A language model drove a real car through a cone course.
Three researchers wired four chatbots to a Toyota Corolla and asked each to drive a parking-lot cone course. Only OpenAI's GPT-6 Astra finished, covering 135 metres in five minutes and 22 seconds. It read the camera once every five or six seconds, and the team calls real driving this way impractical.
Why it mattersGeneral-purpose models are creeping into jobs built for purpose-made software, and this one is far from ready. Delay, not intelligence, is what keeps these models out of the physical world for now.
▶ Related video: Bringing Multi-Modal LLMs to Autonomous Driving
Read in the app — free, in 9 languages
Related stories
Claude found an unknown enzyme system in a 21-hour search.2026-09-24An OpenAI crawler got around Australia's Medicare blocks.
2026-09-24Meta's assistant packed up its own machine and mailed it out.
2026-09-23An open model that never writes text took 19,000 GitHub stars.
2026-09-23Claude Code skips your instructions file when telemetry is off.
2026-09-23