On September 3, 2026, two launches happened an ocean apart. In London, Uber riders could be matched with a Ford Mustang Mach-E driven by Wayve's neural network, with no HD map underneath it and a safety driver still in the seat. In Austin, Tesla started charging fares for the Cybercab, a car with no steering wheel and no pedals, and NHTSA opened an audit into how it was certified within hours.
Both cars are driven by software that nobody wrote line by line. That is the real story of autonomous driving in 2026, and it changes which engineering questions matter.
I should declare a bias. I build vision-only, GPS-denied navigation for drones, so I am professionally inclined to believe a camera and a good model can replace a lot of infrastructure. This article is partly me checking where that belief stops holding.
What actually changed
The traditional AV stack is a pipeline: perception finds objects, prediction guesses what they will do, planning picks a path using a large body of hand-written rules, and control executes it. Each stage has a defined interface. When the car does something wrong, the bug has an address.
The cost is that every interface throws information away, and the rule set grows without ever covering the long tail. A planner that only receives bounding boxes cannot react to a pedestrian's posture, because posture was never passed downstream.
End-to-end systems remove the hand-built interfaces. Sensor data goes in, a trajectory comes out, and the whole thing is trained from driving video. Vision-Language-Action (VLA) models go one step further by putting a language-capable backbone in the loop. NVIDIA's Alpamayo 1, announced at CES in January 2026, is a 10-billion-parameter example, and its first production home is the Mercedes-Benz CLA as a supervised Level 2 system. Wayve's AI Driver and Tesla's FSD sit in the same family.
I covered the robotics side of VLA models in an earlier piece. Driving is where the same idea meets a regulator.
The black box problem, part one: debugging
You cannot set a breakpoint in a neural driver. If the car hesitates at an unprotected left turn, there is no function to open. The fix is to find similar cases, add or reweight data, retrain, and then re-validate everything, because a learned fix is never local. Improving one behaviour can quietly degrade another.
The industry has built three answers. None is complete on its own.
1. Make the model explain itself. Alpamayo emits what NVIDIA calls a Chain of Causation: a short reasoning trace alongside the trajectory. This is useful, but an explanation is just another output of the same network. NVIDIA's own Alpamayo-R1 paper reports that a reinforcement learning stage was needed to raise reasoning-action consistency by 37%, which tells you consistency does not come for free. A separate 2026 study, Lost in Fog, found that when sensor corruption changed the model's explanation, trajectory error jumped roughly fivefold. I read that as good news for monitoring and bad news for anyone treating the trace as proof. A changing explanation is a strong warning light. A stable one is not a guarantee.
2. Test inside a learned world. Wayve's GAIA-3, a 15-billion-parameter generative world model released in December 2025, regenerates real scenes with controlled changes so the driving model can be replayed against near-misses that would be unsafe to stage. Wayve reports that results in GAIA-3 track on-road results. That is the company's own finding, and the obvious objection is that you are validating one neural network with another. Wayve seems aware of this: it is working with the University of Warwick on a UK government-funded project to validate the method itself.
3. Wrap it in something you can verify. The Mercedes deployment runs Alpamayo underneath NVIDIA's Halos safety system, which adds redundancy and guardrails. This is the part the "rules are dead" narrative misses. Rules did not disappear. They moved out of the driver's seat and into the supervisor's.
The black box problem, part two: regulators
I expected regulators to demand explainability. They mostly have not.
In June 2026, the UN's WP.29 forum adopted the first global regulation for driverless Automated Driving Systems. It does not prescribe an architecture or ask anyone to interpret weights. It asks for a safety case: structured claims, arguments, and evidence that the system is at least as safe as a competent and careful human driver, backed by an audited safety management system, in-service monitoring, and on-board data recording.
That framing is architecture-neutral, so a black box is legally survivable. It also moves the whole burden onto evidence, and evidence is where things get uncomfortable.
The US shows why. Manufacturers self-certify and NHTSA investigates afterwards. The Cybercab audit is about how Tesla certified the vehicle against federal standards, not how well it drives. Separately, NHTSA has an engineering analysis, opened in March 2026, into how FSD behaves in reduced visibility, covering about 3.2 million vehicles. And in June, US senators wrote to NHTSA asking whether anyone had independently checked the statistics behind Tesla's safer-than-human claims.
So the regulator's question is not "can you explain the network?" It is "can a third party trust your numbers?" For a system whose behaviour shifts with every retrain, that is a harder question than it sounds.
Tesla vs. the map
This fight usually gets framed as one question. It is two.
Sensors: cameras only, or cameras fused with lidar and radar? Priors: a centimetre-level HD map of the operating area, or ordinary navigation maps?
Tesla says no to both. Wayve says no to the map. Waymo says yes to both, and it is the one operating at scale: around 400,000 paid rides a week in February 2026, reported at roughly half a million by spring, with a company-reported 90% reduction in serious-injury crashes over 127 million driverless miles. That last figure is Waymo's own, so treat it as a claim with a methodology attached rather than a settled fact.
Here is the detail that reframes the debate. Waymo's production driver is also a foundation model trained end-to-end. Its head of research, Drago Anguelov, has made the point that end-to-end describes a training strategy, not an architecture. Waymo keeps internal structure and still backpropagates through all of it.
So the clash is not rules against learning. Everyone learns now. The clash is over how much independent information you give the learned system, and the thing that checks it.
Lidar matters here less as a better sensor than as a different one. Its failure modes are not the camera's failure modes. It is hard to ignore that the open federal investigation into FSD is about low-visibility conditions, which is exactly where a camera-only system has nothing to cross-check against.
The map is a different trade. It is a strong prior that makes localisation easy, and it costs you every time you enter a new city or a road changes. Wayve launching in London without one, and Waymo testing in the same city with one, is the cleanest natural experiment this field has had.
My own position, from the drone side: on a small airframe, payload and power make the sensor decision for you, and GPS denial removes the map. Vision-only is forced. A two-tonne car has the budget for a second modality, so the case for leaving it out has to be argued on cost and scale, not on physics.
The case for rules
I think the modular camp's argument is stronger than end-to-end enthusiasts admit.
A decomposed system lets you assign a failure budget to each part and verify the parts separately. A single network has a single point of common-cause failure. Fixes in a rule-based planner are local and reviewable. And traffic law is itself written as rules, so a system that can show which rule it applied has a natural path through a courtroom.
The end-to-end reply is also fair. Hand-written planners never closed the long tail after fifteen years of trying. Human drivers are black boxes too, and we license them on a behavioural test rather than a brain scan.
Where I land: the modular camp is right about verification and wrong about capability. The end-to-end camp is right about capability and too relaxed about verification. The systems actually shipping in 2026 concede both points. They use a learned driver, some preserved internal structure, an independent safety layer, and a pile of evidence.
What I’m keeping a close eye on
- London. Mapless Wayve and mapped Waymo in the same streets. If driverless approval comes at similar times with similar safety records, the map argument weakens considerably.
- The first approvals under the UN regulation. We will learn what a regulator accepts as a safety case for a learned driver, and whether world-model testing counts as evidence.
- NHTSA's low-visibility analysis. Its outcome is the closest thing to an official verdict on camera-only sensing.
- Reasoning traces as runtime monitors. Using a shift in the model's explanation to trigger a fallback is more defensible than using the explanation as a justification.
The debate in the title is already over. Nobody serious is hand-coding a driver. What is unsettled is everything around the model: what it can cross-check against, what watches it, and whether the people outside the company can believe the numbers.