The Sequence Knowledge - Issue 937: RSI in Post-Training: The Loop That Already Shipped
Inside the RSI post-training pipeline
TL;DR
- The most economically significant AI self-improvement mechanism is the post-training pipeline.
- This pipeline involves a loop where a model generates answers, these answers are graded, and the good ones are used as training data.
- The model then trains on this data, improving its ability to produce better answers.
- This process, known by terms like STaR or RLVR, is responsible for the evolution of frontier models into agents.
- This loop is already being operated at an industrial scale by major AI labs.