Lex Fridman PodcastGeorge Hotz: Tiny Corp, Twitter, AI Safety, Self-Driving, GPT, AGI & God | Lex Fridman Podcast #387
EVERY SPOKEN WORD
150 min read · 30,015 words- 0:00 – 1:39
Introduction
- LFLex Fridman
What possible ideas do you have for the w- how human species ends?
- GHGeorge Hotz
Sure. So I think the most obvious way to me is wire heading. We end up amusing ourselves to death. We end up all staring at that infinite TikTok and forgetting to eat. Maybe, maybe it's even more benign than this. Maybe we all just stop reproducing. Now, to be fair, it's probably hard to get all of humanity.
- LFLex Fridman
Yeah. The interesting thing about humanity is the diversity in it.
- GHGeorge Hotz
Oh, yeah.
- LFLex Fridman
Organisms in general. There's a lot of weirdos out there.
- GHGeorge Hotz
Well-
- LFLex Fridman
Two of them are sitting here.
- GHGeorge Hotz
I mean, diversity in humanity is-
- LFLex Fridman
With due respect. (laughs)
- GHGeorge Hotz
(laughs) I wish I was more weird.
- LFLex Fridman
The following is a conversation with George Hotz, his third time on this podcast. He's the founder of Comma.ai that seeks to solve autonomous driving, and is the founder of a new company called TinyCorp that created TinyGrad, a neural network framework that is extremely simple, with the goal of making it run on any device by any human easily and efficiently. As you know, George also did a large number of fun and amazing things, from hacking the iPhone to recently joining Twitter for a bit as an "intern" in quotes, making the case for refactoring the Twitter code base. In general, he's a fascinating engineer and human being, and one of my favorite people to talk to. This is the Lex Fridman Podcast. To support it, please check out our sponsors in the description. And now, dear friends, here's George Hotz.
- 1:39 – 11:18
Time is an illusion
- LFLex Fridman
You mentioned something in a stream about the philosophical nature of time. So, uh, let's start with a wild question. Do you think time is an illusion?
- GHGeorge Hotz
You know-
- LFLex Fridman
(laughs)
- GHGeorge Hotz
... I sell phone calls, uh, to Comma for $1,000. Uh, and some guy called me, and, uh, like, you know, it's $1,000, so you can talk to me for, for half an hour. And he's like, "Uh, yeah, okay. So like, time doesn't exist, and I really wanted to share this with you." I'm like, "Well, what do you mean time doesn't exist," right? Like, I think time is a useful model whether it exists or not, right? Like, does quantum physics exist? Well, it doesn't matter. I- i- it's about whether it's a useful model to describe reality. Is time maybe compressive?
- LFLex Fridman
Do you think there is an objective reality or is everything just useful models? Like underneath it all, is there an actual thing that we're constructing models for?
- GHGeorge Hotz
I don't know.
- LFLex Fridman
I was hoping you would know.
- GHGeorge Hotz
I don't think it matters.
- LFLex Fridman
I mean, this kind of connects to the models we construct of reality with machine learning, right?
- GHGeorge Hotz
Sure.
- LFLex Fridman
Like, is it just nice to have useful approximations of the world such that we can do something with it?
- GHGeorge Hotz
So there are things that are real. Kolmogorov complexity is real.
- LFLex Fridman
Yeah.
- GHGeorge Hotz
Yeah. The compressive-
- LFLex Fridman
Math.
- GHGeorge Hotz
Math is real, yeah.
- LFLex Fridman
(laughs) Should be a T-shirt.
- GHGeorge Hotz
And I think hard things are actually hard. I don't think P equals NP.
- LFLex Fridman
Ooh, strong words.
- GHGeorge Hotz
Well, I think that's the majority. I do think factoring is in P, but...
- LFLex Fridman
I don't think you're the person that follows the majority in all walks of life, so let's go-
- GHGeorge Hotz
For that, for that one, I do.
- LFLex Fridman
Yeah. In theoretical computer science, you, you're, you're one of the sheep. (laughs)
- GHGeorge Hotz
(laughs)
- LFLex Fridman
All right. But to you, uh, time is a useful model.
- GHGeorge Hotz
Sure.
- LFLex Fridman
Hmm. What were you talking about on the stream with time? Are you made of time?
- GHGeorge Hotz
If I remembered half the things I said on stream.
- LFLex Fridman
Ah.
- GHGeorge Hotz
Someday, someone's gonna make a model of all of it, and it's gonna come back to haunt me.
- 11:18 – 13:55
Memes
- LFLex Fridman
Uh, do you think we have ideas, or ideas have us?
- GHGeorge Hotz
I think that we're gonna get super scary memes once the AIs actually are superhuman.
- LFLex Fridman
Ooh.
- GHGeorge Hotz
Like, the-
- LFLex Fridman
You think AI will generate memes?
- GHGeorge Hotz
Of course.
- LFLex Fridman
You think it'll make humans laugh?
- GHGeorge Hotz
I think it's worse than that. So, um, Infinite Jest, uh, it's introduced in the first 50 pages, is about a tape that you, uh, once you watch it once, you only ever want to watch that tape. Um, in fact, you want to watch the tape so much that someone says, "Okay, here's a hacksaw. Cut off your pinky and then I'll let you watch the tape again," and you'll do it. Uh, so we're actually gonna build that, I think. But it's not gonna be one static tape. I think the human brain is too complex to be stuck in one static tape like that. If you look at, like, ant brains, maybe they can be stuck on a static tape. But we're going to build that using generative models. We're going to build the TikTok that you actually can't look away from.
- LFLex Fridman
So TikTok is already pretty close there with the generation that's done by humans.
- GHGeorge Hotz
Yeah.
- LFLex Fridman
The algorithm is just doing their recommendation, but if it's do- if the algorithm is also able to do the generation...
- GHGeorge Hotz
Well, it's a question about how much intelligence is behind it, right?
- LFLex Fridman
Mm-hmm.
- GHGeorge Hotz
So the content is being generated by, let's say, one humanity worth of intelligence. And you can quantify a humanity, right? That's a ... you know, it's- it's exaflops, yottaflops, uh, but you can quantify it. Once that generation is being done by a hundred humanities, you're done.
- LFLex Fridman
(sighs) So it's actually scale that's the problem, but also speed. Yeah. And what if it's sort of manipulating the very limited human dopamine engine for porn? Imagine it's just TikTok but for porn.
- GHGeorge Hotz
Yeah.
- LFLex Fridman
That's like a brave new world.
- GHGeorge Hotz
I don't even know what it'll look like, right? Like again, y- you can't imagine the behaviors of something smarter than you. But a super intelligent and- and- and- and- and agent that just dominates your intelligence so much will be able to completely manipulate you.
- LFLex Fridman
Is it possible th- that it won't really manipulate, it'll just move past us? It'll just kinda, uh, exist the way water exists or the air exists?
- GHGeorge Hotz
You see? And that's the whole AI safety thing.
- LFLex Fridman
(laughs)
- GHGeorge Hotz
It's not the machine that's gonna do that. It's other humans using the machine that are gonna do that to you.
- LFLex Fridman
Yeah. 'Cause the machine is not interested in hurting humans. It's just-
- GHGeorge Hotz
The machine, the machine is a machine.
- LFLex Fridman
Yeah.
- GHGeorge Hotz
But the human gets the machine, and there's a lot of humans out there very interested in manipulating you.
- 13:55 – 26:19
Eliezer Yudkowsky
- GHGeorge Hotz
- LFLex Fridman
Well, let me bring up Eliezer Yudkowsky, who recently sat where you're sitting. (laughs) He thinks that AI will almost assuredly kill everyone. Do you agree with him or not?
- GHGeorge Hotz
Yes, but maybe for a different reason.
- LFLex Fridman
(laughs) Okay. W- (laughs) and then I'll try to, uh, get you to find hope where we could find a no to that answer. But why yes?
- GHGeorge Hotz
Okay. Why didn't nuclear weapons kill everyone?
- LFLex Fridman
That's a good question.
- GHGeorge Hotz
I think there's an answer. I think it's actually very hard to deploy nuclear weapons tactically.
- LFLex Fridman
Mm-hmm.
- GHGeorge Hotz
It's very hard to accomplish tactical objectives. Great, I can nuke their country and have a irradiated pile of rubble. I don't want that.
- LFLex Fridman
Why not?
- GHGeorge Hotz
Why don't I want an irradiated pile of rubble?
- LFLex Fridman
Yeah.
- GHGeorge Hotz
For all the reasons no one wants an irradiated pile of rubble. (laughs)
- LFLex Fridman
Oh, 'cause you can't use that land for, uh, for resources. You can't populate the land.
- GHGeorge Hotz
Yeah. W- what you want a- a- a total victory in a war is not usually the ir- irradiation and eradication of the people there. It's the subjugation and domination of the people.
- LFLex Fridman
Mm-hmm. Okay. So you can't use this strategically, tactically in a war-
- GHGeorge Hotz
Yeah.
- LFLex Fridman
... to help you, to- so to help, uh, gain a military advantage. It's all complete destruction. All right.
- GHGeorge Hotz
Yeah.
- LFLex Fridman
But there's egos involved. It's still surprising. Still surprising that nobody pressed the big red button.
- GHGeorge Hotz
It's somewhat surprising. But you see, it's the little red button that's gonna be pressed with AI that's gonna ... you know, and that's why we die. It's- it's not because the AI, if there's anything in the nature of AI. It's just the nature of humanity.
- LFLex Fridman
What's the algorithm behind the little red button? Or like, what- what- what possible ideas do you have for the w- how human species ends?
- GHGeorge Hotz
Sure. So I think the most, uh, obvious way to me is wire heading. We end up amusing ourselves to death. We end up all staring at that infinite TikTok and forgetting to eat. Maybe- maybe it's even more benign than this. Maybe we all just stop reproducing. You know, to be fair, it's probably hard- hard to get all of humanity.
- LFLex Fridman
Yeah.
- GHGeorge Hotz
Yeah. (laughs) It probably-
- LFLex Fridman
This is always go ... Like the- the interesting thing about humanity is the diversity in it.
- GHGeorge Hotz
Oh, yeah.
- LFLex Fridman
O- organisms in general. There's a lot of weirdos out there.
- GHGeorge Hotz
Well-
- LFLex Fridman
Two of them are sitting here.
- 26:19 – 32:38
Virtual reality
- GHGeorge Hotz
time.
- LFLex Fridman
You like virtual reality?
- GHGeorge Hotz
I love it.
- LFLex Fridman
Do you want to live there?
- GHGeorge Hotz
Yeah.
- LFLex Fridman
Yeah. Part of me does too. How far away are we, do you think?
- GHGeorge Hotz
Judging from what you can buy today? Far. Very far.
- LFLex Fridman
I gotta tell you that I had the experience of, uh, Meta's Codec avatar, where it's a ultra-high resolution scan. It looked real.
- GHGeorge Hotz
I mean, the headsets just are not quite at, like, eye resolution yet. I, I haven't put on any headset where I'm like, "Oh, I, this could be the real world." Whereas, when I put good headphones on, audio is there. And like, we, we can reproduce audio that I'm like, "I'm actually in a jungle right now." I, I, if I close my eyes, I can't tell I'm not.
- LFLex Fridman
Yeah. But then, there's also smell and all that kind of stuff.
- GHGeorge Hotz
Sure.
- LFLex Fridman
I don't know. I... The, the power of imagination or the power of the, the mechanism in the human mind that fills the gaps, that kind of reaches and wants to make the thing you see in the virtual world real to you, I believe in that power.
- GHGeorge Hotz
Or humans want to believe.
- LFLex Fridman
Yeah. Like, what, what if you're lonely? What if you're sad? What if you're really struggling in life and here's a world where you don't have to struggle anymore?
- GHGeorge Hotz
Humans want to believe so much that people think the large language models are conscious. That's how much humans want to believe.
- LFLex Fridman
Strong words. He's throwing left and right hooks.
- GHGeorge Hotz
(laughs)
- LFLex Fridman
Uh, why do you think large language models are not conscious? Do you think-
- GHGeorge Hotz
I don't think, I don't think I'm conscious.
- LFLex Fridman
Oh. So what is consciousness then, George Hotz?
- GHGeorge Hotz
It's like, what it seems to mean to people, it's just like a word that atheists use for souls.
- LFLex Fridman
Sure. But that doesn't mean that soul is not an interesting word.
- GHGeorge Hotz
If consciousness is a spectrum, I'm definitely way more conscious than the large language models are. I think the large language models are less conscious than a chicken.
- LFLex Fridman
When is the last time you seen a chicken?
- GHGeorge Hotz
Uh, in Miami, like, a couple months ago.
- LFLex Fridman
How... Uh, no. Like a living chicken.
- GHGeorge Hotz
There's living chickens walking around Miami. It's crazy.
- LFLex Fridman
Like, on the street?
- GHGeorge Hotz
Yeah.
- LFLex Fridman
Like a chicken?
- 32:38 – 40:03
AI friends
- GHGeorge Hotz
lot of value in it. Look, I just started my second company. My third company will be AI girlfriends.
- LFLex Fridman
(laughs)
- GHGeorge Hotz
No, like, I mean it.
- LFLex Fridman
I want to find out what your fourth company is after that-
- GHGeorge Hotz
Oh, wow. (laughs)
- LFLex Fridman
... because I think o- once you have AI girlfriends, it's, uh... Oh boy, does it get interesting. Well, maybe let's go there. I mean, the relationships with AI, that's creating human-like organisms, right? And part of being human is being conscious, is being, uh, having the capacity to suffer, having the capacity to experience this life richly in such a way that you can empathize, the AI s- can empathize with you and you can empathize with it, or you can project your, uh, anthropomorphic sense of what the other entity is experiencing. And, and an AI model would need to, um, yeah, to create that experience inside your mind. And it doesn't seem that difficult.
- GHGeorge Hotz
Y- yeah, but... Okay. So here's where it actually gets totally different, right? When you interact with another human, you can make some assumptions.
- LFLex Fridman
Yeah.
- GHGeorge Hotz
When you interact with these models, you can't. You can make some assumptions that that other human experiences suffering and pleasure in a pretty similar way to you do. The golden rule applies.
- LFLex Fridman
Mm-hmm.
- GHGeorge Hotz
With an AI model, this isn't really true, right? Th- these, these large language models are good at fooling people because they were trained on a h- uh, a whole bunch of human data and told to mimic it.
- LFLex Fridman
Yep. But if, if the AI system says, "Hi, my name is Samantha," uh, it has a backstory.
- GHGeorge Hotz
Yeah.
- LFLex Fridman
Went to college here and there.
- GHGeorge Hotz
Yeah.
- LFLex Fridman
Maybe it'll integrate that's an AI system.
- GHGeorge Hotz
I made some chatbots. I gave them backstories. It was lots of fun. I was so happy when LLaMA came out.
- LFLex Fridman
Yeah. Uh, we'll talk about LLaMA. We'll talk about all that, but like, you know, the rock with a smiley face.
- GHGeorge Hotz
Yeah.
- LFLex Fridman
Why... That's pr- it seems pretty natural for, for you to anthropomorphize that thing and then start dating it, and before you know it, you're married and have kids.
- GHGeorge Hotz
With a rock? (laughs)
- LFLex Fridman
(laughs) With a rock.
- GHGeorge Hotz
Oh.
- LFLex Fridman
There's pictures on Instagram with you and a rock and a smiley face.
- GHGeorge Hotz
To be fair, like, you know, something that people generally look for when they're looking for someone to date is intelligence in some form, and the rock doesn't really have intelligence. Only a pretty desperate person would date a rock.
- LFLex Fridman
I think we're all desperate deep down.
- GHGeorge Hotz
Oh, not rock level desperate.
- LFLex Fridman
(laughs) All right. Uh, not rock level desperate, but AI level desperate. I don't know. I think all of us have a deep loneliness. It just feels like the language models are there.
- GHGeorge Hotz
Oh, I agree. And you know what? I won't even say this so cynically. I will actually say this in a way that, like, I want AI friends. I do.
- LFLex Fridman
Yeah.
- 40:03 – 53:24
tiny corp
- LFLex Fridman
All right, let's, uh, before we go to company number three and company number four, let's go to company number two.
- GHGeorge Hotz
All right.
- LFLex Fridman
TinyCorp. Possibly one of the greatest names of all time for a company.
- GHGeorge Hotz
(laughs)
- LFLex Fridman
Uh, you've launched a new company called TinyCorp that leads the development of TinyGrad. What's the origin story of TinyCorp and TinyGrad?
- GHGeorge Hotz
I started TinyGrad as a, like a toy project just to teach myself, okay, like, what is a convolution? Uh, what are all these options you can pass to them? What is the derivative of a convolution, right? Very similar to, uh, uh, Karpathy wrote MicroGrad.
- LFLex Fridman
Mm-hmm.
- GHGeorge Hotz
Um, very similar. And then I started realizing, I started thinking about, like, AI chips.
- LFLex Fridman
Mm-hmm.
- GHGeorge Hotz
I started thinking about chips that run AI and I- I was like, "Well, okay, this is going to be a- a really big problem. If NVIDIA becomes a monopoly here, um, how long before NVIDIA's nationalized?"
- LFLex Fridman
Hmm. So you, uh... One of the reasons to sta- start TinyCorp is to challenge NVIDIA?
- GHGeorge Hotz
It's not so much to challenge NVIDIA. I- I actually, I- I like N- NVIDIA and it's to make sure power stays decentralized.
- LFLex Fridman
Yeah. And here, it's, uh, computational power.
- GHGeorge Hotz
Mm-hmm.
- LFLex Fridman
And to you, NVIDIA is kinda locking down the computational power of the world.
- GHGeorge Hotz
If NVIDIA becomes just, like, 10X better than everything else, you're giving a big advantage to somebody who can secure NVIDIA as a resource.
- LFLex Fridman
Yeah.
- GHGeorge Hotz
In fact, if Jensen watches this podcast, he may wanna consider this. He may wanna consider making sure his company's not nationalized.
- LFLex Fridman
D- do you think that's an actual threat?
- GHGeorge Hotz
Oh, yes.
- LFLex Fridman
No, but there's so much, uh, you know, there's AMD.
- GHGeorge Hotz
Mm-hmm. So we have NVIDIA and AMD, great.
- LFLex Fridman
All right. But you don't- you don't think there's like a push towards, like, selling? Like, Google selling TPUs or something like this, you don't think there's a push for that?
- GHGeorge Hotz
Have you seen it? Google loves to rent you TPUs.
- LFLex Fridman
It doesn't... You can't buy it at Best Buy?
- GHGeorge Hotz
No.
- LFLex Fridman
Hm.
- GHGeorge Hotz
So I started work on a, uh, on a chip. I was like, "Okay, what's it gonna take to make a chip?" And my first notions were all completely wrong about why, about like how you could improve on GPUs. Uh, and I will take this, this is from, uh, Jim Keller, on your podcast.
- LFLex Fridman
Mm-hmm.
- GHGeorge Hotz
Uh, and this is one of my absolute favorite descriptions of computation.Um, so there's three kinds of computation paradigms that are common in the world today. Uh, there's CPUs, and CPUs can do everything. CPUs can do add and multiply, they can do load and store, and they can do compare and branch.
- 53:24 – 56:21
NVIDIA vs AMD
- LFLex Fridman
What's the hope that AMD has? I mean, you did a build with AMD recently that I saw. Uh, how does the, uh, uh, the 7900 XTX compare to the RTX 4090 or 4080?
- GHGeorge Hotz
Oh. Well, let's start with the fact that the 7900 XTX kernel drivers don't work. And if you run demo apps in loops, it panics the kernel.
- LFLex Fridman
Okay. So this is a software issue.
- GHGeorge Hotz
Lisa Su responded to my email.
- LFLex Fridman
Oh.
- GHGeorge Hotz
I reached out. I was like, "This is, you know... Really?"
- LFLex Fridman
Yeah.
- GHGeorge Hotz
Like I understand if your seven by seven transposed Winograd conv is slower than NVIDIA's, but literally when I run demo apps in a loop-
- LFLex Fridman
Yeah.
- GHGeorge Hotz
... the kernel panics?
- LFLex Fridman
So just adding that loop?
- GHGeorge Hotz
Yeah, I just, I just literally took their demo apps and wrote like, "While true; do the app; done" in a bunch of screens.
- LFLex Fridman
Mm-hmm.
- GHGeorge Hotz
Right? This is like, like the most primitive fuzz testing.
- LFLex Fridman
Why do you think that is? They're just not seeing a market in the, in, um, machine learning?
- GHGeorge Hotz
They're changing. They're trying to change. They're trying to change. And I had a pretty positive interaction with them this week. Last week I, I went on YouTube, I was just like, "That's it. I give up on AMD." Like this is the... Their driver doesn't even like, I'm not gonna, I'm not gonna... You know, I'm, I'll, I'll go with Intel GPUs, right? Intel GPUs have better drivers.
- LFLex Fridman
(inhales) So you're kind of spearheading the diversification of, uh, GPUs?
- GHGeorge Hotz
Yeah. And I'd like to extend that diversification to everything. I'd like to diversify the, right, the more... My central thesis about the world is there's things that centralize power and they're bad, and there's things that decentralize power and they're good. Everything I can do to help decentralize power, I'd like to do.
- LFLex Fridman
So you're really worried about the centralization of NVIDIA? That's interesting. And you don't have a fundamental hope for the- the pro- proliferation of ASICs, uh, e- except in the cloud?
- GHGeorge Hotz
I'd like to help them with software. No, actually there's only, the only ASIC that is remotely successful is Google's TPU.
- LFLex Fridman
Yeah.
- GHGeorge Hotz
And the only reason that's successful is because Google wrote a machine learning framework. I- I think that you have to write a competitive machine learning framework in order to be able to build an ASIC.
- LFLex Fridman
Hmm. You think Meta with PyTorch builds a competitor?
- GHGeorge Hotz
I hope so.
- LFLex Fridman
Okay.
- GHGeorge Hotz
They have one. They have an internal one.
- LFLex Fridman
Internal. I mean, uh, public facing with a nice cloud interface and so on.
- GHGeorge Hotz
I don't want a cloud.
- LFLex Fridman
You don't like cloud?
- GHGeorge Hotz
I don't like cloud.
- 56:21 – 1:08:30
tinybox
- LFLex Fridman
AMDs. Uh, well, let me ask, sort of on a tangent to ask you, what, what, um, you've built quite a few PCs. What's your advice on how to build a good custom PC for, uh, let's say for the different applications that you use, for gaming, for, uh, machine learning?
- GHGeorge Hotz
Well, you shouldn't build one. You should buy a box from the Tiny Corp.
- LFLex Fridman
I heard rumors, whispers about this box in the Tiny Corp. What's- what's this thing look like? What is, what is it, what is it called?
- GHGeorge Hotz
Uh, it's called the Tiny Box.
- LFLex Fridman
Tiny Box? Of course it is.
- GHGeorge Hotz
Um, it's $15,000.
- LFLex Fridman
Yep.
- GHGeorge Hotz
And it's almost a petaflop of compute. It's over 100 gigabytes of GPU RAM. It's over five terabytes per second of GPU memory bandwidth. Uh, I'm gonna put, like, four nVMes in- in RAID. You're gonna get, like, 20, 30 gigabytes per second of drive read bandwidth. I'm gonna, I'm gonna build, like, the best deep learning box that I can that plugs into one wall outlet.
- LFLex Fridman
Okay. Can you go through those specs again a little bit from your m- from memory?
- GHGeorge Hotz
Yeah. So it's almost a petaflop of compute.
- LFLex Fridman
So AMD, Intel?
- GHGeorge Hotz
Today, I'm leaning toward AMD.
- LFLex Fridman
Mm-hmm.
- GHGeorge Hotz
Um, but we're pretty agnostic to the type of compute. The- the- the main limiting spec is a 120 volt, 15 amp circuit.
- LFLex Fridman
(laughs) Okay.
- GHGeorge Hotz
Well, I mean it, because in order to, like, like there's a plug over there.
- LFLex Fridman
Mm-hmm.
- GHGeorge Hotz
All right? You have to be able to plug it in. Um, we're also gonna sell the tiny rack, which, like, what's the most power you can get into your house without arousing suspicion? Uh, and one of the, one of the answers is an electric car charger.
- LFLex Fridman
Wait, where does that rack go?
- GHGeorge Hotz
Your garage.
- LFLex Fridman
Interesting. The car charger.
- GHGeorge Hotz
A wall outlet is about 1,500 watts. A car charger is about 10,000 watts.
- LFLex Fridman
Is it? (laughs) What is the most amount of power you can get your hands on without arousing suspicion?
- GHGeorge Hotz
That's right.
- LFLex Fridman
George Hotz. Okay. Uh (laughs) so the- the tiny box, and you said nVMes and RAID.
- GHGeorge Hotz
Yeah.
- LFLex Fridman
Uh, I forget what you said about memory, all that kind of stuff. Okay.
- GHGeorge Hotz
So-
- LFLex Fridman
Uh, what about what GPUs?
- GHGeorge Hotz
Again-
- 1:08:30 – 1:23:09
Self-driving
- GHGeorge Hotz
- LFLex Fridman
Since you mentioned OpenPilot, I'd love to get an update in the, uh, company number one, Comma.ai world. How are things going there in the development of, uh, semi-autonomous driving?
- GHGeorge Hotz
You know, almost no one talks about FSD anymore, and even less people talk about OpenPilot.
- LFLex Fridman
Mm-hmm.
- GHGeorge Hotz
We've solved the problem. Like, we solved it years ago.
- LFLex Fridman
What's the problem exactly?
- GHGeorge Hotz
Well, how do you-
- LFLex Fridman
Or what, what, what does solving it mean?
- GHGeorge Hotz
Solving means how do you build a model that, uh, outputs a human policy for driving?
- LFLex Fridman
Mm-hmm.
- GHGeorge Hotz
How do you build a model that given our, you know, reasonable set of sensors outputs a human policy for driving? Uh, so you have, you know, companies like Waymo and Cruise which are hand-coding these things that are like quasi-human policies.
- LFLex Fridman
Hm.
- GHGeorge Hotz
Then you have Tesla and maybe even to more of an extent, Comma, asking, "Okay, how do we just learn the human policy from data?" The big thing that we're doing now, and we just put it out on Twitter. At the beginning of Comma, we published a paper called Learning a Driving Simulator.
- LFLex Fridman
Mm-hmm.
- GHGeorge Hotz
And the way this thing worked was it's a, it was an autoencoder and then an RNN in the middle, right? Uh, you take an autoencoder, you compress the picture, you use an RNN, predict the next state, and these things work. You know, it was a laughably bad simulator, right? This is 2015 era machine learning technology. Today, we have VQ-VAE and transformers.
- LFLex Fridman
Mm-hmm.
- GHGeorge Hotz
We're building DriveGPT, basically.
- LFLex Fridman
DriveGPT. Okay. Uh, so and it's trained on what? Is it trained in a self-supervised way?
- GHGeorge Hotz
Yeah. It's trained on all the driving data to predict the next frame.
- LFLex Fridman
So really trying to, uh, learn a human policy.
- GHGeorge Hotz
Yeah.
- LFLex Fridman
What would a human do?
- GHGeorge Hotz
Well, actually our simulator is conditioned on the pose. So it's, it's actually a simulator. You can put in like a state action pair and get out the next state.
- LFLex Fridman
Okay.
- GHGeorge Hotz
Um, and then once you have a simulator, you can do RL in the simulator, and RL will get us that human policy.
- LFLex Fridman
So it transfers?
- GHGeorge Hotz
Yeah. RL with a reward function, not asking is this close to the human policy? But asking would a human disengage if you did this behavior?
- LFLex Fridman
Okay. Let me think about the, the distinction there. Would a human disengage?
- GHGeorge Hotz
Yeah.
- LFLex Fridman
Would a human disengage? That, um, correlates, I guess, with human policy, but it could be different. So it's, it, uh, it doesn't just say, "What would a human do?" It says, "What would a good human driver do-"
Episode duration: 3:08:45
Install uListen for AI-powered chat & search across the full episode — Get Full Transcript
Transcript of episode dNrTrx42DGQ
Get more out of YouTube videos.
High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.
Add to Chrome