Latest / Elon Musk Podcast / Tesla hardware limits split the fleet

Transcript

Tesla hardware limits split the fleet · Elon Musk Podcast

• 0:00
  Tesla entirely wiped the word Autopilot from their navigation
• 0:03
  software and officially rebranded the car's processing
• 0:06
  unit from the FSD computer to the AI computer.
• 0:09
  Yeah, and that semantic shift is actually backed by a massive
• 0:13
  underlying code replacement. I mean, the trajectory of this
• 0:15
  conversation moves right through an exploration of a complete
• 0:18
  neural network compiler rewrite, a growing physical hardware
• 0:22
  divide, creating 2 distinct classes of vehicles on the road,
• 0:25
  and the literal gamification of autonomous driving software.
• 0:30
  So does rewriting a vehicle's neural network compiler actually
• 0:34
  make a two ton machine react faster to a squirrel darting
• 0:37
  into the road? Well, to really understand that
• 0:40
  connection we have to look closely at the MLIR compiler
• 0:43
  rewrite. That is the multi level
• 0:44
  intermediate representation compiler that was just pushed
• 0:47
  inside the latest full self driving update, right?
• 0:49
  It completely changes how the software interacts with the
• 0:52
  physical silicon. Hold on, what exactly does an AI
• 0:54
  compiler do inside a car? Think about it like like you
• 0:58
  were trying to translate a massive complicated book into a
• 1:01
  new language using a physical dictionary for every single
• 1:05
  word. Right, exactly like looking up
• 1:07
  each word 1 by 1, you eventually get the meaning of the sentence,
• 1:11
  but the process is incredibly slow and, you know, completely
• 1:14
  disjointed. Exactly.
• 1:16
  The old software was essentially doing that.
• 1:18
  It was constantly looking up instructions to figure out how
• 1:21
  to communicate with the car's hardware.
• 1:23
  Yeah, the code had to be translated step by step before
• 1:27
  the silicon could actually execute a steering command or a
• 1:31
  braking decision. But the new MLIR compiler
• 1:35
  eliminates that dictionary entirely.
• 1:37
  It allows the neural network to fluently speak the native
• 1:40
  language of the car's hardware, and the direct results of that
• 1:44
  fluency is a 20% reduction in reaction time. 20%, yeah, it's.
• 1:48
  Huge. When you reduce latency in the
• 1:51
  system like that, you allow the car to process far more frames
• 1:54
  from the external cameras every single second.
• 1:57
  The physical consequences on the road are immediate and honestly
• 2:00
  highly noticeable for anyone sitting in the driver's seat.
• 2:03
  You. Feel it right away like you get
• 2:05
  significantly smoother stops because the vehicle is digesting
• 2:08
  visual data faster and, you know, planning its breaking
• 2:10
  curve earlier. Definitely, the vehicle also
• 2:13
  improves its handling of complex compound intersections.
• 2:17
  Those are the worst. Right, we are talking about
• 2:20
  those really frustrating, poorly designed city intersections with
• 2:25
  curved roads, faded lane lines, and multiple confusing traffic
• 2:29
  light configurations. The faster the compiler
• 2:32
  processes the geometry of the road well, the more confident
• 2:35
  the steering inputs become. The system even gains the
• 2:39
  ability to maintain control during temporary system
• 2:42
  degradations, like if you are driving directly into harsh
• 2:45
  sunlight and one of the front cameras is temporarily blinded
• 2:48
  by the glare, the vehicle can automatically recover.
• 2:50
  Right, it relies on the remaining sensors.
• 2:52
  Exactly, without immediately screaming at you to intervene
• 2:55
  and take the wheel. Which brings us directly to the
• 2:57
  training process. The engineers actually dedicated
• 3:01
  A reinforcement learning phase specifically to small animals.
• 3:04
  Interesting. Yeah, they fed the system
• 3:06
  incredibly difficult examples of animals darting into the street
• 3:09
  from behind parked cars or bushes.
• 3:11
  The specific focus on those scenarios is fascinating because
• 3:14
  it relies on granting the system digital rewards for proactive
• 3:19
  safety. I mean, the neural network
• 3:21
  operates in a simulation, and every time it successfully
• 3:24
  anticipates a sudden unpredictable movement like from
• 3:27
  a simulated squirrel or dog without causing an erratic
• 3:30
  swerve, it earns A mathematical reward.
• 3:34
  It is literally trained to crave the reward of anticipating the
• 3:37
  unpredictable. But a faster digital brain
• 3:40
  requires sufficient physical muscle to process that data
• 3:43
  flow, and this exposes a permanent physical limitation
• 3:46
  for older vehicles currently on the road.
• 3:49
  Yeah, let's lay out the physical hardware differences stratifying
• 3:52
  the fleet. So the newly named AI computer,
• 3:54
  known internally as Hardware for, boasts A20 core CPU.
• 3:59
  It has 16 gigabytes of RAM and five megapixel cameras.
• 4:02
  We back up a little over 1 megapixel on the older ones.
• 4:05
  Yeah, the older hardware 3 vehicles rely on a much older 12
• 4:08
  Volt architecture and cameras with just over 1 megapixel.
• 4:11
  That is early smartphone camera quality guiding a vehicle at
• 4:15
  highway speeds. It sounds wild when you say it
• 4:18
  out loud, but yes. And the issue goes far deeper
• 4:21
  than just the camera resolution right?
• 4:23
  The older 12 Volt wiring harnesses simply like the
• 4:27
  physical bandwidth to process the uncompressed high resolution
• 4:31
  video streams that the new AI computer demands on its 16 Volt
• 4:35
  system. Yeah, this is just the harsh
• 4:37
  reality of computer science colliding with physical limits.
• 4:40
  I mean, you physically cannot push the massive data required
• 4:43
  by these advanced neural networks through older lower
• 4:46
  capacity wiring. It is a matter of electrical
• 4:49
  gauge and data pipelines. 8 cameras sending uncompressed
• 4:53
  video simultaneously create an enormous stream of information.
• 4:57
  The older 12 Volt pipeline is simply too narrow to carry it
• 5:01
  all without dropping packets of data.
• 5:03
  I understand the physical limits.
• 5:04
  I get the computer science and the bandwidth constraints.
• 5:07
  Yeah, but if you are an owner who purchased one of those older
• 5:10
  vehicles, assuming you possessed all the necessary sensors for
• 5:12
  total autonomy, you are likely incredibly frustrated.
• 5:15
  Oh, absolutely. You were sold a promise of a
• 5:18
  self driving future. Finding out that a literal
• 5:21
  physical wiring harness is holding your car back from the
• 5:24
  latest updates is a really tough pill to swallow.
• 5:28
  I mean, people paid thousands of dollars for those software
• 5:31
  packages. The frustration is completely
• 5:33
  understandable. The unavoidable consequence is
• 5:36
  that this bandwidth limit forces the older fleet to run late
• 5:40
  versions of the software. It permanently caps their
• 5:43
  capability ceiling. The older cars get optimized,
• 5:47
  compressed versions of the neural network just to function,
• 5:50
  while the newer models receive the full uncompressed suite of
• 5:54
  advanced processing. And the most highly visible
• 5:56
  feature, strictly gated behind the newer processors, isn't even
• 6:00
  steering related at all. It is the conversational
• 6:03
  assistant residing in the dashboard.
• 6:05
  Yeah, this involves the full integration of the Grok AI
• 6:08
  Assistant, which is now accessible entirely hands free
• 6:11
  via the Hey Grok Wake Word. The vehicle utilizes its
• 6:14
  upgraded infotainment processor to run local natural language
• 6:18
  processing. That completely changes the
• 6:20
  cabin interface because you no longer need to tap a microphone
• 6:24
  icon or use rigid specific voice commands to get the car to
• 6:28
  understand you. It also has location based
• 6:32
  contextual awareness. Like imagine you are driving and
• 6:36
  you realize you need milk. You can sit in the driver's seat
• 6:39
  and simply ask the car to remind you to pick up groceries when it
• 6:42
  detects you are near your home. Yeah.
• 6:45
  That is so cool. The vehicle uses its GPS
• 6:47
  coordinates to trigger the assistant's memory the moment
• 6:49
  you cross an invisible geographic boundary around your
• 6:52
  house. Along with that sophisticated
• 6:54
  Voice Assistant, we have a completely redesigned self
• 6:56
  driving app situated right on the main display.
• 6:59
  OK, it features a simplified one tap subscription model, removing
• 7:02
  any friction from purchasing the software.
• 7:05
  More importantly, it features a dashboard specifically dedicated
• 7:08
  to tracking self driving streaks alongside your total miles
• 7:11
  driven. It is exactly like the UI design
• 7:14
  you find on fitness trackers. Closing your daily activity
• 7:17
  rings or keeping a language learning streak alive is
• 7:20
  replaced by maintaining your autonomous driving streak.
• 7:23
  They take that fundamental human psychological desire to keep a
• 7:27
  numerical streak alive and apply it directly to vehicle
• 7:30
  operation. Treating the driving experience
• 7:33
  like a daily habit app opens up an entirely new revenue
• 7:36
  approach. When you think about the
• 7:38
  business model, the entire interface is designed to boost
• 7:42
  recurring software subscriptions.
• 7:43
  Yeah, for sure. By keeping drivers constantly
• 7:46
  engaged with their stats, checking their streak, and
• 7:48
  viewing their autonomous mileage, you make the software
• 7:52
  feel indispensable. They also pushed lighter UI
• 7:55
  updates that directly change the physical cabin environment.
• 7:58
  The ambient cabin lights now glue a bright flashing red to
• 8:01
  warn you of blind spot dangers. Oh nice.
• 8:04
  Yeah, integrating safety directly into the interior
• 8:07
  lighting scheme, and pet mode includes new customizable
• 8:11
  avatars for the main screen. If you leave your air
• 8:14
  conditioning on for your pet, you can select a digital dog, a
• 8:17
  cat or a cyberhog wearing pixel sunglasses to display on the
• 8:22
  screen for anyone looking through the window.
• 8:24
  That is funny, but while tracking driving streaks and
• 8:27
  customizing digital hedgehogs entirely alters the digital
• 8:30
  cabin experience for you, the software updates are
• 8:33
  fundamentally rewriting how specific models utilize their
• 8:37
  physical brakes and defend against hackers.
• 8:39
  Right, we see this clearly with a new feature called Comfort
• 8:42
  Braking, which is an exclusive update for newer Model Y
• 8:45
  vehicles. It leverages a brand new dual
• 8:48
  master cylinder to let the AI smoothly modulate brake pressure
• 8:52
  during a complete stop. The dual master cylinder allows
• 8:55
  for an incredibly fine level of physical control when the
• 8:57
  vehicle is coming to a halt. The AI can micro adjust the
• 9:00
  hydraulic pressure on the brake calipers in the final fractions
• 9:03
  of a second. So smooth, it releases just
• 9:05
  enough pressure right as the tires stop rolling, which
• 9:08
  completely eliminates those abrupt, jerky halts that toss
• 9:12
  your head back against the seat. Then you have the Cyber Trek
• 9:15
  exclusives. There is an active Rd. noise
• 9:17
  reduction system that uses microphones planted inside the
• 9:21
  seats themselves. Wait inside the seats.
• 9:23
  Yeah, inside the seats, those microphones pick up the low
• 9:26
  frequency drone of the tires on the highway and instantly
• 9:29
  generate anti noise through the cabin speakers.
• 9:32
  It creates isolated quiet zones for the occupants by literally
• 9:36
  canceling out the sound waves. Wow, and there's also a parked
• 9:41
  blind spot warning system specific to that vehicle.
• 9:44
  If you are parked on a busy St. and the external cameras detect
• 9:47
  an approaching cyclist coming up with along the side, the
• 9:50
  software physically disables the interior door unlatched button
• 9:53
  on your first press it. Forces you to pause, preventing
• 9:56
  you from accidentally throwing your door open directly into the
• 9:59
  path of the cyclist. Wait, hold on, let's reset the
• 10:02
  pace here for a second. O software now dictates physical
• 10:06
  mechanics, but that exact software is under constant
• 10:09
  attack. Yeah, a recent Peen 2 owned
• 10:12
  security competition highlighted this vulnerability perfectly.
• 10:15
  A research team managed to achieve route permissions on the
• 10:18
  vehicle system via a sophisticated USB attack.
• 10:21
  Right, By plugging a device into the physical port, they bypassed
• 10:25
  the user level permissions and gained direct access to the core
• 10:29
  operating system. And once they had route access,
• 10:31
  they exposed private data access the vehicles stored Wi-Fi
• 10:35
  credentials, and potentially even gain the ability to execute
• 10:38
  physical vehicle controls. Scary stuff, yeah.
• 10:42
  Tesla countered the attack by immediately deploying an over
• 10:45
  the air security patch to the entire fleet.
• 10:48
  The ability to patch a car automatically while it is parked
• 10:50
  overnight in your garage drastically limits the lifespan
• 10:53
  of a cyber threat. The vulnerability existed, but
• 10:56
  it was closed almost immediately without anyone needing to visit
• 10:59
  a dealership or plug in a diagnostic tool.
• 11:01
  However, that reality permanently links passenger
• 11:04
  safety to the absolute integrity of a digital firewall.
• 11:08
  You are relying on the manufacturer to find and patch
• 11:11
  those holes faster than malicious actors can exploit
• 11:14
  them. Yeah.
• 11:15
  The physical environment inside the vehicle, like the doors, the
• 11:18
  brakes, the steering, is completely dependent on how
• 11:21
  secure that software layer remains against external
• 11:24
  manipulation. Tesla software updates are
• 11:27
  transforming vehicles into hardware dependent AI notes
• 11:30
  Reaction times, stopping smoothness and voice interfaces
• 11:33
  evolve automatically while parked, assuming the car
• 11:36
  possesses the physical processing power to handle the
• 11:39
  code. As the software demands outgrow
• 11:41
  the physical wiring of millions of cars already on the road,
• 11:44
  will the unified driving experience initially promised by
• 11:47
  the manufacturer permanently fracture into different classes
• 11:50
  of capability? If you're not subscribed yet,
• 11:52
  take a second and hit follow on whatever app you're using.
• 11:55
  It helps us keep making this. We appreciate you being here.