“Features extracted from large multimodal models like Claude 3 Sonnet are often abstract and respond to both images and text that represent the same underlying concept.”