The Economics of Recursive Self-Improvement (RSI)

This paper presents an economic model of Recursive Self-Improvement (RSI), which describes how much faster the performance of the next generation of AI models increases as AI model capabilities improve. The researchers analyze that the key to AI progress lies in the strength of Feedback Loops.

The main points of analysis are as follows:

• Modeling Feedback Loops: The study mathematically shows that the net acceleration of AI capability is determined by the product of Elasticities spanning each feedback loop.
• Narrow vs Broad AI: The model distinguishes between whether AI improvement remains 'Narrow', merely raising benchmark scores, or becomes 'Broad', actually enhancing task capabilities with real economic value.
• Diagnosis of the Current State: After calibrating the model based on existing data, the researchers confirmed that current feedback loops are not yet strong enough to trigger self-sustaining acceleration, but that their strength is gradually increasing.

This research proposes a list of empirical objects that AI companies should disclose in order to measure the strength of feedback loops, and provides a framework for assessing the economic and social impact that accelerating AI progress could bring.