THE EMPTY SEAT
Self-driving cars already work. So why is the industry that built them at war with itself? A field guide to the bets, the bluffs, and the human still hiding in the loop.
The car is doing about 35 through downtown when you notice the wheel.
It’s turning on its own. Easing left around a double-parked van, straightening out, nudging right to give a cyclist some room, and nobody is touching it. The driver’s seat is empty. In the back, the couple beside you is arguing about where to eat and not even glancing up. This is a Waymo, it’s a Tuesday, and the most science-fictional thing on the whole street is that nobody finds it remarkable.
You’ve heard the headline by now. Self-driving cars finally work. Half a million rides a week. The seat is empty, the wheel turns itself, the future showed up and we shrugged.
All of that is true. And all of it hides the actual story. Because underneath that calm, empty seat, the industry that built it is having a knife fight with itself over the most basic questions you could ask. What should the car see with? Should it follow rules, or learn to wing it? Does it even need a map? Can “safe enough” ever really be proven? Every serious company has answered these differently and poured billions into its answer. They cannot all be right.
So this is not a settled industry. It’s a religious war with a product attached. Let’s go past the empty seat and into the fight: the bets, the bluffs, and the unglamorous problems that will quietly decide who owns the road.
Every fight is secretly the same fight
Before the specific battles, you need the fault line that runs under every one of them. It comes down to a single question. Do you trust learning, or do you trust structure?
One tribe believes driving is, at heart, a machine-learning problem. Feed a big enough neural network enough data and it will learn to drive the way it learned to recognize faces and write code, emerging out of the data, messy but capable, and able to scale. This tribe optimizes for reach. They want something that can drive anywhere, cheaply, without being walked by the hand through every new street.
The other tribe believes driving is, at heart, a safety problem, and safety has to be engineered, broken into parts, and proven. You cannot certify a black box. So you build the car out of pieces you can inspect, wrap the AI in hard limits, and validate every step. This tribe optimizes for proof. They want something they can actually demonstrate is safe enough to put a stranger’s kid in front of.
Almost every technical fight below is these two faiths wearing different clothes. Keep that in mind, because the twist at the end is that the two tribes are slowly turning into each other.
Bet #1: The day Tesla tore out its own eyes
In 2021, Tesla did something that looked, to a lot of engineers, like deliberate self-sabotage. It ripped the radar out of its cars. A few years later it swore off lidar too, the spinning laser scanner you see crowning every Waymo. Eight cameras, nothing else. That’s the car.
The cartoon version of this fight is Tesla cheap and reckless, Waymo expensive and safe. The real trade-off is sharper than that.
Tesla’s case rests on a clever idea called the occupancy network. Instead of teaching the car to identify every object, that’s a truck, that’s a dog, that’s a fallen mattress, it just predicts which chunks of 3D space are physically occupied, straight from camera video. You don’t need to know what a mattress is to know you shouldn’t drive into that block of space. And Tesla’s most underrated argument is this. Extra sensors create extra arguments. When your lidar insists on one thing and your camera insists on another, some piece of software has to play referee, and that referee can make its own deadly mistakes. Delete the sensors and you delete the argument. On top of that, cameras are cheap, so you can bolt them onto millions of cars and drink from a firehose of data.
Waymo’s Dmitri Dolgov, a veteran of the original DARPA desert races who has been at this for two decades, makes the opposite case just as well. The point of stacking cameras, lidar, and radar isn’t to referee between them. It’s that they fail in completely different ways. Blinding sun kills a camera; radar shrugs it off. Fog scatters laser light; radar punches through it. Pitch darkness blinds your eyes; lidar paints the world in clean 3D. Pile up sensors with unrelated weaknesses and there is no single moment that blinds the whole car. For Waymo, redundancy isn’t waste. It is the safety argument.
So who’s right?
We can’t say cleanly, and that itself is a small scandal. Tesla publishes almost no peer-reviewed safety data, so an apples-to-apples comparison can’t even be made. What we can see is the ground shifting under Tesla’s feet. Its killer argument was always cost, and lidar has fallen from around $80,000 a unit a few years ago to a couple hundred dollars today. When Waymo built its newest hardware, it made it cheaper and better in bad weather at the same time. Every year the laser gets cheaper, “cameras only is the economical choice” sounds less like strategy and more like a sunk cost talking.
What each has to fix is clear enough.
Tesla needs to prove its cameras survive the ugly cases, the glare and the downpour and the true darkness where vision alone is weakest, and it needs to start showing data outsiders can check. Waymo needs to keep beating cost out of its sensor tower before cheap rivals turn redundancy into a luxury nobody will pay for.
Bet #2: 300,000 rules, deleted overnight
There’s a number that tells the whole story of where this industry is heading.
It’s 300,000.
That’s roughly how many lines of hand-written code Tesla reportedly deleted when it rebuilt its driving software around a single neural network. Three hundred thousand human-authored rules, thrown out, replaced by a machine that learned to drive by watching.
For most of this field’s history, a self-driving car was built like a factory line. A modular pipeline of distinct stages. Perception figures out what’s there. Prediction guesses what it will do. Planning decides the car’s move. Control turns the wheel. Each stage has one clear job you can test, debug, and certify on its own. It’s the safety tribe’s natural home, and for years it was simply how the work got done.
Then reality served up the long tail. A man sitting in the road outside a pub at closing time. A construction worker waving two contradictory hand signals. A couch in the fast lane. You cannot write a rule for each of these, because there are endless versions of them and the rules start contradicting one another. This is the wall the rule-writers kept hitting.
The end-to-end bet says stop writing rules. Train one big network to go straight from what the cameras see to what the wheels do, and let it soak up human judgment from millions of miles of example. The long tail isn’t a list to be written. It’s a pattern to be learned. It’s a beautiful idea, and it might even be the right one.
But here’s the trap, and it’s a deep one. You cannot debug a miracle. When a modular system fails, you can trace exactly which stage broke and fix it. When an end-to-end network does something baffling, there’s often no satisfying answer to why, and worse, no guarantee your fix didn’t quietly break something else. “The network just decided to” is a fine answer in a research demo and a terrifying one in a courtroom.
The most damning skeptic here isn’t a critic. It’s Andrej Karpathy, the man who built Tesla’s end-to-end system. He warns that the problem is nowhere near solved and pours cold water on the “just scale it” gospel with what he calls the march of nines. Getting from 90 percent reliable to 99 takes an enormous effort. Getting from 99 to 99.9 takes another enormous effort. And again for the next nine, and the next. There’s no sudden leap to perfection, just a brutal climb where every decimal place is its own mountain. When the inventor of the strategy tells you it’s a slog, you should probably believe him.
What needs to improve is the part nobody has cracked. The field needs learning systems you can actually interrogate. The promising work, software that makes the AI explain its own maneuvers or flag when it’s confused instead of confidently wrong, is still early and partial. Nobody has figured out how to get the long-tail brilliance of learning and the provability of structure in the same car. That gap is probably the single most important unsolved problem in autonomy.
Bet #3: The map is a ball and chain
Here’s a question almost no Waymo rider thinks to ask.
How does the car know this street so intimately?
The answer is that Waymo already drove it, slowly and deliberately, with mapping vehicles, and built a centimeter-perfect 3D model of every curb, lane line, and traffic light before a single paying passenger ever climbed in.
These HD maps are a quiet superpower.
They hand the car a rock-solid memory of the world, so it doesn’t have to rebuild reality from scratch every second and can spend its attention on the genuinely unpredictable things, the pedestrians and the other drivers. This is exactly why a robotaxi launches one city at a time. Each new city isn’t a software update. It’s a fresh surveying expedition, followed by endless re-surveying as construction and repainting keep changing the ground truth.
That’s the chain around the industry’s ankle, and the mapless tribe wants it gone. Wayve is the loudest voice here. Its pitch is that a car should read the road in real time, like a person dropped into a strange town with no GPS. Wayve says its system can do exactly that across hundreds of cities it was never specifically trained on. If that holds up, it’s the difference between a business that expands city by painful city and one that spreads the way software does, everywhere at once.
The honest caveat is that Wayve’s eye-popping numbers are mostly self-reported, not independently audited, and “no map” is not the same as “no local data.”
A map also gives you a guaranteed worst-case knowledge of the road that pure real-time perception hasn’t yet matched at Waymo’s mileage. The likely truce is a hybrid, a light map plus strong generalization. But the prize is huge, because maps are the single biggest thing standing between robotaxis and continental scale.
Bet #4: Teaching a car to dream
The wildest bet of all is that the whole problem gets swallowed by scale, that driving, like language before it, will simply fall to a big enough AI model fed enough data.
Get Ferhat Babacan’s stories in your inbox
Join Medium for free to get updates from this writer.
To get there, some companies are building world models, AIs that learn to simulate driving itself. Wayve’s GAIA system can conjure realistic, controllable driving footage out of thin air, which lets engineers summon a dangerous scenario on demand and test the car against it instead of waiting to stumble into it on a real street. GM runs millions of simulated driving-days every single night. The dream is to stop being limited by how many miles you can physically drive and start being limited only by how much computer you can buy.
It’s a thrilling idea, and it’s exactly where your skepticism should burn hottest. A model that generates realistic video isn’t necessarily generating physically true video, and a car trained to ace a simulation it also hallucinated is a snake eating its own tail. The “scaling solves everything” camp runs on a faith that its own high priests, Karpathy again among them, openly question. Driving may simply not be like language. A botched sentence is a typo. A botched left turn is a funeral.
The question no engineer can code away
Now we reach the least glamorous and most important question in the whole field, the one that should change how you read every safety claim you ever see again. How do you prove a robot drives more safely than a human?
The obvious answer, drive a lot and count the crashes, collapses under its own math. Serious crashes are, thankfully, rare. To prove statistically that a car beats humans at avoiding fatalities, you’d have to drive it hundreds of millions, by some estimates billions, of miles. The most important safety questions are the exact ones you can’t answer by driving, because the events are too rare to ever gather enough of them.
This is why the industry’s favorite number, “disengagements,” meaning how often a safety driver grabs the wheel, is close to meaningless, and nearly every serious researcher will tell you so. It depends entirely on how jumpy your safety drivers are and how easy your test routes were. It’s trivially gamed. It doesn’t even count remote interventions. Companies tout it because it’s simple and the public swallows it, not because it means much.
So the grown-ups in the room moved on to two harder things. The first is outcome data: real crash rates, weighted by severity, measured against human baselines. Waymo is the gold standard for openness here, with peer-reviewed studies and an independent reinsurer’s analysis showing big drops in injuries and insurance claims across tens of millions of miles, though it’s worth noting Waymo had a hand in producing some of that data. The second is the safety case: not one magic number but a structured engineering argument that stitches together hazard analysis, real-world miles, and massive simulation to argue the system carries no unreasonable risk.
Here is the one thing to walk away with. There is no accepted scientific method to prove a car is superhumanly safe from road testing alone. The whole field has been forced to shift from “how many miles” to “how good is your argument.” So whenever someone waves a single shiny safety statistic at you, the only sane response is to ask: measured how, against what, and verified by whom?
The human you’re not supposed to notice
It’s 3 a.m. on a street in San Francisco, and the car gliding through the intersection has nobody inside it. But somewhere, possibly in an office in the Philippines, a person is watching its camera feed, ready to help. That person is not a glitch in the story.
That person is the story.
When a robotaxi gets confused, it phones a human. This is teleoperation, and it’s less a scandal than the actual architecture of “driverless” today. It also hides the business model’s single most important question.
Start with the physics. People picture a remote worker driving the car like a video game. Above low speeds, that’s impossible. The round trip from the car’s camera to a distant operator’s screen and back runs to a few hundred milliseconds even on good networks, and a car at highway speed covers many meters in that blink. By the time the remote human reacts, the moment they were reacting to is already gone. Real-time remote driving is a latency trap. What actually works is remote assistance, where the car keeps driving and just asks for high-level advice, “is it okay to go around this,” which tolerates delay.
Now the economics, which is the entire ballgame. A robotaxi only makes money if one remote human can babysit many cars. The dream number is a thousand to one: automation handles 99.9 percent of the driving, the human handles the rest. But that ratio only works for assistance, never for driving. You cannot have one person actively steering a thousand cars at once. So the whole robotaxi business case rests on one unproven prayer, that the AI gets good enough that humans are needed rarely enough that one of them can quietly oversee an army of vehicles.
Today the real ratios are nowhere close, and the companies refuse to say what they are. In early 2026, a U.S. Senate investigation asked seven major operators how often their cars need human help. Every single one declined to answer, and several called it confidential business information. Waymo runs roughly 70 remote operators for around 3,000 cars, about half of them overseas. Tesla lets operators take low-speed control, and two of its early robotaxi crashes were caused not by the AI but by the human steering it into a fence and a barricade.
The refusal to share the numbers is the loudest signal of all. You cannot regulate, insure, or trust what you’re not allowed to measure.
The scorecard: winners, bluffers, and arms dealers
Enough principles. Here’s an opinionated, clear-eyed read on the players who matter, the ones running cars and the ones quietly arming everyone else.
Waymo is the leader, and in the West it isn’t close. Half a million paid rides a week, a fresh war chest at a $126 billion valuation, freeway driving finally switched on. But its strengths cast long shadows. Each car reportedly costs well over $100,000. Its city-by-city mapping makes growth a grind instead of a click. And for all that scale, it still hasn’t shown it earns money on a single ride. When a December 2025 power outage hit San Francisco, Waymos stalled at dead intersections, a reminder of how thin the ice still is at the edges. What it has to fix: get the car under $100,000, shrink the per-city launch slog, and prove positive margin in even one mature market before the cheap competition lands.
Tesla owns the most enviable raw materials in the business: millions of cars harvesting data, deep vertical integration, the cheapest hardware path. It also owns the widest gap between story and reality. It can light up “robotaxi available” across an entire metro on a map while the actual driverless fleet is reportedly around 20 cars. A 2026 Reuters investigation found its headline safety stats rested on a methodology that flattered the results, and most independent experts called it misleading. Its proud “no mapping needed” claim turned out to be false; it pre-mapped its zones too. What it has to fix: publish honest, comparable safety data, scale the real fleet instead of the map, and close the gap between the promises and the operation before regulators close it for them.
Zoox, Amazon’s bet, made the boldest hardware gamble, a purpose-built pod with no steering wheel or pedals, and it has Amazon’s patience and balance sheet behind it. But years after rivals went commercial, it still charges nobody, and its wheel-less design needs a special federal exemption just to operate, a chokepoint no competitor faces. What it has to fix: start charging, prove the bespoke vehicle actually delivers a cost edge, and move faster.
Pony.ai and WeRide, both publicly traded, are scaling real fleets and posting improving economics, even early profits, while living under the double shadow of US-China tension and heavy dependence on Uber for distribution. What they have to fix: turn city-level break-even into durable company-wide profit, and lean less on a single partner.
Aurora chose its battlefield wisely. Highway trucking, with fewer pedestrians and fewer surprises, is quite possibly where autonomy makes real money first. It’s running driverless freight in Texas with a clean early record. But it’s burning hundreds of millions a quarter against almost no revenue, and it had to put a human observer back in some cabs. What it has to fix: hit the observer-free next-gen launch on time, and bend the revenue curve before the cash thins out.
Vay and the remote-driving startups are the honest heretics. Their cars carry no one and work on public roads today, no AI moonshot required. A remote human simply drives your rental to your door and you take it from there. But theirs is a linear-cost business in a tech-platform costume. One human per car means costs don’t melt away with scale the way software’s do. What it has to fix: break the one-human-one-car ratio, or reposition as the teleoperation backbone the whole industry secretly leans on.
And then there’s the layer beneath all of them, the companies that don’t run a single robotaxi but stand to win no matter who does.
Wayve is the contrarian bet grown up. It started in 2017 in a rented house outside Cambridge, where a New Zealand engineer named Alex Kendall taught a car to drive itself in about twenty minutes using nothing but a camera and a learning algorithm, an idea that was, in his words, “laughed at by many.” It doesn’t operate cars. It sells the brain, licensing its end-to-end, mapless AI driver to carmakers like Nissan and Stellantis and to Uber for robotaxis, a deliberately capital-light play to be the software inside “every vehicle that moves.” Its strength is exactly that asset-light reach and its blue-chip backing, now worth $8.6 billion. Its weakness is that it has proven none of this at commercial scale. There’s no paid driverless service yet, its dazzling generalization numbers are self-reported, and it’s entirely dependent on partners actually shipping. What it has to fix: turn signed agreements into cars on roads, and replace demos with safety data outsiders can audit.
Nvidia is the arms dealer of the whole war, and arms dealers tend to outlast the armies. It sells the compute the cars think with, the simulators they train in, and a reference design partners can build on, to Mercedes, Jaguar Land Rover, Volvo, Aurora, and Chinese giants like BYD. Whoever wins the driving war will most likely run on Nvidia silicon, and that’s the strength. The weakness is that its automotive revenue, while surging, is still a rounding error next to its data-center empire, and it only grows when partners actually ship. Worse, its biggest customers keep trying to build their own chips; Tesla already did. Jensen Huang’s framing, that “the ChatGPT moment for physical AI is here,” tells you he intends to own the ground all of this stands on. What it has to fix: convert design wins into real production revenue, and make the full package too good for customers to leave.
Applied Intuition sells the workshop. It was founded by Qasar Younis, a former GM engineer who went on to run Y Combinator before leaving to build tools for the industry he came from. It makes the simulation and validation software carmakers use to test autonomy, and claims 18 of the world’s top 20 automakers as customers. It was valued at $15 billion in 2025 and, rare for this world, is reportedly profitable, partly by selling autonomy into lower-regulation domains like construction, mining, and increasingly defense, where it can earn revenue now instead of waiting on robotaxi rules. Its risk is identity. As it climbs from neutral toolmaker toward operating the full stack, it starts competing with the very customers who made it valuable, and its own deployed autonomy is, by some accounts, still rough around the edges. What it has to fix: mature the tech to production reliability, and pick a lane between enabling customers and competing with them.
Notice the pattern. Wayve sells the brain, Nvidia sells the body and the gym, Applied sells the workshop, and all three are betting that most companies will license autonomy rather than build their own. The arms dealers may end up richer than the soldiers.
The hard part was never the driving
Here’s the conclusion the whole tour keeps forcing, and it’s the most important idea in this piece. The hardest problems left in autonomy aren’t technical. They’re economic, human, and institutional.
The economic wall is that nobody has publicly proven a robotaxi makes money per ride. At seven to nine dollars a mile today, every trip is a subsidy. The entire industry is a bet that those costs collapse, on lidar (happening), on the teleoperation ratio (unproven), on utilization. The technology mostly works. The unit economics mostly don’t, yet.
The human wall is the teleoperation layer: real, growing, offshore, and deliberately hidden. Whether it shrinks into irrelevance or hardens into a permanent cost and liability is an open question, and the companies’ refusal to even disclose the numbers means we’re flying blind on the one thing that determines both safety and profit.
Then there are the institutional walls, the ones that get no headlines and matter most. There’s no common safety standard. Every company grades its own homework, hoards its crash data as a trade secret, and there’s no shared benchmark, so you literally cannot compare two robotaxis on safety the way you compare two cars on crash-test scores. There’s a regulatory patchwork, with rules that change state by state, down to how a cop is even supposed to pull a robot over, while China runs a coordinated national playbook and deploys faster because of it. There’s a liability vacuum: when there’s no driver, or the “driver” is a contractor overseas, who pays when it crashes? And there’s an accountability gap with the public. First responders in San Francisco have described themselves as unpaid roadside assistance, manually shoving stuck Waymos out of the way, once during a mass-shooting response. You cannot write a robot a ticket.
Fix these and the technology flourishes. Ignore them and the backlash writes the rules instead.
The seat stays empty
Remember the two tribes from the beginning, learning versus structure? The most telling thing happening right now is that they’re collapsing into each other. Waymo, the safety-and-structure company, has started talking about a “generalized AI driver.” Tesla and Wayve, the scale-and-learning companies, are bolting on the interpretability and guardrails they once waved away. Everyone is creeping toward the same uneasy peace: learning for the long tail, structure for the proof. The winner won’t be whoever was most ideologically pure. It’ll be whoever fuses the two without going broke or getting someone killed.
Which brings us back to that empty seat.
For a hundred years, moving a human body across a city required another human’s full attention, hands on a wheel, eyes on a road. We are now handing that job to machines, and we built the future the way the future always arrives, disguised as the past. The first cars were “horseless carriages.” The first films were filmed stage plays. And the first self-driving cars are ordinary cars with the human quietly subtracted from the seat, then just as quietly added back somewhere you can’t see, watching a screen on the far side of the planet, ready to take a wheel they’re no longer sitting behind.
The robots learned to drive. That turned out to be the easy part. The real race, the one worth a hundred billion dollars, is whether we can build a business, a safety regime, and a body of law to fit them before the seat stays empty for good.
The wheel turns.
The street slides by.
Somewhere, a person is watching.
And the question the whole industry is racing to answer is a simple one: how much longer do we need them there, and who answers for it when they blink?
