Latest / Elon Musk Podcast / Latest Tesla Robotaxi news

Transcript

Latest Tesla Robotaxi news · Elon Musk Podcast

• 0:00
  Welcome to the debate. Today we're looking at a
• 0:03
  controversy that sits right at the bleeding edge of
• 0:07
  transportation technology. It's a dispute that really
• 0:11
  divides the engineering world right down the middle, and the
• 0:14
  outcome is going to determine how and frankly if our vehicles
• 0:19
  drive themselves in the next decade.
• 0:22
  We are talking about the war between vision only systems and
• 0:25
  sensor fusion, specifically regarding the use of Lidar.
• 0:29
  And this isn't just a theoretical argument anymore, is
• 0:31
  it? We're looking at recent and,
• 0:33
  well, quite alarming reports about the deployment of Tesla's
• 0:37
  robo taxi fleet. The headline data suggests a
• 0:39
  crash rate that's significantly higher than human drivers.
• 0:43
  Some reports are saying up to four times higher.
• 0:46
  It it just raises a very uncomfortable question.
• 0:49
  Can a system that relies only on cameras ever truly match the
• 0:52
  reliability of a system that uses active laser sensors?
• 0:56
  And that's really the core of it.
• 0:58
  Can a computer see the world well enough with just video
• 1:01
  feeds to navigate safely? Or does excluding depth sensing
• 1:04
  hardware like Lidar resent an insurmountable barrier?
• 1:08
  I'm the advocate. My position is that visual input
• 1:11
  is, well, theoretically sufficient because it mimics the
• 1:15
  biological model. It mimics us.
• 1:17
  I also suspect the current crash data is being heavily
• 1:20
  misinterpreted because of some serious reporting biases.
• 1:24
  And I'm the dissenter. My position is that abandoning
• 1:27
  Lidar is a dangerous cost cutting measure that ignores the
• 1:31
  kind of redundancy you absolutely need for safety
• 1:34
  critical systems. We're seeing higher crash rates
• 1:37
  not because of a reporting bias, but because of a fundamental
• 1:40
  hardware deficit. When you take away the sensor
• 1:43
  that tells you exactly how far away an object is, you are
• 1:46
  introducing a level of risk that just shouldn't be on public
• 1:48
  roads. OK, so let's get into the
• 1:50
  machinery of this. We really need to start with the
• 1:53
  first principles argument, because this is the hill the
• 1:56
  vision only proponents are willing to die on.
• 1:59
  There's a perspective shared by a contributor to our source
• 2:02
  material, Retroviridae 6, that's basically an existence proof.
• 2:07
  Ah, the humans do it argument. Exactly.
• 2:09
  It's the biological argument. You and I drove here today.
• 2:12
  We navigated traffic. We merged.
• 2:14
  We avoided pedestrians. We did all of that using two
• 2:17
  passive optical sensors, our eyes, and a biological neural
• 2:20
  network, our brain. We don't have lidar in our
• 2:23
  foreheads, we don't emit laser pulses to measure time of
• 2:25
  flight, we don't have radar. We rely entirely on optical flow
• 2:29
  and pattern recognition. Sure.
• 2:31
  So Retrovira day six's point is that if a biological neural net
• 2:35
  can drive a car using only passive optical sensors, then it
• 2:39
  is physically possible for a synthetic neural net to do the
• 2:42
  same thing. The physics allows it.
• 2:45
  Therefore, the argument that Lidar is required is just false.
• 2:49
  Lidar might be a shortcut, but it isn't a necessity.
• 2:53
  The photons entering the camera contain all the information you
• 2:56
  need to drive. I'm sorry but I just don't buy
• 2:58
  that. Let me tell you why that is a
• 3:00
  huge category error. Your conflating otential with
• 3:04
  execution. Just because humans can drive
• 3:06
  with their eyes doesn't mean robots should drive without
• 3:08
  laser precision. But why not if the goal is to
• 3:11
  replicate human capability? Theology is full of flaws we're
• 3:14
  trying to engineer out of the system.
• 3:17
  The promise of autonomy isn't to drive as well as a distracted
• 3:20
  ape, it's to drive perfectly. And to drive perfectly, you need
• 3:24
  data that the human eye simply cannot provide.
• 3:28
  But we're not talking about fatigue.
• 3:29
  We're talking about the sensory input required to build a model
• 3:32
  of the world. But you cannot compete with
• 3:34
  Lidar are using only visual cameras when it comes to what
• 3:37
  I'd call ground truth. As another observer, Wiggly Worm
• 3:41
  pointed out in the materials, Lidar provides absolute depth
• 3:44
  data. So maybe we should break that
• 3:46
  down a bit for anyone who isn't a robotics engineer.
• 3:49
  Right. A camera is a passive sensor.
• 3:52
  It takes in light and it creates a flat 2D image.
• 3:56
  To figure out how far away a car is, the software has to analyze
• 4:00
  the size of that car in the image, compare it to what it
• 4:03
  thinks a car looks like, and then infer the distance it's
• 4:07
  guessing. It's a very educated guess, but
• 4:09
  it is a guess. It's inference based on
• 4:11
  perspective and parallax, yeah. Lidar is active, It shoots out a
• 4:16
  laser pulse, it hits a car, and it measures exactly how long it
• 4:20
  takes for the light to bounce back.
• 4:21
  It's simple physics. Distance equals time multiplied
• 4:24
  by the speed of light. It doesn't guess, it knows.
• 4:27
  It says there is an object 12.4 meters away.
• 4:32
  Wiggly Worm's point is that when you remove that sensor, you're
• 4:35
  forcing the computer to hallucinate death.
• 4:37
  And when you look at the stats, Tesla's robo taxi is reportedly
• 4:41
  crashing at a rate 4 times higher than humans.
• 4:44
  That isn't just the learning curve, that is a failure of
• 4:47
  perception. I think that statistic, the four
• 4:49
  times higher crash rate, is doing a lot of heavy lifting in
• 4:52
  your argument, and I I want to contextualize it.
• 4:56
  We need to be very careful about comparing apples to oranges
• 4:59
  here. A crash is a crash, isn't it?
• 5:02
  Not necessarily. Another analyst, Eskrove 2 He
• 5:07
  noted that this specific headline is based on a really
• 5:10
  small sample size. 5 incidents in Austin in a single month.
• 5:15
  But if you look at the granularity of those incidents,
• 5:18
  the data includes really minor events like a tire touching a
• 5:21
  parking sign or bumping A curb while parking. 5 incidents in a
• 5:25
  month for a small fleet is still high.
• 5:28
  But think about human behavior. If I scrape my rim on a curb
• 5:31
  while I'm parallel parking, or if I tap a plastic Bullard at
• 5:35
  one mile per hour, do I call the police?
• 5:37
  Do I call my insurance company? No.
• 5:39
  It never enters the statistical record.
• 5:41
  It just vanishes. Right, it's unreported.
• 5:44
  Exactly, but for a robo taxi, every single sensor reading is
• 5:49
  logged. Every thump is a reported
• 5:51
  incident. The system self-reports
• 5:54
  everything. So you're comparing reported
• 5:57
  autonomous incidents where every scratch is scrutinized against
• 6:01
  reported human accidents, which are usually only the the one
• 6:04
  severe enough to require tow truck.
• 6:07
  Ask contributors UX Test and Jerkletos pointed out you're
• 6:11
  comparing A microscope to a telescope and then claiming the
• 6:13
  microscope sees more dirt I. Understand the reporting bias
• 6:17
  argument. It's valid to a point, but I
• 6:19
  also think it's a convenient way to wave away failure.
• 6:22
  You call it rubbing a curb. I call it a failure of object
• 6:25
  permanence. That feels like a bit of a
• 6:27
  stretch for a scratched rim. Is it?
• 6:30
  Contrast this with Waymo. We have user experiences from
• 6:33
  San Francisco contributors like Turbo Encapsulator and Luda lol
• 6:37
  who describe Waymo as flawless in the same complex urban
• 6:41
  environments where these vision based systems are struggling.
• 6:44
  Waymo uses Lidar. They have that spinning bucket
• 6:47
  on the roof. They aren't scraping rims.
• 6:49
  They aren't bumping signs. They're also driving in a
• 6:52
  fishbowl. You're driving in San Francisco.
• 6:54
  That's hardly officiable. It's a Geo fenced pre mapped
• 6:57
  environment that Waymo knows exactly where every curb is
• 7:01
  because it has a high definition map stored in its hard drive.
• 7:05
  It's not seeing the curb, it's remembering it.
• 7:07
  Tesla's vision approach is trying to do something much much
• 7:10
  harder. Drive anywhere on any road
• 7:13
  without a map, just like a human.
• 7:15
  Of course it's going to be clumsier in the beginning.
• 7:18
  It's learning general intelligence, not just
• 7:19
  memorizing a map. But that clumsiness has real
• 7:22
  world consequences. The reports of these vision only
• 7:25
  cars hitting stationary objects like parking signs?
• 7:29
  That indicates A fundamental flaw.
• 7:31
  If a vision system cannot calculate the distance to a
• 7:33
  concrete Bullard well enough to avoid hitting it, how can we
• 7:36
  possibly trust it to calculate the velocity of a child running
• 7:40
  into the street? Because the neural networks are
• 7:42
  weighted differently for those tasks, the system is likely
• 7:45
  hyper cautious around pedestrians, but has a higher
• 7:48
  tolerance for static objects to facilitate, say, parking.
• 7:52
  That is an assumption. Lidar solves the static object
• 7:56
  problem instantly. It doesn't need to infer or wait
• 7:59
  anything. It hits the Ballard with a laser
• 8:01
  and it knows it's there. The fact that these vision based
• 8:04
  cars are hitting stationary objects suggests that the
• 8:07
  software is hallucinating free space where there is solid
• 8:10
  matter. That is terrifying.
• 8:13
  It implies the car literally does not know the physical
• 8:15
  boundaries of its own environment.
• 8:17
  I will concede that static object detection is a hurdle
• 8:20
  right now, but identifying these edge cases is exactly how you
• 8:25
  train the network. Every time it hits a parking
• 8:28
  sign at 2 mph, it uploads that failure and the entire fleet
• 8:31
  learns not to do it again. And this brings us directly to
• 8:35
  the concept of systemic risk. We have to talk about the
• 8:38
  multiplier effect. Explain how you view that.
• 8:40
  This was articulated very well by the source contributor Becker
• 8:43
  Hollow. The argument is all about error
• 8:46
  scaling. When a human makes a mistake, it
• 8:48
  causes 1 accident. Human error is stochastic.
• 8:51
  It's random. You might get distracted by a
• 8:53
  text. I might drop my coffee.
• 8:55
  It's isolated. Sure, individual variants.
• 8:58
  But when a vision based software has a flaw in its programming,
• 9:01
  say a specific inability to distinguish a white truck
• 9:04
  against a bright sky, that error is replicated across every
• 9:07
  single device on the road. The centralized bug.
• 9:10
  Exactly. If the software misinterprets a
• 9:12
  specific shadow or glare, thousands of cars become
• 9:15
  dangerous simultaneously in the exact same way.
• 9:18
  You aren't dealing with one bad driver, you're dealing with a
• 9:21
  fleet of clones all sharing the same blind spot.
• 9:24
  That is a systemic risk profile that we have never, ever dealt
• 9:27
  with in automotive history. That's an interesting point,
• 9:30
  though I would frame it differently.
• 9:32
  That logic flips both ways. It's actually the strongest
• 9:35
  argument for autonomous systems. Yes, an error is distributed,
• 9:39
  but so is the solution. If they catch it in time.
• 9:42
  Think about it, when a human driver is bad at merging,
• 9:45
  they're usually bad at merging forever.
• 9:47
  You can't patch their brain, but if you solve the edge case in
• 9:51
  software, if you fix that white truck against the sky bug, you
• 9:54
  instantly fix every car on the road.
• 9:57
  You can upgrade the safety of the entire fleet overnight with
• 10:00
  an over the air update. The multiplier effect applies to
• 10:02
  safety even more than it applies to error.
• 10:05
  You're leveraging the collective learning of millions of miles.
• 10:08
  But. Until that fix arrives, the risk
• 10:10
  is distributed to the public without their consent.
• 10:12
  The public roads are becoming a beta testing environment.
• 10:15
  We're seeing a move fast and break things mentality applied
• 10:19
  to two ton metal projectiles. Beckerhollow's logic holds the
• 10:23
  error rate is a normal human error multiplied by the number
• 10:26
  of devices using the program. If the program is flawed, the
• 10:29
  carnage is scalable. I see why you think that, but
• 10:32
  let me give you a different perspective on the technical
• 10:34
  reliability piece. You keep going back to Lidar as
• 10:37
  this source of truth, but Lidar has its own failure modes.
• 10:41
  It does, but they are different from cameras.
• 10:44
  Lidar struggles with heavy rain. The laser pulses scatter off the
• 10:48
  water droplets. It struggles with fog.
• 10:50
  It can get confused by interference from other lidar
• 10:53
  units. It's not magic.
• 10:55
  And this is where the whole sensor fusion argument gets
• 10:57
  tricky. Go on.
• 10:58
  When you have a camera an A lidar, they will often disagree.
• 11:02
  The camera sees a plastic bag blowing across the road and
• 11:05
  thinks it's nothing. Lidar sees an object and says
• 11:10
  obstacle emergency brake. Now the computer has to decide
• 11:13
  which sensor to trust. This is the sensor fusion
• 11:16
  conflict. By removing Lidar, Tesla is
• 11:20
  arguing that you remove the noise.
• 11:22
  You force the neural net to resolve the visual data just
• 11:25
  like a human does, without getting confused by conflicting
• 11:28
  signals. That sounds like a very
• 11:30
  convenient engineering rationalization for saving
• 11:33
  money. Cameras are cheap, Lidar is
• 11:35
  expensive. It's definitely cheaper, but
• 11:38
  retroverted Z6 mentioned. We went from the horse and buggy
• 11:41
  to the moon in just a few decades.
• 11:44
  Assuming that computer vision can't bridge the gap just
• 11:46
  because it hasn't yet is premature.
• 11:48
  The issue isn't that the camera is blind, it's that the
• 11:51
  processing isn't yet sophisticated enough.
• 11:54
  But processing power is scaling exponentially.
• 11:57
  Software cannot conjure photons where there are none.
• 12:01
  That's the physics problem. But it can interpret context.
• 12:04
  Let's talk about those photons. Cameras are passive.
• 12:08
  They need light. What happens when you drive
• 12:10
  directly into the sunset? We've all done it.
• 12:12
  The visor goes down. You squint.
• 12:14
  You can barely see. Cameras get blinded by sun
• 12:16
  glare. They get obscured by mud.
• 12:18
  We have reports that Tesla is having to employ trailing chase
• 12:22
  cars with human safety monitors for their autonomous taxis.
• 12:26
  If the system is so theoretically sound, why does it
• 12:29
  need a human babysitter in a separate vehicle?
• 12:31
  Every developmental technology has safety protocols during
• 12:35
  testing, but. This is being sold as a future
• 12:37
  that is just around the corner. The reliance on cameras
• 12:41
  introduces A fragility that Lidar solves.
• 12:44
  Lidar cuts through sun glare. It works in total darkness.
• 12:48
  It is a second layer of truth. If the camera sees a shadow and
• 12:52
  thinks it's a hole in the road, the Lidar says no, the ground is
• 12:55
  flat. Removing that sensor removes a
• 12:58
  layer of survival. It's engineering hubris to
• 13:01
  believe you can derive 100% certainty from a sensor that is
• 13:04
  susceptible to optical illusions.
• 13:06
  I'm not convinced by that line of reasoning because it assumes
• 13:09
  we can't solve optical illusions with better AI.
• 13:12
  But let's pivot to the consequences of this, because
• 13:15
  the legal aspect is fascinating. It's a nightmare.
• 13:18
  This leads us to the inevitable question of accountability.
• 13:21
  When these systems do fail, whether it's a clumsy bump or a
• 13:25
  serious collision, who is responsible?
• 13:27
  This is the question posed by Shifty Mennonite in our source
• 13:30
  threads. Who is going to be held
• 13:33
  accountable when these things mow people down?
• 13:35
  It is a legal quagmire. Is it the driver, which in this
• 13:39
  case is the software? Is it the manufacturer or is it
• 13:42
  the limitations of the sensor suite itself?
• 13:44
  I think we need to distinguish between a software bug and a
• 13:48
  design choice. This is crucial.
• 13:51
  If a car crashes because of a line of bad code, that's one
• 13:54
  thing. But if a manufacturer knowingly
• 13:57
  removes a safety sensor like Lidar, a sensor that is industry
• 14:00
  standard for competitors like Waymo, and that removal leads to
• 14:04
  a crash because the camera couldn't estimate depth, that
• 14:07
  feels very distinct from a mere coding error.
• 14:09
  You're suggesting negligence. I'm saying it borders on it.
• 14:13
  There's this sentiment expressed by User Beneficial Soup 3699
• 14:17
  regarding blatant fraud. While that is, you know, strong
• 14:21
  language, the core sentiment is valid.
• 14:24
  If you claim a camera is sufficient and the physics
• 14:26
  suggest it isn't, and the data shows it crashing, at what point
• 14:30
  does adherence to a vision only philosophy become liability?
• 14:33
  But that assumes Lidar was prevented that specific crash.
• 14:36
  We don't know that. Like I said, LIDAR isn't a magic
• 14:40
  bullet. It's not magic, it's redundancy.
• 14:43
  In aviation, we don't fly with one altimeter, we have three.
• 14:47
  Why on earth should we drive with one type of eye?
• 14:49
  If the camera fails due to glare or a bug or mud, there's nothing
• 14:53
  to catch the car. It is a single point of failure
• 14:56
  system. But there's an economic argument
• 14:58
  here too. If you require Lidar, you make
• 15:01
  autonomous cars cost $100,000. They become toys for the rich.
• 15:05
  If you can solve it with vision, the hardware costs 500 $100.
• 15:09
  You can put it in every car. A vision based system that's 99%
• 15:13
  safe and available to everyone might save more total lives than
• 15:17
  a lighter system that's 99.9% safe but only 1000 people can
• 15:21
  afford it. That is a utilitarian calculus
• 15:24
  that works on a spreadsheet, but it doesn't work when you're the
• 15:27
  one crossing the street. The public Rd. should not be a
• 15:30
  testing ground for cost cutting measures, disguises innovation.
• 15:35
  The experiences of users in San Francisco and Austin show a
• 15:37
  clear divide. Waymo with Lidar is providing A
• 15:40
  flawless service while vision only systems are struggling with
• 15:44
  basic static objects. But again, Waymo is on rails.
• 15:47
  It's a local maximum. It's great for San Francisco,
• 15:50
  but it doesn't scale to the rest of the world.
• 15:53
  I'd rather have a safe local maximum than a dangerous global
• 15:56
  beta test. Until vision systems can match
• 15:58
  the redundancy and depth accuracy of Lidar, the safety
• 16:01
  consequences are real and they are statistically proven.
• 16:05
  We cannot verify the safety of a black box neural net without
• 16:08
  ground truth sensors. It ultimately comes down to that
• 16:11
  multiplier effect we talked about.
• 16:13
  It does. We are at a crossroads.
• 16:16
  We can take the safe, expensive route with Lidar, which might
• 16:19
  limit the scalability of the technology but provides that
• 16:22
  warm blanket of redundancy. Or we can push for the vision
• 16:26
  solution, which, if it works, multiply safety exponentially
• 16:31
  across the globe and solves general intelligence.
• 16:34
  But if it fails, it multiplies error.
• 16:36
  It multiplies the risk of a single software blind spot into
• 16:40
  a nationwide. Hazard, and that is the gamble.
• 16:44
  Is the current risk worth the future reward?
• 16:47
  I tend to believe that without taking that risk, we stagnate.
• 16:51
  We'd still be driving horses if we waited for the perfect car.
• 16:55
  And I would argue that safety is not a place for gambling when
• 16:59
  you're moving 2 tons of steel at 60 mph.
• 17:02
  Pretty good isn't good enough. You need absolute truth and
• 17:05
  cameras just don't provide that. A fundamental disagreement on
• 17:08
  the philosophy of engineering. Thank you for listening to the
• 17:12
  debate. We hope this exchange has
• 17:14
  illuminated the complexities behind the sensors.
• 17:17
  Drive safe everyone, and watch out for the robots.
• 17:20
  Goodbye.