LIDAR in self-driving cars

 

Elon Musk

@elonmusk

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

Brivael - FR

@brivael

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.

The Macho Maniac

@Mr_RobotKe

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.

mark young

@forevermcy

·

9h

"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.

The Macho Maniac

@Mr_RobotKe

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.

Inventors Who Didn’t Invent What They Are Famous for Inventing

Adapted from https://www.wsj.com/tech/us-famous-inventions-inventors-wrong-3d5b0a0f?mod=hp_jr_pos1

Inventors Who Didn’t Invent What They Are Famous for Inventing

Any schoolchild knows who came up with the telephone or the sewing machine. But conventional wisdom often has it all wrong.

By Beth DeCarbo

April 12, 2026 12:00 pm ET

Ask any historian—or child, for that matter—to name the “father of the telephone,” and most will say Alexander Graham Bell, who patented his revolutionary invention in 1876.

But long before that, in 1871, inventor Antonio Meucci received a “caveat,” a precursor to a patent, for his version of a telephone, which he called a speaking telegraph. In fact, Meucci already had a working model in his home in New York City’s Staten Island. While tinkering in his basement laboratory, he used it regularly to communicate with his invalid wife, who was confined to her bedroom upstairs.

So why is Bell widely hailed as the father of the telephone and not Meucci? The world will never know for sure, but it might have come down to just $10.

Beset by financial hardship, in part because of a devastating injury in a ferryboat accident, Meucci was unable to afford the $10 fee to renew his caveat, which as a result expired. So on March 7, 1876, Bell received U.S. Patent No. 174,465 for his invention of the telephone. And for posterity, his “Experimental Telephone” can be found at the Smithsonian Institution’s National Museum of American as an artifact of American enterprise.

Illustration of a cutaway drawing of Antonio Meucci's prototype telephone, showing the receiver and transmitter components.

Cutaway drawing of Meucci’s telephone prototype, 1870s. Mondadori/Getty Images

History books are riddled with examples of pioneering inventors, like Meucci, whose work is relegated to the footnotes. In many cases, their creations are still in use. But don’t expect to see bronze monuments memorializing these inventors. Or tour the labs where their creations were birthed. Or behold their prototypes displayed in a museum behind glass.

In truth, few inventors earn immortal fame. It takes an elusive alchemy of skill, knowledge and resources—with a dash of good luck—to introduce a groundbreaking, or even life-changing, invention to the world.

“If I’m not able to get venture capital, get it made to scale, manufacture it, market it, get it out to people, what difference does it make?” asks Eric S. Hintz, acting director of the Lemelson Center for the Study of Invention and Innovation at the Smithsonian’s National Museum of American History.

Down to the wire

In 1867, Alphonso Dabb, Lucien B. Smith and William D. Hunt applied for patents for their nascent designs of barbed-wire fencing, which was much needed at the time as pioneers set up homesteads on the Great Plains, and cattle ranchers moved giant herds to railheads.

But designing barbed wire on paper was easy compared to actually manufacturing it in the quantities necessary to enclose vast tracts of land, and none of those early designs went into full production, according to Farm Collector magazine.

A turning point came when an Illinois farmer, Joseph F. Glidden, saw a barbed-wire exhibit at the 1873 DeKalb County Fair. According to family legend, Glidden was inspired to design his own wire to keep livestock out of his wife’s garden. After some trial and error, he and a local blacksmith retrofitted a hand-cranked coffee mill so that it could efficiently twist wire into uniform barbs. The barbs were placed on a wire, and a second wire was twisted around it to hold the barbs in place.

Advertising poster for Glidden Steel Barb Fence Wire featuring a portrait of inventor Joseph Farwell Glidden and views of the manufacturing works.

An advertisement for Glidden Steel Barb Fence Wire, featuring Glidden. Library of Congress/Corbis/VCG/Getty Images

On Nov. 24, 1874, Glidden was awarded Patent No. 157,124, which he prophetically named “The Winner.” The hand-cranked coffee mill was replaced with horse-powered machinery, and Glidden and a partner, Isaac Ellwood, formed Barb Fence Co.

So who is the father of barbed wire? Inventors like Dabb or Smith, who had an original idea? Or an inventor like Glidden, who had an economically viable idea? Even now, barbed-wire collectors debate the question.

At the Smithsonian, “we try to be a little careful about saying something is first,” Hintz says. “And often, the thing that is first isn’t necessarily the thing that has the most impact.”

By figuring out how to economically mass produce barbed wire, Glidden certainly solved a problem. As a result, Barb Fence Co. gave its founders both riches and a historical legacy. In DeKalb, Ill., today, guided tours of the Ellwood House Museum and Glidden Homestead showcase lavish Victorian-era design.

Following the thread

Much like manufacturing barbed wire, making clothes in the 1860s was a time-consuming effort, with pretty much everything people wore sewn by hand. The clothing industry also employed tens of thousands of people, from skilled tailors and dressmakers in upscale shops to widowed needlewomen working 16-hour days.

One man, Walter Hunt, was the first to figure out how to make the work less onerous. In 1833, he invented a lock-stitch sewing machine that used two threads, one passing through a loop in the other, with both threads interlocking. This was the first machine that did more than merely mimic a hand stitch, according to the National Museum of American History. 

Black-and-white portrait of Walter Hunt, inventor of the safety pin and sewing machine.

Walter Hunt neglected to sew up a patent. Alamy

But Hunt didn’t patent his invention at the time, in part because his daughter owned a corset-making business in Manhattan. In fact, she declined her father’s offer of a sewing machine to speed up the work, fearing that mechanized sewing would eliminate jobs. In John Scott’s 1880 book “The Story of the Sewing Machine,” Hunt’s daughter reportedly said, “The introduction of such a machine into use would be injurious to the interests of hand-sewers.” 

While Walter Hunt’s daughter dissuaded her father from obtaining a sewing-machine patent, Elias Howe Jr.’s wife inspired her husband to get one.

According to the Cambridge Historical Society, in the late 1830s, Howe was working at a shop in Boston that made mariner’s tools and scientific equipment. One day, an aspiring inventor brought in a knitting machine that he was working on, which may have planted a seed in Howe’s mind for his future creation.

In 1843, when he was in his early 20s, Howe was forced from work for a time due to a disability. To pay the family’s bills, his wife took in odd-job sewing work. While watching her sew by hand, Howe realized that a machine could speed up the process, perhaps solving the family’s financial difficulties.

Howe dedicated himself to inventing a sewing machine, a painstaking effort, even with support from a family friend. Finally on Sept. 10, 1846, Howe was awarded U.S. Patent No. 4,750 for his lock-stitch sewing machine.

Then, thinking he would have more success selling his invention overseas, Howe traveled to England. Meanwhile, competitors started selling sewing machines based on Howe’s design in America.

Howe returned to the U.S., broke and unemployed, only to discover that copycat lock-stitch sewing machines had become extremely popular. Howe sent letters to suspected patent-infringers, including Isaac Singer (who would go on to make the sewing machine wildly successful).

An 1860s advertising poster of inventor Elias Howe and his sewing machine, and the first Elias Howe Jr. sewing machine, completed in 1845.

Getty Images (2)

Eventually, Howe ended up in court in a yearslong series of cases dubbed the Sewing Machine War. In the end, he beat both Isaac Singer and Walter Hunt, among others. Singer was forced to pay him royalties for every machine sold, making Howe a millionaire. And later, in 1853, the U.S. patent office denied Hunt’s belated application for his 1830s sewing-machine invention.

Illustration of Walter Hunt's original patent for the first safety pin from April 10, 1849, showing eight different pin designs.

Walter Hunt's original patent for the first safety pin. Library of Congress/Corbis/VCG/Getty Images

In the denial, the patent commissioner scolded Hunt for allowing “his discovery to slumber” for so long, saying he submitted an application only to “strangle” someone else’s invention, presumably Howe’s. (Hunt eventually received a patent for improvements in 1854.)

Today, a bronze statue of Howe stands in Bridgeport, Conn., where the inventor’s factory once cranked out lock-stitch sewing machines. Streets, postage stamps and at least one former school building bear his name.

As for Hunt, he continued to file patents in relative obscurity, never amassing the wealth of other Industrial Age inventors. Adding insult to injury, to settle a debt, Hunt sold the 1849 patent rights for something that could have kept his legacy alive: the modern-day safety pin.

Screwed!

Ditto for a man named John Thompson, whose invention helped keep fumbled fasteners from falling on the floor. On May 9, 1933, Thompson received the patent for the cross-shaped screw head and the hand tool used to drive it. Screws with a cross-shaped head were shown to outperform those with a slotted head because a cross-shaped driver “sticks” inside without slippage. Bonus: This helps ensure the screw will be driven in straight and not at an angle.

But inexplicably, Thompson assigned the rights to Henry F. Phillips, the managing director of Oregon Copper Co., according to the Oregon Encyclopedia. With those rights, Phillips went on to license Thompson’s design to screw manufacturers, making him rich off the royalties. So even though he’s not the inventor, Henry Phillips will forever be remembered for the Phillips-head screw and screwdriver.

Of Thompson’s fate, little is known. The encyclopedia cites a 1939 Sunday Oregonian article stating that Thompson had been an auto mechanic when he invented the screw. Beyond that, merely the date of his death remains: Sept. 4, 1940.