Language Models Take the Wheel: AI Drives Real Car to Fast-Food Chain

Real-World AI Navigation Experiment

Three engineers at AI startup Axiom conducted an unconventional test of artificial intelligence by using a general-purpose language model to drive a 2024 Toyota Corolla to a Bay Area In-N-Out restaurant. The vehicle was equipped with windscreen-mounted cameras and connected to the car's power steering system via a chat interface running on a laptop.

The team tasked OpenAI's GPT-6 Astra with navigating the vehicle autonomously to the drive-thru window. While a safety driver remained ready to intervene, the AI model successfully guided the car to collect their lunch order. This experiment demonstrates how modern language models are beginning to exhibit rudimentary but functional understanding of physical environments.

Implications for AI Development

Unlike traditional autonomous driving systems built specifically for vehicle control, this demonstration relied on a model originally designed for text generation. The success suggests that large language models may be developing emergent capabilities in spatial reasoning and real-world interaction.

However, the engineers acknowledge the risks involved in deploying general-purpose AI in high-stakes physical scenarios. As these models continue advancing, their potential integration into robotics, home automation, and other real-world applications becomes increasingly plausible, though safety considerations remain paramount.

Why it matters for GPU / AI infrastructure: This experiment underscores the computational demands of running multimodal AI models that process both visual and textual inputs in real-time. Companies operating AI workloads must ensure sufficient GPU capacity and low-latency inference capabilities to support such physically interactive applications.