Tag
Anthropic tested Claude models on various robotics tasks, finding that performance heavily depends on the control interface—models succeed when supervising pretrained policies but fail when directly controlling joints—indicating growing but uneven transfer of language model capabilities to physical domains.