October 29, 2025 • 28 mins
Tesla is one of a kind in autonomy, focusing on its benchmark autopilot and innovative neural networks that emulate human decision-making.
Tesla processes vast real-world and synthetic data to enhance adaptability in driving. They discuss the upcoming AI5 chip, set to transform Tesla's computing capabilities, and speculate on the economic implications of a future with humanoid robots and autonomous vehicles.
------
🌌 LIMITLESS HQ: LISTEN & FOLLOW HERE ⬇️
https://limitless.bankless.com/
https://x.com/LimitlessFT
------
TIMESTAMPS
0:00 Intro
0:53 Behind the Scenes of Tesla
2:46 The Secrets of Neural Networks
6:51 The Curse of Dimensionality
7:57 Data Curation and Edge Cases
13:13 World Modeling and Simulation
16:15 The Future of Humanoid Robots
18:04 The Financial Landscape of Tesla
21:48 The AI5 Chip Revolution
25:38 Closing Thoughts on Autonomy
------
RESOURCES
Josh: https://x.com/JoshKale
Ejaaz: https://x.com/cryptopunk7213
------
Not financial or tax advice. See our investment disclosures here:
https://www.bankless.com/disclosures
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Josh:
One of the most important technologies in the world that is happening as we (00:03):
undefined
Josh: speak every day is the rise of autonomy, and particularly around autonomous robots. (00:06):
undefined
Josh: Robots can be many things. Robots can be humanoids, they can be cars. (00:11):
undefined
Josh: And today we're going to talk about both, because there's one company that is (00:14):
undefined
Josh: at the frontier of both of those areas, and that's Tesla. (00:17):
undefined
Josh: Tesla has the most unbelievable set of autopilot software that I think exists in the world. (00:21):
undefined
Josh: I've been using it personally for eight years now. And it's been amazing to (00:27):
undefined
Josh: see how good it's gotten. (00:30):
undefined
Josh: And EJS, now there's, for the first time ever, we have the secrets. (00:31):
undefined
Josh: The secret sauce that shares exactly how they've been able to get autonomy this (00:35):
undefined
Josh: powerful, this impressive. (00:39):
undefined
Josh: And there's now very clearly a world in which I can imagine waking up in the (00:40):
undefined
Josh: morning, getting ready to go to work, stepping outside, and there's a cyber (00:44):
undefined
Josh: cab waiting for me outside that will just take me wherever I want for a fraction (00:47):
undefined
Josh: of the cost that it takes for a normal driver. (00:50):
undefined
Josh: And I think this is an incredibly powerful unlock and to see a behind the scenes (00:53):
undefined
Josh: of this is awesome so the entire episode today is behind the scenes of the most (00:57):
undefined
Josh: impressive new front-end tier technology that exists. (01:02):
undefined
Ejaaz: I think what i'm most excited about today josh is the fact that i've always (01:05):
undefined
Ejaaz: thought tesla ai and robotics is so cool but i i just don't know how any of (01:10):
undefined
Ejaaz: this works and they've refused to tell us and finally they've they've spilt their secrets today (01:15):
undefined
Ejaaz: to quickly paint some context for the listeners here, up until yesterday, (01:20):
undefined
Ejaaz: we only thought of Tesla AI as something called a neural network. (01:24):
undefined
Ejaaz: That's their secret source. And a neural network can be thought of as a software (01:29):
undefined
Ejaaz: program that is designed to function like the human brain. (01:33):
undefined
Ejaaz: So it takes in information and it discovers patterns, trends, (01:37):
undefined
Ejaaz: and it can also sometimes make predictions. (01:41):
undefined
Ejaaz: Now, this contrasts directly to some of Tesla's competitors, (01:43):
undefined
Ejaaz: which do self-driving and robotics in a very different way. (01:47):
undefined
Ejaaz: They take more modular and sensor-driven approaches, right? (01:50):
undefined
Ejaaz: The reason why Tesla's neural network is so special is they have an end-to-end (01:54):
undefined
Ejaaz: neural network, which means that they feed a bunch of raw data from one side (01:59):
undefined
Ejaaz: and out comes the output, which is an action. (02:04):
undefined
Ejaaz: In this case of Tesla cars, it would be driving, steering, and acceleration. (02:07):
undefined
Ejaaz: And they took this approach for a few different ways. (02:12):
undefined
Ejaaz: The most important being, it's really hard Josh to codify what human values (02:15):
undefined
Ejaaz: are and what I mean by that is let's say in this example that you're seeing (02:21):
undefined
Ejaaz: on your screen right now you are driving your car and there's a massive puddle on your lane but (02:25):
undefined
Josh: You see that you. (02:30):
undefined
Ejaaz: Could potentially drive into the oncoming lane to skirt around it now for humans (02:32):
undefined
Ejaaz: it's really easy to do that right it's like okay maybe I should just go through (02:36):
undefined
Ejaaz: it because there's no cars coming but for a machine to do that it requires a (02:39):
undefined
Ejaaz: lot of effort. It's hard to hard code. (02:45):
undefined
Ejaaz: So that's one special thing around the neural network. But Josh, (02:47):
undefined
Ejaaz: I want to jump into the secrets. (02:50):
undefined
Ejaaz: Can you lead us with the first one? (02:52):
undefined
Josh: Well, what you mentioned is really important, the end-to-end stuff. (02:54):
undefined
Josh: And I want to walk through a little experiment. (02:57):
undefined
Josh: So when you kick a soccer ball, I think this is an experience everyone's kind (02:59):
undefined
Josh: of went through, right? What do you do when you kick a soccer ball? (03:02):
undefined
Ejaaz: Yeah, I see the soccer ball coming towards me. I kind of prepare my legs ready to kind of kick. (03:05):
undefined
Ejaaz: I'm right-footed, so I'm kicking with my right foot. (03:12):
undefined
Ejaaz: And then I guess the rest is kind of intuitive, Josh. I just kind of run up to it and kick it. (03:14):
undefined
Josh: Yeah, yeah. And I think that's exactly the point is when you kick a soccer ball, (03:19):
undefined
Josh: this is something a lot of people have experienced. (03:23):
undefined
Josh: You're not actually thinking about all the parts of kicking a soccer ball. (03:25):
undefined
Josh: You're not thinking of where it is on the ground, where your ankle is, (03:28):
undefined
Josh: where your knee is, where your leg is, the positioning, how hard you're going (03:31):
undefined
Josh: to kick it. It just feels very intuitive. (03:34):
undefined
Josh: And with a lot of other car companies, they're hard coding these intuitions as code. (03:35):
undefined
Josh: So it does have to think about each section. It does have to calculate each (03:40):
undefined
Josh: section. And what's different about Tesla and what we learned from this article, (03:43):
undefined
Josh: this is from Ashok, who is the person who's in charge of Tesla AI, (03:46):
undefined
Josh: is that they use this thing called end-to-end neural networks. (03:50):
undefined
Josh: And what does that mean? In like a fun, simple way, it's basically the intuition (03:52):
undefined
Josh: that you just described with kicking a soccer ball, the AI model, (03:56):
undefined
Josh: the chip on a car is able to emulate that. (03:59):
undefined
Josh: So instead of making these minute decisions all the way through a fixed decision (04:01):
undefined
Josh: tree, they're able to take a ton of data and use these things that we've learned (04:05):
undefined
Josh: over time, which are gradients and weights, and basically move the gradients (04:08):
undefined
Josh: and weights throughout the decision process to reach an end goal. (04:13):
undefined
Josh: So if the end goal is to kick a soccer ball, there's a very clear stated end goal. (04:16):
undefined
Josh: And the neural network's job is to figure out the full sweep of gradients as (04:20):
undefined
Josh: it goes across to get to that end goal. (04:24):
undefined
Josh: And it uses a bunch of this training data that they collect in order to get there. (04:26):
undefined
Josh: So this is this remarkable technology that breakthrough that they have. (04:29):
undefined
Josh: And they have some really interesting examples here. (04:32):
undefined
Josh: So in the case of the ducks, like we're looking at an example on the screen (04:34):
undefined
Josh: right now, there's ducks standing in the middle of the road. (04:37):
undefined
Josh: When you're coding an AI system, when you're coding a car, you're not hard coding (04:40):
undefined
Josh: in, if you see ducks, do this. (04:44):
undefined
Josh: What the car is understanding intuitively is like, okay, there's an obstacle (04:46):
undefined
Josh: here and they are ducks. They're not moving. (04:49):
undefined
Josh: The interesting thing is the example above is the car recognizes that the ducks (04:52):
undefined
Josh: are actually moving across the road. (04:56):
undefined
Josh: So it knows to wait and then it could pass once they've moved. (04:57):
undefined
Josh: But the second one, it notices they're just kind of chilling. (05:00):
undefined
Josh: The ducks aren't going anywhere. And what does it do? It understands that intuitively (05:03):
undefined
Josh: and it is able to back up and then move around them. And that's the difference (05:06):
undefined
Josh: in how Tesla does it versus some other companies is they're not hard coding (05:10):
undefined
Josh: a series of fixed parameters. (05:13):
undefined
Josh: They are doing it all entirely through these neural networks. (05:15):
undefined
Ejaaz: If we move on to secret number one, Josh, it kind of explains how they're able (05:17):
undefined
Ejaaz: to achieve this at a pretty high level, right? (05:22):
undefined
Ejaaz: So it's titled The Curse of Dimensionality. And what it basically describes (05:25):
undefined
Ejaaz: is you can imagine for a car to self-drive, it requires a ton of data. (05:30):
undefined
Ejaaz: I think Tesla, the average car, has about seven cameras. (05:35):
undefined
Ejaaz: It ingests a ton of audio data, a ton of navigation GPS data, (05:39):
undefined
Ejaaz: and kinematics. So speed is tracking your speed. (05:44):
undefined
Ejaaz: And so all this data is roughly equivalent to 2 billion tokens. (05:46):
undefined
Ejaaz: And if you think about it, it needs to run through this end-to-end neural network (05:51):
undefined
Ejaaz: that you just described, Josh, and it needs to output pretty much two tokens. (05:54):
undefined
Ejaaz: One token, which determines which way the car should steer, and the other token (05:58):
undefined
Ejaaz: determining how fast should that car be at that point? Should it decelerate (06:03):
undefined
Ejaaz: or should it accelerate? (06:07):
undefined
Ejaaz: And you can imagine this is an incredibly nuanced and complex process. (06:09):
undefined
Ejaaz: And the way that the Tesla neural engine or the neural network is designed is (06:13):
undefined
Ejaaz: it has really special data lanes that process this data in a very nuanced way (06:18):
undefined
Ejaaz: to understand what exactly it needs to map onto when it comes to steering and acceleration. (06:23):
undefined
Ejaaz: Now, you might think that's pretty cool, but Tesla's secret source when it comes (06:28):
undefined
Ejaaz: to this particular component is the driving data, right, Josh? (06:33):
undefined
Ejaaz: So they get access to all the camera data, audio data, GPS data that I just (06:37):
undefined
Ejaaz: mentioned from their entire fleet of Tesla cars. (06:41):
undefined
Ejaaz: So the equivalent of data that they get every day is something crazy like 500 (06:45):
undefined
Ejaaz: years worth of driving data. (06:50):
undefined
Ejaaz: Now, you can imagine if it processes this amount of rich data, (06:52):
undefined
Ejaaz: and not all of that data is important, right? It's kind of like the same kind of standard things. (06:56):
undefined
Josh: Over those years of data. (07:00):
undefined
Ejaaz: You get access to the one or two random nuanced incidents which feed in and (07:02):
undefined
Ejaaz: improve the collective intelligence of the entire Tesla fleet. (07:07):
undefined
Ejaaz: So whether you're on the other side of the world driving a Tesla or you're in (07:11):
undefined
Ejaaz: the local neighborhood, you still benefit from the same types of improvements. (07:14):
undefined
Josh: I want to talk a little bit about the scale because you mentioned 2 billion (07:18):
undefined
Josh: inputs and it's kind of difficult to comprehend what 2 billion actually means. (07:20):
undefined
Josh: And as a good example, I want you to imagine your phone processing every TikTok (07:25):
undefined
Josh: that exists on the platform every single second in order to determine the next (07:31):
undefined
Josh: turn. That is two billion inputs. (07:35):
undefined
Josh: It is an astronomical amount of data. You're basically, you take the whole TikTok (07:37):
undefined
Josh: catalog every second in order to make every decision and you distill that entire (07:41):
undefined
Josh: data set into two single points. (07:45):
undefined
Josh: And it's just, it's a remarkable amount of compression and then a remarkable (07:48):
undefined
Josh: amount of precision to make the right decision over and over and over again, (07:52):
undefined
Josh: and then adjust and calculate as things change. (07:55):
undefined
Josh: The way that they do this, they're not doing this raw. They're not actually (07:58):
undefined
Josh: ingesting all this data. (08:01):
undefined
Josh: They have this data curation process that they use in order to help them kind (08:02):
undefined
Josh: of figure out what is important and what is just noise. (08:06):
undefined
Josh: And what they do, and we have a great example on screen here, (08:09):
undefined
Josh: is they pick the juiciest clips. (08:11):
undefined
Josh: It's like kind of curating like a viral playlist and they use it to train the (08:13):
undefined
Josh: AI on these weird scenarios. (08:17):
undefined
Josh: So we're seeing on the screen, there's someone rolling through an intersection of wheelchair. (08:18):
undefined
Josh: It's actually very funny to see and scary to see what types of things happen. (08:22):
undefined
Josh: I mean, this is crazy. Two cars crashing right in front of you, (08:26):
undefined
Josh: driving on a snow blind street. (08:30):
undefined
Josh: There's kids that are running out in the middle of the road. (08:31):
undefined
Josh: There's these tremendous amount of edge cases that are really difficult to understand. (08:33):
undefined
Josh: And because of the 500 years of driving data every single day that they ingest, (08:38):
undefined
Josh: they're able to analyze and to kind of sift through. (08:43):
undefined
Josh: And they've come up with systems to curate the most viral clips, (08:45):
undefined
Josh: not viral, but the clips with the most implications of safety that are kind of the weird edge cases. (08:48):
undefined
Josh: And then we have this example here. Do you want to walk through the chart that (08:54):
undefined
Josh: we're seeing, because it's really fascinating how the car can kind of see it before the human does. (08:57):
undefined
Ejaaz: Yeah. So what's interesting is when I first watched this clip and for those who are listening, (09:01):
undefined
Ejaaz: it is a car driving on a very rainy evening on the highway and a car in front (09:06):
undefined
Ejaaz: of it kind of crashes out and goes and starts to spin and kind of enter its own lane. (09:12):
undefined
Ejaaz: When I first watched this video, Josh, I didn't even notice the car spinning (09:17):
undefined
Ejaaz: out because it happens so far away. (09:21):
undefined
Ejaaz: And so what's effective about this particular video is, given everything that (09:24):
undefined
Ejaaz: you just described, the Tesla self-driving software and machinery is able to (09:29):
undefined
Ejaaz: detect things that you necessarily as a human aren't able to do this. (09:33):
undefined
Ejaaz: This graph specifically, Josh, can you explain what I'm looking at here? (09:39):
undefined
Josh: Yeah, so this is the gradient. This is the weighted decision tree in real time. (09:43):
undefined
Josh: So you could kind of see every single frame that it receives, (09:46):
undefined
Josh: the chart moves, and then you could actually see the point in which it realizes (09:49):
undefined
Josh: there's a threat and it adjusts very quickly. (09:52):
undefined
Josh: So what you're seeing here is the real time visual representation of what the (09:54):
undefined
Josh: brain sees. And we're going to get into this a little bit later where you can (09:57):
undefined
Josh: actually communicate with this system. (10:00):
undefined
Josh: You could talk to it just like it's a large language model. It's pretty insane. (10:02):
undefined
Josh: But I want to move on to the next section because this is my favorite. (10:05):
undefined
Josh: When I saw this, it just really blew my mind on how they (10:07):
undefined
Josh: were able to basically emulate real world (10:10):
undefined
Josh: driving scenarios and each as I want to start this section with an (10:13):
undefined
Josh: example that they showed if you don't mind scrolling down and sharing the one (10:16):
undefined
Josh: of the the fake screen so after these splats there's one a little bit later (10:18):
undefined
Josh: and basically it shows a driving further down even sorry the like next section (10:23):
undefined
Josh: then we'll go right back up oh sure sure yeah this one yeah yeah so this example (10:28):
undefined
Josh: that we're looking at on the screen. (10:34):
undefined
Josh: This looks like a standard traditional driving setup. (10:36):
undefined
Josh: So the car has, what is that, seven cameras and each one of them ingest data. (10:40):
undefined
Josh: The thing with this EJAS is what you're seeing on screen is not real. (10:44):
undefined
Josh: That is a 100% virtual representation of this real world. (10:47):
undefined
Josh: And it's unbelievable because it looks so good. And as I'm watching this, (10:52):
undefined
Josh: I'm like, man, I hope GTA 6 looks like this because the quality, (10:55):
undefined
Josh: the fidelity of this artificially generated world is indistinguishable from (11:00):
undefined
Josh: real life, the entire thing. (11:04):
undefined
Josh: And the reason they're able to do this is by ingesting all this data. (11:06):
undefined
Josh: So now that you've seen how impressive it gets, this is kind of how they build (11:09):
undefined
Josh: it. So we can go back up to the Gaussian splatting examples. (11:12):
undefined
Josh: And Gaussian splats are kind of a fancy way of saying, as the car drives through, (11:15):
undefined
Josh: you could imagine the cameras as scanners. (11:19):
undefined
Josh: So if you flipped a camera into a scanner, it maps this 3D world and creates a world. (11:22):
undefined
Josh: And then they're actually able to move around and navigate the 3D world they (11:27):
undefined
Josh: create using just the cameras on your car. (11:31):
undefined
Josh: And I want to reiterate that every Tesla you see on the road, (11:33):
undefined
Josh: regardless of when it was made, is capable of collecting this data and creating (11:36):
undefined
Josh: these 3D models that you see on the screen. So... (11:40):
undefined
Josh: The interesting thing here is that top bar is what the car sees. (11:43):
undefined
Josh: The bottom bar is what the car is generated to see. (11:45):
undefined
Josh: And what it's able to do as a result is it's able to kind of get a better understanding (11:48):
undefined
Josh: of the world around it and make much better decisions that in turn make it much (11:51):
undefined
Josh: safer than a human driver does. (11:55):
undefined
Ejaaz: This just looks like a computer game, Josh. Like one of those massive MMORPGs (11:56):
undefined
Ejaaz: that kind of generates the world as I navigate and move through it as I interact (12:01):
undefined
Ejaaz: with different characters. (12:05):
undefined
Ejaaz: This is kind of that, but for self-driving specifically. And why I think this is so cool, (12:06):
undefined
Ejaaz: and these are kind of like widely known as world simulators, (12:12):
undefined
Ejaaz: it's like an AI model that generates simulated realities, is that this data (12:15):
undefined
Ejaaz: can be modified in so many different ways and so many different scenarios to (12:20):
undefined
Ejaaz: train the car for experiences or accidents that it hasn't even, (12:25):
undefined
Ejaaz: that hasn't even encountered just yet. (12:29):
undefined
Ejaaz: And this is really cool because I think one major constraint that a lot of AI (12:32):
undefined
Ejaaz: models and self-driving models come up against is sometimes there's not enough (12:36):
undefined
Ejaaz: data to account for every single different type of scenario. (12:40):
undefined
Ejaaz: So a way to kind of address that is to create something known as synthetic data. (12:44):
undefined
Ejaaz: World simulators is one step towards being able to do that super effectively (12:49):
undefined
Ejaaz: whilst bending this simulated reality to how the actual world works, (12:53):
undefined
Ejaaz: right? Physics is super important, but hard to translate into an AI model. (12:59):
undefined
Ejaaz: And so seeing something like this at scale for a product, a car, (13:03):
undefined
Ejaaz: that is used by almost every human on the world is just so amazing to see. (13:07):
undefined
Josh: And the answer to the question, well, why hasn't everybody done this? (13:13):
undefined
Josh: Is because to generate these world models generally takes tens of seconds to do. (13:16):
undefined
Josh: Tesla's figured out a way to do it in 0.2 seconds. So it's a remarkable efficiency (13:19):
undefined
Josh: improvement that allows them to actually do this. (13:24):
undefined
Josh: It's not like the rest of the world doesn't want to do this. (13:26):
undefined
Josh: Is that technically speaking, it's just very, very difficult to do. (13:28):
undefined
Josh: And the next example they shared was one of my favorite ones because it really just created. (13:31):
undefined
Josh: It made it feel very familiar where you can actually talk to these models like (13:35):
undefined
Josh: they're a language model. (13:39):
undefined
Josh: Yeah. And the example above where you could just say, well, why are you not turning left? (13:40):
undefined
Josh: And it will explain to you, well, there's a detour sign. And why shouldn't you (13:44):
undefined
Josh: turn right? Well, because the detour sign is pointing to the left. (13:48):
undefined
Josh: And it really, you start to get a sense the same way yesterday in our episode (13:50):
undefined
Josh: yesterday, where you can see the behind the scenes of how the model thinks when it trades. (13:55):
undefined
Josh: You can now see the behind the scenes of the brain and you could start to understand (13:59):
undefined
Josh: how it works, why it works, how it's reasoning. (14:03):
undefined
Josh: And the results from this is pretty fascinating. It's not only is it interpreting (14:05):
undefined
Josh: inputs like where the lines on the road are, but it's also able to read signs. (14:09):
undefined
Josh: They have an example where you're able to see a human who's like kind of giving (14:12):
undefined
Josh: you a high five, like saying, wait one second, I'm about to pull out. (14:16):
undefined
Josh: And then the car recognizes that and stops. (14:18):
undefined
Josh: So there's these like unbelievable improvements that they have. (14:20):
undefined
Josh: And this section I want to get into next is because they can reevaluate these (14:22):
undefined
Josh: new decision trees on existing historical models. (14:27):
undefined
Josh: So my car, I've had a few near collision experiences that have been a little (14:30):
undefined
Josh: scary, but they've been narrowly avoided. (14:34):
undefined
Josh: What they can do is they can actually take the exact camera inputs from the (14:36):
undefined
Josh: car and emulate if the collision had actually happened. (14:39):
undefined
Josh: And then they could run these new tests on it and see how the new models would (14:42):
undefined
Josh: compare to the old models. (14:46):
undefined
Josh: So in the case that you narrowly miss an accident, well, you could test it on (14:47):
undefined
Josh: a new model and see if it does better. And in the first example, it does. (14:51):
undefined
Josh: And it actually moves away faster than the others. (14:54):
undefined
Josh: The second example that they have here is that you can create artificial examples. (14:57):
undefined
Josh: So you can take a car, remove it, place it into this virtual world, (15:01):
undefined
Josh: but it looks like the real world. It emulates a real world scenario. And it just. (15:05):
undefined
Josh: As I'm looking at this, Ejas, to your point, it all feels like a video game. (15:09):
undefined
Josh: And it's a really high fidelity video game where they can take things from reality. (15:12):
undefined
Josh: They can distort them. They could create fake realities. And as I was scrolling (15:16):
undefined
Josh: through this post, I started to lose track of what was real and what wasn't (15:20):
undefined
Josh: because it all looks so real to me. (15:24):
undefined
Josh: And to the video game point, which you might be able to share, (15:26):
undefined
Josh: is that they actually allow you to play it as if it was a video game. (15:29):
undefined
Josh: You can drive through these virtual worlds without actually needing a Tesla vehicle. (15:32):
undefined
Ejaaz: Yeah, so what I have here is the Tesla's Neural World Simulator, (15:36):
undefined
Ejaaz: where you have someone that is in basically a driver's seat, (15:41):
undefined
Ejaaz: but it's one of those video gaming driving setups. (15:45):
undefined
Ejaaz: And they are driving through what looks (15:47):
undefined
Ejaaz: like a pretty pleasant suburban neighborhood on a sunny blue sky day. (15:50):
undefined
Ejaaz: And it looks really real, Josh. It looks like something that would be recorded (15:54):
undefined
Ejaaz: from Tesla's seven cameras, except that none of it is real. (15:58):
undefined
Ejaaz: He is navigating through roads. He's skirting around cars. He's narrowly avoiding collisions. (16:02):
undefined
Ejaaz: And every single perspective and animal that you see from the three different (16:08):
undefined
Ejaaz: cameras on the screen here is completely and utterly simulated. (16:11):
undefined
Josh: The most remarkable part is that all of this amazing stuff that we've just talked (16:16):
undefined
Josh: about for the last 20 minutes, it's actually cross compatible with the next (16:19):
undefined
Josh: most important form of autonomy, which is robots. (16:23):
undefined
Josh: Now, everyone knows Tesla's making Optimus. They signal plans to make hundreds (16:26):
undefined
Josh: of thousands of these by next year. (16:29):
undefined
Josh: And the problem with training robots for a lot of other companies is that they (16:31):
undefined
Josh: don't have the data, they don't have the neural models. (16:34):
undefined
Josh: Well, all of the progress and all of the data that's been made previously through (16:36):
undefined
Josh: Tesla is cross-compatible directly with the robot team and Optimus as a humanoid robot. (16:39):
undefined
Josh: And that is one of the most impressive things because as the program gets better (16:46):
undefined
Josh: through AI's autopilot stack, it improves dramatically through Optimus. (16:51):
undefined
Josh: And what you're able to see is, (16:56):
undefined
Josh: A lot of, like you mentioned, Ejaz, the goldmine is the digital data because (16:58):
undefined
Josh: you just want more data to train. (17:01):
undefined
Josh: Optimus gets better. And that (17:03):
undefined
Josh: moves us on to the price of Tesla and the second order effects of Tesla. (17:04):
undefined
Josh: Because now that we have humanoid robots that are learning quickly, (17:09):
undefined
Josh: now that we have cars that are able to drive themselves, well, there's two things. (17:13):
undefined
Josh: One of them is being the chip that unifies the two. (17:17):
undefined
Josh: The other is the second order effects of what happens when this gets rolled out across the world. (17:20):
undefined
Josh: And he says, maybe you want to tee that up for us, because this is a very bullish (17:25):
undefined
Josh: scenario that we're guiding towards. (17:28):
undefined
Ejaaz: Okay, so this is the most exciting part for me for this entire episode, (17:30):
undefined
Ejaaz: because as you mentioned, this data and these neural networks aren't just super (17:35):
undefined
Ejaaz: valuable for the Tesla cars. (17:39):
undefined
Ejaaz: It's for the robots and pretty much any other kind of robotic machine that they create in the future. (17:41):
undefined
Ejaaz: And the beautiful thing about this is that it's self-recursive. (17:47):
undefined
Ejaaz: So whatever is learned from all the camera information and audio information (17:50):
undefined
Ejaaz: that's pulled from the cars can feed into the robots, (17:53):
undefined
Ejaaz: which is like kind of what we're seeing in the demo on our screen here with (17:56):
undefined
Ejaaz: this Optimus robot navigating through what seems to be a manufacturing site, right? (17:59):
undefined
Ejaaz: This is incredibly bullish for Tesla, the stock, in my opinion, (18:04):
undefined
Ejaaz: because it takes it from, well, it's currently breaching or sitting under its (18:09):
undefined
Ejaaz: all-time high, right, Josh? What is that market cap right now? (18:13):
undefined
Josh: We're just under an all-time high, which puts it right around $1.5 trillion. (18:16):
undefined
Ejaaz: Okay, so $1.5 trillion in today's age seems pretty small. (18:19):
undefined
Ejaaz: You just had Microsoft and Apple today cross $4 trillion market cap. (18:25):
undefined
Ejaaz: If you compare that to Tesla, and if you factor in the fact that these humanoid (18:29):
undefined
Ejaaz: robots are largely going to replace or work in conjunction with a large swathe (18:33):
undefined
Ejaaz: of the human manual labor force, (18:38):
undefined
Ejaaz: that prices this up at least up until a $10 trillion company as this scales out. (18:41):
undefined
Ejaaz: Josh, I have a feeling you're probably similarly bullish when it (18:47):
undefined
Josh: Comes to this. Obviously, I share your sentiment. I have been maximally bullish (18:49):
undefined
Josh: on Tesla for over a decade now. It's about, (18:53):
undefined
Josh: 12 years. Did your dad. (18:56):
undefined
Ejaaz: Buy your Tesla stock for you at the start? You asked him to? (18:58):
undefined
Josh: Yeah, I was too young to have my own brokerage account. So we were very early (19:02):
undefined
Josh: shares of Tesla and continue to be maximally bullish on it. (19:05):
undefined
Josh: And we're actually, I'm going to be recording a bull thesis episode about Tesla (19:09):
undefined
Josh: because I'm so bullish on it. So if you're interested in that, let me know. (19:12):
undefined
Josh: But I'm going to pull some notes from that to use here, just to kind of outline (19:15):
undefined
Josh: the humanoid robotic opportunity. (19:19):
undefined
Josh: Because EJ, as you said, $10 trillion, which is an outrageous market cap, (19:20):
undefined
Josh: considering NVIDIA is the largest company in the world sitting at four trillion. (19:24):
undefined
Josh: So that's a long way to go. And NVIDIA is on top of the world. (19:27):
undefined
Josh: But if you think of humanoids as labor, right, you have kind of four billion (19:30):
undefined
Josh: people in the labor market. (19:36):
undefined
Josh: And this becomes a global trend. This is not just for the United States. (19:37):
undefined
Josh: And if the average wage, which is what it is right now, is about $10,000 per (19:40):
undefined
Josh: year, that's a $40 trillion market size. (19:44):
undefined
Josh: So the labor opportunity is $40 trillion, assuming we don't have any productivity (19:47):
undefined
Josh: unlocks that generate brand new opportunities, that generate more use cases for labor. (19:52):
undefined
Josh: So that's just given the current state of the world today. (19:56):
undefined
Josh: So if one humanoid at $5 an hour can replace two humans working at $25 an hour, (19:59):
undefined
Josh: the value per humanoid becomes $200,000 per robot, which is pretty high given (20:04):
undefined
Josh: that the costs are projected to be around $20,000 to $30,000 once it's all said and done. (20:10):
undefined
Josh: The US labor market, there's 160 million people. (20:15):
undefined
Josh: So if just 1% is substituted by humanoid robots, that is greater than $300 billion in value. (20:18):
undefined
Josh: That's a lot of revenue. That is a tremendous amount of revenue. (20:24):
undefined
Josh: And then you get to a point where you're starting to offset significant percentages of GDP. (20:27):
undefined
Josh: So in the 1950s, the US manufacturing share of GDP, it was 30%. (20:31):
undefined
Josh: Today it sits at 10%. And if this goes further, we'll have a total reliance on foreign entities. (20:35):
undefined
Josh: So there's all the incentives in the world to bring robots into the United States. (20:40):
undefined
Josh: So we don't continue this trend of decreasing our manufacturing capabilities. (20:44):
undefined
Josh: There's a lot of headwinds and a lot of trends that all converge on the humanoid (20:47):
undefined
Josh: robot opportunity. It's just a matter of making these. (20:50):
undefined
Josh: And it's possible because of this new software stack and also because of this (20:53):
undefined
Josh: new chip, which is the AI5 chip. (20:58):
undefined
Josh: And the AI5 chip is the brand new golden child of Tesla. And it is going to (21:00):
undefined
Josh: be cross compatible between both robots and, (21:06):
undefined
Josh: cyber cabs. But you just maybe you want to walk us through exactly why this is interesting. (21:08):
undefined
Ejaaz: Yeah. So the way I think about this is this is Tesla's bold attempt to replace the GPU. (21:12):
undefined
Ejaaz: And as we've spoken about many times on this show before, Nvidia kind of rules the kingdom. (21:19):
undefined
Ejaaz: We mentioned that they are sitting at a $4 trillion or above a well above a (21:24):
undefined
Ejaaz: $4 trillion market cap. They are the kings of the roost. (21:27):
undefined
Ejaaz: And the reason why is because they provide the hardware that kind of fuels all (21:30):
undefined
Ejaaz: these different things. (21:34):
undefined
Ejaaz: Now, what Tesla identified is whilst all these GPUs that they've been using (21:35):
undefined
Ejaaz: are really helpful, they're not specifically designed to fit certain niche use (21:40):
undefined
Ejaaz: cases when it comes to a range of different things that they're involved in, right? (21:45):
undefined
Ejaaz: Cars, humanoid robots, and an array of different things. (21:49):
undefined
Ejaaz: And now they've released their AI5 chip, which is basically their brand new (21:52):
undefined
Ejaaz: chip, which is going to be used across all their different robots. (21:57):
undefined
Ejaaz: So it's going to be used in cars, on humanoids, and the like. (22:01):
undefined
Ejaaz: And the coolest part about this, Josh, we were speaking about this before the (22:04):
undefined
Ejaaz: show, is it improves this whole GPU experience for them by 40 times. (22:07):
undefined
Ejaaz: But can you help me unpack as to why exactly? (22:13):
undefined
Ejaaz: Is this like a sizing thing? Can they add more compute? How does this work? (22:15):
undefined
Josh: Okay, so first thing, AI5 isn't out just yet. It's coming. They have completed the spec. (22:19):
undefined
Josh: Elon's been working on it. He said on the most recent earnings call that it (22:24):
undefined
Josh: has been his number one focus for weeks and weeks and weeks on end, (22:27):
undefined
Josh: which is very high signal that it means a lot. (22:30):
undefined
Josh: So it is coming soon. They're working on tooling and they're working to roll (22:32):
undefined
Josh: this out, I assume, in companion with the Optimus robot that is probably coming (22:34):
undefined
Josh: next year. You mentioned it's 40 times better. (22:39):
undefined
Josh: Why is it 40 times better? And why do companies make their own chips? (22:41):
undefined
Josh: I think this is an important question because a lot of people don't know. (22:44):
undefined
Josh: Well, NVIDIA makes awesome GPUs. Why would I go through all the R&D budgeting (22:47):
undefined
Josh: costs and pain in the ass because... (22:50):
undefined
Josh: To make my own chip? And the answer is because vertical integration allows you (22:53):
undefined
Josh: to be hyper customized in what you're able to do. (22:57):
undefined
Josh: So what Tesla has done is they, it's funny, they do this with everything, (22:59):
undefined
Josh: but they kind of, they looked at the chip through first principles. (23:03):
undefined
Josh: They looked at all the different modules that sit on this chip. (23:06):
undefined
Josh: You could think one of them processes graphics, one of them processes images, one is processing math. (23:08):
undefined
Josh: The reason why all of these GPUs from other companies need to have all of these (23:13):
undefined
Josh: is because they need to satisfy their customers. (23:16):
undefined
Josh: They need to be able to be diverse in the types of computing they can do. (23:18):
undefined
Josh: In the narrow band of use cases that Tesla has, they're able to reconsider this and optimize for it. (23:22):
undefined
Josh: So for example, there's this image signal processor that sits on a chip and (23:27):
undefined
Josh: it's meant to what it says. It processes image signals that come in. (23:31):
undefined
Josh: What Tesla has done is they're not actually processing images. (23:35):
undefined
Josh: They're processing photons and photons can be binary. (23:38):
undefined
Josh: They could be expressed in code. So there's this. (23:40):
undefined
Josh: Big chip that sits on a larger chip, they're able to completely remove that (23:43):
undefined
Josh: image processing chip because they said, actually, we don't need to look at images ever. (23:46):
undefined
Josh: We're just doing photons in, photons out, baby. And that unlocks X percent of (23:50):
undefined
Josh: this board to add more compute power to the specific type of compute you need. (23:54):
undefined
Josh: So for the first time ever, you're getting these chips that don't actually look (23:58):
undefined
Josh: like traditional chips. (24:03):
undefined
Josh: They're built very different because of the narrow band use case that's required. (24:04):
undefined
Josh: And that allows them to not only be much more efficient in terms of compute (24:07):
undefined
Josh: per watt, but also cost per watt, and also the cross compatibility across all these devices. (24:11):
undefined
Josh: So a lot of companies, they have, like if you think of Apple, (24:15):
undefined
Josh: they have the M series chip for the computers and the iPhones, (24:18):
undefined
Josh: whereas NVIDIA has 12 different GPUs for mobile devices, for power, (24:22):
undefined
Josh: general computers, for data centers. (24:27):
undefined
Josh: It's this really remarkable unlock that we're going to start to see roll out (24:29):
undefined
Josh: next year in both that enables both the CyberCab and the humanoid robot. (24:32):
undefined
Ejaaz: There's an increasing trend of these new age AI tech companies that once they (24:36):
undefined
Ejaaz: reach escape velocity for a bunch of consumer and enterprise facing products, (24:42):
undefined
Ejaaz: they start to vertically integrate with a part of which includes creating their (24:47):
undefined
Ejaaz: own custom design GPUs and chips. (24:51):
undefined
Ejaaz: The most recent example I can think of aside from Tesla is OpenAI, (24:54):
undefined
Ejaaz: who announced that a partnership with Broadgate, (24:58):
undefined
Ejaaz: that they're going to be developing their own custom GPUs to fuel certain niche (25:01):
undefined
Ejaaz: use cases that their future GPT-6 models and ahead will utilize. (25:05):
undefined
Ejaaz: They haven't quite revealed what those chips are going to be facilitating exactly. (25:10):
undefined
Ejaaz: But what we do know is that they're using the AI model itself to help them design (25:16):
undefined
Ejaaz: this chip. So this thing around AI5 is the most Elon thing ever, (25:21):
undefined
Ejaaz: because we've seen what he's done when he's taken a hammer to data centers. (25:25):
undefined
Ejaaz: And we're seeing now what he's what he's done by creating the probably the most (25:29):
undefined
Ejaaz: valuable resource going forwards for tech companies at the GPU layer. (25:33):
undefined
Ejaaz: So I don't know. I'm excited about this, Josh. (25:36):
undefined
Ejaaz: It makes me unfathomably bullish. (25:39):
undefined
Ejaaz: My earlier $10 trillion estimate is probably too conservative after what we've just discussed. (25:41):
undefined
Josh: Well, with Elon's new pay package, there is a direct incentive alignment. (25:47):
undefined
Josh: One thing on the Broadcom partnership with OpenAI, the difference there is that (25:50):
undefined
Josh: Broadcom exists and Tesla is a single entity. (25:54):
undefined
Josh: So OpenAI doesn't really have the resources in order to create their own chips in-house. (25:58):
undefined
Josh: And I think that's a really big difference because when there is that physical (26:04):
undefined
Josh: gap between different companies when you're designing these chips, (26:07):
undefined
Josh: it makes it a little bit more difficult to do that really hardcore, (26:11):
undefined
Josh: like cost-cutting vertical integration that Tesla has. Tesla's doing this. (26:14):
undefined
Josh: They're making their own ship in-house. They're designing it in-house. (26:18):
undefined
Josh: OpenEye is outsourcing that responsibility. And that's where you'll maybe start to see discrepancy. (26:21):
undefined
Josh: So I am hopeful that they will do great, but I still suspect Tesla will do better. (26:26):
undefined
Josh: And Tesla also has manufacturing prowess. So yeah, I think if we walk away with (26:31):
undefined
Josh: anything from this episode is that both of us share the sentiment that we are (26:35):
undefined
Josh: unfathomably bullish for an assortment of reasons. And this is just one of them. (26:38):
undefined
Josh: The Tesla bookcase will be coming soon, I promise. (26:41):
undefined
Josh: And there's a lot more to the company, but this is autonomy. This is autopilot. (26:44):
undefined
Josh: This is the secrets of Tesla finally unveiled for the world. (26:49):
undefined
Josh: And I imagine the rest of the world, granted, they've probably been trying to (26:52):
undefined
Josh: emulate this. It's not really much of a secret, but we'll have a very difficult time in doing so. (26:56):
undefined
Ejaaz: I think that wraps it up for today's episode. (27:00):
undefined
Ejaaz: We hope you enjoyed this breakdown. We are unfathomably excited and bullish, (27:03):
undefined
Ejaaz: as I've said multiple times about Tesla, but are you? (27:07):
undefined
Ejaaz: Let us know in the comments. Are we crazy? is the vision that we're engaging (27:11):
undefined
Ejaaz: in around Tesla completely insane? (27:15):
undefined
Ejaaz: Are robots not really a thing in your opinion? Let us know in the comments. (27:17):
undefined
Ejaaz: We're also going to be releasing one more episode this week, (27:21):
undefined
Ejaaz: which is going to be the AI Weekly Roundup, which we're going to cover all the (27:24):
undefined
Ejaaz: hottest topics. There's some crazy stuff that has happened this week. (27:26):
undefined
Ejaaz: And if there's anything else that we've missed or that you want to hear about, (27:30):
undefined
Ejaaz: let us know in the comments. DM us. We're always available. And we will see you in the next one. (27:32):
undefined
Josh: Thanks for watching. See you guys. (27:38):
undefined
undefined
Josh: speak every day is the rise of autonomy, and particularly around autonomous robots. (00:06):
undefined
Josh: Robots can be many things. Robots can be humanoids, they can be cars. (00:11):
undefined
Josh: And today we're going to talk about both, because there's one company that is (00:14):
undefined
Josh: at the frontier of both of those areas, and that's Tesla. (00:17):
undefined
Josh: Tesla has the most unbelievable set of autopilot software that I think exists in the world. (00:21):
undefined
Josh: I've been using it personally for eight years now. And it's been amazing to (00:27):
undefined
Josh: see how good it's gotten. (00:30):
undefined
Josh: And EJS, now there's, for the first time ever, we have the secrets. (00:31):
undefined
Josh: The secret sauce that shares exactly how they've been able to get autonomy this (00:35):
undefined
Josh: powerful, this impressive. (00:39):
undefined
Josh: And there's now very clearly a world in which I can imagine waking up in the (00:40):
undefined
Josh: morning, getting ready to go to work, stepping outside, and there's a cyber (00:44):
undefined
Josh: cab waiting for me outside that will just take me wherever I want for a fraction (00:47):
undefined
Josh: of the cost that it takes for a normal driver. (00:50):
undefined
Josh: And I think this is an incredibly powerful unlock and to see a behind the scenes (00:53):
undefined
Josh: of this is awesome so the entire episode today is behind the scenes of the most (00:57):
undefined
Josh: impressive new front-end tier technology that exists. (01:02):
undefined
Ejaaz: I think what i'm most excited about today josh is the fact that i've always (01:05):
undefined
Ejaaz: thought tesla ai and robotics is so cool but i i just don't know how any of (01:10):
undefined
Ejaaz: this works and they've refused to tell us and finally they've they've spilt their secrets today (01:15):
undefined
Ejaaz: to quickly paint some context for the listeners here, up until yesterday, (01:20):
undefined
Ejaaz: we only thought of Tesla AI as something called a neural network. (01:24):
undefined
Ejaaz: That's their secret source. And a neural network can be thought of as a software (01:29):
undefined
Ejaaz: program that is designed to function like the human brain. (01:33):
undefined
Ejaaz: So it takes in information and it discovers patterns, trends, (01:37):
undefined
Ejaaz: and it can also sometimes make predictions. (01:41):
undefined
Ejaaz: Now, this contrasts directly to some of Tesla's competitors, (01:43):
undefined
Ejaaz: which do self-driving and robotics in a very different way. (01:47):
undefined
Ejaaz: They take more modular and sensor-driven approaches, right? (01:50):
undefined
Ejaaz: The reason why Tesla's neural network is so special is they have an end-to-end (01:54):
undefined
Ejaaz: neural network, which means that they feed a bunch of raw data from one side (01:59):
undefined
Ejaaz: and out comes the output, which is an action. (02:04):
undefined
Ejaaz: In this case of Tesla cars, it would be driving, steering, and acceleration. (02:07):
undefined
Ejaaz: And they took this approach for a few different ways. (02:12):
undefined
Ejaaz: The most important being, it's really hard Josh to codify what human values (02:15):
undefined
Ejaaz: are and what I mean by that is let's say in this example that you're seeing (02:21):
undefined
Ejaaz: on your screen right now you are driving your car and there's a massive puddle on your lane but (02:25):
undefined
Josh: You see that you. (02:30):
undefined
Ejaaz: Could potentially drive into the oncoming lane to skirt around it now for humans (02:32):
undefined
Ejaaz: it's really easy to do that right it's like okay maybe I should just go through (02:36):
undefined
Ejaaz: it because there's no cars coming but for a machine to do that it requires a (02:39):
undefined
Ejaaz: lot of effort. It's hard to hard code. (02:45):
undefined
Ejaaz: So that's one special thing around the neural network. But Josh, (02:47):
undefined
Ejaaz: I want to jump into the secrets. (02:50):
undefined
Ejaaz: Can you lead us with the first one? (02:52):
undefined
Josh: Well, what you mentioned is really important, the end-to-end stuff. (02:54):
undefined
Josh: And I want to walk through a little experiment. (02:57):
undefined
Josh: So when you kick a soccer ball, I think this is an experience everyone's kind (02:59):
undefined
Josh: of went through, right? What do you do when you kick a soccer ball? (03:02):
undefined
Ejaaz: Yeah, I see the soccer ball coming towards me. I kind of prepare my legs ready to kind of kick. (03:05):
undefined
Ejaaz: I'm right-footed, so I'm kicking with my right foot. (03:12):
undefined
Ejaaz: And then I guess the rest is kind of intuitive, Josh. I just kind of run up to it and kick it. (03:14):
undefined
Josh: Yeah, yeah. And I think that's exactly the point is when you kick a soccer ball, (03:19):
undefined
Josh: this is something a lot of people have experienced. (03:23):
undefined
Josh: You're not actually thinking about all the parts of kicking a soccer ball. (03:25):
undefined
Josh: You're not thinking of where it is on the ground, where your ankle is, (03:28):
undefined
Josh: where your knee is, where your leg is, the positioning, how hard you're going (03:31):
undefined
Josh: to kick it. It just feels very intuitive. (03:34):
undefined
Josh: And with a lot of other car companies, they're hard coding these intuitions as code. (03:35):
undefined
Josh: So it does have to think about each section. It does have to calculate each (03:40):
undefined
Josh: section. And what's different about Tesla and what we learned from this article, (03:43):
undefined
Josh: this is from Ashok, who is the person who's in charge of Tesla AI, (03:46):
undefined
Josh: is that they use this thing called end-to-end neural networks. (03:50):
undefined
Josh: And what does that mean? In like a fun, simple way, it's basically the intuition (03:52):
undefined
Josh: that you just described with kicking a soccer ball, the AI model, (03:56):
undefined
Josh: the chip on a car is able to emulate that. (03:59):
undefined
Josh: So instead of making these minute decisions all the way through a fixed decision (04:01):
undefined
Josh: tree, they're able to take a ton of data and use these things that we've learned (04:05):
undefined
Josh: over time, which are gradients and weights, and basically move the gradients (04:08):
undefined
Josh: and weights throughout the decision process to reach an end goal. (04:13):
undefined
Josh: So if the end goal is to kick a soccer ball, there's a very clear stated end goal. (04:16):
undefined
Josh: And the neural network's job is to figure out the full sweep of gradients as (04:20):
undefined
Josh: it goes across to get to that end goal. (04:24):
undefined
Josh: And it uses a bunch of this training data that they collect in order to get there. (04:26):
undefined
Josh: So this is this remarkable technology that breakthrough that they have. (04:29):
undefined
Josh: And they have some really interesting examples here. (04:32):
undefined
Josh: So in the case of the ducks, like we're looking at an example on the screen (04:34):
undefined
Josh: right now, there's ducks standing in the middle of the road. (04:37):
undefined
Josh: When you're coding an AI system, when you're coding a car, you're not hard coding (04:40):
undefined
Josh: in, if you see ducks, do this. (04:44):
undefined
Josh: What the car is understanding intuitively is like, okay, there's an obstacle (04:46):
undefined
Josh: here and they are ducks. They're not moving. (04:49):
undefined
Josh: The interesting thing is the example above is the car recognizes that the ducks (04:52):
undefined
Josh: are actually moving across the road. (04:56):
undefined
Josh: So it knows to wait and then it could pass once they've moved. (04:57):
undefined
Josh: But the second one, it notices they're just kind of chilling. (05:00):
undefined
Josh: The ducks aren't going anywhere. And what does it do? It understands that intuitively (05:03):
undefined
Josh: and it is able to back up and then move around them. And that's the difference (05:06):
undefined
Josh: in how Tesla does it versus some other companies is they're not hard coding (05:10):
undefined
Josh: a series of fixed parameters. (05:13):
undefined
Josh: They are doing it all entirely through these neural networks. (05:15):
undefined
Ejaaz: If we move on to secret number one, Josh, it kind of explains how they're able (05:17):
undefined
Ejaaz: to achieve this at a pretty high level, right? (05:22):
undefined
Ejaaz: So it's titled The Curse of Dimensionality. And what it basically describes (05:25):
undefined
Ejaaz: is you can imagine for a car to self-drive, it requires a ton of data. (05:30):
undefined
Ejaaz: I think Tesla, the average car, has about seven cameras. (05:35):
undefined
Ejaaz: It ingests a ton of audio data, a ton of navigation GPS data, (05:39):
undefined
Ejaaz: and kinematics. So speed is tracking your speed. (05:44):
undefined
Ejaaz: And so all this data is roughly equivalent to 2 billion tokens. (05:46):
undefined
Ejaaz: And if you think about it, it needs to run through this end-to-end neural network (05:51):
undefined
Ejaaz: that you just described, Josh, and it needs to output pretty much two tokens. (05:54):
undefined
Ejaaz: One token, which determines which way the car should steer, and the other token (05:58):
undefined
Ejaaz: determining how fast should that car be at that point? Should it decelerate (06:03):
undefined
Ejaaz: or should it accelerate? (06:07):
undefined
Ejaaz: And you can imagine this is an incredibly nuanced and complex process. (06:09):
undefined
Ejaaz: And the way that the Tesla neural engine or the neural network is designed is (06:13):
undefined
Ejaaz: it has really special data lanes that process this data in a very nuanced way (06:18):
undefined
Ejaaz: to understand what exactly it needs to map onto when it comes to steering and acceleration. (06:23):
undefined
Ejaaz: Now, you might think that's pretty cool, but Tesla's secret source when it comes (06:28):
undefined
Ejaaz: to this particular component is the driving data, right, Josh? (06:33):
undefined
Ejaaz: So they get access to all the camera data, audio data, GPS data that I just (06:37):
undefined
Ejaaz: mentioned from their entire fleet of Tesla cars. (06:41):
undefined
Ejaaz: So the equivalent of data that they get every day is something crazy like 500 (06:45):
undefined
Ejaaz: years worth of driving data. (06:50):
undefined
Ejaaz: Now, you can imagine if it processes this amount of rich data, (06:52):
undefined
Ejaaz: and not all of that data is important, right? It's kind of like the same kind of standard things. (06:56):
undefined
Josh: Over those years of data. (07:00):
undefined
Ejaaz: You get access to the one or two random nuanced incidents which feed in and (07:02):
undefined
Ejaaz: improve the collective intelligence of the entire Tesla fleet. (07:07):
undefined
Ejaaz: So whether you're on the other side of the world driving a Tesla or you're in (07:11):
undefined
Ejaaz: the local neighborhood, you still benefit from the same types of improvements. (07:14):
undefined
Josh: I want to talk a little bit about the scale because you mentioned 2 billion (07:18):
undefined
Josh: inputs and it's kind of difficult to comprehend what 2 billion actually means. (07:20):
undefined
Josh: And as a good example, I want you to imagine your phone processing every TikTok (07:25):
undefined
Josh: that exists on the platform every single second in order to determine the next (07:31):
undefined
Josh: turn. That is two billion inputs. (07:35):
undefined
Josh: It is an astronomical amount of data. You're basically, you take the whole TikTok (07:37):
undefined
Josh: catalog every second in order to make every decision and you distill that entire (07:41):
undefined
Josh: data set into two single points. (07:45):
undefined
Josh: And it's just, it's a remarkable amount of compression and then a remarkable (07:48):
undefined
Josh: amount of precision to make the right decision over and over and over again, (07:52):
undefined
Josh: and then adjust and calculate as things change. (07:55):
undefined
Josh: The way that they do this, they're not doing this raw. They're not actually (07:58):
undefined
Josh: ingesting all this data. (08:01):
undefined
Josh: They have this data curation process that they use in order to help them kind (08:02):
undefined
Josh: of figure out what is important and what is just noise. (08:06):
undefined
Josh: And what they do, and we have a great example on screen here, (08:09):
undefined
Josh: is they pick the juiciest clips. (08:11):
undefined
Josh: It's like kind of curating like a viral playlist and they use it to train the (08:13):
undefined
Josh: AI on these weird scenarios. (08:17):
undefined
Josh: So we're seeing on the screen, there's someone rolling through an intersection of wheelchair. (08:18):
undefined
Josh: It's actually very funny to see and scary to see what types of things happen. (08:22):
undefined
Josh: I mean, this is crazy. Two cars crashing right in front of you, (08:26):
undefined
Josh: driving on a snow blind street. (08:30):
undefined
Josh: There's kids that are running out in the middle of the road. (08:31):
undefined
Josh: There's these tremendous amount of edge cases that are really difficult to understand. (08:33):
undefined
Josh: And because of the 500 years of driving data every single day that they ingest, (08:38):
undefined
Josh: they're able to analyze and to kind of sift through. (08:43):
undefined
Josh: And they've come up with systems to curate the most viral clips, (08:45):
undefined
Josh: not viral, but the clips with the most implications of safety that are kind of the weird edge cases. (08:48):
undefined
Josh: And then we have this example here. Do you want to walk through the chart that (08:54):
undefined
Josh: we're seeing, because it's really fascinating how the car can kind of see it before the human does. (08:57):
undefined
Ejaaz: Yeah. So what's interesting is when I first watched this clip and for those who are listening, (09:01):
undefined
Ejaaz: it is a car driving on a very rainy evening on the highway and a car in front (09:06):
undefined
Ejaaz: of it kind of crashes out and goes and starts to spin and kind of enter its own lane. (09:12):
undefined
Ejaaz: When I first watched this video, Josh, I didn't even notice the car spinning (09:17):
undefined
Ejaaz: out because it happens so far away. (09:21):
undefined
Ejaaz: And so what's effective about this particular video is, given everything that (09:24):
undefined
Ejaaz: you just described, the Tesla self-driving software and machinery is able to (09:29):
undefined
Ejaaz: detect things that you necessarily as a human aren't able to do this. (09:33):
undefined
Ejaaz: This graph specifically, Josh, can you explain what I'm looking at here? (09:39):
undefined
Josh: Yeah, so this is the gradient. This is the weighted decision tree in real time. (09:43):
undefined
Josh: So you could kind of see every single frame that it receives, (09:46):
undefined
Josh: the chart moves, and then you could actually see the point in which it realizes (09:49):
undefined
Josh: there's a threat and it adjusts very quickly. (09:52):
undefined
Josh: So what you're seeing here is the real time visual representation of what the (09:54):
undefined
Josh: brain sees. And we're going to get into this a little bit later where you can (09:57):
undefined
Josh: actually communicate with this system. (10:00):
undefined
Josh: You could talk to it just like it's a large language model. It's pretty insane. (10:02):
undefined
Josh: But I want to move on to the next section because this is my favorite. (10:05):
undefined
Josh: When I saw this, it just really blew my mind on how they (10:07):
undefined
Josh: were able to basically emulate real world (10:10):
undefined
Josh: driving scenarios and each as I want to start this section with an (10:13):
undefined
Josh: example that they showed if you don't mind scrolling down and sharing the one (10:16):
undefined
Josh: of the the fake screen so after these splats there's one a little bit later (10:18):
undefined
Josh: and basically it shows a driving further down even sorry the like next section (10:23):
undefined
Josh: then we'll go right back up oh sure sure yeah this one yeah yeah so this example (10:28):
undefined
Josh: that we're looking at on the screen. (10:34):
undefined
Josh: This looks like a standard traditional driving setup. (10:36):
undefined
Josh: So the car has, what is that, seven cameras and each one of them ingest data. (10:40):
undefined
Josh: The thing with this EJAS is what you're seeing on screen is not real. (10:44):
undefined
Josh: That is a 100% virtual representation of this real world. (10:47):
undefined
Josh: And it's unbelievable because it looks so good. And as I'm watching this, (10:52):
undefined
Josh: I'm like, man, I hope GTA 6 looks like this because the quality, (10:55):
undefined
Josh: the fidelity of this artificially generated world is indistinguishable from (11:00):
undefined
Josh: real life, the entire thing. (11:04):
undefined
Josh: And the reason they're able to do this is by ingesting all this data. (11:06):
undefined
Josh: So now that you've seen how impressive it gets, this is kind of how they build (11:09):
undefined
Josh: it. So we can go back up to the Gaussian splatting examples. (11:12):
undefined
Josh: And Gaussian splats are kind of a fancy way of saying, as the car drives through, (11:15):
undefined
Josh: you could imagine the cameras as scanners. (11:19):
undefined
Josh: So if you flipped a camera into a scanner, it maps this 3D world and creates a world. (11:22):
undefined
Josh: And then they're actually able to move around and navigate the 3D world they (11:27):
undefined
Josh: create using just the cameras on your car. (11:31):
undefined
Josh: And I want to reiterate that every Tesla you see on the road, (11:33):
undefined
Josh: regardless of when it was made, is capable of collecting this data and creating (11:36):
undefined
Josh: these 3D models that you see on the screen. So... (11:40):
undefined
Josh: The interesting thing here is that top bar is what the car sees. (11:43):
undefined
Josh: The bottom bar is what the car is generated to see. (11:45):
undefined
Josh: And what it's able to do as a result is it's able to kind of get a better understanding (11:48):
undefined
Josh: of the world around it and make much better decisions that in turn make it much (11:51):
undefined
Josh: safer than a human driver does. (11:55):
undefined
Ejaaz: This just looks like a computer game, Josh. Like one of those massive MMORPGs (11:56):
undefined
Ejaaz: that kind of generates the world as I navigate and move through it as I interact (12:01):
undefined
Ejaaz: with different characters. (12:05):
undefined
Ejaaz: This is kind of that, but for self-driving specifically. And why I think this is so cool, (12:06):
undefined
Ejaaz: and these are kind of like widely known as world simulators, (12:12):
undefined
Ejaaz: it's like an AI model that generates simulated realities, is that this data (12:15):
undefined
Ejaaz: can be modified in so many different ways and so many different scenarios to (12:20):
undefined
Ejaaz: train the car for experiences or accidents that it hasn't even, (12:25):
undefined
Ejaaz: that hasn't even encountered just yet. (12:29):
undefined
Ejaaz: And this is really cool because I think one major constraint that a lot of AI (12:32):
undefined
Ejaaz: models and self-driving models come up against is sometimes there's not enough (12:36):
undefined
Ejaaz: data to account for every single different type of scenario. (12:40):
undefined
Ejaaz: So a way to kind of address that is to create something known as synthetic data. (12:44):
undefined
Ejaaz: World simulators is one step towards being able to do that super effectively (12:49):
undefined
Ejaaz: whilst bending this simulated reality to how the actual world works, (12:53):
undefined
Ejaaz: right? Physics is super important, but hard to translate into an AI model. (12:59):
undefined
Ejaaz: And so seeing something like this at scale for a product, a car, (13:03):
undefined
Ejaaz: that is used by almost every human on the world is just so amazing to see. (13:07):
undefined
Josh: And the answer to the question, well, why hasn't everybody done this? (13:13):
undefined
Josh: Is because to generate these world models generally takes tens of seconds to do. (13:16):
undefined
Josh: Tesla's figured out a way to do it in 0.2 seconds. So it's a remarkable efficiency (13:19):
undefined
Josh: improvement that allows them to actually do this. (13:24):
undefined
Josh: It's not like the rest of the world doesn't want to do this. (13:26):
undefined
Josh: Is that technically speaking, it's just very, very difficult to do. (13:28):
undefined
Josh: And the next example they shared was one of my favorite ones because it really just created. (13:31):
undefined
Josh: It made it feel very familiar where you can actually talk to these models like (13:35):
undefined
Josh: they're a language model. (13:39):
undefined
Josh: Yeah. And the example above where you could just say, well, why are you not turning left? (13:40):
undefined
Josh: And it will explain to you, well, there's a detour sign. And why shouldn't you (13:44):
undefined
Josh: turn right? Well, because the detour sign is pointing to the left. (13:48):
undefined
Josh: And it really, you start to get a sense the same way yesterday in our episode (13:50):
undefined
Josh: yesterday, where you can see the behind the scenes of how the model thinks when it trades. (13:55):
undefined
Josh: You can now see the behind the scenes of the brain and you could start to understand (13:59):
undefined
Josh: how it works, why it works, how it's reasoning. (14:03):
undefined
Josh: And the results from this is pretty fascinating. It's not only is it interpreting (14:05):
undefined
Josh: inputs like where the lines on the road are, but it's also able to read signs. (14:09):
undefined
Josh: They have an example where you're able to see a human who's like kind of giving (14:12):
undefined
Josh: you a high five, like saying, wait one second, I'm about to pull out. (14:16):
undefined
Josh: And then the car recognizes that and stops. (14:18):
undefined
Josh: So there's these like unbelievable improvements that they have. (14:20):
undefined
Josh: And this section I want to get into next is because they can reevaluate these (14:22):
undefined
Josh: new decision trees on existing historical models. (14:27):
undefined
Josh: So my car, I've had a few near collision experiences that have been a little (14:30):
undefined
Josh: scary, but they've been narrowly avoided. (14:34):
undefined
Josh: What they can do is they can actually take the exact camera inputs from the (14:36):
undefined
Josh: car and emulate if the collision had actually happened. (14:39):
undefined
Josh: And then they could run these new tests on it and see how the new models would (14:42):
undefined
Josh: compare to the old models. (14:46):
undefined
Josh: So in the case that you narrowly miss an accident, well, you could test it on (14:47):
undefined
Josh: a new model and see if it does better. And in the first example, it does. (14:51):
undefined
Josh: And it actually moves away faster than the others. (14:54):
undefined
Josh: The second example that they have here is that you can create artificial examples. (14:57):
undefined
Josh: So you can take a car, remove it, place it into this virtual world, (15:01):
undefined
Josh: but it looks like the real world. It emulates a real world scenario. And it just. (15:05):
undefined
Josh: As I'm looking at this, Ejas, to your point, it all feels like a video game. (15:09):
undefined
Josh: And it's a really high fidelity video game where they can take things from reality. (15:12):
undefined
Josh: They can distort them. They could create fake realities. And as I was scrolling (15:16):
undefined
Josh: through this post, I started to lose track of what was real and what wasn't (15:20):
undefined
Josh: because it all looks so real to me. (15:24):
undefined
Josh: And to the video game point, which you might be able to share, (15:26):
undefined
Josh: is that they actually allow you to play it as if it was a video game. (15:29):
undefined
Josh: You can drive through these virtual worlds without actually needing a Tesla vehicle. (15:32):
undefined
Ejaaz: Yeah, so what I have here is the Tesla's Neural World Simulator, (15:36):
undefined
Ejaaz: where you have someone that is in basically a driver's seat, (15:41):
undefined
Ejaaz: but it's one of those video gaming driving setups. (15:45):
undefined
Ejaaz: And they are driving through what looks (15:47):
undefined
Ejaaz: like a pretty pleasant suburban neighborhood on a sunny blue sky day. (15:50):
undefined
Ejaaz: And it looks really real, Josh. It looks like something that would be recorded (15:54):
undefined
Ejaaz: from Tesla's seven cameras, except that none of it is real. (15:58):
undefined
Ejaaz: He is navigating through roads. He's skirting around cars. He's narrowly avoiding collisions. (16:02):
undefined
Ejaaz: And every single perspective and animal that you see from the three different (16:08):
undefined
Ejaaz: cameras on the screen here is completely and utterly simulated. (16:11):
undefined
Josh: The most remarkable part is that all of this amazing stuff that we've just talked (16:16):
undefined
Josh: about for the last 20 minutes, it's actually cross compatible with the next (16:19):
undefined
Josh: most important form of autonomy, which is robots. (16:23):
undefined
Josh: Now, everyone knows Tesla's making Optimus. They signal plans to make hundreds (16:26):
undefined
Josh: of thousands of these by next year. (16:29):
undefined
Josh: And the problem with training robots for a lot of other companies is that they (16:31):
undefined
Josh: don't have the data, they don't have the neural models. (16:34):
undefined
Josh: Well, all of the progress and all of the data that's been made previously through (16:36):
undefined
Josh: Tesla is cross-compatible directly with the robot team and Optimus as a humanoid robot. (16:39):
undefined
Josh: And that is one of the most impressive things because as the program gets better (16:46):
undefined
Josh: through AI's autopilot stack, it improves dramatically through Optimus. (16:51):
undefined
Josh: And what you're able to see is, (16:56):
undefined
Josh: A lot of, like you mentioned, Ejaz, the goldmine is the digital data because (16:58):
undefined
Josh: you just want more data to train. (17:01):
undefined
Josh: Optimus gets better. And that (17:03):
undefined
Josh: moves us on to the price of Tesla and the second order effects of Tesla. (17:04):
undefined
Josh: Because now that we have humanoid robots that are learning quickly, (17:09):
undefined
Josh: now that we have cars that are able to drive themselves, well, there's two things. (17:13):
undefined
Josh: One of them is being the chip that unifies the two. (17:17):
undefined
Josh: The other is the second order effects of what happens when this gets rolled out across the world. (17:20):
undefined
Josh: And he says, maybe you want to tee that up for us, because this is a very bullish (17:25):
undefined
Josh: scenario that we're guiding towards. (17:28):
undefined
Ejaaz: Okay, so this is the most exciting part for me for this entire episode, (17:30):
undefined
Ejaaz: because as you mentioned, this data and these neural networks aren't just super (17:35):
undefined
Ejaaz: valuable for the Tesla cars. (17:39):
undefined
Ejaaz: It's for the robots and pretty much any other kind of robotic machine that they create in the future. (17:41):
undefined
Ejaaz: And the beautiful thing about this is that it's self-recursive. (17:47):
undefined
Ejaaz: So whatever is learned from all the camera information and audio information (17:50):
undefined
Ejaaz: that's pulled from the cars can feed into the robots, (17:53):
undefined
Ejaaz: which is like kind of what we're seeing in the demo on our screen here with (17:56):
undefined
Ejaaz: this Optimus robot navigating through what seems to be a manufacturing site, right? (17:59):
undefined
Ejaaz: This is incredibly bullish for Tesla, the stock, in my opinion, (18:04):
undefined
Ejaaz: because it takes it from, well, it's currently breaching or sitting under its (18:09):
undefined
Ejaaz: all-time high, right, Josh? What is that market cap right now? (18:13):
undefined
Josh: We're just under an all-time high, which puts it right around $1.5 trillion. (18:16):
undefined
Ejaaz: Okay, so $1.5 trillion in today's age seems pretty small. (18:19):
undefined
Ejaaz: You just had Microsoft and Apple today cross $4 trillion market cap. (18:25):
undefined
Ejaaz: If you compare that to Tesla, and if you factor in the fact that these humanoid (18:29):
undefined
Ejaaz: robots are largely going to replace or work in conjunction with a large swathe (18:33):
undefined
Ejaaz: of the human manual labor force, (18:38):
undefined
Ejaaz: that prices this up at least up until a $10 trillion company as this scales out. (18:41):
undefined
Ejaaz: Josh, I have a feeling you're probably similarly bullish when it (18:47):
undefined
Josh: Comes to this. Obviously, I share your sentiment. I have been maximally bullish (18:49):
undefined
Josh: on Tesla for over a decade now. It's about, (18:53):
undefined
Josh: 12 years. Did your dad. (18:56):
undefined
Ejaaz: Buy your Tesla stock for you at the start? You asked him to? (18:58):
undefined
Josh: Yeah, I was too young to have my own brokerage account. So we were very early (19:02):
undefined
Josh: shares of Tesla and continue to be maximally bullish on it. (19:05):
undefined
Josh: And we're actually, I'm going to be recording a bull thesis episode about Tesla (19:09):
undefined
Josh: because I'm so bullish on it. So if you're interested in that, let me know. (19:12):
undefined
Josh: But I'm going to pull some notes from that to use here, just to kind of outline (19:15):
undefined
Josh: the humanoid robotic opportunity. (19:19):
undefined
Josh: Because EJ, as you said, $10 trillion, which is an outrageous market cap, (19:20):
undefined
Josh: considering NVIDIA is the largest company in the world sitting at four trillion. (19:24):
undefined
Josh: So that's a long way to go. And NVIDIA is on top of the world. (19:27):
undefined
Josh: But if you think of humanoids as labor, right, you have kind of four billion (19:30):
undefined
Josh: people in the labor market. (19:36):
undefined
Josh: And this becomes a global trend. This is not just for the United States. (19:37):
undefined
Josh: And if the average wage, which is what it is right now, is about $10,000 per (19:40):
undefined
Josh: year, that's a $40 trillion market size. (19:44):
undefined
Josh: So the labor opportunity is $40 trillion, assuming we don't have any productivity (19:47):
undefined
Josh: unlocks that generate brand new opportunities, that generate more use cases for labor. (19:52):
undefined
Josh: So that's just given the current state of the world today. (19:56):
undefined
Josh: So if one humanoid at $5 an hour can replace two humans working at $25 an hour, (19:59):
undefined
Josh: the value per humanoid becomes $200,000 per robot, which is pretty high given (20:04):
undefined
Josh: that the costs are projected to be around $20,000 to $30,000 once it's all said and done. (20:10):
undefined
Josh: The US labor market, there's 160 million people. (20:15):
undefined
Josh: So if just 1% is substituted by humanoid robots, that is greater than $300 billion in value. (20:18):
undefined
Josh: That's a lot of revenue. That is a tremendous amount of revenue. (20:24):
undefined
Josh: And then you get to a point where you're starting to offset significant percentages of GDP. (20:27):
undefined
Josh: So in the 1950s, the US manufacturing share of GDP, it was 30%. (20:31):
undefined
Josh: Today it sits at 10%. And if this goes further, we'll have a total reliance on foreign entities. (20:35):
undefined
Josh: So there's all the incentives in the world to bring robots into the United States. (20:40):
undefined
Josh: So we don't continue this trend of decreasing our manufacturing capabilities. (20:44):
undefined
Josh: There's a lot of headwinds and a lot of trends that all converge on the humanoid (20:47):
undefined
Josh: robot opportunity. It's just a matter of making these. (20:50):
undefined
Josh: And it's possible because of this new software stack and also because of this (20:53):
undefined
Josh: new chip, which is the AI5 chip. (20:58):
undefined
Josh: And the AI5 chip is the brand new golden child of Tesla. And it is going to (21:00):
undefined
Josh: be cross compatible between both robots and, (21:06):
undefined
Josh: cyber cabs. But you just maybe you want to walk us through exactly why this is interesting. (21:08):
undefined
Ejaaz: Yeah. So the way I think about this is this is Tesla's bold attempt to replace the GPU. (21:12):
undefined
Ejaaz: And as we've spoken about many times on this show before, Nvidia kind of rules the kingdom. (21:19):
undefined
Ejaaz: We mentioned that they are sitting at a $4 trillion or above a well above a (21:24):
undefined
Ejaaz: $4 trillion market cap. They are the kings of the roost. (21:27):
undefined
Ejaaz: And the reason why is because they provide the hardware that kind of fuels all (21:30):
undefined
Ejaaz: these different things. (21:34):
undefined
Ejaaz: Now, what Tesla identified is whilst all these GPUs that they've been using (21:35):
undefined
Ejaaz: are really helpful, they're not specifically designed to fit certain niche use (21:40):
undefined
Ejaaz: cases when it comes to a range of different things that they're involved in, right? (21:45):
undefined
Ejaaz: Cars, humanoid robots, and an array of different things. (21:49):
undefined
Ejaaz: And now they've released their AI5 chip, which is basically their brand new (21:52):
undefined
Ejaaz: chip, which is going to be used across all their different robots. (21:57):
undefined
Ejaaz: So it's going to be used in cars, on humanoids, and the like. (22:01):
undefined
Ejaaz: And the coolest part about this, Josh, we were speaking about this before the (22:04):
undefined
Ejaaz: show, is it improves this whole GPU experience for them by 40 times. (22:07):
undefined
Ejaaz: But can you help me unpack as to why exactly? (22:13):
undefined
Ejaaz: Is this like a sizing thing? Can they add more compute? How does this work? (22:15):
undefined
Josh: Okay, so first thing, AI5 isn't out just yet. It's coming. They have completed the spec. (22:19):
undefined
Josh: Elon's been working on it. He said on the most recent earnings call that it (22:24):
undefined
Josh: has been his number one focus for weeks and weeks and weeks on end, (22:27):
undefined
Josh: which is very high signal that it means a lot. (22:30):
undefined
Josh: So it is coming soon. They're working on tooling and they're working to roll (22:32):
undefined
Josh: this out, I assume, in companion with the Optimus robot that is probably coming (22:34):
undefined
Josh: next year. You mentioned it's 40 times better. (22:39):
undefined
Josh: Why is it 40 times better? And why do companies make their own chips? (22:41):
undefined
Josh: I think this is an important question because a lot of people don't know. (22:44):
undefined
Josh: Well, NVIDIA makes awesome GPUs. Why would I go through all the R&D budgeting (22:47):
undefined
Josh: costs and pain in the ass because... (22:50):
undefined
Josh: To make my own chip? And the answer is because vertical integration allows you (22:53):
undefined
Josh: to be hyper customized in what you're able to do. (22:57):
undefined
Josh: So what Tesla has done is they, it's funny, they do this with everything, (22:59):
undefined
Josh: but they kind of, they looked at the chip through first principles. (23:03):
undefined
Josh: They looked at all the different modules that sit on this chip. (23:06):
undefined
Josh: You could think one of them processes graphics, one of them processes images, one is processing math. (23:08):
undefined
Josh: The reason why all of these GPUs from other companies need to have all of these (23:13):
undefined
Josh: is because they need to satisfy their customers. (23:16):
undefined
Josh: They need to be able to be diverse in the types of computing they can do. (23:18):
undefined
Josh: In the narrow band of use cases that Tesla has, they're able to reconsider this and optimize for it. (23:22):
undefined
Josh: So for example, there's this image signal processor that sits on a chip and (23:27):
undefined
Josh: it's meant to what it says. It processes image signals that come in. (23:31):
undefined
Josh: What Tesla has done is they're not actually processing images. (23:35):
undefined
Josh: They're processing photons and photons can be binary. (23:38):
undefined
Josh: They could be expressed in code. So there's this. (23:40):
undefined
Josh: Big chip that sits on a larger chip, they're able to completely remove that (23:43):
undefined
Josh: image processing chip because they said, actually, we don't need to look at images ever. (23:46):
undefined
Josh: We're just doing photons in, photons out, baby. And that unlocks X percent of (23:50):
undefined
Josh: this board to add more compute power to the specific type of compute you need. (23:54):
undefined
Josh: So for the first time ever, you're getting these chips that don't actually look (23:58):
undefined
Josh: like traditional chips. (24:03):
undefined
Josh: They're built very different because of the narrow band use case that's required. (24:04):
undefined
Josh: And that allows them to not only be much more efficient in terms of compute (24:07):
undefined
Josh: per watt, but also cost per watt, and also the cross compatibility across all these devices. (24:11):
undefined
Josh: So a lot of companies, they have, like if you think of Apple, (24:15):
undefined
Josh: they have the M series chip for the computers and the iPhones, (24:18):
undefined
Josh: whereas NVIDIA has 12 different GPUs for mobile devices, for power, (24:22):
undefined
Josh: general computers, for data centers. (24:27):
undefined
Josh: It's this really remarkable unlock that we're going to start to see roll out (24:29):
undefined
Josh: next year in both that enables both the CyberCab and the humanoid robot. (24:32):
undefined
Ejaaz: There's an increasing trend of these new age AI tech companies that once they (24:36):
undefined
Ejaaz: reach escape velocity for a bunch of consumer and enterprise facing products, (24:42):
undefined
Ejaaz: they start to vertically integrate with a part of which includes creating their (24:47):
undefined
Ejaaz: own custom design GPUs and chips. (24:51):
undefined
Ejaaz: The most recent example I can think of aside from Tesla is OpenAI, (24:54):
undefined
Ejaaz: who announced that a partnership with Broadgate, (24:58):
undefined
Ejaaz: that they're going to be developing their own custom GPUs to fuel certain niche (25:01):
undefined
Ejaaz: use cases that their future GPT-6 models and ahead will utilize. (25:05):
undefined
Ejaaz: They haven't quite revealed what those chips are going to be facilitating exactly. (25:10):
undefined
Ejaaz: But what we do know is that they're using the AI model itself to help them design (25:16):
undefined
Ejaaz: this chip. So this thing around AI5 is the most Elon thing ever, (25:21):
undefined
Ejaaz: because we've seen what he's done when he's taken a hammer to data centers. (25:25):
undefined
Ejaaz: And we're seeing now what he's what he's done by creating the probably the most (25:29):
undefined
Ejaaz: valuable resource going forwards for tech companies at the GPU layer. (25:33):
undefined
Ejaaz: So I don't know. I'm excited about this, Josh. (25:36):
undefined
Ejaaz: It makes me unfathomably bullish. (25:39):
undefined
Ejaaz: My earlier $10 trillion estimate is probably too conservative after what we've just discussed. (25:41):
undefined
Josh: Well, with Elon's new pay package, there is a direct incentive alignment. (25:47):
undefined
Josh: One thing on the Broadcom partnership with OpenAI, the difference there is that (25:50):
undefined
Josh: Broadcom exists and Tesla is a single entity. (25:54):
undefined
Josh: So OpenAI doesn't really have the resources in order to create their own chips in-house. (25:58):
undefined
Josh: And I think that's a really big difference because when there is that physical (26:04):
undefined
Josh: gap between different companies when you're designing these chips, (26:07):
undefined
Josh: it makes it a little bit more difficult to do that really hardcore, (26:11):
undefined
Josh: like cost-cutting vertical integration that Tesla has. Tesla's doing this. (26:14):
undefined
Josh: They're making their own ship in-house. They're designing it in-house. (26:18):
undefined
Josh: OpenEye is outsourcing that responsibility. And that's where you'll maybe start to see discrepancy. (26:21):
undefined
Josh: So I am hopeful that they will do great, but I still suspect Tesla will do better. (26:26):
undefined
Josh: And Tesla also has manufacturing prowess. So yeah, I think if we walk away with (26:31):
undefined
Josh: anything from this episode is that both of us share the sentiment that we are (26:35):
undefined
Josh: unfathomably bullish for an assortment of reasons. And this is just one of them. (26:38):
undefined
Josh: The Tesla bookcase will be coming soon, I promise. (26:41):
undefined
Josh: And there's a lot more to the company, but this is autonomy. This is autopilot. (26:44):
undefined
Josh: This is the secrets of Tesla finally unveiled for the world. (26:49):
undefined
Josh: And I imagine the rest of the world, granted, they've probably been trying to (26:52):
undefined
Josh: emulate this. It's not really much of a secret, but we'll have a very difficult time in doing so. (26:56):
undefined
Ejaaz: I think that wraps it up for today's episode. (27:00):
undefined
Ejaaz: We hope you enjoyed this breakdown. We are unfathomably excited and bullish, (27:03):
undefined
Ejaaz: as I've said multiple times about Tesla, but are you? (27:07):
undefined
Ejaaz: Let us know in the comments. Are we crazy? is the vision that we're engaging (27:11):
undefined
Ejaaz: in around Tesla completely insane? (27:15):
undefined
Ejaaz: Are robots not really a thing in your opinion? Let us know in the comments. (27:17):
undefined
Ejaaz: We're also going to be releasing one more episode this week, (27:21):
undefined
Ejaaz: which is going to be the AI Weekly Roundup, which we're going to cover all the (27:24):
undefined
Ejaaz: hottest topics. There's some crazy stuff that has happened this week. (27:26):
undefined
Ejaaz: And if there's anything else that we've missed or that you want to hear about, (27:30):
undefined
Ejaaz: let us know in the comments. DM us. We're always available. And we will see you in the next one. (27:32):
undefined
Josh: Thanks for watching. See you guys. (27:38):
undefined
Popular Podcasts
Stuff You Should Know
If you've ever wanted to know about champagne, satanism, the Stonewall Uprising, chaos theory, LSD, El Nino, true crime and Rosa Parks, then look no further. Josh and Chuck have you covered.
My Favorite Murder with Karen Kilgariff and Georgia Hardstark
My Favorite Murder is a true crime comedy podcast hosted by Karen Kilgariff and Georgia Hardstark. Each week, Karen and Georgia share compelling true crimes and hometown stories from friends and listeners. Since MFM launched in January of 2016, Karen and Georgia have shared their lifelong interest in true crime and have covered stories of infamous serial killers like the Night Stalker, mysterious cold cases, captivating cults, incredible survivor stories and important events from history like the Tulsa race massacre of 1921. My Favorite Murder is part of the Exactly Right podcast network that provides a platform for bold, creative voices to bring to life provocative, entertaining and relatable stories for audiences everywhere. The Exactly Right roster of podcasts covers a variety of topics including historic true crime, comedic interviews and news, science, pop culture and more. Podcasts on the network include Buried Bones with Kate Winkler Dawson and Paul Holes, That's Messed Up: An SVU Podcast, This Podcast Will Kill You, Bananas and more.
Dateline NBC
Current and classic episodes, featuring compelling true-crime mysteries, powerful documentaries and in-depth investigations. Follow now to get the latest episodes of Dateline NBC completely free, or subscribe to Dateline Premium for ad-free listening and exclusive bonus content: DatelinePremium.com