A week before Tesla pulls the sheet off the production Cybercab in Austin, Waymo published ten things it says it learned from driving 200 million miles with nobody in the driver seat. The first reads less like a lesson than a demolition order aimed at Elon Musk.
“Cameras are incredible, but they aren’t enough,” Waymo wrote on August 26, in a post credited to Srikanth Thirumalai, the company’s vice president of onboard software. “For years, there’s been a debate over whether cameras alone could solve full autonomy. Now, after more than 200 million real-world miles, the data is clear: safe, fully autonomous operations at scale require more.”
The mileage figure holds up, and if anything understates the gap. Waymo’s post claims more than 200 million fully autonomous miles; in an August 4 talk at Y Combinator’s Startup School, co-CEO Dmitri Dolgov put it at roughly 220 million rider-only miles. Tesla has reported on the order of 380,000 driverless miles — nearly three orders of magnitude apart, accumulated across a Waymo fleet of more than 3,500 vehicles in 11 US metro areas.
Waymo’s argument is not that cameras are bad, but that they fail in exactly the conditions where a driverless car has no human fallback. It casts lidar as the wireframe, capturing 3D geometry at millimeter precision; cameras as the semantic overlay that reads signs and traffic-light colors; and radar as a sentinel that measures velocity and sees through rain, fog and dust. Dolgov has cited a Phoenix dust storm in which lidar picked out a pedestrian the cameras could barely register.
Thirumalai extended the case to Axios the same day, framing it as a problem with physical AI generally. “Even the best AI models with trillions of parameters still hallucinate,” he said. “There is no click reboot or reload or refresh [in] physical AI. You have to deal with the consequences of it.” On the end-to-end neural architectures Tesla, Wayve and Waabi have embraced, he was blunt: “What we found is that just pure end-to-end systems are not able to meet our safety bar at the scale that we operate.”
Dolgov’s version is the more quotable one. “Humans of course can drive with just eyes,” he conceded at Startup School. “If the goal were to just approximately match human performance or to build an assist product, that’s a very reasonable way to go.” The trouble starts when the target is superhuman reliability. “Weak sensing just leads to a safety curve that flattens out way too early,” he said, describing autonomy as an exponential ladder of nines where each nine costs ten times the effort of the last. Musk, who called lidar “a fool’s errand” in 2019, has argued the opposite for seven years: stacking sensors introduces conflicting data, and stripping radar and lidar is what makes a robotaxi cheap enough to saturate a city with.
The dispute is now leaking into statute. New Jersey’s S1677 would establish a three-year autonomous vehicle pilot requiring a camera system plus two distinct sensing modalities capable of tracking obstacles if the cameras fail. Operators would need 50,000 miles of supervised in-state testing before removing the safety driver, report every crash, and secure authorization before launching commercially. The bill also favors keeping a steering wheel and pedals, excluding the Cybercab on a second, independent ground. Sponsor Andrew Zwicker, a state senator and Princeton Plasma Physics Laboratory physicist who introduced it after riding in a Phoenix Waymo, told The Verge: “This is not anti-Tesla. I’m pro-New Jersey safety.” Tesla lobbied owners against the bill, generating some 4,000 protest emails in a single day. New York is weighing a near-identical mandate.
Analysis
The sensor fight is a proxy for the biggest open question in physical AI: whether the scaling laws that produced frontier language models transfer to systems that cannot retry. In text, a hallucination costs a bad paragraph and a regenerate button. In two tons of metal at 45 mph, the same failure mode is a collision. Waymo’s answer is architectural rather than statistical. It says it is consolidating toward fewer, larger foundation models, explicitly riding the same scaling curves as the LLM world — but wrapping them in an independent onboard validation layer that checks every proposed trajectory against physics and traffic law. Waymo calls that backstop non-negotiable for Level 4.
Tesla’s bet is that all of this is transitional scaffolding, and that a large enough model on a large enough fleet eventually learns the world well enough that the scaffolding becomes dead weight and cost. That bet is not obviously wrong; it is the same one that beat hand-engineered pipelines in vision, speech and translation. What makes driving different is the denominator. Waymo says its vehicles see roughly 16 times fewer serious-injury crashes than human drivers over comparable miles, and going from 99 percent reliability to fleet-wide reliability is not one more training run.
Neither side has published its most load-bearing number: Waymo has not disclosed a per-vehicle sensor cost, and Tesla has not published a camera-only safety case with miles and crash rates attached. Regulators are filling that vacuum with hardware rules — and if New Jersey’s approach spreads, redundancy stops being an engineering preference and becomes a market-access question.
What to watch
September 3 in Austin is the first checkpoint: whether the production Cybercab arrives purely camera-based, and whether Tesla attaches driverless mileage or crash data to it. The second is Nevada, where the Transportation Authority this month cleared up to 8,000 driverless vehicles in Clark County over 12 months. Tesla drew the largest allocation at roughly 5,000, though its Cybercab chief engineer told regulators the company expects about 2,500 within the year and that the number “has always been a ceiling for us.” Waymo was cleared for up to 1,000, Uber for 1,000, Zoox for 100. Las Vegas will be the first place where camera-only and multimodal robotaxis compete at scale under one regulator, on the same streets. What happens there will outweigh any blog post.
“Even the best AI models with trillions of parameters still hallucinate. There is no click reboot or reload or refresh in physical AI. You have to deal with the consequences of it.”— Srikanth Thirumalai, VP of Onboard Software, Waymo