Most OEMs cannot even make money manufacturing EVs, you add on the cost for Nvidia’s chips and the training cost how are they going to compete with Tesla?
Hans missed the greatest need for real world data: verification. Elon has stated a few times that (even with all their data) one of their biggest challenges is finding enough examples of the remaining corner cases, and verifying in the real world that FSD can handle them reliably.
@@jjj32753 For proper validation you absolutely need real data. No simulation is perfect. You cannot validate safety critical products using simulation, and there is no way authorities would (or should) accept simulation performance stats for satisfying self driving regulations.
@@jjj32753 AI is great at generating good models of reality but not perfect. You need real data to check it against to keep flaws in the generated data from running rampant. Think of the errors GPT makes and imagine basing your FSD on that.
Jensen's entire talk -- NVIDIA's entire approach -- is a validation of Tesla's tech development of FSD and Optimus. But, Wall Street being what it is, is raising NVIDIA's value and decreasing Tesla's. 'Analysts' are truly clueless.
Yes, but give it time and eventually Wall Street will get on board in its entirety. Remember when nobody thought the EV shift was even in the cards? Neither mfgs nor most analysts made the change in direction quickly enough so most are desperate to catch up, and now a good few have. I am saying you are spot on about the validation!!!!!!!!!!!! In a reasonable market probably increasing nvidia valuation while maintaining tsla somewhat baked in valuation would be a better response. Tsla also suffered from stagflation threat news today 😅
This is what will make VR a reality. Since playing Minecraft I knew that we could never code the worlds of VR but needed the computer to generate them.
This is a great explanation of real world data vs simulation data. It's so important for retail investors to understand what is going on with in the autonomous driving space, but most of us don't have the experience to parse these presentations. Thanks guys!
I agree that none of the partners has a vertically integrated stack for producing an SDV using Cosmos & Thor. It will take several years for them to get to where Tesla was 2 years ago. My two biggest asset positions have been Tesla & Nvidia since i started my portfolio precisely because it was clear to me then that Tesla was Apple & Nvidia was Google Android in the race to autonomy. Cosmos fully validates that thesis.
It’s simple - can you use flight simulators exclusively to train pilots. The answer is no. It helps, it allows training in scenarios that can’t be recreated safely or budget wise in real world, but it is not a replacement for actually flying a plane. Not even close. Companies will need a lot of real data and huge compute. This is a brilliant sales strategy to cause some companies to buy compute but they’ll quickly realize the compute they purchased is useless without real world data and even when you have the data you need experienced engineers to build the pipelines and learn how to build the system that trains the system.
The question is not of Cosmos synthetic data is better than Tesla’s real world data, the question is if the Cosmos synthetic data trained autonomous driving cars can be better than human driver enough to be robotaxi
Power requirements for the Hardware Nvidia is going to use is too high. Tesla FSD version 3 only uses about 100w, while Nvidia hardware is probably going to be higher. Nvidia is trying to copy Tesla on their Real World model.....Nvidia won't iterate fast enough and will struggle to keep up with Tesla.
Yes. But it will cost lots more compute which conveniently Nvidia has a whole business selling. Nvidia is happy selling compute, models & inference boards to whoever wants to build robots. They don't need them to be profitable as long as they buy. That's like Google Android. They don't care if Motorola profits from their phones but with each phone it's a new user using Google apps, viewing ads... Using Google search. They get the eye candy without having to manufacturer the frame it's delivered from (save for their own pads & pixel phones)
@@DavidSaintloth Actually, this is the first time Nvidia started selling an Inference Chip/board. Tesla built their own inference chip and the board to go with it first. Nvidia have been selling GPU chips, and system with their software for training which is a general type hardware with their own software.
The minimum inference computer requirements for autonomous driving are: Dual processors for redundancy. Video inputs from all the cameras. Inputs from the driver. Inputs and control of the battery/passenger thermal system. Control outputs for window up/down, breaking, acceleration steering and etc. An INVIDA processor alone is not sufficient.
One possible outcome could also be, that the legacy automakers give their fate into Nvidia's hands but lose in the end time, as they still have to figure the full self driving out on their own, as Nvidia only gives them the tool to maybe do so. Just licensing from Tesla would be the very safe step. Very interesting times, but for sure there is some competition in sight. For Tesla all depends on the fast Robotaxi rollout, because in the end the deciding factor is time of market entry and cost per mile and there Tesla is very well positioned.
Its great that we all understand the diffrence between the brittle artificial data vs the real life data but will hedge funds or regular investors be cognizant of these difffrences ? 😢😢
Like that famous political line from James Carvel, “It’s the total car cost stupid”. Those who produce an autonomous car for the lowest cost and therefore stay ahead of cost per mile, wins. Slowly coming down from current $2-4 / mile to cents just staying ahead of other competitors until they bleed money. Much like Tesla did in China thru price cuts. Why an OEM would choose Tesla OS over NVidia? BECAUSE that will allow them to buy all the modules (48V) from Teslas suppliers and directly as well as how to manufacture their product for less cost.
I upgraded to hw3 in our 2018 model 3 for 1,000. Even if you add the other things, it still sounds like a 3k system for other oems will be more expensive that what tesla spends. Plus the cars need to be able to do ota updates which they still can not do and have not been able to do. Plus the car as a whole needs to be more centralized. It is still going to take these oems years to implement this. Then the training will be years behind at that point even if it can improve at any rate like tesla did without any real world data or gathering from the fleet. They will only be starting at this point years later with data collection. The other oems will still be screwed for a very long time to come. They will probably be bankrupt by then and with robotaxies, the overall fleet will shrink and people will need less of their cars. It will not become viable for them to do all of this profitably. They are screwed either way.
I don't think Elon minds, he gets on well with Jenson. Anyway if you are doing the same thing it's going to look similar. If you ask two blacksmiths to make a Bowie knife they'll look similar even if they never met eachother
@@andrewsaint6581 Who started the EV revolution? Tesla, and Elon. Who started the Robot Revolution? tesla, and Elon. Who started Self Driving? Elon and Tesla. who started Rocket company? Elon and SpaceX. Who started Battery Storage on a Big scale? Elon and Tesla. Everybody else FOLLOWED. Yeah, I know....Some stuff what I stated had others doing it, but on a very small scale. It is Elon who started and got it going on a BIG scale, and is successful at it. I could name few more Elon started on a big scale....
You would need multiple simulations each with its own speciality, from animals, cars, people, activities, weather, emergencies etc. All deployed on a single platform, with limited coordination between each SIM. Then test FSD on this highly unpredictable environment.
Ai can hallucinate just hand a car with all kinds of different sensors and the physics of the car and what is going to happen with rain,fog slate etc. Plus the mass of the cars. Especially different model of cars.
the idea is that if a human can do it, a robot can eventually be designed to do it better - more efficiently, faster response and safer, and far more consistently than a human ability average. The robot won't get tired after driving longer so it won't have stats reductions like a human would, especially important for say, icy road semi-truck drivers
Now the question is whether Tesla can use its huge real-world and correct data to generate much much more synthetic data to deal with almost 100% of all edge scenarios and make Tesla FSD 100% perfect!
so why is it that gm build s10 electric cars, instead of just one really good one, that they can make at scale and speed and maybe get to break even on it?
This doesn’t mean that Tesla has no lead, but if synthetic data can get other car companies to the point of Supervised FSD it’s a huge help. They may not get to robotaxi, but could get to where Tesla is now And yes they will all share their real world video data because it will be a requirement of their contract with nvidia
I wonder if you could use insurance company data based on accidents to generate edge case simulations and then create a multitude of conditional variations to cover for inaccuracies in the data or adjacent edge cases.
Real data is very important as a source of simu data and verification of simu performance. But simu data can accelerate learning dramatically. For example, you drive from one city to another aparting 1000km, it may just have limited situatios that needs special care, which just needs a few miliseconds or a few seconds for handling each special situation. most of time it just drives smoothly, no self-training while in a smooth situation. However, with simu data, unlimited special situations can be generated. Number of miles is not the key, number of special situations is the key. Using simu data, any special situation can be generated along the way, eg. a car entering the road from a side way, a car stops suddenly, or changes lanes. Meaning that a real trip data can be used to simu numerous trips on the way, this will significantly reduce time required to train the driving system. After some point of simu training, the driving system can be verified by new real data. In other words, most of real data are useless except for those special situations periods. So, Tesla shall use nVidia simu tech for fast training or creates its own simu tech.
Chaps the name of the chip for the cars is Thor. Digit mini computer is to allow smallest companies and developers to have AI compute on network/locally.
Well...Nvidia is now trying to copy Tesla's RWM for the Robot, and their Robotaxis. Nvidia is going to struggle to keep up with Tesla's FAST iterations. Noone can keep up with Tesla, except for the Hardwares that Nvidia can come up with. Problem is that Nvidia is trying to be a general purpose Hardware manufacturer, and Tesla is for their specific application. Tesla will win and be No.1, while Nvidia probably will be No.2 with robots.
Nvidia doesn't care if Tesla is no1 as long as they can sell their stuff to all of Tesla's competitors. And they are also selling to Tesla. Tesla is not using their software maybe but they do use the hardware.
Well we had incoming president shit pants talking all kinds of crazy nonsense about invading, Panama and Greenland as well as doubling down on wanting to drill baby drill and at the same time announcing His insistence on wanting to instill immediately inflationary terrorists on the country. So yeah, the markets were quite quite terrified by this. And on top of that was the jobs data that came in from manufacturing showing it's still in inflationary area. Likely almost certainly due to Trump's continued talking about tariffs tariffs tariffs.
@ZoltanBojas-ug6nf The see saw is: Tesla is going to win FSD therefore Nvidia is wasting its time trying to sell AI solutions to car companies. NVDA drops. Or Nvidia will sell working solutions for AI driving to any and all car companies, therefore Tesla will not be the clear winner. TSLA drops.
Good to see folks discussing sim vs real world video for training. I tried to foster such discussion last year, to no avail. It’s a potential giant threat to TSLA, so absolutely worth a lot of discussion.
Great upload and articulation, but to play Devil’s Advocate: 1) maybe Nvidia is more a threat to Optimus rather than FSD ??😮😮 2) Nvidia as a Service, aka NaaS, has its appeal to Tesla competitors who don't want to partner with Tesla or become their customers (in real world competitors do source from each other though of course..) 3) Everywhere else in the world Tesla can maintain their head start advantage, their profitability models, their team and culture etc, but in China: The Chinese companies can catch up and surpass physical data acquision in rate and cost The competitive drivers there will push forward Chinese ai robotics energy etc to be Teslas global competitor, as many of us have been saying and seeing for a decade 4) Herbert and guest kind of gloss past that the idea behind Cosmos (other than to keep and grow Nvidia income stream 😅) is to eventually have a purpose-flexible real-world-physics-environment model that is good enough for approximating the real deal, and which can be scaled up very cheaply relative to future computing costs, and which can be scaled very finely for different uses, and will keep getting better. Also, the more customers use it, the better it becomes, so on, critical mass snowball effect - can surpass in theory, FSD, Optimus, anything imaginable, IN theory, cheaply too or for specific competitive risk advantages. 5) Simulations get too real too soon - human society as we know it collapses and we are in the ICE-dominates hellscape of Mad Max or the AI-prohibited post-great-wars universe of Dune :D
Not really. This gets the competitions from completely hopeless to mostly hopeless. These tools are like a well equipped kitchen, they don't automatically make you a chef.
In this case NVDA is following what Tesla has been doing for years, simulations. NVDA will sell it to other automakers, but it is to late for them to catch up and will be forced to license Tesla FSD anyway because if they do not they will be many years behind in robotaxi.
Very insightful comment from Phil, yep there are just tools and there's going to be a lot of engineering work needed to make them work. Its debatable how far they can get with synthetic data, but putting that aside, NVDA is really giving them a glimpse of what's possible where previously they had no hope and its a great way for NVDA to lock them into using NVDA equipment for the forseeable future. Master move my NVDA and many OEMs will spend a shitload of cash with no usable outcome.
Exactly what I thought, it may even be an advantage for Tesla. Maybe they will come in the end to Tesla and ask for FSD licensing if these comp legacy companies still exist. Legacy auto is in such big trouble that most likely they will not have the money to make the necessary investment to get to autonomy. Furthermore, how should e company that can´t manage to get over the air software updates done to build a sufficient data center to use these Nvidia tools?
i wouldnt flip out legacy auto moves so slow saying they would take 3 years to put AV hardware tech into their PC and simulated data not sure how much that would cover every edge case. But even if everything was shipping today and legacy auto started it would be 3 years from now so 2028 at the very earliest. you always have to see how everything is going to play out.
Yes, in video will be able to help companies robots. Pour the water in the bottle, but did the water go in the bottle just some food for thought a smart man already said it.
What about when the training data is produced by AI systems that have physics built into them as a ground truth. The data that they generate does come from the real world - real world in quotes - because it is based on the principles of physics. Augmentation.
This is the idea behind Nvidia Cosmos, but there is so much that can be simulated, and it is up to the customer to decide for their goals and budgets where and how to focus ;)
So now everyone knows all automakers have access to the hardware, if they can integrate it into their vehicles then the shortcut is an FSD licensing deal with Tesla for the software, real world data, and training compute.
Nvidia do not have a contract with Tesla. All those robots are from the companies that Nvidia have contracts with. I guess those companies are going to use Nvidia RWM, and their Hardware to run those robots.
Let's set the record straight: NVIDIA has been laying down foundation work in its Issac Sim project since 2018, well ahead of Tesla's Optimus revealed in 2021.
To be clear, even if the rest of the industry did pool their data, it will still take them years to even START collecting the data they need to START catching up to where Tesla is NOW, because their current vehicles do not have the capability to capture and transmit that data. They will have to redesign for FSD, which Elon said they told Tesla will take 3-5 years. In other words, Tesla will have fully operational robotaxis years before legacy could even launch a competing effort, which would likely still fail.
Could the prospect for legacy auto to "do it ourselves" with the NVIDIA kit set deter them from partnering with Tesla over autonomy. If so, then this could remove a significant Tesla value driver (that other companies give up). Of course, achieving autonomy with the kitset is another question. Thoughts?
It can but it will be proportional to how far behind they currently are in getting to autonomy. The companies that think they could pick up nvidia's toolkit and start building this. Some of them have already quit. I think GM's Cruise was usually some aspects of nvidia's pipeline. What's new about this presentation is Cosmos which is the actual analog to FSD. It's the actual model that's doing the outputs of the action data against the vehicle in order to make it navigate in the world. Prior to this all they had was Omniverse and various custom models and heuristic code. Much like what Tesla had before it went to end to end neural nets.
@DavidSaintloth Thanks. I think, though, that there is plenty of scope here for analysts to jump to the conclusion that Tesla's position has been significantly eroded. However, the actual proof of the pudding should only be six months away. Once Tesla delivers a robotaxi capability, the actual gap will be obvious. As always, actual capability trumps theory. In the meantime, there could be a lot of FUD.
Jensen Huang has always stated that Tesla is way ahead on autonomous driving. His point is that everybody else will eventually get there as well. So it’s smart to become the platform for all those other guys. We’re talking major $$$ down the road for Nvidia.
Hans did a great job, but please focus on this topic with multiple industry experts. I think we need multiple opinions and full clarity on this development. Clearly licensing FSD (which some of us were hoping) might be out of the question.
Tesla strength is their extremely talented people. Just like you can give me the best supercomputer, I won’t be able to turn into software genius no matter how hard I study, I may be good at it, but let’s be honest genius is born with, not groomed. Elon is good at recruiting extreme talented team, that’s the moat and strength of Tesla, their people. Having one genius to work for you is better than having thousand of mediocre workers work for you , very obvious example is Cariad of VW
I can’t listen to him even though he’s smart and says interesting things. He sounds like a Wookiee talking through a not very good human translator machine at half normal speed.
I’m fairly certain that is a speaking technique to control a stutter. I worked with a guy who talked like that and when he got stressed or needed to speak faster he would begin to stutter.
Yeah, haven’t watched yet, but it sounds dangerous (for driving) because how do we know it’s not missing important edge cases? Edit: watched video, more confused now 😅
Nvidia investor here. We're not giving them anything. They're going to pay for every brick and every ounce of mortar and we will fatten ourselves as they stumble trying to build the house. 🤑
@ NVD are essentially going to sell shelf ware - they will buy and then be unable to do very little with it. I highly doubt NVD will cover their costs of development. It’ll be MUCH cheaper for car manufacturers to license from Tesla. Ps I’m a long term investor in NVD
Elon has said on several occasions that Ai knows the difference between simulation and real world, which is the limitation of simulations. You can see this play out in FSD. I am currently driving FSD 12 on on HW3 and the car was driving very well with summer sun however when winter came with long shadows the car become hesitant struggling with the different sun angle. These subtle differences. Make a huge difference! AI need to learn the real world 🌎
I disagree with the final thought that all companies would have to get along to share driving data to catch up with Tesla. It took Tesla so many years to collect data and reach 7 million vehicles on the road because they had a low base, which is growing all the time. Toyota sells 11 million vehicles a year and if they invested big money they could reach that number in a year. Personally, I see the biggest competition in China because there the tech companies work closely with the auto companies, and the Chinese government has more demands and ambitions than the CEOs of many legacy automakers who don't understand the tech industry and are slow in their decisions. Nvidia unfortunately shows that Tesla, despite being in the lead, will probably have a hard time monopolizing its position in autonomy forever.
Yep, legacy automakers really don’t want to give more of their money to Tesla, which is another reason I don’t expect to see them licensing FSD from Tesla. Now they will try the partner with Nvidia route as it seems to offer a path forward not requiring the vast amount of data they lack. With advances in technology maybe this gets them a decent FSD in five years, but taking robo taxi market from Tesla will be very hard five years from now. The robots is another issue, it is still possible someone will compete well with Optimus.
Time is the critical factor here that determines whether or not they will partner or try their own approach. Keep in mind there are companies that I've already been using nvidia's kit and they're exactly kind of still way behind Tesla. Yes, they probably were using their own custom model and your Erica code and now with Cosmos they could just have their own essentially version of FSD but still it will take them time to integrate this entire system into their vehicles for production more time than the year and a half to 2 years over which Tesla will be debuting a global robo taxi service. The clock is ticking!
@@mikeinsomerset Yes, and other companies also have their humanoid robots in factories already too, at least one being backed by Nvidia. Yes, Tesla has great manufacturing, but there are other competent manufacturing companies in the world. Not saying Tesla won’t win, just saying it is not clear yet and there are several horses in the race with big money and tech behind them.
Cosmos is analogous to FSD. It's the model that generated the action tokens used to drive the embodied robot. I think the description in the keynote could have been less jargon heavy. Lots of people think it's just a simulator and not the actual world model that's embodied into whatever form factor. They had Groot & Omniverse which is the simulated data that trains the model (Cosmos).
Is there a need for concern about Tesla losing out on licensing deals in the future. Are Nvidia and Tesla in a symbiotic relationship, or will Nvidia's advancement with this cannibalize this potential market opportunity?
It depends on how far behind the legacy automaker is and also remember if Tesla releases the robo taxi service starting in 6 months from now and then quickly advances it across the rest of the world that's going to put immediate pressure on any of the existing legacy auto companies because Tesla will be out there selling rides for a fraction of the cost of any of the other services including legacy ride, healing services and so the pressure will be massively on legacy auto companies to come to a deal with Tesla. Basically they just don't have time. It's more likely that they will partner with Tesla then try to do it with nvidia's kit. Unless of course they're a BYD which will likely leverage nvidia's kit because it has no other choice really.
The thing you're missing is that NVIDIA DRIVE Sim provides physically accurate simulations for a variety of sensors, including visual sensors like cameras, as well as active sensors such as radar, lidar, and ultrasonic sensors. Tesla does not use, or train its systems on lidar, radar, or ultrasonic sensors, which limits its ability to train for poor visibility conditions.
I have a question for Hans if he is available. I will be brief so the question May not be very understandable. Please let me know if there's a way I can communicate directly with Hans if you don't mind. I am a mathematical statistician and have a lot of experience with Monte Carlos simulation. I am wondering if this technology could be used to simulate say 100,000 simulations of data input predicting the cash flow of the the payout of the income cash flow for a fixed annuity. I could provide a phone number or email address for further Communications if that might be a possibility of further discussion. Oracle provides a such a simulation model from a product they called Crystal Ball should you require more clarity the type of mathematical statistical modeling for such a Monty Carlo simulation analysis.
As to this argument of is real world data better than simulated data. What? Hans is missing here and what Elon is also missing is that because they use Omniverse, which is a physics-based simulation environment to ground truth, the simulated generations data that's used to train Cosmos much of the scenarios that Cosmos learns from simulated data fully grounded. Even edge cases. You can't have cows floating in the sky, for example, because within a physically based simulation environment objects have mass which force them to fall to the ground... So when you're generating scenarios within the simulation environment, those scenarios must follow the physics and then you can iterate valid physically valid interactions between those objects based on the mass, velocity and friction and other attributes that they're given within the simulated environment. So long as you were able to capture all of the physics associated with reality inside the simulation environment, you will then be able to simulate any possible edge case... Leaving the only distinction between the two training scenarios being the necessity of compute. This is different because in the real world data gathering scenario, you are gathering precisely data that is coming from real world driving scenarios. The grounding to the physics is coming in from those scenarios. Your trying to create a distinction between the two isn't really valid as if you're in the real world for the first time and you experience something, you have no reference as to whether or not that experience is normal for the environment. If I show you a single frame with a wheel in the middle of the frame, you might think that that's normal. But if I show your subsequent frames of that wheel falling to the ground, then you'll understand that that's not normal. It will take you time to understand the physics of the world after viewing enough frames. The same way that it would take Cosmos to understand the various synthesized scenarios and edge cases that would be generated in Omniverse and used for training it. Again, the only distinction remaining being the sparsity of the data where the real world data is very sparse. It's basically exactly what you need because it's coming from a existing fleet of vehicles that are existing experiencing valid driving scenarios... Whereas Omniverse can generate valid driving scenarios, but it can generate a whole bunch of invalid driving scenarios as well, which need to be essentially pruned from the training process for Cosmos.
And that's not true. Omniverse allows simulation of all physically plausible scenarios. There are many odd physically plausible scenarios, every single one of them in the real world... That is an edge case is also plausibly physical by definition. The only difference is the massiveness of the search space of the different trajectories for the possibilities of object interactions between assimilated environment and a real world one. But then that's why you use a physically based simulator like Omniverse to ground the simulations to the physics which is common both to the world and to the simulation environment. Leveraging that stochastic ability to create alternative scenarios ...That's how you reduce the Sim to real Gap.
Legacy have failed at in-house AV development and no chance they can afford a second attempt. They have to rely on outsourcing. The nvidia partnerships, like Toyota's, are DOA.
Outstanding Herbert. Your reaction time is amazing. Thanks to you and Hans. Can Nvidia simulate the physics of human stupidity!? I guess it will have to simulate a million war games scenarios to catch all the edge cases. Let’s not forget that Tesla has other advantages it can deploy in the car such as Starlink, Grok, Supercharging infrastructure, …
Buying a Tesla? Use my referral link and get up to $2,000 off.
ts.la/herbert23392
My daughter: ts.la/amanda443323
Most OEMs cannot even make money manufacturing EVs, you add on the cost for Nvidia’s chips and the training cost how are they going to compete with Tesla?
Government bailouts if a democrat wins the presidency before they all go out of business😂
They aren’t.
@@Richard-cq4kv You new to the world? Both parties do the bailouts bro.
Jensen gave a great presentation. And while one of his backdrops included an array of bots, Optimus was noticeably missing.
Hans makes great points very articulately on the key Tesla differences -- DIY takes the game !!!
Hans missed the greatest need for real world data: verification. Elon has stated a few times that (even with all their data) one of their biggest challenges is finding enough examples of the remaining corner cases, and verifying in the real world that FSD can handle them reliably.
Remaining corner case can be easily generated. You can imagine hundreds of corner cases in your mind , why do think AI can’t do that ?
@@jjj32753AI Is only a BS algorithm, it can't imagine anything.
@@jjj32753 Because AI can only interpolate, not extrapolate. It's not truly creative/innovative.
@@jjj32753 For proper validation you absolutely need real data. No simulation is perfect. You cannot validate safety critical products using simulation, and there is no way authorities would (or should) accept simulation performance stats for satisfying self driving regulations.
@@jjj32753 AI is great at generating good models of reality but not perfect. You need real data to check it against to keep flaws in the generated data from running rampant. Think of the errors GPT makes and imagine basing your FSD on that.
Sounds like nvidia is trying to milk the auto industry for all they are worth before they fully collapse
And it will be absolutely glorious.
lol - absolutely correct. 😂
Absolutely
Right, then merge with tesla to cause said collapse. Idk why the companies don't just merge but I suppose it could be monopoly laws
Jensen's entire talk -- NVIDIA's entire approach -- is a validation of Tesla's tech development of FSD and
Optimus. But, Wall Street being what it is, is raising NVIDIA's value and decreasing Tesla's.
'Analysts' are truly clueless.
Yes, but give it time and eventually Wall Street will get on board in its entirety. Remember when nobody thought the EV shift was even in the cards? Neither mfgs nor most analysts made the change in direction quickly enough so most are desperate to catch up, and now a good few have. I am saying you are spot on about the validation!!!!!!!!!!!! In a reasonable market probably increasing nvidia valuation while maintaining tsla somewhat baked in valuation would be a better response. Tsla also suffered from stagflation threat news today 😅
❤
@@Li-yt7zh it’s a good buying opp - use your edge
This is what will make VR a reality. Since playing Minecraft I knew that we could never code the worlds of VR but needed the computer to generate them.
This is a great explanation of real world data vs simulation data. It's so important for retail investors to understand what is going on with in the autonomous driving space, but most of us don't have the experience to parse these presentations. Thanks guys!
the technique/process vs data I also think is huge... AI is (currently) enhancing the engineering talent, not replacing it
I agree that none of the partners has a vertically integrated stack for producing an SDV using Cosmos & Thor. It will take several years for them to get to where Tesla was 2 years ago.
My two biggest asset positions have been Tesla & Nvidia since i started my portfolio precisely because it was clear to me then that Tesla was Apple & Nvidia was Google Android in the race to autonomy.
Cosmos fully validates that thesis.
It’s simple - can you use flight simulators exclusively to train pilots. The answer is no. It helps, it allows training in scenarios that can’t be recreated safely or budget wise in real world, but it is not a replacement for actually flying a plane. Not even close.
Companies will need a lot of real data and huge compute.
This is a brilliant sales strategy to cause some companies to buy compute but they’ll quickly realize the compute they purchased is useless without real world data and even when you have the data you need experienced engineers to build the pipelines and learn how to build the system that trains the system.
Very good analogy.
Mr. Huang would not disagree with that and nothing he said validates or contradicts that.
@@ThomasBeekBut he does suggest that.
@@fractalelf7760 Because he is an excellent salesman, which is meant as a compliment.
The question is not of Cosmos synthetic data is better than Tesla’s real world data, the question is if the Cosmos synthetic data trained autonomous driving cars can be better than human driver enough to be robotaxi
Power requirements for the Hardware Nvidia is going to use is too high. Tesla FSD version 3 only uses about 100w, while Nvidia hardware is probably going to be higher. Nvidia is trying to copy Tesla on their Real World model.....Nvidia won't iterate fast enough and will struggle to keep up with Tesla.
Yes. But it will cost lots more compute which conveniently Nvidia has a whole business selling.
Nvidia is happy selling compute, models & inference boards to whoever wants to build robots. They don't need them to be profitable as long as they buy.
That's like Google Android. They don't care if Motorola profits from their phones but with each phone it's a new user using Google apps, viewing ads... Using Google search. They get the eye candy without having to manufacturer the frame it's delivered from (save for their own pads & pixel phones)
@@DavidSaintloth Actually, this is the first time Nvidia started selling an Inference Chip/board. Tesla built their own inference chip and the board to go with it first. Nvidia have been selling GPU chips, and system with their software for training which is a general type hardware with their own software.
The minimum inference computer requirements for autonomous driving are:
Dual processors for redundancy.
Video inputs from all the cameras.
Inputs from the driver.
Inputs and control of the battery/passenger thermal system.
Control outputs for window up/down, breaking, acceleration steering and etc.
An INVIDA processor alone is not sufficient.
Tesla will succeed while others are going to struggle to keep up.
One possible outcome could also be, that the legacy automakers give their fate into Nvidia's hands but lose in the end time, as they still have to figure the full self driving out on their own, as Nvidia only gives them the tool to maybe do so. Just licensing from Tesla would be the very safe step. Very interesting times, but for sure there is some competition in sight. For Tesla all depends on the fast Robotaxi rollout, because in the end the deciding factor is time of market entry and cost per mile and there Tesla is very well positioned.
That’s why it’s a change phase transformation. There’s so much with the convergence of multiple technologies that are helping each other grow.
Real World Data 📊 and vertical integration of the most valuable problems to solve for the win! Thanks Hans, Thanks Herbert!
synthetic data vs real world data is akin to digital zoom vs optical zoom
Not even close. It is separated from initial inputs by AI processing.
Thanks Herbert and Hanz!
🙏🏽🙏🏽🙏🏽🙏🏽🙏🏽
⭐️⭐️⭐️⭐️⭐️
Cheers (real world comment, not synthetic comment)
Its great that we all understand the diffrence between the brittle artificial data vs the real life data but will hedge funds or regular investors be cognizant of these difffrences ? 😢😢
Like that famous political line from James Carvel, “It’s the total car cost stupid”. Those who produce an autonomous car for the lowest cost and therefore stay ahead of cost per mile, wins.
Slowly coming down from current $2-4 / mile to cents just staying ahead of other competitors until they bleed money. Much like Tesla did in China thru price cuts.
Why an OEM would choose Tesla OS over NVidia? BECAUSE that will allow them to buy all the modules (48V) from Teslas suppliers and directly as well as how to manufacture their product for less cost.
I upgraded to hw3 in our 2018 model 3 for 1,000. Even if you add the other things, it still sounds like a 3k system for other oems will be more expensive that what tesla spends. Plus the cars need to be able to do ota updates which they still can not do and have not been able to do. Plus the car as a whole needs to be more centralized. It is still going to take these oems years to implement this. Then the training will be years behind at that point even if it can improve at any rate like tesla did without any real world data or gathering from the fleet. They will only be starting at this point years later with data collection. The other oems will still be screwed for a very long time to come. They will probably be bankrupt by then and with robotaxies, the overall fleet will shrink and people will need less of their cars. It will not become viable for them to do all of this profitably. They are screwed either way.
My first thought when I warched the full 2 hour presentation: they blatantly copied Tesla 😂
everyone copies everyone else its the nature of business and capitalism
Yep. Just like they are trying to copy Tesla on the EV front...If they don't copy Tesla, they won't succeed.
I don't think Elon minds, he gets on well with Jenson.
Anyway if you are doing the same thing it's going to look similar.
If you ask two blacksmiths to make a Bowie knife they'll look similar even if they never met eachother
Don’t forget Elon and Jensen are friends…
@@andrewsaint6581 Who started the EV revolution? Tesla, and Elon. Who started the Robot Revolution? tesla, and Elon. Who started Self Driving? Elon and Tesla. who started Rocket company? Elon and SpaceX. Who started Battery Storage on a Big scale? Elon and Tesla. Everybody else FOLLOWED. Yeah, I know....Some stuff what I stated had others doing it, but on a very small scale. It is Elon who started and got it going on a BIG scale, and is successful at it. I could name few more Elon started on a big scale....
You would need multiple simulations each with its own speciality, from animals, cars, people, activities, weather, emergencies etc. All deployed on a single platform, with limited coordination between each SIM. Then test FSD on this highly unpredictable environment.
U go Hans! We do important sh!t ourselves!
Ai can hallucinate just hand a car with all kinds of different sensors and the physics of the car and what is going to happen with rain,fog slate etc. Plus the mass of the cars. Especially different model of cars.
Do not try to force a robotaxi to drive in undrivable conditions. There is no real benefit to risking life and limb to put a taxi on a pedestal.
the idea is that if a human can do it, a robot can eventually be designed to do it better - more efficiently, faster response and safer, and far more consistently than a human ability average. The robot won't get tired after driving longer so it won't have stats reductions like a human would, especially important for say, icy road semi-truck drivers
OP also totally correct to point out you need the physics and 3d models data of the robots ie vehicles themselves for practical training
OP also totally correct to point out you need the physics and 3d models data of the robots ie vehicles themselves for practical training
I actually see Nvidia becoming too extended in their product line leading to jam ups and slowly getting slow downs
Now the question is whether Tesla can use its huge real-world and correct data to generate much much more synthetic data to deal with almost 100% of all edge scenarios and make Tesla FSD 100% perfect!
When the legacy auto get there, then it comes back to manufacturing
so why is it that gm build s10 electric cars, instead of just one really good one, that they can make at scale and speed and maybe get to break even on it?
Best keynote ever!
This doesn’t mean that Tesla has no lead, but if synthetic data can get other car companies to the point of Supervised FSD it’s a huge help. They may not get to robotaxi, but could get to where Tesla is now
And yes they will all share their real world video data because it will be a requirement of their contract with nvidia
I wonder if you could use insurance company data based on accidents to generate edge case simulations and then create a multitude of conditional variations to cover for inaccuracies in the data or adjacent edge cases.
Real data is very important as a source of simu data and verification of simu performance. But simu data can accelerate learning dramatically. For example, you drive from one city to another aparting 1000km, it may just have limited situatios that needs special care, which just needs a few miliseconds or a few seconds for handling each special situation. most of time it just drives smoothly, no self-training while in a smooth situation. However, with simu data, unlimited special situations can be generated. Number of miles is not the key, number of special situations is the key. Using simu data, any special situation can be generated along the way, eg. a car entering the road from a side way, a car stops suddenly, or changes lanes. Meaning that a real trip data can be used to simu numerous trips on the way, this will significantly reduce time required to train the driving system. After some point of simu training, the driving system can be verified by new real data. In other words, most of real data are useless except for those special situations periods. So, Tesla shall use nVidia simu tech for fast training or creates its own simu tech.
No, the balance between average driving and edge cases is important. Elon Musk already spoke about that.
Chaps the name of the chip for the cars is Thor. Digit mini computer is to allow smallest companies and developers to have AI compute on network/locally.
Nvda stock price was down today. So was Tesla. I thought they would be on opposite sides of a see saw today.
Well...Nvidia is now trying to copy Tesla's RWM for the Robot, and their Robotaxis. Nvidia is going to struggle to keep up with Tesla's FAST iterations. Noone can keep up with Tesla, except for the Hardwares that Nvidia can come up with. Problem is that Nvidia is trying to be a general purpose Hardware manufacturer, and Tesla is for their specific application. Tesla will win and be No.1, while Nvidia probably will be No.2 with robots.
Nvidia doesn't care if Tesla is no1 as long as they can sell their stuff to all of Tesla's competitors. And they are also selling to Tesla. Tesla is not using their software maybe but they do use the hardware.
Well we had incoming president shit pants talking all kinds of crazy nonsense about invading, Panama and Greenland as well as doubling down on wanting to drill baby drill and at the same time announcing His insistence on wanting to instill immediately inflationary terrorists on the country. So yeah, the markets were quite quite terrified by this. And on top of that was the jobs data that came in from manufacturing showing it's still in inflationary area. Likely almost certainly due to Trump's continued talking about tariffs tariffs tariffs.
@ZoltanBojas-ug6nf The see saw is:
Tesla is going to win FSD therefore Nvidia is wasting its time trying to sell AI solutions to car companies. NVDA drops.
Or
Nvidia will sell working solutions for AI driving to any and all car companies, therefore Tesla will not be the clear winner. TSLA drops.
Good to see folks discussing sim vs real world video for training. I tried to foster such discussion last year, to no avail. It’s a potential giant threat to TSLA, so absolutely worth a lot of discussion.
And who are you?
agreed
Great upload and articulation, but to play Devil’s Advocate:
1) maybe Nvidia is more a threat to Optimus rather than FSD ??😮😮
2) Nvidia as a Service, aka NaaS, has its appeal to Tesla competitors who don't want to partner with Tesla or become their customers
(in real world competitors do source from each other though of course..)
3) Everywhere else in the world Tesla can maintain their head start advantage, their profitability models, their team and culture etc, but in China:
The Chinese companies can catch up and surpass physical data acquision in rate and cost
The competitive drivers there will push forward Chinese ai robotics energy etc to be Teslas global competitor, as many of us have been saying and seeing for a decade
4) Herbert and guest kind of gloss past that the idea behind Cosmos (other than to keep and grow Nvidia income stream 😅) is to eventually have a purpose-flexible real-world-physics-environment model that is good enough for approximating the real deal, and which can be scaled up very cheaply relative to future computing costs, and which can be scaled very finely for different uses, and will keep getting better. Also, the more customers use it, the better it becomes, so on, critical mass snowball effect - can surpass in theory, FSD, Optimus, anything imaginable, IN theory, cheaply too or for specific competitive risk advantages.
5) Simulations get too real too soon - human society as we know it collapses and we are in the ICE-dominates hellscape of Mad Max or the AI-prohibited post-great-wars universe of Dune :D
Tesla may be ahead, but this is a clear shot across Tesla’s bow.
Not really. This gets the competitions from completely hopeless to mostly hopeless. These tools are like a well equipped kitchen, they don't automatically make you a chef.
Elon tweeted that video out because it validated his opinion at self driving cars are less than a year away
For sure, and unlike Tesla, Nvidia has a focused CEO who is not continually distracted by anything and everything.
In this case NVDA is following what Tesla has been doing for years, simulations. NVDA will sell it to other automakers, but it is to late for them to catch up and will be forced to license Tesla FSD anyway because if they do not they will be many years behind in robotaxi.
@@michelangelobuonarroti916 You’re tough.
Excellent explanations, thank you herbert & hans
This work by NVIDA will help provide some basis of cost for licensing Tesla FSD for other companies.
It is important that Tesla does not have a monopoly for political and competitive development reasons.
Any presentation that has lots of orchestral music in the background is marketing and not science
There's a video about Omniverse on YT channel Connecting the Dots.
An eye opener.
Don’t forget Elon and Jensen are good friends and technical allies… this partnership doesn’t put them in a we vs them competition.
Very insightful comment from Phil, yep there are just tools and there's going to be a lot of engineering work needed to make them work. Its debatable how far they can get with synthetic data, but putting that aside, NVDA is really giving them a glimpse of what's possible where previously they had no hope and its a great way for NVDA to lock them into using NVDA equipment for the forseeable future. Master move my NVDA and many OEMs will spend a shitload of cash with no usable outcome.
Exactly what I thought, it may even be an advantage for Tesla. Maybe they will come in the end to Tesla and ask for FSD licensing if these comp legacy companies still exist. Legacy auto is in such big trouble that most likely they will not have the money to make the necessary investment to get to autonomy. Furthermore, how should e company that can´t manage to get over the air software updates done to build a sufficient data center to use these Nvidia tools?
Thanks Herbert... Appreciate the insight.
Digit doesn't go into vehicles. It's a desktop chip. That's Jetsen Thor & nano inferencing boards.
i use to think that but 5 yeras max behind to cathch up , if they can survive till then.
i wouldnt flip out legacy auto moves so slow saying they would take 3 years to put AV hardware tech into their PC and simulated data not sure how much that would cover every edge case. But even if everything was shipping today and legacy auto started it would be 3 years from now so 2028 at the very earliest. you always have to see how everything is going to play out.
Is this a Tesla Infomercial?
Yea, Hans seems like a fanboy
Thanks guys
Yes, in video will be able to help companies robots. Pour the water in the bottle, but did the water go in the bottle just some food for thought a smart man already said it.
What about when the training data is produced by AI systems that have physics built into them as a ground truth. The data that they generate does come from the real world - real world in quotes - because it is based on the principles of physics. Augmentation.
This is the idea behind Nvidia Cosmos, but there is so much that can be simulated, and it is up to the customer to decide for their goals and budgets where and how to focus ;)
Nvidia is trying to make up real world training data with synthetic data. Will it work as well? I really doubt it.
So now everyone knows all automakers have access to the hardware, if they can integrate it into their vehicles then the shortcut is an FSD licensing deal with Tesla for the software, real world data, and training compute.
That's why no OEM want to licensing FSD, then Elon works on it own Cybercab.
Hans makes me think I need more TSLA stock 😮
I noticed during the Jenson speech when the robots were all lined up Optimus was missing. What is that all about?
Tesla bots don't need to use Nvidia software suite.
Did you notice that they weren't there at all, but a video presentation?
Nvidia do not have a contract with Tesla. All those robots are from the companies that Nvidia have contracts with. I guess those companies are going to use Nvidia RWM, and their Hardware to run those robots.
@ yes I noticed, not the point?
@@keithpeterson9560 optimus wasn't there because he's real.
Some of the video showed by NVIDIA in their post is of a Tesla FSD TH-cam tester.
Would really like to hear from James Douma about this.
Let's set the record straight: NVIDIA has been laying down foundation work in its Issac Sim project since 2018, well ahead of Tesla's Optimus revealed in 2021.
This is very interesting topic.
It would seem to be quicker and cheaper for most of the auto manufacturers to simply license Tesla's system as opposed to trying to create their own
Wow, watch legacy auto spend billions on this just to realize they can’t figure out how to use it.
To be clear, even if the rest of the industry did pool their data, it will still take them years to even START collecting the data they need to START catching up to where Tesla is NOW, because their current vehicles do not have the capability to capture and transmit that data. They will have to redesign for FSD, which Elon said they told Tesla will take 3-5 years. In other words, Tesla will have fully operational robotaxis years before legacy could even launch a competing effort, which would likely still fail.
Could the prospect for legacy auto to "do it ourselves" with the NVIDIA kit set deter them from partnering with Tesla over autonomy. If so, then this could remove a significant Tesla value driver (that other companies give up). Of course, achieving autonomy with the kitset is another question. Thoughts?
It can but it will be proportional to how far behind they currently are in getting to autonomy. The companies that think they could pick up nvidia's toolkit and start building this. Some of them have already quit. I think GM's Cruise was usually some aspects of nvidia's pipeline.
What's new about this presentation is Cosmos which is the actual analog to FSD. It's the actual model that's doing the outputs of the action data against the vehicle in order to make it navigate in the world. Prior to this all they had was Omniverse and various custom models and heuristic code. Much like what Tesla had before it went to end to end neural nets.
@DavidSaintloth Thanks. I think, though, that there is plenty of scope here for analysts to jump to the conclusion that Tesla's position has been significantly eroded. However, the actual proof of the pudding should only be six months away. Once Tesla delivers a robotaxi capability, the actual gap will be obvious. As always, actual capability trumps theory. In the meantime, there could be a lot of FUD.
Jensen Huang has always stated that Tesla is way ahead on autonomous driving. His point is that everybody else will eventually get there as well. So it’s smart to become the platform for all those other guys. We’re talking major $$$ down the road for Nvidia.
Hans did a great job, but please focus on this topic with multiple industry experts. I think we need multiple opinions and full clarity on this development. Clearly licensing FSD (which some of us were hoping) might be out of the question.
Tesla strength is their extremely talented people. Just like you can give me the best supercomputer, I won’t be able to turn into software genius no matter how hard I study, I may be good at it, but let’s be honest genius is born with, not groomed. Elon is good at recruiting extreme talented team, that’s the moat and strength of Tesla, their people. Having one genius to work for you is better than having thousand of mediocre workers work for you , very obvious example is Cariad of VW
NVIDIA sucking dry the desperate legacy auto manufacturers 😂. Like with all generative stuff, if the seed sucks the generated data is limited
Great analysis
So basically, they’re building their own V11
1:48 That's Meituan's autonomous delivery vehicle.
The way Hans drags out his words is very frustrating to listen to
I can’t listen to him even though he’s smart and says interesting things.
He sounds like a Wookiee talking through a not very good human translator machine at half normal speed.
A verbal non-fluency has become a true handicap.
He's gonna need a vocal coach to break this habit.
@@jimmycrack-corn9872 😂
I’m fairly certain that is a speaking technique to control a stutter. I worked with a guy who talked like that and when he got stressed or needed to speak faster he would begin to stutter.
Looks like trying to provide a shortcut avoiding having to collect that massive real world data Tesla has been collecting for years.
Tesla has used simulations in addition to real world data for years
Yeah, haven’t watched yet, but it sounds dangerous (for driving) because how do we know it’s not missing important edge cases?
Edit: watched video, more confused now 😅
Jensen is validating FSD, that it is a real thing and help make it mainstream sooner. It is good for Tesla.
I’m seeing an absolute killer version of Minecraft coming out of this
What would Tesla's data be worth to Nvidia to train their FSD models. I'm thinking 20 to a 100 billion dollars.
How’d work those figures out?
@@michaelholmes8848 bases on estimated revenue from Robotaxi.
They do some cooperation so it is not technology war.
Summary: Giving someone cement and bricks doesn’t mean they know how to build a house
Nvidia investor here. We're not giving them anything. They're going to pay for every brick and every ounce of mortar and we will fatten ourselves as they stumble trying to build the house. 🤑
@ NVD are essentially going to sell shelf ware - they will buy and then be unable to do very little with it. I highly doubt NVD will cover their costs of development. It’ll be MUCH cheaper for car manufacturers to license from Tesla. Ps I’m a long term investor in NVD
Tesla is already working on AI5 for there cars.Wlon said the version 3 is when you can say you achieve what you were aiming for
36:09 good luck on this
Hans, looks like you need a camera upgrade! 🤣
Trillion miles PA. Big TAM!
Car companies can combine to share driving data.
Another way to get real world data is a third party can collect it to share with all.
Elon has said on several occasions that Ai knows the difference between simulation and real world, which is the limitation of simulations. You can see this play out in FSD. I am currently driving FSD 12 on on HW3 and the car was driving very well with summer sun however when winter came with long shadows the car become hesitant struggling with the different sun angle. These subtle differences.
Make a huge difference! AI need to learn the real world 🌎
Does Tesla currently capure Real Word Data for ALL its cars worldwide? How is this data uploaded?
I disagree with the final thought that all companies would have to get along to share driving data to catch up with Tesla. It took Tesla so many years to collect data and reach 7 million vehicles on the road because they had a low base,
which is growing all the time. Toyota sells 11 million vehicles a year and if they invested big money they could reach that number in a year. Personally, I see the biggest competition in China because there the tech companies work closely with the auto companies, and the Chinese government has more demands and ambitions than the CEOs of many legacy automakers who don't understand the tech industry and are slow in their decisions. Nvidia unfortunately shows that Tesla, despite being in the lead, will probably have a hard time monopolizing its position in autonomy forever.
Not true Apple
Yep, legacy automakers really don’t want to give more of their money to Tesla, which is another reason I don’t expect to see them licensing FSD from Tesla. Now they will try the partner with Nvidia route as it seems to offer a path forward not requiring the vast amount of data they lack. With advances in technology maybe this gets them a decent FSD in five years, but taking robo taxi market from Tesla will be very hard five years from now.
The robots is another issue, it is still possible someone will compete well with Optimus.
Re optimus. Its happened already
Time is the critical factor here that determines whether or not they will partner or try their own approach. Keep in mind there are companies that I've already been using nvidia's kit and they're exactly kind of still way behind Tesla. Yes, they probably were using their own custom model and your Erica code and now with Cosmos they could just have their own essentially version of FSD but still it will take them time to integrate this entire system into their vehicles for production more time than the year and a half to 2 years over which Tesla will be debuting a global robo taxi service. The clock is ticking!
@@mikeinsomerset Yes, and other companies also have their humanoid robots in factories already too, at least one being backed by Nvidia. Yes, Tesla has great manufacturing, but there are other competent manufacturing companies in the world. Not saying Tesla won’t win, just saying it is not clear yet and there are several horses in the race with big money and tech behind them.
He just gave them more value as partners, however they are just customers 😂😂😂😂😂 they been with NVIDIA for many years 😂
No. Synthetic training data will be fine. It doesn't need to be better than Tesla's approach. It just needs to be good enough.
Competition is good in capitalist framework.
Cosmos is analogous to FSD. It's the model that generated the action tokens used to drive the embodied robot. I think the description in the keynote could have been less jargon heavy. Lots of people think it's just a simulator and not the actual world model that's embodied into whatever form factor.
They had Groot & Omniverse which is the simulated data that trains the model (Cosmos).
Is there a need for concern about Tesla losing out on licensing deals in the future. Are Nvidia and Tesla in a symbiotic relationship, or will Nvidia's advancement with this cannibalize this potential market opportunity?
It depends on how far behind the legacy automaker is and also remember if Tesla releases the robo taxi service starting in 6 months from now and then quickly advances it across the rest of the world that's going to put immediate pressure on any of the existing legacy auto companies because Tesla will be out there selling rides for a fraction of the cost of any of the other services including legacy ride, healing services and so the pressure will be massively on legacy auto companies to come to a deal with Tesla. Basically they just don't have time. It's more likely that they will partner with Tesla then try to do it with nvidia's kit. Unless of course they're a BYD which will likely leverage nvidia's kit because it has no other choice really.
The thing you're missing is that NVIDIA DRIVE Sim provides physically accurate simulations for a variety of sensors, including visual sensors like cameras, as well as active sensors such as radar, lidar, and ultrasonic sensors. Tesla does not use, or train its systems on lidar, radar, or ultrasonic sensors, which limits its ability to train for poor visibility conditions.
I have a question for Hans if he is available. I will be brief so the question May not be very understandable. Please let me know if there's a way I can communicate directly with Hans if you don't mind.
I am a mathematical statistician and have a lot of experience with Monte Carlos simulation. I am wondering if this technology could be used to simulate say 100,000 simulations of data input predicting the cash flow of the the payout of the income cash flow for a fixed annuity. I could provide a phone number or email address for further Communications if that might be a possibility of further discussion. Oracle provides a such a simulation model from a product they called Crystal Ball should you require more clarity the type of mathematical statistical modeling for such a Monty Carlo simulation analysis.
As to this argument of is real world data better than simulated data.
What? Hans is missing here and what Elon is also missing is that because they use Omniverse, which is a physics-based simulation environment to ground truth, the simulated generations data that's used to train Cosmos much of the scenarios that Cosmos learns from simulated data fully grounded. Even edge cases. You can't have cows floating in the sky, for example, because within a physically based simulation environment objects have mass which force them to fall to the ground... So when you're generating scenarios within the simulation environment, those scenarios must follow the physics and then you can iterate valid physically valid interactions between those objects based on the mass, velocity and friction and other attributes that they're given within the simulated environment. So long as you were able to capture all of the physics associated with reality inside the simulation environment, you will then be able to simulate any possible edge case... Leaving the only distinction between the two training scenarios being the necessity of compute.
This is different because in the real world data gathering scenario, you are gathering precisely data that is coming from real world driving scenarios. The grounding to the physics is coming in from those scenarios.
Your trying to create a distinction between the two isn't really valid as if you're in the real world for the first time and you experience something, you have no reference as to whether or not that experience is normal for the environment. If I show you a single frame with a wheel in the middle of the frame, you might think that that's normal. But if I show your subsequent frames of that wheel falling to the ground, then you'll understand that that's not normal.
It will take you time to understand the physics of the world after viewing enough frames. The same way that it would take Cosmos to understand the various synthesized scenarios and edge cases that would be generated in Omniverse and used for training it.
Again, the only distinction remaining being the sparsity of the data where the real world data is very sparse. It's basically exactly what you need because it's coming from a existing fleet of vehicles that are existing experiencing valid driving scenarios... Whereas Omniverse can generate valid driving scenarios, but it can generate a whole bunch of invalid driving scenarios as well, which need to be essentially pruned from the training process for Cosmos.
If edge cases could be augmented, they wouldn’t be edge cases.😊
And that's not true. Omniverse allows simulation of all physically plausible scenarios. There are many odd physically plausible scenarios, every single one of them in the real world... That is an edge case is also plausibly physical by definition.
The only difference is the massiveness of the search space of the different trajectories for the possibilities of object interactions between assimilated environment and a real world one. But then that's why you use a physically based simulator like Omniverse to ground the simulations to the physics which is common both to the world and to the simulation environment. Leveraging that stochastic ability to create alternative scenarios ...That's how you reduce the Sim to real Gap.
exactly, all cases can be augmented, just a matter of how approximate to reality.... efficiency, cost, once variables are adjusted etc etc
Legacy have failed at in-house AV development and no chance they can afford a second attempt. They have to rely on outsourcing. The nvidia partnerships, like Toyota's, are DOA.
Once $XAI301F breaks key resistance at $1.2 and $1.5 it's flying much higher!! 🚀
Outstanding Herbert. Your reaction time is amazing. Thanks to you and Hans. Can Nvidia simulate the physics of human stupidity!? I guess it will have to simulate a million war games scenarios to catch all the edge cases. Let’s not forget that Tesla has other advantages it can deploy in the car such as Starlink, Grok, Supercharging infrastructure, …