Latest / Elon Musk Podcast / Tesla Cybercab Manufacturing and Autonomous Realities

Transcript

Tesla Cybercab Manufacturing and Autonomous Realities · Elon Musk Podcast

• 0:00
  Imagine walking into a car dealership, dropping 30 grand on
• 0:03
  a brand new car from Tesla and sitting down in the driver's
• 0:06
  seat only to realize they completely forgot to install a
• 0:08
  steering wheel and the pedals are missing too.
• 0:13
  That isn't a factory error, that is the actual intentional design
• 0:18
  of the cyber cab they just officially started producing at
• 0:21
  Gigafactory Texas. Yeah, and Elon Musk claims these
• 0:25
  completely autonomous vehicles will cover 70% of the US
• 0:28
  population. But when you look at the actual
• 0:31
  deployment numbers on the ground and the incredibly creative
• 0:34
  regulatory workarounds they're actively using, a completely
• 0:37
  different reality emerges. How does a company actually
• 0:41
  scale a nationwide fleet of steering wheel less vehicles
• 0:44
  when their current autonomous system struggles with basic rain
• 0:47
  and requires hidden and human operators to function?
• 0:50
  Well, we really have to start by looking at the physical car
• 0:52
  itself because it represents a complete departure from
• 0:55
  traditional automotive logic. You know, we have to look at
• 0:57
  what is actually rolling off that factory floor in Texas.
• 1:00
  Yeah, because the physical design of the Cyber Cab is, I
• 1:03
  mean, it's wild. It's A2 seater with butterfly
• 1:06
  doors that open upward. Like something out of a sci-fi
• 1:09
  movie. Exactly.
• 1:10
  But the exterior engineering goes much deeper than just an
• 1:13
  interesting silhouette. For one, it utilizes inductive
• 1:17
  wireless charging. Oh, like how you charge your
• 1:20
  smartphone by dropping on a little pad on your night stand?
• 1:23
  Yeah, they basically scale that exact concept up for a car.
• 1:27
  There is no charging port where you would plug in a heavy cable.
• 1:30
  Wow. You don't have to get out of the
• 1:32
  car in the cold. The vehicle just parks over a
• 1:34
  magnetic pad on the ground and fills its 35 kWh battery
• 1:39
  completely on its own. That's fascinating.
• 1:42
  And the body itself is made of polyurethane plastic panels that
• 1:45
  are are injected with color during the molding process,
• 1:48
  which you know, entirely avoids the need for a traditional paint
• 1:50
  shop. All of this is engineered
• 1:52
  specifically to hit a sub $30,000 price point.
• 1:55
  Bypassing the paint shop is a brilliant manufacturing move.
• 1:58
  In a traditional car factory, the painting process is just an
• 2:02
  absolute nightmare. Oh really?
• 2:03
  Yeah, it is the most expensive, most environmentally demanding
• 2:08
  and most space consuming step on the entire floor.
• 2:11
  Right, because you have to dip heavy metal chassis in chemical
• 2:15
  anti corrosion baths. Bake them.
• 2:17
  Spray them with primer. Apply to base coat.
• 2:19
  Add a clear coat and then bake them again.
• 2:21
  It's a massive undertaking. By injecting color directly into
• 2:26
  the plastic while it's still hot liquid, they skip all of that
• 2:29
  infrastructure. That makes sense.
• 2:31
  Plus, if you scratch a plastic panel, you don't see a Gray
• 2:34
  primer layer underneath. The color goes all the way
• 2:36
  through. Oh nice.
• 2:37
  But the most significant manufacturing shift here is what
• 2:40
  they call the unboxed process. We really have to look at the
• 2:43
  history of how cars are made to appreciate exactly why this is
• 2:46
  so different. OK.
• 2:47
  So historically, the automotive industry has relied on a linear
• 2:51
  assembly line. Exactly.
• 2:53
  You start with a hollow metal shell and it slowly crawls down
• 2:57
  a massive conveyor belt while workers and robots attach parts
• 3:01
  sequentially. Wiring harnesses, then the
• 3:03
  dashboard. Then the seats.
• 3:04
  Yeah, the unbox process throws that linear concept in the
• 3:08
  trash. Right, so building this car is
• 3:10
  much more like assembling Lego blocks in think chunks and
• 3:13
  different tables before snapping them together rather than
• 3:16
  passing 1 giant piece down a long conveyor belt.
• 3:19
  That is a perfect analogy. Instead of 1 long line, you have
• 3:23
  parallel workstations. High density robotic arms are
• 3:26
  working on the front section, the rear section, the floor and
• 3:30
  the sides all at the exact same time.
• 3:32
  So they just build the module separately.
• 3:34
  Yeah, these distinct modules are fully completed on their own and
• 3:37
  then only merge together at the very final step.
• 3:40
  This parallel manufacturing method cuts capital expenditure
• 3:43
  in half and drastically lowers production costs.
• 3:46
  Which is huge. It is by stripping out the
• 3:49
  steering column, the brake pedals, and all the mechanical
• 3:52
  linkages that connect a human driver to the wheels, you remove
• 3:55
  thousands of dollars from the bill of materials.
• 3:58
  And driving down the production cost to the sub $30,000 level
• 4:02
  changes the entire business model, right?
• 4:04
  Completely, It allows the vehicle to function as a high
• 4:07
  margin, revenue generating robotic asset on the Tesla
• 4:10
  network. You aren't just buying
• 4:12
  transportation to get yourself to the grocery store.
• 4:14
  The economics are designed around purchasing a fleet
• 4:17
  vehicle that can theoretically pick up passengers and own
• 4:20
  passive income while you sleep. But that economic model relies
• 4:24
  entirely on the software who are functioning Florida State
• 4:26
  without any human intervention inside the car, right.
• 4:29
  And that brings us to what is actually happening on the ground
• 4:32
  in reality. Yeah, because despite claims of
• 4:34
  massive population coverage, the unsupervised robo taxi pilot in
• 4:39
  Austin, TX currently operates with roughly 42 vehicles. 42
• 4:45
  That service suffers from availability under 20%, plus it
• 4:50
  completely shuts down during rain due to its vision only
• 4:53
  camera system and recorded a crash rate 9 times worse than
• 4:57
  human drivers in early testing. Yeah, the early metrics are
• 5:00
  rough. Wait, back up.
• 5:02
  They are building a nationwide network, but the cars literally
• 5:04
  stop working if it rains. Yes they do, and that is direct
• 5:07
  result of the specific hardware architecture they chose.
• 5:10
  The system relies entirely on cameras to perceive the
• 5:13
  surrounding environment. Most of their competitors use a
• 5:16
  combination of cameras, radar and lidar.
• 5:19
  Right, and Lidar uses rapid light pulses, lasers essentially
• 5:24
  to map surroundings in three dimensions, and radar uses radio
• 5:28
  waves. Exactly.
• 5:30
  Both of those sensor types can easily penetrate fog, heavy
• 5:33
  rain, and harsh glare from the sun.
• 5:35
  But cameras, on the other hand, act exactly like human eyes.
• 5:39
  Yeah, if you've ever tried to use your phone camera to take a
• 5:42
  picture in the pouring rain, you know exactly why a vision only
• 5:45
  system struggles. Cameras get blinded by direct
• 5:48
  sunlight and water droplets obscure the physical lens.
• 5:51
  It distorts the image processing.
• 5:52
  It does because the cameras lack the redundancy of radar or
• 5:56
  Lidar. The immediate deployment is
• 5:57
  strictly limited to favorable weather locations in the
• 6:00
  Sunbelt. It restricts the technology from
• 6:02
  functioning safely in unpredictable climates.
• 6:05
  If the service only operates a fraction of the time and only
• 6:08
  when the weather is absolutely perfect, it completely changes
• 6:12
  how you view their timeline for a nationwide roll out.
• 6:14
  It really does. It also changes how we look at
• 6:17
  their operations in other more complicated cities.
• 6:21
  Like if the technology fails in a simple Texas rain shower, how
• 6:26
  on earth are they legally operating these cars in a dense,
• 6:29
  foggy city like San Francisco? That is a great point.
• 6:32
  Well, let's look at the regulatory situation in
• 6:34
  California. They are operating these
• 6:36
  vehicles under a charter party carrier permit.
• 6:39
  Which is the exact same permit used by limousine companies.
• 6:42
  Yeah, this legal classification defines the person in the front
• 6:45
  seat as a traditional driver, completely exempting Tesla from
• 6:49
  the safety data reporting required of actual autonomous
• 6:53
  vehicles like Waymo. Because in California, the DMV
• 6:56
  mandates that if you operate under a formal autonomous
• 6:59
  vehicle permit, you are legally obligated to report every single
• 7:02
  intervention, every crash, and every time the software fails to
• 7:06
  the state. Those reports become public
• 7:07
  record. Right by operating under a
• 7:09
  limousine permit instead, the state legally views the human in
• 7:13
  the driver's seat simply as a chauffeur who happens to be
• 7:16
  utilizing advanced driver assistance software.
• 7:18
  It's a massive loophole. I had to say I fundamentally
• 7:21
  disagree with this approach. A company marketing a self
• 7:24
  driving service is legally registering as a chauffeur
• 7:28
  service to hide crash reports and intervention data from the
• 7:31
  public. It feels incredibly deceptive.
• 7:34
  I can see why you'd say that. If you're going to put the word
• 7:37
  autonomous or robo taxi on the marketing materials, you should
• 7:41
  be subject to the exact same transparency laws as the
• 7:45
  company's actively logging millions of miles of public
• 7:48
  safety data. Hiding behind a technicality
• 7:51
  meant for stretch limousines defeats the entire purpose of
• 7:54
  public safety oversight. The.
• 7:55
  Consequence of this regulatory maneuvering is that it Shields
• 7:58
  the company's autonomous capabilities from scrutiny.
• 8:01
  Exactly. It prevents a clear safety
• 8:03
  comparison with competitors who published detailed metrics for
• 8:06
  you to evaluate. When competing services are
• 8:09
  publishing their intervention rates and providing narratives
• 8:12
  of any traffic incidents, they are allowing the public to
• 8:15
  assess the real maturity of their software.
• 8:17
  But by opting out of that reporting structure entirely,
• 8:21
  you are left with a massive blind spot regarding how often
• 8:24
  the human driver actually has to prevent a collision.
• 8:27
  Or grab the steering wheel to avoid an accident.
• 8:29
  Yeah, and we know human intervention is still a huge
• 8:33
  part of this industry across the board.
• 8:35
  Oh, definitely. That is actually the focus of a
• 8:37
  Senate investigation led by Ed Markey, which is probing the
• 8:41
  autonomous vehicle industry's use of remote human operators
• 8:44
  and overseas staffers. The companies refused to
• 8:46
  disclose how frequently these remote workers must intervene to
• 8:50
  reorient the vehicles while simultaneously racing to reach
• 8:54
  10 billion miles of AI training data to proof the system's
• 8:58
  safety. Hold on, So the artificial
• 9:00
  intelligence still relies on remote workers pulling the
• 9:03
  strings from a call center? Exactly when these vehicles
• 9:06
  encounter a situation they simply do not understand, they
• 9:09
  ping a remote call center right? A human operator sitting at a
• 9:13
  desk, sometimes in another country entirely.
• 9:15
  We'll look at the vehicles live camera feeds, manually draw a
• 9:18
  digital path around the obstacle on their screen, or directly
• 9:22
  provide a command to get the car moving safely.
• 9:24
  Again, we call these edge cases. An edge case is basically a
• 9:28
  bizarre, unpredictable scenario the AI hasn't been explicitly
• 9:32
  trained on. Like a person in a chicken suit
• 9:34
  chasing a dog down the highway. Or a flock of wild turkeys
• 9:37
  refusing to cross a crosswalk. This reliance on human in the
• 9:41
  loop systems completely complicates the legal definition
• 9:44
  of full autonomy. Because if an overseas worker is
• 9:48
  quietly resolving edge cases every few miles so the car
• 9:51
  doesn't get stuck, the system isn't actually autonomous.
• 9:54
  Right. It is basically remote
• 9:55
  controlled. Regulators are forced to
• 9:58
  scrutinize whether these vehicles are genuinely driving
• 10:00
  themselves or just being managed from afar.
• 10:03
  And that reliance on remote assistance ties directly back to
• 10:06
  the capability of the software and the physical hardware
• 10:09
  running inside the car. That's so well the artificial
• 10:13
  intelligence processing required to navigate a chaotic city
• 10:16
  street without any human help requires immense computational
• 10:20
  power. You are asking a computer to
• 10:22
  process dozens of high definition video feeds, identify
• 10:25
  objects, predict human behavior, and execute mechanical commands
• 10:29
  in a fraction of a second. That makes sense, and that heavy
• 10:32
  computational demand has created a severe internal hardware
• 10:35
  divide within the fleet. Newer cars equipped with updated
• 10:38
  chips get the full neural network software, while older
• 10:42
  cars only receive a restricted leat version of the driving
• 10:46
  program. Because the older processing
• 10:47
  units simply do not have the horsepower to run the latest end
• 10:50
  to end neural networks. An end to end neural network is
• 10:53
  a huge shift in how AI works. In older systems, programmers
• 10:58
  wrote millions of lines of code with specific rules.
• 11:01
  If you see a stop sign then apply the brakes.
• 11:03
  But an end to end system bypasses human coded rules
• 11:06
  entirely. It takes in raw video data from
• 11:09
  the cameras and directly outputs the steering and braking
• 11:12
  commands, learning curely by analyzing massive amounts of
• 11:15
  driving footage. Running that kind of complex AI
• 11:18
  architecture requires incredibly advanced silicon.
• 11:21
  We're also seeing a divide in the physical form factors being
• 11:23
  developed. Oh yeah, like the clay models
• 11:25
  spotted at Giga, Texas, indicating a three row SUV or
• 11:29
  cyber van. Along with Musk's tease of a
• 11:31
  vehicle way cooler than a minivan to satisfy demands from
• 11:34
  families needing more space. The two seater cyber cab serves
• 11:37
  a very specific urban commuter use case, but large segments of
• 11:42
  the population require higher seating capacity for groceries,
• 11:46
  car seats and daily life. The hardware fragmentation
• 11:49
  limits how much of the existing fleet can actually participate
• 11:53
  in the revenue generating robo taxi network.
• 11:56
  It forces consumers to constantly upgrade their
• 11:58
  physical hardware just to stay relevant in the software
• 12:01
  ecosystem. Think about the car sitting in
• 12:04
  your driveway. If you purchased a vehicle with
• 12:07
  the expectation that it would eventually become a fully
• 12:09
  autonomous asset, but your internal computer chip is now
• 12:12
  obsolete, you are stranded on an older software branch.
• 12:16
  You're locked out of the exact financial model that made the
• 12:19
  purchase appealing in the 1st place it.
• 12:20
  Completely shifts the automotive model away from long term
• 12:23
  hardware ownership and much closer to the smartphone upgrade
• 12:26
  cycle. If the artificial intelligence
• 12:28
  requires a new processor every few years to function safely
• 12:31
  without a driver, the physical chassis of the car becomes
• 12:35
  secondary to the silicon running it.
• 12:37
  You wouldn't expect an older smartphone to run demanding apps
• 12:40
  flawlessly, and we are starting to see that same reality apply
• 12:43
  to $50,000 vehicles. So the push for a driverless
• 12:47
  future involves highly advanced, low cost manufacturing, but the
• 12:51
  actual deployment is severely constrained by weather
• 12:54
  limitations, hidden human operators and regulatory
• 12:57
  workarounds. If the hardware outpaces the
• 13:00
  software's ability to navigate a simple rainstorm, what happens
• 13:03
  when these vehicles encounter truly chaotic urban
• 13:06
  environments? If you're not subscribed yet,
• 13:08
  take a second and hit follow on whatever app you're using.
• 13:10
  It helps us keep making this. We appreciate you being here.