RoBTC 🕳🐇

HIDDEN GEM 💎 Tesla has patented a brilliant two-speed brain that lets AI keep up with every camera frame 🔥 A car camera can deliver a fresh picture of the road every 33 milliseconds. The most accurate AI models can need longer than that to study just one, and the road never pauses while they think. That gap creates a real dilemma. Automotive cameras typically run at 30 frames per second or more, while heavyweight deep learning models may manage only 20 or fewer. For a car that understands the world through its cameras, every frame a model can't get to is valuable information captured and then left unused. Missed frames aren't the only cost. A slow model might spend 200 milliseconds finding the objects in an image, and by then everything in that image has likely moved. Its answer is precise, but it describes a moment that has already passed. Engineers can swap in a lighter, faster model instead. That one keeps pace with the camera, but it pays in accuracy. It becomes more likely to miss a vehicle that is really there or to report one that isn't. Tesla's design refuses to pick a side. It runs two models together, each at the speed that suits it, and wires them so the slow one regularly feeds the fast one. The first model is the deep thinker. It's a heavyweight neural network that reads only some of the frames, perhaps every second or every fifth one. It takes its time, like a meticulous mapmaker, to detect objects and pin down exactly where they are. The second model is the sprinter. It reads frames at the camera's full rate, and it doesn't have to start from scratch. Every time the mapmaker finishes, its result joins the newest image as part of the sprinter's input. That handoff is the heart of the invention. The sprinter holds a precise but slightly old map in one hand and a brand-new photo in the other. Its job is to work out how far everything has shifted and update each object's location to match the present. One version does this with motion flow. The quick model compares the older frame with the newer one and measures how far each pixel has moved. It's like comparing two snapshots of a busy crosswalk to see how far every person has stepped. Next, it averages that motion for each object. It then slides the object's box by that amount to estimate where the object sits right now. The loop keeps turning. When the careful network finishes its next frame, a fresh, accurate map lands with the sprinter, which recalibrates and keeps running. Any drift in the sprinter's estimates gets corrected every time a new careful answer arrives. The sprinter can take other forms too. It can be a support vector machine, a lean classic machine learning tool, or a second neural network. The all-neural version runs on the same rhythm. A careful network needing 80 milliseconds teams up with a fast one needing just 15. The fast network takes the careful prediction from five frames back and refreshes it with the newest frame, producing sharper locations than it could find on its own. The two can even share hardware gracefully. The heavy network can run in the moments when the quick model is idle, or pause between blocks of its layers whenever a new frame needs attention. Then it picks up right where it left off, like a chef who steps away from a slow-simmering sauce to plate a quick order and comes straight back. The payoff shows up in the numbers. In one worked example, a tracker that answers in 5 milliseconds partners with a detector that needs 50, while the camera produces a frame every 33. Together they deliver results at 30 frames per second with just 5 to 20 milliseconds of latency. The detector working alone manages only 15 frames per second at 50 milliseconds. Keeping the deep thinker in the loop can also offer better accuracy than simply shrinking one network down to that speed. Every 33 milliseconds, that camera hands over a fresh picture of the road. The sprinter can now read each one the moment it lands, armed with everything the deep thinker has already worked out. Speed no longer costs accuracy. Accuracy no longer shows up late. The road still never pauses for anyone. With a two-speed brain, the AI no longer needs it to.

Did you know that beating Uber, Lyft and Waymo is too small a goal for Tesla? And one number shows why 🔥 $TSLA’s Robotaxi is growing fast. Cybercabs have been giving paid rides in Austin since September, Dallas follows next month, and unsupervised miles grew eightfold last quarter. So naturally, most people judge Tesla against those three rivals. The debate is usually about who has more cars, cheaper fares and more cities. That’s the wrong scoreboard! The better question is how many miles Americans travel by car each year, and how many of those miles happen in a ride-hailing car at all. The answer is surprisingly lopsided. For every mile Americans ride in an Uber, Lyft, Waymo or any other ride-hailing car, they drive about 70 miles in their own. That’s the number that matters: 70 🔥 Picture all the miles Americans travel by car as a pizza cut into 71 slices. Ride-hailing gets one slice. The other 70 are driven by people behind the wheel of their own cars. So the market Uber, Lyft and Waymo fight over is only a little over 1 percent of American driving. The real prize is everything else, and that's the market Tesla is going after. Once you see that, a lot of Tesla's choices suddenly make sense: the 30-cents-per-mile target, the giant Cybercab production plans, and Elon's claim that Tesla will end up carrying most of the world's passenger miles. 🧮 WHERE THE 70 COMES FROM America's cars, SUVs and pickups covered about 2.9 trillion miles in 2023, according to the Federal Highway Administration. That's 259 million vehicles averaging about 11,100 miles each. A 70x gap puts ride-hailing at roughly 40 billion miles a year. Uber and Lyft's own commissioned study fits that scale. Even in six major metro areas where both services are popular, including San Francisco and Boston, the two companies made up just 1 to 3 percent of all miles driven. In other words, even in its strongest cities, ride-hailing only gets one or two slices of the pizza. Ride-hailing changed how we get to the airport. It barely touched how America actually gets around. 💸 WHY RIDE-HAILING ONLY GETS ONE SLICE The reason is price. A typical 5-mile UberX ride costs about $15 to $22 in most US cities this year. With fees included, that works out to roughly $3 to $4.40 per mile. AAA's 2026 study puts the full cost of owning a new car at about 86 cents per mile. That figure counts depreciation, financing, fuel and maintenance. When riding costs three to five times more than owning, people buy cars. They save ride-hailing for the airport, the bar and the occasional emergency. A big share of every fare pays the human behind the wheel, and that cost doesn't shrink with scale. ⚡ HOW 30 CENTS UNLOCKS THE OTHER 70 Tesla's target for its Robotaxi fleet is 30 cents per mile. On last week's Q3 call, CFO Vaibhav Taneja said Tesla isn't there yet but sees a clear path as the fleet scales. The trick isn't only removing the driver. The average American car drives about 30 miles a day and sits parked the rest of the time. A robotaxi working most of the day could plausibly cover 250 miles or more. That spreads the cost of one vehicle over eight or nine times as many miles, and that is how the price per mile collapses. At 30 cents, riding becomes cheaper than owning, even with a healthy margin on top. That's when the other 70 slices finally come within reach. This isn't about replacing the Tesla already in your garage, since each extra mile in a car you own costs very little. Elon said on the call that the Model 3, Model Y and Roadster remain the cars people will want to own, while Cybercab is mostly Tesla-owned transport-as-a-service. The real fight is over the next car someone decides whether to buy, especially the second one. 🏭 ONE SLICE VS. 70 SLICES This is where the number gets wild. Suppose each robotaxi covers about 100,000 miles a year. Then the one slice ride-hailing serves today would need only around 400,000 vehicles. On the Q3 call, Tesla said its Cybercab line in Texas has a listed capacity of more than 125,000 a year, and Lars Moravy added that the line can actually do significantly more. Even at the listed rate, that single line could build enough cars to cover all of today's US ride-hailing in about three years. The other 70 slices would take closer to 29 million vehicles. At that same rate, one line would need more than two centuries to get there. Tesla's factory ambitions only make sense against that bigger market. Building a Cybercab line just to fight Uber would be like building a stadium for a dinner party. 🔮 THE REAL TARGET On the call, Elon said Tesla will end up doing a majority of all passenger miles on Earth wherever it's allowed to operate freely. Read next to the 70x gap, that isn't a ride-hailing prediction. Ride-hailing holds just one slice out of 71, so a majority can only come from the miles people drive themselves today. For a century, owning a car has been the cheapest way to get door to door on your own schedule. At 30 cents a mile, that would no longer be true 🔥