Tesla, and Elon Musk in particular, has spent years mocking Waymo despite being quite far behind in the autonomous taxi space. Now Waymo is finally clapping back.
Co-CEO Dmitri Dolgov recently laid out why cameras alone will never make a vehicle fully autonomous. He didn’t mention Tesla by name. He didn’t have to. Camera-only is Tesla’s entire bet, and it’s the target of numerous lawsuits.
The remarks came during a talk at Y Combinator’s Startup School 2026, where Dolgov walked through seven lessons from building Waymo over nearly two decades. One of them? You need more than cameras.
“There’s been a long-standing debate about what kind of sensors do you actually need for autonomous driving,” Dolgov started.
Cameras get you to human. That’s the problem
Are cameras enough? Sure, if your goal is matching human performance. Humans drive with two eyeballs, after all. That’s the entire argument for camera-only, and Dolgov concedes it.
But matching humans isn’t the target. Beating them badly is. And that’s where Dolgov says camera-only runs out of road: weak sensing produces a safety curve that “flattens out way too early.”
Waymo uses three sensor types. Cameras for high resolution and color. Lidar to map the 3D structure of the world. Radar for fog, rain, snow and everything else that ruins visibility.
The distinction that matters: cameras are passive, meaning they need light to work. Lidar and radar are active, generating their own signal, so they see just fine in pitch darkness or straight into a blinding sunset.
Dolgov brought receipts. A pedestrian standing at the roadside during a Phoenix dust storm, nearly invisible to cameras but obvious in lidar. Kids chasing a dog across a street with no streetlights. Pedestrians about to climb over a concrete construction barrier.
Each sensor runs its own encoder, and the outputs fuse into a single picture that Dolgov calls “vastly superior to what you get with any one sensor.” Redundancy, in his framing, is a requirement rather than a luxury.
The demo took 18 months. The product took 15 years
That line is my favorite thing Dolgov said, and it explains the whole disagreement.
Waymo had a working self-driving demo in a year and a half. Shipping something the public could actually ride in took another decade and a half. The gap between “it works in the video” and “it works ten million times without killing anyone” is where autonomy companies go to die.
Dolgov calls it an exponential ladder of nines. Musk calls it the march of the nines. Getting to 90% reliable is easy. Getting to 99% is harder. Every additional nine after that costs exponentially more, and you need a lot of nines before you can take the human out of the car.
Right now that ladder has Waymo at 500,000 paid driverless trips a week, over 4 million autonomous miles every seven days, and 17 times fewer serious-injury crashes than human drivers.
So why is Tesla refusing to use lidar?
Tesla pulled radar from its vehicles in 2021, then dropped ultrasonic sensors in 2022. Musk has spent years trashing lidar, memorably calling it a fool’s errand and declaring that anyone relying on it is doomed. What’s left is “Tesla Vision,” which is cameras and nothing else.
I’m going to go out on a limb and assume Musk is insistent on this for two reasons:
- It’s more aesthetic
- It’s cheaper to make
I could be wrong, so take that with a grain of salt. But Dolgov pre-emptively demolished the cost argument anyway.
Waymo is on its sixth generation of hardware, and he says each generation has cut the price dramatically. Betting against lidar based on what it costs today, in his words, means betting on “a number that has a fairly short shelf life.”
Which leaves aesthetics. Cool.
Meanwhile, the robotaxis keep hitting things
Here’s the uncomfortable part. Tesla’s robotaxi service in Austin has logged 14 crashes since launching in June 2025, across roughly 800,000 miles, per NHTSA filings analyzed by Electrek. That’s one crash every 57,000 miles.
Tesla’s own Vehicle Safety Report claims the average American driver has a minor collision every 229,000 miles. So by Tesla’s own math, its robotaxis crash about four times more often than we do. Measured against NHTSA’s police-reported crash data, it’s closer to nine times.
And every one of those vehicles had a safety monitor sitting in the front seat whose entire job was preventing exactly this.
The incident list is almost comedic. A collision with a bus while the Tesla was stationary. Backing into a pole at 1 mph. Hitting a truck at 4 mph. There’s also a July 2025 crash originally filed as property damage only, quietly amended in December to include an injury requiring hospitalization. Five months later, no announcement.
The lawsuits are stacking up too
There is still no public data supporting Musk’s repeated claim that Full Self-Driving is safer than human drivers.
California, meanwhile, has been busy. In December 2025, an administrative law judge ruled Tesla engaged in deceptive marketing over Autopilot and FSD, and the DMV adopted that decision with a 30-day suspension of Tesla’s license to sell and build cars in the state.
Tesla complied rather than lose its Fremont factory. It killed Autopilot as a standalone product in January, slapped “(Supervised)” onto Full Self-Driving, and shifted FSD to a $99 monthly subscription, quietly retiring the $8,000 one-time purchase.
Then, having complied, Tesla sued the DMV in February to get the “false advertiser” finding erased, calling it factually wrong and unconstitutional. Because of course it did.
Add the $243 million Autopilot verdict that survived appeal, at least four more crash lawsuits settled quietly, and an NHTSA investigation into 2.88 million vehicles over FSD running red lights, and you get a picture that doesn’t look much like winning.
So here’s what I keep coming back to. One company is doing half a million driverless rides a week with three kinds of sensors. The other is crashing four times more often than humans while a paid babysitter watches from the passenger seat.
At what point does “cameras are enough” stop being an engineering position and start being a sunk cost?
