These Researchers Made AI Drive a Toyota Corolla to Get In-N-Out (www.wired.com)

Researchers Aditya Ramabadran, Simon Mahns, and Tobias Gessler have successfully navigated a 2024 Toyota Corolla through a drive-thru using OpenAI's GPT-6 Astra, a model primarily designed for generating text. By connecting a laptop interface to the vehicle’s systems and cameras, they demonstrated that AI language models can perform real-world tasks, albeit with a safety driver present. This experiment, aiming to test the physical reasoning capabilities of AI, raises intriguing questions about the potential for general-purpose models to engage with the physical world, suggesting that advancements in multimodal training could lead to unexpected emergent abilities. While self-driving technology typically relies on specific algorithms trained for the task, this stunt highlights a shift toward integrating language models' spatial reasoning in practical scenarios. Their findings could influence the development of future AI applications in robotics and interactive systems. However, a new benchmark introduced by the trio, DrivingBench, indicates these models still have significant limitations in actual driving, with only GPT-6 Astra managing a slow, partial course completion. As AI's understanding of the physical world improves, it could open new avenues and challenges, underscoring a critical frontier in AI and machine learning development.