This is a linkpost for https://forum.effectivealtruism.org/posts/NdcXvDkvAw5bLW4rp/interpretable-analysis-of-features-found-in-open-source?utm_campaign=post_share&utm_source=link
This was an up-skilling project I worked on throughout the past months. Even though I don't think it is anything fancy or highly relevant to the research around SAEs, I find it valuable since I learned a lot and refined my understanding of how mechinterp fits in the holistic, bigger picture of AI Safety.
In the mid-term future I hope to engage in more challenging and impactful projects.
P.D.: brutally honest feedback is completely welcome :p