We’ve lost the Tech

There are some engineering marvels out

there that end up living rentree in my

head for just years on end. And I I go

back and I think about how cool they

are, but there's a new one that's living

in my head. Uh not because of how cool

it is, but because of how ridiculous it

is. I am talking about Claude Code and

their Tui. Now, this tweet caught my

attention and has been just eating me

alive and now I have to make a video

about it. Okay, so let me just read it

to you. Most people's mental model of

Claude Code is that it's just a TUI.

That's what I thought it was at least.

But it's actually really closer to a

small game engine. Okay. Okay. Whoa. I

didn't realize rendering. I mean, I

guess if you if you really think about

it, you know, like if you really dig

down into the nuts and bolts, you can

kind of imagine that a screen every

single one of the characters is kind of

like a pixel, right? Which means it's a

game, right? Except for your bounds are

like 30 by 120 pixels. So, you know,

2400 pixels. We're talking about pretty

easy rendering going on here, right?

Like you could make a decision about

every single pixel many, many times,

right? Wo there, tough guy. For each

frame, our pipeline constructs a graph

from React, then lays elements out,

rasterizes them in 2D, diffs them

against the previous screen, and finally

uses the diff to generate an ANIE

string. If you're not familiar with

ANIE, ANIE effectively is the thing that

allows you a either to color text that's

on the terminal or b you can put your

cursor in places. Uh you can also clear

the screen with an escape sequence. Open

bracket 2J would effectively clear the

entire screen inside your terminal.

There's all these little things. That's

how ANI works, right? It's an escape

character plus an open bracket plus some

sort of code. Don't worry about ANIs.

It's really irrelevant in this situation

other than it just requires a bit more

memory to generate and that's it. And so

they have roughly 16 milliseconds frame

budget. And so we roughly have five

milliseconds to go from React to Ansie.

That means React takes 11 seconds to lay

out something on approximately 2400

pixels. We're going to talk about this.

Okay, I'm going to I'm going to first

off, I'm going to tell you some

engineering management takes that I

have. Okay, and I think they're very

very important. More so, I'm going to

give you some, hey, this is how I think

you should engineer things. And then

lastly, I'm going to show you that I

think whatever they're running up

against is a design of their own making

and a downfall of engineering in the

most truest sense. So, first off, as an

engineering manager, if I had someone

come up to me and say, "Hey, I'm going

to make a TUI and we have all these

constraints and it involves like say

frames per second." My very first

question would be like, "Oh, wow. Okay,

so you must have a bunch of stuff

changing on your scene to require uh

stuff to be rendering at that kind of

rate, right? Like what what's changing?

What's going on? Why do we have to have

a frames per second to ever even begin

with? I want to know. I want to

understand more." And if the answer

isn't we're writing a full a full-blown

game and every single key press and

everything needs to pass through a tick

and we need to have this really verbose

a beautiful way to handle everything. I

would go you got to stop. Something has

gone terribly wrong here. What are you

doing? When I hear this that claude

code, which if you're not familiar with

is just interacting with a model that

lives on the internet that takes seconds

to minutes to respond and leaves little

breadcrumbs along the way, requires 60

fps. This to me feels like a failure in

engineering management and also a

failure in engineering. Now granted I

don't work there. I don't know the

constraints they're working against.

Maybe just maybe there are things that I

don't understand that could make this

very very hard and maybe it's supposed

to make sense if one you know if I were

to understand the difficulty. So I'm

willing to concede that I don't

understand everything. If I were a

manager and I heard this I would say no

way. Back to the drawing board. Now as

far as engineering goes if I was an

engineer attempting to solve this

problem. First off, if I had to generate

text into a fixed box along with the

scroll history background, which is

apparently what they want, you you

should be able to scroll up and see all

your previous stuff. And you want to be

able to have some dynamic elements on

here along with some other stuff.

Obviously, I would have to set up a

situation in which I have some layers,

right? I have my dynamic side of things

right here, and then I have my actual

scroll back going on right here. And

then I have to construct these things to

be able to actually draw what's

happening on screen. Correct? But here's

the deal. I would never set up an FPS

for this. It makes no sense. And here's

why. You only have a few things that

cause a render. One, you have input. So,

when I start typing, I need to see my

characters going on to the prompt,

right? Okay. So, there we go. We have

our characters. For whatever reason,

Claude Code also wants a throbb, right?

That's what they call it. A big old

throber. Something that's like letting

you know like, "Hey, I'm rendering."

Like right here. See, if I say, "Tell me

how to talk to a woman." You can see the

throber right here. It's throberating.

Hey, Blood Code teaching you how to RZ,

just in case you're wondering. Okay,

treat her as an individual person, not a

category. Boom. Nothing says autism like

that answer. [laughter]

Hey there, fellow woman. I hear you like

flower. So, you have the throbb, which

needs a you know that that thing was

rendering like what three, four times a

second. And then lastly, you have output

coming in from Claude. That's really, as

far as I can tell, just looking at the

UI, things that require a change in the

scene to render. So, I'm not even

dogging on React, despite my feelings

about React. I'm just saying this is

what you see. Now, none of these happen

at effectively 60 frames per second.

People aren't typing 60 characters a

second. It's not throbbing plus typing

plus receiving clawed code back at 60

frames per second. So that means you

could make the little diff adjustments

to your screen as each one of these

things come in. There's no running game

loop. It should be completely empty if

nothing is happening. If you're waiting

for Claude, there's no throbbing going

on, no characters being typed. There is

no work to be done. There's no diffing.

There's no reacting. There's no there's

no rasterizing. There's no generating of

antsy. You can say no change, no work.

It's one of these really beautiful

things about writing 2ies is that you

can tell when something has changed.

Therefore, you know when work needs to

be calculated. And so in my head, this

is like what should happen. But

obviously, again, I still may not

understand things and I don't want to

dog on somebody for some engineering

stuff. Uh I I don't understand the

trade-offs they originally used. I don't

understand the React thing. But again,

I'm sure you can explain to me why React

was a good decision. To me, this just

obviously looks like, hey, I want to

write TypeScript. What do you write with

TypeScript? You write React. Okay,

that's what the LLM told me. Hey, I'm an

engineer. I use an LLM. Bada bing, bada

boom. We use React. Or this comes from

some sort of legacy decision that

somebody who thinks that they really

love Tuies, but they also only just want

to write TypeScript was like, yo,

there's this library called Inc. that

uses React, bro. Which is I think what

ended up happening. It doesn't really

matter how you got here. that decision

was made, but the 60 fps is really still

just living rentree because then I

thought, okay, how fast can I render

stuff to the screen? So, I wrote a

little tiny little test right here where

I can say I want to draw a box of one by

one characters and I want to draw it as

fast as possible. So, I'm drawing this

as fast as possible randomly on my

screen. This number right here is me

drawing a thousand of those boxes. How

many milliseconds did it take? So, I

drew a,000 one by one boxes in

effectively 1,000 milliseconds. Okay,

that's pretty cool. I do it again. Let's

do a 2x2. As you can see, I'm writing a

lot more characters. Effectively, no

time has changed. I can do that at

32x32. You can see the screen's going

nut. I bet you YouTube encoding is just

utterly failing at this point because I

have simulated white like white noise

from first principles. Again, I'm also

using Ghosty. This is a modern terminal

with GPU acceleration. So, some of the

older terminals probably couldn't render

as fast as this. I don't actually know

what the real frame rate is because I

don't know if we're actually rendering

at effectively one frame every 1.3

milliseconds, right? We're rendering

some amount at some rate that's very

very astronomically high to the point

that I can't even see what's really

happening on the screen. Just as far as

rendering goes, you can go really really

fast. There clearly is nothing holding

back from just going uh absolutely much

much faster than 60 frames a second even

in a just in a modern terminal. Now

granted again uh legacy terminals I

can't speak for them. I used to use uh

some terminal from 1984. It took a you

know everything was much much slower

when I tried to play Asky Doom. But then

I got curious. So this right here is me

seeing how fast can I render scroll back

history. Okay. So now if I close this, I

can actually take T-Mox and I'm look

like that's the scroll back history

right there. So this is much closer to

what Claude code does is it has an input

area where I can put in some nice input

going on down here and at the same time

you can imagine I'm getting so much

stuff coming back in. Now granted I

didn't use React. This is not nearly as

complicated. I I also included a throbb

that looks pretty similar and along with

like a little Braille little swirly

thing. This is how many new lines I have

created. This is how much of the time

I'm calculating I'm using the CPU in

JavaScript land approximately for every

every frame I'm using a 100 microsconds

and I'm almost idle the entire time. So

again even with this really ridiculous

example which is just you know rendering

at an absurd rate and doing a ridiculous

amount of stuff. I don't know where the

16 milliseconds is going wrong. It

purely seems like a React problem. And

don't just take my word for it. Will

McGugan, uh, if I'm saying that name

correctly, is kind of a longtime

engineer out here that has done a lot of

stuff and I think he has a really

amazing take on this. Sometimes

developers get hyperfocused on tackling

the immediate technical challenges in

front of them that prevents them from

reflecting on past decisions. By the

way, this is one of my greatest

weaknesses. And even after 20 years of

engineering, this right here, it still

rings true today. So much of engineering

is taking a step back and going, why did

I make this decision? Let's let's

actually understand the problem and not

just simply the solution. It's like you

take a fork in a road and find out you

need to climb a mountain. The faster

route to your destination may be to

backtrack and take a more direct route

without the mountain in the way. Tuies

are much more complex than most people

know. Very true. I've written a couple

tuities in my day. You know, I have some

amount of experience in it, but I've

never written a tuy to say to the level

of complexity of claude code. It is a

very impressive tuy. Trust me on this.

But I see nothing nothing in Claude's

output that needs a scene graph. Their

scene is a list of non-interactive

output. It's literally literally a list

of things. Granted, there's a little bit

of like highlighting, a bit of

throbbing, there's some input at the

bottom. They want to preserve scroll

history. Like I mean there's a couple

things. And apparently you can press

like controlB to like expand some stuff.

So there's some like accordion

interactive stuff, though I've never

gotten that to work. Cloud code does not

need to be a small game engine. No game

engine would construct a scene graph and

diff every frame. That would be too

expensive and wouldn't scale. I'm an ex

games developer. Yeah, typically you

just do immediate mode, right? You just

draw the scene, right? Big. You could

draw the entire terminal every single

time. How do I know? I'M DOING IT RIGHT

NOW WITH THE SCENE STORED as a list

tracking what changes rather than

diffing. The caching and see generation

16 milliseconds for a frame wouldn't be

challenging to hit. I don't even think

again I would argue from the beginning

there is no frame rate you need. You

react to changes coming in cuz they're

not going to be that fast. So, by the

way, just in case you're wondering with

input, if you did 60 characters per

second typing, that's 720 words per

minute. Okay? Ain't nobody typing that

fast. Even if I mash my little fingies

as fast as possible, I can't achieve

such speed. I think the mountain is the

React model, a holdover from their ink

days. It allowed TS developers to use a

comfortable pattern for building UIs.

But React was designed to manage the

complexity of a browser UIs, which are

way more complex than Claude's

requirements for a terminal UI.

Absolutely true. And now Anthropic are

expending a lot of engineering effort to

tunnel through this mountain. On top of

it, this also just goes to show,

remember this phrase that people always

say, use the best tool for the job. It

really just goes to show what people

mean when they say that is, I'm going to

use the most comfortable tool for the

job, which is the anti pattern. By the

way, I'm not even upset at JavaScript. I

think JavaScript's perfectly fine here.

I don't even care about that aspect.

Yeah, sure, it's a little bit memory

hoggy. Uh now with RAM becoming the most

expensive thing in the universe, I can

see why that's negative. But

nonetheless, your problem is not

performance. This is not a performance

problem going on here. If there is a

performance problem, it's a U problem.

Perhaps the renderer is too coupled to

the data model and hard to replace. Now,

I wouldn't be surprised if that were

true. But if I were anthropic, I would

have a small team work in parallel to

build a kind of decoupled model and

renderer that you would make if you

didn't have any legacy issues. I would

hire at least one games developer for

this. Honestly, you could hire a games

developer. Just hire a tools developer

in general, not even a games developer.

Hire somebody that doesn't just build

for the web. Really, just anybody else.

You could probably hire half the web

devs and they could just figure it out.

I really think this team would get to

their destination before the team that

is tunneling through the mountain.

Honestly, this is just so true and I

think it's such an interesting point in

engineering, right? Because Anthropic

has been the leading scaremonger about

like the world's going to end, all the

jobs are going to leave. You look at

this and I see a failure of both

engineering management and engineering.

They've got they've dug themselves into

the weirdest place. And it just goes to

show even if they generated all the code

with LLMs and they're saying this is the

best ever, fundamentally the decisions

they have made aren't good. And so

whether the code was generated by hand

or by LLM, it doesn't matter. The

failure already happened. Some more

highlights from Twitter. We apparently

live in a clown universe where a simple

TUI is driven by React and takes 11

milliseconds to lay out a few boxes in

monospace text. By the way, uh, business

right here, excellent blog, excellent

engineer. Layout 0.9 milliseconds.

Rasterization. Still, I still actually

just I've never read what rasterization

means. I'm sure it's really simple. It's

just in my head. I'm like, that's a word

I never know. 2.1 millconds. Diffing 1.2

millconds. Building the React scene

graph 11 milliseconds. I think this one

really is just like this encompasses how

I feel. This company is having a hard

time making a 2D ANC terminal to run at

60 fps and they demand you to take them

seriously when they talk about

revolutionizing programming. I keep

saying this and I feel like this is just

so important. hard skills are still in

demand. The LLM's aren't going to make

the best decisions. They're going to

make the decisions that you tell them to

make crappily. It's not only up to you

to make the decisions, it's up to you to

double check that the decisions you made

were even implemented properly and with

care. So, if you don't have that

foundation, you just might find yourself

in 60 frames per second react diffing

helland hellscape 2i. Lastly, I know

this tweet was made by Tyreek right here

and a bunch of people kind of been

dogpiling him. Hey, my goal wasn't here

to like call him out or single him out

or try to be mean to him. Hey, I've been

in positions in which I've had to make

decisions or follow up with past

decisions that weren't ideal. That did

cause a whole bunch of difficulty. I

just was lucky enough not to have to

make those decisions and explain myself

on Twitter. And I think a lot of you

listening have found yourself in a very

similar situation. So, be nice, be kind.

And all my critiques, as you can see,

are more towards the leadership of

Anthropic, which is their constant

fear-mongering of engineering, saying

it's all over. Because even within their

own company, they suffer from the cost

of legacy. There's nothing more

permanent than a temporary fix. The name

is the primogen. Hey, is that HTTP? Get

that out of here. That's not how we

order coffee. We order coffee via ssh

terminal.shop. Yeah, you want a real

experience. You want real coffee. You

want awesome subscriptions so you never

have to remember again. Oh, you want

exclusive blends with exclusive coffee

and exclusive content? Then check out

Kron. You don't know what SSH is?

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

[singing]

[music]

Living the dream.