Bay Area, Calif. — AI Model Drives a Toyota Corolla to In-N-Out

Three engineers at the AI startup Axiom put a general-purpose language model behind the wheel of a 2024 Toyota Corolla and used it to drive to an In-N-Out Burger drive-thru near the Bay Area, according to Wired. The test, which involved OpenAI's GPT-6 Astra as well as separate attempts with other AI systems, is part of a broader push by researchers to find out whether text-generating models can also reason about the physical world.

How Did the Engineers Rig the Car?

Aditya Ramabadran, Simon Mahns, and Tobias Gessler sat in the Corolla with a laptop connected to a server, Wired reported. The server linked to windshield-mounted cameras and the car's power steering system, letting GPT-6 Astra interpret the road through live video and issue steering commands in a chat interface. A safety driver kept a foot hovering over the brake throughout the drive. The model, which Wired noted is normally used to generate text, code and images, had received no prior training for driving before the trip.

Which AI Models Were Tested, and What Happened?

The engineers tried putting GPT, Claude and Grok in charge of the vehicle; only one of the three completed the drive successfully, Wired reported. GPT-6 Astra guided the Corolla to the take-out window without incident. One of the engineers remarked afterward, "Maybe AGI is here after all," referring to artificial general intelligence, Wired reported.

Why Does Physical Reasoning Matter for AI Companies?

Andrew Dai, chief executive of a startup called Elorian AI and a former researcher at Google DeepMind, told Wired that physical reasoning remains an unsolved problem even as some companies claim general intelligence has arrived. Dai said improved visual and spatial reasoning would open up applications ranging from restaurant service to home robotics. "You can't really imagine home robotics without this," he said, according to Wired.

What New Benchmark Is Measuring This Skill?

Elorian and Scale AI, a company that supplies training data to major AI labs, built a benchmark called Humanity's Sixth Sense to test how well models understand physical scenes, Wired reported. Xingang Guo, a research scientist at Scale AI who worked on the benchmark, said most existing visual AI research focuses narrowly on perception rather than intuitive scene understanding. "We wanted to ask what it would actually take for a model to understand a scene intuitively, the way a person does," Guo told Wired.

What Comes Next for AI-Driven Cars?

Wired framed the In-N-Out run as one of several recent experiments suggesting language models are developing a rudimentary grasp of real-world physics, even as the publication cautioned that handing a general-purpose model control of a moving vehicle carries real risk.

What to Watch

• Whether Axiom or similar labs publish technical details on how GPT-6 Astra processed camera feeds in real time.
• Results from the Humanity's Sixth Sense benchmark as more labs submit models for testing.
• Any safety guidance from regulators on using general-purpose AI models, rather than purpose-built self-driving systems, to operate vehicles.
• Whether Claude's or Grok's developers respond to the reported failures during the same test.