Waymo published a technical post on 26 August setting out ten lessons from more than 200 million fully autonomous miles. It never names Tesla. It does not have to: three of the ten describe the architecture Tesla has chosen, and say it does not arrive.
The author is Srikanth Thirumalai, Waymo's vice president of onboard software, which matters — this is the person responsible for what runs in the car, not a chief executive giving a conference answer.
The claim that goes furthest
"Simply improving a driver-assist system (L2) for full autonomy is a false summit. True L4 maturity can only be safely achieved by a purpose-built system."
This is the argument with teeth, because it is not about sensors. Tesla's entire roadmap is that today's supervised FSD — a Level 2 system with a human legally responsible — becomes tomorrow's unsupervised one through better software on the same fleet. Thirumalai's position is that this path has no summit at the top of it: a system only matures, in his account, once it is solely responsible for the driving task, because that is the only condition that exposes it to the real consequences of its decisions and surfaces situations that human supervision quietly smooths over.
You can run billions of simulated or supervised miles, the post argues, and still not have run the miles that count.
Two more that land on Tesla
HD maps are a prior, not a crutch. Waymo lists them as "a powerful prior" — pre-computed knowledge the car brings to a road before it sees it. Tesla's position has been the reverse for years: general vision without map dependency is what makes a system scale to any road. Waymo's counter is that discarding known-good information to prove a point is not a safety argument.
End-to-end is a black box. On the architecture Tesla has moved toward, the post is blunt: "Pure end-to-end (E2E) neural architectures … run the risk of black box failures", meaning it is hard to understand how a decision was made. Waymo pairs this with a separate lesson — "you can't build trust with a black box" — and a design it calls a Critic, a component whose job is to evaluate the driving model rather than to drive.
The full list
- Multimodal sensors are indispensable
- HD maps are a powerful "prior"
- Fewer, larger models are better
- You can't build trust with a black box
- Closed-loop simulation reveals more edge cases
- Every great driver needs a great Critic
- Vision Language Models improve scene reasoning
- AI is only as effective as the governance that evaluates it
- A data flywheel enables continuous improvement
- There is no substitute for fully autonomous experience
The sensor lesson — "Cameras are incredible, but they aren't enough", with lidar supplying "the wireframe" and cameras "the semantic overlay" — is the familiar half of Waymo's case, and TeslAnt covered the fuller version of it in Dmitri Dolgov's argument about the ceiling camera-only systems hit. What is new here is that the sensor question is one lesson out of ten, and the other nine are about the pipeline.
Why European readers should care
This is marketing as much as engineering — Waymo is talking its own book, and a company with 200 million driverless miles and a lidar bill has every reason to argue that both are mandatory. Tesla's counter-argument, that scale and cost decide who deploys, is not answered by any of the ten.
But the timing is not accidental. Waymo has committed to Munich as its first European robotaxi city by the end of 2027, and European regulators are writing driverless rules now. A document that says "an L2 system cannot be upgraded into an L4 one" is a technical standard an approval authority can pick up and hold against exactly one applicant.