Bridging the VLM and mech interp communities for multimodal interpretability — LessWrong