SPARSE AUTOENCODERS
Sparse Autoencoders
OpenAI Unveils Breakthrough in GPT-4 Interpretability with Sparse Autoencoders
OpenAI's new methods identify 16 million patterns in GPT-4's computations.