[Paper Review] Economic impacts of AI-augmented R&D

Tamay Besiroglu, Nicholas Emery-Xu+1 more|arXiv (Cornell University)|Dec 15, 2022

Innovation Policy and R&D4 citations

TL;DR

This paper demonstrates that AI-augmented R&D, particularly in deep learning-based computer vision, is significantly more capital-intensive than traditional R&D, leading to accelerated idea production. Using empirical data and a semi-endogenous growth model, the authors show that widespread adoption of AI in R&D could nearly double U.S. productivity growth rates by boosting computational capital utilization.

ABSTRACT

Since its emergence around 2010, deep learning has rapidly become the most important technique in Artificial Intelligence (AI), producing an array of scientific firsts in areas as diverse as protein folding, drug discovery, integrated chip design, and weather prediction. As more scientists and engineers adopt deep learning, it is important to consider what effect widespread deployment would have on scientific progress and, ultimately, economic growth. We assess this impact by estimating the idea production function for AI in two computer vision tasks that are considered key test-beds for deep learning and show that AI idea production is notably more capital-intensive than traditional R&D. Because increasing the capital-intensity of R&D accelerates the investments that make scientists and engineers more productive, our work suggests that AI-augmented R&D has the potential to speed up technological change and economic growth.

Motivation & Objective

• To assess the economic impact of AI-augmented R&D on idea production and long-term growth.
• To investigate whether deep learning adoption increases the capital intensity of R&D, thereby accelerating innovation.
• To estimate the elasticity of idea production with respect to computational capital in AI-driven R&D.
• To evaluate whether AI-driven R&D can serve as a General Purpose Technology (GPT) by transforming the innovation process.
• To provide empirical evidence linking AI adoption in R&D to measurable increases in productivity growth.

Proposed method

• The authors estimate an idea production function using microdata from deep learning research in computer vision.
• They measure computational inputs via multiple approximation methods, including hardware specifications and training run metrics, to reduce measurement error.
• A semi-endogenous growth model is used to analyze how increased capital intensity in R&D affects long-term productivity growth.
• The study compares the capital intensity of deep learning R&D to traditional R&D sectors across the U.S. economy.
• Selection bias from reporting practices is mitigated by using standardized conference checklists and multiple estimation techniques.
• The analysis accounts for hyperparameter tuning runs, which may contribute to total computation but are not fully captured in final training estimates.

Experimental results

Research questions

• RQ1 How does the capital intensity of AI-augmented R&D compare to traditional R&D in the U.S.?
• RQ2 To what extent does increased computational capital in deep learning R&D enhance idea production?
• RQ3 Can the widespread adoption of deep learning in R&D lead to permanent acceleration in productivity growth?
• RQ4 Does AI-driven R&D qualify as a General Purpose Technology based on its impact on innovation processes?
• RQ5 What is the elasticity of idea production with respect to computational capital in deep learning-based R&D?

Key findings

• Deep learning-based computer vision research is significantly more capital-intensive than virtually all other R&D sectors in the U.S.
• The capital intensity of AI-augmented R&D implies that it could nearly double the U.S. productivity growth rate if widely adopted.
• The study finds that AI-driven R&D increases the elasticity of idea production with respect to computational capital, suggesting a strong growth effect.
• Measurement errors from approximating total computation may attenuate estimates, implying the true impact could be even greater than estimated.
• Selection bias from selective reporting is likely small due to standardized reporting norms in top-tier machine learning conferences.
• The results support the hypothesis that AI-augmented R&D can act as a General Purpose Technology by enabling sustained, capital-driven acceleration in innovation.