Google DeepMind — Description: Gemma Scope (2024): 400+ openly released sparse autoencoders for Gemma 2 9B and Gemma 2 2B models, intended to support external interpretability research

NoteThe source directly confirms all key elements of the claim: (1) Gemma Scope is a 2024 release (published July 31, 2024); (2) it contains 400+ sparse autoencoders ('more than 400 sparse autoencoders'); (3) they are openly released and freely available ('freely available, open sparse autoencoders'); (4) they are for Gemma 2 9B and Gemma 2 2B models; (5) they are intended to support interpretability research (stated as 'to help researchers understand the inner workings'); (6) Neel Nanda is part of the team (listed in acknowledgements). The source also confirms this was the largest open-source SAE release at the time (prior work 'mainly focused on investigating the inner workings of tiny models or a single layer in larger models'). The date 2024-08 in the claim aligns with the July 31, 2024 publication date.