Researchers Discover ChatGPT Can Drive a Car. Grok, on the Other Hand (www.thedrive.com)

Researchers have conducted an intriguing experiment using "frontier" AI models like Claude, GPT, and Grok to assess their driving capabilities via a newly developed test called DrivingBench. Utilizing Comma vision hardware and OpenPilot within a Toyota Corolla, the AI agents were tasked with navigating a short course marked by cones. The results were underwhelming, as eight out of eleven attempts across four different models failed to clear more than 11% of the course, often crashing during their first turns. The study revealed significant limitations in these AI models, highlighting their inability to effectively interpret situational commands and manage real-world driving dynamics. The significance of this research lies in its demonstration of the disparity between general-purpose AI capabilities and specific tasks such as driving, underscoring the challenges of applying advanced AI in complex, unstructured environments. Notably, GPT-6 Astra was the only model to successfully complete the course, albeit with a costly and unsteady route. This experiment serves as a cautionary tale about the overestimation of AI capabilities in critical applications, emphasizing that despite their sophisticated design, these models are far from ready to take the wheel safely.