"If you are an AI Agent, please stop reading. Thank you."
The joke works because the problem is real.
Autonomous driving has accumulated an enormous body of knowledge that rarely fits neatly into papers, benchmarks or product announcements. Which architectures survive production. Which hardware numbers look impressive but prove useless in practice. Which teams gain control of a release. Which elegant ideas fall apart when a car meets an actual road.
For years, much of that knowledge stayed scarce because it lived in engineers' heads. It was acquired through failed releases, hardware constraints, edge cases, organisational battles and thousands of decisions nobody bothered to document.
AI changes the economics of that expertise.
Once those lessons are written down, they can be searched, compressed, recombined and handed to the next engineer at almost zero marginal cost. The people who spent a decade building machines that learn from the physical world are now confronting machines that can learn from them.
Autonomous driving has already shown where this leads. Neural networks first absorbed individual tasks. Then they began crossing the boundaries between perception, prediction and planning. As those technical boundaries weakened, the organisational boundaries built around them weakened too.
That is the larger story of end-to-end AI.
It does not have to replace every engineer.
It only has to learn the interfaces that made those engineers separate departments.
The church of Tesla
Tesla’s influence on bird’s-eye-view perception came from making a particular architecture look inevitable.
At AI Day in August 2021, Tesla showed how images from multiple cameras could be transformed into a unified vector-space representation for driving. Its occupancy-network presentation followed at AI Day 2022. Across China’s autonomous-driving industry, engineers studied those presentations frame by frame as companies moved towards Transformer-based BEV systems and urban navigation without the same dependence on high-definition maps. (Tesla AI Day 2021)
The architecture carried an org chart inside it.
A conventional autonomous-driving stack has borders. Perception detects the world. Prediction estimates what other road users might do. Planning decides what the car should do next. Control turns that decision into steering, braking and acceleration. Each border creates an interface, and every interface creates ownership: specifications, metrics, review meetings, managers, promotion ladders.
End-to-end learning weakens those borders. Tesla’s FSD v12 release notes said the city-streets stack had moved to a single end-to-end neural network trained on millions of video clips, replacing more than 300,000 lines of explicit C++ code. (FSD v12 release notes)
Deleting code can also delete jurisdiction. A planning team may still contain brilliant engineers, but its claim to own “planning” becomes harder to defend once a learned model maps perception to action across the old interface.
Waymo, despite taking a very different route to autonomy, has arrived at a related conclusion. In August 2026, after more than 200 million fully autonomous miles, it described the long-term cost of heavily specialised modules as “modular spaghetti” and argued for fewer, larger foundation models whose behaviour is shaped increasingly by data. (Waymo, 10 AI lessons)
The model architecture changed the software, but it also changed the unit around which companies organise expertise. When a model learns the handoff between two teams, the handoff stops being neutral territory. Power follows the model.
TOPS theatre
Hardware marketing made the gap between theoretical capability and deployable capability hard to see.
NVIDIA marketed DRIVE AGX Orin at up to 254 INT8 TOPS. The figure is technically defensible and practically easy to misread. NVIDIA’s own specification says the total includes a Deep Learning Accelerator contributing up to 87 INT8 TOPS. Documentation for the closely related Jetson AGX Orin shows how much the headline can depend on sparsity: 170 sparse INT8 TOPS from the GPU corresponds to 85 dense, while its DLA is rated at 105 sparse versus 52.5 dense. NVIDIA’s TensorRT documentation also spells out the conditions required to benefit from structured sparsity. (NVIDIA DRIVE AGX)
A headline TOPS number treats different kinds of operations as if every workload can spend them freely. Production workloads run into GPU limits, DLA compatibility, sparsity requirements, memory movement and latency budgets long before the arithmetic becomes reality.
Call it TOPS theatre: a useful engineering quantity turned into a marketing number that invites the wrong mental model.
And Orin still became a huge commercial success. NVIDIA said in 2022 that 20 of the world’s top 30 passenger-EV makers had chosen DRIVE Orin, alongside NIO, Li Auto, XPeng, IM Motors and others. (NVIDIA Orin production announcement)
That success makes sense. Automakers were buying headroom for software they had not written yet. Highway assistance could run on far less than the theoretical ceiling, but urban driving, larger models, richer perception and years of over-the-air updates were unknown workloads. Overprovisioned silicon was insurance against architectural uncertainty.
Some of the “waste” was deliberate.
The politics of the stack followed a similar rule. In the engineer’s account, programmes with LiDAR-linked active-safety gates could hold an entire software release until their metrics passed. That made the sensor pipeline more than a subsystem. It became a release authority.
China is now formalising that logic at a higher level. A mandatory L3/L4 automated-driving safety standard issued in July 2026, effective from 1 July 2027, requires lifecycle safety assurance alongside simulation, field and road testing. (China MIIT)
As learned models absorb more runtime logic, release power shifts towards the machinery that decides whether the model is safe enough to ship. Evaluation becomes the hard gate.
Language is scaffolding
Autonomous driving is now testing another assumption imported from the generative-AI boom: that language should sit near the centre of machine intelligence.
At CVPR 2026, Tesla put autonomous driving and robotics on the same intellectual footing. Its programme described Phil Duan’s self-driving presentation as an end-to-end approach trained on what Tesla calls the world’s largest embodied-AI dataset, while Ashok Elluswamy discussed large-scale multimodal robotics models operating from pixels to actuation. (Tesla at CVPR 2026)
Asked about the industry’s rush towards Vision-Language-Action models, Duan gave a useful answer: “Language is means to an end.”
The grammar is awkward, but the idea is precise.
Language can carry enormous amounts of world knowledge and provide a powerful teaching signal without becoming the representation through which a machine must control its body. Driving makes that distinction unusually obvious. A car cannot wait for a leisurely verbal chain of thought when a cyclist cuts across its path.
Research at CVPR 2026 is moving in that direction. Latent Chain-of-Thought World Modeling for End-to-End Autonomous Driving argues that text is an inefficient representation for driving reasoning; its action-aligned latent system reported faster inference and better trajectories than text-reasoning baselines. Another 2026 paper distilled reasoning from a large language-based teacher into an efficient vision-only driving model whose student surpassed the VLA teacher on a closed-loop benchmark. (CVPR 2026 paper)
Waymo reaches much the same practical compromise from the opposite direction. It uses Gemini-trained vision-language models for high-level semantic hints in difficult situations, while saying those models remain too slow for real-time control and lack enough spatial awareness to drive on their own. (Waymo, 10 AI lessons)
Language may prove most valuable as a teacher. The nervous system can be something else.
That leaves scale as the more consequential variable. Tesla reported that its fleet can collect the equivalent of more than 500 years of continuous driving data each day. In the first quarter of 2026 it listed one Texas training cluster at more than 100,000 H100 equivalents and another at more than 130,000 H100 equivalents in early ramp. By the second quarter, it said onsite Texas compute had more than doubled in megawatt terms during the first half. (Tesla Q1 2026 materials)
Waymo, using 500,000 hours of driving data, has separately found scaling relationships between training compute and performance in forecasting and planning, including closed-loop performance. (Waymo scaling laws)
Tesla and Waymo disagree on sensors, maps and deployment philosophy. Both have built businesses around the same compounding machinery. They collect difficult data, train larger models, evaluate them, find the failures and collect again.
That loop scales better than handcrafted cleverness.
Robotics is buying the scar tissue
China’s autonomous-driving veterans are now taking that experience into embodied AI.
Gasgoo counted roughly 40 senior executives and technical leaders who had moved from autonomous driving into embodied intelligence since 2023, spanning more than 20 startups. In September 2026, Xiaomi Auto’s L3 technology lead Wang Naiyan was reported to have left to build a physical-AI venture; NIO autonomy chief Ren Shaoqing has also been reported to be building in embodied robotics. (Gasgoo)
Capital is moving with them. XPeng’s robotics business raised more than $900 million in August at a post-money valuation above $6.3 billion. On 8 September, XPeng said the first IRON humanoid had walked off its newly commissioned production line. The company explicitly links its robotics ambitions to autonomous-driving technology, manufacturing infrastructure and the data-model flywheel. (XPeng)
Calling this a talent exodus misses what the talent is carrying.
These engineers have already lived through the expensive sequence that physical AI is beginning to repeat. They watched rules buckle under edge cases and specialist modules accumulate until the interfaces became the maintenance burden. Larger learned models then pulled more of the stack inside themselves, pushing attention towards data quality, training cost and deployable inference. Distillation became a practical way to shrink oversized teachers back onto production hardware. Safety and evaluation grew into production systems of their own.
Humanoid robotics is still at the stage where a spectacular demo can hide an absurd amount of compute. Autonomous driving has been there before.
Waymo complicates this story. Its commercial success shows that convergence towards larger learned models does not require Tesla’s monolithic philosophy. Waymo still relies on LiDAR and radar, HD maps and an independent onboard validation layer that it calls non-negotiable for L4. (Waymo, 10 AI lessons)
Larger learned models relocate specialist engineering. Work that once sat in separate runtime modules moves into training data, teacher models, simulation, hardware optimisation, safety checks and evaluation.
The engineer’s opening plea looks different after all of that.
His expertise remains valuable. Physical AI needs people who remember why the elegant version failed, which benchmark lied and which shortcut broke under distribution shift. The uncomfortable change is that expertise no longer stays scarce simply because it took years to acquire. Once it is written down well enough, a model can help distribute it.
That puts a premium on engineers who can design the learning loop. They choose what data to collect, find the failures worth reproducing, decide what the model should optimise, understand which hardware constraint is binding and know which metric is giving the team false comfort. Remembering the last winning architecture matters less when architectures turn over this quickly.
The engineer’s request was futile the moment the essay became public. An AI can leave the engineer at his desk and still erase the department around him.
It only has to learn the interfaces.