Ashok ????????????

All right. Um I guess let's get started. Uh my name is Ashok. I lead the Tesla AI team uh here at Tesla. Um so excited to be here. Um I couldn't sorry I couldn't be in person. Um it's a bunch of uh other company work here. Uh but nonetheless excited to share our recent progress. This is the work of the Tesla AI team over the last several months. Um so I'm talking on their behalf here. I'd like to start with some of the recent accomplishments that the team has uh achieved. Uh earlier this year um around June July we launched our robot taxi service where uh if you know if someone was in Austin or in the SF Bay area they can hail a robot taxi service from Tesla and then in Austin uh you know below 40 m per hour you can get a car without anyone inside the passenger seat. Uh so it is just driving uh using this cameras and neural networks uh and adapting to real world conditions. Uh we also did this thing where we delivered the first um customer purchasable vehicle from the Tesla factory in Austin directly to a customer's home uh somewhere else in Austin and we covered our um like 20 30 minutes of driving driving across highway, city roads, uh parking lots, etc. to deliver the car. You can see the car driving itself. And this is a production vehicle out of the factory using production cameras, the same computer that's in every car and so on. Um, every Tesla that's manufactured uh in the US also delivers itself from the manufacturing line all the way to the loading docks that's like a couple miles away. So this year we made a lot of strides in autonomy uh and we continue to scale uh the uh the technology to be even more uh robust and resilient so that we can go to even more locations um get no have no one inside the car and so on. Today I like to touch on some of the technolog

?????????????Robotaxi???????????????????

So first of all uh we switched to having a single large end to end neural network that can take in um pixels and other sensor data as an input and then just produce the next action as an output. So this is um this is what's driving the car. Uh it no longer is doing explicit um you know perception of like vehicles uh road boundaries or things like those. They can be implicit and they can be trained as auxiliary things but then the the entire thing the video raw video streams go in and then the actual actions to take uh are produced by the neural network and this it's been so for the last few years um and this has been extremely good. The reason being uh you know we used to work on an explicit modeler approach because it was so much easier to debug and so on. But what we found out was that uh codifying like human values is really difficult. For example, you could um in order to break for an obstacle or go around, you could um go as fast as possible and then break a bit later or um uh or break earlier for more smooth driving. And this really comes down to alignment of preference to human values as opposed to there's one objective value of like how much braking and when to apply. You don't want to break too early either because then be too slow and if you don't you don't want to break too late either then it's too uncomfortable and this varies from situation to situation. The amount uh at like low speeds versus high speeds is very different. So it's really hard to codify what humans actually want and you can't write it in code uh at least not very well. And also the interface between traditional perception and

???????????????????????????????????

planning is very illdefined. Uh it can be very lossy and lose critical information where uh you might need to propagate the uncertaintities u through the entire system. That's really hard. Um going to end to neural network also allows for your compute to be homogeneous. Um because if you have traditional branching compute you can have nondeterministic amounts of latencies and you and in a real-time system you don't want to have the kind of u latencies you you want to have predictable deterministic latencies and such uh single neural network doing the entire thing offers that uh and overall this is what we believe as the uh you know the right path to solving robotics as opposed to this kind of modeler brittle systems that were developed earlier. Here I have to give a few examples of like why this is hard. For example, this vehicle is um driving on like a birectional road and then to avoid this puddle you got to go to the oncoming side. If you have to write this using explicit cost functions, it's going to be really hard because you know typically you don't want to go into the oncoming side. But then you know what is the pain of driving through a puddle versus the you know the risk of going through an oncoming uh lane is um it's like hard to write an explicit code. But here obviously if you look at the scene it's pretty obvious that you can go here. You have a lot of visibility there's no oncoming vehicle coming and you should definitely go into the oncoming side. It's also like hard to understand intent. For example here on the left side uh there's a bunch of chicken crossing the road. I'm not sure if the video is um clear for you guys but um there's a bunch of chicken crossing the road. This is driving on self-driving software and it's just waiting for the

????????????????????????????????

last chicken to cross here. especially waiting the chicken crosses the road and then the car proceeds on its own. This is all just driving on its own. In this case, the car was smart enough to understand that the chicken actually out crossing the road and was patient enough even though there was no collision risk. It was just the right thing to do. >> But in other cases, >> I think it's going to say, look, it's going around him. It's going around him. Oh, look. IT'S CHANGING ITS MIND. IT'S BACKING UP TO GO AROUND THEM. LOOK at that. It backed up to go around the geese. >> This is crazy. This is crazy. >> Yeah. In this case, the the model understands that the geese are actually not crossing, they're just stationary and decided to reverse and then go around them. So, it's not this is something it's very hard to write in explicit um code what is the right thing to do here. But looking at the scene end to end and for humans too it's pretty obvious what was the right thing and that is what that is why end to end is way better than a modeler approach that said it's not it does it's not that easy to develop such a system it's pretty difficult in fact uh and I'd like to like touch on three different like aspects of it that make it really difficult first is the curse of dimensionality um what I mean by that is that the input um context length is quite enormous especially Our cars have seven or eight cameras depending on the vehicle. Uh each produces you know 5 megapixel um camera streams running at like high frame rate. Uh and even if you want to have like you know 30 secondond context or something like that uh along with the other necessary inputs like you know the routes and then the vehicle speed and other kinematic data you can easily be more than uh you know billions of tok

????????????????????????????????

of mapping two billion tokens down to roughly two tokens which is you know what is the next steering and acceleration that the car should take and thankfully Tesla has a huge advantage here uh which is the Tesla fit is quite large so it's basically has access to the Niagara Falls of data uh and it's up to Tesla to use this data to solve this problem. Um, you don't want obviously speurious correlations from the two billion tokens down to the two tokens. You want the right correlations as to why those two output tokens must be the right tokens. Um, so what we do is we re refine this 500 years of driving which obviously is more data than we could ever store on our clusters uh to the the the essential amount of data uh which you know covers the overall spectrum of driving. Uh there's different ways we can collect this data based on like explicit triggers or small neural networks targeting specific scenarios um or like constantly evaluating what the model predicted versus what actually happened uh and things like those. We can use such tricks to extract just the right amount of data uh and the right quality and the kind of data that we need for our training. Here are some examples of like what that kind of like corner case data might look like. These these are examples in manual driving like you know people are just still driving and then they once every now and then they encounter such rare scenarios um and we we can have triggers that catch these uh and then this is very very useful data to train on and this is not normally accessible uh you can't stage this that easily because it it requires the stationary scene the speed of all the vehicles in state space this is really hard to achieve and Tesla has a unique advantage here that it can tap into the entire fleet to get thi

???? + ????????????????????????????????????

the last second but proactively safe. Uh I would like to show a demonstration of like how that manifests like this is again a self-driving system in operation here just on a highway. Um, I'll let you watch the video for a second. Yeah, here the car in front of us spun out and then um hit the barrier and the self-driving system pulled over safely. Uh but what is quite interesting is that if you watch the video again like pause where the car uh the ego car breaks like here if you notice this car here it's like spinning out of control and it's going to hit the barrier and it's going to bounce back into our lane. uh it requires so much intelligence to know that this is not a first order collision here because this at this frame the Tesla already determined that this vehicle there's something wrong here uh and started applying the brakes. It did not wait for the car to hit the barrier and then bounce back for it velocity to change or something like that. This is a second order effect that it needs to model and it's it could have also attributed this to a lane change of this vehicle but it did not. It understood because it was requesting roughly 4 m/s square of braking which is not a light amount of braking. Um and this is only possible if you have a ton of data uh and cover all these corner cases. Uh and that is how you can provide both you know safe and also a smooth ride because a less intelligent system would wait until the direction has changed or actually some bad event has happened but it requires a lot of intelligence to understand that something is going wrong way before it actually goes wrong. uh and that is what you get with an end to-end system with a ton of data. Secondly, um so it's an end to end system. How do you guys debug? How do you actually, you know, deve

???????????????????????????????

some decisions, does it understand the scene and so on. All of this helps uh interpret the model's understanding and also sort of like gives some guarantees around the safety of the entire system. So in practice it sort of looks like this where you can take in arbitrary sensor data and prompt it to produce arbitrary things but in the end the only thing that actually matters in the car is the control actions that it produces. Everything else is auxiliary but they can be quite helpful for predicting the correct control actions. One specific uh task uh that I would like to go over is the uh gshian splatting which you know in the last few years has been uh quite prominent in um the field. What I'm showing here on the left side uh is the traditional gshian splatting. Uh I know it looks really bad and it's not like you know taking some bad example and showing the problem typically is that vehicles drive in sort of a linear fashion uh going forward. Uh, and there's not not a ton of baseline. Um, and because of this, if you just took the same camera views of the vehicle's motion and ran traditional gshian splatting, you know, I don't know, nerf studio or what have you, um, with the same camera views, the views close to the train views look great. But then if you um go to normal views that are far away from the train views, it breaks down quite a bit. But what we have in the middle column here is Tesla's variant of gshian platting where the same model can produce uh gshians. It generally is much better uh with the same limited camera views that the left side has. It can also produce semantics as shown in the third column. And the cool thing is that it can run ridiculously fast uh compared to traditional caution splatting which can take you know tens of minutes. U also requires u

???????????????????? occupancy??????????????????

to um scenes quite rapidly. Um if you use the same traditional splatting a lot of normal views might look blurry or like fuzzed out but as I'm rotating here in 3D space a lot of the structure remains intact and such interpretable representations uh can be used to debug the system you know like if it's going faster or slower you can easily observe okay is it um safely avoiding some obstacle or Like I mentioned earlier, we can also use uh natural language to interact with the model. The same model can sort of like point to things, explain why it made some decision. Um we shouldn't need all of this to drive the car in real time. But then if you do need it, you can always uh think longer and produce um reasoning tokens to then uh produce the right action that is consistent with the reasoning of the entire thing. Um and again it can be the same same model that's running in the car uh but based on the situation can uh use it or can directly produce the control actions if uh it's not required to reason in such detail because if you keep reasoning in detail for every single thing then it's going to take too much latency. Um but then wherever is needed uh it could reason longer to produce the right right answer. The last piece that I would like to touch on is the u evaluation which actually is the most difficult of these three problems uh I I mentioned. If you just train a bunch of data from u you know drivers and then train models um and just like scale it your openloop performance like look look amazing but

Gaussian splatting ?????? 3D ??????????????????

then it may not translate well to how the car actually drives. There's a lot of reasons for this. I don't want to get into into the details there but then it is a very important problem that requires a lot of attention to get good performance out of the system. Um it also requires like thorough uh evaluation sets. It's not just you know if you randomly sample data from your fleet most of it is going to be boring highway driving you don't want to just be evaluating there that's where the early data data uh engine approach that I mentioned uh comes in handy where we can build a balanced uh fully covering eval set which um is extremely important but very tedious work. One uh great thing we figured out was that we can use the um uh cheap to collect uh state action pairs inverted and then um basically build a world world simulator where given past past state and then actions it can synthesize newer states um this condition on the actions. Um this is quite easy to collect uh because you can get it for free. You don't need optimal driving or

??????????????????? reasoning tokens??????????????

something like that. Any kind of trash driving is good enough for this kind of um simulator because it need to simulate the edge cases. Then once you have that, you can connect that with our policy neural network that's driving the car. Uh and then they can both uh run in sequence uh in a loop to uh sort of simulate the world. So here's an example of uh our um roll out through a learned neural network simulator. I let you watch the video for a second here, but the entire thing is generated video. There's eight cameras. Uh the top row is all the front cameras of the car uh driving forward. Uh the middle ones are the left and the right side cameras on the bottom row is the uh rearfacing cameras. All eight cameras are generated simultaneously by a single neural network uh and uh takes action um as an input. So you can steer the network. I'll show you in the next couple of sites uh how that that can work. But yeah, I'm usually blown away by the consistency of generation across different cameras. Uh even the vehicle rims are consistent, the traffic lights. Um and this is over like a minute minute and a half long generation of um eight 5 megapixel uh video streams. In addition, you can take like past issues, you know, like take a year old issue uh and then uh you can rerun the latest neural networks to see how they would perform. On the left side here was the original failure where it was maybe a bit too close to the pedestrian. Um and then on the right side we are evaluating a newer neural neural network that's driving the car. Uh and you can see that the policy uh starts offsetting

???????????????? learned world simulator ???????????????

way earlier as soon as the pedestrian emerges from behind the car. Um and it's super useful for such kind of evaluation because you don't want to just re-evaluate on newer mileage. You just want to if you have a database of older issues, you just want to replay them and verify that you are doing a good job on this past issues. You can also synthetically create new issues. Uh in this case the original video on the left side uh that vehicle was just going forward uh on its own lane. But then on the right side we can make the condition the vehicle to cut across our path and then inject this kind of adversarial event to test the systems corner cases and you can see that the scene the rest of the scene remains consistent. All the other vehicles are moving the same way just that this one vehicle cuts across and that that's how you can make lots of synthetic uh data sets to verify the corner case performance of a self-driving system. If you uh are fine with reducing the test time computer a little bit, uh you can get uh close to realtime rendering performance. This is the same model that's rendering the eight cameras uh with uh lower test time uh compute and you can see that you can actually drive this uh in real time. the FPS is high enough even though it's generating um eight 5 megapixel streams uh in in in parallel. So obviously you can see that the it's responding to the uh users uh driving commands. They can like steer uh and break and then sort of like navigate the world just as if they're driving in the real world. But the entire thing is neural network generated video streams. I think the driver is trying to do something adversarial by you know going on the curb and then just demonstrating that the generalization of the simulator is quite good. Yeah, this entire video

??????????????? rollout??????????????????????

is quite scalable. scales across different vehicle platforms, different uh spatial locations, different weather conditions uh and overall offers a very safe uh comfortable and fast ride to the uh users. There's one more thing which is it's not just scaling self-driving to all the vehicles but also scaling it to other forms of robots. The humanoid robots too. Uh Tesla builds humanoid robots, we call them Optimus. Um and the same technology that we developed here for self-driving transfers most seamlessly to other forms of robots too. Uh I could show one example here like the the video generation uh also works for u u Optimus. So this is Optimus navigating around the Tesla factory and all of these are generated videos and you can see that uh they are uh all quite consistent. You can also just like how we did for the car action condition it. So here the top left is the uh robot just going straight. That was the ground truth. Then uh you can have different actions like going left or right or some other direction and then it can um correctly generate the pixels um for the action u and it's the same neural network um with just some more data added from optimiz generalizes to um other robot form factors too. So yeah in conclusion um Tesla is all in on robotics. The entire company is just focused on uh producing intelligent useful large scale uh robots for helping um everyone in the world. Uh and if you are working on uh AI, robotics, computer vision etc. I'd highly uh you know encourage working at Tesla because it has like extreme real world impact. Uh it's fully in on solving robotics. Uh the team is super pumped. uh you know I think it's a great place and this is this is the best place to work uh uh on AI right now. >> Thank you. Vasculate.

?????????? Optimus???? Tesla ??????????

Hey everyone, my name is Ashok um Eliswami. I work at Tesla. I've been there for the last 12 years and I lead the Tesla AI team. Um thanks for having me today. Uh would like to present about the things that the team has been working on and like you know what is the uh you know the mission of the team and what is the upcoming road map of uh you know the team's work to start out Tesla's mission is producing amazing abundance for the entire world. Um we you know Tesla has been historically known for producing vehicles. Uh but then the more uh important thing is that these vehicles can drive themselves. They have all all the vehicles have uh all the necessary sensors and compute to do autonomous driving. Here you can see our um robot taxi service operating in Austin. Uh there's no one inside the car that's driving the car. This is just a passenger hailing a ride and then the car just uh drives them around all of Austin. This is publicly available. um as of like earlier this month. Um, previously we used to have safety monitors to just like you know supervise FSD but now that's gone and then people can just you know get in the car you know they already have the address punched in so they can hit start trip and then the car just takes them uh wherever they want and these are public roads uh with like you know dense traffic and everything and this is all driven by uh cameras and AI and the future of the company is not just like vehicles we also have like all working on like humanoid robots uh that sort of like go to the next level of functionality. Um like the Tesla vehicles are like low cost scalable solution to transport and humanoid robots are a lowc cost scalable solution to automate like all physical work. The reason for humanoid robots is that they you know they're backw

????? Tesla ????Robotaxi ?????????????????????????

safety, you can get some safety improvements compared to not having any active safety. It's very clear that using self-driving to drive you around as opposed to manually driving the car definitely makes you safer over billions of miles of study. And this is not just on the highway or something like that. Even highway off highway in both cases uh it is at least 2x better than uh the baseline of driving manually. So you know if you have a Tesla you should definitely uh use self-driving. Uh I personally use it all the time. So today we are going to cover you know like what makes a software work u sort of what are the challenges in building such a system uh and like you know how do we evaluate the safety of this first of all um Tesla uses an end toend driving system to u create the self-driving software uh you might think that in the modern era obviously you should have like you know end to end systems but this is uh not uh contemporarily agreed upon you know for example many competitive solutions use uh sort of model approaches that have separate like perception stacks, planning, prediction stacks and so on. Uh and Tesla has foregone all of those systems to have a single end to end neural network that takes in raw sensor inputs. Uh which is predominantly just camera videos from the eight cameras in in the car but also things like you know the navigation instructions uh kinematic states like the vehicle speed, steering etc. uh audio uh to just come up with how the car should behave uh and like what is the next action in terms of you know steering uh or like uh jerk or something like that. uh and the reason for this is that u I guess I can go over uh some of the reasons first of all like codifying uh everything in like you know rules based or like doesn't have to be even rul

??????? + ?????????????AI ?????????

based on independent rules or independent like logical elements here. Like here's one more example. Here's some, you know, this chicken that wants to cross the road on the left side here. Um the car actually on FSD patiently waits for the chickens to cross on the left side. And if you're paying attention, there's like one last chicken. There's a stagler chicken there. Uh it still like waits for it. >> Yeah. Waits for all the chicken to cross and then autonomously proceeds. >> Why are they crossing? >> Yeah. they want to test FSD. And on the right side, um there's a similar scenario. There's a bunch of geese on the road. Um and the there's no one inside the car here. This is using the smart summon software. Uh then the geese, they're not crossing. They're just like, you know, they're just like standing there and the car just backs up and then goes around them. So you have to like see this like subtle cues of how these like you know legs of these birds are moving to know that they're either intending to cross or not intending to cross and do the appropriate action. Like if you had explicit perception prediction and so on what are we going to do like have a chicken leg detector and then like predict how the chicken's going to move and then uh take action. I just think it's too complicated in an end to end system. All of this information can flow from the pixels directly to control. Um it also you know overall uh is has determin deterministic latency like for driving you have to produce actions in the real time. You you can't take um very long time to process regardless of how long we think the world is changing around the car. So the latency is quite important and with neural networks it's way easier to control the latency. You get a lot of determinism um to take actions.

???????????20 ? token ?????????? token?

network. And then in the end you had to produce like two actions you know roughly what is the steering angle for the next time step and what is the uh acceleration on brake uh for the next time step right like so you're compressing down this ridiculous amount of in uh information like two billion tokens input and it produced two actions this be a trivial job if you just randomly took two bytes and then like piped it out right but you had to take the two correct actions and so you have to causally understand what amongst this giant stream of bytes and bits uh is um uh is The reason why I should take these two correct actions that you know for example the human took or like whatever optimize the reward like learning that correct causal mapping is very difficult because the network can learn all kinds of speurious correlations you know I should break here because the tree branch was moving certain way like you don't want to learn speurious correlations you want to learn the correct correlation that I'm breaking right now because the vehicle in front of me has the know turn signal on or something like that um luckily the Tesla fleet is quite large uh and you can collect a significant amount of data from the Tesla fleet um to rough order of magnitude like the entire fleet can produce 500 years of driving data on every single day. But obviously most data is like boring like you know most people are just driving on the highway minding their own business. It's not interesting. So and it'd be a tremendous waste of resources to literally collect all this data and then filter afterwards. So what we do is instead just identify what is interesting data. There's a lot of work that goes into identifying what is interesting. But once you identify the interesting data, then you know, ok

reasoning ???? traces?3D reasoning?????????????

software realized this quite early and then break automatically uh and then like avoided this you know what could have been one more collision here and what I like to emphasize here is like how early it reacts to this stuff and you can see like sort of this like velocity here of the vehicle um it starts breaking already at this time step uh as as seen by uh this like planned velocity. The the the lead vehicle in front of us hasn't even crashed here. It's just like it could have been doing a lane change. It's just going from this lane to this lane but it already realizes that something is wrong here. Something is out of uh something is oddier. So it already starts applying pretty significant amount of break. what you know people might uh be a little bit shook because it anticipates something bad is going to happen many seconds in advance of the actual event happening. Um so you you need the predictive intelligence of okay this the y rate of this other vehicle is more so than necessary for a lane change maybe they're going to crash what if they crash and come into our path hence we must break right now and this kind of like third order intelligence needs to be done uh in real time uh while processing you know like there could be like this these vehicles could be coming here there's like so many other bites here right like it needs to focus on the correct things to make this conclusion to hit the brakes in a confident manner uh and And that that's what all this data is required for not for like you know driving around the block here. So it's end to end system you know like does it mean that you can't understand what it's doing? Um how do you debug if there's a failure? Um even though it's an end toend system there's all kinds of probes you can attach to the system. Uh and

Gaussian splatting ???? 3D ????????????

poorly. So if you took a video clip from the vehicle and you run traditional gshian splatting, you get the video on the left most side and the middle video here is Tesla's own uh like generative gshian splatting which has very good generalization even beyond the uh train views and you can see that uh you the semantics are um quite correct on like which type of vehicle it is and so on. Um and all of this happens way faster than traditional gshian splatting. traditional gshian splatting can take like for example 30 minutes uh or or even more sometimes versus Tesla's own system can run on the order of hundreds of milliseconds um doesn't require any kind of initialization it's all neural network based and this is part of the same neural network that produces the control action so it understands geometry and can produce and explain the 3D geometry around the vehicle um along with the action here one more demonstration of this so it takes in the videos at the top as the the network takes the on the top as an input. I can like entirely generate 3D world that's interactable. You can like move around uh and you can like see what the network thinks the world around it looks like not just in the camera's point of view but in a 3D view. Uh so this really makes it understand um this basically helps with also learning the correct correlations. You know it knows that these pixels in the image space or video belong to this vehicle and so on. In addition to um uh geometric reasoning, you can also do just text based reasoning. Uh for example, here there's some road closure. Uh you can reason out in a mix of language and also pointing to things on the videos. You can say okay the detour sign here I must turn left. Uh it like so it has this comprehensive understanding of the world around i

???????????????????????????????

this to the policy neural network that takes the actions. So the uh world simulator produces the next video frame and then come uh and the other sensor measurements. Uh this is consumed by the policy neural network that then takes the next action and then they just go in round robin fashion simulating the world and these two neural networks are trained uh differently. So they they don't actually like the um the world neural network can use privileged information that the policy does not have access to uh and that is how they can be verified independently. So here's an example of uh video that is generated by the neural network. Um this the neural network produces the um uh eight camera videos uh at 36 fps and 5 megapixel resolution. Uh this is a very long generation. This I think this video is like about a minute long. Uh but all the pixels here are generated by neural network. And you can see how, you know, the vehicle's moving around the van and everything is consistent across different cameras as the other objects move around. Yeah, as I mentioned, every every camera, every pixel here is fully generated. There is no real pixels here, even though it looks quite realistic. The the reason for doing this again is for evaluation purposes. So what we want to do is we add some historical issue. We want to replay them on newer policy models to verify that have they solved the issue. Uh for example on the left side here there was an historical issue where the you know the vehicle was driving a bit close to the pedestrian and then the human driver intervened on him and then now we have a newer neural network that should have improved this and then we try to replay this on the same video clip um and then verify that now have we improved on the performance compared to the origin

??????????????????????????? robotics foundation model?

make this work in a real-time basis. Um here uh you can see that this these are all generated pixels. They're driving in a fully synthetic world uh using just like you know some game engines like steering and um pedal control and you can just like drive around uh sort of in real time and all the pixels here are generated. Obviously, it's running at real time, so it requires some reduced test time compute. Um, but then you it's still pretty reasonable for uh driving in real time. And all of this doesn't uh just uh is needed for uh self-driving like the the end to-end driving neural network is not just driving neural network. It's actually a foundational neural network for robotics. The same for the simulation neural network too. Both of them are trained on common data across all of the robots. Um here you can see that the same video generation network also generalizes to uh generating indoor scenes for optimus to walk around. These are also like controllable. Uh here you can take a single u video clip and then convert it to be an action based one and you can take the action to go straight or turn left or turn right and then you can see that the uh video generation properly reflects the actions. So it water I explained earlier today is not just for uh the vehicle but also for human robots or in fact any robots for that matter. Um you also see it work for uh manipulation. It can you know is this again action condition where it can uh open a drawer and then you can see that the video um sort of like reflects opening a drawer uh or like picking up an object or something like that. Okay, the benefit of using uh generative like world models or like uh neural network based world models is that yeah it can basically like you know everything is green everything is great you shoul

??????????????????

real world. Um thus improving productivity and value for the entire world. And if you're interested in working on such problems, please come join the team. There's a lot of uh interesting problems still. It's u while it works quite well, there's still a lot of other challenges that are technically interesting and challenging and very meaningful to this world. Uh and with that, thank you everyone. >> [applause] >> We have time for a couple of questions. >> Yeah. Go ahead. >> Why are you so convinced that you can solve this self-driving uh challenge only with cameras? I mean, you seem quite convinced. >> Yeah. It's like, how did you get here today? >> With a >> with a Okay. All right. [laughter] But many people drove here with their own eyes. Um, and then, you know, obviously you can like walk around this building with your own eyes. Like it's it's so obvious that you can solve this with cameras. For me at least. Um, like why wouldn't you solve with cameras? It's like 2026. Like it's uh it should be solved with cameras just like how every other human and animal lives around this world. It's just like you know very simple sensors. The self-driving problem is you know thought of as a sensor problem. It's actually not a sensor problem. It's an AI problem like you need to understand the world. You need to understand what other people are going to do. Uh and the cameras have enough information um already. It's just the problem of extracting the information which is an AI problem. Um the the sensor solution was developed back in 2008 or whatever during DARPA DARPA days when there wasn't enough intelligence around back then to extract this information. It was and that's why you um need to depend on all these other sensors back in the day. But nowadays intelligence has grown trem

???????? camera-only???/????????3D ???????????