“Research on Claude 3 Sonnet, detailed in the "Scaling Monosemanticity" paper, provided significant evidence that large language models are substantially explained by linear features.”