I’m broadly interested in anything that helps improve or test the practical usefulness of Parameter Decomposition. This could include but is not limited to:

- extensions to tPD (targeted parameter decomposition) or other ways to make PD orders of magnitude cheaper
- Using PD for practical safety goals - auditing, unlearning, etc.
- Doing a bunch of obvious follow-up experiments to the case studies in “Interpreting Language Model Parameters”
- Variations to the core shape of the technique that use the same fundamental conceptual machinery in new ways, for instance Casual Importance functions trained on different objectives.

Please do not feel discouraged if you don’t fit this whole list. It’s neither entirely sufficient nor entirely necessary.

- Deep understanding of ML basics and strong linear algebra intuitions
- familiarity with Mech interp fundamentals
- Prompting and agent management skills
- Interest in parameter decomposition and engagement with fundamental questions of what interpretability should seek to achieve
- Deep Curiosity about minds, computation, and abstraction
- Strong programming skills