Podcast conversation with George: https://www.youtube.com/watch?v=_L3gNaAVjQ4 Please check out our sponsors to get a discount and to support this podcast:
- Four Sigmatic: https://foursigmatic.com/lex and use code LexPod to get up to 40% & free shipping
- Public Goods: https://publicgoods.com/lex and use code LEX

PODCAST INFO:
Podcast website: https://lexfridman.com/podcast
Apple Podcasts: https://apple.co/2lwqZIr
Spotify: https://spoti.fi/2nEwCF8
RSS: https://lexfridman.com/feed/podcast/
Full episodes playlist: https://www.youtube.com/playlist?list...
Clips playlist: https://www.youtube.com/playlist?list...

CONNECT:
- Subscribe to this YouTube channel
- Twitter: https://twitter.com/lexfridman
- LinkedIn: https://www.linkedin.com/in/lexfridman
- Facebook: https://www.facebook.com/LexFridmanPage
- Instagram: https://www.instagram.com/lexfridman
- Medium: https://medium.com/@lexfridman
- Support on Patreon: https://www.patreon.com/lexfridman

This is a review of programming memes with George Hotz.

Quick mention of 2 sponsors, 4 Sigmatic, the maker of delicious mushroom coffee and Public Goods, my go-to online store for minimalist household products and basic healthy food.

Please check out these sponsors in the description to get a discount and to support this channel.

Let me ask you to do a meme review or inspect some printed black and white memes, and maybe give a rating of pass or fail, like binary.

These are mostly from the programming subreddit.

I mean, you were like, I could do it straight up like PewDiePie meme review.

No, I think, no, I think like, now this 1, this meme's much better.

Never spend 6 minutes doing something by hand when you can spend 6 hours failing to automate it.

Like do you find yourself over-automating?

Of course, I mean, but you learn so much trying to automate.

You know, you could spend 6 minutes just driving there instead of spending, you know, 10 years of your life trying to solve self-driving cars.

Of course, you know, it's like you're trying to overcome laziness and

end up putting in way more effort to overcoming laziness than just overcoming it.

This just, yeah, this is why I trust programmers a lot more than I trust doctors.

Doctors think they have some like divine wisdom because they went to med school for a few years.

Like, you know, I think I remember once when I was, let's say drug shopping And 1 of the doctors referred to Adderall as a Schedule B drug.

And I can't correct this guy, that's drug seeking behavior to an extreme.

But you know, like you're a doctor, you don't know that?

The DEA gave you a pad to prescribe and You don't know about the drug scheduling system?

If you outsource the entirety of your knowledge to Stack Overflow?

You write stuff from scratch pretty well.

They asked me, They asked me on Instagram, they were like, what's the integral of 2 to the x?

Well, the integral of e to the x is e to the x.

You do like user, not user, unit testing?

Our regression testing's gotten a ton better in the last year.

Now we're pretty confident, if the tests pass, that at least it's going to go into someone's car and do something.

Versus in the past, like sometimes we would merge stuff in and then 1 of the processes wouldn't start.

So you think that kind of testing is still really useful for Autonomous?

Not as much unit, but like some kind of, God, bad words today, holistic, like integration testing.

Yeah, that it runs and that it runs in CI and that it doesn't output crazy values.

And actually, we wrote this thing called process replay, which just runs the processes and checks that they give the exact same outputs.

And you're like, this is a useless test, because what if something changes?

This prevents you from, if you're doing something that you claim is a refactor, this will pretty much, it will check that it gives the exact same output every single time, which is good.

And we run that over like, you know, 60 minutes of driving.

But that one's a win for you, you get past the 10.

Yep, yep, yep, yep, yep, yep, yep, that's an 8 out of

Can't give anything a 10 out of 10, that's a good meme.

Yeah, yeah, I mean, you know, this is why I don't use IDEs.

I gotta say, I mean, to me it's kind of funny because it's so annoying.

It is so annoying when I do that's why I'm still using an Emax is because it's like clippy on steroids.

I always still want to, this is a company idea I want to do.

And I've trained, I've trained language models on Python.

You can train these big language models on Python and they'll do a pretty good job reading, checking for bugs and stuff.

So you hope for that, like machine learning approaches for improved IDs.

Yeah, I think machine learning approaches are improved.

Like, look at Google's auto-suggestion in Gmail.

Oh, yeah, even the dumb thing that it does is already awesome.

There's a lot of basic applications of GPT-3 language model that might be kind of cool.

I think that the idea that you're going to somehow get GPT-3s to replace customer service agents, I mean, really, this reflects so badly on customer service agents.

And like, why don't you just give me access to whatever API you have?

No, I'd rather you mine cryptocurrency in my browser than use my speakers.

The only thing I do want to autoplay is like YouTube.

Why do news sites start auto-playing their news?

I still hear about the news because people talk about it.

It's like, I've never been to a news site.

In fact, even on Hacker News, I'll always avoid it if the link is like a New York Times or CNN.

I've read a few things, But yeah, those like the pop-up news sites.

That's at the core why journalism is broken actually.

It's because like the revenue model is broken, but actually the interface.

Like I have to, like I literally, I wanna pay the New York Times and the Wall Street Journal.

I want to pay them money, but they make it to where I have to click so many times.

You have to have a login and have a New York Times password?

Also, they're doing the gym membership thing where they try to make you forget that you signed up for them.

Like just be up front and let it be 5 bucks and then everybody would, it's your idea, which is like just make it cheap, accessible to everyone, don't try to do any tricks, just make it a great product and everybody will love it.

Actually, what I want to see also, I've seen them start to exist, is ad block that also blocks all the recommendeds.

These trending topics have 0 relevance in my life.

They're designed to outrage me and bait me.

You know, 1 interesting thing I've used on Twitter, I had to turn it back on because it was tough, is there's Chrome extensions people could try is it turns off the display of all the likes and retweets and all that, so all numbers are gone.

It's such a different, people should definitely try this.

It's such a different experience, because I've realized that I judge the quality of other people's tweets by the number of likes.

Like it's not, you know, basically if it has no likes, then I'll judge it versus 1 that has like several thousand likes.

And when you turn those numbers off, I'm lost.

At first I'm like, what am I supposed to think?

But it's an interesting experience worth thinking about.

I found that, so actually I do, I block, you can do it with AdBlock.

I have like, like trending topics, anything that's global and applicable to everybody, I definitely do not want.

I do not want to know what's trending in the United States right now.

It's always some like political agitprop or like some, it's agitprop a lot of the times.

But at the same time, like to push back slightly, for me, trending, like if there's a nuclear war that breaks out.

And I just trust that if nuclear war breaks out, someone's gonna come to me.

A friend will text me, like, you know, yo.

Yeah, I did also try turning off YouTube's recommendations, but I realized that they do provide value to me.

As opposed to a general 1, like a trending 1,

So this 1 is about how Stack Overflow can be really rough.

Asking a question on Stack Overflow be like.

No, I've never asked a question on Stack Overflow.

If someone, if there's a comment on the internet, I just want you to know that they probably do not reflect my values Well, I mean there is some truth to that too 1 of the documents that I make everyone everyone a comma has to read it It's linked from the thing is It's Eric Raymond and it's how to ask smart questions.

And it's just like, well, you know, the quality of the answer you get depends a lot on how you will phrase the question.

To be fair, I think, yeah, the auto keyword is not, is a, I think Stack Overflow likes it where it's like, there is an answer.

But some of the most interesting questions are a little bit...

I do this in interviews, I ask ambiguous questions.

And I just see how, I found that that's so much more useful for gauging candidates than like...

To see just like what they think of and like what language they use and whether they show familiarity or not.

What do you, by the way, just as a quick aside, is there any magic formula to a good candidate?

And you can see it through just conversation?

I always wish there was a test that you could just online, like a form that you can fill out.

Well, OK, so there's the old problem of...

Testing for intelligence is easy because, OK, smart people can lie and say they're dumb, but dumb people can't really lie and say they're smart.

But testing for motivation is much harder.

So we do this, we have a micro internship.

We bring people in for 2 days paid to come work on something real in our code base.

And we see, are you just doing this as a job interview?

If you're doing this as, OK, I've got to check this box, then yeah, you're not a good fit.

And then you see, and then you can be picky and see like, if there's a, is there a fire there like that?


Speaker 0: This is a review of programming memes with George Hotz. Quick mention of 2 sponsors, 4 Sigmatic, the maker of delicious mushroom coffee and Public Goods, my go-to online store for minimalist household products and basic healthy food. Please check out these sponsors in the description to get a discount and to support this channel. And now, on to the memes. Let me ask you to do a meme review or inspect some printed black and white memes, and maybe give a rating of pass or fail, like binary.

If this is a win or a lose. These are mostly from the programming subreddit. You don't like this 1. This doesn't pass.

Speaker 1: I mean, you were like, I could do it straight up like PewDiePie meme review. Yeah, this meme gets a 1 out of

Speaker 0: 10. Wow, this is too much text. I only half-assed looked at it.

Speaker 1: Already too much text.

Speaker 0: Are you gonna skip on the text?

Speaker 1: Well, what am I doing? Yeah, I don't know about this. I think this plays on bad stereotypes.

Speaker 0: Isn't that all of memes? I don't know.

Speaker 1: No, I think, no, I think like, now this 1, this meme's much better. Never spend 6 minutes doing something by hand when you can spend 6 hours failing to automate it. Now this 1 I like.

Speaker 0: Crys in PowerShell. What are you doing in PowerShell? Do you do that yourself? Like do you find yourself over-automating?

Speaker 1: Of course, I mean, but you learn so much trying to automate. You know, you could spend 6 minutes just driving there instead of spending, you know, 10 years of your life trying to solve self-driving cars.

Speaker 0: That's true. That's the programming spirit.

Speaker 1: Of course.

Speaker 0: I don't know what that is.

Speaker 1: Of course, you know, it's like you're trying to overcome laziness and

Speaker 0: you

Speaker 1: end up putting in way more effort to overcoming laziness than just overcoming it.

Speaker 0: 78,000 upvotes on that 1? Jesus, Okay.

Speaker 1: Mine doesn't have the, oh wow. Oh, that's up there, I see. This just, yeah, this is why I trust programmers a lot more than I trust doctors.

Speaker 0: You like Googling. You like...

Speaker 1: Doctors think they have some like divine wisdom because they went to med school for a few years. I'm like, dude, I understand. I didn't Google this on WebMD. I actually like I read these 3 papers. Like, you know, I think I remember once when I was, let's say drug shopping And 1 of the doctors referred to Adderall as a Schedule B drug.

And I can't correct this guy, that's drug seeking behavior to an extreme. Schedule II, it's not Schedule B. But you know, like you're a doctor, you don't know that? The DEA gave you a pad to prescribe and You don't know about the drug scheduling system? Like, like...

Speaker 0: Yeah. Yeah, programmers don't have that. They just outsource. Do you think that becomes a problem? If you outsource the entirety of your knowledge to Stack Overflow?

You write stuff from scratch pretty well.

Speaker 1: Yeah, yeah, yeah. They asked me, They asked me on Instagram, they were like, what's the integral of 2 to the x? And I'm like,

Speaker 0: You need to look it up.

Speaker 1: Well, the integral of e to the x is e to the x. So it's like something like that.

Speaker 0: Yeah, the same with Wolfram Alpha. You just start looking it up.

Speaker 1: That's pretty good. That's funny. That's the first 1.

Speaker 0: That's the first funny 1? You do testing? You do like user, not user, unit testing?

Speaker 1: We do unit testing. Our regression testing's gotten a ton better in the last year. Now we're pretty confident, if the tests pass, that at least it's going to go into someone's car and do something. Versus in the past, like sometimes we would merge stuff in and then 1 of the processes wouldn't start.

Speaker 0: Right. So you think that kind of testing is still really useful for Autonomous?

Speaker 1: Not as much unit, but like some kind of, God, bad words today, holistic, like integration testing.

Speaker 0: Sure. Thing runs.

Speaker 1: Yeah, that it runs and that it runs in CI and that it doesn't output crazy values. And actually, we wrote this thing called process replay, which just runs the processes and checks that they give the exact same outputs. And you're like, this is a useless test, because what if something changes? Then we update the process replay hash. This prevents you from, if you're doing something that you claim is a refactor, this will pretty much, it will check that it gives the exact same output every single time, which is good.

And we run that over like, you know, 60 minutes of driving.

Speaker 0: But that one's a win for you, you get past the 10.

Speaker 1: Yep, yep, yep, yep, yep, yep, yep, that's an 8 out of

Speaker 0: 10. 8 out of 10.

Speaker 1: Can't give anything a 10 out of 10, that's a good meme. Breaks the rule. Yeah, yeah, I mean, you know, this is why I don't use IDEs. That's just, this is true.

Speaker 0: It's just true, it's not

Speaker 1: even funny.

Speaker 0: I gotta say, I mean, to me it's kind of funny because it's so annoying. It is so annoying when I do that's why I'm still using an Emax is because it's like clippy on steroids.

Speaker 1: It is, of course. They're going to make good ones though. I always still want to, this is a company idea I want to do. And I've trained, I've trained language models on Python. You can train these big language models on Python and they'll do a pretty good job reading, checking for bugs and stuff.

Speaker 0: So you hope for that, like machine learning approaches for improved IDs.

Speaker 1: Yeah.

Speaker 0: That's very interesting, yeah.

Speaker 1: Yeah, I think machine learning approaches are improved. Like, look at Google's auto-suggestion in Gmail.

Speaker 0: Yeah. Oh, yeah, even the dumb thing that it does is already awesome. There's a lot of basic applications of GPT-3 language model that might be kind of cool.

Speaker 1: I think that the idea that you're going to somehow get GPT-3s to replace customer service agents, I mean, really, this reflects so badly on customer service agents. And like, why don't you just give me access to whatever API you have? You know?

Speaker 0: I don't know. This 1 is,

Speaker 1: Yeah.

Speaker 0: I just saw a picture. I actually wondered if it works at all. That's good.

Speaker 1: That's good. The last line is good.

Speaker 0: Okay, thanks, bye.

Speaker 1: No, I'd rather you mine cryptocurrency in my browser than use my speakers.

Speaker 0: I feel that way about pop-ups too. Just anything.

Speaker 1: I mean, those you can at least block. It's like, you know what? The only thing I do want to autoplay is like YouTube.

Speaker 0: Why do news sites start auto-playing their news?

Speaker 1: I don't get it. I haven't been to a news site in years. I do not read the news. I still hear about the news because people talk about it. It's like, I've never been to a news site.

Speaker 0: Yeah, who goes there?

Speaker 1: Who goes to

Speaker 0: these sites?

Speaker 1: In fact, even on Hacker News, I'll always avoid it if the link is like a New York Times or CNN. I've read a few things, But yeah, those like the pop-up news sites.

Speaker 0: Yeah, those websites are broken. And somebody needs to fix it. That's at the core why journalism is broken actually. It's because like the revenue model is broken, but actually the interface. Like I have to, like I literally, I wanna pay the New York Times and the Wall Street Journal.

I want to pay them money, but they make it to where I have to click so many times. They want to do the-

Speaker 1: You have to have a login and have a New York Times password?

Speaker 0: Yes. Also, they're doing the gym membership thing where they try to make you forget that you signed up for them. God. So that they charge you indefinitely. Like just be up front and let it be 5 bucks and then everybody would, it's your idea, which is like just make it cheap, accessible to everyone, don't try to do any tricks, just make it a great product and everybody will love it.

Speaker 1: Actually, what I want to see also, I've seen them start to exist, is ad block that also blocks all the recommendeds. I don't want trending topics on Twitter. Please get rid of it. These trending topics have 0 relevance in my life. They're designed to outrage me and bait me.

Nope, don't want it.

Speaker 0: You know, 1 interesting thing I've used on Twitter, I had to turn it back on because it was tough, is there's Chrome extensions people could try is it turns off the display of all the likes and retweets and all that, so all numbers are gone. It's such a different, people should definitely try this. It's such a different experience, because I've realized that I judge the quality of other people's tweets by the number of likes. Like it's not, you know, basically if it has no likes, then I'll judge it versus 1 that has like several thousand likes. I'll just see it differently.

And when you turn those numbers off, I'm lost. At first I'm like, what am I supposed to think? It's a beautiful thing.

Speaker 1: That's right. How much is it manipulated?

Speaker 0: Yeah, it totally is. But it's an interesting experience worth thinking about.

Speaker 1: I found that, so actually I do, I block, you can do it with AdBlock. I have like, like trending topics, anything that's global and applicable to everybody, I definitely do not want. I do not want to know what's trending in the United States right now. It's always some like political agitprop or like some, it's agitprop a lot of the times.

Speaker 0: But at the same time, like to push back slightly, for me, trending, like if there's a nuclear war that breaks out.

Speaker 1: That's how they get you. That's how they get you. And I just trust that if nuclear war breaks out, someone's gonna come to me.

Speaker 0: A friend will text you?

Speaker 1: A friend will text me, like, you know, yo.

Speaker 0: Hey, are you in a shelter

Speaker 1: or what? Right? Yeah, okay.

Speaker 0: So if it's bad enough, you'll know.

Speaker 1: Yeah, I did also try turning off YouTube's recommendations, but I realized that they do provide value to me. Like the tailor.

Speaker 0: Customized to you. Yeah, and yeah. As opposed to a general 1, like a trending 1,

Speaker 1: yeah.

Speaker 0: Now YouTube does a pretty good job. So this 1 is about how Stack Overflow can be really rough. Asking a question on Stack Overflow be like.

Speaker 1: No. No, I've never asked a question on Stack Overflow.

Speaker 0: You've never asked? It's rough.

Speaker 1: I don't comment, OK? I've never. If someone, if there's a comment on the internet, I just want you to know that they probably do not reflect my values Well, I mean there is some truth to that too 1 of the documents that I make everyone everyone a comma has to read it It's linked from the thing is It's Eric Raymond and it's how to ask smart questions. And it's just like, well, you know, the quality of the answer you get depends a lot on how you will phrase the question. To be fair, I think, yeah, the auto keyword is not, is a, I think Stack Overflow likes it where it's like, there is an answer.

There's not really an answer to that. Right. That's a great Reddit thread.

Speaker 0: But some of the most interesting questions are a little bit... Ambiguous, yeah. Ambiguous, yeah.

Speaker 1: I do this in interviews, I ask ambiguous questions. And I just see how, I found that that's so much more useful for gauging candidates than like...

Speaker 0: To see how they reason through things?

Speaker 1: To see just like what they think of and like what language they use and whether they show familiarity or not.

Speaker 0: What do you, by the way, just as a quick aside, is there any magic formula to a good candidate?

Speaker 1: Intelligence and motivation.

Speaker 0: Intelligence and motivation, yeah. Motivation, passion, yeah. And you can see it through just conversation? I always wish there was a test that you could just online, like a form that you can fill out.

Speaker 1: Well, OK, so there's the old problem of... Testing for intelligence is easy because, OK, smart people can lie and say they're dumb, but dumb people can't really lie and say they're smart. But testing for motivation is much harder. So we do this, we have a micro internship. We bring people in for 2 days paid to come work on something real in our code base.

And we see, are you just doing this as a job interview? If you're doing this as, OK, I've got to check this box, then yeah, you're not a good fit.

Speaker 0: That's awesome. That's a good idea. And then you see, and then you can be picky and see like, if there's a, is there a fire there like that?

Speaker 1: Yeah, we see, and like now we hire a lot of people you know, who at least like, if you're coming to work at Kama, you're kind of expected to have read the open pilot code. All right, like the best candidates are the ones who have like read the code and asked questions about it. Like it's open source, right? The beauty of Kama is you don't need to be hired by us to start contributing. True.

Speaker 0: Yeah. Are you obsessed about resource constraints?

Speaker 1: Some. Not size. I don't care about bytes, but CPU.

Speaker 0: CPU.

Speaker 1: Well, CPU is watts. And watts is bad.

Speaker 0: Watts is bad.

Speaker 1: Watts is, we gotta pump them somewhere. We're lowering the CPU temp. People complain about the Comma 2's overheating. We have a solution. We're lowering the CPU frequencies.

We're gonna ship that in the next 1, but don't worry, we've made the code more than efficient enough to cover for it.

Speaker 0: Awesome. That's, yeah, that's 1 way to

Speaker 1: go. Beautiful. This is good. I mean, okay, 5 out of 10. Which 1?

My code doesn't work, Let's change nothing. I do that all the time.

Speaker 0: That's the weird thing, and it often works. It's running stuff again. I don't get it. I don't get how this world, this universe is programmed, but it seems to work. Just like restarting a computer, restarting the system.

Like This audio setup, sometimes it's just not working, and then I'll just shut it off and turn it back on and it just works.

Speaker 1: I interpreted this beam totally differently. You interpreted it as in, like, this is a good idea. I interpreted it as in, I find myself doing this all the time.

Speaker 0: And this is dumb.

Speaker 1: I just run it again. I know it's gonna give me the same answer and I'm not gonna be happy with it, but like, you know. I found this, I did Advent of Code last December. It's, Advent of Code's great. It's a programming challenge every night at midnight for December.

Damn, proud of myself, proud of how I did on leaderboard. Yeah, I found myself doing that all the time. I'm like, this has to be right. I don't understand. Yup.

This

Speaker 0: is like, this is kind of, you embody this in some sense. You're able to, I mean, at least I've seen you able to accomplish quite a lot in a very short amount of time. It's interesting.

Speaker 1: Yeah, sometimes, I don't know, some of them, like People don't realize how much I cheated at Twitch Slam. To be fair, I did write that slam from scratch on there. I did not have a copy of it written, but I also spent the previous 6 months studying Slam. I had

Speaker 0: so much

Speaker 1: going into it, whereas with these other things, I don't have anything going into it. And people are like, why isn't it like Twitch Slam? Well, you know.

Speaker 0: So this 1 is just, you find this to be true? Is this funny at all?

Speaker 1: I find it to

Speaker 0: be... 3 out of 10?

Speaker 1: Yeah, 3 out of 10. There's a lot of threes in there.

Speaker 0: This is a machine learning question.

Speaker 1: Yeah, true.

Speaker 0: Yeah. I mean, this is the tension. I get this. I mean, sometimes this meme pisses me off a little bit because people that like learn machine learning for the first time and they're struggling with like object detection and they think how is this supposed to achieve general intelligence? That's because you're just doing a first intro to what software 2.0 as opposed to like when you scale things, I mean you can achieve incredible things.

That would be very surprising.

Speaker 1: I mean, it's the same thing with software 1.0. You know, they sit there and they write hello world and they're like, how is this thing supposed to fly rockets?

Speaker 0: Right, and it does. Yeah. The thing is, it does.

Speaker 1: Just, it just takes a whole lot more effort than you think. I don't know, but I'm very...

Speaker 0: You're bullish on neural networks.

Speaker 1: I'm bullish on neural networks. I think neural networks are definitely going to play a component. I'm bullish on what neural networks are, which is a beautiful, generic way. I was... 1 of our screening questions is, what's the complexity of matrix multiplication?

And I was explaining this, like, why fundamentally this is true. And then I went a little further with it. I'm like, you know, neural networks is just like matrix multiplications interspersed with nonlinearities. And there is something so... The fact that that's such an expressive thing, and it's also differentiable, it's also quasi-convex, like it's convex-y, which means like convex optimizers work on it.

Like, yeah, well this isn't going anywhere. Of course not.

Speaker 0: Yeah. The question is what are the limits of its surprising power?

Speaker 1: I think that we haven't seen the loss function yet. We haven't seen the right loss function. It's going to be something fancy that looks, I think, like compression. And I haven't seen these things really applied. Maybe there was that 1 paper where they tried like surprise to solve Montezuma's revenge.

Speaker 0: Yeah. There's a lot, a lot of fun stuff in RL that, reveals the power of these things that's yet to really be understood, I think. So that's kind of an exciting 1. Thank

See all Lex Fridman transcripts on Youtube

Programming Meme Review with George Hotz