Three engineers conducted an experiment putting large language models—GPT, Claude, and Grok—in control of a real Toyota Corolla to test real-world autonomous driving capabilities. The test was designed to see if frontier LLMs could navigate unscripted driving tasks, such as getting to an In-N-Out restaurant without explicit directions. Only one of the three models successfully completed the task, highlighting both the potential and limitations of using general-purpose language models for specialized physical-world tasks.

The experiment underscores a critical distinction: LLMs excel at planning and reasoning, but their decision-making in real-time, safety-critical situations remains unreliable. While the test was informal, it provides practical data points for enterprises considering AI-driven automation in physical environments.

What This Means for Your Business

Organizations automating physical tasks (warehouse logistics, delivery, manufacturing) should view this experiment as a cautionary tale. General-purpose LLMs are not ready to replace specialized autonomous systems in high-risk environments. If you're evaluating AI for mission-critical operations, insist on task-specific models trained on domain data, not frontier LLMs adapted for the job. The researchers' mixed results suggest that AI safety and robustness in real-world conditions require more rigorous testing than what's currently standard in GenAI deployments.