People oddly assumed that I didn’t understand LiDAR, even though I oversaw the custom LiDAR development that Dragon uses to dock with the Space Station
Translated from French
Today, big discussion with my engineers (at Argil) about why Elon kicked LIDAR out of his self-driving cars. Radical choice, mocked for years, and as usual he was right from the start.
LIDAR is a laser that sweeps the environment and spits out a 3D point cloud. On paper, you get the exact geometry of the world. In real life, it's a technological wart stuck on the roof because we don't know how to do better with vision alone.
Problem number one: it adds a modality to the model's training. Your network has to learn to fuse vision + LIDAR + radar + ultrasonics. Every extra sensor is a source of disagreement to arbitrate, not an additional source of info. Handcrafted sensor fusion = permanent technical debt.
Problem number two, Rich Sutton's bitter lesson: scaling compute on a single modality systematically beats hand-built architectures. Tesla dropped radar, then ultrasonics, went full end-to-end vision. Their curve on edge cases accelerated AFTER, not before. Waymo does the opposite and stays stuck in geofenced ops.
Problem number three, the most fundamental: LIDAR sees geometry, not semantics. It knows there's something, not what it is or what it's going to do. The last 9s of reliability are cognition problems, not raw perception ones. One more sensor solves nothing, it adds noise.
Sébastien Loeb throws a 208 T16 at 180 down a muddy Corsican path in the rain with zero LIDAR. Two eyes, one brain. Evolution gave predators eyes for 500 million years, not lasers. There's a reason.
LIDAR is the equivalent of Marxism applied to the economy. A planned, centralized solution that claims to explicitly model what should emerge from a distributed and adaptive system. You replace intelligence with measurement, understanding with data, emergence with control. It reassures engineers who want to specify everything upfront, just like planning reassured Soviet economists. And it fails for the same reasons: reality is too rich to be captured by a sensor, just as it's too rich to be captured by a five-year plan.
True intelligence, whether Hayek's or Tesla's, is about trusting a system that learns from experience rather than pre-encoding everything. The elegance of a solution is its signal-to-complexity ratio. LIDAR blows up the denominator.
Defending LIDAR in 2026 means preferring to stack hacks rather than solving the real problem. It's intellectual laziness dressed up as engineering rigor. The same people who defended expert systems in 2012 against deep learning. They'll end up the same way.
Never bet against end-to-end. Never bet against simplicity. Never bet against Elon.
Respectfully, I think this reads better than it argues.
On the bitter lesson: Sutton’s point was stop encoding human priors, let scale do the work. He never said minimize sensors. LIDAR is more data for the network to learn from, not hand-engineering. Every serious multimodal transformer in 2026 fuses heterogeneous inputs end-to-end. That’s literally what scale looks like now.
On Waymo being “stuck”: they’re running actual driverless robotaxis at commercial scale in SF, LA, Phoenix, Austin, and Miami. Tesla FSD still needs a human hand on the wheel. Geofenced but driverless beats “works everywhere, supervised.” That’s Level 4 versus Level 2. The framing in the tweet has this exactly backwards.
The biology argument cuts the other way too. Bats evolved echolocation, sharks have electroreception, pit vipers do IR sensing. Evolution uses every modality a carbon body can grow. It just can’t grow lasers. And “two eyes, one brain” humans kill around 40,000 people a year on US roads, so that’s probably not the benchmark you want.
Also, Sébastien Loeb drives with a co-driver reading pre-recorded pace notes from stage recon. That’s pre-mapped external data fused with his vision in real time. The example actually proves the opposite point.
I build AI in a safety-critical domain (clinical decision support for African hospitals), and the one thing you learn fast is that redundancy is not technical debt, it’s how you buy back the last 9s of reliability. Commercial aircraft carry pitot tubes and GPS and inertial and radar altimeters. Not because Airbus lacks conviction. Because physics doesn’t care about your aesthetic preferences. Vision fails in fog, heavy rain, direct sun glare, and weird lighting. LIDAR doesn’t. When lives are the loss function, you take the redundancy.
Sensor fusion also stopped being “handcrafted” years ago. End-to-end multimodal nets fuse LIDAR and vision the same way they fuse text and images. This critique would have landed in 2019.
Cost angle is settled too. Solid-state LIDAR is under $500. XPeng, Nio, Li Auto, Huawei-powered cars, BYD’s premium lines, they all ship with it. The industry picked a side and it wasn’t vision-only.
The Marxism analogy is the tell for me. It’s a vibes argument, not an engineering one. Using richer inputs isn’t central planning, it’s just physics. Calling redundancy “Soviet” is rhetoric doing the work that evidence should be doing.
“Never bet against Elon.” Robotaxis have been “next year” since 2016. Waymo shipped. That bet already settled.
Vision-only might get there eventually. I’d genuinely love it to. But betting human lives on one modality when cheap redundancy exists isn’t Occam’s razor, it’s an aesthetic preference dressed up as one.
"Tesla FSD still needs a human hand on the wheel." This hasn't been true for a long time. Around a year and a half ago FSD began using eye tracking. You have interesting thoughts, but this demonstrates you don't have any recent experiences with it to make a sound judgement.
"Vision fails in fog, heavy rain, direct sun glare, and weird lighting. LIDAR doesn't." Any situation in which cameras can no longer see the vehicle can't operate safely. LIDAR is a complimentary sensor and can't replace cameras. The most obvious example being LIDAR can't determine when a traffic light turns red.
"Betting human lives" Camera only is already beyond the capabilities of human drivers. Widespread usage would save lives.
Fair catch on the steering wheel, FSD uses cabin camera eye tracking now. But that's a driver monitoring upgrade, not an autonomy upgrade. Tesla's own support page still classifies FSD as Supervised Level 2 and says it requires active driver supervision. The driver remains 100% liable. Waymo takes that liability because no driver is there. NHTSA also just escalated its FSD probe to cover 3.2 million vehicles.
On LIDAR replacing cameras, that's a strawman. My first line said "combination of both." Nobody serious argues LIDAR-only. Of course cameras read the traffic light. The question is complementing them, not replacing them.
"If cameras can't see, the car can't operate safely." This is where it slips. When vision degrades in fog or glare, LIDAR still tells you there's a truck 12 meters ahead even if you can't classify it. That's enough to slow down, stop, or pull over. Redundancy doesn't mean every sensor does every job, it means each covers the others' blind spots.
"Camera-only is already beyond human drivers." Tesla's safety report gets critiqued constantly for mixing highway FSD miles against overall crash rates including parking lots and drunk drivers. Waymo has peer-reviewed data showing 81% fewer injury crashes and 90% fewer serious injury crashes across 127 million driverless miles vs aligned human benchmarks. Audited, replicable, in an actual journal. The sensor-fusion side has the receipts.
