back to index

Elon Musk: Tesla Autopilot | Lex Fridman Podcast #18


small model | large model

link |
00:00:00.000
The following is a conversation with Elon Musk.
link |
00:00:03.000
He's the CEO of Tesla, SpaceX, Neuralink,
link |
00:00:06.240
and a cofounder of several other companies.
link |
00:00:09.200
This conversation is part
link |
00:00:10.740
of the artificial intelligence podcast.
link |
00:00:13.200
The series includes leading researchers
link |
00:00:15.640
in academia and industry, including CEOs and CTOs
link |
00:00:19.320
of automotive, robotics, AI, and technology companies.
link |
00:00:24.080
This conversation happened after the release of the paper
link |
00:00:26.880
from our group at MIT on driver functional vigilance
link |
00:00:30.520
during use of Tesla's autopilot.
link |
00:00:32.880
The Tesla team reached out to me,
link |
00:00:34.560
offering a podcast conversation with Mr. Musk.
link |
00:00:37.480
I accepted with full control of questions I could ask
link |
00:00:40.640
and the choice of what is released publicly.
link |
00:00:43.560
I ended up editing out nothing of substance.
link |
00:00:46.840
I've never spoken with Elon before this conversation,
link |
00:00:49.680
publicly or privately.
link |
00:00:51.720
Neither he nor his companies have any influence
link |
00:00:54.360
on my opinion, nor on the rigor and integrity
link |
00:00:57.840
of the scientific method that I practice
link |
00:00:59.760
in my position at MIT.
link |
00:01:01.840
Tesla has never financially supported my research
link |
00:01:04.640
and I've never owned a Tesla vehicle.
link |
00:01:07.320
I've never owned Tesla stock.
link |
00:01:10.160
This podcast is not a scientific paper.
link |
00:01:12.800
It is a conversation.
link |
00:01:14.360
I respect Elon as I do all other leaders
link |
00:01:16.720
and engineers I've spoken with.
link |
00:01:18.680
We agree on some things and disagree on others.
link |
00:01:21.440
My goal is always with these conversations
link |
00:01:23.480
is to understand the way the guest sees the world.
link |
00:01:26.920
One particular point of this agreement
link |
00:01:28.600
in this conversation was the extent
link |
00:01:30.640
to which camera based driver monitoring
link |
00:01:33.240
will improve outcomes and for how long
link |
00:01:36.120
it will remain relevant for AI assisted driving.
link |
00:01:39.960
As someone who works on and is fascinated
link |
00:01:42.240
by human centered artificial intelligence,
link |
00:01:45.200
I believe that if implemented and integrated effectively,
link |
00:01:48.720
camera based driver monitoring is likely to be of benefit
link |
00:01:51.840
in both the short term and the long term.
link |
00:01:55.640
In contrast, Elon and Tesla's focus
link |
00:01:59.240
is on the improvement of autopilot
link |
00:02:01.200
such that its statistical safety benefits
link |
00:02:04.480
override any concern of human behavior and psychology.
link |
00:02:09.040
Elon and I may not agree on everything
link |
00:02:12.040
but I deeply respect the engineering
link |
00:02:13.920
and innovation behind the efforts that he leads.
link |
00:02:16.880
My goal here is to catalyze a rigorous, nuanced
link |
00:02:20.560
and objective discussion in industry and academia
link |
00:02:23.520
on AI assisted driving,
link |
00:02:26.240
one that ultimately makes for a safer and better world.
link |
00:02:30.840
And now here's my conversation with Elon Musk.
link |
00:02:35.600
What was the vision, the dream of autopilot
link |
00:02:38.640
when in the beginning the big picture system level
link |
00:02:41.400
when it was first conceived
link |
00:02:43.680
and started being installed in 2014
link |
00:02:45.960
in the hardware and the cars?
link |
00:02:47.520
What was the vision, the dream?
link |
00:02:49.760
I would characterize the vision or dream
link |
00:02:51.400
simply that there are obviously two
link |
00:02:54.400
massive revolutions in the automobile industry.
link |
00:03:00.120
One is the transition to electrification
link |
00:03:04.440
and then the other is autonomy.
link |
00:03:07.720
And it became obvious to me that in the future
link |
00:03:13.240
any car that does not have autonomy
link |
00:03:16.240
I would be about as useful as a horse.
link |
00:03:19.160
Which is not to say that there's no use, it's just rare
link |
00:03:22.040
and somewhat idiosyncratic
link |
00:03:23.640
if somebody has a horse at this point.
link |
00:03:25.480
It's just obvious that cars will drive themselves completely.
link |
00:03:28.000
It's just a question of time
link |
00:03:29.600
and if we did not participate in the autonomy revolution
link |
00:03:36.920
then our cars would not be useful to people
link |
00:03:40.840
relative to cars that are autonomous.
link |
00:03:43.680
I mean an autonomous car is arguably worth
link |
00:03:47.160
five to 10 times more than a car that which is not autonomous.
link |
00:03:53.760
In the long term.
link |
00:03:55.160
Turns out what you mean by long term,
link |
00:03:56.200
but let's say at least for the next five years
link |
00:03:59.520
perhaps 10 years.
link |
00:04:01.440
So there are a lot of very interesting design choices
link |
00:04:04.080
with autopilot early on.
link |
00:04:05.720
First is showing on the instrument cluster
link |
00:04:09.960
or in the Model 3 on the center stack display
link |
00:04:12.680
what the combined sensor suite sees.
link |
00:04:15.720
What was the thinking behind that choice?
link |
00:04:17.920
Was there a debate?
link |
00:04:18.960
What was the process?
link |
00:04:20.480
The whole point of the display is to provide a health check
link |
00:04:24.840
on the vehicle's perception of reality.
link |
00:04:28.080
So the vehicle's taking information for a bunch of sensors
link |
00:04:31.320
primarily cameras, but also radar and ultrasonics,
link |
00:04:34.680
GPS and so forth.
link |
00:04:37.200
And then that information is then rendered into vector space
link |
00:04:42.200
and that with a bunch of objects with properties
link |
00:04:46.360
like lane lines and traffic lights and other cars.
link |
00:04:49.920
And then in vector space that is re rendered onto a display
link |
00:04:54.920
so you can confirm whether the car knows
link |
00:04:57.400
what's going on or not by looking out the window.
link |
00:05:01.600
Right, I think that's an extremely powerful thing
link |
00:05:04.240
for people to get an understanding
link |
00:05:06.480
to become one with the system
link |
00:05:07.840
and understanding what the system is capable of.
link |
00:05:10.400
Now, have you considered showing more?
link |
00:05:13.600
So if we look at the computer vision,
link |
00:05:16.400
you know, like road segmentation, lane detection,
link |
00:05:18.400
vehicle detection, object detection, underlying the system,
link |
00:05:21.640
there is at the edges some uncertainty.
link |
00:05:24.400
Have you considered revealing the parts
link |
00:05:28.400
that the uncertainty in the system, the sort of problem
link |
00:05:32.400
these associated with say image recognition
link |
00:05:35.000
or something like that?
link |
00:05:35.840
Yeah, so right now it shows like the vehicles
link |
00:05:37.840
and the vicinity of very clean crisp image
link |
00:05:40.840
and people do confirm that there's a car in front of me
link |
00:05:43.840
and the system sees there's a car in front of me
link |
00:05:45.840
but to help people build an intuition
link |
00:05:47.840
of what computer vision is by showing some of the uncertainty.
link |
00:05:51.840
Well, I think it's, in my car,
link |
00:05:53.840
I always look at the sort of the debug view
link |
00:05:56.840
and there's two debug views.
link |
00:05:58.840
One is augmented vision, which I'm sure you've seen
link |
00:06:03.840
where it's basically, we draw boxes and labels
link |
00:06:07.840
around objects that are recognized.
link |
00:06:10.840
And then there's what we call the visualizer,
link |
00:06:14.840
which is basically a vector space representation
link |
00:06:16.840
summing up the input from all sensors.
link |
00:06:21.840
That does not show any pictures,
link |
00:06:23.840
but it shows all of the,
link |
00:06:26.840
it basically shows the cause view of the world in vector space.
link |
00:06:32.840
But I think this is very difficult for normal people to understand.
link |
00:06:36.840
They would not know what they're looking at.
link |
00:06:38.840
So it's almost an HMI challenge.
link |
00:06:40.840
The current things that are being displayed
link |
00:06:42.840
is optimized for the general public understanding
link |
00:06:46.840
of what the system is capable of.
link |
00:06:48.840
It's like if you have no idea how computer vision works
link |
00:06:50.840
or anything, you can still look at the screen
link |
00:06:52.840
and see if the car knows what's going on.
link |
00:06:54.840
And then if you're a development engineer
link |
00:06:57.840
or if you have the development build like I do,
link |
00:07:01.840
then you can see all the debug information.
link |
00:07:05.840
But those would just be total diverse to most people.
link |
00:07:10.840
What's your view on how to best distribute effort?
link |
00:07:13.840
So there's three, I would say, technical aspects of autopilot
link |
00:07:16.840
that are really important.
link |
00:07:18.840
So it's the underlying algorithms,
link |
00:07:19.840
like the neural network architecture.
link |
00:07:21.840
There's the data that's trained on
link |
00:07:23.840
and then there's the hardware development.
link |
00:07:25.840
There may be others.
link |
00:07:26.840
But so look, algorithm, data, hardware.
link |
00:07:31.840
You only have so much money, only have so much time.
link |
00:07:34.840
What do you think is the most important thing
link |
00:07:36.840
to allocate resources to?
link |
00:07:39.840
Do you see it as pretty evenly distributed
link |
00:07:41.840
between those three?
link |
00:07:43.840
We automatically get fast amounts of data
link |
00:07:46.840
because all of our cars have
link |
00:07:50.840
eight external facing cameras and radar
link |
00:07:54.840
and usually 12 ultrasonic sensors, GPS, obviously,
link |
00:07:59.840
and IMU.
link |
00:08:03.840
And so we basically have a fleet that has,
link |
00:08:08.840
we've got about 400,000 cars on the road
link |
00:08:11.840
that have that level of data.
link |
00:08:13.840
I think you keep quite close track of it, actually.
link |
00:08:15.840
Yes.
link |
00:08:16.840
So we're approaching half a million cars
link |
00:08:19.840
on the road that have the full sensor suite.
link |
00:08:22.840
So this is, I'm not sure how many other cars
link |
00:08:26.840
on the road have this sensor suite,
link |
00:08:28.840
but I'd be surprised if it's more than 5,000,
link |
00:08:31.840
which means that we have 99% of all the data.
link |
00:08:35.840
So there's this huge inflow of data.
link |
00:08:37.840
Absolutely, massive inflow of data.
link |
00:08:39.840
And then it's taken about three years,
link |
00:08:43.840
but now we've finally developed our full self driving computer,
link |
00:08:46.840
which can process
link |
00:08:51.840
an order of magnitude as much as the NVIDIA system
link |
00:08:54.840
that we currently have in the cars.
link |
00:08:56.840
And it's really just to use it,
link |
00:08:58.840
you unplug the NVIDIA computer and plug the Tesla computer in.
link |
00:09:01.840
And that's it.
link |
00:09:03.840
And it's, in fact, we're not even,
link |
00:09:06.840
we're still exploring the boundaries of its capabilities,
link |
00:09:09.840
but we're able to run the cameras at full frame rate,
link |
00:09:11.840
full resolution, not even crop of the images,
link |
00:09:14.840
and it's still got headroom, even on one of the systems.
link |
00:09:19.840
The full self driving computer is really two computers,
link |
00:09:22.840
two systems on a chip that are fully redundant.
link |
00:09:25.840
So you could put a bolt through basically any part of that system
link |
00:09:28.840
and it still works.
link |
00:09:29.840
The redundancy, are they perfect copies of each other?
link |
00:09:32.840
Or also it's purely for redundancy
link |
00:09:35.840
as opposed to an arguing machine kind of architecture
link |
00:09:37.840
where they're both making decisions.
link |
00:09:39.840
This is purely for redundancy.
link |
00:09:41.840
I think it's more like, if you have a twin engine aircraft,
link |
00:09:44.840
commercial aircraft,
link |
00:09:46.840
this system will operate best if both systems are operating,
link |
00:09:51.840
but it's capable of operating safely on one.
link |
00:09:55.840
So, but as it is right now, we can just run,
link |
00:09:59.840
we haven't even hit the edge of performance,
link |
00:10:03.840
so there's no need to actually distribute
link |
00:10:08.840
functionality across both SoCs.
link |
00:10:12.840
We can actually just run a full duplicate on each one.
link |
00:10:16.840
You haven't really explored or hit the limit of the system?
link |
00:10:20.840
Not yet, hit the limit now.
link |
00:10:21.840
So the magic of deep learning is that it gets better with data.
link |
00:10:26.840
You said there's a huge inflow of data,
link |
00:10:28.840
but the thing about driving the really valuable data
link |
00:10:33.840
to learn from is the edge cases.
link |
00:10:35.840
So how do you, I mean, I've heard you talk somewhere about
link |
00:10:42.840
autopilot disengagement as being an important moment of time to use.
link |
00:10:46.840
Is there other edge cases or perhaps can you speak to those edge cases,
link |
00:10:51.840
what aspects of them might be valuable,
link |
00:10:53.840
or if you have other ideas,
link |
00:10:55.840
how to discover more and more and more edge cases in driving?
link |
00:10:59.840
Well, there's a lot of things that I learned.
link |
00:11:01.840
There are certainly edge cases where I say somebody's on autopilot
link |
00:11:05.840
and they take over.
link |
00:11:07.840
And then, okay, that's a trigger that goes to a system that says,
link |
00:11:12.840
okay, do they take over for convenience
link |
00:11:14.840
or do they take over because the autopilot wasn't working properly?
link |
00:11:18.840
There's also, like let's say we're trying to figure out
link |
00:11:21.840
what is the optimal spline for traversing an intersection.
link |
00:11:26.840
Then the ones where there are no interventions
link |
00:11:30.840
and are the right ones.
link |
00:11:32.840
So you then say, okay, when it looks like this, do the following.
link |
00:11:36.840
And then you get the optimal spline for a complex,
link |
00:11:40.840
now getting a complex intersection.
link |
00:11:44.840
So that's for, there's kind of the common case.
link |
00:11:48.840
You're trying to capture a huge amount of samples
link |
00:11:51.840
of a particular intersection, how one thing went right.
link |
00:11:54.840
And then there's the edge case where, as you said,
link |
00:11:58.840
not for convenience, but something didn't go exactly right.
link |
00:12:01.840
Somebody took over, somebody asserted manual control from autopilot.
link |
00:12:04.840
And really, like the way to look at this is view all input is error.
link |
00:12:08.840
If the user had to do input, it does something.
link |
00:12:11.840
All input is error.
link |
00:12:13.840
That's a powerful line to think of it that way,
link |
00:12:15.840
because it may very well be error.
link |
00:12:17.840
But if you want to exit the highway,
link |
00:12:19.840
or if you want to, it's a navigation decision
link |
00:12:22.840
that all autopilot is not currently designed to do,
link |
00:12:24.840
then the driver takes over.
link |
00:12:26.840
How do you know the difference?
link |
00:12:28.840
Yeah, that's going to change with navigate and autopilot,
link |
00:12:30.840
which we've just released, and without stall confirm.
link |
00:12:33.840
So the navigation, like lane change based,
link |
00:12:36.840
like asserting control in order to do a lane change,
link |
00:12:39.840
or exit a freeway, or doing highway interchange,
link |
00:12:43.840
the vast majority of that will go away with the release
link |
00:12:47.840
that just went out.
link |
00:12:49.840
Yeah, I don't think people quite understand
link |
00:12:52.840
how big of a step that is.
link |
00:12:54.840
Yeah, they don't.
link |
00:12:55.840
If you drive the car, then you do.
link |
00:12:57.840
So you still have to keep your hands on the steering wheel
link |
00:12:59.840
currently when it does the automatic lane change?
link |
00:13:02.840
What are...
link |
00:13:04.840
So there's these big leaps through the development of autopilot
link |
00:13:07.840
through its history,
link |
00:13:09.840
and what stands out to you as the big leaps?
link |
00:13:12.840
I would say this one,
link |
00:13:14.840
navigate and autopilot without having to confirm,
link |
00:13:19.840
is a huge leap.
link |
00:13:20.840
It is a huge leap.
link |
00:13:21.840
It also automatically overtakes slow cars.
link |
00:13:24.840
So it's both navigation and seeking the fastest lane.
link |
00:13:30.840
So it'll overtake a slow cause and exit the freeway
link |
00:13:36.840
and take highway interchanges.
link |
00:13:39.840
And then we have traffic light recognition,
link |
00:13:46.840
which is introduced initially as a warning.
link |
00:13:49.840
I mean, on the development version that I'm driving,
link |
00:13:51.840
the car fully stops and goes at traffic lights.
link |
00:13:55.840
So those are the steps, right?
link |
00:13:57.840
You just mentioned something sort of
link |
00:13:59.840
including a step towards full autonomy.
link |
00:14:02.840
What would you say are the biggest technological roadblocks
link |
00:14:07.840
to full cell driving?
link |
00:14:09.840
Actually, I don't think...
link |
00:14:10.840
I think we just...
link |
00:14:11.840
the full cell driving computer that we just...
link |
00:14:13.840
that has a...
link |
00:14:14.840
what we call the FSD computer.
link |
00:14:16.840
That's now in production.
link |
00:14:20.840
So if you order any Model SRX or any Model 3
link |
00:14:25.840
that has the full cell driving package,
link |
00:14:28.840
you'll get the FSD computer.
link |
00:14:31.840
That's important to have enough base computation.
link |
00:14:36.840
Then refining the neural net and the control software.
link |
00:14:40.840
But all of that can just be provided as an over there update.
link |
00:14:44.840
The thing that's really profound,
link |
00:14:46.840
and where I'll be emphasizing at the...
link |
00:14:50.840
that investor day that we're having focused on autonomy,
link |
00:14:52.840
is that the cars currently being produced,
link |
00:14:55.840
or the hardware currently being produced,
link |
00:14:57.840
is capable of full cell driving.
link |
00:15:00.840
But capable is an interesting word because...
link |
00:15:03.840
Like the hardware is.
link |
00:15:05.840
And as we refine the software,
link |
00:15:08.840
the capabilities will increase dramatically
link |
00:15:11.840
and then the reliability will increase dramatically
link |
00:15:13.840
and then it will receive regulatory approval.
link |
00:15:15.840
So essentially buying a car today is an investment in the future.
link |
00:15:18.840
You're essentially buying...
link |
00:15:21.840
I think the most profound thing is that
link |
00:15:25.840
if you buy a Tesla today,
link |
00:15:27.840
I believe you are buying an appreciating asset,
link |
00:15:29.840
not a depreciating asset.
link |
00:15:32.840
So that's a really important statement there
link |
00:15:34.840
because if hardware is capable enough,
link |
00:15:36.840
that's the hard thing to upgrade usually.
link |
00:15:39.840
Exactly.
link |
00:15:40.840
So then the rest is a software problem.
link |
00:15:43.840
Yes. Software has no marginal cost, really.
link |
00:15:47.840
But what's your intuition on the software side?
link |
00:15:51.840
How hard are the remaining steps
link |
00:15:55.840
to get it to where...
link |
00:15:58.840
you know, the experience,
link |
00:16:02.840
not just the safety, but the full experience
link |
00:16:05.840
is something that people would enjoy.
link |
00:16:08.840
I think people would enjoy it very much on the highways.
link |
00:16:12.840
It's a total game changer for quality of life,
link |
00:16:16.840
for using Tesla autopilot on the highways.
link |
00:16:20.840
So it's really just extending that functionality to city streets,
link |
00:16:24.840
adding in the traffic light recognition,
link |
00:16:28.840
navigating complex intersections,
link |
00:16:31.840
and then being able to navigate complicated parking lots
link |
00:16:36.840
so the car can exit a parking space
link |
00:16:39.840
and come and find you even if it's in a complete maze of a parking lot.
link |
00:16:45.840
And then you can just drop you off and find a parking spot by itself.
link |
00:16:51.840
Yeah, in terms of enjoyability
link |
00:16:53.840
and something that people would actually find a lot of use from,
link |
00:16:57.840
the parking lot is a really...
link |
00:17:00.840
it's rich of annoyance when you have to do it manually,
link |
00:17:03.840
so there's a lot of benefit to be gained from automation there.
link |
00:17:07.840
So let me start injecting the human into this discussion a little bit.
link |
00:17:11.840
So let's talk about full autonomy.
link |
00:17:14.840
If you look at the current level four vehicles,
link |
00:17:17.840
being Tesla and road like Waymo and so on,
link |
00:17:19.840
they're only technically autonomous.
link |
00:17:22.840
They're really level two systems
link |
00:17:25.840
with just a different design philosophy
link |
00:17:28.840
because there's always a safety driver in almost all cases
link |
00:17:31.840
and they're monitoring the system.
link |
00:17:33.840
Maybe Tesla's full self driving
link |
00:17:37.840
is still for a time to come,
link |
00:17:41.840
requiring supervision of the human being.
link |
00:17:44.840
So its capabilities are powerful enough to drive,
link |
00:17:47.840
but nevertheless requires the human to still be supervising
link |
00:17:50.840
just like a safety driver is in a other fully autonomous vehicles.
link |
00:17:56.840
I think it will require detecting hands on wheel
link |
00:18:01.840
or at least six months or something like that from here.
link |
00:18:08.840
Really it's a question of like,
link |
00:18:11.840
from a regulatory standpoint,
link |
00:18:15.840
how much safer than a person does autopilot need to be
link |
00:18:19.840
for it to be okay to not monitor the car?
link |
00:18:24.840
And this is a debate that one can have.
link |
00:18:27.840
But you need a large amount of data
link |
00:18:31.840
so you can prove with high confidence,
link |
00:18:34.840
statistically speaking,
link |
00:18:36.840
that the car is dramatically safer than a person
link |
00:18:39.840
and that adding in the person monitoring
link |
00:18:42.840
does not materially affect the safety.
link |
00:18:45.840
So it might need to be like two or three hundred percent safer than a person.
link |
00:18:49.840
And how do you prove that?
link |
00:18:51.840
Incidence per mile.
link |
00:18:53.840
So crashes and fatalities.
link |
00:18:56.840
Yeah, fatalities would be a factor,
link |
00:18:58.840
but there are just not enough fatalities
link |
00:19:00.840
to be statistically significant at scale.
link |
00:19:03.840
But there are enough crashes,
link |
00:19:06.840
there are far more crashes than there are fatalities.
link |
00:19:10.840
So you can assess what is the probability of a crash,
link |
00:19:15.840
then there's another step which probability of injury
link |
00:19:19.840
and probability of permanent injury
link |
00:19:21.840
and probability of death.
link |
00:19:23.840
And all of those need to be much better than a person
link |
00:19:27.840
by at least perhaps two hundred percent.
link |
00:19:32.840
And you think there's the ability to have a healthy discourse
link |
00:19:36.840
with the regulatory bodies on this topic?
link |
00:19:39.840
I mean, there's no question that regulators pay
link |
00:19:43.840
disproportionate amount of attention to that which generates press.
link |
00:19:48.840
This is just an objective fact.
link |
00:19:50.840
And Tesla generates a lot of press.
link |
00:19:52.840
So that, you know, in the United States,
link |
00:19:56.840
there's I think almost 40,000 automotive deaths per year.
link |
00:20:00.840
But if there are four in Tesla,
link |
00:20:03.840
they'll probably receive a thousand times more press
link |
00:20:06.840
than anyone else.
link |
00:20:08.840
So the psychology of that is actually fascinating.
link |
00:20:10.840
I don't think we'll have enough time to talk about that,
link |
00:20:12.840
but I have to talk to you about the human side of things.
link |
00:20:16.840
So myself and our team at MIT recently released a paper
link |
00:20:20.840
on functional vigilance of drivers while using autopilot.
link |
00:20:24.840
This is work we've been doing since autopilot was first
link |
00:20:27.840
released publicly over three years ago,
link |
00:20:30.840
collecting video driver faces and driver body.
link |
00:20:34.840
So I saw that you tweeted a quote from the abstract
link |
00:20:38.840
so I can at least guess that you've glanced at it.
link |
00:20:43.840
Can I talk you through what we found?
link |
00:20:46.840
Okay, so it appears that in the data that we've collected
link |
00:20:51.840
that drivers are maintaining functional vigilance
link |
00:20:54.840
such that we're looking at 18,000 disengagement
link |
00:20:57.840
from autopilot, 18,900 and annotating were they able
link |
00:21:02.840
to take over control in a timely manner?
link |
00:21:05.840
So they were there present looking at the road
link |
00:21:07.840
to take over control.
link |
00:21:09.840
Okay, so this goes against what many would predict
link |
00:21:14.840
from the body of literature on vigilance with automation.
link |
00:21:18.840
Now the question is, do you think these results
link |
00:21:21.840
hold across the broader population?
link |
00:21:23.840
So ours is just a small subset.
link |
00:21:26.840
Do you think one of the criticism is that there's
link |
00:21:30.840
a small minority of drivers that may be highly responsible
link |
00:21:34.840
where their vigilance decrement would increase
link |
00:21:37.840
with autopilot use?
link |
00:21:39.840
I think this is all really going to be swept.
link |
00:21:41.840
I mean, the system's improving so much so fast
link |
00:21:46.840
that this is going to be a mood point very soon
link |
00:21:50.840
where vigilance is, if something's many times safer
link |
00:21:56.840
than a person, then adding a person does,
link |
00:22:00.840
the effect on safety is limited.
link |
00:22:04.840
And in fact, it could be negative.
link |
00:22:09.840
That's really interesting.
link |
00:22:11.840
So the fact that a human may, some percent of the population
link |
00:22:16.840
may exhibit a vigilance decrement will not affect
link |
00:22:20.840
overall statistics numbers of safety.
link |
00:22:22.840
No, in fact, I think it will become very, very quickly,
link |
00:22:27.840
maybe even towards the end of this year,
link |
00:22:29.840
but I'd say I'd be shocked if it's not next year,
link |
00:22:32.840
at the latest, that having a human intervene
link |
00:22:36.840
will increase safety.
link |
00:22:39.840
Decrease.
link |
00:22:40.840
I can imagine if you're an elevator.
link |
00:22:42.840
Now, it used to be that there were elevator operators
link |
00:22:45.840
and you couldn't go on an elevator by yourself
link |
00:22:47.840
and work the lever to move between floors.
link |
00:22:51.840
And now, nobody wants an elevator operator
link |
00:22:56.840
because the automated elevator that stops the floors
link |
00:23:00.840
is much safer than the elevator operator.
link |
00:23:03.840
And in fact, it would be quite dangerous
link |
00:23:05.840
if someone with a lever that can move
link |
00:23:07.840
the elevator between floors.
link |
00:23:09.840
So that's a really powerful statement
link |
00:23:12.840
and a really interesting one.
link |
00:23:14.840
But I also have to ask, from a user experience
link |
00:23:16.840
and from a safety perspective,
link |
00:23:18.840
one of the passions for me algorithmically
link |
00:23:20.840
is camera based detection of sensing the human,
link |
00:23:25.840
but detecting what the driver is looking at,
link |
00:23:27.840
cognitive load, body pose.
link |
00:23:29.840
On the computer vision side, that's a fascinating problem,
link |
00:23:31.840
but there's many in industry who believe
link |
00:23:34.840
you have to have camera based driver monitoring.
link |
00:23:37.840
Do you think this could be benefit gained
link |
00:23:39.840
from driver monitoring?
link |
00:23:41.840
If you have a system that's out or below
link |
00:23:45.840
human level reliability, then driver monitoring makes sense.
link |
00:23:49.840
But if your system is dramatically better,
link |
00:23:51.840
more reliable than a human,
link |
00:23:53.840
then driver monitoring is not help much.
link |
00:23:58.840
And like I said, you wouldn't want someone into...
link |
00:24:03.840
You wouldn't want someone in the elevator.
link |
00:24:05.840
If you're in an elevator, do you really want someone
link |
00:24:07.840
with a big lever, some random person operating
link |
00:24:09.840
in the elevator between floors?
link |
00:24:11.840
I wouldn't trust that.
link |
00:24:13.840
I would rather have the buttons.
link |
00:24:16.840
Okay, you're optimistic about the pace
link |
00:24:19.840
of improvement of the system.
link |
00:24:21.840
From what you've seen with the full self driving car,
link |
00:24:23.840
computer.
link |
00:24:25.840
The rate of improvement is exponential.
link |
00:24:27.840
So one of the other very interesting design choices
link |
00:24:30.840
early on that connects to this is the operational
link |
00:24:34.840
design domain of autopilot.
link |
00:24:37.840
So where autopilot is able to be turned on.
link |
00:24:41.840
So contrast another vehicle system that we're studying
link |
00:24:46.840
is the Cadillac SuperCrew system.
link |
00:24:48.840
That's in terms of ODD, very constrained to this particular
link |
00:24:51.840
kinds of highways, well mapped, tested,
link |
00:24:54.840
but it's much narrower than the ODD of Tesla vehicles.
link |
00:24:58.840
What's...
link |
00:25:00.840
It's like ADD.
link |
00:25:02.840
Yeah.
link |
00:25:04.840
That's good. That's a good line.
link |
00:25:07.840
What was the design decision
link |
00:25:10.840
in that different philosophy of thinking where...
link |
00:25:13.840
There's pros and cons.
link |
00:25:15.840
What we see with a wide ODD is Tesla drivers are able
link |
00:25:20.840
to explore more the limitations of the system,
link |
00:25:23.840
at least early on, and they understand together
link |
00:25:26.840
the instrument cluster display.
link |
00:25:28.840
They start to understand what are the capabilities.
link |
00:25:30.840
So that's a benefit.
link |
00:25:32.840
The con is you're letting drivers use it basically anywhere.
link |
00:25:37.840
Well, anyways, I could detect lanes with confidence.
link |
00:25:41.840
Was there a philosophy design decisions that were challenging
link |
00:25:46.840
that were being made there?
link |
00:25:48.840
Or from the very beginning, was that done on purpose
link |
00:25:53.840
with intent?
link |
00:25:55.840
Frankly, it's pretty crazy letting people drive
link |
00:25:58.840
a two ton death machine manually.
link |
00:26:02.840
That's crazy.
link |
00:26:04.840
In the future, people will be like,
link |
00:26:06.840
I can't believe anyone was just allowed to drive
link |
00:26:09.840
one of these two ton death machines
link |
00:26:12.840
and they just drive wherever they wanted,
link |
00:26:14.840
just like elevators.
link |
00:26:16.840
You just move the elevator with the lever wherever you want.
link |
00:26:18.840
It can stop at halfway between floors if you want.
link |
00:26:21.840
It's pretty crazy.
link |
00:26:24.840
So it's going to seem like a mad thing in the future
link |
00:26:29.840
that people were driving cars.
link |
00:26:32.840
So I have a bunch of questions about the human psychology,
link |
00:26:35.840
about behavior and so on.
link |
00:26:37.840
I don't know.
link |
00:26:39.840
Because you have faith in the AI system,
link |
00:26:45.840
not faith, but both on the hardware side
link |
00:26:50.840
and the deep learning approach of learning from data
link |
00:26:52.840
will make it just far safer than humans.
link |
00:26:55.840
Yeah, exactly.
link |
00:26:57.840
Recently, there are a few hackers who tricked autopilot
link |
00:27:00.840
to act in unexpected ways with adversarial examples.
link |
00:27:03.840
So we all know that neural network systems
link |
00:27:06.840
are very sensitive to minor disturbances
link |
00:27:08.840
to these adversarial examples on input.
link |
00:27:10.840
Do you think it's possible to defend against something like this
link |
00:27:13.840
for the industry?
link |
00:27:15.840
Sure.
link |
00:27:17.840
Can you elaborate on the confidence behind that answer?
link |
00:27:22.840
Well, a neural net is just like a basic bunch of matrix math.
link |
00:27:27.840
You have to be like a very sophisticated,
link |
00:27:30.840
somebody who really understands neural nets
link |
00:27:32.840
and basically reverse engineer how the matrix is being built
link |
00:27:37.840
and then create a little thing that just exactly causes
link |
00:27:42.840
the matrix math to be slightly off.
link |
00:27:44.840
But it's very easy to then block that by having
link |
00:27:48.840
basically anti negative recognition.
link |
00:27:51.840
It's like if the system sees something that looks like a matrix hack
link |
00:27:55.840
excluded, it's such an easy thing to do.
link |
00:28:01.840
So learn both on the valid data and the invalid data.
link |
00:28:05.840
So basically learn on the adversarial examples
link |
00:28:07.840
to be able to exclude them.
link |
00:28:09.840
Yeah, you basically want to both know what is a car
link |
00:28:12.840
and what is definitely not a car.
link |
00:28:15.840
You train for this is a car and this is definitely not a car.
link |
00:28:18.840
Those are two different things.
link |
00:28:20.840
People have no idea neural nets really.
link |
00:28:23.840
They probably think neural nets involves like, you know,
link |
00:28:25.840
fishing net or something.
link |
00:28:28.840
So as you know, taking a step beyond just Tesla and autopilot,
link |
00:28:35.840
current deep learning approaches still seem in some ways
link |
00:28:39.840
to be far from general intelligence systems.
link |
00:28:44.840
Do you think the current approaches will take us to general intelligence
link |
00:28:49.840
or do totally new ideas need to be invented?
link |
00:28:55.840
I think we're missing a few key ideas for general intelligence,
link |
00:28:59.840
general, artificial general intelligence.
link |
00:29:04.840
But it's going to be upon us very quickly
link |
00:29:08.840
and then we'll need to figure out what shall we do
link |
00:29:11.840
if we even have that choice.
link |
00:29:15.840
But it's amazing how people can't differentiate between, say,
link |
00:29:18.840
the narrow AI that, you know, allows a car to figure out
link |
00:29:22.840
what a lane line is and, you know,
link |
00:29:25.840
and navigate streets versus general intelligence.
link |
00:29:29.840
Like these are just very different things.
link |
00:29:32.840
Like your toaster and your computer are both machines,
link |
00:29:35.840
but one's much more sophisticated than another.
link |
00:29:38.840
You're confident with Tesla you can create the world's best toaster.
link |
00:29:43.840
The world's best toaster, yes.
link |
00:29:45.840
The world's best self driving.
link |
00:29:48.840
I'm, yes.
link |
00:29:51.840
To me, right now, this seems game set match.
link |
00:29:54.840
I don't, I mean, that's, I don't want to be complacent or overconfident,
link |
00:29:57.840
but that's what it appears.
link |
00:29:59.840
That is just literally what it, how it appears right now.
link |
00:30:02.840
It could be wrong, but it appears to be the case
link |
00:30:06.840
that Tesla is vastly ahead of everyone.
link |
00:30:10.840
Do you think we will ever create an AI system
link |
00:30:13.840
that we can love and loves us back in a deep meaningful way
link |
00:30:17.840
like in the movie, Her?
link |
00:30:20.840
I think AI will be capable of convincing you
link |
00:30:23.840
to fall in love with it very well.
link |
00:30:25.840
And that's different than us humans?
link |
00:30:28.840
You know, we start getting into a metaphysical question
link |
00:30:31.840
and do emotions and thoughts exist in a different realm than the physical.
link |
00:30:35.840
And maybe they do, maybe they don't.
link |
00:30:37.840
I don't know, but from a physics standpoint,
link |
00:30:39.840
I tend to think of things, you know,
link |
00:30:43.840
like physics was my main sort of training.
link |
00:30:47.840
And from a physics standpoint,
link |
00:30:50.840
essentially, if it loves you in a way
link |
00:30:52.840
that you can't tell whether it's real or not, it is real.
link |
00:30:57.840
That's a physics view of love.
link |
00:30:59.840
If you cannot prove that it does not,
link |
00:31:04.840
if there's no test that you can apply
link |
00:31:07.840
that would make it allow you to tell the difference,
link |
00:31:14.840
then there is no difference.
link |
00:31:16.840
And it's similar to seeing our world as simulation.
link |
00:31:20.840
There may not be a test to tell the difference
link |
00:31:22.840
between what the real world and the simulation.
link |
00:31:24.840
And therefore, from a physics perspective,
link |
00:31:26.840
it might as well be the same thing.
link |
00:31:28.840
Yes.
link |
00:31:29.840
There may be ways to test whether it's a simulation.
link |
00:31:32.840
There might be, I'm not saying there aren't,
link |
00:31:35.840
but you could certainly imagine that a simulation could correct
link |
00:31:38.840
that once an entity in the simulation
link |
00:31:40.840
found a way to detect the simulation,
link |
00:31:42.840
it could either restart, you know,
link |
00:31:44.840
pause the simulation, start a new simulation,
link |
00:31:47.840
or do one of many other things that then corrects for that error.
link |
00:31:52.840
So when maybe you or somebody else creates an AGI system
link |
00:31:58.840
and you get to ask her one question,
link |
00:32:02.840
what would that question be?
link |
00:32:16.840
What's outside the simulation?
link |
00:32:21.840
Milan, thank you so much for talking today.
link |
00:32:23.840
All right, thank you.