Waymo's Injury Rate Is Five Times Lower. It Still Loses to a Lyft.

Nathan Labenz invited Timothy B. Lee, author of the *Understanding AI* newsletter, onto The Cognitive Revolution for a two-hour tour of the self-driving industry. Lee has tracked the sector since 2016, through hardware stacks, business models, and the regulatory fights that actually moved the story. Labenz's personal stake is simple: driving is the most dangerous thing he regularly does, and he wants to know when he can sleep in a car on the way to his grandmother's house.

Lee does not find the official SAE levels very helpful. Level 1 is ordinary cruise control; Level 2 is driver assistance; Level 4 is a fully driverless car in a limited area; Level 5 is the mythical vehicle that can go anywhere. Level 3, the shared-responsibility middle, is the one he likes least. The idea is that the car mostly drives while you sit there, then hands control back when it gets confused — a user-experience nightmare and a safety nightmare if you have turned around to buckle a child into a car seat.

His simpler map has two kinds of vehicles. Driver assistance, including Tesla's Full Self-Driving, can do a lot of the work, but the human is still legally responsible: hands on the wheel, eyes on the road. Fully driverless vehicles, like Waymo's, have nobody behind the wheel, and the company is on the hook. So far that second category always comes with a geofence. Lee thinks Level 5 is not a useful concept. There will always be a service area, just as a cell phone is still a real phone even if it has no signal in rural Alaska.

The limit is not connectivity. It is maps, taxi licenses, and the unglamorous fact that if a car with no driver gets stuck, someone has to come service it. Any company offering a service has places it operates and places it does not. A footprint that is literally everywhere, Lee says, is quite far away.

Labenz's last Tesla FSD trip was about a year earlier: an eight-hour round trip taking his grandmother home, the first time either of them had used it. The system kept a short leash. A camera in the rearview mirror watched for alertness, and the steering wheel demanded a bit of torque. Wander, and the car would pull over and shut the feature down. He came away extremely impressed, especially at highway speeds.

The mistakes clustered in ambiguous scenes: a stop sign that could be read as facing either of two ways, and a stop on a highway off-ramp where the car definitely should not have stopped. He also inherited his neighbor's default of 20% over the speed limit and later realized he had been doing 84 in a 70 at night in the rain. Lee's own recent test was 45 minutes in San Francisco in March, on version 12.5 — after Tesla's big shift toward neural networks, before the newest release. He took over twice: once when the car was about to hit plastic bollards in a bike lane, and once when it entered the wrong lane and tried to cut into full traffic while cars honked behind him.

Lee keeps coming back to liability rather than SAE labels. Even if Tesla loosens the driver-monitoring leash, a crash still gets blamed on the person who should have been paying attention. Waymo is the other product. Labenz used an Uber-like app in San Francisco, could not cross the bridge to Oakland Airport, and found the hardest first-timer step was unlocking the door in the app. Then there was nobody in the driver's seat, just a sign not to mess with the controls. Within minutes he was looking at his phone at red lights and had to remind himself to watch.

Over about an hour he called the ride more comfortable than an average Uber, if not quite as aggressive as the best human drivers hunting a light. Lee has taken nine Waymo rides totaling about three hours this year and never saw even a minor mistake. The first couple of trips he filmed and stared; by the fifth or sixth he was reading a book. Crowdsourced Tesla FSD still needs a human takeover about every 300 miles. Waymo, back when it was Google's self-driving project, was already reporting a takeover every 5,000 miles in 2016.

Lee has been covering the industry since 2016, when Waymo was still Google's self-driving project filing annual disengagement reports with California regulators. In February 2017 it reported that safety drivers in 2016 had taken over once every 5,000 miles. There was probably some cherry-picking, Lee notes, but the company sincerely thought it was ready. In 2018 it ordered 60,000 vehicles from Chrysler and Jaguar and expected a nationwide taxi service. The 2018 launch kept safety drivers for two more years. True driverless service in 2020 stayed tiny; serious scale arrived only in the last year or two.

The remaining work is edge cases, especially emergency scenes. On a normal road there are lanes, lights, and rules. At a crash or a fire, a police officer starts directing traffic in a completely different pattern, and every nearby car behaves differently. The system has to reason counterfactually: this vehicle is trying to get out, so I should stop; that hand signal means halt. Those scenes may appear only once every 10,000 to 100,000 miles. With a safety driver they are a shrug. Without one, the car has to handle them.

A driverless vehicle that works 99.9% of the time, Lee says, is still getting stuck constantly, and sometimes hitting people. That is not good enough. Waymo's last five years have been the climb from roughly 99.5% to something that looks like a viable service. Tesla, in his view, is at the start of that same slog. Labenz's neighbor, who uses FSD constantly, said the car will not take unconventional moves — it will not creep forward to give space to a vehicle about to rear-end it. Lee assumes a Tesla that reaches a crash scene will get stuck or ping the driver to take over.

Rarity is the data problem. You need enormous mileage before a learned system sees enough real wrecks, and a test-track officer cannot fake dozens of cars in random configurations. Tesla owners make it worse: their instinct, approaching a chaotic scene, is to switch FSD off before the software can try and fail. Human driving after takeover is still recorded, but it is less useful than a professional safety driver who lets the stack err. The question is often not whether you hit someone in that second, but whether a choice jams traffic three seconds later.

Lee walks through the standard architecture as three layers. Perception turns sensor data into a world model: objects, locations, velocities, types. Prediction asks what the world will look like one, two, or three seconds ahead — essential at an intersection when an oncoming car is not yet in front of you. Planning then searches for a path that reaches the goal without collisions. Naive short-horizon imitation of human steering, he argues, will not work, because the human is already running a predictor.

Five to ten years ago, Waymo and Cruise stacks were thick with hand-coded rules. More of that has been replaced by neural networks, but Lee still thinks the count is in the single digits: a few perception nets, a few planning nets, papers about pieces rather than one blob from pixels to pedals. He is not sure it would even make sense to get there. Driving needs a much higher reliability bar than a language model, and engineers want to inspect the implicit world model: is the car seeing what is actually there, and are its predictions reasonable? Layers can be trained separately against ground truth.

Tesla talks about an end-to-end network. Lee is skeptical that it is literally one net from sensors to steering, though he cannot see the internals. Both companies already show a labeled world model on the dash, which is itself startling to watch. Labenz floated a "guaranteed safe AI" frame — world model, explicit safety spec, verifier — close to Mobileye's Responsible Sensitive Safety standard. Lee likes the instinct and assumes Waymo has implicit rules like do not exceed this acceleration. The hard part is that the car must also arrive on time and not infuriate everyone. Pure conservatism is safe on city streets and still causes jams, blocks fire engines, and fails as a product.

The sensor split is now folklore. Tesla committed to cameras only after dropping radar a year or two ago, with something like eight or twelve cameras. Waymo runs all three modalities: a long-range lidar on the roof, four to six shorter-range wide-angle lidars around the body to catch a child or a pet at the bumper, plus cameras and several radars — 20 to 30 sensors in all. Radar can measure velocity directly, which helps on a freeway. Lidar fires a laser and times the return, so distance is a measurement rather than a guess, and the output is a 3D point cloud. A camera gives a 2D image; 3D structure is inferred.

Lee does not buy the online consensus that cameras have already won. Lidar costs money to buy and maintain, and if cameras were clearly good enough they would be preferable. Nobody has yet shown a system that is *clearly* very safe, so companies are picking points on a safety-versus-cost curve. Waymo chose higher cost and, it hopes, higher safety. The claim that cameras uniquely unlock deep learning is, in his view, a misunderstanding: a network can ingest lidar too. Vision can estimate a point cloud via parallax, and in the average case it is almost as good. The failure mode is the problem. When the algorithm misreads an unfamiliar object, the identity error and the distance error arrive together. Lidar, however clumsy, still says there is something this big this many meters ahead.

If he were CEO of a self-driving company, Lee says he would be "really nervous" about cameras: they might work 99% of the time, and the remaining 1% writes a $10 million loss. It is an empirical question. If Tesla keeps improving rapidly for two or three years, cameras might suffice; they have not made that case yet. The deeper constraint is the business model. In 2016, when Tesla started selling "Full Self-Driving ready" cars, lidar cost on the order of $50,000, and only 10 to 30% of buyers paid for the software. You cannot put a $10,000 sensor in every retail car. Waymo is a vertically integrated taxi with a few hundred vehicles, so expensive lidar is affordable, and the bet is that volume will crash the price.

That also reframes Tesla's data advantage. The fleet is huge, and cars can even search locally for rare images, such as a bicycle strapped to a trunk. Lee thinks Waymo has less data and higher quality: precise lidar plus professional safety drivers who label *why* they took over — a car mistake versus a bathroom break — which is a much cleaner learning signal. Other automakers are planning to add lidar as it gets cheaper. Lee expects Tesla may eventually follow, though Elon Musk is "pretty dug in." Aftermarket kits like Comma.ai, a windshield camera plus a plug into the car's network, impress him as a hobbyist Level 2 project, not as the long-run form factor.

Tesla partisans talk as if Waymo is chained to expensive HD maps while Tesla drives on cameras alone. Lee thinks the gap is smaller than the rhetoric. Classically Waymo kept a fine-grained 3D map of hydrants, trees, and the rest; Tesla did not. The cost critique, he argues, will prove overblown because collection automates. The first pass through a new city still needs a mapping run. After that the fleet is on those streets anyway. When Lee visited Waymo in March, he saw a hard drive in the trunk uploading overnight to headquarters. Cleaning and labeling should get more automated too.

Tesla already collects fleet data to build 3D models for simulation and testing, and its visualization shows lanes that are not even in the current camera view. Both stacks follow a map to an address. Lee expects the approaches to converge. He does not think the cars must stay online to navigate. Waymo is not operating nationwide, and a local copy of San Francisco would fit; longer term he pictures a rolling radius of perhaps 100 miles, downloaded at night. Safety-critical inference, he is pretty sure, stays on the vehicle.

What *does* go over the network is help, not a remote chauffeur. Waymo has said nobody sits with a steering wheel and pedals. When a car is confused, it asks for confirmation: I am planning to do X, is that safe? Lose connectivity for long enough and it may slow to a stop out of caution, which is how some Cruise cars ended up with flashers on, blocking the road. Labenz asked whether live players call out to vision-language models for long-horizon puzzles like finding a drive-through. Lee has not seen that as a safety-critical path. The now-defunct startup Ghost experimented in that direction; today's leaders keep the dangerous loops local.

Cruise was chasing the same robotaxi model as Waymo and growing at a similar pace, but industry insiders spent 2023 slightly nervous. The technology looked behind, which made sense for a younger company, and GM's pockets are deep without being Google-deep. Ford had already shut Argo, its own Cruise analogue. Lee thinks CEO Kyle Vogt felt pressure to prove a commercial business and pushed a little too hard. It was not "one safety incident and a house of cards." The October 2023 crash was still ugly: a human-driven car hit a woman, her body bounced in front of the Cruise, the Cruise did not stop in time, and after halting it pulled over and dragged her several more feet.

What infuriated regulators, in Lee's account, was the aftermath. Officials say Cruise showed the impact clip and not the dragging; Cruise denies it. Earlier, a stuck Cruise was accused of blocking an ambulance. Lee was shown an internal "god's-eye" video on the condition that he take no screenshot and receive no diagram. He describes a company in "defense attorney mode" rather than public-understanding mode. California asked them to halt after the October crash, following an August fire-truck incident that had already cut the fleet in half. Cruise then voluntarily paused Austin and Phoenix. It is now testing with safety drivers in Houston, Dallas, and Phoenix, three or four years behind Waymo, which went driverless in 2020. Lee still sees it as a plausible number two, just much further back.

The rest of the field is earlier. In the West, taking the safety driver out remains the confidence milestone, and only Waymo has it. Amazon-owned Zoox and Hyundai-backed Motional are still testing, not running commercial driverless service. Britain's Wayve raised about a billion dollars on an end-to-end foundation-model bet, with extra text and images so you can talk to the car, hoping to license to Toyota-class OEMs. Aurora pivoted from that licensing story into long-haul trucking and says it will run 18-wheelers between Houston and Dallas by year-end — a case where Lee would be nervous as CEO, because the failure modes are spectacular. Nuro, started by Google self-driving alumni, is building washing-machine-size delivery robots at 20 or 30 mph; the bottleneck is less onboard safety than finding neighborhoods that mix a grocery store with residential streets and no high-speed roads. China, including Baidu, may already offer Waymo-like empty-car rides at meaningful scale, but Lee cannot verify safety records across a language barrier and a constrained press.

Start with the human baseline, because every company number is meaningless without it. Human-driven cars get into a fatal crash about once every 100 million miles. Waymo has driven about 20 million driverless miles with zero fatalities. That sounds pristine and is not yet evidence of superiority: at that mileage you would expect only a fraction of one death anyway. Injury crashes happen on the order of once every few hundred thousand to a couple of million miles, and police-reportable crashes — serious enough for a report, not always an injury — sit in a similar band.

Waymo's December report covered 7 million miles with two or three injury crashes, or about one every 2.4 million miles. Against human drivers on comparable roads in the same metro area, that is roughly once every 350,000 miles — about five times safer on injuries. Police-reported crashes look about two and a half times safer: half a million miles per Waymo crash versus 200,000 for humans. Lee calls that decent evidence, and impressive. He still will not close the book on deaths. Severe crashes are so rare that a random bug every 50 million miles could kill someone without showing up in the fender-bender stats. His guess is that Waymo is safer, probably by a significant margin. He wants 200 to 400 million miles before the fatality claim is statistically solid.

A more persuasive version of the same result came earlier, at 4 million miles, in a study with Swiss Re. The reinsurer holds a multi-year database of U.S. auto insurance claims, so it could match Waymo's actual blocks rather than a national average. It found a similar two-to-three-times edge on injuries and reportable crashes. Lee likes that a firm whose job is to price risk put its name on the paper. Tesla's safety site says Autopilot crashes once every 7 to 7.5 million miles, far better than humans. Lee's objection is the missing mix of miles. Freeways have fewer crashes and worse ones; city streets are messier. Drivers may switch Autopilot on in the easiest environments. Tesla has not handed the data to independent experts the way Waymo has.

Crowdsourced FSD logs currently show a disengagement about every 300 miles. That would wow anyone on a 100-mile drive, and it is still nowhere near human-level — or near Waymo's 5,000-mile figure from eight years ago. Lee keeps meeting people who took a 20-mile test drive, saw no mistakes, and declared the problem solved. Average humans also get 99% of those 20-mile trips right. You need thousands of miles. Even the 300-mile number may be flattered by people who only enable FSD on roads they already trust.

Labenz's prior was that society would demand an order-of-magnitude safety edge before allowing AVs, and that regulators were the handcuffs. Lee's picture is almost the opposite. NHTSA has started requiring crash reports nationwide, which is better than California-only data, but the country still lacks a denominator: miles driven, and miles with FSD on. California recently approved Waymo all the way down the San Francisco Peninsula into Silicon Valley. Months later the company still has not turned it on. Lee's read is that Waymo is not confident on freeways, not that a permit is missing.

U.S. vehicle design still runs on self-certification. NHTSA publishes the Federal Motor Vehicle Safety Standards; an automaker tests against the list and files a report. Waymo buys an already-certified vehicle and adds sensors and computers in the trunk, so it inherits the paperwork. States regulate drivers, licenses, and traffic law. California is among the strictest, with reporting and permits, and those permits have not been that hard to get. Texas is much more open. Cruise is testing there in part because California felt tight, but Lee thinks lawsuit risk still weighs more than a regulator kicking in the door. Cruise's shutdown mixed pressure and choice: California asked it to stop, and the company voluntarily paused other cities.

If the gold-standard numbers are roughly true, why is there no national celebration? Lee does not think demand is the bottleneck. The technology is not quite finished, and a vertically integrated fleet — depots to charge, clean, and repair — expands slowly. Waymo was Phoenix-only two years ago, added San Francisco about 18 months ago, and is now in Los Angeles and Austin. It grew about three to five times in the last year. Keep that up and nationwide coverage is a five-to-ten-year story, similar to Uber and Lyft's crawl out of San Francisco. A Tesla software update that made every car driverless could theoretically move faster; Lee does not think the main obstacle would be legal. He does not think Tesla is technically ready.

People hear "self-driving car" and picture a vehicle they buy, press a button, and ride anywhere. That is not the product, so they file Waymo under research prototype. The sharper product problem is freeways. Waymo's urban safety strategy is: if you are unsure, stop. On city streets that is usually safe, and the remaining job is to stop rarely enough that you do not infuriate traffic. At 70 miles an hour it does not work. You cannot stop in time, you risk a rear-end, and you create a jam. The fallback has to be a minimum-risk condition: the shoulder or the next exit, which may mean driving another half mile with a problem onboard.

Waymo started freeway testing in Phoenix earlier this year, not yet commercially. Until that lands, Lee does not find the service very compelling as everyday transport. Whenever he booked a Waymo in San Francisco he also opened Lyft for the same trip. Waymo was always a little more expensive and always five to ten minutes slower. If Lyft would have taken a freeway, the gap got large. A Phoenix video compared the longest trip inside Waymo's service area: about 20 minutes in a Tesla on the freeway versus 40 minutes in a Waymo on surface streets. Some riders will pay for novelty, or for not sharing a cabin with a stranger. Lee does not think the business is profitable until freeways work. "I don't think they're going to be profitable until they figure freeways out."

Asked what a wartime industrial policy for robotaxis would look like, Lee shrugs off most of it. Preempting local licensing fights might help at the margin; California already overrode cities that wanted their own approval. Rebuilding streets, adding self-driving lanes, or sprinkling infrastructure sensors would not solve the fat-tail events — the crash sites, the wet cement a Cruise once drove into. What *has* helped in San Francisco is operational protocol with first responders: 911 and the fire department can geofence a fire zone that Waymos automatically avoid, and a firefighter can jump in and drive a stuck car. A federal template for those local interfaces would be useful. A public database of stop signs and rules with GPS coordinates would be nice, maybe a couple of percent faster. It is not the main thing holding the industry back.

Lee used to assume self-driving would kill private car ownership. He has cooled on that. People like car seats, golf clubs in the trunk, and the knowledge that a vehicle is in the driveway. Taxis should get cheaper, so the mix should shift toward fewer owned cars and more rides, not toward zero ownership. His market picture is sequential. Waymo expands a taxi footprint nationwide. Once that service area covers perhaps 50 to 70% of the country, it can go to an automaker and sell a built-in subscription — on the order of $100 a month — that makes the family car drive itself inside the map. He thinks that is about a ten-year story. Wayve might get there by licensing instead.

He is skeptical of Tesla's robotaxi pitch. Elon's version is that you buy the car, leave on vacation, and let it earn money. Someone still has to fix a flat tire. Tesla is not in that operations business and, Lee thinks, has not started building the network. It could take years, and he is not sure the culture fits. Google is an odd parent for dirty work too: it is used to high-margin software, top-dollar talent, and not sweating unit cost. He can imagine the stack working perfectly while the service still costs 50% more than Uber would. Lyft already keeps about half of each fare to feed a large engineering org. Possible answers include franchising, an Avis-style depot partner (Waymo tried something like that in Phoenix around 2018), licensing the software, or simply dispatching most Waymos through Uber and Lyft, who already spent a decade and billions building two-sided ride-hail. Amazon, via Zoox, already runs on the order of a million vehicles. Lee would not count it out as a fast follower.

Tesla's $15,000 FSD price, in this telling, was never the main bottleneck. It was a cash raise from a tighter period three to five years ago, and the company is moving to a monthly subscription around $100, which Lee thinks is the right shape for a service. The wild-West funding wave was 2016 to 2018; the sector is now in consolidation, and second-tier companies will not raise again until someone shows a working market. He sees the same pattern, compressed, in large language models: many firms can do 90% of a demo, and the last 10% is where self-driving taught him to be afraid. The difference is that language models do not kill people, so they can be useful while imperfect. Driving has to be close to perfect or it is not a product. Asked whether Labenz's Detroit or Lee's Washington gets robotaxis first, Lee answers immediately: DC. Fewer freeways, testing already underway, and Waymo has gone California–Arizona–Texas, south to north. Washington's few snow days still count as southern enough.