These Researchers Made AI Drive a Toyota Corolla to Get In-N-Out
wired ·
Three AI engineers from Axiom recently demonstrated that large language models can control a physical vehicle by driving a Toyota Corolla to an In-N-Out Burger drive-thru. They connected OpenAI’s GPT-6 Astra, Anthropic’s Claude, and SpaceX’s Grok to the car’s steering system via windscreen-mounted cameras and a laptop interface. While all three models initially refused to issue motion commands, careful prompting allowed them to take control, with only GPT-6 successfully completing the short journey. This experiment suggests that general-purpose text-based models are developing rudimentary physical reasoning skills without specific prior training for driving. The researchers also created a benchmark called DrivingBench to test these capabilities further, revealing significant performance gaps between the models. Only GPT-6 managed to navigate a simple parking lot course, while Claude and Grok achieved partial success. The team believes these vehicular talents may be emergent properties of scaling multimodal training data rather than explicit driving instruction. This stunt highlights both the rapid progress in AI spatial understanding and the high-stakes risks of deploying such systems in the real world.
wired