George Hotz: Hacking the Simulation & Learning to Drive with Neural Nets
技术与编程AI 与机器学习心理与人性商业与创业音乐与艺术
📋 章节目录
暂无章节信息
🔑 关键词
dongoinginterestingbetterdrivingdoinglearningteslahumancommatalkdatacarsablecardriverstuffdoesnrealput
💬 精彩语录
暂无语录
🎙️ 完整对话（4507 条）
Lex Fridman (00:00.000)
The following is a conversation with George Hotz,
以下是与乔治·霍茨的对话，
Lex Fridman (00:02.440)
AKA Geohot, his second time on the podcast.
又名 Geohot，这是他第二次出现在播客上。
Lex Fridman (00:06.520)
He's the founder of Comma AI,
他是Comma AI的创始人，
Lex Fridman (00:09.320)
an autonomous and semi autonomous vehicle technology company
一家自动驾驶和半自动驾驶汽车技术公司
Lex Fridman (00:12.880)
that seeks to be to Tesla Autopilot
旨在成为特斯拉自动驾驶仪
Lex Fridman (00:15.840)
what Android is to the iOS.
Android 之于 iOS。
Lex Fridman (00:18.920)
They sell the Comma 2 device for $1,000
他们以 1,000 美元的价格出售 Comma 2 设备
George Hotz (00:22.760)
that when installed in many of their supported cars
当安装在他们支持的许多汽车上时
Lex Fridman (00:25.600)
can keep the vehicle centered in the lane
可以使车辆保持在车道中央
George Hotz (00:27.960)
even when there are no lane markings.
即使没有车道标记。
Lex Fridman (00:30.760)
It includes driver sensing
它包括驾驶员感应
George Hotz (00:32.400)
that ensures that the driver's eyes are on the road.
确保驾驶员的眼睛始终注视路面。
Lex Fridman (00:35.640)
As you may know, I'm a big fan of driver sensing.
如您所知，我是驾驶员传感技术的忠实粉丝。
George Hotz (00:38.440)
I do believe Tesla Autopilot and others
我确实相信特斯拉自动驾驶仪和其他
Lex Fridman (00:40.560)
should definitely include it in their sensor suite.
绝对应该将其包含在他们的传感器套件中。
George Hotz (00:43.520)
Also, I'm a fan of Android and a big fan of George
另外，我是 Android 的粉丝，也是 George 的忠实粉丝
Lex Fridman (00:47.120)
for many reasons,
由于多种原因，
George Hotz (00:48.240)
including his nonlinear out of the box brilliance
包括他开箱即用的非线性才华
Lex Fridman (00:51.640)
and the fact that he's a superstar programmer
事实上他是一位超级明星程序员
George Hotz (00:55.120)
of a very different style than myself.
与我自己的风格非常不同。
Lex Fridman (00:57.360)
Styles make fights and styles make conversations.
Lex Fridman (01:01.160)
So I really enjoyed this chat
Lex Fridman (01:02.920)
and I'm sure we'll talk many more times on this podcast.
George Hotz (01:06.280)
Quick mention of a sponsor
Lex Fridman (01:07.680)
followed by some thoughts related to the episode.
George Hotz (01:10.160)
First is Four Sigmatic,
Lex Fridman (01:12.240)
the maker of delicious mushroom coffee.
George Hotz (01:15.360)
Second is The Coding Digital,
Lex Fridman (01:17.520)
a podcast on tech and entrepreneurship
George Hotz (01:19.760)
that I listen to and enjoy.
Lex Fridman (01:22.080)
And finally, ExpressVPN,
George Hotz (01:24.520)
the VPN I've used for many years to protect my privacy
Lex Fridman (01:27.800)
on the internet.
George Hotz (01:29.320)
Please check out the sponsors in the description
Lex Fridman (01:31.280)
to get a discount and to support this podcast.
George Hotz (01:34.840)
As a side note, let me say that my work at MIT
Lex Fridman (01:38.080)
on autonomous and semi autonomous vehicles
George Hotz (01:40.480)
led me to study the human side of autonomy
Lex Fridman (01:43.080)
enough to understand that it's a beautifully complicated
Lex Fridman (01:46.600)
and interesting problem space,
Lex Fridman (01:48.600)
much richer than what can be studied in the lab.
George Hotz (01:51.800)
In that sense, the data that Comma AI, Tesla Autopilot
Lex Fridman (01:55.360)
and perhaps others like Cadillac Super Cruiser collecting
George Hotz (01:58.480)
gives us a chance to understand
Lex Fridman (02:00.560)
how we can design safe semi autonomous vehicles
George Hotz (02:03.800)
for real human beings in real world conditions.
Lex Fridman (02:07.560)
I think this requires bold innovation
Lex Fridman (02:09.920)
and a serious exploration of the first principles
Lex Fridman (02:13.000)
of the driving task itself.
George Hotz (02:15.640)
If you enjoyed this thing, subscribe on YouTube,
Lex Fridman (02:17.880)
review it with five stars and up a podcast,
George Hotz (02:20.160)
follow on Spotify, support on Patreon
Lex Fridman (02:22.760)
or connect with me on Twitter at Lex Friedman.
Lex Fridman (02:26.280)
And now here's my conversation with George Hotz.
Lex Fridman (02:31.360)
So last time we started talking about the simulation,
George Hotz (02:34.040)
this time let me ask you,
Lex Fridman (02:35.640)
do you think there's intelligent life out there
Lex Fridman (02:37.480)
in the universe?
Lex Fridman (02:38.600)
I've always maintained my answer to the Fermi paradox.
George Hotz (02:41.640)
I think there has been intelligent life
Lex Fridman (02:44.440)
elsewhere in the universe.
Lex Fridman (02:45.880)
So intelligent civilizations existed
Lex Fridman (02:47.920)
but they've blown themselves up.
Lex Fridman (02:49.240)
So your general intuition is that
Lex Fridman (02:50.760)
intelligent civilizations quickly,
George Hotz (02:54.560)
like there's that parameter in the Drake equation.
Lex Fridman (02:57.760)
Your sense is they don't last very long.
George Hotz (02:59.640)
Yeah.
Lex Fridman (03:00.480)
How are we doing on that?
Lex Fridman (03:01.560)
Like, have we lasted pretty good?
Lex Fridman (03:03.680)
Oh no.
Lex Fridman (03:04.520)
Are we do?
Lex Fridman (03:05.360)
Oh yeah.
George Hotz (03:06.200)
I mean, not quite yet.
Lex Fridman (03:09.280)
Well, it was good to tell you,
George Hotz (03:10.440)
as you'd ask the IQ required to destroy the world
Lex Fridman (03:13.520)
falls by one point every year.
George Hotz (03:15.440)
Okay.
Lex Fridman (03:16.280)
Technology democratizes the destruction of the world.
Lex Fridman (03:21.120)
When can a meme destroy the world?
Lex Fridman (03:23.120)
It kind of is already, right?
George Hotz (03:27.280)
Somewhat.
Lex Fridman (03:28.480)
I don't think we've seen anywhere near the worst of it yet.
George Hotz (03:32.240)
Well, it's going to get weird.
Lex Fridman (03:34.000)
Well, maybe a meme can save the world.
Lex Fridman (03:36.480)
You thought about that?
Lex Fridman (03:37.480)
The meme Lord Elon Musk fighting on the side of good
George Hotz (03:40.800)
versus the meme Lord of the darkness,
Lex Fridman (03:44.560)
which is not saying anything bad about Donald Trump,
Lex Fridman (03:48.280)
but he is the Lord of the meme on the dark side.
Lex Fridman (03:51.720)
He's a Darth Vader of memes.
George Hotz (03:53.760)
I think in every fairy tale they always end it with,
Lex Fridman (03:58.360)
and they lived happily ever after.
Lex Fridman (03:59.920)
And I'm like, please tell me more
Lex Fridman (04:00.960)
about this happily ever after.
George Hotz (04:02.440)
I've heard 50% of marriages end in divorce.
Lex Fridman (04:05.880)
Why doesn't your marriage end up there?
George Hotz (04:07.840)
You can't just say happily ever after.
Lex Fridman (04:09.280)
So it's the thing about destruction
George Hotz (04:12.280)
is it's over after the destruction.
Lex Fridman (04:14.840)
We have to do everything right in order to avoid it.
Lex Fridman (04:18.160)
And one thing wrong,
Lex Fridman (04:20.040)
I mean, actually this is what I really like
George Hotz (04:21.920)
about cryptography.
Lex Fridman (04:22.960)
Cryptography, it seems like we live in a world
George Hotz (04:24.600)
where the defense wins versus like nuclear weapons.
Lex Fridman (04:29.640)
The opposite is true.
George Hotz (04:30.920)
It is much easier to build a warhead
Lex Fridman (04:32.960)
that splits into a hundred little warheads
George Hotz (04:34.520)
than to build something that can, you know,
Lex Fridman (04:36.440)
take out a hundred little warheads.
George Hotz (04:38.880)
The offense has the advantage there.
Lex Fridman (04:41.400)
So maybe our future is in crypto, but.
Lex Fridman (04:44.520)
So cryptography, right.
Lex Fridman (04:45.720)
The Goliath is the defense.
Lex Fridman (04:49.760)
And then all the different hackers are the Davids.
Lex Fridman (04:54.280)
And that equation is flipped for nuclear war.
George Hotz (04:57.840)
Cause there's so many,
Lex Fridman (04:58.800)
like one nuclear weapon destroys everything essentially.
George Hotz (05:01.960)
Yeah, and it is much easier to attack with a nuclear weapon
Lex Fridman (05:06.960)
than it is to like the technology required to intercept
Lex Fridman (05:09.640)
and destroy a rocket is much more complicated
Lex Fridman (05:12.080)
than the technology required to just, you know,
George Hotz (05:13.800)
orbital trajectory, send a rocket to somebody.
Lex Fridman (05:17.480)
So, okay.
George Hotz (05:18.520)
Your intuition that there were intelligent civilizations
Lex Fridman (05:21.880)
out there, but it's very possible
George Hotz (05:24.360)
that they're no longer there.
Lex Fridman (05:26.240)
That's kind of a sad picture.
George Hotz (05:27.640)
They enter some steady state.
Lex Fridman (05:29.520)
They all wirehead themselves.
Lex Fridman (05:31.520)
What's wirehead?
Lex Fridman (05:33.360)
Stimulate, stimulate their pleasure centers
Lex Fridman (05:35.320)
and just, you know, live forever in this kind of stasis.
Lex Fridman (05:39.680)
They become, well, I mean,
Lex Fridman (05:42.560)
I think the reason I believe this is because where are they?
Lex Fridman (05:46.320)
If there's some reason they stopped expanding,
George Hotz (05:50.680)
cause otherwise they would have taken over the universe.
Lex Fridman (05:52.160)
The universe isn't that big.
George Hotz (05:53.440)
Or at least, you know,
Lex Fridman (05:54.280)
let's just talk about the galaxy, right?
George Hotz (05:56.120)
That's 70,000 light years across.
Lex Fridman (05:58.720)
I took that number from Star Trek Voyager.
George Hotz (05:59.920)
I don't know how true it is, but yeah, that's not big.
Lex Fridman (06:04.680)
Right? 70,000 light years is nothing.
George Hotz (06:07.320)
For some possible technology that you can imagine
Lex Fridman (06:10.040)
that can leverage like wormholes or something like that.
George Hotz (06:12.320)
Or you don't even need wormholes.
Lex Fridman (06:13.320)
Just a von Neumann probe is enough.
George Hotz (06:15.120)
A von Neumann probe and a million years of sublight travel
Lex Fridman (06:18.440)
and you'd have taken over the whole universe.
George Hotz (06:20.440)
That clearly didn't happen.
Lex Fridman (06:22.480)
So something stopped it.
Lex Fridman (06:24.000)
So you mean if you, right,
Lex Fridman (06:25.400)
for like a few million years,
George Hotz (06:27.040)
if you sent out probes that travel close,
Lex Fridman (06:29.880)
what's sublight?
Lex Fridman (06:30.720)
You mean close to the speed of light?
Lex Fridman (06:32.280)
Let's say 0.1 C.
Lex Fridman (06:33.720)
And it just spreads.
Lex Fridman (06:34.800)
Interesting.
Lex Fridman (06:35.640)
Actually, that's an interesting calculation, huh?
Lex Fridman (06:38.800)
So what makes you think that we'd be able
Lex Fridman (06:40.640)
to communicate with them?
Lex Fridman (06:42.280)
Like, yeah, what's,
Lex Fridman (06:45.160)
why do you think we would be able to be able
Lex Fridman (06:47.920)
to comprehend intelligent lives that are out there?
George Hotz (06:51.920)
Like even if they were among us kind of thing,
Lex Fridman (06:54.960)
like, or even just flying around?
George Hotz (06:57.600)
Well, I mean, that's possible.
Lex Fridman (07:01.200)
It's possible that there is some sort of prime directive.
George Hotz (07:04.640)
That'd be a really cool universe to live in.
Lex Fridman (07:07.040)
And there's some reason
George Hotz (07:08.000)
they're not making themselves visible to us.
Lex Fridman (07:10.920)
But it makes sense that they would use the same,
George Hotz (07:15.200)
well, at least the same entropy.
Lex Fridman (07:16.960)
Well, you're implying the same laws of physics.
George Hotz (07:18.800)
I don't know what you mean by entropy in this case.
Lex Fridman (07:20.800)
Oh, yeah.
George Hotz (07:21.920)
I mean, if entropy is the scarce resource in the universe.
Lex Fridman (07:25.040)
So what do you think about like Stephen Wolfram
Lex Fridman (07:26.960)
and everything is a computation?
Lex Fridman (07:28.840)
And then what if they are traveling through
Lex Fridman (07:31.440)
this world of computation?
Lex Fridman (07:32.640)
So if you think of the universe
George Hotz (07:34.240)
as just information processing,
Lex Fridman (07:36.600)
then what you're referring to with entropy
Lex Fridman (07:40.840)
and then these pockets of interesting complex computation
Lex Fridman (07:44.240)
swimming around, how do we know they're not already here?
Lex Fridman (07:47.160)
How do we know that this,
Lex Fridman (07:51.040)
like all the different amazing things
George Hotz (07:53.040)
that are full of mystery on earth
Lex Fridman (07:55.080)
are just like little footprints of intelligence
Lex Fridman (07:58.640)
from light years away?
Lex Fridman (08:01.160)
Maybe.
George Hotz (08:02.840)
I mean, I tend to think that as civilizations expand,
Lex Fridman (08:05.760)
they use more and more energy
Lex Fridman (08:07.800)
and you can never overcome the problem of waste heat.
Lex Fridman (08:10.240)
So where is there waste heat?
Lex Fridman (08:11.880)
So we'd be able to, with our crude methods,
Lex Fridman (08:13.560)
be able to see like, there's a whole lot of energy here.
Lex Fridman (08:18.840)
But it could be something we're not,
Lex Fridman (08:20.560)
I mean, we don't understand dark energy, right?
George Hotz (08:22.480)
Dark matter.
Lex Fridman (08:23.560)
It could be just stuff we don't understand at all.
George Hotz (08:26.160)
Or they can have a fundamentally different physics,
Lex Fridman (08:29.080)
you know, like that we just don't even comprehend.
George Hotz (08:32.200)
Well, I think, okay,
Lex Fridman (08:33.440)
I mean, it depends how far out you wanna go.
George Hotz (08:35.120)
I don't think physics is very different
Lex Fridman (08:36.840)
on the other side of the galaxy.
George Hotz (08:39.760)
I would suspect that they have,
Lex Fridman (08:41.920)
I mean, if they're in our universe,
George Hotz (08:43.680)
they have the same physics.
Lex Fridman (08:45.760)
Well, yeah, that's the assumption we have,
Lex Fridman (08:47.600)
but there could be like super trippy things
Lex Fridman (08:50.000)
like our cognition only gets to a slice,
Lex Fridman (08:57.040)
and all the possible instruments that we can design
Lex Fridman (08:59.440)
only get to a particular slice of the universe.
Lex Fridman (09:01.520)
And there's something much like weirder.
Lex Fridman (09:04.080)
Maybe we can try a thought experiment.
George Hotz (09:06.880)
Would people from the past
Lex Fridman (09:10.040)
be able to detect the remnants of our,
Lex Fridman (09:14.000)
or would we be able to detect our modern civilization?
Lex Fridman (09:16.600)
I think the answer is obviously yes.
Lex Fridman (09:18.840)
You mean past from a hundred years ago?
Lex Fridman (09:20.720)
Well, let's even go back further.
Lex Fridman (09:22.080)
Let's go to a million years ago, right?
Lex Fridman (09:24.400)
The humans who were lying around in the desert
George Hotz (09:26.520)
probably didn't even have,
Lex Fridman (09:27.720)
maybe they just barely had fire.
George Hotz (09:31.360)
They would understand if a 747 flew overhead.
Lex Fridman (09:35.200)
Oh, in this vicinity, but not if a 747 flew on Mars.
George Hotz (09:43.200)
Like, cause they wouldn't be able to see far,
Lex Fridman (09:45.080)
cause we're not actually communicating that well
George Hotz (09:47.240)
with the rest of the universe.
Lex Fridman (09:48.920)
We're doing okay.
George Hotz (09:50.160)
Just sending out random like fifties tracks of music.
Lex Fridman (09:54.200)
True.
Lex Fridman (09:55.160)
And yeah, I mean, they'd have to, you know,
Lex Fridman (09:57.800)
we've only been broadcasting radio waves for 150 years.
Lex Fridman (10:02.480)
And well, there's your light cone.
Lex Fridman (10:04.560)
So.
George Hotz (10:05.800)
Yeah. Okay.
Lex Fridman (10:06.920)
What do you make about all the,
George Hotz (10:08.800)
I recently came across this having talked to David Fravor.
Lex Fridman (10:14.720)
I don't know if you caught what the videos
George Hotz (10:16.960)
of the Pentagon released
Lex Fridman (10:18.840)
and the New York Times reporting of the UFO sightings.
Lex Fridman (10:23.480)
So I kind of looked into it, quote unquote.
Lex Fridman (10:26.040)
And there's actually been like hundreds
Lex Fridman (10:30.800)
of thousands of UFO sightings, right?
Lex Fridman (10:33.800)
And a lot of it you can explain away
George Hotz (10:35.960)
in different kinds of ways.
Lex Fridman (10:37.080)
So one is it could be interesting physical phenomena.
George Hotz (10:40.160)
Two, it could be people wanting to believe
Lex Fridman (10:44.640)
and therefore they conjure up a lot of different things
George Hotz (10:46.760)
that just, you know, when you see different kinds of lights,
Lex Fridman (10:48.920)
some basic physics phenomena,
Lex Fridman (10:50.760)
and then you just conjure up ideas
Lex Fridman (10:53.640)
of possible out there mysterious worlds.
George Hotz (10:56.680)
But, you know, it's also possible,
Lex Fridman (10:58.960)
like you have a case of David Fravor,
George Hotz (11:02.480)
who is a Navy pilot, who's, you know,
Lex Fridman (11:06.160)
as legit as it gets in terms of humans
George Hotz (11:08.840)
who are able to perceive things in the environment
Lex Fridman (11:13.440)
and make conclusions,
George Hotz (11:15.320)
whether those things are a threat or not.
Lex Fridman (11:17.600)
And he and several other pilots saw a thing,
George Hotz (11:22.000)
I don't know if you followed this,
Lex Fridman (11:23.440)
but they saw a thing that they've since then called TikTok
George Hotz (11:26.840)
that moved in all kinds of weird ways.
Lex Fridman (11:29.520)
They don't know what it is.
George Hotz (11:30.640)
It could be technology developed by the United States
Lex Fridman (11:36.640)
and they're just not aware of it
Lex Fridman (11:38.040)
and the surface level from the Navy, right?
Lex Fridman (11:40.000)
It could be different kind of lighting technology
George Hotz (11:42.280)
or drone technology, all that kind of stuff.
Lex Fridman (11:45.000)
It could be the Russians and the Chinese,
George Hotz (11:46.560)
all that kind of stuff.
Lex Fridman (11:48.000)
And of course their mind, our mind,
George Hotz (11:51.400)
can also venture into the possibility
Lex Fridman (11:54.160)
that it's from another world.
Lex Fridman (11:56.320)
Have you looked into this at all?
Lex Fridman (11:58.160)
What do you think about it?
George Hotz (11:59.640)
I think all the news is a psyop.
Lex Fridman (12:01.680)
I think that the most plausible.
George Hotz (12:05.240)
Nothing is real.
Lex Fridman (12:06.440)
Yeah, I listened to the, I think it was Bob Lazar
George Hotz (12:10.920)
on Joe Rogan.
Lex Fridman (12:12.360)
And like, I believe everything this guy is saying.
Lex Fridman (12:15.840)
And then I think that it's probably just some like MKUltra
Lex Fridman (12:18.440)
kind of thing, you know?
Lex Fridman (12:20.680)
What do you mean?
Lex Fridman (12:21.520)
Like they, you know, they made some weird thing
Lex Fridman (12:24.720)
and they called it an alien spaceship.
Lex Fridman (12:26.160)
You know, maybe it was just to like
George Hotz (12:27.480)
stimulate young physicists minds.
Lex Fridman (12:29.560)
We'll tell them it's alien technology
Lex Fridman (12:31.040)
and we'll see what they come up with, right?
Lex Fridman (12:33.600)
Do you find any conspiracy theories compelling?
George Hotz (12:36.080)
Like have you pulled at the string
Lex Fridman (12:38.320)
of the rich complex world of conspiracy theories
Lex Fridman (12:42.440)
that's out there?
Lex Fridman (12:43.920)
I think that I've heard a conspiracy theory
George Hotz (12:46.520)
that conspiracy theories were invented by the CIA
Lex Fridman (12:48.960)
in the 60s to discredit true things.
George Hotz (12:52.040)
Yeah.
Lex Fridman (12:53.880)
So, you know, you can go to ridiculous conspiracy theories
George Hotz (12:58.520)
like Flat Earth and Pizza Gate.
Lex Fridman (13:01.000)
And, you know, these things are almost to hide
George Hotz (13:05.440)
like conspiracy theories that like,
Lex Fridman (13:08.000)
you know, remember when the Chinese like locked up
Lex Fridman (13:09.640)
the doctors who discovered coronavirus?
Lex Fridman (13:11.360)
Like I tell people this and I'm like,
George Hotz (13:12.840)
no, no, no, that's not a conspiracy theory.
Lex Fridman (13:14.400)
That actually happened.
Lex Fridman (13:15.920)
Do you remember the time that the money used to be backed
Lex Fridman (13:18.000)
by gold and now it's backed by nothing?
George Hotz (13:20.080)
This is not a conspiracy theory.
Lex Fridman (13:21.680)
This actually happened.
George Hotz (13:23.800)
Well, that's one of my worries today
Lex Fridman (13:26.360)
with the idea of fake news is that when nothing is real,
George Hotz (13:32.880)
then like you dilute the possibility of anything being true
Lex Fridman (13:37.600)
by conjuring up all kinds of conspiracy theories.
Lex Fridman (13:41.000)
And then you don't know what to believe.
Lex Fridman (13:42.400)
And then like the idea of truth of objectivity
George Hotz (13:46.280)
is lost completely.
Lex Fridman (13:47.840)
Everybody has their own truth.
Lex Fridman (13:50.120)
So you used to control information by censoring it.
Lex Fridman (13:53.600)
And then the internet happened and governments were like,
George Hotz (13:55.880)
oh shit, we can't censor things anymore.
Lex Fridman (13:58.360)
I know what we'll do.
George Hotz (14:00.560)
You know, it's the old story of the story of like
Lex Fridman (14:04.160)
tying a flag with a leprechaun tells you his gold is buried
Lex Fridman (14:07.080)
and you tie one flag and you make the leprechaun swear
Lex Fridman (14:09.080)
to not remove the flag.
Lex Fridman (14:10.040)
And you come back to the field later with a shovel
Lex Fridman (14:11.720)
and there's flags everywhere.
Lex Fridman (14:14.640)
That's one way to maintain privacy, right?
Lex Fridman (14:16.320)
It's like in order to protect the contents
George Hotz (14:20.280)
of this conversation, for example,
Lex Fridman (14:21.800)
we could just generate like millions of deep,
George Hotz (14:25.120)
fake conversations where you and I talk
Lex Fridman (14:27.560)
and say random things.
Lex Fridman (14:29.120)
So this is just one of them
Lex Fridman (14:30.440)
and nobody knows which one was the real one.
George Hotz (14:32.720)
This could be fake right now.
Lex Fridman (14:34.440)
Classic steganography technique.
George Hotz (14:37.240)
Okay, another absurd question about intelligent life.
Lex Fridman (14:39.960)
Cause you know, you're an incredible programmer
George Hotz (14:43.960)
outside of everything else we'll talk about
Lex Fridman (14:45.520)
just as a programmer.
Lex Fridman (14:49.320)
Do you think intelligent beings out there,
Lex Fridman (14:52.960)
the civilizations that were out there,
Lex Fridman (14:54.520)
had computers and programming?
Lex Fridman (14:58.240)
Did they, do we naturally have to develop something
George Hotz (15:01.440)
where we engineer machines and are able to encode
Lex Fridman (15:05.480)
both knowledge into those machines
Lex Fridman (15:08.680)
and instructions that process that knowledge,
Lex Fridman (15:11.720)
process that information to make decisions
Lex Fridman (15:14.260)
and actions and so on?
Lex Fridman (15:15.680)
And would those programming languages,
George Hotz (15:18.320)
if you think they exist, be at all similar
Lex Fridman (15:21.320)
to anything we've developed?
Lex Fridman (15:24.120)
So I don't see that much of a difference
Lex Fridman (15:26.600)
between quote unquote natural languages
Lex Fridman (15:29.400)
and programming languages.
Lex Fridman (15:34.160)
Yeah.
George Hotz (15:35.000)
I think there's so many similarities.
Lex Fridman (15:36.680)
So when asked the question,
Lex Fridman (15:39.920)
what do alien languages look like?
Lex Fridman (15:42.380)
I imagine they're not all that dissimilar from ours.
Lex Fridman (15:46.480)
And I think translating in and out of them
Lex Fridman (15:51.560)
wouldn't be that crazy.
George Hotz (15:52.920)
Well, it's difficult to compile like DNA to Python
Lex Fridman (15:57.520)
and then to C.
George Hotz (15:59.160)
There's a little bit of a gap in the kind of languages
Lex Fridman (16:02.000)
we use for touring machines
Lex Fridman (16:06.840)
and the kind of languages nature seems to use a little bit.
Lex Fridman (16:10.160)
Maybe that's just, we just haven't understood
George Hotz (16:13.800)
the kind of language that nature uses well yet.
Lex Fridman (16:16.440)
DNA is a CAD model.
George Hotz (16:19.120)
It's not quite a programming language.
Lex Fridman (16:21.200)
It has no sort of a serial execution.
George Hotz (16:25.280)
It's not quite a, yeah, it's a CAD model.
Lex Fridman (16:29.480)
So I think in that sense,
George Hotz (16:30.920)
we actually completely understand it.
Lex Fridman (16:32.520)
The problem is, well, simulating on these CAD models,
George Hotz (16:37.300)
I played with it a bit this year,
Lex Fridman (16:38.520)
is super computationally intensive.
George Hotz (16:41.120)
If you wanna go down to like the molecular level
Lex Fridman (16:43.720)
where you need to go to see a lot of these phenomenon
George Hotz (16:45.880)
like protein folding.
Lex Fridman (16:48.080)
So yeah, it's not that we don't understand it.
George Hotz (16:52.200)
It just requires a whole lot of compute to kind of compile it.
Lex Fridman (16:55.160)
For our human minds, it's inefficient,
George Hotz (16:56.680)
both for the data representation and for the programming.
Lex Fridman (17:00.520)
Yeah, it runs well on raw nature.
George Hotz (17:02.800)
It runs well on raw nature.
Lex Fridman (17:03.880)
And when we try to build emulators or simulators for that,
George Hotz (17:07.840)
well, they're mad slow, but I've tried it.
Lex Fridman (17:10.640)
It runs in that, yeah, you've commented elsewhere,
George Hotz (17:14.280)
I don't remember where,
Lex Fridman (17:15.780)
that one of the problems is simulating nature is tough.
Lex Fridman (17:20.780)
And if you want to sort of deploy a prototype,
Lex Fridman (17:25.300)
I forgot how you put it, but it made me laugh,
Lex Fridman (17:28.020)
but animals or humans would need to be involved
Lex Fridman (17:31.700)
in order to try to run some prototype code on,
George Hotz (17:38.120)
like if we're talking about COVID and viruses and so on,
Lex Fridman (17:41.180)
if you were trying to engineer
George Hotz (17:42.900)
some kind of defense mechanisms,
Lex Fridman (17:45.020)
like a vaccine against COVID and all that kind of stuff
George Hotz (17:49.580)
that doing any kind of experimentation,
Lex Fridman (17:52.060)
like you can with like autonomous vehicles
George Hotz (17:53.980)
would be very technically and ethically costly.
Lex Fridman (17:59.660)
I'm not sure about that.
George Hotz (18:00.940)
I think you can do tons of crazy biology and test tubes.
Lex Fridman (18:05.060)
I think my bigger complaint is more,
George Hotz (18:08.480)
oh, the tools are so bad.
Lex Fridman (18:11.420)
Like literally, you mean like libraries and?
George Hotz (18:14.540)
I don't know, I'm not pipetting shit.
Lex Fridman (18:16.100)
Like you're handing me a, I got a, no, no, no,
George Hotz (18:20.140)
there has to be some.
Lex Fridman (18:22.580)
Like automating stuff.
Lex Fridman (18:24.140)
And like the, yeah, but human biology is messy.
Lex Fridman (18:28.300)
Like it seems.
Lex Fridman (18:29.140)
But like, look at those Toronto's videos.
Lex Fridman (18:31.240)
They were a joke.
George Hotz (18:32.080)
It's like a little gantry.
Lex Fridman (18:33.460)
It's like little X, Y gantry,
George Hotz (18:34.660)
high school science project with the pipette.
Lex Fridman (18:36.660)
I'm like, really?
George Hotz (18:38.260)
Gotta be something better.
Lex Fridman (18:39.220)
You can't build like nice microfluidics
Lex Fridman (18:41.420)
and I can program the computation to bio interface.
Lex Fridman (18:45.300)
I mean, this is gonna happen.
Lex Fridman (18:47.180)
But like right now, if you are asking me
Lex Fridman (18:50.020)
to pipette 50 milliliters of solution, I'm out.
George Hotz (18:54.460)
This is so crude.
Lex Fridman (18:55.460)
Yeah.
George Hotz (18:56.700)
Okay, let's get all the crazy out of the way.
Lex Fridman (18:59.980)
So a bunch of people asked me,
George Hotz (19:02.260)
since we talked about the simulation last time,
Lex Fridman (19:05.060)
we talked about hacking the simulation.
Lex Fridman (19:06.860)
Do you have any updates, any insights
Lex Fridman (19:09.900)
about how we might be able to go about hacking simulation
Lex Fridman (19:13.740)
if we indeed do live in a simulation?
Lex Fridman (19:17.180)
I think a lot of people misinterpreted
George Hotz (19:19.900)
the point of that South by talk.
Lex Fridman (19:22.420)
The point of the South by talk
George Hotz (19:23.500)
was not literally to hack the simulation.
Lex Fridman (19:26.720)
I think that this is an idea is literally just,
George Hotz (19:33.520)
I think theoretical physics.
Lex Fridman (19:34.580)
I think that's the whole goal, right?
George Hotz (19:39.900)
You want your grand unified theory, but then, okay,
Lex Fridman (19:42.300)
build a grand unified theory search for exploits, right?
George Hotz (19:45.140)
I think we're nowhere near actually there yet.
Lex Fridman (19:47.640)
My hope with that was just more to like,
George Hotz (19:51.500)
are you people kidding me
Lex Fridman (19:52.740)
with the things you spend time thinking about?
Lex Fridman (19:54.980)
Do you understand like kind of how small you are?
Lex Fridman (19:58.020)
You are bytes and God's computer, really?
Lex Fridman (1:00:00.140)
So that's the interesting thing is you're not doing,
Lex Fridman (1:00:03.380)
you don't appear to be, maybe you can correct me,
George Hotz (1:00:05.380)
you don't appear to be doing lane detection
Lex Fridman (1:00:08.700)
or lane marking detection or kind of the segmentation task
George Hotz (1:00:12.420)
or any kind of object detection task.
Lex Fridman (1:00:15.100)
You're doing what's traditionally more called
George Hotz (1:00:17.580)
like end to end learning.
Lex Fridman (1:00:19.420)
So, and trained on actual behavior of drivers
George Hotz (1:00:24.180)
when they're driving the car manually.
Lex Fridman (1:00:27.780)
And this is hard to do.
George Hotz (1:00:29.820)
It's not supervised learning.
Lex Fridman (1:00:32.220)
Yeah, but so the nice thing is there's a lot of data.
Lex Fridman (1:00:34.740)
So it's hard and easy, right?
Lex Fridman (1:00:37.060)
It's a...
George Hotz (1:00:37.900)
We have a lot of high quality data, yeah.
Lex Fridman (1:00:40.020)
Like more than you need in the second.
George Hotz (1:00:41.700)
Well...
Lex Fridman (1:00:42.700)
We have way more than we do.
George Hotz (1:00:43.520)
We have way more data than we need.
Lex Fridman (1:00:44.980)
I mean, it's an interesting question actually,
George Hotz (1:00:47.040)
because in terms of amount, you have more than you need,
Lex Fridman (1:00:50.440)
but the driving is full of edge cases.
Lex Fridman (1:00:54.260)
So how do you select the data you train on?
Lex Fridman (1:00:58.260)
I think this is an interesting open question.
Lex Fridman (1:01:00.540)
Like what's the cleverest way to select data?
Lex Fridman (1:01:04.220)
That's the question Tesla is probably working on.
George Hotz (1:01:07.660)
That's, I mean, the entirety of machine learning can be,
Lex Fridman (1:01:09.900)
they don't seem to really care.
George Hotz (1:01:11.060)
They just kind of select data.
Lex Fridman (1:01:12.260)
But I feel like that if you want to solve,
George Hotz (1:01:14.860)
if you want to create intelligent systems,
Lex Fridman (1:01:16.220)
you have to pick data well, right?
Lex Fridman (1:01:18.900)
And so do you have any hints, ideas of how to do it well?
Lex Fridman (1:01:22.900)
So in some ways that is...
George Hotz (1:01:25.140)
The definition I like of reinforcement learning
Lex Fridman (1:01:27.400)
versus supervised learning.
George Hotz (1:01:29.300)
In supervised learning, the weights depend on the data.
Lex Fridman (1:01:32.720)
Right?
Lex Fridman (1:01:34.560)
And this is obviously true,
Lex Fridman (1:01:35.700)
but in reinforcement learning,
George Hotz (1:01:38.320)
the data depends on the weights.
Lex Fridman (1:01:40.420)
Yeah.
Lex Fridman (1:01:41.380)
And actually both ways.
Lex Fridman (1:01:42.540)
That's poetry.
Lex Fridman (1:01:43.980)
So how does it know what data to train on?
Lex Fridman (1:01:46.260)
Well, let it pick.
George Hotz (1:01:47.540)
We're not there yet, but that's the eventual.
Lex Fridman (1:01:49.460)
So you're thinking this almost like
George Hotz (1:01:51.100)
a reinforcement learning framework.
Lex Fridman (1:01:53.320)
We're going to do RL on the world.
George Hotz (1:01:55.380)
Every time a car makes a mistake, user disengages,
Lex Fridman (1:01:58.140)
we train on that and do RL on the world.
Lex Fridman (1:02:00.140)
Ship out a new model, that's an epoch, right?
Lex Fridman (1:02:03.260)
And for now you're not doing the Elon style promising
George Hotz (1:02:08.540)
that it's going to be fully autonomous.
Lex Fridman (1:02:09.700)
You really are sticking to level two
Lex Fridman (1:02:12.420)
and like it's supposed to be supervised.
Lex Fridman (1:02:15.300)
It is definitely supposed to be supervised
Lex Fridman (1:02:16.900)
and we enforce the fact that it's supervised.
Lex Fridman (1:02:19.740)
We look at our rate of improvement in disengagements.
George Hotz (1:02:23.660)
OpenPilot now has an unplanned disengagement
Lex Fridman (1:02:25.700)
about every a hundred miles.
George Hotz (1:02:27.440)
This is up from 10 miles, like maybe,
Lex Fridman (1:02:32.580)
maybe maybe a year ago.
George Hotz (1:02:36.560)
Yeah.
Lex Fridman (1:02:37.400)
So maybe we've seen 10 X improvement in a year,
Lex Fridman (1:02:38.500)
but a hundred miles is still a far cry
Lex Fridman (1:02:41.740)
from the a hundred thousand you're going to need.
Lex Fridman (1:02:43.880)
So you're going to somehow need to get three more 10 Xs
Lex Fridman (1:02:48.060)
in there.
Lex Fridman (1:02:49.880)
And you're, what's your intuition?
Lex Fridman (1:02:52.300)
You're basically hoping that there's exponential
George Hotz (1:02:54.540)
improvement built into the baked into the cake somewhere.
Lex Fridman (1:02:56.940)
Well, that's even, I mean, 10 X improvement,
Lex Fridman (1:02:58.420)
that's already assuming exponential, right?
Lex Fridman (1:03:00.540)
There's definitely exponential improvement.
Lex Fridman (1:03:02.620)
And I think when Elon talks about exponential,
Lex Fridman (1:03:04.340)
like these things, these systems are going to
George Hotz (1:03:06.340)
exponentially improve, just exponential doesn't mean
Lex Fridman (1:03:09.820)
you're getting a hundred gigahertz processors tomorrow.
George Hotz (1:03:12.660)
Right? Like it's going to still take a while
Lex Fridman (1:03:15.060)
because the gap between even our best system
Lex Fridman (1:03:18.340)
and humans is still large.
Lex Fridman (1:03:20.300)
So that's an interesting distinction to draw.
Lex Fridman (1:03:22.340)
So if you look at the way Tesla is approaching the problem
Lex Fridman (1:03:26.100)
and the way you're approaching the problem,
George Hotz (1:03:28.340)
which is very different than the rest of the self driving
Lex Fridman (1:03:31.780)
car world.
Lex Fridman (1:03:32.720)
So let's put them aside is you're treating most
Lex Fridman (1:03:35.380)
the driving task as a machine learning problem.
Lex Fridman (1:03:37.500)
And the way Tesla is approaching it is with the multitask
Lex Fridman (1:03:40.260)
learning where you break the task of driving into hundreds
George Hotz (1:03:44.100)
of different tasks and you have this multiheaded
Lex Fridman (1:03:47.180)
neural network that's very good at performing each task.
Lex Fridman (1:03:51.660)
And there there's presumably something on top that's
Lex Fridman (1:03:54.620)
stitching stuff together in order to make control
George Hotz (1:03:59.240)
decisions, policy decisions about how you move the car.
Lex Fridman (1:04:02.180)
But what that allows you, there's a brilliance to this
George Hotz (1:04:04.420)
because it allows you to master each task,
Lex Fridman (1:04:08.360)
like lane detection, stop sign detection,
George Hotz (1:04:13.380)
the traffic light detection, drivable area segmentation,
Lex Fridman (1:04:19.180)
you know, vehicle, bicycle, pedestrian detection.
George Hotz (1:04:23.000)
There's some localization tasks in there.
Lex Fridman (1:04:25.420)
Also predicting of like, yeah,
George Hotz (1:04:30.440)
predicting how the entities in the scene are going to move.
Lex Fridman (1:04:34.060)
Like everything is basically a machine learning task.
Lex Fridman (1:04:36.360)
So there's a classification, segmentation, prediction.
Lex Fridman (1:04:40.380)
And it's nice because you can have this entire engine,
George Hotz (1:04:44.460)
data engine that's mining for edge cases for each one of
Lex Fridman (1:04:48.740)
these tasks.
Lex Fridman (1:04:49.580)
And you can have people like engineers that are basically
Lex Fridman (1:04:52.200)
masters of that task,
George Hotz (1:04:53.820)
like become the best person in the world at,
Lex Fridman (1:04:56.600)
as you talk about the cone guy for Waymo,
George Hotz (1:04:59.860)
the becoming the best person in the world at cone detection.
Lex Fridman (1:05:06.700)
So that's a compelling notion from a supervised learning
George Hotz (1:05:10.140)
perspective, automating much of the process of edge case
Lex Fridman (1:05:15.380)
discovery and retraining neural network for each of the
George Hotz (1:05:17.820)
individual perception tasks.
Lex Fridman (1:05:19.780)
And then you're looking at the machine learning in a more
George Hotz (1:05:22.460)
holistic way, basically doing end to end learning on the
Lex Fridman (1:05:27.240)
driving tasks, supervised, trained on the data of the
George Hotz (1:05:31.500)
actual driving of people.
Lex Fridman (1:05:34.420)
They use comma AI, like actual human drivers,
George Hotz (1:05:37.580)
their manual control,
Lex Fridman (1:05:38.700)
plus the moments of disengagement that maybe with some
George Hotz (1:05:44.400)
labeling could indicate the failure of the system.
Lex Fridman (1:05:47.340)
So you have the,
George Hotz (1:05:48.900)
you have a huge amount of data for positive control of the
Lex Fridman (1:05:52.420)
vehicle, like successful control of the vehicle,
George Hotz (1:05:55.560)
both maintaining the lane as,
Lex Fridman (1:05:58.340)
as I think you're also working on longitudinal control of
George Hotz (1:06:01.060)
the vehicle and then failure cases where the vehicle does
Lex Fridman (1:06:04.700)
something wrong that needs disengagement.
Lex Fridman (1:06:08.380)
So like what,
Lex Fridman (1:06:09.820)
why do you think you're right and Tesla is wrong on this?
Lex Fridman (1:06:14.220)
And do you think,
Lex Fridman (1:06:15.660)
do you think you'll come around the Tesla way?
Lex Fridman (1:06:17.540)
Do you think Tesla will come around to your way?
Lex Fridman (1:06:21.400)
If you were to start a chess engine company,
Lex Fridman (1:06:23.920)
would you hire a Bishop guy?
Lex Fridman (1:06:26.080)
See, we have a,
George Hotz (1:06:27.660)
this is Monday morning.
Lex Fridman (1:06:29.060)
Quarterbacking is a yes, probably.
George Hotz (1:06:36.340)
Oh, our Rook guy.
Lex Fridman (1:06:37.400)
Oh, we stole the Rook guy from that company.
George Hotz (1:06:39.420)
Oh, we're going to have real good Rooks.
Lex Fridman (1:06:40.780)
Well, there's not many pieces, right?
George Hotz (1:06:43.900)
You can,
Lex Fridman (1:06:46.220)
there's not many guys and gals to hire.
George Hotz (1:06:48.820)
You just have a few that work in the Bishop,
Lex Fridman (1:06:51.060)
a few that work in the Rook.
George Hotz (1:06:52.860)
Is that not ludicrous today to think about
Lex Fridman (1:06:55.340)
in a world of AlphaZero?
Lex Fridman (1:06:57.520)
But AlphaZero is a chess game.
Lex Fridman (1:06:58.860)
So the fundamental question is,
Lex Fridman (1:07:01.780)
how hard is driving compared to chess?
Lex Fridman (1:07:04.340)
Because, so long term,
George Hotz (1:07:07.200)
end to end,
Lex Fridman (1:07:08.980)
will be the right solution.
Lex Fridman (1:07:10.620)
The question is how many years away is that?
Lex Fridman (1:07:13.460)
End to end is going to be the only solution for level five.
George Hotz (1:07:15.740)
For the only way we'll get there.
Lex Fridman (1:07:17.220)
Of course, and of course,
George Hotz (1:07:18.220)
Tesla is going to come around to my way.
Lex Fridman (1:07:19.740)
And if you're a Rook guy out there, I'm sorry.
George Hotz (1:07:22.980)
The cone guy.
Lex Fridman (1:07:24.980)
I don't know.
George Hotz (1:07:25.820)
We're going to specialize each task.
Lex Fridman (1:07:26.940)
We're going to really understand Rook placement.
George Hotz (1:07:29.100)
Yeah.
Lex Fridman (1:07:30.540)
I understand the intuition you have.
George Hotz (1:07:32.060)
I mean, that,
Lex Fridman (1:07:35.100)
that is a very compelling notion
George Hotz (1:07:36.820)
that we can learn the task end to end,
Lex Fridman (1:07:39.140)
like the same compelling notion you might have
George Hotz (1:07:40.820)
for natural language conversation.
Lex Fridman (1:07:42.620)
But I'm not
George Hotz (1:07:44.800)
sure,
Lex Fridman (1:07:47.060)
because one thing you sneaked in there
George Hotz (1:07:48.940)
is the assertion that it's impossible to get to level five
Lex Fridman (1:07:53.180)
without this kind of approach.
George Hotz (1:07:55.420)
I don't know if that's obvious.
Lex Fridman (1:07:57.140)
I don't know if that's obvious either.
George Hotz (1:07:58.300)
I don't actually mean that.
Lex Fridman (1:08:01.340)
I think that it is much easier
George Hotz (1:08:03.500)
to get to level five with an end to end approach.
Lex Fridman (1:08:05.700)
I think that the other approach is doable,
Lex Fridman (1:08:08.900)
but the magnitude of the engineering challenge
Lex Fridman (1:08:11.380)
may exceed what humanity is capable of.
Lex Fridman (1:08:13.780)
But what do you think of the Tesla data engine approach,
Lex Fridman (1:08:19.140)
which to me is an active learning task,
George Hotz (1:08:21.100)
is kind of fascinating,
Lex Fridman (1:08:22.500)
is breaking it down into these multiple tasks
Lex Fridman (1:08:25.700)
and mining their data constantly for like edge cases
Lex Fridman (1:08:29.500)
for these different tasks.
George Hotz (1:08:30.500)
Yeah, but the tasks themselves are not being learned.
Lex Fridman (1:08:32.460)
This is feature engineering.
George Hotz (1:08:35.700)
Yeah, I mean, it's a higher abstraction level
Lex Fridman (1:08:40.740)
of feature engineering for the different tasks.
George Hotz (1:08:43.340)
Task engineering in a sense.
Lex Fridman (1:08:44.740)
It's slightly better feature engineering,
Lex Fridman (1:08:46.780)
but it's still fundamentally is feature engineering.
Lex Fridman (1:08:49.260)
And if anything about the history of AI
George Hotz (1:08:51.300)
has taught us anything,
Lex Fridman (1:08:52.620)
it's that feature engineering approaches
George Hotz (1:08:54.540)
will always be replaced and lose to end to end.
Lex Fridman (1:08:57.660)
Now, to be fair, I cannot really make promises on timelines,
Lex Fridman (1:09:02.060)
but I can say that when you look at the code for Stockfish
Lex Fridman (1:09:05.700)
and the code for AlphaZero,
George Hotz (1:09:06.940)
one is a lot shorter than the other,
Lex Fridman (1:09:09.060)
a lot more elegant,
George Hotz (1:09:09.900)
required a lot less programmer hours to write.
Lex Fridman (1:09:12.580)
Yeah, but there was a lot more murder of bad agents
George Hotz (1:09:21.620)
on the AlphaZero side.
Lex Fridman (1:09:24.740)
By murder, I mean agents that played a game
Lex Fridman (1:09:29.220)
and failed miserably.
Lex Fridman (1:09:30.420)
Yeah.
George Hotz (1:09:31.460)
Oh, oh.
Lex Fridman (1:09:32.300)
In simulation, that failure is less costly.
George Hotz (1:09:34.740)
Yeah.
Lex Fridman (1:09:35.580)
In real world, it's...
Lex Fridman (1:09:37.300)
Do you mean in practice,
Lex Fridman (1:09:38.260)
like AlphaZero has lost games miserably?
George Hotz (1:09:40.660)
No.
Lex Fridman (1:09:41.500)
Wow.
George Hotz (1:09:42.540)
I haven't seen that.
Lex Fridman (1:09:43.380)
No, but I know, but the requirement for AlphaZero is...
George Hotz (1:09:47.380)
A simulator.
Lex Fridman (1:09:48.220)
To be able to like evolution, human evolution,
George Hotz (1:09:51.500)
not human evolution, biological evolution of life on earth
Lex Fridman (1:09:54.420)
from the origin of life has murdered trillions
George Hotz (1:09:58.700)
upon trillions of organisms on the path thus humans.
Lex Fridman (1:10:02.260)
Yeah.
Lex Fridman (1:10:03.100)
So the question is, can we stitch together
Lex Fridman (1:10:05.860)
a human like object without having to go
Lex Fridman (1:10:07.940)
through the entirety process of evolution?
Lex Fridman (1:10:09.900)
Well, no, but do the evolution in simulation.
George Hotz (1:10:11.940)
Yeah, that's the question.
Lex Fridman (1:10:12.860)
Can we simulate?
Lex Fridman (1:10:13.700)
So do you have a sense that it's possible
Lex Fridman (1:10:15.060)
to simulate some aspect?
George Hotz (1:10:16.220)
MuZero is exactly this.
Lex Fridman (1:10:18.220)
MuZero is the solution to this.
George Hotz (1:10:21.300)
MuZero I think is going to be looked back
Lex Fridman (1:10:23.980)
as the canonical paper.
Lex Fridman (1:10:25.220)
And I don't think deep learning is everything.
Lex Fridman (1:10:26.860)
I think that there's still a bunch of things missing
George Hotz (1:10:28.700)
to get there, but MuZero I think is going to be looked back
Lex Fridman (1:10:31.420)
as the kind of cornerstone paper
George Hotz (1:10:34.260)
of this whole deep learning era.
Lex Fridman (1:10:37.060)
And MuZero is the solution to self driving cars.
George Hotz (1:10:39.580)
You have to make a few tweaks to it,
Lex Fridman (1:10:41.220)
but MuZero does effectively that.
George Hotz (1:10:42.820)
It does those rollouts and those murdering
Lex Fridman (1:10:45.500)
in a learned simulator and a learned dynamics model.
George Hotz (1:10:50.180)
That's interesting.
Lex Fridman (1:10:51.020)
It doesn't get enough love.
George Hotz (1:10:51.860)
I was blown away when I read that paper.
Lex Fridman (1:10:54.220)
I'm like, okay, I've always said a comma.
George Hotz (1:10:57.060)
I'm going to sit and I'm going to wait for the solution
Lex Fridman (1:10:58.460)
to self driving cars to come along.
George Hotz (1:11:00.340)
This year I saw it.
Lex Fridman (1:11:01.180)
It's MuZero.
George Hotz (1:11:05.060)
So.
Lex Fridman (1:11:06.860)
Sit back and let the winning roll in.
Lex Fridman (1:11:09.180)
So your sense, just to elaborate a little bit,
Lex Fridman (1:11:12.300)
it's a link on the topic.
George Hotz (1:11:13.380)
Your sense is neural networks will solve driving.
Lex Fridman (1:11:16.300)
Yes.
George Hotz (1:11:17.140)
Like we don't need anything else.
Lex Fridman (1:11:18.820)
I think the same way chess was maybe the chess
Lex Fridman (1:11:21.260)
and maybe Google are the pinnacle of like search algorithms
Lex Fridman (1:11:25.060)
and things that look kind of like a star.
George Hotz (1:11:28.420)
The pinnacle of this era is going to be self driving cars.
Lex Fridman (1:11:34.700)
But on the path of that, you have to deliver products
Lex Fridman (1:11:38.180)
and it's possible that the path to full self driving cars
Lex Fridman (1:11:42.340)
will take decades.
George Hotz (1:11:44.420)
I doubt it.
Lex Fridman (1:11:45.340)
How long would you put on it?
George Hotz (1:11:47.820)
Like what are we, you're chasing it, Tesla's chasing it.
Lex Fridman (1:11:53.460)
What are we talking about?
George Hotz (1:11:54.340)
Five years, 10 years, 50 years.
Lex Fridman (1:11:56.180)
Let's say in the 2020s.
George Hotz (1:11:58.060)
In the 2020s.
Lex Fridman (1:11:59.540)
The later part of the 2020s.
George Hotz (1:12:03.580)
With the neural network.
Lex Fridman (1:12:05.500)
Well, that would be nice to see.
Lex Fridman (1:12:06.580)
And then the path to that, you're delivering products,
Lex Fridman (1:12:09.140)
which is a nice L2 system.
George Hotz (1:12:10.580)
That's what Tesla's doing, a nice L2 system.
Lex Fridman (1:12:13.060)
Just gets better every time.
George Hotz (1:12:14.380)
L2, the only difference between L2 and the other levels
Lex Fridman (1:12:16.660)
is who takes liability.
Lex Fridman (1:12:17.620)
And I'm not a liability guy, I don't wanna take liability.
Lex Fridman (1:12:20.100)
I'm gonna level two forever.
George Hotz (1:12:22.700)
Now on that little transition,
Lex Fridman (1:12:25.780)
I mean, how do you make the transition work?
Lex Fridman (1:12:29.060)
Is this where driver sensing comes in?
Lex Fridman (1:12:32.780)
Like how do you make the, cause you said a hundred miles,
George Hotz (1:12:35.940)
like, is there some sort of human factor psychology thing
Lex Fridman (1:12:41.340)
where people start to overtrust the system,
George Hotz (1:12:43.140)
all those kinds of effects,
Lex Fridman (1:12:45.020)
once it gets better and better and better and better,
George Hotz (1:12:46.780)
they get lazier and lazier and lazier.
Lex Fridman (1:12:49.340)
Is that, like, how do you get that transition right?
George Hotz (1:12:52.460)
First off, our monitoring is already adaptive.
Lex Fridman (1:12:54.500)
Our monitoring is already seen adaptive.
George Hotz (1:12:56.620)
Driver monitoring is just the camera
Lex Fridman (1:12:58.940)
that's looking at the driver.
George Hotz (1:13:00.060)
You have an infrared camera in the...
Lex Fridman (1:13:03.060)
Our policy for how we enforce the driver monitoring
George Hotz (1:13:06.340)
is seen adaptive.
Lex Fridman (1:13:07.940)
What's that mean?
George Hotz (1:13:08.780)
Well, for example, in one of the extreme cases,
Lex Fridman (1:13:12.580)
if the car is not moving,
Lex Fridman (1:13:14.860)
we do not actively enforce driver monitoring, right?
Lex Fridman (1:13:19.460)
If you are going through a,
George Hotz (1:13:22.380)
like a 45 mile an hour road with lights
Lex Fridman (1:13:25.780)
and stop signs and potentially pedestrians,
George Hotz (1:13:27.980)
we enforce a very tight driver monitoring policy.
Lex Fridman (1:13:30.780)
If you are alone on a perfectly straight highway,
Lex Fridman (1:13:33.860)
and this is, it's all machine learning.
Lex Fridman (1:13:35.620)
None of that is hand coded.
George Hotz (1:13:36.940)
Actually, the stop is hand coded, but...
Lex Fridman (1:13:39.060)
So there's some kind of machine learning
George Hotz (1:13:41.100)
estimation of risk.
Lex Fridman (1:13:42.300)
Yes.
George Hotz (1:13:43.620)
Yeah.
Lex Fridman (1:13:44.460)
I mean, I've always been a huge fan of that.
George Hotz (1:13:45.860)
That's a...
Lex Fridman (1:13:47.500)
Because...
George Hotz (1:13:48.340)
It's difficult to do every step into that direction
Lex Fridman (1:13:53.780)
is a worthwhile step to take.
George Hotz (1:13:55.100)
It might be difficult to do really well.
Lex Fridman (1:13:56.540)
Like us humans are able to estimate risk pretty damn well,
George Hotz (1:13:59.980)
whatever the hell that is.
Lex Fridman (1:14:01.500)
That feels like one of the nice features of us humans.
George Hotz (1:14:06.500)
Cause like we humans are really good drivers
Lex Fridman (1:14:08.980)
when we're really like tuned in
Lex Fridman (1:14:11.180)
and we're good at estimating risk.
Lex Fridman (1:14:12.940)
Like when are we supposed to be tuned in?
George Hotz (1:14:14.900)
Yeah.
Lex Fridman (1:14:15.980)
And, you know, people are like,
George Hotz (1:14:17.580)
oh, well, you know,
Lex Fridman (1:14:18.420)
why would you ever make the driver monitoring policy
Lex Fridman (1:14:20.420)
less aggressive?
Lex Fridman (1:14:21.260)
Why would you always not keep it at its most aggressive?
George Hotz (1:14:23.980)
Because then people are just going to get fatigued from it.
Lex Fridman (1:14:25.780)
Yes.
George Hotz (1:14:26.620)
When they get annoyed.
Lex Fridman (1:14:27.460)
You want them...
George Hotz (1:14:28.300)
Yeah.
Lex Fridman (1:14:29.140)
You want the experience to be pleasant.
George Hotz (1:14:30.900)
Obviously I want the experience to be pleasant,
Lex Fridman (1:14:32.500)
but even just from a straight up safety perspective,
George Hotz (1:14:35.980)
if you alert people when they look around and they're like,
Lex Fridman (1:14:39.740)
why is this thing alerting me?
George Hotz (1:14:41.020)
There's nothing I could possibly hit right now.
Lex Fridman (1:14:42.980)
People will just learn to tune it out.
George Hotz (1:14:45.220)
People will just learn to tune it out,
Lex Fridman (1:14:46.900)
to put weights on the steering wheel,
George Hotz (1:14:48.060)
to do whatever to overcome it.
Lex Fridman (1:14:49.820)
And remember that you're always part
George Hotz (1:14:52.580)
of this adaptive system.
Lex Fridman (1:14:53.580)
So all I can really say about, you know,
Lex Fridman (1:14:55.660)
how this scales going forward is yeah,
Lex Fridman (1:14:57.180)
it's something we have to monitor for.
George Hotz (1:14:59.420)
Ooh, we don't know.
Lex Fridman (1:15:00.260)
This is a great psychology experiment at scale.
George Hotz (1:15:02.060)
Like we'll see.
Lex Fridman (1:15:03.220)
Yeah, it's fascinating.
George Hotz (1:15:04.060)
Track it.
Lex Fridman (1:15:04.900)
And making sure you have a good understanding of attention
George Hotz (1:15:09.140)
is a very key part of that psychology problem.
Lex Fridman (1:15:11.420)
Yeah.
George Hotz (1:15:12.260)
I think you and I probably have a different,
Lex Fridman (1:15:14.100)
come to it differently, but to me,
George Hotz (1:15:16.660)
it's a fascinating psychology problem
Lex Fridman (1:15:19.700)
to explore something much deeper than just driving.
George Hotz (1:15:22.100)
It's such a nice way to explore human attention
Lex Fridman (1:15:26.700)
and human behavior, which is why, again,
George Hotz (1:15:30.180)
we've probably both criticized Mr. Elon Musk
Lex Fridman (1:15:34.100)
on this one topic from different avenues.
Lex Fridman (1:15:38.220)
So both offline and online,
Lex Fridman (1:15:39.740)
I had little chats with Elon and like,
George Hotz (1:15:44.380)
I love human beings as a computer vision problem,
Lex Fridman (1:15:48.620)
as an AI problem, it's fascinating.
George Hotz (1:15:51.020)
He wasn't so much interested in that problem.
Lex Fridman (1:15:53.260)
It's like in order to solve driving,
George Hotz (1:15:56.820)
the whole point is you want to remove the human
Lex Fridman (1:15:58.780)
from the picture.
Lex Fridman (1:16:01.300)
And it seems like you can't do that quite yet.
Lex Fridman (1:16:04.140)
Eventually, yes, but you can't quite do that yet.
Lex Fridman (1:16:07.900)
So this is the moment where you can't yet say,
Lex Fridman (1:16:12.300)
I told you so to Tesla, but it's getting there
George Hotz (1:16:17.700)
because I don't know if you've seen this,
Lex Fridman (1:16:19.220)
there's some reporting that they're in fact
George Hotz (1:16:21.100)
starting to do driver monitoring.
Lex Fridman (1:16:23.260)
Yeah, they shift the model in shadow mode.
George Hotz (1:16:26.260)
With, I believe, only a visible light camera,
Lex Fridman (1:16:29.620)
it might even be fisheye.
George Hotz (1:16:31.780)
It's like a low resolution.
Lex Fridman (1:16:33.260)
Low resolution, visible light.
George Hotz (1:16:34.820)
I mean, to be fair, that's what we have in the Eon as well,
Lex Fridman (1:16:37.100)
our last generation product.
George Hotz (1:16:38.820)
This is the one area where I can say
Lex Fridman (1:16:41.060)
our hardware is ahead of Tesla.
George Hotz (1:16:42.180)
The rest of our hardware, way, way behind,
Lex Fridman (1:16:43.820)
but our driver monitoring camera.
Lex Fridman (1:16:46.020)
So you think, I think on the third row Tesla podcast,
Lex Fridman (1:16:50.940)
or somewhere else, I've heard you say that obviously,
George Hotz (1:16:54.580)
eventually they're gonna have driver monitoring.
Lex Fridman (1:16:57.220)
I think what I've said is Elon will definitely ship
George Hotz (1:16:59.660)
driver monitoring before he ships level five.
Lex Fridman (1:17:01.700)
Before level five.
Lex Fridman (1:17:02.540)
And I'm willing to bet 10 grand on that.
Lex Fridman (1:17:04.660)
And you bet 10 grand on that.
George Hotz (1:17:07.060)
I mean, now I don't wanna take the bet,
Lex Fridman (1:17:08.260)
but before, maybe someone would have,
George Hotz (1:17:09.500)
oh, I should have got my money in.
Lex Fridman (1:17:10.500)
Yeah.
George Hotz (1:17:11.900)
It's an interesting bet.
Lex Fridman (1:17:12.980)
I think you're right.
George Hotz (1:17:16.460)
I'm actually on a human level
Lex Fridman (1:17:19.180)
because he's been, he's made the decision.
George Hotz (1:17:24.380)
Like he said that driver monitoring is the wrong way to go.
Lex Fridman (1:17:27.660)
But like, you have to think of as a human, as a CEO,
George Hotz (1:17:31.260)
I think that's the right thing to say when,
Lex Fridman (1:17:36.460)
like sometimes you have to say things publicly
George Hotz (1:17:40.140)
that are different than when you actually believe,
Lex Fridman (1:17:41.780)
because when you're producing a large number of vehicles
Lex Fridman (1:17:45.420)
and the decision was made not to include the camera,
Lex Fridman (1:17:47.860)
like what are you supposed to say?
George Hotz (1:17:49.460)
Like our cars don't have the thing
Lex Fridman (1:17:51.780)
that I think is right to have.
George Hotz (1:17:54.020)
It's an interesting thing.
Lex Fridman (1:17:55.780)
But like on the other side, as a CEO,
George Hotz (1:17:58.340)
I mean, something you could probably speak to as a leader,
Lex Fridman (1:18:01.220)
I think about me as a human
George Hotz (1:18:04.940)
to publicly change your mind on something.
Lex Fridman (1:18:07.020)
How hard is that?
George Hotz (1:18:08.420)
Especially when assholes like George Haas say,
Lex Fridman (1:18:10.620)
I told you so.
George Hotz (1:18:12.380)
All I will say is I am not a leader
Lex Fridman (1:18:14.740)
and I am happy to change my mind.
Lex Fridman (1:18:17.060)
And I will.
Lex Fridman (1:18:17.900)
You think Elon will?
George Hotz (1:18:20.580)
Yeah, I do.
Lex Fridman (1:18:22.300)
I think he'll come up with a good way
George Hotz (1:18:24.260)
to make it psychologically okay for him.
Lex Fridman (1:18:27.420)
Well, it's such an important thing, man.
George Hotz (1:18:29.700)
Especially for a first principles thinker,
Lex Fridman (1:18:31.420)
because he made a decision that driver monitoring
George Hotz (1:18:34.780)
is not the right way to go.
Lex Fridman (1:18:35.740)
And I could see that decision.
Lex Fridman (1:18:37.660)
And I could even make that decision.
Lex Fridman (1:18:39.260)
Like I was on the fence too.
George Hotz (1:18:41.660)
Like I'm not a,
Lex Fridman (1:18:42.500)
driver monitoring is such an obvious,
George Hotz (1:18:47.140)
simple solution to the problem of attention.
Lex Fridman (1:18:49.980)
It's not obvious to me that just by putting a camera there,
George Hotz (1:18:52.820)
you solve things.
Lex Fridman (1:18:54.260)
You have to create an incredible, compelling experience.
George Hotz (1:18:59.060)
Just like you're talking about.
Lex Fridman (1:19:01.060)
I don't know if it's easy to do that.
George Hotz (1:19:03.300)
It's not at all easy to do that, in fact, I think.
Lex Fridman (1:19:05.980)
So as a creator of a car that's trying to create a product
Lex Fridman (1:19:10.980)
that people love, which is what Tesla tries to do, right?
Lex Fridman (1:19:14.180)
It's not obvious to me that as a design decision,
George Hotz (1:19:18.340)
whether adding a camera is a good idea.
Lex Fridman (1:19:20.940)
From a safety perspective either,
George Hotz (1:19:22.460)
like in the human factors community,
Lex Fridman (1:19:25.100)
everybody says that you should obviously
George Hotz (1:19:27.380)
have driver sensing, driver monitoring.
Lex Fridman (1:19:30.500)
But that's like saying it's obvious as parents,
George Hotz (1:19:36.900)
you shouldn't let your kids go out at night.
Lex Fridman (1:19:39.780)
But okay, but like,
George Hotz (1:19:43.460)
they're still gonna find ways to do drugs.
Lex Fridman (1:19:45.620)
Like, you have to also be good parents.
Lex Fridman (1:19:49.860)
So like, it's much more complicated than just the,
Lex Fridman (1:19:52.580)
you need to have driver monitoring.
George Hotz (1:19:54.140)
I totally disagree on, okay, if you have a camera there
Lex Fridman (1:19:58.740)
and the camera's watching the person,
Lex Fridman (1:20:00.300)
but never throws an alert, they'll never think about it.
Lex Fridman (1:20:03.540)
Right?
George Hotz (1:20:04.380)
The driver monitoring policy that you choose to,
Lex Fridman (1:20:08.380)
how you choose to communicate with the user
George Hotz (1:20:10.180)
is entirely separate from the data collection perspective.
Lex Fridman (1:20:14.220)
Right?
Lex Fridman (1:20:15.060)
Right?
Lex Fridman (1:20:15.900)
So, you know, like, there's one thing to say,
George Hotz (1:20:20.980)
like, you know, tell your teenager they can't do something.
Lex Fridman (1:20:24.500)
There's another thing to like, you know, gather the data.
Lex Fridman (1:20:27.020)
So you can make informed decisions.
Lex Fridman (1:20:28.500)
That's really interesting.
Lex Fridman (1:20:29.340)
But you have to make that,
Lex Fridman (1:20:30.980)
that's the interesting thing about cars.
Lex Fridman (1:20:33.380)
But even true with common AI,
Lex Fridman (1:20:35.020)
like you don't have to manufacture the thing
George Hotz (1:20:37.900)
into the car, is you have to make a decision
Lex Fridman (1:20:40.060)
that anticipates the right strategy longterm.
Lex Fridman (1:20:44.180)
So like, you have to start collecting the data
Lex Fridman (1:20:46.620)
and start making decisions.
George Hotz (1:20:47.780)
Started it three years ago.
Lex Fridman (1:20:49.900)
I believe that we have the best driver monitoring solution
George Hotz (1:20:52.660)
in the world.
Lex Fridman (1:20:54.620)
I think that when you compare it to Super Cruise
George Hotz (1:20:57.220)
is the only other one that I really know that shipped.
Lex Fridman (1:20:59.460)
And ours is better.
Lex Fridman (1:21:01.420)
What do you like and not like about Super Cruise?
Lex Fridman (1:21:06.420)
I mean, I had a few Super Cruise,
George Hotz (1:21:08.740)
the sun would be shining through the window,
Lex Fridman (1:21:12.020)
would blind the camera,
Lex Fridman (1:21:13.180)
and it would say I wasn't paying attention.
Lex Fridman (1:21:14.580)
When I was looking completely straight,
George Hotz (1:21:16.100)
I couldn't reset the attention with a steering wheel touch
Lex Fridman (1:21:19.140)
and Super Cruise would disengage.
George Hotz (1:21:21.060)
Like I was communicating to the car, I'm like, look,
Lex Fridman (1:21:22.980)
I am here, I am paying attention.
Lex Fridman (1:21:24.420)
Why are you really gonna force me to disengage?
Lex Fridman (1:21:26.340)
And it did.
Lex Fridman (1:21:28.620)
So it's a constant conversation with the user.
Lex Fridman (1:21:32.100)
And yeah, there's no way to ship a system
George Hotz (1:21:33.660)
like this if you can OTA.
Lex Fridman (1:21:35.500)
We're shipping a new one every month.
George Hotz (1:21:37.180)
Sometimes we balance it with our users on Discord.
Lex Fridman (1:21:40.140)
Like sometimes we make the driver monitoring
George Hotz (1:21:41.980)
a little more aggressive and people complain.
Lex Fridman (1:21:43.820)
Sometimes they don't.
George Hotz (1:21:45.500)
We want it to be as aggressive as possible
Lex Fridman (1:21:47.060)
where people don't complain and it doesn't feel intrusive.
Lex Fridman (1:21:49.180)
So being able to update the system over the air
Lex Fridman (1:21:51.100)
is an essential component.
George Hotz (1:21:52.540)
I mean, that's probably to me, you mentioned,
Lex Fridman (1:21:56.620)
I mean, to me that is the biggest innovation of Tesla,
George Hotz (1:22:01.060)
that it made people realize that over the air updates
Lex Fridman (1:22:04.860)
is essential.
George Hotz (1:22:06.460)
Yeah.
Lex Fridman (1:22:07.460)
I mean, was that not obvious from the iPhone?
George Hotz (1:22:10.140)
The iPhone was the first real product that OTA'd, I think.
Lex Fridman (1:22:13.060)
Was it actually, that's brilliant, you're right.
Lex Fridman (1:22:15.220)
I mean, the game consoles used to not, right?
Lex Fridman (1:22:17.060)
The game consoles were maybe the second thing that did.
George Hotz (1:22:18.980)
Wow, I didn't really think about one of the amazing features
Lex Fridman (1:22:22.260)
of a smartphone isn't just like the touchscreen
George Hotz (1:22:26.180)
isn't the thing, it's the ability to constantly update.
Lex Fridman (1:22:30.740)
Yeah, it gets better.
George Hotz (1:22:31.980)
It gets better.
Lex Fridman (1:22:35.100)
Love my iOS 14.
George Hotz (1:22:36.820)
Yeah.
Lex Fridman (1:22:38.300)
Well, one thing that I probably disagree with you
George Hotz (1:22:41.540)
on driver monitoring is you said that it's easy.
Lex Fridman (1:22:46.700)
I mean, you tend to say stuff is easy.
George Hotz (1:22:48.940)
I'm sure the, I guess you said it's easy
Lex Fridman (1:22:52.700)
relative to the external perception problem.
Lex Fridman (1:22:58.180)
Can you elaborate why you think it's easy?
Lex Fridman (1:23:00.940)
Feature engineering works for driver monitoring.
George Hotz (1:23:03.500)
Feature engineering does not work for the external.
Lex Fridman (1:23:05.860)
So human faces are not, human faces and the movement
George Hotz (1:23:10.700)
of human faces and head and body is not as variable
Lex Fridman (1:23:14.740)
as the external environment, is your intuition?
George Hotz (1:23:17.140)
Yes, and there's another big difference as well.
Lex Fridman (1:23:20.140)
Your reliability of a driver monitoring system
George Hotz (1:23:22.500)
doesn't actually need to be that high.
Lex Fridman (1:23:24.380)
The uncertainty, if you have something that's detecting
George Hotz (1:23:27.220)
whether the human's paying attention and it only works
Lex Fridman (1:23:29.140)
92% of the time, you're still getting almost all
George Hotz (1:23:31.700)
the benefit of that because the human,
Lex Fridman (1:23:33.620)
like you're training the human, right?
George Hotz (1:23:35.500)
You're dealing with a system that's really helping you out.
Lex Fridman (1:23:39.100)
It's a conversation.
Lex Fridman (1:23:40.180)
It's not like the external thing where guess what?
Lex Fridman (1:23:43.460)
If you swerve into a tree, you swerve into a tree, right?
George Hotz (1:23:46.140)
Like you get no margin for error there.
Lex Fridman (1:23:48.340)
Yeah, I think that's really well put.
George Hotz (1:23:49.620)
I think that's the right, exactly the place
Lex Fridman (1:23:54.020)
where comparing to the external perception,
George Hotz (1:23:58.660)
the control problem, the driver monitoring is easier
Lex Fridman (1:24:01.500)
because you don't, the bar for success is much lower.
George Hotz (1:24:05.260)
Yeah, but I still think like the human face
Lex Fridman (1:24:09.100)
is more complicated actually than the external environment,
Lex Fridman (1:24:12.140)
but for driving, you don't give a damn.
Lex Fridman (1:24:14.380)
I don't need, yeah, I don't need something,
George Hotz (1:24:15.620)
I don't need something that complicated
Lex Fridman (1:24:18.300)
to have to communicate the idea to the human
George Hotz (1:24:22.220)
that I want to communicate, which is,
Lex Fridman (1:24:23.980)
yo, system might mess up here.
George Hotz (1:24:25.740)
You gotta pay attention.
Lex Fridman (1:24:26.940)
Yeah, see, that's my love and fascination is the human face.
Lex Fridman (1:24:32.500)
And it feels like this is a nice place to create products
Lex Fridman (1:24:38.420)
that create an experience in the car.
Lex Fridman (1:24:40.100)
So like, it feels like there should be
Lex Fridman (1:24:42.640)
more richer experiences in the car, you know?
George Hotz (1:24:47.540)
Like that's an opportunity for like something like On My Eye
Lex Fridman (1:24:51.500)
or just any kind of system like a Tesla
George Hotz (1:24:53.900)
or any of the autonomous vehicle companies
Lex Fridman (1:24:56.220)
is because software is, there's much more sensors
Lex Fridman (1:24:59.180)
and so much is on our software
Lex Fridman (1:25:00.660)
and you're doing machine learning anyway,
George Hotz (1:25:02.940)
there's an opportunity to create totally new experiences
Lex Fridman (1:25:06.340)
that we're not even anticipating.
Lex Fridman (1:25:08.260)
You don't think so?
Lex Fridman (1:25:10.140)
Nah.
George Hotz (1:25:10.980)
You think it's a box that gets you from A to B
Lex Fridman (1:25:12.900)
and you want to do it chill?
George Hotz (1:25:14.920)
Yeah, I mean, I think as soon as we get to level three
Lex Fridman (1:25:16.940)
on highways, okay, enjoy your candy crush,
George Hotz (1:25:19.300)
enjoy your Hulu, enjoy your, you know, whatever, whatever.
Lex Fridman (1:25:23.660)
Sure, you get this, you can look at screens basically
George Hotz (1:25:26.260)
versus right now where you have music and audio books.
Lex Fridman (1:25:28.700)
So level three is where you can kind of disengage
George Hotz (1:25:31.020)
in stretches of time.
Lex Fridman (1:25:34.860)
Well, you think level three is possible?
George Hotz (1:25:37.460)
Like on the highway going for 100 miles
Lex Fridman (1:25:39.180)
and you can just go to sleep?
George Hotz (1:25:40.500)
Oh yeah, sleep.
Lex Fridman (1:25:43.620)
So again, I think it's really all on a spectrum.
George Hotz (1:25:47.340)
I think that being able to use your phone
Lex Fridman (1:25:50.060)
while you're on the highway and like this all being okay
Lex Fridman (1:25:53.500)
and being aware that the car might alert you
Lex Fridman (1:25:55.360)
and you have five seconds to basically.
Lex Fridman (1:25:57.180)
So the five second thing is you think is possible?
Lex Fridman (1:25:59.060)
Yeah, I think it is, oh yeah.
Lex Fridman (1:26:00.420)
Not in all scenarios, right?
Lex Fridman (1:26:02.180)
Some scenarios it's not.
George Hotz (1:26:03.940)
It's the whole risk thing that you mentioned is nice
Lex Fridman (1:26:06.300)
is to be able to estimate like how risky is this situation?
George Hotz (1:26:10.660)
That's really important to understand.
Lex Fridman (1:26:12.620)
One other thing you mentioned comparing KAMA
Lex Fridman (1:26:15.380)
and Autopilot is that something about the haptic feel
Lex Fridman (1:26:20.380)
of the way KAMA controls the car when things are uncertain.
George Hotz (1:26:25.920)
Like it behaves a little bit more uncertain
Lex Fridman (1:26:27.760)
when things are uncertain.
George Hotz (1:26:29.240)
That's kind of an interesting point.
Lex Fridman (1:26:31.080)
And then Autopilot is much more confident always
George Hotz (1:26:34.080)
even when it's uncertain until it runs into trouble.
Lex Fridman (1:26:39.200)
That's a funny thing.
George Hotz (1:26:40.920)
I actually mentioned that to Elon, I think.
Lex Fridman (1:26:42.700)
And then the first time we talked, he wasn't biting.
George Hotz (1:26:46.280)
It's like communicating uncertainty.
Lex Fridman (1:26:48.880)
I guess KAMA doesn't really communicate uncertainty
George Hotz (1:26:51.880)
explicitly, it communicates it through haptic feel.
Lex Fridman (1:26:55.040)
Like what's the role of communicating uncertainty
Lex Fridman (1:26:57.420)
do you think?
Lex Fridman (1:26:58.260)
Oh, we do some stuff explicitly.
George Hotz (1:26:59.840)
Like we do detect the lanes when you're on the highway
Lex Fridman (1:27:01.800)
and we'll show you how many lanes we're using to drive with.
George Hotz (1:27:04.380)
You can look at where it thinks the lanes are.
Lex Fridman (1:27:06.200)
You can look at the path.
Lex Fridman (1:27:08.480)
And we want to be better about this.
Lex Fridman (1:27:10.440)
We're actually hiring, want to hire some new UI people.
George Hotz (1:27:12.900)
UI people, you mentioned this.
Lex Fridman (1:27:14.320)
Cause it's such an, it's a UI problem too, right?
George Hotz (1:27:17.300)
We have a great designer now, but you know,
Lex Fridman (1:27:19.720)
we need people who are just going to like build this
Lex Fridman (1:27:21.160)
and debug these UIs, QT people.
Lex Fridman (1:27:23.800)
QT.
Lex Fridman (1:27:24.640)
Is that what the UI is done with, is QT?
Lex Fridman (1:27:26.880)
The new UI is in QT.
Lex Fridman (1:27:29.320)
C++ QT?
Lex Fridman (1:27:32.000)
Tesla uses it too.
George Hotz (1:27:33.300)
Yeah.
Lex Fridman (1:27:34.200)
We had some React stuff in there.
Lex Fridman (1:27:37.760)
React JS or just React?
Lex Fridman (1:27:39.440)
React is his own language, right?
George Hotz (1:27:41.160)
React Native, React is a JavaScript framework.
Lex Fridman (1:27:44.480)
Yeah.
Lex Fridman (1:27:45.320)
So it's all based on JavaScript, but it's, you know,
Lex Fridman (1:27:48.960)
I like C++.
Lex Fridman (1:27:51.480)
What do you think about Dojo with Tesla
Lex Fridman (1:27:55.080)
and their foray into what appears to be
Lex Fridman (1:28:00.240)
specialized hardware for training your own nets?
Lex Fridman (1:28:05.120)
I guess it's something, maybe you can correct me,
George Hotz (1:28:07.360)
from my shallow looking at it,
Lex Fridman (1:28:10.000)
it seems like something like Google did with TPUs,
Lex Fridman (1:28:12.120)
but specialized for driving data.
Lex Fridman (1:28:15.680)
I don't think it's specialized for driving data.
George Hotz (1:28:18.360)
It's just legit, just TPU.
Lex Fridman (1:28:20.120)
They want to go the Apple way,
George Hotz (1:28:22.160)
basically everything required in the chain is done in house.
Lex Fridman (1:28:25.640)
Well, so you have a problem right now,
Lex Fridman (1:28:27.840)
and this is one of my concerns.
Lex Fridman (1:28:31.740)
I really would like to see somebody deal with this.
George Hotz (1:28:33.800)
If anyone out there is doing it,
Lex Fridman (1:28:35.280)
I'd like to help them if I can.
George Hotz (1:28:38.000)
You basically have two options right now to train.
Lex Fridman (1:28:40.620)
One, your options are NVIDIA or Google.
Lex Fridman (1:28:45.980)
So Google is not even an option.
Lex Fridman (1:28:50.120)
Their TPUs are only available in Google Cloud.
George Hotz (1:28:53.180)
Google has absolutely onerous
Lex Fridman (1:28:55.140)
terms of service restrictions.
George Hotz (1:28:58.060)
They may have changed it,
Lex Fridman (1:28:59.280)
but back in Google's terms of service,
George Hotz (1:29:00.620)
it said explicitly you are not allowed to use Google Cloud ML
Lex Fridman (1:29:03.740)
for training autonomous vehicles
George Hotz (1:29:05.340)
or for doing anything that competes with Google
Lex Fridman (1:29:07.360)
without Google's prior written permission.
George Hotz (1:29:09.300)
Wow, okay.
Lex Fridman (1:29:10.340)
I mean, Google is not a platform company.
George Hotz (1:29:14.140)
I wouldn't touch TPUs with a 10 foot pole.
Lex Fridman (1:29:16.760)
So that leaves you with the monopoly.
George Hotz (1:29:19.300)
NVIDIA? NVIDIA.
Lex Fridman (1:29:21.020)
So, I mean.
George Hotz (1:29:22.340)
That you're not a fan of.
Lex Fridman (1:29:23.900)
Well, look, I was a huge fan of in 2016 NVIDIA.
George Hotz (1:29:28.540)
Jensen came sat in the car.
Lex Fridman (1:29:31.880)
Cool guy.
George Hotz (1:29:32.820)
When the stock was $30 a share.
Lex Fridman (1:29:35.540)
NVIDIA stock has skyrocketed.
George Hotz (1:29:38.220)
I witnessed a real change
Lex Fridman (1:29:39.920)
in who was in management over there in like 2018.
Lex Fridman (1:29:43.580)
And now they are, let's exploit.
Lex Fridman (1:29:46.700)
Let's take every dollar we possibly can
George Hotz (1:29:48.460)
out of this ecosystem.
Lex Fridman (1:29:49.580)
Let's charge $10,000 for A100s
George Hotz (1:29:51.740)
because we know we got the best shit in the game.
Lex Fridman (1:29:54.180)
And let's charge $10,000 for an A100
George Hotz (1:29:57.920)
when it's really not that different from a 3080,
Lex Fridman (1:30:00.100)
which is 699.
George Hotz (1:30:03.500)
The margins that they are making
Lex Fridman (1:30:05.100)
off of those high end chips are so high
George Hotz (1:30:08.640)
that, I mean, I think they're shooting themselves
Lex Fridman (1:30:10.260)
in the foot just from a business perspective.
George Hotz (1:30:12.220)
Because there's a lot of people talking like me now
Lex Fridman (1:30:14.980)
who are like, somebody's gotta take NVIDIA down.
George Hotz (1:30:19.060)
Yeah.
Lex Fridman (1:30:19.900)
Where they could dominate it.
George Hotz (1:30:21.020)
NVIDIA could be the new Intel.
Lex Fridman (1:30:22.460)
Yeah, to be inside everything essentially.
Lex Fridman (1:30:26.940)
And yet the winners in certain spaces
Lex Fridman (1:30:30.660)
like autonomous driving, the winners,
George Hotz (1:30:33.780)
only the people who are like desperately falling back
Lex Fridman (1:30:36.620)
and trying to catch up and have a ton of money,
George Hotz (1:30:38.540)
like the big automakers are the ones
Lex Fridman (1:30:40.660)
interested in partnering with NVIDIA.
George Hotz (1:30:43.220)
Oh, and I think a lot of those things
Lex Fridman (1:30:44.820)
are gonna fall through.
George Hotz (1:30:45.940)
If I were NVIDIA, sell chips.
Lex Fridman (1:30:49.340)
Sell chips at a reasonable markup.
George Hotz (1:30:52.260)
To everybody.
Lex Fridman (1:30:53.100)
To everybody.
George Hotz (1:30:53.920)
Without any restrictions.
Lex Fridman (1:30:54.940)
Without any restrictions.
George Hotz (1:30:56.140)
Intel did this.
Lex Fridman (1:30:57.360)
Look at Intel.
George Hotz (1:30:58.260)
They had a great long run.
Lex Fridman (1:30:59.940)
NVIDIA is trying to turn their,
George Hotz (1:31:01.580)
they're like trying to productize their chips
Lex Fridman (1:31:04.020)
way too much.
George Hotz (1:31:05.620)
They're trying to extract way more value
Lex Fridman (1:31:07.880)
than they can sustainably.
George Hotz (1:31:09.380)
Sure, you can do it tomorrow.
Lex Fridman (1:31:10.740)
Is it gonna up your share price?
George Hotz (1:31:12.140)
Sure, if you're one of those CEOs
Lex Fridman (1:31:13.540)
who's like, how much can I strip mine this company?
Lex Fridman (1:31:15.300)
And I think, you know, and that's what's weird about it too.
Lex Fridman (1:31:17.860)
Like the CEO is the founder.
George Hotz (1:31:19.380)
It's the same guy.
Lex Fridman (1:31:20.300)
Yeah.
George Hotz (1:31:21.120)
I mean, I still think Jensen's a great guy.
Lex Fridman (1:31:22.340)
He is great.
Lex Fridman (1:31:23.300)
Why do this?
Lex Fridman (1:31:25.180)
You have a choice.
George Hotz (1:31:26.660)
You have a choice right now.
Lex Fridman (1:31:27.940)
Are you trying to cash out?
Lex Fridman (1:31:28.820)
Are you trying to buy a yacht?
Lex Fridman (1:31:30.660)
If you are, fine.
Lex Fridman (1:31:32.100)
But if you're trying to be
Lex Fridman (1:31:34.220)
the next huge semiconductor company, sell chips.
George Hotz (1:31:37.240)
Well, the interesting thing about Jensen
Lex Fridman (1:31:40.140)
is he is a big vision guy.
Lex Fridman (1:31:42.060)
So he has a plan like for 50 years down the road.
Lex Fridman (1:31:48.700)
So it makes me wonder like.
Lex Fridman (1:31:50.500)
How does price gouging fit into it?
Lex Fridman (1:31:51.780)
Yeah, how does that, like it's,
George Hotz (1:31:54.000)
it doesn't seem to make sense as a plan.
Lex Fridman (1:31:57.060)
I worry that he's listening to the wrong people.
George Hotz (1:31:59.260)
Yeah, that's the sense I have too sometimes.
Lex Fridman (1:32:02.540)
Because I, despite everything, I think NVIDIA
George Hotz (1:32:07.940)
is an incredible company.
Lex Fridman (1:32:09.020)
Well, one, so I'm deeply grateful to NVIDIA
George Hotz (1:32:12.420)
for the products they've created in the past.
Lex Fridman (1:32:13.740)
Me too.
Lex Fridman (1:32:14.580)
Right?
Lex Fridman (1:32:15.400)
And so.
George Hotz (1:32:16.240)
The 1080 Ti was a great GPU.
Lex Fridman (1:32:18.000)
Still have a lot of them.
George Hotz (1:32:18.840)
Still is, yeah.
Lex Fridman (1:32:21.840)
But at the same time, it just feels like,
George Hotz (1:32:26.860)
feels like you don't want to put all your stock in NVIDIA.
Lex Fridman (1:32:29.380)
And so like Elon is doing, what Tesla is doing
George Hotz (1:32:32.940)
with Autopilot and Dojo is the Apple way is,
Lex Fridman (1:32:37.260)
because they're not going to share Dojo with George Hott's.
George Hotz (1:32:40.300)
I know.
Lex Fridman (1:32:42.340)
They should sell that chip.
George Hotz (1:32:43.780)
Oh, they should sell that.
Lex Fridman (1:32:44.700)
Even their accelerator.
George Hotz (1:32:46.400)
The accelerator that's in all the cars, the 30 watt one.
Lex Fridman (1:32:49.060)
Sell it, why not?
Lex Fridman (1:32:51.580)
So open it up.
Lex Fridman (1:32:52.700)
Like make, why does Tesla have to be a car company?
George Hotz (1:32:55.820)
Well, if you sell the chip, here's what you get.
Lex Fridman (1:32:58.060)
Yeah.
George Hotz (1:32:59.020)
Make some money off the chips.
Lex Fridman (1:33:00.260)
It doesn't take away from your chip.
George Hotz (1:33:02.080)
You're going to make some money, free money.
Lex Fridman (1:33:03.860)
And also the world is going to build an ecosystem
George Hotz (1:33:07.380)
of tooling for you.
Lex Fridman (1:33:09.020)
Right?
George Hotz (1:33:09.860)
You're not going to have to fix the bug in your 10H layer.
Lex Fridman (1:33:12.860)
Someone else already did.
George Hotz (1:33:15.140)
Well, the question, that's an interesting question.
Lex Fridman (1:33:16.780)
I mean, that's the question Steve Jobs asked.
George Hotz (1:33:18.740)
That's the question Elon Musk is perhaps asking is,
Lex Fridman (1:33:24.940)
do you want Tesla stuff inside other vehicles?
George Hotz (1:33:28.060)
Inside, potentially inside like a iRobot vacuum cleaner.
Lex Fridman (1:33:32.620)
Yeah.
George Hotz (1:33:34.860)
I think you should decide where your advantages are.
Lex Fridman (1:33:37.160)
I'm not saying Tesla should start selling battery packs
George Hotz (1:33:39.260)
to automakers.
Lex Fridman (1:33:40.380)
Because battery packs to automakers,
George Hotz (1:33:41.720)
they are straight up in competition with you.
Lex Fridman (1:33:43.660)
If I were Tesla, I'd keep the battery technology totally.
George Hotz (1:33:46.060)
Yeah.
Lex Fridman (1:33:46.900)
As far as we make batteries.
Lex Fridman (1:33:47.940)
But the thing about the Tesla TPU is anybody can build that.
Lex Fridman (1:33:53.160)
It's just a question of, you know,
Lex Fridman (1:33:54.600)
are you willing to spend the money?
Lex Fridman (1:33:57.460)
It could be a huge source of revenue potentially.
Lex Fridman (1:34:00.220)
Are you willing to spend a hundred million dollars?
Lex Fridman (1:34:02.420)
Anyone can build it.
Lex Fridman (1:34:03.660)
And someone will.
Lex Fridman (1:34:04.680)
And a bunch of companies now are starting
George Hotz (1:34:06.680)
trying to build AI accelerators.
Lex Fridman (1:34:08.020)
Somebody is going to get the idea right.
Lex Fridman (1:34:10.200)
And yeah, hopefully they don't get greedy
Lex Fridman (1:34:13.700)
because they'll just lose to the next guy who finally,
Lex Fridman (1:34:15.780)
and then eventually the Chinese are going to make knockoff
Lex Fridman (1:34:17.540)
and video chips and that's.
George Hotz (1:34:19.580)
From your perspective,
Lex Fridman (1:34:20.400)
I don't know if you're also paying attention
George Hotz (1:34:21.820)
to stay on Tesla for a moment.
Lex Fridman (1:34:24.180)
Dave, Elon Musk has talked about a complete rewrite
George Hotz (1:34:28.820)
of the neural net that they're using.
Lex Fridman (1:34:31.700)
That seems to, again, I'm half paying attention,
Lex Fridman (1:34:34.800)
but it seems to involve basically a kind of integration
Lex Fridman (1:34:39.000)
of all the sensors to where it's a four dimensional view.
George Hotz (1:34:44.180)
You know, you have a 3D model of the world over time.
Lex Fridman (1:34:47.540)
And then you can, I think it's done both for the,
George Hotz (1:34:52.100)
for the actually, you know,
Lex Fridman (1:34:53.280)
so the neural network is able to,
George Hotz (1:34:55.260)
in a more holistic way,
Lex Fridman (1:34:56.920)
deal with the world and make predictions and so on,
Lex Fridman (1:34:59.340)
but also to make the annotation task more, you know, easier.
Lex Fridman (1:35:04.780)
Like you can annotate the world in one place
Lex Fridman (1:35:08.180)
and then kind of distribute itself across the sensors
Lex Fridman (1:35:10.500)
and across a different,
George Hotz (1:35:12.860)
like the hundreds of tasks that are involved
Lex Fridman (1:35:15.180)
in the Hydro Net.
Lex Fridman (1:35:16.540)
What are your thoughts about this rewrite?
Lex Fridman (1:35:19.140)
Is it just like some details that are kind of obvious
George Hotz (1:35:22.420)
that are steps that should be taken,
Lex Fridman (1:35:24.060)
or is there something fundamental
George Hotz (1:35:26.100)
that could challenge your idea
Lex Fridman (1:35:27.560)
that end to end is the right solution?
George Hotz (1:35:31.120)
We're in the middle of a big rewrite now as well.
Lex Fridman (1:35:33.160)
We haven't shipped a new model in a bit.
Lex Fridman (1:35:34.860)
Of what kind?
Lex Fridman (1:35:36.400)
We're going from 2D to 3D.
George Hotz (1:35:38.240)
Right now, all our stuff, like for example,
Lex Fridman (1:35:39.740)
when the car pitches back,
George Hotz (1:35:40.980)
the lane lines also pitch back
Lex Fridman (1:35:43.020)
because we're assuming the flat world hypothesis.
George Hotz (1:35:47.160)
The new models do not do this.
Lex Fridman (1:35:48.420)
The new models output everything in 3D.
Lex Fridman (1:35:50.420)
But there's still no annotation.
Lex Fridman (1:35:53.580)
So the 3D is, it's more about the output.
George Hotz (1:35:56.500)
Yeah.
Lex Fridman (1:35:57.340)
We have Zs in everything.
George Hotz (1:36:00.100)
We've...
Lex Fridman (1:36:00.940)
Zs.
George Hotz (1:36:01.760)
Yeah.
Lex Fridman (1:36:02.600)
We had a Zs.
George Hotz (1:36:03.420)
We had a Zs.
Lex Fridman (1:36:04.260)
We unified a lot of stuff as well.
George Hotz (1:36:06.620)
We switched from TensorFlow to PyTorch.
Lex Fridman (1:36:10.720)
My understanding of what Tesla's thing is,
George Hotz (1:36:13.740)
is that their annotator now annotates
Lex Fridman (1:36:15.660)
across the time dimension.
George Hotz (1:36:16.960)
Mm hmm.
Lex Fridman (1:36:19.980)
I mean, cute.
Lex Fridman (1:36:22.000)
Why are you building an annotator?
Lex Fridman (1:36:24.400)
I find their entire pipeline.
George Hotz (1:36:28.280)
I find your vision, I mean,
Lex Fridman (1:36:30.560)
the vision of end to end very compelling,
Lex Fridman (1:36:32.860)
but I also like the engineering of the data engine
Lex Fridman (1:36:35.880)
that they've created.
George Hotz (1:36:37.400)
In terms of supervised learning pipelines,
Lex Fridman (1:36:41.560)
that thing is damn impressive.
George Hotz (1:36:43.560)
You're basically, the idea is that you have
Lex Fridman (1:36:47.480)
hundreds of thousands of people
George Hotz (1:36:49.280)
that are doing data collection for you
Lex Fridman (1:36:51.200)
by doing their experience.
Lex Fridman (1:36:52.400)
So that's kind of similar to the Comma AI model.
Lex Fridman (1:36:55.220)
And you're able to mine that data
George Hotz (1:36:59.580)
based on the kind of edge cases you need.
Lex Fridman (1:37:02.940)
I think it's harder to do in the end to end learning.
George Hotz (1:37:07.360)
The mining of the right edge cases.
Lex Fridman (1:37:09.520)
Like that's where feature engineering
George Hotz (1:37:11.440)
is actually really powerful
Lex Fridman (1:37:14.120)
because like us humans are able to do
George Hotz (1:37:17.280)
this kind of mining a little better.
Lex Fridman (1:37:19.800)
But yeah, there's obvious, as we know,
George Hotz (1:37:21.980)
there's obvious constraints and limitations to that idea.
Lex Fridman (1:37:25.880)
Carpathia just tweeted, he's like,
George Hotz (1:37:28.280)
you get really interesting insights
Lex Fridman (1:37:29.640)
if you sort your validation set by loss
Lex Fridman (1:37:33.720)
and look at the highest loss examples.
Lex Fridman (1:37:36.400)
Yeah.
Lex Fridman (1:37:37.560)
So yeah, I mean, you can do,
Lex Fridman (1:37:39.180)
we have a little data engine like thing.
George Hotz (1:37:42.040)
We're training a segment.
Lex Fridman (1:37:43.560)
I know it's not fancy.
George Hotz (1:37:44.560)
It's just like, okay, train the new segment,
Lex Fridman (1:37:48.280)
run it on 100,000 images
Lex Fridman (1:37:50.080)
and now take the thousand with highest loss.
Lex Fridman (1:37:52.160)
Select a hundred of those by human,
George Hotz (1:37:54.160)
put those, get those ones labeled, retrain, do it again.
Lex Fridman (1:37:57.840)
And so it's a much less well written data engine.
Lex Fridman (1:38:01.480)
And yeah, you can take these things really far
Lex Fridman (1:38:03.600)
and it is impressive engineering.
Lex Fridman (1:38:06.600)
And if you truly need supervised data for a problem,
Lex Fridman (1:38:09.920)
yeah, things like data engine are at the high end
Lex Fridman (1:38:12.560)
of what is attention?
Lex Fridman (1:38:14.960)
Is a human paying attention?
George Hotz (1:38:15.960)
I mean, we're going to probably build something
Lex Fridman (1:38:17.960)
that looks like data engine
George Hotz (1:38:18.940)
to push our driver monitoring further.
Lex Fridman (1:38:21.120)
But for driving itself,
George Hotz (1:38:22.920)
you have it all annotated beautifully by what the human does.
Lex Fridman (1:38:26.400)
Yeah, that's interesting.
George Hotz (1:38:27.240)
I mean, that applies to driver attention as well.
Lex Fridman (1:38:30.040)
Do you want to detect the eyes?
Lex Fridman (1:38:31.200)
Do you want to detect blinking and pupil movement?
Lex Fridman (1:38:33.500)
Do you want to detect all the like face alignments
George Hotz (1:38:36.680)
or landmark detection and so on,
Lex Fridman (1:38:38.720)
and then doing kind of reasoning based on that?
George Hotz (1:38:41.560)
Or do you want to take the entirety of the face over time
Lex Fridman (1:38:43.880)
and do end to end?
George Hotz (1:38:45.320)
I mean, it's obvious that eventually you have to do end
Lex Fridman (1:38:48.480)
to end with some calibration, some fixes and so on,
Lex Fridman (1:38:51.360)
but it's like, I don't know when that's the right move.
Lex Fridman (1:38:55.760)
Even if it's end to end, there actually is,
George Hotz (1:38:58.420)
there is no kind of, you have to supervise that with humans.
Lex Fridman (1:39:03.380)
Whether a human is paying attention or not
George Hotz (1:39:05.480)
is a completely subjective judgment.
Lex Fridman (1:39:08.560)
Like you can try to like automatically do it
George Hotz (1:39:11.040)
with some stuff, but you don't have,
Lex Fridman (1:39:13.160)
if I record a video of a human,
George Hotz (1:39:15.080)
I don't have true annotations anywhere in that video.
Lex Fridman (1:39:18.400)
The only way to get them is with,
George Hotz (1:39:21.120)
you know, other humans labeling it really.
Lex Fridman (1:39:22.840)
Well, I don't know.
George Hotz (1:39:26.080)
If you think deeply about it,
Lex Fridman (1:39:28.540)
you could, you might be able to just,
George Hotz (1:39:30.160)
depending on the task,
Lex Fridman (1:39:31.000)
maybe a discover self annotating things like,
George Hotz (1:39:34.320)
you know, you can look at like steering wheel reverse
Lex Fridman (1:39:36.920)
or something like that.
George Hotz (1:39:37.760)
You can discover little moments of lapse of attention.
Lex Fridman (1:39:41.200)
I mean, that's where psychology comes in.
George Hotz (1:39:44.540)
Is there indicate,
Lex Fridman (1:39:45.480)
cause you have so much data to look at.
Lex Fridman (1:39:48.000)
So you might be able to find moments when there's like,
Lex Fridman (1:39:51.140)
just inattention that even with smartphone,
George Hotz (1:39:54.320)
if you want to detect smartphone use,
Lex Fridman (1:39:56.700)
you can start to zoom in.
George Hotz (1:39:57.880)
I mean, that's the gold mine, sort of the comma AI.
Lex Fridman (1:40:01.480)
I mean, Tesla is doing this too, right?
George Hotz (1:40:02.920)
Is they're doing annotation based on,
Lex Fridman (1:40:06.920)
it's like a self supervised learning too.
George Hotz (1:40:10.500)
It's just a small part of the entire picture.
Lex Fridman (1:40:13.440)
That's kind of the challenge of solving a problem
George Hotz (1:40:17.760)
in machine learning.
Lex Fridman (1:40:18.660)
If you can discover self annotating parts of the problem,
Lex Fridman (1:40:24.000)
right?
Lex Fridman (1:40:25.020)
Our driver monitoring team is half a person right now.
George Hotz (1:40:27.760)
I would, you know, once we have,
Lex Fridman (1:40:29.280)
once we have two, three people on that team,
George Hotz (1:40:33.280)
I definitely want to look at self annotating stuff
Lex Fridman (1:40:35.280)
for attention.
George Hotz (1:40:38.200)
Let's go back for a sec to a comma and what,
Lex Fridman (1:40:43.720)
you know, for people who are curious to try it out,
Lex Fridman (1:40:46.240)
how do you install a comma in say a 2020 Toyota Corolla
Lex Fridman (1:40:51.120)
or like, what are the cars that are supported?
Lex Fridman (1:40:53.400)
What are the cars that you recommend?
Lex Fridman (1:40:55.500)
And what does it take?
George Hotz (1:40:57.880)
You have a few videos out, but maybe through words,
Lex Fridman (1:41:00.000)
can you explain what's it take to actually install a thing?
Lex Fridman (1:41:02.880)
So we support, I think it's 91 cars, 91 makes the models.
Lex Fridman (1:41:08.080)
We've got to 100 this year.
George Hotz (1:41:10.160)
Nice.
Lex Fridman (1:41:11.000)
The, yeah, the 2020 Corolla, great choice.
George Hotz (1:41:16.680)
The 2020 Sonata, it's using the stock longitudinal.
Lex Fridman (1:41:21.200)
It's using just our lateral control,
Lex Fridman (1:41:23.280)
but it's a very refined car.
Lex Fridman (1:41:25.140)
Their longitudinal control is not bad at all.
Lex Fridman (1:41:28.200)
So yeah, Corolla, Sonata,
Lex Fridman (1:41:31.720)
or if you're willing to get your hands a little dirty
Lex Fridman (1:41:34.240)
and look in the right places on the internet,
Lex Fridman (1:41:35.940)
the Honda Civic is great,
Lex Fridman (1:41:37.520)
but you're going to have to install a modified EPS firmware
Lex Fridman (1:41:40.600)
in order to get a little bit more torque.
Lex Fridman (1:41:42.160)
And I can't help you with that.
Lex Fridman (1:41:43.380)
Comma does not officially endorse that,
Lex Fridman (1:41:45.960)
but we have been doing it.
Lex Fridman (1:41:47.560)
We didn't ever release it.
George Hotz (1:41:49.800)
We waited for someone else to discover it.
Lex Fridman (1:41:51.440)
And then, you know.
Lex Fridman (1:41:52.880)
And you have a Discord server where people,
Lex Fridman (1:41:55.680)
there's a very active developer community, I suppose.
Lex Fridman (1:42:00.360)
So depending on the level of experimentation
Lex Fridman (1:42:04.000)
you're willing to do, that's the community.
George Hotz (1:42:07.600)
If you just want to buy it and you have a supported car,
Lex Fridman (1:42:11.240)
it's 10 minutes to install.
George Hotz (1:42:13.920)
There's YouTube videos.
Lex Fridman (1:42:15.400)
It's Ikea furniture level.
George Hotz (1:42:17.040)
If you can set up a table from Ikea,
Lex Fridman (1:42:19.040)
you can install a Comma 2 in your supported car
Lex Fridman (1:42:21.200)
and it will just work.
Lex Fridman (1:42:22.600)
Now you're like, oh, but I want this high end feature
George Hotz (1:42:24.920)
or I want to fix this bug.
Lex Fridman (1:42:26.160)
Okay, well, welcome to the developer community.
Lex Fridman (1:42:29.520)
So what, if I wanted to,
Lex Fridman (1:42:31.000)
this is something I asked you offline like a few months ago.
George Hotz (1:42:34.680)
If I wanted to run my own code to,
Lex Fridman (1:42:39.800)
so use Comma as a platform
Lex Fridman (1:42:43.420)
and try to run something like OpenPilot,
Lex Fridman (1:42:46.040)
what does it take to do that?
Lex Fridman (1:42:48.320)
So there's a toggle in the settings called enable SSH.
Lex Fridman (1:42:51.840)
And if you toggle that, you can SSH into your device.
George Hotz (1:42:54.600)
You can modify the code.
Lex Fridman (1:42:55.620)
You can upload whatever code you want to it.
George Hotz (1:42:58.260)
There's a whole lot of people.
Lex Fridman (1:42:59.160)
So about 60% of people are running stock comma.
George Hotz (1:43:03.040)
About 40% of people are running forks.
Lex Fridman (1:43:05.480)
And there's a community of,
George Hotz (1:43:07.280)
there's a bunch of people who maintain these forks
Lex Fridman (1:43:10.320)
and these forks support different cars
George Hotz (1:43:13.040)
or they have different toggles.
Lex Fridman (1:43:15.700)
We try to keep away from the toggles
George Hotz (1:43:17.360)
that are like disabled driver monitoring,
Lex Fridman (1:43:18.920)
but there's some people might want that kind of thing
Lex Fridman (1:43:21.720)
and like, yeah, you can, it's your car.
Lex Fridman (1:43:24.560)
I'm not here to tell you.
George Hotz (1:43:29.400)
We have some, we ban,
Lex Fridman (1:43:31.080)
if you're trying to subvert safety features,
George Hotz (1:43:32.920)
you're banned from our Discord.
Lex Fridman (1:43:33.760)
I don't want anything to do with you,
Lex Fridman (1:43:35.260)
but there's some forks doing that.
Lex Fridman (1:43:37.920)
Got it.
Lex Fridman (1:43:39.880)
So you encourage responsible forking.
Lex Fridman (1:43:42.880)
Yeah, yeah.
George Hotz (1:43:43.720)
We encourage, some people, yeah, some people,
Lex Fridman (1:43:46.080)
like there's forks that will do,
George Hotz (1:43:48.160)
some people just like having a lot of readouts on the UI,
Lex Fridman (1:43:52.040)
like a lot of like flashing numbers.
Lex Fridman (1:43:53.440)
So there's forks that do that.
Lex Fridman (1:43:55.120)
Some people don't like the fact that it disengages
George Hotz (1:43:57.200)
when you press the gas pedal.
Lex Fridman (1:43:58.320)
There's forks that disable that.
George Hotz (1:44:00.480)
Got it.
Lex Fridman (1:44:01.320)
Now the stock experience is what like,
Lex Fridman (1:44:04.920)
so it does both lane keeping
Lex Fridman (1:44:06.240)
and longitudinal control all together.
Lex Fridman (1:44:08.960)
So it's not separate like it is in autopilot.
Lex Fridman (1:44:11.020)
No, so, okay.
George Hotz (1:44:12.520)
Some cars we use the stock longitudinal control.
Lex Fridman (1:44:15.040)
We don't do the longitudinal control in all the cars.
George Hotz (1:44:17.400)
Some cars, the ACCs are pretty good in the cars.
Lex Fridman (1:44:19.560)
It's the lane keep that's atrocious in anything
George Hotz (1:44:21.360)
except for autopilot and super cruise.
Lex Fridman (1:44:23.420)
But, you know, you just turn it on and it works.
Lex Fridman (1:44:27.880)
What does this engagement look like?
Lex Fridman (1:44:29.480)
Yeah, so we have, I mean,
George Hotz (1:44:30.960)
I'm very concerned about mode confusion.
Lex Fridman (1:44:32.960)
I've experienced it on super cruise and autopilot
George Hotz (1:44:36.960)
where like autopilot, like autopilot disengages.
Lex Fridman (1:44:39.820)
I don't realize that the ACC is still on.
George Hotz (1:44:42.400)
The lead car moves slightly over
Lex Fridman (1:44:44.680)
and then the Tesla accelerates
George Hotz (1:44:46.160)
to like whatever my set speed is super fast.
Lex Fridman (1:44:48.000)
I'm like, what's going on here?
George Hotz (1:44:51.320)
We have engaged and disengaged.
Lex Fridman (1:44:53.720)
And this is similar to my understanding, I'm not a pilot,
Lex Fridman (1:44:56.380)
but my understanding is either the pilot is in control
Lex Fridman (1:45:00.080)
or the copilot is in control.
Lex Fridman (1:45:02.080)
And we have the same kind of transition system.
Lex Fridman (1:45:05.120)
Either open pilot is engaged or open pilot is disengaged.
George Hotz (1:45:08.620)
Engage with cruise control,
Lex Fridman (1:45:10.160)
disengage with either gas brake or cancel.
George Hotz (1:45:13.320)
Let's talk about money.
Lex Fridman (1:45:14.560)
What's the business strategy for Kama?
George Hotz (1:45:17.360)
Profitable.
Lex Fridman (1:45:18.840)
Well, so you're.
George Hotz (1:45:19.680)
We did it.
Lex Fridman (1:45:20.520)
So congratulations.
George Hotz (1:45:23.200)
What, so basically selling,
Lex Fridman (1:45:25.760)
so we should say Kama cost a thousand bucks, Kama two?
George Hotz (1:45:29.960)
200 for the interface to the car as well.
Lex Fridman (1:45:31.640)
It's 1200, I'll send that.
George Hotz (1:45:34.360)
Nobody's usually upfront like this.
Lex Fridman (1:45:36.400)
Yeah, you gotta add the tack on, right?
George Hotz (1:45:38.200)
Yeah.
Lex Fridman (1:45:39.040)
I love it.
George Hotz (1:45:39.860)
I'm not gonna lie to you.
Lex Fridman (1:45:41.080)
Trust me, it will add $1,200 of value to your life.
George Hotz (1:45:43.840)
Yes, it's still super cheap.
Lex Fridman (1:45:45.560)
30 days, no questions asked, money back guarantee,
Lex Fridman (1:45:47.880)
and prices are only going up.
Lex Fridman (1:45:50.400)
If there ever is future hardware,
George Hotz (1:45:52.320)
it could cost a lot more than $1,200.
Lex Fridman (1:45:53.840)
So Kama three is in the works.
George Hotz (1:45:56.960)
It could be.
Lex Fridman (1:45:57.800)
All I will say is future hardware
George Hotz (1:45:59.640)
is going to cost a lot more than the current hardware.
Lex Fridman (1:46:02.900)
Yeah, the people that use,
George Hotz (1:46:05.260)
the people I've spoken with that use Kama,
Lex Fridman (1:46:07.880)
that use open pilot,
George Hotz (1:46:10.320)
first of all, they use it a lot.
Lex Fridman (1:46:12.120)
So people that use it, they fall in love with it.
George Hotz (1:46:14.400)
Oh, our retention rate is insane.
Lex Fridman (1:46:16.840)
It's a good sign.
George Hotz (1:46:17.740)
Yeah.
Lex Fridman (1:46:18.580)
It's a really good sign.
George Hotz (1:46:19.680)
70% of Kama two buyers are daily active users.
Lex Fridman (1:46:23.720)
Yeah, it's amazing.
George Hotz (1:46:27.760)
Oh, also, we don't plan on stopping selling the Kama two.
Lex Fridman (1:46:30.520)
Like it's, you know.
Lex Fridman (1:46:31.960)
So whatever you create that's beyond Kama two,
Lex Fridman (1:46:36.600)
it would be potentially a phase shift.
George Hotz (1:46:40.760)
Like it's so much better that,
Lex Fridman (1:46:42.840)
like you could use Kama two
Lex Fridman (1:46:44.200)
and you can use Kama whatever.
Lex Fridman (1:46:45.760)
Depends what you want.
George Hotz (1:46:46.600)
It's 3.41, 42.
Lex Fridman (1:46:48.320)
Yeah.
George Hotz (1:46:49.160)
You know, autopilot hardware one versus hardware two.
Lex Fridman (1:46:52.200)
The Kama two is kind of like hardware one.
George Hotz (1:46:53.600)
Got it, got it.
Lex Fridman (1:46:54.440)
You can still use both.
George Hotz (1:46:55.260)
Got it, got it.
Lex Fridman (1:46:56.320)
I think I heard you talk about retention rate
George Hotz (1:46:58.000)
with the VR headsets that the average is just once.
Lex Fridman (1:47:01.240)
Yeah.
George Hotz (1:47:02.080)
Just fast.
Lex Fridman (1:47:02.900)
I mean, it's such a fascinating way
George Hotz (1:47:03.880)
to think about technology.
Lex Fridman (1:47:05.760)
And this is a really, really good sign.
Lex Fridman (1:47:07.420)
And the other thing that people say about Kama
Lex Fridman (1:47:09.000)
is like they can't believe they're getting this 4,000 bucks.
Lex Fridman (1:47:12.060)
Right?
Lex Fridman (1:47:12.900)
It seems like some kind of steal.
George Hotz (1:47:17.020)
So, but in terms of like longterm business strategies
Lex Fridman (1:47:20.040)
that basically to put,
Lex Fridman (1:47:21.640)
so it's currently in like a thousand plus cars.
Lex Fridman (1:47:27.560)
1,200.
George Hotz (1:47:28.520)
More, more.
Lex Fridman (1:47:30.560)
So yeah, dailies is about, dailies is about 2,000.
George Hotz (1:47:35.560)
Weeklys is about 2,500, monthlys is over 3,000.
Lex Fridman (1:47:38.960)
Wow.
George Hotz (1:47:39.800)
We've grown a lot since we last talked.
Lex Fridman (1:47:42.120)
Is the goal, like can we talk crazy for a second?
Lex Fridman (1:47:44.800)
I mean, what's the goal to overtake Tesla?
Lex Fridman (1:47:48.620)
Let's talk, okay, so.
George Hotz (1:47:49.960)
I mean, Android did overtake iOS.
Lex Fridman (1:47:51.480)
That's exactly it, right?
Lex Fridman (1:47:52.520)
So they did it.
Lex Fridman (1:47:55.360)
I actually don't know the timeline of that one.
Lex Fridman (1:47:57.720)
But let's talk, because everything is in alpha now.
Lex Fridman (1:48:02.160)
The autopilot you could argue is in alpha
George Hotz (1:48:03.960)
in terms of towards the big mission
Lex Fridman (1:48:05.840)
of autonomous driving, right?
Lex Fridman (1:48:07.680)
And so what, yeah, is your goal to overtake
Lex Fridman (1:48:11.600)
millions of cars essentially?
George Hotz (1:48:13.920)
Of course.
Lex Fridman (1:48:15.520)
Where would it stop?
George Hotz (1:48:16.760)
Like it's open source software.
Lex Fridman (1:48:18.000)
It might not be millions of cars
George Hotz (1:48:19.280)
with a piece of comma hardware, but yeah.
Lex Fridman (1:48:21.360)
I think open pilot at some point
George Hotz (1:48:24.320)
will cross over autopilot in users,
Lex Fridman (1:48:26.920)
just like Android crossed over iOS.
Lex Fridman (1:48:29.200)
How does Google make money from Android?
Lex Fridman (1:48:31.240)
It's complicated.
George Hotz (1:48:34.840)
Their own devices make money.
Lex Fridman (1:48:37.400)
Google, Google makes money
George Hotz (1:48:39.480)
by just kind of having you on the internet.
Lex Fridman (1:48:42.240)
Yes.
George Hotz (1:48:43.080)
Google search is built in, Gmail is built in.
Lex Fridman (1:48:45.620)
Android is just a shill
George Hotz (1:48:46.540)
for the rest of Google's ecosystem.
Lex Fridman (1:48:48.200)
Yeah, but the problem is Android is not,
George Hotz (1:48:50.680)
is a brilliant thing.
Lex Fridman (1:48:52.440)
I mean, Android arguably changed the world.
Lex Fridman (1:48:55.080)
So there you go.
Lex Fridman (1:48:56.440)
That's, you can feel good ethically speaking.
Lex Fridman (1:49:00.800)
But as a business strategy, it's questionable.
Lex Fridman (1:49:04.320)
Or sell hardware.
George Hotz (1:49:05.720)
Sell hardware.
Lex Fridman (1:49:06.560)
I mean, it took Google a long time to come around to it,
Lex Fridman (1:49:08.120)
but they are now making money on the Pixel.
Lex Fridman (1:49:10.000)
You're not about money, you're more about winning.
George Hotz (1:49:13.240)
Yeah, of course.
Lex Fridman (1:49:14.160)
No, but if only 10% of open pilot devices
George Hotz (1:49:18.320)
come from comma AI.
Lex Fridman (1:49:19.920)
They still make a lot.
George Hotz (1:49:20.820)
That is still, yes.
Lex Fridman (1:49:21.660)
That is a ton of money for our company.
Lex Fridman (1:49:22.840)
But can't somebody create a better comma using open pilot?
Lex Fridman (1:49:27.120)
Or are you basically saying, well, I'll compete them?
George Hotz (1:49:28.840)
Well, I'll compete you.
Lex Fridman (1:49:29.680)
Can you create a better Android phone than the Google Pixel?
George Hotz (1:49:32.280)
Right.
Lex Fridman (1:49:32.800)
I mean, you can, but like, you know.
George Hotz (1:49:34.680)
I love that.
Lex Fridman (1:49:35.520)
So you're confident, like, you know
Lex Fridman (1:49:37.360)
what the hell you're doing.
Lex Fridman (1:49:38.480)
Yeah.
George Hotz (1:49:40.040)
It's confidence and merit.
Lex Fridman (1:49:43.520)
I mean, our money comes from, we're
George Hotz (1:49:44.960)
a consumer electronics company.
Lex Fridman (1:49:46.160)
Yeah.
Lex Fridman (1:49:46.660)
And put it this way.
Lex Fridman (1:49:48.040)
So we sold like 3,000 comma twos.
George Hotz (1:49:51.600)
2,500 right now.
Lex Fridman (1:49:54.080)
And like, OK, we're probably going
George Hotz (1:49:59.720)
to sell 10,000 units next year.
Lex Fridman (1:50:01.920)
10,000 units, even just $1,000 a unit, OK,
George Hotz (1:50:04.520)
we're at 10 million in revenue.
Lex Fridman (1:50:09.400)
Get that up to 100,000, maybe double the price of the unit.
George Hotz (1:50:12.100)
Now we're talking like 200 million revenue.
Lex Fridman (1:50:13.560)
We're talking like series.
George Hotz (1:50:14.440)
Yeah, actually making money.
Lex Fridman (1:50:15.840)
One of the rare semi autonomous or autonomous vehicle companies
George Hotz (1:50:19.360)
that are actually making money.
Lex Fridman (1:50:21.080)
Yeah.
George Hotz (1:50:22.600)
You know, if you look at a model,
Lex Fridman (1:50:24.880)
and we were just talking about this yesterday.
George Hotz (1:50:26.680)
If you look at a model, and like you're AB testing your model,
Lex Fridman (1:50:29.760)
and if you're one branch of the AB test,
George Hotz (1:50:32.360)
the losses go down very fast in the first five epochs.
Lex Fridman (1:50:35.320)
That model is probably going to converge
George Hotz (1:50:37.520)
to something considerably better than the one
Lex Fridman (1:50:39.360)
where the losses are going down slower.
Lex Fridman (1:50:41.360)
Why do people think this is going to stop?
Lex Fridman (1:50:43.000)
Why do people think one day there's
George Hotz (1:50:44.500)
going to be a great like, well, Waymo's eventually
Lex Fridman (1:50:46.840)
going to surpass you guys?
George Hotz (1:50:49.600)
Well, they're not.
Lex Fridman (1:50:52.000)
Do you see like a world where like a Tesla or a car
George Hotz (1:50:55.800)
like a Tesla would be able to basically press a button
Lex Fridman (1:50:59.160)
and you like switch to open pilot?
George Hotz (1:51:01.920)
You know, you load in.
Lex Fridman (1:51:04.480)
No, so I think so first off, I think
George Hotz (1:51:06.840)
that we may surpass Tesla in terms of users.
Lex Fridman (1:51:10.560)
I do not think we're going to surpass Tesla ever
George Hotz (1:51:12.520)
in terms of revenue.
Lex Fridman (1:51:13.520)
I think Tesla can capture a lot more revenue per user
George Hotz (1:51:16.380)
than we can.
Lex Fridman (1:51:17.320)
But this mimics the Android iOS model exactly.
George Hotz (1:51:20.520)
There may be more Android devices,
Lex Fridman (1:51:22.000)
but there's a lot more iPhones than Google Pixels.
Lex Fridman (1:51:24.320)
So I think there'll be a lot more Tesla cars sold
Lex Fridman (1:51:26.360)
than pieces of common hardware.
Lex Fridman (1:51:30.280)
And then as far as a Tesla owner being
Lex Fridman (1:51:34.420)
able to switch to open pilot, does iPhones run Android?
George Hotz (1:51:40.920)
No, but it doesn't make sense.
Lex Fridman (1:51:42.400)
You can if you really want to do it,
Lex Fridman (1:51:43.480)
but it doesn't really make sense.
Lex Fridman (1:51:44.440)
Like it's not.
George Hotz (1:51:45.320)
It doesn't make sense.
Lex Fridman (1:51:46.240)
Who cares?
Lex Fridman (1:51:46.740)
What about if a large company like automakers, Ford, GM,
Lex Fridman (1:51:51.640)
Toyota came to George Hots?
George Hotz (1:51:53.720)
Or on the tech space, Amazon, Facebook, Google
Lex Fridman (1:51:58.000)
came with a large pile of cash?
Lex Fridman (1:52:01.080)
Would you consider being purchased?
Lex Fridman (1:52:07.360)
Do you see that as a one possible?
George Hotz (1:52:10.500)
Not seriously, no.
Lex Fridman (1:52:12.360)
I would probably see how much shit they'll entertain for me.
Lex Fridman (1:52:19.680)
And if they're willing to jump through a bunch of my hoops,
Lex Fridman (1:52:22.080)
then maybe.
Lex Fridman (1:52:22.960)
But no, not the way that M&A works today.
Lex Fridman (1:52:25.200)
I mean, we've been approached.
Lex Fridman (1:52:26.520)
And I laugh in these people's faces.
Lex Fridman (1:52:28.000)
I'm like, are you kidding?
George Hotz (1:52:31.000)
Yeah.
Lex Fridman (1:52:31.600)
Because it's so demeaning.
George Hotz (1:52:33.680)
The M&A people are so demeaning to companies.
Lex Fridman (1:52:36.960)
They treat the startup world as their innovation ecosystem.
Lex Fridman (1:52:41.340)
And they think that I'm cool with going along with that,
Lex Fridman (1:52:43.640)
so I can have some of their scam fake Fed dollars.
George Hotz (1:52:46.440)
Fed coin.
Lex Fridman (1:52:47.680)
What am I going to do with more Fed coin?
George Hotz (1:52:49.680)
Fed coin.
Lex Fridman (1:52:50.320)
Fed coin, man.
George Hotz (1:52:51.400)
I love that.
Lex Fridman (1:52:52.120)
So that's the cool thing about podcasting,
George Hotz (1:52:54.040)
actually, is people criticize.
Lex Fridman (1:52:56.200)
I don't know if you're familiar with Spotify giving Joe Rogan
George Hotz (1:53:00.560)
$100 million.
Lex Fridman (1:53:01.920)
I don't know about that.
Lex Fridman (1:53:03.720)
And they respect, despite all the shit
Lex Fridman (1:53:08.200)
that people are talking about Spotify,
George Hotz (1:53:11.440)
people understand that podcasters like Joe Rogan
Lex Fridman (1:53:15.300)
know what the hell they're doing.
Lex Fridman (1:53:17.200)
So they give them money and say, just do what you do.
Lex Fridman (1:53:21.160)
And the equivalent for you would be like,
George Hotz (1:53:25.440)
George, do what the hell you do, because you're good at it.
Lex Fridman (1:53:28.440)
Try not to murder too many people.
George Hotz (1:53:31.400)
There's some kind of common sense things,
Lex Fridman (1:53:33.480)
like just don't go on a weird rampage of it.
George Hotz (1:53:37.400)
Yeah.
Lex Fridman (1:53:38.080)
It comes down to what companies I could respect, right?
Lex Fridman (1:53:43.400)
Could I respect GM?
Lex Fridman (1:53:44.440)
Never.
George Hotz (1:53:46.520)
No, I couldn't.
Lex Fridman (1:53:47.480)
I mean, could I respect a Hyundai?
George Hotz (1:53:50.840)
More so.
Lex Fridman (1:53:52.520)
That's a lot closer.
Lex Fridman (1:53:53.560)
Toyota?
Lex Fridman (1:53:54.720)
What's your?
George Hotz (1:53:55.840)
Nah.
Lex Fridman (1:53:56.840)
Nah.
George Hotz (1:53:57.520)
Korean is the way.
Lex Fridman (1:53:59.400)
I think that the Japanese, the Germans, the US, they're all
George Hotz (1:54:02.440)
too, they're all too, they all think they're too great.
Lex Fridman (1:54:05.720)
What about the tech companies?
Lex Fridman (1:54:07.520)
Apple?
Lex Fridman (1:54:08.480)
Apple is, of the tech companies that I could respect,
George Hotz (1:54:11.000)
Apple's the closest.
Lex Fridman (1:54:12.120)
Yeah.
George Hotz (1:54:12.600)
I mean, I could never.
Lex Fridman (1:54:13.560)
It would be ironic.
George Hotz (1:54:14.440)
It would be ironic if Comma AI is acquired by Apple.
Lex Fridman (1:54:19.840)
I mean, Facebook, look, I quit Facebook 10 years ago
George Hotz (1:54:21.840)
because I didn't respect the business model.
Lex Fridman (1:54:24.320)
Google has declined so fast in the last five years.
Lex Fridman (1:54:28.480)
What are your thoughts about Waymo and its present
Lex Fridman (1:54:32.080)
and its future?
George Hotz (1:54:33.200)
Let me start by saying something nice, which is I've
Lex Fridman (1:54:39.240)
visited them a few times and have ridden in their cars.
Lex Fridman (1:54:45.240)
And the engineering that they're doing,
Lex Fridman (1:54:49.760)
both the research and the actual development
Lex Fridman (1:54:51.720)
and the engineering they're doing
Lex Fridman (1:54:53.360)
and the scale they're actually achieving
George Hotz (1:54:55.160)
by doing it all themselves is really impressive.
Lex Fridman (1:54:58.400)
And the balance of safety and innovation.
Lex Fridman (1:55:01.480)
And the cars work really well for the routes they drive.
Lex Fridman (1:55:07.520)
It drives fast, which was very surprising to me.
George Hotz (1:55:10.960)
It drives the speed limit or faster than the speed limit.
Lex Fridman (1:55:14.800)
It goes.
Lex Fridman (1:55:16.120)
And it works really damn well.
Lex Fridman (1:55:17.800)
And the interface is nice.
George Hotz (1:55:19.160)
In Chandler, Arizona, yeah.
Lex Fridman (1:55:20.360)
Yeah, in Chandler, Arizona, very specific environment.
Lex Fridman (1:55:22.560)
So it gives me enough material in my mind
Lex Fridman (1:55:27.360)
to push back against the madmen of the world,
George Hotz (1:55:30.360)
like George Hotz, to be like, because you kind of imply
Lex Fridman (1:55:36.040)
there's zero probability they're going to win.
Lex Fridman (1:55:38.760)
And after I've used, after I've ridden in it, to me,
Lex Fridman (1:55:43.560)
it's not zero.
George Hotz (1:55:44.440)
Oh, it's not for technology reasons.
Lex Fridman (1:55:46.760)
Bureaucracy?
George Hotz (1:55:48.000)
No, it's worse than that.
Lex Fridman (1:55:49.520)
It's actually for product reasons, I think.
George Hotz (1:55:51.840)
Oh, you think they're just not capable of creating
Lex Fridman (1:55:53.840)
an amazing product?
George Hotz (1:55:55.600)
No, I think that the product that they're building
Lex Fridman (1:55:58.280)
doesn't make sense.
Lex Fridman (1:56:01.040)
So a few things.
Lex Fridman (1:56:03.320)
You say the Waymo's are fast.
George Hotz (1:56:05.840)
Benchmark a Waymo against a competent Uber driver.
Lex Fridman (1:56:09.040)
Right.
Lex Fridman (1:56:09.600)
Right?
Lex Fridman (1:56:10.080)
The Uber driver's faster.
George Hotz (1:56:11.080)
It's not even about speed.
Lex Fridman (1:56:12.200)
It's the thing you said.
George Hotz (1:56:13.240)
It's about the experience of being stuck at a stop sign
Lex Fridman (1:56:16.280)
because pedestrians are crossing nonstop.
George Hotz (1:56:20.120)
I like when my Uber driver doesn't come to a full stop
Lex Fridman (1:56:22.160)
at the stop sign.
George Hotz (1:56:22.960)
Yeah.
Lex Fridman (1:56:23.520)
You know?
Lex Fridman (1:56:24.440)
And so let's say the Waymo's are 20% slower than an Uber.
Lex Fridman (1:56:31.320)
Right?
George Hotz (1:56:33.000)
You can argue that they're going to be cheaper.
Lex Fridman (1:56:35.160)
And I argue that users already have the choice
George Hotz (1:56:37.640)
to trade off money for speed.
Lex Fridman (1:56:39.360)
It's called UberPool.
George Hotz (1:56:42.280)
I think it's like 15% of rides are UberPools.
Lex Fridman (1:56:45.560)
Right?
George Hotz (1:56:46.040)
Users are not willing to trade off money for speed.
Lex Fridman (1:56:49.480)
So the whole product that they're building
George Hotz (1:56:52.320)
is not going to be competitive with traditional ride sharing
Lex Fridman (1:56:56.200)
networks.
George Hotz (1:56:56.720)
Right.
Lex Fridman (1:56:59.560)
And also, whether there's profit to be made
George Hotz (1:57:04.360)
depends entirely on one company having a monopoly.
Lex Fridman (1:57:07.360)
I think that the level four autonomous ride sharing
George Hotz (1:57:11.200)
vehicles market is going to look a lot like the scooter market
Lex Fridman (1:57:14.800)
if even the technology does come to exist, which I question.
Lex Fridman (1:57:18.680)
Who's doing well in that market?
Lex Fridman (1:57:20.400)
It's a race to the bottom.
George Hotz (1:57:22.320)
Well, it could be closer like an Uber and a Lyft,
Lex Fridman (1:57:25.680)
where it's just one or two players.
George Hotz (1:57:28.000)
Well, the scooter people have given up
Lex Fridman (1:57:31.160)
trying to market scooters as a practical means
George Hotz (1:57:34.760)
of transportation.
Lex Fridman (1:57:35.480)
And they're just like, they're super fun to ride.
George Hotz (1:57:37.640)
Look at wheels.
Lex Fridman (1:57:38.360)
I love those things.
Lex Fridman (1:57:39.160)
And they're great on that front.
Lex Fridman (1:57:40.560)
Yeah.
Lex Fridman (1:57:41.040)
But from an actual transportation product
Lex Fridman (1:57:43.840)
perspective, I do not think scooters are viable.
Lex Fridman (1:57:46.240)
And I do not think level four autonomous cars are viable.
Lex Fridman (1:57:49.200)
If you, let's play a fun experiment.
George Hotz (1:57:51.560)
If you ran, let's do a Tesla and let's do Waymo.
Lex Fridman (1:57:56.960)
If Elon Musk took a vacation for a year, he just said,
George Hotz (1:58:01.880)
screw it, I'm going to go live on an island, no electronics.
Lex Fridman (1:58:05.400)
And the board decides that we need to find somebody
George Hotz (1:58:07.520)
to run the company.
Lex Fridman (1:58:09.160)
And they did decide that you should run the company
George Hotz (1:58:11.520)
for a year.
Lex Fridman (1:58:12.240)
How do you run Tesla differently?
George Hotz (1:58:14.920)
I wouldn't change much.
Lex Fridman (1:58:16.240)
Do you think they're on the right track?
George Hotz (1:58:17.920)
I wouldn't change.
Lex Fridman (1:58:18.680)
I mean, I'd have some minor changes.
Lex Fridman (1:58:21.640)
But even my debate with Tesla about end
Lex Fridman (1:58:25.520)
to end versus SegNets, that's just software.
Lex Fridman (1:58:29.120)
Who cares?
Lex Fridman (1:58:30.720)
It's not like you're doing something terrible with SegNets.
George Hotz (1:58:33.920)
You're probably building something that's
Lex Fridman (1:58:35.600)
at least going to help you debug the end to end system a lot.
George Hotz (1:58:39.440)
It's very easy to transition from what they have
Lex Fridman (1:58:42.200)
to an end to end kind of thing.
Lex Fridman (1:58:45.840)
And then I presume you would, in the Model Y
Lex Fridman (1:58:50.520)
or maybe in the Model 3, start adding driver
George Hotz (1:58:52.480)
sensing with infrared.
Lex Fridman (1:58:53.600)
Yes, I would add infrared camera, infrared lights
George Hotz (1:58:58.000)
right away to those cars.
Lex Fridman (1:59:02.120)
And start collecting that data and do all that kind of stuff,
George Hotz (1:59:04.920)
yeah.
Lex Fridman (1:59:05.440)
Very much.
George Hotz (1:59:06.080)
I think they're already kind of doing it.
Lex Fridman (1:59:07.780)
It's an incredibly minor change.
George Hotz (1:59:09.320)
If I actually were CEO of Tesla, first off,
Lex Fridman (1:59:11.160)
I'd be horrified that I wouldn't be able to do
George Hotz (1:59:13.080)
a better job as Elon.
Lex Fridman (1:59:14.000)
And then I would try to understand
George Hotz (1:59:16.480)
the way he's done things before.
Lex Fridman (1:59:17.760)
You would also have to take over his Twitter.
George Hotz (1:59:20.720)
I don't tweet.
Lex Fridman (1:59:22.200)
Yeah, what's your Twitter situation?
Lex Fridman (1:59:24.240)
Why are you so quiet on Twitter?
Lex Fridman (1:59:25.880)
Since Dukama is like what's your social network presence like?
George Hotz (1:59:30.280)
Because on Instagram, you do live streams.
Lex Fridman (1:59:34.280)
You understand the music of the internet,
Lex Fridman (1:59:39.240)
but you don't always fully engage into it.
Lex Fridman (1:59:41.520)
You're part time.
George Hotz (1:59:42.800)
Well, I used to have a Twitter.
Lex Fridman (1:59:44.160)
Yeah, I mean, Instagram is a pretty place.
George Hotz (1:59:47.600)
Instagram is a beautiful place.
Lex Fridman (1:59:49.120)
It glorifies beauty.
George Hotz (1:59:49.960)
I like Instagram's values as a network.
Lex Fridman (1:59:53.360)
Twitter glorifies conflict, glorifies shots,
Lex Fridman (20:02.540)
And the things that people get worked up about, you know?
Lex Fridman (20:06.700)
So basically, it was more a message
George Hotz (20:10.060)
of we should humble ourselves.
Lex Fridman (20:12.540)
That we get to, like what are we humans in this byte code?
George Hotz (20:19.460)
Yeah, and not just humble ourselves,
Lex Fridman (20:22.380)
but like I'm not trying to like make people guilty
George Hotz (20:24.900)
or anything like that.
Lex Fridman (20:25.740)
I'm trying to say like, literally,
Lex Fridman (20:27.260)
look at what you are spending time on, right?
Lex Fridman (20:30.200)
What are you referring to?
Lex Fridman (20:31.040)
You're referring to the Kardashians?
Lex Fridman (20:32.460)
What are we talking about?
Lex Fridman (20:34.140)
Twitter?
Lex Fridman (20:34.980)
No, the Kardashians, everyone knows that's kind of fun.
Lex Fridman (20:38.060)
I'm referring more to like the economy, you know?
Lex Fridman (20:42.980)
This idea that we gotta up our stock price.
Lex Fridman (20:50.720)
Or what is the goal function of humanity?
Lex Fridman (20:55.380)
You don't like the game of capitalism?
George Hotz (20:57.600)
Like you don't like the games we've constructed
Lex Fridman (20:59.340)
for ourselves as humans?
George Hotz (21:00.660)
I'm a big fan of capitalism.
Lex Fridman (21:02.860)
I don't think that's really the game we're playing right now.
George Hotz (21:05.120)
I think we're playing a different game
Lex Fridman (21:07.260)
where the rules are rigged.
George Hotz (21:10.300)
Okay, which games are interesting to you
Lex Fridman (21:12.580)
that we humans have constructed and which aren't?
Lex Fridman (21:14.780)
Which are productive and which are not?
Lex Fridman (21:18.380)
Actually, maybe that's the real point of the talk.
George Hotz (21:21.880)
It's like, stop playing these fake human games.
Lex Fridman (21:25.100)
There's a real game here.
George Hotz (21:26.780)
We can play the real game.
Lex Fridman (21:28.660)
The real game is, you know, nature wrote the rules.
George Hotz (21:31.260)
This is a real game.
Lex Fridman (21:32.540)
There still is a game to play.
Lex Fridman (21:35.180)
But if you look at, sorry to interrupt,
Lex Fridman (21:36.940)
I don't know if you've seen the Instagram account,
George Hotz (21:38.420)
nature is metal.
Lex Fridman (21:40.220)
The game that nature seems to be playing
George Hotz (21:42.820)
is a lot more cruel than we humans want to put up with.
Lex Fridman (21:47.340)
Or at least we see it as cruel.
George Hotz (21:49.520)
It's like the bigger thing eats the smaller thing
Lex Fridman (21:53.660)
and does it to impress another big thing
Lex Fridman (21:58.180)
so it can mate with that thing.
Lex Fridman (22:00.460)
And that's it.
George Hotz (22:01.300)
That seems to be the entirety of it.
Lex Fridman (22:04.040)
Well, there's no art, there's no music,
George Hotz (22:07.260)
there's no comma AI, there's no comma one,
Lex Fridman (22:10.860)
no comma two, no George Hots with his brilliant talks
George Hotz (22:14.940)
at South by Southwest.
Lex Fridman (22:17.020)
I disagree, though.
George Hotz (22:17.860)
I disagree that this is what nature is.
Lex Fridman (22:19.620)
I think nature just provided basically a open world MMORPG.
George Hotz (22:26.780)
And, you know, here it's open world.
Lex Fridman (22:29.860)
I mean, if that's the game you want to play,
George Hotz (22:31.100)
you can play that game.
Lex Fridman (22:32.260)
But isn't that beautiful?
George Hotz (22:33.780)
I don't know if you played Diablo.
Lex Fridman (22:35.580)
They used to have, I think, cow level where it's...
Lex Fridman (22:39.420)
So everybody will go just, they figured out this,
Lex Fridman (22:44.420)
like the best way to gain like experience points
George Hotz (22:48.420)
is to just slaughter cows over and over and over.
Lex Fridman (22:52.180)
And so they figured out this little sub game
George Hotz (22:55.900)
within the bigger game that this is the most efficient way
Lex Fridman (22:58.800)
to get experience points.
Lex Fridman (22:59.900)
And everybody somehow agreed
Lex Fridman (23:01.860)
that getting experience points in RPG context
George Hotz (23:04.460)
where you always want to be getting more stuff,
Lex Fridman (23:06.500)
more skills, more levels, keep advancing.
George Hotz (23:09.140)
That seems to be good.
Lex Fridman (23:10.440)
So might as well sacrifice actual enjoyment
George Hotz (23:14.740)
of playing a game, exploring a world,
Lex Fridman (23:17.640)
and spending like hundreds of hours of your time
George Hotz (23:21.580)
at cow level.
Lex Fridman (23:22.420)
I mean, the number of hours I spent in cow level,
George Hotz (23:26.400)
I'm not like the most impressive person
Lex Fridman (23:28.140)
because people have spent probably thousands of hours there,
Lex Fridman (23:30.460)
but it's ridiculous.
Lex Fridman (23:31.580)
So that's a little absurd game that brought me joy
George Hotz (23:35.220)
in some weird dopamine drug kind of way.
Lex Fridman (23:37.500)
So you don't like those games.
George Hotz (23:40.060)
You don't think that's us humans feeling the nature.
Lex Fridman (23:46.500)
I think so.
Lex Fridman (23:47.340)
And that was the point of the talk.
Lex Fridman (23:49.640)
Yeah.
Lex Fridman (23:50.480)
So how do we hack it then?
Lex Fridman (23:51.460)
Well, I want to live forever.
Lex Fridman (23:52.740)
And I want to live forever.
Lex Fridman (23:55.820)
And this is the goal.
George Hotz (23:56.780)
Well, that's a game against nature.
Lex Fridman (23:59.220)
Yeah, immortality is the good objective function to you?
George Hotz (24:03.540)
I mean, start there and then you can do whatever else
Lex Fridman (24:05.100)
you want because you got a long time.
Lex Fridman (24:07.380)
What if immortality makes the game just totally not fun?
Lex Fridman (24:10.740)
I mean, like, why do you assume immortality
Lex Fridman (24:13.860)
is somehow a good objective function?
Lex Fridman (24:18.160)
It's not immortality that I want.
George Hotz (24:19.940)
A true immortality where I could not die,
Lex Fridman (24:22.580)
I would prefer what we have right now.
Lex Fridman (24:25.020)
But I want to choose my own death, of course.
Lex Fridman (24:27.180)
I don't want nature to decide when I die,
George Hotz (24:29.840)
I'm going to win.
Lex Fridman (24:30.780)
I'm going to be you.
Lex Fridman (24:33.100)
And then at some point, if you choose commit suicide,
Lex Fridman (24:36.920)
like how long do you think you'd live?
George Hotz (24:41.700)
Until I get bored.
Lex Fridman (24:43.140)
See, I don't think people like brilliant people like you
George Hotz (24:48.060)
that really ponder living a long time
Lex Fridman (24:52.380)
are really considering how meaningless life becomes.
George Hotz (24:58.620)
Well, I want to know everything and then I'm ready to die.
Lex Fridman (25:03.620)
As long as there's...
George Hotz (25:04.460)
Yeah, but why do you want,
Lex Fridman (25:05.280)
isn't it possible that you want to know everything
Lex Fridman (25:06.940)
because it's finite?
Lex Fridman (25:09.700)
Like the reason you want to know quote unquote everything
George Hotz (25:12.220)
is because you don't have enough time to know everything.
Lex Fridman (25:16.380)
And once you have unlimited time,
Lex Fridman (25:18.780)
then you realize like, why do anything?
Lex Fridman (25:22.220)
Like why learn anything?
George Hotz (25:25.140)
I want to know everything and then I'm ready to die.
Lex Fridman (25:27.100)
So you have, yeah.
George Hotz (25:28.460)
It's not a, like, it's a terminal value.
Lex Fridman (25:30.940)
It's not in service of anything else.
George Hotz (25:34.740)
I'm conscious of the possibility, this is not a certainty,
Lex Fridman (25:37.900)
but the possibility of that engine of curiosity
George Hotz (25:41.800)
that you're speaking to is actually
Lex Fridman (25:47.100)
a symptom of the finiteness of life.
George Hotz (25:49.780)
Like without that finiteness, your curiosity would vanish.
Lex Fridman (25:55.060)
Like a morning fog.
George Hotz (25:57.000)
All right, cool.
Lex Fridman (25:57.840)
Bukowski talked about love like that.
George Hotz (25:59.300)
Then let me solve immortality
Lex Fridman (26:01.340)
and let me change the thing in my brain
George Hotz (26:02.900)
that reminds me of the fact that I'm immortal,
Lex Fridman (26:04.700)
tells me that life is finite shit.
George Hotz (26:06.260)
Maybe I'll have it tell me that life ends next week.
Lex Fridman (26:09.060)
Right?
George Hotz (26:10.660)
I'm okay with some self manipulation like that.
Lex Fridman (26:12.660)
I'm okay with deceiving myself.
George Hotz (26:14.420)
Oh, Rika, changing the code.
Lex Fridman (26:17.020)
Yeah, well, if that's the problem, right?
George Hotz (26:18.300)
If the problem is that I will no longer have that,
Lex Fridman (26:20.580)
that curiosity, I'd like to have backup copies of myself,
George Hotz (26:24.460)
which I check in with occasionally
Lex Fridman (26:27.460)
to make sure they're okay with the trajectory
Lex Fridman (26:29.240)
and they can kind of override it.
Lex Fridman (26:31.000)
Maybe a nice, like, I think of like those wave nets,
George Hotz (26:33.180)
those like logarithmic go back to the copies.
Lex Fridman (26:35.180)
Yeah, but sometimes it's not reversible.
George Hotz (26:36.700)
Like I've done this with video games.
Lex Fridman (26:39.980)
Once you figure out the cheat code
George Hotz (26:41.620)
or like you look up how to cheat old school,
Lex Fridman (26:43.940)
like single player, it ruins the game for you.
George Hotz (26:46.860)
Absolutely.
Lex Fridman (26:47.700)
It ruins that feeling.
Lex Fridman (26:48.540)
But again, that just means our brain manipulation
Lex Fridman (26:51.900)
technology is not good enough yet.
George Hotz (26:53.260)
Remove that cheat code from your brain.
Lex Fridman (26:54.700)
Here you go.
Lex Fridman (26:55.820)
So it's also possible that if we figure out immortality,
Lex Fridman (27:00.500)
that all of us will kill ourselves
George Hotz (27:03.460)
before we advance far enough
Lex Fridman (27:06.100)
to be able to revert the change.
George Hotz (27:08.820)
I'm not killing myself till I know everything, so.
Lex Fridman (27:11.900)
That's what you say now, because your life is finite.
George Hotz (27:15.020)
You know, I think yes, self modifying systems gets,
Lex Fridman (27:19.060)
comes up with all these hairy complexities
Lex Fridman (27:21.020)
and can I promise that I'll do it perfectly?
Lex Fridman (27:23.020)
No, but I think I can put good safety structures in place.
Lex Fridman (27:27.180)
So that talk and your thinking here
Lex Fridman (27:29.740)
is not literally referring to a simulation
Lex Fridman (27:36.180)
and that our universe is a kind of computer program
Lex Fridman (27:40.520)
running on a computer.
George Hotz (27:42.240)
That's more of a thought experiment.
Lex Fridman (27:45.180)
Do you also think of the potential of the sort of Bostrom,
George Hotz (27:51.780)
Elon Musk and others that talk about an actual program
Lex Fridman (27:57.820)
that simulates our universe?
George Hotz (27:59.700)
Oh, I don't doubt that we're in a simulation.
Lex Fridman (28:01.980)
I just think that it's not quite that important.
George Hotz (28:05.300)
I mean, I'm interested only in simulation theory
Lex Fridman (28:06.940)
as far as like it gives me power over nature.
Lex Fridman (28:09.740)
If it's totally unfalsifiable, then who cares?
Lex Fridman (28:13.140)
I mean, what do you think that experiment would look like?
George Hotz (28:15.300)
Like somebody on Twitter asks,
Lex Fridman (28:17.740)
asks George what signs we would look for
George Hotz (28:20.780)
to know whether or not we're in the simulation,
Lex Fridman (28:22.980)
which is exactly what you're asking is like,
George Hotz (28:25.880)
the step that precedes the step of knowing
Lex Fridman (28:29.500)
how to get more power from this knowledge
George Hotz (28:32.200)
is to get an indication that there's some power to be gained.
Lex Fridman (28:35.220)
So get an indication that there,
George Hotz (28:37.900)
you can discover and exploit cracks in the simulation
Lex Fridman (28:42.100)
or it doesn't have to be in the physics of the universe.
George Hotz (28:45.340)
Yeah.
Lex Fridman (28:46.720)
Show me, I mean, like a memory leak could be cool.
Lex Fridman (28:51.620)
Like some scrying technology, you know?
Lex Fridman (28:54.000)
What kind of technology?
Lex Fridman (28:55.260)
Scrying?
Lex Fridman (28:56.220)
What's that?
George Hotz (28:57.060)
Oh, that's a weird,
Lex Fridman (28:58.460)
scrying is the paranormal ability to like remote viewing,
George Hotz (29:03.460)
like being able to see somewhere where you're not.
Lex Fridman (29:08.220)
So, you know, I don't think you can do it
George Hotz (29:10.100)
by chanting in a room,
Lex Fridman (29:11.220)
but if we could find, it's a memory leak, basically.
George Hotz (29:16.300)
It's a memory leak.
Lex Fridman (29:17.300)
Yeah, you're able to access parts you're not supposed to.
George Hotz (29:19.980)
Yeah, yeah, yeah.
Lex Fridman (29:20.820)
And thereby discover a shortcut.
George Hotz (29:22.100)
Yeah, maybe memory leak means the other thing as well,
Lex Fridman (29:24.720)
but I mean like, yeah,
Lex Fridman (29:25.560)
like an ability to read arbitrary memory, right?
Lex Fridman (29:28.100)
And that one's not that horrifying, right?
George Hotz (29:29.820)
The right ones start to be horrifying.
Lex Fridman (29:31.420)
Read, right.
George Hotz (29:32.260)
It's the reading is not the problem.
Lex Fridman (29:34.900)
Yeah, it's like Heartfleet for the universe.
George Hotz (29:37.220)
Oh boy, the writing is a big, big problem.
Lex Fridman (29:40.740)
It's a big problem.
George Hotz (29:43.060)
It's the moment you can write anything,
Lex Fridman (29:44.620)
even if it's just random noise.
George Hotz (29:47.740)
That's terrifying.
Lex Fridman (29:49.180)
I mean, even without that,
George Hotz (29:51.560)
like even some of the, you know,
Lex Fridman (29:52.580)
the nanotech stuff that's coming, I think is.
George Hotz (29:57.140)
I don't know if you're paying attention,
Lex Fridman (29:58.340)
but actually Eric Weistand came out
George Hotz (2:00:00.320)
taking shots of people.
Lex Fridman (2:00:01.400)
And it's like, you know, Twitter and Donald Trump
George Hotz (2:00:05.280)
are perfectly, they're perfect for each other.
Lex Fridman (2:00:08.440)
So Tesla's on the right track in your view.
George Hotz (2:00:12.800)
OK, so let's try, let's really try this experiment.
Lex Fridman (2:00:16.600)
If you ran Waymo, let's say they're,
George Hotz (2:00:19.720)
I don't know if you agree, but they
Lex Fridman (2:00:21.160)
seem to be at the head of the pack of the kind of,
Lex Fridman (2:00:25.640)
what would you call that approach?
Lex Fridman (2:00:27.040)
Like it's not necessarily lighter based
George Hotz (2:00:29.000)
because it's not about lighter.
Lex Fridman (2:00:30.320)
Level four robotaxi.
George Hotz (2:00:31.520)
Level four robotaxi, all in before making any revenue.
Lex Fridman (2:00:37.040)
So they're probably at the head of the pack.
George Hotz (2:00:38.720)
If you were said, hey, George, can you
Lex Fridman (2:00:42.840)
please run this company for a year, how would you change it?
George Hotz (2:00:47.160)
I would go.
Lex Fridman (2:00:47.760)
I would get Anthony Levandowski out of jail,
Lex Fridman (2:00:49.800)
and I would put him in charge of the company.
Lex Fridman (2:00:56.400)
Well, let's try to break that apart.
Lex Fridman (2:00:58.200)
Why do you want to destroy the company by doing that?
Lex Fridman (2:01:01.600)
Or do you mean you like renegade style thinking that pushes,
George Hotz (2:01:09.400)
that throws away bureaucracy and goes
Lex Fridman (2:01:11.560)
to first principle thinking?
Lex Fridman (2:01:12.840)
What do you mean by that?
Lex Fridman (2:01:14.080)
I think Anthony Levandowski is a genius,
Lex Fridman (2:01:16.040)
and I think he would come up with a much better idea of what
Lex Fridman (2:01:19.640)
to do with Waymo than me.
Lex Fridman (2:01:22.360)
So you mean that unironically.
Lex Fridman (2:01:23.840)
He is a genius.
George Hotz (2:01:24.800)
Oh, yes.
Lex Fridman (2:01:25.400)
Oh, absolutely.
George Hotz (2:01:26.640)
Without a doubt.
Lex Fridman (2:01:27.600)
I mean, I'm not saying there's no shortcomings,
Lex Fridman (2:01:30.960)
but in the interactions I've had with him, yeah.
Lex Fridman (2:01:34.760)
What?
George Hotz (2:01:35.720)
He's also willing to take, like, who knows
Lex Fridman (2:01:38.200)
what he would do with Waymo?
George Hotz (2:01:39.400)
I mean, he's also out there, like far more out there
Lex Fridman (2:01:41.400)
than I am.
George Hotz (2:01:41.880)
Yeah, there's big risks.
Lex Fridman (2:01:43.480)
What do you make of him?
George Hotz (2:01:44.480)
I was going to talk to him on this podcast,
Lex Fridman (2:01:47.040)
and I was going back and forth.
George Hotz (2:01:48.880)
I'm such a gullible, naive human.
Lex Fridman (2:01:51.720)
Like, I see the best in people.
Lex Fridman (2:01:53.840)
And I slowly started to realize that there
Lex Fridman (2:01:56.720)
might be some people out there that, like,
George Hotz (2:02:02.400)
have multiple faces to the world.
Lex Fridman (2:02:05.280)
They're, like, deceiving and dishonest.
George Hotz (2:02:08.160)
I still refuse to, like, I just, I trust people,
Lex Fridman (2:02:13.040)
and I don't care if I get hurt by it.
George Hotz (2:02:14.640)
But, like, you know, sometimes you
Lex Fridman (2:02:16.520)
have to be a little bit careful, especially platform
George Hotz (2:02:18.760)
wise and podcast wise.
Lex Fridman (2:02:21.040)
What do you, what am I supposed to think?
Lex Fridman (2:02:23.160)
So you think, you think he's a good person?
Lex Fridman (2:02:26.400)
Oh, I don't know.
George Hotz (2:02:27.840)
I don't really make moral judgments.
Lex Fridman (2:02:30.000)
It's difficult to.
George Hotz (2:02:30.840)
Oh, I mean this about the Waymo.
Lex Fridman (2:02:32.480)
I actually, I mean that whole idea very nonironically
George Hotz (2:02:34.960)
about what I would do.
Lex Fridman (2:02:36.200)
The problem with putting me in charge of Waymo
Lex Fridman (2:02:38.160)
is Waymo is already $10 billion in the hole, right?
Lex Fridman (2:02:41.720)
Whatever idea Waymo does, look, commas profitable, commas
George Hotz (2:02:44.840)
raised $8.1 million.
Lex Fridman (2:02:46.680)
That's small, you know, that's small money.
George Hotz (2:02:48.180)
Like, I can build a reasonable consumer electronics company
Lex Fridman (2:02:50.840)
and succeed wildly at that and still never be able to pay back
George Hotz (2:02:54.200)
Waymo's $10 billion.
Lex Fridman (2:02:55.800)
So I think the basic idea with Waymo, well,
George Hotz (2:02:58.680)
forget the $10 billion because they have some backing,
Lex Fridman (2:03:00.880)
but your basic thing is, like, what can we do
Lex Fridman (2:03:04.000)
to start making some money?
Lex Fridman (2:03:05.640)
Well, no, I mean, my bigger idea is, like,
George Hotz (2:03:07.920)
whatever the idea is that's gonna save Waymo,
Lex Fridman (2:03:10.320)
I don't have it.
George Hotz (2:03:11.440)
It's gonna have to be a big risk idea
Lex Fridman (2:03:13.480)
and I cannot think of a better person
George Hotz (2:03:15.280)
than Anthony Levandowski to do it.
Lex Fridman (2:03:17.840)
So that is completely what I would do as CEO of Waymo.
George Hotz (2:03:20.240)
I would call myself a transitionary CEO,
Lex Fridman (2:03:22.640)
do everything I can to fix that situation up.
George Hotz (2:03:24.880)
I'm gonna see.
Lex Fridman (2:03:25.720)
Yeah.
George Hotz (2:03:27.920)
Yeah.
Lex Fridman (2:03:28.760)
Because I can't do it, right?
George Hotz (2:03:29.760)
Like, I can't, I mean, I can talk about how
Lex Fridman (2:03:33.440)
what I really wanna do is just apologize
George Hotz (2:03:35.260)
for all those corny, you know, ad campaigns
Lex Fridman (2:03:38.120)
and be like, here's the real state of the technology.
George Hotz (2:03:40.160)
Yeah, that's, like, I have several criticism.
Lex Fridman (2:03:42.280)
I'm a little bit more bullish on Waymo
George Hotz (2:03:44.320)
than you seem to be, but one criticism I have
Lex Fridman (2:03:48.000)
is it went into corny mode too early.
George Hotz (2:03:50.880)
Like, it's still a startup.
Lex Fridman (2:03:52.200)
It hasn't delivered on anything.
Lex Fridman (2:03:53.660)
So it should be, like, more renegade
Lex Fridman (2:03:56.320)
and show off the engineering that they're doing,
George Hotz (2:03:59.280)
which just can be impressive,
Lex Fridman (2:04:00.560)
as opposed to doing these weird commercials
George Hotz (2:04:02.320)
of, like, your friendly car company.
Lex Fridman (2:04:07.040)
I mean, that's my biggest snipe at Waymo is always,
George Hotz (2:04:10.080)
that guy's a paid actor.
Lex Fridman (2:04:11.440)
That guy's not a Waymo user.
George Hotz (2:04:12.800)
He's a paid actor.
Lex Fridman (2:04:13.640)
Look here, I found his call sheet.
George Hotz (2:04:15.360)
Do kind of like what SpaceX is doing
Lex Fridman (2:04:17.400)
with the rocket launches.
George Hotz (2:04:18.960)
Just put the nerds up front, put the engineers up front,
Lex Fridman (2:04:22.200)
and just, like, show failures too, just.
George Hotz (2:04:25.720)
I love SpaceX's, yeah.
Lex Fridman (2:04:27.880)
Yeah, the thing that they're doing is right,
Lex Fridman (2:04:29.840)
and it just feels like the right.
Lex Fridman (2:04:31.720)
But.
George Hotz (2:04:32.560)
We're all so excited to see them succeed.
Lex Fridman (2:04:34.440)
Yeah.
Lex Fridman (2:04:35.260)
I can't wait to see when it won't fail, you know?
Lex Fridman (2:04:37.320)
Like, you lie to me, I want you to fail.
George Hotz (2:04:39.400)
You tell me the truth, you be honest with me,
Lex Fridman (2:04:41.200)
I want you to succeed.
George Hotz (2:04:42.160)
Yeah.
Lex Fridman (2:04:44.120)
Ah, yeah, and that requires the renegade CEO, right?
George Hotz (2:04:50.200)
I'm with you, I'm with you.
Lex Fridman (2:04:51.480)
I still have a little bit of faith in Waymo
George Hotz (2:04:54.000)
for the renegade CEO to step forward, but.
Lex Fridman (2:04:57.860)
It's not, it's not John Kraftik.
George Hotz (2:05:00.600)
Yeah, it's, you can't.
Lex Fridman (2:05:02.980)
It's not Chris Hormiston.
Lex Fridman (2:05:04.980)
And those people may be very good at certain things.
Lex Fridman (2:05:07.680)
Yeah.
Lex Fridman (2:05:08.520)
But they're not renegades.
Lex Fridman (2:05:10.360)
Yeah, because these companies are fundamentally,
George Hotz (2:05:12.040)
even though we're talking about billion dollars,
Lex Fridman (2:05:14.260)
all these crazy numbers,
George Hotz (2:05:15.640)
they're still, like, early stage startups.
Lex Fridman (2:05:19.300)
I mean, and I just, if you are pre revenue
Lex Fridman (2:05:21.840)
and you've raised 10 billion dollars,
Lex Fridman (2:05:23.120)
I have no idea, like, this just doesn't work.
George Hotz (2:05:26.520)
You know, it's against everything Silicon Valley.
Lex Fridman (2:05:28.040)
Where's your minimum viable product?
Lex Fridman (2:05:29.880)
You know, where's your users?
Lex Fridman (2:05:31.600)
Where's your growth numbers?
George Hotz (2:05:33.320)
This is traditional Silicon Valley.
Lex Fridman (2:05:36.280)
Why do you not apply it to what you think
Lex Fridman (2:05:38.040)
you're too big to fail already, like?
Lex Fridman (2:05:41.760)
How do you think autonomous driving will change society?
Lex Fridman (2:05:45.780)
So the mission is, for comma, to solve self driving.
Lex Fridman (2:05:52.520)
Do you have, like, a vision of the world
Lex Fridman (2:05:54.240)
of how it'll be different?
Lex Fridman (2:05:57.940)
Is it as simple as A to B transportation?
George Hotz (2:06:00.100)
Or is there, like, cause these are robots.
Lex Fridman (2:06:03.640)
It's not about autonomous driving in and of itself.
George Hotz (2:06:05.880)
It's what the technology enables.
Lex Fridman (2:06:09.760)
It's, I think it's the coolest applied AI problem.
George Hotz (2:06:12.200)
I like it because it has a clear path to monetary value.
Lex Fridman (2:06:17.720)
But as far as that being the thing that changes the world,
George Hotz (2:06:21.440)
I mean, no, like, there's cute things we're doing in common.
Lex Fridman (2:06:25.480)
Like, who'd have thought you could stick a phone
George Hotz (2:06:26.960)
on the windshield and it'll drive.
Lex Fridman (2:06:29.080)
But like, really, the product that you're building
George Hotz (2:06:31.160)
is not something that people were not capable
Lex Fridman (2:06:33.920)
of imagining 50 years ago.
Lex Fridman (2:06:35.760)
So no, it doesn't change the world on that front.
Lex Fridman (2:06:37.800)
Could people have imagined the internet 50 years ago?
George Hotz (2:06:39.480)
Only true genius visionaries.
Lex Fridman (2:06:42.560)
Everyone could have imagined autonomous cars 50 years ago.
George Hotz (2:06:45.120)
It's like a car, but I don't drive it.
Lex Fridman (2:06:47.000)
See, I have this sense, and I told you, like,
George Hotz (2:06:49.920)
my longterm dream is robots with which you have deep,
Lex Fridman (2:06:55.880)
with whom you have deep connections, right?
Lex Fridman (2:06:59.320)
And there's different trajectories towards that.
Lex Fridman (2:07:03.600)
And I've been thinking,
Lex Fridman (2:07:04.440)
so I've been thinking of launching a startup.
Lex Fridman (2:07:07.060)
I see autonomous vehicles
George Hotz (2:07:09.820)
as a potential trajectory to that.
Lex Fridman (2:07:11.420)
That's not where the direction I would like to go,
Lex Fridman (2:07:16.580)
but I also see Tesla or even Comma AI,
Lex Fridman (2:07:19.100)
like, pivoting into robotics broadly defined
George Hotz (2:07:24.900)
at some stage in the way, like you're mentioning,
Lex Fridman (2:07:27.140)
the internet didn't expect.
George Hotz (2:07:29.580)
Let's solve, you know, when I say a comma about this,
Lex Fridman (2:07:32.580)
we could talk about this,
Lex Fridman (2:07:33.420)
but let's solve self driving cars first.
Lex Fridman (2:07:35.740)
You gotta stay focused on the mission.
George Hotz (2:07:37.140)
Don't, don't, don't, you're not too big to fail.
Lex Fridman (2:07:39.260)
For however much I think Comma's winning,
George Hotz (2:07:41.380)
like, no, no, no, no, no, you're winning
Lex Fridman (2:07:43.420)
when you solve level five self driving cars.
Lex Fridman (2:07:45.020)
And until then, you haven't won.
Lex Fridman (2:07:46.780)
And you know, again, you wanna be arrogant
George Hotz (2:07:48.900)
in the face of other people, great.
Lex Fridman (2:07:50.620)
You wanna be arrogant in the face of nature, you're an idiot.
George Hotz (2:07:53.780)
Stay mission focused, brilliantly put.
Lex Fridman (2:07:56.420)
Like I mentioned, thinking of launching a startup,
George Hotz (2:07:58.660)
I've been considering, actually, before COVID,
Lex Fridman (2:08:01.220)
I've been thinking of moving to San Francisco.
George Hotz (2:08:03.340)
Ooh, ooh, I wouldn't go there.
Lex Fridman (2:08:06.020)
So why is, well, and now I'm thinking
George Hotz (2:08:09.020)
about potentially Austin and we're in San Diego now.
Lex Fridman (2:08:13.660)
San Diego, come here.
Lex Fridman (2:08:14.940)
So why, what, I mean, you're such an interesting human.
Lex Fridman (2:08:20.540)
You've launched so many successful things.
Lex Fridman (2:08:23.060)
What, why San Diego?
Lex Fridman (2:08:26.220)
What do you recommend?
Lex Fridman (2:08:27.060)
Why not San Francisco?
Lex Fridman (2:08:29.300)
Have you thought, so in your case,
George Hotz (2:08:31.820)
San Diego with Qualcomm and Snapdragon,
Lex Fridman (2:08:33.780)
I mean, that's an amazing combination.
George Hotz (2:08:36.860)
But.
Lex Fridman (2:08:37.700)
That wasn't really why.
Lex Fridman (2:08:38.540)
That wasn't the why?
Lex Fridman (2:08:39.500)
No, I mean, Qualcomm was an afterthought.
George Hotz (2:08:41.340)
Qualcomm was, it was a nice thing to think about.
Lex Fridman (2:08:42.900)
It's like, you can have a tech company here.
George Hotz (2:08:45.140)
Yeah.
Lex Fridman (2:08:45.980)
And a good one, I mean, you know, I like Qualcomm, but.
George Hotz (2:08:48.300)
No.
Lex Fridman (2:08:49.140)
Well, so why San Diego better than San Francisco?
Lex Fridman (2:08:50.540)
Why does San Francisco suck?
Lex Fridman (2:08:51.900)
Well, so, okay, so first off,
George Hotz (2:08:53.020)
we all kind of said like, we wanna stay in California.
Lex Fridman (2:08:55.300)
People like the ocean.
George Hotz (2:08:57.620)
You know, California, for its flaws,
Lex Fridman (2:09:00.020)
it's like a lot of the flaws of California
George Hotz (2:09:02.380)
are not necessarily California as a whole,
Lex Fridman (2:09:03.900)
and they're much more San Francisco specific.
George Hotz (2:09:05.740)
Yeah.
Lex Fridman (2:09:06.700)
San Francisco, so I think first tier cities in general
George Hotz (2:09:09.860)
have stopped wanting growth.
Lex Fridman (2:09:13.140)
Well, you have like in San Francisco, you know,
George Hotz (2:09:15.460)
the voting class always votes to not build more houses
Lex Fridman (2:09:18.740)
because they own all the houses.
Lex Fridman (2:09:19.980)
And they're like, well, you know,
Lex Fridman (2:09:21.820)
once people have figured out how to vote themselves
George Hotz (2:09:23.860)
more money, they're gonna do it.
Lex Fridman (2:09:25.180)
It is so insanely corrupt.
George Hotz (2:09:27.900)
It is not balanced at all, like political party wise,
Lex Fridman (2:09:31.620)
you know, it's a one party city and.
George Hotz (2:09:34.700)
For all the discussion of diversity,
Lex Fridman (2:09:38.700)
it stops lacking real diversity of thought,
George Hotz (2:09:42.140)
of background, of approaches, of strategies, of ideas.
Lex Fridman (2:09:48.540)
It's kind of a strange place
George Hotz (2:09:51.180)
that it's the loudest people about diversity
Lex Fridman (2:09:54.380)
and the biggest lack of diversity.
Lex Fridman (2:09:56.420)
I mean, that's what they say, right?
Lex Fridman (2:09:58.620)
It's the projection.
George Hotz (2:10:00.340)
Projection, yeah.
Lex Fridman (2:10:02.140)
Yeah, it's interesting.
Lex Fridman (2:10:02.980)
And even people in Silicon Valley tell me
Lex Fridman (2:10:04.500)
that's like high up people,
George Hotz (2:10:07.940)
everybody is like, this is a terrible place.
Lex Fridman (2:10:10.100)
It doesn't make sense.
George Hotz (2:10:10.940)
I mean, and coronavirus is really what killed it.
Lex Fridman (2:10:13.220)
San Francisco was the number one exodus
George Hotz (2:10:17.340)
during coronavirus.
Lex Fridman (2:10:18.900)
We still think San Diego is a good place to be.
George Hotz (2:10:21.660)
Yeah.
Lex Fridman (2:10:23.460)
Yeah, I mean, we'll see.
George Hotz (2:10:24.740)
We'll see what happens with California a bit longer term.
Lex Fridman (2:10:29.780)
Like Austin's an interesting choice.
George Hotz (2:10:32.100)
I wouldn't, I don't have really anything bad to say
Lex Fridman (2:10:33.940)
about Austin either,
George Hotz (2:10:35.740)
except for the extreme heat in the summer,
Lex Fridman (2:10:37.940)
which, but that's like very on the surface, right?
George Hotz (2:10:40.180)
I think as far as like an ecosystem goes, it's cool.
Lex Fridman (2:10:43.700)
I personally love Colorado.
George Hotz (2:10:45.540)
Colorado's great.
Lex Fridman (2:10:47.100)
Yeah, I mean, you have these states that are,
George Hotz (2:10:49.140)
like just way better run.
Lex Fridman (2:10:51.380)
California is, you know, it's especially San Francisco.
George Hotz (2:10:55.380)
It's not a tie horse and like, yeah.
Lex Fridman (2:10:58.460)
Can I ask you for advice to me and to others
Lex Fridman (2:11:02.620)
about what's it take to build a successful startup?
Lex Fridman (2:11:07.540)
Oh, I don't know.
George Hotz (2:11:08.380)
I haven't done that.
Lex Fridman (2:11:09.220)
Talk to someone who did that.
George Hotz (2:11:10.740)
Well, you've, you know,
Lex Fridman (2:11:14.980)
this is like another book of years
George Hotz (2:11:16.460)
that I'll buy for $67, I suppose.
Lex Fridman (2:11:18.860)
So there's, um.
George Hotz (2:11:20.700)
One of these days I'll sell out.
Lex Fridman (2:11:24.140)
Yeah, that's right.
George Hotz (2:11:24.980)
Jailbreaks are going to be a dollar
Lex Fridman (2:11:26.020)
and books are going to be 67.
Lex Fridman (2:11:27.860)
How I jailbroke the iPhone by George Hots.
Lex Fridman (2:11:32.060)
That's right.
Lex Fridman (2:11:32.900)
How I jail broke the iPhone and you can too.
Lex Fridman (2:11:35.620)
You can too.
George Hotz (2:11:36.860)
67 dollars.
Lex Fridman (2:11:37.700)
In 21 days.
George Hotz (2:11:39.020)
That's right.
Lex Fridman (2:11:39.860)
That's right.
George Hotz (2:11:40.700)
Oh God.
Lex Fridman (2:11:41.540)
Okay, I can't wait.
Lex Fridman (2:11:42.380)
But quite, so you have an introspective,
Lex Fridman (2:11:44.940)
you have built a very unique company.
George Hotz (2:11:49.460)
I mean, not you, but you and others.
Lex Fridman (2:11:53.180)
But I don't know.
George Hotz (2:11:55.620)
There's no, there's nothing.
Lex Fridman (2:11:56.940)
You have an introspective,
George Hotz (2:11:57.780)
you haven't really sat down and thought about like,
Lex Fridman (2:12:01.780)
well, like if you and I were having a bunch of,
George Hotz (2:12:04.180)
we're having some beers
Lex Fridman (2:12:06.180)
and you're seeing that I'm depressed
Lex Fridman (2:12:08.100)
and whatever, I'm struggling.
Lex Fridman (2:12:09.860)
There's no advice you can give?
George Hotz (2:12:11.580)
Oh, I mean.
Lex Fridman (2:12:13.100)
More beer?
Lex Fridman (2:12:13.940)
More beer?
Lex Fridman (2:12:15.100)
Um, yeah, I think it's all very like situation dependent.
George Hotz (2:12:23.340)
Here's, okay, if I can give a generic piece of advice,
Lex Fridman (2:12:25.860)
it's the technology always wins.
George Hotz (2:12:28.140)
The better technology always wins.
Lex Fridman (2:12:30.740)
And lying always loses.
George Hotz (2:12:35.180)
Build technology and don't lie.
Lex Fridman (2:12:38.580)
I'm with you.
George Hotz (2:12:39.420)
I agree very much.
Lex Fridman (2:12:40.420)
The long run, long run.
George Hotz (2:12:41.340)
Sure.
Lex Fridman (2:12:42.180)
That's the long run, yeah.
George Hotz (2:12:43.020)
The market can remain irrational longer
Lex Fridman (2:12:44.860)
than you can remain solvent.
George Hotz (2:12:46.540)
True fact.
Lex Fridman (2:12:47.380)
Well, this is an interesting point
George Hotz (2:12:49.260)
because I ethically and just as a human believe that
Lex Fridman (2:12:54.380)
like hype and smoke and mirrors is not
George Hotz (2:12:58.980)
at any stage of the company is a good strategy.
Lex Fridman (2:13:02.740)
I mean, there's some like, you know,
George Hotz (2:13:04.900)
PR magic kind of like, you know.
Lex Fridman (2:13:07.020)
Oh, hype around a new product, right?
George Hotz (2:13:08.540)
If there's a call to action,
Lex Fridman (2:13:09.780)
if there's like a call to action,
George Hotz (2:13:10.980)
like buy my new GPU, look at it.
Lex Fridman (2:13:13.060)
It takes up three slots and it's this big.
George Hotz (2:13:14.940)
It's huge.
Lex Fridman (2:13:15.780)
Buy my GPU.
George Hotz (2:13:16.620)
Yeah, that's great.
Lex Fridman (2:13:17.460)
If you look at, you know,
George Hotz (2:13:18.300)
especially in the AI space broadly,
Lex Fridman (2:13:20.780)
but autonomous vehicles,
George Hotz (2:13:22.780)
like you can raise a huge amount of money on nothing.
Lex Fridman (2:13:26.540)
And the question to me is like, I'm against that.
George Hotz (2:13:30.060)
I'll never be part of that.
Lex Fridman (2:13:31.500)
I don't think, I hope not, willingly not.
Lex Fridman (2:13:36.820)
But like, is there something to be said
Lex Fridman (2:13:40.020)
to essentially lying to raise money,
Lex Fridman (2:13:44.980)
like fake it till you make it kind of thing?
Lex Fridman (2:13:47.860)
I mean, this is Billy McFarland in the Fyre Festival.
George Hotz (2:13:50.140)
Like we all experienced, you know,
Lex Fridman (2:13:53.460)
what happens with that.
George Hotz (2:13:54.300)
No, no, don't fake it till you make it.
Lex Fridman (2:13:57.460)
Be honest and hope you make it the whole way.
George Hotz (2:14:00.540)
The technology wins.
Lex Fridman (2:14:01.940)
Right, the technology wins.
Lex Fridman (2:14:02.980)
And like, there is, I'm not used to like the anti hype,
Lex Fridman (2:14:06.820)
you know, that's a Slava KPSS reference,
Lex Fridman (2:14:08.900)
but hype isn't necessarily bad.
Lex Fridman (2:14:13.140)
I loved camping out for the iPhones, you know,
Lex Fridman (2:14:17.580)
and as long as the hype is backed by like substance,
Lex Fridman (2:14:21.020)
as long as it's backed by something I can actually buy,
Lex Fridman (2:14:23.780)
and like it's real, then hype is great
Lex Fridman (2:14:26.740)
and it's a great feeling.
George Hotz (2:14:28.700)
It's when the hype is backed by lies
Lex Fridman (2:14:30.860)
that it's a bad feeling.
George Hotz (2:14:32.060)
I mean, a lot of people call Elon Musk a fraud.
Lex Fridman (2:14:34.540)
How could he be a fraud?
George Hotz (2:14:35.620)
I've noticed this, this kind of interesting effect,
Lex Fridman (2:14:37.740)
which is he does tend to over promise
Lex Fridman (2:14:42.300)
and deliver, what's the better way to phrase it?
Lex Fridman (2:14:45.660)
Promise a timeline that he doesn't deliver on,
George Hotz (2:14:49.220)
he delivers much later on.
Lex Fridman (2:14:51.740)
What do you think about that?
George Hotz (2:14:52.820)
Cause I do that, I think that's a programmer thing too.
Lex Fridman (2:14:56.220)
I do that as well.
Lex Fridman (2:14:57.540)
You think that's a really bad thing to do or is that okay?
Lex Fridman (2:15:01.300)
I think that's, again, as long as like,
George Hotz (2:15:03.900)
you're working toward it and you're gonna deliver on it,
Lex Fridman (2:15:06.940)
it's not too far off, right?
Lex Fridman (2:15:10.260)
Right?
Lex Fridman (2:15:11.100)
Like, you know, the whole autonomous vehicle thing,
George Hotz (2:15:14.900)
it's like, I mean, I still think Tesla's on track
Lex Fridman (2:15:18.620)
to beat us.
George Hotz (2:15:19.740)
I still think even with their missteps,
Lex Fridman (2:15:21.900)
they have advantages we don't have.
George Hotz (2:15:25.660)
You know, Elon is better than me
Lex Fridman (2:15:28.100)
at like marshaling massive amounts of resources.
George Hotz (2:15:33.060)
So, you know, I still think given the fact
Lex Fridman (2:15:36.260)
they're maybe making some wrong decisions,
George Hotz (2:15:38.220)
they'll end up winning.
Lex Fridman (2:15:39.300)
And like, it's fine to hype it
Lex Fridman (2:15:42.620)
if you're actually gonna win, right?
Lex Fridman (2:15:44.900)
Like if Elon says, look, we're gonna be landing rockets
George Hotz (2:15:47.220)
back on earth in a year and it takes four,
Lex Fridman (2:15:49.340)
like, you know, he landed a rocket back on earth
Lex Fridman (2:15:53.540)
and he was working toward it the whole time.
Lex Fridman (2:15:55.420)
I think there's some amount of like,
George Hotz (2:15:57.060)
I think when it becomes wrong is if you know
Lex Fridman (2:15:59.260)
you're not gonna meet that deadline.
George Hotz (2:16:00.460)
If you're lying.
Lex Fridman (2:16:01.540)
Yeah, that's brilliantly put.
George Hotz (2:16:03.260)
Like this is what people don't understand, I think.
Lex Fridman (2:16:06.860)
Like Elon believes everything he says.
George Hotz (2:16:09.500)
He does, as far as I can tell, he does.
Lex Fridman (2:16:12.140)
And I detected that in myself too.
George Hotz (2:16:14.780)
Like if I, it's only bullshit
Lex Fridman (2:16:17.620)
if you're like conscious of yourself lying.
George Hotz (2:16:21.460)
Yeah, I think so.
Lex Fridman (2:16:22.500)
Yeah.
Lex Fridman (2:16:23.420)
Now you can't take that to such an extreme, right?
Lex Fridman (2:16:25.660)
Like in a way, I think maybe Billy McFarland
George Hotz (2:16:27.620)
believed everything he said too.
Lex Fridman (2:16:30.020)
Right, that's how you start a cult
Lex Fridman (2:16:31.460)
and everybody kills themselves.
Lex Fridman (2:16:33.820)
Yeah.
George Hotz (2:16:34.660)
Yeah, like it's, you need, you need,
Lex Fridman (2:16:36.020)
if there's like some factor on it, it's fine.
Lex Fridman (2:16:39.300)
And you need some people to like, you know,
Lex Fridman (2:16:41.820)
keep you in check, but like,
George Hotz (2:16:44.580)
if you deliver on most of the things you say
Lex Fridman (2:16:46.580)
and just the timelines are off, yeah.
George Hotz (2:16:48.780)
It does piss people off though.
Lex Fridman (2:16:50.300)
I wonder, but who cares?
George Hotz (2:16:53.300)
In a long arc of history, the people,
Lex Fridman (2:16:55.260)
everybody gets pissed off at the people who succeed,
George Hotz (2:16:58.100)
which is one of the things
Lex Fridman (2:16:59.260)
that frustrates me about this world,
George Hotz (2:17:01.820)
is they don't celebrate the success of others.
Lex Fridman (2:17:07.660)
Like there's so many people that want Elon to fail.
George Hotz (2:17:12.900)
It's so fascinating to me.
Lex Fridman (2:17:14.860)
Like what is wrong with you?
George Hotz (2:17:18.180)
Like, so Elon Musk talks about like people shorting,
Lex Fridman (2:17:21.620)
like they talk about financial,
Lex Fridman (2:17:23.580)
but I think it's much bigger than the financials.
Lex Fridman (2:17:25.340)
I've seen like the human factors community,
George Hotz (2:17:27.860)
they want, they want other people to fail.
Lex Fridman (2:17:31.540)
Why, why, why?
George Hotz (2:17:32.940)
Like even people, the harshest thing is like,
Lex Fridman (2:17:36.740)
you know, even people that like seem
George Hotz (2:17:38.140)
to really hate Donald Trump, they want him to fail
Lex Fridman (2:17:41.980)
or like the other president
George Hotz (2:17:43.020)
or they want Barack Obama to fail.
Lex Fridman (2:17:45.900)
It's like.
George Hotz (2:17:47.140)
Yeah, we're all on the same boat, man.
Lex Fridman (2:17:49.740)
It's weird, but I want that,
George Hotz (2:17:51.660)
I would love to inspire that part of the world to change
Lex Fridman (2:17:54.100)
because damn it, if the human species is gonna survive,
George Hotz (2:17:58.540)
we should celebrate success.
Lex Fridman (2:18:00.460)
Like it seems like the efficient thing to do
George Hotz (2:18:02.780)
in this objective function that we're all striving for
Lex Fridman (2:18:06.300)
is to celebrate the ones that like figure out
Lex Fridman (2:18:09.060)
how to like do better at that objective function
Lex Fridman (2:18:11.820)
as opposed to like dragging them down back into the mud.
George Hotz (2:18:16.580)
I think there is, this is the speech I always give
Lex Fridman (2:18:19.100)
about the commenters on Hacker News.
Lex Fridman (2:18:21.700)
So first off, something to remember
Lex Fridman (2:18:23.260)
about the internet in general is commenters
George Hotz (2:18:26.140)
are not representative of the population.
Lex Fridman (2:18:29.300)
I don't comment on anything.
George Hotz (2:18:31.460)
You know, commenters are representative
Lex Fridman (2:18:34.180)
of a certain sliver of the population.
Lex Fridman (2:18:36.660)
And on Hacker News, a common thing I'll see
Lex Fridman (2:18:39.540)
is when you'll see something that's like,
George Hotz (2:18:42.100)
you know, promises to be wild out there and innovative.
Lex Fridman (2:18:47.500)
There is some amount of, you know,
George Hotz (2:18:49.700)
checking them back to earth,
Lex Fridman (2:18:50.940)
but there's also some amount of if this thing succeeds,
George Hotz (2:18:55.260)
well, I'm 36 and I've worked
Lex Fridman (2:18:57.580)
at large tech companies my whole life.
George Hotz (2:19:02.300)
They can't succeed because if they succeed,
Lex Fridman (2:19:05.020)
that would mean that I could have done something different
George Hotz (2:19:07.180)
with my life, but we know that I couldn't have,
Lex Fridman (2:19:09.220)
we know that I couldn't have,
Lex Fridman (2:19:10.140)
and that's why they're gonna fail.
Lex Fridman (2:19:11.940)
And they have to root for them to fail
George Hotz (2:19:13.100)
to kind of maintain their world image.
Lex Fridman (2:19:15.940)
So tune it out.
Lex Fridman (2:19:17.860)
And they comment, well, it's hard, I, so one of the things,
Lex Fridman (2:19:21.860)
one of the things I'm considering startup wise
George Hotz (2:19:25.260)
is to change that.
Lex Fridman (2:19:27.620)
Cause I think the, I think it's also a technology problem.
George Hotz (2:19:31.700)
It's a platform problem.
Lex Fridman (2:19:33.060)
I agree.
George Hotz (2:19:33.900)
It's like, because the thing you said,
Lex Fridman (2:19:35.980)
most people don't comment.
George Hotz (2:19:39.660)
I think most people want to comment.
Lex Fridman (2:19:42.940)
They just don't because it's all the assholes
George Hotz (2:19:45.180)
who are commenting.
Lex Fridman (2:19:46.020)
Exactly, I don't want to be grouped in with them.
George Hotz (2:19:47.620)
You don't want to be at a party
Lex Fridman (2:19:49.420)
where everyone is an asshole.
Lex Fridman (2:19:50.780)
And so they, but that's a platform problem.
Lex Fridman (2:19:54.500)
I can't believe what Reddit's become.
George Hotz (2:19:56.100)
I can't believe the group thinking, Reddit comments.
Lex Fridman (2:20:00.580)
There's a, Reddit is an interesting one
George Hotz (2:20:02.700)
because they're subreddits.
Lex Fridman (2:20:05.100)
And so you can still see, especially small subreddits
George Hotz (2:20:09.260)
that like, that are a little like havens
Lex Fridman (2:20:11.940)
of like joy and positivity and like deep,
George Hotz (2:20:16.140)
even disagreement, but like nuanced discussion.
Lex Fridman (2:20:18.860)
But it's only like small little pockets,
Lex Fridman (2:20:21.700)
but that's emergent.
Lex Fridman (2:20:23.460)
The platform is not helping that or hurting that.
Lex Fridman (2:20:26.780)
So I guess naturally something about the internet,
Lex Fridman (2:20:31.100)
if you don't put in a lot of effort to encourage
George Hotz (2:20:34.380)
nuance and positive, good vibes,
Lex Fridman (2:20:37.180)
it's naturally going to decline into chaos.
George Hotz (2:20:41.140)
I would love to see someone do this well.
Lex Fridman (2:20:42.860)
Yeah.
George Hotz (2:20:43.780)
I think it's, yeah, very doable.
Lex Fridman (2:20:45.540)
I think actually, so I feel like Twitter
George Hotz (2:20:49.740)
could be overthrown.
Lex Fridman (2:20:52.060)
Yashua Bach talked about how like,
George Hotz (2:20:55.220)
if you have like and retweet,
Lex Fridman (2:20:58.020)
like that's only positive wiring, right?
George Hotz (2:21:02.120)
The only way to do anything like negative there
Lex Fridman (2:21:05.580)
is with a comment.
Lex Fridman (2:21:08.300)
And that's like that asymmetry is what gives,
Lex Fridman (2:21:12.420)
you know, Twitter its particular toxicness.
George Hotz (2:21:15.620)
Whereas I find YouTube comments to be much better
Lex Fridman (2:21:18.060)
because YouTube comments have an up and a down
Lex Fridman (2:21:21.540)
and they don't show the downvotes.
Lex Fridman (2:21:23.500)
Without getting into depth of this particular discussion,
George Hotz (2:21:26.820)
the point is to explore possibilities
Lex Fridman (2:21:29.580)
and get a lot of data on it.
George Hotz (2:21:30.860)
Because I mean, I could disagree with what you just said.
Lex Fridman (2:21:34.780)
The point is it's unclear.
George Hotz (2:21:36.100)
It hasn't been explored in a really rich way.
Lex Fridman (2:21:39.860)
Like these questions of how to create platforms
George Hotz (2:21:44.700)
that encourage positivity.
Lex Fridman (2:21:47.100)
Yeah, I think it's a technology problem.
Lex Fridman (2:21:49.580)
And I think we'll look back at Twitter as it is now.
Lex Fridman (2:21:51.980)
Maybe it'll happen within Twitter,
Lex Fridman (2:21:53.900)
but most likely somebody overthrows them
Lex Fridman (2:21:56.460)
is we'll look back at Twitter and say,
George Hotz (2:22:00.500)
can't believe we put up with this level of toxicity.
Lex Fridman (2:22:03.220)
You need a different business model too.
George Hotz (2:22:05.100)
Any social network that fundamentally has advertising
Lex Fridman (2:22:07.700)
as a business model, this was in The Social Dilemma,
George Hotz (2:22:10.460)
which I didn't watch, but I liked it.
Lex Fridman (2:22:11.740)
It's like, you know, there's always the, you know,
George Hotz (2:22:12.940)
you're the product, you're not the,
Lex Fridman (2:22:15.380)
but they had a nuanced take on it that I really liked.
Lex Fridman (2:22:17.980)
And it said, the product being sold is influence over you.
Lex Fridman (2:22:24.660)
The product being sold is literally your,
George Hotz (2:22:27.140)
you know, influence on you.
Lex Fridman (2:22:29.820)
Like that can't be, if that's your idea, okay.
Lex Fridman (2:22:33.660)
Well, you know, guess what?
Lex Fridman (2:22:35.420)
It can't not be toxic.
George Hotz (2:22:37.060)
Yeah, maybe there's ways to spin it,
Lex Fridman (2:22:39.420)
like with giving a lot more control to the user
Lex Fridman (2:22:42.540)
and transparency to see what is happening to them
Lex Fridman (2:22:44.660)
as opposed to in the shadows, it's possible,
Lex Fridman (2:22:47.300)
but that can't be the primary source of.
Lex Fridman (2:22:49.380)
But the users aren't, no one's gonna use that.
George Hotz (2:22:51.980)
It depends, it depends, it depends.
Lex Fridman (2:22:54.020)
I think that the, you're not going to,
George Hotz (2:22:57.140)
you can't depend on self awareness of the users.
Lex Fridman (2:23:00.220)
It's a longer discussion because you can't depend on it,
Lex Fridman (2:23:04.460)
but you can reward self awareness.
Lex Fridman (2:23:09.740)
Like if for the ones who are willing to put in the work
George Hotz (2:23:12.340)
of self awareness, you can reward them and incentivize
Lex Fridman (2:23:16.060)
and perhaps be pleasantly surprised how many people
George Hotz (2:23:20.180)
are willing to be self aware on the internet.
Lex Fridman (2:23:23.260)
Like we are in real life.
George Hotz (2:23:24.380)
Like I'm putting in a lot of effort with you right now,
Lex Fridman (2:23:26.940)
being self aware about if I say something stupid or mean,
George Hotz (2:23:30.300)
I'll like look at your like body language.
Lex Fridman (2:23:32.740)
Like I'm putting in that effort.
George Hotz (2:23:33.700)
It's costly for an introvert, very costly.
Lex Fridman (2:23:36.260)
But on the internet, fuck it.
George Hotz (2:23:39.620)
Like most people are like, I don't care if this hurts
Lex Fridman (2:23:42.940)
somebody, I don't care if this is not interesting
George Hotz (2:23:46.220)
or if this is, yeah, it's a mean or whatever.
Lex Fridman (2:23:48.660)
I think so much of the engagement today on the internet
George Hotz (2:23:50.980)
is so disingenuine too.
Lex Fridman (2:23:53.060)
You're not doing this out of a genuine,
George Hotz (2:23:54.620)
this is what you think.
Lex Fridman (2:23:55.500)
You're doing this just straight up to manipulate others.
George Hotz (2:23:57.660)
Whether you're in, you just became an ad.
Lex Fridman (2:23:59.660)
Yeah, okay, let's talk about a fun topic,
George Hotz (2:24:02.780)
which is programming.
Lex Fridman (2:24:04.420)
Here's another book idea for you.
George Hotz (2:24:05.780)
Let me pitch.
Lex Fridman (2:24:07.300)
What's your perfect programming setup?
Lex Fridman (2:24:09.700)
So like this by George Hots.
Lex Fridman (2:24:12.660)
So like what, listen, you're.
George Hotz (2:24:17.500)
Give me a MacBook Air, sit me in a corner of a hotel room
Lex Fridman (2:24:20.340)
and you know I'll still ask you.
Lex Fridman (2:24:21.180)
So you really don't care.
Lex Fridman (2:24:22.020)
You don't fetishize like multiple monitors, keyboard.
George Hotz (2:24:27.020)
Those things are nice and I'm not gonna say no to them,
Lex Fridman (2:24:30.260)
but did they automatically unlock tons of productivity?
George Hotz (2:24:33.300)
No, not at all.
Lex Fridman (2:24:34.140)
I have definitely been more productive on a MacBook Air
George Hotz (2:24:36.300)
in a corner of a hotel room.
Lex Fridman (2:24:38.300)
What about IDE?
Lex Fridman (2:24:41.620)
So which operating system do you love?
Lex Fridman (2:24:45.980)
What text editor do you use IDE?
George Hotz (2:24:49.020)
What, is there something that is like the perfect,
Lex Fridman (2:24:53.020)
if you could just say the perfect productivity setup
George Hotz (2:24:57.020)
for George Hots.
Lex Fridman (2:24:57.860)
It doesn't matter.
George Hotz (2:24:58.700)
It literally doesn't matter.
Lex Fridman (2:25:00.420)
You know, I guess I code most of the time in Vim.
George Hotz (2:25:03.100)
Like literally I'm using an editor from the 70s.
Lex Fridman (2:25:05.020)
You know, you didn't make anything better.
George Hotz (2:25:07.460)
Okay, VS code is nice for reading code.
Lex Fridman (2:25:09.060)
There's a few things that are nice about it.
George Hotz (2:25:10.820)
I think that you can build much better tools.
Lex Fridman (2:25:13.460)
How like IDA's xrefs work way better than VS codes, why?
Lex Fridman (2:25:18.660)
Yeah, actually that's a good question, like why?
Lex Fridman (2:25:20.760)
I still use, sorry, Emacs for most.
George Hotz (2:25:25.320)
I've actually never, I have to confess something dark.
Lex Fridman (2:25:28.860)
So I've never used Vim.
George Hotz (2:25:32.620)
I think maybe I'm just afraid
Lex Fridman (2:25:36.460)
that my life has been like a waste.
George Hotz (2:25:39.340)
I'm so, I'm not evangelical about Emacs.
Lex Fridman (2:25:43.500)
I think this.
George Hotz (2:25:44.620)
This is how I feel about TensorFlow versus PyTorch.
Lex Fridman (2:25:47.820)
Having just like, we've switched everything to PyTorch now.
George Hotz (2:25:50.300)
Put months into the switch.
Lex Fridman (2:25:51.900)
I have felt like I've wasted years on TensorFlow.
George Hotz (2:25:54.580)
I can't believe it.
Lex Fridman (2:25:56.300)
I can't believe how much better PyTorch is.
George Hotz (2:25:58.380)
Yeah.
Lex Fridman (2:25:59.620)
I've used Emacs and Vim, doesn't matter.
George Hotz (2:26:01.500)
Yeah, it's still just my heart.
Lex Fridman (2:26:03.060)
Somehow I fell in love with Lisp.
George Hotz (2:26:04.700)
I don't know why.
Lex Fridman (2:26:05.540)
You can't, the heart wants what the heart wants.
George Hotz (2:26:08.140)
I don't understand it, but it just connected with me.
Lex Fridman (2:26:10.480)
Maybe it's the functional language
George Hotz (2:26:11.700)
that first I connected with.
Lex Fridman (2:26:13.260)
Maybe it's because so many of the AI courses
George Hotz (2:26:15.860)
before the deep learning revolution
Lex Fridman (2:26:17.300)
were taught with Lisp in mind.
George Hotz (2:26:19.500)
I don't know.
Lex Fridman (2:26:20.340)
I don't know what it is, but I'm stuck with it.
Lex Fridman (2:26:22.340)
But at the same time, like,
Lex Fridman (2:26:23.540)
why am I not using a modern ID
Lex Fridman (2:26:25.100)
for some of these programming?
Lex Fridman (2:26:26.300)
I don't know.
George Hotz (2:26:27.260)
They're not that much better.
Lex Fridman (2:26:28.500)
I've used modern IDs too.
Lex Fridman (2:26:30.180)
But at the same time, so to just,
Lex Fridman (2:26:32.040)
well, not to disagree with you,
Lex Fridman (2:26:33.060)
but like, I like multiple monitors.
Lex Fridman (2:26:35.880)
Like I have to do work on a laptop
Lex Fridman (2:26:38.380)
and it's a pain in the ass.
Lex Fridman (2:26:41.340)
And also I'm addicted to the Kinesis weird keyboard.
George Hotz (2:26:45.420)
You could see there.
Lex Fridman (2:26:46.740)
Yeah, yeah, yeah.
George Hotz (2:26:48.540)
Yeah, so you don't have any of that.
Lex Fridman (2:26:50.100)
You can just be on a MacBook.
George Hotz (2:26:51.820)
I mean, look at work.
Lex Fridman (2:26:53.020)
I have three 24 inch monitors.
George Hotz (2:26:55.220)
I have a happy hacking keyboard.
Lex Fridman (2:26:56.580)
I have a Razer Death Hatter mouse, like.
Lex Fridman (2:26:59.820)
But it's not essential for you.
Lex Fridman (2:27:01.300)
No.
George Hotz (2:27:02.120)
Let's go to a day in the life of George Hots.
Lex Fridman (2:27:04.680)
What is the perfect day productivity wise?
Lex Fridman (2:27:08.780)
So we're not talking about like Hunter S. Thompson drugs.
Lex Fridman (2:27:12.380)
Yeah, yeah, yeah.
Lex Fridman (2:27:13.220)
And let's look at productivity.
Lex Fridman (2:27:16.420)
Like what's the day look like, like hour by hour?
George Hotz (2:27:19.820)
Is there any regularities that create
Lex Fridman (2:27:23.260)
a magical George Hots experience?
George Hotz (2:27:25.940)
I can remember three days in my life.
Lex Fridman (2:27:28.300)
And I remember these days vividly
George Hotz (2:27:30.300)
when I've gone through kind of radical transformations
Lex Fridman (2:27:36.620)
to the way I think.
Lex Fridman (2:27:37.900)
And what I would give, I would pay $100,000
Lex Fridman (2:27:40.140)
if I could have one of these days tomorrow.
George Hotz (2:27:42.980)
The days have been so impactful.
Lex Fridman (2:27:44.900)
And one was first discovering Eliezer Yudkowsky
George Hotz (2:27:47.780)
on the singularity and reading that stuff.
Lex Fridman (2:27:50.740)
And like, you know, my mind was blown.
George Hotz (2:27:54.260)
The next was discovering the Hutter Prize
Lex Fridman (2:27:57.820)
and that AI is just compression.
George Hotz (2:27:59.900)
Like finally understanding AIXI and what all of that was.
Lex Fridman (2:28:03.820)
You know, I like read about it when I was 18, 19,
George Hotz (2:28:05.300)
I didn't understand it.
Lex Fridman (2:28:06.300)
And then the fact that like lossless compression
George Hotz (2:28:08.380)
implies intelligence, the day that I was shown that.
Lex Fridman (2:28:12.240)
And then the third one is controversial.
George Hotz (2:28:14.140)
The day I found a blog called Unqualified Reservations.
Lex Fridman (2:28:17.340)
And read that and I was like.
Lex Fridman (2:28:20.620)
Wait, which one is that?
Lex Fridman (2:28:21.500)
That's, what's the guy's name?
George Hotz (2:28:22.980)
Curtis Yarvin.
Lex Fridman (2:28:24.020)
Yeah.
Lex Fridman (2:28:25.020)
So many people tell me I'm supposed to talk to him.
Lex Fridman (2:28:27.640)
Yeah, the day.
George Hotz (2:28:28.480)
He looks, he sounds insane.
Lex Fridman (2:28:30.140)
Definitely. Or brilliant,
Lex Fridman (2:28:31.300)
but insane or both, I don't know.
Lex Fridman (2:28:33.100)
The day I found that blog was another like,
George Hotz (2:28:35.140)
this was during like Gamergate
Lex Fridman (2:28:37.180)
and kind of the run up to the 2016 election.
Lex Fridman (2:28:39.020)
And I'm like, wow, okay, the world makes sense now.
Lex Fridman (2:28:42.860)
This is like, I had a framework now to interpret this.
George Hotz (2:28:45.620)
Just like I got the framework for AI
Lex Fridman (2:28:47.200)
and a framework to interpret technological progress.
George Hotz (2:28:49.420)
Like those days when I discovered these new frameworks were.
Lex Fridman (2:28:52.820)
Oh, interesting.
Lex Fridman (2:28:53.660)
So it's not about, but what was special about those days?
Lex Fridman (2:28:57.160)
How did those days come to be?
Lex Fridman (2:28:58.780)
Is it just, you got lucky?
Lex Fridman (2:28:59.980)
Like, you just encountered a hotter prize
Lex Fridman (2:29:04.900)
on Hacker News or something like that?
Lex Fridman (2:29:09.740)
But you see, I don't think it's just,
George Hotz (2:29:11.840)
see, I don't think it's just that like,
Lex Fridman (2:29:13.260)
I could have gotten lucky at any point.
George Hotz (2:29:14.820)
I think that in a way.
Lex Fridman (2:29:16.260)
You were ready at that moment.
George Hotz (2:29:17.820)
Yeah, exactly.
Lex Fridman (2:29:18.660)
To receive the information.
Lex Fridman (2:29:21.460)
But is there some magic to the day today
Lex Fridman (2:29:24.300)
of like eating breakfast?
Lex Fridman (2:29:27.220)
And it's the mundane things.
Lex Fridman (2:29:29.060)
Nah.
George Hotz (2:29:29.900)
Nothing.
Lex Fridman (2:29:30.720)
Nah, I drift through life.
George Hotz (2:29:32.900)
Without structure.
Lex Fridman (2:29:34.180)
I drift through life hoping and praying
George Hotz (2:29:36.220)
that I will get another day like those days.
Lex Fridman (2:29:38.540)
And there's nothing in particular you do
George Hotz (2:29:40.320)
to be a receptacle for another, for day number four.
Lex Fridman (2:29:46.060)
No, I didn't do anything to get the other ones.
Lex Fridman (2:29:48.060)
So I don't think I have to really do anything now.
Lex Fridman (2:29:51.100)
I took a month long trip to New York
Lex Fridman (2:29:53.420)
and the Ethereum thing was the highlight of it,
Lex Fridman (2:29:56.260)
but the rest of it was pretty terrible.
George Hotz (2:29:57.980)
I did a two week road trip
Lex Fridman (2:29:59.860)
and I got, I had to turn around.
George Hotz (2:30:01.740)
I had to turn around driving in Gunnison, Colorado.
Lex Fridman (2:30:06.620)
I passed through Gunnison
Lex Fridman (2:30:08.060)
and the snow starts coming down.
Lex Fridman (2:30:10.660)
There's a pass up there called Monarch Pass
George Hotz (2:30:12.160)
in order to get through to Denver,
Lex Fridman (2:30:13.000)
you gotta get over the Rockies.
Lex Fridman (2:30:14.700)
And I had to turn my car around.
Lex Fridman (2:30:16.800)
I couldn't, I watched a F150 go off the road.
George Hotz (2:30:20.260)
I'm like, I gotta go back.
Lex Fridman (2:30:21.620)
And like that day was meaningful.
George Hotz (2:30:24.620)
Cause like, it was real.
Lex Fridman (2:30:26.040)
Like I actually had to turn my car around.
George Hotz (2:30:28.900)
It's rare that anything even real happens in my life.
Lex Fridman (2:30:31.240)
Even as, you know, mundane as the fact that,
George Hotz (2:30:34.300)
yeah, there was snow, I had to turn around,
Lex Fridman (2:30:36.100)
stay in Gunnison and leave the next day.
George Hotz (2:30:37.700)
Something about that moment felt real.
Lex Fridman (2:30:40.260)
Okay, so actually it's interesting to break apart
George Hotz (2:30:43.180)
the three moments you mentioned, if it's okay.
Lex Fridman (2:30:45.380)
So I always have trouble pronouncing his name,
Lex Fridman (2:30:48.540)
but Alousa Yurkowski.
Lex Fridman (2:30:53.300)
So what, how did your worldview change
George Hotz (2:30:57.820)
in starting to consider the exponential growth of AI
Lex Fridman (2:31:02.900)
and AGI that he thinks about
Lex Fridman (2:31:05.020)
and the threats of artificial intelligence
Lex Fridman (2:31:07.620)
and all that kind of ideas?
Lex Fridman (2:31:09.340)
Can you, is it just like, can you maybe break apart
Lex Fridman (2:31:12.460)
like what exactly was so magical to you?
Lex Fridman (2:31:15.660)
Is it transformational experience?
Lex Fridman (2:31:17.380)
Today, everyone knows him for threats and AI safety.
George Hotz (2:31:20.340)
This was pre that stuff.
Lex Fridman (2:31:22.260)
There was, I don't think a mention of AI safety on the page.
George Hotz (2:31:25.980)
This is, this is old Yurkowski stuff.
Lex Fridman (2:31:27.860)
He'd probably denounce it all now.
George Hotz (2:31:29.060)
He'd probably be like,
Lex Fridman (2:31:29.900)
that's exactly what I didn't want to happen.
George Hotz (2:31:32.220)
Sorry, man.
Lex Fridman (2:31:33.060)
Is there something specific you can take from his work
Lex Fridman (2:31:37.700)
that you can remember?
Lex Fridman (2:31:38.700)
Yeah, it was this realization
George Hotz (2:31:40.780)
that computers double in power every 18 months
Lex Fridman (2:31:45.980)
and humans do not, and they haven't crossed yet.
Lex Fridman (2:31:50.140)
But if you have one thing that's doubling every 18 months
Lex Fridman (2:31:52.700)
and one thing that's staying like this, you know,
George Hotz (2:31:55.100)
here's your log graph, here's your line, you know,
Lex Fridman (2:31:58.940)
calculate that.
Lex Fridman (2:31:59.860)
And then the data opened the door
Lex Fridman (2:32:03.540)
to the exponential thinking, like thinking that like,
George Hotz (2:32:06.420)
you know what, with technology,
Lex Fridman (2:32:07.740)
we can actually transform the world.
George Hotz (2:32:11.580)
It opened the door to human obsolescence.
Lex Fridman (2:32:13.740)
It opened the door to realize that in my lifetime,
George Hotz (2:32:16.980)
humans are going to be replaced.
Lex Fridman (2:32:20.620)
And then the matching idea to that of artificial intelligence
George Hotz (2:32:23.740)
with the Hutter prize, you know, I'm torn.
Lex Fridman (2:32:27.940)
I go back and forth on what I think about it.
George Hotz (2:32:30.220)
Yeah.
Lex Fridman (2:32:31.380)
But the basic thesis is it's a nice compelling notion
George Hotz (2:32:36.180)
that we can reduce the task of creating
Lex Fridman (2:32:38.340)
an intelligent system, a generally intelligent system
George Hotz (2:32:41.380)
into the task of compression.
Lex Fridman (2:32:43.940)
So you can think of all of intelligence in the universe,
George Hotz (2:32:46.300)
in fact, as a kind of compression.
Lex Fridman (2:32:50.220)
Do you find that, was that just at the time
George Hotz (2:32:52.420)
you found that as a compelling idea
Lex Fridman (2:32:53.900)
or do you still find that a compelling idea?
George Hotz (2:32:56.980)
I still find that a compelling idea.
Lex Fridman (2:32:59.180)
I think that it's not that useful day to day,
Lex Fridman (2:33:02.180)
but actually one of maybe my quests before that
Lex Fridman (2:33:06.260)
was a search for the definition of the word intelligence.
Lex Fridman (2:33:09.340)
And I never had one.
Lex Fridman (2:33:10.860)
And I definitely have a definition of the word compression.
George Hotz (2:33:14.700)
It's a very simple, straightforward one.
Lex Fridman (2:33:18.420)
And you know what compression is,
George Hotz (2:33:19.740)
you know what lossless, it's lossless compression,
Lex Fridman (2:33:21.100)
not lossy, lossless compression.
Lex Fridman (2:33:22.940)
And that that is equivalent to intelligence,
Lex Fridman (2:33:25.620)
which I believe, I'm not sure how useful
George Hotz (2:33:27.540)
that definition is day to day,
Lex Fridman (2:33:28.980)
but like I now have a framework to understand what it is.
Lex Fridman (2:33:32.340)
And he just 10X the prize for that competition
Lex Fridman (2:33:36.300)
like recently a few months ago.
Lex Fridman (2:33:37.700)
You ever thought of taking a crack at that?
Lex Fridman (2:33:39.580)
Oh, I did.
George Hotz (2:33:41.100)
Oh, I did.
Lex Fridman (2:33:41.940)
I spent the next, after I found the prize,
George Hotz (2:33:44.660)
I spent the next six months of my life trying it.
Lex Fridman (2:33:47.820)
And well, that's when I started learning everything about AI.
Lex Fridman (2:33:51.300)
And then I worked at Vicarious for a bit
Lex Fridman (2:33:53.340)
and then I read all the deep learning stuff.
Lex Fridman (2:33:55.340)
And I'm like, okay, now I like I'm caught up to modern AI.
Lex Fridman (2:33:58.460)
Wow.
Lex Fridman (2:33:59.300)
And I had a really good framework to put it all in
Lex Fridman (2:34:01.380)
from the compression stuff, right?
George Hotz (2:34:04.420)
Like some of the first deep learning models I played with
Lex Fridman (2:34:07.340)
were GPT basically, but before transformers,
George Hotz (2:34:12.020)
before it was still RNNs to do character prediction.
Lex Fridman (2:34:17.900)
But by the way, on the compression side,
George Hotz (2:34:19.940)
I mean, especially with neural networks,
Lex Fridman (2:34:22.940)
what do you make of the lossless requirement
Lex Fridman (2:34:25.500)
with the Hutter prize?
Lex Fridman (2:34:26.540)
So, you know, human intelligence and neural networks
George Hotz (2:34:31.460)
can probably compress stuff pretty well,
Lex Fridman (2:34:33.300)
but it would be lossy.
George Hotz (2:34:35.220)
It's imperfect.
Lex Fridman (2:34:36.660)
You can turn a lossy compression
George Hotz (2:34:37.860)
to a lossless compressor pretty easily
Lex Fridman (2:34:39.620)
using an arithmetic encoder, right?
George Hotz (2:34:41.060)
You can take an arithmetic encoder
Lex Fridman (2:34:42.540)
and you can just encode the noise with maximum efficiency.
Lex Fridman (2:34:45.980)
Right?
Lex Fridman (2:34:46.820)
So even if you can't predict exactly
Lex Fridman (2:34:48.900)
what the next character is,
Lex Fridman (2:34:50.460)
the better a probability distribution,
George Hotz (2:34:52.500)
you can put over the next character.
Lex Fridman (2:34:54.260)
You can then use an arithmetic encoder to, right?
George Hotz (2:34:57.340)
You don't have to know whether it's an E or an I,
Lex Fridman (2:34:59.300)
you just have to put good probabilities on them
Lex Fridman (2:35:01.060)
and then, you know, code those.
Lex Fridman (2:35:03.020)
And if you have, it's a bits of entropy thing, right?
Lex Fridman (2:35:06.460)
So let me, on that topic,
Lex Fridman (2:35:07.740)
it'd be interesting as a little side tour.
Lex Fridman (2:35:10.260)
What are your thoughts in this year about GPT3
Lex Fridman (2:35:13.580)
and these language models and these transformers?
George Hotz (2:35:16.140)
Is there something interesting to you as an AI researcher,
Lex Fridman (2:35:20.660)
or is there something interesting to you
Lex Fridman (2:35:22.540)
as an autonomous vehicle developer?
Lex Fridman (2:35:24.740)
Nah, I think it's overhyped.
George Hotz (2:35:27.460)
I mean, it's not, like, it's cool.
Lex Fridman (2:35:29.020)
It's cool for what it is, but no,
George Hotz (2:35:30.820)
we're not just gonna be able to scale up to GPT12
Lex Fridman (2:35:33.500)
and get general purpose intelligence.
George Hotz (2:35:35.340)
Like, your loss function is literally just,
Lex Fridman (2:35:38.700)
you know, cross entropy loss on the character, right?
George Hotz (2:35:41.460)
Like, that's not the loss function of general intelligence.
Lex Fridman (2:35:44.860)
Is that obvious to you?
George Hotz (2:35:45.940)
Yes.
Lex Fridman (2:35:47.380)
Can you imagine that, like,
George Hotz (2:35:51.460)
to play devil's advocate on yourself,
Lex Fridman (2:35:53.300)
is it possible that you can,
George Hotz (2:35:55.860)
the GPT12 will achieve general intelligence
Lex Fridman (2:35:58.580)
with something as dumb as this kind of loss function?
George Hotz (2:36:01.980)
I guess it depends what you mean by general intelligence.
Lex Fridman (2:36:05.060)
So there's another problem with the GPTs,
Lex Fridman (2:36:07.860)
and that's that they don't have a,
Lex Fridman (2:36:11.060)
they don't have longterm memory.
George Hotz (2:36:13.020)
Right.
Lex Fridman (2:36:13.860)
So, like, just GPT12,
George Hotz (2:36:18.580)
a scaled up version of GPT2 or GPT3,
Lex Fridman (2:36:22.460)
I find it hard to believe.
George Hotz (2:36:26.060)
Well, you can scale it in,
Lex Fridman (2:36:28.780)
so it's a hard coded length,
Lex Fridman (2:36:32.060)
but you can make it wider and wider and wider.
Lex Fridman (2:36:34.300)
Yeah.
George Hotz (2:36:36.380)
You're gonna get cool things from those systems,
Lex Fridman (2:36:40.820)
but I don't think you're ever gonna get something
George Hotz (2:36:44.820)
that can, like, you know, build me a rocket ship.
Lex Fridman (2:36:47.580)
What about solved driving?
George Hotz (2:36:49.420)
So, you know, you can use Transformer with video,
Lex Fridman (2:36:53.140)
for example.
Lex Fridman (2:36:54.740)
You think, is there something in there?
Lex Fridman (2:36:57.220)
No, because, I mean, look, we use a GRU.
George Hotz (2:37:01.260)
We use a GRU.
Lex Fridman (2:37:02.100)
We could change that GRU out to a Transformer.
George Hotz (2:37:05.900)
I think driving is much more Markovian than language.
Lex Fridman (2:37:09.500)
So, Markovian, you mean, like, the memory,
Lex Fridman (2:37:11.460)
which aspect of Markovian?
Lex Fridman (2:37:13.660)
I mean that, like, most of the information
George Hotz (2:37:16.340)
in the state at T minus one is also in state T.
Lex Fridman (2:37:19.540)
I see, yeah.
George Hotz (2:37:20.380)
Right, and it kind of, like, drops off nicely like this,
Lex Fridman (2:37:22.820)
whereas sometime with language,
George Hotz (2:37:23.900)
you have to refer back to the third paragraph
Lex Fridman (2:37:25.900)
on the second page.
George Hotz (2:37:27.260)
I feel like.
Lex Fridman (2:37:28.140)
There's not many, like, you can say, like,
George Hotz (2:37:30.060)
speed limit signs, but there's really not many things
Lex Fridman (2:37:32.420)
in autonomous driving that look like that.
Lex Fridman (2:37:33.860)
But if you look at, to play devil's advocate,
Lex Fridman (2:37:37.140)
is the risk estimation thing that you've talked about
George Hotz (2:37:39.740)
is kind of interesting.
Lex Fridman (2:37:41.180)
Is, it feels like there might be some longer term
George Hotz (2:37:45.700)
aggregation of context necessary to be able to figure out,
Lex Fridman (2:37:49.020)
like, the context.
George Hotz (2:37:51.900)
Yeah, I'm not even sure I'm believing my devil's advocate.
Lex Fridman (2:37:55.820)
We have a nice, like, vision model,
George Hotz (2:37:58.260)
which outputs, like, a one or two,
Lex Fridman (2:38:00.180)
four dimensional perception space.
Lex Fridman (2:38:03.300)
Can I try Transformers on it?
Lex Fridman (2:38:04.740)
Sure, I probably will.
George Hotz (2:38:06.580)
At some point, we'll try Transformers,
Lex Fridman (2:38:08.260)
and then we'll just see.
Lex Fridman (2:38:09.100)
Do they do better?
Lex Fridman (2:38:09.940)
Sure, I'm.
Lex Fridman (2:38:10.780)
But it might not be a game changer, you're saying?
Lex Fridman (2:38:12.180)
No, well, I'm not.
George Hotz (2:38:13.020)
Like, might Transformers work better than GRUs
Lex Fridman (2:38:15.300)
for autonomous driving?
George Hotz (2:38:16.140)
Sure.
Lex Fridman (2:38:16.980)
Might we switch?
George Hotz (2:38:17.820)
Sure.
Lex Fridman (2:38:18.660)
Is this some radical change?
George Hotz (2:38:19.500)
No.
Lex Fridman (2:38:20.340)
Okay, we use a slightly different,
George Hotz (2:38:21.540)
you know, we switch from RNNs to GRUs.
Lex Fridman (2:38:23.020)
Like, okay, maybe it's GRUs to Transformers,
Lex Fridman (2:38:24.980)
but no, it's not.
Lex Fridman (2:38:26.540)
Yeah.
George Hotz (2:38:27.900)
Well, on the topic of general intelligence,
Lex Fridman (2:38:30.500)
I don't know how much I've talked to you about it.
George Hotz (2:38:32.140)
Like, what, do you think we'll actually build
Lex Fridman (2:38:36.540)
an AGI?
George Hotz (2:38:38.060)
Like, if you look at Ray Kurzweil with Singularity,
Lex Fridman (2:38:40.740)
do you have like an intuition about,
George Hotz (2:38:43.380)
you're kind of saying driving is easy.
Lex Fridman (2:38:45.420)
Yeah.
Lex Fridman (2:38:46.260)
And I tend to personally believe that solving driving
Lex Fridman (2:38:52.500)
will have really deep, important impacts
George Hotz (2:38:56.580)
on our ability to solve general intelligence.
Lex Fridman (2:38:59.300)
Like, I think driving doesn't require general intelligence,
Lex Fridman (2:39:03.380)
but I think they're going to be neighbors
Lex Fridman (2:39:05.220)
in a way that it's like deeply tied.
George Hotz (2:39:08.380)
Cause it's so, like driving is so deeply connected
Lex Fridman (2:39:11.540)
to the human experience that I think solving one
George Hotz (2:39:15.020)
will help solve the other.
Lex Fridman (2:39:17.420)
But, so I don't see, I don't see driving as like easy
Lex Fridman (2:39:20.980)
and almost like separate than general intelligence,
Lex Fridman (2:39:23.540)
but like, what's your vision of a future with a Singularity?
Lex Fridman (2:39:26.700)
Do you see there'll be a single moment,
Lex Fridman (2:39:28.180)
like a Singularity where it'll be a phase shift?
Lex Fridman (2:39:30.620)
Are we in the Singularity now?
Lex Fridman (2:39:32.380)
Like what, do you have crazy ideas about the future
Lex Fridman (2:39:34.540)
in terms of AGI?
Lex Fridman (2:39:35.740)
We're definitely in the Singularity now.
Lex Fridman (2:39:38.020)
We are?
Lex Fridman (2:39:38.860)
Of course, of course.
George Hotz (2:39:40.180)
Look at the bandwidth between people.
Lex Fridman (2:39:41.500)
The bandwidth between people goes up, right?
George Hotz (2:39:44.820)
The Singularity is just, you know, when the bandwidth, but.
Lex Fridman (2:39:47.300)
What do you mean by the bandwidth of people?
George Hotz (2:39:48.620)
Communications, tools, the whole world is networked.
Lex Fridman (2:39:51.340)
The whole world is networked
Lex Fridman (2:39:52.260)
and we raise the speed of that network, right?
Lex Fridman (2:39:54.740)
Oh, so you think the communication of information
George Hotz (2:39:57.380)
in a distributed way is an empowering thing
Lex Fridman (2:40:00.980)
for collective intelligence?
George Hotz (2:40:02.300)
Oh, I didn't say it's necessarily a good thing,
Lex Fridman (2:40:03.660)
but I think that's like,
George Hotz (2:40:04.540)
when I think of the definition of the Singularity,
Lex Fridman (2:40:06.500)
yeah, it seems kind of right.
George Hotz (2:40:08.180)
I see, like it's a change in the world
Lex Fridman (2:40:12.060)
beyond which like the world be transformed
George Hotz (2:40:14.940)
in ways that we can't possibly imagine.
Lex Fridman (2:40:16.780)
No, I mean, I think we're in the Singularity now
George Hotz (2:40:18.380)
in the sense that there's like, you know,
Lex Fridman (2:40:19.700)
one world and a monoculture and it's also linked.
George Hotz (2:40:22.340)
Yeah, I mean, I kind of share the intuition
Lex Fridman (2:40:24.800)
that the Singularity will originate
George Hotz (2:40:27.700)
from the collective intelligence of us ands
Lex Fridman (2:40:31.220)
versus the like some single system AGI type thing.
George Hotz (2:40:35.340)
Oh, I totally agree with that.
Lex Fridman (2:40:37.020)
Yeah, I don't really believe in like a hard take off AGI
George Hotz (2:40:40.460)
kind of thing.
Lex Fridman (2:40:45.140)
Yeah, I don't even think AI is all that different in kind
George Hotz (2:40:49.500)
from what we've already been building.
Lex Fridman (2:40:52.060)
With respect to driving,
George Hotz (2:40:53.380)
I think driving is a subset of general intelligence
Lex Fridman (2:40:56.340)
and I think it's a pretty complete subset.
George Hotz (2:40:58.060)
I think the tools we develop at Kama
Lex Fridman (2:41:00.300)
will also be extremely helpful
George Hotz (2:41:02.300)
to solving general intelligence
Lex Fridman (2:41:04.020)
and that's I think the real reason why I'm doing it.
George Hotz (2:41:06.620)
I don't care about self driving cars.
Lex Fridman (2:41:08.420)
It's a cool problem to beat people at.
Lex Fridman (2:41:10.920)
But yeah, I mean, yeah, you're kind of, you're of two minds.
Lex Fridman (2:41:14.540)
So one, you do have to have a mission
Lex Fridman (2:41:16.340)
and you wanna focus and make sure you get there.
Lex Fridman (2:41:19.020)
You can't forget that but at the same time,
George Hotz (2:41:22.280)
there is a thread that's much bigger
Lex Fridman (2:41:26.060)
than that connects the entirety of your effort.
George Hotz (2:41:28.460)
That's much bigger than just driving.
Lex Fridman (2:41:31.180)
With AI and with general intelligence,
George Hotz (2:41:33.380)
it is so easy to delude yourself
Lex Fridman (2:41:35.220)
into thinking you've figured something out when you haven't.
George Hotz (2:41:37.300)
If we build a level five self driving car,
Lex Fridman (2:41:39.820)
we have indisputably built something.
George Hotz (2:41:42.660)
Yeah.
Lex Fridman (2:41:43.500)
Is it general intelligence?
George Hotz (2:41:44.800)
I'm not gonna debate that.
Lex Fridman (2:41:45.940)
I will say we've built something
George Hotz (2:41:47.520)
that provides huge financial value.
Lex Fridman (2:41:49.740)
Yeah, beautifully put.
George Hotz (2:41:50.580)
That's the engineering credo.
Lex Fridman (2:41:51.940)
Like just build the thing.
George Hotz (2:41:53.660)
It's like, that's why I'm with Elon
Lex Fridman (2:41:57.120)
on go to Mars.
George Hotz (2:41:58.780)
Yeah, that's a great one.
Lex Fridman (2:41:59.740)
You can argue like who the hell cares about going to Mars.
Lex Fridman (2:42:03.700)
But the reality is set that as a mission, get it done.
Lex Fridman (2:42:07.220)
Yeah.
Lex Fridman (2:42:08.060)
And then you're going to crack some problem
Lex Fridman (2:42:09.540)
that you've never even expected
George Hotz (2:42:11.580)
in the process of doing that, yeah.
Lex Fridman (2:42:13.900)
Yeah, I mean, no, I think if I had a choice
George Hotz (2:42:16.220)
between humanity going to Mars
Lex Fridman (2:42:17.460)
and solving self driving cars,
George Hotz (2:42:18.500)
I think going to Mars is better, but I don't know.
Lex Fridman (2:42:21.900)
I'm more suited for self driving cars.
George Hotz (2:42:23.580)
I'm an information guy.
Lex Fridman (2:42:24.420)
I'm not a modernist, I'm a postmodernist.
George Hotz (2:42:26.540)
Postmodernist, all right, beautifully put.
Lex Fridman (2:42:29.620)
Let me drag you back to programming for a sec.
Lex Fridman (2:42:32.220)
What three, maybe three to five programming languages
Lex Fridman (2:42:35.140)
should people learn, do you think?
George Hotz (2:42:36.580)
Like if you look at yourself,
Lex Fridman (2:42:38.220)
what did you get the most out of from learning?
George Hotz (2:42:42.700)
Well, so everybody should learn C and assembly.
Lex Fridman (2:42:45.980)
We'll start with those two, right?
Lex Fridman (2:42:47.380)
Assembly?
Lex Fridman (2:42:48.220)
Yeah, if you can't code an assembly,
George Hotz (2:42:49.940)
you don't know what the computer's doing.
Lex Fridman (2:42:51.720)
You don't understand like,
George Hotz (2:42:53.340)
you don't have to be great in assembly,
Lex Fridman (2:42:54.820)
but you have to code in it.
Lex Fridman (2:42:56.580)
And then like, you have to appreciate assembly
Lex Fridman (2:42:58.940)
in order to appreciate all the great things C gets you.
Lex Fridman (2:43:02.020)
And then you have to code in C
Lex Fridman (2:43:03.380)
in order to appreciate all the great things Python gets you.
Lex Fridman (2:43:06.340)
So I'll just say assembly C and Python,
Lex Fridman (2:43:07.860)
we'll start with those three.
George Hotz (2:43:09.740)
The memory allocation of C and the fact that,
Lex Fridman (2:43:14.700)
so assembly gives you a sense
George Hotz (2:43:16.340)
of just how many levels of abstraction
Lex Fridman (2:43:18.460)
you get to work on in modern day programming.
George Hotz (2:43:20.660)
Yeah, yeah, yeah, yeah, graph coloring for assignment,
Lex Fridman (2:43:22.900)
register assignment and compilers.
George Hotz (2:43:24.740)
Like, you know, you gotta do,
Lex Fridman (2:43:25.940)
you know, the compiler,
George Hotz (2:43:26.780)
the computer only has a certain number of registers,
Lex Fridman (2:43:28.340)
yet you can have all the variables you want in a C function.
Lex Fridman (2:43:31.140)
So you get to start to build intuition about compilation,
Lex Fridman (2:43:34.460)
like what a compiler gets you.
Lex Fridman (2:43:37.380)
What else?
Lex Fridman (2:43:38.500)
Well, then there's kind of a,
Lex Fridman (2:43:41.220)
so those are all very imperative programming languages.
Lex Fridman (2:43:45.800)
Then there's two other paradigms for programming
George Hotz (2:43:47.700)
that everybody should be familiar with.
Lex Fridman (2:43:49.220)
And one of them is functional.
George Hotz (2:43:51.260)
You should learn Haskell and take that all the way through,
Lex Fridman (2:43:54.140)
learn a language with dependent types like Coq,
George Hotz (2:43:57.260)
learn that whole space,
Lex Fridman (2:43:58.980)
like the very PL theory, heavy languages.
Lex Fridman (2:44:02.740)
And Haskell is your favorite functional?
Lex Fridman (2:44:04.860)
Is that the go to, you'd say?
George Hotz (2:44:06.500)
Yeah, I'm not a great Haskell programmer.
Lex Fridman (2:44:08.580)
I wrote a compiler in Haskell once.
George Hotz (2:44:10.580)
There's another paradigm,
Lex Fridman (2:44:11.460)
and actually there's one more paradigm
George Hotz (2:44:12.540)
that I'll even talk about after that,
Lex Fridman (2:44:14.180)
that I never used to talk about
George Hotz (2:44:15.100)
when I would think about this,
Lex Fridman (2:44:15.940)
but the next paradigm is learn Verilog of HDL.
George Hotz (2:44:20.240)
Understand this idea of all of the instructions
Lex Fridman (2:44:22.280)
execute at once. If I have a block in Verilog
Lex Fridman (2:44:26.860)
and I write stuff in it, it's not sequential.
Lex Fridman (2:44:29.900)
They all execute at once.
Lex Fridman (2:44:33.740)
And then think like that, that's how hardware works.
Lex Fridman (2:44:36.700)
To be, so I guess assembly doesn't quite get you that.
George Hotz (2:44:40.020)
Assembly is more about compilation,
Lex Fridman (2:44:42.260)
and Verilog is more about the hardware,
George Hotz (2:44:44.260)
like giving a sense of what actually
Lex Fridman (2:44:46.420)
is the hardware is doing.
George Hotz (2:44:48.580)
Assembly, C, Python are straight,
Lex Fridman (2:44:50.480)
like they sit right on top of each other.
George Hotz (2:44:52.420)
In fact, C is, well, C is kind of coded in C,
Lex Fridman (2:44:55.660)
but you could imagine the first C was coded in assembly,
Lex Fridman (2:44:57.900)
and Python is actually coded in C.
Lex Fridman (2:45:00.020)
So you can straight up go on that.
George Hotz (2:45:03.620)
Got it, and then Verilog gives you, that's brilliant.
Lex Fridman (2:45:06.940)
Okay.
Lex Fridman (2:45:07.780)
And then I think there's another one now.
Lex Fridman (2:45:09.820)
Everyone, Carpathia calls it programming 2.0,
George Hotz (2:45:12.660)
which is learn a, I'm not even gonna,
Lex Fridman (2:45:16.500)
don't learn TensorFlow, learn PyTorch.
Lex Fridman (2:45:18.620)
So machine learning.
Lex Fridman (2:45:20.180)
We've got to come up with a better term
George Hotz (2:45:21.540)
than programming 2.0, or, but yeah.
Lex Fridman (2:45:26.100)
It's a programming language, learn it.
George Hotz (2:45:29.900)
I wonder if it can be formalized a little bit better.
Lex Fridman (2:45:32.660)
It feels like we're in the early days
George Hotz (2:45:34.900)
of what that actually entails.
Lex Fridman (2:45:37.100)
Data driven programming?
George Hotz (2:45:39.220)
Data driven programming, yeah.
Lex Fridman (2:45:41.700)
But it's so fundamentally different
George Hotz (2:45:43.080)
as a paradigm than the others.
Lex Fridman (2:45:44.900)
Like it almost requires a different skillset.
Lex Fridman (2:45:50.420)
But you think it's still, yeah.
Lex Fridman (2:45:53.740)
And PyTorch versus TensorFlow, PyTorch wins.
George Hotz (2:45:56.580)
It's the fourth paradigm.
Lex Fridman (2:45:57.460)
It's the fourth paradigm that I've kind of seen.
George Hotz (2:45:59.380)
There's like this, you know,
Lex Fridman (2:46:01.460)
imperative functional hardware.
George Hotz (2:46:04.840)
I don't know a better word for it.
Lex Fridman (2:46:06.340)
And then ML.
Lex Fridman (2:46:08.180)
Do you have advice for people that wanna,
Lex Fridman (2:46:13.180)
you know, get into programming, wanna learn programming?
George Hotz (2:46:16.180)
You have a video,
Lex Fridman (2:46:19.940)
what is programming noob lessons, exclamation point.
Lex Fridman (2:46:22.860)
And I think the top comment is like,
Lex Fridman (2:46:24.620)
warning, this is not for noobs.
Lex Fridman (2:46:27.820)
Do you have a noob, like a TLDW for that video,
Lex Fridman (2:46:32.560)
but also a noob friendly advice
Lex Fridman (2:46:38.020)
on how to get into programming?
Lex Fridman (2:46:39.540)
We're never going to learn programming
George Hotz (2:46:41.400)
by watching a video called Learn Programming.
Lex Fridman (2:46:44.640)
The only way to learn programming, I think,
Lex Fridman (2:46:46.660)
and the only one is that the only way
Lex Fridman (2:46:48.060)
everyone I've ever met who can program well,
George Hotz (2:46:50.140)
learned it all in the same way.
Lex Fridman (2:46:51.940)
They had something they wanted to do
Lex Fridman (2:46:54.960)
and then they tried to do it.
Lex Fridman (2:46:56.580)
And then they were like, oh, well, okay.
George Hotz (2:47:00.020)
This is kind of, you know, it'd be nice
Lex Fridman (2:47:01.380)
if the computer could kind of do this.
Lex Fridman (2:47:02.740)
And then, you know, that's how you learn.
Lex Fridman (2:47:04.440)
You just keep pushing on a project.
Lex Fridman (2:47:09.020)
So the only advice I have for learning programming
Lex Fridman (2:47:10.900)
is go program.
George Hotz (2:47:12.140)
Somebody wrote to me a question like,
Lex Fridman (2:47:14.680)
we don't really, they're looking to learn
George Hotz (2:47:17.100)
about recurring neural networks.
Lex Fridman (2:47:19.060)
And he's saying, like, my company's thinking
George Hotz (2:47:20.540)
of using recurring neural networks for time series data,
Lex Fridman (2:47:24.140)
but we don't really have an idea of where to use it yet.
George Hotz (2:47:27.420)
We just want to, like, do you have any advice
Lex Fridman (2:47:28.980)
on how to learn about, these are these kind of
George Hotz (2:47:31.780)
general machine learning questions.
Lex Fridman (2:47:33.340)
And I think the answer is, like,
George Hotz (2:47:36.780)
actually have a problem that you're trying to solve.
Lex Fridman (2:47:39.020)
And just.
George Hotz (2:47:40.220)
I see that stuff.
Lex Fridman (2:47:41.220)
Oh my God, when people talk like that,
George Hotz (2:47:42.660)
they're like, I heard machine learning is important.
Lex Fridman (2:47:45.500)
Could you help us integrate machine learning
Lex Fridman (2:47:47.500)
with macaroni and cheese production?
Lex Fridman (2:47:51.300)
You just, I don't even, you can't help these people.
Lex Fridman (2:47:54.000)
Like, who lets you run anything?
Lex Fridman (2:47:55.980)
Who lets that kind of person run anything?
George Hotz (2:47:58.400)
I think we're all, we're all beginners at some point.
Lex Fridman (2:48:02.620)
So.
George Hotz (2:48:03.460)
It's not like they're a beginner.
Lex Fridman (2:48:04.860)
It's like, my problem is not that they don't know
George Hotz (2:48:07.340)
about machine learning.
Lex Fridman (2:48:08.620)
My problem is that they think that machine learning
George Hotz (2:48:10.840)
has something to say about macaroni and cheese production.
Lex Fridman (2:48:14.740)
Or like, I heard about this new technology.
Lex Fridman (2:48:17.020)
How can I use it for why?
Lex Fridman (2:48:19.860)
Like, I don't know what it is, but how can I use it for why?
George Hotz (2:48:23.500)
That's true.
Lex Fridman (2:48:24.340)
You have to build up an intuition of how,
George Hotz (2:48:26.140)
cause you might be able to figure out a way,
Lex Fridman (2:48:27.620)
but like the prerequisites,
George Hotz (2:48:29.300)
you should have a macaroni and cheese problem to solve first.
Lex Fridman (2:48:32.300)
Exactly.
Lex Fridman (2:48:33.460)
And then two, you should have more traditional,
Lex Fridman (2:48:36.940)
like the learning process should involve
George Hotz (2:48:39.340)
more traditionally applicable problems
Lex Fridman (2:48:41.940)
in the space of whatever that is, machine learning,
Lex Fridman (2:48:44.540)
and then see if it can be applied to mac and cheese.
Lex Fridman (2:48:47.060)
At least start with, tell me about a problem.
George Hotz (2:48:49.180)
Like if you have a problem, you're like,
Lex Fridman (2:48:50.740)
you know, some of my boxes aren't getting
George Hotz (2:48:52.500)
enough macaroni in them.
Lex Fridman (2:48:54.500)
Can we use machine learning to solve this problem?
George Hotz (2:48:56.820)
That's much, much better than how do I apply
Lex Fridman (2:48:59.380)
machine learning to macaroni and cheese?
George Hotz (2:49:01.600)
One big thing, maybe this is me talking
Lex Fridman (2:49:05.380)
to the audience a little bit, cause I get these days
Lex Fridman (2:49:07.860)
so many messages, advice on how to like learn stuff, okay?
Lex Fridman (2:49:15.060)
My, this is not me being mean.
George Hotz (2:49:18.160)
I think this is quite profound actually,
Lex Fridman (2:49:20.540)
is you should Google it.
George Hotz (2:49:22.780)
Oh yeah.
Lex Fridman (2:49:23.820)
Like one of the like skills that you should really acquire
George Hotz (2:49:29.700)
as an engineer, as a researcher, as a thinker,
Lex Fridman (2:49:33.100)
like one, there's two complementary skills.
George Hotz (2:49:36.660)
Like one is with a blank sheet of paper
Lex Fridman (2:49:39.100)
with no internet to think deeply.
Lex Fridman (2:49:41.660)
And then the other is to Google the crap
Lex Fridman (2:49:44.740)
out of the questions you have.
George Hotz (2:49:45.940)
Like that's actually a skill people often talk about,
Lex Fridman (2:49:49.180)
but like doing research, like pulling at the thread,
George Hotz (2:49:52.020)
like looking up different words,
Lex Fridman (2:49:53.860)
going into like GitHub repositories with two stars
Lex Fridman (2:49:58.060)
and like looking how they did stuff,
Lex Fridman (2:49:59.700)
like looking at the code or going on Twitter,
George Hotz (2:50:03.460)
seeing like there's little pockets of brilliant people
Lex Fridman (2:50:05.780)
that are like having discussions.
George Hotz (2:50:07.620)
Like if you're a neuroscientist,
Lex Fridman (2:50:09.860)
go into signal processing community.
George Hotz (2:50:11.620)
If you're an AI person going into the psychology community,
Lex Fridman (2:50:15.860)
like switch communities.
George Hotz (2:50:18.020)
I keep searching, searching, searching,
Lex Fridman (2:50:19.980)
because it's so much better to invest
George Hotz (2:50:23.860)
in like finding somebody else who already solved your problem
Lex Fridman (2:50:27.260)
than it is to try to solve the problem.
Lex Fridman (2:50:30.900)
And because they've often invested years of their life,
Lex Fridman (2:50:34.380)
like entire communities are probably already out there
George Hotz (2:50:37.420)
who have tried to solve your problem.
Lex Fridman (2:50:39.180)
I think they're the same thing.
George Hotz (2:50:40.980)
I think you go try to solve the problem.
Lex Fridman (2:50:44.180)
And then in trying to solve the problem,
George Hotz (2:50:46.180)
if you're good at solving problems,
Lex Fridman (2:50:47.740)
you'll stumble upon the person who solved it already.
Lex Fridman (2:50:50.260)
But the stumbling is really important.
Lex Fridman (2:50:52.300)
I think that's a skill that people should really put,
George Hotz (2:50:54.260)
especially in undergrad, like search.
Lex Fridman (2:50:57.700)
If you ask me a question,
Lex Fridman (2:50:58.700)
how should I get started in deep learning, like especially?
Lex Fridman (2:51:04.100)
Like that is just so Googleable.
George Hotz (2:51:07.260)
Like the whole point is you Google that
Lex Fridman (2:51:10.060)
and you get a million pages and just start looking at them.
George Hotz (2:51:13.980)
Start pulling at the threads, start exploring,
Lex Fridman (2:51:16.420)
start taking notes, start getting advice
George Hotz (2:51:19.300)
from a million people that already like spent their life
Lex Fridman (2:51:22.820)
answering that question, actually.
George Hotz (2:51:25.060)
Oh, well, yeah, I mean, that's definitely also, yeah,
Lex Fridman (2:51:26.460)
when people like ask me things like that, I'm like, trust me,
George Hotz (2:51:28.500)
the top answer on Google is much, much better
Lex Fridman (2:51:30.340)
than anything I'm going to tell you, right?
George Hotz (2:51:32.860)
Yeah.
Lex Fridman (2:51:34.460)
People ask, it's an interesting question.
George Hotz (2:51:38.100)
Let me know if you have any recommendations.
Lex Fridman (2:51:39.940)
What three books, technical or fiction or philosophical,
Lex Fridman (2:51:43.700)
had an impact on your life or you would recommend perhaps?
Lex Fridman (2:51:49.180)
Maybe we'll start with the least controversial,
George Hotz (2:51:51.100)
Infinite Jest, Infinite Jest is a...
Lex Fridman (2:51:57.500)
David Foster Wallace.
George Hotz (2:51:58.860)
Yeah, it's a book about wireheading, really.
Lex Fridman (2:52:03.820)
Very enjoyable to read, very well written.
George Hotz (2:52:07.540)
You know, you will grow as a person reading this book,
Lex Fridman (2:52:11.180)
its effort, and I'll set that up for the second book,
George Hotz (2:52:14.740)
which is pornography, it's called Atlas Shrugged,
Lex Fridman (2:52:17.500)
which...
George Hotz (2:52:21.020)
Atlas Shrugged is pornography.
Lex Fridman (2:52:22.580)
I mean, it is, I will not defend the,
George Hotz (2:52:25.540)
I will not say Atlas Shrugged is a well written book.
Lex Fridman (2:52:28.380)
It is entertaining to read, certainly, just like pornography.
George Hotz (2:52:31.380)
The production value isn't great.
Lex Fridman (2:52:33.540)
You know, there's a 60 page monologue in there
George Hotz (2:52:36.220)
that Ann Rand's editor really wanted to take out.
Lex Fridman (2:52:38.740)
And she paid, she paid out of her pocket
George Hotz (2:52:42.580)
to keep that 60 page monologue in the book.
Lex Fridman (2:52:45.060)
But it is a great book for a kind of framework
George Hotz (2:52:53.220)
of human relations.
Lex Fridman (2:52:54.660)
And I know a lot of people are like,
George Hotz (2:52:55.960)
yeah, but it's a terrible framework.
Lex Fridman (2:52:58.060)
Yeah, but it's a framework.
George Hotz (2:53:00.500)
Just for context, in a couple of days,
Lex Fridman (2:53:02.360)
I'm speaking for probably four plus hours
George Hotz (2:53:06.260)
with Yaron Brook, who's the main living,
Lex Fridman (2:53:10.200)
remaining objectivist, objectivist.
George Hotz (2:53:13.340)
Interesting.
Lex Fridman (2:53:14.580)
So I've always found this philosophy quite interesting
George Hotz (2:53:19.380)
on many levels.
Lex Fridman (2:53:20.260)
One of how repulsive some percent of,
George Hotz (2:53:24.020)
large percent of the population find it,
Lex Fridman (2:53:26.300)
which is always, always funny to me
George Hotz (2:53:29.140)
when people are like unable to even read a philosophy
Lex Fridman (2:53:32.980)
because of some, I think that says more
George Hotz (2:53:36.700)
about their psychological perspective on it.
Lex Fridman (2:53:40.580)
But there is something about objectivism
Lex Fridman (2:53:45.300)
and Ann Rand's philosophy that's deeply connected
Lex Fridman (2:53:48.780)
to this idea of capitalism,
George Hotz (2:53:50.740)
of the ethical life is the productive life
Lex Fridman (2:53:56.620)
that was always compelling to me.
George Hotz (2:54:00.740)
It didn't seem as, like I didn't seem to interpret it
Lex Fridman (2:54:03.160)
in the negative sense that some people do.
George Hotz (2:54:05.660)
To be fair, I read that book when I was 19.
Lex Fridman (2:54:07.980)
So you had an impact at that point, yeah.
George Hotz (2:54:09.620)
Yeah, and the bad guys in the book have this slogan
Lex Fridman (2:54:13.700)
from each according to their ability
George Hotz (2:54:15.340)
to each according to their need.
Lex Fridman (2:54:17.300)
And I'm looking at this and I'm like,
George Hotz (2:54:19.740)
these are the most cart,
Lex Fridman (2:54:20.580)
this is team rocket level cartoonishness, right?
George Hotz (2:54:22.940)
No bad guy.
Lex Fridman (2:54:23.820)
And then when I realized that was actually the slogan
George Hotz (2:54:25.780)
of the communist party, I'm like, wait a second.
Lex Fridman (2:54:29.940)
Wait, no, no, no, no, no.
Lex Fridman (2:54:31.660)
You're telling me this really happened?
Lex Fridman (2:54:34.100)
Yeah, it's interesting.
George Hotz (2:54:34.920)
I mean, one of the criticisms of her work
Lex Fridman (2:54:36.660)
is she has a cartoonish view of good and evil.
George Hotz (2:54:39.220)
Like the reality, as Jordan Peterson says,
Lex Fridman (2:54:44.300)
is that each of us have the capacity for good and evil
George Hotz (2:54:47.420)
in us as opposed to like, there's some characters
Lex Fridman (2:54:49.940)
who are purely evil and some characters that are purely good.
Lex Fridman (2:54:52.220)
And that's in a way why it's pornographic.
Lex Fridman (2:54:55.220)
The production value, I love it.
George Hotz (2:54:57.020)
Like evil is punished and there's very clearly,
Lex Fridman (2:55:01.020)
there's no, just like porn doesn't have character growth.
George Hotz (2:55:06.020)
Well, you know, neither does Alice Shrugged, like.
Lex Fridman (2:55:09.540)
Really, well put.
Lex Fridman (2:55:10.860)
But at 19 year old George Hots, it was good enough.
Lex Fridman (2:55:14.220)
Yeah, yeah, yeah, yeah.
Lex Fridman (2:55:15.380)
What's the third?
Lex Fridman (2:55:16.940)
You have something?
George Hotz (2:55:18.620)
I could give, these two I'll just throw out.
Lex Fridman (2:55:21.520)
They're sci fi.
George Hotz (2:55:22.420)
Perputation City.
Lex Fridman (2:55:24.280)
Great thing to start thinking about copies of yourself.
Lex Fridman (2:55:26.620)
And then the...
Lex Fridman (2:55:27.460)
Who's that by?
George Hotz (2:55:28.280)
Sorry, I didn't catch that.
Lex Fridman (2:55:29.120)
That is Greg Egan.
George Hotz (2:55:31.980)
He's a, that might not be his real name.
Lex Fridman (2:55:33.540)
Some Australian guy, might not be Australian.
George Hotz (2:55:35.740)
I don't know.
Lex Fridman (2:55:36.700)
And then this one's online.
George Hotz (2:55:38.740)
It's called The Metamorphosis of Prime Intellect.
Lex Fridman (2:55:43.020)
It's a story set in a post singularity world.
George Hotz (2:55:45.420)
It's interesting.
Lex Fridman (2:55:46.720)
Is there, can you, either of the worlds,
Lex Fridman (2:55:49.180)
do you find something philosophical interesting in them
Lex Fridman (2:55:51.540)
that you can comment on?
George Hotz (2:55:53.660)
I mean, it is clear to me that
Lex Fridman (2:55:57.860)
Metamorphosis of Prime Intellect is like written by
George Hotz (2:56:00.620)
an engineer, which is,
Lex Fridman (2:56:03.780)
it's very almost a pragmatic take on a utopia, in a way.
Lex Fridman (2:56:12.620)
Positive or negative?
Lex Fridman (2:56:15.260)
That's up to you to decide reading the book.
Lex Fridman (2:56:17.940)
And the ending of it is very interesting as well.
Lex Fridman (2:56:21.580)
And I didn't realize what it was.
George Hotz (2:56:23.660)
I first read that when I was 15.
Lex Fridman (2:56:25.260)
I've reread that book several times in my life.
Lex Fridman (2:56:27.540)
And it's short, it's 50 pages.
Lex Fridman (2:56:29.220)
Everyone should go read it.
George Hotz (2:56:30.740)
What's, sorry, it's a little tangent.
Lex Fridman (2:56:33.100)
I've been working through the foundation.
George Hotz (2:56:34.700)
I've been, I haven't read much sci fi my whole life
Lex Fridman (2:56:37.060)
and I'm trying to fix that the last few months.
George Hotz (2:56:40.180)
That's been a little side project.
Lex Fridman (2:56:42.180)
What's to you as the greatest sci fi novel
Lex Fridman (2:56:46.180)
that people should read?
Lex Fridman (2:56:47.740)
Or is that?
George Hotz (2:56:49.220)
I mean, I would, yeah, I would say like, yeah,
Lex Fridman (2:56:51.180)
Permutation City, Metamorphosis of Prime Intellect.
George Hotz (2:56:53.820)
I don't know.
Lex Fridman (2:56:54.660)
I didn't like Foundation.
George Hotz (2:56:56.240)
I thought it was way too modernist.
Lex Fridman (2:56:58.820)
You like Dune and all of those.
George Hotz (2:57:00.780)
I've never read Dune.
Lex Fridman (2:57:01.780)
I've never read Dune.
George Hotz (2:57:02.820)
I have to read it.
Lex Fridman (2:57:04.580)
Fire Upon the Deep is interesting.
George Hotz (2:57:09.140)
Okay, I mean, look, everyone should read,
Lex Fridman (2:57:10.540)
everyone should read Neuromancer.
George Hotz (2:57:11.380)
Everyone should read Snow Crash.
Lex Fridman (2:57:12.820)
If you haven't read those, like start there.
George Hotz (2:57:15.500)
Yeah, I haven't read Snow Crash.
Lex Fridman (2:57:16.340)
You haven't read Snow Crash?
George Hotz (2:57:17.460)
Oh, it's, I mean, it's very entertaining.
Lex Fridman (2:57:19.980)
Go to Lesher Bach.
Lex Fridman (2:57:20.820)
And if you want the controversial one,
Lex Fridman (2:57:22.220)
Bronze Age Mindset.
George Hotz (2:57:25.340)
All right, I'll look into that one.
Lex Fridman (2:57:27.740)
Those aren't sci fi, but just to round out books.
Lex Fridman (2:57:30.360)
So a bunch of people asked me on Twitter
Lex Fridman (2:57:34.360)
and Reddit and so on for advice.
Lex Fridman (2:57:36.880)
So what advice would you give a young person today
Lex Fridman (2:57:39.440)
about life?
George Hotz (2:57:40.560)
In other words, what, yeah, I mean, looking back,
Lex Fridman (2:57:47.480)
especially when you were younger, you did,
Lex Fridman (2:57:50.480)
and you continued it.
Lex Fridman (2:57:51.560)
You've accomplished a lot of interesting things.
George Hotz (2:57:54.840)
Is there some advice from those,
Lex Fridman (2:57:57.840)
from that life of yours that you can pass on?
George Hotz (2:58:01.880)
If college ever opens again,
Lex Fridman (2:58:03.760)
I would love to give a graduation speech.
George Hotz (2:58:07.640)
At that point, I will put a lot of somewhat satirical effort
Lex Fridman (2:58:11.360)
into this question.
George Hotz (2:58:12.320)
Yeah, at this, you haven't written anything at this point.
Lex Fridman (2:58:15.880)
Oh, you know what?
George Hotz (2:58:16.760)
Always wear sunscreen.
Lex Fridman (2:58:18.240)
This is water.
George Hotz (2:58:19.560)
Pick your plagiarizing.
Lex Fridman (2:58:21.160)
I mean, you know, but that's the,
George Hotz (2:58:23.960)
that's the like clean your room.
Lex Fridman (2:58:26.240)
You know, yeah, you can plagiarize from all of this stuff.
Lex Fridman (2:58:28.600)
And it's, there is no,
Lex Fridman (2:58:35.920)
self help books aren't designed to help you.
George Hotz (2:58:37.680)
They're designed to make you feel good.
Lex Fridman (2:58:40.080)
Like whatever advice I could give, you already know.
George Hotz (2:58:44.120)
Everyone already knows.
Lex Fridman (2:58:45.840)
Sorry, it doesn't feel good.
Lex Fridman (2:58:50.240)
Right?
Lex Fridman (2:58:51.080)
Like, you know, you know,
George Hotz (2:58:53.040)
if I tell you that you should, you know,
Lex Fridman (2:58:56.880)
eat well and read more and it's not gonna do anything.
George Hotz (2:59:01.920)
I think the whole like genre
Lex Fridman (2:59:03.480)
of those kinds of questions is meaningless.
George Hotz (2:59:07.480)
I don't know.
Lex Fridman (2:59:08.320)
If anything, it's don't worry so much about that stuff.
George Hotz (2:59:10.560)
Don't be so caught up in your head.
Lex Fridman (2:59:12.440)
Right.
George Hotz (2:59:13.280)
I mean, you're, yeah.
Lex Fridman (2:59:14.280)
In a sense that your whole life,
George Hotz (2:59:16.560)
your whole existence is like moving version of that advice.
Lex Fridman (2:59:20.680)
I don't know.
George Hotz (2:59:23.840)
There's something, I mean,
Lex Fridman (2:59:25.480)
there's something in you that resists
George Hotz (2:59:27.080)
that kind of thinking and that in itself is,
Lex Fridman (2:59:30.880)
it's just illustrative of who you are.
Lex Fridman (2:59:34.480)
And there's something to learn from that.
Lex Fridman (2:59:36.760)
I think you're clearly not overthinking stuff.
George Hotz (2:59:41.280)
Yeah.
Lex Fridman (2:59:42.120)
And you know what?
George Hotz (2:59:42.960)
There's a gut thing.
Lex Fridman (2:59:43.800)
Even when I talk about my advice,
George Hotz (2:59:45.000)
I'm like, my advice is only relevant to me.
Lex Fridman (2:59:47.400)
It's not relevant to anybody else.
George Hotz (2:59:48.720)
I'm not saying you should go out.
Lex Fridman (2:59:49.960)
If you're the kind of person who overthinks things
George Hotz (2:59:51.640)
to stop overthinking things, it's not bad.
Lex Fridman (2:59:54.080)
It doesn't work for me.
George Hotz (2:59:54.960)
Maybe it works for you.
Lex Fridman (2:59:55.800)
I don't know.
George Hotz (2:59:57.960)
Let me ask you about love.
Lex Fridman (2:59:59.400)
Yeah.
George Hotz (30:00.500)
with the theory of everything.
Lex Fridman (30:02.260)
I mean, that came out.
George Hotz (30:03.580)
He's been working on a theory of everything
Lex Fridman (30:05.460)
in the physics world called geometric unity.
Lex Fridman (30:08.060)
And then for me, from computer science person like you,
Lex Fridman (30:11.700)
Steven Wolfram's theory of everything,
George Hotz (30:14.220)
of like hypergraphs is super interesting and beautiful,
Lex Fridman (30:17.660)
but not from a physics perspective,
Lex Fridman (30:19.460)
but from a computational perspective.
Lex Fridman (30:20.940)
I don't know, have you paid attention to any of that?
Lex Fridman (30:23.020)
So again, like what would make me pay attention
Lex Fridman (30:26.440)
and like why like I hate string theory is,
Lex Fridman (30:29.540)
okay, make a testable prediction, right?
Lex Fridman (30:31.780)
I'm only interested in,
George Hotz (30:33.700)
I'm not interested in theories for their intrinsic beauty.
Lex Fridman (30:36.060)
I'm interested in theories
George Hotz (30:37.060)
that give me power over the universe.
Lex Fridman (30:39.900)
So if these theories do, I'm very interested.
Lex Fridman (30:43.220)
Can I just say how beautiful that is?
Lex Fridman (30:45.100)
Because a lot of physicists say,
George Hotz (30:47.140)
I'm interested in experimental validation
Lex Fridman (30:49.980)
and they skip out the part where they say
George Hotz (30:52.940)
to give me more power in the universe.
Lex Fridman (30:55.500)
I just love the.
George Hotz (30:57.500)
No, I want. The clarity of that.
Lex Fridman (30:59.780)
I want 100 gigahertz processors.
George Hotz (31:02.020)
I want transistors that are smaller than atoms.
Lex Fridman (31:04.120)
I want like power.
George Hotz (31:08.100)
That's true.
Lex Fridman (31:10.580)
And that's where people from aliens
George Hotz (31:12.460)
to this kind of technology where people are worried
Lex Fridman (31:14.420)
that governments, like who owns that power?
Lex Fridman (31:19.320)
Is it George Harts?
Lex Fridman (31:20.780)
Is it thousands of distributed hackers across the world?
Lex Fridman (31:25.020)
Is it governments?
Lex Fridman (31:26.580)
Is it Mark Zuckerberg?
George Hotz (31:28.660)
There's a lot of people that,
Lex Fridman (31:32.500)
I don't know if anyone trusts any one individual with power.
Lex Fridman (31:35.540)
So they're always worried.
Lex Fridman (31:37.380)
It's the beauty of blockchains.
George Hotz (31:39.340)
That's the beauty of blockchains, which we'll talk about.
Lex Fridman (31:43.140)
On Twitter, somebody pointed me to a story,
George Hotz (31:46.220)
a bunch of people pointed me to a story a few months ago
Lex Fridman (31:49.300)
where you went into a restaurant in New York.
Lex Fridman (31:51.660)
And you can correct me if any of this is wrong.
Lex Fridman (31:53.660)
And ran into a bunch of folks from a company
George Hotz (31:56.580)
in a crypto company who are trying to scale up Ethereum.
Lex Fridman (32:01.780)
And they had a technical deadline
George Hotz (32:03.300)
related to solidity to OVM compiler.
Lex Fridman (32:07.380)
So these are all Ethereum technologies.
Lex Fridman (32:09.660)
So you stepped in, they recognized you,
Lex Fridman (32:14.700)
pulled you aside, explained their problem.
Lex Fridman (32:16.220)
And you stepped in and helped them solve the problem,
Lex Fridman (32:19.540)
thereby creating legend status story.
Lex Fridman (32:23.100)
So can you tell me the story in a little more detail?
Lex Fridman (32:28.980)
It seems kind of incredible.
Lex Fridman (32:31.540)
Did this happen?
Lex Fridman (32:32.380)
Yeah, yeah, it's a true story, it's a true story.
George Hotz (32:34.060)
I mean, they wrote a very flattering account of it.
Lex Fridman (32:39.340)
So Optimism is the company called Optimism,
George Hotz (32:43.820)
spin off of Plasma.
Lex Fridman (32:45.420)
They're trying to build L2 solutions on Ethereum.
Lex Fridman (32:47.820)
So right now, every Ethereum node
Lex Fridman (32:52.620)
has to run every transaction on the Ethereum network.
Lex Fridman (32:56.420)
And this kind of doesn't scale, right?
Lex Fridman (32:58.540)
Because if you have N computers,
George Hotz (33:00.020)
well, if that becomes two N computers,
Lex Fridman (33:02.340)
you actually still get the same amount of compute, right?
George Hotz (33:05.420)
This is like all of one scaling
Lex Fridman (33:09.120)
because they all have to run it.
George Hotz (33:10.140)
Okay, fine, you get more blockchain security,
Lex Fridman (33:12.840)
but like, blockchain is already so secure.
Lex Fridman (33:15.380)
Can we trade some of that off for speed?
Lex Fridman (33:17.780)
So that's kind of what these L2 solutions are.
George Hotz (33:20.500)
They built this thing, which kind of,
Lex Fridman (33:23.020)
kind of sandbox for Ethereum contracts.
Lex Fridman (33:26.300)
So they can run it in this L2 world
Lex Fridman (33:28.200)
and it can't do certain things in L world, in L1.
Lex Fridman (33:30.900)
Can I ask you for some definitions?
Lex Fridman (33:32.420)
What's L2?
George Hotz (33:33.420)
Oh, L2 is layer two.
Lex Fridman (33:34.840)
So L1 is like the base Ethereum chain.
Lex Fridman (33:37.180)
And then layer two is like a computational layer
Lex Fridman (33:40.920)
that runs elsewhere,
Lex Fridman (33:44.420)
but still is kind of secured by layer one.
Lex Fridman (33:47.680)
And I'm sure a lot of people know,
Lex Fridman (33:49.720)
but Ethereum is a cryptocurrency,
Lex Fridman (33:51.940)
probably one of the most popular cryptocurrency
George Hotz (33:53.720)
second to Bitcoin.
Lex Fridman (33:55.320)
And a lot of interesting technological innovation there.
George Hotz (33:58.800)
Maybe you could also slip in whenever you talk about this
Lex Fridman (34:03.240)
and things that are exciting to you in the Ethereum space.
Lex Fridman (34:06.380)
And why Ethereum?
Lex Fridman (34:07.800)
Well, I mean, Bitcoin is not Turing complete.
George Hotz (34:12.260)
Ethereum is not technically Turing complete
Lex Fridman (34:13.820)
with the gas limit, but close enough.
Lex Fridman (34:16.040)
With the gas limit?
Lex Fridman (34:16.880)
What's the gas limit, resources?
George Hotz (34:19.080)
Yeah, I mean, no computer is actually Turing complete.
Lex Fridman (34:21.180)
Right.
Lex Fridman (34:23.160)
You're gonna find out RAM, you know?
Lex Fridman (34:24.720)
I can actually solve the whole thing.
Lex Fridman (34:25.560)
What's the word gas limit?
Lex Fridman (34:26.720)
You just have so many brilliant words.
George Hotz (34:28.660)
I'm not even gonna ask.
Lex Fridman (34:29.500)
That's not my word, that's Ethereum's word.
George Hotz (34:32.220)
Gas limit.
Lex Fridman (34:33.060)
Ethereum, you have to spend gas per instruction.
Lex Fridman (34:35.360)
So like different op codes use different amounts of gas
Lex Fridman (34:37.820)
and you buy gas with ether to prevent people
George Hotz (34:40.680)
from basically DDoSing the network.
Lex Fridman (34:42.740)
So Bitcoin is proof of work.
Lex Fridman (34:45.440)
And then what's Ethereum?
Lex Fridman (34:47.120)
It's also proof of work.
George Hotz (34:48.360)
They're working on some proof of stake,
Lex Fridman (34:49.960)
Ethereum 2.0 stuff.
Lex Fridman (34:51.040)
But right now it's proof of work.
Lex Fridman (34:52.720)
It uses a different hash function from Bitcoin.
George Hotz (34:54.700)
That's more ASIC resistance, because you need RAM.
Lex Fridman (34:57.340)
So we're all talking about Ethereum 1.0.
Lex Fridman (34:59.960)
So what were they trying to do to scale this whole process?
Lex Fridman (35:03.840)
So they were like, well, if we could run contracts elsewhere
Lex Fridman (35:07.800)
and then only save the results of that computation,
Lex Fridman (35:13.120)
well, we don't actually have to do the compute on the chain.
George Hotz (35:14.680)
We can do the compute off chain
Lex Fridman (35:15.680)
and just post what the results are.
George Hotz (35:17.440)
Now, the problem with that is,
Lex Fridman (35:18.800)
well, somebody could lie about what the results are.
Lex Fridman (35:21.120)
So you need a resolution mechanism.
Lex Fridman (35:23.240)
And the resolution mechanism can be really expensive
George Hotz (35:26.500)
because you just have to make sure
Lex Fridman (35:29.000)
that the person who is saying,
George Hotz (35:31.140)
look, I swear that this is the real computation.
Lex Fridman (35:33.800)
I'm staking $10,000 on that fact.
Lex Fridman (35:36.640)
And if you prove it wrong,
Lex Fridman (35:39.000)
yeah, it might cost you $3,000 in gas fees to prove wrong,
Lex Fridman (35:42.820)
but you'll get the $10,000 bounty.
Lex Fridman (35:44.800)
So you can secure using those kinds of systems.
Lex Fridman (35:47.740)
So it's effectively a sandbox, which runs contracts.
Lex Fridman (35:52.500)
And like, it's like any kind of normal sandbox,
George Hotz (35:55.600)
you have to like replace syscalls
Lex Fridman (35:57.960)
with calls into the hypervisor.
George Hotz (36:02.920)
Sandbox, syscalls, hypervisor.
Lex Fridman (36:05.080)
What do these things mean?
George Hotz (36:06.680)
As long as it's interesting to talk about.
Lex Fridman (36:09.240)
Yeah, I mean, you can take like the Chrome sandbox
Lex Fridman (36:11.280)
is maybe the one to think about, right?
Lex Fridman (36:12.720)
So the Chrome process that's doing a rendering,
George Hotz (36:16.000)
can't, for example, read a file from the file system.
Lex Fridman (36:18.800)
It has, if it tries to make an open syscall in Linux,
George Hotz (36:21.800)
the open syscall, you can't make it open syscall,
Lex Fridman (36:23.720)
no, no, no.
George Hotz (36:24.780)
You have to request from the kind of hypervisor process
Lex Fridman (36:29.160)
or like, I don't know what it's called in Chrome,
Lex Fridman (36:31.680)
but the, hey, could you open this file for me?
Lex Fridman (36:36.400)
And then it does all these checks
Lex Fridman (36:37.520)
and then it passes the file handle back in
Lex Fridman (36:39.200)
if it's approved.
Lex Fridman (36:41.160)
So that's, yeah.
Lex Fridman (36:42.640)
So what's the, in the context of Ethereum,
Lex Fridman (36:45.320)
what are the boundaries of the sandbox
Lex Fridman (36:47.200)
that we're talking about?
George Hotz (36:48.400)
Well, like one of the calls that you,
Lex Fridman (36:50.800)
actually reading and writing any state
George Hotz (36:53.960)
to the Ethereum contract,
Lex Fridman (36:55.480)
or to the Ethereum blockchain.
George Hotz (36:58.720)
Writing state is one of those calls
Lex Fridman (37:01.060)
that you're going to have to sandbox in layer two,
George Hotz (37:04.500)
because if you let layer two just arbitrarily write
Lex Fridman (37:08.120)
to the Ethereum blockchain.
Lex Fridman (37:09.480)
So layer two is really sitting on top of layer one.
Lex Fridman (37:15.120)
So you're going to have a lot of different kinds of ideas
George Hotz (37:17.160)
that you can play with.
Lex Fridman (37:18.680)
And they're all, they're not fundamentally changing
George Hotz (37:21.360)
the source code level of Ethereum.
Lex Fridman (37:25.140)
Well, you have to replace a bunch of calls
George Hotz (37:28.900)
with calls into the hypervisor.
Lex Fridman (37:31.120)
So instead of doing the syscall directly,
George Hotz (37:33.840)
you replace it with a call to the hypervisor.
Lex Fridman (37:37.360)
So originally they were doing this
George Hotz (37:39.320)
by first running the, so Solidity is the language
Lex Fridman (37:43.520)
that most Ethereum contracts are written in.
George Hotz (37:45.440)
It compiles to a bytecode.
Lex Fridman (37:47.320)
And then they wrote this thing they called the transpiler.
Lex Fridman (37:50.040)
And the transpiler took the bytecode
Lex Fridman (37:52.440)
and it transpiled it into OVM safe bytecode.
George Hotz (37:56.000)
Basically bytecode that didn't make any
Lex Fridman (37:57.520)
of those restricted syscalls
Lex Fridman (37:58.800)
and added the calls to the hypervisor.
Lex Fridman (38:01.680)
This transpiler was a 3000 line mess.
Lex Fridman (38:05.640)
And it's hard to do.
Lex Fridman (38:07.100)
It's hard to do if you're trying to do it like that,
George Hotz (38:09.060)
because you have to kind of like deconstruct the bytecode,
Lex Fridman (38:12.200)
change things about it, and then reconstruct it.
Lex Fridman (38:15.360)
And I mean, as soon as I hear this, I'm like,
Lex Fridman (38:17.760)
well, why don't you just change the compiler, right?
Lex Fridman (38:20.440)
Why not the first place you build the bytecode,
Lex Fridman (38:22.340)
just do it in the compiler.
Lex Fridman (38:25.560)
So yeah, I asked them how much they wanted it.
Lex Fridman (38:29.160)
Of course, measured in dollars and I'm like, well, okay.
Lex Fridman (38:33.040)
And yeah.
Lex Fridman (38:34.600)
And you wrote the compiler.
George Hotz (38:35.920)
Yeah, I modified, I wrote a 300 line diff to the compiler.
Lex Fridman (38:39.040)
It's open source, you can look at it.
George Hotz (38:40.840)
Yeah, it's, yeah, I looked at the code last night.
Lex Fridman (38:43.000)
It's, yeah, exactly.
George Hotz (38:46.680)
Cute is a good word for it.
Lex Fridman (38:49.440)
And it's C++.
George Hotz (38:52.000)
C++, yeah.
Lex Fridman (38:54.320)
So when asked how you were able to do it,
George Hotz (38:57.040)
you said, you just gotta think and then do it right.
Lex Fridman (39:03.080)
So can you break that apart a little bit?
Lex Fridman (39:04.680)
What's your process of one, thinking and two, doing it right?
Lex Fridman (39:09.920)
You know, the people that I was working for
George Hotz (39:12.440)
were amused that I said that.
Lex Fridman (39:13.480)
It doesn't really mean anything.
George Hotz (39:14.800)
Okay.
Lex Fridman (39:16.480)
I mean, is there some deep, profound insights
Lex Fridman (39:19.640)
to draw from like how you problem solve from that?
Lex Fridman (39:23.600)
This is always what I say.
Lex Fridman (39:24.640)
I'm like, do you wanna be a good programmer?
Lex Fridman (39:26.060)
Do it for 20 years.
George Hotz (39:27.840)
Yeah, there's no shortcuts.
Lex Fridman (39:29.600)
No.
Lex Fridman (39:31.360)
What are your thoughts on crypto in general?
Lex Fridman (39:33.440)
So what parts technically or philosophically
Lex Fridman (39:38.080)
do you find especially beautiful maybe?
Lex Fridman (39:40.040)
Oh, I'm extremely bullish on crypto longterm.
George Hotz (39:42.800)
Not any specific crypto project, but this idea of,
Lex Fridman (39:48.680)
well, two ideas.
George Hotz (39:50.320)
One, the Nakamoto Consensus Algorithm
Lex Fridman (39:54.240)
is I think one of the greatest innovations
George Hotz (39:57.320)
of the 21st century.
Lex Fridman (39:58.600)
This idea that people can reach consensus.
George Hotz (3:00:02.240)
I think last time we talked about the meaning of life
Lex Fridman (3:00:05.120)
and it was kind of about winning.
George Hotz (3:00:08.640)
Of course.
Lex Fridman (3:00:10.880)
I don't think I've talked to you about love much,
George Hotz (3:00:13.120)
whether romantic or just love
Lex Fridman (3:00:15.000)
for the common humanity amongst us all.
Lex Fridman (3:00:18.120)
What role has love played in your life?
Lex Fridman (3:00:21.400)
In this quest for winning, where does love fit in?
George Hotz (3:00:26.360)
Well, the word love, I think means several different things.
Lex Fridman (3:00:29.840)
There's love in the sense of, maybe I could just say,
George Hotz (3:00:32.960)
there's like love in the sense of opiates
Lex Fridman (3:00:34.400)
and love in the sense of oxytocin
Lex Fridman (3:00:37.880)
and then love in the sense of,
Lex Fridman (3:00:43.080)
maybe like a love for math.
George Hotz (3:00:44.480)
I don't think it fits into either
Lex Fridman (3:00:45.640)
of those first two paradigms.
Lex Fridman (3:00:49.160)
So each of those, have they given something to you
Lex Fridman (3:00:55.560)
in your life?
George Hotz (3:00:56.920)
I'm not that big of a fan of the first two.
Lex Fridman (3:01:00.800)
Why?
George Hotz (3:01:03.600)
The same reason I'm not a fan of,
Lex Fridman (3:01:06.360)
the same reason I don't do opiates and don't take ecstasy.
Lex Fridman (3:01:09.880)
And there were times, look, I've tried both.
Lex Fridman (3:01:14.200)
I liked opiates way more than I liked ecstasy,
Lex Fridman (3:01:18.400)
but they're not, the ethical life is the productive life.
Lex Fridman (3:01:24.400)
So maybe that's my problem with those.
Lex Fridman (3:01:27.080)
And then like, yeah, a sense of, I don't know,
Lex Fridman (3:01:29.440)
like abstract love for humanity.
George Hotz (3:01:32.200)
I mean, the abstract love for humanity,
Lex Fridman (3:01:34.520)
I'm like, yeah, I've always felt that.
Lex Fridman (3:01:36.200)
And I guess it's hard for me to imagine
Lex Fridman (3:01:39.680)
not feeling it and maybe there's people who don't.
Lex Fridman (3:01:41.560)
And I don't know.
Lex Fridman (3:01:43.560)
Yeah, that's just like a background thing that's there.
George Hotz (3:01:46.520)
I mean, since we brought up drugs, let me ask you,
Lex Fridman (3:01:51.760)
this is becoming more and more a part of my life
George Hotz (3:01:54.000)
because I'm talking to a few researchers
Lex Fridman (3:01:55.520)
that are working on psychedelics.
George Hotz (3:01:57.680)
I've eaten shrooms a couple of times
Lex Fridman (3:02:00.440)
and it was fascinating to me that like the mind can go,
George Hotz (3:02:04.680)
like just fascinating the mind can go to places
Lex Fridman (3:02:08.200)
I didn't imagine it could go.
Lex Fridman (3:02:09.480)
And it was very friendly and positive and exciting
Lex Fridman (3:02:12.720)
and everything was kind of hilarious in the place.
George Hotz (3:02:16.160)
Wherever my mind went, that's where I went.
Lex Fridman (3:02:18.200)
Is, what do you think about psychedelics?
Lex Fridman (3:02:20.960)
Do you think they have, where do you think the mind goes?
Lex Fridman (3:02:24.680)
Have you done psychedelics?
Lex Fridman (3:02:25.920)
Where do you think the mind goes?
Lex Fridman (3:02:28.640)
Is there something useful to learn about the places it goes
Lex Fridman (3:02:32.240)
once you come back?
Lex Fridman (3:02:33.840)
I find it interesting that this idea
George Hotz (3:02:38.040)
that psychedelics have something to teach
Lex Fridman (3:02:40.400)
is almost unique to psychedelics, right?
George Hotz (3:02:43.800)
People don't argue this about amphetamines.
Lex Fridman (3:02:46.560)
And I'm not really sure why.
George Hotz (3:02:50.240)
I think all of the drugs have lessons to teach.
Lex Fridman (3:02:53.800)
I think there's things to learn from opiates.
George Hotz (3:02:55.160)
I think there's things to learn from amphetamines.
Lex Fridman (3:02:56.600)
I think there's things to learn from psychedelics,
George Hotz (3:02:58.160)
things to learn from marijuana.
Lex Fridman (3:03:02.320)
But also at the same time recognize
George Hotz (3:03:05.200)
that I don't think you're learning things about the world.
Lex Fridman (3:03:07.400)
I think you're learning things about yourself.
George Hotz (3:03:09.160)
Yes.
Lex Fridman (3:03:10.680)
And, you know, what's the, even, it might've even been,
George Hotz (3:03:15.760)
might've even been a Timothy Leary quote.
Lex Fridman (3:03:17.240)
I don't wanna misquote him,
Lex Fridman (3:03:18.200)
but the idea is basically like, you know,
Lex Fridman (3:03:20.240)
everybody should look behind the door,
Lex Fridman (3:03:21.600)
but then once you've seen behind the door,
Lex Fridman (3:03:22.760)
you don't need to keep going back.
George Hotz (3:03:26.120)
So, I mean, and that's my thoughts on all real drug use too.
Lex Fridman (3:03:29.960)
Except maybe for caffeine.
George Hotz (3:03:32.680)
It's a little experience that is good to have, but.
Lex Fridman (3:03:37.160)
Oh yeah, no, I mean, yeah, I guess,
George Hotz (3:03:39.240)
yes, psychedelics are definitely.
Lex Fridman (3:03:41.880)
So you're a fan of new experiences, I suppose.
George Hotz (3:03:43.920)
Yes.
Lex Fridman (3:03:44.760)
Because they all contain a little,
George Hotz (3:03:45.880)
especially the first few times,
Lex Fridman (3:03:47.000)
it contains some lessons that can be picked up.
George Hotz (3:03:49.720)
Yeah, and I'll revisit psychedelics maybe once a year.
Lex Fridman (3:03:55.720)
Usually smaller doses.
George Hotz (3:03:58.760)
Maybe they turn up the learning rate of your brain.
Lex Fridman (3:04:01.440)
I've heard that, I like that.
George Hotz (3:04:03.160)
Yeah, that's cool.
Lex Fridman (3:04:04.280)
Big learning rates have pros and cons.
George Hotz (3:04:07.360)
Last question, and this is a little weird one,
Lex Fridman (3:04:09.440)
but you've called yourself crazy in the past.
George Hotz (3:04:14.120)
First of all, on a scale of one to 10,
Lex Fridman (3:04:16.120)
how crazy would you say are you?
George Hotz (3:04:18.040)
Oh, I mean, it depends how you, you know,
Lex Fridman (3:04:19.480)
when you compare me to Elon Musk and Anthony Levandowski,
George Hotz (3:04:21.680)
not so crazy.
Lex Fridman (3:04:23.520)
So like a seven?
George Hotz (3:04:25.920)
Let's go with six.
Lex Fridman (3:04:27.000)
Six, six, six.
Lex Fridman (3:04:29.320)
What?
Lex Fridman (3:04:31.440)
Well, I like seven, seven's a good number.
George Hotz (3:04:32.960)
Seven, all right, well, I'm sure day by day it changes,
Lex Fridman (3:04:36.320)
right, so, but you're in that area.
George Hotz (3:04:42.440)
In thinking about that,
Lex Fridman (3:04:43.440)
what do you think is the role of madness?
George Hotz (3:04:45.760)
Is that a feature or a bug
Lex Fridman (3:04:48.520)
if you were to dissect your brain?
George Hotz (3:04:51.680)
So, okay, from like a mental health lens on crazy,
Lex Fridman (3:04:57.040)
I'm not sure I really believe in that.
George Hotz (3:04:59.040)
I'm not sure I really believe in like a lot of that stuff.
Lex Fridman (3:05:02.840)
Right, this concept of, okay, you know,
George Hotz (3:05:05.080)
when you get over to like hardcore bipolar and schizophrenia,
Lex Fridman (3:05:09.560)
these things are clearly real, somewhat biological.
Lex Fridman (3:05:13.160)
And then over here on the spectrum,
Lex Fridman (3:05:14.480)
you have like ADD and oppositional defiance disorder
Lex Fridman (3:05:18.360)
and these things that are like,
Lex Fridman (3:05:20.760)
wait, this is normal spectrum human behavior.
George Hotz (3:05:22.800)
Like this isn't, you know, where's the line here
Lex Fridman (3:05:28.640)
and why is this like a problem?
Lex Fridman (3:05:31.320)
So there's this whole, you know,
Lex Fridman (3:05:33.280)
the neurodiversity of humanity is huge.
George Hotz (3:05:35.840)
Like people think I'm always on drugs.
Lex Fridman (3:05:37.640)
People are saying this to me on my streams.
Lex Fridman (3:05:38.920)
And I'm like, guys, you know,
Lex Fridman (3:05:39.760)
like I'm real open with my drug use.
George Hotz (3:05:41.320)
I'd tell you if I was on drugs and yeah,
Lex Fridman (3:05:44.080)
I had like a cup of coffee this morning,
Lex Fridman (3:05:45.640)
but other than that, this is just me.
Lex Fridman (3:05:47.240)
You're witnessing my brain in action.
Lex Fridman (3:05:51.400)
So the word madness doesn't even make sense
Lex Fridman (3:05:55.600)
in the rich neurodiversity of humans.
George Hotz (3:05:59.680)
I think it makes sense, but only for like
Lex Fridman (3:06:04.600)
some insane extremes.
George Hotz (3:06:07.040)
Like if you are actually like visibly hallucinating,
Lex Fridman (3:06:11.720)
you know, that's okay.
Lex Fridman (3:06:15.000)
But there is the kind of spectrum on which you stand out.
Lex Fridman (3:06:17.440)
Like that's like, if I were to look, you know,
George Hotz (3:06:22.000)
at decorations on a Christmas tree or something like that,
Lex Fridman (3:06:25.080)
like if you were a decoration, that would catch my eye.
George Hotz (3:06:28.760)
Like that thing is sparkly, whatever the hell that thing is.
Lex Fridman (3:06:35.360)
There's something to that.
George Hotz (3:06:37.360)
Just like refusing to be boring
Lex Fridman (3:06:42.120)
or maybe boring is the wrong word,
Lex Fridman (3:06:43.920)
but to yeah, I mean, be willing to sparkle, you know?
Lex Fridman (3:06:52.440)
It's like somewhat constructed.
George Hotz (3:06:54.200)
I mean, I am who I choose to be.
Lex Fridman (3:06:57.000)
I'm gonna say things as true as I can see them.
George Hotz (3:07:01.000)
I'm not gonna lie.
Lex Fridman (3:07:04.480)
But that's a really important feature in itself.
Lex Fridman (3:07:06.600)
So like whatever the neurodiversity of your,
Lex Fridman (3:07:09.080)
whatever your brain is, not putting constraints on it
George Hotz (3:07:13.800)
that force it to fit into the mold of what society is like,
Lex Fridman (3:07:18.720)
defines what you're supposed to be.
Lex Fridman (3:07:20.640)
So you're one of the specimens
Lex Fridman (3:07:22.360)
that doesn't mind being yourself.
George Hotz (3:07:27.800)
Being right is super important,
Lex Fridman (3:07:31.720)
except at the expense of being wrong.
George Hotz (3:07:37.240)
Without breaking that apart,
Lex Fridman (3:07:38.480)
I think it's a beautiful way to end it.
George Hotz (3:07:40.400)
George, you're one of the most special humans I know.
Lex Fridman (3:07:43.080)
It's truly an honor to talk to you.
George Hotz (3:07:44.600)
Thanks so much for doing it.
Lex Fridman (3:07:45.800)
Thank you for having me.
George Hotz (3:07:47.640)
Thanks for listening to this conversation with George Hotz
Lex Fridman (3:07:50.360)
and thank you to our sponsors,
George Hotz (3:07:52.320)
Four Sigmatic, which is the maker
Lex Fridman (3:07:54.640)
of delicious mushroom coffee,
George Hotz (3:07:57.160)
Decoding Digital, which is a tech podcast
Lex Fridman (3:07:59.880)
that I listen to and enjoy,
Lex Fridman (3:08:02.120)
and ExpressVPN, which is the VPN I've used for many years.
Lex Fridman (3:08:07.080)
Please check out these sponsors in the description
George Hotz (3:08:09.240)
to get a discount and to support this podcast.
Lex Fridman (3:08:13.120)
If you enjoy this thing, subscribe on YouTube,
George Hotz (3:08:15.520)
review it with Five Stars and Apple Podcast,
Lex Fridman (3:08:17.920)
follow on Spotify, support on Patreon,
George Hotz (3:08:20.680)
or connect with me on Twitter at Lex Friedman.
Lex Fridman (3:08:24.480)
And now, let me leave you with some words
George Hotz (3:08:27.040)
from the great and powerful Linus Torvalds.
Lex Fridman (3:08:30.720)
Talk is cheap, show me the code.
George Hotz (3:08:33.120)
Thank you for listening and hope to see you next time.
Lex Fridman (40:01.260)
You can reach a group consensus.
George Hotz (40:03.400)
Using a relatively straightforward algorithm is wild.
Lex Fridman (40:08.220)
And like, you know, Satoshi Nakamoto,
George Hotz (40:14.120)
people always ask me who I look up to.
Lex Fridman (40:15.580)
It's like, whoever that is.
Lex Fridman (40:17.800)
Who do you think it is?
Lex Fridman (40:19.120)
I mean, Elon Musk?
Lex Fridman (40:21.020)
Is it you?
Lex Fridman (40:22.320)
It is definitely not me.
Lex Fridman (40:24.040)
And I do not think it's Elon Musk.
Lex Fridman (40:26.060)
But yeah, this idea of groups reaching consensus
George Hotz (40:31.060)
in a decentralized yet formulaic way
Lex Fridman (40:34.940)
is one extremely powerful idea from crypto.
George Hotz (40:40.540)
Maybe the second idea is this idea of smart contracts.
Lex Fridman (40:45.860)
When you write a contract between two parties,
George Hotz (40:49.100)
any contract, this contract, if there are disputes,
Lex Fridman (40:53.720)
it's interpreted by lawyers.
George Hotz (40:56.140)
Lawyers are just really shitty overpaid interpreters.
Lex Fridman (41:00.020)
Imagine you had, let's talk about them in terms of a,
Lex Fridman (41:02.060)
in terms of like, let's compare a lawyer to Python, right?
Lex Fridman (41:05.260)
So lawyer, well, okay.
George Hotz (41:07.140)
That's really, I never thought of it that way.
Lex Fridman (41:10.020)
It's hilarious.
Lex Fridman (41:11.620)
So Python, I'm paying even 10 cents an hour.
Lex Fridman (41:15.900)
I'll use the nice Azure machine.
George Hotz (41:17.180)
I can run Python for 10 cents an hour.
Lex Fridman (41:19.660)
Lawyers cost $1,000 an hour.
Lex Fridman (41:21.480)
So Python is 10,000 X better on that axis.
Lex Fridman (41:25.900)
Lawyers don't always return the same answer.
George Hotz (41:31.100)
Python almost always does.
Lex Fridman (41:36.700)
Cost.
George Hotz (41:37.740)
Yeah, I mean, just cost, reliability,
Lex Fridman (41:40.940)
everything about Python is so much better than lawyers.
Lex Fridman (41:43.860)
So if you can make smart contracts,
Lex Fridman (41:46.100)
this whole concept of code is law.
George Hotz (41:50.400)
I love, and I would love to live in a world
Lex Fridman (41:53.180)
where everybody accepted that fact.
Lex Fridman (41:55.620)
So maybe you can talk about what smart contracts are.
Lex Fridman (42:01.180)
So let's say, let's say, you know,
George Hotz (42:05.020)
we have a, even something as simple
Lex Fridman (42:08.700)
as a safety deposit box, right?
George Hotz (42:11.700)
Safety deposit box that holds a million dollars.
Lex Fridman (42:14.780)
I have a contract with the bank that says
George Hotz (42:17.040)
two out of these three parties must be present
Lex Fridman (42:22.620)
to open the safety deposit box and get the money out.
Lex Fridman (42:25.280)
So that's a contract for the bank,
Lex Fridman (42:26.300)
and it's only as good as the bank and the lawyers, right?
George Hotz (42:29.020)
Let's say, you know, somebody dies and now,
Lex Fridman (42:32.820)
oh, we're gonna go through a big legal dispute
George Hotz (42:34.580)
about whether, oh, well, was it in the will,
Lex Fridman (42:36.140)
was it not in the will?
George Hotz (42:37.540)
What, like, it's just so messy,
Lex Fridman (42:39.780)
and the cost to determine truth is so expensive.
George Hotz (42:44.620)
Versus a smart contract, which just uses cryptography
Lex Fridman (42:47.220)
to check if two out of three keys are present.
George Hotz (42:50.100)
Well, I can look at that, and I can have certainty
Lex Fridman (42:53.860)
in the answer that it's going to return.
Lex Fridman (42:55.980)
And that's what, all businesses want is certainty.
Lex Fridman (42:58.220)
You know, they say businesses don't care.
George Hotz (42:59.620)
Viacom, YouTube, YouTube's like,
Lex Fridman (43:02.020)
look, we don't care which way this lawsuit goes.
George Hotz (43:04.100)
Just please tell us so we can have certainty.
Lex Fridman (43:07.220)
Yeah, I wonder how many agreements in this,
George Hotz (43:09.100)
because we're talking about financial transactions only
Lex Fridman (43:12.720)
in this case, correct, the smart contracts?
George Hotz (43:15.620)
Oh, you can go to anything.
Lex Fridman (43:17.380)
You can put a prenup in the Ethereum blockchain.
Lex Fridman (43:21.820)
A married smart contract?
Lex Fridman (43:23.020)
Sorry, divorce lawyer, sorry.
George Hotz (43:24.900)
You're going to be replaced by Python.
Lex Fridman (43:29.580)
Okay, so that's another beautiful idea.
Lex Fridman (43:34.980)
Do you think there's something that's appealing to you
Lex Fridman (43:37.700)
about any one specific implementation?
Lex Fridman (43:40.260)
So if you look 10, 20, 50 years down the line,
Lex Fridman (43:45.100)
do you see any, like, Bitcoin, Ethereum,
Lex Fridman (43:48.140)
any of the other hundreds of cryptocurrencies winning out?
Lex Fridman (43:51.140)
Is there, like, what's your intuition about the space?
George Hotz (43:53.340)
Or are you just sitting back and watching the chaos
Lex Fridman (43:55.380)
and look who cares what emerges?
George Hotz (43:57.380)
Oh, I don't.
Lex Fridman (43:58.220)
I don't speculate.
George Hotz (43:59.060)
I don't really care.
Lex Fridman (43:59.900)
I don't really care which one of these projects wins.
George Hotz (44:02.620)
I'm kind of in the Bitcoin as a meme coin camp.
Lex Fridman (44:05.500)
I mean, why does Bitcoin have value?
George Hotz (44:07.000)
It's technically kind of, you know,
Lex Fridman (44:11.740)
not great, like the block size debate.
George Hotz (44:14.380)
When I found out what the block size debate was,
Lex Fridman (44:16.140)
I'm like, are you guys kidding?
Lex Fridman (44:18.020)
What's the block size debate?
Lex Fridman (44:21.180)
You know what?
George Hotz (44:22.020)
It's really, it's too stupid to even talk.
Lex Fridman (44:23.820)
People can look it up, but I'm like, wow.
George Hotz (44:27.140)
You know, Ethereum seems,
Lex Fridman (44:28.420)
the governance of Ethereum seems much better.
George Hotz (44:31.300)
I've come around a bit on proof of stake ideas.
Lex Fridman (44:35.360)
You know, very smart people thinking about some things.
George Hotz (44:37.740)
Yeah, you know, governance is interesting.
Lex Fridman (44:40.340)
It does feel like Vitalik,
George Hotz (44:44.100)
like it does feel like an open,
Lex Fridman (44:46.100)
even in these distributed systems,
George Hotz (44:48.020)
leaders are helpful
Lex Fridman (44:51.220)
because they kind of help you drive the mission
Lex Fridman (44:54.580)
and the vision and they put a face to a project.
Lex Fridman (44:58.420)
It's a weird thing about us humans.
George Hotz (45:00.140)
Geniuses are helpful, like Vitalik.
Lex Fridman (45:02.820)
Yeah, brilliant.
George Hotz (45:06.540)
Leaders are not necessarily, yeah.
Lex Fridman (45:10.320)
So you think the reason he's the face of Ethereum
George Hotz (45:15.340)
is because he's a genius.
Lex Fridman (45:17.100)
That's interesting.
George Hotz (45:18.060)
I mean, that was,
Lex Fridman (45:21.460)
it's interesting to think about
George Hotz (45:22.940)
that we need to create systems
Lex Fridman (45:25.860)
in which the quote unquote leaders that emerge
George Hotz (45:30.340)
are the geniuses in the system.
Lex Fridman (45:33.060)
I mean, that's arguably why
George Hotz (45:35.000)
the current state of democracy is broken
Lex Fridman (45:36.980)
is the people who are emerging as the leaders
George Hotz (45:39.420)
are not the most competent,
Lex Fridman (45:40.980)
are not the superstars of the system.
Lex Fridman (45:43.340)
And it seems like at least for now
Lex Fridman (45:45.020)
in the crypto world oftentimes
George Hotz (45:47.220)
the leaders are the superstars.
Lex Fridman (45:49.380)
Imagine at the debate they asked,
Lex Fridman (45:51.720)
what's the sixth amendment?
Lex Fridman (45:53.640)
What are the four fundamental forces in the universe?
Lex Fridman (45:56.300)
What's the integral of two to the X?
Lex Fridman (45:59.940)
I'd love to see those questions asked
Lex Fridman (46:01.740)
and that's what I want as our leader.
Lex Fridman (46:03.380)
It's a little bit.
Lex Fridman (46:04.220)
What's Bayes rule?
Lex Fridman (46:07.160)
Yeah, I mean, even, oh wow, you're hurting my brain.
George Hotz (46:10.900)
It's that my standard was even lower
Lex Fridman (46:15.020)
but I would have loved to see
George Hotz (46:17.660)
just this basic brilliance.
Lex Fridman (46:20.600)
Like I've talked to historians.
George Hotz (46:22.180)
There's just these, they're not even like
Lex Fridman (46:23.900)
they don't have a PhD or even education history.
George Hotz (46:26.720)
They just like a Dan Carlin type character
Lex Fridman (46:30.040)
who just like, holy shit.
Lex Fridman (46:32.780)
How did all this information get into your head?
Lex Fridman (46:35.420)
They're able to just connect Genghis Khan
George Hotz (46:38.460)
to the entirety of the history of the 20th century.
Lex Fridman (46:41.820)
They know everything about every single battle that happened
Lex Fridman (46:46.180)
and they know the game of Thrones
Lex Fridman (46:51.020)
of the different power plays and all that happened there.
Lex Fridman (46:55.100)
And they know like the individuals
Lex Fridman (46:56.700)
and all the documents involved
Lex Fridman (46:58.580)
and they integrate that into their regular life.
Lex Fridman (47:02.060)
It's not like they're ultra history nerds.
George Hotz (47:03.980)
They're just, they know this information.
Lex Fridman (47:06.380)
That's what competence looks like.
George Hotz (47:08.020)
Yeah.
Lex Fridman (47:09.060)
Cause I've seen that with programmers too, right?
George Hotz (47:10.660)
That's what great programmers do.
Lex Fridman (47:12.580)
But yeah, it would be, it's really unfortunate
George Hotz (47:15.780)
that those kinds of people aren't emerging as our leaders.
Lex Fridman (47:19.340)
But for now, at least in the crypto world
George Hotz (47:21.820)
that seems to be the case.
Lex Fridman (47:23.260)
I don't know if that always, you could imagine
Lex Fridman (47:26.700)
that in a hundred years, it's not the case, right?
Lex Fridman (47:28.820)
Crypto world has one very powerful idea going for it
Lex Fridman (47:31.940)
and that's the idea of forks, right?
Lex Fridman (47:35.020)
I mean, imagine, we'll use a less controversial example.
George Hotz (47:42.940)
This was actually in my joke app in 2012.
Lex Fridman (47:47.140)
I was like, Barack Obama, Mitt Romney,
Lex Fridman (47:49.460)
let's let them both be president, right?
Lex Fridman (47:51.060)
Like imagine we could fork America
Lex Fridman (47:52.940)
and just let them both be president.
Lex Fridman (47:54.500)
And then the Americas could compete
Lex Fridman (47:56.060)
and people could invest in one,
Lex Fridman (47:58.180)
pull their liquidity out of one, put it in the other.
George Hotz (48:00.500)
You have this in the crypto world.
Lex Fridman (48:02.420)
Ethereum forks into Ethereum and Ethereum classic.
Lex Fridman (48:05.540)
And you can pull your liquidity out of one
Lex Fridman (48:07.340)
and put it in another.
Lex Fridman (48:08.980)
And people vote with their dollars,
Lex Fridman (48:11.980)
which forks, companies should be able to fork.
Lex Fridman (48:16.420)
I'd love to fork Nvidia, you know?
Lex Fridman (48:20.260)
Yeah, like different business strategies
Lex Fridman (48:22.780)
and then try them out and see what works.
Lex Fridman (48:26.180)
Like even take, yeah, take comma AI that closes its source
Lex Fridman (48:34.780)
and then take one that's open source and see what works.
Lex Fridman (48:38.300)
Take one that's purchased by GM
Lex Fridman (48:41.060)
and one that remains Android Renegade
Lex Fridman (48:43.540)
and all these different versions and see.
George Hotz (48:45.300)
The beauty of comma AI is someone can actually do that.
Lex Fridman (48:47.900)
Please take comma AI and fork it.
George Hotz (48:50.620)
That's right, that's the beauty of open source.
Lex Fridman (48:53.020)
So you're, I mean, we'll talk about autonomous vehicle space,
Lex Fridman (48:56.420)
but it does seem that you're really knowledgeable
Lex Fridman (49:02.100)
about a lot of different topics.
Lex Fridman (49:03.940)
So the natural question a bunch of people ask this,
Lex Fridman (49:06.100)
which is how do you keep learning new things?
Lex Fridman (49:09.980)
Do you have like practical advice
Lex Fridman (49:12.420)
if you were to introspect, like taking notes,
George Hotz (49:15.820)
allocate time, or do you just mess around
Lex Fridman (49:19.220)
and just allow your curiosity to drive?
George Hotz (49:21.340)
I'll write these people a self help book
Lex Fridman (49:23.060)
and I'll charge $67 for it.
Lex Fridman (49:25.060)
And I will write, I will write,
Lex Fridman (49:28.500)
I will write on the cover of the self help book.
George Hotz (49:30.300)
All of this advice is completely meaningless.
Lex Fridman (49:32.500)
You're gonna be a sucker and buy this book anyway.
Lex Fridman (49:34.740)
And the one lesson that I hope they take away from the book
Lex Fridman (49:38.860)
is that I can't give you a meaningful answer to that.
George Hotz (49:42.580)
That's interesting.
Lex Fridman (49:44.020)
Let me translate that.
George Hotz (49:45.420)
Is you haven't really thought about what it is you do
Lex Fridman (49:51.580)
systematically because you could reduce it.
Lex Fridman (49:53.900)
And there's some people, I mean, I've met brilliant people
Lex Fridman (49:56.980)
that this is really clear with athletes.
George Hotz (50:00.180)
Some are just, you know, the best in the world
Lex Fridman (50:03.580)
at something and they have zero interest
George Hotz (50:06.820)
in writing like a self help book,
Lex Fridman (50:09.500)
or how to master this game.
Lex Fridman (50:11.620)
And then there's some athletes who become great coaches
Lex Fridman (50:15.620)
and they love the analysis, perhaps the over analysis.
Lex Fridman (50:18.820)
And you right now, at least at your age,
Lex Fridman (50:20.740)
which is an interesting, you're in the middle of the battle.
George Hotz (50:23.180)
You're like the warriors that have zero interest
Lex Fridman (50:25.620)
in writing books.
Lex Fridman (50:27.820)
So you're in the middle of the battle.
Lex Fridman (50:29.140)
So you have, yeah.
George Hotz (50:30.580)
This is a fair point.
Lex Fridman (50:31.740)
I do think I have a certain aversion
George Hotz (50:34.060)
to this kind of deliberate intentional way of living life.
Lex Fridman (50:39.060)
You're eventually, the hilarity of this,
George Hotz (50:41.820)
especially since this is recorded,
Lex Fridman (50:43.620)
it will reveal beautifully the absurdity
George Hotz (50:47.620)
when you finally do publish this book.
Lex Fridman (50:49.620)
I guarantee you, you will.
George Hotz (50:51.540)
The story of comma AI, maybe it'll be a biography
Lex Fridman (50:56.060)
written about you.
George Hotz (50:57.060)
That'll be better, I guess.
Lex Fridman (50:58.740)
And you might be able to learn some cute lessons
George Hotz (51:00.740)
if you're starting a company like comma AI from that book.
Lex Fridman (51:03.700)
But if you're asking generic questions,
Lex Fridman (51:05.660)
like how do I be good at things?
Lex Fridman (51:07.740)
How do I be good at things?
George Hotz (51:10.260)
Dude, I don't know.
Lex Fridman (51:11.900)
Do them a lot.
George Hotz (51:14.380)
Do them a lot.
Lex Fridman (51:15.220)
But the interesting thing here is learning things
George Hotz (51:18.580)
outside of your current trajectory,
Lex Fridman (51:22.020)
which is what it feels like from an outsider's perspective.
George Hotz (51:28.140)
I don't know if there's advice on that,
Lex Fridman (51:30.900)
but it is an interesting curiosity.
George Hotz (51:33.340)
When you become really busy, you're running a company.
Lex Fridman (51:38.020)
Hard time.
George Hotz (51:40.340)
Yeah.
Lex Fridman (51:41.580)
But there's a natural inclination and trend.
George Hotz (51:46.100)
Just the momentum of life carries you
Lex Fridman (51:48.980)
into a particular direction of wanting to focus.
Lex Fridman (51:51.060)
And this kind of dispersion that curiosity can lead to
Lex Fridman (51:55.140)
gets harder and harder with time.
George Hotz (51:58.220)
Because you get really good at certain things
Lex Fridman (52:00.940)
and it sucks trying things that you're not good at,
George Hotz (52:03.820)
like trying to figure them out.
Lex Fridman (52:05.340)
When you do this with your live streams,
George Hotz (52:07.380)
you're on the fly figuring stuff out.
Lex Fridman (52:10.060)
You don't mind looking dumb.
George Hotz (52:11.780)
No.
Lex Fridman (52:14.140)
You just figure it out pretty quickly.
George Hotz (52:16.660)
Sometimes I try things and I don't figure them out quickly.
Lex Fridman (52:19.180)
My chest rating is like a 1400,
George Hotz (52:20.980)
despite putting like a couple of hundred hours in.
Lex Fridman (52:23.300)
It's pathetic.
George Hotz (52:24.340)
I mean, to be fair, I know that I could do it better
Lex Fridman (52:26.860)
if I did it better.
George Hotz (52:27.700)
Like don't play five minute games,
Lex Fridman (52:29.860)
play 15 minute games at least.
George Hotz (52:31.380)
Like I know these things, but it just doesn't,
Lex Fridman (52:34.260)
it doesn't stick nicely in my knowledge stream.
George Hotz (52:37.260)
All right, let's talk about Comma AI.
Lex Fridman (52:39.300)
What's the mission of the company?
George Hotz (52:42.140)
Let's like look at the biggest picture.
Lex Fridman (52:44.860)
Oh, I have an exact statement.
George Hotz (52:46.820)
Solve self driving cars
Lex Fridman (52:48.660)
while delivering shippable intermediaries.
Lex Fridman (52:51.540)
So longterm vision is have fully autonomous vehicles
Lex Fridman (52:56.300)
and make sure you're making money along the way.
George Hotz (52:59.060)
I think it doesn't really speak to money,
Lex Fridman (53:00.220)
but I can talk about what solve self driving cars means.
George Hotz (53:03.260)
Solve self driving cars of course means
Lex Fridman (53:06.700)
you're not building a new car,
George Hotz (53:08.020)
you're building a person replacement.
Lex Fridman (53:10.460)
That person can sit in the driver's seat
Lex Fridman (53:12.340)
and drive you anywhere a person can drive
Lex Fridman (53:14.580)
with a human or better level of safety,
George Hotz (53:17.980)
speed, quality, comfort.
Lex Fridman (53:21.420)
And what's the second part of that?
George Hotz (53:23.260)
Delivering shippable intermediaries is well,
Lex Fridman (53:26.620)
it's a way to fund the company, that's true.
Lex Fridman (53:28.180)
But it's also a way to keep us honest.
Lex Fridman (53:31.980)
If you don't have that, it is very easy
George Hotz (53:34.900)
with this technology to think you're making progress
Lex Fridman (53:39.020)
when you're not.
George Hotz (53:40.180)
I've heard it best described on Hacker News as
Lex Fridman (53:43.420)
you can set any arbitrary milestone,
George Hotz (53:46.660)
meet that milestone and still be infinitely far away
Lex Fridman (53:49.500)
from solving self driving cars.
Lex Fridman (53:51.740)
So it's hard to have like real deadlines
Lex Fridman (53:53.820)
when you're like Cruz or Waymo when you don't have revenue.
George Hotz (54:02.100)
Is that, I mean, is revenue essentially
Lex Fridman (54:06.060)
the thing we're talking about here?
George Hotz (54:07.820)
Revenue is, capitalism is based around consent.
Lex Fridman (54:11.460)
Capitalism, the way that you get revenue
George Hotz (54:13.100)
is real capitalism comes in the real capitalism camp.
Lex Fridman (54:16.340)
There's definitely scams out there,
Lex Fridman (54:17.340)
but real capitalism is based around consent.
Lex Fridman (54:19.580)
It's based around this idea that like,
George Hotz (54:20.740)
if we're getting revenue, it's because we're providing
Lex Fridman (54:22.980)
at least that much value to another person.
George Hotz (54:24.780)
When someone buys $1,000 comma two from us,
Lex Fridman (54:27.460)
we're providing them at least $1,000 of value
George Hotz (54:29.100)
or they wouldn't buy it.
Lex Fridman (54:30.180)
Brilliant, so can you give a whirlwind overview
George Hotz (54:32.940)
of the products that Comma AI provides,
Lex Fridman (54:34.940)
like throughout its history and today?
George Hotz (54:38.180)
I mean, yeah, the past ones aren't really that interesting.
Lex Fridman (54:40.780)
It's kind of just been refinement of the same idea.
George Hotz (54:45.180)
The real only product we sell today is the Comma 2.
Lex Fridman (54:48.300)
Which is a piece of hardware with cameras.
George Hotz (54:50.500)
Mm, so the Comma 2, I mean, you can think about it
Lex Fridman (54:54.700)
kind of like a person.
George Hotz (54:57.100)
Future hardware will probably be
Lex Fridman (54:58.340)
even more and more personlike.
Lex Fridman (55:00.300)
So it has eyes, ears, a mouth, a brain,
Lex Fridman (55:07.460)
and a way to interface with the car.
Lex Fridman (55:09.540)
Does it have consciousness?
Lex Fridman (55:10.980)
Just kidding, that was a trick question.
George Hotz (55:13.500)
I don't have consciousness either.
Lex Fridman (55:15.140)
Me and the Comma 2 are the same.
Lex Fridman (55:16.420)
You're the same?
Lex Fridman (55:17.260)
I have a little more compute than it.
George Hotz (55:18.660)
It only has like the same compute as a B, you know.
Lex Fridman (55:23.260)
You're more efficient energy wise
George Hotz (55:25.300)
for the compute you're doing.
Lex Fridman (55:26.460)
Far more efficient energy wise.
George Hotz (55:29.020)
20 petaflops, 20 watts, crazy.
Lex Fridman (55:30.580)
Do you lack consciousness?
George Hotz (55:32.220)
Sure.
Lex Fridman (55:33.060)
Do you fear death?
George Hotz (55:33.900)
You do, you want immortality.
Lex Fridman (55:35.500)
Of course I fear death.
Lex Fridman (55:36.340)
Does Comma AI fear death?
Lex Fridman (55:38.100)
I don't think so.
George Hotz (55:39.580)
Of course it does.
Lex Fridman (55:40.500)
It very much fears, well, it fears negative loss.
George Hotz (55:42.980)
Oh yeah.
Lex Fridman (55:43.820)
Okay, so Comma 2, when did that come out?
Lex Fridman (55:49.220)
That was a year ago?
Lex Fridman (55:50.500)
No, two.
George Hotz (55:52.100)
Early this year.
Lex Fridman (55:53.540)
Wow, time, it feels like, yeah.
George Hotz (55:56.900)
2020 feels like it's taken 10 years to get to the end.
Lex Fridman (56:00.820)
It's a long year.
George Hotz (56:01.820)
It's a long year.
Lex Fridman (56:03.140)
So what's the sexiest thing about Comma 2 feature wise?
Lex Fridman (56:08.140)
So, I mean, maybe you can also link on like, what is it?
Lex Fridman (56:14.340)
Like what's its purpose?
George Hotz (56:15.740)
Cause there's a hardware, there's a software component.
Lex Fridman (56:18.540)
You've mentioned the sensors,
Lex Fridman (56:20.020)
but also like what is it, its features and capabilities?
Lex Fridman (56:23.060)
I think our slogan summarizes it well.
George Hotz (56:25.340)
Comma slogan is make driving chill.
Lex Fridman (56:28.860)
I love it, okay.
George Hotz (56:30.660)
Yeah, I mean, it is, you know,
Lex Fridman (56:33.020)
if you like cruise control, imagine cruise control,
Lex Fridman (56:35.340)
but much, much more.
Lex Fridman (56:36.660)
So it can do adaptive cruise control things,
George Hotz (56:41.020)
which is like slow down for cars in front of it,
Lex Fridman (56:42.980)
maintain a certain speed.
Lex Fridman (56:44.220)
And it can also do lane keeping.
Lex Fridman (56:46.260)
So staying in the lane and doing it better
Lex Fridman (56:48.700)
and better and better over time.
Lex Fridman (56:50.020)
It's very much machine learning based.
Lex Fridman (56:53.060)
So this camera is, there's a driver facing camera too.
Lex Fridman (57:01.100)
What else is there?
Lex Fridman (57:02.100)
What am I thinking?
Lex Fridman (57:02.940)
So the hardware versus software.
Lex Fridman (57:04.380)
So open pilot versus the actual hardware of the device.
Lex Fridman (57:09.020)
What's, can you draw that distinction?
Lex Fridman (57:10.580)
What's one, what's the other?
Lex Fridman (57:11.580)
I mean, the hardware is pretty much a cell phone
George Hotz (57:13.820)
with a few additions.
Lex Fridman (57:14.700)
A cell phone with a cooling system
Lex Fridman (57:16.940)
and with a car interface connected to it.
Lex Fridman (57:20.980)
And by cell phone, you mean like Qualcomm Snapdragon.
George Hotz (57:25.620)
Yeah, the current hardware is a Snapdragon 821.
Lex Fridman (57:29.380)
It has wifi radio, it has an LTE radio, it has a screen.
George Hotz (57:32.180)
We use every part of the cell phone.
Lex Fridman (57:35.980)
And then the interface with the car
George Hotz (57:37.540)
is specific to the car.
Lex Fridman (57:38.460)
So you keep supporting more and more cars.
George Hotz (57:41.460)
Yeah, so the interface to the car,
Lex Fridman (57:42.660)
I mean, the device itself just has four CAN buses.
George Hotz (57:45.340)
It has four CAN interfaces on it
Lex Fridman (57:46.860)
that are connected through the USB port to the phone.
Lex Fridman (57:49.420)
And then, yeah, on those four CAN buses,
Lex Fridman (57:53.300)
you connect it to the car.
Lex Fridman (57:54.460)
And there's a little harness to do this.
Lex Fridman (57:56.460)
Cars are actually surprisingly similar.
Lex Fridman (57:58.460)
So CAN is the protocol by which cars communicate.
Lex Fridman (58:01.500)
And then you're able to read stuff and write stuff
George Hotz (58:04.260)
to be able to control the car depending on the car.
Lex Fridman (58:06.900)
So what's the software side?
Lex Fridman (58:08.260)
What's OpenPilot?
Lex Fridman (58:10.380)
So I mean, OpenPilot is,
George Hotz (58:11.740)
the hardware is pretty simple compared to OpenPilot.
Lex Fridman (58:13.740)
OpenPilot is, well, so you have a machine learning model,
George Hotz (58:21.020)
which it's in OpenPilot, it's a blob.
Lex Fridman (58:24.540)
It's just a blob of weights.
George Hotz (58:25.620)
It's not like people are like, oh, it's closed source.
Lex Fridman (58:27.420)
I'm like, it's a blob of weights.
Lex Fridman (58:28.860)
What do you expect?
Lex Fridman (58:29.780)
So it's primarily neural network based.
George Hotz (58:33.380)
You, well, OpenPilot is all the software
Lex Fridman (58:36.020)
kind of around that neural network.
George Hotz (58:37.660)
That if you have a neural network that says,
Lex Fridman (58:39.060)
here's where you wanna send the car,
George Hotz (58:40.860)
OpenPilot actually goes and executes all of that.
Lex Fridman (58:44.780)
It cleans up the input to the neural network.
George Hotz (58:46.900)
It cleans up the output and executes on it.
Lex Fridman (58:49.060)
So it connects, it's the glue
George Hotz (58:50.860)
that connects everything together.
Lex Fridman (58:51.860)
Runs the sensors, does a bunch of calibration
George Hotz (58:54.420)
for the neural network, deals with like,
Lex Fridman (58:58.100)
if the car is on a banked road,
George Hotz (59:00.140)
you have to counter steer against that.
Lex Fridman (59:02.060)
And the neural network can't necessarily know that
George Hotz (59:03.820)
by looking at the picture.
Lex Fridman (59:06.420)
So you do that with other sensors
Lex Fridman (59:08.100)
and Fusion and Localizer.
Lex Fridman (59:09.900)
OpenPilot also is responsible
George Hotz (59:11.780)
for sending the data up to our servers.
Lex Fridman (59:14.780)
So we can learn from it, logging it, recording it,
George Hotz (59:17.940)
running the cameras, thermally managing the device,
Lex Fridman (59:21.500)
managing the disk space on the device,
George Hotz (59:23.180)
managing all the resources on the device.
Lex Fridman (59:24.780)
So what, since we last spoke,
George Hotz (59:26.860)
I don't remember when, maybe a year ago,
Lex Fridman (59:28.460)
maybe a little bit longer,
Lex Fridman (59:30.180)
how has OpenPilot improved?
Lex Fridman (59:33.100)
We did exactly what I promised you.
George Hotz (59:34.860)
I promised you that by the end of the year,
Lex Fridman (59:36.760)
where you'd be able to remove the lanes.
George Hotz (59:40.500)
The lateral policy is now almost completely end to end.
Lex Fridman (59:46.060)
You can turn the lanes off and it will drive,
George Hotz (59:48.600)
drive slightly worse on the highway
Lex Fridman (59:49.980)
if you turn the lanes off,
Lex Fridman (59:51.060)
but you can turn the lanes off and it will drive well,
Lex Fridman (59:54.260)
trained completely end to end on user data.
Lex Fridman (59:57.220)
And this year we hope to do the same
Lex Fridman (59:58.700)
for the longitudinal policy.
🔗 相关节目