0:00

Hi listeners and welcome back to

0:02

No Priors. Today we're joined by Kyle

0:04

Vote, a serial entrepreneur who has

0:06

helped build some of the

0:09

most influential tech companies. He

0:11

co-founded Twitch, shaping live streaming,

0:13

Cruise, the Autonomous Vehicle Company,

0:15

acquired by GM for a

0:17

billion dollars, and now Kyle's

0:19

launched the bot company, a

0:21

start focused on building consumer

0:24

robots. Kyle, welcome to No

0:26

Priors. Awesome, let's get going. Obviously,

0:28

you've done a variety of different

0:30

things over time. Everything from co-founding

0:32

Twitch, you started a cruise, you're

0:34

not working on a new startup.

0:36

Can you tell us a little

0:38

bit more about your cruise experience?

0:40

Because I think that whole

0:42

era was incredibly formative for

0:44

everything that's happening that's happening

0:46

that's happening today in AI. And we'd

0:49

just love to get your perspective on

0:51

why Star Cruz when you did, how that's

0:53

informing what you're doing. working on their

0:55

self-driving car project and rumor had it

0:57

where they had spent like a hundred

0:59

million dollars and they had the world's

1:01

best engineers and so going after something

1:03

like that was a little bit crazy

1:05

and you know even after having worked

1:07

on twitch. You'd think that'd be enough credibility that as

1:09

a repeat founder I could go back and raise money, but

1:11

it turned out even that was like kind of a crazy

1:14

enough idea and to which hadn't been acquired yet that I

1:16

had a hard time. I had to scrape the bottom of

1:18

the barrel to raise money. I think I pitched like 120

1:20

investors, you know, over the course of probably a couple years

1:22

to raise all the money we needed. But our thesis back

1:24

then was very simple. We were going to start by finding,

1:27

you know, you know, instead of going directly after what

1:29

Google was going directly after what Google was

1:31

doing directly after what Google was doing, what

1:33

Google was doing. you know, self-driving cars that

1:35

were trying to make the ultimate self-driving car,

1:37

I think, as a moonshot, we took an approach

1:39

of what's the lean startup approach of this. Can

1:42

you build something that has the minimum quantum of

1:44

utility that is maybe a lower cost or easier

1:46

to execute so you can get to market more

1:48

quickly and move from there? And so we

1:50

started with a retrofit system where we would

1:53

take, you know, a regular car, put some

1:55

sensors on a computer in the back and

1:57

get it in the back and get it

1:59

to the back and get it to the

2:02

back and get it to the version of

2:04

Tesla full-stop driving. I think that was for

2:06

just one car model too, right? That was

2:09

like a BMW or something in the time?

2:11

That is the challenge with a retrofit business.

2:13

It's like without the blessing of the carmakers

2:15

you have to kind of reverse engineer protocols

2:18

and figure out how to attach motors to

2:20

steering out how to attach motors to steering

2:22

wheels. So it wasn't necessarily sustainable, but we

2:24

were still going to try to figure that

2:27

product for about a year and a half.

2:29

and then realized that we had done enough

2:31

technically that maybe we didn't have to do

2:34

this lean startup approach, maybe we had just

2:36

go straight after the big fish, which would

2:38

be building robotaxies. And around that time, Hoover

2:40

and Lyft had sort of risen in popularity

2:43

and now we're becoming these household names. and

2:45

like you know talk of going public and

2:47

all this kind of stuff and they had

2:49

this big hole in their unit economics which

2:52

is paying the drivers and so suddenly there

2:54

was like a strong bulk market poll for

2:56

self-driving technology whereas before I had been seen

2:59

as just kind of a cool sci-fi thing

3:01

and so we were able to raise some

3:03

money from spark capital and go straight into

3:05

that within a year of that I think

3:08

we were required by GM. You know we

3:10

had working prototypes driving prototypes driving around San

3:12

Francisco, obeying traffic lights changing lanes you know

3:14

going from point A to point A to

3:17

point B with an iPhone app. you know,

3:19

back in 2015. That's pretty amazing. How do

3:21

you think about the different approaches that people

3:24

are taking today? So there's Tesla on one

3:26

side with a very specific approach, kind of

3:28

moving everything to things that are a bit

3:30

more camera-centric, but training on sort of a

3:33

richer set of sensors and approaches. There's sort

3:35

of the Waymo approach, which is much heavier

3:37

on the hardware side in terms of what's

3:39

actually on the vehicle. Both seem to be

3:42

doing very interesting things. One is... Robo Taxis,

3:44

one is kind of still building mainly cars.

3:46

How do you think about the different approaches

3:49

both from a business model perspective, but also

3:51

from a technology perspective? To be fair, Elon

3:53

nailed it from a business model perspective. He's

3:55

been making billions of dollars of profit while

3:58

developing self-driving cars, or as everyone else has

4:00

been, burning billions of dollars to try to

4:02

get to basically the same point. I think

4:04

in the end, when you start with custom

4:07

vehicles, lots of sensors, lots of sensors, really

4:09

expensive, work in a constrained environment. And Tesla

4:11

starts with unconstrained environment, low-cost sensors, but it

4:14

doesn't quite work without a driver. They're all

4:16

trying to get to the same spot in

4:18

the end, which is low-cost, works everywhere, you

4:20

know, commodity sensors. Different paths to get there.

4:23

I think Elon won that hands-down. What do

4:25

you think of their criticism that you can't

4:27

get there? You can't get to full self-driving

4:29

from like mostly self-driving. That statement is almost

4:32

certainly wrong, given a long enough timespan. Again,

4:34

going back to what Elon's approach was, he

4:36

doesn't have to like finish by a certain

4:39

date or run out of money. He's like

4:41

making money along the way. And so, you

4:43

know, I think the only risk is that,

4:45

you know, customers get fed up and, you

4:48

know, sort of rage quit his program, but

4:50

they're getting something they like along the way.

4:52

Tesla Full self-driving. So I think that's the

4:55

right approach. But in 2013, for sure, in

4:57

2015, even 2018, it really wasn't viable. I

4:59

think now if you take a fresh look

5:01

at where we are today with large language

5:04

models and that, you know, generative models and

5:06

other things that that sort of class of

5:08

technology applied to the classical challenges of perception

5:10

for autonomous driving, even motion planning for autonomous

5:13

driving, completely changed the game in terms of

5:15

the magnitude of compute that you needed, the

5:17

expense of that. And I think now from

5:20

cameras, as long as you have sufficient redundancy

5:22

and low light sensitivity and, you know, some

5:24

robustness there. You can extract from a single

5:26

camera image, not even stereo, you can get

5:29

beautiful depth data, really accurate, and those models

5:31

are getting better every day. And so, you

5:33

know, if you're making a bet on the

5:35

right technical approach in 2025, it does not

5:38

involve a bunch of expensive lyders or exotic

5:40

sensors. It involves the most commodity of sort

5:42

of, you know, high-balling readily available sensors you

5:45

can get, and probably just several more of

5:47

them than you would find on a typical

5:49

driver assistance system. So I think that's the

5:51

path from here on out. Is there anything

5:54

else you think is lacking from a technology

5:56

perspective either in terms of hardware or just

5:58

scaling models or as you know better than

6:00

anyone everybody started moving towards end-to-end deep learning

6:03

over the last you know year or two

6:05

and that's really made a big difference? But

6:07

is it just scaling that up or is

6:10

something else lacking? And approach will be an

6:12

in-to-in type model. I mean, it's sort of

6:14

hard to put it in a bucket of

6:16

in-to-end or smaller models because there's such a

6:19

spectrum in between and everything I've seen as

6:21

a mix and match of various technologies. If

6:23

I look at the limiting factors, at least

6:25

on the hardware side, at least previously, it's

6:28

been hard to get high performance compute in

6:30

an automotive high temperature range, safety critical environment.

6:32

And so. You know, cruise made custom chips.

6:35

I'm sure Waymo makes custom chips and sort

6:37

of piecing together things from the supply chain

6:39

to solve that is a little challenging. And

6:41

so there is room for more high performance

6:44

compute automotive silicon and I've seen some things

6:46

happening in that space. That's one. And then

6:48

I'd say the other piece, you know, most

6:50

robot taxi deployments I've seen rely on some

6:53

form of remote assistance. And so there's a

6:55

question of like how you get reliable. connectivity

6:57

to a vehicle from anywhere. Using multiple cell

7:00

networks like what Cruise is done and Waymore

7:02

I'm sure others works, provided you have cell

7:04

phone coverage. I think the missing piece may

7:06

be Starlink or something similar where you can

7:09

have always on connectivity between Starlink and maybe

7:11

a cell phone and some other fallback. I

7:13

think that really opens up the opportunity and

7:15

the number of places you can deploy robot

7:18

taxis, whereas before is sort of an open

7:20

question, what you do if you're driving down

7:22

Highway 1 in California and there's no cell

7:25

coverage there, like should you still have an

7:27

avi on that road, given, if there's an

7:29

avi on that road, if there's an issue,

7:31

or if there's an issue, or if a

7:34

customer needs some help, you've, was there like

7:36

a right ear to start a self-driving pipeline

7:38

and everything? So probably like... you know, circuit

7:40

2020 or so would be the right time

7:43

to get started. And then, you know, around

7:45

now, I think you'd be having like a

7:47

combination of hardware and software that's mature enough.

7:50

If you have a nimble enough engineering team

7:52

that's able to adopt new technologies when they

7:54

pop up and quickly pull them into the

7:56

pipeline, you're actually well positioned even if you

7:59

started a while ago and your tech stack

8:01

was based on similar technologies. If you have

8:03

over the infrastructure in place for validation and

8:06

testing and training models and deploying on... on

8:08

public roads with test drivers, I think you

8:10

can go a lot faster, even if you

8:12

have to like sort of rip out and

8:15

change some of your tech staff to adapt

8:17

with the times. One thing I've heard, two

8:19

opposing viewpoints on, is the autonomous vehicle market

8:21

in China. And one point of view is,

8:24

well, it's not that real, and it's mainly

8:26

teleop, and it's a little bit more sizzled

8:28

than steak, and then the other opposing view

8:31

is, well, actually they've advanced dramatically, really rapidly,

8:33

do more, try more, etc. in that models

8:35

and approaches there are at least a parody

8:37

with the leading contenders in the Western world.

8:40

Which of those two views do you subscribe

8:42

to or how do you think that market

8:44

will evolve? But what I've seen so far

8:46

and I don't have you know a lot

8:49

of inside information just from what I've seen

8:51

videos online and other things, it does still

8:53

seem like there's a lot of teleoperation, like

8:56

I think even even someone like Tesla may

8:58

start off with like a one-to-one ratio of

9:00

remote operators to vehicles. Cruise and Waymo probably

9:02

started off pretty close to that. And then

9:05

I think over time moved to a smaller

9:07

ratio. And so I think in the interim

9:09

to get the deployment numbers up to get

9:11

more experience to accelerate data collection, people are

9:14

brute forcing it, which means there's probably a

9:16

lot of remote operation. I actually think that's

9:18

fine because it doesn't take much, once you

9:21

get to like, you know, 50% remote assistance

9:23

or even 25% remote assistance, you've already reduced

9:25

the labor costs, you know, 75% and so

9:27

you're almost already at the diminishing returns point.

9:30

And it sounds on one hand, it kind

9:32

of crazy to say like, oh, if there's

9:34

a, you know, a thousand AVs out there,

9:36

maybe 250 people monitoring them, but that's actually

9:39

not crazy from a cost from a unit

9:41

economic standpoint. It actually makes. trivial I think

9:43

with today's technology over time you could get

9:46

to 20 to 1 or 50 to 1

9:48

but you know you're just talking single-digit points

9:50

of margin at that point. The real benefit

9:52

is just going you know anywhere less than

9:55

one full human in a car makes the

9:57

economics of this really good and I think

9:59

you know puts you on a pathway towards

10:01

you know better safety for the vehicles because

10:04

they're primarily driven by a robot that has

10:06

great reflexes. and is going to avoid situations,

10:08

but then also lower cost to consumers over

10:11

time. You did amazing work on cruise, and

10:13

then you decided to start another company, which

10:15

I always think is a really brave endeavor,

10:17

because anybody who's been through multiple startups knows

10:20

how painful and terrible it is. Could you

10:22

tell us a little bit more about the

10:24

impetus behind the bot company and what you're

10:26

doing there? Yeah, well we talked about this

10:29

a little bit when I was making that

10:31

decision, what to do next? And I did

10:33

some soul searching and determined that I'm just

10:36

a builder. I like building things and I

10:38

like building things and I'm like building things

10:40

and I'm sitting on the sidelines or helping,

10:42

I like building things and I'm sitting on

10:45

the sidelines or helping other start-up. And the

10:47

first one was, you know, Twitch and Justin

10:49

TV, and that was straight out of college,

10:51

and that was just doing anything. Doing a

10:54

startup and trying to make it work, was

10:56

the priority. And that ended up being video

10:58

games and entertainment. The second time around for

11:01

crews, I decided I want to focus on

11:03

impact. So like, what's something where we can

11:05

use technology to meaningfully improve people's lives? Self-driven

11:07

cars, they save lives, they give you tons

11:10

of time back, that was like squarely in

11:12

the time back, that was like, I definitely

11:14

care about impact, but also fun. So it's

11:16

like working with people I like on problems

11:19

I like, really challenging technical problems and building

11:21

amazing products. And so we're building home robots.

11:23

And the impact side of that is one

11:26

of those things that's hidden in plain sight.

11:28

There's only 24 hours in a day. And

11:30

if you're sleeping for eight hours, working for

11:32

eight to 10, there's precious few hours left

11:35

that are actually your time. And people spend

11:37

a surprising amount of that remaining time doing

11:39

like. essentially unskilled labor acting like robots every

11:42

day like making the bed doing the dishes

11:44

folding the laundry like picking up toys after

11:46

your kids these are not things that make

11:48

us human these are actually things that detract

11:51

from our humanity and they're the perfect criteria

11:53

for that reason to be automated by machines

11:55

and I think you know when you describe

11:57

that to people like oh you don't have

12:00

to do all those things There's a machine

12:02

that could do this for you. It clicks

12:04

instantly. People like that is so obvious, to

12:07

the point where I think in five years,

12:09

maybe ten years, it will seem as insane

12:11

to have like a house without multiple home

12:13

robots as it would be to have a

12:16

house without a house without a house without

12:18

a house without a house without a house

12:20

without a house without a sink or like

12:22

a laundry machine or like a toilet. Like

12:25

these are going to be like critical things

12:27

that, you know, if you can afford them

12:29

and we want to make them really affordable.

12:32

are just going to seem like extremely common

12:34

sense to like why wouldn't I want to

12:36

have the time in my home be my

12:38

time not you know consumed by these chores.

12:41

I just thought that analogy was really interesting

12:43

because we never really think about plumbing as

12:45

a technology and it's a technology and because

12:47

we never really think about plumbing as a

12:50

technology and it is a technology and it

12:52

is and to your point up until recently

12:54

recently you know most of human history we

12:57

had no running water. You'd like walk down

12:59

a hill with a bucket and you'd just

13:01

take it off for granted. Yeah, so plumbing

13:03

and electricity were sort of turn-of-the-century things, and

13:06

then I'd say in the 1950s and 60s

13:08

there's a resurgence, but around home appliances. There's

13:10

some great advertisements from the 50s and 60s

13:12

to look back. It's like, this is 1950s

13:15

time, but it's like, this is 1950s time,

13:17

but it's like, you know, the housewife standing

13:19

in the kitchen, and the dishwasher, and all

13:22

these like new appliances, like, 70 years. So

13:24

I think it's time to revisit that. And

13:26

going back to the robotics side, like actually

13:28

pulling this off, it's basically been the dream

13:31

for people working on robots since the dawn

13:33

of robotics to build a robot that can

13:35

go to the fridge and get you a

13:37

beer or something like that. You know, you

13:40

said on the couch, it's like the dream

13:42

for nerds working on robots. And that's like

13:44

obviously a tiny subset of what you'd want

13:47

a household robot to do. But that's really

13:49

hard. And the reality is it's similar to

13:51

self-driving cars in some regard, where it's a

13:53

very unstructured environment, like every home is different,

13:56

everybody has organizes their home in a different

13:58

way. The layout is... different, the objects in

14:00

their home, the different, how they live in

14:02

their home is different. And so having a

14:05

robot that sort of lives in this

14:07

unstructured environment, it's like the polar opposite

14:09

of a factory assembly line where everything

14:11

is rigid and repetitive and precise in

14:13

a home, it's like sloppy and changes

14:15

every day. And so using classical approaches,

14:17

we have computer vision and you're trying

14:19

to like reconstruct 3D objects or fit

14:21

to a map, would make this like

14:23

a really, really challenging and computationally intensive

14:25

problem. Moving to more modern

14:28

techniques like end-to-end learning or

14:30

imitation learning, even reinforcement learning.

14:32

Now if you can tell the operator

14:34

a robot and demonstrate how to do

14:36

something or collect data from humans in

14:38

some way or from internet videos, you

14:40

can kind of imbue a robot with

14:43

a sense of common sense and ability

14:45

to make sense of these unstructured environments.

14:47

And on top of that, you can

14:49

talk to the robot in natural language

14:51

using your voice rather than typing into

14:53

an app or on a keyboard. And you asked

14:55

when the time is to start a robots

14:58

or a self-driving company, maybe that was 2020.

15:00

I think for home robots, it feels a

15:02

little bit early, so like now is definitely

15:04

the time in my view. How much of

15:06

why you did it, or that people have

15:09

learned at places like Cruz or Waymo or

15:11

others, is also useful in the context of

15:13

home robots? In other words, what sorts of

15:15

things overlap? And then what things are

15:17

just completely different or new, because

15:19

people would often talk about driving

15:22

environments is similarly chaotic and messy.

15:24

And, you know, the canonical example

15:26

is always like the kid chasing the

15:28

ball across the road suddenly or things like

15:30

that. Is it even more difficult in the home?

15:32

Is it less, is it more structured? I'm so

15:35

bit curious about the analogies that could

15:37

be pulled there if any. To start with,

15:39

the big difference between the two is that,

15:41

you know, for a driverless robot taxi. you

15:44

basically have no product until it achieves superhuman

15:46

safety performance, whatever you establish that is, like

15:48

just slightly better than humans or 10 times

15:50

better. I think most driverless cars that

15:52

are on the road today fall somewhere in that

15:55

category. And to get there, it means there's no

15:57

MVP, there's no launching with something that's partially useful.

15:59

It's like. you have to reach that human

16:01

safety performance. And on public roads, it's hard

16:03

to constrain the environment to the point where

16:05

you make the problem much easier. You can

16:07

operate at night or in sparsely populated areas,

16:09

but the reality is, just like you said,

16:11

at any moment, anywhere, the kid could dart

16:14

out in front of that vehicle. And so

16:16

you need like a high number of nines

16:18

of reliability to have any sort of product.

16:20

I think in the home and most consumer

16:22

applications and even most industrial applications. Safety is

16:24

still critically important, but the bar that you

16:26

need to reach for the functionality that you

16:28

need to reach for the constraints you could

16:30

put on the system and able to use

16:32

to launch a product much more quickly. And

16:34

so I think that's one big difference. On

16:37

this topic, how do you imagine deployment to

16:39

work? You're obviously like, hey, Tesla had the

16:41

right path here. Is there a testament where

16:43

you make billions along the way? It's not

16:45

obvious. Are there constraints you can put around

16:47

it where you have teleop? Right, or just

16:49

a couple tasks or a more constrained environment

16:51

in the messiness of a home? There's a

16:53

number of ways to attack that. Approaches I've

16:55

seen are like you sell a really high-priced

16:58

robot today, like a humanoid or something resembling

17:00

a human, that's like fully teleopt, and you

17:02

just tell someone this is going to cost,

17:04

I don't know, something crazy, like $50,000 and

17:06

$1,000 a month, but it's the first robot

17:08

you can buy that will like do stuff

17:10

in your stuff in your house. I think

17:12

that will like do stuff in your stuff

17:14

in your house. you know, make money along

17:16

the way. I think your market size is

17:19

pretty small doing that, but that's a viable

17:21

approach. And the other side would be to

17:23

sell robots that don't fulfill like the promise

17:25

of a household robot that does all your

17:27

chores, but do like little bits of useful

17:29

things. I just saw at CES, this year,

17:31

this year, I just saw at CES this

17:33

year, they have like little bits of useful

17:35

things. I just saw at CES this year,

17:37

they have like little bits of useful things.

17:39

I just to sell products of useful things,

17:42

but do like little bits of useful things,

17:44

like little bits of useful, that ladder I

17:46

guess to the Holy Grail which would be

17:48

like you know a robot that that takes

17:50

the place of your butler and your housekeeper

17:52

and just just about anything else that you

17:54

would ever want you know if you could

17:56

have an infinite staff of people or robots

17:58

doing things in your home. I guess we'll

18:00

run the analogy to self-driving if you look

18:03

at what happened from a market structure perspective

18:05

there. There are originally dozens of startups that

18:07

raised collectively billions of dollars. And one could

18:09

argue that the end winners of the things

18:11

that actually somehow worked in the market were

18:13

two incumbents, Tesla, and Waymo, Cruz, slash GM,

18:15

and then maybe to a secondary extent applied

18:17

to tuition, which is building more general software

18:19

for cars and things like that. And then

18:21

most of that market kind of didn't end

18:23

up with the outcomes on what I hope

18:26

for. Do you think there's going to be

18:28

a similar sort of shakeout here in robotics?

18:30

And do you think there will be incumbent

18:32

bias? Do you think there's a lot of

18:34

room for startups? How do you think about

18:36

how that market will evolve? I think for

18:38

sure, both both, both in AI generally, like,

18:40

you know, just pure software companies, and then

18:42

also I think the next way of starting

18:44

on robotics, there will be like that bubble

18:47

effect, and this. There's either a couple big

18:49

rounds that get everyone excited and then other

18:51

investors start throwing money into the same space

18:53

because they see the markups happening quickly. And

18:55

that following effect kind of floods the market

18:57

and when there's a lot of investors talking

18:59

about funding these companies more people kind of

19:01

drop out of their PhD programs or quit

19:03

their job to start a company. And I

19:05

think the majority of those companies I would

19:07

say are like low quality in that. they're

19:10

a founder that's like half into it, or

19:12

a founding team that's like half into it

19:14

and half hedging going back to work or

19:16

whatever, or maybe they're hoping it's a get-rich-quick

19:18

thing, or they're just the wrong, you know,

19:20

like, founder market fit or founder product fit

19:22

isn't there, even though they're smart technically. I

19:24

saw this in the self-driving wave, like, people

19:26

who are really brilliant, academically, not a product-centric

19:28

mentality, they will leave their academic program because

19:31

they want to commercialize their You're saying I'm

19:33

not going to be flexible on how I

19:35

solve the problem. I'm going to force my

19:37

solution into the, you know, square peg into

19:39

the round hole. And I think that can

19:41

be very problematic for a startup when you're

19:43

constantly wrong. I need to adapt to whatever

19:45

you see. So most of those companies will

19:47

be a low quality and as a result,

19:49

we'll say that the bubble popped and like,

19:52

you know, there's a huge wipe out in

19:54

industry, inevitably, whether it's robotics or AI. But

19:56

I think really what it was, is there

19:58

was, is there was, is a handful of

20:00

good companies, is a handful of good companies,

20:02

is a handful of good companies, that time

20:04

and before and those companies did it just

20:06

fine I don't think they'll be affected by

20:08

the collapse it's just all the noise and

20:10

sort of the follow-on and you know sort

20:12

of the hype and mania that follows that

20:15

sort of gives the impression that you know

20:17

these things are collapsing are not viable when

20:19

in reality I think there are lots you

20:21

know a handful of companies doing really good

20:23

work I don't know if they're necessarily limited

20:25

to the incumbents but that is is is

20:27

a acquired by Amazon, and so has the

20:29

resources to keep going? Cruise fell into that

20:31

category. So these were not incumbents. These were

20:33

companies that were started from scratch during the

20:36

beginning of that self-driving car cycle and are

20:38

enduring, and I hopefully do well. Thanks for

20:40

that overview, and an explanation, what happened, and

20:42

I hopefully do well. Thanks for that explanation,

20:44

what happened in the industry, and others, being

20:46

part of the things that either had an

20:48

exit or worked in different ways over time

20:50

or time. Is they ascribe personality or they

20:52

kind of project personhood onto these machines? Is

20:54

that something that you think is worth winning

20:56

into? Is it something that's worth avoiding like

20:59

the anthropomization, I can never say that word,

21:01

the humanization of these devices? Like how do

21:03

you think about that as somebody who's actually

21:05

building things that will be in the home

21:07

with consumers and, you know, may get interpreted

21:09

in different ways by the customer? You know,

21:11

to start with analogy in the self-driving car.

21:13

industry was interesting because we named our cars

21:15

every car had a name and people would

21:17

personify it when it came up and you

21:20

know a car in itself doesn't look like

21:22

a creature it looks like a car or

21:24

something that you drive but once it starts

21:26

moving on its own your brain plays some

21:28

tricks on you and starts like treating it

21:30

like it's an entity or a creature of

21:32

some kind and so you can try to

21:34

pretend that that doesn't exist and then you

21:36

have this weird cognitive dissonative dissonance where you're

21:38

saying it's a you know, consciousness or life

21:40

force in some way, or you can lean

21:43

into it and acknowledge that and find a

21:45

way to integrate it in the right way.

21:47

I think with anthropomorphism, like I said that

21:49

word, right? You're just showing off now. Yeah,

21:51

seriously. That was the hard part of the

21:53

bot company. Yeah, yeah. Yeah, but too much

21:55

anthropomorphism can, yeah, I screwed up there, can

21:57

imply like a set of human-like behaviors that

21:59

don't exist in that product, and I think

22:01

Rodney Brooks wrote an essay about this basically

22:04

saying with robots in particular, the appearance of

22:06

them sets the expectation for what that product

22:08

will do. And so if you make something

22:10

that looks exactly like a human, and in

22:12

fact the more human like you make it,

22:14

the higher the expectations I think the average

22:16

person will have for that machine. It looks

22:18

like me, it walks like me, it has

22:20

a face and talks like me, so it

22:22

must be capable of doing all the things

22:25

that I can do. And today, 2025, I

22:27

think it would be a leap of faith

22:29

for any company to sell a humanoid robot

22:31

or something like that. and imply that it

22:33

can do all of these things, because we're

22:35

still, I think, many years out from that,

22:37

at least from meeting those expectations. And so

22:39

I think that there's a lot of thought

22:41

that can go into the design of a

22:43

robot, the form of a robot and other

22:45

things to try to match the expectations that

22:48

you have for a robot when you see

22:50

it, to what it can actually do, or

22:52

even go the other direction instead of overpromising

22:54

by showing a humanoid, maybe you can do

22:56

something. in the other direction and surprise people

22:58

with how much it can do. And that's

23:00

kind of how I think about it personally.

23:02

I'd like to surprise and delight customers rather

23:04

than set them up for disappointment. Maybe on

23:06

this front of just consumer acceptance and expectations,

23:09

are there like lessons that transfer from self-driving

23:11

to home robots? One thing I saw in

23:13

self-driving, which I guess you could say it's

23:15

intuitive, but it was still very striking, was

23:17

that most people. on the whole, we're very

23:19

skeptical of self-driving cars. Like it was like

23:21

75, 80% of people, like I've never, you

23:23

know, I'm never going to trust one of

23:25

those things. That dropped, like, dropped to like

23:27

20 or 30% after one ride. Amazing. And

23:29

so it's one of those things where it's

23:32

like, you just simply do not believe it.

23:34

The more transformative and the more science fiction

23:36

technology. feels, the higher the skepticism will be

23:38

for that kind of thing. And so I

23:40

think anytime you're doing something new, whether it's

23:42

self-driving or a home robot, which let's be

23:44

honest, that sounds like sci-fi, I'd love to

23:46

have that, but like, you know, can this

23:48

be real, is the question. And I think

23:50

this be real, is the question. And I

23:53

think there will be a barrier. And I

23:55

think this be real, is the question. And

23:57

I think there will be a barrier. There

23:59

will be a barrier. There will be a

24:01

barrier. There will be skepticism. There will be

24:03

a barrier. There will be a barrier. and

24:05

saying, no, no, no, I wrote in this

24:07

thing or I tried this thing and it's

24:09

real, you've got to try it. And so

24:11

I think leaning into too much sort of

24:14

classical marketing and trying to like tell people

24:16

like what this thing will do or what

24:18

its specs are and all that is very

24:20

different than like hearing from someone you trust

24:22

that I have this thing in my home,

24:24

one of your most intimate spaces, and it's

24:26

working. And it's like, and I love it.

24:28

And so that's kind of how I think

24:30

about it for something about it for something

24:32

like it for something like it for something

24:34

like this for something like this where something

24:37

like this where something like this where it

24:39

for something like this where it for something

24:41

like this where it for something like this

24:43

where it for something like this where it

24:45

for something like this where it's real, where

24:47

it for something like this where it's real,

24:49

where it's real, just don't believe that a

24:51

sci-fi think can exist and try to convince

24:53

them through any other medium other than just

24:55

trying it themselves. What's the timeline for that?

24:58

Like you mentioned in passing like a pretty

25:00

important claim that said like maybe five, ten,

25:02

twenty years until everyone who can afford them

25:04

expects robots in the house like they expect

25:06

home appliances? Like what changes that timeline between

25:08

the five to twenty year span or whatever

25:10

it ends up being? Well, I'm on my

25:12

office, so it's basically when I get off

25:14

this podcast and go back to work, that'll

25:16

let me know. Okay, so we should let

25:18

you go. I was interested. Yeah, no, no,

25:21

no, no, no. No, it's, it's, I think,

25:23

to me, and I felt this way in

25:25

2013 when I started Cruz, it seems like

25:27

the basic building blocks are there. I can

25:29

point to all the challenges for a low

25:31

cost home robot, and low cost is important

25:33

to me, because I want a lot of

25:35

people to a lot of people to a

25:37

lot of people to have to work, and

25:39

I can point to a technology we've either

25:42

built, you know, the last year, or some

25:44

reaches that came out, or like a product,

25:46

like a chip that's coming in the pipeline,

25:48

whatever it is, I can see, you know,

25:50

from where I sit today, the path to

25:52

put these things together and assemble a great

25:54

product and a great product experience. And so

25:56

I think it comes down to like... execution,

25:58

how quickly and how well those things are

26:00

put together. And then the big question, which

26:02

you always face in something like this, is

26:05

what are the unknown unknowns? Like what can

26:07

I sit here today and just simply not

26:09

see because we haven't put enough robots in

26:11

homes and haven't tried this out? And I

26:13

think there could be like something, some new

26:15

discovery that happens, maybe it turns out that people are

26:17

not happy with a home robot unless it does X or. whenever

26:19

a robot does this other thing, it makes people never want to

26:21

use it again. And so we're kind of early in our own

26:24

journey to figure out what those things are. And as far as

26:26

I've seen, there's no one else really doing this. And so it's

26:28

a big, the cloud of uncertainty is, there's a lot of it

26:30

in front of us, and that's why I can't be more specific

26:32

on the timeline. It could be like one year or it could

26:34

be like 20 years. And so the best way to figure out

26:36

is to figure that I was just charge forward, I was just

26:38

charge forward, I was just charge forward, I just charge forward and

26:40

try to discover to discover it as just charge forward and try

26:42

to discover it as quickly as quickly as quickly as quickly as

26:44

quickly as quickly as quickly as possible. All of

26:46

those are really exciting as timelines. We

26:48

talked about Chinese. Avi, how do you

26:51

think about manufacturing and supply chain given

26:53

competition with China and like Chinese robotics

26:55

companies? Yeah, I spent a lot of

26:57

time thinking about that. There is this

26:59

sentiment in robotics or I'd say in

27:01

the hardware company space that I would

27:03

say the sentiment that I've heard is

27:05

almost like you shouldn't even try. Yeah.

27:07

Because if you're just making a widget, you'll

27:09

make it using US engineering, which is

27:12

100 plus thousand dollars a year for

27:14

engineer. you're going to go to

27:16

US-based machine shops and US-based tooling,

27:18

all this kind of stuff, and

27:20

it's going to be slower, you're

27:22

not going to iterate it as

27:24

fast, or you're going to pay more

27:26

for it, and the quality may or may

27:29

not be as good. So don't even try.

27:31

And I think it is possible to be

27:33

a global company, but it may not be

27:35

as good. So don't even try. And I

27:37

think it is possible, and you have to

27:40

be willing to travel and go, like, like,

27:42

you know, like, like, like, for a US-based

27:44

company, you're going to have a hard time

27:46

competing on pure engineering services. And so if

27:49

that's all you've got, if you're just like

27:51

doing mechanical engineering and cranking out products, you

27:53

can have a hard time on the margin side and have

27:55

to build a brand instead in order to create that margin.

27:57

I think when you get into more complex machines,

27:59

And I saw this with self-driving. I think

28:02

even though they may be doing a lot

28:04

of teleoperation and other things there, I do

28:06

think that China is still used behind the

28:08

best U.S. companies for self-driving. And it's been

28:11

my experience that any time you have a

28:13

sufficiently complicated technical problem on the software side,

28:15

where you can't just copy it by measuring

28:17

something and then recreating it in CAD, or

28:20

it has to do with taste, like the

28:22

product experience is more than just like a

28:24

light switch where you flip it on and

28:26

off. Then it becomes a little bit harder

28:29

to quickly copy that and commoditize it. You

28:31

know, even if you're a fast follower, if

28:33

you are aggressive enough as a company and

28:35

can maintain a lead and keep innovating and

28:38

keep building new products, I think there's room

28:40

to be a hardware company in the US,

28:42

you know, provided that you take steps to

28:44

ensure that if your cost is higher than

28:47

a potential Chinese competitor or somewhere in Asia,

28:49

it's not that much higher, to the point

28:51

where you can win on the merits of

28:53

your product and brand and brand and brand

28:55

and other things. This is like the Wild

28:58

West. There's no established stuff to copy. We've

29:00

got to build a lot of stuff ourselves.

29:02

And I think it'll be interesting to see

29:04

how this plays out too, especially if these

29:07

end up being connected devices and they're constantly

29:09

getting new software updates with better models on

29:11

them or whatever it is. I can see

29:13

a bunch of angles in which these companies

29:16

are very durable, whether it's us or another

29:18

one. You've already been through the ringer once

29:20

on the regulatory front with AV. If you

29:22

could wave a magic wand, we just call

29:25

Trump right now. What do you think is

29:27

the right policy approach to make sure we

29:29

have a competitive domestic robotics industry, if that's

29:31

relevant? First of all, I mean, I think

29:34

there should be tons of regulation on AV.

29:36

It reminds me what's happening in the AV

29:38

space and sort of, you know, companies like

29:40

Cruz got wiped out, companies like Waymo are

29:43

growing or expanding, I think much more slowly

29:45

than they should, based on the safety of

29:47

their technology. And it reminds me of like

29:49

when the first airlines were formed in the

29:51

US and there was no FAA there was

29:54

no regulation if you had basically any kind

29:56

of plane crash You would get sued out

29:58

of existence and you just wiped out and

30:00

many of the first airlines are no longer

30:03

in existence because of this and so the

30:05

FAA was created because you know the government

30:07

decided that actually we should have airlines and

30:09

if they keep going out of business no

30:12

one's gonna start an airline anymore and so

30:14

create the FAA monitor air plane manufacturers in

30:16

airlines make sure they meet safety criteria and

30:18

in exchange give them you know reasonable protections

30:21

and limits on liabilities so they can actually

30:23

operate you know in a society like the

30:25

US. That hasn't happened for self-driving cars and

30:27

so I think the only Canada today the

30:30

only companies that stand a chance. are the

30:32

ones where they can afford to take on

30:34

that liability, because they're a giant tech company

30:36

that makes money in other places. Otherwise, it's

30:39

very bleak. So I would recommend that for

30:41

80s. And for home robots, I think a

30:43

similar thing is true. There are no similar

30:45

thing is true. There are no regulations right

30:48

now on cybersecurity, for example. You can have

30:50

a Chinese manufactured robot in your home with

30:52

cameras and a microphone running and sending that

30:54

data. Who knows where? Who knows where? On

30:56

the safety side, I would love to see

30:59

that too. There's lots of best practices from

31:01

the industrial robot industry. They're not a great

31:03

fit for home robots. There's lots of good

31:05

best practices for other consumer products, but particularly

31:08

home robots is a bit of a vacuum.

31:10

And I think that generally regulation is a

31:12

very good thing for companies operating in environments

31:14

like this that are especially ones that are

31:17

unpredictable. It encourages discussion of best practices. It

31:19

encourages oversight, all of these things lead to

31:21

better outcomes for... I think both the companies

31:23

and for consumers. So in terms of regulation,

31:26

even though like Trump administration's anti-regulation, we actually,

31:28

it's like a necessary enabler to get these

31:30

industries going as odd as that sounds. Seems

31:32

like there's some circumstances where you're pointing a

31:35

clear regulatory framework helps a lot. Like for

31:37

the crypto community during the Biden administration, a

31:39

lot of what they wanted from the SEC

31:41

was just guidance in terms of what to

31:44

do, and then everybody's going to go do

31:46

it, and it was that ambiguity that ambiguity

31:48

that really hurt. is due to the FAA,

31:50

the US is now behind on the drone

31:53

side, and China was really able to get

31:55

a leg up if you look at everything

31:57

from drone shows, it's just what DGI drones

31:59

can do. at sort of very low cost,

32:01

and they're being used for military and other

32:04

applications in Ukraine and other places. And so

32:06

there are the argument as well, the FAA

32:08

or the regulated drones and airspace, and that

32:10

kind of prevented the US market from evolving

32:13

down a proper route. So I'm a little

32:15

bit curious how you think about that balance

32:17

between too much and too little regulation and

32:19

how that may apply to the robotics world,

32:22

given what's happened in drones. Yeah, it's a

32:24

good point. I mean, I agree. I would

32:26

love to see more transportation innovation generally, but

32:28

aviation, like, you know, electric takeoff and landing

32:31

airplanes, there were tons of startups in that

32:33

space, and they've all kind of like, sort

32:35

of fizzled out, it seems like, or bumped

32:37

into these like FAA certification challenges. And as

32:40

far as I understand, there's still no pathway

32:42

even today to make a pilotless plane of

32:44

any kind. There's like baby steps that we're

32:46

taking, but I would love to see these

32:49

sort of two track approaches where you have

32:51

a very mature track for existing, you know,

32:53

industries and technologies like, you know, passenger airlines

32:55

today, and then an innovation track where basically

32:58

you're almost encouraged. I mean, there's grants or

33:00

other programs to like spur innovation in this

33:02

category, and then maybe a phased release process

33:04

to go from like a working prototype to,

33:06

you know, being regulated, but being able to

33:09

operate at scale. Like I think it would

33:11

be okay for us to sit down and

33:13

say, here's what we expect to see at

33:15

each level of maturity. and then provided you

33:18

demonstrate to us that you meet that level

33:20

of maturity, we'll progressively open up, you know,

33:22

the regulations, or the areas that you can

33:24

operate. And this is like a standard thing.

33:27

This is like, you know, what we did

33:29

in self-driving cars, but also like, even new

33:31

airplanes, like the boom super sonic jet that

33:33

just made history. they started off with like

33:36

a low and slow flight and then like

33:38

gradually as they as they saw the data

33:40

check out ratcheted up until they hit supersonic.

33:42

And so I think regulations are there to

33:45

prevent irresponsible from people from going zero to

33:47

supersonic, but there are plenty of responsible people

33:49

willing to take responsible people willing to take

33:51

this stepped approach provided that there is a

33:54

pathway to do so. And so I think

33:56

that's the right balance of having a cruise

33:58

journey besides like a. I guess, have more

34:00

fun. Don't sell it to GM. Yeah, I

34:02

learned a lot. You know, big companies that have

34:05

their, have their, you know, core business in

34:07

another domain doing an acquisition that's

34:09

not in that domain, selling cars

34:11

to people who buy pickup trucks

34:13

and SUVs in the Midwest versus

34:15

Midwest, versus robot taxis and urban

34:17

environments. These are not compatible things.

34:20

And when push comes to shove,

34:22

they're going to pick that one

34:24

over the other. And we got

34:26

completely decimated by GMs like GMs

34:28

like LAC. Like LAC. ever. And so

34:30

like, you know, maybe a IPO or something like that,

34:32

but it will never be an acquisition in my life

34:34

again. The reality is, like, if I'm working on

34:36

something today, I'm working on it because I think

34:39

it's important and I care about it. And the day

34:41

you sell the company is the day that you have

34:43

to let that go. And so, you know, if I'm,

34:45

by definition, if this is something I care about, I'm

34:47

not going to let it go. The second thing is,

34:49

like team size. I along with many other people I

34:51

think who started companies around the same time fell

34:53

victim to the Silicon Valley dogma

34:56

of traditional engineering management for Silicon

34:58

Valley companies and so that's like hire

35:00

the the VPs first and then they hire the

35:02

senior managers who hire the senior managers who hire

35:04

the managers who hire the managers who hire the managers

35:06

and there can't be a ratio of more than

35:08

eight to one for fan outs and you do

35:10

performance review cycles and all these things that

35:12

creates bureaucracy and structure creates

35:14

like communication gaps between the people

35:16

actually doing the work and the people

35:19

making decisions. Keep companies very small. Like

35:21

have fewer employees. Make every seat count.

35:23

Get the absolute best person in every

35:25

role in every seat. And just never grow

35:27

the company to be so large that

35:29

it sort of like crumbles under all

35:31

the structuring and bureaucracy and politics that

35:34

see thin. These are natural and bureaucracy

35:36

and politics that see thin. These are

35:38

natural things that happen to large. That

35:40

it sort of like crumbles under all

35:42

the structuring and bureaucracy and politics. If

35:44

you want to have to have small ambitions.

35:47

You have to have small ambitions. coding

35:49

assistance, that it continue to get better, things

35:51

like deep research. Like I found that nearly every

35:53

job function can be partially automated and I think

35:55

that trend is going to continue. And as someone

35:58

who spends like a lot of time programming. I felt

36:00

my ability to take on things where I

36:02

would have had to hire a team of

36:04

people or like a specialist in iOS development

36:06

or a specialist in like, you know, low-level

36:08

rust programming for motor drivers. If you have

36:10

a couple good engineers, they can adapt and

36:12

do all those things when sitting next to

36:14

an LLLM and some good coding tools. So

36:16

I think it actually is viable to do

36:18

what I want to do, which is like

36:21

lesson learning to keep the company small to

36:23

build like a company that has grand ambitions,

36:25

but a very tiny team. It's pretty amazing.

36:27

I think you've had a really amazing career

36:29

arc overall, right? You've taken on three high-risk

36:31

complex companies back-to-back with little downtown in between.

36:33

You've run seven marathons and seven countenance and

36:35

three and a half days. Where does your

36:37

drive and stamina come from? Is there a

36:39

supplement we should all be taking? Is there

36:41

like some Brian Johnson style? Like, we should

36:43

inject ourselves with young blood? Like, what's the

36:45

deal? I haven't tried that. If you do

36:47

that, let me know how it goes. I

36:49

have a problem, which is that once I

36:52

get an idea in my head, it just,

36:54

it just like, burns a hole in my

36:56

brain. I cannot, I cannot sleep, I cannot

36:58

do anything, I can't focus until I like,

37:00

you know, see this idea through. And this, for

37:02

what a reason, it doesn't happen, it's not like, you

37:04

know, 20 ideas a day. It's like, I'll get this

37:06

idea that, that, hey, the star is aligned, this thing,

37:08

this thing should happen, this thing should happen, now is

37:10

the time, now is the time, now is the time,

37:12

now is the time, now is the time, now is

37:14

the time, now is the time, now is the time,

37:16

I need to, I need to, I need to, I

37:18

need to, I need to, I need to, I need

37:20

to, I need to, I need to, I need to,

37:22

I need to, I need to, And then once I'm

37:24

on that track, I just cannot let go until it's

37:27

done. And so I think I just latch on to

37:29

these problems and have this like sense of delayed gratification

37:31

where I want to work on it for a long

37:33

time and get the result. But that's also really satisfying

37:35

to me, the notion of like putting in a ton

37:37

of effort, building something really complicated and hard, and making

37:39

a little bit of progress each day. That is motivating

37:41

to me. And so I know you could say starting

37:43

these companies in high risk and high risks. but I

37:45

enjoy every day and so I can't imagine doing anything

37:47

else. That's awesome. I think that's what makes Silicon Valley

37:49

so great. So thank you so much for joining us

37:51

today. Thanks for having me. Find us on Twitter at

37:53

No Prior's Pod. Subscribe to our YouTube channel if you

37:55

want to see your face. Follow the show on

37:57

the show on Apple Spotify, or

38:00

Spotify, or wherever you listen. way

38:02

That way you get a

38:04

new episode every week. week. And

38:06

sign up for for find

38:08

transcripts for every episode for every .com.