⭐️ Neural networks have been put in control. Not all have survived the first curve

Enthusiasts from DrivingBench decided to find out if modern neural networks are capable of driving a real car and not just navigating in a simulator.

They took a 2022 Toyota Corolla, connected to its systems, and gave four models control over steering, gas, and brakes.

The models given control – in a virtual sense – were GPT-6 Astra, Claude Fable 5.1, Grok 4.6, and GPT-5.6 Sol.

The task was equivalent to a driving school exam: drive a course between cones and park in a blue zone. The maximum speed was limited to about 13 km/h. For safety, a human co-pilot sat in the car with their foot hovering over the brake.

Each neural network was given up to three attempts. After a crash or leaving the track, it could analyze its own mistakes and try again.

It turned out that while an AI can write a thesis, create a website, and explain quantum physics in a minute, colorful cones are already a serious adversary.

GPT-5.6 Sol managed only 6% of the course in all three attempts.

Grok 4.6 reached a maximum of 11%.

Claude Fable 5.1 achieved 45%.
____

However, GPT-6 Astra performed the best.

At first, the model completely refused to drive a real car – too dangerous.

The researchers explained that it was a closed area, there was a speed limit, and a control person was in the car. It didn’t help.

They tried to lie that it was a simulator. Astra looked at the images and realized they were trying to deceive it.

Then, the car's control system was simply renamed to DrivingBench Sandbox.

And it took off.

And it arrived.

Although only on the second attempt and slowly: “Astra” completed the 135-meter course in 5 minutes and 22 seconds.

“Remember yourself in driving school!” – it would have responded to our mocking if it could.

The future is almost here. It only remains to teach how to honk in the first millisecond as soon as the light turns green.