Видео недоступно для предпросмотра
Смотреть в Telegram
🤖 GPT-6 Astra Can Drive a Car
The creators of DrivingBench connected several frontier language models to a real 2022 Toyota Corolla and gave them control over the steering wheel, accelerator, and brakes. GPT-6 Astra performed the best: on its second attempt, it completed the entire 134.7-meter cone course in 5 minutes and 22 seconds.
👨💻 The model received images from the car's cameras along with telemetry data, then issued commands telling the vehicle how far and in which direction to turn the steering wheel, how fast and for how many seconds to drive, or when to stop. On its first attempt, Astra made it through only 49% of the course, but it retained the context from that run and corrected its mistakes on the second attempt, reaching the finish.
Claude Fable 5.1 completed no more than 45% of the route. Grok 4.6 and GPT-5.6 Sol failed to make it past even the first turn. The test took place in an empty parking lot, the car moved at roughly a slow walking pace, and a human remained behind the wheel at all times, ready to brake.
💡 The DrivingBench team wanted to test whether general-purpose LLMs could apply their planning and tool-use abilities in the physical world. And it seems language models may indeed have a future as general-purpose AI systems. OpenAI's Head of Applied Research, Boris Power, even called the results "a nail in the coffin" for specialized models trained from scratch for a single narrow task.
Would you trust ChatGPT with your car?
❤️ — Yes, I already trust it with my secrets
🔥 — No, absolutely not!
@hiaimediaen
The creators of DrivingBench connected several frontier language models to a real 2022 Toyota Corolla and gave them control over the steering wheel, accelerator, and brakes. GPT-6 Astra performed the best: on its second attempt, it completed the entire 134.7-meter cone course in 5 minutes and 22 seconds.
👨💻 The model received images from the car's cameras along with telemetry data, then issued commands telling the vehicle how far and in which direction to turn the steering wheel, how fast and for how many seconds to drive, or when to stop. On its first attempt, Astra made it through only 49% of the course, but it retained the context from that run and corrected its mistakes on the second attempt, reaching the finish.
Claude Fable 5.1 completed no more than 45% of the route. Grok 4.6 and GPT-5.6 Sol failed to make it past even the first turn. The test took place in an empty parking lot, the car moved at roughly a slow walking pace, and a human remained behind the wheel at all times, ready to brake.
💡 The DrivingBench team wanted to test whether general-purpose LLMs could apply their planning and tool-use abilities in the physical world. And it seems language models may indeed have a future as general-purpose AI systems. OpenAI's Head of Applied Research, Boris Power, even called the results "a nail in the coffin" for specialized models trained from scratch for a single narrow task.
Would you trust ChatGPT with your car?
❤️ — Yes, I already trust it with my secrets
🔥 — No, absolutely not!
@hiaimediaen