Carsonight

Dear Mr. Anderson, Tesla never ignored physics. I'm writing because of what you told The Verge on October 27, when you called dropping radar from a self-driving car "a weird religious stance that ignores physics." That line got a lot of attention among Tesla owners, and I think it deserves a direct answer. You know this territory better than most. You ran Tesla's Autopilot program from late 2015 to late 2016, cofounded Aurora, and now lead product at GM. Your new employer is betting the other way. A 2028 Cadillac Escalade IQ will launch eyes-off highway driving with cameras, radar and lidar all working together. As I understand it, your case rests on four claims. Lidar helps camera systems learn faster, cameras can't see through dense fog, cameras can't measure instantaneous velocity the way radar can, and radar is so cheap that removing it makes no sense. I agree that physics is real, and nobody at Tesla is pretending otherwise. What I'd like to show you is that Tesla's own engineering has been answering these exact points for years, and two of your arguments describe things Tesla already does. A lot has also changed since you left. The in-house FSD computer arrived in 2019, camera-only Tesla Vision followed in 2021, and the end-to-end neural network that drives the car today came in 2024. So let me take your physics one claim at a time. 🎓 ON LIDAR AS A TEACHER, TESLA HAS BEEN DOING THAT FOR YEARS You said lidar works on a wavelength close to visible light, so it helps camera systems learn faster. You're right, and Tesla agrees so strongly that it built an entire training pipeline around that idea. I like to think of lidar as a tutor holding a perfect answer key. It measures distance directly by timing laser pulses, so it can grade every guess a camera network makes about how far away something is. Tesla has designed a system where a special test fleet carries radar and lidar alongside the regular cameras. Every reading is synchronized, so the precise depth and speed from the expensive sensors get attached automatically to the objects the cameras see. The labeling even uses hindsight. When two cars blur into one blob early in a clip and separate later, the training servers rewind the footage and use the clear moment to label the confusing one correctly. Tesla applies the same teacher to 3D space itself. Its occupancy network learns from lidar-backed ground truth to predict, for every small cube of space around the car, how far that cube sits from the nearest surface. That's how plain cameras learn to sketch a sloped curb or a curved bumper smoothly instead of as a pile of blocky Lego bricks. Even the grading got faster. Tesla designed a cross-labeling trick that draws a box around a car in a flat photo, then projects that box into the laser point cloud as a pyramid of vision, so the software only searches that slice instead of the whole 3D world. You probably know DeepScale, the startup Tesla acquired in 2019. It was already working out in 2016 how to train cheap sensors to mimic premium lidar data, and Tesla is still pursuing that work today. Here's the step where our views split. Once the student has learned, the teacher can stay home. The test fleet keeps the expensive hardware, and the cars customers actually buy ship with cameras only. Tesla has put real money behind this. In early 2024, lidar maker Luminar named Tesla its largest customer for the quarter, and Elon soon replied on X, "We don't need them even for that anymore." The student is even learning to write its own answer key. To map the shape of the road surface, Tesla has designed a method that can triangulate ground truth points straight from camera images instead of relying on lidar. Co-training requires lidar in the classroom, not in every car on the road. With respect, GM's plan keeps the teacher riding along forever. ⚡ ON SPEED, CAMERAS ARE QUICKER THAN THEY LOOK Your sharpest point is about speed. Radar can measure how fast something is moving in a single snapshot, and a camera can't. Radar's trick is the Doppler effect, the same physics that makes an ambulance siren sound higher as it rushes toward you and lower as it drives away. Radar reads that pitch shift in its own radio echo to measure closing speed instantly. A camera has to work speed out from motion. It compares where an object sits in one frame with where it sits in the next, much like you judge a thrown ball by watching it fly. That sounds slower, but the gap is tiny. Tesla's own designs describe cameras running at rates such as 36 frames per second, which puts a fresh frame in front of the network about every 28 milliseconds, quicker than a single blink. The physics also cuts both ways. Doppler only measures motion toward or away from the sensor, so a car crossing your path sideways at an intersection produces little to no Doppler shift. To catch that sideways motion, a radar has to track the car across several scans, which is exactly what a camera does with frames. On that kind of motion, the "instantaneous" advantage disappears. Radar has another well-known weak spot. Owner's manuals for radar-based cruise control have long warned that the system may not brake for a stopped vehicle, partly because a parked car, an overhead sign and a bridge can all look like the same motionless metal. Now here's what Tesla has built since you left. Tesla has designed its vision network to output a full 3D velocity vector for every object, covering speed, direction and even height. It learned that skill from the radar on its test fleet, which turns ordinary cameras into something that behaves like a Doppler radar made of pixels. Tesla's virtual camera invention goes further by measuring speed in two different ways. Cars are tracked relative to the moving Tesla itself, while pedestrians are tracked against the fixed ground, so the system knows a person's true walking path even while the car accelerates or turns. When things get fast, the network speeds up its memory. A frame selector pulls video snapshots more rapidly as speeds rise and blends them with the car's own acceleration and yaw rate, which is how quickly the car is rotating, to build a smooth picture of motion over time. There's also a dedicated sentinel for the car directly ahead. A super-narrow model watches only the closest vehicle in the lane through the narrowest forward lens and keeps estimating its distance, speed and acceleration, which is exactly what you need to brake hard and early. So yes, a camera infers velocity instead of measuring it in one shot. But inference refreshed every few dozen milliseconds, trained on radar's own answers, and able to see sideways motion that Doppler misses is not a camera ignoring physics. 🌫️ ON FOG, I'LL GIVE YOU THAT ONE, BUT HEAR ME OUT No camera sees through dense fog, and neither does the human eye. Radar does. But look closely at your own argument. The near-visible wavelength that you said makes lidar a great teacher is the same reason lidar struggles in fog, because the tiny water droplets that scatter visible light also scatter laser light. So the fog argument is really a radar argument, and radar alone can't finish the job. It can't read lane paint, a traffic light, a speed limit sign or a flagger's hand signal, so even a car loaded with radar still leans on its cameras when the world turns gray. That's why I think the real question isn't whether a car can see through fog. The real question is whether the car knows when its vision is degraded, and whether it behaves like a careful driver when it is. This is where Tesla's visibility invention comes in. Tesla has designed a separate neural network whose only job is grading its own eyesight. It splits every camera feed into a fine grid and scores each cell from clear to blacked out, building a live heat map of where the car can and can't trust what it sees. Each camera position gets its own specialist. A model judging the left pillar camera was trained on images from that same pillar location across many other cars, because a windshield camera behind wipers and a fender camera facing tire spray get dirty in very different ways. The system also diagnoses the cause. It can tell a smear of mud from fog, heavy snow or a hardware fault, then reach for the right fix, such as running the wipers or warming the glass. When a lens is blinded for a moment, the car doesn't forget the road. Tesla anchors what it saw to fixed spots in the physical world, so a pothole hidden by a truck's spray stays pinned on the car's mental map for those few seconds. And when conditions really are too much, the car responds the way a good human driver would. It slows down to widen its reaction window, can move to a safer lane, and can hand control back with a plain explanation such as "Due to Heavy Snow, Autopilot is Reducing Speed." Tesla is also sharpening the eyes themselves. Its designs feed raw, unprocessed sensor data straight into the AI chip, the same idea behind the "photon count" images Elon posted in May, when he said this is why FSD can see so well at night or through extreme glare. Around the lens, Tesla has designed micro-cone glare shields, corrective windshield patches, pulsed laser cleaning and even gas-absorbing plastics that keep a hazy chemical film off the inside of the glass. Each one attacks a different way a camera can lose its view. Fog will always be fog. Tesla's bet is a car that measures how blind it is, remembers what it saw, and slows down exactly when a wise driver would. 💸 YES, RADAR IS CHEAP, BUT THAT WAS NEVER THE POINT You're right that a radar unit doesn't cost much. But Elon's case against it was never about the sticker price. In April 2021, as Tesla prepared to remove radar, Elon explained it this way: "Sensors are a bitstream and cameras have several orders of magnitude more bits/sec than radar (or lidar)." A second sensor, he added, has to meaningfully improve the signal to be worth the complexity of integrating it. In August 2025 he put it even more bluntly: "Lidar and radar reduce safety due to sensor contention. If lidars/radars disagree with cameras, which one wins?" In the same post, he wrote that Tesla turned off the radars in its cars to increase safety. Picture two witnesses to the same moment telling different stories. The radar reports a solid object ahead, the cameras see an empty road under a steel overpass, and the car has a fraction of a second to decide who's wrong. Trust the radar every time and you get phantom braking under bridges. Trust the cameras every time and the radar was only along for the ride. Then there's the cost that never shows up on a parts list. Every extra sensor needs calibration, its own data pipeline, fusion software and validation for every way the sensors can disagree, and that burden grows with each new sensor type. Tesla's own history shows how heavy that lift is. The DeepScale design for mimicking lidar needed separate neural branches for each sensor type just to absorb their different quirks, timing and mounting positions before merging them. I'd argue the biggest hidden cost is data. Tesla's FSD (Supervised) fleet has logged more than 16.7 billion miles, nearly all of it seen through cameras alone, and a radar channel added today would arrive with none of that history to learn from. Look at what happened at Volvo this summer. In July, the company confirmed it was dropping lidar from the EX90 and ES90 after cutting ties with its supplier Luminar, which went bankrupt, and Norwegian owners are receiving NOK 18,000 per car for lidar features that never arrived. Elon's reply was short: "I did try to warn them. Humans drive using neural nets and optical sensors. Same is true for robot cars." Cheap hardware can turn into an expensive argument, and back in 2021 Tesla decided it would rather not have the argument at all. 🛣️ WHILE YOUR SYSTEM WAITS FOR 2028, CAMERA-ONLY TESLAS ARE ALREADY LOGGING THE MILES Your plan for handing control back deserves real credit. You said the Escalade won't toss the wheel to a surprised driver, and will instead give a countdown of around ten minutes and pull off at an exit if nobody takes over. That philosophy should sound familiar, because Tesla's visibility invention describes the same graceful exit. The car slows down first, looks for a safer lane, explains the reason on the screen, and only then hands control back. The real difference is timing. GM's eyes-off system is still a 2028 promise, while Tesla's camera-only cars are collecting evidence right now. Tesla's latest FSD (Supervised) safety report shows 6,006,200 miles per major collision in North America, compared with a U.S. average of 660,164 miles. That's roughly nine times as far between airbag-level crashes, measured over a rolling twelve months. Tesla's new Robotaxi safety report goes a step further. Driverless Teslas with no safety monitor inside have covered 3.1 million miles in Texas and Florida this year, including 862,000 miles after dark and 113,000 miles in rain, with zero people injured by a Robotaxi and zero airbag deployments, based on the crash reports Tesla files with NHTSA. And that Robotaxi fleet runs on the latest Model Y and the Cybercab, neither of which carries radar or lidar. I find that hard to square with the idea that camera-only driving ignores physics. 🔭 WHERE I THINK WE REALLY DISAGREE You framed this as physics versus faith, but I think both sides accept the same physics. What we really disagree about is where the intelligence should live. GM puts more of it in the hardware, stacking sensors so each one covers another's weak spot. Tesla puts it in the neural network, using radar and lidar as teachers in the classroom and then trusting a trained brain behind the cameras. There's a quiet twist hidden in that choice. A multi-sensor car can only learn from fleets that carry the same sensors, while every camera-equipped Tesla on the road feeds the same classroom. Physics decides what a sensor can see. Data decides what a car understands. Best regards, Ming P.S. If you're ever in Austin, a driverless Tesla might make my case better than I can, and it won't need lidar to find you.