At an Anthropic Research Salon event in San Francisco, four of our researchers—Alex Tamkin, Jan Leike, Amanda Askell and Josh Batson—discussed alignment science, interpretability, and the future of AI research.
Further reading: Anthropic’s research: https://anthropic.com/research Claude’s character: https://www.anthropic.com/news/claude-character Evaluating feature steering: https://www.anthropic.com/research/evaluating-feature-steering
0:00 Introduction 0:30 An overview of alignment 4:48 Challenges of scaling 8:08 Role of interpretability 12:02 How models can help 14:31 Signs of whether alignment is easy or hard 18:28 Q&A — Multi-agent deliberation 20:38 Q&A — Model alignment epiphenomenon 23:43 Q&A — What solving alignment could look like
