Introduction

This is a process in which a machine learning model operates on an autonomous loop, rewriting itself in a way that improves itself, and this new improvement will allow for more efficient improvements, resulting in a kind of recursive growth.

RSI exists on a spectrum; different levels of RSI demands varying levels of human intervention. Complete RSI is a goal we have yet to achieve, where an AI completely builds the next improved version of itself without any human involvement.

Self-Improvement vs Recursive Self-Improvement

A typical modern AI model already self-improves by learning from large datasets. Its self-improvement is enabled by methods such as gradient-descent and backpropagation.

Recursive self-improvement on the other hand, is the idea of a model improving the rate at which it improves. This creates a compounding effect, that leads to exponential growth, and a "foom" that propels us to superintelligence.

Forms of RSI

There are a few ways in which this can be achieved:

• Using an LLM to generate code for an improved version of the LLM
  • This is an existing form of RSI, because many AI engineers already do make use of AI to code. Hence, the AI generates code that builds the next improved version of itself.
• Getting AI to improve its own model architecture
• Kaparthy Loop