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