George Hotz: Comma.ai, OpenPilot, and Autonomous Vehicles
技术与编程AI 与机器学习心理与人性政治与社会商业与创业
🤖
AI 智能总结
乔治·霍兹谈Comma.ai、自动驾驶与黑客精神
这是 Lex Fridman 与传奇黑客、Comma.ai 创始人 George Hotz 的对话。Hotz 分享了他对自动驾驶的独特方法、对 AI 行业的犀利批评，以及他作为独立黑客对抗大公司的哲学。
自动驾驶Comma.aiOpenPilot黑客文化AI工程特斯拉
George Hotz（geohot）是美国传奇黑客，以第一个破解 iPhone 和 PlayStation 3 而闻名，后创立 Comma.ai，开发开源自动驾驶系统 OpenPilot，是硅谷最具个性的技术人物之一。
📌 核心观点
- Comma.ai 的方法论：Hotz 认为自动驾驶不需要高精地图、激光雷达或复杂的规则系统，只需要端到端的深度学习——就像人类驾驶一样，只用眼睛和大脑。OpenPilot 是这一理念的实践。
- 对特斯拉 Autopilot 的看法：Hotz 认为特斯拉的方向是正确的（纯视觉方案），但他们的执行速度太慢，官僚主义阻碍了创新。他相信小团队可以比大公司更快地解决问题。
- 黑客精神与独立性：Hotz 是第一个破解 iPhone 和 PlayStation 3 的人，他将这种黑客精神带入了自动驾驶领域。他认为真正的创新来自不受约束的个人，而非大型组织。
- 对 AI 行业的批评：Hotz 对 AI 安全研究持怀疑态度，认为很多 AI 安全担忧是夸大的，是为了阻止竞争对手。他更关注实际的工程问题而非假设性的存在性风险。
- 关于 AGI：Hotz 认为 AGI 会到来，但时间线比大多数人预测的要长。他认为当前的 LLM 是强大的工具，但距离真正的通用智能还有很长的路要走。
✨ 金句摘录
Hotz：自动驾驶不需要高精地图或激光雷达——人类只用眼睛就能开车，AI 也应该能做到。
Hotz：我是第一个破解 iPhone 的人，我相信小团队可以比大公司更快地解决问题。
Hotz：很多 AI 安全担忧是夸大的——它们更多是为了阻止竞争对手，而非真正的安全考虑。
📋 章节目录
暂无章节信息
🔑 关键词
donlanecardriverdrivingdatahumanteslagoingcarsrealdoingsayingwaymomonitoringputlearningbettercodestuff
💬 精彩语录
暂无语录
🎙️ 完整对话（3069 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 (00:02.480)
He's the founder of Kama AI,
他是Kama AI的创始人，
Lex Fridman (00:04.440)
a machine learning based vehicle automation company.
一家基于机器学习的车辆自动化公司。
Lex Fridman (00:07.360)
He is most certainly an outspoken personality
他无疑是一个直言不讳的人
Lex Fridman (00:10.160)
in the field of AI and technology in general.
在人工智能和一般技术领域。
George Hotz (00:13.120)
He first gained recognition for being the first person
他首先因成为第一人而获得认可
Lex Fridman (00:16.200)
to carry or unlock an iPhone.
携带或解锁 iPhone。
Lex Fridman (00:18.360)
And since then, he's done quite a few interesting things
从那时起，他做了很多有趣的事情
Lex Fridman (00:21.240)
at the intersection of hardware and software.
处于硬件和软件的交叉点。
George Hotz (00:24.360)
This is the Artificial Intelligence Podcast.
这是人工智能播客。
Lex Fridman (00:27.400)
If you enjoy it, subscribe on YouTube,
如果您喜欢，请在 YouTube 上订阅，
George Hotz (00:29.560)
give it five stars on iTunes, support it on Patreon,
在 iTunes 上给它五颗星，在 Patreon 上支持它，
Lex Fridman (00:32.880)
or simply connect with me on Twitter
或者直接在 Twitter 上与我联系
George Hotz (00:34.880)
at Lex Friedman, spelled F R I D M A N.
在 Lex Friedman，拼写为 F R I D M A N。
Lex Fridman (00:39.120)
And I'd like to give a special thank you
我想特别感谢你
George Hotz (00:40.960)
to Jennifer from Canada
詹妮弗（来自加拿大）
Lex Fridman (00:43.240)
for her support of the podcast on Patreon.
以表彰她对 Patreon 播客的支持。
George Hotz (00:45.840)
Merci beaucoup, Jennifer.
谢谢，詹妮弗。
Lex Fridman (00:47.680)
She's been a friend and an engineering colleague
她是我的朋友和工程同事
George Hotz (00:50.600)
for many years since I was in grad school.
自从我读研究生以来已经很多年了。
Lex Fridman (00:52.760)
Your support means a lot
Lex Fridman (00:54.320)
and inspires me to keep this series going.
Lex Fridman (00:57.880)
And now, here's my conversation with George Hotz.
Lex Fridman (01:02.680)
Do you think we're living in a simulation?
Lex Fridman (01:06.440)
Yes, but it may be unfalsifiable.
Lex Fridman (01:10.040)
What do you mean by unfalsifiable?
Lex Fridman (01:12.400)
So if the simulation is designed in such a way
George Hotz (01:16.800)
that they did like a formal proof
Lex Fridman (01:19.600)
to show that no information can get in and out,
Lex Fridman (01:22.280)
and if their hardware is designed
Lex Fridman (01:24.040)
for the anything in the simulation
George Hotz (01:25.960)
to always keep the hardware in spec,
Lex Fridman (01:27.840)
it may be impossible to prove
George Hotz (01:29.440)
whether we're in a simulation or not.
Lex Fridman (01:32.560)
So they've designed it such that it's a closed system
George Hotz (01:35.640)
you can't get outside the system.
Lex Fridman (01:37.160)
Well, maybe it's one of three worlds.
George Hotz (01:38.760)
We're either in a simulation which can be exploited,
Lex Fridman (01:41.360)
we're in a simulation which not only can't be exploited,
Lex Fridman (01:44.160)
but like the same thing's true about VMs.
Lex Fridman (01:46.400)
A really well designed VM,
George Hotz (01:48.120)
you can't even detect if you're in a VM or not.
Lex Fridman (01:51.360)
That's brilliant.
Lex Fridman (01:52.480)
So the simulation is running on a virtual machine.
Lex Fridman (01:56.760)
But now in reality, all VMs have ways to detect.
George Hotz (01:59.400)
That's the point.
Lex Fridman (02:00.240)
I mean, you've done quite a bit of hacking yourself.
Lex Fridman (02:04.800)
So you should know that really any complicated system
Lex Fridman (02:08.600)
will have ways in and out.
Lex Fridman (02:10.960)
So this isn't necessarily true going forward.
Lex Fridman (02:15.240)
I spent my time away from Comma,
George Hotz (02:18.040)
I learned Coq, it's a dependently typed,
Lex Fridman (02:21.800)
it's a language for writing math proofs in.
Lex Fridman (02:24.320)
And if you write code that compiles in a language like that,
Lex Fridman (02:28.160)
it is correct by definition.
George Hotz (02:30.800)
The types check its correctness.
Lex Fridman (02:33.520)
So it's possible that the simulation
George Hotz (02:34.960)
is written in a language like this, in which case, yeah.
Lex Fridman (02:39.600)
Yeah, but that can't be sufficiently expressive
George Hotz (02:42.640)
a language like that.
Lex Fridman (02:43.720)
Oh, it can.
Lex Fridman (02:44.560)
It can be?
Lex Fridman (02:45.400)
Oh, yeah.
George Hotz (02:46.240)
Okay, well, so all right, so.
Lex Fridman (02:48.880)
The simulation doesn't have to be Turing complete
George Hotz (02:50.600)
if it has a scheduled end date.
Lex Fridman (02:52.280)
Looks like it does actually with entropy.
George Hotz (02:54.560)
I mean, I don't think that a simulation
Lex Fridman (02:58.560)
that results in something as complicated as the universe
Lex Fridman (03:03.040)
would have a form of proof of correctness, right?
Lex Fridman (03:08.240)
It's possible, of course.
George Hotz (03:09.800)
We have no idea how good their tooling is.
Lex Fridman (03:12.720)
And we have no idea how complicated
George Hotz (03:14.600)
the universe computer really is.
Lex Fridman (03:16.240)
It may be quite simple.
Lex Fridman (03:17.880)
It's just very large, right?
Lex Fridman (03:19.640)
It's very, it's definitely very large.
Lex Fridman (03:22.120)
But the fundamental rules might be super simple.
Lex Fridman (03:24.440)
Yeah, Conway's getting a life kind of stuff.
George Hotz (03:26.200)
Right.
Lex Fridman (03:28.040)
So if you could hack,
Lex Fridman (03:30.280)
so imagine a simulation that is hackable,
Lex Fridman (03:32.360)
if you could hack it,
Lex Fridman (03:35.040)
what would you change about the,
Lex Fridman (03:37.880)
like how would you approach hacking a simulation?
George Hotz (03:41.640)
The reason I gave that talk.
Lex Fridman (03:44.320)
By the way, I'm not familiar with the talk you gave.
George Hotz (03:46.640)
I just read that you talked about escaping the simulation
Lex Fridman (03:50.120)
or something like that.
Lex Fridman (03:51.240)
So maybe you can tell me a little bit about the theme
Lex Fridman (03:53.640)
and the message there too.
George Hotz (03:55.320)
It wasn't a very practical talk
Lex Fridman (03:57.640)
about how to actually escape a simulation.
George Hotz (04:00.560)
It was more about a way of restructuring
Lex Fridman (04:03.280)
an us versus them narrative.
George Hotz (04:05.080)
If
Lex Fridman (04:08.160)
we continue on the path we're going with technology,
George Hotz (04:12.320)
I think we're in big trouble,
Lex Fridman (04:14.120)
like as a species and not just as a species,
Lex Fridman (04:16.720)
but even as me as an individual member of the species.
Lex Fridman (04:19.440)
So if we could change rhetoric
George Hotz (04:22.080)
to be more like to think upwards,
Lex Fridman (04:26.200)
like to think about that we're in a simulation
Lex Fridman (04:29.040)
and how we could get out,
Lex Fridman (04:30.320)
already we'd be on the right path.
Lex Fridman (04:32.560)
What you actually do once you do that,
Lex Fridman (04:34.760)
well, I assume I would have acquired way more intelligence
George Hotz (04:37.320)
in the process of doing that.
Lex Fridman (04:38.440)
So I'll just ask that.
Lex Fridman (04:39.720)
So the thinking upwards,
Lex Fridman (04:42.040)
what kind of ideas,
Lex Fridman (04:43.720)
what kind of breakthrough ideas
Lex Fridman (04:44.800)
do you think thinking in that way could inspire?
Lex Fridman (04:47.280)
And why did you say upwards?
Lex Fridman (04:49.760)
Upwards.
Lex Fridman (04:50.600)
Into space?
Lex Fridman (04:51.440)
Are you thinking sort of exploration in all forms?
George Hotz (04:54.040)
The space narrative
Lex Fridman (04:57.400)
that held for the modernist generation
George Hotz (04:59.800)
doesn't hold as well for the postmodern generation.
Lex Fridman (05:04.480)
What's the space narrative?
George Hotz (05:05.400)
Are we talking about the same space,
Lex Fridman (05:06.440)
the three dimensional space?
George Hotz (05:07.280)
No, no, no, space, like going on space,
Lex Fridman (05:08.720)
like building like Elon Musk,
George Hotz (05:09.960)
like we're going to build rockets,
Lex Fridman (05:11.080)
we're going to go to Mars,
George Hotz (05:11.960)
we're going to colonize the universe.
Lex Fridman (05:13.440)
And the narrative you're referring,
George Hotz (05:14.640)
I was born in the Soviet Union,
Lex Fridman (05:15.960)
you're referring to the race to space.
George Hotz (05:17.960)
The race to space, yeah.
Lex Fridman (05:18.880)
Explore, okay.
George Hotz (05:19.720)
That was a great modernist narrative.
Lex Fridman (05:21.760)
Yeah.
George Hotz (05:23.320)
It doesn't seem to hold the same weight in today's culture.
Lex Fridman (05:27.640)
I'm hoping for good postmodern narratives that replace it.
Lex Fridman (05:32.120)
So let's think, so you work a lot with AI.
Lex Fridman (05:35.520)
So AI is one formulation of that narrative.
George Hotz (05:39.040)
There could be also,
Lex Fridman (05:40.040)
I don't know how much you do in VR and AR.
George Hotz (05:42.280)
Yeah.
Lex Fridman (05:43.120)
That's another, I know less about it,
Lex Fridman (05:45.120)
but every time I play with it in our research,
Lex Fridman (05:47.600)
it's fascinating, that virtual world.
Lex Fridman (05:49.600)
Are you interested in the virtual world?
Lex Fridman (05:51.800)
I would like to move to virtual reality.
Lex Fridman (05:55.280)
In terms of your work?
Lex Fridman (05:56.400)
No, I would like to physically move there.
George Hotz (05:58.720)
The apartment I can rent in the cloud
Lex Fridman (06:00.200)
is way better than the apartment
George Hotz (06:01.120)
I can rent in the real world.
Lex Fridman (06:03.200)
Well, it's all relative, isn't it?
George Hotz (06:04.720)
Because others will have very nice apartments too,
Lex Fridman (06:07.240)
so you'll be inferior in the virtual world as well.
Lex Fridman (06:09.080)
No, but that's not how I view the world, right?
Lex Fridman (06:11.280)
I don't view the world,
George Hotz (06:12.400)
I mean, it's a very almost zero sum ish way
Lex Fridman (06:15.600)
to view the world.
George Hotz (06:16.440)
Say like, my great apartment isn't great
Lex Fridman (06:18.760)
because my neighbor has one too.
George Hotz (06:20.360)
No, my great apartment is great
Lex Fridman (06:21.600)
because look at this dishwasher, man.
Lex Fridman (06:24.280)
You just touch the dish and it's washed, right?
Lex Fridman (06:26.640)
And that is great in and of itself
George Hotz (06:28.680)
if I have the only apartment
Lex Fridman (06:30.080)
or if everybody had the apartment.
George Hotz (06:31.480)
I don't care.
Lex Fridman (06:32.360)
So you have fundamental gratitude.
George Hotz (06:34.720)
The world first learned of George Hots
Lex Fridman (06:39.080)
in August 2007, maybe before then,
Lex Fridman (06:42.240)
but certainly in August 2007
Lex Fridman (06:44.040)
when you were the first person to unlock,
George Hotz (06:46.760)
carry unlock an iPhone.
Lex Fridman (06:48.840)
How did you get into hacking?
Lex Fridman (06:50.480)
What was the first system
Lex Fridman (06:51.520)
you discovered vulnerabilities for and broke into?
Lex Fridman (06:56.200)
So that was really kind of the first thing.
Lex Fridman (07:01.600)
I had a book in 2006 called Grey Hat Hacking.
Lex Fridman (07:06.600)
And I guess I realized that if you acquired
Lex Fridman (07:12.240)
these sort of powers, you could control the world.
Lex Fridman (07:16.160)
But I didn't really know that much
Lex Fridman (07:18.960)
about computers back then.
George Hotz (07:20.560)
I started with electronics.
Lex Fridman (07:22.160)
The first iPhone hack was physical.
George Hotz (07:24.200)
Cardware.
Lex Fridman (07:25.040)
You had to open it up and pull an address line high.
Lex Fridman (07:28.160)
And it was because I didn't really know
Lex Fridman (07:29.960)
about software exploitation.
George Hotz (07:31.360)
I learned that all in the next few years
Lex Fridman (07:32.960)
and I got very good at it.
Lex Fridman (07:33.920)
But back then I knew about like how memory chips
Lex Fridman (07:37.640)
are connected to processors and stuff.
George Hotz (07:38.960)
You knew about software and programming.
Lex Fridman (07:40.960)
You just didn't know.
Lex Fridman (07:43.160)
Oh really?
Lex Fridman (07:44.000)
So your view of the world and computers
George Hotz (07:46.800)
was physical, was hardware.
Lex Fridman (07:49.280)
Actually, if you read the code that I released
George Hotz (07:51.760)
with that in August 2007, it's atrocious.
Lex Fridman (07:55.720)
What language was it?
George Hotz (07:56.720)
C.
Lex Fridman (07:57.560)
C, nice.
Lex Fridman (07:58.400)
And in a broken sort of state machine ask C.
Lex Fridman (08:01.440)
I didn't know how to program.
George Hotz (08:02.920)
Yeah.
Lex Fridman (08:04.040)
So how did you learn to program?
Lex Fridman (08:07.400)
What was your journey?
Lex Fridman (08:08.320)
Cause I mean, we'll talk about it.
George Hotz (08:10.000)
You've live streamed some of your programming.
Lex Fridman (08:12.640)
This chaotic, beautiful mess.
Lex Fridman (08:14.360)
How did you arrive at that?
Lex Fridman (08:16.440)
Years and years of practice.
George Hotz (08:18.600)
I interned at Google after the summer
Lex Fridman (08:22.720)
after the iPhone unlock.
Lex Fridman (08:24.760)
And I did a contract for them where I built hardware
Lex Fridman (08:27.960)
for Street View and I wrote a software library
George Hotz (08:30.640)
to interact with it.
Lex Fridman (08:31.760)
And it was terrible code.
Lex Fridman (08:34.360)
And for the first time I got feedback from people
Lex Fridman (08:36.440)
who I respected saying, no, like don't write code like this.
George Hotz (08:42.120)
Now, of course, just getting that feedback is not enough.
Lex Fridman (08:45.120)
The way that I really got good was I wanted to write
George Hotz (08:51.080)
this thing like that could emulate and then visualize
Lex Fridman (08:56.480)
like arm binaries.
George Hotz (08:57.880)
Cause I wanted to hack the iPhone better.
Lex Fridman (08:59.480)
And I didn't like that I couldn't like see
Lex Fridman (09:01.120)
what the, I couldn't single step through the processor
Lex Fridman (09:03.720)
because I had no debugger on there,
George Hotz (09:05.120)
especially for the low level things like the boot rum
Lex Fridman (09:06.560)
and the bootloader.
Lex Fridman (09:07.480)
So I tried to build this tool to do it.
Lex Fridman (09:10.880)
And I built the tool once and it was terrible.
George Hotz (09:13.400)
I built the tool a second time, it was terrible.
Lex Fridman (09:15.080)
I built the tool a third time.
George Hotz (09:16.320)
This was by the time I was at Facebook, it was kind of okay.
Lex Fridman (09:18.600)
And then I built the tool a fourth time
George Hotz (09:20.520)
when I was a Google intern again in 2014.
Lex Fridman (09:22.520)
And that was the first time I was like,
George Hotz (09:24.320)
this is finally usable.
Lex Fridman (09:25.840)
How do you pronounce this Kira?
George Hotz (09:27.080)
Kira, yeah.
Lex Fridman (09:28.360)
So it's essentially the most efficient way
George Hotz (09:31.800)
to visualize the change of state of the computer
Lex Fridman (09:35.680)
as the program is running.
George Hotz (09:37.160)
That's what you mean by debugger.
Lex Fridman (09:38.840)
Yeah, it's a timeless debugger.
Lex Fridman (09:41.720)
So you can rewind just as easily as going forward.
Lex Fridman (09:45.040)
Think about if you're using GDB,
George Hotz (09:46.200)
you have to put a watch on a variable.
Lex Fridman (09:47.840)
If you wanna see if that variable changes.
George Hotz (09:49.640)
In Kira, you can just click on that variable
Lex Fridman (09:51.480)
and then it shows every single time
George Hotz (09:53.840)
when that variable was changed or accessed.
Lex Fridman (09:56.480)
Think about it like Git for your computers, the run log.
Lex Fridman (09:59.760)
So there's like a deep log of the state of the computer
Lex Fridman (10:05.600)
as the program runs and you can rewind.
Lex Fridman (10:07.800)
Why isn't that, maybe it is, maybe you can educate me.
Lex Fridman (10:11.440)
Why isn't that kind of debugging used more often?
George Hotz (10:14.600)
Cause the tooling's bad.
Lex Fridman (10:16.280)
Well, two things.
George Hotz (10:17.120)
One, if you're trying to debug Chrome,
Lex Fridman (10:19.320)
Chrome is a 200 megabyte binary
George Hotz (10:22.840)
that runs slowly on desktops.
Lex Fridman (10:25.400)
So that's gonna be really hard to use for that.
Lex Fridman (10:27.640)
But it's really good to use for like CTFs
Lex Fridman (10:30.080)
and for boot roms and for small parts of code.
Lex Fridman (10:33.120)
So it's hard if you're trying to debug like massive systems.
Lex Fridman (10:36.280)
What's a CTF and what's a boot rom?
George Hotz (10:38.120)
A boot rom is the first code that executes
Lex Fridman (10:40.400)
the minute you give power to your iPhone.
George Hotz (10:42.160)
Okay.
Lex Fridman (10:43.440)
And CTF where these competitions
George Hotz (10:45.520)
that I played capture the flag.
Lex Fridman (10:46.880)
Capture the flag, I was gonna ask you about that.
Lex Fridman (10:48.480)
What are those, look at,
Lex Fridman (10:49.840)
I watched a couple of videos on YouTube,
George Hotz (10:51.320)
those look fascinating.
Lex Fridman (10:52.840)
What have you learned about maybe
George Hotz (10:54.760)
at the high level of vulnerability of systems
Lex Fridman (10:56.680)
from these competitions?
George Hotz (11:00.760)
I feel like in the heyday of CTFs,
Lex Fridman (11:04.160)
you had all of the best security people in the world
George Hotz (11:08.080)
challenging each other and coming up
Lex Fridman (11:10.640)
with new toy exploitable things over here.
Lex Fridman (11:13.600)
And then everybody, okay, who can break it?
Lex Fridman (11:15.320)
And when you break it, you get like,
George Hotz (11:17.080)
there's like a file on the server called flag.
Lex Fridman (11:19.280)
And then there's a program running,
George Hotz (11:20.920)
listening on a socket that's vulnerable.
Lex Fridman (11:22.600)
So you write an exploit, you get a shell,
Lex Fridman (11:24.920)
and then you cat flag, and then you type the flag
Lex Fridman (11:27.080)
into like a web based scoreboard and you get points.
Lex Fridman (11:29.440)
So the goal is essentially,
Lex Fridman (11:31.520)
to find an exploit in the system
George Hotz (11:32.920)
that allows you to run shell,
Lex Fridman (11:35.200)
to run arbitrary code on that system.
George Hotz (11:37.960)
That's one of the categories.
Lex Fridman (11:40.120)
That's like the pwnable category.
Lex Fridman (11:43.480)
Pwnable?
Lex Fridman (11:44.320)
Yeah, pwnable.
George Hotz (11:45.160)
It's like, you know, you pwn the program.
Lex Fridman (11:47.520)
It's a program that's, yeah.
George Hotz (11:48.920)
Yeah, you know, first of all, I apologize.
Lex Fridman (11:54.160)
I'm gonna say it's because I'm Russian,
Lex Fridman (11:56.240)
but maybe you can help educate me.
Lex Fridman (12:00.080)
Some video game like misspelled own way back in the day.
George Hotz (12:02.760)
Yeah, and it's just, I wonder if there's a definition.
Lex Fridman (12:06.280)
I'll have to go to Urban Dictionary for it.
George Hotz (12:08.240)
It'll be interesting to see what it says.
Lex Fridman (12:09.720)
Okay, so what was the heyday of CTF, by the way?
Lex Fridman (12:12.720)
But was it, what decade are we talking about?
Lex Fridman (12:15.440)
I think like, I mean, maybe unbiased
George Hotz (12:18.360)
because it's the era that I played,
Lex Fridman (12:21.040)
but like 2011 to 2015,
George Hotz (12:27.120)
because the modern CTF scene
Lex Fridman (12:30.240)
is similar to the modern competitive programming scene.
George Hotz (12:32.560)
You have people who like do drills.
Lex Fridman (12:34.200)
You have people who practice.
Lex Fridman (12:35.800)
And then once you've done that,
Lex Fridman (12:36.920)
you've turned it less into a game of generic computer skill
Lex Fridman (12:39.960)
and more into a game of, okay,
Lex Fridman (12:41.680)
you drill on these five categories.
Lex Fridman (12:44.560)
And then before that, it wasn't,
Lex Fridman (12:48.880)
it didn't have like as much attention as it had.
George Hotz (12:52.760)
I don't know, they were like,
Lex Fridman (12:53.600)
I won $30,000 once in Korea for one of these competitions.
George Hotz (12:56.040)
Holy crap.
Lex Fridman (12:56.880)
Yeah, they were, they were, that was.
Lex Fridman (12:57.880)
So that means, I mean, money is money,
Lex Fridman (12:59.480)
but that means there was probably good people there.
George Hotz (13:02.240)
Exactly, yeah.
Lex Fridman (13:03.520)
Are the challenges human constructed
Lex Fridman (13:06.720)
or are they grounded in some real flaws and real systems?
Lex Fridman (13:10.720)
Usually they're human constructed,
Lex Fridman (13:13.000)
but they're usually inspired by real flaws.
Lex Fridman (13:15.720)
What kind of systems are imagined
George Hotz (13:17.240)
is really focused on mobile.
Lex Fridman (13:19.040)
Like what has vulnerabilities these days?
Lex Fridman (13:20.880)
Is it primarily mobile systems like Android?
Lex Fridman (13:25.080)
Oh, everything does.
George Hotz (13:26.480)
Still. Yeah, of course.
Lex Fridman (13:28.040)
The price has kind of gone up
George Hotz (13:29.320)
because less and less people can find them.
Lex Fridman (13:31.200)
And what's happened in security
George Hotz (13:32.760)
is now if you want to like jailbreak an iPhone,
Lex Fridman (13:34.480)
you don't need one exploit anymore, you need nine.
Lex Fridman (13:37.880)
Nine chained together, what would it mean?
Lex Fridman (13:39.600)
Yeah, wow.
George Hotz (13:40.560)
Okay, so it's really,
Lex Fridman (13:42.680)
what's the benefit speaking higher level
Lex Fridman (13:46.120)
philosophically about hacking?
Lex Fridman (13:48.160)
I mean, it sounds from everything I've seen about you,
George Hotz (13:50.320)
you just love the challenge
Lex Fridman (13:51.840)
and you don't want to do anything.
George Hotz (13:54.960)
You don't want to bring that exploit out into the world
Lex Fridman (13:58.040)
and do any actual, let it run wild.
George Hotz (14:01.600)
You just want to solve it
Lex Fridman (14:02.680)
and then you go on to the next thing.
George Hotz (14:05.320)
Oh yeah, I mean, doing criminal stuff's not really worth it.
Lex Fridman (14:08.360)
And I'll actually use the same argument
George Hotz (14:10.440)
for why I don't do defense for why I don't do crime.
Lex Fridman (14:15.320)
If you want to defend a system,
Lex Fridman (14:16.720)
say the system has 10 holes, right?
Lex Fridman (14:19.160)
If you find nine of those holes as a defender,
George Hotz (14:22.120)
you still lose because the attacker
Lex Fridman (14:23.800)
gets in through the last one.
George Hotz (14:25.400)
If you're an attacker,
Lex Fridman (14:26.240)
you only have to find one out of the 10.
Lex Fridman (14:28.600)
But if you're a criminal,
Lex Fridman (14:30.680)
if you log on with a VPN nine out of the 10 times,
Lex Fridman (14:34.680)
but one time you forget, you're done.
Lex Fridman (14:37.640)
Because you're caught, okay.
George Hotz (14:39.240)
Because you only have to mess up once
Lex Fridman (14:41.040)
to be caught as a criminal.
George Hotz (14:42.760)
That's why I'm not a criminal.
Lex Fridman (14:45.800)
But okay, let me,
George Hotz (14:46.960)
because I was having a discussion with somebody
Lex Fridman (14:49.400)
just at a high level about nuclear weapons actually,
Lex Fridman (14:52.640)
why we're having blown ourselves up yet.
Lex Fridman (14:56.120)
And my feeling is all the smart people in the world,
George Hotz (14:59.680)
if you look at the distribution of smart people,
Lex Fridman (15:04.000)
smart people are generally good.
Lex Fridman (15:06.600)
And then this other person I was talking to,
Lex Fridman (15:07.920)
Sean Carroll, the physicist,
Lex Fridman (15:09.320)
and he was saying, no, good and bad people
Lex Fridman (15:11.280)
are evenly distributed amongst everybody.
George Hotz (15:13.960)
My sense was good hackers are in general good people
Lex Fridman (15:17.960)
and they don't want to mess with the world.
Lex Fridman (15:20.280)
What's your sense?
Lex Fridman (15:21.640)
I'm not even sure about that.
George Hotz (15:25.800)
Like,
Lex Fridman (15:28.880)
I have a nice life.
George Hotz (15:30.400)
Crime wouldn't get me anything.
Lex Fridman (15:34.160)
But if you're good and you have these skills,
Lex Fridman (15:36.400)
you probably have a nice life too, right?
Lex Fridman (15:38.600)
Right, you can use it for other things.
Lex Fridman (15:40.040)
But is there an ethical,
Lex Fridman (15:41.000)
is there a little voice in your head that says,
George Hotz (15:46.000)
well, yeah, if you could hack something
Lex Fridman (15:48.920)
to where you could hurt people
Lex Fridman (15:52.720)
and you could earn a lot of money doing it though,
Lex Fridman (15:54.840)
not hurt physically perhaps,
Lex Fridman (15:56.200)
but disrupt their life in some kind of way,
Lex Fridman (16:00.080)
isn't there a little voice that says?
George Hotz (16:03.240)
Well, two things.
Lex Fridman (16:04.440)
One, I don't really care about money.
Lex Fridman (16:06.680)
So like the money wouldn't be an incentive.
Lex Fridman (16:08.600)
The thrill might be an incentive.
Lex Fridman (16:10.560)
But when I was 19, I read Crime and Punishment.
Lex Fridman (16:14.320)
And that was another great one
George Hotz (16:16.040)
that talked me out of ever really doing crime.
Lex Fridman (16:19.320)
Cause it's like, that's gonna be me.
George Hotz (16:21.640)
I'd get away with it, but it would just run through my head.
Lex Fridman (16:25.000)
Even if I got away with it, you know?
Lex Fridman (16:26.400)
And then you do crime for long enough,
Lex Fridman (16:27.560)
you'll never get away with it.
George Hotz (16:28.880)
That's right.
Lex Fridman (16:29.720)
In the end, that's a good reason to be good.
George Hotz (16:32.600)
I wouldn't say I'm good.
Lex Fridman (16:33.440)
I would just say I'm not bad.
George Hotz (16:34.840)
You're a talented programmer and a hacker
Lex Fridman (16:38.080)
in a good positive sense of the word.
George Hotz (16:40.920)
You've played around,
Lex Fridman (16:42.400)
found vulnerabilities in various systems.
Lex Fridman (16:44.720)
What have you learned broadly
Lex Fridman (16:46.120)
about the design of systems and so on
Lex Fridman (16:49.480)
from that whole process?
Lex Fridman (16:53.280)
You learn to not take things
George Hotz (16:59.280)
for what people say they are,
Lex Fridman (17:02.000)
but you look at things for what they actually are.
George Hotz (17:07.000)
Yeah.
Lex Fridman (17:07.880)
I understand that's what you tell me it is,
Lex Fridman (17:10.040)
but what does it do?
Lex Fridman (17:11.440)
Right.
Lex Fridman (17:12.920)
And you have nice visualization tools
Lex Fridman (17:14.560)
to really know what it's really doing.
George Hotz (17:16.680)
Oh, I wish.
Lex Fridman (17:17.760)
I'm a better programmer now than I was in 2014.
George Hotz (17:20.040)
I said, Kira, that was the first tool
Lex Fridman (17:21.840)
that I wrote that was usable.
George Hotz (17:23.400)
I wouldn't say the code was great.
Lex Fridman (17:25.320)
I still wouldn't say my code is great.
Lex Fridman (17:28.800)
So how was your evolution as a programmer except practice?
Lex Fridman (17:31.480)
So you started with C.
Lex Fridman (17:33.840)
At which point did you pick up Python?
Lex Fridman (17:35.520)
Because you're pretty big in Python now.
George Hotz (17:37.040)
Now, yeah, in college.
Lex Fridman (17:39.920)
I went to Carnegie Mellon when I was 22.
George Hotz (17:42.480)
I went back.
Lex Fridman (17:43.320)
I'm like, all right,
George Hotz (17:44.160)
I'm gonna take all your hardest CS courses.
Lex Fridman (17:46.600)
We'll see how I do, right?
George Hotz (17:47.600)
Like, did I miss anything
Lex Fridman (17:48.520)
by not having a real undergraduate education?
George Hotz (17:51.480)
Took operating systems, compilers, AI,
Lex Fridman (17:54.200)
and they're like a freshman wheat or math course.
George Hotz (17:58.560)
And...
Lex Fridman (18:00.480)
Operating systems, some of those classes
George Hotz (18:02.120)
you mentioned are pretty tough, actually.
Lex Fridman (18:04.160)
They're great.
George Hotz (18:05.560)
At least the 2012, circa 2012,
Lex Fridman (18:08.600)
operating systems and compilers were two of the,
George Hotz (18:12.200)
they were the best classes I've ever taken in my life.
Lex Fridman (18:14.360)
Because you write an operating system
Lex Fridman (18:15.560)
and you write a compiler.
Lex Fridman (18:18.040)
I wrote my operating system in C
Lex Fridman (18:19.720)
and I wrote my compiler in Haskell,
Lex Fridman (18:21.320)
but somehow I picked up Python that semester as well.
George Hotz (18:26.360)
I started using it for the CTFs, actually.
Lex Fridman (18:28.040)
That's when I really started to get into CTFs
Lex Fridman (18:30.280)
and CTFs, you're all, it's a race against the clock.
Lex Fridman (18:33.320)
So I can't write things in C.
George Hotz (18:35.080)
Oh, there's a clock component.
Lex Fridman (18:36.200)
So you really want to use the programming languages
Lex Fridman (18:38.000)
so you can be fastest.
Lex Fridman (18:38.920)
48 hours, pwn as many of these challenges as you can.
George Hotz (18:41.400)
Pwn.
Lex Fridman (18:42.240)
Yeah, you got like a hundred points of challenge.
George Hotz (18:43.920)
Whatever team gets the most.
Lex Fridman (18:46.280)
You were both at Facebook and Google for a brief stint.
George Hotz (18:50.200)
Yeah.
Lex Fridman (18:51.040)
With Project Zero actually at Google for five months
George Hotz (18:54.880)
where you developed Kira.
Lex Fridman (18:56.920)
What was Project Zero about in general?
George Hotz (18:59.840)
What, I'm just curious about the security efforts
Lex Fridman (19:03.960)
in these companies.
George Hotz (19:05.160)
Well, Project Zero started the same time I went there.
Lex Fridman (19:08.760)
What years are there?
George Hotz (19:11.080)
2015.
Lex Fridman (19:12.320)
2015.
Lex Fridman (19:13.160)
So that was right at the beginning of Project Zero.
Lex Fridman (19:15.040)
It's small.
George Hotz (19:16.200)
It's Google's offensive security team.
Lex Fridman (19:21.840)
I'll try to give the best public facing explanation
George Hotz (19:25.640)
that I can.
Lex Fridman (19:26.480)
So the idea is basically these vulnerabilities
George Hotz (19:31.760)
exist in the world.
Lex Fridman (19:33.200)
Nation states have them.
George Hotz (19:35.200)
Some high powered bad actors have them.
Lex Fridman (19:39.800)
Sometime people will find these vulnerabilities
Lex Fridman (19:44.160)
and submit them in bug bounties to the companies.
Lex Fridman (19:47.800)
But a lot of the companies don't really care.
George Hotz (19:49.480)
They don't even fix the bug.
Lex Fridman (19:51.160)
It doesn't hurt for there to be a vulnerability.
Lex Fridman (19:53.840)
So Project Zero is like, we're going to do it different.
Lex Fridman (19:55.920)
We're going to announce a vulnerability
Lex Fridman (19:57.880)
and we're going to give them 90 days to fix it.
Lex Fridman (19:59.680)
And then whether they fix it or not,
George Hotz (1:00:01.620)
to what you're describing,
Lex Fridman (1:00:02.980)
which is really turning into not some kind of modular thing,
Lex Fridman (1:00:08.300)
but really do formulate it as a learning problem
Lex Fridman (1:00:11.140)
and solve the learning problem with scale.
Lex Fridman (1:00:13.380)
So how many years, put one is how many years
Lex Fridman (1:00:17.700)
would it take to solve this problem
Lex Fridman (1:00:18.860)
or just how hard is this freaking problem?
Lex Fridman (1:00:21.660)
Well, the cool thing is I think there's a lot of value
George Hotz (1:00:27.780)
that we can deliver along the way.
Lex Fridman (1:00:30.820)
I think that you can build lane keeping assist actually
George Hotz (1:00:37.260)
plus adaptive cruise control, plus, okay, looking at ways,
Lex Fridman (1:00:42.260)
extends to like all of driving.
Lex Fridman (1:00:45.980)
Yeah, most of driving, right?
Lex Fridman (1:00:47.900)
Oh, your adaptive cruise control treats red lights
George Hotz (1:00:49.740)
like cars, okay.
Lex Fridman (1:00:51.180)
So let's jump around.
George Hotz (1:00:52.980)
You mentioned that you didn't like navigate an autopilot.
Lex Fridman (1:00:55.740)
What advice, how would you make it better?
Lex Fridman (1:00:57.740)
Do you think as a feature that if it's done really well,
Lex Fridman (1:01:00.540)
it's a good feature?
George Hotz (1:01:02.340)
I think that it's too reliant on like hand coded hacks
Lex Fridman (1:01:07.460)
for like, how does navigate an autopilot do a lane change?
George Hotz (1:01:10.380)
It actually does the same lane change every time
Lex Fridman (1:01:13.340)
and it feels mechanical.
George Hotz (1:01:14.260)
Humans do different lane changes.
Lex Fridman (1:01:15.820)
Humans sometime will do a slow one,
George Hotz (1:01:17.300)
sometimes do a fast one.
Lex Fridman (1:01:18.860)
Navigate an autopilot, at least every time I use it,
George Hotz (1:01:20.820)
it is the identical lane change.
Lex Fridman (1:01:22.980)
How do you learn?
George Hotz (1:01:24.220)
I mean, this is a fundamental thing actually
Lex Fridman (1:01:26.740)
is the braking and then accelerating
George Hotz (1:01:30.340)
something that's still, Tesla probably does it better
Lex Fridman (1:01:33.900)
than most cars, but it still doesn't do a great job
George Hotz (1:01:36.740)
of creating a comfortable natural experience.
Lex Fridman (1:01:39.900)
And navigate an autopilot is just lane changes
Lex Fridman (1:01:42.620)
and extension of that.
Lex Fridman (1:01:44.060)
So how do you learn to do a natural lane change?
Lex Fridman (1:01:49.180)
So we have it and I can talk about how it works.
Lex Fridman (1:01:52.980)
So I feel that we have the solution for lateral.
George Hotz (1:01:58.860)
We don't yet have the solution for longitudinal.
Lex Fridman (1:02:00.700)
There's a few reasons longitudinal is harder than lateral.
George Hotz (1:02:03.420)
The lane change component,
Lex Fridman (1:02:05.180)
the way that we train on it very simply
George Hotz (1:02:08.060)
is like our model has an input
Lex Fridman (1:02:10.900)
for whether it's doing a lane change or not.
Lex Fridman (1:02:14.100)
And then when we train the end to end model,
Lex Fridman (1:02:16.420)
we hand label all the lane changes,
George Hotz (1:02:18.420)
cause you have to.
Lex Fridman (1:02:19.580)
I've struggled a long time about not wanting to do that,
Lex Fridman (1:02:22.460)
but I think you have to.
Lex Fridman (1:02:24.300)
Or the training data.
Lex Fridman (1:02:25.340)
For the training data, right?
Lex Fridman (1:02:26.540)
Oh, we actually, we have an automatic ground truther
George Hotz (1:02:28.380)
which automatically labels all the lane changes.
Lex Fridman (1:02:30.580)
Was that possible?
Lex Fridman (1:02:31.700)
To automatically label the lane changes?
Lex Fridman (1:02:32.780)
Yeah.
Lex Fridman (1:02:33.620)
Yeah, detect the lane, I see when it crosses it, right?
Lex Fridman (1:02:34.820)
And I don't have to get that high percent accuracy,
Lex Fridman (1:02:36.700)
but it's like 95, good enough.
Lex Fridman (1:02:38.980)
Now I set the bit when it's doing the lane change
George Hotz (1:02:43.220)
in the end to end learning.
Lex Fridman (1:02:44.860)
And then I set it to zero when it's not doing a lane change.
Lex Fridman (1:02:47.940)
So now if I wanted to do a lane change at test time,
Lex Fridman (1:02:49.740)
I just put the bit to a one and it'll do a lane change.
George Hotz (1:02:52.380)
Yeah, but so if you look at the space of lane change,
Lex Fridman (1:02:54.660)
you know, some percentage, not a hundred percent
George Hotz (1:02:57.340)
that we make as humans is not a pleasant experience
Lex Fridman (1:03:01.140)
cause we messed some part of it up.
George Hotz (1:03:02.860)
It's nerve wracking to change the look,
Lex Fridman (1:03:04.940)
you have to see, it has to accelerate.
Lex Fridman (1:03:06.940)
How do we label the ones that are natural and feel good?
Lex Fridman (1:03:09.940)
You know, that's the, cause that's your ultimate criticism.
George Hotz (1:03:13.380)
The current navigate and autopilot
Lex Fridman (1:03:15.860)
just doesn't feel good.
George Hotz (1:03:16.940)
Well, the current navigate and autopilot
Lex Fridman (1:03:18.460)
is a hand coded policy written by an engineer in a room
George Hotz (1:03:21.660)
who probably went out and tested it a few times on the 280.
Lex Fridman (1:03:25.020)
Probably a more, a better version of that, but yes.
George Hotz (1:03:29.420)
That's how we would have written it at Comma AI.
Lex Fridman (1:03:31.020)
Yeah, yeah, yeah.
George Hotz (1:03:31.860)
Maybe Tesla did, Tesla, they tested it in the end.
Lex Fridman (1:03:33.420)
That might've been two engineers.
George Hotz (1:03:35.100)
Two engineers, yeah.
Lex Fridman (1:03:37.380)
No, but so if you learn the lane change,
George Hotz (1:03:40.060)
if you learn how to do a lane change from data,
Lex Fridman (1:03:42.420)
just like you have a label that says lane change
Lex Fridman (1:03:44.660)
and then you put it in when you want it
Lex Fridman (1:03:46.380)
to do the lane change,
George Hotz (1:03:48.020)
it'll automatically do the lane change
Lex Fridman (1:03:49.620)
that's appropriate for the situation.
George Hotz (1:03:51.580)
Now, to get at the problem of some humans
Lex Fridman (1:03:54.700)
do bad lane changes,
George Hotz (1:03:57.380)
we haven't worked too much on this problem yet.
Lex Fridman (1:03:59.900)
It's not that much of a problem in practice.
George Hotz (1:04:03.100)
My theory is that all good drivers are good in the same way
Lex Fridman (1:04:06.140)
and all bad drivers are bad in different ways.
Lex Fridman (1:04:09.340)
And we've seen some data to back this up.
Lex Fridman (1:04:11.300)
Well, beautifully put.
Lex Fridman (1:04:12.380)
So you just basically, if that's true hypothesis,
Lex Fridman (1:04:16.540)
then your task is to discover the good drivers.
George Hotz (1:04:19.860)
The good drivers stand out because they're in one cluster
Lex Fridman (1:04:23.300)
and the bad drivers are scattered all over the place
Lex Fridman (1:04:25.140)
and your net learns the cluster.
Lex Fridman (1:04:27.180)
Yeah, that's, so you just learn from the good drivers
Lex Fridman (1:04:30.740)
and they're easy to cluster.
Lex Fridman (1:04:33.140)
In fact, we learned from all of them
Lex Fridman (1:04:33.980)
and the net automatically learns the policy
Lex Fridman (1:04:35.780)
that's like the majority,
Lex Fridman (1:04:36.860)
but we'll eventually probably have to filter them out.
Lex Fridman (1:04:38.500)
If that theory is true, I hope it's true
George Hotz (1:04:41.500)
because the counter theory is there is many clusters,
Lex Fridman (1:04:49.420)
maybe arbitrarily many clusters of good drivers.
George Hotz (1:04:53.620)
Because if there's one cluster of good drivers,
Lex Fridman (1:04:55.780)
you can at least discover a set of policies.
George Hotz (1:04:57.540)
You can learn a set of policies,
Lex Fridman (1:04:58.940)
which would be good universally.
George Hotz (1:05:00.580)
Yeah.
Lex Fridman (1:05:01.620)
That would be a nice, that would be nice if it's true.
Lex Fridman (1:05:04.540)
And you're saying that there is some evidence that.
Lex Fridman (1:05:06.540)
Let's say lane changes can be clustered into four clusters.
George Hotz (1:05:09.740)
Right. Right.
Lex Fridman (1:05:10.580)
There's this finite level of.
George Hotz (1:05:12.020)
I would argue that all four of those are good clusters.
Lex Fridman (1:05:15.260)
All the things that are random are noise and probably bad.
Lex Fridman (1:05:18.420)
And which one of the four you pick,
Lex Fridman (1:05:20.340)
or maybe it's 10 or maybe it's 20.
George Hotz (1:05:21.900)
You can learn that.
Lex Fridman (1:05:22.740)
It's context dependent.
George Hotz (1:05:23.780)
It depends on the scene.
Lex Fridman (1:05:24.980)
And the hope is it's not too dependent on the driver.
George Hotz (1:05:31.380)
Yeah. The hope is that it all washes out.
Lex Fridman (1:05:34.220)
The hope is that there's, that the distribution's not bimodal.
George Hotz (1:05:36.980)
The hope is that it's a nice Gaussian.
Lex Fridman (1:05:39.100)
So what advice would you give to Tesla,
Lex Fridman (1:05:41.660)
how to fix, how to improve navigating autopilot?
Lex Fridman (1:05:44.980)
That's the lessons that you've learned from Comm AI?
George Hotz (1:05:48.260)
The only real advice I would give to Tesla
Lex Fridman (1:05:50.580)
is please put driver monitoring in your cars.
Lex Fridman (1:05:52.940)
With respect to improving it?
Lex Fridman (1:05:55.100)
You can't do that anymore.
George Hotz (1:05:55.940)
I decided to interrupt, but you know,
Lex Fridman (1:05:58.220)
there's a practical nature of many of hundreds of thousands
George Hotz (1:06:01.740)
of cars being produced that don't have
Lex Fridman (1:06:04.180)
a good driver facing camera.
George Hotz (1:06:05.780)
The Model 3 has a selfie cam.
Lex Fridman (1:06:07.500)
Is it not good enough?
Lex Fridman (1:06:08.660)
Did they not put IR LEDs for night?
Lex Fridman (1:06:10.780)
That's a good question.
Lex Fridman (1:06:11.620)
But I do know that it's fisheye
Lex Fridman (1:06:13.340)
and it's relatively low resolution.
Lex Fridman (1:06:15.780)
So it's really not designed.
Lex Fridman (1:06:16.740)
It wasn't.
George Hotz (1:06:17.580)
It wasn't designed for driver monitoring.
Lex Fridman (1:06:18.740)
You can hope that you can kind of scrape up
Lex Fridman (1:06:21.740)
and have something from it.
Lex Fridman (1:06:24.180)
Yeah.
Lex Fridman (1:06:25.020)
But why didn't they put it in today?
Lex Fridman (1:06:27.500)
Put it in today.
George Hotz (1:06:28.340)
Put it in today.
Lex Fridman (1:06:29.500)
Every time I've heard Karpathy talk about the problem
Lex Fridman (1:06:31.500)
and talking about like software 2.0
Lex Fridman (1:06:33.220)
and how the machine learning is gobbling up everything,
George Hotz (1:06:35.220)
I think this is absolutely the right strategy.
Lex Fridman (1:06:37.420)
I think that he didn't write navigate on autopilot.
George Hotz (1:06:40.140)
I think somebody else did
Lex Fridman (1:06:41.540)
and kind of hacked it on top of that stuff.
George Hotz (1:06:43.220)
I think when Karpathy says, wait a second,
Lex Fridman (1:06:45.700)
why did we hand code this lane change policy
Lex Fridman (1:06:47.420)
with all these magic numbers?
Lex Fridman (1:06:48.340)
We're gonna learn it from data.
George Hotz (1:06:49.340)
They'll fix it.
Lex Fridman (1:06:50.180)
They already know what to do there.
George Hotz (1:06:51.060)
Well, that's Andrei's job
Lex Fridman (1:06:53.380)
is to turn everything into a learning problem
Lex Fridman (1:06:55.780)
and collect a huge amount of data.
Lex Fridman (1:06:57.500)
The reality is though,
George Hotz (1:06:59.540)
not every problem can be turned into a learning problem
Lex Fridman (1:07:02.780)
in the short term.
George Hotz (1:07:04.100)
In the end, everything will be a learning problem.
Lex Fridman (1:07:07.300)
The reality is like if you wanna build L5 vehicles today,
George Hotz (1:07:12.940)
it will likely involve no learning.
Lex Fridman (1:07:15.460)
And that's the reality is,
Lex Fridman (1:07:17.420)
so at which point does learning start?
Lex Fridman (1:07:20.340)
It's the crutch statement that LiDAR is a crutch.
George Hotz (1:07:23.500)
At which point will learning
Lex Fridman (1:07:24.860)
get up to part of human performance?
George Hotz (1:07:27.260)
It's over human performance on ImageNet,
Lex Fridman (1:07:30.980)
classification, on driving, it's a question still.
George Hotz (1:07:34.060)
It is a question.
Lex Fridman (1:07:35.820)
I'll say this, I'm here to play for 10 years.
George Hotz (1:07:39.260)
I'm not here to try to,
Lex Fridman (1:07:40.340)
I'm here to play for 10 years and make money along the way.
George Hotz (1:07:43.020)
I'm not here to try to promise people
Lex Fridman (1:07:45.100)
that I'm gonna have my L5 taxi network
George Hotz (1:07:47.060)
up and working in two years.
Lex Fridman (1:07:48.300)
Do you think that was a mistake?
George Hotz (1:07:49.500)
Yes.
Lex Fridman (1:07:50.580)
What do you think was the motivation behind saying that?
George Hotz (1:07:53.540)
Other companies are also promising L5 vehicles
Lex Fridman (1:07:56.700)
with very different approaches in 2020, 2021, 2022.
George Hotz (1:08:01.940)
If anybody would like to bet me
Lex Fridman (1:08:03.740)
that those things do not pan out, I will bet you.
George Hotz (1:08:06.940)
Even money, even money, I'll bet you as much as you want.
Lex Fridman (1:08:09.780)
Yeah.
Lex Fridman (1:08:10.900)
So are you worried about what's going to happen?
Lex Fridman (1:08:13.660)
Cause you're not in full agreement on that.
George Hotz (1:08:16.140)
What's going to happen when 2022, 21 come around
Lex Fridman (1:08:19.180)
and nobody has fleets of autonomous vehicles?
George Hotz (1:08:22.900)
Well, you can look at the history.
Lex Fridman (1:08:25.060)
If you go back five years ago,
George Hotz (1:08:26.740)
they were all promised by 2018 and 2017.
Lex Fridman (1:08:29.980)
But they weren't that strong of promises.
George Hotz (1:08:32.260)
I mean, Ford really declared pretty,
Lex Fridman (1:08:36.260)
I think not many have declared as like definitively
George Hotz (1:08:40.640)
as they have now these dates.
Lex Fridman (1:08:42.660)
Well, okay, so let's separate L4 and L5.
George Hotz (1:08:45.100)
Do I think that it's possible for Waymo to continue to kind
Lex Fridman (1:08:49.480)
of like hack on their system
Lex Fridman (1:08:51.020)
until it gets to level four in Chandler, Arizona?
Lex Fridman (1:08:53.460)
Yes.
Lex Fridman (1:08:55.060)
When there's no safety driver?
Lex Fridman (1:08:56.860)
Chandler, Arizona?
George Hotz (1:08:57.700)
Yeah.
Lex Fridman (1:08:59.580)
By, sorry, which year are we talking about?
George Hotz (1:09:02.540)
Oh, I even think that's possible by like 2020, 2021.
Lex Fridman (1:09:06.180)
But level four, Chandler, Arizona,
George Hotz (1:09:08.460)
not level five, New York City.
Lex Fridman (1:09:10.340)
Level four, meaning some very defined streets,
George Hotz (1:09:15.980)
it works out really well.
Lex Fridman (1:09:17.460)
Very defined streets.
Lex Fridman (1:09:18.300)
And then practically these streets are pretty empty.
Lex Fridman (1:09:20.720)
If most of the streets are covered in Waymo's,
George Hotz (1:09:24.700)
Waymo can kind of change the definition of what driving is.
Lex Fridman (1:09:28.420)
Right?
George Hotz (1:09:29.260)
If your self driving network
Lex Fridman (1:09:30.980)
is the majority of cars in an area,
George Hotz (1:09:33.460)
they only need to be safe with respect to each other
Lex Fridman (1:09:35.740)
and all the humans will need to learn to adapt to them.
George Hotz (1:09:38.660)
Now go drive in downtown New York.
Lex Fridman (1:09:41.140)
Well, yeah, that's.
George Hotz (1:09:42.220)
I mean, already you can talk about autonomy
Lex Fridman (1:09:44.780)
and like on farms, it already works great
George Hotz (1:09:46.980)
because you can really just follow the GPS line.
Lex Fridman (1:09:51.300)
So what does success look like for common AI?
Lex Fridman (1:09:55.640)
What are the milestones?
Lex Fridman (1:09:57.900)
Like where you can sit back with some champagne
Lex Fridman (1:09:59.820)
and say, we did it, boys and girls?
Lex Fridman (1:10:04.140)
Well, it's never over.
George Hotz (1:10:06.260)
Yeah, but.
Lex Fridman (1:10:07.300)
You must drink champagne and celebrate.
Lex Fridman (1:10:10.420)
So what is a good, what are some wins?
Lex Fridman (1:10:13.180)
A big milestone that we're hoping for
George Hotz (1:10:17.780)
by mid next year is profitability of the company.
Lex Fridman (1:10:23.580)
And we're gonna have to revisit the idea
George Hotz (1:10:27.680)
of selling a consumer product,
Lex Fridman (1:10:30.320)
but it's not gonna be like the comma one.
George Hotz (1:10:32.740)
When we do it, it's gonna be perfect.
Lex Fridman (1:10:35.320)
Open pilot has gotten so much better in the last two years.
George Hotz (1:10:39.640)
We're gonna have a few features.
Lex Fridman (1:10:41.720)
We're gonna have a hundred percent driver monitoring.
George Hotz (1:10:43.800)
We're gonna disable no safety features in the car.
Lex Fridman (1:10:47.120)
Actually, I think it'd be really cool
Lex Fridman (1:10:48.280)
what we're doing right now.
Lex Fridman (1:10:49.160)
Our project this week is we're analyzing the data set
Lex Fridman (1:10:51.640)
and looking for all the AEB triggers
Lex Fridman (1:10:53.240)
from the manufacturer systems.
George Hotz (1:10:55.640)
We have better data set on that than the manufacturers.
Lex Fridman (1:10:59.440)
How much, just how many,
George Hotz (1:11:00.920)
does Toyota have 10 million miles of real world driving
Lex Fridman (1:11:03.360)
to know how many times their AEB triggered?
Lex Fridman (1:11:05.320)
So let me give you, cause you asked, right?
Lex Fridman (1:11:08.400)
Financial advice.
George Hotz (1:11:09.560)
Yeah.
Lex Fridman (1:11:10.880)
Cause I work with a lot of automakers
Lex Fridman (1:11:12.400)
and one possible source of money for you,
Lex Fridman (1:11:15.800)
which I'll be excited to see you take on
George Hotz (1:11:18.040)
is basically selling the data.
Lex Fridman (1:11:24.600)
So, which is something that most people,
Lex Fridman (1:11:29.120)
and not selling in a way where here, here at Automaker,
Lex Fridman (1:11:31.800)
but creating, we've done this actually at MIT,
George Hotz (1:11:34.360)
not for money purposes,
Lex Fridman (1:11:35.480)
but you could do it for significant money purposes
Lex Fridman (1:11:37.760)
and make the world a better place by creating a consortia
Lex Fridman (1:11:41.360)
where automakers would pay in
Lex Fridman (1:11:44.200)
and then they get to have free access to the data.
Lex Fridman (1:11:46.960)
And I think a lot of people are really hungry for that
Lex Fridman (1:11:52.400)
and would pay significant amount of money for it.
Lex Fridman (1:11:54.200)
Here's the problem with that.
George Hotz (1:11:55.400)
I like this idea all in theory.
Lex Fridman (1:11:56.880)
It'd be very easy for me to give them access to my servers
Lex Fridman (1:11:59.660)
and we already have all open source tools
Lex Fridman (1:12:01.480)
to access this data.
George Hotz (1:12:02.320)
It's in a great format.
Lex Fridman (1:12:03.440)
We have a great pipeline,
Lex Fridman (1:12:05.640)
but they're gonna put me in the room
Lex Fridman (1:12:07.120)
with some business development guy.
Lex Fridman (1:12:10.140)
And I'm gonna have to talk to this guy
Lex Fridman (1:12:12.560)
and he's not gonna know most of the words I'm saying.
George Hotz (1:12:15.200)
I'm not willing to tolerate that.
Lex Fridman (1:12:17.400)
Okay, Mick Jagger.
George Hotz (1:12:18.960)
No, no, no, no, no.
Lex Fridman (1:12:19.800)
I think I agree with you.
George Hotz (1:12:21.120)
I'm the same way, but you just tell them the terms
Lex Fridman (1:12:23.040)
and there's no discussion needed.
George Hotz (1:12:24.720)
If I could just tell them the terms,
Lex Fridman (1:12:28.080)
Yeah.
Lex Fridman (1:12:28.920)
and like, all right, who wants access to my data?
Lex Fridman (1:12:31.720)
I will sell it to you for, let's say,
Lex Fridman (1:12:36.800)
you want a subscription?
Lex Fridman (1:12:37.720)
I'll sell to you for 100K a month.
George Hotz (1:12:40.800)
Anyone.
Lex Fridman (1:12:41.640)
100K a month.
George Hotz (1:12:42.480)
100K a month.
Lex Fridman (1:12:43.300)
I'll give you access to this data subscription.
George Hotz (1:12:45.160)
Yeah.
Lex Fridman (1:12:46.000)
Yeah, I think that's kind of fair.
George Hotz (1:12:46.820)
Came up with that number off the top of my head.
Lex Fridman (1:12:48.080)
If somebody sends me like a three line email
George Hotz (1:12:50.200)
where it's like, we would like to pay 100K a month
Lex Fridman (1:12:52.600)
to get access to your data.
George Hotz (1:12:54.040)
We would agree to like reasonable privacy terms
Lex Fridman (1:12:56.180)
of the people who are in the data set.
George Hotz (1:12:58.360)
I would be happy to do it,
Lex Fridman (1:12:59.560)
but that's not going to be the email.
George Hotz (1:13:01.200)
The email is going to be, hey,
Lex Fridman (1:13:02.880)
do you have some time in the next month
George Hotz (1:13:04.680)
where we can sit down and we can,
Lex Fridman (1:13:06.000)
I don't have time for that.
George Hotz (1:13:06.880)
We're moving too fast.
Lex Fridman (1:13:07.880)
Yeah.
George Hotz (1:13:08.720)
You could politely respond to that email,
Lex Fridman (1:13:10.080)
but not saying, I don't have any time for your bullshit.
George Hotz (1:13:13.280)
You say, oh, well, unfortunately these are the terms.
Lex Fridman (1:13:15.480)
And so this is, we try to,
George Hotz (1:13:17.720)
we brought the cost down for you
Lex Fridman (1:13:19.840)
in order to minimize the friction and communication.
George Hotz (1:13:22.480)
Absolutely.
Lex Fridman (1:13:23.320)
Here's the, whatever it is,
George Hotz (1:13:24.520)
one, two million dollars a year and you have access.
Lex Fridman (1:13:28.960)
And it's not like I get that email from like,
Lex Fridman (1:13:31.460)
but okay, am I going to reach out?
Lex Fridman (1:13:32.720)
Am I going to hire a business development person
Lex Fridman (1:13:34.200)
who's going to reach out to the automakers?
Lex Fridman (1:13:35.920)
No way.
George Hotz (1:13:36.740)
Yeah. Okay.
Lex Fridman (1:13:37.580)
I got you.
George Hotz (1:13:38.400)
If they reached into me, I'm not going to ignore the email.
Lex Fridman (1:13:40.640)
I'll come back with something like,
George Hotz (1:13:41.960)
yeah, if you're willing to pay 100K a month
Lex Fridman (1:13:43.920)
for access to the data, I'm happy to set that up.
George Hotz (1:13:46.120)
That's worth my engineering time.
Lex Fridman (1:13:48.240)
That's actually quite insightful of you.
George Hotz (1:13:49.560)
You're right.
Lex Fridman (1:13:50.480)
Probably because many of the automakers
George Hotz (1:13:52.520)
are quite a bit old school,
Lex Fridman (1:13:54.200)
there will be a need to reach out and they want it,
Lex Fridman (1:13:57.680)
but there'll need to be some communication.
Lex Fridman (1:13:59.800)
You're right.
George Hotz (1:14:00.640)
Mobileye circa 2015 had the lowest R&D spend
Lex Fridman (1:14:04.520)
of any chip maker, like per, per,
Lex Fridman (1:14:08.360)
and you look at all the people who work for them
Lex Fridman (1:14:10.640)
and it's all business development people
George Hotz (1:14:12.120)
because the car companies are impossible to work with.
Lex Fridman (1:14:15.340)
Yeah.
Lex Fridman (1:14:16.180)
So you're, you have no patience for that
Lex Fridman (1:14:17.880)
and you're, you're legit Android, huh?
Lex Fridman (1:14:20.040)
I have something to do, right?
Lex Fridman (1:14:21.400)
Like, like it's not like, it's not like,
George Hotz (1:14:22.520)
I don't, like, I don't mean to like be a dick
Lex Fridman (1:14:23.760)
and say like, I don't have patience for that,
Lex Fridman (1:14:25.120)
but it's like that stuff doesn't help us
Lex Fridman (1:14:28.280)
with our goal of winning self driving cars.
George Hotz (1:14:30.560)
If I want money in the short term,
Lex Fridman (1:14:33.800)
if I showed off like the actual,
George Hotz (1:14:36.080)
like the learning tech that we have,
Lex Fridman (1:14:38.080)
it's, it's somewhat sad.
George Hotz (1:14:39.580)
Like it's years and years ahead of everybody else's.
Lex Fridman (1:14:42.720)
Not to, maybe not Tesla's.
George Hotz (1:14:43.720)
I think Tesla has some more stuff to us actually.
Lex Fridman (1:14:45.280)
Yeah.
George Hotz (1:14:46.120)
I think Tesla has similar stuff,
Lex Fridman (1:14:46.940)
but when you compare it to like
Lex Fridman (1:14:47.940)
what the Toyota Research Institute has,
Lex Fridman (1:14:50.800)
you're not even close to what we have.
George Hotz (1:14:53.480)
No comments.
Lex Fridman (1:14:54.360)
But I also can't, I have to take your comments.
George Hotz (1:14:58.480)
I intuitively believe you,
Lex Fridman (1:15:01.640)
but I have to take it with a grain of salt
George Hotz (1:15:03.240)
because I mean, you are an inspiration
Lex Fridman (1:15:06.200)
because you basically don't care about a lot of things
George Hotz (1:15:09.040)
that other companies care about.
Lex Fridman (1:15:10.840)
You don't try to bullshit in a sense,
George Hotz (1:15:15.560)
like make up stuff.
Lex Fridman (1:15:16.640)
So to drive up valuation, you're really very real
Lex Fridman (1:15:19.720)
and you're trying to solve the problem
Lex Fridman (1:15:20.920)
and admire that a lot.
Lex Fridman (1:15:22.280)
What I don't necessarily fully can't trust you on,
Lex Fridman (1:15:25.960)
with all due respect, is how good it is, right?
George Hotz (1:15:28.440)
I can only, but I also know how bad others are.
Lex Fridman (1:15:32.460)
And so.
Lex Fridman (1:15:33.300)
I'll say two things about, trust but verify, right?
Lex Fridman (1:15:36.680)
I'll say two things about that.
George Hotz (1:15:38.040)
One is try, get in a 2020 Corolla
Lex Fridman (1:15:42.360)
and try open pilot 0.6 when it comes out next month.
George Hotz (1:15:46.680)
I think already you'll look at this
Lex Fridman (1:15:48.400)
and you'll be like, this is already really good.
Lex Fridman (1:15:51.200)
And then I could be doing that all with hand labelers
Lex Fridman (1:15:54.280)
and all with like the same approach that Mobileye uses.
George Hotz (1:15:57.960)
When we release a model that no longer has the lanes in it,
Lex Fridman (1:16:01.440)
that only outputs a path,
George Hotz (1:16:04.960)
then think about how we did that machine learning
Lex Fridman (1:16:08.680)
and then right away when you see,
Lex Fridman (1:16:10.080)
and that's gonna be an open pilot,
Lex Fridman (1:16:11.240)
that's gonna be an open pilot before 1.0.
George Hotz (1:16:13.000)
When you see that model,
Lex Fridman (1:16:14.080)
you'll know that everything I'm saying is true
Lex Fridman (1:16:15.400)
because how else did I get that model?
Lex Fridman (1:16:16.840)
Good.
George Hotz (1:16:17.680)
You know what I'm saying is true about the simulator.
Lex Fridman (1:16:19.240)
Yeah, yeah, this is super exciting, that's super exciting.
Lex Fridman (1:16:22.680)
But like, you know, I listened to your talk with Kyle
Lex Fridman (1:16:25.760)
and Kyle was originally building the aftermarket system
Lex Fridman (1:16:30.460)
and he gave up on it because of technical challenges,
Lex Fridman (1:16:34.920)
because of the fact that he's gonna have to support
George Hotz (1:16:38.160)
20 to 50 cars, we support 45,
Lex Fridman (1:16:40.520)
because what is he gonna do
Lex Fridman (1:16:41.480)
when the manufacturer ABS system triggers?
Lex Fridman (1:16:43.460)
We have alerts and warnings to deal with all of that
Lex Fridman (1:16:45.520)
and all the cars.
Lex Fridman (1:16:46.600)
And how is he going to formally verify it?
George Hotz (1:16:48.440)
Well, I got 10 million miles of data,
Lex Fridman (1:16:49.840)
it's probably better,
George Hotz (1:16:50.680)
it's probably better verified than the spec.
Lex Fridman (1:16:53.240)
Yeah, I'm glad you're here talking to me.
George Hotz (1:16:57.000)
This is, I'll remember this day,
Lex Fridman (1:17:00.280)
because it's interesting.
George Hotz (1:17:01.120)
If you look at Kyle's from cruise,
Lex Fridman (1:17:04.140)
I'm sure they have a large number
George Hotz (1:17:05.320)
of business development folks
Lex Fridman (1:17:07.400)
and you work with, he's working with GM,
George Hotz (1:17:10.200)
you could work with Argo AI, working with Ford.
Lex Fridman (1:17:13.240)
It's interesting because chances that you fail,
George Hotz (1:17:17.560)
business wise, like bankrupt, are pretty high.
Lex Fridman (1:17:20.160)
Yeah.
Lex Fridman (1:17:21.080)
And yet, it's the Android model,
Lex Fridman (1:17:23.880)
is you're actually taking on the problem.
Lex Fridman (1:17:26.340)
So that's really inspiring, I mean.
Lex Fridman (1:17:28.200)
Well, I have a long term way for Comma to make money too.
Lex Fridman (1:17:30.920)
And one of the nice things
Lex Fridman (1:17:32.180)
when you really take on the problem,
George Hotz (1:17:34.400)
which is my hope for Autopilot, for example,
Lex Fridman (1:17:36.760)
is things you don't expect,
George Hotz (1:17:39.560)
ways to make money or create value
Lex Fridman (1:17:41.840)
that you don't expect will pop up.
George Hotz (1:17:43.960)
Oh, I've known how to do it since kind of,
Lex Fridman (1:17:46.640)
2017 is the first time I said it.
Lex Fridman (1:17:48.520)
Which part, to know how to do which part?
Lex Fridman (1:17:50.440)
Our long term plan is to be a car insurance company.
George Hotz (1:17:52.480)
Insurance, yeah, I love it, yep, yep.
Lex Fridman (1:17:55.280)
I make driving twice as safe.
George Hotz (1:17:56.640)
Not only that, I have the best data
Lex Fridman (1:17:57.600)
such to know who statistically is the safest drivers.
Lex Fridman (1:17:59.840)
And oh, oh, we see you, we see you driving unsafely,
Lex Fridman (1:18:03.700)
we're not gonna insure you.
Lex Fridman (1:18:05.320)
And that causes a bifurcation in the market
Lex Fridman (1:18:08.960)
because the only people who can't get Comma insurance
George Hotz (1:18:10.880)
are the bad drivers, Geico can insure them,
Lex Fridman (1:18:12.740)
their premiums are crazy high,
George Hotz (1:18:13.860)
our premiums are crazy low.
Lex Fridman (1:18:15.320)
We'll win car insurance, take over that whole market.
George Hotz (1:18:18.040)
Okay, so.
Lex Fridman (1:18:19.920)
If we win, if we win.
Lex Fridman (1:18:21.240)
But that's what I'm saying,
Lex Fridman (1:18:22.080)
how do you turn Comma into a $10 billion company?
George Hotz (1:18:24.080)
It's that.
Lex Fridman (1:18:24.920)
That's right.
Lex Fridman (1:18:25.740)
So you, Elon Musk, who else?
Lex Fridman (1:18:29.960)
Who else is thinking like this and working like this
Lex Fridman (1:18:32.700)
in your view?
Lex Fridman (1:18:33.540)
Who are the competitors?
George Hotz (1:18:34.760)
Are there people seriously,
Lex Fridman (1:18:36.120)
I don't think anyone that I'm aware of
George Hotz (1:18:38.280)
is seriously taking on lane keeping,
Lex Fridman (1:18:42.960)
like where it's a huge business
George Hotz (1:18:45.100)
that turns eventually into full autonomy
Lex Fridman (1:18:47.160)
that then creates, yeah, like that creates other businesses
George Hotz (1:18:52.000)
on top of it and so on.
Lex Fridman (1:18:53.400)
Thinks insurance, thinks all kinds of ideas like that.
Lex Fridman (1:18:56.460)
Do you know anyone else thinking like this?
Lex Fridman (1:19:00.480)
Not really.
George Hotz (1:19:02.140)
That's interesting.
Lex Fridman (1:19:02.980)
I mean, my sense is everybody turns to that
George Hotz (1:19:05.320)
in like four or five years.
Lex Fridman (1:19:07.760)
Like Ford, once the autonomy doesn't fall through.
George Hotz (1:19:10.400)
Yeah.
Lex Fridman (1:19:11.240)
But at this time.
George Hotz (1:19:12.560)
Elon's the iOS.
Lex Fridman (1:19:14.100)
By the way, he paved the way for all of us.
George Hotz (1:19:16.680)
It's the iOS, true.
Lex Fridman (1:19:17.960)
I would not be doing Comma AI today
George Hotz (1:19:20.840)
if it was not for those conversations with Elon.
Lex Fridman (1:19:23.440)
And if it were not for him saying like,
George Hotz (1:19:27.080)
I think he said like,
Lex Fridman (1:19:27.900)
well, obviously we're not gonna use LiDAR,
George Hotz (1:19:29.120)
we use cameras, humans use cameras.
Lex Fridman (1:19:31.260)
So what do you think about that?
Lex Fridman (1:19:32.560)
How important is LiDAR?
Lex Fridman (1:19:33.880)
Everybody else on L5 is using LiDAR.
Lex Fridman (1:19:36.920)
What are your thoughts on his provocative statement
Lex Fridman (1:19:39.200)
that LiDAR is a crutch?
George Hotz (1:19:41.320)
See, sometimes he'll say dumb things,
Lex Fridman (1:19:43.040)
like the driver monitoring thing,
Lex Fridman (1:19:44.040)
but sometimes he'll say absolutely, completely,
Lex Fridman (1:19:46.240)
100% obviously true things.
George Hotz (1:19:48.360)
Of course LiDAR is a crutch.
Lex Fridman (1:19:50.800)
It's not even a good crutch.
George Hotz (1:19:53.020)
You're not even using it.
Lex Fridman (1:19:53.860)
Oh, they're using it for localization.
George Hotz (1:19:56.240)
Yeah.
Lex Fridman (1:19:57.080)
Which isn't good in the first place.
George Hotz (1:19:58.140)
If you have to localize your car to centimeters
Lex Fridman (1:20:00.480)
in order to drive, like that's not driving.
George Hotz (1:20:04.280)
Currently not doing much machine learning
Lex Fridman (1:20:06.280)
I thought for LiDAR data.
George Hotz (1:20:07.560)
Meaning like to help you in the task of,
Lex Fridman (1:20:11.320)
general task of perception.
George Hotz (1:20:12.840)
The main goal of those LiDARs on those cars
Lex Fridman (1:20:15.320)
I think is actually localization more than perception.
George Hotz (1:20:18.840)
Or at least that's what they use them for.
Lex Fridman (1:20:20.080)
Yeah, that's true.
George Hotz (1:20:20.920)
If you want to localize to centimeters,
Lex Fridman (1:20:22.480)
you can't use GPS.
George Hotz (1:20:23.680)
The fanciest GPS in the world can't do it.
Lex Fridman (1:20:25.120)
Especially if you're under tree cover and stuff.
George Hotz (1:20:26.920)
With LiDAR you can do this pretty easily.
Lex Fridman (1:20:28.440)
So you really, they're not taking on,
George Hotz (1:20:30.200)
I mean in some research they're using it for perception,
Lex Fridman (1:20:33.160)
but, and they're certainly not, which is sad,
George Hotz (1:20:35.800)
they're not fusing it well with vision.
Lex Fridman (1:20:38.660)
They do use it for perception.
George Hotz (1:20:40.520)
I'm not saying they don't use it for perception,
Lex Fridman (1:20:42.360)
but the thing that, they have vision based
Lex Fridman (1:20:45.440)
and radar based perception systems as well.
Lex Fridman (1:20:47.640)
You could remove the LiDAR and keep around
George Hotz (1:20:51.400)
a lot of the dynamic object perception.
Lex Fridman (1:20:54.000)
You want to get centimeter accurate localization?
George Hotz (1:20:56.280)
Good luck doing that with anything else.
Lex Fridman (1:20:59.080)
So what should Cruz, Waymo do?
Lex Fridman (1:21:02.840)
Like what would be your advice to them now?
Lex Fridman (1:21:06.360)
I mean Waymo is actually,
George Hotz (1:21:08.480)
they're, I mean they're doing, they're serious.
Lex Fridman (1:21:11.640)
Waymo out of the ball of them are quite
Lex Fridman (1:21:14.120)
so serious about the long game.
Lex Fridman (1:21:16.280)
If L5 is a lot, requires 50 years,
George Hotz (1:21:20.800)
I think Waymo will be the only one left standing at the end
Lex Fridman (1:21:24.160)
with the, given the financial backing that they have.
George Hotz (1:21:26.840)
Buku Google bucks.
Lex Fridman (1:21:28.800)
I'll say nice things about both Waymo and Cruz.
George Hotz (1:21:32.560)
Let's do it.
Lex Fridman (1:21:33.640)
Nice is good.
George Hotz (1:21:35.880)
Waymo is by far the furthest along with technology.
Lex Fridman (1:21:39.360)
Waymo has a three to five year lead on all the competitors.
George Hotz (1:21:44.000)
If that, if the Waymo looking stack works,
Lex Fridman (1:21:48.720)
maybe three year lead.
George Hotz (1:21:49.760)
If the Waymo looking stack works,
Lex Fridman (1:21:51.320)
they have a three year lead.
George Hotz (1:21:52.880)
Now I argue that Waymo has spent too much money
Lex Fridman (1:21:55.840)
to recapitalize, to gain back their losses
George Hotz (1:21:59.280)
in those three years.
Lex Fridman (1:22:00.200)
Also self driving cars have no network effect like that.
George Hotz (1:22:03.680)
Uber has a network effect.
Lex Fridman (1:22:04.840)
You have a market, you have drivers and you have riders.
George Hotz (1:22:07.160)
Self driving cars, you have capital and you have riders.
Lex Fridman (1:22:09.960)
There's no network effect.
George Hotz (1:22:11.480)
If I want to blanket a new city in self driving cars,
Lex Fridman (1:22:13.880)
I buy the off the shelf Chinese knockoff self driving cars
Lex Fridman (1:22:16.080)
and I buy enough of them in the city.
Lex Fridman (1:22:17.240)
I can't do that with drivers.
Lex Fridman (1:22:18.400)
And that's why Uber has a first mover advantage
Lex Fridman (1:22:20.920)
that no self driving car company will.
Lex Fridman (1:22:24.040)
Can you disentangle that a little bit?
Lex Fridman (1:22:26.600)
Uber, you're not talking about Uber,
George Hotz (1:22:28.200)
the autonomous vehicle Uber.
Lex Fridman (1:22:29.280)
You're talking about the Uber car, the, yeah.
George Hotz (1:22:31.640)
I'm Uber.
Lex Fridman (1:22:32.480)
I open for business in Austin, Texas, let's say.
George Hotz (1:22:36.000)
I need to attract both sides of the market.
Lex Fridman (1:22:38.880)
I need to both get drivers on my platform
Lex Fridman (1:22:41.320)
and riders on my platform.
Lex Fridman (1:22:42.880)
And I need to keep them both sufficiently happy, right?
George Hotz (1:22:45.400)
Riders aren't gonna use it
Lex Fridman (1:22:46.640)
if it takes more than five minutes for an Uber to show up.
George Hotz (1:22:49.080)
Drivers aren't gonna use it
Lex Fridman (1:22:50.240)
if they have to sit around all day and there's no riders.
Lex Fridman (1:22:52.280)
So you have to carefully balance a market.
Lex Fridman (1:22:54.600)
And whenever you have to carefully balance a market,
George Hotz (1:22:56.400)
there's a great first mover advantage
Lex Fridman (1:22:58.400)
because there's a switching cost for everybody, right?
George Hotz (1:23:01.120)
The drivers and the riders
Lex Fridman (1:23:02.240)
would have to switch at the same time.
George Hotz (1:23:04.200)
Let's even say that, you know, let's say a Luber shows up
Lex Fridman (1:23:08.960)
and Luber somehow, you know, agrees to do things
George Hotz (1:23:12.640)
at a bigger, you know, we're just gonna,
Lex Fridman (1:23:15.800)
we've done it more efficiently, right?
George Hotz (1:23:17.520)
Luber is only takes 5% of a cut
Lex Fridman (1:23:19.880)
instead of the 10% that Uber takes.
George Hotz (1:23:21.680)
No one is gonna switch
Lex Fridman (1:23:22.840)
because the switching cost is higher than that 5%.
Lex Fridman (1:23:25.000)
So you actually can, in markets like that,
Lex Fridman (1:23:27.280)
you have a first mover advantage.
George Hotz (1:23:28.640)
Yeah.
Lex Fridman (1:23:30.240)
Autonomous vehicles of the level five variety
George Hotz (1:23:32.800)
have no first mover advantage.
Lex Fridman (1:23:34.600)
If the technology becomes commoditized,
George Hotz (1:23:36.840)
say I wanna go to a new city, look at the scooters.
Lex Fridman (1:23:39.600)
It's gonna look a lot more like scooters.
George Hotz (1:23:41.560)
Every person with a checkbook
Lex Fridman (1:23:44.080)
can blanket a city in scooters.
Lex Fridman (1:23:45.800)
And that's why you have 10 different scooter companies.
Lex Fridman (1:23:47.960)
Which one's gonna win?
George Hotz (1:23:48.800)
It's a race to the bottom.
Lex Fridman (1:23:49.680)
It's a terrible market to be in
George Hotz (1:23:51.120)
because there's no market for scooters.
Lex Fridman (1:23:55.000)
And the scooters don't get a say
George Hotz (1:23:56.600)
in whether they wanna be bought and deployed to a city
Lex Fridman (1:23:58.240)
or not. Right.
Lex Fridman (1:23:59.080)
So the, yeah.
Lex Fridman (1:24:00.120)
We're gonna entice the scooters
George Hotz (1:24:01.360)
with subsidies and deals and.
Lex Fridman (1:24:03.920)
So whenever you have to invest that capital,
George Hotz (1:24:05.920)
it doesn't.
Lex Fridman (1:24:06.840)
It doesn't come back.
George Hotz (1:24:07.760)
Yeah.
Lex Fridman (1:24:08.960)
That can't be your main criticism of the Waymo approach.
George Hotz (1:24:12.400)
Oh, I'm saying even if it does technically work.
Lex Fridman (1:24:14.920)
Even if it does technically work, that's a problem.
George Hotz (1:24:17.120)
Yeah.
Lex Fridman (1:24:18.400)
I don't know if I were to say,
George Hotz (1:24:21.000)
I would say you're already there.
Lex Fridman (1:24:23.600)
I haven't even thought about that,
Lex Fridman (1:24:24.640)
but I would say the bigger challenge
Lex Fridman (1:24:26.600)
is the technical approach.
George Hotz (1:24:28.000)
The.
Lex Fridman (1:24:29.800)
So Waymo's cruises.
Lex Fridman (1:24:31.880)
And not just the technical approach,
Lex Fridman (1:24:33.000)
but of creating value.
George Hotz (1:24:34.800)
I still don't understand how you beat Uber,
Lex Fridman (1:24:40.760)
the human driven cars.
George Hotz (1:24:43.480)
In terms of financially,
Lex Fridman (1:24:44.920)
it doesn't make sense to me
George Hotz (1:24:47.160)
that people wanna get in an autonomous vehicle.
Lex Fridman (1:24:50.120)
I don't understand how you make money.
George Hotz (1:24:52.800)
In the longterm, yes.
Lex Fridman (1:24:54.280)
Like real longterm.
Lex Fridman (1:24:56.440)
But it just feels like there's too much
Lex Fridman (1:24:58.600)
capital investment needed.
George Hotz (1:24:59.960)
Oh, and they're gonna be worse than Ubers
Lex Fridman (1:25:01.160)
because they're gonna stop for every little thing,
George Hotz (1:25:04.040)
everywhere.
Lex Fridman (1:25:06.280)
I'll say a nice thing about cruise.
George Hotz (1:25:07.280)
That was my nice thing about Waymo.
Lex Fridman (1:25:08.360)
They're three years ahead.
Lex Fridman (1:25:09.200)
Wait, what was the nice?
Lex Fridman (1:25:10.040)
Oh, because they're three.
George Hotz (1:25:10.880)
They're three years technically ahead of everybody.
Lex Fridman (1:25:12.400)
Their tech stack is great.
George Hotz (1:25:14.720)
My nice thing about cruise is GM buying them
Lex Fridman (1:25:17.840)
was a great move for GM.
George Hotz (1:25:20.520)
For $1 billion,
Lex Fridman (1:25:22.200)
GM bought an insurance policy against Waymo.
George Hotz (1:25:25.520)
They put, cruise is three years behind Waymo.
Lex Fridman (1:25:30.920)
That means Google will get a monopoly on the technology
George Hotz (1:25:33.240)
for at most three years.
Lex Fridman (1:25:36.800)
And if technology works,
Lex Fridman (1:25:38.800)
so you might not even be right about the three years,
Lex Fridman (1:25:40.760)
it might be less.
George Hotz (1:25:41.800)
Might be less.
Lex Fridman (1:25:42.640)
Cruise actually might not be that far behind.
George Hotz (1:25:44.240)
I don't know how much Waymo has waffled around
Lex Fridman (1:25:47.280)
or how much of it actually is just that long tail.
George Hotz (1:25:49.720)
Yeah, okay.
Lex Fridman (1:25:50.560)
If that's the best you could say in terms of nice things,
George Hotz (1:25:53.520)
that's more of a nice thing for GM
Lex Fridman (1:25:55.160)
that that's the smart insurance policy.
George Hotz (1:25:58.560)
It's a smart insurance policy.
Lex Fridman (1:25:59.640)
I mean, I think that's how,
George Hotz (1:26:01.840)
I can't see cruise working out any other.
Lex Fridman (1:26:05.160)
For cruise to leapfrog Waymo would really surprise me.
George Hotz (1:26:10.360)
Yeah, so let's talk about
Lex Fridman (1:26:12.000)
the underlying assumption of everything is.
George Hotz (1:26:13.600)
We're not gonna leapfrog Tesla.
Lex Fridman (1:26:17.520)
Tesla would have to seriously mess up for us
George Hotz (1:26:19.440)
because you're.
Lex Fridman (1:26:20.400)
Okay, so the way you leapfrog, right?
George Hotz (1:26:23.200)
Is you come up with an idea
Lex Fridman (1:26:26.080)
or you take a direction perhaps secretly
George Hotz (1:26:28.560)
that the other people aren't taking.
Lex Fridman (1:26:31.640)
And so the cruise, Waymo,
George Hotz (1:26:35.000)
even Aurora.
Lex Fridman (1:26:38.080)
I don't know Aurora, Zooks is the same stack as well.
George Hotz (1:26:40.080)
They're all the same code base even.
Lex Fridman (1:26:41.720)
And they're all the same DARPA Urban Challenge code base.
Lex Fridman (1:26:45.360)
So the question is,
Lex Fridman (1:26:46.800)
do you think there's a room for brilliance and innovation
Lex Fridman (1:26:48.960)
that will change everything?
Lex Fridman (1:26:50.360)
Like say, okay, so I'll give you examples.
George Hotz (1:26:53.880)
It could be if revolution and mapping, for example,
Lex Fridman (1:26:59.600)
that allow you to map things,
George Hotz (1:27:03.000)
do HD maps of the whole world,
Lex Fridman (1:27:05.800)
all weather conditions somehow really well,
George Hotz (1:27:08.040)
or revolution and simulation
Lex Fridman (1:27:13.040)
to where the all the way you said before becomes incorrect.
George Hotz (1:27:20.480)
That kind of thing.
Lex Fridman (1:27:21.320)
Any room for breakthrough innovation?
Lex Fridman (1:27:24.920)
What I said before about,
Lex Fridman (1:27:25.960)
oh, they actually get the whole thing.
George Hotz (1:27:27.160)
Well, I'll say this about,
Lex Fridman (1:27:30.480)
we divide driving into three problems
Lex Fridman (1:27:32.640)
and I actually haven't solved the third yet,
Lex Fridman (1:27:33.800)
but I haven't had you how to do it.
Lex Fridman (1:27:34.800)
So there's the static.
Lex Fridman (1:27:36.120)
The static driving problem is assuming
Lex Fridman (1:27:38.000)
you are the only car on the road, right?
Lex Fridman (1:27:40.120)
And this problem can be solved 100%
George Hotz (1:27:41.960)
with mapping and localization.
Lex Fridman (1:27:43.920)
This is why farms work the way they do.
George Hotz (1:27:45.680)
If all you have to deal with is the static problem
Lex Fridman (1:27:48.360)
and you can statically schedule your machines, right?
George Hotz (1:27:50.120)
It's the same as like statically scheduling processes.
Lex Fridman (1:27:52.640)
You can statically schedule your tractors
Lex Fridman (1:27:53.960)
to never hit each other on their paths, right?
Lex Fridman (1:27:56.080)
Cause they know the speed they go at.
Lex Fridman (1:27:57.440)
So that's the static driving problem.
Lex Fridman (1:28:00.080)
Maps only helps you with the static driving problem.
George Hotz (1:28:03.880)
Yeah, the question about static driving,
Lex Fridman (1:28:06.880)
you've just made it sound like it's really easy.
George Hotz (1:28:08.720)
Static driving is really easy.
Lex Fridman (1:28:11.880)
How easy?
George Hotz (1:28:13.040)
How, well, cause the whole drifting out of lane,
Lex Fridman (1:28:16.440)
when Tesla drifts out of lane,
George Hotz (1:28:18.720)
it's failing on the fundamental static driving problem.
Lex Fridman (1:28:22.000)
Tesla is drifting out of lane?
George Hotz (1:28:24.440)
The static driving problem is not easy for the world.
Lex Fridman (1:28:27.720)
The static driving problem is easy for one route.
George Hotz (1:28:31.840)
One route and one weather condition
Lex Fridman (1:28:33.920)
with one state of lane markings
Lex Fridman (1:28:37.920)
and like no deterioration, no cracks in the road.
Lex Fridman (1:28:40.880)
No, I'm assuming you have a perfect localizer.
Lex Fridman (1:28:42.600)
So that's solved for the weather condition
Lex Fridman (1:28:44.200)
and the lane marking condition.
Lex Fridman (1:28:45.600)
But that's the problem is,
Lex Fridman (1:28:46.600)
how do you have a perfect localizer?
George Hotz (1:28:48.400)
Perfect localizers are not that hard to build.
Lex Fridman (1:28:50.560)
Okay, come on now, with LIDAR?
George Hotz (1:28:53.320)
With LIDAR, yeah.
Lex Fridman (1:28:54.160)
Oh, with LIDAR, okay.
Lex Fridman (1:28:55.000)
With LIDAR, yeah, but you use LIDAR, right?
Lex Fridman (1:28:56.400)
Like use LIDAR, build a perfect localizer.
George Hotz (1:28:58.600)
Building a perfect localizer without LIDAR,
Lex Fridman (1:29:02.960)
it's gonna be hard.
George Hotz (1:29:04.280)
You can get 10 centimeters without LIDAR,
Lex Fridman (1:29:05.720)
you can get one centimeter with LIDAR.
George Hotz (1:29:07.200)
I'm not even concerned about the one or 10 centimeters.
Lex Fridman (1:29:09.240)
I'm concerned if every once in a while,
George Hotz (1:29:11.160)
you're just way off.
Lex Fridman (1:29:12.640)
Yeah, so this is why you have to carefully make sure
George Hotz (1:29:17.920)
you're always tracking your position.
Lex Fridman (1:29:19.960)
You wanna use LIDAR camera fusion,
Lex Fridman (1:29:21.680)
but you can get the reliability of that system
Lex Fridman (1:29:24.400)
up to 100,000 miles,
Lex Fridman (1:29:27.960)
and then you write some fallback condition
Lex Fridman (1:29:29.640)
where it's not that bad if you're way off, right?
George Hotz (1:29:32.120)
I think that you can get it to the point,
Lex Fridman (1:29:33.720)
it's like ASLD that you're never in a case
George Hotz (1:29:36.760)
where you're way off and you don't know it.
Lex Fridman (1:29:38.440)
Yeah, okay, so this is brilliant.
Lex Fridman (1:29:40.200)
So that's the static. Static.
Lex Fridman (1:29:42.240)
We can, especially with LIDAR and good HG maps,
George Hotz (1:29:45.920)
you can solve that problem. Easy.
Lex Fridman (1:29:47.680)
No, I just disagree with your word easy.
George Hotz (1:29:50.440)
The static problem's so easy.
Lex Fridman (1:29:51.760)
It's very typical for you to say something is easy.
George Hotz (1:29:54.000)
I got it. No.
Lex Fridman (1:29:54.840)
It's not as challenging as the other ones, okay.
George Hotz (1:29:56.880)
Well, okay, maybe it's obvious how to solve it.
Lex Fridman (1:29:58.760)
The third one's the hardest.
Lex Fridman (1:30:00.320)
And a lot of people don't even think about the third one
Lex Fridman (1:30:01.880)
and even see it as different from the second one.
Lex Fridman (1:30:03.640)
So the second one is dynamic.
Lex Fridman (1:30:05.720)
The second one is like, say there's an obvious example
Lex Fridman (1:30:08.520)
is like a car stopped at a red light, right?
Lex Fridman (1:30:10.360)
You can't have that car in your map
George Hotz (1:30:12.520)
because you don't know whether that car
Lex Fridman (1:30:13.720)
is gonna be there or not.
Lex Fridman (1:30:14.880)
So you have to detect that car in real time
Lex Fridman (1:30:17.960)
and then you have to do the appropriate action, right?
George Hotz (1:30:21.600)
Also, that car is not a fixed object.
Lex Fridman (1:30:24.800)
That car may move and you have to predict
Lex Fridman (1:30:26.600)
what that car will do, right?
Lex Fridman (1:30:28.680)
So this is the dynamic problem.
George Hotz (1:30:30.840)
Yeah.
Lex Fridman (1:30:31.680)
So you have to deal with this.
George Hotz (1:30:32.800)
This involves, again, like you're gonna need models
Lex Fridman (1:30:36.640)
of other people's behavior.
George Hotz (1:30:39.080)
Are you including in that,
Lex Fridman (1:30:40.320)
I don't wanna step on the third one.
George Hotz (1:30:42.320)
Oh.
Lex Fridman (1:30:43.160)
But are you including in that your influence on people?
George Hotz (1:30:46.920)
Ah, that's the third one.
Lex Fridman (1:30:48.240)
Okay.
George Hotz (1:30:49.080)
That's the third one.
Lex Fridman (1:30:49.920)
We call it the counterfactual.
George Hotz (1:30:51.840)
Yeah, brilliant.
Lex Fridman (1:30:52.680)
And that.
George Hotz (1:30:53.520)
I just talked to Judea Pearl
Lex Fridman (1:30:54.360)
who's obsessed with counterfactuals.
Lex Fridman (1:30:55.800)
And the counterfactual.
Lex Fridman (1:30:56.640)
Oh yeah, yeah, I read his books.
Lex Fridman (1:30:58.600)
So the static and the dynamic
Lex Fridman (1:31:00.760)
Yeah.
George Hotz (1:31:01.960)
Our approach right now for lateral
Lex Fridman (1:31:04.720)
will scale completely to the static and dynamic.
George Hotz (1:31:07.560)
The counterfactual, the only way I have to do it yet,
Lex Fridman (1:31:10.720)
the thing that I wanna do once we have all of these cars
George Hotz (1:31:13.960)
is I wanna do reinforcement learning on the world.
Lex Fridman (1:31:16.760)
I'm always gonna turn the exploiter up to max.
George Hotz (1:31:18.880)
I'm not gonna have them explore.
Lex Fridman (1:31:20.440)
But the only real way to get at the counterfactual
George Hotz (1:31:22.760)
is to do reinforcement learning
Lex Fridman (1:31:24.080)
because the other agents are humans.
Lex Fridman (1:31:27.760)
So that's fascinating that you break it down like that.
Lex Fridman (1:31:30.080)
I agree completely.
George Hotz (1:31:31.720)
I've spent my life thinking about this problem.
Lex Fridman (1:31:33.680)
It's beautiful.
Lex Fridman (1:31:34.520)
And part of it, because you're slightly insane,
Lex Fridman (1:31:37.840)
it's good.
George Hotz (1:31:39.080)
Because.
Lex Fridman (1:31:41.240)
Not my life.
George Hotz (1:31:42.080)
Just the last four years.
Lex Fridman (1:31:43.120)
No, no.
George Hotz (1:31:43.960)
You have some nonzero percent of your brain
Lex Fridman (1:31:48.920)
has a madman in it, which is good.
George Hotz (1:31:51.520)
That's a really good feature.
Lex Fridman (1:31:52.360)
But there's a safety component to it
George Hotz (1:31:55.920)
that I think sort of with counterfactuals and so on
Lex Fridman (1:31:59.040)
that would just freak people out.
Lex Fridman (1:32:00.280)
How do you even start to think about just in general?
Lex Fridman (1:32:03.320)
I mean, you've had some friction with NHTSA and so on.
George Hotz (1:32:07.600)
I am frankly exhausted by safety engineers.
Lex Fridman (1:32:14.280)
The prioritization on safety over innovation
George Hotz (1:32:21.360)
to a degree where it kills, in my view,
Lex Fridman (1:32:23.720)
kills safety in the long term.
Lex Fridman (1:32:26.200)
So the counterfactual thing,
Lex Fridman (1:32:28.080)
they just actually exploring this world
George Hotz (1:32:31.560)
of how do you interact with dynamic objects and so on.
Lex Fridman (1:32:33.880)
How do you think about safety?
George Hotz (1:32:34.840)
You can do reinforcement learning without ever exploring.
Lex Fridman (1:32:38.080)
And I said that, so you can think about your,
George Hotz (1:32:40.400)
in reinforcement learning,
Lex Fridman (1:32:41.520)
it's usually called a temperature parameter.
Lex Fridman (1:32:44.280)
And your temperature parameter
Lex Fridman (1:32:45.320)
is how often you deviate from the argmax.
George Hotz (1:32:48.080)
I could always set that to zero and still learn.
Lex Fridman (1:32:50.680)
And I feel that you'd always want that set to zero
George Hotz (1:32:52.800)
on your actual system.
Lex Fridman (1:32:54.040)
Gotcha.
Lex Fridman (1:32:54.880)
But the problem is you first don't know very much.
Lex Fridman (1:32:58.120)
And so you're going to make mistakes.
Lex Fridman (1:32:59.520)
So the learning, the exploration happens through mistakes.
Lex Fridman (1:33:02.360)
Yeah, but okay.
Lex Fridman (1:33:03.720)
So the consequences of a mistake.
Lex Fridman (1:33:06.080)
Open pilot and autopilot are making mistakes left and right.
George Hotz (1:33:09.400)
We have 700 daily active users,
Lex Fridman (1:33:12.560)
a thousand weekly active users.
George Hotz (1:33:14.040)
Open pilot makes tens of thousands of mistakes a week.
Lex Fridman (1:33:18.920)
These mistakes have zero consequences.
George Hotz (1:33:21.160)
These mistakes are,
Lex Fridman (1:33:22.560)
oh, I wanted to take this exit and it went straight.
Lex Fridman (1:33:26.840)
So I'm just going to carefully touch the wheel.
Lex Fridman (1:33:28.520)
The humans catch them.
George Hotz (1:33:29.360)
The humans catch them.
Lex Fridman (1:33:30.680)
And the human disengagement is labeling
George Hotz (1:33:33.120)
that reinforcement learning
Lex Fridman (1:33:34.160)
in a completely consequence free way.
Lex Fridman (1:33:37.280)
So driver monitoring is the way you ensure they keep.
Lex Fridman (1:33:39.880)
Yes.
George Hotz (1:33:40.720)
They keep paying attention.
Lex Fridman (1:33:42.160)
How is your messaging?
George Hotz (1:33:43.280)
Say I gave you a billion dollars,
Lex Fridman (1:33:45.280)
you would be scaling it now.
George Hotz (1:33:47.840)
Oh, I couldn't scale it with any amount of money.
Lex Fridman (1:33:49.760)
I'd raise money if I could, if I had a way to scale it.
George Hotz (1:33:51.680)
Yeah, you're now not focused on scale.
Lex Fridman (1:33:53.360)
I don't know how to do,
George Hotz (1:33:54.200)
oh, like I guess I could sell it to more people,
Lex Fridman (1:33:55.840)
but I want to make the system better.
George Hotz (1:33:57.040)
Better, better.
Lex Fridman (1:33:57.880)
And I don't know how to, I mean.
Lex Fridman (1:33:58.920)
But what's the messaging here?
Lex Fridman (1:34:01.160)
I got a chance to talk to Elon and he basically said
George Hotz (1:34:06.320)
that the human factor doesn't matter.
Lex Fridman (1:34:09.360)
You know, the human doesn't matter
George Hotz (1:34:10.440)
because the system will perform,
Lex Fridman (1:34:12.360)
there'll be sort of a, sorry to use the term,
Lex Fridman (1:34:14.840)
but like a singular,
Lex Fridman (1:34:15.680)
like a point where it gets just much better.
Lex Fridman (1:34:17.880)
And so the human, it won't really matter.
Lex Fridman (1:34:20.880)
But it seems like that human catching the system
George Hotz (1:34:25.040)
when it gets into trouble is like the thing
Lex Fridman (1:34:29.440)
which will make something like reinforcement learning work.
Lex Fridman (1:34:32.800)
So how do you think messaging for Tesla,
Lex Fridman (1:34:35.680)
for you should change,
Lex Fridman (1:34:36.880)
for the industry in general should change?
Lex Fridman (1:34:39.120)
I think our messaging is pretty clear.
George Hotz (1:34:40.880)
At least like our messaging wasn't that clear
Lex Fridman (1:34:43.120)
in the beginning and I do kind of fault myself for that.
George Hotz (1:34:45.240)
We are proud right now to be a level two system.
Lex Fridman (1:34:48.520)
We are proud to be level two.
George Hotz (1:34:50.400)
If we talk about level four,
Lex Fridman (1:34:51.680)
it's not with the current hardware.
George Hotz (1:34:53.240)
It's not gonna be just a magical OTA upgrade.
Lex Fridman (1:34:55.960)
It's gonna be new hardware.
George Hotz (1:34:57.360)
It's gonna be very carefully thought out.
Lex Fridman (1:34:59.600)
Right now, we are proud to be level two
Lex Fridman (1:35:01.640)
and we have a rigorous safety model.
Lex Fridman (1:35:03.400)
I mean, not like, okay, rigorous, who knows what that means,
Lex Fridman (1:35:06.680)
but we at least have a safety model
Lex Fridman (1:35:08.720)
and we make it explicit as in safety.md in OpenPilot.
Lex Fridman (1:35:11.920)
And it says, seriously though, safety.md.
Lex Fridman (1:35:17.040)
This is brilliant, this is so Android.
George Hotz (1:35:18.680)
Well, this is the safety model
Lex Fridman (1:35:21.880)
and I like to have conversations like,
George Hotz (1:35:25.600)
sometimes people will come to you and they're like,
Lex Fridman (1:35:27.240)
your system's not safe.
Lex Fridman (1:35:29.320)
Okay, have you read my safety docs?
Lex Fridman (1:35:31.160)
Would you like to have an intelligent conversation
Lex Fridman (1:35:32.760)
about this?
Lex Fridman (1:35:33.600)
And the answer is always no.
George Hotz (1:35:34.440)
They just like scream about, it runs Python.
Lex Fridman (1:35:38.280)
Okay, what?
Lex Fridman (1:35:39.120)
So you're saying that because Python's not real time,
Lex Fridman (1:35:41.600)
Python not being real time never causes disengagements.
George Hotz (1:35:44.320)
Disengagements are caused by, the model is QM.
Lex Fridman (1:35:47.720)
But safety.md says the following,
George Hotz (1:35:49.840)
first and foremost,
Lex Fridman (1:35:50.680)
the driver must be paying attention at all times.
George Hotz (1:35:55.400)
I still consider the software to be alpha software
Lex Fridman (1:35:57.760)
until we can actually enforce that statement,
Lex Fridman (1:36:00.120)
but I feel it's very well communicated to our users.
Lex Fridman (1:36:03.320)
Two more things.
George Hotz (1:36:04.560)
One is the user must be able to easily take control
Lex Fridman (1:36:09.120)
of the vehicle at all times.
Lex Fridman (1:36:10.920)
So if you step on the gas or brake with OpenPilot,
Lex Fridman (1:36:14.480)
it gives full manual control back to the user
George Hotz (1:36:16.440)
or press the cancel button.
Lex Fridman (1:36:18.720)
Step two, the car will never react so quickly,
George Hotz (1:36:23.280)
we define so quickly to be about one second,
Lex Fridman (1:36:26.000)
that you can't react in time.
Lex Fridman (1:36:27.640)
And we do this by enforcing torque limits,
Lex Fridman (1:36:29.480)
braking limits and acceleration limits.
Lex Fridman (1:36:31.520)
So we have like our torque limits way lower than Tesla's.
Lex Fridman (1:36:36.520)
This is another potential.
George Hotz (1:36:39.080)
If I could tweak Autopilot,
Lex Fridman (1:36:40.240)
I would lower their torque limit
Lex Fridman (1:36:41.320)
and I would add driver monitoring.
Lex Fridman (1:36:42.960)
Because Autopilot can jerk the wheel hard.
George Hotz (1:36:46.240)
OpenPilot can't.
Lex Fridman (1:36:47.960)
We limit, and all this code is open source, readable.
Lex Fridman (1:36:52.080)
And I believe now it's all Misra C compliant.
Lex Fridman (1:36:54.880)
What's that mean?
George Hotz (1:36:57.080)
Misra is like the automotive coding standard.
Lex Fridman (1:37:00.400)
At first, I've come to respect.
George Hotz (1:37:03.400)
I've been reading like the standards lately
Lex Fridman (1:37:05.000)
and I've come to respect them.
George Hotz (1:37:05.880)
They're actually written by very smart people.
Lex Fridman (1:37:07.800)
Yeah, they're brilliant people actually.
George Hotz (1:37:09.880)
They have a lot of experience.
Lex Fridman (1:37:11.320)
They're sometimes a little too cautious,
Lex Fridman (1:37:13.360)
but in this case, it pays off.
Lex Fridman (1:37:16.800)
Misra is written by like computer scientists.
Lex Fridman (1:37:18.440)
And you can tell by the language they use.
Lex Fridman (1:37:19.840)
You can tell by the language they use,
George Hotz (1:37:21.120)
they talk about like whether certain conditions in Misra
Lex Fridman (1:37:24.480)
are decidable or undecidable.
Lex Fridman (1:37:26.560)
And you mean like the halting problem?
Lex Fridman (1:37:28.360)
And yes, all right, you've earned my respect.
George Hotz (1:37:31.640)
I will read carefully what you have to say
Lex Fridman (1:37:33.120)
and we wanna make our code compliant with that.
George Hotz (1:37:35.760)
All right, so you're proud level two, beautiful.
Lex Fridman (1:37:38.160)
So you were the founder and I think CEO of Kama AI,
George Hotz (1:37:42.360)
then you were the head of research.
Lex Fridman (1:37:44.320)
What the heck are you now?
Lex Fridman (1:37:46.080)
What's your connection to Kama AI?
Lex Fridman (1:37:47.440)
I'm the president, but I'm one of those
George Hotz (1:37:49.400)
like unelected presidents of like a small dictatorship
Lex Fridman (1:37:53.040)
country, not one of those like elected presidents.
George Hotz (1:37:55.160)
Oh, so you're like Putin when he was like the,
Lex Fridman (1:37:57.160)
yeah, I got you.
Lex Fridman (1:37:59.920)
So there's a, what's the governance structure?
Lex Fridman (1:38:02.080)
What's the future of Kama AI?
George Hotz (1:38:04.800)
I mean, yeah, it's a business.
Lex Fridman (1:38:07.440)
Do you want, are you just focused on getting things
George Hotz (1:38:10.000)
right now, making some small amount of money in the meantime
Lex Fridman (1:38:14.880)
and then when it works, it works and you scale.
George Hotz (1:38:17.520)
Our burn rate is about 200K a month
Lex Fridman (1:38:20.440)
and our revenue is about 100K a month.
Lex Fridman (1:38:23.000)
So we need to forex our revenue,
Lex Fridman (1:38:24.880)
but we haven't like tried very hard at that yet.
Lex Fridman (1:38:28.160)
And the revenue is basically selling stuff online.
Lex Fridman (1:38:30.120)
Yeah, we sell stuff shop.kama.ai.
George Hotz (1:38:32.320)
Is there other, well, okay,
Lex Fridman (1:38:33.880)
so you'll have to figure out the revenue.
George Hotz (1:38:35.720)
That's our only, see, but to me,
Lex Fridman (1:38:37.840)
that's like respectable revenues.
George Hotz (1:38:40.360)
We make it by selling products to consumers
Lex Fridman (1:38:42.640)
who are honest and transparent about what they are.
Lex Fridman (1:38:45.960)
Most actually level four companies, right?
Lex Fridman (1:38:50.680)
Cause you could easily start blowing up like smoke,
George Hotz (1:38:54.240)
like overselling the hype and feeding into,
Lex Fridman (1:38:57.040)
getting some fundraisers.
George Hotz (1:38:59.000)
Oh, you're the guy, you're a genius
Lex Fridman (1:39:00.440)
because you hacked the iPhone.
George Hotz (1:39:01.760)
Oh, I hate that, I hate that.
Lex Fridman (1:39:03.280)
Yeah, well, I can trade my social capital for more money.
George Hotz (1:39:06.640)
I did it once, I almost regret it doing it the first time.
Lex Fridman (1:39:10.280)
Well, on a small tangent,
George Hotz (1:39:11.600)
what's your, you seem to not like fame
Lex Fridman (1:39:16.560)
and yet you're also drawn to fame.
Lex Fridman (1:39:18.880)
Where are you on that currently?
Lex Fridman (1:39:24.560)
Have you had some introspection, some soul searching?
George Hotz (1:39:27.200)
Yeah, I actually,
Lex Fridman (1:39:29.240)
I've come to a pretty stable position on that.
George Hotz (1:39:32.200)
Like after the first time,
Lex Fridman (1:39:33.880)
I realized that I don't want attention from the masses.
George Hotz (1:39:36.840)
I want attention from people who I respect.
Lex Fridman (1:39:40.280)
Who do you respect?
George Hotz (1:39:41.960)
I can give a list of people.
Lex Fridman (1:39:43.960)
So are these like Elon Musk type characters?
Lex Fridman (1:39:47.200)
Yeah, well, actually, you know what?
Lex Fridman (1:39:50.000)
I'll make it more broad than that.
George Hotz (1:39:51.240)
I won't make it about a person, I respect skill.
Lex Fridman (1:39:54.040)
I respect people who have skills, right?
Lex Fridman (1:39:56.840)
And I would like to like be, I'm not gonna say famous,
Lex Fridman (1:40:01.400)
but be like known among more people who have like real skills.
Lex Fridman (1:40:06.880)
Who in cars do you think have skill, not do you respect?
Lex Fridman (1:40:15.000)
Oh, Kyle Vogt has skill.
George Hotz (1:40:17.760)
A lot of people at Waymo have skill and I respect them.
Lex Fridman (1:40:20.840)
I respect them as engineers.
George Hotz (1:40:23.760)
Like I can think, I mean,
Lex Fridman (1:40:24.920)
I think about all the times in my life
George Hotz (1:40:26.280)
where I've been like dead set on approaches
Lex Fridman (1:40:27.960)
and they turn out to be wrong.
George Hotz (1:40:29.760)
So, I mean, this might, I might be wrong.
Lex Fridman (1:40:31.720)
I accept that.
George Hotz (1:40:32.600)
I accept that there's a decent chance that I'm wrong.
Lex Fridman (1:40:36.600)
And actually, I mean,
George Hotz (1:40:37.440)
having talked to Chris Hermsons, Sterling Anderson,
Lex Fridman (1:40:39.480)
those guys, I mean, I deeply respect Chris.
George Hotz (1:40:43.360)
I just admire the guy.
Lex Fridman (1:40:46.040)
He's legit.
George Hotz (1:40:47.400)
When you drive a car through the desert
Lex Fridman (1:40:48.960)
when everybody thinks it's impossible, that's legit.
Lex Fridman (1:40:52.440)
And then I also really respect the people
Lex Fridman (1:40:53.840)
who are like writing the infrastructure of the world,
George Hotz (1:40:55.680)
like the Linus Torvalds and the Chris Lattiners.
Lex Fridman (1:40:57.760)
They were doing the real work.
George Hotz (1:40:59.080)
I know, they're doing the real work.
Lex Fridman (1:41:00.520)
This, having talked to Chris,
George Hotz (1:41:03.000)
like Chris Lattiners, you realize,
Lex Fridman (1:41:04.600)
especially when they're humble,
George Hotz (1:41:05.720)
it's like you realize, oh, you guys,
Lex Fridman (1:41:07.720)
we're just using your,
George Hotz (1:41:09.680)
Oh yeah.
Lex Fridman (1:41:10.520)
All the hard work that you did.
George Hotz (1:41:11.560)
Yeah, that's incredible.
Lex Fridman (1:41:13.160)
What do you think, Mr. Anthony Lewandowski,
Lex Fridman (1:41:18.480)
what do you, he's another mad genius.
Lex Fridman (1:41:21.680)
Sharp guy, oh yeah.
Lex Fridman (1:41:22.920)
What, do you think he might long term become a competitor?
Lex Fridman (1:41:27.680)
Oh, to comma?
George Hotz (1:41:28.880)
Well, so I think that he has the other right approach.
Lex Fridman (1:41:32.440)
I think that right now there's two right approaches.
George Hotz (1:41:35.320)
One is what we're doing, and one is what he's doing.
Lex Fridman (1:41:37.720)
Can you describe, I think it's called Pronto AI.
George Hotz (1:41:39.840)
He started a new thing.
Lex Fridman (1:41:40.960)
Do you know what the approach is?
George Hotz (1:41:42.400)
I actually don't know.
Lex Fridman (1:41:43.240)
Embark is also doing the same sort of thing.
George Hotz (1:41:45.080)
The idea is almost that you want to,
Lex Fridman (1:41:47.360)
so if you're, I can't partner with Honda and Toyota.
George Hotz (1:41:51.840)
Honda and Toyota are like 400,000 person companies.
Lex Fridman (1:41:56.840)
It's not even a company at that point.
George Hotz (1:41:58.640)
I don't think of it like, I don't personify it.
Lex Fridman (1:42:00.600)
I think of it like an object,
Lex Fridman (1:42:01.440)
but a trucker drives for a fleet,
Lex Fridman (1:42:06.280)
maybe that has like, some truckers are independent.
George Hotz (1:42:09.480)
Some truckers drive for fleets with a hundred trucks.
Lex Fridman (1:42:11.320)
There are tons of independent trucking companies out there.
George Hotz (1:42:14.160)
Start a trucking company and drive your costs down
Lex Fridman (1:42:17.400)
or figure out how to drive down the cost of trucking.
George Hotz (1:42:23.040)
Another company that I really respect is Nato.
Lex Fridman (1:42:25.800)
Actually, I respect their business model.
George Hotz (1:42:27.800)
Nato sells a driver monitoring camera
Lex Fridman (1:42:31.040)
and they sell it to fleet owners.
George Hotz (1:42:33.360)
If I owned a fleet of cars
Lex Fridman (1:42:35.840)
and I could pay 40 bucks a month to monitor my employees,
George Hotz (1:42:41.840)
this is gonna, it like reduces accidents 18%.
Lex Fridman (1:42:45.040)
It's so like that, in the space,
George Hotz (1:42:48.400)
that is like the business model that I like most respect.
Lex Fridman (1:42:52.840)
Cause they're creating value today.
George Hotz (1:42:54.800)
Yeah, which is a, that's a huge one.
Lex Fridman (1:42:57.880)
How do we create value today with some of this?
Lex Fridman (1:42:59.680)
And the lane keeping thing is huge.
Lex Fridman (1:43:01.720)
And it sounds like you're creeping in
George Hotz (1:43:03.840)
or full steam ahead on the driver monitoring too,
Lex Fridman (1:43:06.720)
which I think actually where the short term value,
George Hotz (1:43:09.280)
if you can get it right.
Lex Fridman (1:43:10.520)
I still, I'm not a huge fan of the statement
George Hotz (1:43:12.840)
that everything has to have driver monitoring.
Lex Fridman (1:43:15.160)
I agree with that completely,
Lex Fridman (1:43:16.160)
but that statement usually misses the point
Lex Fridman (1:43:18.720)
that to get the experience of it right is not trivial.
George Hotz (1:43:21.960)
Oh no, not at all.
Lex Fridman (1:43:22.880)
In fact, like, so right now we have,
George Hotz (1:43:26.200)
I think the timeout depends on speed of the car,
Lex Fridman (1:43:29.600)
but we want to depend on like the scene state.
George Hotz (1:43:32.520)
If you're on like an empty highway,
Lex Fridman (1:43:35.440)
it's very different if you don't pay attention
George Hotz (1:43:37.720)
than if like you're like coming up to a traffic light.
Lex Fridman (1:43:42.080)
And longterm, it should probably learn from the driver
George Hotz (1:43:45.720)
because that's to do, I watched a lot of video.
Lex Fridman (1:43:48.120)
We've built a smartphone detector
George Hotz (1:43:49.520)
just to analyze how people are using smartphones
Lex Fridman (1:43:51.600)
and people are using it very differently.
George Hotz (1:43:53.400)
It's a texting styles.
Lex Fridman (1:43:57.720)
There's.
George Hotz (1:43:58.560)
We haven't watched nearly enough of the videos.
Lex Fridman (1:44:00.280)
We haven't, I got millions of miles
George Hotz (1:44:01.760)
of people driving cars.
Lex Fridman (1:44:02.960)
In this moment, I spent a large fraction of my time
George Hotz (1:44:05.920)
just watching videos because it's never fails to learn.
Lex Fridman (1:44:10.840)
Like it never, I've never failed
George Hotz (1:44:12.280)
from a video watching session
Lex Fridman (1:44:13.440)
to learn something I didn't know before.
George Hotz (1:44:15.360)
In fact, I usually like when I eat lunch,
Lex Fridman (1:44:18.440)
I'll sit, especially when the weather is good
Lex Fridman (1:44:20.640)
and just watch pedestrians with an eye to understand
Lex Fridman (1:44:24.560)
like from a computer vision eye,
George Hotz (1:44:26.400)
just to see can this model, can you predict,
Lex Fridman (1:44:29.280)
what are the decisions made?
Lex Fridman (1:44:30.480)
And there's so many things that we don't understand.
Lex Fridman (1:44:33.000)
This is what I mean about the state vector.
George Hotz (1:44:34.760)
Yeah, it's, I'm trying to always think like,
Lex Fridman (1:44:37.880)
cause I'm understanding in my human brain,
Lex Fridman (1:44:40.280)
how do we convert that into,
Lex Fridman (1:44:43.000)
how hard is the learning problem here?
George Hotz (1:44:44.960)
I guess is the fundamental question.
Lex Fridman (1:44:46.960)
So something that's from a hacking perspective,
George Hotz (1:44:51.760)
this is always comes up, especially with folks.
Lex Fridman (1:44:54.160)
Well, first the most popular question
Lex Fridman (1:44:55.520)
is the trolley problem, right?
Lex Fridman (1:44:58.400)
So that's not a sort of a serious problem.
George Hotz (1:45:01.920)
There are some ethical questions I think that arise.
Lex Fridman (1:45:06.080)
Maybe you wanna, do you think there's any ethical,
Lex Fridman (1:45:09.600)
serious ethical questions?
Lex Fridman (1:45:11.280)
We have a solution to the trolley problem at Comm.ai.
George Hotz (1:45:14.040)
Well, so there is actually an alert in our code,
Lex Fridman (1:45:16.520)
ethical dilemma detected.
George Hotz (1:45:18.000)
It's not triggered yet.
Lex Fridman (1:45:18.960)
We don't know how yet to detect the ethical dilemmas,
Lex Fridman (1:45:21.040)
but we're a level two system.
Lex Fridman (1:45:22.320)
So we're going to disengage
Lex Fridman (1:45:23.480)
and leave that decision to the human.
Lex Fridman (1:45:25.280)
You're such a troll.
George Hotz (1:45:26.640)
No, but the trolley problem deserves to be trolled.
Lex Fridman (1:45:28.720)
Yeah, that's a beautiful answer actually.
George Hotz (1:45:32.040)
I know, I gave it to someone who was like,
Lex Fridman (1:45:34.400)
sometimes people will ask,
George Hotz (1:45:35.360)
like you asked about the trolley problem,
Lex Fridman (1:45:36.560)
like you can have a kind of discussion about it.
George Hotz (1:45:38.040)
Like you get someone who's like really like earnest about it
Lex Fridman (1:45:40.760)
because it's the kind of thing where,
George Hotz (1:45:43.560)
if you ask a bunch of people in an office,
Lex Fridman (1:45:45.560)
whether we should use a SQL stack or a no SQL stack,
George Hotz (1:45:48.280)
if they're not that technical, they have no opinion.
Lex Fridman (1:45:50.560)
But if you ask them what color they want to paint the office,
George Hotz (1:45:52.360)
everyone has an opinion on that.
Lex Fridman (1:45:54.040)
And that's why the trolley problem is...
George Hotz (1:45:56.040)
I mean, that's a beautiful answer.
Lex Fridman (1:45:57.240)
Yeah, we're able to detect the problem
Lex Fridman (1:45:59.200)
and we're able to pass it on to the human.
Lex Fridman (1:46:01.960)
Wow, I've never heard anyone say it.
George Hotz (1:46:03.720)
This is your nice escape route.
Lex Fridman (1:46:06.120)
Okay, but...
George Hotz (1:46:07.320)
Proud level two.
Lex Fridman (1:46:08.680)
I'm proud level two.
George Hotz (1:46:09.760)
I love it.
Lex Fridman (1:46:10.600)
So the other thing that people have some concern about
George Hotz (1:46:14.400)
with AI in general is hacking.
Lex Fridman (1:46:17.800)
So how hard is it, do you think,
George Hotz (1:46:20.120)
to hack an autonomous vehicle,
Lex Fridman (1:46:21.400)
either through physical access
George Hotz (1:46:23.800)
or through the more sort of popular now,
Lex Fridman (1:46:25.680)
these adversarial examples on the sensors?
George Hotz (1:46:28.200)
Okay, the adversarial examples one.
Lex Fridman (1:46:30.720)
You want to see some adversarial examples
Lex Fridman (1:46:32.280)
that affect humans, right?
Lex Fridman (1:46:34.880)
Oh, well, there used to be a stop sign here,
Lex Fridman (1:46:38.000)
but I put a black bag over the stop sign
Lex Fridman (1:46:40.000)
and then people ran it, adversarial, right?
George Hotz (1:46:43.520)
Like there's tons of human adversarial examples too.
Lex Fridman (1:46:48.360)
The question in general about like security,
George Hotz (1:46:51.480)
if you saw something just came out today
Lex Fridman (1:46:53.360)
and like there are always such hypey headlines
George Hotz (1:46:55.080)
about like how navigate on autopilot
Lex Fridman (1:46:57.560)
was fooled by a GPS spoof to take an exit.
George Hotz (1:47:00.960)
Right.
Lex Fridman (1:47:01.800)
At least that's all they could do was take an exit.
George Hotz (1:47:03.920)
If your car is relying on GPS
Lex Fridman (1:47:06.360)
in order to have a safe driving policy,
George Hotz (1:47:09.000)
you're doing something wrong.
Lex Fridman (1:47:10.240)
If you're relying,
Lex Fridman (1:47:11.080)
and this is why V2V is such a terrible idea.
Lex Fridman (1:47:14.560)
V2V now relies on both parties getting communication right.
George Hotz (1:47:19.760)
This is not even, so I think of safety,
Lex Fridman (1:47:26.040)
security is like a special case of safety, right?
George Hotz (1:47:28.440)
Safety is like we put a little, you know,
Lex Fridman (1:47:31.840)
piece of caution tape around the hole
Lex Fridman (1:47:33.320)
so that people won't walk into it by accident.
Lex Fridman (1:47:35.520)
Security is like put a 10 foot fence around the hole
Lex Fridman (1:47:38.200)
so you actually physically cannot climb into it
Lex Fridman (1:47:40.100)
with barbed wire on the top and stuff, right?
Lex Fridman (1:47:42.320)
So like if you're designing systems that are like unreliable,
Lex Fridman (1:47:45.800)
they're definitely not secure.
George Hotz (1:47:48.440)
Your car should always do something safe
Lex Fridman (1:47:51.240)
using its local sensors.
Lex Fridman (1:47:53.400)
And then the local sensor should be hardwired.
Lex Fridman (1:47:55.240)
And then could somebody hack into your CAN bus
Lex Fridman (1:47:57.360)
and turn your steering wheel on your brakes?
Lex Fridman (1:47:58.600)
Yes, but they could do it before common AI too, so.
George Hotz (1:48:02.800)
Let's think out of the box on some things.
Lex Fridman (1:48:04.640)
So do you think teleoperation has a role in any of this?
Lex Fridman (1:48:09.360)
So remotely stepping in and controlling the cars?
Lex Fridman (1:48:13.880)
No, I think that if the safety operation by design
George Hotz (1:48:22.300)
requires a constant link to the cars,
Lex Fridman (1:48:26.160)
I think it doesn't work.
Lex Fridman (1:48:27.600)
So that's the same argument you're using for V2I, V2V?
Lex Fridman (1:48:31.120)
Well, there's a lot of non safety critical stuff
George Hotz (1:48:34.300)
you can do with V2I.
Lex Fridman (1:48:35.140)
I like V2I, I like V2I way more than V2V.
George Hotz (1:48:37.440)
Because V2I is already like,
Lex Fridman (1:48:39.280)
I already have internet in the car, right?
George Hotz (1:48:40.880)
There's a lot of great stuff you can do with V2I.
Lex Fridman (1:48:44.240)
Like for example, you can, well, I already have V2I,
Lex Fridman (1:48:47.280)
Waze is V2I, right?
Lex Fridman (1:48:48.880)
Waze can route me around traffic jams.
George Hotz (1:48:50.500)
That's a great example of V2I.
Lex Fridman (1:48:52.720)
And then, okay, the car automatically talks
George Hotz (1:48:54.420)
to that same service, like it works.
Lex Fridman (1:48:55.260)
So it's improving the experience,
Lex Fridman (1:48:56.800)
but it's not a fundamental fallback for safety.
Lex Fridman (1:48:59.440)
No, if any of your things that require wireless communication
George Hotz (1:49:04.440)
are more than QM, like have an ASL rating, it shouldn't be.
Lex Fridman (1:49:10.600)
You previously said that life is work
Lex Fridman (1:49:15.400)
and that you don't do anything to relax.
Lex Fridman (1:49:17.440)
So how do you think about hard work?
Lex Fridman (1:49:20.960)
What do you think it takes to accomplish great things?
Lex Fridman (1:49:24.680)
And there's a lot of people saying
George Hotz (1:49:25.820)
that there needs to be some balance.
Lex Fridman (1:49:28.160)
You need to, in order to accomplish great things,
George Hotz (1:49:31.120)
you need to take some time off,
Lex Fridman (1:49:32.200)
you need to reflect and so on.
George Hotz (1:49:33.920)
Now, and then some people are just insanely working,
Lex Fridman (1:49:37.920)
burning the candle on both ends.
Lex Fridman (1:49:39.680)
How do you think about that?
Lex Fridman (1:49:41.400)
I think I was trolling in the Siraj interview
George Hotz (1:49:43.440)
when I said that.
Lex Fridman (1:49:44.880)
Off camera, right before I smoked a little bit of weed,
Lex Fridman (1:49:47.280)
like, you know, come on, this is a joke, right?
Lex Fridman (1:49:49.840)
Like I do nothing to relax.
Lex Fridman (1:49:50.880)
Look where I am, I'm at a party, right?
Lex Fridman (1:49:52.600)
Yeah, yeah, yeah, that's true.
Lex Fridman (1:49:55.240)
So no, no, of course I don't.
Lex Fridman (1:49:58.080)
When I say that life is work though,
George Hotz (1:49:59.840)
I mean that like, I think that what gives my life meaning is work.
Lex Fridman (1:50:04.200)
I don't mean that every minute of the day
George Hotz (1:50:05.720)
you should be working.
Lex Fridman (1:50:06.560)
I actually think this is not the best way to maximize results.
George Hotz (1:50:09.800)
I think that if you're working 12 hours a day,
Lex Fridman (1:50:12.040)
you should be working smarter and not harder.
George Hotz (1:50:14.900)
Well, so work gives you meaning.
Lex Fridman (1:50:17.880)
For some people, other sorts of meaning
George Hotz (1:50:20.520)
is personal relationships, like family and so on.
Lex Fridman (1:50:24.560)
You've also, in that interview with Siraj,
George Hotz (1:50:27.200)
or the trolling, mentioned that one of the things
Lex Fridman (1:50:30.680)
you look forward to in the future is AI girlfriends.
Lex Fridman (1:50:34.280)
So that's a topic that I'm very much fascinated by,
Lex Fridman (1:50:38.720)
not necessarily girlfriends,
Lex Fridman (1:50:39.760)
but just forming a deep connection with AI.
Lex Fridman (1:50:42.880)
What kind of system do you imagine
George Hotz (1:50:44.320)
when you say AI girlfriend,
Lex Fridman (1:50:46.160)
whether you were trolling or not?
George Hotz (1:50:47.720)
No, that one I'm very serious about.
Lex Fridman (1:50:49.640)
And I'm serious about that on both a shallow level
Lex Fridman (1:50:52.280)
and a deep level.
Lex Fridman (1:50:53.600)
I think that VR brothels are coming soon
Lex Fridman (1:50:55.600)
and are going to be really cool.
Lex Fridman (1:50:57.760)
It's not cheating if it's a robot.
George Hotz (1:50:59.680)
I see the slogan already.
Lex Fridman (1:51:03.120)
But there's, I don't know if you've watched,
George Hotz (1:51:06.160)
or just watched the Black Mirror episode.
Lex Fridman (1:51:08.320)
I watched the latest one, yeah.
George Hotz (1:51:09.320)
Yeah, yeah.
Lex Fridman (1:51:11.320)
Oh, the Ashley 2 one?
George Hotz (1:51:15.080)
No, where there's two friends
Lex Fridman (1:51:16.920)
who are having sex with each other in...
George Hotz (1:51:20.160)
Oh, in the VR game.
Lex Fridman (1:51:21.000)
In the VR game.
George Hotz (1:51:22.720)
It's just two guys,
Lex Fridman (1:51:23.560)
but one of them was a female, yeah.
George Hotz (1:51:27.200)
Which is another mind blowing concept.
Lex Fridman (1:51:29.520)
That in VR, you don't have to be the form.
George Hotz (1:51:33.240)
You can be two animals having sex.
Lex Fridman (1:51:37.160)
It's weird.
George Hotz (1:51:38.000)
I mean, I'll see how nice that the software
Lex Fridman (1:51:38.920)
maps the nerve endings, right?
George Hotz (1:51:40.240)
Yeah, it's huge.
Lex Fridman (1:51:41.600)
I mean, yeah, they sweep a lot of the fascinating,
George Hotz (1:51:44.440)
really difficult technical challenges under the rug,
Lex Fridman (1:51:46.400)
like assuming it's possible
George Hotz (1:51:48.320)
to do the mapping of the nerve endings, then...
Lex Fridman (1:51:51.120)
I wish, yeah, I saw that,
George Hotz (1:51:51.960)
the way they did it with the little like stim unit
Lex Fridman (1:51:53.800)
on the head, that'd be amazing.
George Hotz (1:51:56.800)
So, well, no, no, on a shallow level,
Lex Fridman (1:51:58.760)
like you could set up like almost a brothel
George Hotz (1:52:01.680)
with like real dolls and Oculus Quests,
Lex Fridman (1:52:05.160)
write some good software.
George Hotz (1:52:06.200)
I think it'd be a cool novelty experience.
Lex Fridman (1:52:09.280)
But no, on a deeper, like emotional level,
George Hotz (1:52:12.840)
I mean, yeah, I would really like to fall in love
Lex Fridman (1:52:17.000)
with a machine.
Lex Fridman (1:52:18.120)
Do you see yourself having a long term relationship
Lex Fridman (1:52:25.000)
of the kind monogamous relationship that we have now
George Hotz (1:52:28.800)
with a robot, with a AI system even,
Lex Fridman (1:52:31.360)
not even just a robot?
Lex Fridman (1:52:32.680)
So I think about maybe my ideal future.
Lex Fridman (1:52:38.120)
When I was 15, I read Eliezer Yudkowsky's early writings
George Hotz (1:52:43.120)
on the singularity and like that AI
Lex Fridman (1:52:49.120)
is going to surpass human intelligence massively.
George Hotz (1:52:53.600)
He made some Moore's law based predictions
Lex Fridman (1:52:55.440)
that I mostly agree with.
Lex Fridman (1:52:57.360)
And then I really struggled
Lex Fridman (1:52:59.320)
for the next couple of years of my life.
Lex Fridman (1:53:01.320)
Like, why should I even bother to learn anything?
Lex Fridman (1:53:03.320)
It's all gonna be meaningless when the machines show up.
George Hotz (1:53:06.080)
Right.
Lex Fridman (1:53:07.960)
Maybe when I was that young,
George Hotz (1:53:10.480)
I was still a little bit more pure
Lex Fridman (1:53:11.960)
and really like clung to that.
Lex Fridman (1:53:13.120)
And then I'm like, well,
Lex Fridman (1:53:13.960)
the machines ain't here yet, you know,
Lex Fridman (1:53:14.960)
and I seem to be pretty good at this stuff.
Lex Fridman (1:53:16.720)
Let's try my best, you know,
George Hotz (1:53:18.440)
like what's the worst that happens.
Lex Fridman (1:53:21.320)
But the best possible future I see
George Hotz (1:53:24.000)
is me sort of merging with the machine.
Lex Fridman (1:53:26.760)
And the way that I personify this
George Hotz (1:53:28.760)
is in a long term monogamous relationship with a machine.
Lex Fridman (1:53:32.880)
Oh, you don't think there's a room
George Hotz (1:53:34.040)
for another human in your life,
Lex Fridman (1:53:35.680)
if you really truly merge with another machine?
George Hotz (1:53:39.160)
I mean, I see merging.
Lex Fridman (1:53:40.840)
I see like the best interface to my brain
George Hotz (1:53:46.280)
is like the same relationship interface
Lex Fridman (1:53:48.680)
to merge with an AI, right?
Lex Fridman (1:53:49.920)
What does that merging feel like?
Lex Fridman (1:53:52.840)
I've seen couples who've been together for a long time.
Lex Fridman (1:53:55.920)
And like, I almost think of them as one person,
Lex Fridman (1:53:58.400)
like couples who spend all their time together and...
George Hotz (1:54:01.840)
That's fascinating.
Lex Fridman (1:54:02.680)
You're actually putting,
Lex Fridman (1:54:03.840)
what does that merging actually looks like?
Lex Fridman (1:54:06.040)
It's not just a nice channel.
George Hotz (1:54:08.120)
Like a lot of people imagine it's just an efficient link,
Lex Fridman (1:54:12.160)
search link to Wikipedia or something.
George Hotz (1:54:14.280)
I don't believe in that.
Lex Fridman (1:54:15.200)
But it's more,
George Hotz (1:54:16.040)
you're saying that there's the same kind of relationship
Lex Fridman (1:54:18.520)
you have with another human,
George Hotz (1:54:19.400)
that's a deep relationship.
Lex Fridman (1:54:20.760)
That's what merging looks like.
George Hotz (1:54:22.880)
That's pretty...
Lex Fridman (1:54:24.400)
I don't believe that link is possible.
George Hotz (1:54:26.600)
I think that that link,
Lex Fridman (1:54:27.680)
so you're like, oh, I'm gonna download Wikipedia
George Hotz (1:54:29.160)
right to my brain.
Lex Fridman (1:54:30.080)
My reading speed is not limited by my eyes.
George Hotz (1:54:33.280)
My reading speed is limited by my inner processing loop.
Lex Fridman (1:54:36.720)
And to like bootstrap that sounds kind of unclear
Lex Fridman (1:54:40.680)
how to do it and horrifying.
Lex Fridman (1:54:42.360)
But if I am with somebody and I'll use a somebody
George Hotz (1:54:46.480)
who is making a super sophisticated model of me
Lex Fridman (1:54:51.280)
and then running simulations on that model,
George Hotz (1:54:53.120)
I'm not gonna get into the question
Lex Fridman (1:54:54.040)
whether the simulations are conscious or not.
George Hotz (1:54:55.840)
I don't really wanna know what it's doing.
Lex Fridman (1:54:58.200)
But using those simulations
George Hotz (1:55:00.080)
to play out hypothetical futures for me,
Lex Fridman (1:55:01.840)
deciding what things to say to me,
George Hotz (1:55:04.840)
to guide me along a path.
Lex Fridman (1:55:06.240)
And that's how I envision it.
Lex Fridman (1:55:08.680)
So on that path to AI of superhuman level intelligence,
Lex Fridman (1:55:15.080)
you've mentioned that you believe in the singularity,
George Hotz (1:55:16.840)
that singularity is coming.
Lex Fridman (1:55:18.600)
Again, could be trolling, could be not,
George Hotz (1:55:20.200)
could be part, all trolling has truth in it.
Lex Fridman (1:55:23.000)
I don't know what that means anymore.
Lex Fridman (1:55:24.120)
What is the singularity?
Lex Fridman (1:55:25.920)
Yeah, so that's really the question.
Lex Fridman (1:55:28.040)
How many years do you think before the singularity,
Lex Fridman (1:55:30.520)
what form do you think it will take?
Lex Fridman (1:55:32.080)
Does that mean fundamental shifts in capabilities of AI?
Lex Fridman (1:55:35.440)
Or does it mean some other kind of ideas?
George Hotz (1:55:39.400)
Maybe that's just my roots, but.
Lex Fridman (1:55:41.360)
So I can buy a human beings worth of compute
George Hotz (1:55:43.880)
for like a million bucks today.
Lex Fridman (1:55:46.000)
It's about one TPU pod V3.
George Hotz (1:55:47.680)
I want like, I think they claim a hundred pay to flops.
Lex Fridman (1:55:50.120)
That's being generous.
George Hotz (1:55:50.960)
I think humans are actually more like 20.
Lex Fridman (1:55:52.240)
So that's like five humans.
George Hotz (1:55:53.080)
That's pretty good.
Lex Fridman (1:55:53.960)
Google needs to sell their TPUs.
Lex Fridman (1:55:56.720)
But I could buy, I could buy, I could buy GPUs.
Lex Fridman (1:55:58.560)
I could buy a stack of like, I'd buy 1080 TIs,
George Hotz (1:56:02.200)
build data center full of them.
Lex Fridman (1:56:03.760)
And for a million bucks, I can get a human worth of compute.
Lex Fridman (1:56:08.040)
But when you look at the total number of flops in the world,
Lex Fridman (1:56:12.160)
when you look at human flops,
George Hotz (1:56:14.400)
which goes up very, very slowly with the population
Lex Fridman (1:56:17.040)
and machine flops, which goes up exponentially,
Lex Fridman (1:56:19.760)
but it's still nowhere near.
Lex Fridman (1:56:22.360)
I think that's the key thing to talk about
George Hotz (1:56:24.560)
when the singularity happened.
Lex Fridman (1:56:25.880)
When most flops in the world are Silicon and not biological,
George Hotz (1:56:29.760)
that's kind of the crossing point.
Lex Fridman (1:56:32.280)
Like they're now the dominant species on the planet.
Lex Fridman (1:56:35.480)
And just looking at how technology is progressing,
Lex Fridman (1:56:38.720)
when do you think that could possibly happen?
Lex Fridman (1:56:40.360)
You think it would happen in your lifetime?
Lex Fridman (1:56:41.680)
Oh yeah, definitely in my lifetime.
George Hotz (1:56:43.600)
I've done the math.
Lex Fridman (1:56:44.440)
I like 2038 because it's the Unix timestamp rollover.
George Hotz (1:56:49.880)
Yeah, beautifully put.
Lex Fridman (1:56:52.640)
So you've said that the meaning of life is to win.
George Hotz (1:56:57.960)
If you look five years into the future,
Lex Fridman (1:56:59.520)
what does winning look like?
George Hotz (1:57:02.640)
So,
Lex Fridman (1:57:08.560)
there's a lot of,
George Hotz (1:57:10.120)
I can go into like technical depth
Lex Fridman (1:57:12.680)
to what I mean by that, to win.
George Hotz (1:57:15.760)
It may not mean, I was criticized for that in the comments.
Lex Fridman (1:57:18.280)
Like, doesn't this guy wanna like save the penguins
George Hotz (1:57:20.520)
in Antarctica or like,
Lex Fridman (1:57:22.520)
oh man, listen to what I'm saying.
George Hotz (1:57:24.920)
I'm not talking about like I have a yacht or something.
Lex Fridman (1:57:27.560)
But I am an agent.
George Hotz (1:57:30.520)
I am put into this world.
Lex Fridman (1:57:32.920)
And I don't really know what my purpose is.
Lex Fridman (1:57:37.480)
But if you're an intelligent agent
Lex Fridman (1:57:40.280)
and you're put into a world,
Lex Fridman (1:57:41.400)
what is the ideal thing to do?
Lex Fridman (1:57:43.160)
Well, the ideal thing mathematically,
George Hotz (1:57:44.800)
you can go back to like Schmidt Hoover theories about this,
Lex Fridman (1:57:47.080)
is to build a compressive model of the world.
George Hotz (1:57:50.480)
To build a maximally compressive,
Lex Fridman (1:57:51.840)
to explore the world such that your exploration function
George Hotz (1:57:55.600)
maximizes the derivative of compression of the past.
Lex Fridman (1:57:58.880)
Schmidt Hoover has a paper about this.
Lex Fridman (1:58:00.720)
And like, I took that kind of
Lex Fridman (1:58:02.040)
as like a personal goal function.
Lex Fridman (1:58:05.360)
So what I mean to win, I mean like,
Lex Fridman (1:58:07.720)
maybe this is religious,
Lex Fridman (1:58:09.080)
but like I think that in the future,
Lex Fridman (1:58:11.320)
I might be given a real purpose
George Hotz (1:58:13.040)
or I may decide this purpose myself.
Lex Fridman (1:58:14.680)
And then at that point,
George Hotz (1:58:16.160)
now I know what the game is and I know how to win.
Lex Fridman (1:58:18.240)
I think right now,
George Hotz (1:58:19.080)
I'm still just trying to figure out what the game is.
Lex Fridman (1:58:20.720)
But once I know,
Lex Fridman (1:58:21.800)
so you have imperfect information,
Lex Fridman (1:58:26.440)
you have a lot of uncertainty about the reward function
Lex Fridman (1:58:28.600)
and you're discovering it.
Lex Fridman (1:58:29.720)
Exactly.
Lex Fridman (1:58:30.560)
But the purpose is...
Lex Fridman (1:58:31.400)
That's a better way to put it.
George Hotz (1:58:33.120)
The purpose is to maximize it
Lex Fridman (1:58:34.440)
while you have a lot of uncertainty around it.
Lex Fridman (1:58:37.960)
And you're both reducing the uncertainty
Lex Fridman (1:58:39.400)
and maximizing at the same time.
George Hotz (1:58:41.160)
Yeah.
Lex Fridman (1:58:42.000)
And so that's at the technical level.
Lex Fridman (1:58:44.240)
What is the, if you believe in the universal prior,
Lex Fridman (1:58:47.440)
what is the universal reward function?
George Hotz (1:58:49.360)
That's the better way to put it.
Lex Fridman (1:58:51.320)
So that win is interesting.
George Hotz (1:58:53.680)
I think I speak for everyone in saying that
Lex Fridman (1:58:57.280)
I wonder what that reward function is for you.
Lex Fridman (1:59:01.920)
And I look forward to seeing that in five years,
Lex Fridman (1:59:05.920)
in 10 years.
George Hotz (1:59:07.040)
I think a lot of people, including myself,
Lex Fridman (1:59:08.680)
are cheering you on, man.
Lex Fridman (1:59:09.840)
So I'm happy you exist and I wish you the best of luck.
Lex Fridman (1:59:14.280)
Thanks for talking to me, man.
George Hotz (1:59:15.360)
Thank you.
Lex Fridman (1:59:16.200)
Have a good one.
George Hotz (20:00.840)
we're going to drop the zero day.
Lex Fridman (20:03.240)
Oh, wow.
George Hotz (20:04.120)
We're going to drop the weapon.
Lex Fridman (20:04.960)
That's so cool.
George Hotz (20:05.920)
That is so cool.
Lex Fridman (20:07.520)
I love the deadlines.
George Hotz (20:09.240)
Oh, that's so cool.
Lex Fridman (20:10.080)
Give them real deadlines.
George Hotz (20:10.920)
Yeah.
Lex Fridman (20:12.360)
And I think it's done a lot for moving the industry forward.
George Hotz (20:15.800)
I watched your coding sessions on the streamed online.
Lex Fridman (20:20.360)
You code things up, the basic projects,
George Hotz (20:22.720)
usually from scratch.
Lex Fridman (20:24.000)
I would say sort of as a programmer myself,
George Hotz (20:28.200)
just watching you that you type really fast
Lex Fridman (20:30.360)
and your brain works in both brilliant and chaotic ways.
George Hotz (20:34.520)
I don't know if that's always true,
Lex Fridman (20:35.800)
but certainly for the live streams.
Lex Fridman (20:37.600)
So it's interesting to me because I'm more,
Lex Fridman (20:40.360)
I'm much slower and systematic and careful.
Lex Fridman (20:43.520)
And you just move, I mean,
Lex Fridman (20:44.920)
probably in order of magnitude faster.
Lex Fridman (20:48.040)
So I'm curious, is there a method to your madness?
Lex Fridman (20:51.080)
Is it just who you are?
George Hotz (20:53.040)
There's pros and cons.
Lex Fridman (20:54.720)
There's pros and cons to my programming style.
Lex Fridman (20:58.080)
And I'm aware of them.
Lex Fridman (20:59.440)
Like if you ask me to like get something up
Lex Fridman (21:03.560)
and working quickly with like an API
Lex Fridman (21:05.400)
that's kind of undocumented,
George Hotz (21:06.760)
I will do this super fast
Lex Fridman (21:08.160)
because I will throw things at it until it works.
George Hotz (21:10.200)
If you ask me to take a vector and rotate it 90 degrees
Lex Fridman (21:14.720)
and then flip it over the XY plane,
George Hotz (21:19.320)
I'll spam program for two hours and won't get it.
Lex Fridman (21:22.320)
Oh, because it's something that you could do
George Hotz (21:23.920)
with a sheet of paper, think through design,
Lex Fridman (21:26.280)
and then just, do you really just throw stuff at the wall
Lex Fridman (21:30.440)
and you get so good at it that it usually works?
Lex Fridman (21:34.640)
I should become better at the other kind as well.
George Hotz (21:36.960)
Sometimes I'll do things methodically.
Lex Fridman (21:39.440)
It's nowhere near as entertaining on the Twitch streams.
George Hotz (21:41.160)
I do exaggerate it a bit on the Twitch streams as well.
Lex Fridman (21:43.520)
The Twitch streams, I mean,
Lex Fridman (21:44.680)
what do you want to see a game or you want to see
Lex Fridman (21:45.840)
actions per minute, right?
George Hotz (21:46.840)
I'll show you APM for programming too.
Lex Fridman (21:48.160)
Yeah, I recommend people go to it.
George Hotz (21:50.280)
I think I watched, I watched probably several hours
Lex Fridman (21:53.400)
of you, like I've actually left you programming
George Hotz (21:56.240)
in the background while I was programming
Lex Fridman (21:59.080)
because you made me, it was like watching
George Hotz (22:02.040)
a really good gamer.
Lex Fridman (22:03.160)
It's like energizes you because you're like moving so fast.
George Hotz (22:06.280)
It's so, it's awesome.
Lex Fridman (22:07.600)
It's inspiring and it made me jealous that like,
George Hotz (22:12.320)
because my own programming is inadequate
Lex Fridman (22:14.320)
in terms of speed.
George Hotz (22:15.520)
Oh, I was like.
Lex Fridman (22:17.000)
So I'm twice as frantic on the live streams
George Hotz (22:20.520)
as I am when I code without them.
Lex Fridman (22:22.680)
It's super entertaining.
Lex Fridman (22:23.760)
So I wasn't even paying attention to what you were coding,
Lex Fridman (22:26.440)
which is great.
George Hotz (22:27.280)
It's just watching you switch windows and Vim I guess
Lex Fridman (22:30.840)
is the most.
George Hotz (22:31.680)
Yeah, there's Vim on screen.
Lex Fridman (22:33.080)
I've developed the workload at Facebook
Lex Fridman (22:34.440)
and stuck with it.
Lex Fridman (22:35.640)
How do you learn new programming tools,
Lex Fridman (22:37.360)
ideas, techniques these days?
Lex Fridman (22:39.480)
What's your like a methodology for learning new things?
Lex Fridman (22:42.120)
So I wrote for comma, the distributed file systems
Lex Fridman (22:48.800)
out in the world are extremely complex.
George Hotz (22:50.720)
Like if you want to install something like like like Ceph,
Lex Fridman (22:55.280)
Ceph is I think the like open infrastructure
George Hotz (22:58.760)
distributed file system,
Lex Fridman (23:00.320)
or there's like newer ones like seaweed FS,
Lex Fridman (23:04.000)
but these are all like 10,000 plus line projects.
Lex Fridman (23:06.880)
I think some of them are even a hundred thousand line
Lex Fridman (23:09.480)
and just configuring them as a nightmare.
Lex Fridman (23:11.120)
So I wrote, I wrote one, it's 200 lines
Lex Fridman (23:16.440)
and it's, it uses like NGINX and volume servers
Lex Fridman (23:18.880)
and has this little master server that I wrote in Go.
Lex Fridman (23:21.640)
And the way I go, this,
Lex Fridman (23:24.200)
if I would say that I'm proud per line of any code I wrote,
George Hotz (23:27.240)
maybe there's some exploits that I think are beautiful.
Lex Fridman (23:29.160)
And then this, this is 200 lines.
Lex Fridman (23:31.320)
And just the way that I thought about it,
Lex Fridman (23:33.720)
I think was very good.
Lex Fridman (23:34.600)
And the reason it's very good is because
Lex Fridman (23:35.880)
that was the fourth version of it that I wrote.
Lex Fridman (23:37.600)
And I had three versions that I threw away.
Lex Fridman (23:39.320)
You mentioned, did you say Go?
George Hotz (23:40.960)
I wrote in Go, yeah.
Lex Fridman (23:41.800)
In Go.
Lex Fridman (23:42.640)
Is that a functional language?
Lex Fridman (23:43.840)
I forget what Go is.
George Hotz (23:45.240)
Go is Google's language.
Lex Fridman (23:47.120)
Right.
George Hotz (23:48.200)
It's not functional.
Lex Fridman (23:49.440)
It's some, it's like in a way it's C++, but easier.
George Hotz (23:56.160)
It's, it's strongly typed.
Lex Fridman (23:58.160)
It has a nice ecosystem around it.
George Hotz (23:59.720)
When I first looked at it, I was like, this is like Python,
Lex Fridman (24:02.680)
but it takes twice as long to do anything.
George Hotz (24:04.560)
Yeah.
Lex Fridman (24:05.560)
Now that I've, OpenPilot is migrating to C,
Lex Fridman (24:09.560)
but it still has large Python components.
Lex Fridman (24:10.960)
I now understand why Python doesn't work
George Hotz (24:12.720)
for large code bases and why you want something like Go.
Lex Fridman (24:15.800)
Interesting.
Lex Fridman (24:16.640)
So why, why doesn't Python work for,
Lex Fridman (24:18.640)
so even most, speaking for myself at least,
George Hotz (24:21.680)
like we do a lot of stuff,
Lex Fridman (24:23.360)
basically demo level work with autonomous vehicles
Lex Fridman (24:26.480)
and most of the work is Python.
Lex Fridman (24:28.240)
Yeah.
Lex Fridman (24:29.200)
Why doesn't Python work for large code bases?
Lex Fridman (24:32.400)
Because, well, lack of type checking is a big part.
Lex Fridman (24:37.960)
So errors creep in.
Lex Fridman (24:39.360)
Yeah.
Lex Fridman (24:40.200)
And like, you don't know,
Lex Fridman (24:41.920)
the compiler can tell you like nothing, right?
Lex Fridman (24:45.320)
So everything is either, you know,
Lex Fridman (24:48.440)
like, like syntax errors, fine.
Lex Fridman (24:49.880)
But if you misspell a variable in Python,
Lex Fridman (24:51.800)
the compiler won't catch that.
George Hotz (24:53.000)
There's like linters that can catch it some of the time.
Lex Fridman (24:56.600)
There's no types.
George Hotz (24:57.560)
This is really the biggest downside.
Lex Fridman (25:00.520)
And then, well, Python's slow, but that's not related to it.
George Hotz (25:02.640)
Well, maybe it's kind of related to it, so it's lack of.
Lex Fridman (25:04.800)
So what's, what's in your toolbox these days?
Lex Fridman (25:06.600)
Is it Python?
Lex Fridman (25:07.440)
What else?
George Hotz (25:08.280)
I need to move to something else.
Lex Fridman (25:10.120)
My adventure into dependently typed languages,
George Hotz (25:12.840)
I love these languages.
Lex Fridman (25:14.200)
They just have like syntax from the 80s.
Lex Fridman (25:18.480)
What do you think about JavaScript?
Lex Fridman (25:21.080)
ES6, like the modern, or TypeScript?
George Hotz (25:23.960)
JavaScript is,
Lex Fridman (25:26.080)
the whole ecosystem is unbelievably confusing.
George Hotz (25:28.960)
Right.
Lex Fridman (25:29.800)
NPM updates a package from 0.2.2 to 0.2.5,
Lex Fridman (25:32.800)
and that breaks your Babel linter,
Lex Fridman (25:34.520)
which translates your ES5 into ES6,
George Hotz (25:37.040)
which doesn't run on, so.
Lex Fridman (25:39.920)
Why do I have to compile my JavaScript again, huh?
George Hotz (25:42.440)
It may be the future, though.
Lex Fridman (25:44.000)
You think about, I mean,
George Hotz (25:45.760)
I've embraced JavaScript recently,
Lex Fridman (25:47.360)
just because, just like I've continually embraced PHP,
George Hotz (25:52.280)
it seems that these worst possible languages
Lex Fridman (25:54.840)
live on for the longest, like cockroaches never die.
George Hotz (25:57.440)
Yeah.
Lex Fridman (25:58.480)
Well, it's in the browser, and it's fast.
George Hotz (26:00.720)
It's fast.
Lex Fridman (26:01.680)
Yeah.
George Hotz (26:02.520)
It's in the browser, and compute might stay,
Lex Fridman (26:04.880)
become, you know, the browser.
George Hotz (26:06.440)
It's unclear what the role of the browser is
Lex Fridman (26:09.000)
in terms of distributed computation in the future, so.
George Hotz (26:13.600)
JavaScript is definitely here to stay.
Lex Fridman (26:15.240)
Yeah.
George Hotz (26:16.080)
It's interesting if autonomous vehicles
Lex Fridman (26:18.160)
will run on JavaScript one day.
George Hotz (26:19.480)
I mean, you have to consider these possibilities.
Lex Fridman (26:21.800)
Well, all our debug tools are JavaScript.
George Hotz (26:24.280)
We actually just open sourced them.
Lex Fridman (26:26.040)
We have a tool, Explorer,
George Hotz (26:27.400)
which you can annotate your disengagements,
Lex Fridman (26:29.200)
and we have a tool, Cabana,
George Hotz (26:30.080)
which lets you analyze the can traffic from the car.
Lex Fridman (26:32.920)
So basically, anytime you're visualizing something
George Hotz (26:35.240)
about the log, you're using JavaScript.
Lex Fridman (26:37.720)
Well, the web is the best UI toolkit by far, so.
Lex Fridman (26:41.280)
And then, you know what?
Lex Fridman (26:42.120)
You're coding in JavaScript.
George Hotz (26:42.960)
We have a React guy.
Lex Fridman (26:43.800)
He's good.
George Hotz (26:44.640)
React, nice.
Lex Fridman (26:46.080)
Let's get into it.
Lex Fridman (26:46.920)
So let's talk autonomous vehicles.
Lex Fridman (26:48.800)
Yeah.
George Hotz (26:49.640)
You founded Comma AI.
Lex Fridman (26:51.760)
Let's, at a high level,
Lex Fridman (26:54.920)
how did you get into the world of vehicle automation?
Lex Fridman (26:57.840)
Can you also just, for people who don't know,
Lex Fridman (26:59.880)
tell the story of Comma AI?
Lex Fridman (27:01.360)
Sure.
Lex Fridman (27:02.880)
So I was working at this AI startup,
Lex Fridman (27:06.080)
and a friend approached me,
Lex Fridman (27:08.160)
and he's like, dude, I don't know where this is going,
Lex Fridman (27:12.040)
but the coolest applied AI problem today
George Hotz (27:15.160)
is self driving cars.
Lex Fridman (27:16.480)
I'm like, well, absolutely.
George Hotz (27:18.800)
You want to meet with Elon Musk,
Lex Fridman (27:20.520)
and he's looking for somebody to build a vision system
George Hotz (27:24.560)
for autopilot.
Lex Fridman (27:27.560)
This is when they were still on AP1.
George Hotz (27:29.320)
They were still using Mobileye.
Lex Fridman (27:30.840)
Elon, back then, was looking for a replacement,
Lex Fridman (27:33.680)
and he brought me in,
Lex Fridman (27:36.160)
and we talked about a contract
George Hotz (27:37.320)
where I would deliver something
Lex Fridman (27:39.040)
that meets Mobileye level performance.
George Hotz (27:41.360)
I would get paid $12 million if I could deliver it tomorrow,
Lex Fridman (27:43.920)
and I would lose $1 million
George Hotz (27:45.280)
for every month I didn't deliver.
Lex Fridman (27:46.720)
Yeah.
Lex Fridman (27:47.720)
So I was like, okay, this is a great deal.
Lex Fridman (27:49.080)
This is a super exciting challenge.
Lex Fridman (27:52.360)
You know what?
Lex Fridman (27:53.200)
Even if it takes me 10 months,
George Hotz (27:54.400)
I get $2 million.
Lex Fridman (27:55.360)
It's good.
George Hotz (27:56.200)
Maybe I can finish up in five.
Lex Fridman (27:57.120)
Maybe I don't finish it at all,
Lex Fridman (27:58.120)
and I get paid nothing,
Lex Fridman (27:58.960)
and I can still work for 12 months for free.
Lex Fridman (28:00.840)
So maybe just take a pause on that.
Lex Fridman (28:02.920)
I'm also curious about this
George Hotz (28:04.240)
because I've been working in robotics for a long time,
Lex Fridman (28:06.200)
and I'm curious to see a person like you
George Hotz (28:07.640)
just step in and sort of somewhat naive,
Lex Fridman (28:11.040)
but brilliant, right?
Lex Fridman (28:11.960)
So that's the best place to be
Lex Fridman (28:13.960)
because you basically full steam take on a problem.
Lex Fridman (28:17.200)
How confident, how, from that time,
Lex Fridman (28:19.680)
because you know a lot more now,
George Hotz (28:21.280)
at that time, how hard do you think it is
Lex Fridman (28:23.440)
to solve all of autonomous driving?
George Hotz (28:25.840)
I remember I suggested to Elon in the meeting
Lex Fridman (28:30.560)
putting a GPU behind each camera
George Hotz (28:33.080)
to keep the compute local.
Lex Fridman (28:35.120)
This is an incredibly stupid idea.
George Hotz (28:38.000)
I leave the meeting 10 minutes later,
Lex Fridman (28:39.400)
and I'm like, I could have spent a little bit of time
George Hotz (28:41.520)
thinking about this problem before I went in.
Lex Fridman (28:42.760)
Why is it a stupid idea?
George Hotz (28:44.160)
Oh, just send all your cameras to one big GPU.
Lex Fridman (28:46.240)
You're much better off doing that.
George Hotz (28:48.200)
Oh, sorry.
Lex Fridman (28:49.040)
You said behind every camera have a GPU.
George Hotz (28:50.200)
Every camera have a small GPU.
Lex Fridman (28:51.360)
I was like, oh, I'll put the first few layers
George Hotz (28:52.720)
of my comms there.
Lex Fridman (28:54.040)
Ugh, why'd I say that?
George Hotz (28:56.040)
That's possible.
Lex Fridman (28:56.880)
It's possible, but it's a bad idea.
George Hotz (28:58.960)
It's not obviously a bad idea.
Lex Fridman (29:00.440)
Pretty obviously bad,
Lex Fridman (29:01.280)
but whether it's actually a bad idea or not,
Lex Fridman (29:02.920)
I left that meeting with Elon, beating myself up.
Lex Fridman (29:05.240)
I'm like, why'd I say something stupid?
Lex Fridman (29:07.000)
Yeah, you haven't at least thought through
George Hotz (29:10.720)
every aspect of it, yeah.
Lex Fridman (29:12.200)
He's very sharp too.
George Hotz (29:13.360)
Usually in life, I get away with saying stupid things
Lex Fridman (29:15.760)
and then kind of course,
George Hotz (29:16.920)
oh, right away he called me out about it.
Lex Fridman (29:18.520)
And usually in life, I get away with saying stupid things
Lex Fridman (29:21.080)
and then a lot of times people don't even notice
Lex Fridman (29:26.080)
and I'll correct it and bring the conversation back.
Lex Fridman (29:28.200)
But with Elon, it was like, nope, okay, well.
Lex Fridman (29:31.840)
That's not at all why the contract fell through.
George Hotz (29:33.520)
I was much more prepared the second time I met him.
Lex Fridman (29:35.520)
Yeah, but in general, how hard did you think it is?
George Hotz (29:39.640)
Like 12 months is a tough timeline.
Lex Fridman (29:43.680)
Oh, I just thought I'd clone Mobileye IQ3.
George Hotz (29:45.720)
I didn't think I'd solve level five self driving
Lex Fridman (29:47.560)
or anything.
Lex Fridman (29:48.400)
So the goal there was to do lane keeping, good lane keeping.
Lex Fridman (29:52.760)
I saw, my friend showed me the outputs from a Mobileye
Lex Fridman (29:55.480)
and the outputs from a Mobileye was just basically
Lex Fridman (29:57.080)
two lanes at a position of a lead car.
George Hotz (29:59.360)
I'm like, I can gather a data set and train this net
Lex Fridman (30:02.160)
in weeks and I did.
George Hotz (30:04.760)
Well, first time I tried the implementation of Mobileye
Lex Fridman (30:07.520)
in a Tesla, I was really surprised how good it is.
George Hotz (30:11.200)
It's going incredibly good.
Lex Fridman (30:12.240)
Cause I thought it's just cause I've done a lot
George Hotz (30:14.280)
of computer vision, I thought it'd be a lot harder
Lex Fridman (30:17.960)
to create a system that that's stable.
Lex Fridman (30:20.960)
So I was personally surprised, you know,
Lex Fridman (30:24.200)
have to admit it.
George Hotz (30:25.040)
Cause I was kind of skeptical before trying it.
Lex Fridman (30:27.800)
Cause I thought it would go in and out a lot more.
George Hotz (30:31.160)
It would get disengaged a lot more and it's pretty robust.
Lex Fridman (30:36.160)
So what, how hard is the problem when you tackled it?
Lex Fridman (30:42.080)
So I think AP1 was great.
Lex Fridman (30:44.480)
Like Elon talked about disengagements on the 405 down in LA
George Hotz (30:49.000)
with like the lane marks are kind of faded
Lex Fridman (30:51.000)
and the Mobileye system would drop out.
George Hotz (30:53.920)
Like I had something up and working that I would say
Lex Fridman (30:57.760)
was like the same quality in three months.
Lex Fridman (31:02.480)
Same quality, but how do you know?
Lex Fridman (31:04.720)
You say stuff like that confidently, but you can't,
Lex Fridman (31:07.360)
and I love it, but the question is you can't,
Lex Fridman (31:12.080)
you're kind of going by feel cause you test it out.
George Hotz (31:14.400)
Absolutely, absolutely.
Lex Fridman (31:15.440)
Like I would take, I borrowed my friend's Tesla.
George Hotz (31:18.320)
I would take AP1 out for a drive
Lex Fridman (31:20.600)
and then I would take my system out for a drive.
Lex Fridman (31:22.160)
And it seems reasonably like the same.
Lex Fridman (31:25.920)
So the 405, how hard is it to create something
Lex Fridman (31:30.320)
that could actually be a product that's deployed?
Lex Fridman (31:34.040)
I mean, I've read an article where Elon,
George Hotz (31:37.120)
this respondent said something about you saying
Lex Fridman (31:40.640)
that to build autopilot is more complicated
George Hotz (31:46.920)
than a single George Hodge level job.
Lex Fridman (31:51.720)
How hard is that job to create something
Lex Fridman (31:55.360)
that would work across globally?
Lex Fridman (31:58.800)
Why don't think globally is the challenge?
Lex Fridman (32:00.480)
But Elon followed that up by saying
Lex Fridman (32:02.080)
it's gonna take two years in a company of 10 people.
Lex Fridman (32:04.760)
And here I am four years later with a company of 12 people.
Lex Fridman (32:07.760)
And I think we still have another two to go.
George Hotz (32:09.800)
Two years, so yeah.
Lex Fridman (32:11.160)
So what do you think about how Tesla is progressing
Lex Fridman (32:15.840)
with autopilot of V2, V3?
Lex Fridman (32:19.080)
I think we've kept pace with them pretty well.
George Hotz (32:23.960)
I think navigate and autopilot is terrible.
Lex Fridman (32:26.720)
We had some demo features internally of the same stuff
Lex Fridman (32:31.000)
and we would test it.
Lex Fridman (32:32.080)
And I'm like, I'm not shipping this
George Hotz (32:33.320)
even as like open source software to people.
Lex Fridman (32:35.160)
Why do you think it's terrible?
George Hotz (32:37.280)
Consumer Reports does a great job of describing it.
Lex Fridman (32:39.480)
Like when it makes a lane change,
George Hotz (32:41.160)
it does it worse than a human.
Lex Fridman (32:43.520)
You shouldn't ship things like autopilot, open pilot.
George Hotz (32:46.880)
They lane keep better than a human.
Lex Fridman (32:49.680)
If you turn it on for a stretch of a highway,
George Hotz (32:53.360)
like an hour long, it's never gonna touch a lane line.
Lex Fridman (32:56.600)
Human will touch probably a lane line twice.
George Hotz (32:58.960)
You just inspired me.
Lex Fridman (33:00.000)
I don't know if you're grounded in data on that.
George Hotz (33:02.120)
I read your paper.
Lex Fridman (33:03.200)
Okay, but that's interesting.
George Hotz (33:05.320)
I wonder actually how often we touch lane lines
Lex Fridman (33:10.560)
in general, like a little bit,
George Hotz (33:11.960)
because it is.
Lex Fridman (33:13.440)
I could answer that question pretty easily
George Hotz (33:14.920)
with the common data set.
Lex Fridman (33:15.760)
Yeah, I'm curious.
George Hotz (33:16.960)
I've never answered it.
Lex Fridman (33:17.800)
I don't know.
George Hotz (33:18.640)
I just, two is like my personal.
Lex Fridman (33:19.960)
It feels right.
George Hotz (33:21.720)
That's interesting.
Lex Fridman (33:22.560)
Because every time you touch a lane,
George Hotz (33:23.800)
that's a source of a little bit of stress
Lex Fridman (33:26.720)
and kind of lane keeping is removing that stress.
George Hotz (33:29.280)
That's ultimately the biggest value add honestly
Lex Fridman (33:32.320)
is just removing the stress of having to stay in lane.
Lex Fridman (33:35.520)
And I think honestly, I don't think people fully realize,
Lex Fridman (33:39.000)
first of all, that that's a big value add,
Lex Fridman (33:41.920)
but also that that's all it is.
Lex Fridman (33:44.960)
And that not only, I find it a huge value add.
George Hotz (33:48.560)
I drove down when we moved to San Diego,
Lex Fridman (33:50.400)
I drove down in a enterprise rental car and I missed it.
Lex Fridman (33:53.320)
So I missed having the system so much.
Lex Fridman (33:55.440)
It's so much more tiring to drive without it.
George Hotz (34:00.280)
It is that lane centering.
Lex Fridman (34:02.920)
That's the key feature.
George Hotz (34:04.800)
Yeah.
Lex Fridman (34:06.560)
And in a way, it's the only feature
George Hotz (34:08.920)
that actually adds value to people's lives
Lex Fridman (34:11.000)
in autonomous vehicles today.
George Hotz (34:12.160)
Waymo does not add value to people's lives.
Lex Fridman (34:13.800)
It's a more expensive, slower Uber.
George Hotz (34:15.840)
Maybe someday it'll be this big cliff where it adds value,
Lex Fridman (34:18.600)
but I don't usually believe it.
George Hotz (34:19.440)
It is fascinating.
Lex Fridman (34:20.280)
I haven't talked to, this is good.
George Hotz (34:22.520)
Cause I haven't, I have intuitively,
Lex Fridman (34:25.760)
but I think we're making it explicit now.
George Hotz (34:28.240)
I actually believe that really good lane keeping
Lex Fridman (34:35.440)
is a reason to buy a car.
George Hotz (34:37.200)
Will be a reason to buy a car and it's a huge value add.
Lex Fridman (34:39.680)
I've never, until we just started talking about it,
George Hotz (34:41.720)
I haven't really quite realized it.
Lex Fridman (34:43.840)
That I've felt with Elon's chase of level four
George Hotz (34:49.400)
is not the correct chase.
Lex Fridman (34:52.320)
It was on, cause you should just say Tesla has the best
George Hotz (34:55.920)
as if from a Tesla perspective, say,
Lex Fridman (34:58.280)
Tesla has the best lane keeping.
George Hotz (35:00.560)
Comma AI should say, Comma AI is the best lane keeping.
Lex Fridman (35:04.120)
And that is it.
George Hotz (35:05.560)
Yeah. Yeah.
Lex Fridman (35:06.400)
So do you think?
George Hotz (35:07.960)
You have to do the longitudinal as well.
Lex Fridman (35:09.880)
You can't just lane keep.
George Hotz (35:10.880)
You have to do ACC,
Lex Fridman (35:12.880)
but ACC is much more forgiving than lane keep,
George Hotz (35:15.760)
especially on the highway.
Lex Fridman (35:17.360)
By the way, are you Comma AI's camera only, correct?
George Hotz (35:21.920)
No, we use the radar.
Lex Fridman (35:23.680)
From the car, you're able to get the, okay.
Lex Fridman (35:25.440)
Hmm?
Lex Fridman (35:26.960)
We can do a camera only now.
George Hotz (35:28.800)
It's gotten to the point,
Lex Fridman (35:29.640)
but we leave the radar there as like a, it's fusion now.
George Hotz (35:33.440)
Okay, so let's maybe talk through some of the system specs
Lex Fridman (35:36.520)
on the hardware.
Lex Fridman (35:37.920)
What's the hardware side of what you're providing?
Lex Fridman (35:42.880)
What's the capabilities on the software side
Lex Fridman (35:44.720)
with OpenPilot and so on?
Lex Fridman (35:46.800)
So OpenPilot, as the box that we sell, that it runs on,
George Hotz (35:52.200)
it's a phone in a plastic case.
Lex Fridman (35:54.440)
It's nothing special.
George Hotz (35:55.280)
We sell it without the software.
Lex Fridman (35:56.680)
So you buy the phone, it's just easy.
George Hotz (35:59.360)
It'll be easy set up, but it's sold with no software.
Lex Fridman (36:03.960)
OpenPilot right now is about to be 0.6.
George Hotz (36:07.040)
When it gets to 1.0,
Lex Fridman (36:08.280)
I think we'll be ready for a consumer product.
George Hotz (36:10.120)
We're not gonna add any new features.
Lex Fridman (36:11.600)
We're just gonna make the lane keeping really, really good.
George Hotz (36:14.120)
Okay, I got it.
Lex Fridman (36:15.560)
So what do we have right now?
George Hotz (36:16.560)
It's a Snapdragon 820.
Lex Fridman (36:20.000)
It's a Sony IMX 298 forward facing camera.
George Hotz (36:24.080)
Driver monitoring camera,
Lex Fridman (36:26.040)
it's just a selfie camera on the phone.
Lex Fridman (36:27.960)
And a CAN transceiver,
Lex Fridman (36:31.640)
maybe there's a little thing called PANDAS.
Lex Fridman (36:33.840)
And they talk over USB to the phone
Lex Fridman (36:36.560)
and then they have three CAN buses
George Hotz (36:37.720)
that they talk to the car.
Lex Fridman (36:39.960)
One of those CAN buses is the radar CAN bus.
George Hotz (36:42.280)
One of them is the main car CAN bus
Lex Fridman (36:44.320)
and the other one is the proxy camera CAN bus.
George Hotz (36:46.240)
We leave the existing camera in place
Lex Fridman (36:48.400)
so we don't turn AEB off.
George Hotz (36:50.720)
Right now, we still turn AEB off
Lex Fridman (36:52.320)
if you're using our longitudinal,
Lex Fridman (36:53.600)
but we're gonna fix that before 1.0.
Lex Fridman (36:55.600)
Got it.
George Hotz (36:56.440)
Wow, that's cool.
Lex Fridman (36:57.280)
And it's CAN both ways.
Lex Fridman (36:59.040)
So how are you able to control vehicles?
Lex Fridman (37:03.320)
So we proxy,
George Hotz (37:05.440)
the vehicles that we work with
Lex Fridman (37:06.760)
already have a lane keeping assist system.
Lex Fridman (37:10.160)
So lane keeping assist can mean a huge variety of things.
Lex Fridman (37:13.800)
It can mean it will apply a small torque to the wheel
George Hotz (37:17.800)
after you've already crossed a lane line by a foot,
Lex Fridman (37:21.160)
which is the system in the older Toyotas
George Hotz (37:23.920)
versus like, I think Tesla still calls it
Lex Fridman (37:26.520)
lane keeping assist,
George Hotz (37:27.560)
where it'll keep you perfectly
Lex Fridman (37:28.840)
in the center of the lane on the highway.
George Hotz (37:32.320)
You can control, like with the joystick, the car.
Lex Fridman (37:35.080)
So these cars already have the capability of drive by wire.
Lex Fridman (37:37.920)
So is it trivial to convert a car that it operates with?
Lex Fridman (37:45.400)
OpenPILOT is able to control the steering?
George Hotz (37:48.480)
Oh, a new car or a car that we,
Lex Fridman (37:49.720)
so we have support now for 45 different makes of cars.
Lex Fridman (37:52.800)
What are the cars in general?
Lex Fridman (37:54.880)
Mostly Hondas and Toyotas.
George Hotz (37:56.360)
We support almost every Honda and Toyota made this year.
Lex Fridman (38:01.680)
And then a bunch of GMs, a bunch of Subarus,
George Hotz (38:04.480)
a bunch of Chevys.
Lex Fridman (38:05.320)
It doesn't have to be like a Prius,
George Hotz (38:06.160)
it could be a Corolla as well.
Lex Fridman (38:07.320)
Oh, the 2020 Corolla is the best car with OpenPILOT.
George Hotz (38:10.760)
It just came out.
Lex Fridman (38:11.720)
The actuator has less lag than the older Corolla.
George Hotz (38:15.800)
I think I started watching a video with your,
Lex Fridman (38:18.240)
I mean, the way you make videos is awesome.
George Hotz (38:21.400)
You're just literally at the dealership streaming.
Lex Fridman (38:24.220)
Yeah, I had my friend on the phone,
Lex Fridman (38:26.060)
I'm like, bro, you wanna stream for an hour?
Lex Fridman (38:27.520)
Yeah, and basically, like if stuff goes a little wrong,
George Hotz (38:31.100)
you're just like, you just go with it.
Lex Fridman (38:33.120)
Yeah, I love it.
George Hotz (38:33.940)
Well, it's real.
Lex Fridman (38:34.780)
Yeah, it's real.
George Hotz (38:35.600)
That's so beautiful and it's so in contrast
Lex Fridman (38:39.760)
to the way other companies
George Hotz (38:42.960)
would put together a video like that.
Lex Fridman (38:44.680)
Kind of why I like to do it like that.
George Hotz (38:46.080)
Good.
Lex Fridman (38:46.920)
I mean, if you become super rich one day and successful,
George Hotz (38:49.840)
I hope you keep it that way
Lex Fridman (38:50.800)
because I think that's actually what people love,
George Hotz (38:53.200)
that kind of genuine.
Lex Fridman (38:54.760)
Oh, it's all that has value to me.
George Hotz (38:56.520)
Money has no, if I sell out to like make money,
Lex Fridman (38:59.920)
I sold out, it doesn't matter.
Lex Fridman (39:01.320)
What do I get?
Lex Fridman (39:02.160)
Yacht?
George Hotz (39:03.000)
I don't want a yacht.
Lex Fridman (39:04.560)
And I think Tesla's actually has a small inkling
George Hotz (39:09.240)
of that as well with Autonomy Day.
Lex Fridman (39:11.320)
They did reveal more than, I mean, of course,
George Hotz (39:14.080)
there's marketing communications, you could tell,
Lex Fridman (39:15.760)
but it's more than most companies would reveal,
George Hotz (39:17.720)
which is, I hope they go towards that direction more,
Lex Fridman (39:21.440)
other companies, GM, Ford.
George Hotz (39:23.080)
Oh, Tesla's gonna win level five.
Lex Fridman (39:25.440)
They really are.
Lex Fridman (39:26.600)
So let's talk about it.
Lex Fridman (39:27.840)
You think, you're focused on level two currently?
George Hotz (39:32.280)
Currently.
Lex Fridman (39:33.120)
We're gonna be one to two years behind Tesla
George Hotz (39:36.200)
getting to level five.
Lex Fridman (39:37.200)
Okay.
Lex Fridman (39:38.560)
We're Android, right?
Lex Fridman (39:39.400)
We're Android.
George Hotz (39:40.220)
You're Android.
Lex Fridman (39:41.060)
I'm just saying, once Tesla gets it,
George Hotz (39:42.280)
we're one to two years behind.
Lex Fridman (39:43.800)
I'm not making any timeline on when Tesla's
George Hotz (39:45.640)
gonna get it. That's right.
Lex Fridman (39:46.480)
You did, that was brilliant.
George Hotz (39:47.300)
I'm sorry, Tesla investors,
Lex Fridman (39:48.400)
if you think you're gonna have an autonomous
George Hotz (39:49.880)
Robo Taxi fleet by the end of the year.
Lex Fridman (39:52.460)
Yeah, so that's.
George Hotz (39:53.300)
I'll bet against that.
Lex Fridman (39:54.960)
So what do you think about this?
George Hotz (39:57.740)
The most level four companies
Lex Fridman (40:02.000)
are kind of just doing their usual safety driver,
George Hotz (40:07.280)
doing full autonomy kind of testing.
Lex Fridman (40:08.800)
And then Tesla does basically trying to go
George Hotz (40:12.000)
from lane keeping to full autonomy.
Lex Fridman (40:15.600)
What do you think about that approach?
Lex Fridman (40:16.840)
How successful would it be?
Lex Fridman (40:18.400)
It's a ton better approach.
George Hotz (40:20.720)
Because Tesla is gathering data on a scale
Lex Fridman (40:23.980)
that none of them are.
George Hotz (40:25.240)
They're putting real users behind the wheel of the cars.
Lex Fridman (40:29.560)
It's, I think, the only strategy that works.
George Hotz (40:33.120)
The incremental.
Lex Fridman (40:34.480)
Well, so there's a few components to Tesla approach
George Hotz (40:36.980)
that's more than just the incrementalists.
Lex Fridman (40:38.920)
What you spoke with is the ones, the software,
Lex Fridman (40:41.440)
so over the air software updates.
Lex Fridman (40:43.760)
Necessity.
George Hotz (40:44.840)
I mean Waymo crews have those too.
Lex Fridman (40:46.480)
Those aren't.
George Hotz (40:47.660)
But.
Lex Fridman (40:48.500)
Those differentiate from the automakers.
George Hotz (40:49.880)
Right, no lane keeping systems have,
Lex Fridman (40:52.040)
no cars with lane keeping system have that except Tesla.
George Hotz (40:54.840)
Yeah.
Lex Fridman (40:55.800)
And the other one is the data, the other direction,
George Hotz (40:59.840)
which is the ability to query the data.
Lex Fridman (41:01.920)
I don't think they're actually collecting
George Hotz (41:03.560)
as much data as people think,
Lex Fridman (41:04.580)
but the ability to turn on collection and turn it off.
Lex Fridman (41:09.520)
So I'm both in the robotics world
Lex Fridman (41:12.120)
and the psychology human factors world.
George Hotz (41:15.080)
Many people believe that level two autonomy is problematic
Lex Fridman (41:18.540)
because of the human factor.
George Hotz (41:20.120)
Like the more the task is automated,
Lex Fridman (41:23.380)
the more there's a vigilance decrement.
George Hotz (41:26.080)
You start to fall asleep.
Lex Fridman (41:27.240)
You start to become complacent,
George Hotz (41:28.600)
start texting more and so on.
Lex Fridman (41:30.560)
Do you worry about that?
George Hotz (41:32.320)
Cause if we're talking about transition from lane keeping
Lex Fridman (41:35.080)
to full autonomy, if you're spending 80% of the time,
George Hotz (41:39.880)
not supervising the machine,
Lex Fridman (41:42.840)
do you worry about what that means
Lex Fridman (41:45.480)
to the safety of the drivers?
Lex Fridman (41:47.140)
One, we don't consider open pilot to be 1.0
George Hotz (41:49.640)
until we have 100% driver monitoring.
Lex Fridman (41:52.360)
You can cheat right now, our driver monitoring system.
George Hotz (41:55.040)
There's a few ways to cheat it.
Lex Fridman (41:56.080)
They're pretty obvious.
George Hotz (41:58.200)
We're working on making that better.
Lex Fridman (41:59.720)
Before we ship a consumer product that can drive cars,
George Hotz (42:02.560)
I want to make sure that I have driver monitoring
Lex Fridman (42:04.240)
that you can't cheat.
Lex Fridman (42:05.480)
What's like a successful driver monitoring system look like?
Lex Fridman (42:07.840)
Is it all about just keeping your eyes on the road?
George Hotz (42:11.720)
Well, a few things.
Lex Fridman (42:12.760)
So that's what we went with at first for driver monitoring.
George Hotz (42:16.640)
I'm checking, I'm actually looking at
Lex Fridman (42:18.040)
where your head is looking.
George Hotz (42:19.040)
The camera's not that high resolution.
Lex Fridman (42:20.440)
Eyes are a little bit hard to get.
George Hotz (42:21.880)
Well, head is this big.
Lex Fridman (42:22.920)
I mean, that's.
George Hotz (42:23.760)
Head is good.
Lex Fridman (42:24.680)
And actually a lot of it, just psychology wise,
George Hotz (42:28.740)
to have that monitor constantly there,
Lex Fridman (42:30.760)
it reminds you that you have to be paying attention.
Lex Fridman (42:33.480)
But we want to go further.
Lex Fridman (42:35.120)
We just hired someone full time
George Hotz (42:36.400)
to come on to do the driver monitoring.
Lex Fridman (42:37.960)
I want to detect phone in frame
Lex Fridman (42:40.400)
and I want to make sure you're not sleeping.
Lex Fridman (42:42.600)
How much does the camera see of the body?
George Hotz (42:44.880)
This one, not enough.
Lex Fridman (42:47.480)
Not enough.
George Hotz (42:48.440)
The next one, everything.
Lex Fridman (42:50.760)
Well, it's interesting, Fisheye,
George Hotz (42:51.600)
because we're doing just data collection, not real time.
Lex Fridman (42:55.240)
But Fisheye is a beautiful,
George Hotz (42:57.600)
being able to capture the body.
Lex Fridman (42:59.080)
And the smartphone is really like the biggest problem.
George Hotz (43:03.280)
I'll show you.
Lex Fridman (43:04.120)
I can show you one of the pictures from our new system.
George Hotz (43:06.360)
Awesome, so you're basically saying
Lex Fridman (43:09.680)
the driver monitoring will be the answer to that.
George Hotz (43:13.120)
I think the other point
Lex Fridman (43:14.320)
that you raised in your paper is good as well.
George Hotz (43:16.960)
You're not asking a human to supervise a machine
Lex Fridman (43:20.480)
without giving them the,
George Hotz (43:21.680)
they can take over at any time.
Lex Fridman (43:23.220)
Right.
George Hotz (43:24.060)
Our safety model, you can take over.
Lex Fridman (43:25.800)
We disengage on both the gas or the brake.
George Hotz (43:27.880)
We don't disengage on steering.
Lex Fridman (43:28.900)
I don't feel you have to.
Lex Fridman (43:30.020)
But we disengage on gas or brake.
Lex Fridman (43:31.760)
So it's very easy for you to take over
Lex Fridman (43:34.320)
and it's very easy for you to reengage.
Lex Fridman (43:36.440)
That switching should be super cheap.
George Hotz (43:39.380)
The cars that require,
Lex Fridman (43:40.240)
even autopilot requires a double press.
George Hotz (43:42.440)
That's almost, I see, I don't like that.
Lex Fridman (43:44.400)
And then the cancel, to cancel in autopilot,
George Hotz (43:48.080)
you either have to press cancel,
Lex Fridman (43:49.040)
which no one knows what that is, so they press the brake.
Lex Fridman (43:51.040)
But a lot of times you don't actually want
Lex Fridman (43:52.120)
to press the brake.
George Hotz (43:53.380)
You want to press the gas.
Lex Fridman (43:54.560)
So you should cancel on gas.
George Hotz (43:55.920)
Or wiggle the steering wheel, which is bad as well.
Lex Fridman (43:57.960)
Wow, that's brilliant.
George Hotz (43:58.920)
I haven't heard anyone articulate that point.
Lex Fridman (44:01.440)
Oh, this is all I think about.
George Hotz (44:03.480)
It's the, because I think,
Lex Fridman (44:06.960)
I think actually Tesla has done a better job
George Hotz (44:09.800)
than most automakers at making that frictionless.
Lex Fridman (44:12.920)
But you just described that it could be even better.
George Hotz (44:16.600)
I love Super Cruise as an experience once it's engaged.
Lex Fridman (44:21.160)
I don't know if you've used it,
Lex Fridman (44:22.040)
but getting the thing to try to engage.
Lex Fridman (44:25.040)
Yeah, I've used the, I've driven Super Cruise a lot.
Lex Fridman (44:27.520)
So what's your thoughts on the Super Cruise system?
Lex Fridman (44:29.440)
You disengage Super Cruise and it falls back to ACC.
Lex Fridman (44:32.680)
So my car's like still accelerating.
Lex Fridman (44:34.640)
It feels weird.
George Hotz (44:36.280)
Otherwise, when you actually have Super Cruise engaged
Lex Fridman (44:39.040)
on the highway, it is phenomenal.
George Hotz (44:41.200)
We bought that Cadillac.
Lex Fridman (44:42.320)
We just sold it.
Lex Fridman (44:43.320)
But we bought it just to like experience this.
Lex Fridman (44:45.620)
And I wanted everyone in the office to be like,
George Hotz (44:47.320)
this is what we're striving to build.
Lex Fridman (44:49.400)
GM pioneering with the driver monitoring.
Lex Fridman (44:52.800)
You like their driver monitoring system?
Lex Fridman (44:55.000)
It has some bugs.
George Hotz (44:56.400)
If there's a sun shining back here, it'll be blind to you.
Lex Fridman (45:00.280)
Right.
Lex Fridman (45:01.960)
But overall, mostly, yeah.
Lex Fridman (45:03.340)
That's so cool that you know all this stuff.
George Hotz (45:05.920)
I don't often talk to people that,
Lex Fridman (45:08.460)
because it's such a rare car, unfortunately, currently.
George Hotz (45:10.980)
We bought one explicitly for this.
Lex Fridman (45:12.700)
We lost like 25K in the deprecation,
Lex Fridman (45:15.020)
but I feel it's worth it.
Lex Fridman (45:16.700)
I was very pleasantly surprised that GM system
George Hotz (45:21.260)
was so innovative and really wasn't advertised much,
Lex Fridman (45:26.320)
wasn't talked about much.
George Hotz (45:27.460)
Yeah.
Lex Fridman (45:28.460)
And I was nervous that it would die,
George Hotz (45:30.420)
that it would disappear.
Lex Fridman (45:31.860)
Well, they put it on the wrong car.
George Hotz (45:33.500)
They should have put it on the Bolt
Lex Fridman (45:34.580)
and not some weird Cadillac that nobody bought.
George Hotz (45:36.620)
I think that's gonna be into,
Lex Fridman (45:38.420)
they're saying at least it's gonna be
George Hotz (45:40.020)
into their entire fleet.
Lex Fridman (45:41.820)
So what do you think about,
George Hotz (45:43.820)
as long as we're on the driver monitoring,
Lex Fridman (45:45.940)
what do you think about Elon Musk's claim
Lex Fridman (45:49.280)
that driver monitoring is not needed?
Lex Fridman (45:51.940)
Normally, I love his claims.
George Hotz (45:53.700)
That one is stupid.
Lex Fridman (45:55.560)
That one is stupid.
George Hotz (45:56.580)
And, you know, he's not gonna have his level five fleet
Lex Fridman (46:00.320)
by the end of the year.
George Hotz (46:01.340)
Hopefully he's like, okay, I was wrong.
Lex Fridman (46:04.900)
I'm gonna add driver monitoring.
George Hotz (46:06.260)
Because when these systems get to the point
Lex Fridman (46:08.260)
that they're only messing up once every thousand miles,
George Hotz (46:10.340)
you absolutely need driver monitoring.
Lex Fridman (46:14.060)
So let me play, cause I agree with you,
Lex Fridman (46:15.900)
but let me play devil's advocate.
Lex Fridman (46:17.340)
One possibility is that without driver monitoring,
George Hotz (46:22.340)
people are able to monitor, self regulate,
Lex Fridman (46:26.420)
monitor themselves.
George Hotz (46:28.260)
You know, that, so your idea is.
Lex Fridman (46:30.500)
You've seen all the people sleeping in Teslas?
George Hotz (46:33.860)
Yeah, well, I'm a little skeptical
Lex Fridman (46:37.340)
of all the people sleeping in Teslas
George Hotz (46:38.860)
because I've stopped paying attention to that kind of stuff
Lex Fridman (46:44.260)
because I want to see real data.
George Hotz (46:45.660)
It's too much glorified.
Lex Fridman (46:47.180)
It doesn't feel scientific to me.
Lex Fridman (46:48.660)
So I want to know how many people are really sleeping
Lex Fridman (46:52.500)
in Teslas versus sleeping.
George Hotz (46:54.620)
I was driving here sleep deprived in a car
Lex Fridman (46:58.060)
with no automation.
George Hotz (46:59.420)
I was falling asleep.
Lex Fridman (47:00.980)
I agree that it's hypey.
Lex Fridman (47:02.060)
It's just like, you know what?
Lex Fridman (47:04.780)
If you want to put driver monitoring,
George Hotz (47:06.020)
I rented a, my last autopilot experience
Lex Fridman (47:08.420)
was I rented a model three in March and drove it around.
George Hotz (47:12.140)
The wheel thing is annoying.
Lex Fridman (47:13.500)
And the reason the wheel thing is annoying,
George Hotz (47:15.340)
we use the wheel thing as well,
Lex Fridman (47:16.700)
but we don't disengage on wheel.
George Hotz (47:18.620)
For Tesla, you have to touch the wheel just enough
Lex Fridman (47:21.620)
to trigger the torque sensor, to tell it that you're there,
Lex Fridman (47:25.260)
but not enough as to disengage it,
Lex Fridman (47:28.340)
which don't use it for two things.
George Hotz (47:30.380)
Don't disengage on wheel.
Lex Fridman (47:31.300)
You don't have to.
George Hotz (47:32.340)
That whole experience, wow, beautifully put.
Lex Fridman (47:35.300)
All of those elements,
George Hotz (47:36.300)
even if you don't have driver monitoring,
Lex Fridman (47:38.180)
that whole experience needs to be better.
George Hotz (47:41.060)
Driver monitoring, I think would make,
Lex Fridman (47:43.700)
I mean, I think Super Cruise is a better experience
George Hotz (47:46.140)
once it's engaged over autopilot.
Lex Fridman (47:48.340)
I think Super Cruise is a transition
George Hotz (47:50.900)
to engagement and disengagement are significantly worse.
Lex Fridman (47:53.900)
Yeah.
George Hotz (47:54.900)
Well, there's a tricky thing,
Lex Fridman (47:56.340)
because if I were to criticize Super Cruise is,
George Hotz (47:59.660)
it's a little too crude.
Lex Fridman (48:00.740)
And I think like six seconds or something,
George Hotz (48:03.580)
if you look off road, it'll start warning you.
Lex Fridman (48:05.980)
It's some ridiculously long period of time.
Lex Fridman (48:09.020)
And just the way,
Lex Fridman (48:12.740)
I think it's basically, it's a binary.
George Hotz (48:15.740)
It should be adaptive.
Lex Fridman (48:17.180)
Yeah, it needs to learn more about you.
George Hotz (48:19.820)
It needs to communicate what it sees about you more.
Lex Fridman (48:24.380)
Tesla shows what it sees about the external world.
George Hotz (48:27.100)
It would be nice if Super Cruise would tell us
Lex Fridman (48:29.060)
what it sees about the internal world.
George Hotz (48:30.780)
It's even worse than that.
Lex Fridman (48:31.900)
You press the button to engage
Lex Fridman (48:33.260)
and it just says Super Cruise unavailable.
Lex Fridman (48:35.380)
Yeah. Why?
Lex Fridman (48:36.220)
Why?
Lex Fridman (48:37.740)
Yeah, that transparency is good.
George Hotz (48:41.420)
We've renamed the driver monitoring packet to driver state.
Lex Fridman (48:45.300)
Driver state.
George Hotz (48:46.140)
We have car state packet, which has the state of the car.
Lex Fridman (48:48.220)
And you have driver state packet,
George Hotz (48:49.380)
which has the state of the driver.
Lex Fridman (48:50.940)
So what is the...
George Hotz (48:52.060)
Estimate their BAC.
Lex Fridman (48:53.980)
What's BAC?
George Hotz (48:54.820)
Blood alcohol content.
Lex Fridman (48:57.260)
You think that's possible with computer vision?
George Hotz (48:59.100)
Absolutely.
Lex Fridman (49:03.300)
To me, it's an open question.
George Hotz (49:04.420)
I haven't looked into it too much.
Lex Fridman (49:06.580)
Actually, I quite seriously looked at the literature.
George Hotz (49:08.380)
It's not obvious to me that from the eyes and so on,
Lex Fridman (49:10.780)
you can tell.
George Hotz (49:11.620)
You might need stuff from the car as well.
Lex Fridman (49:13.140)
Yeah.
Lex Fridman (49:13.980)
You might need how they're controlling the car, right?
Lex Fridman (49:15.700)
And that's fundamentally at the end of the day,
Lex Fridman (49:17.340)
what you care about.
Lex Fridman (49:18.620)
But I think, especially when people are really drunk,
George Hotz (49:21.620)
they're not controlling the car nearly as smoothly
Lex Fridman (49:23.620)
as they would look at them walking, right?
George Hotz (49:25.460)
The car is like an extension of the body.
Lex Fridman (49:27.220)
So I think you could totally detect.
Lex Fridman (49:29.380)
And if you could fix people who are drunk, distracted,
Lex Fridman (49:31.340)
asleep, if you fix those three.
George Hotz (49:32.820)
Yeah, that's huge.
Lex Fridman (49:35.460)
So what are the current limitations of open pilot?
Lex Fridman (49:38.220)
What are the main problems that still need to be solved?
Lex Fridman (49:41.700)
We're hopefully fixing a few of them in 06.
George Hotz (49:45.420)
We're not as good as autopilot at stop cars.
Lex Fridman (49:49.460)
So if you're coming up to a red light at 55,
Lex Fridman (49:55.180)
so it's the radar stopped car problem, which
Lex Fridman (49:57.060)
is responsible for two autopilot accidents,
George Hotz (49:59.180)
it's hard to differentiate a stopped car from a signpost.
Lex Fridman (50:03.580)
Yeah, a static object.
Lex Fridman (50:05.300)
So you have to fuse.
Lex Fridman (50:06.300)
You have to do this visually.
George Hotz (50:07.500)
There's no way from the radar data to tell the difference.
Lex Fridman (50:09.580)
Maybe you can make a map, but I don't really
George Hotz (50:11.540)
believe in mapping at all anymore.
Lex Fridman (50:13.820)
Wait, wait, wait, what, you don't believe in mapping?
George Hotz (50:16.020)
No.
Lex Fridman (50:16.660)
So you basically, the open pilot solution
George Hotz (50:20.660)
is saying react to the environment as you see it,
Lex Fridman (50:22.660)
just like human beings do.
Lex Fridman (50:24.460)
And then eventually, when you want
Lex Fridman (50:25.820)
to do navigate on open pilot, I'll
George Hotz (50:28.380)
train the net to look at ways.
Lex Fridman (50:29.940)
I'll run ways in the background, I'll
George Hotz (50:31.460)
train a confident way.
Lex Fridman (50:32.300)
Are you using GPS at all?
George Hotz (50:34.540)
We use it to ground truth.
Lex Fridman (50:35.940)
We use it to very carefully ground truth the paths.
George Hotz (50:38.340)
We have a stack which can recover relative to 10
Lex Fridman (50:40.980)
centimeters over one minute.
Lex Fridman (50:42.940)
And then we use that to ground truth exactly where
Lex Fridman (50:45.020)
the car went in that local part of the environment,
Lex Fridman (50:47.420)
but it's all local.
Lex Fridman (50:48.700)
How are you testing in general, just for yourself,
Lex Fridman (50:50.780)
like experiments and stuff?
Lex Fridman (50:53.220)
Where are you located?
George Hotz (50:54.940)
San Diego.
Lex Fridman (50:55.540)
San Diego.
George Hotz (50:56.140)
Yeah.
Lex Fridman (50:56.780)
OK.
Lex Fridman (50:58.660)
So you basically drive around there, collect some data,
Lex Fridman (51:01.420)
and watch the performance?
George Hotz (51:03.060)
We have a simulator now.
Lex Fridman (51:04.300)
And we have, our simulator is really cool.
George Hotz (51:06.420)
Our simulator is not, it's not like a Unity based simulator.
Lex Fridman (51:09.660)
Our simulator lets us load in real state.
Lex Fridman (51:12.820)
What do you mean?
Lex Fridman (51:13.620)
We can load in a drive and simulate
Lex Fridman (51:16.700)
what the system would have done on the historical data.
Lex Fridman (51:20.260)
Ooh, nice.
George Hotz (51:22.460)
Interesting.
Lex Fridman (51:23.460)
So what, yeah.
George Hotz (51:24.260)
Right now we're only using it for testing,
Lex Fridman (51:26.060)
but as soon as we start using it for training, that's it.
George Hotz (51:29.140)
That's all that matters.
Lex Fridman (51:30.780)
What's your feeling about the real world versus simulation?
Lex Fridman (51:33.020)
Do you like simulation for training,
Lex Fridman (51:34.420)
if this moves to training?
Lex Fridman (51:35.700)
So we have to distinguish two types of simulators, right?
Lex Fridman (51:40.020)
There's a simulator that is completely fake.
George Hotz (51:44.620)
I could get my car to drive around in GTA.
Lex Fridman (51:47.740)
I feel that this kind of simulator is useless.
George Hotz (51:51.780)
You're never, there's so many.
Lex Fridman (51:54.580)
My analogy here is like, OK, fine.
George Hotz (51:56.940)
You're not solving the computer vision problem,
Lex Fridman (51:59.860)
but you're solving the computer graphics problem.
George Hotz (52:02.300)
Right.
Lex Fridman (52:02.780)
And you don't think you can get very far by creating
Lex Fridman (52:05.300)
ultra realistic graphics?
Lex Fridman (52:07.980)
No, because you can create ultra realistic graphics
George Hotz (52:10.340)
of the road, now create ultra realistic behavioral models
Lex Fridman (52:13.140)
of the other cars.
George Hotz (52:14.500)
Oh, well, I'll just use myself driving.
Lex Fridman (52:16.860)
No, you won't.
George Hotz (52:18.180)
You need actual human behavior, because that's
Lex Fridman (52:22.180)
what you're trying to learn.
George Hotz (52:23.860)
Driving does not have a spec.
Lex Fridman (52:25.820)
The definition of driving is what humans do when they drive.
George Hotz (52:29.860)
Whatever Waymo does, I don't think it's driving.
Lex Fridman (52:32.700)
Right.
George Hotz (52:33.220)
Well, I think actually Waymo and others,
Lex Fridman (52:36.380)
if there's any use for reinforcement learning,
George Hotz (52:38.980)
I've seen it used quite well.
Lex Fridman (52:40.340)
I study pedestrians a lot, too, is
George Hotz (52:42.020)
try to train models from real data of how pedestrians move,
Lex Fridman (52:45.500)
and try to use reinforcement learning models to make
George Hotz (52:47.540)
pedestrians move in human like ways.
Lex Fridman (52:49.980)
By that point, you've already gone so many layers,
Lex Fridman (52:53.500)
you detected a pedestrian?
Lex Fridman (52:55.660)
Did you hand code the feature vector of their state?
George Hotz (53:00.180)
Did you guys learn anything from computer vision
Lex Fridman (53:02.860)
before deep learning?
George Hotz (53:04.580)
Well, OK, I feel like this is.
Lex Fridman (53:07.140)
So perception to you is the sticking point.
Lex Fridman (53:10.820)
I mean, what's the hardest part of the stack here?
Lex Fridman (53:13.780)
There is no human understandable feature vector separating
George Hotz (53:20.500)
perception and planning.
Lex Fridman (53:23.060)
That's the best way I can put that.
George Hotz (53:25.100)
There is no, so it's all together,
Lex Fridman (53:26.780)
and it's a joint problem.
Lex Fridman (53:29.540)
So you can take localization.
Lex Fridman (53:31.460)
Localization and planning, there is
George Hotz (53:33.260)
a human understandable feature vector between these two
Lex Fridman (53:35.300)
things.
George Hotz (53:35.980)
I mean, OK, so I have like three degrees position,
Lex Fridman (53:38.540)
three degrees orientation, and those derivatives,
George Hotz (53:40.540)
maybe those second derivatives.
Lex Fridman (53:41.980)
That's human understandable.
George Hotz (53:43.140)
That's physical.
Lex Fridman (53:45.460)
Between perception and planning, so like Waymo
George Hotz (53:50.780)
has a perception stack and then a planner.
Lex Fridman (53:53.620)
And one of the things Waymo does right
George Hotz (53:55.580)
is they have a simulator that can separate those two.
Lex Fridman (54:00.020)
They can like replay their perception data
Lex Fridman (54:02.900)
and test their system, which is what
Lex Fridman (54:04.380)
I'm talking about about like the two
George Hotz (54:05.380)
different kinds of simulators.
Lex Fridman (54:06.500)
There's the kind that can work on real data,
Lex Fridman (54:08.220)
and there's the kind that can't work on real data.
Lex Fridman (54:10.860)
Now, the problem is that I don't think you can hand code
Lex Fridman (54:14.900)
a feature vector, right?
Lex Fridman (54:16.140)
Like you have some list of like, oh, here's
George Hotz (54:17.740)
my list of cars in the scenes.
Lex Fridman (54:19.100)
Here's my list of pedestrians in the scene.
George Hotz (54:21.220)
This isn't what humans are doing.
Lex Fridman (54:23.180)
What are humans doing?
George Hotz (54:24.860)
Global.
Lex Fridman (54:27.180)
And you're saying that's too difficult to hand engineer.
George Hotz (54:31.860)
I'm saying that there is no state vector given a perfect.
Lex Fridman (54:35.020)
I could give you the best team of engineers in the world
George Hotz (54:37.300)
to build a perception system and the best team
Lex Fridman (54:39.060)
to build a planner.
George Hotz (54:40.580)
All you have to do is define the state vector
Lex Fridman (54:42.660)
that separates those two.
George Hotz (54:43.860)
I'm missing the state vector that separates those two.
Lex Fridman (54:48.580)
What do you mean?
Lex Fridman (54:49.300)
So what is the output of your perception system?
Lex Fridman (54:53.860)
Output of the perception system, it's, OK, well,
George Hotz (55:00.500)
there's several ways to do it.
Lex Fridman (55:01.620)
One is the SLAM components localization.
George Hotz (55:03.780)
The other is drivable area, drivable space.
Lex Fridman (55:05.780)
Drivable space, yeah.
Lex Fridman (55:06.580)
And then there's the different objects in the scene.
Lex Fridman (55:10.860)
And different objects in the scene over time,
George Hotz (55:15.340)
maybe, to give you input to then try
Lex Fridman (55:17.660)
to start modeling the trajectories of those objects.
George Hotz (55:21.500)
Sure.
Lex Fridman (55:22.140)
That's it.
George Hotz (55:22.740)
I can give you a concrete example
Lex Fridman (55:24.060)
of something you missed.
Lex Fridman (55:25.060)
What's that?
Lex Fridman (55:25.780)
So say there's a bush in the scene.
George Hotz (55:28.580)
Humans understand that when they see this bush
Lex Fridman (55:30.860)
that there may or may not be a car behind that bush.
George Hotz (55:34.580)
Drivable area and a list of objects does not include that.
Lex Fridman (55:37.180)
Humans are doing this constantly at the simplest intersections.
Lex Fridman (55:40.820)
So now you have to talk about occluded area.
Lex Fridman (55:44.900)
But even that, what do you mean by occluded?
George Hotz (55:47.740)
OK, so I can't see it.
Lex Fridman (55:49.500)
Well, if it's the other side of a house, I don't care.
George Hotz (55:51.740)
What's the likelihood that there's
Lex Fridman (55:53.100)
a car in that occluded area?
Lex Fridman (55:55.180)
And if you say, OK, we'll add that,
Lex Fridman (55:57.940)
I can come up with 10 more examples that you can't add.
George Hotz (56:01.620)
Certainly, occluded area would be something
Lex Fridman (56:03.860)
that Simulator would have because it's
George Hotz (56:05.860)
simulating the entire occlusion is part of it.
Lex Fridman (56:11.180)
Occlusion is part of a vision stack.
Lex Fridman (56:12.580)
But what I'm saying is if you have a hand engineered,
Lex Fridman (56:16.500)
if your perception system output can
George Hotz (56:19.420)
be written in a spec document, it is incomplete.
Lex Fridman (56:22.980)
Yeah, I mean, certainly, it's hard to argue with that
George Hotz (56:27.740)
because in the end, that's going to be true.
Lex Fridman (56:30.100)
Yeah, and I'll tell you what the output of our perception
George Hotz (56:32.260)
system is.
Lex Fridman (56:32.740)
What's that?
George Hotz (56:33.300)
It's a 1,024 dimensional vector, trained by neural net.
Lex Fridman (56:37.940)
Oh, you know that.
George Hotz (56:38.980)
No, it's 1,024 dimensions of who knows what.
Lex Fridman (56:43.460)
Because it's operating on real data.
George Hotz (56:45.060)
Yeah.
Lex Fridman (56:46.940)
And that's the perception.
George Hotz (56:48.340)
That's the perception state.
Lex Fridman (56:50.380)
Think about an autoencoder for faces.
George Hotz (56:53.500)
If you have an autoencoder for faces and you say
Lex Fridman (56:56.660)
it has 256 dimensions in the middle,
Lex Fridman (56:59.260)
and I'm taking a face over here and projecting it
Lex Fridman (57:01.260)
to a face over here.
Lex Fridman (57:02.820)
Can you hand label all 256 of those dimensions?
Lex Fridman (57:06.300)
Well, no, but those have to generate automatically.
Lex Fridman (57:09.260)
But even if you tried to do it by hand,
Lex Fridman (57:11.380)
could you come up with a spec between your encoder
Lex Fridman (57:15.580)
and your decoder?
Lex Fridman (57:17.660)
No, because it wasn't designed, but there.
George Hotz (57:20.740)
No, no, no, but if you could design it.
Lex Fridman (57:23.620)
If you could design a face reconstructor system,
Lex Fridman (57:26.460)
could you come up with a spec?
Lex Fridman (57:29.260)
No, but I think we're missing here a little bit.
George Hotz (57:32.340)
I think you're just being very poetic about expressing
Lex Fridman (57:35.980)
a fundamental problem of simulators,
George Hotz (57:38.940)
that they're going to be missing so much that the feature
Lex Fridman (57:44.300)
vector will just look fundamentally different
George Hotz (57:47.500)
in the simulated world than the real world.
Lex Fridman (57:51.260)
I'm not making a claim about simulators.
George Hotz (57:53.820)
I'm making a claim about the spec division
Lex Fridman (57:57.060)
between perception and planning, even in your system.
George Hotz (58:00.780)
Just in general.
Lex Fridman (58:01.980)
Just in general.
George Hotz (58:03.300)
If you're trying to build a car that drives,
Lex Fridman (58:05.620)
if you're trying to hand code the output of your perception
George Hotz (58:08.340)
system, like saying, here's a list of all the cars
Lex Fridman (58:10.340)
in the scene, here's a list of all the people,
George Hotz (58:11.860)
here's a list of the occluded areas,
Lex Fridman (58:13.060)
here's a vector of drivable areas, it's insufficient.
Lex Fridman (58:16.540)
And if you start to believe that,
Lex Fridman (58:17.900)
you realize that what Waymo and Cruz are doing is impossible.
George Hotz (58:20.780)
Currently, what we're doing is the perception problem
Lex Fridman (58:24.220)
is converting the scene into a chessboard.
Lex Fridman (58:29.140)
And then you reason some basic reasoning
Lex Fridman (58:31.660)
around that chessboard.
Lex Fridman (58:33.340)
And you're saying that really, there's a lot missing there.
Lex Fridman (58:38.380)
First of all, why are we talking about this?
Lex Fridman (58:40.180)
Because isn't this a full autonomy?
Lex Fridman (58:42.740)
Is this something you think about?
George Hotz (58:44.580)
Oh, I want to win self driving cars.
Lex Fridman (58:47.540)
So your definition of win includes?
George Hotz (58:51.940)
Level four or five.
Lex Fridman (58:53.060)
Level five.
George Hotz (58:53.900)
I don't think level four is a real thing.
Lex Fridman (58:55.740)
I want to build the AlphaGo of driving.
Lex Fridman (59:01.060)
So AlphaGo is really end to end.
Lex Fridman (59:06.060)
Yeah.
George Hotz (59:06.900)
Is, yeah, it's end to end.
Lex Fridman (59:09.780)
And do you think this whole problem,
George Hotz (59:12.420)
is that also kind of what you're getting at
Lex Fridman (59:14.620)
with the perception and the planning?
George Hotz (59:16.580)
Is that this whole problem, the right way to do it
Lex Fridman (59:19.380)
is really to learn the entire thing.
George Hotz (59:21.540)
I'll argue that not only is it the right way,
Lex Fridman (59:23.620)
it's the only way that's going to exceed human performance.
George Hotz (59:27.620)
Well.
Lex Fridman (59:28.540)
It's certainly true for Go.
George Hotz (59:29.940)
Everyone who tried to hand code Go things
Lex Fridman (59:31.460)
built human inferior things.
Lex Fridman (59:33.420)
And then someone came along and wrote some 10,000 line thing
Lex Fridman (59:36.180)
that doesn't know anything about Go that beat everybody.
George Hotz (59:39.780)
It's 10,000 lines.
Lex Fridman (59:41.060)
True, in that sense, the open question then
George Hotz (59:44.500)
that maybe I can ask you is driving is much harder than Go.
Lex Fridman (59:53.460)
The open question is how much harder?
Lex Fridman (59:56.260)
So how, because I think the Elon Musk approach here
Lex Fridman (59:59.500)
with planning and perception is similar
🔗 相关节目