The following is a conversation with Rodney Brooks, one of the greatest roboticists in history.

He led the Computer Science and Artificial Intelligence Laboratory at MIT,

then cofounded iRobot, which is one of the most successful robotics companies ever.

Then he cofounded Rethink Robotics that created some amazing collaborative robots like Baxter

and Sawyer. Finally, he cofounded Robust.ai, whose mission is to teach robots common sense,

which is a lot harder than it sounds. To support this podcast,

please check out our sponsors in the description.

As a side note, let me say that Rodney is someone I've looked up to for many years in my now over

two decade journey in robotics because, one, he's a legit great engineer of real world systems,

and two, he's not afraid to state controversial opinions that challenge the way we see the AI

world. But of course, while I agree with him on some of his critical views of AI, I don't agree

with some others, and he's fully supportive of such disagreement. Nobody ever built anything great

by being fully agreeable. There's always respect and love behind our interactions, and when a

conversation is recorded like it was for this podcast, I think a little bit of disagreement is

fun. This is the Lex Friedman Podcast, and here is my conversation with Rodney Brooks.

What is the most amazing or beautiful robot that you've ever had the chance to work with?

I think it was Domo, which was made by one of my grad students, Aaron Edsinger. It now sits in

Daniela Russo's office, director of CSAIL, and it was just a beautiful robot. Aaron was really

clever. He didn't give me a budget ahead of time. He didn't tell me what he was going to do.

He just started spending money. He spent a lot of money. He and Jeff Weber, who is a mechanical

engineer who Aaron insisted he bring with him when he became a grad student, built this beautiful,

gorgeous robot, Domo, which is an upper torso humanoid, two arms with fingers, three fingered

hands, and face eyeballs. Not the eyeballs, but everything else, series elastic actuators.

You can interact with it. Cable driven. All the motors are inside, and it's just gorgeous.

The eyeballs are actuated too, or no?

Oh yeah, the eyeballs are actuated with cameras, so it had a visual attention mechanism,

looking when people came in and looking in their face and talking with them.

Wow, was it amazing?

The beauty of it.

You said what was the most beautiful?

What is the most beautiful?

It's just mechanically gorgeous. As everything Aaron builds,

there's always been mechanically gorgeous. It's just exquisite in the detail.

We're talking about mechanically, like literally the amount of actuators.

The actuators, the cables, he anodizes different parts, different colors,

and it just looks like a work of art.

What about the face? Do you find the face beautiful in robots?

When you make a robot, it's making a promise for how well it will be able to interact,

so I always encourage my students not to overpromise.

Even with its essence, like the thing it presents, it should not overpromise.

Yeah, so the joke I make, which I think you'll get, is if your robot looks like Albert Einstein,

it should be as smart as Albert Einstein.

So the only thing in Domo's face is the eyeballs, because that's all it can do.

It can look at you and pay attention.

It's not like one of those Japanese robots that looks exactly like a person at all.

But see, the thing is, us humans and dogs, too, don't just use eyes as attentional mechanisms.

They also use it to communicate, as part of the communication.

Like a dog can look at you, look at another thing, and look back at you,

and that designates that we're going to be looking at that thing together.

Yeah, or intent, you know, on both Baxter and Sawyer at Rethink Robotics,

they had a screen with, you know, graphic eyes,

so it wasn't actually where the cameras were pointing, but the eyes would look in the direction

it was about to move its arm, so people in the factory nearby were not surprised by its motions,

because it gave that intent away.

Before we talk about Baxter, which I think is a beautiful robot, let's go back to the beginning.

When did you first fall in love with robotics?

We're talking about beauty and love to open the conversation.

This is great.

I was born in the end of 1954, and I grew up in Adelaide, South Australia,

and I have these two books that are dated 1961, so I'm guessing my mother found them in a store

in 62 or 63, How and Why Wonder Books.

How and Why Wonder Book of Electricity, and a How and Why Wonder Book of Giant Brains and Robots.

And I learned how to build circuits, you know, when I was eight or nine, simple circuits,

and I read, you know, learned the binary system, and saw all these drawings, mostly, of robots,

and then I tried to build them for the rest of my childhood.

Wait, 61, you said?

This was when the two books, I've still got them at home.

What does the robot mean in that context?

Some of the robots that they had were arms, you know, big arms to move nuclear material around,

but they had pictures of welding robots that looked like humans under the sea, welding stuff

underwater.

So they weren't real robots, but they were, you know, what people were thinking about for robots.

What were you thinking about?

Were you thinking about humanoids?

Were you thinking about arms with fingers?

Were you thinking about faces or colors?

Were you thinking about faces or cars?

No, actually, to be honest, I realized my limitation on building mechanical stuff.

So I just built the brains, mostly, out of different technologies as I got older.

I built a learning system which was chemical based, and I had this ice cube tray.

Each well was a cell, and by applying voltage to the two electrodes, it would build up a

copper bridge.

So over time, it would learn a simple network so I could teach it stuff.

And mostly, things were driven by my budget, and nails as electrodes and an ice cube tray

was about my budget at that stage.

Later, I managed to buy transistors, and I could build gates and flip flops and stuff.

So one of your first robots was an ice cube tray?

Yeah, it was very cerebral because it learned to add.

Very nice.

Well, just a decade or so before, in 1950, Alan Turing wrote a paper that formulated

the Turing Test, and he opened that paper with the question, can machines think?

So let me ask you this question.

Can machines think?

Can your ice cube tray one day think?

Certainly, machines can think because I believe you're a machine, and I'm a machine, and I

believe we both think.

I think any other philosophical position is sort of a little ludicrous.

What does think mean if it's not something that we do?

And we are machines.

So yes, machines can, but do we have a clue how to build such machines?

That's a very different question.

Are we capable of building such machines?

Are we smart enough?

We think we're smart enough to do anything, but maybe we're not.

Maybe we're just not smart enough to build stuff like us.

The kind of computer that Alan Turing was thinking about, do you think there is something

fundamentally or significantly different between the computer between our ears, the biological

computer that humans use, and the computer that he was thinking about from a sort of

high level philosophical?

Yeah, I believe that it's very wrong.

In fact, I'm halfway through a, I think it'll be about a 480 page book, the working title

is Not Even Wrong.

And if I may, I'll tell you a bit about that book.

Yes, please.

So there's two, well, three thrusts to it.

One is the history of computation, what we call computation.

It goes all the way back to some manuscripts in Latin from 1614 and 1620 by Napier and

Kepler through Babbage and Lovelace.

And then Turing's 1936 paper is what we think of as the invention of modern computation.

And that paper, by the way, did not set out to invent computation.

It set out to negatively answer one of Hilbert's three later set of problems.

He called it an effective way of getting answers.

And Hilbert really worked with rewriting rules, as did Church, who also, at the same time,

a month earlier than Turing, disproved Hilbert's one of these three hypotheses.

The other two had already been disproved by Gödel.

Turing set out to disprove it, because it's always easier to disprove these things than

to prove that there is an answer.

And so he needed, and it really came from his professor while I was an undergrad at

Cambridge, who turned it into, is there a mechanical process?

So he wanted to show a mechanical process that could calculate numbers, because that

was a mechanical process that people used to generate tables.

They were called computers, the people at the time.

And they followed a set of rules where they had paper, and they would write numbers down,

and based on the numbers, they'd keep writing other numbers.

And they would produce numbers for these tables, engineering tables, that the more iterations

they did, the more significant digits came out.

And so Turing, in that paper, set out to define what sort of machine could do that, mechanical

machine, where it could produce an arbitrary number of digits in the same way a human computer

did.

And he came up with a very simple set of constraints where there was an infinite supply

of paper.

This is the tape of the Turing machine, and each Turing machine came with a set of instructions

that, as a person, could do with pencil and paper, write down things on the tape and erase

them and put new things there.

And he was able to show that that system was not able to do something that Hilbert had

hypothesized, so he disproved it.

But he had to show that this system was good enough to do whatever could be done, but couldn't

do this other thing.

And there he said, and he says in the paper, I don't have any real arguments for this,

but based on intuition.

So that's how he defined computation.

And then if you look over the next, from 1936 up until really around 1975, you see people

struggling with, is this really what computation is?

And so Marvin Minsky, very well known in AI, but also a fantastic mathematician, in his

book Finite and Infant Machines from the mid-'60s, which is a beautiful, beautiful mathematical

book, says at the start of the book, well, what is computation?

Turing says it's this, and yeah, I sort of think it's that.

It doesn't really matter whether the stuff's made of wood or plastic.

It's just that relatively cheap stuff can do this stuff.

And so yeah, seems like computation.

And Donald Knuth, in his first volume of his Art of Computer Programming in around 1968,

says, well, what's computation?

It's this stuff, like Turing says, that a person could do each step without too much

trouble.

And so one of his examples of what would be too much trouble was a step which required

knowing whether Fermat's Last Theorem was true or not, because it was not known at the

time.

And that's too much trouble for a person to do as a step.

And Hopcroft and Ullman sort of said a similar thing later that year.

And by 1975, in the A.H.O.

Hopcroft and Ullman book, they're saying, well, you know, we don't really know what

computation is, but intuition says this is sort of about right, and this is what it is.

That's computation.

It's a sort of agreed upon thing which happens to be really easy to implement in silicon.

And then we had Moore's Law, which took off, and it's been an incredibly powerful tool.

I certainly wouldn't argue with that.

The version we have of computation, incredibly powerful.

Can we just take a pause?

So what we're talking about is there's an infinite tape with some simple rules of how

to write on that tape, and that's what we're kind of thinking about.

This is computation.

Yeah, and it's modeled after humans, how humans do stuff.

And I think it's, Turing says in the 36th paper, one of the critical facts here is that

a human has a limited amount of memory.

So that's what we're going to put onto our mechanical computers.

So, you know, I'm like mass.

I'm like mass or charge or, you know, it's not given by the universe.

It was, this is what we're going to call computation.

And then it has this really, you know, it had this really good implementation, which

has completely changed our technological world.

That's computation.

Second part of the book, or argument in the book, I have this two by two matrix with science.

In the top row, engineering in the bottom row, left column is intelligence, right column

is life.

So in the bottom row, the engineering, there's artificial intelligence and artificial life.

In the top row, there's neuroscience and abiogenesis.

How does living matter turn in?

How does nonliving matter become living matter?

Four disciplines.

These four disciplines all came into the current form in the period 1945 to 1965.

That's interesting.

There was neuroscience before, but it wasn't effective neuroscience.

It was, you know, there were these ganglia and there's electrical charges, but no one

knows what to do with it.

And furthermore, there are a lot of players who are common across them.

I've identified common players except for artificial intelligence and abiogenesis.

I don't have, but for any other pair, I can point to people who work them.

And a whole bunch of them, by the way, were at the research lab for electronics at MIT

where Warren McCulloch held forth.

In fact, McCulloch, Pitts, Letvin, and Maturana wrote the first paper on functional neuroscience

called What the Frog's Eye Tells the Frog's Brain, where instead of it just being this

bunch of nerves, they sort of showed what different anatomical components were doing

and telling other anatomical components and, you know, generating behavior in the frog.

Would you put them as basically the fathers or one of the early pioneers of what are now

called artificial neural networks?

Yeah, I mean, McCulloch and Pitts.

Pitts was a much younger than him.

In 1943, had written a paper inspired by Bertrand Russell on a calculus for the ideas eminent

in neural systems where they had tried to, without any real proof, they had tried to

give a formalism for neurons basically in terms of logic and gates or gates and not

gates with no real evidence that that was what was going on, but they talked about it

and that was picked up by Minsky for his 1953 dissertation on, which was a neural

network, we call it today.

It was picked up by John von Neumann when he was designing the Edbeck computer in 1945.

He talked about its components being neurons based on, and in references, he's only got

three references and one of them is the McCulloch Pitts paper.

So all these people and then the AI people and the artificial life people, which was

John von Neumann originally, there's like overlap between all, they're all going around

the same time.

And three of these four disciplines turned to computation as their primary metaphor.

So I've got a couple of chapters in the book.

One is titled, wait, computers are people?

Because that's where our computers came from.

Yeah.

And, you know, from people who were computing stuff.

And then I've got another chapter, wait, people are computers?

Which is about computational neuroscience.

Yeah.

So there's this whole circle here.

And that computation is it.

And, you know, I have talked to people about, well, maybe it's not computation that goes

on in the head.

Of course it is.

Yeah.

Okay, well, when Elon Musk's rocket goes up, is it computing?

Is that how it gets into orbit?

By computing?

But we've got this idea, if you want to build an AI system, you write a computer program.

Yeah, so the word computation very quickly starts doing a lot of work that it was not

initially intended to do.

It's the second and same if you talk about the universe as essentially performing a

computation.

Yeah, right.

Wolfram does this.

He turns it into computation.

You don't turn rockets into computation.

Yeah.

By the way, when you say computation in our conversation, do you tend to think of computation

narrowly in the way Turing thought of computation?

It's gotten very, you know, squishy.

Yeah.

Squishy.

But computation in the way Turing thinks about it and the way most people think about it

actually fits very well with thinking like a hunter gatherer.

There are places and there can be stuff in places and the stuff in places can change

and it stays there until someone changes it.

And it's this metaphor of place and container, which, you know, is a combination of our place

cells in our hippocampus and our cortex.

But this is how we use metaphors for mostly to think about.

And when we get outside of our metaphor range, we have to invent tools which we can sort

of switch on to use.

So calculus is an example of a tool.

It can do stuff that our raw reasoning can't do, and we've got conventions of when you

can use it or not.

But sometimes, you know, people try to all the time, we always try to get physical metaphors

for things, which is why quantum mechanics has been such a problem for a hundred years.

Because it's a particle.

No, it's a wave.

It's got to be something we understand.

And I say, no, it's some weird mathematical logic that's different from those, but we

want that metaphor.

Well, you know, I suspect that, you know, a hundred years or 200 years from now, neither

quantum mechanics nor dark matter will be talked about in the same terms, you know,

in the same way that Flogerson's theory eventually went away.

Because it just wasn't an adequate explanatory metaphor, you know.

That metaphor was the stuff, there is stuff in the burning, the burning is in the matter.

As it turns out, the burning was outside the matter, it was the oxygen.

So our desire for metaphor and combined with our limited cognitive capabilities gets us

into trouble.

That's my argument in this book.

Now, and people say, well, what is it then?

And I say, well, I wish I knew that, right, the book about that.

But I, you know, I give some ideas.

But so there's the three things.

Computation is sort of a particular thing we use.

Oh, can I tell you one beautiful thing, one beautiful thing I found?

So, you know, I used an example of a thing that's different from computation.

You hit a drum and it vibrates, and there are some stationary points on the drum surface,

you know, because the waves are going up and down the stationary points.

Now, you could compute them to arbitrary precision, but the drum just knows them.

The drum doesn't have to compute.

What was the very first computer program ever written by Ada Lovelace?

To compute Bernoulli numbers, and the Bernoulli numbers are exactly what you need to find those

stable points in the drum surface.

Wow.

And there was a bug in the program.

The arguments to divide were, I don't know, I don't know.

The arguments to divide were reversed in one place.

And it still worked?

Well, no, she's never got to run it.

They never built the analytical engine.

She wrote the program without it, you know.

So the computation?

Computation is sort of, you know, a thing that's become dominant as a metaphor, but

is it the right metaphor?

All three of these four fields adopted computation.

And, you know, a lot of it swirls around Warren McCulloch and all his students, and he funded

a lot of people.

And our human metaphors, our limitations to human thinking, all play into this.

Those are the three themes of the book.

So I have a little to say about computation.

So you're saying that there is a gap between the computer or the machine that performs

computation and this machine that appears to have consciousness and intelligence.

Yeah, that piece of meat in your head.

Piece of meat.

And maybe it's not just the meat in your head, it's the rest of you too.

I mean, you actually have a neural system in your gut.

I tend to also believe, not believe, but we're now dancing around things we don't know, but

I tend to believe other humans are important.

Like, so we're almost like, I just don't think we would ever have achieved the level

of intelligence we have with other humans.

I'm not saying so confidently, but I have an intuition that some of the intelligence

is in the interaction.

Yeah, and I think it seems to be very likely, again, this is speculation, but we, our species,

and probably neanderthals to some extent, because you can find old bones where they

seem to be counting on them by putting notches that were neanderthals, we are able to put

some of our stuff outside our body into the world.

And then other people can share it.

And then we get these tools that become shared tools.

And so there's a whole coupling that would not occur in the single deep learning network,

which was fed all of literature or something.

Yeah, the neural network can't step outside of itself.

But is there some, can we explore this dark room a little bit and try to get at something?

What is the magic?

Where does the magic come from in the human brain that creates the mind?

What's your sense as scientists that try to understand it and try to build it?

What are the directions it followed might be productive?

Is it creative, interactive robots?

Is it creating large deep neural networks that do like self supervised learning and

just like we'll discover that when you make something large enough, some interesting things

will emerge?

Is it through physics and chemistry, biology, like artificial life angle?

Like we'll sneak up in this four quadrant matrix that you mentioned.

Is there anything you're most, if you had to bet all your money, financial?

I wouldn't.

So every intelligence we know, animal intelligence, dog intelligence,

octopus intelligence, which is a very different sort of architecture from us.

All the intelligences we know perceive the world in some way and then have action in

the world, but they're able to perceive objects in a way which is actually pretty damn phenomenal

and surprising.

We tend to think that the box over here between us, which is a sound box, I think is a blue

box, but blueness is something that we construct with color constancy.

The blueness is not a direct function of the photons we're receiving.

It's actually context, which is why you can turn, maybe seeing the examples where someone

turns a stop sign into some other sort of sign by just putting a couple of marks on

them and the deep learning system gets it wrong.

And everyone says, but the stop sign's red.

Why is it thinking it's the other sort of sign?

Because redness is not intrinsic in just the photons.

It's actually a construction of an understanding of the whole world and the relationship between

objects to get color constancy.

But our tendency, in order that we get an archive paper really quickly, is you just

show a lot of data and give the labels and hope it figures it out.

But it's not figuring it out in the same way we do.

We have a very complex perceptual understanding of the world.

Dogs have a very different perceptual understanding based on smell.

They go smell a post, they can tell how many different dogs have visited it in the last

10 hours and how long ago.

There's all sorts of stuff that we just don't perceive about the world.

And just taking a single snapshot is not perceiving about the world.

It's not seeing the registration between us and the object.

And registration is a philosophical concept.

Brian Cantwell Smith talks about it a lot.

Very difficult, squirmy thing to understand.

But I think none of our systems do that.

We've always talked in AI about the symbol grounding problem, how our symbols that we

talk about are grounded in the world.

And when deep learning came along and started labeling images, people said, ah, the grounding

problem has been solved.

No, the labeling problem was solved with some percentage accuracy, which is different from

the grounding problem.

So you agree with Hans Marvick and what's called the Marvick's paradox that highlights

this counterintuitive notion that reasoning is easy, but perception and mobility are hard.

Yeah.

We shared an office when I was working on computer vision and he was working on his

first mobile robot.

What were those conversations like?

They were great.

So do you still kind of, maybe you can elaborate, do you still believe this kind of notion that

perception is really hard?

Like, can you make sense of why we humans have this poor intuition about what's hard

and not?

Well, let me give us sort of another story.

Sure.

If you go back to the original teams working on AI from the late 50s into the 60s, and

you go to the AI lab at MIT, who was it that was doing that?

It was a bunch of really smart kids who got into MIT and they were intelligent.

So what's intelligence about?

Well, the stuff they were good at, playing chess, doing integrals, that was hard stuff.

But, you know, a baby could see stuff, that wasn't intelligent, anyone could do that,

that's not intelligence.

And so, you know, there was this intuition that the hard stuff is the things they were

good at and the easy stuff was the stuff that everyone could do.

Yeah.

And maybe I'm overplaying it a little bit, but I think there's an element of that.

Yeah, I mean, I don't know how much truth there is to, like chess, for example, was

for the longest time seen as the highest level of intellect, right?

Until we got computers that were better at it than people.

And then we realized, you know, if you go back to the 90s, you'll see, you know, the

stories in the press around when Kasparov was beaten by Deep Blue.

Oh, this is the end of all sorts of things.

Computers are going to be able to do anything from now on.

And we saw exactly the same stories with Alpha Zero, the Go Playing program.

Yeah.

But still, to me, reasoning is a special thing.

And perhaps...

No, actually, we're really bad at reasoning.

We just use these analogies based on our hunter gatherer intuitions.

But why is that not, don't you think the ability to construct metaphor is a really powerful

thing?

Oh, yeah, it is.

Tell stories.

It is.

It's the constructing the metaphor and registering that something constant in our brains.

Like, isn't that what we're doing with vision too?

And we're telling our stories.

We're constructing good models of the world.

Yeah, yeah.

But I think we jumped between what we're capable of and how we're doing it right there.

It was a little confusion that went on as we were telling each other stories.

Yes, exactly.

Trying to delude each other.

No, I just think I'm not exactly so.

I'm trying to pull apart this Moravec's paradox.

I don't view it as a paradox.

What did evolution spend its time on?

Yes.

It spent its time on getting us to perceive and move in the world.

That was 600 million years as multi cell creatures doing that.

And then it was relatively recent that we were able to hunt or gather or even animals hunting.

That's much more recent.

And then anything that we, speech, language, those things are a couple of hundred thousand

years probably, if that long.

And then agriculture, 10,000 years.

All that stuff was built on top of those earlier things, which took a long time to develop.

So if you then look at the engineering of these things, so building it into robots,

what's the hardest part of robotics?

Do you think as the decades that you worked on robots in the context of what we're talking

about, vision, perception, the actual sort of the biomechanics of movement, I'm kind

of drawing parallels here between humans and machines always.

Like what do you think is the hardest part of robotics?

I just want to think all of them.

I just want to think all of them.

There are no easy parts to do well.

We sort of go reductionist and we reduce it.

If only we had all the location of all the points in 3D, things would be great.

If only we had labels on the images, things would be great.

But as we see, that's not good enough.

Some deeper understanding.

But if I came to you and I could solve one category of problems in robotics instantly,

what would give you the greatest pleasure?

I mean, you look at robots that manipulate objects, what's hard about that?

You know, is it the perception, is it the reasoning about the world, that common sense

reasoning, is it the actual building a robot that's able to interact with the world?

Is it like human aspects of a robot that's interacting with humans in that game theory

of how they work well together?

Well, let's talk about manipulation for a second because I had this really blinding

moment, you know, I'm a grandfather, so grandfathers have blinding moments.

Just three or four miles from here, last year, my 16 month old grandson was in his new house

for the first time, right?

First time in this house.

And he'd never been able to get to a window before, but this had some low windows.

And he goes up to this window with a handle on it that he's never seen before.

And he's got one hand pushing the window and the other hand turning the handle to open

the window.

He knew two different hands, two different things he knew how to put together.

And he's 16 months old.

And there you are watching in awe.

In an environment he'd never seen before, a mechanism he'd never seen.

How did he do that?

Yes, that's a good question.

How did he do that?

That's why.

It's like, okay, like you could see the leap of genius from using one hand to perform a

task to combining, doing, I mean, first of all, in manipulation, that's really difficult.

It's like two hands, both necessary to complete the action.

And completely different.

And he'd never seen a window open before, but he inferred somehow handle open something.

Yeah, there may have been a lot of slightly different failure cases that you didn't see.

Not with a window, but with other objects of turning and twisting and handles.

There's a great counter to reinforcement learning.

We'll just give the robot plenty of time to try everything.

Can I tell a little side story here?

Yeah, so I'm in DeepMind in London, this is three, four years ago, where there's a big

Google building, and then you go inside and you go through this more security, and then

you get to DeepMind where the other Google employees can't go.

And I'm in a conference room, a conference room with some of the people, and they tell

me about their reinforcement learning experiment with robots, which are just trying stuff out.

And they're my robots.

They're Sawyer's.

We sold them.

And they really like them because Sawyer's are compliant and can sense forces, so they

don't break when they're bashing into walls.

They stop and they do all this stuff.

So you just let the robot do stuff, and eventually it figures stuff out.

By the way, Sawyer, we're talking about robot manipulation, so robot arms and so on.

Yeah, Sawyer's a robot.

What's Sawyer?

Sawyer's a robot arm that my company Rethink Robotics built.

Thank you for the context.

Sorry.

Okay, cool.

So we're in DeepMind.

And it's in the next room, these robots are just bashing around to try and use reinforcement

learning to learn how to act.

Can I go see them?

Oh no, they're secret.

They were my robots.

They were secret.

That's hilarious.

Okay.

Anyway, the point is, you know, this idea that you just let reinforcement learning figure

everything out is so counter to how a kid does stuff.

So again, story about my grandson.

I gave him this box that had lots of different lock mechanisms.

He didn't randomly, you know, and he was 18 months old, he didn't randomly try to touch

every surface or push everything.

He found he could see where the mechanism was, and he started exploring the mechanism

for each of these different lock mechanisms.

And there was reinforcement, no doubt, of some sort going on there.

But he applied a pre filter, which cut down the search space dramatically.

I wonder to what level we're able to introspect what's going on.

Because what's also possible is you have something like reinforcement learning going

on in the mind in the space of imagination.

So like you have a good model of the world you're predicting and you may be running those

tens of thousands of like loops, but you're like, as a human, you're just looking at yourself

trying to tell a story of what happened.

And it might seem simple, but maybe there's a lot of computation going on.

Whatever it is, but there's also a mechanism that's being built up.

It's not just random search.

Yeah, that mechanism prunes it dramatically.

Yeah, that pruning, that pruning stuff, but it doesn't, it's possible that that's, so

you don't think that's akin to a neural network inside a reinforcement learning algorithm.

Is it possible?

It's, yeah, until it's possible.

It's possible, but I'll be incredibly surprised if that happens.

I'll also be incredibly surprised that after all the decades that I've been doing this,

where every few years someone thinks, now we've got it.

Now we've got it.

Four or five years ago, I was saying, I don't think we've got it yet.

And everyone was saying, you don't understand how powerful AI is.

I had people tell me, you don't understand how powerful it is.

I sort of had a track record of what the world had done to think, well, this is no different

from before.

Or we have bigger computers.

We had bigger computers in the 90s and we could do more stuff.

But okay, so let me push back because I'm generally sort of optimistic and try to find

the beauty in things.

I think there's a lot of surprising and beautiful things that neural networks, this new generation

of deep learning revolution has revealed to me, has continually been very surprising

the kind of things it's able to do.

Now, generalizing that over saying like this, we've solved intelligence.

That's another big leap.

But is there something surprising and beautiful to you about neural networks that were actually

you said back and said, I did not expect this?

Oh, I think their performance on ImageNet was shocking.

The computer vision in those early days was just very like, wow, okay.

That doesn't mean that they're solving everything in computer vision we need to solve or in

vision for robots.

What about AlphaZero and self play mechanisms and reinforcement learning?

Yeah, that was all in the 90s.

Yeah, that was all in Donald Mickey's 1961 paper.

Everything that was there, which introduced reinforcement learning.

No, but come on.

So no, you're talking about the actual techniques.

But isn't it surprising to you the level it's able to achieve with no human supervision

of chess play?

Like, to me, there's a big, big difference between Deep Blue and...

Maybe what that's saying is how overblown our view of ourselves is.

You know, the chess is easy.

Yeah, I mean, I came across this 1946 report that, and I'd seen this as a kid in one of

those books that my mother had given me actually.

The 1946 report, which pitted someone with an abacus against an electronic calculator,

and he beat the electronic calculator.

You know, so there at that point was, well, humans are still better than machines at calculating.

Are you surprised today that a machine can, you know, do a billion floating point operations

a second and, you know, you're puzzling for minutes through one?

I mean, I don't know, but I am certainly surprised there's something, to me, different about

learning, so a system that's able to learn.

Learning.

See, now you're getting into one of the deadly sins.

Because of using terms overly broadly.

Yeah, I mean, there's so many different forms of learning.

Yeah.

So many different forms.

You know, I learned my way around the city.

I learned to play chess.

I learned Latin.

I learned to ride a bicycle.

All of those are, you know, very different capabilities.

Yeah.

And if someone, you know, has a, you know, in the old days, people would write a paper

about learning something.

Now the corporate press office puts out a press release about how Company X is leading

the world because they have a system that can...

Yeah, but here's the thing.

Okay.

So what is learning?

When I refer to...

Learning is many things.

But...

It's a suitcase word.

It's a suitcase word, but loosely, there's a dumb system, and over time, it becomes smart.

Well, it becomes less dumb at the thing that it's doing.

Smart is a loaded word.

Yes, less dumb at the thing it's doing.

It gets better performance under some measure, under some set of conditions at that thing.

And most of these learning algorithms, learning systems, fail when you change the conditions

just a little bit in a way that humans don't.

So I was at DeepMind, the AlphaGo had just come out, and I said, what would have happened

if you'd given it a 21 by 21 board instead of a 19 by 19 board?

They said, fail totally.

But a human player would actually be able to play.

And actually, funny enough, if you look at DeepMind's work since then, they're presenting

a lot of algorithms that would do well at the bigger board.

So they're slowly expanding this generalization.

I mean, to me, there's a core element there.

I think it is very surprising to me that even in a constrained game of chess or Go, that

through self play, by a system playing itself, that it can achieve superhuman level performance

through learning alone.

Okay, so you didn't like it when I referred to Donald Mickey's 1961 paper.

There, in the second part of it, which came a year later, they had self play on an electronic

computer at tic tac toe, okay, but it learned to play tic tac toe through self play.

And it learned to play optimally.

What I'm saying is, okay, I have a little bit of a bias, but I find ideas beautiful,

but only when they actually realize the promise.

That's another level of beauty.

For example, what Bezos and Elon Musk are doing with rockets.

We had rockets for a long time, but doing reusable cheap rockets, it's very impressive.

In the same way, I would have not predicted.

First of all, when I started and fell in love with AI, the game of Go was seen to be impossible

to solve.

Okay, so I thought maybe, you know, maybe it'd be possible to maybe have big leaps in

a Moore's law style of way, in computation, I'll be able to solve it.

But I would never have guessed that you can learn your way, however, I mean, in the narrow

sense of learning, learn your way to beat the best people in the world at the game of

Go without human supervision, not studying the game of experts.

Okay, so using a different learning technique, Arthur Samuel in the early 60s, and he was

the first person to use machine learning, had a program that could beat the world champion

at checkers.

And that at the time was considered amazing.

By the way, Arthur Samuel had some fantastic advantages.

Do you want to hear Arthur Samuel's advantages?

Two things.

One, he was at the 1956 AI conference.

I knew Arthur later in life.

He was at Stanford when I was a graduate student there.

He wore a tie and a jacket every day, the rest of us didn't.

Delightful man, delightful man.

It turns out Claude Shannon, in a 1950 Scientific American article, on chess playing, outlined

the learning mechanism that Arthur Samuel used, and they had met in 1956.

I assume there was some communication, but I don't know that for sure.

But Arthur Samuel had been a vacuum tube engineer, getting reliability of vacuum tubes, and then

had overseen the first transistorized computers at IBM.

And in those days, before you shipped a computer, you ran it for a week to get early failures.

So he had this whole farm of computers running random code for hours and hours for each computer.

He had a whole bunch of them.

So he ran his chess learning program with self play on IBM's production line.

He had more computation available to him than anyone else in the world, and then he was

able to produce a chess playing program, I mean a checkers playing program, that could

beat the world champion.

So that's amazing.

The question is, what I mean surprised, I don't just mean it's nice to have that accomplishment,

is there is a stepping towards something that feels more intelligent than before.

Yeah, but that's in your view of the world.

Okay, well let me then, it doesn't mean I'm wrong.

No, no it doesn't.

So the question is, if we keep taking steps like that, how far that takes us?

Are we going to build a better recommender systems?

Are we going to build a better robot?

Or will we solve intelligence?

So, you know, I'm putting my bet on, but still missing a whole lot.

A lot.

And why would I say that?

Well, in these games, they're all, you know, 100% information games, but again, but each

of these systems is a very short description of the current state, which is different from

registering and perception in the world, which gets back to Marovec's paradox.

I'm definitely not saying that chess is somehow harder than perception or any kind of, even

any kind of robotics in the physical world, I definitely think is way harder than the

game of chess.

So I was always much more impressed by the workings of the human mind.

It's incredible.

The human mind is incredible.

I believe that from the very beginning, I wanted to be a psychiatrist for the longest

time.

I always thought that's way more incredible in the game of chess.

I think the game of chess is, I love the Olympics.

It's just another example of us humans picking a task and then agreeing that a million humans

will dedicate their whole life to that task.

And that's the cool thing that the human mind is able to focus on one task and then compete

against each other and achieve like weirdly incredible levels of performance.

That's the aspect of chess that's super cool.

Not that chess in itself is really difficult.

It's like the Fermat's last theorem is not in itself to me that interesting.

The fact that thousands of people have been struggling to solve that particular problem

is fascinating.

So can I tell you my disease in this way?

Sure.

Which actually is closer to what you're saying.

So as a child, I was building various, I called them computers.

They weren't general purpose computers.

Ice cube tray.

The ice cube tray was one.

But I built other machines.

And what I liked to build was machines that could beat adults at a game and the adults

couldn't beat my machine.

Yeah.

So you were like, that's powerful.

That's a way to rebel.

Oh, by the way, when was the first time you built something that outperformed you?

Do you remember?

Well, I knew how it worked.

I was probably nine years old and I built a thing that was a game where you take turns

in taking matches from a pile and either the one who takes the last one or the one who

doesn't take the last one wins.

I forget.

And so it was pretty easy to build that out of wires and nails and little coils that were

like plugging in the number and a few light bulbs.

The one I was proud of, I was 12 when I built a thing out of old telephone switchboard switches

that could always win at tic tac toe.

And that was a much harder circuit to design.

But again, it was no active components.

It was just three position switches, empty, X, zero, O.

And nine of them and a light bulb on which move it wanted next.

And then the human would go and move that.

See, there's magic in that creation.

There was.

Yeah, yeah.

I tend to see magic in robots that like I also think that intelligence is a little bit

overrated.

I think we can have deep connections with robots very soon.

And well, we'll come back to connections for sure.

But I do want to say, I think too many people make the mistake of seeing that magic and

thinking, well, we'll just continue.

But each one of those is a hard fought battle for the next step, the next step.

Yes.

The open question here is, and this is why I'm playing devil's advocate, but I often

do when I read your blog post in my mind because I have like this eternal optimism, is it's

not clear to me.

So I don't do what obviously the journalists do or they give into the hype, but it's not

obvious to me how many steps away we are from a truly transformational understanding of

what it means to build intelligent systems or how to build intelligent systems.

I'm also aware of the whole history of artificial intelligence, which is where your deep grounding

of this is, is there has been an optimism for decades and that optimism, just like reading

old optimism is absurd because people were like, this is, they were saying things are

trivial for decades since the sixties, they're saying everything is true.

Computer vision is trivial, but I think my mind is working crisply enough to where, I

mean, we can dig into if you want.

I'm really surprised by the things DeepMind has done.

I don't think they're so, they're yet close to solving intelligence, but I'm not sure

it's not 10 to 10 years away.

What I'm referring to is interesting to see when the engineering, it takes that idea to

scale and the idea works.

And no, it fools people.

Okay.

Honestly, Rodney, if it was you, me and Demis inside a room, forget the press, forget all

those things, just as a scientist, as a roboticist, that wasn't surprising to you that at scale.

So we're talking about very large now, okay, let's pick one.

That's the most surprising to you.

Okay.

Please don't yell at me.

GPT three, okay.

Hold on, hold on, I was going to say, okay, alpha zero, alpha go, alpha go, zero, alpha

zero, and then alpha fold one and two.

So do any of these kind of have this core of, forget usefulness or application and so

on, which you could argue for alpha fold, like, as a scientist, was those surprising

to you that it worked as well as it did?

Okay, so if we're going to make the distinction between surprise and usefulness, and I have

to explain this, I would say alpha fold, and one of the problems at the moment with alpha

fold is, you know, it gets a lot of them right, which is a surprise to me, because they're

a really complex thing, but you don't know which ones it gets right, which then is a

bit of a problem.

Now they've come out with a recent...

You mean the structure of the proteins, it gets a lot of those right.

Yeah, it's a surprising number of them right, it's been a really hard problem.

So that was a surprise how many it gets right.

So far, the usefulness is limited, because you don't know which ones are right or not,

and now they've come out with a thing in the last few weeks, which is trying to get a useful

tool out of it, and they may well do it.

In that sense, at least alpha fold is different, because your alpha fold tool is different,

because now it's producing data sets that are actually, you know, potentially revolutionizing

competition biology, like they will actually help a lot of people, but...

You would say potentially revolutionizing, we don't know yet, but yeah.

That's true, yeah.

But they're, you know, but I got you.

I mean, this is...

Okay, so you know what, this is gonna be so fun, so let's go right into it.

Speaking of robots that operate in the real world, let's talk about self driving cars.

Oh, okay.

Okay, because you have built robotics companies, you're one of the greatest roboticists in

history, and that's not just in the space of ideas, we'll also probably talk about that,

but in the actual building and execution of businesses that make robots that are useful

for people and that actually work in the real world and make money.

You also sometimes are critical of Mr. Elon Musk, or let's more specifically focus on

this particular technology, which is autopilot inside Teslas.

What are your thoughts about Tesla autopilot, or more generally vision based machine learning

approach to semi autonomous driving?

These are robots, they're being used in the real world by hundreds of thousands of people,

and if you want to go there, I can go there, but that's not too much, which they're...

Let's say they're on par safety wise as humans currently, meaning human alone versus human

plus robot.

Okay, so first let me say I really like the car I came in here today.

Which is?

2021 model, Mercedes E450.

I am impressed by the machine vision, sonar, other things.

I'm impressed by what it can do.

I'm really impressed with many aspects of it.

It's able to stay in lane, is it?

Oh yeah, it does the lane stuff.

It's looking on either side of me, it's telling me about nearby cars.

For blind spots and so on.

Yeah, when I'm going in close to something in the park, I get this beautiful, gorgeous,

top down view of the world.

I am impressed up the wazoo of how registered and metrical that is.

So it's like multiple cameras and it's all ready to go to produce the 360 view kind of

thing?

360 view, it's synthesized so it's above the car, and it is unbelievable.

I got this car in January, it's the longest I've ever owned a car without digging it.

So it's better than me.

Me and it together are better.

So I'm not saying technology's bad or not useful, but here's my point.

Yes, it's a replay of the same movie.

Okay, so maybe you've seen me ask this question before.

But when did the first car go over 55 miles an hour for over 10 miles on a public freeway

with other traffic around driving completely autonomously?

When did that happen?

Was it CMU in the 80s or something?

It was a long time ago.

It was actually in 1987 in Munich at the Bundeswehr.

So they had it running in 1987.

When do you think, and Elon has said he's going to do this, when do you think we'll

have the first car drive coast to coast in the US, hands off the wheel, feet off the

pedals, coast to coast?

As far as I know, a few people have claimed to do it.

1995, that was Carnegie Mellon.

I didn't know, but oh, that was the, they didn't claim, did they claim 100%?

Not 100%, not 100%.

And then there's a few marketing people who have claimed 100% since then.

My point is that, you know, what I see happening again is someone sees a demo and they overgeneralize

and say, we must be almost there.

But we've been working on it for 35 years.

So that's demos.

But this is going to take us back to the same conversation with AlphaZero.

Are you not, okay, I'll just say what I am because I thought, okay, when I first started

interacting with the Mobileye implementation of Tesla Autopilot, I've driven a lot of car,

you know, I've been in Google self driving car since the beginning.

I thought there was no way before I sat and used Mobileye, I thought they're just knowing

computer vision.

I thought there's no way it could work as well as it was working.

So my model of the limits of computer vision was way more limited than the actual implementation

of Mobileye.

I was so that's one example.

I was really surprised.

It's like, wow, that was that was incredible.

The second surprise came when Tesla threw away Mobileye and started from scratch.

I thought there's no way they can catch up to Mobileye.

I thought what Mobileye was doing was kind of incredible, like the amount of work and

the annotation.

Yeah, well, Mobileye was started by Amnon Shashua and used a lot of traditional, you

know, hard fought computer vision techniques.

But they also did a lot of good sort of like non research stuff, like actual like just

good, like what you do to make a successful product, right?

Scale, all that kind of stuff.

And so I was very surprised when they from scratch were able to catch up to that.

That's very impressive.

And I've talked to a lot of engineers that was involved.

This is that was impressive.

That was impressive.

And the recent progress, especially under the involvement of Andrej Karpathy, what they

were what they're doing with the data engine, which is converting into the driving task

into these multiple tasks and then doing this edge case discovery when they're pulling back

like the level of engineering made me rethink what's possible.

I don't I still, you know, I don't know to that intensity, but I always thought it was

very difficult to solve autonomous driving with all the sensors, with all the computation.

I just thought it's a very difficult problem.

But I've been continuously surprised how much you can engineer.

First of all, the data acquisition problem, because I thought, you know, just because

I worked with a lot of car companies and they're they're so a little a little bit old school

to where I didn't think they could do this at scale like AWS style data collection.

So when Tesla was able to do that, I started to think, OK, so what are the limits of this?

I still believe that driver like sensing and the interaction with the driver and like studying

the human factor psychology problem is essential.

It's it's always going to be there.

It's always going to be there, even with fully autonomous driving.

But I've been surprised what is the limit, especially a vision based alone, how far that

can take us.

So that's my levels of surprise now.

OK, can you explain in the same way you said, like Alpha Zero, that's a homework problem

that's scaled large in its chest, like who cares?

Go with here's actual people using an actual car and driving.

Many of them drive more than half their miles using the system.

Right.

So, yeah, they're doing well with with pure vision for your vision.

Yeah.

And, you know, and now no radar, which is I suspect that can't go all the way.

And one reason is without without new cameras that have a dynamic range closer to the human

eye, because human eye has incredible dynamic range.

And we make use of that dynamic range in its 11 orders of magnitude or some crazy number

like that.

The cameras don't have that, which is why you see the the the bad cases where the sun

on a white thing and it blinds it in a way it wouldn't blind the person.

I think there's a bunch of things to think about before you say this is so good, it's

just going to work.

OK, and I'll come at it from multiple angles.

And I know you've got a lot of time.

Yeah.

OK, let's let's I have thought about these things.

Yeah, I know.

You've been writing a lot of great blog posts about it for a while before Tesla had autopilot.

Right.

So you've been thinking about autonomous driving for a while from every angle.

So so a few things, you know, in the US, I think that the death rate for autonomous driving

death rate from motor vehicle accidents is about thirty five thousand a year,

which is an outrageous number, not outrageous compared to covid deaths.

But, you know, there is no rationality.

And that's part of the thing people have said.

Engineers say to me, well, if we cut down the number of deaths by 10 percent by having

autonomous driving, that's going to be great.

Everyone will love it.

And my prediction is that if autonomous vehicles kill more than 10 people a year, they'll be

screaming and hollering, even though thirty five thousand people a year have been killed

by human drivers.

It's not rational.

It's a different set of expectations.

And that will probably continue.

So there's that aspect of it.

The other aspect of it is that when we introduce new technology, we often change the rules

of the game.

So when we introduced cars first into our daily lives, we completely rebuilt our cities

and we changed all the laws.

Yeah, jaywalking was not an offense that was pushed by the car companies so that people

would stay off the road so there wouldn't be deaths from pedestrians getting hit.

We completely changed the structure of our cities and had these foul smelling things

everywhere around us.

And now you see pushback in cities like Barcelona is really trying to exclude cars, et cetera.

So I think that to get to self driving, we will, large adoption, it's not going to be

just take the current situation, take out the driver and put the same car doing the

same stuff because the end case is too many.

Here's an interesting question.

How many fully autonomous train systems do we have in the U.S.?

I mean, do you count them as fully autonomous?

I don't know because they're usually as a driver, but they're kind of autonomous, right?

No, let's get rid of the driver.

Okay.

I don't know.

It's either 15 or 16.

Most of them are in airports.

There's a few that are fully autonomous.

Seven are in airports, there's a few that go about five, two that go about five kilometers

out of airports.

When is the first fully autonomous train system for mass transit expected to operate fully

autonomously with no driver in a U.S.

City?

It's expected to operate in 2017 in Honolulu.

Oh, wow.

It's delayed, but they will get there.

BART, by the way, was originally going to be autonomous here in the Bay Area.

I mean, they're all very close to fully autonomous, right?

Yeah, but getting that close is the thing.

And I've often gone on a fully autonomous train in Japan, one that goes out to that

fake island in the middle of Tokyo Bay.

I forget the name of that.

And what do you see when you look at that?

What do you see when you go to a fully autonomous train in an airport?

It's not like regular trains.

At every station, there's a double set of doors so that there's a door of the train

and there's a door off the platform.

And this is really visible in this Japanese one because it goes out in amongst buildings.

The whole track is built so that people can't climb onto it.

Yeah.

So there's an engineering that then makes the system safe and makes them acceptable.

I think we'll see similar sorts of things happen in the U.S.

What surprised me, I thought, wrongly, that we would have special purpose lanes on 101

in the Bay Area, the leftmost lane, so that it would be normal for Teslas or other cars

to move into that lane and then say, okay, now it's autonomous and have that dedicated lane.

I was expecting movement to that.

Five years ago, I was expecting we'd have a lot more movement towards that.

We haven't.

And it may be because Tesla's been overpromising by saying this, calling their system fully

self driving, I think they may have been gotten there quicker by collaborating to change the

infrastructure.

This is one of the problems with long haul trucking being autonomous.

I think it makes sense on freeways at night for the trucks to go autonomously, but then

is that how do you get onto and off of the freeway?

What sort of infrastructure do you need for that?

Do you need to have the human in there to do that or can you get rid of the human?

So I think there's ways to get there, but it's an infrastructure argument because the

long tail of cases is very long and the acceptance of it will not be at the same level as human

drivers.

So I'm with you still, and I was with you for a long time, but I am surprised how well

how many edge cases of machine learning and vision based methods can cover.

This is what I'm trying to get at is I think there's something fundamentally different

with vision based methods and Tesla Autopilot and any company that's trying to do the same.

Okay, well, I'm not going to argue with you because, you know, we're speculating.

Yes, but, you know, my gut feeling tells me it's going to be things will speed up when

there is engineering of the environment because that's what happened with every other technology.

I'm a bit, I don't know about you, but I'm a bit cynical that infrastructure is going

to rely on government to help out in these cases.

If you just look at infrastructure in all domains, it's just a government always drags

behind on infrastructure.

There's like there's so many just well in this country in the future.

Sorry.

Yes, in this country.

And of course, there's many, many countries that are actually much worse on infrastructure.

Oh, yes, many of the much worse and there's some that are much worse.

You know, like high speed rail, the other countries are much better.

I guess my question is, like, which is at the core of what I was trying to think through

here and ask is like, how hard is the driving problem as it currently stands?

So you mentioned, like, we don't want to just take the human out and duplicate whatever

the human was doing.

But if we were to try to do that, what, how hard is that problem?

Because I used to think is way harder.

Like, I used to think it's with vision alone, it would be three decades, four decades.

Okay, so I don't know the answer to this thing I'm about to pose, but I do notice that on

Highway 280 here in the Bay Area, which largely has concrete surface rather than blacktop

surface, the white lines that are painted there now have black boundaries around them.

And my lane drift system in my car would not work without those black boundaries.

Interesting.

So I don't know whether they started doing it to help the lane drift, whether it is an

instance of infrastructure following the technology, but my car would not perform as well as the

lane, my car would not perform as well without that change in the way they paint the line.

Unfortunately, really good lane keeping is not as valuable.

Like, it's orders of magnitude more valuable to have a fully autonomous system.

Like, yeah, but for me, lane keeping is really helpful because I'm more healthy at it.

But you wouldn't pay 10 times.

Like, the problem is there's not financial, like, it doesn't make sense to revamp the

infrastructure to make lane keeping easier.

It does make sense to revamp the infrastructure.

If you have a large fleet of autonomous vehicles, now you change what it means to own cars,

you change the nature of transportation.

But for that, you need autonomous vehicles.

Let me ask you about Waymo then.

I've gotten a bunch of chances to ride in a Waymo self driving car.

And they're, I don't know if you'd call them self driving, but.

Well, I mean, I rode in one before they were called Waymo when I was still at X.

So there's currently, there's a big leap, another surprising leap I didn't think would

happen, which is they have no driver currently.

Yeah, in Chandler.

In Chandler, Arizona.

And I think they're thinking of doing that in Austin as well.

But they're expanding.

Although, you know, and I do an annual checkup on this.

So as of late last year, they were aiming for hundreds of rides a week, not thousands.

And there is no one in the car, but there's certainly safety people in the loop.

And it's not clear how many, you know, what the ratio of cars to safety people is.

It wasn't, obviously, they're not 100% transparent about this.

None of them are 100% transparent.

They're very untransparent.

But at least the way they're, I don't want to make definitively, but they're saying

there's no teleoperation.

So like, they're, I mean, okay.

And that sort of fits with YouTube videos I've seen of people being trapped in the car

by a red cone on the street.

And they do have rescue vehicles that come, and then a person gets in and drives it.

Yeah.

But isn't it incredible to you, it was to me, to get in a car with no driver and watch

the steering wheel turn, like for somebody who has been studying, at least certainly

the human side of autonomous vehicles for many years, and you've been doing it for way

longer, like it was incredible to me that this was actually could happen.

I don't care if that scale is 100 cars.

This is not a demo.

This is not, this is me as a regular human.

The argument I have is that people make interpolations from that.

Interpolations.

That, you know, it's here, it's done.

You know, it's just, you know, we've solved it.

No, we haven't yet.

And that's my argument.

Okay.

So I'd like to go to, you keep a list of predictions on your amazing blog post.

It'd be fun to go through them.

But before then, let me ask you about this.

You have a harshness to you sometimes in your criticisms of what is perceived as hype.

And so like, because people extrapolate, like you said, and they kind of buy into the hype

and then they kind of start to think that the technology is way better than it is.

But let me ask you maybe a difficult question.

Sure.

Do you think if you look at history of progress, don't you think to achieve the quote impossible,

you have to believe that it's possible?

Oh, absolutely.

Yeah.

Look, his two great runs, great, unbelievable, 1903, first human power, human, you know,

human, you know, heavier than their flight.

Yeah.

1969, we land on the moon.

That's 66 years.

I'm 66 years old in my lifetime, that span of my lifetime, barely, you know, flying,

I don't know what it was, 50 feet, the length of the first flight or something to landing

on the moon.

Unbelievable.

Fantastic.

But that requires, by the way, one of the Wright brothers, both of them, but one of

them didn't believe it's even possible like a year before.

Right.

So, like, not just possible soon, but like ever.

So, you know.

How important is it to believe and be optimistic is what I guess.

Oh, yeah, it is important.

It's when it goes crazy, when I, you know, you said that, what was the word you used

for my bad?

Harshness.

Harshness.

Yes.

I just get so frustrated.

Yes.

When people make these leaps and tell me that I'm, that I don't understand, you know, yeah.

There's just from iRobot, which I was co founder of.

Yeah.

I don't know the exact numbers now because I haven't, it's 10 years since I stepped

off the board, but I believe it's well over 30 million robots cleaning houses from that

one company.

And now there's lots of other companies.

Yes.

Was that a crazy idea that we had to believe in 2002 when we released it?

Yeah, that was, we had, we had to, you know, believe that it could be done.

Let me ask you about this.

So iRobot, one of the greatest robotics companies ever in terms of creating a robot that actually

works in the real world, probably the greatest robotics company ever.

You were the co founder of it.

If, if the Rodney Brooks of today talked to the Rodney of back then, what would you tell

him?

Cause I have a sense that would you pat him on the back and say, well, you're doing is

going to fail, but go at it anyway.

That's what I'm referring to with the harshness.

You've accomplished an incredible thing there.

One of the several things we'll talk about was, you know, you know, you know, you've

done several things we'll talk about.

Well, like that's what I'm trying to get at that line.

No, it's, it's when my harshness is reserved for people who are not doing it, who claim

it's just, well, this shows that it's just going to happen.

But here, here's the thing.

This shows.

But you have that harshness for Elon too.

And no, no, it's a different harshness.

No, it's, it's a different argument with Elon.

I think SpaceX is an amazing company.

On the other hand, you know, I, in one of my blog posts, I said, what's easy and what's

hard.

I said, yeah, space X vertical landing rockets.

It had been done before.

Grid fins had been done since the sixties.

Every Soyuz has them.

Reusable space DCX reuse those rockets that landed vertically.

There's a whole insurance industry in place for rocket launches.

There are all sorts of infrastructure that was doable.

It took a great entrepreneur, a great personal expense.

He almost drove himself, you know, bankrupt doing it, a great belief to do it.

Whereas Hyperloop, there's a whole bunch more stuff that's never been thought about and

never been demonstrated.

So my estimation is Hyperloop is a long, long, long, a lot further off.

But, and if I've got a criticism of, of, of Elon, it's that he doesn't make distinctions

between when the technology's coming along and ready.

And then he'll go off and mouth off about other things, which then people go and compete

about and try and do.

And so this is where I, I, I, I understand what you're saying.

I tend to draw a different distinction.

I, I have a similar kind of harshness towards people who are not telling the truth, who

are basically fabricating stuff to make money or to, well, he believes what he says.

I just think that's a very important difference because I think in order to fly, in order

to get to the moon, you have to believe even when most people tell you you're wrong and

most likely you're wrong, but sometimes you're right.

I mean, that's the same thing I have with Tesla autopilot.

I think that's an interesting one.

I was, especially when I was at MIT and just the entire human factors in the robotics community

were very negative towards Elon.

It was very interesting for me to observe colleagues at MIT.

I wasn't sure what to make of that.

That was very upsetting to me because I understood where that, where that's coming from.

And I agreed with them and I kind of almost felt the same thing in the beginning until

I kind of opened my eyes and realized there's a lot of interesting ideas here that might

be over hype.

You know, if you focus yourself on the idea that you shouldn't call a system full self

driving when it's obviously not autonomous, fully autonomous, you're going to miss the

magic.

Oh, yeah, you are going to miss the magic.

But at the same time, there are people who buy it, literally pay money for it and take

those words as given.

So it's, but I haven't.

So that I take words as given is one thing.

I haven't actually seen people that use autopilot that believe that the behavior is really important,

like the actual action.

So like, this is to push back on the very thing that you're frustrated about, which

is like journalists and general people buying all the hype and going out in the same way.

I think there's a lot of hype about the negatives of this, too, that people are buying without

using people use the way this is what this was.

This opened my eyes.

Actually, the way people use a product is very different than the way they talk about

it.

This is true with robotics, with everything.

Everybody has dreams of how a particular product might be used or so on.

And then when it meets reality, there's a lot of fear of robotics, for example, that

robots are somehow dangerous and all those kinds of things.

But when you actually have robots in your life, whether it's in the factory or in the

home, making your life better, that's going to be that's way different.

Your perceptions of it are going to be way different.

And so my just tension was like, here's an innovator.

Supercruise from Cadillac was super interesting, too.

That's a really interesting system.

We should be excited by those innovations.

OK, so can I tell you something that's really annoyed me recently?

It's really annoyed me that the press and friends of mine on Facebook are going, these

billionaires and their space games, why are they doing that?

And that really, really pisses me off.

I must say, I applaud that.

I applaud it.

It's the taking and not necessarily the people who are doing the things, but, you know, that

I keep having to push back against unrealistic expectations when these things can become

real.

Yeah, I this was interesting on because there's been a particular focus for me is autonomous

driving, Elon's prediction of when certain milestones will be hit.

There's several things to be said there that I always I thought about, because whenever

you said them, it was obvious that's not going to me as a person that kind of not inside

the system is obvious.

It's unlikely to hit those.

There's two comments I want to make.

One, he legitimately believes it.

And two, much more importantly, I think that having ambitious deadlines drives people to

do the best work of their life, even when the odds of those deadlines are very low.

To a point, and I'm not talking about anyone here, I'm just saying.

So there's a line there, right?

You have to have a line because you overextend and it's demoralizing.

It's demoralizing, but I will say that there's an additional thing here that those words

also drive the stock market.

And we have because of the way that rich people in the past have manipulated the rubes through

investment, we have developed laws about what you're allowed to say.

And you know, there's an area here which is I tend to be maybe I'm naive, but I tend to

believe that like engineers, innovators, people like that, they're not they're my they don't

think like that, like manipulating the price of the stock price.

But it's possible that I'm I'm certain it's possible that I'm wrong.

It's a very cynical view of the world because I think most people that run companies, especially

original founders, they yeah, I'm not saying that's the intent.

I'm saying it's eventually it's kind of you you you you fall into that kind of behavior

pattern.

I don't know.

I tend to I wasn't saying I wasn't saying it's falling into that intent.

It's just you also have to protect investors in this environment.

In this market.

Yeah.

OK, so you have first of all, you have an amazing blog that people should check out.

But you also have this in that blog, a set of predictions.

Such a cool idea.

I don't know how long ago you started, like three, four years ago.

It was January 1st, 2018.

18.

And I made these predictions and I said that every January 1st, I was going to check back

on how my predictions.

That's such a great thought experiment.

For 32 years.

Oh, you said 32 years.

I said 32 years because it's still that'll be January 1st, 2050.

I'll be I will just turn ninety.

Five, you know, and so people know that your predictions, at least for now, are in the

space of artificial intelligence.

Yeah, I didn't say I was going to make new predictions.

I was just going to measure this set of predictions that I made because I was sort of I was sort

of annoyed that everyone could make predictions.

They didn't come true and everyone forgot.

So I should hold myself to a high standard.

Yeah, but also just putting years and like date ranges on things.

It's a good thought exercise.

Yeah, like and like reasoning your thoughts out.

And so the topics are artificial intelligence, autonomous vehicles and space.

Yeah.

I was wondering if we could just go through some that stand out maybe from memory.

I can just mention to you some.

Let's talk about self driving cars, like some predictions that you're particularly proud

of or are particularly interesting from flying cars to the other element here is like how

widespread the location where the deployment of the autonomous vehicles is.

And there's also just a few fun ones.

Is there something that jumps to mind that you remember from the predictions?

Well, I think I did put in there that there would be a dedicated self driving lane on

101 by some year, and I think I was over optimistic on that one.

Yeah, actually.

Yeah, I actually do remember that.

But you I think you were mentioning like difficulties at different cities.

Yeah.

Cambridge, Massachusetts, I think was an example.

Yeah, like in Cambridge Port, you know, I lived in Cambridge Port for a number of years

and you know, the roads are narrow and getting getting anywhere as a human driver is incredibly

frustrating when you start to put and people drive the wrong way on one way streets there.

It's just your prediction was driverless taxi services operating on all streets in

Cambridge Port, Massachusetts in 2035.

Yeah.

And that may have been too optimistic.

You think so?

You know, I've gotten a little more pessimistic since I made these internally on some of these

things.

So what can you put a year to a major milestone of deployment of a taxi service in in a few

major cities like something where you feel like autonomous vehicles are here.

So let's let's take the grid streets of San Francisco north of market.

Okay.

Okay.

Relatively benign environment, the streets are wide, the major problem is delivery trucks

stopping everywhere, which made things more complicated.

Taxi system there with somewhat designated pickup and drop offs, unlike with Uber and

Lyft, where you can sort of get to any place and the drivers will figure out how to get

in there.

We're still a few years away.

I, you know, I live in that area.

So I see, you know, the self driving car companies cars, multiple multiple ones every day.

Now if they're cruise, Zooks less often, Waymo all the time, different and different ones

come and go.

And there's always a driver.

There's always a driver at the moment, although I have noticed that sometimes the driver does

not have the authority to take over without talking to the home office, because they will

sit there waiting for a long time, and clearly something's going on where the home office

is making a decision.

So they're, you know, and, and so you can see whether they've got their hands on the

wheel or not.

And, and it's the incident resolution time that tells you, gives you some clues.

So what year do you think, what's your intuition?

What date range are you currently thinking San Francisco would be?

Are you currently thinking San Francisco would be autonomous taxi service from any point

A to any point B without a driver?

Are you still, are you thinking 10 years from now, 20 years from now, 30 years from now?

Certainly not 10 years from now.

It's going to be longer.

If you're allowed to go south of market way longer.

And unless it's reengineering of roads.

By the way, what's the biggest challenge?

You mentioned a few.

Is it, is it the delivery trucks?

Is it the edge cases, the computer perception, well, here's a case that I saw outside my

house a few weeks ago, about 8pm on a Friday night, it was getting dark, it was before

the solstice.

It was a cruise vehicle come down the hill, turned right and stopped dead, covering the

crosswalk.

Why did it stop dead?

Because there was a human just two feet from it.

Now, I just glanced, I knew what was happening.

The human was a woman was at the door of her car trying to unlock it with one of those

things that, you know, when you don't have a key.

That car thought, oh, she could jump out in front of me any second.

As a human, I could tell, no, she's not going to jump out.

She's busy trying to unlock her.

She's lost her keys.

She's trying to get in the car.

And it stayed there for, until I got bored.

And so the human driver in there did not take over.

But here's the kicker to me.

A guy comes down the hill with a stroller, I assume there's a baby in there, and now

the crosswalk's blocked by this cruise vehicle.

What's he going to do?

Cleverly, I think, he decided not to go in front of the car.

But he had to go behind it.

He had to get off the crosswalk, out into the intersection, to push his baby around

this car, which was stopped there.

And no human driver would have stopped there for that length of time.

They would have got out and out of the way.

And that's another one of my pet peeves, that safety is being compromised for individuals

who didn't sign up for having this happen in their neighborhood.

Now you can say that's an edge case, but...

Yeah, well, I'm in general not a fan of anecdotal evidence for stuff like this is one of my

biggest problems with the discussion of autonomous vehicles in general, people that criticize

them or support them are using edge cases, are using anecdotal evidence, but I got you.

Your question is, when is it going to happen in San Francisco?

I say not soon, but it's going to be one of them.

But where it is going to happen is in limited domains, campuses of various sorts, gated

communities where the other drivers are not arbitrary people.

They're people who know about these things, they've been warned about them, and at velocities

where it's always safe to stop dead.

You can't do that on the freeway.

That I think we're going to start to see, and they may not be shaped like current cars,

they may be things like May Mobility has those things and various companies have these.

Yeah, I wonder if that's a compelling experience.

To me, it's not just about automation, it's about creating a product that makes your...

It's not just cheaper, but it's fun to ride.

One of the least fun things is for a car that stops and waits.

There's something deeply frustrating for us humans for the rest of the world to take advantage

of us as we wait.

But think about not you as the customer, but someone who's in their 80s in a retirement

village whose kids have said, you're not driving anymore, and this gives you the freedom to

go to the market.

That's a hugely beneficial thing, but it's a very few orders of magnitude less impact

on the world.

It's just a few people in a small community using cars as opposed to the entirety of the

world.

I like that the first time that a car equipped with some version of a solution to the trolley

problem is...

What's NIML stand for?

Not in my life.

Not in my life.

I define my lifetime as up to 2050.

You know, I ask you, when have you had to decide which person shall I kill?

No, you put the brakes on and you break as hard as you can.

You're not making that decision.

I do think autonomous vehicles or semi autonomous vehicles do need to solve the whole pedestrian

problem that has elements of the trolley problem within it, but it's not...

Yeah, well, and I talk about it in one of the articles or blog posts that I wrote, and

people have told me, one of my coworkers has told me he does this.

He tortures autonomously driven vehicles and pedestrians will torture them.

Now, once they realize that putting one foot off the curb makes the car think that they

might walk into the road, teenagers will be doing that all the time.

I, by the way, one of my, and this is a whole nother discussion, because my main interest

with robotics is HRI, human robot interaction.

I believe that robots that interact with humans will have to push back.

Like they can't just be bullied because that creates a very uncompelling experience for

the humans.

Yeah, well, you know, Waymo, before it was called Waymo, discovered that, you know, they

had to do that at four way intersections.

They had to nudge forward to give the cue that they were going to go, because otherwise

the other drivers would just beat them all the time.

So you cofounded iRobot, as we mentioned, one of the most successful robotics companies

ever.

What are you most proud of with that company and the approach you took to robotics?

Well, there's something I'm quite proud of there, which may be a surprise, but, you know,

I was still on the board when this happened, it was March 2011, and we sent robots to Japan

and they were used to help shut down the Fukushima Daiichi nuclear power plant, which was, everything

was, I've been there since, I was there in 2014, and the robots, some of the robots were

still there.

I was proud that we were able to do that.

Why were we able to do that?

And, you know, people have said, well, you know, Japan is so good at robotics.

It was because we had had about 6,500 robots deployed in Iraq and Afghanistan, teleopt,

but with intelligence, dealing with roadside bombs.

So we had, it was at that time, nine years of in field experience with the robots in

harsh conditions, whereas the Japanese robots, which were, you know, getting, this goes back

to what annoys me so much, getting all the hype, look at that, look at that Honda robot,

it can walk, wow, the future's here, couldn't do a thing because they weren't deployed,

but we had deployed in really harsh conditions for a long time, and so we're able to do

something very positive in a very bad situation.

What about just the simple, and for people who don't know, one of the things that iRobot

has created is the Roomba vacuum cleaner.

What about the simple robot that, that is the Roomba, quote unquote, simple, that's

deployed in tens of millions of, in tens of millions of homes?

What do you think about that?

Well, I make the joke that I started out life as a pure mathematician and turned into a

vacuum cleaner salesman, so if you're going to be an entrepreneur, be ready for, be ready

to do anything, but I was, you know, there was a, there was a wacky lawsuit that I got

opposed for not too many years ago, and I was the only one who had emailed from the

1990s, and no one in the company had it, so I went and went through my email, and it

reminded me of, you know, the joy of what we were doing, and what was I doing?

What was I doing at the time we were building, building the Roomba?

One of the things was we had this, you know, incredibly tight budget because we wanted

to put it on the shelves at $200.

There was another home cleaning robot at the time, it was the Electrolux Trilobite, which

sold for 2,000 euros, and to us that was not going to be a consumer product, so we had

reason to believe that $200 was a, was a thing that people would buy at.

That was our aim, but that meant we had, you know, that's on the shelf making profit.

That means the cost of goods has to be minimal, so I find all these emails of me going, you

know, I'd be in Taipei for a MIT meeting, and I'd stay a few extra days and go down

to Hsinchu and talk to these little tiny companies, lots of little tiny companies outside of TSMC,

Taiwan Semiconductor Manufacturing Corporation, which let all these little companies be fabulous.

They didn't have to have their own fab so they could innovate, and they were building,

their innovations were to build, strip down 6802s, 6802 was what was in an Apple I, get

rid of half the silicon and still have it be viable, and I'd previously got some of

those for some earlier failed products of iRobot, and that was in Hong Kong going to

all these companies that built, you know, they weren't gaming in the current sense,

there were these handheld games that you would play, or birthday cards, because we had about

a 50 cent budget for computation, so I'm trekking from place to place looking at their chips,

looking at what they'd removed, ah, their interrupt handling is too weak for a general

purpose, so I was going deep technical detail, and then I found this one from a company called

Winbond, which had, and I'd forgotten it had this much RAM, it had 512 bytes of RAM,

and it was in our budget, and it had all the capabilities we needed.

Yeah, and you were excited.

Yeah, and I was reading all these emails, Colin, I found this, so.

Did you think, did you ever think that you guys could be so successful?

Like, eventually this company would be so successful, could you possibly have imagined?

No, we never did think that.

We'd had 14 failed business models up to 2002, and then we had two winners the same year.

No, and then, you know, we, I remember the board, because by this time we had some venture

capital in, the board went along with us building some robots for, you know, aiming at the Christmas

2002 market, and we went three times over what they authorized and built 70,000 of them,

and sold them all in that first, because we released on September 18th, and they were

all sold by Christmas.

So it was, so we were gutsy, but.

But yeah, you didn't think this will take over the world.

Well, this is, so a lot of amazing robotics companies have gone under over the past few

decades.

Why do you think it's so damn hard to run a successful robotics company?

There's a few things.

One is expectations of capabilities by the founders that are off base.

The founders, not the consumer, the founders.

Yeah, expectations of what can be delivered.

Sure.

Mispricing, and what a customer thinks is a valid price, is not rational, necessarily.

Yeah.

And expectations of customers, and just the sheer hardness of getting people to adopt a

new technology.

And I've suffered from all three of these, you know.

I've had more failures than successes, in terms of companies.

I've suffered from all three.

So, do you think one day there will be a robotics company, and by robotics company, I mean, where

your primary source of income is from robots, that will be a trillion plus dollar company?

And if so, what would that company do?

I can't, you know, because I'm still starting robot companies.

Yeah.

I'm not making any such predictions in my own mind.

I'm not thinking about a trillion dollar company.

And by the way, I don't think, you know, in the 90s, anyone was thinking that Apple would

ever be a trillion dollar company.

So, these are, these are, you know, these are, you know, these are, you know, these

would be a trillion dollar company, so these are, these are very hard to predict.

But, sorry to interrupt, but don't you, because I kind of have a vision in a small way, and

it's a big vision in a small way, that I see that there would be robots in the home,

at scale, like Roomba, but more.

And that's trillion dollar.

Right.

And I think there's a real market pull for them because of the demographic inversion,

you know, who's going to do all the stuff for the older people?

There's too many, you know, I'm leading here.

There's going to be too many of us.

But we don't have capable enough robots to make that economic argument at this point.

Do I expect that that will happen?

Yes, I expect it will happen.

But I got to tell you, we introduced the Roomba in 2002, and I stayed another

nine years.

We were always trying to find what the next home robot would be, and still today, the

primary product of 20 years late, almost 20 years later, 19 years later, the primary product

is still the Roomba.

So iRobot hasn't found the next one.

Do you think it's possible for one person in the garage to build it versus, like, Google

launching Google self driving car that turns into Waymo?

Do you think this is almost like what it takes to build a successful robotics company?

Do you think it's possible to go from the ground up, or is it just too much capital

investment?

Yeah, so it's very hard to get there without a lot of capital.

And we're starting to see, you know, fair chunks of capital for some robotics companies.

You know, Series B's, I saw one yesterday for $80 million, I think it was, for Covariant.

But it can take real money to get into these things, and you may fail along the way.

I've certainly failed at Rethink Robotics, and we lost $150 million in capital there.

So, okay, so Rethink Robotics is another amazing robotics company you cofounded.

So what was the vision there?

What was the dream?

And what are you most proud of with Rethink Robotics?

I'm most proud of the fact that we got robots out of the cage in factories that were safe,

absolutely safe, for people and robots to be next to each other.

So these are robotic arms.

Robotic arms.

Able to pick up stuff and interact with humans.

Yeah, and that humans could retask them without writing code.

And now that's sort of become an expectation for a lot of other little companies and big

companies, our advertising they're doing.

That's both an interface problem and also a safety problem.

Yeah, yeah.

So I'm most proud of that.

I completely, I let myself be talked out of what I wanted to do.

And, you know, you always got, you know, I can't replay the tape.

I can't replay it.

Maybe, you know, if I'd been stronger on, and I remember the day, I remember the exact

meeting.

Can you take me through that meeting?

Yeah.

So I'd said that I'd set as a target for the company that we were going to build $3,000

robots with force feedback that was safe for people to be around.

Wow.

That was my goal.

And we built, so we started in 2008, and we had prototypes built of plastic, plastic

gearboxes, and at a $3,000, you know, lifetime, or $3,000, I was saying, we're going to go

after not the people who already have robot arms in factories, the people who would never

have a robot arm.

We're going to go after a different market.

So we don't have to meet their expectations.

And so we're going to build it out of plastic.

It doesn't have to have a $35,000 lifetime.

It's going to be so cheap that it's OpEx, not CapEx.

And so we had a prototype that worked reasonably well, but the control engineers were complaining

about these plastic gearboxes with a beautiful little planetary gearbox that we could use

something called series elastic actuators.

We embedded them in there.

We could measure forces.

We knew when we hit something, et cetera.

The control engineers were saying, yeah, but there's this torque ripple because these plastic

gears, they're not great gears, and there's this ripple, and trying to do force control

around this ripple is so hard.

And I'm not going to name names, but I remember one of the mechanical engineers saying, we'll

just build a metal gearbox with spur gears, and it'll take six weeks.

We'll be done.

Problem solved.

Two years later, we got the spur gearbox working.

We cost reduced it every possible way we could, but now the price went up too.

And then the CEO at the time said, well, we have to have two arms, not one arm.

So our first robot product, Baxter, now cost $25,000, and the only people who were going

to look at that were people who had arms in factories because that was somewhat cheaper

for two arms than arms in factories.

But they were used to 0.1 millimeter reproducibility of motion and certain velocities, and I kept

thinking, but that's not what we're giving you.