In an empty parking lot, researchers tested a 2022 Toyota Corolla using a 130-meter cone course as part of a broader effort to integrate general-purpose AI models—like large language models (LLMs) and multimodal models—into achieving level 5 autonomous driving. Level 5 autonomy refers to fully self-driving vehicles that require no human intervention in any situation. The car was equipped with a third-party hardware box from comma.ai, which connected directly to the vehicle’s CAN bus, allowing external software to control the steering, acceleration, and braking. To ensure safety, the vehicle was limited to a speed of 3.5 meters per second (around 12 km/h), and a human driver was present to manually apply the emergency brake if needed. The AI model used in the experiment, called GPT-6 Astra, did not have direct, continuous control over the vehicle's physical movements but instead communicated with the car through an API using text and visual data. During the first attempt, Astra failed at the 67-meter mark by veering off course and hitting a white line due to an unexpected turn. After analyzing its mistake, the AI adapted its strategy for a second run, making smaller, slower movements and steering more precisely to navigate tight turns. This time, it successfully completed the course in 5 minutes and 22 seconds, achieving a 100% success rate. According to one report, Astra was the only model tested that completed the course, highlighting the progress being made in using AI for autonomous driving. This experiment is part of a larger trend in autonomous vehicle research. For example, Waymo, a leader in self-driving technology, recently introduced EMMA, an end-to-end multimodal model designed for autonomous driving. Unlike Waymo’s commercial vehicles, which use LiDAR and radar sensors, EMMA relies solely on video streams from onboard cameras. It was trained using Google’s computational resources, including TPUs, and required only standard navigation data rather than ultra-detailed maps. Despite being a research model, EMMA performed well on benchmark tests, showing the ability to avoid obstacles it had not been explicitly trained on. Experts suggest that general-purpose AI models, like GPT-6 Astra and EMMA, have strong capabilities in understanding visual and textual information but still struggle with precise spatial awareness and dynamic physics—skills essential for real-time driving decisions. While these models are seen as a crucial step toward achieving full autonomy, they must be combined with secure, modular systems to ensure safety. Some industry leaders, like NVIDIA’s Jensen Huang, have compared the shift toward using AI in autonomous vehicles to the transformative impact of ChatGPT on digital communication. However, many car manufacturers are focusing on level 4 autonomy—where the vehicle can handle most driving tasks but still requires a human to take control in rare situations—due to the complexity and cost of achieving full level 5 autonomy.