RSI Simulator

Paradigm has released a web game and model explorer that allows users to experience the economics of Recursive Self-Improvement (RSI).

Players run an AI lab, investing labor, compute, and data into R&D to experience the process of reaching self-sustaining research acceleration. The game is based on the economic model from the paper The Economics of Recursive Self-Improvement, but parameters have been adjusted for educational purposes rather than prediction.

The explorer visualizes the model and key parameters from the paper, allowing users to examine a dynamic model that incorporates the following elements:

• Compute-optimal training
• R&D-based growth model
• Scale-dependent algorithmic progress
• Weak links in automation

Key insights presented during the development process are as follows:

• Since AI research requires human researchers, compute, and data in a complementary manner, the scarcest resource can become the bottleneck.
• Recursive self-improvement may not be a continuous explosion but can repeat phases of acceleration and stagnation.
• A narrow intelligence explosion specialized in AI research or optimization is more likely to appear before general intelligence.
• Results depend heavily on the elasticity parameter, which indicates how sensitive the rate of discovery is to current model capabilities.