HomeVideos

I Think They Are Lying To You

Now Playing

I Think They Are Lying To You

Transcript

354 segments

0:04

All right, people. It's time for a

0:06

passion yapping. I hope that you're

0:07

buckled in and sitting down because this

0:09

is going to be a bit intense. Now, I

0:11

want you to look right here. Do you see

0:12

this YouTube video? I'm not going to

0:14

play that YouTube video because

0:15

apparently someone has already been uh

0:17

DMCAD for showing any part of it. So,

0:19

I'm not going to have that happen. But

0:20

here's a quote from that video. Now it's

0:24

actually leveled up again to the next

0:26

abstraction where I don't prompt Claude

0:29

anymore. Now what what the heck are we

0:31

talking about? Well, this is Boris from

0:33

Anthropic, the man who started or

0:35

created Claude code talking about his

0:37

coding habits. Now 6 months ago, he

0:40

uninstalled his IDE and now he doesn't

0:43

even prompt Claude anymore. Okay,

0:45

prompting Claude is the old way we do

0:47

things. Now I set a goal. I let it run

0:50

in a loop. It burns just a cajillion

0:52

tokens until a win condition is met. And

0:54

bada bing, bada boom, we're done with

0:56

it. Okay, none of this hand coding. Ew,

0:59

gross. None of this prompting. Also ew

1:01

and also gross. Okay, we loop. And the

1:04

quote goes on. I have loops that are

1:07

running. They are the ones that are

1:09

prompting Claude and kind of figuring

1:11

out what to do. My job is to write

1:14

loops. And this is the kind of next

1:16

transition I think we're going to see in

1:18

the next few months, maybe through the

1:20

rest of the year. Now, this quote was

1:22

said and everybody, wow, deep insight.

1:26

And everybody kind of clapped about it.

1:27

And if you haven't been following

1:29

Boris's journey, he's also said phrases

1:31

a little while back, coding is mostly

1:34

solved. Now, he just says coding is

1:36

solved. And even more, he says coding is

1:39

the easy part. The hard parts like

1:41

infrastructure or hardware or uh

1:43

listening to users or getting feedback.

1:45

Those are the hard parts. But coding

1:47

easy peasy. It is solved. Lemon squeezy.

1:51

Not a big issue. Now I take a little bit

1:54

of issue with that. Okay. In fact, I

1:55

would like to say that I take well a lot

1:58

of bit of issue with that. And let me

1:59

kind of build up a case and I'm going to

2:01

explain to you why I think they're all

2:04

lying. I think they are all lying to

2:06

you. And I know that's going to kind of

2:09

ruffle some feathers. Well, first off,

2:10

you need to see this tweet right here

2:12

cuz it's very important as part of this

2:14

entire concept. This is Anthropic

2:16

saying, "Hey, we've greatly accelerated,

2:19

okay? We are faster. We are stronger.

2:22

We're better than we've ever been

2:23

before. In fact, in this last quarter,

2:26

Q2 2026, we are shipping 8x the amount

2:30

of code per employee than we were on the

2:34

average before 2025." In other words,

2:37

they are shipping two years worth of

2:39

code every single quarter for every

2:42

single employee. Hence the reason why

2:44

coding is the easy part. Coding is

2:45

solved, bro. We don't we don't write

2:47

code no more. Okay? We prompt the

2:49

prompters. All right? Big difference. We

2:51

write loops. We do not code ourselves.

2:56

Review code. It's all done

2:58

automatically. Hey, me from the future

3:00

here. So when I say they are lying, what

3:03

I mean is either A they are

3:05

intentionally lying to you or B they so

3:08

believe what they say that even in the

3:12

face of contrary evidence they go now

3:14

I'm correct. I no I am right it doesn't

3:16

Nope. The sky is in fact not blue. No

3:18

it's not will not accepted. Tell me I'm

3:21

wrong. I don't think I am. All right

3:22

back to past me. By the way I get even

3:25

more passionate. So you're welcome. I

3:27

have been using linear lately. And look

3:28

at this. Okay. I just clicked on a

3:30

ticket. It loaded. I go back. I go

3:32

forward. Look at how fast this is. I can

3:34

drag them between columns. Absolutely

3:36

creamy smooth. I know this is unusual.

3:39

This is not what you expect from

3:40

software, but this is some of the best

3:42

software I've been using, and I use it

3:43

to track all the things I am working on,

3:45

and you should, too. You should try out

3:47

Linear at linear.app. Now, to the main

3:49

thesis. I think they are lying, and I

3:52

think that they're hurting people. I

3:54

think they're genuinely hurting people.

3:55

First off, I get lots of messages about

3:57

people genuinely worried about the

3:59

future and hey, maybe there's no room

4:01

for me left. Hey, I think everything's

4:02

going to be destroyed. People's jobs are

4:04

becoming this really frustrating

4:07

experience where they just have people

4:08

yeeting code into production. There's

4:10

absolutely no controls because this is

4:12

what they've been told to do. Run the

4:14

loops, code the coders, just let it run.

4:16

Let it fly. And so, just even on a

4:18

personal level, people are feeling

4:20

really burnt out and freaked out. At

4:22

least that's the general gist that I'm

4:25

getting from a ton of people. This has

4:27

not been so good. And a lot of people's

4:30

well-being and how kind of their outlook

4:31

in life has actually gone very negative.

4:34

I think what anthropic and more

4:37

specifically Boris is doing is very

4:40

destructive. Now, I know I'm going to

4:42

get, "Hey, man, Boris, he's a he's a

4:44

nice guy." Okay? Like, in person, very,

4:46

very nice. And you know what? I'm sure

4:48

you're right. In person, humble and

4:49

nice. But here's the deal. humble and

4:52

nice people can do bad things. Now, let

4:55

me make the case to you why I think that

4:57

they are lying and I think it's very

5:00

very obvious. So, the first thing you

5:01

know to understand is that Claude code

5:04

was released for researchers in February

5:06

2025. Now, if you're unfamiliar, Claude

5:09

Code is a little terminal application in

5:12

which you can prompt Claude to go and do

5:14

some angentic work and produce code. All

5:16

right. Well, within two weeks of

5:18

releasing that GitHub issue 392, hey

5:23

bro, screen's flickering. A little bit

5:25

afterwards, April 11th, 2025 in

5:29

progress. Screen's flickering,

5:31

everybody. And this continues on. 1913,

5:34

terminal flickering, and there's like so

5:35

many other issues devoted to terminal

5:37

flickering. It's almost as if Anthropic

5:41

thought people needed a little bit more

5:42

epileptic training so that they can

5:44

watch their favorite animes. But

5:46

nonetheless, this is a very well-known

5:48

issue and it was reported almost

5:51

immediately upon research release.

5:54

Meaning that this bug most likely

5:56

existed not even in 2025, potentially in

5:59

2024 while it was all being developed

6:01

and they actively knew about it is my

6:04

personal guess. If you used the product

6:06

at all during that time period, it was

6:08

so evident. I remember having my own

6:10

experiences with it, being shocked

6:12

people even used the product at all.

6:14

Well, let's fast forward a little bit.

6:16

They finally respond to it publicly

6:18

December 17th, 2025.

6:22

And I quote, "We've rewritten Claude

6:24

Code's terminal rendering system to

6:26

reduce flickering by roughly 85%." Now,

6:29

first off, that's impressive. I have

6:31

never ever in my lifetime seen a bug fix

6:35

that only works 85% of the time. Like,

6:37

what kind of stochastic process are is

6:40

going on out there? Like, that is

6:42

insane. And then they go on to explain

6:44

how difficult it was. They're building a

6:46

game engine brothers to render some text

6:48

to a terminal. This was said out loud

6:52

couple days later, December 18th. We're

6:55

rolling back a few changes we shipped to

6:56

Claude Code this week to make sure

6:58

things are stable heading into the

6:59

holidays. In other words, there, hey,

7:01

coding is largely solved. Turns out it's

7:04

not largely solved. Now, you're probably

7:06

wondering, why am I harping so hard on

7:07

this screen flickering issue? Of course,

7:09

being known for like nine months at that

7:10

point, why am I harping so hard on it?

7:12

Well, here's the thing. Screen

7:13

flickering, that's a pure software

7:16

problem. There's no hardware involved in

7:18

this. Okay? There's no, "Oh, the

7:20

infrastructure wasn't quite quite right.

7:22

Oh, the capacity. We just didn't have

7:24

the capacity for non-flickering terminal

7:26

experience." No, it's the user feedback.

7:28

We just didn't have the feedback. We

7:30

didn't even know users didn't like

7:33

flickering. No, they knew. They were

7:36

well aware. They probably even

7:38

prioritized a flick flickering

7:40

considering they also did a whole

7:42

rewriting of the rendering system. But

7:44

ultimately the first swing, the first

7:46

rewrite failed failed epically failed

7:49

right next to Christmas and they had the

7:51

poll out. And so hey, is software

7:54

largely solved? I'm not buying it

7:55

currently. But here's the thing is if

7:57

the story ended there, maybe we could

7:59

move on, but it doesn't. Okay, March

8:02

25th, 2026. Now we are officially over a

8:06

year into first known reported issue on

8:10

GitHub. Over a year. Just let that sink

8:13

in. Over a year of the terminal

8:15

flickering. Text characters laid out in

8:19

a grid. Couldn't help but to flicker.

8:24

I

8:25

do. I feel crazy that this is a crazy

8:27

hard problem. No. This cannot I refuse

8:30

to believe that this is that hard of a

8:32

problem. Well, guess what? It was

8:34

finally asked, "Please fix the

8:36

uncontrollable scrolling/flickering

8:38

before the next 3,000 features." People

8:41

were sick of it. And on April 1st, not

8:44

as a joke, Boris releases no flicker

8:46

mode for Claude Code. This is that new

8:48

rendering that they talk about. It's

8:50

called Claude Code, no flicker. I

8:52

believe it takes advantage of alternate

8:54

screen as opposed to doing this direct

8:55

print down because the direct print down

8:57

does take a little bit more effort to

8:59

get right. Whereas alternate mode,

9:00

that's like how Vim works, right? It's

9:02

beautiful. I I love I love Vim. Okay.

9:05

Hey, we're talking about Vim. I could I

9:07

could I could sing some praises about

9:09

Justin and his running of Neo Vim. But

9:11

back to the thing. I want you once again

9:13

to think about this. This was only two

9:16

months and a little bit of change ago

9:17

and they still haven't fixed flickering.

9:20

It's been over a year and they have a

9:23

feature branch or a feature flagged

9:25

feature in which solves flickering via a

9:28

completely different path. Now, I want

9:29

you to think about that. If coding is

9:31

solved, do you need feature branches?

9:34

No.

9:35

No. No. No. No. You don't. Somebody's

9:39

lying to you. Because if coding was

9:40

easy, if coding was solved, no, you

9:42

don't need that. Who is telling the

9:45

truth right now? Because it does not

9:46

seem like Anthropic or Boris is leading

9:49

you to the correct path. Now, I don't

9:51

know if they're actually actively lying

9:53

to you, but in my head, I feel like they

9:56

are. To me, it seems apparent that I

9:59

think they are liars. Now, lastly, we

10:02

get all the way up to May 27th with

10:04

Claude Dev's official account even

10:07

saying, "Hey, guess what? First off, I

10:09

just want to let everybody know. New

10:12

terminal dropping, right? This is the

10:14

official Claude Dev account saying,

10:16

"Hey, we're doing it. We're dropping the

10:18

big deal." But even more so, if you

10:19

scroll a little bit down, you'll see

10:21

this fewer mysterious error messages.

10:25

They have error messages in a product

10:28

that they're just like, "I don't know.

10:33

I guess it's not working right now.

10:35

We're not really sure what the heck's

10:37

happening." Are you kidding me? Yo, bro,

10:39

why don't you launch a loop? Why don't

10:40

you launch a loop like 6 months ago?

10:42

Bro, write a loop and fix that. Don't

10:45

have a mysterious error. Let people

10:47

know. Oh, whoopsies. Turns out the

10:49

connection terminated. You would you

10:51

like to try again or would you like to

10:53

start over? Oh, hey. Whoopsies. It turns

10:57

out that you're out of credits. Sorry,

11:00

you can't continue. Like, there's only

11:02

so many problems here. Write out the

11:05

problems and the errors. Loop it. Loop

11:08

it, brother. Anyways, I say all this

11:10

because even as I'm confused about this

11:13

entire thing, Boris even responds to me

11:15

saying, "Coding is the easy part.

11:17

Everything else is not yet solved, but

11:20

is also becoming increasingly

11:22

automated." saying to me, "Coding is

11:24

solved.

11:26

It's the easy part." And yet during all

11:29

of this, it just feels like they have an

11:32

abundance of obvious failures. And if

11:34

you just look at their clawed status,

11:36

their claw status, it's like at 98 some

11:39

percent up and they're having just

11:41

errors elevated with this model, errors

11:43

with this model. Like they're not like,

11:45

"Oh, we ran out of capacity. Sorry,

11:48

overloaded with requests." No, it's

11:50

like, bro, we actually like there's

11:52

actual errors. Not only that, but for

11:54

the last year, you can see GitHub

11:57

issues, Reddit posts, hacker news posts

12:00

saying, okay, I was prompting and it

12:02

looks like I got somebody else's

12:03

feedback or I got somebody else's prompt

12:06

result cuz I'm getting like a legal

12:07

document, but um there's none of us are

12:10

talking about legal documents. What the

12:12

heck's going on, Claude? And there's

12:13

this whole thing where people are

12:14

accusing them of leaking other people's

12:16

session. And I'm not even saying they

12:18

are or they aren't or that this is some

12:20

crazy hallucination case, but they can't

12:23

and don't know how to solve that issue.

12:25

They don't even really know what's

12:27

happening with that one because it

12:28

continues to happen. And so it's like,

12:30

yes, is coding hard? Sure, coding can be

12:33

hard. That's okay to say. No, nobody's

12:36

going to be upset about that. Have we

12:38

accelerated the rate in which we can

12:40

produce code? Yes. Have we accelerated

12:42

the rate in which we could produce more

12:44

correct code? Do we have tools in which

12:47

can help people write code? Sure. Yes.

12:50

Beautiful, glorious, knock yourself out.

12:52

But this idea that everybody should be

12:54

writing loops, everybody should be

12:55

spending $10,000 a day on my company

12:58

salesman selling his items and just

13:01

telling everybody this is the way it

13:02

should be when they get their tokens

13:05

free and unlimited while at the exact

13:07

same time making demonstrabably false

13:10

claims. Drives me bonkers. And I just

13:14

had to talk about it. because they in my

13:16

opinion they're lying to your face. They

13:18

are peeing on your leg and saying it's

13:20

raining. I can hear it's raining right

13:22

now. That's why I said that. Anyways,

13:25

that's the video. That's what I wanted

13:27

to say. Hey, the name is the primogen

Interactive Summary

The video presents a critical perspective on the claims made by Anthropic and Boris, particularly the assertion that 'coding is solved' and that developers should rely on autonomous loops instead of manual coding. The speaker argues that these claims are misleading and potentially harmful, citing persistent technical issues with the Claude Code tool—such as terminal flickering and mysterious errors—as evidence that coding remains a complex and unsolved challenge. The speaker questions the integrity of these claims, suggesting that they are being used to push a narrative that prioritizes automated output over stable, high-quality software development.

Suggested questions

3 ready-made prompts