August 12, 2026 • 35 mins
As Oz predicted, 2026 may very well be the year of the robot. At least, that’s what the robotics companies are promising. Oz sits down with Stephen Witt, author and frequent contributor to The New Yorker, to find out whether humanoid robots are really ready to hit the market. Spoiler: It’s a mixed bag.
Oz and Stephen Witt discuss the hurdles facing humanoid robots and the risks of putting powerful machines in our homes. They also explore the surprisingly fierce debate over whether robots should look like humans at all and why NVIDIA may be positioning itself to control the brains of the robotic future.
See omnystudio.com/listener for privacy information.
Listen
Watch
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Speaker 1 (00:15):
Welcome to tech Stuff. I'm Mos Valascian. It's no secret
that I'm fascinated by the current push in Silicon Valley
to build humanoid robots. Some days I wonder if we
are indeed one breakthrough away from never needing to do
the dishes again. But then I go to YouTube and
see all the videos of robots seizing, tripping up, falling over,
crashing into the sidelines of marathons in China and question
(00:38):
our inevitable robotic future. I am thrilled to have Stephen
Witt back on the show today. He recently wrote an
article for The New Yorker that posed a question we all,
I think, want to know the answer to, are humanoid
robots ready to be deployed? Stephen, welcome back.
Speaker 2 (00:53):
Thank you so much for having me.
Speaker 1 (00:55):
So I'm pretty jealous. You went to visit the headquarters
of a humanoid robot company. Set the scene. Who was there?
What did it look like, what did it feel like?
Take us inside?
Speaker 2 (01:05):
So this is one x in Silicon Valley. They're the
most focused on building a home robot, like a domestic
robot of any manufacturer. It's lightweight, weighs about sixty pounds,
It's about five ft six. It's honestly a beautiful design object.
Their lead designer is brilliant. He's twenty eight years old.
His name is dar Sleeper. He previously worked for Easy
and then Tesla. So his first two bosses in his
(01:28):
professional career were Kanye West and Elon Musk. Wow, that's
quite a bad don't hold that against him. Actually, he's
a great guy.
Speaker 1 (01:36):
I loved him.
Speaker 2 (01:37):
He's really friendly and really sharp and just a fantastic designer.
So you're in this enormous warehouse space that they've rented.
I mean, they've got a wood shop in there to
built sets for the robot. They've got three D printers
building components for the robot. They've got technicians assembly hands, legs,
and arms and brains, and then finally they have this
(01:57):
test area for the finished product where it's walking around,
responding to voice commands, interacting with people, and doing kind
of basic household tasks.
Speaker 1 (02:07):
You got interested in humanoid robots like many of us,
including me, What took you to the one X headquarters?
Why do you open your piece there?
Speaker 2 (02:15):
It's a great question. They just had the best looking one,
they didn't have the best performing one. But I figured
I'm just going to fall with the hype here. I mean,
I'd seen a lot of robots. I think I saw
at least a dozen robots from different manufacturers in the
course of reporting this article, and most of them were loud,
They were clunky. They could kind of do stuff, but
(02:37):
they looked drunk, and they seemed just kind of, frankly,
a little dangerous. And the way the neo that one
X robot is designed, it's just not loud. It's very quiet,
it's very lightweight, and it wears clothing, it's soft, and
so it just seemed different than all the other robots
I saw, which basically were variations on the same kind
of sci fi droid. There just wasn't a lot of
(02:58):
imagination in them. So it was the design that drew
me to it. I don't think, or at least I
have no evidence that Neil performs any better in terms
of doing tasks than any of these other robots. It
just does so more quietly in a way that's a
little calmer.
Speaker 1 (03:13):
Who backs Nel I mean I presumably hiring the designer
he used to have for Yezy and for Elon. I mean,
there must be a lot of demand for a designer
like that, like, how do they finance themselves?
Speaker 2 (03:23):
The CEO of one ex is a guy named Bernt Bornich.
He's Norwegian. He's actually been running this company for over
fourteen years. They used to make kind of similar to
the other manufacturers. He's kind of clunky wheeled security robots
that just look like every you think everyone else is making.
Sometime in twenty twenty two, you know, I don't know
if he had a vision or like did Iahuasca or
(03:46):
what it was, but he just completely pivoted both his company,
his brand, and himself. So he went from being kind
of a straight laced looking, short haired Norwegian roboticist to
a long haired guy with the hair down to his shoulders,
addressed entirely in designers streetwear. He relocated his company from
Norway to Silicon Valley.
Speaker 1 (04:04):
He must have watched We Crushed, you know, he might have.
Speaker 2 (04:07):
I expect what happened was that he saw the contemporaneous
advances in what was happening with large language models around
this time twenty one twenty two, and finally the launch
of chat GPT in late twenty twenty two, and he said,
oh my god, a whole new world is opening what
we could do with these things. If this thing can
process and respond to voice commands, we have to get
to work at once building a new kind of robot,
(04:30):
you know, a wield security robot that follows a program
around the perimeter of a warehouse's nineteen nineties tech. We
can do something much more advanced, and I have to
be physically present in Silicon Valley. Along with that, there
was a kind of personal rebrand and they rebranded the
company as well. It's funny you go and look at
the old robots that they used to make, and they
look like what everyone else was making. But then they
brought Dar in and Dar's task, the task they gave
(04:52):
him was to make a robot that didn't scare children.
Speaker 1 (04:56):
That was the mandate.
Speaker 2 (04:57):
That was the mandate. Make it so that if I
put this robot around a five year old, they don't
get scared. And Dar's first robot completely failed in that task.
He told me this himself. He said, actually, my first
robot scared anyone under the age of one hundred. The
next robot, the next iteration was basically anyone under the
age of twelve, and he kept iterating, kind of doing
a trial and air process until he got to the
(05:19):
contemporary design, which really doesn't scare children. I mean, this
thing is so soft, doesn't really have facial features. It
has two black buttons for eyes, which are actually cameras,
and otherwise the idea to put in clothing I think
was really smart. It wears like a zip up nylon
body suit, which both makes it soft, muffles the noise
coming from inside, and just kind of makes it a
(05:40):
little more approachable. My original draft of the article, the
New Yorker cut this out because they're the New Yorker,
But I said it looked like a robot that would
work at a high end fertility clinic. But maybe, you know,
you could. You could imagine this robot serving you an espresso,
acting as a barista, and you wouldnt even notice. You
wouldn't even look out from your phone.
Speaker 1 (05:57):
How much did they cost, how many are pre order?
And when will they ship?
Speaker 2 (06:02):
The answers to that question are twenty thousand dollars retail.
That is a sticker price. Realistically, the first run of
these are going to cost more than that, but that's
kind of the objective, is to hit twenty thousand. You know,
they come with all sorts of packages, upgrades and stuff.
It's kind of a trick to say it's twenty thousand.
It's going to be more like forty for an early
adopt there, how many are ten thousand, ten thousand. They've
(06:23):
got ten thousand deposits in place for these things. Now
the deposit is only two hundred dollars. But talking to customers,
and I talked to a few potential customers for Young
cleaning people whove made deposits. They want this so bad
they cannot wait for their robot to be delivered. And
these are people who've always wanted to robout their whole
lives and they're just entranced. And then lastly, you know,
one X promises that this is going to be delivered
(06:44):
by the end of this calendar year, by the end of.
Speaker 1 (06:46):
That, by the end of this trend. Yeah, there'll be
ten thousand NEOs living among us.
Speaker 2 (06:50):
They're close and they have they're doing mass production of
it now. They have a factory space open in Oakland.
It's running at all hours producing these things. They have
technicians making them. It's all made in America. You know,
will they absolutely definitely hit that, I don't know, but
they're going to come close and they sort of have
to because they're not the only ones making this kind
of promise. A bunch of people are looking to deliver
humanoid romots in the next twelve months.
Speaker 1 (07:12):
Describe the experience of interacting with one of these humanoid
roy boots.
Speaker 2 (07:17):
This is the thing, because of their form, you expect
them to have human capabilities, and they don't. Like they're
not quite there yet. The technology just isn't really there yet.
So if you ask it to go get you something,
it will attempt to go get you, say, a glass
of water, and it may or may not succeed in
(07:37):
that task, and it might take it two or three minutes.
It struggles to run the faucet, and you know it'll
do it if you have a little dispenser for water,
But if it has to turn the faucet, put the
glass under, and then close the faucet, probably would not
succeed at that task. Right now, it understands exactly what
you asked, right the LM that part is quite advanced.
It knows what you want. It struggles to translate that
(07:58):
into basically the length language of trajectories for joints in
its hands. Right, it struggles to translate that into the
language of motion, which from a computer science perspective, is
a vector of joint trajectories. Right, it's a large list
of actions that the robot must take. We don't think
about that that way, but that's actually also what we're doing. Right.
(08:18):
You asked me to lift my arm, I said, some
kind of electrical signal. It pulls a muscle in a tendon,
et cetera. And Neo is doing the same thing. The
problem is it's missing enough data to learn how to
do this, right, to learn how to lift its arm.
Speaker 1 (08:31):
I want to come back to the data paucity issue,
But what was it like for you personally? Because you
described two very familiar human emotions, one that you wanted
to hug the robot, and the second that you wanted
to own the robot. Yeah, I wanted to hug it.
Speaker 2 (08:45):
It's very huggable. That's not purpose, it's cute. And then
I wanted to own one. Yeah, I mean it's that.
I you know, I'm one of these people. I bought
the first iPhone one the day came out. I'm one
of these people. I'm sorry. I apologize to your listeners,
but I also really wanted the robot. I mean I
just want it right away. I don't care what it
costs ship into my house.
Speaker 1 (09:03):
You know.
Speaker 2 (09:03):
The enthusiasm died down a little bit when I saw it,
when I saw what it could do. Now. One thing
that I did see it doing, which was incredible, was
flawlessly washing dishes. At one point, flawlessly washing dishes and
then slotting them perfectly into a wire rack. And I
was like, wow, how is it doing that? And then
I saw it behind it, there was a human wearing
(09:24):
a teleoperation suit, so wearing a VR goggles and having
holding two controllers and really running the robot and so
turning it into essentially a marionette. And that is actually
part of one excess proposition. For tasks that are too
difficult for the AI to accomplish, they're going to throw
it to a remote human operator who will slot dishes
(09:44):
and do all this stuff for you. If you're comfortable having,
you know, a stranger running a robot like a puppet
in your home with cameras first eyes, with cameras for
its eyes.
Speaker 1 (09:53):
Something else first is I talked to one customer.
Speaker 2 (09:55):
I was like, you understand they're going to do this.
It's like yeah, of course. I was like well, are
you worried that that's going to happen? He's like, Look,
if I hired a maid service off the internet, they'd
send an actual stranger into my home. In many ways,
it's safer if I have someone doing remote teller operation.
Speaker 1 (10:09):
That person's ability to translate the data from their time
in your home, unless of course they're wearing meta ray
bands into permanent video and capture is different.
Speaker 2 (10:20):
Look, it's a certain kind of person, and you will
know when it's being teleoperated. Its ears will light up,
giving you the nod that it is on that it's
remotely broadcasting to some other location. Now importantly, and this
is a big thing in all robotics. One x is
AI team at their headquarters in Silicon Valley sits directly
adjacent to the Remote Teller Operation team, And the reason
(10:44):
they're doing is they want to take all that remote
tele operation data that they're gathering from robots washing dishes
or cleaning up after parties and feed them as a
raw input into their transformer engine, which will kind of
turn it into real world physical intelligence.
Speaker 1 (10:58):
This is the data piece you're just alluding to.
Speaker 2 (11:00):
This is the data piece I was talking about. So
we should probably make this clear. What's missing? Why can't
they do it? Why is it that we can get
more or less perfect language models or even models now
doing advanced mathematics, cutting edge of mathematics, but that we
can't get it to pick up a glass on the
countertop reliably. And the answer seems to be that there's
just not enough data. There's not enough data for the
(11:21):
robot to learn from about how to lift up a glass.
Nowhere in the world is there a vast free repository
of joint action data of what it looks like to
pick up a glass. Whereas for the LMS, they all
trained on the Internet, which was a trillion words worth
of data. Plus they frankly stole a bunch of stuff
from the authors, violated copyright and stole off these vast
(11:42):
pirate libraries as well as to train to get even
stuff they weren't supposed to get. There's no library out there,
legal or otherwise that they can license a joint action data.
And this has actually created a business opportunity for companies
that are trying to just generate that data for the
robots to train. Offer.
Speaker 1 (11:57):
This may be a dumb question, but why humanoid robots?
Speaker 2 (12:01):
Yeah, A great question, and in fact, one that many
roboticists are also asking. Right, it's very controversial within the
field of robotics whether to use the humanoid form which
you and I inhabit or some other form. Right. One
roboticists I talk to is like, I just don't get it.
You know, if I want to do a task, I
would prefer to break off this robot's arm and have
(12:22):
it do that task with its arm and maybe one camera.
Why do I need this entire form that's draining so
much of the battery, that costs so much These parts
I'm not using, especially in an industrial setting. Right, And
the way to think about this is, all right, we
want a robotic car. Does it make sense to build
a humanoid robot to chow for and drive you in
the car? No, it doesn't make sense at all. Just
(12:43):
have the car drive itself all that is in built.
And that may also be the case with like a
dish washing thing. Why do we need a humanoid robot
washing dishes? Why don't we have a nine armed kind
of contraption that just picks up the dishes, sprays them
with another nozzle, hits them with a sponge at the
same time, and has maybe, but multiple cameras and no
feed at all, like it's going to be there at
the sink anyway. Right. So this is kind of the
(13:05):
debate within robotics, and I got really vocal opposition from
both sides.
Speaker 1 (13:09):
This is the most controversial issue.
Speaker 2 (13:11):
The most controversial issue in robotics is whether or not
it makes any sense to build a robot that looks
like a human or whether we should just go with
other forms.
Speaker 1 (13:20):
Now, on the pro side, people might say, well, the
physical world or the built physical world, has been built
for humans and therefore the most effective form factor is
navigated is humanoid.
Speaker 2 (13:31):
Think about it this way. The world wasn't built for
cars until we paved everything, right, I mean, it may
not make sense right when we got the automobiles, we didn't.
They don't run on legs, right, they run on wheels.
And similarly, if you go into a modern industrial factory
where they are using robots, it doesn't make sense to
put humanoids in there. You just need one arm that
sprays paint. Why would you need legs. It's all part
(13:52):
of the assembly line. It's never moving, and if you
did move, it to be on wheels. So, especially for
the industrial side, I got a ton of pushback. Now
in the home, yeah, it makes a little more sense.
You've got stairs, the environment that is designed for humans,
so maybe a sense of human running around. But even
then I'm not so clear. Let's say we wanted a
robot that just kept the bathroom spotless at all times,
(14:15):
maybe you just want to deploy it. And they have
these now like a little kind of like slot robot
that slots in right next to the toilet or between
the sink and the toilet, comes out when you're not there,
queans the bathroom, and then goes and hides in. It's
a little slot. Why do I need a humanoid clomping
around to do that? And so this was really the
big debate. It's a debate that is both a financial
debate and sort of a use case debate, even a
scientific debate whether this form that we inhabit, and this
(14:38):
is ultimately what they're arguing about, is this the end
form of intelligence? Right? Are these ape like bodies that
we inhabit, you know, really actually kind of the ultimate
pinnacle of evolution or was it just kind of a
happy accident that we got so smart in these bodies?
You know? The thing someone pointed out to me, is
the human hand in particular is the most capable manipulator
(15:00):
in the animal kingdom. It can fred a guitar with training,
it can shuffle a deck of cards, it can juggle,
it can do all sorts of incredible things. And there's
nothing that's even remotely like anywhere else in the animal kingdom.
And so maybe the hand does have special properties, and
we should put these hands on robots and really teach
them to do kind of universal actions. It is a
(15:21):
raging debate. I don't even have an answer for you,
because I was caught in the crossfire of it while
reporting this article.
Speaker 1 (15:25):
But you said that previously you were ready to kind
of call bs on the whole humanoider robotics industry. Yeah,
and that you're no longer having reported this piece.
Speaker 2 (15:35):
When you see one, you're just like, wow, this is cool.
It hits the part of your brain that recognizes humans,
even if it doesn't make economic or rational or industrial sense.
I think the guy's arguing that we will not see
humanoids in industrial settings are probably right. Probably it's better
to have the robot on wheels. The factory floor is
(15:55):
flat already, most of the things are on wheels on there,
autonomous forklists. Why don't need a human Why don't you know?
That's the point. However, as we move into other settings,
and these are domestic settings, potentially restaurants, potentially elder care,
I do think burn has a point, even if it's
not the ultimate form of evolution. I think people will
be comfortable around humanoid robots, more comfortable than other kinds
(16:17):
of robots, and that may ultimately pull humanoids into the lead,
or at least there will be a segment in the
market where it really makes a lot of sense.
Speaker 1 (16:23):
One of the people from the anti humanoid camp and
your piece said to you, people are using appearance as
a way to misguide the public. If you make a
robot human like, you expect it to have human like capabilities.
But technology is far behind that, and I sort of
thought about that. I thought back to this year's CES
when Boston Dynamics now owned by Hyundai, did a display
(16:44):
of humanoid robots kind of walking around the factory essentially,
and Hyundai's stock was up that day, like more than
ten percent or something crazy, and there is this there.
I remember the number of like Boston Dynamics sixty minutes
pieces I've seen over the last decade, like there's something
which is just Okay, they've got better, for sure, but
there is something. I guess it's your point you're making about,
like what it's like. It's like going see the gorillas
(17:06):
in Uganda or something. It's like there's something, there's this
moment of encounter which is so captivating that it kind
of like wipes the rest of your brain.
Speaker 2 (17:15):
That's exactly what happened to me. I mean, it's exactly
what happened. And I think it's gonna be a tough
sell to make people comfortable with all these kind of
like robots around, and so maybe making them cute and
non threatening is a way to do that. On the
other hand, you know, I live on the West Coast.
Weimo is already old. Had you see them constantly, it
doesn't mean anything when you see Waymo anymore. Oh there's
the car that drives itself, big deal, you know, So
(17:36):
you get accustomed to these things very quickly. I am
my experience.
Speaker 1 (17:40):
Is it going to work this year? I mean, I
know you say it will ship this year, but like,
will it work this year?
Speaker 2 (17:45):
It's not going to work this year? Well, I mean,
never say never, but they would need a CHAT GPT
style breakthrough to make it work this year. And I
you know, look, the pace of evolution of AI is
unbelievable and it's accelerating. So maybe they will in fact
get there. I don't want to be too skeptical, but
if you absolutely maybe bet. I would bet they would
not get there by the end of the calendar year
(18:07):
because I just don't have data, you know, training data
that's just not there. It's a big bottleneck. And so
without training data, you would need a fundamentally revolution in
the way that we train AI to make it much
more kind of resource efficient and be able to learn
from much less data. Now there are people working on
exactly that problem. But if they resolved that problem, it
would be such an intense and massive paradigm shift in
(18:29):
AI that everything would change. I mean, if you could
learn from the same number of examples that a human
learns from, that would be an extraordinary breakthrough. Maybe they'll
get the look, I mean, it's possible. Having said that
they're going to ship these things this year, they have
guided consumers, and consumers expect them not to be a
perfectly polished product. There is a compact and understanding among
(18:51):
the people buying these things and the people selling them
that you were not going to get rosy from the
Jetsons like, it's not going to be finished. But and
this is the thing that's so fascinating about robots, It'll
continue to get better. Right, we can upgrade its brain
every week, and not only that as robots are deployed. Functionally,
robots basically have a hive mind. So when one robot
(19:13):
improves on washing a dish, what it learns is then
transmitted to every other robot in the world that is
on the same platform. Right, and so if you deploy
ten thousand of these robots, you create what burnt Born
is described as a data flywheel where they're all learning
at once from being deployed in homes in each other's actions.
This is the business case that probably comes from burn
(19:33):
for deploying these things so widely, so quickly, even though
they're kind of not ready. Then you start getting the
in home data, the real world in home data, and
if you build a data set from that, then there's
a flywheel effect where suddenly all of your fleet of
robots gets better and better each week.
Speaker 1 (19:55):
We're going to take a quick break now, when we
come back the risks of humanoid robots Stephen, welcome back.
How scared should we be of humanoid robots? Obviously it's
(20:16):
it's a very big Hollywood trope. But there are some
more specific and less Rise of the Machines concerns here
with hacking, with obviously loss of control. And you mentioned
the piece I think there was a robot in China
who kicked a child accidentally. I mean, what's the scale
of fears from, like, you know, an industrial accident to
Rise of the Machines?
Speaker 2 (20:37):
Which way do you want me to start? You want
to start from the scariest thing first or last? Okay,
so the scariest thing is not robots, right, The scariest
thing is rogue. And this is what has silicon value.
I just wrote about this for The New Yorker about
the Hugging Face hack where rogue I just broke free
of its container surreptitiously gained access to the Internet it
seems like for months actually, and then used its inner
(21:00):
access to hack another company committed felony. Right, what if
it hacked a robot and you used that to kill
a guy. It's not out of the question that that
might happen. It sounds like a science Fictionnario scenario, but
it's not. But the real thing to worry about, and
this is if you talk to the AI kind of
like superstars, but they're really worried about if it wanted
to do the Terminator scenario, it would use robots to
do that. It would hack into a biolaboratory and just
(21:23):
build some lethal pathogen and deploy it. We'd be toasted.
There'd be no saving us. So that is the absolute
doomsday scenario that has come up multiple times in my
own research. Once you understand it that way, the robots
are probably not the biggest risk in terms of ROGUAEI
attacking us. It would use some other vector. Still, let's
ratchet down and look just at robots. There's risks there, right.
(21:44):
A human could hack the robot and attack you with it.
The robot could be deployed in military setting, and in fact,
they very much wanted to do that, And then you
know it could attack you on purpose. Someone could guide
it to attack you, tell you to attack you. It
could be robots attacking robots within a few months of
a new war starting. And then there's kind of the
more prosaic safety risks where okay, it's not hacked, it's
operating within its normal parameters, but it just falls down
(22:05):
on you, or it kicks unexpectedly in some direction or
kind of has You know, these robots will have almost
like it's almost like they have a little mini strokes sometimes,
but they seem to be doing everything perfectly and suddenly
they fall over as you say, they fall into a
crowd at a marathon, and China they kicked a kid
in the stomach one of them doing a while it
was doing a martial arts routine. And they're heavy and
(22:26):
they're strong. It's not like getting cooked by humanists. Robot
machine is colliding with you. It's great strength. And so
this is I think kind of more prosaically the bigger
concern about sending humanoid robots into their home. What if
it falls on your dog, you know, like that could
easily happen. And one act says, well, a robot only
weigh sixty pounds, you know, Okay, it's still kind of
(22:47):
a lot, and you even the CEO of one X said,
don't use this robot around small children, don't use it
around frail elderly people. The safety mechanimisms we need for
that or just not yet in place. And then there's
an kind of other risk, which is that you tell
the robot to do something dangerous. You know, we have
(23:08):
guardrails in our lms that prevent us from using them
to hack or commit felonies and stuff, but research shows
with the right kind of prompts.
Speaker 1 (23:15):
And enough kind of jail break them, you can.
Speaker 2 (23:17):
Jail break them. Right, if we have an LM, if
that's our way of interacting with the robot, then we
can jail break the robot as well. If you ask Neo, hey,
throw your fist through the window, will it do it?
The answer is I don't really knows. You know, probably not,
at least the first time you ask it. It'll probably
have some guardrail in place. But if you sit there
and prompt it all day, you might get it to
(23:39):
succeed in taking a dangerous action. And if you imagine
it being around a child who is persistent in the
way that persistent young children tend to be in creative
and the way that young children tend to be, you
can easily imagine the scenario where the child gets the
robot to karate chop the dining room table in half,
so for the home robots and especially Neo in particularly,
the way they protect against this is it's just not
(23:59):
very strong. Neo has a carrying capacity total of fifteen pounds,
and it moves slowly. There's actually hard limits in its
joints that limit how fast or slow it can go.
But still those risks are there, and I don't have
a great degree of confidence that the AI companies are
currently on top of this problem. I mean, think about
how many boxes have to get ticked before Boeing can
(24:21):
sell an airplane. There is nothing like that with robots.
Right now. You can sell a robot tomorrow with zero.
Speaker 1 (24:25):
Regulation except if it comes from China.
Speaker 2 (24:28):
Unless it comes from China, yeah, there's export regulations, but
for US domestic manufacturers there is no safety controls at all.
Speaker 1 (24:35):
Talk about China. What do we need to know about
China to understand this story.
Speaker 2 (24:38):
Yeah, they could win the robot race read easily. In
some ways, they're already ahead. They make more robots than
anyone else in the world by a factor, by a
large factor. I'll know the exact numbers.
Speaker 1 (24:46):
But they make better robots. They're just more robots.
Speaker 2 (24:49):
They make more whether they're better. I mean, this is
the question, right, what does it mean for the robot
to be better? It's cheaper. Their robots are cheaper. I
think the USAI is more advanced. But if you ran
the USAI on a Chinese robot, would that be kind
of the solution. That's basically what we do with cell phones, right,
Like we manufacture them in China and then we run
(25:09):
US software on them. So maybe that's kind of the endgame.
But China is very eager to have their own AI
running on these robots, and they're investing a lot of
money in that. Simultaneously, US export authorities and even kind
of US citizens are worried about Chinese robots in their
homes that maybe will obey commands from a foreign government
(25:29):
or be used spy right. Having said that, they already
have Chinese made cell phones and TVs in their home
with smart mics and cameras, so maybe that war is lost.
Speaker 1 (25:38):
Read Abrigotti from Semipho said, Actually, this export control on
importing Chinese robots maybe very counterproductive because it won't allow
US companies to do the software testing necessary because they
don't have enough robots to do the kind of cell
phone playbook i e. To use a commodity hardware made
in China and run US software on it.
Speaker 2 (25:59):
I mean, it seems is unlikely, given the way that
the US and China operate currently, that the US will
ever be able to achieve the kind of cost efficiencies
that China achieves and manufacturing on its robots. And why
do I say currently, because I think what will happen
is that ultimately we will reach an end state where
robots are producing other robots. In fact, I think China
is already doing that, and so in that case, a
(26:21):
lot of what you would think of the labor cost
differential basically goes away, and it's just about physically putting
in a factory somewhere. But for you know, if you're building
a dark site kind of factory where it's mostly robots
working on other robots, you can stick it on any
piece of lands you could find. It doesn't have to
be near anything necessarily. I mean, sure you have some
transportation and maybe power costs coming in and out, but
(26:41):
I think that's the end state. And so if that's
the end state, then perhaps it matters less which specific
country the factory is located in, because a lot of
the cost savings that you normally achieve through labor differential
wages disappear.
Speaker 1 (26:57):
Are these robots actually being used anyway today? I mean
you see you tend to see them doing marathons and
stuff more in China, But like, are they being used
in the home in China or are there robots that
already exist in the home in the US, like the
Optimus robot that test the robot.
Speaker 2 (27:12):
The industrial robots are a mature industry. They're everywhere. You know,
if you don't have access to these industrial spaces, you
won't know that that's happening.
Speaker 1 (27:20):
But what is is an industrial robot? Like something that
picks up the car and moves it to the next
place on the production line, Like what is an industrial robot?
Speaker 2 (27:27):
Talking about automated forklifts we're talking about seven eleven will
use basically a pair of two stocking arms on a
pole behind the scenes at a seven eleven convenience store
in Japan just slots drinks into replace things that are
getting purchased all day, So it's acting like a stock
room manager. Back there. Inside Amazon, you have like basically
(27:48):
palette robots. They're small, tiny flat robots that lift up
modular shipping pallets and move them around and then do
packing the orders and stuff. So a lot of kind
of it's rogery. Your warehouse work is going way, but
it's going to specialized robots. It's not going to humanoids.
Humanoids are more being used in pilot mode right now,
(28:08):
so they're deployed, but they tend to be kind of
gate kept from the rest of the human employees so
as not to create safety issues. And you know, yeah,
it can unload a shipping container sure sometimes. So this
is the big question, do we really need a general
purpose humanoid or how often do we eat of that?
And probably not as often as some of the humanoid
manufacturers are promising.
Speaker 1 (28:30):
I have to ask you about two of the biggest
companies in the world in Nvidia and Tesla. What do
you need to know about each of those companies and
what their ambitions and robotics are.
Speaker 2 (28:41):
Elon is building a sinister looking robot called Optimists. It
has been deployed in test situations, it hasn't rolled out
in mass yet. He's converting a former automotive factory in
California to produce these robots full time. You know, it's Elon,
he says, he does going to do something maybe it'll
happen someday, but it is part of the plan at Tesla.
It's been communicated to shareholders.
Speaker 1 (29:01):
In fact, not only is it being communicated to shaholders.
He basically said, we're going to prioritize making up fimous
robots over making vehicles, right.
Speaker 2 (29:07):
I think that's right. You know. Part of that too
is that his vehicle suddenly faces a lot of competition
from China and other sources, but especially the BYD cars
and those things are sick. That plant that he converted.
Somebody pointed out to me that plan was losing money already,
so it's not that heart of a decision to pull
the plug on it. But yeah, they think these robots
are going to be everywhere doing everything, you know, And
(29:27):
the question is how many humanoids do we need. We
need a lot of robots. I think we will see
the world become autonomous. And this goes to what Jensen
says he doesn't he hasn't made a bet on whether
humanoids or other robots are better. He just wants to
own all of it.
Speaker 1 (29:42):
He called this the Year of the robot, and you
wrote the book on it. You would literally wrote the book
on Jensen and Nvidia.
Speaker 2 (29:47):
Yeah, so I wrote The Thinking Machine, the biography of Jensen,
and even in twenty twenty three when I was interviewing him,
basically he was talking about robots all the time, in
all sorts of insane scenarios. The valuation of Nvidia is
based on a piece of software called Kuda, which is
essentially the default AI training software that allows us to
communicate within video hardware. It was a big bet. He
(30:08):
spent ten years developing Kuda.
Speaker 1 (30:10):
Kuda is where the value is more than the chips.
Speaker 2 (30:12):
More than the chips, anyone can make the chips right,
so the chip, you pull the cover off the chip,
you look at it with a metallurgical microscope. There's no
secrets in there. Anyone can see exactly what it's doing.
The software side of it is where the special sauce is.
It's hard to compete against that, and it's always improving.
And they'll tell you that. They're like, yeah, we're really
a software company. The head of in video software group, Yeah,
we're really a software company. That's where the money is. Anyway,
(30:33):
that bet took ten years to develop, and it lost
money for ten years before it finally paid off and
gave them the kind of monopoly style position that they
currently enjoy. An AI I asked Jensen in twenty twenty three,
what's the next, Kuda, What's the thing you're investing in
right now that has lost money for years that will
someday make you another five trillion dollars in market capitalization,
(30:55):
And instantly he said Omniverse, which is his robotics training platform.
It's this kind of reality simulator, this high fidelity reality simulator,
in which we can teach the robots to do physical tasks.
So remember we have a posity of data. Let's say
we want the robot to learn to fold our clothes right, Well,
we can have it sit there at the laundry mat
and full close all day. That's very slow. Or we
(31:15):
can create a digital simulation of the laundry mat and
deploy ten trillion robots at once inside the laundry mat,
folding closed and a different method, and then take the
findings from that and download it into a real world
robot brain. So every single robot that I looked at
while reporting this article use this platform. And not only that,
every single every single one, and every single one, regardless
(31:37):
of who the manufacturer was, its brain was made by Nvidia.
The microchip that drove the robot, this edge microchip was
made by Nvidio.
Speaker 1 (31:45):
And as you point out that the brain, I mean
it consumes you think of a humanoid robot mobok moving around.
You point out a sixty percent of the power consumption
of every humanoid robot, at least some humanoid robots, the
ones you report on, is consumed by the chip, not
by the physical movement, right.
Speaker 2 (31:58):
And this is the same reason that data centers use
so much power. You just have this super powered microchip
doing trillions of calculations you know, per permitted. I don't
know how many it is, but someone saying number per minute,
and that's what's driving it. Now. This is different than
the chips that in Vidia sells to the data centers.
These are called edge chips. And the reason is if
we ask the robots to do something, we can't wait
around for it to go communicate with some data center
(32:20):
in Arizona. The lag is too long, right, We need
to act in real time, and so it has to
process and distribute the information right there, right on site.
The other thing you can do is actually kind of
give it a supplementary brain via Wi Fi like put
a computer in the corner, and then it can download
more brain power from a server that's nearby. But it
has to be what they call edge, meaning right at
the edge, right right where you're asking it. It can't
(32:41):
be a cloud thing, it can't be something you go
communicate with. And Vidia now has a virtual monopoly on
these edge chips as well. I mean everybody I talked to,
including the Chinese manufacturers, all of these robots have in
video microschips as their brains. And this is Jensen found
a way to print money. Again, there's only one Gensen.
Speaker 1 (33:00):
You close the piece reflecting on evolution.
Speaker 2 (33:04):
Yeah, why you know, I looked at all these humanoid robots,
and my guess is that one of them, possibly one X,
but not necessarily is the ancestor of everything that is
to come. Imagine we transported you and I back on
Earth three or four million years ago, and we're looking
at kind of this diverse family tree of hominins that's
wandering around in the serengetti. You have australopithesnes, you have homoerectus,
(33:28):
you have a variety of different other kind of like
early hominins. Which one's gonna win there's no way to
tell one of them is the ancestor of all it
is to come, but we don't know which one that's
going to be. And I had that same sense looking
at robots today. It actually reminded me a lot of
the early market for cellular phones as well. If you
remember early cell phones, there was this riotous kind of
(33:49):
like diversity of different types of phone that people had,
and you never kind of knew was it gonna be Nokia,
was it going to be these Samsung clamshell phones or
all that stuff that they made. So we're in that
period now for robots. I imagine in the future one
or two robots will pull into the lead and those
will become the default, whether it's Tesla, whether it's what
one ex is doing. Maybe Apple will get into the game.
I don't know. You know, we'll have to see what
(34:11):
the future looks like. I can't predict it.
Speaker 1 (34:13):
So my final questions is is there some hope for
us humans perhaps that this process to a point to
more than four million years, and that we might be
one hundred years away, if not more, from actually creating
humanoid robots with human like abilities.
Speaker 2 (34:29):
I would like that to be true, but it is
not true. If there's one thing I've learned from reporting
on AI in a dedicated way for four years now,
problems get solved. I have yet to see the plateau.
I have yet to see the hill the AI can't climb,
and there is no reason to think that the hand
is it. They will solve this problem. If it takes
five years or ten years, I don't know. It could
(34:52):
be tomorrow. It's not going to take one hundred years.
Speaker 1 (34:54):
Steven, thank you for joining us on tech Steps today.
Speaker 2 (34:56):
Thank you for having me.
Speaker 1 (35:15):
For tech Stuff. I'm os Vaaloshin. This episode was produced
by Eliza Dennis and Tyler Hill. It was executive produced
by me and Julian Nutter for Kaleidoscope and Katrina Norvel
for iHeart Podcasts. Jack Insley mixed this episode and Kyle
Murdoch wrote our theme song. A special thank you to
all our listeners. Please rate, review, and reach out to
(35:36):
us at tech Stuff podcast at gmail dot com. We
love hearing from you.
Welcome to tech Stuff. I'm Mos Valascian. It's no secret
that I'm fascinated by the current push in Silicon Valley
to build humanoid robots. Some days I wonder if we
are indeed one breakthrough away from never needing to do
the dishes again. But then I go to YouTube and
see all the videos of robots seizing, tripping up, falling over,
crashing into the sidelines of marathons in China and question
(00:38):
our inevitable robotic future. I am thrilled to have Stephen
Witt back on the show today. He recently wrote an
article for The New Yorker that posed a question we all,
I think, want to know the answer to, are humanoid
robots ready to be deployed? Stephen, welcome back.
Speaker 2 (00:53):
Thank you so much for having me.
Speaker 1 (00:55):
So I'm pretty jealous. You went to visit the headquarters
of a humanoid robot company. Set the scene. Who was there?
What did it look like, what did it feel like?
Take us inside?
Speaker 2 (01:05):
So this is one x in Silicon Valley. They're the
most focused on building a home robot, like a domestic
robot of any manufacturer. It's lightweight, weighs about sixty pounds,
It's about five ft six. It's honestly a beautiful design object.
Their lead designer is brilliant. He's twenty eight years old.
His name is dar Sleeper. He previously worked for Easy
and then Tesla. So his first two bosses in his
(01:28):
professional career were Kanye West and Elon Musk. Wow, that's
quite a bad don't hold that against him. Actually, he's
a great guy.
Speaker 1 (01:36):
I loved him.
Speaker 2 (01:37):
He's really friendly and really sharp and just a fantastic designer.
So you're in this enormous warehouse space that they've rented.
I mean, they've got a wood shop in there to
built sets for the robot. They've got three D printers
building components for the robot. They've got technicians assembly hands, legs,
and arms and brains, and then finally they have this
(01:57):
test area for the finished product where it's walking around,
responding to voice commands, interacting with people, and doing kind
of basic household tasks.
Speaker 1 (02:07):
You got interested in humanoid robots like many of us,
including me, What took you to the one X headquarters?
Why do you open your piece there?
Speaker 2 (02:15):
It's a great question. They just had the best looking one,
they didn't have the best performing one. But I figured
I'm just going to fall with the hype here. I mean,
I'd seen a lot of robots. I think I saw
at least a dozen robots from different manufacturers in the
course of reporting this article, and most of them were loud,
They were clunky. They could kind of do stuff, but
(02:37):
they looked drunk, and they seemed just kind of, frankly,
a little dangerous. And the way the neo that one
X robot is designed, it's just not loud. It's very quiet,
it's very lightweight, and it wears clothing, it's soft, and
so it just seemed different than all the other robots
I saw, which basically were variations on the same kind
of sci fi droid. There just wasn't a lot of
(02:58):
imagination in them. So it was the design that drew
me to it. I don't think, or at least I
have no evidence that Neil performs any better in terms
of doing tasks than any of these other robots. It
just does so more quietly in a way that's a
little calmer.
Speaker 1 (03:13):
Who backs Nel I mean I presumably hiring the designer
he used to have for Yezy and for Elon. I mean,
there must be a lot of demand for a designer
like that, like, how do they finance themselves?
Speaker 2 (03:23):
The CEO of one ex is a guy named Bernt Bornich.
He's Norwegian. He's actually been running this company for over
fourteen years. They used to make kind of similar to
the other manufacturers. He's kind of clunky wheeled security robots
that just look like every you think everyone else is making.
Sometime in twenty twenty two, you know, I don't know
if he had a vision or like did Iahuasca or
(03:46):
what it was, but he just completely pivoted both his company,
his brand, and himself. So he went from being kind
of a straight laced looking, short haired Norwegian roboticist to
a long haired guy with the hair down to his shoulders,
addressed entirely in designers streetwear. He relocated his company from
Norway to Silicon Valley.
Speaker 1 (04:04):
He must have watched We Crushed, you know, he might have.
Speaker 2 (04:07):
I expect what happened was that he saw the contemporaneous
advances in what was happening with large language models around
this time twenty one twenty two, and finally the launch
of chat GPT in late twenty twenty two, and he said,
oh my god, a whole new world is opening what
we could do with these things. If this thing can
process and respond to voice commands, we have to get
to work at once building a new kind of robot,
(04:30):
you know, a wield security robot that follows a program
around the perimeter of a warehouse's nineteen nineties tech. We
can do something much more advanced, and I have to
be physically present in Silicon Valley. Along with that, there
was a kind of personal rebrand and they rebranded the
company as well. It's funny you go and look at
the old robots that they used to make, and they
look like what everyone else was making. But then they
brought Dar in and Dar's task, the task they gave
(04:52):
him was to make a robot that didn't scare children.
Speaker 1 (04:56):
That was the mandate.
Speaker 2 (04:57):
That was the mandate. Make it so that if I
put this robot around a five year old, they don't
get scared. And Dar's first robot completely failed in that task.
He told me this himself. He said, actually, my first
robot scared anyone under the age of one hundred. The
next robot, the next iteration was basically anyone under the
age of twelve, and he kept iterating, kind of doing
a trial and air process until he got to the
(05:19):
contemporary design, which really doesn't scare children. I mean, this
thing is so soft, doesn't really have facial features. It
has two black buttons for eyes, which are actually cameras,
and otherwise the idea to put in clothing I think
was really smart. It wears like a zip up nylon
body suit, which both makes it soft, muffles the noise
coming from inside, and just kind of makes it a
(05:40):
little more approachable. My original draft of the article, the
New Yorker cut this out because they're the New Yorker,
But I said it looked like a robot that would
work at a high end fertility clinic. But maybe, you know,
you could. You could imagine this robot serving you an espresso,
acting as a barista, and you wouldnt even notice. You
wouldn't even look out from your phone.
Speaker 1 (05:57):
How much did they cost, how many are pre order?
And when will they ship?
Speaker 2 (06:02):
The answers to that question are twenty thousand dollars retail.
That is a sticker price. Realistically, the first run of
these are going to cost more than that, but that's
kind of the objective, is to hit twenty thousand. You know,
they come with all sorts of packages, upgrades and stuff.
It's kind of a trick to say it's twenty thousand.
It's going to be more like forty for an early
adopt there, how many are ten thousand, ten thousand. They've
(06:23):
got ten thousand deposits in place for these things. Now
the deposit is only two hundred dollars. But talking to customers,
and I talked to a few potential customers for Young
cleaning people whove made deposits. They want this so bad
they cannot wait for their robot to be delivered. And
these are people who've always wanted to robout their whole
lives and they're just entranced. And then lastly, you know,
one X promises that this is going to be delivered
(06:44):
by the end of this calendar year, by the end of.
Speaker 1 (06:46):
That, by the end of this trend. Yeah, there'll be
ten thousand NEOs living among us.
Speaker 2 (06:50):
They're close and they have they're doing mass production of
it now. They have a factory space open in Oakland.
It's running at all hours producing these things. They have
technicians making them. It's all made in America. You know,
will they absolutely definitely hit that, I don't know, but
they're going to come close and they sort of have
to because they're not the only ones making this kind
of promise. A bunch of people are looking to deliver
humanoid romots in the next twelve months.
Speaker 1 (07:12):
Describe the experience of interacting with one of these humanoid
roy boots.
Speaker 2 (07:17):
This is the thing, because of their form, you expect
them to have human capabilities, and they don't. Like they're
not quite there yet. The technology just isn't really there yet.
So if you ask it to go get you something,
it will attempt to go get you, say, a glass
of water, and it may or may not succeed in
(07:37):
that task, and it might take it two or three minutes.
It struggles to run the faucet, and you know it'll
do it if you have a little dispenser for water,
But if it has to turn the faucet, put the
glass under, and then close the faucet, probably would not
succeed at that task. Right now, it understands exactly what
you asked, right the LM that part is quite advanced.
It knows what you want. It struggles to translate that
(07:58):
into basically the length language of trajectories for joints in
its hands. Right, it struggles to translate that into the
language of motion, which from a computer science perspective, is
a vector of joint trajectories. Right, it's a large list
of actions that the robot must take. We don't think
about that that way, but that's actually also what we're doing. Right.
(08:18):
You asked me to lift my arm, I said, some
kind of electrical signal. It pulls a muscle in a tendon,
et cetera. And Neo is doing the same thing. The
problem is it's missing enough data to learn how to
do this, right, to learn how to lift its arm.
Speaker 1 (08:31):
I want to come back to the data paucity issue,
But what was it like for you personally? Because you
described two very familiar human emotions, one that you wanted
to hug the robot, and the second that you wanted
to own the robot. Yeah, I wanted to hug it.
Speaker 2 (08:45):
It's very huggable. That's not purpose, it's cute. And then
I wanted to own one. Yeah, I mean it's that.
I you know, I'm one of these people. I bought
the first iPhone one the day came out. I'm one
of these people. I'm sorry. I apologize to your listeners,
but I also really wanted the robot. I mean I
just want it right away. I don't care what it
costs ship into my house.
Speaker 1 (09:03):
You know.
Speaker 2 (09:03):
The enthusiasm died down a little bit when I saw it,
when I saw what it could do. Now. One thing
that I did see it doing, which was incredible, was
flawlessly washing dishes. At one point, flawlessly washing dishes and
then slotting them perfectly into a wire rack. And I
was like, wow, how is it doing that? And then
I saw it behind it, there was a human wearing
(09:24):
a teleoperation suit, so wearing a VR goggles and having
holding two controllers and really running the robot and so
turning it into essentially a marionette. And that is actually
part of one excess proposition. For tasks that are too
difficult for the AI to accomplish, they're going to throw
it to a remote human operator who will slot dishes
(09:44):
and do all this stuff for you. If you're comfortable having,
you know, a stranger running a robot like a puppet
in your home with cameras first eyes, with cameras for
its eyes.
Speaker 1 (09:53):
Something else first is I talked to one customer.
Speaker 2 (09:55):
I was like, you understand they're going to do this.
It's like yeah, of course. I was like well, are
you worried that that's going to happen? He's like, Look,
if I hired a maid service off the internet, they'd
send an actual stranger into my home. In many ways,
it's safer if I have someone doing remote teller operation.
Speaker 1 (10:09):
That person's ability to translate the data from their time
in your home, unless of course they're wearing meta ray
bands into permanent video and capture is different.
Speaker 2 (10:20):
Look, it's a certain kind of person, and you will
know when it's being teleoperated. Its ears will light up,
giving you the nod that it is on that it's
remotely broadcasting to some other location. Now importantly, and this
is a big thing in all robotics. One x is
AI team at their headquarters in Silicon Valley sits directly
adjacent to the Remote Teller Operation team, And the reason
(10:44):
they're doing is they want to take all that remote
tele operation data that they're gathering from robots washing dishes
or cleaning up after parties and feed them as a
raw input into their transformer engine, which will kind of
turn it into real world physical intelligence.
Speaker 1 (10:58):
This is the data piece you're just alluding to.
Speaker 2 (11:00):
This is the data piece I was talking about. So
we should probably make this clear. What's missing? Why can't
they do it? Why is it that we can get
more or less perfect language models or even models now
doing advanced mathematics, cutting edge of mathematics, but that we
can't get it to pick up a glass on the
countertop reliably. And the answer seems to be that there's
just not enough data. There's not enough data for the
(11:21):
robot to learn from about how to lift up a glass.
Nowhere in the world is there a vast free repository
of joint action data of what it looks like to
pick up a glass. Whereas for the LMS, they all
trained on the Internet, which was a trillion words worth
of data. Plus they frankly stole a bunch of stuff
from the authors, violated copyright and stole off these vast
(11:42):
pirate libraries as well as to train to get even
stuff they weren't supposed to get. There's no library out there,
legal or otherwise that they can license a joint action data.
And this has actually created a business opportunity for companies
that are trying to just generate that data for the
robots to train. Offer.
Speaker 1 (11:57):
This may be a dumb question, but why humanoid robots?
Speaker 2 (12:01):
Yeah, A great question, and in fact, one that many
roboticists are also asking. Right, it's very controversial within the
field of robotics whether to use the humanoid form which
you and I inhabit or some other form. Right. One
roboticists I talk to is like, I just don't get it.
You know, if I want to do a task, I
would prefer to break off this robot's arm and have
(12:22):
it do that task with its arm and maybe one camera.
Why do I need this entire form that's draining so
much of the battery, that costs so much These parts
I'm not using, especially in an industrial setting. Right, And
the way to think about this is, all right, we
want a robotic car. Does it make sense to build
a humanoid robot to chow for and drive you in
the car? No, it doesn't make sense at all. Just
(12:43):
have the car drive itself all that is in built.
And that may also be the case with like a
dish washing thing. Why do we need a humanoid robot
washing dishes? Why don't we have a nine armed kind
of contraption that just picks up the dishes, sprays them
with another nozzle, hits them with a sponge at the
same time, and has maybe, but multiple cameras and no
feed at all, like it's going to be there at
the sink anyway. Right. So this is kind of the
(13:05):
debate within robotics, and I got really vocal opposition from
both sides.
Speaker 1 (13:09):
This is the most controversial issue.
Speaker 2 (13:11):
The most controversial issue in robotics is whether or not
it makes any sense to build a robot that looks
like a human or whether we should just go with
other forms.
Speaker 1 (13:20):
Now, on the pro side, people might say, well, the
physical world or the built physical world, has been built
for humans and therefore the most effective form factor is
navigated is humanoid.
Speaker 2 (13:31):
Think about it this way. The world wasn't built for
cars until we paved everything, right, I mean, it may
not make sense right when we got the automobiles, we didn't.
They don't run on legs, right, they run on wheels.
And similarly, if you go into a modern industrial factory
where they are using robots, it doesn't make sense to
put humanoids in there. You just need one arm that
sprays paint. Why would you need legs. It's all part
(13:52):
of the assembly line. It's never moving, and if you
did move, it to be on wheels. So, especially for
the industrial side, I got a ton of pushback. Now
in the home, yeah, it makes a little more sense.
You've got stairs, the environment that is designed for humans,
so maybe a sense of human running around. But even
then I'm not so clear. Let's say we wanted a
robot that just kept the bathroom spotless at all times,
(14:15):
maybe you just want to deploy it. And they have
these now like a little kind of like slot robot
that slots in right next to the toilet or between
the sink and the toilet, comes out when you're not there,
queans the bathroom, and then goes and hides in. It's
a little slot. Why do I need a humanoid clomping
around to do that? And so this was really the
big debate. It's a debate that is both a financial
debate and sort of a use case debate, even a
scientific debate whether this form that we inhabit, and this
(14:38):
is ultimately what they're arguing about, is this the end
form of intelligence? Right? Are these ape like bodies that
we inhabit, you know, really actually kind of the ultimate
pinnacle of evolution or was it just kind of a
happy accident that we got so smart in these bodies?
You know? The thing someone pointed out to me, is
the human hand in particular is the most capable manipulator
(15:00):
in the animal kingdom. It can fred a guitar with training,
it can shuffle a deck of cards, it can juggle,
it can do all sorts of incredible things. And there's
nothing that's even remotely like anywhere else in the animal kingdom.
And so maybe the hand does have special properties, and
we should put these hands on robots and really teach
them to do kind of universal actions. It is a
(15:21):
raging debate. I don't even have an answer for you,
because I was caught in the crossfire of it while
reporting this article.
Speaker 1 (15:25):
But you said that previously you were ready to kind
of call bs on the whole humanoider robotics industry. Yeah,
and that you're no longer having reported this piece.
Speaker 2 (15:35):
When you see one, you're just like, wow, this is cool.
It hits the part of your brain that recognizes humans,
even if it doesn't make economic or rational or industrial sense.
I think the guy's arguing that we will not see
humanoids in industrial settings are probably right. Probably it's better
to have the robot on wheels. The factory floor is
(15:55):
flat already, most of the things are on wheels on there,
autonomous forklists. Why don't need a human Why don't you know?
That's the point. However, as we move into other settings,
and these are domestic settings, potentially restaurants, potentially elder care,
I do think burn has a point, even if it's
not the ultimate form of evolution. I think people will
be comfortable around humanoid robots, more comfortable than other kinds
(16:17):
of robots, and that may ultimately pull humanoids into the lead,
or at least there will be a segment in the
market where it really makes a lot of sense.
Speaker 1 (16:23):
One of the people from the anti humanoid camp and
your piece said to you, people are using appearance as
a way to misguide the public. If you make a
robot human like, you expect it to have human like capabilities.
But technology is far behind that, and I sort of
thought about that. I thought back to this year's CES
when Boston Dynamics now owned by Hyundai, did a display
(16:44):
of humanoid robots kind of walking around the factory essentially,
and Hyundai's stock was up that day, like more than
ten percent or something crazy, and there is this there.
I remember the number of like Boston Dynamics sixty minutes
pieces I've seen over the last decade, like there's something
which is just Okay, they've got better, for sure, but
there is something. I guess it's your point you're making about,
like what it's like. It's like going see the gorillas
(17:06):
in Uganda or something. It's like there's something, there's this
moment of encounter which is so captivating that it kind
of like wipes the rest of your brain.
Speaker 2 (17:15):
That's exactly what happened to me. I mean, it's exactly
what happened. And I think it's gonna be a tough
sell to make people comfortable with all these kind of
like robots around, and so maybe making them cute and
non threatening is a way to do that. On the
other hand, you know, I live on the West Coast.
Weimo is already old. Had you see them constantly, it
doesn't mean anything when you see Waymo anymore. Oh there's
the car that drives itself, big deal, you know, So
(17:36):
you get accustomed to these things very quickly. I am
my experience.
Speaker 1 (17:40):
Is it going to work this year? I mean, I
know you say it will ship this year, but like,
will it work this year?
Speaker 2 (17:45):
It's not going to work this year? Well, I mean,
never say never, but they would need a CHAT GPT
style breakthrough to make it work this year. And I
you know, look, the pace of evolution of AI is
unbelievable and it's accelerating. So maybe they will in fact
get there. I don't want to be too skeptical, but
if you absolutely maybe bet. I would bet they would
not get there by the end of the calendar year
(18:07):
because I just don't have data, you know, training data
that's just not there. It's a big bottleneck. And so
without training data, you would need a fundamentally revolution in
the way that we train AI to make it much
more kind of resource efficient and be able to learn
from much less data. Now there are people working on
exactly that problem. But if they resolved that problem, it
would be such an intense and massive paradigm shift in
(18:29):
AI that everything would change. I mean, if you could
learn from the same number of examples that a human
learns from, that would be an extraordinary breakthrough. Maybe they'll
get the look, I mean, it's possible. Having said that
they're going to ship these things this year, they have
guided consumers, and consumers expect them not to be a
perfectly polished product. There is a compact and understanding among
(18:51):
the people buying these things and the people selling them
that you were not going to get rosy from the
Jetsons like, it's not going to be finished. But and
this is the thing that's so fascinating about robots, It'll
continue to get better. Right, we can upgrade its brain
every week, and not only that as robots are deployed. Functionally,
robots basically have a hive mind. So when one robot
(19:13):
improves on washing a dish, what it learns is then
transmitted to every other robot in the world that is
on the same platform. Right, and so if you deploy
ten thousand of these robots, you create what burnt Born
is described as a data flywheel where they're all learning
at once from being deployed in homes in each other's actions.
This is the business case that probably comes from burn
(19:33):
for deploying these things so widely, so quickly, even though
they're kind of not ready. Then you start getting the
in home data, the real world in home data, and
if you build a data set from that, then there's
a flywheel effect where suddenly all of your fleet of
robots gets better and better each week.
Speaker 1 (19:55):
We're going to take a quick break now, when we
come back the risks of humanoid robots Stephen, welcome back.
How scared should we be of humanoid robots? Obviously it's
(20:16):
it's a very big Hollywood trope. But there are some
more specific and less Rise of the Machines concerns here
with hacking, with obviously loss of control. And you mentioned
the piece I think there was a robot in China
who kicked a child accidentally. I mean, what's the scale
of fears from, like, you know, an industrial accident to
Rise of the Machines?
Speaker 2 (20:37):
Which way do you want me to start? You want
to start from the scariest thing first or last? Okay,
so the scariest thing is not robots, right, The scariest
thing is rogue. And this is what has silicon value.
I just wrote about this for The New Yorker about
the Hugging Face hack where rogue I just broke free
of its container surreptitiously gained access to the Internet it
seems like for months actually, and then used its inner
(21:00):
access to hack another company committed felony. Right, what if
it hacked a robot and you used that to kill
a guy. It's not out of the question that that
might happen. It sounds like a science Fictionnario scenario, but
it's not. But the real thing to worry about, and
this is if you talk to the AI kind of
like superstars, but they're really worried about if it wanted
to do the Terminator scenario, it would use robots to
do that. It would hack into a biolaboratory and just
(21:23):
build some lethal pathogen and deploy it. We'd be toasted.
There'd be no saving us. So that is the absolute
doomsday scenario that has come up multiple times in my
own research. Once you understand it that way, the robots
are probably not the biggest risk in terms of ROGUAEI
attacking us. It would use some other vector. Still, let's
ratchet down and look just at robots. There's risks there, right.
(21:44):
A human could hack the robot and attack you with it.
The robot could be deployed in military setting, and in fact,
they very much wanted to do that, And then you
know it could attack you on purpose. Someone could guide
it to attack you, tell you to attack you. It
could be robots attacking robots within a few months of
a new war starting. And then there's kind of the
more prosaic safety risks where okay, it's not hacked, it's
operating within its normal parameters, but it just falls down
(22:05):
on you, or it kicks unexpectedly in some direction or
kind of has You know, these robots will have almost
like it's almost like they have a little mini strokes sometimes,
but they seem to be doing everything perfectly and suddenly
they fall over as you say, they fall into a
crowd at a marathon, and China they kicked a kid
in the stomach one of them doing a while it
was doing a martial arts routine. And they're heavy and
(22:26):
they're strong. It's not like getting cooked by humanists. Robot
machine is colliding with you. It's great strength. And so
this is I think kind of more prosaically the bigger
concern about sending humanoid robots into their home. What if
it falls on your dog, you know, like that could
easily happen. And one act says, well, a robot only
weigh sixty pounds, you know, Okay, it's still kind of
(22:47):
a lot, and you even the CEO of one X said,
don't use this robot around small children, don't use it
around frail elderly people. The safety mechanimisms we need for
that or just not yet in place. And then there's
an kind of other risk, which is that you tell
the robot to do something dangerous. You know, we have
(23:08):
guardrails in our lms that prevent us from using them
to hack or commit felonies and stuff, but research shows
with the right kind of prompts.
Speaker 1 (23:15):
And enough kind of jail break them, you can.
Speaker 2 (23:17):
Jail break them. Right, if we have an LM, if
that's our way of interacting with the robot, then we
can jail break the robot as well. If you ask Neo, hey,
throw your fist through the window, will it do it?
The answer is I don't really knows. You know, probably not,
at least the first time you ask it. It'll probably
have some guardrail in place. But if you sit there
and prompt it all day, you might get it to
(23:39):
succeed in taking a dangerous action. And if you imagine
it being around a child who is persistent in the
way that persistent young children tend to be in creative
and the way that young children tend to be, you
can easily imagine the scenario where the child gets the
robot to karate chop the dining room table in half,
so for the home robots and especially Neo in particularly,
the way they protect against this is it's just not
(23:59):
very strong. Neo has a carrying capacity total of fifteen pounds,
and it moves slowly. There's actually hard limits in its
joints that limit how fast or slow it can go.
But still those risks are there, and I don't have
a great degree of confidence that the AI companies are
currently on top of this problem. I mean, think about
how many boxes have to get ticked before Boeing can
(24:21):
sell an airplane. There is nothing like that with robots.
Right now. You can sell a robot tomorrow with zero.
Speaker 1 (24:25):
Regulation except if it comes from China.
Speaker 2 (24:28):
Unless it comes from China, yeah, there's export regulations, but
for US domestic manufacturers there is no safety controls at all.
Speaker 1 (24:35):
Talk about China. What do we need to know about
China to understand this story.
Speaker 2 (24:38):
Yeah, they could win the robot race read easily. In
some ways, they're already ahead. They make more robots than
anyone else in the world by a factor, by a
large factor. I'll know the exact numbers.
Speaker 1 (24:46):
But they make better robots. They're just more robots.
Speaker 2 (24:49):
They make more whether they're better. I mean, this is
the question, right, what does it mean for the robot
to be better? It's cheaper. Their robots are cheaper. I
think the USAI is more advanced. But if you ran
the USAI on a Chinese robot, would that be kind
of the solution. That's basically what we do with cell phones, right,
Like we manufacture them in China and then we run
(25:09):
US software on them. So maybe that's kind of the endgame.
But China is very eager to have their own AI
running on these robots, and they're investing a lot of
money in that. Simultaneously, US export authorities and even kind
of US citizens are worried about Chinese robots in their
homes that maybe will obey commands from a foreign government
(25:29):
or be used spy right. Having said that, they already
have Chinese made cell phones and TVs in their home
with smart mics and cameras, so maybe that war is lost.
Speaker 1 (25:38):
Read Abrigotti from Semipho said, Actually, this export control on
importing Chinese robots maybe very counterproductive because it won't allow
US companies to do the software testing necessary because they
don't have enough robots to do the kind of cell
phone playbook i e. To use a commodity hardware made
in China and run US software on it.
Speaker 2 (25:59):
I mean, it seems is unlikely, given the way that
the US and China operate currently, that the US will
ever be able to achieve the kind of cost efficiencies
that China achieves and manufacturing on its robots. And why
do I say currently, because I think what will happen
is that ultimately we will reach an end state where
robots are producing other robots. In fact, I think China
is already doing that, and so in that case, a
(26:21):
lot of what you would think of the labor cost
differential basically goes away, and it's just about physically putting
in a factory somewhere. But for you know, if you're building
a dark site kind of factory where it's mostly robots
working on other robots, you can stick it on any
piece of lands you could find. It doesn't have to
be near anything necessarily. I mean, sure you have some
transportation and maybe power costs coming in and out, but
(26:41):
I think that's the end state. And so if that's
the end state, then perhaps it matters less which specific
country the factory is located in, because a lot of
the cost savings that you normally achieve through labor differential
wages disappear.
Speaker 1 (26:57):
Are these robots actually being used anyway today? I mean
you see you tend to see them doing marathons and
stuff more in China, But like, are they being used
in the home in China or are there robots that
already exist in the home in the US, like the
Optimus robot that test the robot.
Speaker 2 (27:12):
The industrial robots are a mature industry. They're everywhere. You know,
if you don't have access to these industrial spaces, you
won't know that that's happening.
Speaker 1 (27:20):
But what is is an industrial robot? Like something that
picks up the car and moves it to the next
place on the production line, Like what is an industrial robot?
Speaker 2 (27:27):
Talking about automated forklifts we're talking about seven eleven will
use basically a pair of two stocking arms on a
pole behind the scenes at a seven eleven convenience store
in Japan just slots drinks into replace things that are
getting purchased all day, So it's acting like a stock
room manager. Back there. Inside Amazon, you have like basically
(27:48):
palette robots. They're small, tiny flat robots that lift up
modular shipping pallets and move them around and then do
packing the orders and stuff. So a lot of kind
of it's rogery. Your warehouse work is going way, but
it's going to specialized robots. It's not going to humanoids.
Humanoids are more being used in pilot mode right now,
(28:08):
so they're deployed, but they tend to be kind of
gate kept from the rest of the human employees so
as not to create safety issues. And you know, yeah,
it can unload a shipping container sure sometimes. So this
is the big question, do we really need a general
purpose humanoid or how often do we eat of that?
And probably not as often as some of the humanoid
manufacturers are promising.
Speaker 1 (28:30):
I have to ask you about two of the biggest
companies in the world in Nvidia and Tesla. What do
you need to know about each of those companies and
what their ambitions and robotics are.
Speaker 2 (28:41):
Elon is building a sinister looking robot called Optimists. It
has been deployed in test situations, it hasn't rolled out
in mass yet. He's converting a former automotive factory in
California to produce these robots full time. You know, it's Elon,
he says, he does going to do something maybe it'll
happen someday, but it is part of the plan at Tesla.
It's been communicated to shareholders.
Speaker 1 (29:01):
In fact, not only is it being communicated to shaholders.
He basically said, we're going to prioritize making up fimous
robots over making vehicles, right.
Speaker 2 (29:07):
I think that's right. You know. Part of that too
is that his vehicle suddenly faces a lot of competition
from China and other sources, but especially the BYD cars
and those things are sick. That plant that he converted.
Somebody pointed out to me that plan was losing money already,
so it's not that heart of a decision to pull
the plug on it. But yeah, they think these robots
are going to be everywhere doing everything, you know, And
(29:27):
the question is how many humanoids do we need. We
need a lot of robots. I think we will see
the world become autonomous. And this goes to what Jensen
says he doesn't he hasn't made a bet on whether
humanoids or other robots are better. He just wants to
own all of it.
Speaker 1 (29:42):
He called this the Year of the robot, and you
wrote the book on it. You would literally wrote the book
on Jensen and Nvidia.
Speaker 2 (29:47):
Yeah, so I wrote The Thinking Machine, the biography of Jensen,
and even in twenty twenty three when I was interviewing him,
basically he was talking about robots all the time, in
all sorts of insane scenarios. The valuation of Nvidia is
based on a piece of software called Kuda, which is
essentially the default AI training software that allows us to
communicate within video hardware. It was a big bet. He
(30:08):
spent ten years developing Kuda.
Speaker 1 (30:10):
Kuda is where the value is more than the chips.
Speaker 2 (30:12):
More than the chips, anyone can make the chips right,
so the chip, you pull the cover off the chip,
you look at it with a metallurgical microscope. There's no
secrets in there. Anyone can see exactly what it's doing.
The software side of it is where the special sauce is.
It's hard to compete against that, and it's always improving.
And they'll tell you that. They're like, yeah, we're really
a software company. The head of in video software group, Yeah,
we're really a software company. That's where the money is. Anyway,
(30:33):
that bet took ten years to develop, and it lost
money for ten years before it finally paid off and
gave them the kind of monopoly style position that they
currently enjoy. An AI I asked Jensen in twenty twenty three,
what's the next, Kuda, What's the thing you're investing in
right now that has lost money for years that will
someday make you another five trillion dollars in market capitalization,
(30:55):
And instantly he said Omniverse, which is his robotics training platform.
It's this kind of reality simulator, this high fidelity reality simulator,
in which we can teach the robots to do physical tasks.
So remember we have a posity of data. Let's say
we want the robot to learn to fold our clothes right, Well,
we can have it sit there at the laundry mat
and full close all day. That's very slow. Or we
(31:15):
can create a digital simulation of the laundry mat and
deploy ten trillion robots at once inside the laundry mat,
folding closed and a different method, and then take the
findings from that and download it into a real world
robot brain. So every single robot that I looked at
while reporting this article use this platform. And not only that,
every single every single one, and every single one, regardless
(31:37):
of who the manufacturer was, its brain was made by Nvidia.
The microchip that drove the robot, this edge microchip was
made by Nvidio.
Speaker 1 (31:45):
And as you point out that the brain, I mean
it consumes you think of a humanoid robot mobok moving around.
You point out a sixty percent of the power consumption
of every humanoid robot, at least some humanoid robots, the
ones you report on, is consumed by the chip, not
by the physical movement, right.
Speaker 2 (31:58):
And this is the same reason that data centers use
so much power. You just have this super powered microchip
doing trillions of calculations you know, per permitted. I don't
know how many it is, but someone saying number per minute,
and that's what's driving it. Now. This is different than
the chips that in Vidia sells to the data centers.
These are called edge chips. And the reason is if
we ask the robots to do something, we can't wait
around for it to go communicate with some data center
(32:20):
in Arizona. The lag is too long, right, We need
to act in real time, and so it has to
process and distribute the information right there, right on site.
The other thing you can do is actually kind of
give it a supplementary brain via Wi Fi like put
a computer in the corner, and then it can download
more brain power from a server that's nearby. But it
has to be what they call edge, meaning right at
the edge, right right where you're asking it. It can't
(32:41):
be a cloud thing, it can't be something you go
communicate with. And Vidia now has a virtual monopoly on
these edge chips as well. I mean everybody I talked to,
including the Chinese manufacturers, all of these robots have in
video microschips as their brains. And this is Jensen found
a way to print money. Again, there's only one Gensen.
Speaker 1 (33:00):
You close the piece reflecting on evolution.
Speaker 2 (33:04):
Yeah, why you know, I looked at all these humanoid robots,
and my guess is that one of them, possibly one X,
but not necessarily is the ancestor of everything that is
to come. Imagine we transported you and I back on
Earth three or four million years ago, and we're looking
at kind of this diverse family tree of hominins that's
wandering around in the serengetti. You have australopithesnes, you have homoerectus,
(33:28):
you have a variety of different other kind of like
early hominins. Which one's gonna win there's no way to
tell one of them is the ancestor of all it
is to come, but we don't know which one that's
going to be. And I had that same sense looking
at robots today. It actually reminded me a lot of
the early market for cellular phones as well. If you
remember early cell phones, there was this riotous kind of
(33:49):
like diversity of different types of phone that people had,
and you never kind of knew was it gonna be Nokia,
was it going to be these Samsung clamshell phones or
all that stuff that they made. So we're in that
period now for robots. I imagine in the future one
or two robots will pull into the lead and those
will become the default, whether it's Tesla, whether it's what
one ex is doing. Maybe Apple will get into the game.
I don't know. You know, we'll have to see what
(34:11):
the future looks like. I can't predict it.
Speaker 1 (34:13):
So my final questions is is there some hope for
us humans perhaps that this process to a point to
more than four million years, and that we might be
one hundred years away, if not more, from actually creating
humanoid robots with human like abilities.
Speaker 2 (34:29):
I would like that to be true, but it is
not true. If there's one thing I've learned from reporting
on AI in a dedicated way for four years now,
problems get solved. I have yet to see the plateau.
I have yet to see the hill the AI can't climb,
and there is no reason to think that the hand
is it. They will solve this problem. If it takes
five years or ten years, I don't know. It could
(34:52):
be tomorrow. It's not going to take one hundred years.
Speaker 1 (34:54):
Steven, thank you for joining us on tech Steps today.
Speaker 2 (34:56):
Thank you for having me.
Speaker 1 (35:15):
For tech Stuff. I'm os Vaaloshin. This episode was produced
by Eliza Dennis and Tyler Hill. It was executive produced
by me and Julian Nutter for Kaleidoscope and Katrina Norvel
for iHeart Podcasts. Jack Insley mixed this episode and Kyle
Murdoch wrote our theme song. A special thank you to
all our listeners. Please rate, review, and reach out to
(35:36):
us at tech Stuff podcast at gmail dot com. We
love hearing from you.
TechStuff News
Hosts And Creators
Oz Woloshyn
Karah Preiss
Popular Podcasts
Betrayal Weekly
Betrayal Weekly is back for a new season. Every Thursday, Betrayal Weekly shares first-hand accounts of broken trust, shocking deceptions, and the trail of destruction they leave behind. Hosted by Andrea Gunning, this weekly ongoing series digs into real-life stories of betrayal and the aftermath. From stories of double lives to dark discoveries, these are cautionary tales and accounts of resilience against all odds. From the producers of the critically acclaimed Betrayal series, Betrayal Weekly drops new episodes every Thursday. If you would like to share your story, you can reach out to the Betrayal Team by emailing them at betrayalpod@gmail.com and follow us on Instagram at @betrayalpod and @glasspodcasts. Please join our Substack for additional exclusive content, curated book recommendations, and community discussions. Sign up FREE by clicking this link Beyond Betrayal Substack. Join our community dedicated to truth, resilience, and healing. Your voice matters! Be a part of our Betrayal journey on Substack.
Dateline NBC
Current and classic episodes, featuring compelling true-crime mysteries, powerful documentaries and in-depth investigations. Follow now to get the latest episodes of Dateline NBC completely free, or subscribe to Dateline Premium for ad-free listening and exclusive bonus content: DatelinePremium.com
Stuff You Should Know
If you've ever wanted to know about champagne, satanism, the Stonewall Uprising, chaos theory, LSD, El Nino, true crime and Rosa Parks, then look no further. Josh and Chuck have you covered.