AI News Feed
Market watch
Large Language Models

DrivingBench Links General-Purpose Chatbots to a Toyota Corolla in Parking-Lot Test

Tech workers in San Francisco calling themselves DrivingBench connected GPT-6 Astra, Claude Fable 5.1, Grok 4.6 and GPT-5.6 Sol to a Toyota Corolla and let them steer, accelerate and brake on a cone course. Only GPT-6 Astra completed it, according to a Slashdot report citing 404 Media.

DrivingBench posted its code, prompts and videos of the experiment online, according to the report. The models tested were GPT-6 Astra, Claude Fable 5.1, Grok 4.6 and GPT-5.6 Sol. In the parking-lot course, Grok, Sol and Fable drove only a few meters and did not finish, the report said. GPT-6 Astra eventually learned to navigate the course and completed it, but not without what the group described as a lot of troubleshooting.

The experiment drew a contrast with purpose-built autonomous-driving systems such as those from Waymo and Tesla, which rely on millions of hours of training data. The chatbot driving the Corolla was a general-purpose model running on a laptop, rather than a system designed specifically for driving, according to the Slashdot report. That distinction was central to the stunt: the organizers said they wanted to find out whether off-the-shelf frontier AIs could control a real car.

Mahns, who was quoted in the report, said the outcome was an interesting demonstration of potential emergent capabilities. “It can fail a turn, and then the next turn, next try, it will be able to do that turn and other turns that it hasn't seen before,” Mahns said, according to the report. “Not necessarily saying LLMs are gonna put Waymo out of business or something, just an interesting demonstration of where things are, and things are just moving fast.”

Mahns also said most people familiar with ChatGPT have used it in white-collar, computer-based settings, and that the capability is nearing the point where it will start having more impact in the physical world. “So I think seeing the sparks of this happening, the shift of this utility coming into the physical world, is also somewhat exciting,” Mahns said. DrivingBench's code, prompts and videos were posted online, the report said. The experiment was conducted in a parking lot, not on public roads, and involved a cone course rather than open traffic.

Editor's Summary

DrivingBench, a group of San Francisco tech workers, connected four general-purpose chatbots to a Toyota Corolla and tested them on a parking-lot cone course, with GPT-6 Astra the only model to complete it. The experiment highlights both the limits and emergent potential of off-the-shelf large language models in physical-world control tasks, while remaining far from the trained autonomous-driving systems used by Waymo and Tesla.