The story’s score and claims

Comments on No. 63 of 137 · ethahtz.github.io · edition of Friday 13 November 2026

Are we interpreting the model of interest?

Summary

A blog post by an interpretability researcher reports that enabling output_attentions=True silently switched a vision-language model from SDPA to eager attention, cutting benchmark accuracy from about 70% to 20%.

Comments

Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules