Chris Anderson: Elon Musk,
great to see you. How are you? Elon Musk: Good. How are you? CA: We're here at the Texas Gigafactory
the day before this thing opens. It's been pretty crazy out there. Thank you so much
for making time on a busy day. I would love you to help us,
kind of, cast our minds, I don't know, 10, 20,
30 years into the future. And help us try to picture
what it would take to build a future that's worth
getting excited about. The last time you spoke at TED, you said that that was really
just a big driver. You know, you can talk about lots of other
reasons to do the work you're doing, but fundamentally, you want
to think about the future and not think that it sucks. EM: Yeah, absolutely. I think in general, you know, there's a lot of discussion of like,
this problem or that problem. And a lot of people are sad
about the future and they're …
Pessimistic. And I think … this is … This is not great. I mean, we really want
to wake up in the morning and look forward to the future. We want to be excited
about what's going to happen. And life cannot simply be about sort of, solving one miserable
problem after another. CA: So if you look forward 30 years,
you know, the year 2050 has been labeled by scientists as this, kind of, almost like this
doomsday deadline on climate. There's a consensus of scientists,
a large consensus of scientists, who believe that if we haven't
completely eliminated greenhouse gases or offset them completely by 2050, effectively we're inviting
climate catastrophe.
Do you believe there is a pathway
to avoid that catastrophe? And what would it look like? EM: Yeah, so I am not one
of the doomsday people, which may surprise you. I actually think we're on a good path. But at the same time, I want to caution against complacency. So, so long as we are not complacent, as long as we have a high sense of urgency about moving towards
a sustainable energy economy, then I think things will be fine. So I can't emphasize that enough, as long as we push hard
and are not complacent, the future is going to be great. Don't worry about it. I mean, worry about it, but if you worry about it, ironically,
it will be a self-unfulfilling prophecy.
So, like, there are three elements
to a sustainable energy future. One is of sustainable energy generation,
which is primarily wind and solar. There's also hydro, geothermal, I'm actually pro-nuclear. I think nuclear is fine. But it's going to be primarily
solar and wind, as the primary generators of energy. The second part is you need batteries
to store the solar and wind energy because the sun
doesn't shine all the time, the wind doesn't blow all the time. So it's a lot of stationary battery packs.
And then you need electric transport. So electric cars, electric planes, boats. And then ultimately, it’s not really possible
to make electric rockets, but you can make
the propellant used in rockets using sustainable energy. So ultimately, we can have a fully
sustainable energy economy. And it's those three things: solar/wind, stationary
battery pack, electric vehicles. So then what are the limiting
factors on progress? The limiting factor really will be
battery cell production.
So that's going to really be
the fundamental rate driver. And then whatever the slowest element of the whole lithium-ion
battery cells supply chain, from mining and the many steps of refining to ultimately creating a battery cell and putting it into a pack, that will be the limiting factor
on progress towards sustainability. CA: All right, so we need to talk
more about batteries, because the key thing
that I want to understand, like, there seems to be
a scaling issue here that is kind of amazing and alarming. You have said that you have calculated that the amount of battery production
that the world needs for sustainability is 300 terawatt hours of batteries. That's the end goal? EM: Very rough numbers, and I certainly would invite others
to check our calculations because they may arrive
at different conclusions. But in order to transition, not just
current electricity production, but also heating and transport, which roughly triples the amount
of electricity that you need, it amounts to approximately 300 terawatt
hours of installed capacity. CA: So we need to give people
a sense of how big a task that is.
I mean, here we are at the Gigafactory. You know, this is one of the biggest
buildings in the world. What I've read, and tell me
if this is still right, is that the goal here is to eventually
produce 100 gigawatt hours of batteries here a year eventually. EM: We will probably do more than that, but yes, hopefully we get there
within a couple of years. CA: Right. But I mean, that is one — EM: 0.1 terrawat hours. CA: But that's still 1/100
of what's needed. How much of the rest of that 100
is Tesla planning to take on let's say, between now and 2030, 2040, when we really need to see
the scale up happen? EM: I mean, these are just guesses.
So please, people shouldn't
hold me to these things. It's not like this is like some — What tends to happen
is I'll make some like, you know, best guess and then people, in five years, there’ll be some jerk
that writes an article: "Elon said this would happen,
and it didn't happen. He's a liar and a fool." It's very annoying when that happens. So these are just guesses,
this is a conversation. CA: Right. EM: I think Tesla probably ends up
doing 10 percent of that. Roughly. CA: Let's say 2050 we have this amazing, you know,
100 percent sustainable electric grid made up of, you know, some mixture
of the sustainable energy sources you talked about. That same grid probably
is offering the world really low-cost energy, isn't it, compared with now.
And I'm curious about like, are people entitled to get
a little bit excited about the possibilities of that world? EM: People should be optimistic
about the future. Humanity will solve sustainable energy. It will happen if we, you know,
continue to push hard, the future is bright and good
from an energy standpoint. And then it will be possible to also use
that energy to do carbon sequestration. It takes a lot of energy to pull
carbon out of the atmosphere because in putting it in the atmosphere
it releases energy.
So now, you know, obviously
in order to pull it out, you need to use a lot of energy. But if you've got a lot of sustainable
energy from wind and solar, you can actually sequester carbon. So you can reverse the CO2 parts
per million of the atmosphere and oceans. And also you can really have
as much fresh water as you want. Earth is mostly water. We should call Earth “Water.” It's 70 percent water by surface area. Now most of that’s seawater, but it's like we just happen to be
on the bit that's land. CA: And with energy,
you can turn seawater into — EM: Yes. CA: Irrigating water
or whatever water you need.
EM: At very low cost. Things will be good. CA: Things will be good. And also, there's other benefits
to this non-fossil fuel world where the air is cleaner — EM: Yes, exactly. Because, like, when you burn fossil fuels, there's all these side reactions and toxic gases of various kinds. And sort of little particulates
that are bad for your lungs. Like, there's all sorts
of bad things that are happening that will go away. And the sky will be cleaner and quieter. The future's going to be good. CA: I want us to switch now to think
a bit about artificial intelligence. But the segue there, you mentioned how annoying it is
when people call you up for bad predictions in the past. So I'm possibly going to be annoying now, but I’m curious about your timelines
and how you predict and how come some things are so amazingly
on the money and some aren't.
So when it comes to predicting sales
of Tesla vehicles, for example, you've kind of been amazing, I think in 2014 when Tesla
had sold that year 60,000 cars, you said, "2020, I think we will do
half a million a year." EM: Yeah, we did
almost exactly a half million. CA: You did almost exactly half a million. You were scoffed in 2014
because no one since Henry Ford, with the Model T, had come close
to that kind of growth rate for cars. You were scoffed, and you actually
hit 500,000 cars and then 510,000 or whatever produced. But five years ago,
last time you came to TED, I asked you about full self-driving, and you said, “Yeah, this very year, I'm confident that we will have a car
going from LA to New York without any intervention." EM: Yeah, I don't want to blow your mind,
but I'm not always right. CA: (Laughs) What's the difference between those two? Why has full self-driving in particular
been so hard to predict? EM: I mean, the thing that really got me, and I think it's going to get
a lot of other people, is that there are just so many
false dawns with self-driving, where you think you've got the problem, have a handle on the problem, and then it, no, turns out
you just hit a ceiling.
Because if you were to plot the progress, the progress looks like a log curve. So it's like a series of log curves. So most people don't know
what a log curve is, I suppose. CA: Show the shape with your hands. EM: It goes up you know,
sort of a fairly straight way, and then it starts tailing off and you start getting diminishing returns. And you're like, uh oh, it was trending up and now
it's sort of, curving over and you start getting to these,
what I call local maxima, where you don't realize
basically how dumb you were. And then it happens again. And ultimately… These things, you know,
in retrospect, they seem obvious, but in order to solve
full self-driving properly, you actually have to solve real-world AI.
Because what are the road networks
designed to work with? They're designed to work
with a biological neural net, our brains, and with vision, our eyes. And so in order to make it
work with computers, you basically need to solve
real-world AI and vision. Because we need cameras and silicon neural nets in order to have self-driving work for a system that was designed
for eyes and biological neural nets. You know, I guess
when you put it that way, it's sort of, like, quite obvious that the only way
to solve full self-driving is to solve real world AI
and sophisticated vision. CA: What do you feel
about the current architecture? Do you think you have an architecture now where there is a chance for the logarithmic curve
not to tail off any anytime soon? EM: Well I mean, admittedly
these may be infamous last words, but I actually am confident
that we will solve it this year.
That we will exceed — The probability of an accident, at what point do you exceed
that of the average person? I think we will exceed that this year. CA: What are you seeing behind the scenes
that gives you that confidence? EM: We’re almost at the point
where we have a high-quality unified vector space. In the beginning, we were trying
to do this with image recognition on individual images. But if you get one image out of a video, it's actually quite hard to see
what's going on without ambiguity. But if you look at a video segment
of a few seconds of video, that ambiguity resolves. So the first thing we had to do
is tie all eight cameras together so they're synchronized, so that all the frames
are looked at simultaneously and labeled simultaneously by one person, because we still need human labeling.
So at least they’re not labeled
at different times by different people in different ways. So it's sort of a surround picture. Then a very important part
is to add the time dimension. So that you’re looking at surround video, and you're labeling surround video. And this is actually quite difficult to do
from a software standpoint. We had to write our own labeling tools and then create auto labeling, create auto labeling software to amplify
the efficiency of human labelers because it’s quite hard to label.
In the beginning,
it was taking several hours to label a 10-second video clip. This is not scalable. So basically what you have to have
is you have to have surround video, and that surround video has to be
primarily automatically labeled with humans just being editors and making slight corrections
to the labeling of the video and then feeding back those corrections
into the future auto labeler, so you get this flywheel eventually where the auto labeler
is able to take in vast amounts of video and with high accuracy, automatically label the video
for cars, lane lines, drive space. CA: What you’re saying is … the result of this is that you're
effectively giving the car a 3D model of the actual objects
that are all around it. It knows what they are, and it knows how fast they are moving. And the remaining task is to predict what the quirky behaviors are
that, you know, that when a pedestrian is walking
down the road with a smaller pedestrian, that maybe that smaller pedestrian
might do something unpredictable or things like that.
You have to build into it
before you can really call it safe. EM: You basically need to have
memory across time and space. So what I mean by that is … Memory can’t be infinite, because it's using up a lot
of the computer's RAM basically. So you have to say how much
are you going to try to remember? It's very common
for things to be occluded. So if you talk about say,
a pedestrian walking past a truck where you saw the pedestrian start
on one side of the truck, then they're occluded by the truck.
You would know intuitively, OK, that pedestrian is going to pop out
the other side, most likely. CA: A computer doesn't know it. EM: You need to slow down. CA: A skeptic is going to say
that every year for the last five years, you've kind of said, well, no this is the year, we're confident that it will be there
in a year or two or, you know, like it's always been about that far away.
But we've got a new architecture now, you're seeing enough improvement
behind the scenes to make you not certain,
but pretty confident, that, by the end of this year, what in most, not in every city,
and every circumstance but in many cities and circumstances, basically the car will be able
to drive without interventions safer than a human. EM: Yes. I mean, the car currently
drives me around Austin most of the time with no interventions. So it's not like … And we have over 100,000 people in our full self-driving beta program. So you can look at the videos
that they post online. CA: I do. And some of them are great,
and some of them are a little terrifying. I mean, occasionally
the car seems to veer off and scare the hell out of people. EM: It’s still a beta. CA: But you’re behind the scenes,
looking at the data, you're seeing enough improvement to believe that a this-year
timeline is real.
EM: Yes, that's what it seems like. I mean, we could be here
talking again in a year, like, well, another year went by,
and it didn’t happen. But I think this is the year. CA: And so in general,
when people talk about Elon time, I mean it sounds like
you can't just have a general rule that if you predict that something
will be done in six months, actually what we should imagine
is it’s going to be a year or it’s like two-x or three-x,
it depends on the type of prediction.
Some things, I guess,
things involving software, AI, whatever, are fundamentally harder
to predict than others. Is there an element that you actually deliberately make
aggressive prediction timelines to drive people to be ambitious? Without that, nothing gets done? EM: Well, I generally believe,
in terms of internal timelines, that we want to set the most aggressive
timeline that we can. Because there’s sort of like
a law of gaseous expansion where, for schedules, where
whatever time you set, it's not going to be less than that. It's very rare
that it'll be less than that. But as far as our predictions
are concerned, what tends to happen in the media is that they will report
all the wrong ones and ignore all the right ones. Or, you know, when writing
an article about me — I've had a long career
in multiple industries.
If you list my sins, I sound
like the worst person on Earth. But if you put those
against the things I've done right, it makes much more sense, you know? So essentially like,
the longer you do anything, the more mistakes
that you will make cumulatively. Which, if you sum up those mistakes, will sound like I'm the worst
predictor ever.
But for example, for Tesla vehicle growth, I said I think we’d do 50 percent,
and we’ve done 80 percent. CA: Yes. EM: But they don't mention that one. So, I mean, I'm not sure what my exact
track record is on predictions. They're more optimistic than pessimistic,
but they're not all optimistic. Some of them are exceeded
probably more or later, but they do come true. It's very rare that they do not come true. It's sort of like, you know, if there's some radical
technology prediction, the point is not
that it was a few years late, but that it happened at all. That's the more important part. CA: So it feels like
at some point in the last year, seeing the progress on understanding, the Tesla AI understanding
the world around it, led to a kind of, an aha moment at Tesla. Because you really surprised people
recently when you said probably the most important
product development going on at Tesla this year
is this robot, Optimus.
EM: Yes. CA: Many companies out there
have tried to put out these robots, they've been working on them for years. And so far no one has really cracked it. There's no mass adoption
robot in people's homes. There are some in manufacturing,
but I would say, no one's kind of, really cracked it. Is it something that happened in the development of full self-driving
that gave you the confidence to say, "You know what, we could do
something special here." EM: Yeah, exactly. So, you know, it took me a while
to sort of realize that in order to solve self-driving, you really needed to solve real-world AI.
And at the point of which you solve
real-world AI for a car, which is really a robot on four wheels, you can then generalize that
to a robot on legs as well. The two hard parts I think — like obviously companies
like Boston Dynamics have shown that it's possible
to make quite compelling, sometimes alarming robots. CA: Right. EM: You know, so from a sensors
and actuators standpoint, it's certainly been demonstrated by many that it's possible to make
a humanoid robot.
The things that are currently missing
are enough intelligence for the robot to navigate the real world
and do useful things without being explicitly instructed. So the missing things are basically
real-world intelligence and scaling up manufacturing. Those are two things
that Tesla is very good at. And so then we basically just need
to design the specialized actuators and sensors that are needed
for humanoid robot. People have no idea,
this is going to be bigger than the car. CA: So let's dig into exactly that. I mean, in one way, it's actually
an easier problem than full self-driving because instead of an object
going along at 60 miles an hour, which if it gets it wrong,
someone will die. This is an object that's engineered
to only go at what, three or four or five miles an hour. And so a mistake,
there aren't lives at stake. There might be embarrassment at stake. EM: So long as the AI doesn't take it over
and murder us in our sleep or something.
CA: Right. (Laughter) So talk about — I think the first applications
you've mentioned are probably going to be manufacturing, but eventually the vision is to have
these available for people at home. If you had a robot that really understood
the 3D architecture of your house and knew where every object
in that house was or was supposed to be, and could recognize all those objects, I mean, that’s kind of amazing, isn’t it? Like the kind of thing
that you could ask a robot to do would be what? Like, tidy up? EM: Yeah, absolutely. Make dinner, I guess, mow the lawn. CA: Take a cup of tea to grandma
and show her family pictures. EM: Exactly. Take care
of my grandmother and make sure — CA: It could obviously recognize
everyone in the home. It could play catch with your kids. EM: Yes. I mean, obviously,
we need to be careful this doesn't become a dystopian situation. I think one of the things
that's going to be important is to have a localized
ROM chip on the robot that cannot be updated over the air. Where if you, for example, were to say,
“Stop, stop, stop,” if anyone said that, then the robot would stop,
you know, type of thing.
And that's not updatable remotely. I think it's going to be important
to have safety features like that. CA: Yeah, that sounds wise. EM: And I do think there should be
a regulatory agency for AI. I've said that for many years. I don't love being regulated, but I think this is an important
thing for public safety. CA: Let's come back to that. But I don't think many people
have really sort of taken seriously the notion of, you know, a robot at home. I mean, at the start
of the computing revolution, Bill Gates said there's going to be
a computer in every home. And people at the time said, yeah,
whatever, who would even want that. Do you think there will be basically
like in, say, 2050 or whatever, like a robot in most homes,
is what there will be, and people will love them
and count on them? You’ll have your own butler basically. EM: Yeah, you'll have your sort of
buddy robot probably, yeah. CA: I mean, how much of a buddy? How many applications have you thought, you know, can you have
a romantic partner, a sex partner? EM: It's probably inevitable.
I mean, I did promise the internet
that I’d make catgirls. We could make a robot catgirl. CA: Be careful what
you promise the internet. (Laughter) EM: So, yeah, I guess it'll be
whatever people want really, you know. CA: What sort of timeline
should we be thinking about of the first models
that are actually made and sold? EM: Well, you know, the first units
that we intend to make are for jobs that are dangerous,
boring, repetitive, and things that people don't want to do. And, you know, I think we’ll have like
an interesting prototype sometime this year. We might have something useful next year, but I think quite likely
within at least two years. And then we'll see
rapid growth year over year of the usefulness
of the humanoid robots and decrease in cost
and scaling up production. CA: Initially just selling to businesses, or when do you picture
you'll start selling them where you can buy your parents one
for Christmas or something? EM: I'd say in less than ten years. CA: Help me on the economics of this.
So what do you picture the cost
of one of these being? EM: Well, I think the cost is actually
not going to be crazy high. Like less than a car. Initially, things will be expensive
because it'll be a new technology at low production volume. The complexity and cost of a car
is greater than that of a humanoid robot. So I would expect that it's going
to be less than a car, or at least equivalent to a cheap car. CA: So even if it starts at 50k,
within a few years, it’s down to 20k or lower or whatever. And maybe for home
they'll get much cheaper still. But think about the economics of this. If you can replace a $30,000, $40,000-a-year worker, which you have to pay every year, with a one-time payment of $25,000 for a robot that can work longer hours, a pretty rapid replacement
of certain types of jobs. How worried should
the world be about that? EM: I wouldn't worry about the sort of,
putting people out of a job thing.
I think we're actually going to have,
and already do have, a massive shortage of labor. So I think we will have … Not people out of work, but actually still a shortage
labor even in the future. But this really will be
a world of abundance. Any goods and services will be available
to anyone who wants them. It'll be so cheap to have goods
and services, it will be ridiculous. CA: I'm presuming it should be possible
to imagine a bunch of goods and services that can't profitably be made now
but could be made in that world, courtesy of legions of robots. EM: Yeah. It will be a world of abundance. The only scarcity
that will exist in the future is that which we decide to create
ourselves as humans. CA: OK. So AI is allowing us to imagine
a differently powered economy that will create this abundance.
What are you most worried
about going wrong? EM: Well, like I said,
AI and robotics will bring out what might be termed the age of abundance. Other people have used this word, and that this is my prediction: it will be an age of abundance
for everyone. But I guess there’s … The dangers would be
the artificial general intelligence or digital superintelligence decouples
from a collective human will and goes in the direction
that for some reason we don't like. Whatever direction it might go. You know, that’s sort of
the idea behind Neuralink, is to try to more tightly couple
collective human world to digital superintelligence. And also along the way solve a lot
of brain injuries and spinal injuries and that kind of thing. So even if it doesn't succeed
in the greater goal, I think it will succeed in the goal
of alleviating brain and spine damage. CA: So the spirit there is
that if we're going to make these AIs that are so vastly intelligent,
we ought to be wired directly to them so that we ourselves can have
those superpowers more directly. But that doesn't seem to avoid
the risk that those superpowers might …
Turn ugly in unintended ways. EM: I think it's a risk, I agree. I'm not saying that I have
some certain answer to that risk. I’m just saying like maybe one of the things
that would be good for ensuring that the future
is one that we want is to more tightly couple the collective human world
to digital intelligence. The issue that we face here
is that we are already a cyborg, if you think about it. The computers are
an extension of ourselves. And when we die, we have,
like, a digital ghost. You know, all of our text messages
and social media, emails. And it's quite eerie actually, when someone dies but everything
online is still there. But you say like, what's the limitation? What is it that inhibits
a human-machine symbiosis? It's the data rate. When you communicate,
especially with a phone, you're moving your thumbs very slowly. So you're like moving
your two little meat sticks at a rate that’s maybe 10 bits per second,
optimistically, 100 bits per second.
And computers are communicating
at the gigabyte level and beyond. CA: Have you seen evidence
that the technology is actually working, that you've got a richer, sort of,
higher bandwidth connection, if you like, between like external
electronics and a brain than has been possible before? EM: Yeah. I mean, the fundamental principles
of reading neurons, sort of doing read-write on neurons
with tiny electrodes, have been demonstrated for decades. So it's not like the concept is new. The problem is that there is
no product that works well that you can go and buy. So it's all sort of, in research labs. And it's like some cords
sticking out of your head. And it's quite gruesome,
and it's really … There's no good product
that actually does a good job and is high-bandwidth and safe and something actually that you could buy
and would want to buy.
But the way to think
of the Neuralink device is kind of like a Fitbit
or an Apple Watch. That's where we take out
sort of a small section of skull about the size of a quarter, replace that with what, in many ways really is very much like
a Fitbit, Apple Watch or some kind of smart watch thing. But with tiny, tiny wires, very, very tiny wires. Wires so tiny, it’s hard to even see them. And it's very important
to have very tiny wires so that when they’re implanted,
they don’t damage the brain. CA: How far are you from putting
these into humans? EM: Well, we have put in
our FDA application to aspirationally do the first
human implant this year.
CA: The first uses will be
for neurological injuries of different kinds. But rolling the clock forward and imagining when people
are actually using these for their own enhancement, let's say, and for the enhancement of the world, how clear are you in your mind as to what it will feel like
to have one of these inside your head? EM: Well, I do want to emphasize
we're at an early stage. And so it really will be
many years before we have anything approximating
a high-bandwidth neural interface that allows for AI-human symbiosis. For many years, we will just be solving
brain injuries and spinal injuries. For probably a decade. This is not something
that will suddenly one day it will have this incredible
sort of whole brain interface. It's going to be, like I said, at least a decade of really
just solving brain injuries and spinal injuries. And really, I think you can solve
a very wide range of brain injuries, including severe depression,
morbid obesity, sleep, potentially schizophrenia, like, a lot of things that cause
great stress to people.
