With the rapid technological progress we're seeing today, the AI Alignment Problem is more pertinent than ever. Paired with the fact that neural networks are still black boxes, I believe this makes AI Safety (specifically, interpretability) one of the most important subfields of ML.

• Understanding how the concept of neutrality is represented in LLMs

  Independent research in mechanistic interpretability to understand how the ideas of "neutrality" and opinion are represented in the internal activation space of LLMs. Techniques include various types of steering and comparision with other types of "refusal", using Pytorch and TransformerLens.