Mechanistic Interpretability

Compiling a Calculator Into AI Weights: A New Path to Decode Transformers

A dev compiled an RPN interpreter into Transformer weights. The 1.1GB basic-math model's value: offering a new way to bypass training and decode AI in