Waymo told Axios in an exclusive published Wednesday that there is no artificial-intelligence shortcut to full self-driving, pushing back on rivals and AI commentators who argue that scaling end-to-end neural networks alone can solve autonomy. [1] Waymo co-CEO Dmitri Dolgov framed the company's position bluntly: "At Waymo, 200 million-plus fully autonomous miles have taught us that there are no shortcuts to building and scaling physical AI safely." [1]
The comments land as a direct answer to a live argument inside AI circles this year — that autonomous driving is fundamentally a data-and-scale problem the same way large language models are, and that a big enough end-to-end model trained on enough driving footage will eventually match or exceed layered, simulation-heavy engineering stacks like Waymo's. [1] Waymo's public position is that this framing understates what safety at scale actually requires: years of simulation, redundant sensor systems, and an engineering stack built specifically for physical-world failure modes that a chat-style model was never designed to handle. [1]
The rebuttal comes from a position of commercial strength Waymo did not have even a year ago. The company has been racing to make its expansion self-sustaining, rolling out a Chinese-made robotaxi to riders in Los Angeles, Phoenix and San Francisco earlier this year as it works to cut hardware costs while scaling ride volume. [2] That commercial momentum is precisely what gives Dolgov's comments teeth: Waymo is not arguing from the sidelines against a hypothetical AI shortcut, but from a company running hundreds of millions of real autonomous miles while competitors chase a more software-only approach.
On X, Dolgov's quote has circulated widely since the Axios story published, split between engineers agreeing that physical AI cannot skip the hard simulation and validation work, and skeptics arguing Waymo has financial incentive to dismiss cheaper end-to-end approaches that could undercut its lidar-heavy hardware stack. One reply thread noted that lidar's advantage — seeing through weather and darkness that defeats camera-only systems — is exactly the kind of physical constraint an end-to-end model trained purely on video cannot shortcut around, reinforcing Dolgov's point from a different angle. Other accounts read the timing as defensive, coming as vision-only autonomy bets from Tesla and camera-first startups draw fresh venture funding.
Axios did not name a specific competitor Dolgov was responding to, framing the comments as addressed to the broader AI-autonomy discourse rather than any single rival. [1] That omission is itself part of the story: Waymo's rebuttal works as a general statement of engineering philosophy precisely because it applies to any company betting that a large enough model, rather than a purpose-built safety stack, can close the gap to full autonomy. Whether 200 million miles of layered engineering beats a bet on scale alone remains an empirical question neither side can yet settle — Waymo's mileage keeps accumulating, and so does the compute behind the end-to-end approaches it is arguing against.
-- THEO KAPLAN, San Francisco