Pragmatic AI with Matt Stauffer

The Pragmatic AI Podcast is sponsored by Tighten.

T-I-G-H-T-E-N.

We will take your AI ideas, prototypes, or even vibe-coded apps,

and we'll take them to production.

Scalable and secure. Check us out at tighten.com.

Hey, and welcome back to Pragmatic AI,

where we talk about using AI in the real
world.

What works, how to use it well,

and when it causes more harm than good.

Practical tools and real trade-offs for
builders and business leaders.

And my guest today is my friend who is a
builder and a business leader,

John O'Nolan, founder of Ghost.

John, would you mind saying hi to the
people and tell them who are you

and what do you do?

Sure. Hello, hello. my name's John.

as you already said, work on a product
called Ghost.

Ghost is an open source publishing
platform for blogs,

websites, newsletters, subscriptions,

payments. some people think of it as a
more modern WordPress,

some people think of it as an open source
substack.

The truth is somewhere in the middle and
in between.

Yeah. I I'm glad you said that 'cause I
was going to after you finished saying
your

pitch, I was gonna say, in my mind when I
tell people,

I'm just like it's Substack but not evil.

Like and I I know that you probably don't
want to say that,

but that that's the open source version is
is a nice way of saying it.

But yeah, anybody who ever has wanted
something like Substack and not wanted

to deal with all the issues of Substack,

Ghost obviously is way more than that,

but that is definitely one of the easiest
kind of mindsets to think about for it.

And fascinatingly, Ghost has been around

For quite a while, but you have like a
different way of running it than

a lot of other organizations. Would you
mind kind of talking a little

bit about the organization of Ghost and
how it's different?

We do, yeah. We've made some unusual
decisions over the years and along the
way.

most notably I mean there's a lot
outside of being just open source in
general,

but we're actually a nonprofit
foundation.

So a tech company that's a nonprofit is
fairly unusual.

And we have this idea of being
artificially headcount capped.

Yeah.

so we never want to have more than
roughly fifty people working at Ghost.

And there's lots of reasons for that we
could go into,

but the

The kind of overarching theme of,

you know, why do that, which is the most
common question I get,

why the hell would you do that?

That sounds weird. is as an experiment
in what would happen if you tried

to do a company that's set out to be
different from every other company

and its eventuality. So if you think of
every other corporation in the world,

the main ambition is, through one means or
another,

is to enrich its owners, its
shareholders.

You know, we want to make a great product,

Mm-hmm.

we want to change the world, blah,

blah, blah, blah, blah.

As long as shareholders benefit from it.

Yeah. Yeah.

And that's a legal obligation.

You have a fiduciary duty to shareholders
to do what's in their best interests.

And Eric Ries actually has a wonderful
new book about this that just came

out called Incorruptible, which is his
follow-on to the lean startup that looks

at how companies are structured.

And once you've, you know, lean startup
your way to some success,

how do you protect it? How do you stop a
company from inshitifying?

And in many ways, Ghost is a 13-year
experiment in

Yeah.

Trying exactly that, trying to make a
company that doesn't intritify,

that serves its users, first and
foremost,

and practically solely, by having a
company that can never be bought or sold,

that does not have shareholders,

all revenue we make is reinvested into the
product,

reinvested into making things better
towards our mission.

Our mission is to promote and further
journalism,

open source, and everything that's kind of
in that sphere.

Yeah. it's wild because Ghost is
incredibly successful.

It is omnipresent is in the word.

There's a much less biblical word than
that.

But it's it's I can meet people and
mention Ghost,

and there's a very good chance they've
heard of it if they're at

all in the publishing world, if they've
ever considered anything like Substack.

And so you are, by objective measures,

one of the most successful people I know
from it from the tech world.

And normally we pair that together with
and you're also,

you know, living in a lavish house in the
mountains.

And I'm not saying you don't have a
wonderful life,

you have an amazing life, right?

You have a great life, but the fact that
success does not mean an individual

or a board or a set of investors are
financially just kind of like what

is the post post fine post whatever the
Mullin White guy says.

He I don't think it's post financial,

but basically you don't worry about money
anymore.

post economic.

Yeah, it's post economic, right?

Like that's not the end goal of Ghost.

Ha ha.

And as an owner of a company in the tech
world,

we talked a little bit before the call
about some of the financial aspects

of kind of like being an owner of a
company and the way your your personal
finances

are entangled. it's really fascinating
the way that even the best intended people

can find themselves

Compromised?

The

yes, that's the word. Because you can have
the best intentions of the world,

but as the company grows, or as you talk,

take on different types of financing,

or as the company shrinks and all of a
sudden you have to figure out what things

you want to do and how much are your
personal finances being impacted from this
kind

of rate, this spot that you've hit of
comfort or whatever,

there's a lot of hard decisions to make
that you didn't expect to make on day one.

And there are no protections to ensure
you're gonna do the right thing

or that you're gonna care for other people
of yourself.

And you have those protections.

And

Did you know on day one what you were
doing?

Like if as you look back on it,

you know, so many years later,

like d you know, did that young,

impressionable, you know, idealistic young
kid who set Ghost Up originally,

like, understand where you're gonna be
today?

And if so, like how do you feel about it?

I this is an AI podcast, but I'm just so
curious to hear about this.

Ha ha that's fine. partially,

I think is is the honest answer.

So my background prior to Ghost was I was
a core contributor to WordPress.

And you know, everyone lots of people are
are quite familiar with

the more recent dramas in WordPress world
and some of the conflicts

of interest that exists between Automatic
and the WordPress Foundation

and Matt's personal website and everything
in between.

a lot of that stuff existed in 2012 when I
was contributing and

it really bothered me having seen it.

How it worked really from the inside,

from the core team perspective,

of how decisions were made weren't
necessarily always what was best

for the end user. There was this big,

you know, multi-hundred million at the
time dollar funded business that

had a lot of sway, shall we say.

And so, in part, some of my ideas came
from thinking about what would WordPress
look

like if it lived up to what I thought it
was.

You know, this idea of democratizing
publishing,

Yeah.

this idea of

creating open technology for everyone,

not just to get rich off it.

And in many ways it's,

you know, an exercise in kind of doing the
opposite of of some of the

the the paths that that WordPress followed
in in that regard.

And then the other, which was on a more
personal note,

was eventually I don't know what how old I
was,

maybe was twenty three or something.

I was sitting on a beach somewhere in Asia
and I was hanging out with

a few friends and we were kiteboarding.

And in between kiteboarding I was sat on
my laptop

just coding and submitting code to to
WordPress as an open source contributor.

And I'd kind of been playing this like,

what do I want to do when I grow up game
in the form of like if I

won a hundred million dollars,

what would I do with it? kind of thing.

And I always say that game's really easy
in the beginning.

You know, you buy a Ferrari, get a big
house,

give money to all your friends and family,

pay off the mortgages, whatever,

and then you go, Okay, you done that,

now what? travel the worlds, give to good
causes.

Okay.

Like you've done all that now what?

Learn some languages. and I don't know,

like the the Grand Theft Auto of Life as I
imagine it,

I think you could burn through those like
materialistic,

Yeah.

capitalist, consumerist things.

Not to say they wouldn't be fun,

I'm sure they'd very fun. But let's say
you burn through the novelty

of those things wears off after,

I don't know, a year or two, maybe five.

So then what? Then you wake up on a
Saturday morning.

what do you do now?

You don't have to be anywhere or do
anything,

but what do you choose to do? And I
realized my answer to that was I would
just

be doing this. I'd be traveling,

hanging out with people I like,

working on open source code. And I at the
time was making maybe $30,000 a year.

And I was like, interesting. So the goal
of like having this extreme wealth

now to me seems kind of silly if five
years down the line what

I really actually truly want to be doing
is the thing that requires

no wealth whatsoever.

and so now it's like, okay, so now my
goals can change.

I don't need to come up with like a
billion dollar idea.

I need to come up with a $30,000 a year
idea or an idea that can pay me a

you know a reasonable tech salary.

And tech salary is incredible.

I have a wonderful quality of life.

I don't need the 50,

100, billion, like any of those amounts,

like a regular tech salary. I can find
happiness very easily within.

And so then like blending these two things
of what if you did the opposite

of an average corporation and then what if
your goal was not

to become extremely wealthy, but rather to
to build something that allows

you to live the life you want?

that's where the the genesis of the idea
was born.

And I think now with hindsight,

yeah, there was a lot of naivety.

There was a lot of things that turned out
to be harder than I expected them to be.

But I'm very proud of what it is today and
that it is true to its values.

And I think with all the ways in which

The tech industry has portrayed people
over the last 15 years in particular.

it's more and more relevant to be a
shining kind of counterexample to that.

And to to at least have something to hold
up and say,

look, there is another potential path you
could take.

You don't have to take it, but there is an
option that works.

Yeah. It's been proven. Yeah.

Like this is possible. You you can do it.

I love that.

I really appreciate you starting with
this in part because I want people

to hear it. but in part because I do
think that any aspect of 'cause like

you are one of the most pro AI people I
know,

especially in the early days, and I was
supposed to keep that a secret until

you know I I ask you about that.

But I it is it is relevant because guy
owns a tech company and is a big fan of
AI,

slots you in a particular category in
people's minds.

Right. And what I want is to be able to
hear your perspectives,

understanding that you're not the classic
guy who owns a tech company and

is a big fan of AI. Right. And so I think
that I wanted this kind

of framework for like the conversations at
AI.

So since we're in an AI podcast,

let's talk about it.

so I did mention that you have been one
of the most pro AI people I know.

Love it.

I remember sitting down with you at a
conference last year and you were just
like,

Matt, you need to be doing more,

you need to try more things, that's gonna
blow your mind.

And it wasn't any of this like or you'll
lose your job kind of pressure,

right? It was more just like as a creator,

as an entrepreneur, see all the what's
gonna bring to you.

So, what was your kind of origin story of
bringing AI into your day to

day life and into your business?

Yeah, these things connect all the topics
we're talking about connect

in really fun ways. So openai dot com,

Okay.

the website, started out long before it
was the that you know the mega giant

is today as an AI research publication
where they would just share these long

reports on machine learning models and
what they were doing and what they were

achieving. Early on it was mostly about
how well they could get them

to play video games and what was working,

what wasn't.

And that research publication ran on
Ghost.

so all of OpenAI.com was hosted by us.

and we continue to host it up until
around GPT four.

at a certain point, we were serving more
traffic for open AI than

all other Ghost sites in the world that
we host combined.

and it so that was pretty spicy.

So we had like very early awareness of
OpenAI at Ghost just by proxy,

kind of reading these posts. They did

most beautiful site design that kind of
blew our minds in terms of what

was possible with Ghost theming.

and so as soon as GPT came out,

you know, we were right in there and
trying it just in terms of like write

me a poem, you know, the things we used to
do very early.

Yeah, yeah, yeah.

And then by around GPT three,

I think, I was starting to pair it into VS
Code.

I mean

One form or another and and try out like
what does this do?

People were just starting to talk about
using it for coding and it could do
coding.

And the autocomplete version of it was not
very interesting to me.

But as soon as you could have a
conversation with it and say,

hey, I want you to write me code that does
this,

it got very interesting. And it got
particularly interesting to me in two
ways,

which is one, I'm primarily a designer and
a front-end developer,

and then very much a product person.

So I can talk.

structurally about an entire code base,

but if you ask me to go and write an
algorithm,

I cannot. There is I just I my brain does
not function in the method

of memorizing syntax and reasoning about
code architecture,

but very much does function the way of
knowing what it needs to do and

how it needs to work and what all various
components are and how they talk

to each other. So the writing of the code
was my impediment to being able

to do things myself. And as Ghost has
grown,

you we've employed more and more

engineers and built a team around having
multidisciplinary people working together

to kind of bring their specialist skills.

But what I found very quickly with GPT-3
was suddenly my ability

to create just absolutely exploded.

Because even a small app or a small simple
thing that I could

not have produced myself in the past,

suddenly I could. And then as AI has
become more powerful,

it has created a really interesting new
dynamic for us as a company,

which is, you know, we

We're a nonprofit, we have no funding,

we have no ability to get external
funding,

and we have the silly idea of never hiring
more than 50 people.

Every competitor we have has between 100
and 3,000 employees,

and between 100 million and 1.1 billion in
funding.

And so suddenly what had always seemed
like we're just going

to be David versus Goliath forever now
feels like if

You play your cards right. If you think
strategically and you

can organize agents well, headcount might
not be the biggest advantage it used

to be anymore. Small companies,

I think, increasingly will have more and
more ability to compete with big
companies.

and my my spoiler alert here,

but my long-term hypothesis for AI that I
hope comes true.

I think it's the sort of utopian future is
will have a much larger long tail of

smaller, more niche businesses serving
smaller markets and fewer of these

giant horizontal platforms serving
absolutely everything.

I love that. See, this is why I need you
on this podcast,

because that gives me hope and excitement
about the impact of AI,

right? Like and you know, like one of the
things I ask everybody like,

Hell yeah.

how do you feel about the future?

And I'm like, That future I love,

right? I love this future because often
what we hear is,

well, people are gonna build their own
software so that SaaS is gonna go away,

or people are gonna build their own
software and software engineers

are not gonna have jobs or whatever.

But if you tell me we are gonna have these
mega corporations that

are not really serving anybody well,

they're

Finances are not really being distributed
to anybody other than

the ultra-rich people at the top really
well.

And I don't think their existence has any
benefit to the common good

of humanity if you're saying like they are
going to lose power

and individual business owners and
individual small businesses

and individual entrepreneurs are gonna
have more opportunity to take market share

from them. I'm like, sign me up for the
revolution.

You know what I mean? I'm like,

Ha ha.

hell yeah. Okay, cool, cool, cool,

cool. So you guys started using that
pretty early.

What was the adoption like on the team?

I know that you were pro AI. Did did you
have a mixture where some people

are on board and some people aren't?

Was it pretty universal? What was it like?

yeah, big mix. I was I was definitely one
of the first,

if not the first, within within ghosts to
start kind of using it in anger

in in day to day work. and then let's see
what've been twenty twenty

Five, four, God, I c I don't know,

time is a flat circle. a couple of years
ago I kind of came to

the conclusion that this wasn't going
anywhere.

I think it was relatively shortly after,

you know, Toby had done his big Shopify
moment of reflexive AI usage

is now a baseline expectation at Shopify.

And at our next team retreat which we do
annually,

I said, look, this is we need to get
everyone to the same point because

we can't have this this big spectrum of

knowledge and ability and skill level.

So we did an entire week, the whole
company with the people

who had already experimented with figured
out stuff about AI,

kind of talking about helping,

teaching everyone else about what they
should try out,

and then spent the week kind of running
experiments,

building little projects, going like,

this didn't work at all. It's terrible.

this worked great. And of course within
that we had some really significant
skeptics

on the team. there was a a couple for
sure.

Mm-hmm.

We had some really significant
enthusiasts,

of which I was one, and then a you know,

a large swath in the middle of people who
were open-minded but skeptical

to various degrees. and with hindsight,

I'm so glad we did that at that moment,

because had it been a year later at the
next retreat,

it would have been too late. this was I
think we were kind of GPT 4.0,

Claude 4.0 sort of point. It was

Mm-hmm.

pre-Cambrian explosion that sort of came
after the like four point fives.

Yes. Yep.

And you in my mind to really take
advantage of of that moment as it came up,

you kind of needed to be ready to r to
receive it,

if you will. had we had we waited and just
had skepticism for another year,

I think we would be two years behind
probably.

Mm. Okay. so in your day to day right
now,

what is your personal interact with
interaction with AI?

Are you, you know, is all the code that
comes out of your desk all through

AI or are you using it for organizing your
business and your life and your email?

Like how deeply embedded are you?

Yeah. so it varies with my job,

depending on the month, the things that
are on my plate can be very different,

you know. one point it might be all
quarterly planning,

another point I might have a couple of
weeks where there's not that much
business

stuff on my plate. So then I'm doing more
code.

Other points I might be needing to go and
give a presentation at a conference.

And so the types of things I work on very
wildly.

And so my

type of use and my intensity of usage of
AI kind of varies depending

on the season to some extent. But my
largest subset of use is still code

and code related things. codex desktop

is still my MVP right now. And nothing has
surpassed that.

It's been about six months. I'm kind of
surprised.

Nothing has unseated that yet,

but that's my my daily driver across
two different Macs

simultaneously with screen sharing for
arcane arcane reasons.

Okay. okay. Yeah, so can talk me through

talk me through your coding setup because
like I said,

when we talked at that conference,

you were doing more than most people I
knew.

And also almost everybody I know uses
Claude,

so you're one of the few codex peoples I
know.

And you're also using desktop,

Mm-hmm. Yeah.

not CLI, and you got the multiple
machines.

So can you kind of talk me through what's
your coding setup?

Yeah, for sure. So I think i these
abstractions that happen are really
interesting

in how the tools move. So we had you
know,

the chat interface and then the chat
interface moved in VS Code,

and then we went like, what if that was a
product and then that was cursor

and then we weren't actually hang on,

the CLI has way more power, so we're gonna
move over to the CLI.

And so now I use Codex Desktop.

And what Codex Desktop represents to me is
the power of the CLI without

the UI of the CLI. You go, it can do
everything a CLI can do.

But also you can click stuff and that
turns out to be convenient.

It's pretty nice. Yeah.

And you can do things in one window,

not four, because if you work in a CLI,

you typically, okay, now I want to at the
end review the code.

So I'm going to pull up some code editor
or a diff viewer of some description.

Okay, now I need to get it into GitHub.

So that's another window for that tab to
see what it's doing.

And then we have like various points where
we need to go back and forth.

And then, you want to work on more than
one thing at the same time?

Multiply all those windows by two or three
or four.

and so Codex Desktop, I think,

is a reimagined what if you had the
workflow you were using with Claude Code

CLI, but it was put into a graphical user
interface that does all

the things that you're trying to do,

as as that's the problem it's solving.

It's not solving code editing,

it's not solving the traditional things
you're doing,

it's solving the the new workflow that has
emerged out of how people

are interacting with things like Claude
code.

And bringing all of that front and center.

So you have a sidebar of the projects that
are running.

You can see a spinny thing for which ones
are working and which ones

are waiting for a review. You can indicate
what's a work tree or what's just running

on your local machine, what's running on a
remote machine or a cloud machine.

They're all there. And then when you click
between them,

you've got a mix of the, you know,

the agent and what it's doing in a
traditional conversation view.

But you've also got an integrated browser,

so you can see the agent clicking around.

No playwright, nothing. You can see it
clicking around testing its work.

You can get in that browser and also
interact with it and test it.

You can annotate parts of it or take
screenshots.

When it's done with the work, it has the
sidebar can turn into a

a code review plane, which with which has
a diff viewer.

And then when you click open pull request,

it integrates with GitHub and it shows you
pull requests has been opened,

tests are pending, they're yellow or they
turn green when they're done.

And then you even have a button for like
resolve comments.

So if you have something like CodeRabbit
or Codex doing comments on your GitHub PR.

It's just a button, you go like go deal
with those comments and like update the
PR.

So it it just accelerates all those
things.

And I think the most significant
difference to me is just the number

of things I'm able to do in parallel and
keep track of is greatly increased.

Yeah.

Because with five Claude codes and five
CLIs,

sooner or later I forget which one is
which.

And it's too hard to like reload what was
going on again,

because there's not enough fidelity
there for me to kind of have a a sense of

yeah, I remember and go back into the flow
of it.

Whereas Codex does give me that.

you just taught me things that Codex
desktop does that I had no idea,

which makes sense because you're at the
forefront of everything.

It's so cool.

so you're just just from a sorry for the
non programmers in the audience.

So if I've got three like full stack
applications running locally,

needing to be served by something like
Laravel Herd it can be

in their directories or even in work trees
for them,

doing work and previewing them at whatever
URL I want them to be.

Wow. Cause usually when people say that
that whole promise,

it's because it's all front end stuff.

So they're like, yeah, we can spin up this
JavaScript directly in our browser,

but we're not actually integrated and
touching the local file system.

Well, that's fascinating. you also
mentioned another computer,

so talk about that.

Yeah, so unfortunately some code bases
work well with work trees and there's

not too much to figure out other than,

you know, you want your live reload server
not port conflicting and so on.

other code bases, particularly larger
ones,

of which Ghost Ghost is these days,

were not designed to work with work trees,

and so you can't really run these parallel
tasks at once.

but I have a Spare Mac mini.

And Codex Desktop has this really cool
feature where if you have Codex Desktop

installed on two Macs, you can just say
talk to each other,

and then the project of one shows up in
the project of the other.

Really?

And so you can remote control it natively
through the Codex Desktop app.

You don't even need screen sharing.

it all will just work. but then it has
a fully isolated environment.

Yeah.

So it's effectively like two hardware work
trees,

if you will, and there's no risk of
anything conflicting.

Even if your project wasn't designed to
work with work trees.

So I I use it for that to do parallel work
in Ghost,

because I mean the the way I interact
with an agent now versus maybe

six months ago is six months ago it was do
the next thing and then you know

it will come back ten minutes later and
then you review it and go and then

do the next thing. That's not how I work
with agents anymore at all.

Now it's we're we're making a plan,

we're setting a goal, and the agent will
work

Anywhere between two hours to I think my
longest session has been like seven days.

and at the end, then I'll go and test it
and talk to it and

huh.

see what's gone on. But man, that's a long
time to wait.

Like what am I supposed to do?

Yeah. Yeah.

Just sit here? No, I want to do something
else.

So like parallelisation, the like the
importance of it has steadily increased

as that as those sessions have kind of
become longer.

That's nuts. I want to hear more about
this workflow.

'Cause historically when people have said
I I set it and forget it,

they're using Ralph Wigan bloops or other
things that are just kind of like

you set a spec and then just say,

Iterate until you're done. Is that what
you're working on or is that

a different way of working?

So it's kind of that. I like to say or
describe the way I work as I'm I

try to be one to two steps behind whatever
the frontier is.

And the frontier is whatever people are
talking about on Twitter.

And it's it changes week to week.

Yep. Yep.

It sounds crazy most of the time.

And if you ever attempt to dabble with
whatever that thing is,

you're like, this is so complicated.

God, it's taken me like six hours to
figure out like what you're talking about.

Yeah.

And another three to like have it not
work.

there is value in being the first mover
there,

but I think unless you have you know all
the free time,

you can just end up down so many dead ends
of the things that turned out

Right.

not to be the next thing. so I like to
be one to two steps back.

And so what both Claude and Codex have
now is a feature called goal,

goal mode. You can do slash goal and you
set a goal,

and the goal can be.

Follow the plan that's in whatever.md.

work step by step in incremental slices,

Markdown file or whatever, yeah.

use red green test driven development.

keep a progress log of everything that
you've built,

and in particular anywhere you've deviated
from the plan,

and complete each piece of work with a
commit with a descriptive message before

moving on to the next. And so what agents
will do then is effectively what people

used to call a Ralph Wiggum loop.

But you don't need to know anything about
what the hell a Ralph Wiggham loop is.

You just slash goal, you write a prompt,

Good. huh.

it's referencing a plan. And through every
compaction or new session

or sub-agent or whatever, it's still just
going and referencing the same plan

and cross-referencing against the progress
file that it keeps for itself

of what's already been completed.

And so then it's able to keep going.

And provided you that you have a you know
a test suite it can use

to validate its work against, and a clear
enough set of parameters of

you know, what's in the plan of what the
outcome is that you're looking for,

the results are surprisingly good these
days.

It's it's nuts. And it just kinda it blows
me away that this is

is still possible just because of how much
less stressful it is.

There they go I think so people talk about
AI psychosis

in particular for developers, right?

Mm-hmm.

And there's various definitions,

but the one that

I use and like Aaron Francis, we talk
about this,

is you discover AI, you for programming
work,

you go, my god, this is incredible.

And the next thing you know, you're there
at 2 a.m.

going, hit that dopamine button,

Yeah. Yeah. Mm-hmm.

hit that do one more prompt, one more
prompt,

one more prompt. And what I've seen time
and time again is this phase,

this AI psychosis phase, lasts kind of
somewhere between two and six months.

I would say usually three. And then you
kind of just get exhausted of like,

Yes, you get burnt out.

I don't, I don't want to just I

The novelty is worn off. It the the
results are still good,

but I'm tired. And so this getting into
the goal mode or what people call being

on the loop instead of in the loop,

heals that. It heals that really
significantly in that you hand off some
work

and you'll find out in an hour or a day
if that went somewhere.

But you don't have to be babysitting.

You can have a life. You can step away and
watch some Netflix or go for

a walk or spend time with a significant
other.

and it's still

doing amazing things. And so I I think
that mental shift is very healthy
actually.

Yeah. I have I'm so if you're two steps
behind the the lead,

I'm three steps because I've I've done
almost everything you're saying.

I don't use slash goal and obviously need
to,

but I have walked long running projects
through,

but one of my personal idiosyncrasies is
that I don't trust my

AI to touch Git. So you can read from Git,

but you can't write to Git. And so I'm
like,

do all the work, hit the stopping point,

here's the criteria for the stopping
point.

I'll review all the code, I'll commit the
code,

and then I'll kick you off to the next
one.

And it I've found that's a lot easier for
me,

A, to trust that it's not gonna go wild
and kill my get or whatever,

but also B to understand it. Because if I
come back after it's been two days,

there's no way I'm gonna comprehend all
the code that I'm I'm looking at.

And so that's one of the things I wanted
to ask you was do you feel like when
you've

got a seven day loop or a two two day loop
or even a seven hour loop,

do you feel like you're able to come in
and have the discipline,

to c to read through and comprehend
everything that you're shipping?

Or are you at the point where you're like,

you know what, like I understand that it
accomplishes the goals and

if the code breaks, I'll just use AI to
fix it.

Both, and it depends on the project.

it it depends on the size of the change
set,

Okay.

and it depends on the project a bit.

So for example, if we were touching
something to do with Stripe payments,

complicated. In Ghost, huge code base.

Yeah. Yeah.

the AI AI runs rampant, does some stuff.

That's gonna have at least two,

if not three, levels of code review before
we consider putting on staging,

let alone production, right? but for

Internal code bases of which we have many
for small projects of which

I have a really clear understanding for
small changes of which the change

set isn't actually going to be that
significant,

even if it's like a chunky feature.

no, let's go. Like ship,

if there's a problem, AI will figure it
out.

and the risk, it's all about like what
is the risk?

What's the worst thing that could happen
if this goes terribly wrong?

Yeah, yeah.

If this goes terribly wrong and like,

it's gonna inconvenience a couple of
people on our team.

Yeah. Okay.

Then we'll fix that when

that happens. if it's gonna take down
customer sites and cause them

to lose revenue, i yeah, that's maybe we
slow down a bit,

you know, pump the brakes.

We and we have the exact same metrics and
I love that response.

I mean, like for me, I'm not coding on
client sites right now,

and my developers are not using long
running loops or anything like that.

But and they also ensure that they
understand every single line of code.

And they're still writing a large portion
of it,

but when they use AI, they're they're
understanding everything.

Whereas when I'm coding, if the CEO is
coding,

it's probably going to be prompt in
between meetings,

review in between other meetings,

or I'm working nights and weekends,

and I'm almost always it's either an
internal tool.

Or it's this Laricon puzzle I'm working on
that is completely nonsensical.

And if it breaks, it literally does not
matter at all.

And it's my ADHD hyper obsession right
now,

but does not matter at all. And so I love
your metric.

You're like, d what happens if it breaks?

Which which I've always told clients and
and programmers as they're learning.

I'm like, what do you need to test?

You need to test the things that if it
breaks,

you know, you're on the front page of the
New York Times for all the wrong reasons,

or your clients are all angry at you.

And this is a similar thing here.

Like, if what what happens if I don't
understand this code and something goes
wrong?

that then that can have a big impact on
importance of you understanding the code.

Okay. What things outside of code do you
find yourself regularly using LLMs for?

Yeah, exactly.

vast majority is conversational,

deep research, self-education.

there's a concept I'm trying to figure
out,

Okay.

there's a plan I'm trying to make,

there's topic that I'm curious about or
trying to reason about.

I would say that's the vast majority of
I mean there's you'll know as

a a company founder or even someone
who's a head of products,

there's a lot of just thinking and
investigation into products.

Yes.

forming ideas and validating or
invalidating them.

And I find it just incredibly useful for
that.

For the types of things that, you know,

three or four years ago you might have
Googled to find a starting point,

then gone and read a bunch of stuff about,

and then kind of made a long set of notes
and then gone and done more research about

it just compresses that into a
conversation.

but for the mo that's probably the most
significant thing.

What one area that I

have not gone particularly deep or spent a
lot of time

is in non-coding agent workflows that
where people are getting

it to do crazy things with emails or
monitoring their inbox or managing their

Mm-hmm.

to do lists. I haven't done I do have an
open claw running.

You don't have an open claw running.

yeah, but

You do.

It's it's limited open claw. You know,

it gets lots and lots of hype.

It can do some things, but it makes tons
and tons and tons of mistakes.

and so it's utility for my life has been
relatively limited.

The most useful thing it does,

this is just a fun one, is I have it I run
like a home open source home automation

system called Home Assistant, which sits
on a little PC in my in my house

and you can SSH into it and configure it
or whatever.

Now I don't want to SSH into it and
configure it.

So whenever something's not working in my
house,

previously I would have to like log in and
go and like figure out what's going

Uh-huh.

on with the settings, all this kind of
stuff.

And now I tell OpenClaw like, hey,

my I came home and the lights didn't come
on,

like figure it out. And then go and fix
it.

Like they're supposed to, yeah.

You know. And it does. It can it can fully
like go and reconfigure

whatever's wrong with the automation
that I forgot an edge case of.

And so that's really fun. But I haven't
got too many uses for it outside of that,

other than

Experimenting and playing with like,

I wonder if this is possible.

Yeah. Hmm. my dad and my brothers all love
home assistant

and my dad would regularly share these.

You know, 'cause sometimes you worry when
you got boomer parents,

sometimes you worry about like,

you know, are they gonna be the ones who
get soaked up in these terrible Facebook

conspiracies that turn them into like
right wing extremists or are that you
know,

Sure.

whatever, fake news. And my dad constantly
is one of the most critical people

of experience. And this you know,

he's been a programmer in the past,

he's been a president, he was an
electrician,

he understands

technical stuff. And so it's been very fun
watching him do exactly

the process you're doing. And he's like,

Yeah, freaking Claude told me this.

And you know it's and he's like rolling
his eyes and I'm like,

It's so funny 'cause people are always
like,

There's hope for the next generation.

I'm like, I want hope for the previous
generation.

And watching him interact with Claude,

Ha ha.

I'm like, we're good. We're good.

He's totally fine. but it has made it
has made things like home assistant more

more accessible to me. I w I would love to
do it,

but I just don't have time and energy to
to become an expert enough to use it.

But I'm like, I do have time enough and
energy to buy a Raspberry Pi

and then have Claude walk me through the
project of setting,

you know, setting the thing up.

Yes.

It is ex it's my favorite open source
projects outside of outside of Ghosts.

Yeah.

Home Assistant is wonderful. It's very,

very fun. And the to me the most fun thing
about it is,

Yeah.

you know, lots of these, whether it's like
a Philips Hue system

or Apple's nonsense or whatever,

they're all just very expensive.

So, you know, one door sensor is like
fifty bucks or something,

and then the whole thing doesn't play
nicely and then Google's

Yeah.

not compatible with the Lexa's not
compatible with all the other

Assistants who I won't name for fear of
setting off people's smart speakers.

Home Assistant just is open. It works with
any components,

typically, you know, five dollar sensors,

eight dollar sensors. And that you can
just make it all play together

and kind of make it work how you want.

And you're not within this annoying
ecosystem that doesn't quite do what you
want.

You can kind of do whatever you want.

Yeah. Yeah. I love it.

and I I always say as if you're a
programmer,

it's like bringing your programming skills
to the physical world.

Because everything you know about like

Yes.

logic and automations and systems thinking
and sensors and data,

like suddenly that's in three dimensional
space and you can like do stuff with it.

which is very satisfying if you spend your
life in a glass rectangle just looking

Right.

at stuff in a in two dimensions,

it turns out.

Why so many

of us become carpenters? They're like,

I just need to do something in the real
world.

the COVID years, yes.

Yes. Okay. So you travel more than
probably anybody I know.

because your travel's not just visiting.

you do visit more than anybody I know,

but you also go to conferences
internationally,

you're just gonna live in other place for
a while.

Are there elements of whether it's AI or
not,

of your kind of pr productivity,

work, computer setup?

That are really big boons to you as
someone who travels.

Like for example, d does you said you have
this other computer,

does it sit at home all the time and
you're SSHing across the world?

What do you have going on that makes it
more viable for you to work from wherever?

Yeah, I don't know that AI has helped me
as much as I would like it to yet in this.

Yet. Yeah.

my my dream AI use case would be to
never interact with

an airline website ever again.

which I just they're all bad.

Yeah.

None of them work. it's the number of
hours of my life that have been spent

Uh-huh.

on a checkout failed screen of an airline
website because it broke,

right?

is astonishing. And then you have to start
all over,

of course, because you couldn't possibly
try again.

You now have to search for the flight
again and start from scratch.

So like an an airline MCP, where I can
just say,

I wanna put I wanna be on a flight from
here to here on roughly this date.

I want roughly this price range.

And you know all my preferences about like
window,

aisle, root, layover times, like all the
those things.

Make it happen. God, that would be
wonderful.

Yeah. I need I need

to get you a VA. It's exactly what VAs do.

Ha ha ha.

That's amazing.

I I have tried that in the past
unsuccessfully for the type of travel I
do,

which is often one way connecting trips
and flights,

Uh-huh.

and it just gets very nuanced.

It's the type of thing where weirdly I
think an AI on an LLM could probably hold

all of that context more easily than a
than a human.

but we'll see. We'll see.

Human. Yeah. Okay.

But yes, for the the computer at home,

it's it sits in a cupboard, it has tail
scale on it,

and then Codex does its own cloud
connection between machines that goes
through

OpenAI's thing. so that that all works
like just relatively easily.

But yeah, no more travel AI related things
would be great.

when they come when they come.

I think I will benefit from that.

Yeah.

Okay. Airline MCP. Someone's gonna make a
lot of money if they can figure that out.

Or not. okay.

Yeah. Yeah, I know they will, but they
gotta disrupt the whole airline industry,

Yeah. Yeah. And that's why I was like,

which is like

they're they're not gonna make money
'cause it's not gonna happen.

okay, so I one of the things I told you
that we were gonna talk about

a little bit is is things that everybody
else should be doing.

And I'm I you know, again, you sat me down
at a conference and said

you should be doing this from a code
perspective.

And so maybe we've already covered
everything,

but if

people, especially people in your setting,

whether they are people who run
nonprofits,

people who run tech companies,

or people who are programmers or
designers,

is there one thing you're like,

everybody should at least be doing this?

Maybe not everybody has to maybe or maybe
Codex Desktop is yours,

but maybe not not your full setup.

Is there one piece where you're like,

if you're not doing this, you should at
least be trying this out?

Yeah, I I mean we already touched on it,

but I think it's goal mode in in whichever
app you you prefer,

whether it's Claude or or Codex,

I think almost all of them have some
approximation for it now.

and it's it's surprising to me because
even the

the people I know who use AI lots and lots
and lots,

it's a very, very small percentage who
are using goal mode right now.

They're they're in the Yeah, and I and I s
I see this and I noticed this

Yeah, I'm not.

in myself of

learning a new tool comes with some amount
of friction,

even if it's an iteration of an an
existing AI tool that you're familiar
with.

and so you have you you have this sort
of internal,

either conscious or subconscious question
to yourself each day of Am I gonna

go and spend an hour figuring out this
thing that might not work?

Or am I gonna get an hour of work done
using the workflow that I know works super

well, which is the thing that I did six
months ago and the habit I've gotten into

Yeah.

of like

I'm in my comfort zone, this is working,

I'm shipping, it's going good.

And I think you have to come up for air
every now and then on that and find

a good pace for it. I think there's a risk
in getting too stuck in

the workflow that you had from three to
six months ago because things

do move forwards. And the people on
Twitter are usually too far forwards,

I think. You'll be wasting time if you try
and chase them.

Yes. There's a risk there as well,

right? Yeah, huh.

Yeah. But you want to try and be somewhere
in the middle.

And where I see at least in my sphere,

Yeah.

people are a bit behind the middle right
now is not using goal mode.

a good plan plus goal mode is the the
biggest like step

change in in workflow that I've had in the
last few months.

when you're building these plans,

when I build plans, I go to Claude and I
say,

Turn on plan mode. Here's what I'm trying
to do,

here's the technical constraints I want.

Usually pops up a couple questions for me.

it gives me the thing. I almost always
fix at least one of the things

in its plan and you know, we'll maybe go
back and forth.

And eventually I have a plan and then I
say,

Yes, build it. Is your and there have been
times where I have said,

you know what, there's a multi stage plan
or whatever,

throw it in a markdown document and then,

you know, build something to keep track,

similar to what you're talking about.

Are you using that or are you building it
in a different environment?

Are you making your plans in the desktop?

I guess because I was like, I guess you
use what what I was going to

ask is are you making your plans in Chat
GPT and then pasting them in a Claude?

But I guess you're doing everything in one
space,

right? So what is it like for Codex

Yeah.

Yeah, it's all all in one space.

It can be a conversation. It can go and
write some code.

It can go and click buttons on your
computer.

It's like a very it can do whatever the
hell you want kind of thing.

Mm-hmm.

for the most part,

it's the same. It's like, hey,

I wanna I wanna do this thing.

You describe the plan, go make a plan for
it.

The trick,

tweak, technique, the thing that has
that I have noticed

really meaningfully improve the quality of
the plan as measured by

the quality of the resulting output from
the agent that uses that plan is

to give it a larger set of reference
material to work from that

is representative of approximately what
you want.

I want something like that. And so in the
context of product work,

Okay. Yeah.

that might be a competitor's
implementation of the same feature.

including their help desk documentation,

their API documentation, their marketing
announcements about it,

Mm. Okay.

which typically emphasizes, you know,

what is and is not important by what's at
the top and what's at the bottom.

and finding a handful of those and saying,

like, look, we're building this.

I want you to also go and look at how
company X,

Y, Z have done it, and use that as a
baseline for the kind

of expectations or scope of this feature
and find patterns between them and so on.

And the resulting plan I find is much more
rooted in sort of base reality

as opposed to the agent kind of
hallucinating its interpretation of what

it thinks you meant by the plan that you
asked for.

When it goes out and looks at stuff and
goes like,

here's more context, here's more context,

here's more context, then it it does a
thing that is in line with that versus
just

kind of like Magic Pixie does,

here's the plan, you know. so that's
usually what I will do.

Yes. Yeah.

The

level of planning, it corresponds to the
size of the piece of work.

If I'm mapping out a a big new feature
that's a whole new data model,

I'll do a big plan. I will go and look at
like a lot of reference material,

go and I'll go and have it research and
read and validate its ideas

and then maybe have another agent kind of
red team it so like find problems with it.

But if it's we have this feature and I
want to add a a button to it

or we're gonna add a new setting,

yeah, make a small plan and then go and do
the thing.

Like we'll keep it simple.

So it's it's always, you know,

appropriate to the size or trying to be.

I I love that you noted the idea of road
teaming it of saying all right,

you have been in this one mindset and now
somebody with a fresh perspective,

please go look at this and see,

does this make sense? Are we gonna be able
to act on this?

What are the ramifications? And maybe you
give it a different set

of source material you're to use to
criticize it or something.

Yes. And you can also do this in the flow
of work.

This is a this is another fun trick is I
think a reasonable number

of people do this at the planning stage.

They make a plan, have another agent red
team it.

but in goal mode, you can, you know,

you tell it to work in these incremental
slices of work following the plan.

You can also say, as a part of the goal
prompt,

at the end of each slice of work before
committing it to a branch locally,

it's not going to GitHub, spin up a sub
agent at maximum reasoning effort.

And do an adversarial self-review.

And so what that will do is every time
it's completed something,

it will kick it off to another agent in
a new context space.

So it's not going to be polluted by all
the prior stuff from the main agent

who will pick holes in it, give a report
back to the main agent,

who then has to deal with that report
prior to doing the commit.

And so it has a self-reinforcing little
red team loop whilst it's implementing

Yeah. That's helpful. One

things along the way as well.

of the things I like when doing that is
often you can find that

the critical agent is just looking to have
reasons to poke at something.

And so if you just take its notes without
considering it,

Mm-hmm.

you're gonna be like, Everything's
terrible and we need to make a million
changes.

But often if you feed it back to the
original agent,

the original agent will be like,

Well, it was right in these ways,

but it's wrong in these ways. So you
actually be like,

Hey, be critical of the criticism and
hopefully we meet to some kind

Yeah.

of reasonable place in the middle.

Yeah, that's that's still an issue.

I find it it's gotten better. I think
things like CodeRabbit now

Yeah.

on on GitHub pull request reviews are
better at flagging things that are
nitpicks,

like little suggestions of like
technically this should be a semicolon

one versus something that's an actual
bug.

Yes.

Like actual problem. Yeah, for sure.

yeah. Yeah.

That's good. Okay. So I mean, I think the
answer to this is easy to assume,

but I want to hear it anyway. What do you
feel about the future of humanity,

the future of tech, the future of
publishing,

the future of work and people as a
result of the advent of AI and LLMs

I don't know. I think it's

I think I've I've heard every argument
every persuasive version of let's

say most of the arguments for why it's all
doom or why it's all utopia.

And I suspect that the right answer is is
likely somewhere in the middle.

I'm very much not of

Right. Yeah.

the opinion that the SaaS pocalypse is a
real thing and we're all gonna

be out of a job and it's gonna eat us all
next year.

I think that's just

Yeah.

sci-fi thinking, sci-fi wishful thinking
at its best.

I'm also not of the opinion that it's

all complete nonsense generated by an
autocomplete text engine and none

of it means anything and it's all fake and
nothing's really gonna change ever.

I think we're s we're somewhere in the
middle.

I think it will probably take longer to
materialize

in all of the

world changing ways than we think it will.

I think as as people who work in tech,

we have a delusionally optimistic view of
of how long it takes technology

to roll out and impact the world.

It typically takes far longer than we
always think it's going to.

Every prediction that Elon Musk's ever
made has been,

you know, five years ahead of time.

He's been right a lot, but he's I don't
think ever once been on time with

Yeah.

his predictions. And, you know,

there's an old saying in financial markets
that being early is the same thing

as being wrong.

And so I I think we're we're too
optimistic.

I think it probably will change.

I think for the most part I expect the net
result

of AI on humanity over a long enough time
period to be positive.

which is not to say there won't be bumps
along the way,

but my vibe, if you will,

because this is all just speculation,

is far more optimistic than it is worried.

Yeah. Hmm. you've got

me thinking that I've asked this question
and that's

a really similar answer to the people
who've really considered it.

the most common answer is it's somewhere
in between.

And I'm just looking at it and I'm like,

Yeah, when is the last time I asked
anybody like a meaningful,

life changing question about programming?

And their answer is anything other than it
depends.

If if it's a senior person, if they you
know

Yeah.

You got

the the young people who are just getting
started and say the answer is

you must do this, that, and the other,

and that's always it, you know?

And you get somebody who understands
nuance and they're always gonna say it
depends.

Yeah.

so I appreciate the awareness that it
can be all of the above,

right? It can't there can be negative
ramifications and there can

be overly you know, optimistic
predictions,

and we can see an improvement in the human
race,

and also there can be costs and they can
just all coexist.

For

sure. I I think the closest parallel
still,

and this is not a new metaphor,

it's one people have used before,

is just the advent of of mass
manufacturing,

not quite industrial revolution level,

like a little bit later. Do you remember
maybe what was it,

late eighties, early nineties,

when everything was made in Taiwan?

to the extent that made in Taiwan became
a slur for low quality.

Yes.

That's what that meant, is handmade
goods from proper companies good.

made in Taiwan, which later would be
subsumed by made in China,

became bat because it was a proxy for low
quality.

Yeah.

But guess what? A couple of decades
later,

the highest quality products that you
own are made in China,

right? The quality changed over time.

And so this this slop era, it very much
feels to me like the

the revolution of mass manufacturing where
the first thing you could produce

was lots and lots of low quality things.

And then over time it turns out what what
does the market want?

Well, the market doesn't want low quality
things.

So the mass manufacturing reshaped itself
around achieving more quality,

but still at a high production level.

And so now you have Apple products,

you have cars, you have chips,

you have monitors, we have the cameras
that you and I both looking

at each other through, all of which are
made in China,

Yeah. Yeah.

or at the very least, the vast,

vast majority. And I like to think of us
as two discerning individuals.

I would not describe them as as mass
manufactured slop,

right? I'd say they're pretty good
products.

Yeah.

And so I think we're in this horrible,

you know, the dot com era has maybe had a
similar path of like the early

majority is this slopification of new
thing.

And then the market sort of settles into
pe well,

people don't want that. They demand
better.

And so the the industry has to then
readapt to how do we do this,

which has a new sense of scale?

But how do we do it with quality?

And I think we're gonna get there.

I think that's the direction we're heading
in,

but I'm

I remain optimistic about it. And the same
outside of code,

you know, like the all the Sora stuff.

Yes, this is the worst version that it
will ever be.

You you are correct. It is all bad,

Yeah.

people who hate it. I agree with you,

but it's not going to be like that forever
because clearly it's not what

any of us want. but I think we're going
in a good path.

Love that. before we get to the community
contributions,

is there any topics that you hoped or
expected we would get to that we didn't

get a chance to talk about today?

only just to turn this around on you a
little bit and tell like tell

Okay.

me where since we had our conversation in
it was I think it was Laracon

US last year. And I think at the time,

Yes, it was.

correct me if I'm wrong, you had dabbled
in experiments with AI.

but it wasn't part of your daily workflow
yet,

I don't think. Now you're running an AI
podcast.

Like what has to just

You don't need to tell me your whole AI
workflow because you've

I'm sure you've touched on that in other
episodes that people have listened to.

But think looking at the contrast between
a year ago to today,

like what what does that look like?

What is that difference in where you were
versus where you are?

That's great question.

Yeah. And the funniest thing is I
started the podcast in large part because

I'm not an expert, I'm not at the
bleeding edge,

and I don't think there are enough
conversations happening,

not at the bleeding edge. I think all the
conversations are

either completely contrarian,

this is terrible, or this is the the
future of humanity.

And there's so there's so many the the
vast majority of people are in the middle,

and the vast majority of people are not
being heard.

they're not being

given platforms. so I also found
myself sharing whatever I

was learning and people would DM me and
later and say,

you know, I was so relieved or hey can we
have a call?

And then I tell them what we're doing.

They're like, that's what we're doing too.

Okay. You know, like so people are
secretly afraid that they're not keeping
up,

that they're not doing enough,

or that they're doing too much.

They just don't know and there's not a lot
of voices that are being platformed

of just saying, hey, look, as a normal
human being,

here's what I use and here's what I don't
use.

and so I just wanted to make a space for
that.

And that's something that is not because
I'm using AI more or less.

It's because I think it's something that
should be out there.

My general AI workflow started from a
couple of years

ago doing, you know, chat GPT and tab
completions and really simple stuff like

a lot of people. And I was pretty critical
of of it being of any value,

especially as a full time programmer.

and most of the time when I'm considering
these things,

I'm considering what should my team be
doing.

Right. And so I was like, I wasn't even
thinking about myself as an entrepreneur.

Mm-hmm.

I was thinking about my team of
programmers,

what makes the most sense for them so they
are delivering maximum value

to our clients without losing their souls
and their happiness

and actually doing their jobs.

Right. There's some balance there of like
them enjoying what they're doing

and not feeling like they're just prompt
in reviewing robots as their

job instead of craft and creativity,

while also not sticking our heads in the
sand and just being unwilling

to take advantage of this new stuff.

So that was

That was my focus. So at that point,

before you and I spoke, I was really
consistently focused on what workflow
makes

the most sense for senior level engineers
who are focused on one client at a time,

focused on doing the best work for those
clients.

And I had yet to see any world in which
that was actually better

by them really heavily using AI.

And they I I they had plenty of people who
were at the forefront and using

all the models and trying all the things.

So like it's not that they weren't
interested.

It's just we hadn't found a workflow that
really made sense for them

to have their jobs be any better or faster
once you're delivering quality work.

You can be faster and deliver slop.

And I knew a lot of agencies that were
doing that.

They're like, We're 10 times faster.

And I said, Do you read the code you're
delivering?

Do you feel confident that you know that
what you're being paid to

do is actually being done? Is there a
quality level that has changed?

And every single one of them that were
like,

We're now 10 times faster, all of them
were like sheepishly like,

well.

You know, like they weren't actually
delivering at the level they should.

Mm-hmm.

And I was like, Titan will not deliver at
a lower level because we're using AI.

We will either not use AI or we'll use AI
in a way that allows us to deliver

at the same level. And at that point,

that wasn't the case. A year and a half
ago,

I think, one of my friends was like,

Hey Matt, I think that if you're gonna
have all these opinions,

you should use it yourself, because it's
gonna guide you and not just

be reading other people's code or
whatever,

but you should actually build an entire
application using it.

and so I was like, Okay, that's

That that makes sense. I've been wanting
to build this internal tool

for managing our our company. d I I
started in a cursor and then moved very
quickly

into Cloud Code. And I was able to find
what I think it's good at

and what it's bad at. And that was
December the December of twenty twenty
four.

So when every you you talked about the
Cambrian explosion.

Like that was right when I started using
it,

which was when everybody discovered,

you know, I used again, I had used AI to
write code years prior to that,

but that was when I built my own project.

And what I discovered is what is what I've
mentioned often in this podcast

is that ADHD, entrepreneur, running a
company with a bunch of other things going
on.

AI-based coding gives me the ability to
create in a way that I could never have

before because I'm not a full-time coder.

Yes.

I am a full-time business runner who has
coding ideas,

but no time to work on them. And so AI
enables me to be able to say,

Yeah.

I'm building this mass. I mean,

this puzzle that I'm building for LarCon
this year is.

hundreds of thousands of lines of code.

It's when I show it to somebody,

every single person's response is just
kind of like jaw drop.

Because it has been every night and
weekend and spare second for three

Mm-hmm.

or four months. No way I ever could have
done this because I don't have time,

because I'm running a freaking company.

At best, I could have prompted a couple of
my team members,

paid them to not work on client work for
several weeks,

and we would have gotten an eighth of what
we're working on here.

So it enables me to do things I never
could have done before.

I'm I'm sorry, I'm going longer than I
want to,

but basically

No, no, I love

it.

Okay, so you and I spoke, you know,

when I was six months into that journey.

So I was coding regularly. I was fully
comfortable using different agents

and stuff like that, but I was still
pretty limited to one or two projects.

And I was still pretty limited.

And our conversation, I just saw that you
were just going freaking wild.

You were just like, I'm building this and
I've got this running and blah,

blah, blah, blah. And it was just like a
next level of interaction with

it that I said, Well, now that I've opened
it up to there is a value in me as a,

again, a

technical, you know, programmer of 20
years who has no time to code.

So I've got aspects and interactions with
this AI that another person might not.

And it is okay for me to have an approach
to AI that's different than

my full-time team members to approach,

that's different than a junior developer
trying to learn.

And each of us should use them
differently.

And once I kind of had that freedom that I
think really unlocked when

we had the conversation, I was like,

I don't have to worry about is what I'm
doing what I want my developers

to do when they're working full time for a
full week on a client work

and we're delivering Titan level work.

That can be a different answer than what
I'm doing as the CEO building

a freaking video game on the internet for
people to play at LarCon

or building an internal tool that if it
breaks,

that just means that I can't use it until
I fix it,

right? Complete and so that opened up this
whole world.

Mm-hmm, mm-hmm.

so yeah, so that that conversation was
very valuable to me to see that

you and I and Ian Landsman and Aaron
Francis and all our other friends

who are like going wild building the stuff
that makes sense for us.

There's a there's a whole

nerdy world of having ten projects running
at a time and,

you know, like having your Mac Mini doing
other things.

that does not mean programmers should
lose their jobs.

That does not mean that's what my
developer should be doing.

And it's been so freeing and so fun.

I I love the love all of that,

first of all. but the point on quality
resonates particularly for

me because it's one of the things we pride
ourselves on,

rightly or wrongly, is I think we have a
pretty high quality product

and we care about quality a lot.

And I'm curious if this resonates for you.

one of the kind of perspective shifts that
I've I've pushed for within

the company a lot of the times is I
think early on when I AI

was producing low quality code.

might say, okay, well, we're not going to
use AI for everything.

Sometimes we have to handwrite it because
we have to maintain quality.

And one of the things, one of the sort
of messages I've tried to push

is if AI produces badly written code,

your reflex will be to say the AI is bad.

But my challenge to you would be,

actually you're bad at prompting the AI.

The results that the AI has given you of
low quality code are a result

of the inputs that you have given the AI.

And your role as a developer now is not
the skill of writing code.

Your role as a developer is the skill of
prompting an AI to produce good code.

And if you start to think about it in that
way,

then the perspective of of how you
interact with these models changes.

It's less of a blame thing of like,

it's stupid, it did the wrong thing.

And it's more of a what do I need to
change to make it do the right thing?

Because that I tell you, there are people
out there,

Yeah.

we both know them, who are somehow getting
AI to do the right thing.

And I think it's very easy to sort of
blame AI for like being bad or dumb.

When actually the the whole skill set now
is not how much code can you

as a human memorize, it's how much systems
thinking can you do and how well

can you communicate your ideas within a
structure or a workflow that allows

an LLM to thrive. Because that is the
skill set that a year from now,

two years from now, five years from now,

is going to pay exactly the salary you
have,

if not more, probably more. But the skill
set of memorizing all

of the functions that you used to have and
which order to do things in and

How to explode an array, or I was never a
good programmer,

so I don't know if that's like a good
example.

the the market value of of that skill set
is probably gonna be half what

Sure, yeah.

it was a year from now. And it's probably
gonna be significantly less than that

a year after that. and I I always think
that's a really useful framing.

you know, if you think of building a
product as moving a a box from point

A to point B, we used to walk and carry it
and we go,

like, here's the product.

And AI is now like a bicycle, right?

In all the ways. Like the computer was the
bicycle for mind,

AI is the bicycle for productivity.

And initially you fall off that bike a lot
and you go,

Well, this bike fucking sucks.

Like, no, you just haven't learned to ride
it yet.

And then you go, okay, well, I'm gonna
ride it slowly,

but not on this bumpy terrain,

because then I fall off and the bike
sucks.

Like, no, you need to learn to ride it on
the bumpy terrain.

Like, if the more we were able to see
everything as an us problem

and kind of re-ask, like, what can I be
doing?

Like

to change the outcome of what's happening.

I think that's just the the fastest path
to adapting your skill set

Mm-hmm.

for the reality of the world around you
versus wishing that your skill

set could stay the same and that the world
would stop changing.

Yeah. And I like that a lot. And one of
the things that I told my team

is I don't actually care how much you do
or don't use AI.

but every single one of you needs to be
an expert and how to use AI well,

which means you're going to understand
what it's good at and what it's not good
at,

and you're gonna be able to know when that
shifts.

'Cause for example, like there were times
where you could do the best prompt

in the world and it just can't do things.

But that happens less and less frequently.

But that still exists, right? There's the
AI's still not perfect.

A perfect prompt does not give you perfect
answer all the time.

But we're moving more and more towards the
world where prompting is the major,

like prompting ability is the major well,

prompting ability and understanding what
it's a good tool for and what it's not.

so I was like, that can't be the thing
holding you back.

You can't choose not to use AI because you
don't know how to prompt it

or you don't understand what it's good at.

You can choose to use it because you said,

I know what it's good at, and this is not
it.

And that's totally fine. That's a that is
an acceptable answer,

just like it has been for any other thing
where a client comes to

us asking for something and we say,

No, because we're the experts and that's
not what you need right now,

versus

No, because I don't understand it and I've
got my head in the sand.

Right. So that's been my big thing with my
team is be experts and then

Yeah. Yeah.

I trust your decision making abilities
once you're the experts.

But until you tell me I'm an expert in
prompting and identifying

Yeah.

the systems and knowing what makes sense
here,

then that's your job is to become the
expert.

There's there's a funny sub-layer to that
as well,

because there's like prompting ability is
is one obvious big topic.

Like how well do you prompt it will
determine how much of a good job it does.

But another one we found like very quickly
was,

as I mentioned a little bit earlier,

Ghost is a very big code base.

And unfortunately it's a very big code
base not built with a framework,

because when we started no JS frameworks
didn't exist.

So you have a sprawling code base that is
not shaped like any other code base that

it's been trained on. And so particularly
with the early models,

but still to this day to some extent.

you would see the agent just go like,

I don't understand this. I need to explore
further and get more context

and it'll read more files and it would
read more files and it'd read more files.

And so its context would fill up and we
go,

Yeah. Yep.

like, okay, no matter how well we prompt
it within this environment,

it's not producing a good result because
its context getting flooded.

But that's an interesting thing to
understand and articulate because

now you can say, well, there would be
economic benefit for the whole company.

If we actually decoupled some of this code
base into modules or services,

because now we're going to get better
results out of the AI as a result

of shifting the architecture a little bit.

And then the weird thing, you know,

that turns out is changing a code base to
be more AI friendly also makes

it more human friendly. And the developers
are actually like,

Yeah, I was gonna say the same thing.

We've been saying this for years.

You know, you're like, okay, well,

like, guess what? You get what you want.

and there's a bunch, there's a bunch of
things like that where getting better

Yep. Yep.

results out of the AI is a lot to do with
the prompt.

Yes.

But it's also to do with the

Yeah.

environment that the agent's working
within,

the tools that it has access to,

the amount of context that it has
outside of the prompt,

its ability to find additional context
beyond the scope of the prompt,

and how you interact with it,

of course. And so like having a good
understanding of the ways in which

you can influence its output just gives
you so much more power to do more with it.

Yeah.

And part of good prompting is
understanding how to communicate well

to agents and what agents are good or bad
at.

But part of good prompting is just good
communication.

And just as as you said, one of the
tangential benefits of making

a system better for AI is making it better
for the humans.

I also think communicating a prompt better
means you're more capable

and familiar with and comfortable with
communicating better with the rest

of your team about what you want.

And we've always said it, Titan,

we don't think the answer for clients is
to come to us for with a fifty page spec

document and say, we did all the planning
up front and now please build it,

Yeah.

because you have no idea what's gonna
happen.

But communicating extremely well and
efficiently and effectively about what
you're

gonna be doing in small spaces is really
valuable.

And it's something that is a skill that a
lot of developers don't have

is here's what I'm gonna do, here's the
things I wanna make sure

are not gonna happen. And that

That often reflects your developer
expertise that you've built over last 15
years.

Exactly. Yeah.

Here are the security issues I know that
are going to come up here.

Here are the scale issues that I think
will probably get touched on that one.

Mm-hmm.

Here's a gotcha, whatever. You're
communicating all those things.

And so as you become an expert here,

you are still honing in on the things that
make you uniquely qualified

to do this thing, but you're communicating
them.

Exactly.

You know, I one of the things I tell
people often is it's it's as if you've
taken

a job as a lead developer versus a staff
developer,

right? Like you're still.

responsible for understanding and
reviewing the work,

but you have delegated the work of typing
characters to somebody else.

And that somebody else may become your AI
more and more versus a junior developer

or a mid developer, but it's the shift
that you're going to make in your career

at some point anyway. So as long as we see
it like this was a natural part

of my development anyway, then it becomes
less scary.

Yep. Exactly. Couldn't have said about it.

Love it.

I love talking to you. Okay.

so we have a community thing where we
asked the people in

Ha ha.

the community how they're using AI in
their day-to-day life.

My friend Stephen Grant says we use chat
to plan our ill-fated,

never happened trip to Sydney in
September.

Unfortunately, it couldn't tell you the
state of our visa requirements getting
into

Australia. So you travel,

like I said, more than anybody I know.

are there ways? You mentioned you'd love
to have the the

The MCP for flights, but are there ways
where you are using AI for your your
travel

planning or anything else like that that
you were previously having

to do it manually?

Mm, honestly, not really. Not that much.

Yeah.

I mean I I don't know where our our
community was it Steven?

I don't know where he's from. But I'm very

Steven, yeah. He's from Scotland and

he was he's moved to Australia since but
i in the in the interim he

was trying to to kinda make that trip.

And so

so

it was like visa requirements for moving
there,

not not for like a vacation.

I imagine this was a

initial trip trying it out. I don't
remember the exact timeline that he did
this.

Yeah. Okay. Okay.

I think this was a vacation originally.

So

Yeah, yeah, yeah. Depending on I'm very
fortunate,

I have British Irish citizenship.

So for me, go like going to Australia as a
as a going to most places honestly

is a case of just show up and not really
having to think about visa at all.

Like there's a visa on arrival situation,

Yeah.

which is just a very lucky, fortunate
position to be in.

So I rarely in the past couple of
decades had to research very much about

visas and and how to resolve that.

But I do use it quite extensively

for the deep nuanced sort

of tangled edge cases that I often find
myself in.

Like I'm living in Thailand, but I'm
working remotely,

but my company's based in Singapore and
I'm staying 60 days.

Is that cool? Which one of these options?

Like I I can say, like, here's what I know
so far.

What might I be missing?

and I found it quite useful for for
answering those types of things.

for planning a trip in terms of like
where to go and what to see

and where to stay, I'd be very,

very skeptical of it. But for things that
are like laws or regulations

or things like that, where there's
concrete,

like there is a right answer. It's not
like we're gonna go and make some things
up.

I found it pretty useful for that.

Right.

And that's the most most interesting part
about Steven's kind of point here

is it's a you d used it for where to go,

but not the visa requirements aspect,

which is I flipped from what I would
definitely expect it to be good at.

I thought it would be really good at the
visa part.

So I gotta reach out to him and be like,

Tell me more of this about the story.

Yeah, yeah, I'd be curious.

So yeah. And I did just look and I don't
want to misspeak,

they'd live in New Zealand, not Australia
now.

And I know that that's a been for very
important for folks in New Zealand

to di differentiate the two. So

I thought New Zealand and Australia had
a free movement thing.

clearly I'm neither, so

Well, so at the time

that he sent this message, he was still
living in Scotland.

So So they were living in Scotland,

Got it. Stephen's man after my own heart.

they they wanted yes,

Geographically complicated.

yes. And he's a he's a sweetheart too.

yeah, they were living there,

they wanted to visit Australia,

eventually moved to New Zealand.

So well, awesome. So if people are
fascinated by John O'Nolan,

Got it gotta go.

as they should be, and also potentially
want to give him money,

but at least listen to things he's doing.

How do they find you? How do they sign up
for Ghost?

What else are you working on that they
should be?

Following.

Sure. find me on Twitter or as Aaron
calls it,

X The Everything App. my username is
John O'Nolan.

and check out Ghost. If you have ever
considered sending a newsletter

or would like to see some wonderful
software that is great

for running a subscription publishing
business,

if that has ever tickled your fancy.

Also great for podcasts. have a look at
Ghost dot org and and check it out.

Yeah.

And otherwise you can find me various
places all over the internet.

Brilliant.

And we will have links to everything that
we've mentioned here so far

in the show notes. So please check that
out.

John, I had such a good time and I also
talked more than I think I have

on any of these episodes ever.

So thank you so much for your curiosity
and participating in the conversation.

And just thanks for hanging out,

man.

Absolutely. It's been an absolute
pleasure.

Let's do it again.

Yeah, me as well. it sounds good.

And for the rest of you, we will see you
all next time.