The Bounded Parallelizability of R&D: Theory and Application to AI
- Working Paper
Technological development is typically modeled as Cobb-Douglas in (a) the quality of existing technology and (b) the quantity of research inputs, such as researchers or lab equipment. In fact, technology and research inputs are often gross complements: even infinite quantities of old equipment cannot drive progress as quickly as finite quantities of new tools which allow more of the research process to occur simultaneously. I model this and consider its implications for the automation of research in AI.
If relieving parallelization constraints is no harder than growing the stock of effective research inputs, then automating AI R&D will cause AI development to accelerate almost as quickly as in the Cobb-Douglas case. Otherwise, parallelization constraints will eventually bind, and automating AI R&D may have little impact on the speed of AI development in the long run.