Researchers Discover ChatGPT Can Drive a Car. Grok, on the Other Hand...

Researchers discovered that frontier AI models like Grok, Claude, and GPT are not well-suited for driving tasks, as they struggle with navigation, perception, and decision-making in real-time. DrivingBench, a test using Comma vision hardware and OpenPilot, showed that these models often fail to complete even a small portion of a course, crashing or misinterpreting obstacles. The test revealed that AI agents lack the contextual understanding and real-time processing needed for autonomous driving, even when trained on complex datasets.

• DrivingBench found that eight out of eleven runs failed to complete more than 11% of the course, with many crashing in the first turn.
• GPT-6 Astra was the only agent to complete the course fully on its second attempt, though its route was unconventional and risky.
• The test highlighted the limitations of frontier AI models in handling complex driving tasks, despite their advanced capabilities in other areas.
• The cost difference between Astra’s successful run and its previous attempt underscores the resource-intensive nature of training and executing such tasks.

AI-generated summary of the original article — see the source for the full story.