Episode 7: AI for Infrastructure Engineers

Oh man, it's it it's these weeks have
flown by and there's so many
developments. Uh I don't know where to
start even like like uh Okay. What do
you think of the Gemini 3 Pro thing?
Have you managed to get on it?
>> Um I'm using anti-gravity and I got some
Gemini 3 Pro for free, so obviously my
limits are pretty bad. Um, and it has
been I mean I started I'm like a day I'm
shifted in in like 12 hours compared to
when they announced it, right? But that
also means that I'm online when most
Americans that try it out are offline.
So I I have to fight with other
Southeast Asianbased people which not
everyone is following the latest Gemini
tree release. So the first day I had a
decent amount of work done by Gemini 3.
It was pretty good. The experience in
anti-gravity was also pretty nice. It
gives you a couple of new like things
reviewing files. I haven't used cursor
so I wasn't too familiar with like a
proper IDE integration. I've given up on
Visual Studio Code Copilot for a while
already. I'm purely in CLI land. So
using an IDE and you know it's it's
[snorts] been a long time before the LM
was like rewriting the whole file and
you had to watch it rewrite and you go
like no don't change that. Ah and now
it's it just instantly pops up with like
um this has changed this has changed you
can review this file. So it's pretty
pretty cool. Yeah
>> because um I don't know if you saw that
link I sent you yesterday to some AR AI
summit with Steve Jger and uh
>> it was it was live right? I I didn't
watch it live
>> because I like to fast forward and so
on. So,
>> well, I I just someone shared the link
and I just tapped on it and like Steve
Jger was giving his uh his his spiel
live and uh it was pretty good. It
>> It was a pretty good
>> I definitely want to see that now. You
told me that he's talking about it. I
see a lot of improvements on the
>> but like one of his like he says like,
"Oh, I needed to give a hot take and
here's my hot take. You're a bad
engineer if you use an IDE by the end of
the year. Thought
>> by the end of the year. Oh, that's a hot
take.
>> Interesting.
>> Like but like an IDE.
>> Is he sponsored by Warp now? [laughter]
>> Well, the thing is the AI summit was
like sponsored by I think Google and
Google was like anti-gravity this,
anti-gravity that. So I thought it was a
little bit rude of him to to basically
drop a bombshell. Don't use an IDE. But
like I think it perhaps what he's trying
to say is that you you don't need to
edit code. You don't need to edit code.
>> Yeah. Because you definitely want to see
those diffs and and and to be honest
even Claude Code is integrated within
Visual Studio Code or whatever IDE you
run it. Every other ID is a V VS code
fork anyway.
>> Um so so it it's really nicely
integrated like it opens up the de for
you. You can clearly see it. those
things still need a proper and and what
anti-gravity does really well is and I'm
still learning but it gives you like a
proposal and I was like great like I
don't see where I need to click but then
there was a small button at the top
which is like review and then you can
like highlight which they show I've seen
on the product demo and you can
highlight like here it made an
assumption I say well this is a wrong
assumption and then I feed back the it's
like a PR review you can just highlight
different sections and put a comment on
that
>> it's like a different take on that whole
plan mode include cuz I really like plan
mode in the sense that it's like
prompting you this this that like making
>> now it's new right now now when it's
planning it goes like I have a couple of
clarify questions and it it asks those
questions to you which is something that
speckit also did and now they kind of
build it into Claude Code like it goes it
goes off into a planning agent which is
running Sony 4.5 goes through your code
with the requirements and comes back
with like a plan and some clarifications
and it asks you like I have five
clarifying questions and then you answer
them and you submit it and then comes up
with a plan which is a very nice uh
experience.
>> Yeah. I mean I'm more CLI gang because I
feel like you know moving between the
keyboard moving like ABC like I can be a
little bit faster cuz the ide can be a
bit clumsy but it's it feels like we're
always going to go back we it feels like
we're going we're in this golden age
where it's like terminal claw code. it's
probably going to last a few more months
and then then they all go back to some
IDE uh you know noob land or something
like that
>> if they can address I guess the heart
like the problems that people have with
IDs I know a friend of mine really likes
cursor and he swears by cursor it's
mostly front end maybe that's why I
don't know u but like another thing that
I really like when you were saying about
curs um Claude Code um terminal based
view Gemini CLI did a big upgrade on
their terminal all um you can do Ctrl S
now to go into selection mode and select
things and copy paste them and the
scroll experience they say we fixed the
flickering like you know Claude Code
starts to disco um Gemini CLI has this
new way of of scrolling I found it slow
my machine is not slow but when I was
scrolling it was sometimes a bit laggy
but if it fixes some of the like
scrolling problems because if you're
reading an LLM and it puts out a message
and it scrolls all the way back down and
you have to go back to where you were to
like the thinking. It's so annoying.
>> These little things are like so
important, aren't they? Like that.
>> Well, Gemini is doing a good like I used
it two weeks ago. It was horrible. I
posted it. It was on LinkedIn and I was
like that was a horrible experience.
>> We were living in a golden age where
like this this competition between
Gemini and Claude. I don't know where I
I I did use codeex the other day, but
like I was I was flabbergasted because
codeex doesn't have like support for
project based configuration and things
like this.
>> Yeah. No, you Yeah, you want the prompt.
>> I mean, how easy is that a feature to
implement? Like
>> horrible.
>> And you don't have a shell also like you
can't exclamation mark. And I've tried
to search how do I run terminal shell
commands in inside codeex and it's like
you can't um which is
>> Codex is basically off the map at this
point. I mean what?
>> Yeah. Which is but then OpenAI says okay
we give Codex uh JGBD 5.1 CEX Pro to
everyone. Right after Gemini was
announced Gemini 3. They go like hey
everyone access to this now. [laughter]
>> But I mean I don't really see the value
in I mean I'm a clawed 4.5 gang still.
I don't I don't get anything out of uh
open AI because like that that's the
trouble like people I think a couple of
chats ago you were saying that you you
put on auto but I'm I'm now thinking
like no I need claw 4.5 or or this is
not going to work
>> auto
um like you you let it like auto approve
and
>> well auto choose the models like I'm
choose I want to choose the model
>> yeah I also change um for Gemini CLI
definitely I don't want it to go into
flash. So, I don't choose it to I don't
let it choose. I I think that's mainly
why I was having so many frustrations
with Gemini CI because they were going
into Flash.
>> So, I mean like would you like if you
were ranking the models, would you say
Claude is still top or would you say
Gemini is three 3 is now Eclipse?
>> It's very hard to to say because my
access to Gemini 3 has been quite
limited. Like I said, um I haven't tried
it in preview mode on on the paid
subscription of Gemini CLI. I'm using it
in anti-gravity free. So, um
>> I think I think you can assume you got
the you got the uh tasty experience
there.
>> Yeah, I did some really cool stuff. Um
but again, like I don't I'm not doing
benchmarks cuz I find this is the
hardest thing for me to make an opinion
about these things is I I Okay, so
basically what happened to me this week
was
>> benchmarks are always like weirdly weird
and manipulated. I think I think it it
is that to go on fields basically.
>> I don't think so because it's so hard to
to basically I saw some Reddit posts and
people were like I gave all three models
the same prompt. This is how how they
fared. I think that's a valid thing but
I don't have time for that because like
what happened to me was I I promised to
get something deployed um by this by
today like two hours ago. uh supposed to
finish it and I started on Monday and
then I realized I kind of underestimated
it as usual as all software engineers do
all the time [laughter]
and maybe maybe not all just me then I
guess um but
>> I get it wrong.
>> Yeah. So, so then what happened is I I
had to get things done and I was like
watch watching my weekly budget on
entropic carefully because I cannot run
out of tokens when I wanted to send it
off.
>> It's like a drug, isn't it? It's like,
oh, I check my supply. Yeah, I I kind of
choose what jobs I what I I will do
myself and like I don't let the model do
some stuff. But if I wanted to do like a
heavy troubleshoot going across a 100
files, I'd let the model do that, right?
Uh it's going to be a lot faster than me
trying to do it. If I have to rename a
couple of files and and and and do
things like that, I'm not going to say,
"Yeah, go ahead and go do that and
waste, you know, a million tokens
renaming files or or or you know,
reading a whole file and then outputting
the same file into a new directory."
That's ridiculous. Yeah. So anyway, I
was kind of hitting close to the limits
by by Wednesday and there was all this
speculation about like Gemini coming
out, these batting screenshots getting
shared. It was pretty cool to and funny
to see all of that, but I was like
betting on it that it come out and then
I have access to a free LLM on top of
everything.
>> And so
>> the moment it came out, I literally have
like oh and I also have the Claude Code
web version and I thought what you know
I never tried it. I have $250 credits
there. Uh, and I thought given the API
calls a day for me, easily around $100.
>> So, you have a different credit set for
the different
>> Yeah. So, so with Entropic subscription,
you you get your Claude Code access, but
then they give everyone that has a
subscription, depending on their tier, a
number of credits. So, like if you're in
a prom max 200, whatever, you got like
$1,000 to try out their web interface.
So, you can go to cloud.ai/codes AI/ uh
code and it's a web version of cloud
code. So you don't have all of your
CLIs, you don't have like your
credentials, everything there. It just
have access to your git repo and it can
do some tasks. And I was like, I I want
to try out a new um interface like a
React library of components. Uh and I
was like, you don't need any tools for
that. So I have $250 credits there.
Given the amount of credits I spend a
day, $100, I think in three days, um I
will be able to spend $250. I I again I
thought that this was going to be like
API because the API is charged very
heavily whereas you pay $20 a month but
it telling me that I'm using over like
$600 of API calls. That's how the tool
tracks my calls and says this is how
many you are actually spending if you
were actually paying for every API call.
So So yeah. So so 250 I thought was
going to be over in a minute, right? And
it literally spent $8 only and it
finished. Why? Why is it because is it
because the web one is more efficient?
>> I don't know. Maybe it's because it's I
don't I I have no idea why the web was
so efficient.
>> So, are you are you on the max plan or
the pro plan?
>> I'm on the pro. I pray $20 a month. I I
I'm I'm diversifying, right? I have put
$20 in OpenAI, $20 in, and $20. And just
to reiterate what you were saying, so
you have you're on like a weekly limit
by the sounds of things for the
>> pretty much every provider gives you um
a 5 hour session token budget and then a
weekly token budget. Before they only
gave you 5 hours session token budget
and then people were running it and
sharing keys and they were just
maximizing it um by by having pools of
keys and so now they put uh like weekly
budget um which basically equates to
about 10 sessions that you 10 5 hour
sessions.
Besides like you know tracking your your
tokens and your credits, have you been
trying to make your your prompts more
tur? I mean I think we I think we agreed
that that MCP sucks. So
>> yeah,
>> have you dropped you you don't use any
MCP servers?
>> I I just run MCTP on my cloud
coordinates has zero MCPs connected.
Maybe it's because you just updated.
>> What other token optimizations have you
been making?
>> Okay, so
I find these prompts are so difficult to
get right unless you are running an eval
suite that evaluates your use cases to
to make sure that the prompts still work
with the model and all that. I I I don't
want to write them myself. I don't want
to maintain them myself. So I'm using
spec kit and I've customized it a tiny
bit. I removed anything that says you're
gonna write out everything uh into a
different markdown file. I'm gonna make
sure that you use uh beats Steve's thing
to do the the issues, right?
>> Be kit and and Steve.
>> Yeah. And and it's going all right. I've
tested it out. I I started recording. I
say this every time we talk like I have
got a recording, but it's like I'm still
not happy with it. [laughter]
Um but yeah, it's going all right.
>> Oh, by the way, can you can you take
take the framing out? It auto frames
your your face all the time.
>> I don't know how to do that. Uh it's
>> Are you Are you on me last time?
>> I think I think it's called center
stage. You just click the green icon and
you turn off center stage.
>> Uh it's on Zoom.
>> I'm not too sure. I think it's I think
it's a a Mac feature actually.
>> But where do I disable it then? Because
I'm on a Zoom call. [snorts]
>> Yeah. Well, in in my in my in my Mac OS
machine, I have this green icon when
when it's like showing the usage of the
>> and then [snorts] you just click the
green icon and
>> enter stage.
>> Oh, it was not on.
>> Now it's Now it's off. Now it's on.
>> Okay, there's something.
>> Maybe. Okay, maybe go video settings. I
It's probably in video settings
and zoom then. Yeah, it's zoom. Of
course.
>> Zoom doom doom.
>> Video settings.
>> Yeah. Autoframing. New. Turn that off.
>> Video settings.
Virtual background appearance.
>> It's It's an appearance.
>> Autoframing. Oh,
okay. It's off. Wait. My video is off
too now.
>> Okay. It stops, right? Hopefully.
>> Yes.
>> Okay. But you were saying do you like
how do you keep your prompts st
your token usage down
>> cuz like I'm like for example when I'm
working with
>> uh Claude
I need to give quite a lot of context
cuz I'm working on a huge code base to
be honest and like I'm never too sure if
it's like looking through my git logs.
Um, I I know it kind of shows the
thinking, but I'm like it doesn't like
every stage when it does the thinking,
uh, it doesn't really give a good
feedback about how many tokens it's used
or what it's actually consumed. So I so,
so I'm always in this like gray zone of
like am I using my tokens efficiently or
not? But
>> so there is this
>> Yeah,
>> there's this one person ship hu I think
she wrote a book about AI cookbook or
something. She posted on LinkedIn a
script that parses through your cloud
locks, like your Claude Code locks
because, so this is interesting. When
you run Claude Code, it stores all of
your conversations on JSON files in
somewhere home directory or somewhere,
right? Including every single
screenshot. So if you take a screenshot
and you paste it, that gets checked some
u uh what's it called? Uh B 64 encoded
and written down into this JSON file. So
they she said no not she some people
said that there were like gigabytes of
of of logs in there but maybe if you're
really curious about your token usage
you could try and she posted a script
that she did to parse it and it
basically in her case she wanted to look
at how many times does um a tool use or
a command fail for example it tries to
read the file and says I can't find it
oh I'm in the wrong directory I need to
go up a directory and so on right and
she says is like she analyzed how many
times which commands failed and then she
updated her cloud autom to like specify
like if like focus on this focus on this
>> there's definitely like people who are
starting AI consultancies right now and
I can't help but think that would be
like a key tool some key tooling in your
AI consultancy like you would you you
know if you're trying to upskill a team
you would want these logs or you would
want to run your your tool over those
logs
to better uh train to make people better
at using the AI.
>> So I think this is a very good
observation that the fact that as an
engineer we must know how to use these
tools. Using these tools is not easy.
they changes so fast. Um, and there's a
lot of
AI slop being generated as solutions
around these tools.
>> Like the the signal the signal and noise
is pretty intense.
>> Actually, come to think of which which
Reddit are you on? because I I posted
like a genuine AI question on the VS
Code Reddit and everyone was replying to
me like I was like they didn't even
understand what I was saying.
>> The VS Code doesn't sound like somewhere
you would get I I'm usually on the
entropic um on the LLM devs. There's so
many. I I have no idea. Like sometimes
and but Reddit also suggests posts that
do not belong to any subreddit that
you're subscribed to, but because you
visited one Reddit, it will sub um it
will suggest these post because they're
similar to what you normally look at. So
>> usually if I see something nice, I I
subscribe to that uh subreddit and it's
like Claude Code, cloud dev, llmdev. So I
like a whole list over time.
But but Reddit as as a platform seems to
be the the best one for you still.
You're not in you're not one of these
discord disco bunnies. Nitro
>> course is a bit like slack where you you
can have like a thousand slacks
connected and you don't remember. You
don't know what
>> so busy so noisy.
>> Also I have like agents of AI uh AI
agents AI agents AI gamedev AI prompt
programming.
So you you just hop between them and
just just try to I mean
>> it's on my front feet usually sl
artificial
artificial intelligence.
>> The other person I the other person I
like to follow is is Simon Willis. Do
you ever like check out his blogs? I
mean he blogs about AI like three times
a day almost.
>> You shared it with me, didn't you? You
share me a blog post of him.
>> Yeah, I like I like his work a lot. He
and the thing that blew me away about
the Gemini 3 release is that he was
talks he he has his like own unofficial
or official test of drawing a pelican on
a bicycle and the Gemini 3 did actually
a really good job compared to all the
other ones. [snorts]
>> There's also Chpt Coding and Chpt Pro
just I thought those and yeah and Cloud
AI cloud Claude Code.
>> So you're in like 20 Reddits. Oh my god.
You're you're living. You're you're
you're deep diving, man.
>> No, I'm only listing the AI related
ones. Not [laughter] the other one. Not
the one.
>> Cursed AI. Oh, no. Not Not that one. Not
cursed.
>> What about all the Don't you have any
other interests like uh kite surfing or
whatever the thing?
>> I'm not listing those.
Um,
>> I did did you see that thing on hacker
news about
uh Europe relaxing their GDPR rules
around uh AI? And um when I looked into
the comments, most most people sort of
interpreted this as a victory for
cookies. So we don't have cookie banners
anymore in Europe. It's funny, right? Uh
this whole cookie thing. But but coming
back to the the the Simon Willis uh test
of asking you to draw a pelican on a
bike. Um we were talking earlier about
like benchmarks that that people run and
I think that's a type of benchmark right
that's that's interesting that other
people um you know evaluate a model
against um there was something
uh I forgot because
yeah usually I also just go off of like
what other people say about like I've
given if this test definitely if they
are providing like receipts right if
they're saying here's is a repo. Here's
the assignment. Here's how they did.
This is what I saw. And then also if
it's not like overexaggerated and
repeated cuz then it's like clearly AI
generated and it's so painful to read
through because it just repeats itself
and you're like, "Oh my god, can you
just get to the point?"
>> That's actually a good telltale sign of
AI. Sometimes it repeats, doesn't it?
So, that's funny because a lot of
subreddits um immediately ban a post if
it has too many emojis cuz it's like
classic AI to generate emojis or m
dashes and things like that.
>> So, um yeah.
>> Yeah.
When I'm using Claude, I I I get it I
get it to give me the solution and then
when I validate it's working and then I
feel like I have to go through this
uh step or stage or phase where where
I'm basically saying less code, suck
less,
make it more tight
because it's
>> Yeah. So it's pretty verbose by default.
So with the crunch that I'm under for
delivering something right now, I have
and you were saying earlier you have a
really large code base and and you then
went on to say that it's very hard to
control how much token it's using and so
on. But when we go back to the very
large core bait part um when you said
one of the issues I have I I just give
it like a really big like seven-year-old
Golang monor repo that I adopted like
forked and added my own features on top
and usually hacked them in. So, it's
super messy code and I'm always like a
little bit, you know, reluctant or I
always try to just find the different
paths. I kind of know the codebase where
it needs to go and change things and I
was like, I don't know. I'm going to do
this myself. But then I tried
anti-gravity Gemini tree. It's a really
light 7-year-old codebase and it pretty
it was I didn't have to give it anything
like I didn't have an agent. MD or
anything like that and I said like look
I I'm adding a new CLI. So the root
right now main.go is at the root but we
need to move it refactor it so that the
the CLI are separate directory so we can
mint or build the different binaries
from it and and then we need to like
have a shared interface between those
two binaries because one will write
files on this the other one reads it. So
um where like I suggest we put it here
there I went a bit back and forth with
HGBT on the web actually while in the
gym which is pretty cool. You can just
talk and then you know brainstorm and
then I took the prompt gave it to the
anti-gravity and it did an amazing job.
It did a really good job. And in terms
of code, the test suite on that repo is
quite extensive. I went in, I wrote my
own tests in the in that one area. So,
it's kind of like a golden.
>> How do you write your test your in Go
tests or what?
>> Yeah. Um,
it's funny because when I first started
working with this repository, I was very
surprised by their test approach. They
have like this golden files approach
which is kind of very more popular now
as snapshot testing when they just
>> Oh, yeah. Um, Hashimoto talked about
that.
>> I think you shared me a post or or it
was on Reddit.
>> Hashimoto did a did a call did a talk
back in 2007 of how they wrote no 2017
of how they wrote Terraform and how they
wrote wrote their test suites and he did
different testing mechanisms and then I
saw he did a newer one uh
>> it was linked in the same YouTube video.
So I I I looked at the new one as well.
Very interesting. But that's how the
tests were written. So they were all
golden tests there. It's a directory
with all the poss like different
scenarios of of the config file, which
is ugly YL. I hate Yaml, but anyway,
it's a like I said, seven-year-old
thing. It uses a YAML file as an input
file. So, I just added a couple new
fields in the config, and then I
activated those fields in the test uh in
the test data. I ran the generate all
the golden files. I reviewed it if it
makes sense. And then I just let uh
Gemini and the LLM do its thing. No
matter like the code I I saw like
there's a whole bunch of lines there and
it's like duplicated and I committed
everything and I said like hey can we
look at this and clean this up and it
says done it's cleaned up. I ran the
golden test fail error. I was like okay
there's something wrong because you
duplicate like all of these additional
lines are here that shouldn't be there
and then the LLM says oh yeah my
refactor did not take into account this
edge case and that add case. And I just
went like revert everything. I stick to
the commit. I [laughter] don't care my
tests are passing. Um, I'm not going to
bother too much. Again, under crunch, I
can understand that over time. Um, it
can result in a horrible spaghetti code
and hard to maintain problems.
>> So, it's kind of like the balance.
Classic tech.
>> Speaking of reverting a commit, I'm very
nervous about letting AI make a commit.
>> No. Yeah, I didn't revert to commit. I
on stage like reverted all the changes
>> cuz sometimes claude
>> last commit
>> I feel takes some liberties because like
I don't want it to uh do get commits. I
should put that
>> absolutely not. No,
>> I should put that in my claw. MD, but
sometimes it does it. I'm like dude
stop.
>> Well, so I am very reluctant to let the
LLM do a commit. I usually I usually
stage the files and then I let I I
prompt the LLM and then I look at what
it did against the stage files and if
I'm happy I then stage additional files.
So that's like I I should be committing
more aggressively myself but I'm just
staging files until I'm happy and I know
that things work and then I want to do
commit because I like my commits to
work.
>> What is your workflow? cuz like for me
it's like I um I'm using VS Code at work
to be honest but but if I'm if I'm at
home I'm using claude code and then I
have another terminal where I'm like
checking the things that it does and
then in that terminal I'm also doing get
ads but I'm just curious if you are some
in turbo lazy get mode or something.
>> Yeah, so I have I have VS Code and I use
a shell the terminal in VS Code uh to
run Claude Code which has an ID
integration. So that means Claude Code
will open files into VS code to to show
the diff. If I didn't allow it to to to
make changes itself.
>> Uh also I don't know if it always
captures it but if you select something
then when you do a prompt it will say
and you have also the selected so it
reads your
>> Yeah. Um and then what I do is I stage
files that I like know that I'm happy
with the changes it made. Um just in the
UI on the left and you're just gonna
>> Yeah, I st but when I do the commits I
always do lazy get because I used to use
G g I used to like sign my commits and
for some reason it doesn't I I don't
have it now with my latest laptop but
I've got the habit now to always commit
through the terminal um because the
>> so you stage in VS Code and then you
jump into Lazy G to commit.
>> Yeah. Yeah,
>> it seems a bit more audible, but uh
>> no. Um
>> I I kind of like the VS Code commit
because it does the automatic commit
message summary.
>> H you know um
my commit messages are kind of like just
usually one headline of the feature or
the chore or the fix and then some
comments. Usually I look at the stage
files that I look at like I go through
the files on the left while I'm looking
at like oh we also added this oh we also
added that and then I just um yeah it's
conventional commits and and some
message body um but that's my usual
workflow
>> but but then when you were saying
earlier we're afraid of letting it
commit but when you work with Claude Code
web and I I just like told it here's my
repo um integrate this new react
component library to replace my custom
components that that are kind of bad
that you wrote by the way. [laughter]
Um, and then Claude Code just created a
branch and of course it has the only way
to get its work is by it committing and
it was committing all the time and uh it
was pretty cool. I just pulled the
changes to my local and then made some
changes, pushed it and
>> I guess that's like a that's like the
thing I want to control. I can get yes
or no. I can get yes or no.
>> Oh [snorts] my god. So with the with the
PR workflow that I'm setting up the
release workflow now I'm I'm I'm doing
release please with um with go releaser
and I want the PRs to be merged uh
squash merged merged so no matter how
many commits it did on the branch it
will all become one commit when the PR
is merged.
>> Right. Right. Right.
>> And that works well with with like
release please to find what has changed
and then it creates a release PR keeping
track of every uh PR merge that
happened.
>> Yeah. I mean when I'm devel I mean I was
developing like a Terraform workflow
yesterday
I I kind of wish that I could have got
AI to do it but it was one of those
things where this [clears throat] a this
workflow was part of all these other
workflows and these all these other
secrets so you could only trigger it
through GitHub workflow. So I was going
really really really slowly because I
was like get commit uh you know - a - m
fix uh and and git push and then and
then triggering a workflow.
One good thing is that I got AI to to be
able to read the the workflow error log.
>> Yeah.
>> And then it can help me iterate on that
thing. But it's still it was slow to
iterate and it was uh yeah
>> CI/CD is always slow.
>> That's why you need you need project
like earthly or what's it called? Dagger
from the Docker Solomon.
>> I've been I've been dipping into Dagger
but I never got much out of it. I mean
have you got any joy out of
>> Did I say Docker Solomon? I meant
Solomon hikes from Docker. Yeah, I
haven't.
>> [snorts]
>> So the I really like the idea behind
Dagger.
>> Um because Doc Solomon says
we failed. If Docker is just a runtime
for containers, Docker failed um because
the idea was to improve the the
ecosystem dev ecosystem, right? um
Docker Docker Mac uh is basically a
transparent way of um getting things to
work on your machine and that their
focus was always dev tooling.
>> Not super transparent to be honest, but
yeah. Okay,
>> that was their idea, right? The whole
idea of um cloud when they released
Docker, their internal container stuff
was dev tooling and they really
optimized the dev workflow, you know, of
building and having it run across
different machines, right? build chip
run was always uh
>> their their moto right um and so with
Dagger says I'm focused on the dev
workflow like CI/CD is horrible
experience we want to be the ability to
um reproducibly have our CI/CD pipeline
and at the same time there was
earthly.dev dev and there was another
one as well that were trying to do the
same thing
>> but but but I mean they must have failed
because I don't know anyone who's doing
this stuff locally everyone GitHub
workflows have won the day
>> and uh I mean I have this really
>> this is a horrible experience um but I
had the same with AI though it's really
nice I used the GitHub CLI did you is
that how you integrated with the with
>> Yeah yeah yeah I have this like MCP
snippet where I just expose the the
actions tool tool set because the MCP
>> is too big
>> is too big from GitHub.
>> So this is where I do few shop samples,
right? I I basically push my commit to
the PR the the workflows run and I I go
into a Claude Code. I run the shell. I
say GitHub dash uh or GitHub PR check
number three. And then it shows these
are all the checks that run and these
succeeded. This how how much they took.
This is a job uh description or the link
to the job ID. Um and then you just go
uh GitHub CLI um PR or is it run view
and then you can do either run id or or
with the argument flag d- job and you
just pass that job from that PR check in
there. And when I first show to cloud
code I do these commands myself. So it
sees the output and you can you can do
d-lock failed and whatever. Then it
started using jq to like you know JSON
outputs to parse through things. Um, but
it was I don't know. It did things that
I thought wasn't really necessary. But I
don't need an MCP. I don't need
anything. I'm just showing it the GitHub
CLI command
>> and then it just goes in a loop. It just
like, oh, I see it's failing here. Oh, I
see this problem.
>> So, I'm in I'm in that team showing
GitHub what to do. But like
>> not repeatedly, not every time, all the
time.
>> My my colleague my colleague asked u
a good question in a sense. is like he
asked like what do I uh I need to use
the GitHub CLI what should I use the
GitHub MCP
or some other approach and then
basically I chimed in and said like well
the GitHub C MCP sucks because it's too
verbose and it wastes your tokens but if
you use the GitHub um API or CLI sorry
if you use the GH CLI you have to tell
it how what to use. So my my colleague
is a bit of a noob.
>> I I think to be honest he wasn't he was
a little bit confused by what I was
saying like you have to teach
uh GHLI you have to you have to teach
the the the
better way is to teaching the AI how to
use the command. I mean that kind of
blew his brain.
>> Cloud skills are full right?
>> I need Okay, then I need a skill to use
the GitHub CLI. That's a good
[clears throat] weekend project. That's
a good weekend project.
>> For sure there's one.
>> There's actually people that are writing
LLM like agents to create skills and you
give it a couple of prompts of what it
does and then it will explore the CLI
and it will write an a like an agent
skill which I think is a bit dumb right
uh but like if it works it works. The
whole idea of the of cloud skills is
progressive disclosure. So you start
with like the first basic flows and then
if you need more do this, if you need
more, do that. And then cloud is
supposed to like parse through it and
and and progressively
read more to like be able to do more.
And and yes, that should be the way that
you don't always have to show how to use
a GH CLI every time, but you
>> You're right. You're right. I should
have thought about that cuz I think my
colleague just was thinking I was an
idiot or something like that. Like
>> Yeah. Yeah. Okay. I need to I need Do do
you have you created any skills?
>> No, I've only like played with the beats
skill, right? But I don't like I
integrated it so it's living somewhere
in my I think it's in your home
directory.
>> The the problem is in my world like I
have my personal claude code space which
is fun but daytoday in my in my
workplace I'm using VS code and VS code
doesn't support skills. I believe
>> they they have something in speckit they
have like they have a different way now
to expose different agents sub agents
into your VS code as well when when I
look at the specit latest presentation
from from the guy who who uh is like one
of the main advocates on YouTube he
showed how they have like integrated
speckit into vs code settings like
copilotcopilot/s
settings.json JSON or something and they
have like declared all of the agents
that the specket provides. So then when
you are in the chat window of VS code
you can switch between these agents and
one of them is like the speckit plan
>> better at VS code or something like
that.
>> Look at the the speckit YouTube video
channel and then one of the latest
videos he shows how how they they are
they are exporting the LLM specket
agents into VS code. So I think if you
can do that, you have dedicated agents,
maybe there's a way to like have the
agent skills as well. Um cuz again like
they are constantly feeding off each
other, right? One one company brings
something out, another one copies it.
>> Oh, sorry. I'm getting distracted. My
wife is messaging me about the school
run.
What time do you have to go back to work
in 20 minutes or something?
>> Yeah, something like that.
You said earlier something about um when
your colleague asked and teaching it how
to use ah um I think that's where if you
are a terminal like somebody who likes
to live in the terminal tools like cloud
code kind of feels very cool because you
are then you already are used to create
pull requests using gh right you already
go ghbrr create-fill
whatever u you push your branches from
with git so you're you're not someone
that goes to to the web and then clicks
on a button to create and so and so on.
Right?
>> So that's where I think that transfers
easily into these other
>> going back right back to what Steve Jo's
talk was about was that like he was
saying that the statistics in companies
where people are adopting AI and they
are are are by all measures uh becoming
more productive and then there's the
senior staff engineers who go who who
who there there people in a company who
are not adopting it.
>> So what do you do with those people who
are not adopting it? And to be honest, I
see that in my workplace. It's like I'm
using AI like uh like like you, I
suppose, and I see the benefit, but
there's some there's there's there's
definitely uh members of my team who are
just completely not interested in in
picking up these new workflows. I'm
like,
>> so I had I I went to a Q Q&A conference
back in June and one of I did I was
invited on the panel and one of the
audience questions was exactly this like
as a team lead because QA and and and
and LLMs it's quite a big a big changing
there right automation of of LLM
generating test cases and all sorry
automation of of tests and generating
test cases and test data um so one of
the the people there was like how can I
help or ensure everyone within the team
um picks up on these LLMs um and and the
reply from the the panel members and
some some somebody worked at Singapore
Airlines and he says we we create a
prompt database we do regular brown bags
we we um you know we share people with
different workflows on on how it helps
uh but at the same time when you're
saying like there's these senior
engineer who is really reluctant to to
pick us up I think maybe you need
someone of those in your company, right?
Maybe you need someone. Uh, and that
that makes me think of a Reddit post I
saw on the DevOps subreddit, I think
this week, which was like, I'm so
frustrated. Everyone in my team is using
LLM and I'm like the only one that's
actually doing manual reviews of all
these PRs and I'm reading all these PRs
and I'm like the I feel like I'm the
only one like commenting on these lines
of code and saying like why is this like
that and so on. So, so um
>> but still I mean maybe yeah, I see your
point. There's definitely
definitely like an argument to be made.
Okay, here's another problem that I hit
out of work, right, is that strictly
speaking, we have a co-pilot license
and I think there's basically the JBT
models exposed there,
>> but I want to use I want to use Claude.
So, VS Code gives me an ability and to
paste in an API key for uh Claude as
well as Google Gemini and all the other
stuff.
>> But then you then you have this
conundrum like hm the company hasn't
expressly approved this model and that
model.
>> Can I use it?
>> Yeah. And and then furthermore, there's
a whole uh topic around um the you know,
who's going to pay for this.
>> Um
>> I feel like these are unresolved issues,
but maybe maybe let's get back to I
think I think I need to do the school
run. It's it's it's icy out here. My
wife doesn't want to drive in that in
the
>> I need to take the kids to school. Okay.
Yeah. Um just on that point right for me
currently
I'm I'm using I'm this is I think if
you're in in infrastructure and DevOps
I've always said these are to me are the
most interesting fields to be in because
it's the same everywhere. It's not
secret sauce. It's not this the the the
the secret terraform everywhere.
>> What do you mean?
>> Yeah. I mean the ideas of running
infrastructure on the cloud and setting
up uh you know observability dashboards
and monitors and CI/CD pipelines it's
not a core differentiator or secret
sauce of an organization it's not like
something that's propri saying yeah
right the skills are transferable
>> exactly so I used to coach or work with
product engineers and then try to
convince them like why do you you know
bother with working purely on business
um um business feature uh tickets, user
stories and so on. I isn't it way more
interesting to be in the DevOps space
and then be like exposed to something
that's you much more transferable
between organizations when you want to
change as well. Um but that was then
until I got into a position where DevOps
was really seen as a cost center and
then you are like completely blocked to
do any type of refactoring because why
would you want to do this and and um you
know
>> this sounds like my life right now to be
honest.
>> Yeah. Yeah. Yeah. And and I was always
super happy to be DevOps. Uh but then
you have the ne negatives which are like
um you know unthankful because nobody
cares everything's running. and people
are wondering what is this guy doing
until there's an outage and then
everyone's you know breathing down your
neck and why is it down [laughter] and
then uh I had a few companies where
there was like a celebrity uh but that
was a pretty dire state right because
the whole website was down and you
managed to recover it within x amount of
time and then you're like
>> this hero [laughter] this hero thing
like yeah you saved the day dude you got
the website back up maybe it shouldn't
have been down in the first place for
Christ's sake
>> yeah it depends also if you're the fire
starter right if you're the ones putting
fires all the time that you are actually
the fire starter [laughter] but yeah
yeah yeah um yeah it's that was
>> industry isn't it
>> so that's where I think using LLMs for
DevOps related tasks is easier because
it's like it's not so proprietary and a
lot of the things I do I actually take
an open source library and I I work on
adding features there so I'm saying to
everyone like hey don't I I this rule
doesn't apply to me because I'm taking
an Apache 2 library I'm putting it as an
open source fork and all of my cloud
code generated code uh is not related to
any company data. But but like do you
think like for example do you think
Google uh trains
the Gemini model on their own
proprietary uh code base? I I I think
there would be afraid that there would
be excilation of data, right? That I
think there's clear cases of of
>> So basically like you so if you went to
Gemini say create me a Google search
engine, they would it just wouldn't
>> I'm pretty sure that there's a whole
[snorts] like actors trying to
excfiltrate as much. I mean they all
they always say that Quen when it was
like trained and like how many% of the
total cost that OpenAI and other
companies spend quen AI from China was
developed with like a fraction of the
cost and they say it's trained on
synthetic data. trained exfiltrated uh
from all of the um you know open AI
models and so on right so so there's
definitely organizations trying to
excfiltrate any type of proprietary
information from companies and I think
Google must separate the training data
for
>> but I don't think other companies have
that luxury I think that I think the um
I mean
okay
obviously people are trying to keep
their own proprietary source code um
safe and obviously their their private
GitHub repos are not being indexed
by um by a model. But nonetheless,
um,
you know, I'm sure if you ask a CTO of
of a company with a big code base and
they and and you asked and you said to
them like, uh, would you like a model
that knows about your coding style and
your uh, and your existing code? They
would jump at it, right?
>> Yeah, there's companies doing that,
right?
>> So, there's a bit of a a chasm there,
like how do you bridge that? So, so
source graph is had very has a a lot of
interesting YouTube videos because
source graph positioned themselves as
the originally enterprise solution for
code like search right so they have big
contracts with enterprises for you know
propriety self-hosted deployments um to
index the code and for you know JP
Morgan whatever huge clients to to for
their software engineers to quickly find
um code or relevant um microservices for
certain functionality across all of the
code base right so [snorts] then they
they are also having their LLM on top of
this now like Annie I think it's called
or uh
>> okay so there there are products for
this okay
>> yeah yeah um and and a lot of yeah when
earlier
>> I mean I mean GitHub GitHub needs to do
this next seriously I mean like
>> so when we talked about I'm using VS
Code and um and then I'm not sure which
model I have access to I can give it an
API key and what's the contract like
when I talk to someone from GitHub he
was visiting in Vietnam. He told me that
they have very strict um contracts in
place with enterprises and how they are
using the code. And also when you sign
with with the with Microsoft and on the
GitHub enterprise, you decide like
for example, one of the risks that
companies have is that they get um a
lawsuit for copying code, right? if they
use an AI, how can GitHub copilot
deployed across all of these people that
are using the AI guarantee that I'm not
generating code that is proprietary or
that is copyrighted and that's now in my
codebase. Like imagine it's a GPL and
now I have to open source everything. Um
so so that's a risk. So so Microsoft has
these contracts in place that says first
off it it sources different it generates
solutions
>> mess up like what stops one of my team.
>> Hold on. You haven't listened how they
solved it yet. So he told me that they
are generating different code solutions
and then they do a cross check to any
like existing and licensed code in their
database and then they're you know
rejecting anything that could be and
then they're sending you only the one
that is passing all of that. Sometimes
co-pilot gives you code and then
suddenly gets removed says like this is
a copyrighted code I can't give you that
>> and then the second thing is what he
told me is that Microsoft actually will
sponsor you in if there's a lawsuit if
you if you prove that this code is
generated by co-pilot and Microsoft will
be part of the lawsuit that's why he
told me like part of the contract
stipulates that if your company gets um
gets what's it called um
>> extra lawsuit yeah no gets a lawsuit
against you like um then
>> Oh, okay. Because you're using source
code that might be interpreted as being
proprietary.
>> Yeah. So, so basically you said it's
very hard but he was saying that's one
of the things Microsoft has to convince
enterprise users that they are
protecting. seems like a a good tact but
but I mean at the end of the day that's
that's Microsoft that's GitHub they they
know about your proprietary code but
like
>> like how how about this scenario like
there's a new there's a new edit there's
a new IDE let's call it anti-gravity and
someone like oh wow this can help me do
my job and then he just launches on his
proprietary codebase and then all of a
sudden
>> so that's where Um it's it's important
for organizations to put in place AI
policies and and and books um what's it
called? Runbooks or guides because if if
this happens, right, it means that the
user hasn't really been informed um or
or [snorts] he's like maliciously doing
it. He knows that it's not allowed but
he's going to do it or he's not know he
doesn't know about the risk and he just
installs it and then he hasn't been
trained. Right? So, so in my
organization we had um to sign a
contract that says that I will not copy
proprietary and this is we had to first
learn what is propriety and confidential
information. who like no be no no be uh
no before trainings and then also sign a
contract saying like
>> you know if they find out that I do this
but I'm not doing it [laughter]
>> then I will be
>> like one thing in in some workplaces
that I like to do is to encourage the
use of code spaces do you know GitHub
GitHub code space
>> yeah I haven't used it but yeah
>> but the cool thing about that is that
like the code is there and it's not on
your local machine and it's there's a
nice clear division And
yeah, I'm just finding it hard like
because
like yeah, like we we like I I we have a
special API key with u with a particular
vendor. I'm not going to say who, which
gives us private confidentiality with
the code that we use on it. But then
it's not clear if we have permission
from the client if we can
use uh
>> snippets of the like like to be honest
I'm not even using
the code base right it's it's like what
you were saying with the Terraform it's
like uh I mean not Terraform like
infrastructures code it's like maybe I'm
trying to debug a GitHub workflow that's
no different to anyone else's Terraform
plan and apply and I would want to use
that that that paid for model and and
and then I'm always a little bit like hm
I'm not sure I can do this because
technically what I've just copy and
pasted here is proprietary code but it
isn't really and yeah all this ambiguity
makes makes the whole experience a
little bit crappy
right um but but I was going to say that
you are somebody that then worries about
that either because you've been trained
or informed about these risks. Um so you
know that's because your company has put
in place policies so that you are
knowledgeable about these things right
if if you look at many other people
>> but the policies are not clear they they
just create FUD you know what I mean
like and FUD is worse than
>> fear uncertainty and doubt
>> which is a crypto term like crypto
>> I mean [sighs]
how many how many companies have really
clear policies with AI like you can only
use this model and that model and that's
it. I don't I don't think I've seen any
I don't seen I don't think I've seen
any.
>> I think that's where your organization
is providing you with access to AI and
my organization is giving us access to
GitHub copilot. I I am naming it. Um but
um that also at the enterprise level the
organization level they control which
models the users have access to. In our
case, we can only select the ones that
that they have approved in our
organization.
>> Okay. In in code in VS Code basically,
right?
>> And I only know this I don't manage
this, but I know this because somebody
in in the chat asked like, "Hey, they
just released Sun 4.5. We can only
access four. Can you give us access to
this new model and then somebody has to
go into it and then give access to it?"
>> You say that and I think it's I think
what you're saying is true, but at the
same time, VS Code gives me the option
just to paste in an API key. So like
>> Yeah. I don't know like I'm not using
the VS code one and again I'm I'm
excusing my not using it by only working
on on repositories that are not private
[laughter]
and um and basically if I do something
um yeah anyway so so I don't know if I
can add the key but earlier you also say
if you use VS code in the first place
right um the guy who was working at
Microsoft was telling me that the plugin
system at the time which is more than a
year ago had no control about like VS
Code did not have any any constraints on
what the plug-in could read. So a
plug-in could just read anything from
your your your folder, right? And I
think they have improved that now
because now whenever you install an
extension, it tells you like, "Watch
out. Do you trust these people?" And
that's all they did basically.
>> Yeah. I mean, yeah. V extensions.
>> Yeah. Like
>> Yeah. I mean, that's a good that's a
good uh segue because, you know, people
go like, "Oh, did you get approval to
use these model thing like that?"
They're like, "It's there. I mean, I
didn't uh you know, if they wanted to
prohibit it, they Yeah, but you should
get approval before you use anything.
It's like, how many extensions do you
have installed? You're like, how many
tools do you have installed? Did you get
permission for every one of those
extensions or tools that you get have
installed? I mean, you're being
ridiculous. So,
>> and that's one of the things the
Microsoft guy told me as well. This is
something that they as the uh copilot
extension implementation uh provider he
he said it was actually GitHub acquired
by Microsoft right he said that they
guarantee that it will only read files
that you have open in the editor it will
not read anything from the disk so um
that's the type of guarantees that you
like I don't know that until he told me
right it's probably somewhere in the
fine print I don't know about the MD
gravity right now uh from Google I just
installed it I gave it the code base
which is open source uh to work on and
it probably read everything anyway and
it must be able to because if it can't
it wouldn't wouldn't have been able to
do such a good job to find exactly where
it needed to change what right
>> yeah
well okay I I wanted to end the the
conversation on something a little bit
more positive but um I guess the
positive thing is that the all these uh
we're living in a really exciting time
really I'm really digging the fact that
I'm alive I mean it's amazing I mean
like the future's is is bonkers, isn't
it? It's incredible.
>> So, I agree absolutely with what you
just said, but my second the second part
about the future is where I'm like, for
how long? [laughter]
For how long?
>> I do feel I do feel this stuff for $20 a
month.
>> I'm probably like in
positive. I'm I'm in that that maybe the
plateau camp like I do feel like we are
hitting a surge cuz some people say like
oh this thing is just going to go like
this like like like forever and I'm like
I think the models are improving but not
as fast as you would think cuz the
models I I feel have a limited training
supply right they're not they're not
going to invent new information that
it's going to plateau essentially by how
much good stuff there is for it to be
trained on, right?
>> If it's like software,
um
where it's just an ever expanding gas,
it takes up whatever resources are
available. Um I think with with AI right
now, it's taking up every resource it
can in terms of electricity and power
and um you know, GPU capacity. Um, it's
very hard to predict because
before the release of Gemini 3 this
week, everyone was like, "What happened
to Google? Gemini 2.5 Pro, me included,
in March was great. Now, what's what
really have to like be really good? Is
it delay?" Because is there a disarray?
There was so many articles about cuz I
searched Gemini 3.0 release pretty much
every day to find out if there was any
news and every article you're like,
"When is the next?"
>> Yeah. because I had to get things done
by Wednesday and I was like hoping
Gemini 2.3 zero release I get some free
credits.
>> Your claw tokens or your claw budget was
almost expired.
>> Every article I read before Wednesday
was like
>> the truth and the disarray within
Google. Google is completely unable to
handle it. Google has lost has lost it.
You know what is about these TPUs?
there's a TPU supposedly rolled out and
give them the edge across the world but
the infrastructure is delayed that's why
Gemini 3 is delayed and blah blah blah
but then Gemini came out and everyone's
like Google did it Google is the market
leader look at at um how Google Gemini
3.0 zero is like this high in the
benchmarks. Does that give like um what
they called it again? Because you have a
much small smarter model within your
organization and you have full access to
it. That means that you have the ability
to accelerate much faster than the other
companies, right? So with the
announcement of Gemini 3 and in how good
it is compared to GPT 5.1 and 74.5 that
gives them not just like a small edge
that gives them a massive edge because
it means internally they can have like a
hundreds of Gemini trees running and
improving on the product.
>> Yeah. Yeah. I mean like I definitely
felt that uh maybe as long as a year ago
that that all of a sudden like Google
Sheets and uh Google Docs basically got
a whole bunch of new features and like
previous 5 years
Google Docs and the Google Suite felt
frankly unmaintained. And I was thinking
to myself like why all of a sudden is
there so many cool new features
>> after they lay off 30,000 people or
something. [laughter] Why suddenly
Google Sheets have all these new
features? H
>> well okay there could be other
mitigating factors but I did I did I did
basically occur to me that like maybe
they have some technology that we don't
have or something or or do you think
it's the fact that they laid off people?
>> No. No, no, no. I was just correlating
with the fact that they laid off people
and and so it must be llams or
something, right? How can you suddenly
have a whole bunch of new kettle of fish
>> after you lay off all these people?
>> Yeah, let's let's get back to that. I I
think I've got to wind this up. Anyway,
it's been it was a great chat. I'll
upload it. I'll share the link and I ask
everyone to
please like and subscribe and share the
podcast.
>> And and and and comment below like if we
said something stupid or something
better if we could have said something
better or then let let us know that
>> this comment will be Google did not lay
off 30,000 people. It was more like
100,000 people.
>> Yeah.
>> Okay. We're hoping to get in intelligent
comments. To be honest, the comments
section is always not as good as I hoped
it to be. All right.
So, you're instigating right now.
>> I wouldn't I wouldn't take it.
>> I speak the truth only.
>> I wouldn't I wouldn't take it.
>> Okay. See you. See you, Vincent. Bye.