At the 2025 Consumer Electronics Show, Nvidia showed off its new consumer graphics cards, home-scale compute machines, and commercial AI offerings. One of these offerings included the new Nvidia Cosmos training system.
Nvidia is a close partner of Tesla - in fact, they produce and supply the GPUs that Tesla uses to train FSD - the H100s and soon-to-be H200s, located at the new Cortex Supercomputing Cluster at Giga Texas. Nvidia will also challenge Tesla’s lead in developing and deploying synthetic training data for an autonomous driving system - something Tesla is already doing.
However, this is far more important for other manufacturers. We’re going to take a look at what Nvidia is offering and how it compares to what Tesla is already doing. We’ve done a few deep dives into how Tesla’s FSD works, how Tesla streamlines FSD, and, more recently, how they optimize FSD. If you want to get familiar with a bit of the lingo and the background knowledge, we recommend reading those articles before continuing, but we’ll do our best to explain how all this synthetic data works.
Nvidia Cosmos
Nvidia’s Cosmos is a generative AI model created to accelerate the development of physical AI systems, including robots and autonomous vehicles. Remember - Tesla’s FSD is also the same software that powers their humanoid robot, Optimus. Nvidia is aiming to tackle physical, real-world deployments of AI anywhere from your home, your street, or your workplace, just like Tesla.
Cosmos is a physics-aware engine that learns from real-world video and builds simulated video inputs. It tokenizes data to help AI systems learn quicker, all based on the video that is input into the system. Sound familiar? That’s exactly how FSD learns as well.
Cosmos also has the capability to do sensor-fused simulations. That means it can take multiple input sources - video, LiDAR, audio, or whatever else the user intends, and fuse them together into a single-world simulation for your AI model to learn from. This helps train, test, and validate autonomous vehicle behavior in a safe, synthetic format while also providing a massive breadth of data.
Data Scaling
Of course, Cosmos itself still requires video input - the more video you feed it, the more simulations it can generate and run. Data scaling is a necessity for AI applications, as you’ll need to feed it an infinite amount of data to build an infinite amount of scenarios for it to train itself on.
Synthetic data also has a problem - is it real? Can it predict real-world situations? In early 2024, Elon Musk commented on this problem, noting that data scales infinitely both in the real world and in simulated data. A better way to gather testing data is through real-world data. After all, no AI can predict the real world just yet - in fact, that’s an excellent quantum computing problem that the brightest minds are working on.
Yun-Ta Tsai, an engineer at Tesla’s AI team, also mentioned that writing code or generating scenarios doesn’t cover what even the wildest AI hallucinations might come up with. There are lots of optical phenomena and real-world situations that don’t necessarily make sense in the rigid training sets that AI would develop, so real-world data is absolutely essential to build a system that can actually train a useful real-world AI.
Tesla has billions of miles of real-world video that can be used for training, according to Tesla’s Social Media Team Lead Viv. This much data is essential because even today, FSD encounters “edge cases” that can confuse it, slow it down, or render it incapable of continuing, throwing up the dreaded red hands telling the user to take over.
Cosmos was trained on approximately 20 million hours of footage, including human activities like walking and manipulating objects. On the other hand, Tesla’s fleet gathers approximately 2,380 recorded minutes of real-world video per minute. Every 140 hours - just shy of 6 days - Tesla’s fleet gathers 20 million hours of footage. That was a little bit of back-of-the-napkin math, calculated at 60 mph as the average speed.
Generative Worlds
Both Tesla’s FSD and Nvidia’s Cosmos can generate highly realistic, physics-based worlds. These worlds are life-like environments and simulate the movement of people and traffic and the real-life position of obstacles and objects, including curbs, fences, buildings, and other objects.
Tesla uses a combination of real-world data and synthetic data, but the combination of data is heavily weighted to real-world data. Meanwhile, companies who use Cosmos will be weighting their data heavily towards synthetically created situations, drastically limiting what kind of cases they may see in their training datasets.
As such, while generative worlds may be useful to validate an AI quickly, we would argue that these worlds aren’t as useful as real-world data to do the training of an AI.
Overall, Cosmos is an exciting step - others are clearly following in Tesla’s footsteps, but they’re extremely far behind in real-world data. Tesla has built a massive first-mover advantage in AI and autonomy, and others are now playing catch-up.
We’re excited to see how Tesla’s future deployment of its Dojo Supercomputer for Data Labelling adds to its pre-existing lead, and how Cortex will be able to expand, as well as what competitors are going to be bringing to the table. After all, competition breeds innovation - and that’s how Tesla innovated in the EV space to begin with.
Subscribe
Subscribe to our newsletter to stay up to date on the latest Tesla news, upcoming features and software updates.
The Super Manifold is Tesla’s solution to reducing the complexity of a heat pump system for an EV. Tesla showed off its engineering chops back with the original Model Y in 2019, where it introduced a new 8-way valve (the Octovalve) and a new heat pump alongside the uniquely designed Super Manifold to improve efficiency.
Now, Tesla is launching an improved version with the refreshed Model Y - the Super Manifold V2. We got to hear about it thanks to Sandy Munro’s interview with Tesla’s Lars Moravy (Vice President of Vehicle Engineering) and Franz Von Holzhausen (Chief of Vehicle Design). You can watch the video further below.
What Is The Super Manifold?
The Super Manifold (get it, Superman?), is an all-in-one package that brings in all the components of a heat pump system into one component. The Super Manifold packs all the refrigerant and coolant components around a 2-layer PCB (printed circuit board).
This Super Manifold would normally have 15 or 20 separate components, but Tesla managed to integrate them all into one nice package. That presented Tesla with a new challenge: how to integrate a heat pump—capable of both heating and cooling—into a single, efficient platform?
Several years ago, Tesla designed the Octovalve. It combines inlets and outlets and can variably change between heating or cooling on the fly - without needing to be plumbed in different directions. This is especially important for EVs, which may need to heat the battery with the waste heat generated from the motors or the heat pump while also cooling the cabin - or vice versa.
Original Super Manifold V1.1
Tesla launched the Super Manifold V1.1 back in 2022, and it provided some minor improvements to the waste heat processing of the heat exchange system. It also tightened up the Octovalve, preventing the leakage of oils into the HVAC loop that could cause it to freeze at extremely low temperatures.
Tesla has been using the V1.1 for several years now, and it has really solved the vast majority of issues with the heat pump system that many older Model Ys experienced.
Super Manifold V2 Coming Soon
Now, Tesla is introducing the Super Manifold V2 in the new Model Y. It will improve the overall cooling capacity provided by the original Super Manifold, but unfortunately, not every single new Model Y will come with it equipped. Tesla will be introducing it slowly across the lineup and at different rates at different factories, depending on part availability.
Eventually, the Super Manifold V2 will also make its way to other vehicles, potentially including the upcoming refresh for the Model S and Model X, but initially, it’ll be exclusive to the new Model Y. Tesla expects to have the new manifold in every new Model Y later this year.
If you’re interested in checking out the whole video, we’ve got it for you below.
Mark Rober, of glitter bomb package fame, recently released a video titled Can You Fool A Self-Driving Car? (posted below). Of course, the vehicle featured in the video was none other than a Tesla - but there’s a lot wrong with this video that we’d like to discuss.
We did some digging and let the last couple of days play out before making our case. Mark Rober’s Wile E. Coyote video is fatally flawed.
The Premise
Mark Rober wanted to prove whether or not it was possible to fool a self-driving vehicle, using various test scenarios. These included a wall painted to look like a road, low-lying fog, mannequins, hurricane-force rain, and bright beams.
All of these individual “tests” had their own issues - not least because Mark didn’t adhere to any sort of testing methodology, but because he was looking for a result - and edited his tests until he was sure of it.
Interestingly, many folks on X were quick to spot that Mark had been previously sponsored by Google to use a Pixel phone - but was using an iPhone to record within the vehicle - which he had edited to look like a Pixel phone for some reason. This, alongside other poor edits and cuts, led many, including us, to believe that Mark’s testing was edited and flawed.
Flaw 1: Autopilot, Not FSD
Let’s take a look at the first flaw. Mark tested Autopilot - not FSD. Autopilot is a driving aid for lane centering and speed control - and is not the least bit autonomous. It cannot take evasive maneuvers outside the lane it is in, but it can use the full stable of Tesla’s extensive features, including Automatic Emergency Braking, Forward Collision Warnings, Blind Spot Collision Warnings, and Lane Departure Avoidance.
On the other hand, FSD is allowed and capable of departing the lane to avoid a collision. That means that even if Autopilot tried to stop and was unable to, it would still impact whatever obstacle was in front of it - unlike FSD.
As we continue with the FSD argument - remember that Autopilot is running on a 5-year-old software stack that hasn’t seen updates. Sadly, this is the reality of Tesla not updating the Autopilot stack for quite some time. It seems likely that they’ll eventually bring a trimmed-down version of FSD to replace Autopilot, but that hasn’t happened yet.
Mark later admitted that he used Autopilot rather than FSD because “You cannot engage FSD without putting in a destination,” which is also incorrect. It is possible to engage FSD without a destination, but FSD chooses its own route. Where it goes isn’t within your control until you select a destination, but it tends to navigate through roads in a generally forward direction.
The whole situation, from not having FSD on the vehicle to not knowing you can activate FSD without a destination, suggests Mark is rather unfamiliar with FSD and likely has limited exposure to the feature.
Let’s keep in mind that FSD costs $99 for a single month, so there’s no excuse for him not using it in this video.
Flaw 2: Cancelling AP and Pushing Pedals
Many people on X also followed up with reports that Mark was pushing the pedals or pulling on the steering wheel. When you tap on the brake pedal or pull or jerk the steering wheel too much, Autopilot will disengage. For some reason, during each of his “tests,” Mark closely held the steering wheel of the vehicle.
This comes off as rather odd - at the extremely short distances he was enabling AP at, there wouldn’t be enough time for a wheel nag or takeover warning required. In addition, we can visibly see him pulling the steering wheel before “impact” in multiple tests.
Over on X, techAU breaks it down excellently on a per-test basis. Mark did not engage AP in several tests, and he potentially used the accelerator pedal during the first test - which means that Automatic Emergency Braking is overridden. In another test, Mark admitted to using the pedals.
Flaw 3: Luminar Sponsored
This video was potentially sponsored by a LiDAR manufacturer - Luminar. Although Mark says that this isn’t the case. Interestingly, Luminar makes LiDAR rigs for Tesla - who uses them to test ground truth accuracy for FSD. Just as interesting, Luminar’s Earnings Call was also coming up at the time of the video’s posting.
Luminar had linked the video at the top of their homepage but has since taken it down. While Mark did not admit to being sponsored by Luminar, there appear to be more distinct conflicts of interest, as Mark’s charity foundation has received donations from Luminar’s CEO.
Given the positivity of the results for Luminar, it seems that the video had been well-designed and well-timed to take advantage of the current wave of negativity against Tesla, while also driving up Luminar’s stock.
Flaw 4: Vision-based Depth Estimation
The next flaw to address is the fact that humans and machines can judge depth using vision. On X, user Abdou ran the “invisible wall” through a monocular depth estimation model (DepthAnythingV2) - one that uses a single image with a single angle. This fairly simplified model can estimate the distance and depth of items inside an image - and it was able to differentiate the fake wall from its surroundings easily.
Tesla’s FSD uses a far more advanced multi-angle, multi-image tool that stitches together and creates a 3D model of the environment around it and then analyzes the result for decision-making and prediction. Tesla’s more refined and complex model would be far more able to easily detect such an obstacle - and these innovations are far more recent than the 5-year-old Autopilot stack.
While detecting distances is more difficult in a single image, once you have multiple images, such as in a video feed, you can more easily decipher between objects and determine distances by tracking the size of each pixel as the object approaches. Essentially, if all pixels are growing at a constant rate, then that means it’s a flat object — like a wall.
Case in Point: Chinese FSD Testers
To make the case stronger - some Chinese FSD testers took to the streets and put up a semi-transparent sheet - which the vehicle refused to drive through or drive near. It would immediately attempt to maneuver away each time the test was engaged - and refused to advance with a pedestrian standing in the road.
Would FSD hit a transparent film wall? This test showed it just avoids it.
Thanks to Douyin and Aaron Li for putting this together, as it makes an excellent basic example of how FSD would handle such a situation in real life.
Flaw 5: The Follow-Up Video and Interview
Following the community backlash, Mark released a video on X, hoping to resolve the community’s concerns. However, this also backfired. It turned out Mark’s second video was of an entirely different take than the one in the original video - this was at a different speed, angle, and time of initiation.
Mark then followed up with an interview with Philip DeFranco (below), where he said that there were multiple takes and that he used Autopilot because he didn’t know that FSD could be engaged without a destination. He also answered here that Luminar supposedly did not pay him for the video - even with their big showing as the “leader in LiDAR technology” throughout the video.
Putting It All Together
Overall, Mark’s video was rather duplicitous - he recorded multiple takes to get what he needed, prevented Tesla’s software from functioning properly by intervening, and used an outdated feature set that isn’t FSD - like his video is titled.
Upcoming Videos
Several other video creators are already working to replicate what Mark “tried” to test in this video.
To get a complete picture, we need to see unedited takes, even if they’re included at the end of the video. The full vehicle specifications should also be disclosed. Additionally, the test should be conducted using Tesla’s latest hardware and software—specifically, an HW4 vehicle running FSD v13.2.8.
In Mark’s video, Autopilot was engaged just seconds before impact. However, for a proper evaluation, FSD should be activated much earlier, allowing it time to react and, if capable, stop before hitting the wall.
A wave of new videos is likely on the way—stay tuned, and we’ll be sure to cover the best ones.
![Tesla Updates Energy App in Update 2025.8.3: Adds Reset and Distance Filter [VIDEO]](https://www.notateslaapp.com/img/containers/article_images/2025/updated_energy_app.jpg/979943fb17df5ae4e9fa668b8cd61fca/updated_energy_app.jpg)
_300w.png)
![A Look at the Tesla Cybertruck’s Crumple Zones [VIDEO]](https://www.notateslaapp.com/img/containers/article_images/cybertruck/cybertruck-crash-test.webp/e43ef2433bd6e44ae8db63421604a4fb/cybertruck-crash-test.jpg)
