Interpretability for Control and Alignment

Planned topics include applications of SAEs and linear probes, J-space and the logit lens, activation steering, persona alignment, and monitoring safety-relevant representations.

Add lecture notes, readings, and transcripts here as the course progresses.

Further reading

• A Mathematical Framework for Transformer Circuits — mechanistic interpretability foundations.
• Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet — large-scale feature extraction and analysis.
• Sky sphere representation in language models — preprint by Yevgeny Liokumovich and A. Berdnikov, listed on the instructor's page.
• Golden Gate Claude — an Anthropic research demo that amplified a Golden Gate Bridge feature in Claude 3 Sonnet; the demo was available for 24 hours.
• Chapter 3.2: Monitoring and Chapter 3.4: Alignment — broader context on monitoring and controlling AI systems.