Bloom et al. (2020): in mature fields, productivity per researcher declines. Keeping up Moore's law today requires ~18× more researchers than in the 1970s.
Ideas get harder to find. The low-hanging fruit gets picked. Progress requires exponentially more researchers / capital / compute just to stay on trend.
If AI research itself becomes a saturating field, then AGI→ASI requires exponentially increasing R&D investment to maintain capability gains.
The paper makes a sharp counterargument: research-gets-harder is measured against a constant population of human researchers. AI flips this.
"Compute build-out, e.g., for a particular area of research, can be done very rapidly, flexibly, and on much shorter time-scales and much more cost-effective compared to training additional human researchers and developers, which takes years and is cost intensive."
The math: - Sustaining Moore's law today ≈ needing 18× more researchers vs 1970s - Increasing effective compute by 18× to run 18× more artificial researchers ≈ ~14 months at 10×/yr (slightly faster than that, technically, but the discount for hardware investment offsets) - And artificial researchers can be multiplied with no training time
So unless AI is stopped before it can become a useful research assistant, this friction inverts: AI demolishes it.
This is the most important sleight-of-hand to notice in the paper — it argues a bottleneck into a tailwind.
"Note that this argument only applies to the cognitive aspect of research — artificial researchers will still need to run experiments and collect data."
Experiments take real time. Physical experiments take real-world time (cell cultures, fab runs, particle colliders). This is the seam back to 13 - Bottleneck — Abstraction Barrier — the Embodied Bottleneck.
Chan et al. (2026) — "Measuring AI R&D Automation" — the proposed paper-tracking framework for quantifying how much AI is actually accelerating AI research. This will be the empirical signal that decides whether this friction reverses.