DrivingBench lets ChatGPT, Claude, and Grok drive a real Toyota Corolla; only one finishes the course
pwlot · x · 2026-09-23
The DrivingBench team wired ChatGPT, Claude, and Grok into a real Toyota Corolla, giving the models direct control of steering, gas, and brakes with no human driving (just a foot hovering over the brake as a safety net). The models drove a full driving course on a real road, and only one managed to complete the entire course. The team is introducing DrivingBench as a new benchmark for real-world driving capability.
More from Fun
- 100% code-made Van Gogh short built with Claude Opus 5.5 — sahilypatel · 2026-09-23
- On Terry Tao 'losing his job' takes: grief for a field, not self-interest — zetalyrae · 2026-09-23
- Utilitarianism debate: many objections are just puzzles people prefer to ignore — eigenrobot · 2026-09-23
- Meme: The Average Vibe Coder's Strategy for Building a $1B Startup — _jaydeepkarale · 2026-09-23
- Dev builds 3D Spider-Man game from scratch with Opus 5.5 in Three.js — TawohAwa · 2026-09-23
- User Jokes Meta's Muse Access Controls Make Anthropic's Risk Control Look Lenient — op7418 · 2026-09-23