0:02

Programing Throwdown,

0:05

Episode 171,

0:09

Compilers and

0:11

Interpreters. Take

0:22

it away, Jason. Hey everybody!

0:27

This is super exciting. We

0:29

got this request from a

0:31

listener. Maybe later on

0:33

the show I'll find out, remind myself

0:35

who it was. But they said, how

0:37

about you folks do compilers and interpreters?

0:40

I can't believe we haven't done

0:42

this episode. It seems so pivotal

0:44

to everything we do in

0:48

software engineering. But

0:51

yeah, here we are to lead

0:53

the way. I have to

0:56

confess, I don't know as much about

0:58

compilers and interpreters as Patrick. Patrick's really

1:00

going to carry the water here, but

1:03

I will add color commentary as appropriate.

1:08

Before we get into that, I wanted to talk about

1:11

monitor setups for a little bit. I

1:15

was working next to somebody who

1:17

had three monitors, side by

1:20

side by side. They

1:22

had three different things going on. They

1:24

had one was for coding, the other

1:26

one was for email. I

1:29

thought this was really compelling. I thought, oh,

1:31

this is a great idea. I

1:34

just can't get myself to do it. Basically

1:36

I just have to have just one thing in

1:39

front of me and just

1:42

keep all tabbing to go to different

1:44

things. I haven't

1:46

figured out how to drag from one

1:48

desk screen to the other and do all of

1:51

that. What is

1:53

your setup? Do you have

1:55

an array of monitors? What

1:57

do you do? Yeah, eight.

2:00

No, I don't. Seriously? I

2:03

do all my programming in VR, so I

2:05

just have as many as I want. I actually do look

2:07

for it. I think that'll be cool one day, one day. That

2:10

day is not today. I just

2:12

have one big one, but I also have iPad

2:15

off to the side that I use

2:17

for my conferencing, for video conferencing. That

2:20

works well for me, so I guess it's

2:22

technically two, but I'm on big one for

2:24

sort of similar reasons to what you're saying.

2:27

Normally if you have two monitors, you just get them

2:29

extra wide or on top of each other, which is

2:31

too much, versus one big monitor, I feel. You don't

2:34

end up with an awkward split in the middle or

2:36

have to shift one far to the side. Right.

2:39

I will say for people who do simulation race

2:41

games or flight games, having three for a gaming

2:43

setup would be really cool, because you're

2:46

one thing over three, and the people who

2:48

do, of course, the stock trading, where

2:50

they're actually just mostly focused on

2:53

one task. All

2:55

of it is related. There's just

2:57

lots of dials and metrics. It's akin to,

2:59

I think, a cockpit and an airplane, where

3:02

there's many, many indicators. Some are lower

3:04

priority, but they're kind of very related

3:07

to a singular task. To

3:09

me, it seems a little bit different. When

3:11

you're doing your programming, email is...

3:15

Maybe that's not fair. I don't know. I've never

3:17

been those other things, but it feels like a

3:19

different task. Slack or email, they feel different to

3:21

the task of coding. I'm

3:24

with you. I don't feel compelled to have them up

3:26

on the side. Yeah. I

3:30

used to have my laptop. My

3:33

laptop right now is just mirroring

3:35

the screen of

3:37

the main display. I

3:39

used to have it, kind of like what you

3:41

were saying, where it was sort of secondary screen, but

3:45

I found myself not paying attention in

3:47

meetings, even when I should be. I

3:51

kind of couldn't do that. Isn't

3:54

this like a trope that people can't

3:56

actually multitask? sequentially

4:00

singular task. Yeah. Yeah,

4:03

I mean I just But

4:06

yeah, I mean I do envy the people who can have

4:09

like three different screens on and not be distracted

4:11

I will say

4:13

I do find I it's

4:16

not all bad I think if you

4:18

occasionally need like a paper or document or

4:20

you're implementing an algorithm and you want to

4:23

consult something or The

4:25

few times I've had the kind of I'm

4:27

generating some kind of output or whatever. That's visual

4:31

It's very cool to have a extra

4:33

real estate where when I update something and click,

4:35

you know Test I can sort

4:37

of still see the code but

4:39

see the results and you know up inside

4:41

it You know screen real estate can be

4:43

very very useful So I'll never sort of

4:45

say something bad against someone who who wants

4:47

to have more screens But the variety is

4:49

yeah, it is very interesting Yeah,

4:52

I wonder You

4:55

know now that I'm working from home, it's less

4:57

of an issue But when I was going into

4:59

the office, I was running into problems where you

5:01

my home Monitor

5:04

and my office monitor aren't the same and

5:07

so, you know It wouldn't quite extend the desktop

5:09

the right way and things would get all messed

5:11

up That's the

5:14

thing I've been picky about is I have

5:16

my monitor pretty high. I'm I'm fairly tall.

5:18

So I Want you

5:20

know my monitor nice and high so I either put

5:22

little plastic risers or put it on a monitor arm

5:24

and lift it up So I have it higher than

5:26

I think what I see a lot

5:28

of other people do just because it feels unnatural to me

5:30

All right. We have to have a we have to have

5:32

a bake-off here. My monitor is on

5:35

a I'm trying to

5:38

eyeball it maybe like an eight inch riser

5:41

What about you? Oh, I'm sorry. Look on

5:43

the bottom of my screen. I'm like at a like

5:45

a foot Like oh wow.

5:47

Okay, that's even bigger. I might be Yeah,

5:50

a little bit more than eight inches Yeah, that

5:52

is pretty tall so like the when you look

5:54

at your monitor at eye level are you looking

5:56

at the center of it? I

5:59

wish it's not I'm not quite that high, the

6:01

monitor, I'm gonna go quite that high. So it's, yeah,

6:03

I'm looking at like just above, somewhere like 60% up

6:05

from the bottom. Yeah, okay, makes

6:07

sense. Yeah, I'm about maybe like 75 or

6:10

so up from the bottom. But yeah,

6:12

if I didn't have any riser, the monitor, like

6:15

my eye level would be above the top of

6:17

the monitor. It's just on you. But I see

6:19

people do that all the time. I

6:21

can't. Yeah, it's

6:23

like, I think your neck would get

6:25

hurt. I think if you do, if you keep it

6:27

like that. What I want to

6:29

try is one of the very wide

6:31

curved monitors. I

6:34

think that would be really cool. Yeah,

6:36

you know, I've tried it. I have

6:38

one at work

6:40

and it's pretty good. It's pretty good.

6:43

I will say that the

6:46

one I have, the refresh

6:48

rate isn't as good. It's

6:52

not like a deal breaker

6:54

or anything, but you do like, if you

6:56

have one monitor, has twice the refresh rate

6:58

of the other, it can kind of like,

7:00

it's like when you're at the store and

7:02

all the differences between the TVs are so

7:04

noticeable then you take it home, totally

7:07

lose your frame of reference, right? So

7:09

if you have like a four to

7:11

three aspect ratio or whatever, like not

7:13

1080p widescreen 16 to nine, but

7:15

like if it's curved left and right, does

7:17

it also curve up and down if you

7:19

get a nice tall one, like, or is

7:22

like on a sphere? I've never looked it up. I guess I should. I'm

7:25

looking at it right now. It's

7:29

curved, but just, you know, it's just curved

7:31

like a cylinder. It's curved like a cylinder,

7:33

okay. Yeah, but it's very

7:35

wide and not very tall. So the

7:38

aspect ratio is not, it's not HD

7:40

aspect ratio. It's much wider. Oh,

7:42

okay. I think I would want one that, yeah.

7:45

I think I just want like a very large, like probably 8K,

7:48

but like, you know, big, but then I wouldn't actually

7:50

run it. You know, I would, everything would be scaled

7:52

up just so it's a comfortable eyesight thing. I

7:55

think 8K would be too small at like minimum

7:57

resolution or whatever. Yeah, right.

8:00

Right. But yeah, folks out

8:02

there, let us know your monitor setup. I'm curious. I

8:04

do feel like I could do better. And

8:08

yeah, curious to know what people's setups

8:10

are out there. Just chat us, email

8:12

us, post in Discord. Discord

8:15

is getting more popular, so people posting some

8:17

funny jokes. There is a video of this

8:19

professor pranking his class, which

8:21

I thought was hilarious. Wait, what? A

8:23

professor rained to their class? Okay, I had to go

8:25

on Discord. But pranking their class.

8:28

Oh, pranking. I said ranking. I

8:31

was like, oh. Yes,

8:34

we all know that they do that,

8:37

but I was supposed to actually do

8:39

it. ratemystudents.com, your website. All

8:42

right, well, we'll jump over to the news of

8:44

the show. Starting off since last

8:46

time, I mentioned this and then Jason

8:48

was excited about it as well. So I figure we'll

8:51

talk about it here. The

8:54

headlines, everyone's probably seen them, but I feel like

8:56

it's pretty cool for something that is a

8:59

memorable part of my childhood, which

9:01

is being absolutely crushed and being

9:03

terrible at the game of Tetris.

9:07

But now there has been someone

9:09

who has beaten Tetris. Actually,

9:11

as I'm recording this show, I think two people

9:13

have done it. So

9:15

the world record didn't last very long after lasting for

9:18

years and years. But Blue

9:20

Scooty, I think first name is Willis, beat

9:23

Tetris first, not at the earliest possible

9:25

time, which was accomplished now, but

9:27

did beat it. And here actually, the reason why I

9:29

think it's worth mentioning on the show

9:31

aside from just being a cool thing, and I linked

9:34

the video there that is the first place I saw

9:36

it, but it's interesting

9:38

how they define beat here. So when

9:40

Tetris was originally built, they didn't actually

9:43

have an end game in mind, which

9:45

is kind of unusual. It would just play

9:47

faster and faster. And I guess the assumption

9:49

would just, which is so fast, you couldn't

9:51

possibly keep going. And so eventually it would

9:53

die. A lot of games

9:55

actually that era are like that. Like Donkey Kong

9:58

actually, In Donkey Kong,

10:01

the amount of time you have to

10:03

finish the level goes down every time

10:05

you have completed a full

10:07

revolution of all the levels. And

10:10

basically, you get to a point

10:12

where there just literally isn't enough

10:14

time to finish the first

10:16

level, even if you're frame

10:18

perfect, and that's just how it ends. So

10:22

Tetris, I guess, gets faster

10:24

and faster. And then at some point, it

10:26

basically kind of caps out. It just doesn't

10:28

go any faster. And for a long time,

10:31

people would get to that area, that

10:34

speed, and then just sort of die pretty quickly.

10:36

If their row count was low, they

10:38

would be able to manage

10:40

to maybe get a few more lines cleared and

10:42

then die. But then recently, people discovered a new

10:44

way of pounding, tapping

10:48

the controller faster and

10:50

faster with this crazy

10:52

two handed technique. And they've

10:56

been able to go faster and faster and began to discover,

10:58

which I guess people had figured out via tool

11:00

assisted speed running. So they had kind of known

11:02

this would happen. But humans began to be able

11:04

to get there, which is levels where incrementing

11:07

the color palette left the predefined

11:09

color palette and would start using

11:12

actually machine instructions as the

11:14

color selection, which would result in just really

11:16

ill formed colors that made it even more

11:18

difficult. But then eventually, a

11:21

portion of memory would be accessed

11:23

that would just cause basically, you

11:25

know, a corruption of

11:27

the data. And so when that happens, the game,

11:29

the game just crashes and stops and just hangs.

11:32

And so this is what Blue Scooty managed

11:35

to do was to basically cause the game

11:37

to try to execute code that wasn't code,

11:39

and then hung the game. And so first

11:41

person to do that is just it's crazy.

11:43

It's very fascinating. It's also, and

11:46

it's the same, you know, as Jason was

11:48

mentioning, sort of frame perfect analysis for Super

11:51

Mario Brothers is some any sort of

11:53

speed running game that is from around

11:55

that era, the amount of analysis

11:57

and reverse engineering that goes into it and there

11:59

are even patch versions of the game so

12:01

that it doesn't get weird and you can keep

12:03

playing or just other things

12:05

where people really, really understand the

12:08

mechanisms that take place at

12:10

this end game and the amount of analysis. I'm sure

12:12

the programmers at the time had no idea that for

12:14

30 years, 40 years, people would

12:18

be thinking about what

12:20

that code that they wrote in the back of a

12:22

cubicle subware and turned out.

12:25

Yeah, it's totally

12:27

wild, right? I

12:30

think that

12:32

whole thing where they do this crazy, I mean,

12:34

you really have to see it almost in slow

12:37

motion this way with two hands. The

12:40

idea is every frame, in Tetris,

12:43

there's no inertia

12:45

or momentum or anything like that

12:47

with the pieces. So, every frame,

12:49

if you could get a press

12:51

left and a stop pressing left

12:54

in those two frames, then you could move it

12:57

over one whole

12:59

column. So yeah,

13:02

they do this weird thing where I guess they

13:04

hit the joystick from the bottom. So

13:08

in other words, if you imagine just pressing

13:10

with your thumb on a button, there's

13:13

only so fast your thumb can press

13:15

and release this button. But then

13:17

I guess they're saying, well, part of that

13:19

is just the mechanics of your body. So

13:22

what if you had another thumb on the

13:25

bottom of the controller? And

13:27

so one thumb is pressing and releasing, but

13:29

then the other thumb is pressing

13:31

the bottom into the first thumb.

13:34

And I guess by doing that, you can

13:36

get more presses than you could with just

13:39

one thumb. It's totally

13:41

crazy. You have to see it. I

13:43

mean, I guess Tetris lends

13:45

itself to that because I don't know.

13:49

We'd have to learn more. So I guess you only have

13:51

to move either left or right, as long as you don't

13:53

overshoot. And you only

13:55

have to rotate the piece three times in

13:57

a direction, and then you would just be

14:00

back. So you can only rotate 90 degrees. So

14:02

there's like a finite number of presses.

14:05

If you knew exactly what you were doing, your

14:07

path planning, you could be very

14:09

minimal inputs. Yep, yep,

14:12

totally. All right, on

14:14

to PowerWorld, accused of being an

14:16

AI product. So

14:19

yeah, so PowerWorld is wild. Patrick and I

14:21

were talking about it before the show. It

14:24

really is a great area,

14:27

a great theme for

14:29

a game. Basically, it's adult-themed

14:32

Pokemon with guns.

14:40

I think they're also just really aggressive looking.

14:44

You could actually catch

14:47

people in the Pokeballs and

14:49

sell them. So

14:53

it's basically a crazy, much

14:56

more like Mad Max. It's Mad

14:59

Max meets Pokemon. It's

15:02

a great idea of a theme. This

15:04

article is really interesting. Patrick actually found

15:06

this. It's

15:09

claiming that a lot of

15:11

the assets are AI-generated, but

15:14

they have no hard evidence.

15:18

There's really like two things that I

15:21

took away from this. One

15:23

is that the generative AI

15:26

is really here to stay. It's not

15:28

going anywhere. The days of downloading 30

15:30

gigabytes of MP3s when you

15:35

buy a game, I feel like those

15:37

days are very short-lived. I think

15:39

even if you did hire a

15:43

voice actor, you almost

15:45

certainly want to do some

15:48

type of capture of their

15:50

voice, and if nothing else, just

15:52

to reduce the storage costs and

15:55

generate their voice dynamically. So

15:59

that's definitely And then the other

16:01

thing is, it's amazing that they don't

16:03

know because I've seen a lot of

16:06

papers and other reports

16:09

where you can see how

16:11

something is generated by an

16:13

AI. But

16:16

I guess they have a point that it's really hard

16:18

to prove it. I

16:21

think the examples

16:23

I saw were interesting. For

16:26

example, if you see a... and this isn't a pal world

16:28

thing, but if you see a picture of a face and

16:33

the earrings don't match. If

16:35

it's a woman's or man's face, anyone's face,

16:37

and the earrings, they're wearing earrings and they

16:39

don't match, that's a sign that it's generative

16:42

AI because it's hard for the

16:44

AI to get

16:47

that symmetry right. But

16:51

they haven't really found anything like this here.

16:53

It might be just that

16:56

the person is starting with

16:58

an AI model

17:00

and then just as inspiration, but then

17:02

putting a human touch on it makes

17:04

it very hard to track. So

17:08

okay, a couple things. If you're going

17:10

to commit a crime, put on mismatched earrings

17:12

and do all the tips that are generative.

17:14

So if someone captures you on the camera,

17:17

you'd be like, clearly not me. Clearly

17:20

an AI. So

17:23

I sense a trend where people do

17:25

this. Second, a hat tip here to

17:27

Ace Rola. Ace Rola does

17:29

a YouTube video on shaders actually. Anyways,

17:32

and on Twitter, oh, X is where I

17:34

sort of first saw this controversy rising up.

17:38

But if you've never checked out him, check

17:40

it out, Ace Rola. And

17:42

then yeah, I think Jason's right. First of

17:44

all, does it matter? Like

17:47

if it's generative AI or not? And the

17:49

answer is maybe. It

17:53

sort of gets into one of those gray areas.

17:55

If it's fun, it's fun. But at the same

17:58

time, there are people trying to put work in there. into the

18:00

craft, right? It goes back to that debate,

18:02

which, you know, I don't think there is

18:04

any singular answer to, but, you know,

18:06

the difference between IP of furniture and, you know,

18:08

I heard, and I don't take medicine, real wood,

18:11

you know, piece of furniture. They're

18:13

both furniture, but, you know,

18:16

you could kind of talk at

18:18

length about, you know, various pros

18:20

or cons of each, and so,

18:24

I think it's, and, you know, as, I

18:27

don't know that even they have to be forthright

18:29

or honest about what they did, they just, you

18:31

know, probably shouldn't go

18:33

around claiming that it's high art, or, you know,

18:35

that, you know, they somehow, that there's, you know,

18:38

tons of people behind it, right? If you tried

18:40

to say, hey, you should support this game because

18:42

this took a lot of work, and

18:44

it didn't, then, you know, there's like a

18:46

sort of ethics problem, but as far

18:48

as, you know, as long as if I think right now,

18:51

they're not saying anything, they're just not

18:53

talking, it's, you know, been a kind

18:55

of overnight sensation, and I personally, yeah, it's

18:57

kind of interesting, but I don't know that

18:59

I have a specific problem with it. Yeah,

19:04

I think, as

19:08

a coder, it's hard to really relate, right? Like,

19:11

all of our code that we've done

19:13

on open source, you know, like, in

19:15

my case, like, Eternal Terminal and MAME

19:17

Hub and all these things, they're

19:20

all public repos, and so,

19:22

you know, AI is scanning those and

19:25

using them to build GitHub co-pilot and all these

19:27

things, right? So in a way, they're kind of

19:29

like copying it from me,

19:31

but it's okay, because I copied it

19:33

from Stack Overflow. Anyway. Ha ha

19:35

ha ha. So,

19:40

but artists don't feel

19:43

the same way. But

19:45

that's not, but that's not. But if I'm

19:47

not, I'm not an artist, so I can't

19:49

speak authoritatively, but it's very common when people

19:52

are learning to take inspiration to practice copying

19:54

and to understand what it takes, you

19:56

know, to make the paintings the paintings, and

19:59

it's not good to copy. copy too much, but

20:01

to be derivative, or you can even

20:03

see the lineage of how artists build

20:05

on other artists' concepts and extend them

20:08

and push them forward. There's

20:11

this very fine line between ripping

20:13

off Pokemon characters,

20:15

although if you look at other

20:17

places like Digimon, there's a lot

20:19

of Digimon monsters that look very

20:21

similar, and the same kind

20:24

of brouhaha didn't get it real. I don't know,

20:26

music is the same way. Everyone's

20:33

taking riffs and ideas from each other,

20:35

and there's been court cases recently where

20:38

it's, I don't know the answer.

20:40

It sure sounds really similar, but then again,

20:42

it's a nice sounding thing. If

20:46

you just sit there and play music until you find

20:48

something nice, you may play something that someone else has

20:50

already figured out before. This

20:55

is maybe, I'm trying to

20:57

think if there's other examples before

20:59

I say this, but I don't

21:01

really think so. I mean, let's

21:03

say this is all AI generated.

21:05

Somebody basically threw all the Pokemon

21:07

characters of all time into some

21:09

AI model, and now

21:11

they're able to spit out more Pokemon

21:13

characters. This

21:17

is really the first time where

21:19

we've seen some huge

21:22

commercial success come out

21:24

of some generative AI thing. People

21:27

are so worried about fake things,

21:30

like you're trying to impersonate

21:32

a politician or something like

21:34

that, but

21:37

this is not really that. I mean,

21:40

in terms of the incentive

21:42

is not to impersonate Pokemon.

21:45

Incentive is actually the opposite. It's to generate

21:48

something that has plausible

21:50

deniability that it came from

21:52

Pokemon. they,

22:00

you know, crash can.

22:02

So true. Right, like, and you know, but

22:05

to Jason's point, those don't experience the sort

22:07

of commercial success. The numbers I've seen are

22:09

absolutely crazy for how many people have bought

22:11

this. I haven't bought it yet. I'm

22:13

empty, because it actually does look hilarious. But,

22:17

yeah. My neighbors are actually going to play

22:19

it, so I'm feeling kind of

22:21

compelled to buy it. All

22:25

right. Well, hard pivoting off

22:27

of this one, I found a blog article

22:30

here, which is something that's been a

22:32

sort of, I don't even know how

22:34

to call it, not a meme, and like a picture

22:36

with an image macro or whatever pasted over top of

22:38

it. But just a sort of

22:40

meme, like, going around, which is kind

22:43

of absolutely silly ways to

22:45

solve determining whether a number

22:47

is even or odd. And

22:50

so I linked the blog here. The title

22:52

of the blog, it says the blog's name

22:54

is Blabin. But this is the only post

22:56

I saw on this blog. So I think

22:58

a new blog. But anyway, the link will

23:00

be in the show notes. But

23:02

this person was saying, based on TikTok, someone

23:04

get kind of ripped, which, again, I'm not

23:06

on TikTok, so this

23:08

is third hand at this point. But

23:11

getting ripped for saying that they determined

23:13

if a number is even or odd

23:15

by basically a bunch of if else

23:17

statements. So if number is one, odd.

23:20

If number is two, that's even. If number is

23:22

three, that's odd. I'm going to mess it up

23:24

already. Most people would

23:26

say, oh, would you just use a

23:28

modulus operator if you know what a

23:30

modulus operator is? Or oh, you use

23:32

bitwise and with the first

23:35

least significant bit just being one. And then

23:37

that'll tell you. There's a number of ways

23:39

to kind of do this check.

23:42

And most people would be expected as a

23:44

test of your language knowledge to know what

23:46

the modulus operator is at minimum or bit

23:48

masking, depending on the context. But

23:51

this person in the blog decided to take it

23:53

on themselves to for a 32-bit number,

23:55

which is 4

23:57

billion different numbers, write if statement.

24:00

And the interesting part and the reason I bring it up

24:02

is first just the absurdity of it. They

24:04

decided to write it in C with what

24:06

I sort of sounds like some tongue-in-cheek justification

24:10

initially. But what they

24:12

did, which is an actual interesting

24:14

thing to learn to do, and

24:16

on topic a bit for us

24:19

today, is to do code generation.

24:21

And so they wrote a simple

24:23

program, far simpler than a four

24:25

billion line program, to generate all

24:27

of the if statements automatically. And

24:30

that way they could produce that output.

24:32

And what does that take? Which is a useful

24:34

skill to be able to know how to do because there are

24:36

times when doing, not that, but

24:39

other code generation. Writing code to make

24:41

code is very useful. That didn't work.

24:43

The compiler, C compiler didn't like that

24:46

many lines. So then they decided to

24:48

output raw assembly and actually did

24:50

get it to work. And I guess that it was very

24:52

performant. If you think about it, it's

24:55

almost just a big lookup table and

24:57

you can just, there are jump tables, so you

25:00

can just basically jump to a set offset. So

25:02

if you know each if else

25:04

statement is two bytes and you had

25:06

a very clever assembler or whatever, you

25:08

could just jump based on the

25:11

number that many memory

25:13

offsets times two or whatever. So I

25:15

can imagine it being very, very efficient.

25:18

And then just having a table where it either outputs zero

25:20

or one. But yeah, so they actually

25:22

did manage to make assembly that they say works. Of

25:24

course, it's a, I guess we can

25:26

figure out four gigabyte file roughly. So

25:28

yeah, it's like, it's executable or

25:31

gigabyte. Ridiculous.

25:35

In that like there, this is of course, I don't

25:37

want to say this is a very silly way to

25:40

solve whether a number is even or odd, but

25:43

actually leads to some kind

25:46

of interesting computer science exercises

25:48

and doing this. And then of course, people responded, can

25:50

you do it for 64 bit numbers? And

25:52

the answer is no. Yeah.

25:57

I love how this, this is the

25:59

only post. in this person's blog. That's

26:02

okay, are you gonna post something up when you have a

26:04

cool idea and you wanna show people? Like, I don't fault

26:06

them. No, yeah, no,

26:08

I mean, I didn't mean that sarcastic

26:11

me. I literally love that this person,

26:13

you know, it's not like 10 posts

26:15

about, like, hire

26:18

me or self-aggrandizing or any of that.

26:20

It's just literally like, here's this cool

26:22

thing that I did, boom, blog. I

26:24

will say though, that's like a tough act to follow. So I don't know

26:26

what you make post number two. That's true,

26:28

yeah. I mean, you really can't. Just

26:31

started noodle. It's been a month now and

26:33

he hasn't made a post. That

26:38

is awesome. All right,

26:40

my next news is Seamless M4T.

26:43

Have you seen this, Patrick? No, uh-uh.

26:46

This thing is freaking wild. So this

26:48

is what I use to translate

26:50

the episode. So if you haven't done

26:52

this yet, go

26:55

to programmingthrowdown.com. Depending

26:57

on a bunch of factors, the very latest

27:00

episode might not have this. Just go back

27:02

an episode and you can listen

27:04

to us in Mandarin

27:06

Chinese. So

27:10

it's totally freaking wild.

27:12

So basically the way this works

27:14

is it

27:17

does the speech to

27:20

text. Now, like, you have to have a

27:22

separate file for each person. They have to

27:24

have a way of separating, or

27:27

in our case, we have a separate

27:29

MP3 for each

27:31

person on the show. But

27:33

it converts that to text and

27:36

then it translates the text to

27:38

other languages and then

27:41

it goes from that language's

27:43

text to one of several

27:45

voices that they have

27:47

for each language. And so

27:49

the reason why

27:52

they're calling it Seamless is

27:55

I don't think that it's end-to-end learned,

27:57

but it might be. But

28:00

it is separate sort of modules,

28:04

but the API and all

28:06

of that is extremely simple. You

28:08

basically just say, here's audio

28:10

and I want to get back text, or

28:13

I want to get back text and audio

28:15

in another language. I

28:18

tried at first translating English to English,

28:21

so it went from us to robot us.

28:24

That was really entertaining, and

28:27

then I started picking other languages.

28:30

This is amazing. I mean, it's amazing that it's

28:32

free. It's totally open source,

28:35

anyone can use it. I

28:37

actually bought for Black Friday a few

28:39

months ago, a GPU at

28:41

16 gigs of VRAM. That's

28:45

the bare minimum you need to run this model. But

28:49

if you don't have that, it's totally fine. I

28:53

also got this run on Google Cloud

28:55

Platform, and you pay about 50 cents

28:58

an hour, I think. Oh, wow. One

29:01

of those. Yeah, very reasonable.

29:04

You can even now they have something where

29:06

if it's idle for so many

29:08

hours, it shuts itself off, so you don't have to

29:10

worry about forgetting about it and getting a $500 bill

29:14

or something at the end of the day. Yeah,

29:18

all of that is super

29:20

convenient. You can

29:22

run it on Google Cloud. There's also Google CoLab.

29:26

You can run it there. There's some 16

29:29

gig VRAM GPUs there. I don't know if

29:31

the free plan would work, but the

29:34

paid plan is only 10 bucks a month. So

29:37

yeah, check this out. I mean, it's amazing. I

29:39

will say that there's two

29:42

modes. There's a

29:44

batch mode and a streaming mode. The

29:46

batch mode loads everything into

29:49

memory. So you can really only

29:51

use it for like a 10 second

29:53

file or something like that. The

29:56

streaming mode is really where it's at. The

29:59

document. The documentation is still pretty

30:03

nascent. What I

30:05

found really useful was they gave

30:07

a tutorial at NURIPS, which is

30:10

a AI conference. They

30:13

did a whole tutorial on

30:15

seamless M4T and you

30:18

can watch the video and follow along as

30:20

if you were at the conference for free.

30:22

I was able to find that. So

30:25

check it out, check out this tutorial. I

30:28

feel like this is really on

30:30

the vanguard of something super exciting.

30:34

I'm looking into, can we clone our

30:36

voices? Because if we could actually, if

30:39

it could be Mandarin Chinese, but it

30:41

sounds like us, that would be even

30:43

better. It's

30:45

like a football, your earrings match. Ha

30:49

ha ha ha. It's

30:51

gonna be like the deep fake Jason, but

30:53

yeah, speaking Mandarin. Yeah,

30:55

so if in a few months you

30:58

might hear Patrick and I in Mandarin,

31:00

maybe we'll swap out an episode

31:02

just for fun and see if people notice. I

31:05

hope people notice. No, no,

31:07

an old episode. Oh, I see, I see. People

31:12

focus the machine learning stuff

31:14

on the open AI and

31:17

I understand why and the approach to AGI, I

31:23

guess, general intelligence. You're

31:25

right, there's a lot of stuff happening out

31:28

in the call it periphery that feels

31:30

like we're really moving forward, especially once

31:32

we start combining some of these things.

31:35

There's been a long work separating back speakers.

31:37

People say they can't tell the difference between

31:39

us. So I don't know, maybe that'll have

31:41

to be Gen 2 of that. But having

31:43

a single file separating it back and then

31:47

running it through the system and combining it. But

31:49

also, I don't know that we've talked about

31:52

it, the sort of nerfs

31:54

and Gaussian splatting around making

31:56

3D models from pictures.

31:58

And right now they need. kind of a

32:00

lot of pictures of a specific kind, but

32:02

that'll get progress as well. And

32:04

then you can imagine the sort

32:06

of stable diffusion feeding into those

32:08

things. Like there's like

32:11

a lot of interesting sort

32:13

of like one, two iteration away

32:17

and probably requiring some leaps there because

32:19

it's not sort of like gluing normal

32:21

software together, but yeah,

32:23

there's a lot of exciting stuff happening that I

32:25

think it's missed for the focus on LLMs. Yeah,

32:28

I actually am more excited about

32:31

diffusion models and LLMs. We

32:34

should do a whole show on diffusion models.

32:36

All right, well, we know a person who probably speak to

32:38

that. It's not me. Yeah,

32:43

I was wondering if we knew a third person.

32:46

But yeah, have

32:49

you seen the diffusion model with Bomberman? No.

32:53

Yeah, so basically, okay, let me just do

32:55

a really quick and then we'll dedicate a

32:57

show to this. The way diffusion models work

33:00

is you have a bunch

33:02

of references. So in the case of Mario

33:04

or Bomberman or any of these games, you

33:06

have levels that were created by humans, right?

33:08

That's your reference. And you

33:10

wanna create more of those, right? So

33:13

what you do is you corrupt

33:15

the references, in

33:22

this case, the Bomberman levels, you corrupt them, but

33:25

the way you corrupt them is reversible. So

33:28

whatever you do, you know the exact way

33:30

to undo it. And

33:32

then you train a model to

33:35

reverse the corruption. All

33:37

right? And then once

33:39

you have that model trained, you

33:41

give that model random noise,

33:44

like levels that are complete garbage,

33:47

and you just tell it to reverse the corruption,

33:49

and you tell it to do it again, and

33:51

again, and again. And you tell it to reverse

33:53

the corruption until

33:55

it gets stuck in a cycle. When

33:58

it gets stuck in a cycle, you say, done and

34:01

that's your ballroom in level. It

34:08

should have guarantees on playability, like

34:10

not having something you can't reach.

34:13

So it's

34:15

all just AI, so there's no

34:17

guarantees of anything. You'd have to

34:19

do some fine tuning afterwards, but

34:22

it's like scarily true

34:24

to form. I

34:27

can't remember if they wrote something

34:30

to throw out the levels.

34:35

They wrote basically a validator and threw out all

34:38

the bad levels. Or if

34:42

they didn't actually make a real game with it, it

34:44

was a research paper. So they might have not bothered

34:46

to do that. But you see

34:48

the levels and it's like, yeah, that

34:51

looks pretty cool. I could see how that

34:53

could be kind of a fun ballroom in

34:55

level. We

34:57

should save it for the diffusion, but

34:59

I guess there's some randomness that goes

35:01

into how you select each operation. So

35:04

I guess almost backtracking, you could just

35:06

naively get towards the end and try

35:08

to go back and assign some attempts

35:11

or whatever and do

35:13

a validation at the end. Did I get a good level?

35:15

Did I get a validatable

35:17

level? Yep, yep,

35:20

exactly. Oh man, it's going to

35:22

be cool. The future is scary. We're all going to be

35:24

out of jobs. The future is

35:26

super scary. I definitely wouldn't want to

35:28

be in the games industry. As an

35:30

engineer, you would be fine. But if

35:32

you're a concept artist

35:34

or a voice actor

35:36

or something like that, it's going to be

35:40

different. There are some heads of

35:42

studios and stuff coming out and

35:44

basically forecasting reduced need for

35:46

a lot of those things. People

35:49

talking about cold

35:51

calling and sales calls and informational

35:54

sessions, maybe reduced need for folks

35:56

there. And then also some places

35:58

are going to be in the industry. getting flooded, there's

36:00

always been a sort

36:03

of like low bar

36:05

Kindle book, ebook, you

36:07

know, coloring book, drawing,

36:09

just like, you know,

36:12

low effort sort of written word kind of stuff. And

36:15

all of that, too, I think is this

36:17

AI stuff is sort of competing with not

36:19

the top tier stuff, but the sort of

36:21

like low tier low effort, people

36:23

turning those out, hoping one or two people buy them

36:26

or whatever, which has proven successful.

36:28

I think there's a threat to

36:30

those not that that's going to be a big

36:33

problem. But once they start targeting people who are,

36:35

you know, higher up the food chain there,

36:37

that's that's going to be a struggle. And

36:40

I will weirdly segue that into our

36:42

book of the show from high quality

36:44

books that are they're definitely not AI

36:46

generated. That's right. So

36:49

my first one, I tried, I might have

36:51

recommended it before if I did. Well, it's

36:53

worth a second recommendation. I tried looking, but

36:55

after 171 shows, it's a it's a bit

36:57

of a process to try to search for

36:59

it. And that is

37:01

foundation by Isaac as as mob,

37:04

which is a actually many book

37:06

series. And

37:08

the first one is, you know, great place to start. And

37:11

recently, this became a TV show, which I

37:13

think renewed some interest. The TV show doesn't

37:15

follow the book super closely.

37:17

But if you've never read

37:20

some of the classic science fiction, it's

37:22

really like a fascinating trip to see

37:24

how close

37:27

some of the stuff got from people

37:29

writing long before, you know, computers were

37:31

a thing, much less mobile phones, some

37:34

stuff is just like laughably wrong. And

37:36

some stuff is just like, wow, that's

37:38

eerily the same issues we're dealing with

37:40

today. And so reading classic science fiction

37:43

is something we're doing if you've never

37:45

tried it before. And

37:47

the foundation series is a I feel

37:49

like a pretty formative series of science

37:51

fiction. And just to give

37:53

people flavor, since I've not said anything actually about the

37:55

book is the idea

37:57

is there's a an empire

38:00

are ruling the galaxy and there's

38:03

concern that these

38:07

people raise that the empire is going

38:09

to collapse and it's going to cause

38:11

humanity basically to be set back and

38:13

really be a struggle and everyone's going

38:15

to be bad off. And so they've created a

38:18

sort of way to help shortcut that to

38:20

make it better. But of course, the empire

38:22

itself is very unhappy with this. And

38:25

the whole prediction that this was going to happen is

38:27

this, of course,

38:29

fake, but it sounds very convincing

38:31

psycho history, which is this idea

38:33

that you cannot predict individuals, but

38:36

that, you know, sort of like

38:38

I think the analogy they use is like

38:40

a gas. You can predict a lot

38:42

of statistics about gases, even though an individual

38:44

molecule in a gas is, you

38:47

know, very, very difficult to predict. And so

38:49

through that, they can kind of predict the

38:51

course of human history and how things will

38:53

go and make updates and help keep it

38:55

on, you know, a plan that will, you

38:57

know, sort of benefit the most humans. And

38:59

it's just a very interesting read in the

39:01

dynamics. And so if you've never read it

39:03

before, and if that sounds at all interesting,

39:05

although I probably did a poor pitch of it,

39:07

or if you've watched the TV show and maybe

39:10

didn't realize it was also a book, I'll

39:13

pitch Foundation by Isaac Asimov. Yeah,

39:16

did you read it recently? Or is that when

39:18

you read? I read it actually as a teenager.

39:20

So I read it a long time ago. Yeah,

39:22

same here. I have to confess, I don't remember that

39:25

much of it. But I do remember having a big

39:27

impression on me when I read it. I

39:30

think I also had kind of forgotten I read it. And

39:33

then I like looked up the sort of summary and

39:35

I was like, Oh, I remember. And

39:37

then I actually have watched, I think there's a couple

39:39

of seasons now of the TV show. And

39:42

I will say, you know, it's a pretty dark,

39:44

dark thing to not watch with kids,

39:46

but pursuing

39:48

the same kinds of ideas. And I thought

39:51

I thought I like it, even though it

39:53

doesn't follow the book that closely. It uses

39:55

some of the concepts and the sort of

39:57

setup to kind of examine some of different

40:00

different ideas but in a similar vein. Cool,

40:05

yeah, definitely. Folks haven't read that. Definitely give

40:08

it a read. It's a great series. My

40:11

book of the show is Propaganda

40:13

by Edward Bernays. So

40:17

this is a fascinating book. It's

40:21

so wild that I almost, you

40:23

know, I really, I

40:26

wouldn't take it all as ground

40:28

truth, really. I mean, a

40:31

book about propaganda might be propaganda. You know, it's

40:33

kind of like meta, but. But

40:37

basically the premise, which they get right

40:40

into in the book, is that, and

40:46

again, this is not my theory

40:48

on life, but that basically, you

40:51

know, democracy can't work.

40:54

And so you

40:57

really need to have just like a few

40:59

different ideas to follow. It can't just be

41:01

a total free-for-all. I

41:04

mean, a good example is the app

41:07

store that Patrick was talking about, where there's

41:09

like a million knockoffs. Like you could search

41:11

pretty much any word in the dictionary, and

41:13

you're going to find a video game based

41:15

on that word. And they're all kind of cookie

41:18

cutter. And there's no, it's very hard to get noticed.

41:24

In the app store or to find good content. And

41:28

so you're really relying on the, you know, editors picks. And

41:31

so in a way, it's like you have this really democratic

41:33

platform, but what it really degenerates

41:35

to is just

41:37

like a handful of editors picks and

41:39

folks who have a big audience already and

41:41

all of that. And

41:44

so the claim by the book by the

41:46

propaganda book is that, you know, there's just

41:48

so much content, and this was written 100

41:50

years ago, 19 years ago. Yeah,

41:54

there's just, there's so many clubs that you could

41:56

be a part of. There's just so much, so

41:59

much. many choices, so many

42:01

brands, so many of this, that

42:04

really propaganda is the way

42:06

by which 90%

42:10

of your options get filtered out and you

42:12

can just focus on the top 10%. So

42:14

it almost becomes a

42:18

way to ante in. It's

42:20

like, okay, you have to build

42:23

this propaganda machine to

42:26

be part of of the system

42:29

and everyone who doesn't do that ends up

42:31

just getting kind of lost, right? And

42:35

it dives into how

42:37

propaganda works as a bunch of examples.

42:40

Some of them were pretty mind-blowing. Yeah,

42:45

I won't spoil them, but maybe

42:48

I'll spoil one. There's this

42:50

cigarette company that they

42:55

were under a lot of pressure. There were a lot

42:57

of moms who didn't want their sons

43:00

and daughters smoking because they had

43:02

heard about all these issues. I

43:04

don't know if lung cancer was a really known entity

43:06

in the 1920s, but they

43:08

thought there were health issues. And

43:11

so they were able

43:13

to start up this

43:16

Women's Liberation March sponsored

43:19

by Marlboro. And

43:22

they basically tapped into the

43:24

women's liberation movement of the

43:26

1920s as

43:29

sort of like a

43:31

back pressure against this

43:34

kind of like Mothers Against Smoking

43:36

group. And the

43:38

whole thing is just blowing my mind.

43:41

I mean, it goes without saying that. I

43:43

think that's fair to say. Patrick and I

43:45

are task oriented in

43:47

a sense. We're definitely not master manipulators.

43:55

We're not like these kind of people. Patrick, do

43:57

these even

44:00

have a social media account. I barely use

44:02

my social media account. And

44:05

so for me, this is like a

44:08

whole different dimension to, you

44:10

know, humanity, really. I mean,

44:12

it's something that I never would

44:14

have really considered or thought of if I

44:17

hadn't been recommended this book. So I

44:20

found it absolutely fascinating, highly

44:22

recommend it. Very interesting.

44:26

Yeah, I think sometimes

44:28

words get

44:33

like artificially intense negative things. So I

44:35

think like some of the things you're

44:37

saying, the other one I'll kind of

44:39

by analogy or by similarity talk about

44:41

is politics. So like people have I

44:43

don't want politics in the office. Like, what

44:46

you need to be aware, regardless

44:48

of whether you want to quote

44:50

unquote play politics, politics is a

44:52

like human way of, you

44:55

know, that things happen and decisions get made

44:57

and, you know, back door, back room decisions,

44:59

you know, these kinds of you have to

45:01

be aware of them. If

45:03

you, you know, want to kind

45:05

of like, at least have some modicum

45:07

of control over what happens. And I

45:10

think by extension, propaganda similar, it's, you

45:13

know, it has a very, very negative tone to

45:15

it. But it, I mean, it's, it's

45:17

a way of lots of people doing stuff, they

45:19

may not see it as propaganda, right? They just

45:21

see it as, hey, I'm trying to get my

45:23

ideas across or find a way to make people

45:25

more open to this. But sort of

45:27

calling it out and putting a label on it and saying how common

45:29

it is, is a way of sort

45:31

of engaging with it. Yeah,

45:33

and it turns out actually the

45:36

negative connotation to prop to the

45:38

word propaganda is also propaganda. So

45:42

the word they actually talk about this,

45:45

the root of the word propaganda comes

45:47

from the

45:50

Catholic Church in like the

45:53

1600s had a like

45:56

a part of seminary where they were

45:58

kind of teaching priests

46:01

and bishops and cardinals

46:05

about how to kind of spread

46:07

Catholicism. And I think

46:10

that's where the word actually comes from. And then

46:13

I think, and so at

46:15

that time, they had a

46:17

positive connotation. And I think the

46:19

connotation flipped over time as a

46:21

society we got kind

46:23

of like, you know, not happy

46:26

with like proselytizing. And then

46:28

now it's like really like

46:30

commercial propaganda, and then all of

46:32

that. So yeah, like

46:34

the propaganda, I think collided headfirst into

46:37

the Enlightenment and lost when it comes

46:39

to, you know, reputation as a word.

46:43

But yeah, it's, it's, yeah, but the book

46:45

is totally mind blowing, highly recommend people, people

46:47

read and to Patrick's point, and

46:50

we talked about this in the marketing episode,

46:52

we had an episode of the gentleman from

46:54

marketing, you know, like, you have

46:56

to, I'll give you a really

46:59

concrete example. You know, I made

47:01

that game AI hero, right? I

47:04

made it totally free. There's

47:07

no ads, there's no in-app purchases or

47:09

anything like that. You know, I told

47:11

my friends and family about it, all of that. And

47:14

it has I didn't check

47:16

recently, but it has, I

47:18

think, like 100 downloads. And

47:21

now, you know, I made it, you know, to

47:23

be frank, I kind of made it for myself.

47:25

I mean, so I'm not out there trying to

47:27

get a bunch of downloads. But I was even

47:29

still surprised that it got

47:32

such, such, such little notice.

47:35

And it really just doubles down on

47:37

that point that, you know, you

47:40

have to get a

47:42

propaganda cycle going a marketing cycle

47:44

going for anything that you want

47:46

to have mass appeal. Otherwise, there's

47:51

just too much content out there. So

47:54

yeah, it's thought of propaganda, marketing, you know, these aren't

47:56

words that you should be afraid of. I'll

48:01

say that it's not something that at least

48:03

I'm, I don't know about Patrick, but it's not something

48:05

that I personally am very good at. It's definitely something

48:07

where I try to partner with

48:09

other people who are much better at

48:11

it than I am. But

48:13

it's something that I think we all

48:15

should really appreciate. It is very important.

48:18

If I did want AI

48:20

Hero to get a ton

48:22

of people, I

48:25

would have to go about things very

48:27

differently. I think I'd have to be

48:29

posting regularly on Reddit. I'd have

48:31

to figure out some way to maybe

48:34

gin up some controversy. Maybe I'd make

48:36

a fake post saying that it's generative

48:38

AI, pretending to be somebody else attacking

48:42

me for using generative

48:44

AI. I don't know, but this is

48:46

just from reading the book. There's a million things you could

48:48

do to amp up

48:51

your thing, whatever it is. But if you

48:53

don't do that, probably

48:55

no one's going to pay attention. That's

48:57

true in politics. If you

48:59

make a political party, nobody's really going to care.

49:02

It's true in most

49:05

things. It's just a

49:07

fact of life. Yeah, there's

49:09

a lot to say here, but probably just

49:11

move on for now as they get it.

49:13

All right, yeah. Keep moving forward. Time for

49:16

Tool of the Show. All

49:18

right, what's your tool? Not a

49:20

tool, a game, but classic Patrick Koppout.

49:22

Mine is, I've been getting into these recently,

49:24

but this is the one that I encountered

49:26

first, which is The Room. This is pretty

49:29

old at this point. Most people

49:31

probably have played it before. If not,

49:33

I would encourage you to check one

49:35

of them out. Often, they're very cheap

49:37

or free on Android, iOS. I think

49:39

it's just a Steam version, probably a

49:41

web version somewhere. But the idea is

49:43

basically like an escape room in an

49:45

app. Just like without

49:48

the time limit, you're just like a series of

49:50

puzzles, and you

49:52

interact with them. For me, I've actually

49:54

never done an escape room. I

49:57

want to. It just hasn't been a thing that I've been

49:59

able to do. to do, but I'm sort of fascinated

50:01

by the concept and these games are just like an

50:03

easy way to spend a few minutes sort of playing

50:06

with something and trying to figure it out. I will

50:08

say in the genre some games do a better or

50:10

worse and I have my own particular sort

50:12

of like I don't like things that

50:14

you're just supposed to try random stuff until they happen.

50:16

There should be some clue that you

50:18

can look for and it can be obscure or it

50:21

could be that you have to pay attention or whatever,

50:23

but sometimes they fall into like how

50:26

was I supposed to know that? Like

50:28

there's no indication that you're supposed to

50:30

touch the top of the windowsill and

50:32

it's going to be movable. It's

50:35

like okay, but sometimes you fall into this.

50:37

But anyways, I guess

50:39

a pitch for these kinds of

50:42

puzzle escape room games in general

50:44

and there's many in the genre,

50:46

but for me the first one

50:48

that I sort of bumped into was the room

50:50

series. I think there's like now four or five

50:52

in the series, but shout out

50:54

to the original one and if you've never

50:56

played one before, you know, check

50:59

it out. Yeah,

51:01

totally. I've played The Room. I think

51:03

it's great, a lot of fun. It

51:05

taps into your common sense. You do

51:07

have some intuition into what the

51:09

answer to the puzzle is. My

51:12

tool of the show is also somewhat of

51:15

a cop-out because we've had them on the

51:17

show before. We did in fact had a

51:19

whole episode dedicated to Incredibuild, but

51:21

I was building a new

51:23

version of Mamehub on my

51:26

house today, I guess, because it's like using all the desktops

51:28

in the house. It's kind of the point of the tool.

51:32

And I just thought to myself, man, this is such

51:34

an amazing tool. I've used it for so many years

51:36

for free. If you're

51:38

using it for your home use, it's

51:40

not for professional use. It's totally free. I

51:43

felt like they deserve another shout out

51:45

after so many years. We've gotten tons

51:47

of emails from folks who have

51:51

got their company to buy Incredibuild and

51:53

all of that. We have been able

51:55

to drive some business to them, which

51:57

I'm really proud of. That's

52:00

Incredibilt. Basically, the way it works

52:02

is it hooks into your build

52:04

system. It even has plugins for

52:07

Visual Studio. So if you're building on Visual

52:09

Studio, it works. If not, it

52:12

hooks into GCC and all that stuff. And

52:15

instead of choosing Build

52:17

Solution for Visual Studio,

52:19

you choose Start Incredibilt.

52:22

And I have the Incredibilt

52:24

agent running on all the desktops here in the

52:26

house. And the fans all start roaring and they

52:29

all start working together

52:31

to build stuff. So Mame

52:33

Hub takes, I think, like two

52:35

and a half hours to build. But

52:39

with Incredibilt, I can get it done in

52:42

like 10 minutes or something. So

52:45

it's super nice and highly

52:47

recommend it. Very

52:50

nice. Very nice. All

52:52

right. Well, I think it's time

52:54

for our topic, which is compilers

52:56

and interpreters. Yeah, this

52:58

is a request by Jessica W. I

53:00

won't say your last name. I know some people

53:02

might census that. But thank you so much, Jessica,

53:04

for this idea. It's a great one. All

53:08

right. I didn't know Jason sort of bowed out. It's

53:10

not my background. So it kind

53:14

of fell to me. And I struggled a bit with

53:16

the tact to take here because there's a lot to

53:19

be said. This is a big

53:21

request. So I decided to take a

53:24

stab at what I'll say

53:26

is a sampler, kind of going through

53:28

a bunch of different topics in

53:30

the area with some

53:33

sort of high level overview. But

53:35

choosing an interpreted versus

53:37

compiled language. I guess you could say choosing a

53:39

compiler versus interpreter. But most of the time, if

53:41

you choose a language in the pros and cons

53:43

of language, you're sort of choosing

53:45

one of these for that. But then in

53:47

each of these, there's just an incredible depth

53:51

you can go into or not, you

53:53

know, most of the time, I guess,

53:56

we're big believers and you know, the right tool

53:58

for the job. despite the fact that

54:00

I may throw stones at

54:02

Python, you know, a lot of

54:05

times, you're only going to run an analysis once

54:07

or, you know, it doesn't really even really matter how

54:09

long it takes. And so you

54:11

can get this just really big worry

54:13

about speed or performance or, you know,

54:15

selecting exactly the best tool. But I

54:17

will say, most of the programs I

54:19

write are probably run less than a

54:21

dozen times ever, even to

54:23

check them in and put them in at work,

54:25

you know, and you, I may

54:27

run it a few times, someone else may use

54:30

it, we may copy and paste the code around,

54:32

but any given instances is just not that. And

54:34

then, and then a very few are, you

54:36

know, run hundreds of times every day, and you know,

54:39

or thousands of times, interloops and calls

54:41

and those ones. And so there's

54:43

a really, a really wide range. And

54:46

so find a few things

54:48

that you're you're sort of proficient in. And I'll

54:50

say about way of disclaimer before we dive in

54:52

here is just sort of like, you'll

54:55

see a lot of internet flamewars about,

54:57

you know, oh, that that's slow, because it's, you

54:59

know, scripting, or, you know, don't if you're going

55:02

to build something in Godot, don't use as a

55:04

GD script, because it's, you know, it's going to

55:06

be slow. And it's, I

55:08

don't, I mean, if you're finding yourself

55:10

in a problem, and maybe you want to listen

55:12

to some of that, or being aware is good.

55:14

But I feel a lot of people hesitate to

55:17

take the first step and get what you know,

55:19

call analysis paralysis, you know,

55:21

from making a decision and actually better

55:23

to just build something and let it suck

55:25

and be slow and, and introspect why at

55:27

the end of getting somewhere with it and

55:29

then you know, make a follow up choice,

55:31

then to never take the first step. Yep,

55:34

totally. There's a balance there. But

55:36

okay, off of my soapbox. All

55:38

right. I guess

55:40

anytime we run a program, we're talking about, you

55:42

know, a processor. And we were talking in the

55:45

Christmas episode, holiday episode, Jason

55:48

was mentioning, oh, and I already forgot if it

55:50

was risk V or risk five, I think risk,

55:52

risk five. Yeah, resounded. Okay, good. Risk five and

55:54

risk five is an instruction

55:57

set architecture, a sort of way of

56:00

of having computers run and it has, I guess,

56:02

what you would call machine code. So out

56:05

of, you know, wherever it's

56:07

loading the program, probably RAM, and then

56:09

it sort of loads a set of

56:12

bits, right? If it's a 64-bit processor,

56:14

it loads, you know, 64 bits and can

56:16

kind of look through there. And part of

56:18

that is the operation to perform in some

56:20

instructions about to load it from a register

56:22

or to load it from a place

56:25

in memory or whatever. And then the

56:27

processor sort of handles that. And that

56:29

is machine code. If you ever do

56:31

like NAND to Tetris, or there's another

56:33

game where you build logic gates out

56:35

of NAND on your web browser, or

56:37

you take a class in university where

56:39

they're sort of teaching you about sort

56:42

of building up from like VHDL, you know, how

56:44

to kind of like, you know, implement a processor,

56:47

you'll end up realizing it really is

56:49

just, you know, hex numbers on a

56:51

screen that you, you know, end up

56:53

getting executed in their, that's your program.

56:56

And every kind

56:58

of family of processors has a different set of

57:00

instructions, a different way of organizing those operations, and

57:03

to write a program in it, you know, you

57:05

would have to sit there and really write, write,

57:07

you know, a hex, you know, or you could

57:09

write binary too, but, you know, most people would

57:11

write it in hex. And that's the way we

57:13

sort of think about it, because it gives you

57:15

the nice byte delineation. And

57:18

you would not do that after about

57:21

maybe one time of doing that, you would not

57:23

want to do that anymore. And so as an

57:25

example, like when we did Motorola 68k

57:28

processors, which is just, you know, one of these

57:30

families, when I was in university, and we had

57:32

to do this as a task, like, you know,

57:35

write some, you know, machine code. And I did

57:37

that once before sort of realizing, wait a minute,

57:39

I'm just going to write a program to be

57:41

like a crappy assembler, and,

57:43

you know, do that. And so then that moves us

57:45

to assembly language. And so assembly

57:48

language is like a very

57:50

light skin over this

57:52

machine code. So you're pretty much

57:54

having one to one correspondence, but

57:56

you're writing English symbols. So instead

57:58

of hex code, code 47

58:01

being the move operator where

58:03

you move one set

58:05

of bytes to another location, you would

58:07

say MOV, move, right? That's a lot

58:09

easier. And then the registers each have

58:12

names rather than just numbers like R8

58:15

or whatever for register 8. And

58:18

so this is the first time we encounter

58:20

what would you would call a compiler. And

58:22

the compiler takes those English

58:24

letters or whatever language you're writing in

58:27

letters and maps them directly and emits

58:29

the bytecode. So

58:31

does that mean that it's reversible? Like, can

58:33

you go from machine code to assembly code?

58:37

Yes, great, great question. So that you

58:39

can it's called disassembly. And for

58:41

assembly, and I'll talk about a caveat here

58:43

in a second, it's pretty straightforward. So if

58:45

you take a program that lives and you

58:47

haven't gone, we had, oh, that was way

58:49

that was many years ago. But we had

58:51

the folks from I think Intel in here

58:53

talking about obfuscating your program codes that people

58:55

can't do that. And they

58:58

can't sort of like patch it or do

59:00

what would in video games would be hacking

59:02

or, you know, and more critical software trying

59:04

to get your keys or something. But yeah,

59:06

so you can do disassembly. So you can

59:08

take the bytes flowing through your processor and

59:10

sort of convert them back

59:12

to assembly because yeah, it's

59:14

a very direct mapping. Got

59:17

it. Okay. So the

59:19

caveat there that I was already alluding to, and

59:21

you can can help here, it's not too bad

59:23

in assembly, is what I've described so

59:25

far, you kind of do in a single pass.

59:27

But that also turns out to be easier

59:30

to do better than and you know, it's frustrating, so you wouldn't

59:32

want to stay there very long, which is you

59:34

may want, you may want to have an example where

59:36

you have a loop. And so when you're

59:38

on a loop, you want to go back in your

59:40

program to somewhere previous, very pretty good forward if you

59:43

have an if statement. And

59:45

if you were doing what, you know, I described initially,

59:47

and that's kind of like a single pass compiler, you

59:50

would have to calculate, you know, go down, figure

59:52

it out as eight instructions for my loop. So

59:54

I need to go back to that branch statement

59:56

and say jump eight instructions down, that'd be really

59:58

annoying. So What people do

1:00:01

is implement the ability to add labels so you can put

1:00:03

a label You know end of

1:00:05

loop and you can say you know

1:00:07

jump JMP and then you can put end of

1:00:09

loop The issue is when

1:00:11

you're doing a single pass compiler you're scanning down.

1:00:13

It says jump end of loop But

1:00:15

it's never seen the symbol end of loop So

1:00:18

it does where to jump to right and

1:00:20

so now you get what we call sort

1:00:22

of a two pass compiler So now you

1:00:24

start to get what you would kind of

1:00:26

think about variable names and labels and jump

1:00:28

statements It's not really that

1:00:30

much extra But it your productivity goes

1:00:32

way higher and for many years people

1:00:35

coded an assembly with two pass compilers

1:00:37

and the two pass just means first scans through

1:00:40

to find all of the symbols that are names

1:00:42

and Their you know locations and

1:00:44

then on the second pass does the actual

1:00:46

emit the machine code and now it knows

1:00:48

Oh this offset is you know this many

1:00:51

instructions down and it's able to produce the

1:00:53

machine code That's it. Does it does it

1:00:55

literally do it in two passes because I

1:00:57

could also imagine like anytime you see a

1:01:00

jump forward You just keep track of it

1:01:02

and then you go back and fill it

1:01:04

in I Mean,

1:01:06

but that sure but that's still you can't do

1:01:08

it in a single It's not one pass because

1:01:10

you have to go back up right and right

1:01:12

in it So I guess

1:01:14

like the big O notation kind of thing like

1:01:16

how do you want to call it? I'm sure

1:01:19

I'm sure they'd get pretty sophisticated. There's probably you

1:01:21

know, it's just calling out that the

1:01:23

progression I guess in here, but still

1:01:25

these Assembly languages

1:01:27

you're writing in, you know, there's kind

1:01:30

of some asterisk here, but for the

1:01:32

most part you're writing it per You

1:01:35

know kind of processor So if you want to write

1:01:37

it on a you know if you're gonna use an

1:01:39

Arduino, which is I think based on a pic chip,

1:01:41

then you need to use the pic flavor of assembly

1:01:44

if you want to do this in You

1:01:47

know Motorola 68 K x86

1:01:50

processor arm processor if you're you

1:01:52

know gonna run on a phone Every

1:01:54

one of these has a different set

1:01:56

of instructions and therefore a different assembly

1:01:58

language. And so again that did it's

1:02:00

really crappy if you want to ever

1:02:02

write for more than one thing or

1:02:05

you want to go faster because you're

1:02:07

pretty much still basically writing one line

1:02:09

of code per instruction, which

1:02:12

is going to make you very easily

1:02:14

able to sort of intuit about sort of

1:02:16

like runtime and complexity of things a

1:02:18

little easier than a high level language. But

1:02:21

again, it gets kind of frustrating. And there

1:02:23

are concepts for still everything that's

1:02:25

in a high level language sort of has to

1:02:27

be capable of being done in assembly language. So

1:02:29

you still see things like doing function calls or

1:02:32

jumping to a function and pushing stuff onto the

1:02:34

stack. For some

1:02:37

processors, those things are strictly implemented

1:02:39

by a higher language compiler. But

1:02:41

some processors actually have support for

1:02:43

that even at the assembly language

1:02:45

level. The newest sort of

1:02:47

start to call those things complex instruction

1:02:50

sets, Sys versus risk reduced instruction set

1:02:53

computers. And so there's a whole

1:02:55

separate topic. We'll sort of leave this

1:02:57

here because we'll go forever and move

1:02:59

up the family tree,

1:03:02

I guess, to higher level

1:03:04

languages. Yeah,

1:03:08

that makes sense.

1:03:10

So every programming language

1:03:13

that you read about, C, C++, Java, Python,

1:03:17

they're all considered high level. Yeah,

1:03:20

yeah, yeah. I mean, I

1:03:22

don't, there's degrees to that, I guess, you

1:03:25

know, but at least

1:03:27

when I sort of started learning about this, and you know,

1:03:29

they shift over time, right? There was early

1:03:31

on, it was you writing an assembly or

1:03:33

you're writing in, you know, something that wasn't assembly

1:03:35

and anything that wasn't assembly was, you know,

1:03:38

high level. And the idea there being,

1:03:40

in my mind, at least the delineation,

1:03:43

and it may be it's moved over time. But

1:03:45

if you write a line of

1:03:48

code, and it produces sort of

1:03:50

an arbitrary number of machine instructions,

1:03:53

then you're now writing in a

1:03:55

high level language. Got

1:03:57

it. Okay. So in a line of

1:03:59

C code. just as an example or

1:04:01

Java or whatever, you can write A plus

1:04:04

B divided by C times D

1:04:07

to the power two, whatever, right?

1:04:11

That's one line, but it's many, many, many assembly operations

1:04:14

to do that depending on architecture. And

1:04:17

the compiler now is able

1:04:19

to target different backends. So if you write, and

1:04:21

we'll talk about CC++ here for

1:04:24

a second, if you're writing in

1:04:26

CC or C++, you can compile against

1:04:28

many different processors. So you could write

1:04:30

one program that runs on ARM or

1:04:32

on x86 because the compiler

1:04:35

is picking up your code and we

1:04:38

won't go into it, but building, sort

1:04:40

of finding the symbols, building a syntax

1:04:42

tree, and it knows how to emit

1:04:44

the proper machine code for executing

1:04:47

that program. And it does, it can

1:04:49

do that with many passes, depending on

1:04:52

the compilers, but many compilers also emit

1:04:54

a sort of intermediate representation. So even

1:04:57

different frontends, so you could say,

1:05:00

like an LLVM, you'll see

1:05:02

things that can target, you

1:05:05

know, the LLVM IR, intermediate

1:05:07

representation, which is something,

1:05:09

which means the sort of machine code

1:05:12

assembly and the high level language that

1:05:14

Watts and lots of frontends can compile

1:05:16

into. And then there's a specific set

1:05:19

of operations and libraries and team

1:05:21

working on taking that intermediate representation

1:05:24

and compiling it down to your

1:05:26

specific final machine. And so

1:05:28

then that's called the backend. So you have

1:05:30

the frontend, compiling it to intermediate representation and

1:05:33

then the backend, taking it from intermediate representation

1:05:35

to a specific class of

1:05:37

processor. Got it,

1:05:39

I see. So if you wanted to write a new

1:05:41

language, you

1:05:44

could create a language that

1:05:46

compiles to the LLVM immediate

1:05:49

representation, and then you'd

1:05:51

be guaranteed that it would run on

1:05:54

all these different processors. And

1:05:56

not only that, you're also gonna benefit from

1:05:58

a bunch of optimizations that. the

1:06:00

sort of like imagine loop unrolling. So

1:06:02

you write a loop in this intermediate

1:06:05

representation and now the sophisticated folks across

1:06:07

all these languages, economies of scale kind

1:06:09

of thing, can implement loop unrolling or

1:06:11

some stuff we're gonna talk about later,

1:06:13

but it can implement it at the IR

1:06:16

level and then you can benefit from it. Got

1:06:18

it, cool, that makes sense. So

1:06:20

now you mentioned Java as well. Java

1:06:23

gets a special shout out

1:06:25

here, which is that

1:06:29

when Java is compiled and this starts

1:06:31

to say, you know, these are words

1:06:33

we use, but you're

1:06:35

no requirement that it has

1:06:37

to map cleanly. So Java

1:06:39

also targets sort of an

1:06:41

intermediate representation called the JVM,

1:06:44

the Java Virtual Machine byte code. And

1:06:47

so it does get compiled and it has

1:06:49

all this optimization and stuff that

1:06:51

can happen. And then it produces

1:06:54

basically a machine code, but the

1:06:56

machine it targets is the Java Virtual

1:06:58

Machine, not an individual processor. And when

1:07:00

it was first developed, this was a

1:07:03

huge boon because now you

1:07:05

talk about portability, you can run on anything

1:07:07

that there was a JVM for, the JVM

1:07:09

for, and LLVM wasn't really a big time.

1:07:12

So, you know, again, times change, but

1:07:14

at the time, this was like, you know, it's like a

1:07:16

really big thing and Sun Microsystems

1:07:18

at the time, you know, we're sort

1:07:20

of pushing that, hey, Java really, you

1:07:22

know, runs on, I forget, I had

1:07:24

an ad, you know, some obscene number

1:07:26

of devices and, you know, has this

1:07:28

ability. And so there you're compiling it

1:07:30

down. And then the Java Virtual Machine

1:07:33

has an implementation per, you know,

1:07:35

kind of architecture. So it has an ARM

1:07:38

implementation and it has an

1:07:40

x86 implementation. And the implementation

1:07:43

of that virtual machine is to

1:07:45

take that byte code and pretend

1:07:47

it emulates, right? Emulates this sort

1:07:50

of conceptual computer that is like

1:07:52

a, almost like a lowest common

1:07:54

denominator, but on some processors,

1:07:57

it may be able to be very efficient with a

1:07:59

certain option. and others not, but you don't

1:08:01

have to worry about that because it's

1:08:04

sort of just handled for you.

1:08:06

And all of the standard libraries

1:08:08

can have sort of special treatment

1:08:11

or efficiencies added. And so this

1:08:14

was a very interesting approach, but it sort

1:08:16

of threads the line here where you have,

1:08:19

it's not a scripting language, we'll talk about that

1:08:21

in a second, but it's not a true compile

1:08:23

in that there's still an emulator, a virtual machine

1:08:25

as it were, that sits in the middle here

1:08:28

and sort of that's the program that's actually

1:08:30

running. So that program is running your program.

1:08:34

And so the compiler is not targeting a machine,

1:08:36

the compiler is targeting, well it is targeting a

1:08:38

machine, but it's not targeting a physical machine. So

1:08:41

if I understand correctly, right? So the

1:08:43

difference is with LLVM, you're

1:08:46

going to this immediate representation,

1:08:48

but then you're instantly going to

1:08:51

machine code. And

1:08:55

so now you have machine code that

1:08:58

can only run on that architecture.

1:09:01

But with Java, you're

1:09:03

still going to this intermediate format, but

1:09:06

you're not doing that second step to

1:09:08

go to machine code at

1:09:10

compile time, you're doing it at runtime.

1:09:13

Yeah, so the intermediate representation is

1:09:15

not really executed, right? There's no

1:09:17

program, well, again, you could

1:09:20

write one, but the intention isn't that

1:09:22

there's a program that runs the intermediate

1:09:25

representation. There is a

1:09:27

program, the JVM, that runs your

1:09:29

JVM byte code. Got

1:09:33

it. Okay, so

1:09:35

then now we get to sort of

1:09:37

scripting. And scripting and

1:09:40

interpreters, the idea is

1:09:42

sort of similar, it's very rhyming,

1:09:45

which basic was this way, other

1:09:47

things are this way. And that

1:09:49

is rather than a compiler, you're

1:09:52

running a program that

1:09:54

needs to work per

1:09:56

architecture, but that program is

1:09:58

doing all of the... work at

1:10:00

runtime of executing your script.

1:10:03

So it's converting

1:10:05

the symbols you wrote into

1:10:09

the letters you wrote into symbols. It's

1:10:11

building whatever it needs to

1:10:13

be able to execute it. So if you're

1:10:15

setting a variable name like Jason was mentioning,

1:10:17

it's keeping track of that. But it's doing

1:10:19

it sort of fresh every run. And so

1:10:21

it's opening it up. And if you have,

1:10:23

as an example, you know, a

1:10:25

typo halfway down, you're going to find

1:10:28

out about it out about it, a

1:10:30

down your program's execution or whatever. And

1:10:32

so it's not attempting to in this

1:10:34

sort of level

1:10:36

zero, it's not attempting to do anything other than

1:10:38

just sort of execute your script, right. So if

1:10:40

you've ever written Python, and you have a, you

1:10:42

know, typo at the bottom of your program, and

1:10:44

your whole program runs, and it doesn't output, you

1:10:46

know, the answer because it crashes at the end,

1:10:48

this is happening, it's not trying to compile,

1:10:51

it's not trying to look ahead, it's just there

1:10:53

is the program that's running on per

1:10:56

architecture is parsing

1:10:58

the script input at runtime.

1:11:02

Yeah, I mean, you can, this is

1:11:04

starting to really boggle my mind

1:11:06

that I'm thinking about it, like,

1:11:08

in Python, you can open the

1:11:10

interactive interpreter. And you can actually have,

1:11:12

you can write if statements and for loops

1:11:14

and anything you can write in Python, you

1:11:16

can write in there. And

1:11:18

so you can actually like write

1:11:21

a for loop in that for

1:11:23

loop have a break, if

1:11:26

a certain condition is met, and

1:11:28

then keep going. And so like,

1:11:30

Python has to kind of

1:11:32

remember that like you put that break

1:11:34

there, but it doesn't really know what's going to

1:11:37

happen after the break, because like you literally haven't

1:11:39

written it yet. And so

1:11:41

that's the kind of keep track of all of

1:11:43

these things while you're while it's interpreting what you're

1:11:45

writing. It's also not that

1:11:49

it's impossible other places, but it becomes much more

1:11:51

straightforward to things like Python

1:11:53

loves to do, which is like modified

1:11:55

definitions on the fly, right? So like

1:11:58

changing variable names to another. type

1:12:00

or to even modifying functions that

1:12:02

are on a class because all

1:12:05

those things are sort of held in memory,

1:12:07

right? Like the sort of processing,

1:12:10

the compilation goes into memory and is

1:12:12

just held and is done sort of

1:12:14

in real time like at execution. And

1:12:16

so modifying that stuff is more

1:12:20

obviously, you know, straightforward to implement.

1:12:23

It's not obvious, but you can kind of

1:12:25

envision how you would do that if it's

1:12:27

like a thing sitting in RAM versus if

1:12:29

you've compiled it into something that you admitted

1:12:31

to the disk, right? And you're sort of

1:12:33

executing out of read-only memory. And

1:12:35

again, there's lots of things that blur

1:12:37

these lines. But yeah, so scripting, when

1:12:41

people sort of think about it, it can be slower

1:12:43

because at runtime, you have to

1:12:45

do more, right? You have to convert the

1:12:48

mapping. You don't need

1:12:50

to know that, you know, your variable name

1:12:52

is some complex, you know, set of underscores.

1:12:55

When you're doing the compile, you just map it to, this

1:12:57

is the first variable I've seen. So it's

1:12:59

variable lookup table entry zero, right? It's literally

1:13:01

just a pointer or just an index zero.

1:13:04

And then every time you see that, you're keeping track

1:13:06

of it. But in the final program, it

1:13:09

all gets taken away, right? Because it's just everything

1:13:11

points to the right spot. And

1:13:13

all those things are resolved. When you're

1:13:15

doing scripting, though, when you're marching down the program,

1:13:18

that table has to live in memory. And you have to

1:13:20

go do that lookup, right? Jason's

1:13:23

most complicated, very long named, but

1:13:25

not meaningful variable, that has

1:13:27

to get parse, which is a lot of characters, and

1:13:30

has to get that mapping to take place

1:13:32

every time it seems sort of like at

1:13:34

runtime. And so this gets to why people

1:13:36

kind of say scripting would be slower. But

1:13:38

as Jason mentioned, the advantage is because execution

1:13:40

is happening at like, you know,

1:13:42

kind of while you're doing it, and all these

1:13:44

steps are you can also modify the steps. And

1:13:46

it becomes very easy to do and to interactive

1:13:48

and to kind of see what's happening. And,

1:13:51

you know, very straightforward. Yeah,

1:13:53

that makes sense. I believe

1:13:56

if you your

1:13:59

machine code. is loaded

1:14:02

into the process memory that you

1:14:04

can't touch. If

1:14:06

you start writing in

1:14:08

the chunk of memory where your machine

1:14:10

code went, they'll

1:14:13

assume that you're trying to do something nefarious

1:14:15

or something or a mistake, but

1:14:18

in Python there's probably no way to catch that. A

1:14:21

lot of early viruses did this

1:14:23

self-modifying code. So people screen

1:14:25

them, they look normal, but then they get modified to

1:14:27

do something that is nefarious

1:14:30

at execution. Yeah,

1:14:32

a lot of people, a lot of systems will

1:14:34

try to stop or prevent that and assume that

1:14:36

you were making a mistake and at best preventing

1:14:39

you from doing something you really weren't supposed to

1:14:41

do. Yeah, that

1:14:43

makes sense. But yeah, that's mind-bendy stuff when you

1:14:45

start talking about that. Okay, two couple

1:14:47

other things I wanted to cover briefly. You'll

1:14:49

hear about, so one is a JIT, which

1:14:52

is just in time. So for scripting languages,

1:14:54

all is not lost about these runtimes. So

1:14:56

in Python as an example or other things,

1:14:58

you can have a JIT,

1:15:00

which says, hey, I've noticed

1:15:03

that this piece of script

1:15:05

has been executed, this function has been called

1:15:07

many, many, many times. So I'm going to,

1:15:09

rather than every time I get to it,

1:15:12

do all the parsing and all the work

1:15:14

again, it can effectively do a compilation step.

1:15:16

It can go ahead and resolve all of

1:15:18

those things, it could even go could do,

1:15:21

implementing them all the way out to machine code or

1:15:23

to operations or whatever, right? Java

1:15:25

has this as well for byte code,

1:15:28

but basically at runtime, sort of observing

1:15:30

that you're about to do something or

1:15:32

that you've done it before and

1:15:35

pre-computing and storing those results so that the

1:15:37

next time you do it, you

1:15:39

get your execution. That's sort of the

1:15:41

concept, I'm probably doing a bad job,

1:15:43

but of just in time, right? It's

1:15:45

just in the nick of time, I've

1:15:47

provided you a optimized thing to run.

1:15:50

Right, and you can do the loop unrolling and

1:15:53

all of that stuff. If you know that this

1:15:55

for loop's only been executed four times or

1:15:57

you don't know it, but you're confident, then

1:16:00

you could unroll it. Yep. And

1:16:03

you could even do code analysis to

1:16:05

do it right again, that you know,

1:16:08

extend these all have like research directions

1:16:10

or implementations and various, you know, features.

1:16:12

And then the last thing I want to talk about

1:16:14

that also blurs the line though, is something

1:16:16

that's actually interesting you've ever seen it, which

1:16:19

is profile guided optimization. So some

1:16:22

of these drag races, I guess you

1:16:24

call them between sort of scripting languages

1:16:27

and compile languages will show, Hey, actually

1:16:29

this Python code is a lot faster

1:16:31

because in this case, the, you know,

1:16:33

during runtime, Python, the interpreter observed

1:16:35

these behaviors and perform this optimization

1:16:37

that you couldn't have known a

1:16:39

priori, it had to be known

1:16:41

sort of like at runtime and

1:16:44

it gains this advantage. It's

1:16:46

true. That can happen. That is a thing. And

1:16:49

it can be really cool. In compiled languages,

1:16:52

we're starting to see some work

1:16:54

around profile guided optimization, which is

1:16:56

run your program and

1:16:58

you record a bunch of statistics about

1:17:00

what, what, how many times various loops

1:17:02

by default was this if statement skipped

1:17:05

99% of the time and even compilers

1:17:07

do some of this, sorry, even processors

1:17:09

do some of this to help with

1:17:11

optimization, but sort of recording statistics about

1:17:13

what was executed, how much and various

1:17:15

sort of things. And, and that the sort

1:17:17

of call it happy path, you make a guess, I'm

1:17:20

just going to assume this is going to be true

1:17:22

and I'll add a little check. If

1:17:24

it's true, I'm going to gain this boost. And

1:17:26

if it's not true, then I'll have to execute something

1:17:28

a little slower, right? No free lunch. But

1:17:31

it does this via a profiling. So

1:17:33

you've profiled your code. And now this

1:17:35

optimization uses that profile of the code

1:17:37

as a sort of almost second compilation.

1:17:39

So you run the compilation, you insert

1:17:41

all these sort of like, you know,

1:17:44

extra hooks to measure stuff. You

1:17:47

look at the output and then you run compiling again,

1:17:49

that uses that as

1:17:51

input and it's able to sort of

1:17:53

be more thoughtful about how to perform

1:17:56

certain optimization steps of compiling. Also

1:18:00

if people are really excited about

1:18:03

compilers and interpreters and

1:18:05

they want to work on the. I

1:18:08

thought of the mr the c++

1:18:10

like gcc or something where

1:18:12

does most of that work take place but

1:18:14

i guess the pilot government. Programs

1:18:16

and giant companies and stuff like what

1:18:18

would be your career advice so

1:18:21

there are folks working up you know big

1:18:23

company big tech companies that you sort of

1:18:25

here right if we if we sort of

1:18:27

name some you know dark you know i

1:18:30

google or swift at apple or i think

1:18:33

java is still at oracle you know

1:18:35

obviously those places have teams that work

1:18:37

on these things and you can go

1:18:40

work on them. But

1:18:42

I think a lot of it also happens

1:18:44

as a hobbyist do stuff you

1:18:47

know in university, this

1:18:49

is compilers and interpreters is,

1:18:51

of course, you can take. If

1:18:54

you want to do it as a job. What

1:18:57

I don't know is like how you as a

1:18:59

job without working at a big company that has

1:19:01

a vested interest in it you sort of get

1:19:03

paid to make a very cool. lvm

1:19:06

and intermediate representation optimization definitely

1:19:09

something you can do. And

1:19:11

if you have a company that will sponsor their work which is you

1:19:13

know maybe a hard thing to find i don't

1:19:16

know that it's not like writing a game we just do it and

1:19:18

then people pay you for a good outfit.

1:19:20

Very difficult to write like a new

1:19:22

language that someone would pay

1:19:24

you just to like consult on my

1:19:28

opinion maybe i'm wrong. Yeah

1:19:31

i think you're right i think even if if

1:19:33

you set out to create a

1:19:35

new language i think

1:19:38

even then i would probably start by.

1:19:41

Working for one of these big companies i think like

1:19:44

i want to say. What

1:19:47

is the military labs is like the.

1:19:50

Keep person on gcc like send you

1:19:52

are one of these labs i don't

1:19:54

remember but basically you know for all

1:19:56

of these things lvms

1:19:58

with gcc. You know,

1:20:01

look at who their biggest sponsors

1:20:03

are, right? Look at

1:20:05

who the domain

1:20:07

name of the email addresses that are

1:20:09

most prevalent on the mailing list, right?

1:20:12

It's not hard to figure out who

1:20:14

is actually putting in most,

1:20:17

you know, of the effort for

1:20:20

that particular compiler or interpreter. And

1:20:22

then once you know that,

1:20:24

then it's a matter of, you

1:20:27

know, creating changes. You

1:20:29

know, basically doing that job kind of

1:20:31

pro bono for a little while and

1:20:33

proving yourself to those folks, you

1:20:36

know, becoming ingratiated by those

1:20:39

folks and then reaching out to them and saying, hey, I

1:20:41

want to do this full time. And

1:20:43

so if that's something that interests you, that's the path for

1:20:46

it to do that. You don't need to be a

1:20:48

master propagandist or anything like

1:20:51

that. It's a very simple formula.

1:20:54

Just put some sweat equity

1:20:56

into that. You

1:20:58

know, talk to the folks who are also putting

1:21:01

in that sweat equity. I'll

1:21:04

give two shout outs quickly, you

1:21:07

know, for I don't think this is the first

1:21:09

thing you do, right? Write programs for a while

1:21:11

before attempting to write your own programming language. Right.

1:21:14

Probably a good move. But two resources

1:21:16

that I think are really

1:21:19

useful here, many more, but, you know,

1:21:21

people would have recommended the Dragon compiler book

1:21:24

before, but I never made it through it.

1:21:27

Anyways, the first one is by, and

1:21:29

I might not say the last name

1:21:32

correctly, but Robert Neistrum, Neistrum, and he

1:21:34

wrote a game design pattern, like design

1:21:36

patterns for games that's very useful. Definitely

1:21:39

check that out. I think you can read it

1:21:41

free online, but also read free online crafting interpreters.

1:21:44

So craftinginterpreters.com. I

1:21:46

think there's a web version, but I think there's actually a

1:21:49

physical book as well. And I

1:21:51

haven't sort of gone through it and implemented it myself.

1:21:53

It's on my to-do list. But definitely

1:21:55

a way of talking about how

1:21:57

would you add an interpreter, implement Polymer?

1:22:00

How would you handle variables and sort of building

1:22:02

up your own programming language and sort of, you

1:22:05

know, not going to be, I don't think state

1:22:07

of the art, you know, pushing the envelope, but

1:22:10

definitely giving you a flavor for this, which I

1:22:12

think is really useful to know how some of

1:22:14

this stuff works. And then in

1:22:17

the sort of opposite side of that, I

1:22:19

mentioned it, but nandgame.com. There's

1:22:22

also NANDtetetris, but nandgame.com is really easy

1:22:24

to jump in and start showing you,

1:22:26

like, one

1:22:29

touch too low, maybe, for what we talked about

1:22:32

today, where how do you build a NANDgate? How

1:22:34

do you make an OR gate from a NANDgate?

1:22:36

But quickly jumps in, basically, how do you build,

1:22:39

you know, the actual

1:22:41

processor components, the

1:22:44

parts that do math? How do

1:22:46

the bits get added? And why

1:22:48

is it that a certain machine

1:22:50

code byte evokes a certain behavior

1:22:54

from the pieces of the computer, the actual silicon? Anyways, and

1:22:56

so you can just, like, play it in your web browser

1:22:58

and, you know, just sort of go at it. You might

1:23:00

have to Google some help. I did. Sorry.

1:23:03

But, you know, how do you implement registers? How do address

1:23:05

buses work? Anyways, that's nandgame.com open

1:23:07

up both in the show notes, but two

1:23:09

useful resources. There's a modern

1:23:12

take on this game too called

1:23:14

Turing Complete. Oh, I didn't

1:23:16

want to pitch it because I haven't tried it. I

1:23:18

haven't tried it either. So I

1:23:20

haven't tried NANDtetris either, but it

1:23:23

has amazing reviews. I just

1:23:25

looked it up. It's overwhelmingly

1:23:28

positive. Steve reviews,

1:23:30

2,200 reviews. So

1:23:33

that is a really good endorsement

1:23:35

by the community. Yeah, I've heard

1:23:37

a lot of people say it's another good resource, so I'll put

1:23:40

it in the show notes too. Cool.

1:23:43

All right. Thank you, Patrick, for

1:23:45

going into a ton of detail

1:23:47

here. This is an awesome episode.

1:23:49

Thank you, Jessica, for recommending the

1:23:52

topic, phenomenal topic. And

1:23:55

thank you to all of our

1:23:57

patrons who are kind of supporting

1:23:59

the show. No helping us reach

1:24:01

new people and. I. Was

1:24:03

looking through my email the other day

1:24:05

about something kind of unrelated and I

1:24:08

stumbled upon a. Email. From

1:24:10

a listener. I'm. A.

1:24:12

While back, who are who we are able

1:24:14

to help them. You'll find their first job.

1:24:16

I mean that's it's really inspiring to Jesus

1:24:18

and that to us. We read all of

1:24:20

them such as read him twice and does

1:24:22

he find him years later? read of again,

1:24:25

On So thank you everybody out

1:24:27

there for listening for supporting the

1:24:29

south and and we'll see you

1:24:31

next time. I

1:24:45

know. More. Forgiving.

1:24:48

Third Down is distributed under a

1:24:50

Creative Commons. Attributes Sarah like to

1:24:53

point out allies: You're free to

1:24:55

share copy to services with the

1:24:57

work. The Remix adapt the work

1:24:59

but you must provide attributes and.

1:25:03

I and I sarah in

1:25:06

time.