Full automation of AI R&D probably yields a large speed up even without a software-only singularity
Alignment Forum · ryan_greenblatt · 2026-05-27
Ryan Greenblatt argues that full automation of AI R&D would produce a substantial one-time speedup in AI progress—on the order of 2.5–3.5 years of progress in a single year—even if a recursive software-only intelligence explosion does not occur.
Extraction
Topics: ai-takeoffai-rd-automationintelligence-explosioncompute-scalingai-forecasting
Claims
- Full automation of AI R&D produces a large one-time speedup estimated at roughly 2.5–3.5 years of progress in the first post-automation year, even without a software-only singularity.
- After AI systems automate AI R&D, additional compute yields higher returns than before because it simultaneously improves the AI labor force and can scale the number of AI researchers, creating a subcritical but meaningful feedback loop.
- Even with a subcritical recursive improvement coefficient (r=0.7 or lower), the automation-compute feedback loop may roughly double to quadruple AI progress rates compared to human-bottlenecked baselines.
- The rate of compute scaling may be substantially lower by the time full AI R&D automation is achieved, dampening the absolute magnitude of the speedup relative to current trends.
- Indirect effects—increased investment, compute consolidation from trailing companies, and AI-accelerated hardware R&D—may further boost progress around the time of full automation.
Key quotes
With my median parameters but r=0.7, the model...indicates you get 3.5 years of progress in the first year after full automation of AI R&D while assuming you aren't scaling up compute at all in this period. This is a huge amount of progress!
Even if this software-only feedback loop is subcritical...it still means every increase in compute will now drive more progress.
Even after the one-time speed up, increasing the available quantity of compute now has larger returns than it did when humans were the core source of AI R&D labor.