Nomi Banday @nomibanday

#NewProfilePic

In consumer marketing, there is a concept called “renting share” with which every assistant product manager is familiar. When a company puts a temporary price promotion on a product (e.g. potato chips, sneakers, furniture), the company will see an increase in purchases pulled forward from future periods as would-be consumers try to lock in the temporary discounted price. When the current and future periods are netted together, one tends to see no change in sales unless there is a permanent change in perception of the product. All evidence so far suggests this will be true of $TSLA ‘s 10-day promotion of Cybertruck after which TSLA will likely raise the price. Once the promotion is over, the heightened interest goes back to normal.

Software engineers realizing the code they’re writing will make them unemployed soon

Anthropic CEO: Software engineering will be completely obsolete in 6-12 months…

@SawyerMerritt They will. The Past predicts the future. It would be limited to new car purchases only.

READ IT TO BELIEVE IT 🚨 TESLA'S LATEST PATENT HAS SOLVED THE BIGGEST ANXIETY—THE HARDWARE OBSOLESCENCE—FOR ALL HW3 and HW4 OWNERS 🔥 If you are driving a Hardware 3 vehicle, you might feel like the clock is ticking on your Full Self-Driving capabilities. If you just bought a shiny new AI4 model, you are likely looking over your shoulder at the rumors of the upcoming AI5, wondering if your cutting-edge car is about to become a legacy artifact. But on January 15, 2026, Tesla quietly published a document that should let every owner breathe a massive sigh of relief. Patent US20260017503A1 is not just a collection of circuit diagrams; it is a technical lifeline. It reveals a sophisticated method for "BIT-AUGMENTED ARITHMETIC CONVOLUTION" that essentially allows Tesla to push future-generation AI models onto current-generation silicon. For the Hardware 3 owners feeling abandoned, this is your consolation prize—proof that Tesla is engineering deep-tech workarounds to keep your car in the game. For the AI4 buyers, this is your safety net, ensuring that the neural networks of the future won't break your investment. This patent confirms that Tesla isn't just relying on new chips to solve autonomy; they are rewriting the rules of math to ensure the car in your driveway today can "think" with the precision of the cars being built tomorrow. 🛑 The problem: The hardware-software mismatch Designing hardware for autonomous vehicles is a brutal balancing act because the product lifecycle of a car spans decades, while the lifecycle of an AI algorithm is measured in weeks or months. When Tesla designs a chip, like the FSD computer, they have to lock in the hardware architecture years before the car hits the road. They might fill the chip with 8-bit multipliers because they are fast, small, and power-efficient. In the world of computing, "bits" are just the amount of detail a computer can handle at once. Think of an 8-bit system like a box of 256 crayons. It is great for basic coloring, but it lacks nuance. However, three years later, the AI team might develop a massive new Transformer model or an Occupancy Network that performs best with 16-bit or 32-bit data. These newer models are like using a box of 65,000 crayons—they can see subtle shades and textures that the 8-bit box simply misses. Suddenly, you have a mismatch. The new software is too "heavy" and detailed for the old hardware, and you cannot physically upgrade the chips in millions of cars already on the road. The traditional solution is to "quantize" the model. This basically means taking that high-definition 16-bit picture and crunching it down to a grainy 8-bit version to fit the chip. But "dumbing down" the data degrades performance and accuracy—something you absolutely cannot afford when the car is trying to distinguish a stop sign from a red balloon at 60 mph. 💡 Tesla's solution: Virtual high-precision arithmetic Tesla’s engineers have devised a brilliant workaround that essentially uses software logic to hack the physical limitations of the hardware. Instead of dumbing down the data, they break the high-precision data into smaller, manageable chunks that the hardware can understand. If the hardware can only eat "bite-sized" 8-bit pieces, but the data comes in a "king-sized" 16-bit format, this system splits the 16-bit data into two 8-bit packages. They call this process "deplaning." The system separates the Most Significant Byte (the big numbers) from the Least Significant Byte (the fine details). Think of this like looking at a price tag of $1,500. The "Most Significant" part is the $1,000 (the big picture), and the "Least Significant" part is the $500 (the specific detail). These separate "planes" are then fed through the standard, lower-precision hardware independently. Once the hardware processes these chunks, the system cleverly stitches the results back together at the end, applying the necessary mathematical shifts to reconstruct a high-precision output. It is like moving a grand piano through a narrow door by taking the legs off first and reassembling it inside. 🔪 The genius of deplaning: Weaponizing the MAC array The genius of this patent lies in how Tesla re-purposes the existing hardware to handle this splitting process. Normally, preparing data—like splitting 16-bit numbers into 8-bit chunks—is a housekeeping task left to the Central Processing Unit (CPU) or a simple Arithmetic Logic Unit (ALU). The problem is that the Neural Network Accelerator is a beast that eats data faster than a CPU can spoon-feed it. If you ask the CPU to split every pixel in a 4K video feed into high and low bytes, the accelerator will starve waiting for the data. Tesla’s solution is to perform this "surgery" using the accelerator itself. They realized that a mathematical convolution is just a fancy way of multiplying and adding. "Convolution" is just the math word for sliding a filter over an image to find patterns, like edges or colors. So, they designed a specific "kernel"—which acts like a digital stencil or cookie cutter—that tricks the convolution engine into acting like a data splitter. The patent details a specific "deplaning" operation that avoids the memory bottleneck entirely via a "stride-two" hack. Imagine a stream of 16-bit raw data entering the chip. Instead of stopping the flow to cut these numbers in half, Tesla runs a 1x2 convolutional kernel over them with a specific stride of two. A "stride" is just how many steps the filter skips as it moves across the data. They define a kernel of [1 0] and another of [0 1]. When the hardware convolves the [1 0] kernel with a stride of two, it mathematically multiplies the first byte (the big number) by 1 and the second byte (the detail number) by 0. The stride of two then forces the window to jump over the next byte, effectively snatching only the "upper" halves of the numbers. A parallel process runs the inverse [0 1] kernel to snatch the "lower" halves. This effectively turns the massive array of multipliers—usually reserved for finding lane lines or stop signs—into a high-speed data shredder, separating the high-precision signal into two digestable streams without ever slowing down the pipeline. 🧮 Doing the math: The four-way cross-multiplication Once the data is split into these lower-precision planes, the system runs the neural network layers on them. This involves convolving weights (the learned parameters of the AI) across the input data. If the weights themselves are also high-precision, they get split up too. The patent describes a reconstruction method that mirrors the "FOIL" method (First, Outer, Inner, Last) you might remember from high school algebra. To multiply a 16-bit input by a 16-bit weight using only 8-bit hardware, the system actually performs four distinct operations: High-byte x High-byte, High-byte x Low-byte, Low-byte x High-byte, and Low-byte x Low-byte. Because the hardware just sees 8-bit integers and doesn't know that a "High-byte" is worth 256 times more than a "Low-byte," the patent specifies a post-processing shift. The results involving a High-byte are conceptually "left-shifted" or multiplied by 256 to restore their magnitude. Think of this like adding zeros to the end of a check. A "1" in the millions place is worth way more than a "1" in the ones place, so the system has to add those zeros back in to make the math work. The system then sums these four partial products in a larger accumulator register to get the final, mathematically perfect high-precision answer. This allows the 8-bit silicon to punch way above its weight class, delivering 16-bit or even 32-bit accuracy by simply running four times as many cheap operations rather than one expensive operation. 💾 Data storage: The logarithmic packing format Perhaps the most specific detail in the patent is how they handle the massive numbers generated by this process. When you stitch two 16-bit operations back together, you often end up with a 32-bit result. However, moving 32-bit numbers around the chip consumes expensive memory bandwidth—it clogs the pipes. Tesla’s solution, described in the later sections of the patent, is a custom data format that compresses this 32-bit value back into a 16-bit container without losing the "meaning" of the data. They utilize a custom floating-point format that drastically skews the balance between the "exponent" and the "mantissa." In scientific notation (like 1.23 x 10^5), the exponent is the "10^5" part that tells you how big the number is, and the mantissa is the "1.23" part that gives you the precision. Standard formats might split the bits evenly, but Tesla’s patent proposes using a massive 10-bit exponent and a tiny 5-bit mantissa. This is a deliberate engineering trade-off. In autonomous driving—specifically for "Occupancy Networks" that map 3D space—knowing the scale of an object (is it 5 meters away or 50?) is often more critical than knowing the precise millimeter detail. By allocating 10 bits to the exponent, the format can represent a gargantuan dynamic range of values, preventing the "overflow" errors that crash standard integer math, while packing the data tight enough to keep the memory bus running cool and fast. ➕ Handling the sign: Zero-padding vs sign-extension The document also touches on the subtle but headache-inducing problem of negative numbers. In binary math, the "sign" of a number (positive or negative) is usually determined by the first bit. When you slice a 16-bit number in half, the lower 8 bits lose their context—the hardware doesn't know if they belong to a positive or negative whole. The patent details a logic flow for "Sign Extension." If the original number was signed (like a velocity vector, which needs to know if the car is moving forward or backward), the system has to intelligently fill the empty bits of the upper planes with ones or zeros to preserve that negative value. Conversely, for "natural numbers" like pixel intensity (which are always positive because you can't have negative light), it uses "Zero Padding." This dynamic switching ensures that the system works equally well for image data (unsigned) and motion vectors (signed), making it a universal accelerator for the entire Full Self-Driving stack. ❄️ Efficiency: Why this beats bigger chips You might wonder why Tesla does not just put bigger, more powerful chips in the cars to begin with. The answer lies in the constraints of an electric vehicle: power and heat. A 16-bit multiplier takes up significantly more silicon area and consumes much more electricity than an 8-bit multiplier. It creates heat that is hard to get rid of. By sticking with the smaller, narrower hardware, Tesla keeps the chip size down and the power consumption low, which preserves battery range and simplifies cooling. This patent allows them to have their cake and eat it too; they get the thermal and physical benefits of efficient, low-precision hardware while retaining the ability to run high-precision, cutting-edge AI models whenever the safety case demands it. 🚀 How this patent contributes to Tesla's now and future This patent is the strategic linchpin that prevents Tesla’s older fleet from becoming obsolete as their AI ambitions grow. Consider the millions of vehicles currently on the road running Hardware 3. These chips were designed in an era when standard Convolutional Neural Networks running 8-bit integer math were the gold standard. However, the industry has rapidly shifted toward massive "End-to-End" Transformers and complex Occupancy Networks that crave the stability and dynamic range of 16-bit or even 32-bit floating-point precision. Without the technology described in this document, Tesla would face a brutal fork in the road: either cap the intelligence of their newest models to accommodate the old hardware, or leave millions of customers behind. This "bit-augmented" breakthrough provides a third option, effectively emulating next-generation precision on current-generation silicon. By unlocking virtual high-precision arithmetic, Tesla unshackles its AI training team. Engineers training models on massive H100 or Dojo clusters can now push for higher fidelity—using that custom 10-bit exponent to perfectly track distinct objects at 300 meters—without worrying that the car’s computer will fail to execute the math. It allows for a unified software stack where a single, sophisticated model architecture can be deployed across Hardware 3, Hardware 4, and the upcoming AI5. The system simply adjusts how many "passes" the hardware makes to achieve the required precision. This not only saves the company billions of dollars by avoiding a logistical nightmare of hardware retrofits but also ensures that the "Full Self-Driving" capability sold years ago can actually be delivered using the cutting-edge transformer models of today and tomorrow.

@ChuckCook I think Tesla should allow the owners who purchased it to be able to transfer to a vehicle till unsupervised FSD with all its features is available in that vehicles HW configuration .

@Tesla_Built @wholemars @elonmusk Assuming Tesla will continue selling cars to the consumer, this is a great demand lever that I believe they will use again for end of quarter push when needed.

@elonmusk Make it $69 monthly or $699 yearly and watch your subscribers double.

Tesla will stop selling FSD after Feb 14. FSD will only be available as a monthly subscription thereafter.

The Evolution of Games from 1970 📹AI x.com/MarchUnofficia…

@farzyness I would buy it.

#FedDecision J.pow is caught in a tough spot given the dual mandate. Inflation is up, employment is down. So I believe it’s analogous to saying our house is on fire and flooded….at the same time.

@elonmusk Sums up technology life cycle 😀

@AskPayPal we are facing issues with using the virtual terminal. Do you know when we should expect the issues to be resolved?

@elonmusk @wholemars It’s already saved my life.

@elonmusk 🎉 Happy Birthday, Elon Musk! 🚀🎂 🤖🔋 Wishing a stellar birthday to the one and only Elon Musk, the real-life Iron Man of the tech world! 🚀🌟 Your ambition, innovation, and out-of-this-world ideas have truly left us all in awe.

Impressions after nearly 600 miles on 11.4.3 with Elon mode (could not get a non-Tesla car to try in time). It went much better than the prior experiment obviously. Many contributing factors. I was not as late so I did not mind as much (still ended up 5 minutes late solely

@elonmusk Holding on to my 2017 Model S till my Cyber Truck reservation will drive out of giga Texas! Excited and can't wait to buy my second Tesla! Version 11.3 is as close to perfect no intervention drives as I have ever experienced before. It's gonna happen! 🤖🚕🌍

@elonmusk Once the balance between the left and the right becomes even, twitter will begin settling down. Twitter will be the place for both parties to spread their messages to their constituents and that's how Twitter will have the opportunity to grow further.

@elonmusk In about 3-4 years, Twitter will be the best PR/marketing for Tesla and all other companies that Elon owns. First is to find out who's been messing with the flow of information and next is to clean the house. Expose the problems and then solve them. When all these are done.