more extreme growth possibilities than normal macro models.
But AI researchers are not actually free copies
Another model instance needs GPUs, energy, data etc. And in science, many useful observations are rival too — there may be one microscope, one robot lab, one patient cohort, one long-running experiment, one clinical trial that takes years. This creates a split:
I keep coming back to this in Two ICML 2026 AI For Science Papers: if AI generates one million hypotheses and the lab can validate 100, the bottleneck is experiment selection and verification, and that changes what another AI agent is worth.
The scarce thing may become research taste
Romer mostly treats R&D as an input that produces ideas probabilistically. With AI, I think we need to split R&D into more pieces:
AI may make the first two almost free, but that doesn’t make the whole pipeline free — if anything, it raises the value of deciding which branch deserves expensive evidence. This is close to Categories of AI Research Ideas: generating another plausible project is not the main constraint once plausible projects are abundant. Choosing a useful one is.
The Romer question for AI is empirical
I see three possibilities:
- AI mostly lowers the cost of routine R&D work. We get a productivity bump, then diminishing returns continue.
- AI creates massive effective research labor. Progress accelerates until compute, experiments or other inputs bind.
- AI improves the process that improves AI research itself. Then we can get a feedback loop much closer to recursive growth.
People often jump directly to 3, but current evidence is much stronger for 1 — we’re only beginning to see pieces of 2.