We build research software tools which enable not just our researchers but also future researchers to investigate AI race dynamics. On the one hand, they allow us to computationally explore possible future scenarios and the effects of policy proposals; on the other hand, their implementation acts as quality assurance for our model formalizations. We also created a write-up elaborating our approach to implementing models and addressing issues of reliability, composability, and sustainability in computational science.
In 2016, Armstrong, Bostrom, and Shulman published the paper Racing to the Precipice which presents a static AI race model. We implemented the model and built an easy-to-use web app which allows anyone to run the model, enter custom parameters, choose between different agent behaviors, and visualize as well as download the results.
Building upon the static game presented in Racing to the Precipice, we formalized and implemented a dynamic AI race model which combines best practices from game theory, economics, technological races, and AI safety literatures. The model is intended to be more representative of real-world AI development dynamics and can be explored using our web app.