GEMMA SCOPE

Summary from the TLDR AI newsletter

A few weeks ago, DeepMind released a number of sparse autoencoders on the Gemma 2 suite of models. This is now the companion paper where researchers discuss the training paradigm and some interesting results.