A fine-tuned three-billion-parameter vision-language model learned to halt all traffic at simulated intersections when ...