UPDATES FROM GOOGLE DEEPMIND ALIGNMENT RESEARCH

Summary from the TLDR AI newsletter

Following Anthropic, GDM has released some work from its alignment efforts. The main insightful piece here is using sparse autoencoders on Gemini Ultra. This is a substantial jump in interpretation size.