Episode 20: cfgmgmtcamp 2026 debrief

Hello. Where are you?
>> Can you recognize?
>> That looks like a ugly city.
>> Sure is.
>> Yeah. I I miss Kent.
>> Yeah, it's actually it's pretty good. I
just I just been enjoying walking around
actually. It's been fun. And the
architecture is pretty amazing. Like I'm
confused cuz I I always thought to
myself that like
I don't know that that um
that Europe is
dying.
>> Well, yeah. Exactly. But like the
buildings there are quite wellmaintained
and rich. You know what I mean? It's
like confused.
>> If you believe certain politicians, it's
all you know run over by crime. And
>> maybe maybe that's still the case.
>> Yeah. And there's a lot of like um you
know young beautiful people walking
around and uh
Yeah. It's quite it's quite different to
London. I'd say
>> cuz I'm old now. See why hair cannot be
there anymore.
>> I'm I'm freaking Gandalf right now, man.
>> Yeah. So yeah, Belgian's been great and
yeah, the conflict management camp was
good. I mean, I think it's actually the
best infrastructure conference on the
planet, I dare say.
>> Is it now?
>> And you missed out on
>> Yeah, I did uh fortunately um after some
information around holiday period, I had
to
>> lock in some uh contracts and deals and
it's been going great. So,
unfortunately, I couldn't make it to
config management camp, but I was able
to do different things.
>> So, this is all CDK ter was it CD? what
do you call your your
>> DDK terrain it's called now hey actually
your your your proposal of the Golang um
setup of a project and hosting on AWS
where on one side you evaluate cloud
form versus the AWS CLI and AWS CDK uh
but you're not evaluating Terraform in
there um you're comparing it against
cloud for
>> well that that plan that I sent you like
I
>> I I thought I wouldn't influence it that
much but I asked Claude
>> to give me what it wants to do.
>> And funny enough, it didn't mention
Terraform. Like I was I started the
prompt saying like
>> that's interesting.
>> Like what how would you
um and I I think I I should have shared
the prompt, shouldn't I? I think I was
opt I was optimizing for speed and
operational simplicity
and that's what and and CDK came up top.
So yeah.
>> Yeah. But like
>> even if you look at the data of that
like um report that was shared by Matt
Gawi um but it's it's AWS usage data
report about how people deploy on top of
AWS. Cloud formation was by far the the
highest used and then there was
Terraform right after and then there was
AWS CDK. You know, I I I find it funny
that you don't combine AWS CDK and cloud
for because it's the same, right? It's
cloud form. Uh except you have a better
interface when you use AWS CDK.
>> Well, I mean, not you like the the
people that that support.
>> Yeah. The I guess the winner for CDK is
the is the you know the shift left early
feedback because at at compile time
you should kind of know if if your code
is working, right?
>> Yeah.
>> Yeah. I mean and I think that
>> not 100% but like good enough.
>> You you I you you must have influenced
cloud to consider this as one of the
criteria fact of the compile time
checking or it's influenced by your
previous discussions.
>> Maybe maybe maybe
>> so I had a chat with the Palumi guys
at the desk at config management camp
yesterday. I was pretty impressed with
what they
>> they're doing really cool stuff. Yeah.
No, no. I have to say like if you
consider CDKTF and Pulumi as as like
competitors, I don't think that's true
because I I mean I was naive. I I did
originally consider them. Uh but then
somebody showed you know wisen me up and
and showed me some of the really cool
stuff because Pulumi is a host engine
that allows your um your infrastructure
code to hook into some really advanced
features that you can't do with
Terraform declarative configuration or
you can't do even with CDKTF unless you
have some type of like the the runtime
the host environment in which your your
TypeScript or Python or Golang you know
the host is is giving the
infrastructures code a lot of like
dynamic capabilities of lookup and AWS
CDK does some of that because you you do
have the CLI able to host
>> what do you mean the host like the their
SAS application what do you mean by host
>> no no no like the puli host so when you
run the when you run your infrastructure
configuration application it's an
application right when you run it in
puli it runs inside a host that runs
your
>> like the kubernetes model like desired
state is it why is it like that?
>> Uh no it's like you know because
ultimately puli uses the terapform
providers so under the hood at least
originally. So, so what what it does
give the program, you know, because your
Terraform, sorry, your TypeScript is a
program that executes, right? And it
exposes some functionality to that
program that you write that can do
things like
custom or tailor make tailor my
Terraform or my my provider resource
configuration to the target. So it does
like it it's called environment agnostic
or um environment aware setups. If you
give it information about the target
environment, it can actually uh provide
um details. Oh, sorry. Uh I just have to
reply to someone.
>> So so that that it does it does that's
the power of puli to me. um that you're
>> the the question in my mind when I
approached the Palumi desk was like I
guess I was coming from an AI
uh perspective in the sense that like
since Palumi offers different language
bindings and of course it does straight
up YAML what would be the best language
for
uh AI to use given all those choices and
the guy interestingly said like it
doesn't matter it doesn't matter it
doesn't matter
which I thought was quite curious
>> he said it doesn't matter like what is
the best language to like the front end
the the interface that you use to define
your infrastructure
>> yeah like he said he was basically
implying that it wasn't it didn't matter
if it was go yaml
>> okay he was a saleserson
>> well I think he kind of he kind of made
some arguments to me like I think that
essentially when it all goes through
paleumi It gets
>> no because they even have like a full
full declarative interface, right? Which
is um
>> yeah, you can even define but to me
that's weird because YAML is always
going to be like this domain specific
language type of approach.
>> I guess he was implying that once you've
>> thrown it into the Palumi whatever it
gets it gets found out. I don't really
know but yeah I thought it was
interesting.
>> I I also don't really know the details
there because one of the major drawbacks
about AWS CDK and the gsi ecosystem is
that you get a lot more capabilities
when you write in Typescript because you
can make libraries or reusable
components in Typescript and then you
can consume them in Python or Node. Oh
sorry in Python or Java or Golang or Net
so you can generate
>> but you can't make you can't go the
other way.
Yeah, because all that gsi does is it
just creates a binding uh from the other
languages that under the hood calls
JavaScript and you can't go the other
way. You can't create a binding from the
other languages back into TypeScript. So
like when I was evaluating it for like
teams to work with, one of the things
that we do with CDK AWS uh Terraform
CDK, what we do is we we're building
reusable components like L2s. like I I I
have a common L2 library, but then the
teams they're actually building type of
very customized L3s which are like very
focused on what they're trying to do and
they can do that in Typescript and then
we are publishing and consuming it and
transpiling it into uh Terapform
configuration so it lands straight into
our existing Terapform pipelines. So we
have all of our existing Atlantis, all
of our existing Terraform modules and
they just neatly land side by side with
no migration required. So if I
understand you correctly, Vincent, you
need to write your your shared
uh level two or three construct in in
Typescript for it to be shared.
>> Yes. Because like my whole my whole
pipeline is like you publish um an MPM
like code artifact private registry
package of your like I'm a product team,
right? I have my um service definition
which is an XJS or whatever uh or maybe
it's just an expressjs or yes we are
typescript shop sorry um and and then or
a lambdas right because actually it
applies much more to lambdas like let's
say that I am a product team and I have
a very serverless um solution that uses
SQS SNS um and all this like bucket
notifications and lambdas and all that
and and so I need to produce re reusable
artifact that I can then deploy in my
dev environ environment uh first you
know my sandbox evaluation environment
and then my staging environment with
more integration points with other um
you know existing uh staging resources
so I am producing a bundle uh an
artifact that is actually a library it's
an it's an IC little library and and the
integration
>> IC IC library what what's that mean
>> infrastructure as code library
>> oh okay sir
>> like a TF module is is kind of like a
bundle or library right except you um
you don't really have the capabilities
of defining strong interfaces and also
the library can't the module can't
really u modify itself depending on the
target environment right whereas if I am
um in in Typescript
>> the interfaces that I get are like
strongly typed and I have like I I can
even create imperative type of
reconfiguration like that the whole
thing can can adjust itself towards its
target and there's pros and cons right
because the more flexible the more you
get like um you know validation like
mechanisms now I need to I have like a
bigger interface right because I have
more ways I can be configured then more
things I need to test for the the the
different configuration
>> exposure
>> the
>> the other thing that was demoed to me
with the Palumi just getting back to the
Palumi thing was that
>> um yeah their SAS offering seemed pretty
robust you know they had secrets
management they had stack management
>> I mean I I I don't know if Terraform I
mean the last time I used Terraform
cloud or Terraform I don't know
workspaces or whatever it's called
nowadays
they feel I mean is IBM even supporting
it anymore because it doesn't seem to
change
>> this was always the biggest problem of
>> I don't get it like why why buy
Terraform if you're not going to
>> yeah I mean even the ter the cash cow
you know what I mean like
>> I don't get it
>> well IBM just been iterating on it
recently and Maybe part of deprecating
CDKTF is to ensure that they can focus
on other areas. Although that there was
a skeleton crew on CDKTF. I don't think
it actually took any of their headcount.
But it's a good thing they got rid of of
that that team. Not they got rid of the
team. They got rid of the product. Uh
the team is still there. The team moved
on. They work on more of the like core
features.
>> Yeah.
>> Yeah. But but go back to the Terraform
cloud uh and the Terraform enterprise.
One of the biggest issues with Hashior
Corp always and even Grab in Singapore
was leveraging Atlantis an open source
solution against Hashi Cororb saying
like look you're asking for this much
money and I can get like most of it out
of Atlantis in a more customizable way.
>> Yeah. So, so they did hire the guy um
you know Luke Kaslau and um the other
guy I forgot. Uh he's he's he's actually
head of product of of one component. um
Mikail he did I saw his presentation on
on Hashi Kate but but what I'm trying to
say is that they always lacked in in the
execution level of of the cloud solution
and Pulumi is very strong there you know
our by
>> yeah it's really good the I did ask them
that like so you have multiple stacks do
you have like a dependency tree and the
guy conceded that they didn't to my
surprise the outputs are strings
>> like what what what grid does you mean
something or or or what's it scon or
whatever you call it.
>> No, what I do with grit which is like
>> Yeah. Yeah. Well, you do grit but you
you were inspired by tacons or something
weren't you or something. I mean it's
terri ticons is my project. I was doing
grit. Tons I was inspired by open taco.
>> Open taco. That's that's the one. Yeah.
But like but yeah I I mean we talked
about it last podcast and I like
strongly typed shift left like we have
that now. What's the next thing? Better
dependency management with
infrastructure, right? I mean,
>> yes,
>> maybe.
>> But it's very hard because like
>> the argument that I hear the most from
Terra team and state graph is that this
is very hard to achieve with
infrastructure because you can't
validate that your network rules work
correct unless you actually deploy it
into that particular environment. And
there's always a difference because of
the side range differences and all. And
so that's kind of like where they they
say you can't fast your infrastructure.
It's not the same as software delivery.
And I can see that in a way. But I think
if you have a strong language and you
have like strong interfaces, you can do
a lot more you can get a lot more
confidence in in the fact that you know
it's going to work the way you expected.
>> Well, I think it's yeah problem yet to
be solved. At least I haven't seen it
solved. I mean I've
>> if you believe I should try your grid uh
your vibe code.
>> Yeah.
the grid is great. Um,
>> are you using it in production?
>> I guess your your your attention is has
been starved. You know, you're focusing
>> so I have not I mean I I I have it
deployed in a PC repo with the intention
of other people to validate it um in my
organization but we haven't been able to
promote it again because of like you
know
>> cost cutting exercises across the board
in every company that I talked to. I did
I didn't see the the flow in Palumi
actually. I was cuz it was like you know
didn't you didn't have a computer. We're
just talking face to face. The other but
the the demo that kind of blew me away
was not a plume
>> system initiative.
>> No no not the system initiative.
>> I can talk we can talk about that a
little bit later.
>> The demo that blew me away was the um so
so qlang is like pickle. you define your
config in a fancy fancy language.
>> You got to choose which one. You don't
know. You can't you can't
>> Well, I'm to be honest, I'm I'm leaning
towards Qang just because of some of
quality of life aspect. So, the cool
thing about Qang is like I don't know
that it changed the config, you know,
like scaled the memory, had more CPUs,
and then it and then of course it can
hook into pretty much anything. But they
showed a demo when it it did a terraform
plan. And the cool thing about that was
that when the Terraform plan came back,
it fed back into the Qang. So you could
so you could actually see the change
without like look, you know, like when
you do a plan, it's just like bloody
insane. And the cool thing about it, I
hope I'm trying to explain it correctly,
was that is that it summarized the plan
in a way that was like operationally
sane like yeah, you've changed this and
that's and and that's because of this
config change you and this plan cor like
works out how it is and then I I think
they've I'm not too sure if they
implemented it uh quite yet, but then
there's things in the works like the
approval gates and things like I mean
this This is something that we struggle
with at my workplace is like we're
changing Terraform like we have to
control how it's done the audited
whether it's deleted whether it's
updated replaced we have to we have to
understand what's going on because
reading plans is just or there is this
TF summarizer that we use but it's still
a nightmare. So the cool thing about
this was that it in the PR it like spun
up a Q lang thing. I didn't even know
you could do this and it was just really
fast and and nice. It was like
>> but this is a framework on Qang. This is
some like this is a bunch of tools that
they built on top of Qang.
>> Yeah, something like that. It's like a Q
lang. I think they call it Q hub.
>> Okay. I haven't heard about this.
>> That that demo was impressive. But then
and then some some other guy from
Ubisoft was giving a demo where I've
never quite seen this concept before but
he gave like I mean it was a bit
contrived but he gave the scenario where
he needs he's imagining like a thousand
users and then given a thousand users
then you can can say that given what the
application does you need you know a
thousand megabytes of RAM and then and
then um and then basically
this model of how your application works
becomes config and then also that model
gets confirmed in the observability. So
you have this sort of feedback loop
between what's deployed, what's being
observed, what's your model, what's
being deployed and you you go you go
from there. And then of course he uses Q
lang as a way of marking that all up so
that everything comes together and and
and further validates the model. I
thought that was I thought that was
really interesting. I've never I've
never seen it done that way before. But
like
>> that's interesting. But like my problem
with all of these custom like I
understand that a lot of organizations
like Google they even invented Qang
specifically to manage the problem of
configuration. um meaning that you don't
have a full-fledged programming language
for configuration management. That's why
they built Quolang. That's why Apple has
Pickle. Um
>> but I still I mean I think that's not
not a bad idea, right? Um the fact that
then you can use QAN to manage
configuration like the same reason why
JSONet exists, right? JSONet is a
language to basically, you know,
imperatively kind of control the JSON
configurations that you're trying to
generate.
>> Um but
My issue and I think I tried to tell
this to you as well when you send me
some of this this stuff is like great
you know that's yet another way of doing
the same thing on top of ter
resources. Yeah, you're inventing a new
thing. But the same way like with the
system initiative, you're just creating
another way of, you know, doing CRUD on
resources and that's not the problem,
right? Doing CRUD on resources.
>> Well, I wouldn't I mean that's a bit
unfair because I I mean I think Q is
trying to get that one pager that over
what that one pager of your of your
application system thing, right? Sorry,
I can't I'm not too sure. Sorry, I can
do I can do one page.
that with my hands
it's not trying to uh do CRUD operations
of course like the interesting thing I
was talking to some of the QAN guys like
you know
>> but you you delegate the problem you
delegate the CRUD to the to the to the
terraform provider right as long as you
don't as you leverageing providers you
get all of the
>> CR but in an ideal world things would
would great I mean this is the way I
understood what he was saying in an
ideal world things would vastly improve
if you could actually give that whole
config to the the cloud service provider
instead of doing all these CRUD
operations because there's so much
complexity in that Terraform provider.
There's so much stuff going on with all
those CRUD operations. In an ideal
world, you could just give the
declarative config and that's that I I
mean that's what we want to move
>> which is what kind of like Kubernetes
promises, right? You give it config
>> and that's what crossplane also
promises. You have to have a whole cloud
control layer like within Kubernetes. I
was chatting with a Kubernetes guy and I
mean we all know this stuff but like but
I just think it was I don't know maybe
it's just a gig that I'm on like he was
saying like yeah when you when you put
your CPUs and your memory definitions
and your deployment uh YAML the idea is
that you deploy it into your environment
and then you constantly refine it
knowing what the application
>> okay was that was that was that um is
that the guy okay I can guess who Daniel
was that Danielle
>> no it wasn't Daniel it was Okay,
>> it was another guy. But but I thought to
myself, it's like this is so weird
because or foreign to me because we have
so like in my workplace we have so many
environments
and there's a pro promotion environment
and like it's very very hard to nail
down the the resource limits. Anyway,
>> yeah. So I just thought there's this
cube learn uh from Daniel. They do some
really interesting like visualizations
of how Kubernetes worked. They even got
some gigs where they they trained Apple
uh engineers to on on communities. And
one of the recent things that he was
really into was to like how do you
evaluate the actual you know resource
requirements of your deployments and he
asked a lot of questions and spend a lot
of time trying to see because they added
now they added virtual pod autoscaling
as well. So before you could only like
add more pods, but now you can actually
with the control group V2, you can
actually resize the number of CPU and
memory you give a pot without recreating
the pot, which is insane, right? And and
and if you have those capabilities, then
you can like dynamically resize your CPU
and memory.
>> I still I still think it's pretty
insane in a way. I mean like
>> Yeah, but um I mean that's just one
application level. Hold on. Wait up.
Wait up. Wait up. Wait up. I just need
some water.
But these are all the dials and knobs
that you get with Kubernetes, right? And
the problem for me is YAML the way to
declaratively define it's insane. Then
you build languages on top like Q that
basically I I don't know do they use
because cuz doing a demo from what you
just said, right? He said Q is great if
only we could give it directly to the
provider uh directly to the cloud
service right and without having to go
through these providers with all their
legacy assumptions. And then we
discussed Kubernetes is a cloud
controller interface that gives you the
ability to declaratively define a very
you know
>> yeah I mean the guy behind Q is like the
the same guy behind uh Kubernetes
actually. So I guess that's why he says
stuff like that.
>> Yeah. Yeah. Well, if they built a
configuration front end that uses Qang
and then is able to manage the YAML um I
guess great but um that's still
I mean you still need a cluster to
deploy it and then yes maybe you can be
uh like like up upbound which is the
crossplane SAS um basically we provide
you the cluster so you don't have to run
your own clusters and then you just work
purely with crossplane so you just um
define all
>> no not crossplane Oh, getting
sidetracked here. Okay, let's
>> No, because you were talking about um
you know basically a control loop on top
of the cloud service and an uniform
control loop and that's what crossplane
is, right? They're a unifor uniform
control loop on top of the um
>> Yeah, but Kubernetes or something. I
mean, if you want to go
>> what are you talking about when you like
if you need to provision SQSQs and
topics and buckets?
>> Oh, okay. You want to still do AWS?
Okay. Yeah. Okay. I think I got you. I
got you. Yeah.
>> Not necessarily AWS. I mean, any cloud
that you're using, you're still using
managed services, right?
>> Uh oh, God. Sorry. Well, let let's let's
go back to AI. Did Did you Did you Did
you see
>> because that was part of your prompt
like
>> did you see Adam did you see Adam Jacobs
talk? Did I did I share it with you?
>> Uh is are they online? I don't know why
I didn't.
>> Yeah, it's online. Yeah, it was it was
good.
>> Okay, that's good. Yeah. No, but like
you give me some notes where you you
said like um it's going to be so much
code that we cannot control it and our
role is more towards validation. I mean
I've hear the same thing for
>> spot checking
>> everyone I talk to.
>> Yeah. Yeah.
>> I think those are like Yeah. If you use
the AI you kind of like you got to be
blind if you don't see that.
>> Yeah.
>> I think everyone's coming to this
conclusion.
>> Yeah. the I you know what one thing I've
always found very polarizing in my
career is sllock like source lines of
code and he basically just came up guns
firing saying that like that with AI you
can output 50k lines of code and he I
the point he was trying to make is that
like the force multiplier the utility
the uh the what AI can do is absolutely
insane
and um I I mean 50k I mean what what is
50k? It means nothing. It just but it
means it's just it's just a number.
>> I mean people just refer to that like
>> huh
>> Boris when he talks about all of the
code to plot code he he just says like
500 PRs a week or something. It's all
managed different.
>> So yeah, he was he was talking about the
velocity. The velocity Yeah. My take
away from his talk was that the velocity
is now absolutely bananas
and now the bottleneck is
infrastructures code because because
because all that velocity needs to run
on some service. So we need to match
that velocity with the way that we
provide infrastructure as code. So yeah,
it had me thinking like what is slow in
my organization like oh my god you know
PRs are slow um our CDK stack
deployments and everything like that the
coordination and the releases is slow
the promotion is slow
>> there's so much slow slowness in at
least
in my in my workplace I find um for
different re not not just technological
reasons I would say mostly
organizational and approval
reasons everything is so slow in my in
my client.
>> Mhm. So yeah, it got me thinking and
this is why I was asking questions like
like
now I'm now I'm looking at my
infrastructures code pipeline and just
thinking like like how can I make this
fast
so that um so I mean like I don't when
you develop with AI are you doing a
Terraform plan and apply or or CDK
deploy and apply and and like don't you
feel that that's a bit slow or maybe
we're we're looking at the Oh
>> no, because one of the things that you
highlighted in that Golang architecture
for AWS or that that cloud gave you was
like deployment time with AWS CDK is
fast just CDK deploy. I think that's one
of the main criticisms I hear about CDK
is that you know cloud formation is so
extremely slow. Um you know because it
goes and and it gives you this
transactional uh guarantees like it's
all or nothing. So it starts to deploy,
it starts to create backups, it try
starts to do all of this additional
stuff and then if anything goes wrong,
it does guarantee the full roll back
which is super frustrating
>> which never seems
>> super frustrating work because then you
have something left behind then
something you know prevents it from
being destroyed and and it doesn't work
right. Um
>> but
>> but that's the thing that you don't get
with terraform right when you run a fly
um if it doesn't go through totally it
just left like that right and then
you're left to to figure it out it goes
very fast right it going to create
destroy everything and then it go like
whoop uh actually we're stuck good luck
and there's no guarantee on on the roll
back right which is what you don't get
uh with terafform it does make
everything feel much faster because you
have these individual resources all
getting created super fast but you're
then you don't have the transactional
guant guarantee right so that was one of
the things that I was trying to solve
with my friend when we were uh because
he heavily relied on this cloud
formation guarantee uh doing cross
regional deploys because that's another
great thing about cloud formation it's
basically available in every region so
you want to do a cross regional deploy
just kick it off and it goes um but then
when we looking at like moving over to
Terraform we we are like oh [ __ ] now we
are responsible because we're like like
the fly.io IO right now we are
responsible for like what if the
deployment failed how do we roll back
and I'm like yeah I can't tell you you
know could be so many reasons why ter I
just had the thought that um I made a
video about this a month ago or two I
just had the thought that like
data aside what probably works best for
AI is just having its own AWS account
which it just bootstraps from from
scratch and when that AWS account
becomes comes crofty. It just basically
deletes it AWS account.
>> Nukes it.
>> Nukes it and starts a new one. What is
that? There's probably a name for it.
It's uh I don't know like
>> ephemeral.
>> Pardon?
>> Sounds ephemeral.
>> Yeah. Yeah. Ephemeral. Yeah. I think AI
needs more ephemeral environments. Oh,
but in my current workplace, just
getting an account vended is really
really really really hard. No. And I I
don't know how we can even close it. But
yeah, a firmware environment surely
seems like an an AI friendly pattern.
>> Well, everything has data. You know, the
biggest problem with environments is
data. You know, how do you like even
replicate or even when I want to do a
full integration test with with full QA
automation end to end? How do I get
realistic data? Do I have seats for
everything? Do I replicate production
data? Do I, you know, mask out all of
the information like some
question. Yeah, I don't think anyone
really talked about data and AI, but
that was probably the missing piece of
the puzzle for me.
>> Another we have a job. If this was all
solved, we wouldn't have a job,
>> I suppose. God, you sound really
Belgian. Um, maybe it's all the Flemish
I'm hearing. I one other demo before
before I need to probably run out
screaming is um which I found really
interesting and that really proves the
cat is out the bag was this guy was
showing that his new framework laptop
with the AI Max chipset and
interestingly he had like he had like I
don't know
>> open clocks
>> he had 180i he had 128 gigs of RAM or
something and the cool thing he said is
that this new AMD Mac AI max one the
USPS. One of the key features is that
you can dedicate your RAM to the GPU and
make your GPU even more powerful.
>> So he basically KVMED into his machine
which didn't have any internet. And then
he he's also the author of era which is
like this uh anible playbook um sort of
UI where it like logs how long the AI
playbook took to run and what the inputs
were and and what the logs were. So
basically he managed to configure his
his framework desktop to diagnose
failing playbooks
and he was just using you know Alama and
uh I don't know the GP GPT
whatever or the Z or what I don't
whatever local models
>> GM
>> and it was just ch it was just steaming
through all these playbook errors and
fixing things you know missing updating
>> local models nice
>> yeah updating Postgress libraries,
fixing errors.
>> It wasn't a prepared demo at all.
>> Huh?
>> It was probably Postgress from 5 years
ago. Well, within the training database
of the of the local maybe, maybe.
>> Just kidding. I'm just kidding.
>> But the the demo was compelling because
u you know how it is. It's like or you
don't want to use AI because it's you
know uh it's only used to be become
dependent on anthropic and open AI. But
no, hell no. This shows that we have the
ability to run it locally. Of course,
the model the model is not as capable as
claude, but it's damn well the cat is
>> but think about it.
>> Those local models now are like 04
capabilities and 04 like chain of think
of of of thinking.
>> Why you say 04? I mean, okay, whatever.
because they they they do compare like
the the uh GPT OSS are equal to 04 is
GPT04 like the the first
>> reason I think 34 I mean the whole
bloody versioning thing doesn't really
make sense but like
>> yeah it was 01 04 because they went GPT3
GPT4 then 01 then 04
>> even even six month or one year old
models at this point are actually pretty
capable for some for some tasks
>> so yeah like so like Of course, anyone
that's listening to us right now, I
would say that like the best model at
the moment is the Claude Anthropic 4.5
Opus or or SA. It's
>> So, you just got a sponsorship, is it?
>> No, this is # notsponsored. It's It's
like It's $20 a month to get on pro. Of
course, you get limited quite heavily if
you do anything hardcore, but like like
guys, the sun is shining. The sun is
shining. Make hay.
>> Wait, you're in uh in Belgium and the
sun is shining. I can imagine you're
happy.
>> I think there's something something
going weird. Is Is the AI Did the AI do
this? Is the AI listening?
>> So, there's this that happened. Um
Oh [ __ ] Is there sensitive information
here? Probably just
>> your your locations maybe. Hmc.
>> Um what is this? Oh,
you now you you're now a big boy wearing
big boy pants.
>> Just up just upgraded this morning. I
felt like I'm
>> Have you Have you already hit the limits
knowing you?
>> No, no, no, no, no, no. I was I was
running out of limits on pro every week
like within 3 days and I was like that's
it. I need I need for the next few weeks
I have quite a lot of things to I have
to do and I need
>> I'd be curious if you hit the limits cuz
it's like it's that classic isn't it?
It's like you get the back now. Now
you're over overprovisioned and now
you're like thinking, "Oh, I'm not using
it enough."
>> Yeah, I know. This is definitely going
to lead to some burnout.
So, yeah, I'm um I'm probably going to
you're probably going to have to cut
some vision like put some black boxes
some places because I just showed part
of my credit card. I don't know if that
was going to help you with the last four
digits of your credit card.
>> Yeah, it's true. And it's also there's
no money on it.
But that's not how credit cards work.
>> Max plan. I can just imagine like people
panhandling on the streets.
>> I need money for my max pan.
>> Hey, you're kidding. But I do think that
that's what's going to happen next.
>> No, no, no. I mean, I was just going to
become elite.
>> What?
>> It's going to become only accessible to
some people.
Well, when you talk about panhandling, I
I I really think some people are not
going to get access anymore.
>> Come on. My earlier story was supposed
to point out that for for four grand,
you can basically get a less capable
model, but run it locally.
>> For four grand of hardware,
>> that's only four grand he spent.
>> Yeah, the framework laptop is 4,000
Canadian or something.
>> Okay. Okay, that's not too bad. So, you
could argue that, you know, with the
money I spend every month, I could have
just bought and depreciated it. I still
don't think it's the same, though. Um,
>> yeah, it's difficult like, yeah,
>> you know what? What convinced me to to
bump it up? Not just that I was hitting
the limits on Pro, but also that the
recent Claude Code announcements. So,
they added um like task dependencies.
So, you basically have beats built into
Claude Code, right? The ability
>> Oh, that's all that's all in Claude Code
right now. Yeah. Like since I've been
since I've been in a conference for a
week, I feel like I missed a month's
worth of announcements.
>> What did I miss?
>> So So I I should share with you the
video from the guy who used to do
speckit at Microsoft and he now he's now
anthropic, right? Dan I forgot his last
name, Dan de Demen something. Um and he
just broke down the future of cloud
code, right? He he's he's obviously very
experienced with like he highlights some
of the techniques that you use and they
are very very heavily used in in speckit
which is um the templating um so
basically you give the model a template
of what a plan looks like right and and
speckit does that when you run the
command it executes a script and that
script prepares it template in the right
place and then it tells the model read
it these are the things you need to do
and it follows a template right so you
get very consistent
>> why do you call it a template I mean
that could be a skill couldn't Yeah. I
mean, why do you
>> what's the difference between a template
and a skill?
>> It's a combination because like now
skill and commands, they're very
similar. You can like invoke skills the
same way that you can invoke commands.
So, what used to be a command has been
renamed to skills. So, skills and he
calls it the templates approach because
on one side you have a skill that gives
you instructions of the steps to follow,
but on the other side you have a
template of the actual artifact that you
want to produce. If you're talking
specifically about plans and and tasks
and execution of what your agents are
going to do, you need to have a certain
um you know level of information within
these specifications that the agents can
do it correctly without um confusion.
Right? So the the biggest issue is um if
you do VIPE coding, you are missing
intent and you're missing some of the
like you're you're letting the model
make assumptions and you're leaving some
of the edge cases out without your
decision and that leads to unexpected
results. So when you are following this
template model, you are capturing those
requirements and you are eliminating
those assumptions and those um blurry
lines. So so that's part of the spec
kit. Uh and
>> I guess the template implies that things
need to be captured perhaps.
>> Yes, the template has your minimum
requirements and he doesn't do only the
template because he also adds on top
that you have the hook validation hook.
So when the agent walked worked through
the template and captured all of the
information according to the template
and the task and the user specification
then there's a validation hook like he
uses actually um the one that you
mentioned like the the Astra TY the type
thing. He actually validates that the
the template has the attributes um
deterministically like this particular
specification documents or requirements
documents has this section this section
this section are all valid it doesn't
you know validate the contents but it
validates the section exists uh and
that's also what ko does right when you
are you know building your your your
plan to to to feed it to agents to
execute because ko is all about the
upfront planning everyone that uses AI
lately they all say you spend 80% time
up front planning 20% of the time is is
the agent executing right that that
execution load like what what Adam just
said as well it's just generated just
goes on right and your job as a human
with your limited amount of effective
hours and attention span you are um in
in in charge of like making the critical
decisions
>> Adam Adam was saying that we all need to
read the u the domain driven design book
which I think is written by like um a
thought worker um 100% or ex thought
worker. I mean have you have you ever
actually dived into domain uh driven
design yourself?
>> Yeah there so many times I've wanted to
and I've never really spent a lot of
time on it. So like anyway with with all
the specit stuff we I mean you missed
the specit to me did you find the stuff
any improving I guess it's just more and
more integrated with clawed code right
it's nothing
>> yeah so the guy who built who who wrote
large parts of specket and you know
iterated through a lot of the prompts
and set this all up aside from the idea
he worked with someone at Microsoft that
came up like that that that lead led
with the idea he's the guy that actually
built the specket now at Mantropic he's
introducing a lot of these
features on top of the Claude Code suite.
So if you watch his video, it's not on
his um official channel like he has a
separate channel which is like um how
you call like indie dan. Yeah, that's
his because like independent it's it's
like
>> oh
>> and yeah and there he actually goes down
he talks about the validation hooks he
talks about the template uh approach and
then he talks about my god
>> yes and then he talks about the agent
system because you were able to do
agents but the the co coordination
system was bad right I told you I tried
I I had all of my plans worked out I had
identified parallel tasks perfect for
parallel agents right because there's no
way that these agents will step on each
other's toes if they're completely
working on separate parallel tasks.
Wrong. Cuz there was assumptions in
there in in the in the tasks in the
context that they got which was like
there's this shared library and each one
of them implemented the shared library
and things like that. Um so that
requires coordination between agents uh
and and like Gasttown and others they
use agent mail and more traditional
systems of like agents messaging each
other and I think they're building
Gasttown now directly into Claude Code.
So that's why I upgraded because I felt
like Opus is not going to like my my my
basic access to Opus is not going to cut
it.
>> Okay. There's a there's a lot going on
here. So do you still use beads is what
is was at the top of my mind?
>> At the moment I am uh because what I'm
working with is something Claude Code
independent but I also feel that I
upgraded because I wanted to see what
this is going to be. you know this thing
that then uh showcased where he
highlighted some of the core concepts
that I see in in specket reapplied into
Claude Code and the concept of agent
swarm directly in Claude Code I felt that
um I don't know if I will be using beats
in that scenario right uh I talked to
other people they they use GitHub issues
and other stuff like that um
>> so does this t I mean have you played
with this crawl code or has it just been
announced is it live like does it save
the tasks to to some to some database or
to some
>> it has its task like it has it memory
system so it has the ability to save
task across sessions that's one of the
features of this new task system
>> because it is it able to do what beads
does like you know
>> it has dependency it has the ability to
define this task locks test task and
things like that
>> yeah it they build all of that into
Claude Code now so if you if you consider
all of that the reason to use beats
becomes a loss less if you are go full
in onto Claude Code So that's kind of
like where I'm I'm I'm feeling out.
That's why I want to explore
>> because I remember talking to you and
you were like getting uh Gemini to
review the output from claw code. Those
sort of workflows are going to I guess
become harder and harder to do if you go
all in, right?
>> One of the thing the points that Theo
made about the move from Entropic to not
allow people to use their subscription
outside of Claude Code and and
effectively shut down anyone that does
and ban their accounts. um is because
they want to make it very sticky, right?
They not the model alone is not the and
I think we know Opus by itself is just
on par with the toolkit that they want
you to stick to. So that's why they are
like building the toolkit.
>> Yeah, it seems sad though and
>> yeah, it would be very like I also I
always said like I keep my options open.
I don't go following full in on any of
them. But with the recent evolution and
the way that Antropic is leading and I
also think like I think KO is actually
because AWS has invested into Antropic I
don't know how Kir is affected by this
um you know this this decision of how
you get to use the model but KO uses
entropic models right um so I think Kiru
might have an edge there but I still
feel you know I still feel Claude Code is
like ahead even though that most people
hate the leadership of Antropic
with with all of his, you know, like
even even
>> Well, it's a bit anti-competitive, but I
guess they're trying to build they're
trying to build a moat. Trying to build
a moat.
>> Yeah. And they're like, you know, don't
sell GPU to China. That's an like and
all that. Like they're very vocal about
those things. And
>> what about other announcements? I mean,
I I mean, I can't help but think be a
>> Yeah, I'm being a shenan friend to open
AI. They don't seem to do anything that
seems to capture my attention lately.
>> Yeah. So, I shared with a friend that I
just went in on on Max and he's like,
"What are you doing?" Because OpenAI
just uh you know, double.
>> Yeah.
>> Yeah. They just released this new UI and
and and Tio did like a very nice uh
breakdown um comparing it. Uh like I I
used to complain a lot about Teio, but I
think I I
>> Did he did he do an honest thing and say
this is junk and open AI and Claude is
>> No. Yeah, he he had to he had to frame
his video very heavily with like I am
not being paid by OpenAI blah blah blah
blah blah but Codex is amazing.
>> Um so it was interesting to see he
basically said they took all of the good
stuff of anti-gravity which is the agent
management and they made it accessible
in codeex so that you can run multiple
agents across work trees. They made the
ability to use work trees more intuitive
so you can see all of the work trees
that exist and switch between them very
easily.
>> Have you tried Have you tried it
yourself?
>> No, I only watched
It looked like a very It looked like a
co-work ripoff on the face of it. I
mean, it looked
>> okay. I haven't tried co-work.
>> It just looked like another like fat
electron app. It looked very uninspiring
to me.
>> Okay. So, a lot of like So, somebody
told me like, "Hey, Open AI just doubled
the usage limits because I was like, but
when I I went from Pro, I went to max,
which is five times my usage limit, but
of course at um three times the cost. Um
but um Oh, no. Is it It is actually
five. Wait,
>> it said five I think it said 5x on your
earlier share.
>> Yeah, it says 5x usage, but the cost is
also not triple. I think it's also more
than anyway doesn't matter.
>> 5x decision
productivity.
>> It's like no buyers remorse here.
>> No buyers remorse here. No impulse
decision.
>> And
>> okay, cool. I have to go also.
>> Okay, you have to go. But like talking
speaking of buyers remorse, I I did find
it very interesting. A colleague said to
me that like Claude is basically making
you feel like you're at a gambling
machine. It make it's making it
addictive to do the thing. And I think
you might be addicted, Vincent. You
might be hooked by
>> is this an intervention? If the next
call I join and there's five people on
the call, I'm going and my wife
>> when says it was a great idea. Maybe it
wasn't Vincent. That's all I'm saying.
When you when you started saying using
cloud is like a gambling machine, my
initial reaction was this guy never
heard of specit. This guy doesn't know
how to how to you know get the quality
control if he cannot control.
>> Oh okay. No no I mean it is very
addictive this whole stuff. Okay. See
you. Thanks for the chat. I'll publish
this sometime.
>> Enjoy K. Enjoy the sunshine.
>> Thank you. Bye.