Exclusive: Waymo says there's no AI shortcut to self-driving

Breakthroughs in AI are fueling hopes that more data and smarter models can provide a shortcut to self-driving vehicles.

Waymo, after more than 15 years and 200 million driverless miles, says there is none.

Why it matters: The answer could determine whether the AV race is a long, expensive slog that rewards Waymo's head start — or a new track that lets newer rivals hit the road faster.

Driving the news: Srikanth Thirumalai, Waymo's vice president of onboard software, said in an exclusive interview with Axios that it takes more than just better AI to safely deploy AVs at scale.

Waymo's goal from the start, he said, has been "demonstrably safe AI." But a growing reliance from competitors on single "end-to-end" AI systems introduces risks.

"Even the best AI models with trillions of parameters still hallucinate...

We don't have a choice to say, 'Let's hit refresh' ...

There is no click reboot or reload or refresh [in] physical AI.

You have to deal with the consequences of it." In a blog post published Wednesday, Thirumalai goes deeper on 10 AI lessons Waymo has gleaned from its first 200 million autonomous miles.

Catch up quick: Waymo got its start as Google's self-driving car project in 2009, well before the modern deep-learning revolution.

In the early days, it relied on many specialized models — one for pedestrian detection, another to tell when a light turns green, for example.

Gradually it shifted toward fewer, larger foundation models, "riding the AI wave," Thirumalai said.

Now companies like Tesla, Wayve and Waabi are going even further, developing AV 2.0 systems they say are capable of humanlike reasoning — "end-to-end" neural networks that process raw sensor data and directly output steering commands.

Waymo says it has tried the same tools but concluded that there aren't enough safety guardrails.