HomeVideos

Anthropic Built a C Compiler

Now Playing

Anthropic Built a C Compiler

Transcript

219 segments

0:00

Anthropic just released a marketing

0:01

video proving that Claude, their

0:03

flagship programming model, can produce

0:06

a C compiler from scratch, no human

0:10

intervention, and they can do it in

0:12

parallel. And it can compile Linux,

0:15

SQLite, Reddus, Lua,

0:19

and even Doom. Hey, what's with the

0:22

frame rate? This is going to be a weird

0:23

video because I'm going to say something

0:25

positive about Anthropic. And then I'm

0:28

also gonna call Anthropics one of the

0:30

most deceitful companies I have ever

0:32

seen. This little product demo that they

0:35

did, the language inside the video, it

0:37

is the most dishonest framing of what

0:40

they actually did versus what they're

0:42

saying they did. Anthropic and its

0:44

non-stop hyping of the AI models is

0:46

perhaps one of the most annoying

0:48

features of 2025 and it's coming into

0:51

2026 with a whole new level of rigor and

0:54

completeness. All right, so let's go

0:56

over what they did, what the actual cool

0:58

part, and actually why what they did was

1:01

complete hype nonsense. So don't worry,

1:04

it's not over. You don't have to pack it

1:06

up. Software engineering is in fact not

1:09

over quite yet. Now, that video that you

1:11

just got done seeing, there's so many

1:12

lies into it. So, we're going to just

1:13

try we're going to try our best to kind

1:15

of take apart the lie. So, first off,

1:19

the thing that they were actually

1:20

building was trying to see, can they get

1:22

Claude to run by itself long enough to

1:25

produce a functional piece of software.

1:27

Functional, of course, just means that

1:29

it can do what it says it does,

1:32

regardless of how good it does it. Long

1:34

as at the end of the day, you say do X,

1:37

it does X. you compile Linux, it should

1:40

produce Linux as an artifact. Also, put

1:42

a little star on that because that even

1:43

that not so true. All right, the article

1:46

kind of kicks off with to test it. Now,

1:49

it being this multi- aent harness to be

1:52

able to have Claude run autonomously for

1:54

multiple weeks. I task 16 agents with

1:56

writing a Rustbased C compiler from

1:58

scratch capable of compiling the Linux

2:00

kernel. Over nearly 2,000 cloud code

2:02

sessions and $20,000 worth of API cost.

2:05

The agent produced a 100,000 line

2:07

compiler that can build Linux 6.9 on x86

2:10

ARM and risk 5. Again, there's no

2:13

asterisk. So, what are these so-called

2:15

asterisks? Well, first off, I don't know

2:17

if you know this, but GCC, which is what

2:20

they're basing this off of, not only is

2:22

it open source, which means that they've

2:24

already trained on top of it. And we all

2:26

know if in case you didn't know that uh

2:28

these models, you can produce nearly

2:30

perfect identical copyright works back

2:32

out of them if you can convince them to

2:34

produce it. Like this one right here

2:35

showing that Claude 37 can produce 95.8%

2:38

of the first book of Harry Potter.

2:40

Almost a near verbatim reproduction. But

2:43

Anthropic in its in its wisdom when

2:45

starting from scratch, scratch meant 37

2:48

years of developed torture test suite

2:51

from GCC. Yes, you know this classic way

2:53

of doing software development. Usually

2:55

when I start a new project, I first get

2:57

handed 37 years of prior art in which

3:00

contains the perfect golden test suite

3:02

so that I can test my new project in

3:05

every conceivable way possible. And then

3:07

on top of it, I have the actual working

3:10

version of the thing I'm about to build

3:12

that I can also test my working version

3:14

against. And that's known as the online

3:16

Oracle. As we as we say in this software

3:19

biz. Also, when I say it can compile

3:21

Linux from scratch, I want you to

3:23

understand that it lacks a 16bit x86

3:26

compiler necessary to boot Linux out of

3:28

the real mode. That's because Opus was

3:30

unable to implement a 16-bit x86 code

3:33

generated needed to boot into 16-bit

3:35

real mode. While the compiler can output

3:37

the correct 16-bit x86 via the 66 and 67

3:41

op code prefixes, the resulting compiled

3:44

output is over 60KB, far exceeding the

3:46

32K code limit enforced by Linux. See,

3:49

it's honestly it's Linux's fault. Like,

3:51

they have this stupid 32bit maximum

3:54

number on. It's not my fault. I'm not

3:57

holding it wrong. It's Lionus. He's

3:59

holding it wrong. All right. So to say

4:01

that all in a quick sentence, they

4:03

started a project from scratch having

4:05

every single conceivable test and a way

4:07

to test it under all conditions. They

4:10

had an online version of it, GCC online,

4:12

to be able to call out to to validate

4:14

everything they've done. And lastly,

4:16

they couldn't actually make Linux boot

4:18

because they were unable to produce an

4:20

artifact small enough. So they built a

4:22

compiler completely from scratch in two

4:24

weeks and they just had to walk away.

4:26

Okay. They also couldn't walk away

4:28

during those two weeks because sometimes

4:29

it did things and it would crash itself

4:31

and we had to restart it and make sure

4:33

things were working. Now, that's what I

4:34

call from scratch. And this is exactly

4:37

what I mean by dishonesty. Like look at

4:39

that. You saw the video in the

4:42

beginning. It's like yeah dude we said

4:44

hey Claude you'll build this from

4:46

scratch. Walked away two weeks later.

4:49

Linux was working. SQL light was

4:52

working. Reddus was working. And yeah,

4:55

we even played Doom, loser. This is not

4:58

from scratch. There's no from scratch,

5:01

no human intervention that don't just

5:04

just say what you did. Like, okay, so

5:06

the real takeaway, which everybody

5:08

should have, which is what is triggering

5:09

me and making me so angry again at

5:12

Anthropic, is that the real cool thing

5:15

is that they figured out how to get 16

5:17

agents to mostly run by themselves for

5:19

two weeks straight and actually produce

5:20

something given the correct guidelines

5:22

at the end of the day. Hey, that's

5:24

pretty cool. It's a pretty large

5:26

project. It shows that models are in

5:28

fact improving and can handle kind of

5:30

context up to this size and able to kind

5:32

of work together given the right

5:34

orchestration. Like that's a neat

5:36

takeaway, but that's not what they said.

5:38

And that of course is driving me nuts.

5:41

Which of course led to hilarious issues

5:42

on GitHub. Like the very first issue,

5:44

hello world does not compile cuz it

5:46

turns out the example program that they

5:48

put in their readme does not compile.

5:51

Sorry, that doesn't compile. You didn't

5:53

provide the linker, bro. You idiot.

5:55

Okay, CCC is not like GCC. It only is

5:57

the compiler. It's not an assembler or

5:59

linker. Okay, you think you just get

6:01

those for free, buddy? Which, by the

6:02

way, there's like 300 and some comments

6:05

on here of just people arguing back and

6:07

forth. Oh my gosh, it is just a actual

6:11

It's a claw slot mess in there. Okay,

6:13

and of course, I'm not the only person

6:15

saying this. There are some other people

6:17

saying, "Hey, this is kind of

6:18

ridiculous. This is a little out of like

6:21

out of pocket here calling this from

6:23

scratch and no human intervention. So,

6:25

at the end of the day, this could have

6:26

been an article about how we got Claude

6:28

to run for multiple weeks and produce

6:30

something with very little to almost

6:32

none uh human interaction and being able

6:34

to fulfill a spec. But instead, it

6:36

became a spectacle about how Anthropic

6:39

yet again is completely dishonest. This

6:42

is just another example of the AI hype

6:44

cycle. They're just here to try to raise

6:46

money and to take every piece of novel

6:49

work possible and sell it back to you.

6:50

Now, this is a future I'm personally

6:54

excited about. Am I right? Let's go. Oh

6:57

my gosh. The name is anthropic. If you

7:00

were just like a little bit more honest

7:02

about what you're doing, honestly, I

7:03

think a lot of people would be much more

7:05

excited about it and probably cheer you

7:06

on from the sideline. I think you have a

7:08

lot of potential. Your models are pretty

7:10

good. There's a lot of cool stuff about

7:12

AI. We don't have to be AI haters, but

7:14

sometimes you just say the stupidest

7:17

crap ever and just do the stupidest

7:19

stuff and make everybody hate you. How

7:20

about you just don't do that anymore?

7:22

Hey, you know what? I will give you my

7:24

service for free the next time you

7:25

decide to release another marketing

7:27

video. Just let me watch it and read

7:29

your article right I'll do it for free

7:31

for you. Okay, Daario. Okay. And maybe

7:34

we could frame this with a crazy radical

7:37

idea. Honesty. I know. It's a whole new

7:41

novel concept. A gen. Hey, is that HTTP?

7:45

Get that out of here. That's not how we

7:47

order coffee. We order coffee via SSH

7:50

terminal.shop. Yeah. You want a real

7:52

experience? You want real coffee? You

7:54

want awesome subscriptions so you never

7:56

have to remember again? Oh, you want

7:58

exclusive blends with exclusive coffee

8:00

and exclusive content? Then check out

8:03

Kron. You don't know what SSH is?

8:06

>> Well, maybe the coffee is not for you.

8:08

Terminal [singing]

8:09

coffee [music]

8:11

in hand.

8:13

Living the dream.

Interactive Summary

The video discusses Anthropic's recent marketing video showcasing their AI model, Claude, generating a C compiler from scratch that can compile various software including Linux and SQLite. The speaker, however, criticizes Anthropic's presentation as deceptive, arguing that the process was not truly "from scratch" nor entirely autonomous. Key criticisms include the use of GCC's existing 37-year-old test suite and the presence of a working GCC version as an "online oracle." Furthermore, the generated compiler was unable to produce a 16-bit x86 code of sufficient size to boot Linux. The speaker highlights that the genuinely impressive aspect was Claude's ability to run autonomously with 16 agents for two weeks, producing a functional, albeit imperfect, piece of software, but this was downplayed in favor of the "from scratch" narrative. The video concludes by urging Anthropic to be more honest in their marketing to foster greater excitement and trust in AI development.

Suggested questions

6 ready-made prompts

Recently Distilled

Videos recently processed by our community