Researcher Critiques 'Latent Space Archaeology': Compression Artifacts Are Not a Window into Digital Minds

Gerard Sans cautions that many interpretability studies amount to 'latent space archaeology,' arguing that compression artifacts are merely data fossils rather than windows into a model's digital mind. He emphasizes that pretraining and curriculum-based data curation are what truly matter for understanding how underlying distributions form.

2026-11-07 ~ 2026-11-07 · 2 related posts