GPT-6 Astra has gained the ability to drive a car(drivingbench.com)
307 points by plurby 20 days ago | 244 comments
tl;dr: Researchers gave frontier LLMs direct control of a Toyota Corolla's steering, throttle, and brakes, then benchmarked them on a fixed cone course with up to three attempts per chat. A public leaderboard ranks models by course progress, distance, finish time, command counts, and token cost, with trace videos available for each run.
HN Discussion:
  • ~Latency makes real-world deployment infeasible despite benchmark success
  • ↑Astra's vision/spatial capabilities are genuinely impressive across multiple benchmarks
  • ↑This signals the bitter lesson coming for self-driving, replacing complex stacks with end-to-end LLMs
  • ↓LLMs are the wrong tool for spatial/continuous control problems
  • ↓Questioning the value or methodology of this as a benchmark