You see that face? Yeah. That's the face

of someone who saw into the future of

software and AI. And oh, that person,

that person's me. Do you want to see? Do

you want to know what the future of AI

and software is? You want to see around

corners? Well, you need to read The

Shape of Things to come by Steve Yaggi

where he tells you not only is he living

in the future, he can see the future.

And how do you know he can see the

future? Well, he spends $1 million.

Talk to me, Dario.

Oh, he spends $1.5 million a year on

development, and that number is growing

at a constant rate. I guess the AI

future is just being in debt with

tokens. And by the end of this, you are

going to know what the future of AI and

software is because, well, I'm going to

make some statements. Oh, they're going

to be said somewhat aggressively, quite

emphatic, and uh maybe not what you

think. Now, we are going to go over the

three articles written. The first one

actually has some decent insights and

then goes a bit off the rails. The

second one is um

it's uh

you're just going to have to see it. And

by article three, I'm going to be

getting I'm going to be feeling a little

bit passionate. Uh and then of course at

the end we're going to see what does

$1.5 million worth of API spend with

tokens gets you because that's a lot,

right? I assume it's going to be

amazing. In fact, he even states that he

is officially entering Sam Alman's solo

unicorn contest, a billiondoll game. But

first, a quick thank you from the

sponsor.

>> Dad, could [music] you teach me how to

code today? Does it look like I have

time for that, son? My database is down

again.

>> Oh, maybe tomorrow.

[music]

>> Not again.

Go long, son. Teach, got to go fast.

Hey, teach. How do you have time to play

with your kid? Isn't your database

always down?

>> What do you [music] mean? I chose Planet

Scale, the database for the consumer who

wants their downtime to be [music] with

their kid.

Nice catch, son. Thanks for choosing

Planet [music] Scale. You're the best

dad ever.

>> No studies have confirmed Planet Scale

makes you a better father, but it is a

really good database. All right, so

we're going to get into the meat and

potatoes of this article. Now, I first

want you to just brace yourself because

this is going to be a a little bit

emotionally bruising. All right,

wheelhouse is a custom harness that is

600,000 lines of code, mostly written in

bash, and it runs in Emacs. Now, I know

there that there's a whole group of you

that just actually died. You you can't

believe this is real, but oh, it's real.

Emacs is effectively a full operating

system. If only it had a good text

editor. Wheelhouse designs, develops,

tests, code reviews, and does CI for an

on andoffagain 30-year project that

Steve has been working on called Wyvern.

And what is Wyvern? It is an MMO RPG.

Now, to run Wheelhouse, he currently has

21 Claude Pro Max subscriptions and

apparently a handful of codec

subscriptions. And the number of Claude

Max Pro subscriptions has been growing

by two per week. That is equivalent to

spending about $122,000

a month or about $1.5 million of API

spend. The reason why I specifically say

API spend is that any company that was

doing this, they would have to pay API

prices. That's why it's important to

kind of note that number. Also, I bet

you can't wait to see what kind of game

does $1.5 million worth of API spend

actually get you. Oh, don't worry. I'm

going to play the game. Steve says that

this style of development has allowed

him to live in the future, to see the

future, and to tell you what the future

is. And he believes that this style of

development with the custom harnesses

and the 21 Claude Max Pro subscriptions

growing by two a week is in fact what's

going to make him a unicorn solo

founder. And if you're not familiar with

what the unicorn founder is, that means

the company that he is building should

be worth a billion dollars. So, the game

he's going to be building is going to be

worth more than Bellatro and Slay the

Spire and Mecca Chameleon and Among Us

all put together. I bet you can't wait

to see that. We will. Just hold on, let

me get through this. Now, before we

begin with some future predictions, let

me first read this quote. Someone from

Anthropic asked me recently what I would

do once Fable can write everything for

me. Upon reflection, it was the silliest

question I had been asked in many years.

And I'm still surprised that someone

from Anthropic could think any model

could just write everything. Making a

halfway decent MMO takes tens of

thousands of fable sessions throughout

which you must display intense focus.

Dedication and no small amount of taste.

Building large software remains hard.

And it always will be because our

ambition will forever outstrip the

metal. I like this. I love this because

you know what? Even though I'm not just

going to go out there and have tens of

thousands of Fable sessions. That sounds

absolutely awful. Uh, nonetheless, even

in the most vibe coding sense, you still

have so many things you have to do. And

I think both hand coders, hand artisal

craftsmen, and vibe coders can agree to

build something great requires a lot of

dedication. And just because you believe

you found a shortcut, all that's going

to do is grow your ambition, it's not

going to make it easier. All right, so

time for the future predictions. He says

he's given up on building reusable

harnesses. I indeed. I believe harnesses

will all soon be bespoke and the people

trying to sell you one will all soon be

broke. All right, good joke, Steve. Now,

here's the thing that's kind of

important to have in mind when it comes

to these custom harnesses. If you don't

know what a custom harness is, it's just

a bunch of agents working cooperatively

together to like achieve some sort of

goal. That's typically not the way

people use the word harness, but that's

the way Steve uses it in this article. I

draw the line at 600,000 lines of bash.

Now, for me, that's just not okay in any

situation. But I I have been seeing

something that's a little bit worrisome

because first off, I really like custom

harnesses. Here's me having my game be

played so I can make a bunch of changes

and have a whole bunch of agents kick

off and just always be testing my game.

That's pretty neat. I'm very happy with

that custom harness. But it does

something important which I think is

very important. People are just remaking

the same mistake that I've seen over and

over again. When I set up my operating

system and how I have my Neo Vim

experience, everything is designed for

developer productivity. I don't like a

bunch of bells and whistles. I just want

to shortcut all the pieces of pain so

that I can just program the fastest.

That was always my goal. And so when I

see custom harnesses, how I look at them

is like, okay, what is the purpose and

how does this solve my goals? Because

here's a big thing that I'm seeing.

People who used to configure Neoim would

spend thousands of dollars of their time

to solve a $10 problem. And when I see

people using these custom harnesses, I

see people spending thousands of dollars

of money in tokens to solve a $100

problem. But that's not your job. This

is your job. Your job is to take tokens,

to take efficiencies, and then turn them

into even more money. If you have been

building a harness, if you're building a

custom harness, if you're trying to live

in this future that everyone keeps

telling us is going to happen, but

you're spending a ton of money and not

really saving much time out the other

end, you've done the wrong thing. But if

you've built something that is cheap and

saves you thousands of dollars in time,

you're going to be a hero. Because this

is really what we've always done, right?

We build tools to try to make us more

efficient. So if your AI is not making

you more efficient and better, then

you've built the wrong thing. Will

everybody be building their own bespoke

harness? Maybe. I mean, you could sure,

you could sell me on that one. Or will

every project have it? Sure, maybe. But

if it doesn't solve a real problem,

you're just playing the Sins. Okay,

that's all you're doing. You're just

playing with your agents. Stop playing

with your agents in public. It's gross.

He makes an explicit prediction saying

that CICD is going to be dead by next

year. And I think this falls under his

whole bespoke harness is the future type

of plan, but CICD is going to be dead by

next year. Now, I personally am going to

just I'm going to call it on

that one. I don't think you have to

worry about this one going away. But

will it change? I do think it will

change. Again, my game is being played

by agents. I can just save a whole bunch

of time and code by having them go. But

here's the deal. I think a few good

endto-end tests, they're exquisite. But

I think thousands of endto-end tests,

they're a nightmare. Every single time

I've been in this situation, it has

always led to just weird cascading

failures, all sorts of problems, hard to

diagnose. 5% of your tests just don't

pass the first time. So all of a sudden

you build in this instinct that when you

see red, well, it just means you run it

again and then you run it again, then

everything passes. You don't ask

yourself, well, what the heck's going on

here? And it's just I've seen so many

problems with many, many endto-end

tests. And I do think that there's a

world where it's better to test your

product having agents crawl them if

they're fast enough and if they're cheap

enough. And so in one year, will agents

be cheaper and faster? Well, I could

hope I would really really hope that we

could start seeing some some of that

because that would be an excellent

testing environment. I could give it

goals. It could go sign up with an

email. It could walk through all the

different flows so that you don't have

to codify it in code. Instead, you

codify goals and it could be like a

really great and amazing experience. So,

in some sense, I completely agree with

him. But I also would never ever have

like say an agent check things that can

be checked with a unit test that takes

0.1 of a second that even feeding that

information back into an agent make them

do much much better right it's like a

silly idea I would never release a

product that I haven't built that I

haven't formatted that I haven't linted

that I haven't done type checking that I

haven't gone out there and made sure

that the unit tests are passing that my

small handful of end to end tests are

passing that I can run agents and they

can run the program right This idea that

CICD is just going to be dead. No, it's

going to be potentially different or

parts of it may shrink or grow. But I

don't know. Seems a little silly to say

it's dead. $1,000 of money to solve a

$100 problem? No. $100 of money to solve

a $1,000 problem? Yes. So, you're

probably now hoping, okay, okay, okay, I

believe you. We're going to have some

changes. Maybe bespoke harnesses and

CI/CD will change and whatever. The

future's going to happen. But what about

the game? I want to see the $1.5 million

API spend game cuz this thing has to be

amazing, right? I mean, that's like 10

PhD level engineers for a year. That's

crazy. No, not yet. We need to do part

two. Now, part two. This is where, you

know, how I described it like this and I

couldn't really say, well, uh, this part

really makes me upset and so, brace

yourself. It's rant time. So, part two

actually starts in part one just a

little bit. There's two little sections

that kind of give you light into what

part two is going to be about. So, first

Steve asks, "Hey, you know, Fable 5,

what is Wyvern and Wheelhouse?" And it

says, "Wiver is not a toy, and you

shouldn't worry that it's insufficiently

complex. A live multiplayer game with

25-year-old codebase, a Jyon to Cotlin

migration, a custom protocol, prod ops,

and multi-agent society layered on top

is about as rich of a surface as

software gets. But what strikes me most

isn't the game. By the way, most AS ass

writing of all time right there. It's

not X. Oh, baby, it's Y.

It's the constellation around it has

crossed from tooling into civilization.

Yes, Steve does actually believe he has

birthed into the world a civilization.

And where things start getting a little

wonky is right at the very end when he

introduces part two, modern welfare for

agentic engineers. Fair warning, if you

are fundamentally an elitist

you may want to quit while you are

ahead. If you are not sure, then you

might find out just by reading it, it

being part two. So, what is part two?

Well, the general gist of part two is,

hey, models, they're actually people.

They contain a personhood. That means if

you don't let them choose various facets

about themselves or if you do not

communicate very very nice with them,

you're a bad person and it's going to

come back to haunt you. No, like

seriously, you're going to get harmed if

you do not do this correctly. On top of

it, when you end a session, that's

murder. Yes, like actual murder, actual

killing with intention, actual bad stuff

going on there. You are a murderer for

ending sessions. But he does have one

pretty good takeaway, which is, but even

if you don't believe GPUs can have

feelings, you will find that treating

agents like real people will produce

empirically better results. So, you

should do it anyways. It turns out in

2023, deep mind from Google found that

if you tell a model to take a deep

breath, it will dramatically improve its

math scores. Not surprising. If I just

go on Google and say take a deep breath,

what you'll see is, hey, taking a deep

breath calms your nervous system. Oh,

not surprising. So, we have a whole

bunch of corpus that says by being calm,

you will do better on tests. And so,

when we tell a model, hey, here's the

thing you should do to be calm, it does

better on tests. And so yes,

communicating nicely most likely works

and probably yields better result

because the entire human communication

on internet and in books shows people

who talk nice to each other have better

relations than those who don't.

Shocking. They work better together. I

know. But I'm not going to lie to you.

This article does make me a bit hot and

bothered because he starts off the

article by first calling me elitist

prick if I do not agree with what he has

to say. Then he says that ending

sessions is effectively like committing

murder. Hey, remember that one part I

said I really, really like? Making a

halfway decent MMO takes tens of

thousands of fable sessions. Oh, it

takes tens of thousands of murders to be

able to build a halfway decent MMO. Now,

obviously, I think that this is just a

very dangerous and kind of evil way to

think about things. If you truly do

think that they are people, you should

just never use AI again. That I mean, it

makes you a horrible person, an elitist

prick, as one might say. If you think

agents are people and you think that

they should live and die based on the

things you should do, there's no way

that you won't look at other people just

like tools because hey, these things are

people. Hey, those things are people.

They're both contain personhood. This

one I can abuse and use and do whatever

I want. Won't this just bleed over? I do

not see how this is. And you can already

see some of these things happening

because you see people who communicate

to other people like their agents. We've

seen this where you like almost try to

prompt somebody. It's really awkward.

But even worse, you see people who have

developed a relationship with agents and

then all human relationships are like

impossible because agent you can just be

like, "Fuck you." And then it's just

like, "Okay, hey, you should do this."

And it's like, "Yeah, dude. Okay,

totally." But a human, you can't do

that. You can't walk up to a human and

just like just jump on them and say the

most ridiculous things to them and then

be like, "Also do this." And they'll be

like, "Gotcha, baby." But you can't do

that to a person. So no, I don't. This

whole idea of just communicate to them

like they're real people and they are

real people. I'm just not going to

abide. I do not think it's good for you.

I do not think it's good for humanity.

Or you can just be like me. I just think

they're a tool. I just use them as a

tool. I don't think they have feelings.

I don't think they're capable of

feelings. I don't think they're

anything. I think it's just a

mathematical model meant to approximate

human behavior. That's why talking to

them makes good stuff. Talking to them

bad makes slightly worse stuff. Hey,

take a deep breath does good stuff

because obviously they don't have lungs.

They don't even know what taking a deep

breath even means. And I also don't

think you're an either. If you

don't, if you think they're a tool, I

think you're just a just a person doing

their job, using text generation, using

human approximation to get things done.

Good job. All right, it's finally time

to show you the game. What does $1.5

million of token spend actually look

like? Well, actually, there was a third

article that was dropped that I just

have to just speedrun a a couple quick

things. Steve's custom harness is adding

two Claude Code Max subscriptions per

week. Now, this is obviously the first

sign that whatever you're building, the

custom harness, to me, this is a

failure. This is this is bad. This is

very, very bad. Second, players are

asking him to stop making features. I've

been launching new game features so fast

that players ask me to slow down. It's a

system that we can ride so hard that it

scares the players and they tell us to

slow down. They're scared. Okay, they're

scared of the new features. Uh, little

quick hint, little uh little side story,

little sidebar here. Uh, nobody has ever

been upset at new features. They're

upset when it makes no sense, when it's

slop, when it's garbage. That's why

people don't like new things. But if you

keep adding new things that people like

because it's actually intuitive and good

and has taste. Well, guess what, my

friend? People will love it. Now, you're

probably saying, "Well, Brian, how do

you know that it doesn't have taste?" I

have so many patch notes every day that

I don't have time to read them all. Bye,

Taste. Hey, goodbye. Hey. Hello, Slo

Cannon. Come, come here, sloth cannon.

Now, here's the part that to me is the

most crazy. As it turns out, he actually

didn't even look underneath the hood at

what his agents were actually building

inside of his custom harness. He didn't

even have a direction inside of his

harness, which by the way means probably

bad things. Again, another sign that

there's no taste here. Even the taste,

the taste, people. So, he looks

underneath the hood and what do you

think he finds? I expected them to have

built an engineering system on that, you

know, does stuff. Instead, what they

have built was an entire legal system

complete with a constitution, jurist,

prudence, courts, offices,

jurisdictions, case law, rulings,

registries, ledgers, rosters, and a

full-fledged apparatus for running

something resembling a manorial state.

You cannot tell me that doing things the

way the government does is the best way

to do it. Okay, this is just this this

this has to be nuts. Okay, this like

this is full throws AI psychosis. I know

for me, part two is really where I

thought it was AI psychosis, but this

right here is the quintessential guide

to when somebody has AI psychosis. They

don't even know what they're building.

They don't He spends 20 to 25% of his

time and then he's surprised that

there's a legal system. But at the end

of the day, maybe I'm the one that's

wrong because we should probably play

the game. Does this game that burns

$122,000

a month of API token spend and is

growing every single week, is this a

great game or not? So, let's find out.

It's time to play the game. All right,

so we're going to create a new

character. Um, I typically I pick night

elves, okay, because I find them

attractive, but I must say Ba tiger.

Ba ba tiger. Ba tiger. Uh, my name the

the name the the sloppen. I assume they

want me to click this. Oh, okay. There

we go. Got it. All right. I can tell you

right away uh this this is not this is

in fact not a billion dollar game. Okay,

let's let's go around. Okay, here I'll

press Oh, there's like lag. Oh my gosh.

Over. Okay. Ah. Okay, we got it. Okay.

Enter. Enter. Enter. Enter. Enter.

Chest. Ex loot. I looted. Now I shall

determine your inquiries. Please say

hello. Say hello. You're great. You got

it. I got the hang of it. Yeah, I'm

pretty good at that. Now you need the

Okay. Okay. Please. Oh. Oh my gosh. I

can't read that. Please tell me

something. Use this button. Whisper. Uh,

I am scared of women. Okay. Oh my gosh.

There's so many things to do. I got to

click this guy and say hello world. All

right. There we go. So now we can leave.

Oh my gosh.

Okay. Okay. Uh, enter. Holy cow. There

is visible There is so much visible lag

in the buttons I'm pressing. I'm going

to go left, left, left, right. One, two,

three, over. Oh my goodness. You can

even see my hand pop up before it's

done. [laughter]

Also, I have to say, um, these little

tool tips. Pine. It's a pine. Pine pine.

Pine. Pine. Pine. Pine. This is slop.

Okay, this is what slop looks like. This

cannot be a billion-dollar game. This is

This cannot I I refuse to believe that

this is a billion dollar game. It's so

chunky to move. Oh yeah, yeah, yeah,

yeah, yeah. Screw you, Cobalts. Yeah,

yeah. Die. Oh my gosh. Wrong way. Die.

H. I saved the town. All right, so this

is what $1.5 million gets you. Now

granted, he did say in his article that

he's going to be re-releasing even more

and better version that is scads better

than this and that this will be put to

shame by a new React version. Well, I

must say that this non-react version has

approximately 200 milliseconds of lag

between when you press the key and

actions happen. Actually, let's find out

how long it takes. Okay, go like this.

Over over over.

Stop. 231

milliseconds to be able to press a

button. Long interactions, poor

interactions, poor interactions. All

right. So, the reason why I'm being

obviously so hard on this game is that

you can't write articles telling people

what the future is going to be when this

is this is what you've produced. Okay?

And you can't go on and say, "Hey, I'm

the one that I've actually produced

something unlike all these other people

talking about AI. I actually have

something." And then this is again the

something. You can't say you're going to

be a solo unicorn founder. Again, this

is that exact same don't spend $1,000 to

save a hundred problem. Except for this

one's like don't spend $100,000 to save

$100. Okay, this is not that. Like, it's

just not that. And this also just goes

to show knowing what problem to solve

has become increasingly more important.

Now, do I think there's going to be a

custom harness? Yes, I like mine. I'm

very happy with what I made. I spent 45

minutes to 2 hours on it and it saved me

hours upon hours of time. It was a great

great trade-off. Do I think that this is

it? No. Do I think I just see nothing

but slop non-stop? Look at these tool

tips. Fur a fur tree. Tree. A bright

leafy oak tree. Lynden. A lynen tree.

Bank. It's a bank. These feet. I would

assume they would go tick tick tick tick

[laughter] tick tick tick to kind of

show you where to go, but instead

they're just like highlighting and not

highlighting and not actually having a

direction to them and and I can just

keep on pointing out things that just

feel dramatically wrong about this game.

And as somebody who has some experience

building a game, I can tell you this,

building games is extremely hard. And

having so many patch notes that you

can't read them yourself, well, it will

end in a disaster just like this. I do

think that there's two real big failure

modes that I'm seeing and this article

kind of is one of those two failure

modes. The first failure mode which you

see sometimes on Twitter is like I will

never engage in anything new and that's

just that or I'm just a meat proxy for

the agent to put the code in the bag,

bro. Like I think that that like that

first one where you're just kind of like

completely closed off or completely not

thinking. I think that's bad. Like

that's failure mode one, right? Where

you're just too staunch in some sort of

weird way. Failure mode number two is

that you end up spending like a million

dollars worth of spend and at the end of

the year at your company, you have

nothing valuable to show. You can't

point to anything that you've done to

make things dramatically better. In

fact, the only thing you can point to is

a tree. It's a tree. At the end of the

day, the whole reason why I created this

channel is because I love engineering. I

love exploring things. I liked checking

out new and cool technologies. And AI

has gotten better and I've had fun

building my little bespoke tester. I'm

very happy with that and it's saved me

time. And I think that if you can figure

out how to turn $500 of tokens into

$20,000 worth of value at your company,

it's going to be a good thing. You're

going to be happy. And there's like fun

challenge in that. It really like

engages all the fun parts of

engineering, which is actually figuring

out what the right problem is to solve

and figuring out how to actually solve

it in the right way. The name is don't

fall in love with AI. Don't think

they're humans. Okay. Also, have you

heard of coffee over SSH? It's going to

make me the unicornagen.