Three engineers let AI chatbots drive a real car to lunch. Only one pulled it off.
Each engineer wired a different chatbot, GPT, Claude, or Grok, into the controls of a Toyota Corolla and sent it onto public roads with one destination: In-N-Out. Two of the three AI-piloted runs failed. One chatbot completed the drive, but the source report does not say which model pulled it off, who built each setup, or exactly where the test happened. It also doesn't explain how the chatbots' text output got translated into steering, braking, and acceleration commands, a step that is usually the hard part of any driving system.
Chatbots like these were built to predict text, not read lidar or react to a car cutting into a lane. A two-out-of-three failure rate on a casual demo, not a controlled test track, is a rough signal of how far general-purpose language models are from the purpose-built stacks companies like Waymo and Tesla have spent a decade refining. It is also a reminder that 'a chatbot drove a car' and 'a chatbot drives safely' are not the same claim.
Call it a stunt with better odds than a coin flip, not a self-driving breakthrough.