Aditya Ramabadran, Simon Mahns and Tobias Gessler, three AI engineers from a startup called Axiom, decided to put their latest AI model to the test. The trio asked OpenAI’s GPT-6 Astra to take the wheel of a 2024 Toyota Corolla, guiding the car to In-N-Out Burger for a casual lunch.

The experiment, while simple, suggests that language-based AI models are starting to gain a rudimentary understanding of the physical world. The model, normally tasked with generating text and code, slowly navigated the vehicle to the take-out window, much to the engineers’ amazement. They remarked, 'Maybe AGI is here after all,' referring to artificial general intelligence.

Despite the lack of overtly alarming incidents, the engineers admit that putting a general-purpose model in charge of a fast-moving car is a high-risk operation. Their experiment, done outside of their work at Axiom, aimed to measure how well models can operate in the real world, especially in the messy and unpredictable environment outside.

The engineers noted that big AI companies are focusing on improving the spatial reasoning capabilities of their models, but they doubt this means the companies are training them specifically for driving. The latest models may have developed the ability to drive as an emergent capability during training focused on 3D reasoning. However, these models still have a long way to go before they could pass a driving test, as demonstrated by a benchmark called DrivingBench.

While the experiment is a fun thought experiment, it also raises questions about the future of AI in our daily lives. Will our cars soon choose our meals, or is this just another snack for thought?