Tesla’s AI Day is here. In a few minutes, Tesla watchers would be seeing executives like Elon Musk provide an in-depth discussion on the company’s AI efforts on not just its automotive business but on its energy business and beyond as well. AI Day promises to be yet another tour-de-force of technical information from the electric car manufacturer. Thus, it is no surprise that there is a lot of excitement from the EV community heading into the event.
Tesla has kept the details of AI Day behind closed doors, so the specifics of the actual event are scarce. That being said, an AI Day agenda sent to attendees indicated that they could expect to hear Elon Musk speak during a live keynote, speak with Andrej Karpathy and the rest of Tesla’s AI engineers, and participate in breakout sessions with the teams behind Tesla’s AI development.
Similar to Autonomy Day and Battery Day, Teslarati would be following along on AI Day’s discussions to provide you with an updated account of the highly-anticipated event. Please refresh this page from time to time, as notes, details, and quotes from Elon Musk’s keynote and its following discussions will be posted here.
Simon 19:40 PT – A question about the use cases for the Tesla Bot was asked. Musk notes that the Tesla Bot would start with boring, repetitive, work, or work that people would least like to do.
Simon 19:25 PT – A question about AI and manufacturing is asked and how it potentially relates to the “Alien Dreadnaught” concept. Musk notes that most of Tesla’s manufacturing today is already automated. Musk also noted that humanoid robots would be done either way, so it would be great for Tesla to do this project, and safely as well. “We’re making the pieces that would be useful for a humanoid robot, so we should probably make it. If we don’t someone else will — and we want to make sure it’s safe,” Musk said.
Simon 19:15 PT – And the Q&A starts. First question involves open-sourcing Tesla’s innovations. Musk notes that it’s pretty expensive to develop all this tech, so he’s not sure how things could be open-sourced. But if other car companies would like to license the system, that could be done.
Simon 19:11 PT – There will really be a “Tesla Bot.” It would be built by humans, for humans. It would be friendly, and it would eliminate dangerous, repetitive, boring tasks. This is still petty darn unreal. It uses the systems that are currently being developed for the company’s vehicles. “There will be profound applications for the economy,” Musk said.
Simon 19:06 PT – New products! A whole Tesla suit?! After a fun skit, Elon says the “Tesla Bot” would eventually be real.
Simon 19:00 PT – What is crazy is that Dojo is not even done. This is just what it is today. Dojo is still evolving, and it is going to be way more powerful in the future. Now, it’s Elon Musk’s turn. What’s next for Tesla beyond vehicles.
Simon 19:00 PT – Venkataramanan teases the ExaPOD. Yet another revolutionary solution from Tesla. With all this, it is evident that Tesla’s approach to autonomy is on a whole other level. It would not be surprising if it takes Wall Street and the market a few days to fully absorb what is happening here.
Simon 18:55 PT – The specs of Dojo are insane. Behind its beastly specs, it seems that Dojo’s full potential lies in the fact that all this power is being used to do one thing: to make autonomous cars possible. Dojo is a pure learning machine, with more than 500,000 training nodes being built together. Nine petaflops of compute per tile, 36 terabytes per second of off-tile bandwidth. But this is just the tip of the iceberg for Dojo.
Simon 18:50 PT – Ganesh Venkataramanan, Project Dojo’s lead, takes the stage. He states that Elon Musk wanted a super-fast training computer to train Autopilot. And thus Project Dojo was born. Dojo is a distributed compute architecture connected by network fabric. It also has a large compute plane, extremely high bandwidth with low latencies, and big networks that are partitioned and mapped, to name a few.
Simon 18:45 PT – Milan Kovac, Tesla’s Director of Autopilot Engineering takes the stage. He notes that he would discuss how neural networks are run in the company’s cars. He notes that Tesla’s systems require supercomputers.
Simon 18:40 PT – Ashok notes that simulations have helped Tesla a lot already. It has, for example, helped the company identify pedestrian, bicycle, and vehicle detection and kinematics. The networks in the vehicles were traded to 371 million simulated images and 480 million cuboids.
Simon 18:35 PT – Ashok notes that these strategies ultimately helped Tesla retire radar from its FSD and Autopilot suite and adopt a pure vision model. A comparison between a radar+camera system and pure vision shows just how much more refined the company’s current strategy is. The executive also touched on how simulations help Tesla develop its self-driving systems. He states that simulations help when data is difficult to source, difficult to label, or in a closed loop.
Simon 18:30 PT – Ashok returns to discuss Auto Labeling. Simply put, there is so much labeling that needs to be done that it’s impossible to be done manually. He shows how roads and other items on the road are “reconstructed” from a single car that’s driving. This effectively allowed Tesla to label data much faster, while allowing vehicles to navigate safely and accurately even when occlusions are present.
Simon 18:25 PT – Karpathy returns to talk about manual labeling. He notes that manual labeling that’s outsourced to third-party firms is not optimal. Thus, in the spirit of vertical integration, Tesla opted to establish its own labeling team. Karpathy notes that in the beginning, that Tesla was using 2D image labeling. Eventually, Tesla transitioned to 4D labeling, where the company could label in vector space. But even this was not enough, and thus, auto labeling was developed.
Simon 18:23 PT – The executive states that traffic behavior is extremely complicated, especially in several parts of the world. Ashok notes that this partly illustrated by parking lots and how they are actually complex. Summoning a car from a parking lot, for example, used to utilize 400k notes to navigate, resulting in a system whose performance left much to be desired.
Simon 18:18 PT – Ashok notes that when driving alongside other cars, Autopilot must not only think about how they would drive, they must also think about how other cars would operate. He shows a video of a Tesla navigating a road and dealing with multiple vehicles to demonstrate this point.
Simon 18:15 PT – Director of Autopilot Software Ashok Elluswamy takes the stage. He starts off by discussing some key problems in planning in both non-convex and high-dimensional action spaces. He also shows Tesla’s solution to these issues, a “Hybrid Planning System.” He demonstrates this by showing how Autopilot performs a lane change.
Simon 18:10 PT – Karpathy’s discussion notes that today, Tesla’s FSD strategy is a lot more cohesive. This is demonstrated by the fact that the company’s vehicles could effectively draw a map in real-time as it drives. This is a massive difference compared to the pre-mapped strategies employed by rivals in both the automotive and software field like Super Cruise and Waymo.
To solve several problems encountered over the last few years with the previous suite, Tesla re-engineered their NN learning from the ground up and utilized a multi-head route, camera calibrations, caching, queues, and optimizations to streamline all tasks.
(heavily simplified) pic.twitter.com/LG2TRgjxip
— Teslascope (@teslascope) August 20, 2021
Simon 18:05 PT – The AI Director discusses how Tesla practically re-engineered their neural network learning from the ground-up and utilized a multi-head route. These include camera calibrations, caching, queues, and optimizations to streamline all tasks. Do note that this is an extremely simplified iteration of Karpathy’s discussion so far.
Simon 18:00 PT – Karpathy covers more challenges that are involved in even the basics of perception. Needless to say, AI Day is quickly proving to be Tesla’s most technical event right off the bat. That said, multi-camera networks are amazing. They’re just a ton of work, but it may very well be a silver bullet for Tesla’s predictive efforts.
Simon 17:56 PT – Karpathy showcases a video of how Tesla used to process its image data in the past. He shows a popular video for FSD that has been shared in the past. He notes that while great, such a system proved to be inadequate, and this is something that Tesla learned when it launched Smart Summon. While per-camera detection is great, the vector space proves inadequate.
Simon 17:55 PT – Karpathy noted that when Tesla designs the visual cortex in its car, the company is modeling it to how a biological vision is perceived by eyes. He also touches on how Tesla’s visual processing strategies have evolved over the years, and how it is done today. The AI Director also touches on Tesla’s “HydraNets,” on account of their multi-task learning capabilities.
Simon 17:51 PT – Karpathy starts off by discussing the visual component of Tesla’s AI, as characterized by the eight cameras used in the company’s vehicles. The AI director notes that AI could be considered like a biological being, and it’s built from the ground up, including its synthetic visual cortex.
Simon 17:48 PT – Elon Musk takes the stage. He apologizes for the event’s delay. He jokes that Tesla probably needs AI to solve these “technical difficulties.” The CEO highlights that AI Day is a recruitment event. He calls Tesla’s head of AI Andrej Karpathy. There’s no better person to discuss AI.
Simon 17:45 PT – We’re here watching the AI Day FSD preview video and we can’t help but notice that… are those Waypoints?!
Simon 17:38 PT – Looks like we’ve got an Elon sighting! And a preview video too! Here we go, folks!
We’ve got an Elon sighting
— Rob Maurer (@TeslaPodcast) August 20, 2021
Simon 17:30 PT – A 30-minute delay. We haven’t seen this much delay in quite a bit.
Simon 17:20 PT – It’s a good thing that Tesla has great taste in music. Did Grimes mix this track?
Simon 17:15 PT – We’re 15 minutes in. “Elon Time” is going strong on AI Day. To be honest, though, this music would fit the “Rave Cave” in Giga Berlin this coming October.
Simon 17:10 PT – A good thing to keep in mind is that AI Day is a recruitment event. Some food for thought just in case the discussions take a turn for the extremely technical. AI Day is designed to attract individuals who speak Tesla’s language in its rawest form. We’re just fortunate enough to come along for the ride.
Tesla Board Member Hiro Mizuno sums it up in this tweet pretty well.
Anybody passionate about real world AI !! https://t.co/ydaWQlkE4O
— HIRO MIZUNO (@hiromichimizuno) August 20, 2021
Simon 17:05 PT – I guess AI Day is starting on “Elon Time?” We’re on to the next track of chill music.
Simon 17:00 PT – And with 5 p.m. PST here, the music is officially live on the AI Day live stream. Looks like we’re in for some wait. Wonder how many minutes it would take before it starts? Gotta love this chill music though.
Simon 16:58 PT – While waiting, I can’t help but think that a ton of TSLA bears and Wall Street would likely not understand the nuances of what Tesla would be discussing today. Will Tesla go three-for-three? It was certainly the case with Battery Day and Autonomy Day.
Made it pic.twitter.com/aAWqxgf0bP
— Johnna (@JohnnaCrider1) August 19, 2021
Simon 16:55 PT – T-minus 5 minutes. Some attendees of AI Day are now posting some photos on Twitter, but it seems like photos and videos are not allowed on the actual venue of the event. Pretty much expected, I guess.
Simon 16:50 PT – Greetings, everyone, and welcome to another Live Blog. This is Tesla’s most technical event yet, so I expect this one to go extremely in-depth on the company’s AI efforts and the technology behind it. We’re pretty excited.
Don’t hesitate to contact us with news tips. Just send a message to tips@teslarati.com to give us a heads up.
Tesla has spent years trying to engineer its way out of the automotive paint shop, one of the most expensive, space-consuming, and environmentally costly steps in vehicle manufacturing. With the Cybercab, Tesla confirmed on X this week that a new reaction injection molding process will embed color directly into the panel itself during production.
“Our new reaction injection molding (RIM) process shrinks Cybercab paint cycles from hours to minutes. This cuts those parts’ manufacturing and supply chain emissions by 35% and eliminating 100% of paint volatile organic compounds (VOCs) emitted in traditional paint methods.” noted Tesla.
While the RIM process isn’t necessarily new and has existed since the 1960s, what makes Tesla’s application notable is how it is being used specifically for exterior body panels that traditionally required a separate paint process after forming.
Tesla’s RIM approach integrates the color directly into the panel material during the molding process itself. The pigment is part of the polymer mix injected into the mold, meaning the panel comes out of the mold already colored, with no separate paint application required. The clear coat or protective layer can be applied at the mold stage or through a much faster post-process than traditional multi-stage painting. Tesla claims this compresses what was a multi-hour paint cycle into minutes per panel.
Tesla’s obsession with killing the paint shop is one of the most consistent threads running through the company’s manufacturing philosophy going back years. As far back as 2018, Musk was trimming paint color options to simplify production, tweeting at the time: “Moving 2 of 7 Tesla colors off menu on Wednesday to simplify manufacturing.” Two years later, in a 2020 Automotive News interview, Musk laid out his broader vision, saying he believed Tesla factories could one day be 1,000 times more efficient than conventional plants, and pointing to the paint shop as one of the biggest sources of waste, cost, and complexity. The Cybertruck was the most extreme expression of that thinking. Tesla chose an unpainted stainless steel exterior partly because it would eliminate the need for a $200 million paint facility at Gigafactory Texas. The stainless approach proved harder and more expensive than anticipated, but the underlying ambition never changed. The Cybercab is what happens when that same ambition meets a manufacturing process that delivers on it.
Lifestyle
Tesla app update makes Robotaxi ownership make a lot more sense
Tesla’s app now shows a live indicator when your car is actively driving itself.
A recent Tesla app update, released last week (4.58.5), gives visibility on whether a vehicle is navigating in its semi-autonomous mode or being drive by a human driver. The updated app now displays a live “Self-Driving” indicator in bright blue text directly beneath the vehicle’s speed readout whenever Full Self-Driving is actively engaged, along with the signature glowing blue navigation path that FSD users see on the main touchscreen. It is a small visual update with meaningful implications for how Tesla owners monitor their vehicles remotely.
The feature was first spotted in the wild by X user Jordan Camina, who shared video of a Hardware 3 Model S displaying the new animation through the app while driving. That detail is significant because it confirms the update is not limited to newer HW4 vehicles. It works across hardware generations, and Tesla confirmed it will eventually support all vehicles regardless of chip platform once both the app and vehicle software are updated. The vehicle side requires software version 2026.20.6.1, which has reached nearly 40% of the fleet so far, as monitored by NotaTeslaApp.
The feature makes the most practical sense when viewed through the lens of Tesla’s expanding robotaxi operation. In a robotaxi context, the owner of a vehicle generating ride revenue has a direct financial and safety interest in knowing whether their car is operating under autonomous control at any given moment. The app’s new FSD indicator gives fleet owners exactly that visibility, the same way a logistics company monitors whether a delivery driver is following the planned route. It also carries implications for Tesla’s insurance model. Tesla’s own insurance product prices premiums in part based on FSD engagement rates, and real-time visibility into when FSD is active creates a feedback loop that could eventually tie directly into policy pricing. For individual owners who have opted their personal vehicles into the robotaxi network, the update effectively turns the Tesla app into a fleet management dashboard, one that tells you whether your car is earning money, whether it is driving itself to do it, and whether everything is operating the way it should from wherever you happen to be.
Tesla expands Robotaxi to Florida, marking its third state for autonomy
As Teslarati has reported, Tesla launched unsupervised robotaxi rides in Miami this summer, a milestone that makes a remote FSD status indicator significantly more practical than a cosmetic feature. When a vehicle is operating as a robotaxi without a driver present, the owner or fleet operator needs a reliable way to confirm autonomy is engaged. The app now provides exactly that.
As noted by NotATeslaApp, The update also arrived alongside a hint buried in the same app version that Tesla plans to use the cabin camera to verify driver identity before FSD can be activated. Pairing identity verification with a live autonomy status indicator points toward the infrastructure Tesla is building for a fleet of driverless vehicles that owners can monitor the way you would track a package delivery.
Elon Musk
California snubs Tesla in its newly passed EV incentive that favors Rivian and Lucid
California passed a $135 million EV incentive that rewards Rivian and Lucid while sidelining Tesla
California just drew a line in the EV incentive sand to put Tesla on the wrong side of it. The state recently passed a $135 million program offering first-time electric vehicle buyers a direct incentive with no application required, but the rules were written in a way that leaves Tesla at a structural disadvantage compared to Rivian and Lucid.
The program caps eligible vehicles at $50,000 for new EVs and $25,000 for used ones. That pricing threshold rules out a significant portion of Tesla’s lineup, though some lower-priced Model 3 and Model Y configurations would still qualify. California-based automakers are exempt from the price cap entirely, regardless of what their vehicles cost. Rivian, headquartered in Irvine, and Lucid, based in the San Francisco Bay Area, both benefit from that exemption. Rivian’s R2 starts at roughly $45,000 but has versions above the cap. Lucid’s Air and Gravity start at $70,990 and $79,990 respectively, well above any threshold a non-California company would face.
California hits Tesla Cybercab and Robotaxi driverless cars with new law
Tesla built its reputation and a significant portion of its early market share in California, where EV adoption has consistently led the nation. The company operates its original factory in Fremont, California, and the state was home to Tesla’s headquarters for most of its existence. That changed in 2021 when Tesla moved its corporate headquarters to Austin, Texas. Since then, the relationship between the company and California Governor Gavin Newsom has been openly adversarial, with Musk and Newsom trading public criticism on multiple occasions.
California’s EV incentive landscape has shifted repeatedly in recent years, and Tesla has previously lost eligibility for state-level programs as its vehicles exceeded income-adjusted price thresholds. The federal $7,500 EV tax credit, which Tesla models have qualified for and lost depending on policy cycles, is no longer available after it expired without renewal, making state-level programs more meaningful to buyers than they have been in years.
The practical impact for buyers is more nuanced than the headline suggests. California residents purchasing a Tesla under $50,000 for the first time can still access the incentive. But the exemption written for California-based manufacturers is a structural advantage that rewards where a company plants its headquarters flag rather than where it builds its products, and Tesla moved that flag to Texas.
Tesla FSD is about to know your specific house and neighborhood better than any map
The secret behind Tesla’s Cybercab Gold goes well beyond just the color
