Map of the Shallow Review of Technical AI Safety

Someone recently pointed out that I underestimated the amount of interpretability researchers when I thought about the most important and neglected research directions, so I made a semantic map and clustering of recent AI safety papers to get a better understanding of what AI safety researchers are working on.

This map was created by taking the title and abstract of each paper, embedding with BGE-small-v1.5, reducing to 5 dimensions with UMAP, clustering with HBDSCAN, and then reducing to 2 dimensions. The data is from the Shallow Reviews of Technical AI Safety (2025, 2024, 2023). The idea is from this old LessWrong post.