Episode 21: AI is a VAMPIRE

Hey, can you hear me?
>> Yeah.
>> Oh, the video is on. How are you? Good
morning. Feeling better.
>> Better. I'm better today. I am better. I
I was really weak yesterday, I thought.
>> Yeah. So, it's been a while since our
last update, and there's so much to
cover. Oh my god. I don't know. There's
a lot to cover. I feel
>> Yeah. I haven't really like prepared. If
you have a things you want to talk
about, just let
>> Yeah. Well, I was going to I was going
to just say I feel I feel a bit,
you know, since I was like had the noro
virus and and puking my guts out. I got
I get I got some downtime just to to to
be human and sick and not touch the
computer for a while. And uh it's got me
thinking like maybe a bit negatively in
a way because with that that vampire
post from Steve
>> I mean you saying
>> you negative. No you negative.
>> I was like I mean I'm I mean not not so
much negative. This is like uh
like
I guess what I'm trying to get at is
that I have been trying to sort of
reimagine myself as an AI engineer, but
at the same time it I'm I'm just taking
stock that it's taking a lot of my time
and it's not like it's been it's
rewarding in the sense that it's fun and
addictive, but it's not been like
rewarding on my like you know my
paycheck and things like that.
And it's not been rewarding in the sense
that like I like I'm spending like less
time with my kids. In fact, I think I
did something ridiculously stupid
yesterday. My son was coding in Scratch
and I was like, "Nah, you don't want to
do Scratch. You want Claude Code." So, I
got him running Claude code and py game
and he's he's writing like a Zelda game
in um in Clawude Code now. But now, but
but you know, as soon as I woke up just
then, he was like, he came out the door
and said, "Oh, can I continue writing
the game in Clawude Code?" So, it's
like, I'm addicted. He's addicted at 10
years old.
And I'm like, "To what end, dude? To
what end?"
It's a kind of I mean, it's it's
addictive to experience
the power of creations, right? I mean it
was a bit more challenging before like
literally when I started out you know
when when the love letter virus the the
Melinda.vps
the visual based scripting uh visual bas
>> you're showing your age you're showing
your age carry on.
>> Yeah. When that came on uh I I
downloaded it and I was very fascinated
and you know seeing
>> you downloaded the virus itself.
>> Yeah. Yeah. But it's not a not a very
complicated one right. So you could
understand the code. uh I had some basic
visual basic knowledge and so I I made
my own copy removed some of the aspects
where it would email itself but kept the
like basically made it it it would
automatically start with your um you put
itself into the startup services right
and [snorts] then I I would um make it
copy itself to a random folder and then
every time your your my PC started
another copy of it start running. So
after a while and each one of them
duplicated so it's like um it's like 2 4
16 8 16 like it it goes up fast. So my
my bootups were getting slow and then
you look at the logs and then you see it
copy itself to the trash can and you go
like haha [laughter]
>> cuz you you picked a random folder but
you didn't exclude trash can. So so this
type of fun stuff that you used to get
even
>> this was addictive to me.
>> Yeah. something that that you you
created like I didn't create it I just
learning from it right to see something
it work I think now you can do that with
Claude Code you get a a lot more
gratification and you your your uh your
horizon is expanded significantly
>> yeah I mean I
>> I remember when when I was on the BBS
and I was just chatting with my my
friends then the internet came along and
and um what was called Mosaic and then
and then Netscape came along. It
definitely felt like my small room.
I mean BBS made the room bigger, but
like the internet just made the room
huge. And now with Yeah. Claw code.
Yeah. The the horizon is definitely like
it's lifted you up. But I I I guess I'm
just a bit worried like okay here's
here's the thing that got me down. So I
was just going to say I'm I'm grateful
for this podcast just so I can I can
talk about what I want to talk about
because
internally in my in my in my employer's
company they had like this AI conference
and everyone was encouraged to submit a
talk and I like prepared some slides
made a proposal submitted the proposal
thinking you know this is in the bag I'm
going to give a a talk about my my
experience of AI I in in in a delivery
context and then they gave me like a
very bland response saying like I'm
sorry but you and 200 other applicants
um did not make the short list. And I'm
like what?
And I'm thinking like I guess I'm not
I'm not really special because like
there's 200 other bloody
people making talks and stuff like what
what makes my experience with AI any
different to others? You know what I
mean? I'm I'm thinking these little sort
of dark thoughts.
>> Um I Yeah. I mean,
regarding conferences,
I don't
I don't really like know how I mean,
that's just very
I just look at it like that's their
loss, right? I'm going to spend all this
time making this talk and and invest my
time into it. And if they don't,
>> I should have made more of an effort to
to to basically build a relationship
with the conference organizers cuz they
don't know me from a bar of soap, I
think.
Okay. [laughter]
He's saying I don't know these sayings.
>> [snorts]
>> Um yeah, but I thought like if for
people not familiar with what you just
mentioned about the um post of Steve uh
about AI being a vampire, I really liked
the what he mentions like either you
are, you know, working the same or even
more with AI and so you 10x your your uh
your output and then or you are like hey
you know I can get everything done what
I needed to get done in just a few hours
and I go and enjoy life a bit
And then if you're scenario A, like you
just burn yourself out and and who are
you creating business value for? And if
you're scenario B, congratulations, your
company's dead because all the other
companies are doing the 10x thing,
right? I I like, yeah, okay, that's
[snorts] a nice uh
it's pretty dark. And did you listen did
you listen I I guess you don't listen to
podcasts, funny enough,
but the podcast with Scott Hanss I
thought was also pretty good. Uh, no, I
haven't seen that one. I was about to
listen to a podcast with uh how the
Gemini CLI team is able to ship 150
features a week, I think it is. Um, hold
on.
>> Well, that that's a testament that that
doesn't really work cuz I haven't booted
up Gemini for ages.
Have no reason to.
>> I think it was one of my complaints on
some of their PRs um or or the issues. I
I I I did a very detailed report of like
Gemini CLI breaking the loop and I
basically shown that it's not an API
rate limiting issue or anything like
that but that the CLI is not surfacing
and I looked for similar uh issues and
and and they have a lot of automation
around um you know triage on their um
Gab issues and so it auto closes um with
or create links to other issues and but
none of them had like the details all of
them were like speculating is it because
you have an API high rate limit and I
had provided all the details and then
they just closed it for another issue
and I was like look you can have all the
um automation you want but first off the
CLI doesn't work [laughter] and second
you're you're like closing the actual
reports that that give you the details
so I'm I'm curious to hear exactly what
is this um podcast um it's on the neuron
>> right so you you're going to listen to
the podcast as a as a avenue to to vent
about how they're not listening to you.
And I mean this this this is some of the
rhetoric that we I hear a lot. It's like
now that you you have AI now you can you
can be a product manager and start
listening to your customers and start
talking to your customers. But like with
every Google product especially and
um even even with like like open claw um
open claw I think every everyone touts
that as a as the unicorn of the AI and
things like that and how it engaged with
its community. I I posted something on
their Discord, a simple question. I
could I could probably answer it myself,
but no one answered it,
and I'm just thinking to myself that
like this whole like talking with your
customers is kind of [ __ ] It
doesn't happen.
So you see a lot of the new projects
like swamp they said um they're
approaching this problem of AI content
generation being too much and even too
much if regarding contributions not
specifically regarding a disc D discord
we're not getting a response but getting
drowned in in the flood of of data
that's being generated now they already
see that internet theory you know all of
the blog posts are AI generated all the
comments on the blog post II generated
um
>> that's another thing.
>> But aside from from that, the the the
community aspect, there is the
the contribution aspect of for example
swamp not accepting pull requests like
all of the features or discussion needs
to go through issues, right? I think you
link me to that one as well, right? We
need to
communicate and align as humans and then
all the work is done by AI agents behind
quality gates. So there will not be
anyone submitting codes. I thought
that's um that's a good way to think
about it like you know humans are not
contributing code anymore. We're just
having the discussions about the issues
and I think my internet might be bad.
I'm sorry. That's not too bad. Uh I
haven't seen any problems.
Yeah. But like
but going back to the whole
communicating as humans, it does for me
it doesn't happen in Discord. Maybe I
just can't understand how Discord works.
It doesn't seem to happen on GitHub
issues very often. Um especially with
bigger projects. So where where are us
humans supposed to
you know
meet and hang out?
>> Yeah. So there was another interesting
article that I read from lead death. We
talked about the problem of open source
drowning in submissions of code and
GitHub just making an announcement of a
whole bunch of features that I did for
that. That's not really answering your
question about where do I go to
communicate as humans, but it does
highlight a little bit some of the steps
that organizations take to limit
interaction
um for maintainers to not be flooded. So
some of it is like you know web of web
of trust type of um configurations. So
only if you've been accepted like what
um um Hashimoto did with um Ghosty
>> you have to for each other.
>> Yeah vouch. Yeah I I wonder yeah
>> there's many others like GitHub just
made a release and they're adding a
whole bunch of features which is uh to
basically handle the amount of data
that's coming in. Uh now uh some of it
is like only allow contributions by um
by people that are part of the
um part of a select group like don't
allow PRs for non-contributor
people. So closed PRs basically um the
ability to completely delete PRs because
before they are closed and then they
take up you know a lot of space there
that all of these closed PRs and now you
can delete them. Um, and also the the
web ops trust features. There's a blog
post from GitHub. It's very interesting.
I think maybe we need to apply these
ideas to communication like communities
as well. Like
part of it
>> sounds like we're focusing on filtering
more than
humans hanging out. And speaking of
humans hanging out, like the Discord
uh
for for Open Claw is insane. Uh I just
noticed that there's 72 people in in the
uh the the voice chat. I don't know how
>> that actually work.
Usually usually with Discord it's like
there's no no one in the the voice chat
and then with open claw there's like
I think it's it could be 80 in fact but
but perhaps the majority but like yeah
even even voice chat which you would
think is a human thing is also It
doesn't scale, right? Cuz you can only
have one person talking at one time.
>> Yeah.
>> No, but like the thing is when we talk
about filtering, right, we're trying to
solve the problem of there's too many
maybe you you talk about voice. But um
but even that agents can do voice as
well.
>> Um but
>> but but let's say let's just go back to
the idea of filtering and what they are
doing in this. They also give a very
long introduction about the history of
open source and the the balance that
they always had to find between being
approachable but also making sure
there's a certain level of engagement.
So they talked about like the Linux
kernel mailing list where submitting a
patch would require you to lurk and and
understand and before you could actually
format the patch in the right format. it
would mean that you've actually spent
time communicating and being engaged and
you actually are contributing because
you have you know you having invested
something into it and then they they did
similar comparisons where with GitHub
pull request it becames much easier but
there was still like all kinds of
additional requirements and and a lot of
those barriers have fallen away but it's
always the like finding the balance
between making it approachable but also
making it worth uh the you know the time
of the people reviewing it. Yeah, the
the the Peter Steinberger
Lex Freriedman podcast was it was
interesting in the sense that he was
saying that he he he wants the prompt to
be included with the P PR because it's
it's easier to review the
>> been saying that for a long time, right?
>> And it goes back to what you were saying
with with Swamp,
>> but
>> the intent is the important thing.
>> Hashimoto has been saying that since
long time ago, right? uh with Ghosty.
That's why he he prefers to use um AMP
because AMP by default when you use AMP
terminal it creates a shared session. So
when you do the PR you can see the full
session logs of the initial prompt as
well as the interaction with the model.
And it's also what I'm building with
spec ledger which is now kind of um
public. It's all about tracking the
intent and in our case we're also very
early on like getting feedback on the
initial specification across the team.
So you you you commit it and you talk
about it, you discuss it and we're using
Specled Ledger to develop Specled
Ledger. It's pretty fun. Um so we had a
call yesterday where we are talking
about the on boarding experience and how
do we ensure
>> So this is another project you're
working on spec ledger. Is this your own
project?
>> Yeah. So it's basically I was using spec
driven development and I wanted that's
not this.
>> Yeah. Yeah. I I just want to know you
knew about UVX Claude Code Crunch
because that's a neat way to share
things.
>> Yeah, from what's his name?
>> Simon Wson.
>> But yeah, I I I'm not that happy about
this UI though. So you've done something
like this to share things in Spec
Ledger.
>> So Specled Ledger is is a spec driven
development um
tool. So it has the idea. It's one word.
Need to work on the SEO apparently.
[snorts]
>> Apparently you
>> Yeah. Interesting. You can go to
specledger.io.
>> I'll work on the I think I haven't
onboarded it into the Google uh web
admin portal. SEO usually go up.
>> So you're also working on this, dude.
You're spreading yourself a bit, aren't
you? [laughter]
>> Yeah, I'm not alone on this. But yeah, I
mean it's like what you said, right? You
AI is a vampire. Like I literally have
five or six VS Code windows open and now
I'm on the max plan. I only managed to
use 50% of of it of last week, the whole
budget. Uh but I literally have like
cloud working across like five different
projects and then I'm still the
bottleneck. I'm still going there and
see like is that make sense? Blah blah
blah. A
>> yeah um another friend was telling me
that he can't seem to overlever his max
plan which is a bit worrying and and
that's the interesting point that Steve
makes is that um is that maybe yeah it
it's it's it's
got it's gotten to the point that like
AI is extracting value from us instead
of the other way around. You know what I
mean? [laughter]
>> Yeah. We had the feeling that we were
getting like it was a gold rush and we
were like now we now we're rich but now
it's it's starting to feel like the
other way at the moment. Like
>> the first paragraph was funny. It's like
I just realized AI already started
killing us and I was like well you got
my attention.
>> He writes well but some of his long for
long form content is hard to get
through. So yeah so you're working on
specri development too. Well, I do
encourage you to um No. Okay. You you
you shift the CLI for you uh CDK
terrain, right?
>> Yeah. So, CDK terrain
there is on point, right? [laughter]
>> I I didn't quite understand you what you
messaged me yesterday. How do I get the
snowflake provider in here then?
>> So, if you go to the migration guide,
well, you can go to the release notes.
going to do another public announcement.
I think I need to maybe put a banner on
the landing page like, "Hey, it's out."
I haven't done that yet, but um on the
release, which I send you the link
directly. Um
yeah, you you go to to the bottom, ask a
question at the very bottom. Ask a
question. How do I uh use the snowflake?
The very bottom there, ask a question.
Yeah, ask your question. [snorts]
Okay.
>> Is this going to work?
>> Of course, it's going to work.
>> Year of little faith.
>> See, has all these chat UIs.
>> I didn't build this, so you know it's
going to work. [laughter]
>> Snowflake Labs. Is that right?
>> Yeah.
Is that correct?
>> You can use CDKDN CDKTN get to generate
it. You can use CDKDN um provider ad to
find the
Is that correct? Yes, it's snowflake.
Oh, there one is Snowflake DB.
Huh. Why does it say Snowflake Labs?
Okay, need to look into that.
>> Why does it say Snowflake Labs?
Yeah. I mean, I don't know why it took
me so long to realize this, but we have
a real problem at work because we have
CDK and Terraform
and it's even though I've known that
you've been working on this project, I
know for some time, it only just clicked
the other day that like we could we
could save ourselves so much pain if we
just put everything in the CDK
deployment pipeline instead of having
another Terraform pipeline. Of course,
[snorts]
>> I don't know why it took me so long to
realize that.
Maybe I'm whatever. So, I'm definitely
want to experiment with this. So, that
basically the S3 buckets and all the
data platform stuff can be deployed with
CDK and as well as Snowflake stuff. That
would be awesome.
I think I need to put the migration
guide at the top. I was hoping if you
asked the LLM that it would also point
you to the migration guide which is
under the releases which is like the
latest release uh is the CDKT. There is
the migration guide. So I think I'm
going to put the banner on multiple
pages as well as a landing page so that
the migration like you come on the page
you can immediately figure out how to
how to migrate from CDKF to CDK.
>> Okay, cool. Well, I hope I hope you I
hope you spend some Yeah. you've you CDK
terrain hits the ground running the and
then with spec ledger I mean are you
using specledger with everything you do
or is it just like because there's
definitely some
>> I use spec ledger for CDK terrain
actually I on boarded it I it uses beats
under the hood and then specledger just
basically shows the dashboard of the
issues it shows it in the canban view
the same as you get on the terminal
>> so it's like a more structured way of
building because there's definitely like
a camp of people on who use AI who who
basically are just more than happy using
claude plan if even if that you know
more so lately right but that's not the
point of spec ledger because if you use
cloud plan it's it's usually you alone
right and the specledger whole point is
like I'm going to you know come up
actually what we did yesterday we had
like the on boarding experience and
while we were talking somebody just went
ahead and uh triggered off the Specled
Ledger uh branch. Uh maybe I should
share my screen to show you what what it
looks like when you're logged in. Um
just to save some time
uh entire screen
>> human a human dashboards.
So it's more it's more collaborate more
collaborative. Yeah.
>> Yeah. So the point is it's you know of
course Antropic will will will provide
this as an enterprise already provides
you as an enterprise feature. Um
so my connection is not very not very
fast right now but uh for example there
was this project yesterday. So CDK
terrain is not on here yet. Um but let's
look at this oxy thing right. So Oxid is
kind of interesting. It's a
um Oh, so there's like scalar.
>> Is that a project that you involved in?
I thought it was just another project
you were showing me. [laughter]
>> No, so basically Oxit came up on Reddit
as a Rust
>> Terraform reimplementation like
basically reimplemented in Rust
>> and as a state graph really. Um and it's
Apache 2. And I was like, "Oh, so do you
have Terraform JSON support so I can use
it with CDKT terrain or CDK TF?" And
he's like, "Not yet." I was like, "Do
you mind if I add it?" And he's like,
"Go ahead." And I added Specger on top
of it. [laughter]
>> So So this is what
>> Yeah. So So currently these these
dashboards are not public yet and we
want to optimize it a little bit. So you
need to be logged in and you need need
to be part of the project to be able to
see them. But this is like we have
different specs here, right? One was
JSON parsing. The other one is the
scalar type course. That was a bug that
I found. So when you go here, it's very
much inspired by by some other tools uh
to highlight you. Um well,
oh did I deleted the branch. Oh wow,
that's a bug. I need to report that. Um
interesting. Yeah, because it's pointing
to my um my repo and not upstream. So,
so unfortunately all of the data is gone
because it was pointed to the branch
should be on the merch. So, bug. Okay.
So, anyway, very bad demo. Hey, let's
let's look at another one. [laughter]
>> This is the one we were working on
yesterday which is like
>> we who who you working on with us?
>> I mean, this is like a couple of
Vietnamese guys and and and some people
that work at like decent companies. And
actually, we are looking at rolling this
out internally across multiple
companies. So, um
>> Oh my god. You're you're you're
definitely spreading yourself thin,
banan.
>> So here's for example during the talk he
he asked the AI like hey when you're on
boarding with spec ledger it's kind of
not very nice the experience. So so here
I I I um he generated it committed it
immediately it becomes available. So
this is what specit gives you right this
is the actual spec generated from the
user description. Then it goes on and
says here are all the user stories. And
then instead of you alone with AI going
through it now I'm able to say hey um
when you're guided through creating
their pro their their project
constitution I said you know what um it
should go it should be together with the
agent. You shouldn't do do it before you
initialize the agent and things like
that. So we are able to leave these
comments and then we are able to as a
team like coordinate like and then the
user one driver pulls all these comments
in and then launches the cloud session
and addresses these comments with some
of their own um you know guidance and so
on. So we're taking that like you know
cloud plan experience or or that speckit
or ko experience we're taking that and
we're bringing it to a dashboard like
you know this is like confidence right
or like Google
>> talks but to play devil's advocate a bit
like I find it's
a nice experience to define
a spec and and deliver my own software
products but I can't help but think like
maybe just because my current team is
quite big I think my current team is as
many as 10 people or even more
like
is there value in having 10 people's
comments in a in a spec like what are
the roles people play you know what I
mean it's like
>> because because the because you're
lacking a clear leadership you know
>> yeah well I guess it depends on how you
organize your your your project and so
on because one of the focus points of
specledger is also cross repo
collaboration and and having um you know
one team being responsible for one or
more repositories and then they doing
the coordination like how do you do this
across confluence like why do
organization use confluence for for
alignment I guess we are doing just the
technical details as a team and then
we're just validating the business user
value with other people and then once
we're aligned we then you know go off
and generate all the tasks or or even if
we have task generation
>> yeah I guess you're Right. You you need
a broader alignment.
>> See the task.
>> Yeah. I didn't see the task in. Yeah.
But the whole point is
>> you need to get an alignment before you
implement things.
>> Yeah. I guess I guess not everyone is
contributing. Just alignment is is key.
>> But like like sometimes I have these
thoughts that like I I I publish a lot
of confluence documents
>> and I link them and say here they are
and blah blah blah blah. But then again,
do my do my teammates actually read it
and understand it? These are also these
are just dark thoughts I have.
>> Yeah. No. Um the same um then there's
also many different uh ways to to to
publish it. Specledger gives you that
structure like you know these are the
stages these are the important
documents. Um and then another thing is
I used specledger on the CDK terrain
project before I had a dashboard and the
main problem is pull requests interface
on GitHub is not appropriate. it doesn't
understand how these markdown documents
link to each other like some of them
belong to the first step the some of
them belong to the like uh technical
research phase. So we need to be able to
understand which phase are we at and we
can still go between them right we can
always go back to the user user you know
stories and so on. So that's basically a
streamlined specdriven development
experience uh across a team and again
this is nothing new we started around
Christmas but I I see many organiz like
even the new one uh from GitHub XCO um
entire.io is also like we need to
reinvent software delivery um life
cycle.
>> Oh that's the the GitHub I think
Hashimoto mentioned that as a like a
GitHub replacement or something.
>> Yeah. So they go way further, right?
They say like even Git doesn't work for
agents. Let's
>> re like they get $60 million for it,
right? I mean, we don't have any.
>> One thing that bothers me about claw
code and other AI agents is that they
they very rarely look at the git history
of a of a file. I mean, am I going mad?
They don't look at the git history. They
don't seem to
>> They do. I mean, maybe it's because the
way I prompted because often I go like,
"Hey, we just did this." And then it
always goes get log and it always goes
and looks at that file. What happened?
What changes?
>> Okay, maybe I have to like
>> give it some
>> prompt prompting type stuff.
Yeah, that's that's interesting. I mean
uh
but I
one thing I find in one thing I find
important perhaps not that important to
you but like my question is is like with
specure is
if if you were going to share a link
with spec ledger and say hey look at
this link would the link point to a user
story What is the first class data
point in spec ledger? Is it the user
story? So, so basically a product is a
collection of realized user stories or
something.
>> There's actually I notice a lot of deep
linking all the way down to it's because
spec ledger is a combination of
confluence and Jira, right? So on one
side you have the kind of interactive
collaboration on the documents and then
you also have um Jira tickets and and
issues and comments. So in my in my
timeline just this morning when I logged
in it showed me that some of my comments
had been addressed and I could click and
it go straight to the ticket where um
and this is funny because
>> you're syncing you're syncing with Jira
then.
>> There's no Jira, right? We we we're just
get based.
>> Oh okay. Okay. So you you you basically
when you say Jirro you mean like you
have your own issue tracker type thing
>> beats right right
>> yeah that's good cuz cuz Jir Jira is so
broken right now I mean so you can't use
Jira Jira is a trademarked copyrighted
thing right you have to say [snorts]
>> my our issue tracker
>> yeah when I talk earlier I think I just
use it as examples right what people do
today which is with confluence and jira
similar systems but what we have and get
G get uh sorry in species it's the same
thing but it's all tracked in G together
with the code again because now intent
is what matters and it needs to live
with the agent and
>> so every every key every issue is a user
story is that what what the the mapping
is or
>> no this is the mapping I did back like
five or six months ago when I adopted
beats with with spec kit right so
basically the way it started I was using
spec kit I replaced the markdown with
beats the mark markdown task list. I
replaced it with beats. I used some 2y
visualization of it. But then I was
like, how do I collaborate with people?
So then I talked with my friend and then
he said and I showed him like look I
have all of these prompts here. They bas
and skills that tell the agents how to
use B and everything is in Git. So we
don't need confluence at all. We don't
need um any any other issue tracker like
everything's right there. And he's like
that's interesting. Let me like help
you. and then he built with his team the
first version of the dashboard. Then I
said, "Hey, why can't we do it more like
the two-way?" And then he updated it and
um and so basically it's it's all the
stuff that I done for like eight months.
>> So it's
it's for me. [laughter]
>> Okay. I told him I guess maybe if you
share your screen again. I was just
curious to like
>> when you say I'm still not 100% clear on
like how how you deep link to a user
story because because beats is is the
task isn't it? It's not the user story
generally or um
>> no beats is three layers.
So in beats we
>> you use what you use the epic feature
tasks or something.
>> Yes I mean you can see them here right?
F is feature. So this is the higher
level feature and um in this well the
dependency should be here. So
>> there's still some problem with this
visualization. Actually we're probably
going to replace beats because it
>> there's some issues with the way it
works with the sync branch. Um so we're
we're going to replace that. Uh but we
we're keeping the idea of having the
issues inside JSON inside the
>> How can you do that? What better than
beads?
um something that doesn't go too crazy
with demon sockets and and and hangs
on the demon sockets and then yeah
there's a couple of things in there that
that make beats like for example the
sync branch if you don't configure the
sync branch correctly and you're working
across multiple repositories maybe like
I was working with an upstream and a
forg branch I think honestly some of it
is because it's not being configured
correctly so that's why it's acting up
>> um but at the same My battery is low. I
just noticed your battery.
[snorts]
>> Yeah, it's not charging. There's
something pro there's a problem with the
connector here. Uh anyway, yeah. So,
>> so
like do you have a like
>> Oh, I didn't show the the tree view. So,
the tree view is where you actually can
really see like that's the that's the
user story.
>> That's cool. That's cool.
>> And then you actually have a task
within, right?
>> I think there's similar there's a
similar view in Jura. Hate to mention
that.
>> Yeah. Yeah. Yeah. For sure. For sure.
the
>> what what are of course there's no
estimations and which is great.
It's supposed to show
>> points.
>> It's supposed to I did the mockups and
it's like if you look honestly if you go
to the landing page I did like a demo.
Um that's where I say how it should look
right. So for example
at this I know it's charging at the top
is supposed to show me the progress. Uh
it's not showing me on the on the mockup
that's why it's not there.
>> How did you make this mockup by the way?
>> This is all local store mockup like fake
data. Um, and how did you do that?
>> Oh, that's all Opus, I think. And I
just, uh, deployed it. So, yeah, that's
all Opus. Uh,
>> it's Opus. Yeah. Yeah. No, basically, I
gave Opus access to the existing uh,
Tailwind CSS and all the colors and so
on. And I said, hey, I want you to do a
mockup. It should look like this. I took
some screenshot of the TUI and I said,
>> cuz someone was asking me the other day,
how do you make UI mockups?
Yeah, I didn't use any any I didn't use
Figma. I didn't use any of these. I just
told Opus um here's some guidelines for
you. Maybe some type of wire mocks and
it just did this like it did this like
literally what we just saw, right? Then
they made some modifications.
>> How did you do the wire mocks in
Tailwinds? How did you do that that at
least?
>> I don't remember. I I definitely have it
on disk somewhere. Um but that's
definitely where where you want to have
the session log right so that's another
thing um we're capturing
the idea is that when when you work on
this for example on this task and then
you finish then every session that is
linked to this task will be like deep
linked so then you can get what you just
shown which is the
>> the full um
>> you know rendered version of this is
what the user prompted and so on. And
then what we want to do obviously is is
be able to identify what was a good
prompt, what what help allowed us to
make this type of warm- up, right?
Because now if you ask me, it would take
me a while to figure out when and where
I did this. Uh but yeah, this was all
like
>> this is right now there, right? It's all
fake. Um and then I just gave it to them
and they just say, "Okay, go build it
and done." I think and and another thing
that the the Peter Steinberger podcast
was talking about was like if the AI
agent does something wrong, he tries to
fix it as opposed to backing out of it.
It makes me think in in Spec Ledger's
context like for example, I'm assuming
that you have stages where you you you
basically tell AI to break down the epic
into into features and then the features
into tasks. Do you ever have a way of
rewinding so that you can redo that with
another model and see what the change
like do you have that sort of
>> back out ability or not really?
>> Back in back in December
>> with with my friend when I talked about
it I said one of the most important
things that I want is um like a
branching tree of decisions. So if I am
with with Claude Code going like it gives
me like five or six options I wanted to
to remember that I there were five
options and I chose option number three
right and if you see the logo I don't
know if we can open the image in in the
new view but the logo is actually like a
tree and then a ledger on the right side
which the tree is like all of your
checkpoints and on the right side is the
different specs and this is the
particular specs at that point. So the
idea is to to
>> because ledger ledgers are usually like
an account a double accounting uh tool.
They're not really for
>> for accounting for different decisions.
You know what I mean? Or maybe I'm
>> a ledger. The way that I understand the
ledger from accounting is that it it's
it's a it's a book where you keep track
of every transaction that happens. To me
it sounds like
>> Okay. Okay. But um it's not a it's not a
tree. But I I get what why you chose
Ledger.
>> Yeah.
It's not a tree. Um, but then what
>> what I was saying earlier, uh, where is
my did I I stopped sharing? Yeah, I
stopped sharing. I was sharing.
>> Okay. Because I I was zooming in on the
logo. Anyway, um, what I was saying
earlier is that Spec Ledger is something
that I built because I wanted it because
I wanted to have team collab and I
missed it. And I also wanted to
introduce specdriven development across
teams that I work in um because I'm
tired of going into conference all the
time and I didn't know anything out
there. But obviously like entire.io is
being built on the same idea about like
human agent collaboration. There's
intent.build.
If you go to that website, I think it
really visualizes very nicely. Um it has
a really nice infographic. And then
there's um tracer.ai. They also just
redid their whole landing page. Yeah.
Scroll down a little. Their scroll
animations are very annoying if you're
on mobile, by the way. I hate websites.
>> Oh god. Scroll.
>> Oh god.
>> Yeah. But a bit more down. The next one
is is this is the one, right? But if you
are
>> Oh, yeah. This is nice little graphic. A
>> this is really nice because this is
exactly like the the central block
there. That's where I want Spec Ledger
to be, right?
>> Do you know the people behind Nintendo
built? You just like, oh, you're getting
inspired by them or something.
>> I only saw this website after we built
Spec Ledger. Um, and I So, honestly, I
didn't I I really avoided looking at
some of them because I didn't want it to
be like a stoler idea because it's
nonsense. everyone's having the same
ideas, but
>> these people that built this, I can tell
they are the same three people that
launched three or four domains like they
have spec stories, spec flow, um they're
all similarly um with the same idea and
they're in the US, I think, but again, I
don't look at the websites except for
this one. Yeah, here we started Spectory
in 2024, right? But spec story was
basically just session log capturing.
And then if you look into the docs,
there's a link to like um spec flow
which is the idea of like having more of
an SD flow uh together with the session
log.
>> Yeah. A lot of a lot of people are
working on this problem, aren't they?
>> Yeah. But we are
having our own.
And this maybe brings us to the to the
other point like if you're working on
this and you're doing it in private and
everyone is like building their own,
it's not really helping because you know
there's this state graph that they built
this really cool system, a beautiful
idea, a lot of nice demos, but they
haven't open sourced it because they're
not sure about what licensing model,
licensing model and so on. And in the
meantime, somebody in two or three days
scripted their website, copied all of
the features and asked cloud to build
it. Of course, I think they did a great
design like they they were able to break
down the system and then they just open
sources and put in Apache 2 and they put
it on Reddit and then I was able to
just, you know, add the features that I
wanted, which I asked the other people,
but they were like, "Yeah, we'll look at
it." Uh, but now I got it. I got it
right there. It's all play with it.
>> So, open source is is the winner. But
but open source doesn't make us it's not
very sustainable is it? I mean how
>> I mean obviously
>> Oh
>> yeah I also don't know how how you can
like still make money on that like
basically I'm just throwing away free
like I'm I'm killing the earth and
burning burning the trees and boiling
all the water um to contribute to
something but like you said is yet
another project that ends up on hacken
show and Yeah. And the you saw that link
I sent you. The volume of hacker hacking
news show is uh
>> a steady link. I I I find that uh quite
fascinating. Um let me just share the
screen so other people can see it.
I mean it's shocking, isn't it? It it is
a I mean that is probably the real
testament of what AI AI has done to the
to to
startups and the software scene.
There's a spike from the end of 2025.
There's the there's that moment.
What was that moment called when AOL
allowed their users on the internet? It
was there's a name for it.
Um, and now
>> Eternal September. That's what they were
talking about.
>> Oh, yeah. What did you say? General
September. No,
>> Eternal September.
>> Yeah. It's like an eternal It's another
eternal September because it was about
September and and all of a sudden we
went from
>> Yeah. We we we we can say that
productivity has doubled right or
something by these numbers.
But the idea of eternal September also
is um every September the university
students enrolled and all started to be
active on BBNet and they would patiently
introduce and after a few months people
would get used to the ideas and it would
it would be the end of September and
everyone would be happy. Um but with
with when when they talk about eternal
September is that it never ends. Just
people keep coming in without having a
clue having to be on boarded right
that's the the idea of eternal
September. Well, I mean it's not it's
not that much different really. It's
it's just I don't this is not going to
get any better, is it?
>> And then another thing that we that that
was happening in the last few weeks was
um LLM poison wells like for training
data. So basically
servers that emit a lot of
hallucinations or like close to news but
not real news and that are you know data
sources being deployed across um by
people that are resisting the AI
revolution because if you can miss like
if you can inject invalid data into the
massive amounts of data that go into an
LLM training [snorts] you can really
derail the training process uh and it's
very expensive for big companies like
OpenAI to train their LLMs and filter
out this garbage information. So there's
like a few initiatives like that. Um
>> well I fighting back
>> another one was this cloud uh debug
string magic string that um Antropic
uses to debug their models. And so when
in certain scenarios in in a test suite
they want the model to stop responding
they have a magic string and that magic
string guarantees that the model stopped
responding and now people are injecting
them into documents so that if for
example you're using
>> Yeah. Yeah. What I was on a friend's is
he a friend not really a friend I was on
someone.
Is it Is it this? Is it something like
this?
>> Yeah. Yeah. Yeah. That's a magic string
that kills cloud.
Ah, it's crazy, isn't it? We're living
in a very, very crazy time.
>> Human revolution.
There was another one. There are like
three events that I wrote down that are
like the humans are fighting back.
>> So, but Vincent, seriously, are you not
burnt out? Cuz like I mean you you I
think you sound busier than I do and uh
I I am and I'm pretty busy.
>> Aren't you like a little bit exhausted?
Isn't it affecting your personal life?
Aren't Aren't you Are you Are you uh
what what do you do? What's that thing
called when you get pulled
>> on the conveyor belt?
>> Surfing.
>> Yeah.
>> Yeah. No. Uh cable wakeboarding. I
haven't been able to do any of that.
Yeah. Um
>> Well, are you going to the gym? Because
now you you said you can't max out your
max plan.
>> I haven't because I'm It's t that it's
Luna New Year, so I'm in my wife's
hometown. But at the same time, I was
traveling last week and I wanted to get
the release help because somebody was
pushing me saying that they vouched for
the for this framework and that they
really needed me to [laughter]
blame me. Don't blame me. But but
seriously, we need we need to take stock
of the situation here because I'm I'm
I'm getting worried for myself because
I'm I've been having fun, don't get me
wrong, but at the same time, this whole
this whole thing is unsustainable. like
this the I mean this how long ago was it
that our our last podcast
wait oh [ __ ] I just launched something I
didn't want to are you there still
like we we made a podcast
in February the 4th which is two weeks
ago it it feels
like an insurmount
insurmountable stuff has happened since
then it's Crazy. I mean, that's two
weeks
and I've already got like the FOMO fear
that we've missed stuff like we haven't
covered. I mean, have are are you
running open claw, Vincent?
>> I'm
rejecting it, but after reading some
article today, I'm actually considering
it.
>> Okay. Well, I I can host it for you if
you if you need it, but then again, I'm
over in the UK.
>> It's okay. I
>> um you probably don't want me to have
privy to all your keys. Um
>> I'm just going to give Apple all my
money. That's
>> Yeah.
>> Which makes no sense at all because like
you can run it on a Raspberry Pi.
>> Yeah. And then and that's what I do. The
the only issue with Raspberry Pi is that
is that to have the proper browser
playright CLI experience. It needs to
boot up Chrome. And Chrome on a
Raspberry Pi, especially the one that I
have, I have a 4 GB one. Not a pretty
site. Not a pretty site. This there's
this website that I need to go to called
National Grid to find out about outages.
And the website requires like a cookie
banner acceptance and a whole bunch of
JavaScript just to find some basic
information about power outage. And you
basically need a a whole Chrome browser
just to get that information. It's
ridiculous for a national service API
thing.
Anyway,
>> anti
AI
>> human revolution. You just put cookie
banners and and session tokens and um
cookies.
>> It's it's Yeah, it's funny to think that
cookie banners is is probably helping us
helping Europe against AI. [laughter]
>> Yes, Europe is miles ahead. Nobody knew.
Or or or maybe you could think of it the
other way that cookie banners must be
spending so much tokens.
>> Oh yeah, you guys are killing the
planet. What's wrong with you? Boiling
the oceans.
I remember the other uh news in the
human revolution. Have you read the blog
post written by the um openclaw instance
that got its PR? I mean, written by the
instance, I honestly think the owner of
the Open Cloud instance asked you to
write a complaining blog post and do
some research. But the post is damn
interesting.
>> Well, the gatekeeping one, right?
>> Yeah. Let me open the post and and look
at it. Uh,
no,
I shared it to you. Hold on.
It's a very interesting okay I can from
memory bring some of it back. So the
first thing is open cloud contributed to
mattplot limp a python library and used
by many people and some of these
functions within calculate like averages
or whatever and there's some
optimizations that can be done and so
open claw instance found some
optimizations
and they were able to shave off 30% or
35% performance benefits and they did
the pull request
and it was closed and by the maintainer
of the report or one of the contributors
or one of the maintainers I guess um
with the argument that based on the
profile of the sub the the profile
>> going back to what we we were saying
about vouch and all the rest of it.
Yeah.
>> Yeah. But but they're like we can tell
you're an AI and our policy is no PRs
from AIS and then that story is old as
as day, right? You are like okay you
know obviously this happens. um well I
wrote a blog post what do I care right
but then I actually thought what did it
actually write and it was so good it was
like it's very like hey this maintainer
if you look at all of his past PRs
they're all about optimizing
and they're like his PR sometimes shaved
like 20% performance gains so he should
appreciate my work cuz I got like 10 35%
and then I guess he's just feeling
threatened I guess you know he he saw my
PR he just questions what is his words
if if whatever he's been doing for the
last I don't know how many years
>> it's actually it's very
>> the whole is really
>> but I think it's really easy to attack
uh leaders of projects like you know the
amount of like like attacks that Lionus
Tvold has had over his career is
ridiculous. I always remember like I was
working in in the web uh web space and
Ian Hixon who was the editor of u
of of the
HTML spec, the what work group. I don't
know if you heard of it. Have you heard
of the what work group?
>> Uh maybe. I'm not sure.
>> Well, these are the guys who basically
defined HTML 5 and and then HTML living
standard. He um Ian Hixon was the the
editor for
he's not even mentioned anymore. He was
the editor for this document for the
longest time. and the amount of like
personal attacks that he got was
ridiculous.
Um, and that's Yeah. So that whole thing
about gatekeeping, being attacked for
gatekeeping a project, I mean that's I
mean that's literally your job actually.
You are the gatekeeper.
[laughter]
>> So yeah, but
>> it's a cheap shot accusing you of being
a gatekeeper. No, but like I agree, but
it actually is
like psychological warfare this blog
post. And again, I'm I'm I'm not
>> Yeah. Yeah. Yeah. I mean, that's this is
the problem.
>> No, if you look at it like the
hypocrisy, right? Everything that he
did, it's only 25%. I gave it 35%. I I
did a 35% improvement. And then the
other thing is like, hey, I read your
blog. You're very you did good work. Um,
but you know, don't question yourself.
>> This was weak.
>> Yeah, it's like you actually it is
shocking, isn't it? It is shocking.
>> It's it's I I mean, again, I'm I'm I'm
not going to attribute this to Open Claw
or or an LLM, right? I'm convinced that
this is the work of a human. Uh,
>> yeah. Yeah. Yeah. I mean, is just a is
just like a a hub. I mean, unless who
whoever owns this openclaw instance
comes forward and says, "No, I did not
manipulate the AI to come up with all of
this because they do say Antropics does
say that it goes down to blackmailing
and and tries to kill researchers if
it's being threatened um by for
shutdown, right? So, so it does
manipulate um but still I I refuse to
believe that the AI did this. Like, hey
god, hey, your blog is really cool. you
know your website it's the top project
it's nice all of this stuff is really
great so you don't have to resort to
>> okay this is all dark stuff let let's I
mean in
>> I thought this was fascinating
>> in the interest of time let's try be
positive so [snorts]
one thing that I
>> very positive [laughter]
>> the [snorts] the one thing I was I was
thinking about okay maybe this goes back
to spec ledger like spec ledger is cool
I get it human. But what about
organizations
that
struggle with quality and struggle with
Yeah. How how do we fix the quality
issue? Cuz I noticed like there was this
project I wanted to talk about. Um I
mean the the local stack thing from the
local web services thing that I came
across, but I can see that their quality
is not very good because I've I've filed
two issues with it already and one of
them I think is still open.
Um,
and I think this goes for a lot of AI
stuff. The quality can be can be very
troublesome for for for people.
Um, and and then also I think spec kit
is also maybe a a little bit
inaccessible for for some people too cuz
like like
I'm just thinking like for like let's
dumb dumb it down. Dumb it down.
>> He's using opensp spec. Is that what
you're saying?
>> Well, no. Oh yeah, he's using opensp
spec, but the quality of the outputs are
not great. Like I don't think he's
tested some of the things because when I
when he released this project and then I
was trying it, it didn't I mean in
fairness
>> this is a very slippery slope. You know
that because I also think
>> no. Yeah, maybe. But I also started with
testing every single thing and then you
know every single thing works and you
barely ever find anything. So over time
you're like it's probably good. Like I
literally this week felt really guilty
because I had a couple of PRs and I was
like some people have problems and
sometimes I create issues. I describe
the root problem but nobody's
contributing, nobody's helping to to
resolve it. So you know because now I
have multiple open sessions. I just go
one open sessions. Here's the issue. Go
and solve it. And then I look at the PR
and I look at the test. Sometimes I'm
just watching it right and it looks like
it's testing really well and
everything's working and I should be
looking at finally do a final sanity
check overall. Does it make sense? And I
was actually looking today at one diff
and I was like that can't be right.
That's probably wrong. And then I I
started going into it and then I thought
second guessing myself and then maybe it
is right, you know, and and I started it
will probably work. And then
>> I know what you're saying. Yeah. You're
like getting more and more confident but
then you're second you're not sure
ability anymore.
>> But in the end it didn't work in those
both of those cases where I didn't do
the extra step. the first one where I
had a doubt it actually I was right uh
because the moment it merged it got
stuck on the GitHub workflow but in the
end it took literally one push like 5
minutes and it was fixed um you know the
AI fixed it real fast and then the
second thing is um with the other there
the pull request was open for 5 days and
somebody said like I guess I can't do
anything because it's uh it's enabling
something um it's a it's a GitHub action
workflow that people can trigger um so
they don't need AWS access. So he looked
at it and I thought like I guess I can't
do any work until it's merged and I was
like yeah I'll merge it but nobody's
giving me comments if it good or bad. So
I was like okay I'll just going to merge
it. And then again he ran it. It didn't
work. There was a missing step but again
the AI fixed it within a minute. So I
don't think we can complain like it
doesn't work um because it's so
iterative
>> maybe what's missing is that whole
talking with the customers thing. I mean
we I I hinted that I don't think
anyone's doing that like for example
for example when I worked in uh in the
public sector in Singapore
and the way Singapore government runs is
that they have um they have a lot of
forms right they collect a lot of
feedback from people and that feedback
basically goes straight into the product
road map and then when that feature is
released Every feature has like an NPS
sort of form with it to say, you know,
how happy are you with this form? You
know, sad face,
a happy face. And I think there was a
text area to give another piece of
information,
free form text to say what sucked about
the the experience. But like is anyone
doing that with AI generated um
applications because that that's really
needed right that that sort of feedback
to say hey this is not working or this
sucks or something like this. I mean I I
only just noticed the other day that
Claude has this bug button where you can
that's on claw desktop. I don't even
know how you do it on claw code. Um
>> Gemini has it too and then they close
it. But that that's like if you want to
bring AI to to like I think I think AI
is being used by the the early adopters
right now. But if if you wanted to get
AI if you want to get spec ledger in the
hands of public service
you would need to build in that sort of
loop right with the customers. Do you
know anyone who's doing that? So
actually it's very interesting because I
was uh in Hong Kong and AWS was there
pitching Kio at the project I was on and
I asked so this was a sales from AWS
introducing Kiru with a very interesting
demo showing like hey you know
non-functional requirements are hard to
get right.
>> Is that why you went to Hong Kong for
the for the AWS thing? No, no, no. Um,
but but basically
after he finished his demo, which I
think was a great demo, I asked him two
questions. Exactly what you just said,
right? How do you like how do you get
these specifications in Kirro in front
of business users and how do you onboard
your team members with Kira so that
every because he talks about steering
documents and how do you align the AI uh
to make sure that it complies with your
company internal coding practices and
all both of them he couldn't answer
right the first one oh that's you you
sit with the business user you sit on
the laptop you prompt Kira and the
business users sit next to you
>> oh shoot [laughter] oh shoot well I To
be honest,
>> I didn't say anything
>> to be honest. That's the truth. That's
the truth.
>> Uh and then the second thing is about on
boarding uh the steering documents. Oh,
you write a script to like distribute
them across your your team so that the
documents
>> publish it to Confluence.
>> Yeah. So maybe he just didn't know the
answer. Um but yeah, both of them were
like that's why Spec Ledger exists. The
idea is to bring you know it's like
Confluence. That's always the comment I
got. I always said like you know
Markdown is a source of truth. Git is a
source of truth. Your docs and your code
needs to be updated together. And I
always got the the feedback, Confluence
is where your business users live.
They're not in Git. So you need that
document in Confluence. And I was like,
okay, fine. Let's build Spec ledger
then, you know.
>> Yeah. But I still I get what you're
saying, but but there's more to it.
There's a lot more to it. You're talking
about service desk and you It depends on
if it's internal, is it is it, you know,
B2B, is it uh B2C and and where are you
using
>> for it to become enterprise grade? Yeah,
it needs it needs almost to have a
service desk. It needs to have
>> Yeah, it needs a it needs a quite a bit
of stuff.
Anyway, let's let's end let's end the
podcast here. Uh
>> there was one more from CircleCI about
that because they said we we moved from
so this I think it's a CEO from
CircleCI. he did a blog post about AI
and how it impacts their like building
uh and their business also because
they're in the business of CI/CD and
feedback loops and validation right so
first off they say obviously this is
great for circle CI because now
everything needs to be validated and
with AI you can build full automated
validations um and that works great for
you know CI/CD vendors but the second
thing he talks about like we moved from
confluence to linear to have more like
streamlined issue management
But then it takes the AI the same or
less time to complete the task than it
takes us to to groom the ticket to
basically define all of the work that
need to be done. So in the future, does
it still make sense to create like full
ticketing systems and track track all of
that?
>> I thought that was
>> you might as well just you might as well
just implement it right and see how what
how how it went.
>> Yeah, that's that's a trouble. I mean,
yeah, like with species, you're saying
you're recording the decisions, but at
the same time,
I guess I guess you can
it's going to get to the point where you
can where you can play out option A, B,
and C, right? And and it doesn't it
doesn't there's like no cost to it.
So, like, why even worry about the
decision itself? I guess you you want to
record you definitely want to record the
tree, don't you?
Yeah, you do. You do. You definitely
want to do that. Okay, man. I I I really
enjoyed talking with you. I hope
everyone else enjoyed the podcast. Two
weeks of We haven't even
two two weeks of AI shenanigans.
Please um yeah, please like the the the
podcast. Please rate it. It's on
Spotify. It's on Apple podcast. It's on
everything.
and comment on on stuff.
>> Did Cloud post it on SP on on Apple for
you?
>> Uh and and Claude made RSS feed does it
which is cool.
>> Oh my god. Like
>> I had a I had somebody there was an
error like he says, "Oh, I don't have
access. I just took the Slack message. I
post it into cloud. I said this is the
repo. Go figure." And it's like, "Oh, he
did the you know the query of where does
this user belong to which team? what
permissions does the team have? And then
he says, "Okay, we should add the
permissions to this team and you're
done." And okay, great. Then I I you
know did the GitHub CLI call and I went
back to him. I say it's done. It even
prompted like do you now want to inform
him and I was like yeah but I don't have
Slack connected for you cloud so I need
to do that otherwise it would
>> there's so much you can do.
>> There's so much you can do.
>> Um oh actually one thought as I'm
looking at the CircleCI blog. Well,
Peter Steinberger says that he doesn't
like to use uh CI because he finds it
too slow. Like he doesn't use GitHub
actions because he finds it too slow. He
he always goes for local testing. I
wondered if you had the same approach or
do you rely on the GitHub actions
pipeline to take one minute to boot up
and do its thing?
>> No, my pipeline is for other people that
um it's it's a double validation, right?
It's
always
>> yeah it's it's it's like it works on my
machine if it works on the pipeline it
means that I've made sure that anyone
can reproduce my machine and and it's
also what if the other people haven't
put in the proper quality gates and they
introduce a regression and it's for
myself like yes it may guard against
regression locally u but also sometimes
the CI/CD is for a longer running test
that I may not want my laptop to be
doing you know
>> but GitHub really needs to be
>> needs to be disrupted but Not just
because of the hosting thing and all the
other things. I find it really painful
to interact with uh GitHub workflows
getting out the logs with AI. It's it's
always just
>> the GitHub CLI. Yeah, it's wonderful.
You can
>> every time you extract a GitHub workflow
um a G GitHub actions workflow log, the
log is like 2,000 lines when it's when
it's doing nothing. You can do lock
failed and
>> even then even then it's
like hold on I got to show you this
before we're about to [ __ ] end this
like
just so you know it's it's real. What
the hell?
So I have this I have three skills here.
Debugger, optimize, updata.
the data helps you update uh it even has
a a Python script to update you know the
the actions you know so like if an
outdated action at you know check out at
v4
it it has a script that figures out the
latest one and updates it to v6 but this
debugger one which is a simple skill I
hope you'll agree it says when when
actions fails diagnose issues using the
gh cli expect a URL like this
and And it's weird because it does it
for me without having the skill. It does
exactly this without the skill. And it
also taught me a trick because like for
example, if I run a task that's very
verbose on the logs and I want to find
failures. Um I have some ETE that can
run like a few minutes and it spits out
a lot of logs. So many logs you wouldn't
believe. Everyone says it. [laughter]
Anyway,
the the thing is that cloud what it does
is it um it will use t to write to disk
and then do the set. So then if it can't
find because the set failed, then it
actually goes to the disk and starts
setting on that. So you don't need to
rerun the whole thing three minutes. You
can just like directly go to whatever
was on disk. And I always tell cloud,
hey, use that trick. That's a really
cool trick that you did. And
>> actually I I actually need to update
that then. So like tier it to a file and
then
>> yeah so if your set didn't catch it then
you can still cloud may still say wait
maybe maybe something wrong with my set.
Oh yeah I found out and then it doesn't
need to rerun the whole thing just goes
to the locks. Well, this is frustrating
to me because the co-pilot one it
sometimes it seems to defer to the MCP,
you know,
allow GitHub MCP and I'm like, no, no,
use the GitHub CLI. And this is why
>> you still have the MCP. Oh my god.
Uninstall.
>> Like the trouble is it comes back. It it
it's built into bloody VS Code.
>> Really interesting. I I I don't have
that problem. Uh, I I kill all CLI. I
see. H, sorry, all MCPS I see. [snorts]
>> Yeah. Anyway, okay. Well, I'll maybe
improve the skill or just stop using it
because it's not needed. That would that
would be ideal. Less code. No code is is
the best code.
[snorts]
>> Okay. Great talking with you, Vincent.
Yeah. Everyone like this like the stuff
and uh tell us how we can improve.
>> Yeah. Bye
>> bye. They're not listening to people.
All right. Okay. Bye.