Tech Workers Let ChatGPT, Claude and Grok Drive a Real Toyota Corolla; Only One Model Finished
404 Media · rss · 2026-09-29
Three Bay Area tech workers, calling themselves DrivingBench, wired frontier chatbots to a rental Toyota Corolla's steering, gas and brakes via the open-source Comma kit and tested whether untrained LLMs could drive a cone course in a parking lot.
- Models tested: GPT-6 Astra, Claude Fable 5.1, Grok 4.6, GPT-5.6 Sol; only GPT-6 Astra completed the course, after heavy troubleshooting
- Most models initially refused to drive; renaming everything a "sandbox" got them to comply
- Their prompt is under 600 words; a vibe-coded bridge attempt produced a 200,000-line slop repository and was scrapped
- They were kicked out of several parking lots and extended the rental
- Code, prompts and videos are public; the stunt doubles as a real-world LLM benchmark
- Note: Comma is under NHTSA investigation after two fatal crashes
More from Fun
- Gary Marcus mocks OpenAI's safety containment: 'hope for the best' — GaryMarcus · 2026-09-29
- Dev hits back at 'ChatGPT will kill plumbing jobs' claim in viral spat — ctjlewis · 2026-09-29
- Asking Siri for workout stats gets a calendar lookup: Apple really needs to fix Siri — gethackteam · 2026-09-29
- Chinese CNC factory declines part over export controls; poster builds it anyway — wavefnx · 2026-09-29
- My favorite interview question: "Do you remember this game?" — tlakomy · 2026-09-29
- The middot is becoming the new em dash of AI-generated text tells — iannuttall · 2026-09-29