Latest / Agentic Conversations (formally mlops.community) / The Modern Software Engineer

Transcript

The Modern Software Engineer · Agentic Conversations (formally mlops.community)

0:00
Nobody watches these anyways, as you know.
0:02
OK, OK. Well, well then what am I doing?
0:05
Here now, enjoy the side quest because now we all of a sudden
0:19
have the ability to do so many things and go down so many
0:22
rabbit holes that we didn't necessarily before.
0:25
Like if I wanted to learn how to fucking code front end shit
0:30
before it would take months but now I'm like.
0:33
I saw this skill that someone was making online and I forget,
0:37
I forget her name, but she's actually on quite a bit of good
0:38
educational content and she has this this other thing inside
0:43
there Claude skill. Or maybe like, I think it's a
0:45
skill that every time any implementation is made, it'll
0:49
it'll sort of like summarize the learnings the like the
0:52
implementation details that are actually that that, that you
0:55
should commit to your own memory, right?
0:58
And so you. As a human, not you as the
1:00
human. Exactly.
1:01
So different from like Claude Ception where it creates a
1:04
skill. That's right.
1:05
Yeah, exactly. It's not Claude Ception.
1:06
I mean that that there's enough that almost feels that that's
1:09
clearly top of mind for people is how do these agents like self
1:12
learn, become more autonomous that that that's almost obvious
1:15
in my opinion. But but I think what is less
1:18
obvious and interesting here is hey, yeah, I think this is
1:21
especially true for more junior engineers.
1:23
It's there's this huge question of, hey, the seniors are going
1:27
to keep having jobs because they're seniors and they've been
1:29
around a while. The juniors are now in the
1:33
workforce. They're like, holy, you know,
1:35
crap, what are we going to do? I need to, I need to actually
1:37
get a job here, but no one wants to hire me because I can just be
1:39
replaced by cursor. At least that's the running
1:41
hypothesis. And that just creates this huge
1:45
gap of basically like a training gap where people, you know,
1:49
typically you, you, you hire juniors, you kind of train them
1:51
into the senior roles. Right.
1:53
And you pair program you. Pair program with them, you like
1:55
review their code and and and you effectively are creating
1:57
this gap of knowledge between one generation and the the most
2:01
senior engineer and and and that generation like that gap has to
2:04
be filled some way. So I like this idea of using the
2:06
agents to actually fill that gap, because in the same way
2:10
that they can generate code, they can also understand code,
2:12
probably, you know, arguably better, if not at the same
2:14
level. Actually, there's something to
2:16
be said for that because somebody was just talking about
2:19
in the community Slack, how before they'll go into a new
2:24
area that they don't know about, they'll go back and forth with
2:28
just the chatbot to learn about it, and then when they implement
2:32
it, they can be better suited. For that's right, that's right
2:35
one maybe of those side quests you're alluding to, which is
2:37
hey, I just want to keep learning about the craft,
2:39
learning about the discipline and all the things that are that
2:41
are that the agent help me do more quickly.
2:43
And then I think there's also any I don't know what the
2:46
distribution is of these two. There's also an equally big
2:49
movement here maybe of of just yeah, I don't know if you ever
2:53
want to look at this. I am fine to just exist in the
2:56
world of ignorance and and you know we can talk about what that
2:59
means and how you can make sure that isn't just a recipe for
3:02
disaster at the level of like financial crisis when you start
3:05
having that, you know working bank that happened in banking
3:07
institutions. But but yeah, I mean there's
3:11
there's a lot of implications there.
3:12
And I think of that, that second category of we're going to
3:15
forget about it is more like non-technical people.
3:18
What do you do when you encounter a issue where the
3:25
agent will implement something and it's not the way you would
3:29
have implemented it, but it still works?
3:31
And so you're like, I guess I'm a little out of my depth here.
3:35
I guess this might be the better way.
3:37
So fuck it. Like I've come to just accept
3:40
everything and eventually it figures it out and if it doesn't
3:44
then I'll prompt it again. Yeah, Yeah, I, I think that's,
3:49
it depends on what task you're working on.
3:52
And and I think there's, there's somewhere you do want a stronger
3:56
say on architecture, on design, on having a handle on these
4:00
things. That's probably true.
4:01
Like core infrastructure stuff, You know, I would be careful to
4:07
have no control over some like core piece of elites at this
4:11
level, at the, you know, at this current point in time of, of
4:13
having no understanding of what's happening there.
4:17
That being said, like I, I do think that, you know, people may
4:20
not call it creativity. So, you know, don't people on
4:23
online shouldn't, shouldn't hate me for this, you know, saying
4:25
that the agents are creative. Yes, they're pattern matchers
4:28
that glorify pattern matchers that extrapolate.
4:30
But I think that there is something to be said about
4:32
creativity to respect what you understand at least, right.
4:34
So like you don't know this implementation, this way of
4:36
doing something and it's new to me.
4:38
And in that way, it's as if somebody had come to me and
4:40
presented something that didn't exist before.
4:41
And so I actually like that aspect.
4:43
I think that should be controlled, but but encouraged
4:48
in its own way. Exactly.
4:49
And so like, yeah, I don't I don't know next JS tremendously,
4:54
I don't know view. I don't know these things.
4:55
But if it got to that point and I don't have very strong
4:58
opinions, yeah, like it's it's clearly generating this because
5:03
it's pretty well supported. It exists in the world.
5:05
It's seen this a good number of times.
5:07
And so at least use this for a starting off point.
5:09
I'm not saying you should literally just let it go off the
5:11
rails but but you know, as far as a prior is concerned, if you
5:15
don't have an opinion then then it's not a bad prior to start
5:17
with. I heard from chip stock that she
5:20
did at our agents in production a few months back.
5:24
She was saying like, what do you think the hardest programming
5:29
languages for agents? And it had, you know, Russ,
5:32
Python, TypeScript, JavaScript. It was JavaScript.
5:38
You know why? Because there's so much shitty
5:42
JavaScript on GitHub. That's true.
5:44
That's very. True that it trains on all that
5:46
and then next thing you know it's like it's outputting what
5:50
it trained on. Yeah, Yeah, I think that's fair.
5:54
And where I think this is one of those places where I was
5:58
actually talking with Eno who's who's the CTO factory a little
6:01
while ago and he. Well, dude, you just dropped
6:04
something over there, Did I? Yeah, I think it's a name.
6:12
Nice, nice. I'm going to use that one.
6:15
That one I like. I was like, I didn't see
6:18
anything in my periphery. That's good.
6:24
It was also convincing. Good on the acting Good on the
6:26
acting. You have a you were a
6:27
salesperson once the time, weren't you?
6:29
You you know how to how to sell the story.
6:33
Hilarious. I gotcha.
6:35
Got me All right. No, you.
6:37
Were talking to you know. Well, I mean, not, not for any
6:39
like I'm not really not trying to to to name drop here.
6:42
I'm just, I was talking to him like legitimately not a little
6:44
while ago and and. My favorite part?
6:47
I'm the worst culprit. If I can name drop anybody I
6:50
will so. I'm a little That's why no one
6:52
can. No one takes you seriously
6:54
because you just lost all your legitimacy when you when you
6:56
your credibility. Yeah, exactly.
6:58
No, but but I actually no affiliation to factory at all.
7:03
And you know, I'm not getting endorsed by them in any respect.
7:06
But but I'm they actually do like a lot of really cool work.
7:08
And I I'm a big fan of like, you know, in particular, and I've
7:11
also spoken to they're both very smart.
7:13
You know, I spend more time with, but very, very thoughtful
7:15
about how he thinks about like coding agents and, and just, you
7:17
know, develop the future of AI software systems.
7:20
I think what he is his. Core hypothesis and thesis
7:24
around how they build factory androids, which is their their
7:28
coding agent is that the most crucial piece that you have to
7:31
get right as a developer is like the, the, the validations, like
7:36
the the testing harness, like the checks to make sure that
7:38
what if you let the agent go off autonomously on one small task
7:42
on many, many tasks for single agent, on many tasks for many
7:45
agents, that at the very least there's some set of checks you
7:47
put in place that the agent can validate that it's correct.
7:50
It opens up a browser. It tries playing around with the
7:52
website inside a little bit. It has some some tests that
7:54
exist in back end, you know, so some way of checking that it
7:57
that it's not going completely off the rails.
7:59
And that's I think the strongest thing we can do right now as
8:01
humans is like being still in the loop for those validations.
8:05
But the validation can only work in some use cases, right?
8:09
You can't validate everything, and that's probably one of the
8:14
hardest pieces of it. Yeah, I think that's absolutely
8:16
right. And, and it's a good nuance that
8:18
you bring up because he, you know, I was talking to him in
8:21
the context of this new thing. They, they're, they're rolling
8:24
out called missions, which is essentially their answer to this
8:27
question of long running autonomous agents.
8:30
You know, we see some blog posts and cursor, we see some blog
8:31
posts, Anthropic, I think probably even one from, from
8:33
Devon as well. And so, you know, there, there,
8:36
there's a lot of experimentation happening in how long can we
8:39
have these agents run? Can I launch them for an hour,
8:41
multiple hours, multiple days and weeks?
8:43
And what is possible if you remove the constraint of time.
8:47
And so mission is their kind of answer.
8:49
They're, they're, they're attempted at least positing a
8:51
future. And you know, what I think was
8:54
just interesting was as we were talking about missions, he, he
8:57
was acknowledging that there is something where they're reaching
9:00
almost like a, the frontier of what is verifiable or what is
9:03
like easily validatable. That's a word.
9:06
And, and, and you know, and we're trying to define what that
9:10
line is a little bit. Is it stuff related to
9:12
infrastructure? Is it stuff related to when I
9:14
have to go in like interact with some to get API keys from
9:17
somewhere, you know, like then yeah, you probably have a human
9:20
there if I'm deploying DNS things.
9:22
I mean, that's like tricky, right?
9:23
Because you're talking about how, how traffic is getting
9:25
routed through the web. And, you know, maybe that won't
9:28
always be the case, but, but at least for now, there's
9:31
definitely stuff and, and, and systems that we interact with
9:34
the software developers when we're creating production code
9:36
that are not yet verifiable in an easy way.
9:40
And and that's probably where coding agents even today will
9:43
like hit their limits a little bit.
9:45
But that's not to say we shouldn't keep pushing that
9:46
frontier. And I think the frontier will
9:48
keep getting. It's funny, if you look at the
9:51
graph of how long a coding agent can run without human
9:57
intervention, it's just like Moore's Law type of thing where
10:02
it's doubling every six months and you're seeing it go up and
10:05
up and up, right? So this episode is brought to
10:08
you by the good folks at ML Flow, the open source platform
10:12
for developers who want to build production ready AI
10:15
applications, enhance your AI applications with end to end AI
10:20
observability, all in a single integrated platform.
10:23
With ML Flow's Gen. AI capabilities, you can
10:26
evaluate AI applications using a suite of built in or custom
10:30
judges, visualize trace executions and agentic
10:34
analytics, and continuous mostly monitor evaluations, all while
10:39
tracking every run in one place. Ship better agents faster.
10:44
You know that's the name of the game.
10:46
Get started at mlflow.org. I notice that I will try and
10:54
kick off these tasks where I'm like, oh, this is a meaty 1.
10:58
And then next thing you know, it's like 10 minutes later or
11:01
even 5 minutes later, I'm getting hit with yeah, all
11:04
right, it's all done. Right, right.
11:06
That's right. Fuck.
11:07
That's right. Where are these long running
11:09
tasks that we're talking about because I can't find them.
11:11
Yeah, yeah, I, I think I think that's a really, really
11:14
interesting point. And, and what what is worth
11:18
really drilling down there a little bit is, is it's actually
11:20
important to find them. You know, that's almost like a
11:23
first point that I would make is like, it's actually important to
11:25
find those because you, you know, one of the questions that
11:29
people have now with how, how good agents have become and, and
11:32
especially as you think about scaling to multiple agent
11:34
workflows, it's, it's like, where does deep work still
11:37
exist? What does deep work mean for a
11:38
developer? Because deep work almost
11:40
requires at least like, I mean, definitionally when you know,
11:43
you know, like kind of brain science is like at least like 20
11:44
to 30 minutes or something, just get into something and then flow
11:48
Exactly. And multiple hours of that in,
11:50
in the, in the to actually get some really, really many tasks
11:52
done. If you can't find those points,
11:55
there's a gentle cast, you know, you ideally like an agenda cast
11:57
that is in that sweet spot of time, so you can switch to
12:00
something else and then kick off a new thing and then switch to
12:02
another thing in here. And you need to, you need to
12:03
like be able to identify those things.
12:05
But I think what, what is true to your point is, is what
12:08
constitutes those media tasks is, is getting constantly
12:12
shifting in terms of definitions.
12:13
And we're going to have to keep kind of pushing that.
12:15
What 1. So that's like thought one.
12:17
And the second thing that may be interesting here is I I find
12:20
that it's it's probably more true in like brownfield code
12:24
bases where the verify the verification step, the
12:28
validation is actually takes a lot of time, right?
12:30
If you imagine running making a change in Google as an example.
12:34
And you and I don't work at Google or anything like that.
12:38
So so take what I'm saying with a grain of salt, but I'm just
12:39
using, let's say Mughal. If you don't want to use a name
12:43
Mughal. No.
12:44
And you have these like massive multi million lines of code and
12:47
then you make like even the smallest change.
12:49
And that's certainly is before that even gets reviewed by
12:51
human. You wanted to go to like the
12:52
extensive set of testing suite that exists.
12:55
Takes hours. Takes hours now meanwhile,
12:57
you're not going to let go and you know you don't want to like
12:59
spend their baby time babysitting that and and and
13:01
fixing a little error here. And I wonder if the tests and I,
13:05
I have no idea what it's like so we could ask somebody that works
13:08
at Google, but I guess Google, yeah, their their tests probably
13:13
happen where you do a few. Does it pass that and then it
13:17
gradually increases? That that's correct.
13:19
That is correct, yeah. So I mean their build systems
13:21
are sophisticated at I don't know about Google, but certainly
13:23
Moogle the build systems are are sophisticated enough that they
13:26
that they are they're good at identifying sort of what what is
13:29
impacted by the change that you've made.
13:31
And so that is true and that but you know, that's all to say that
13:34
that iterative loop of try a thing, build a thing, test the
13:37
thing, fix the thing based on some error that it detects and
13:40
and your errors can have many different flavors.
13:42
That is something you that that loop can happen multiple times
13:45
as as as a system gets built. And I think that's I I think in
13:49
my in practice, I've seen that's where it starts getting a bit
13:51
meteor. Yeah, like a one shot, make a
13:53
whole thing. That's fine.
13:55
But if you don't have like, the tests are where you have to keep
13:57
kind of hitting. That's where the rubber hits the
13:59
pavement. That's true.
14:00
That's right. What have you been seeing when
14:03
it comes to teams working together?
14:05
Because in our lunch and learns last week, we were talking about
14:10
how companies and specifically teams need to start or are
14:19
thinking more about like, I'm not going to say like governance
14:23
structures because that might be a little bit too strong of a
14:25
word. But are we using skills?
14:29
How are we making sure that if we are committing skills back to
14:32
it, it's not just like constantly doing skills?
14:35
When do we have a skill that we commit back versus we keep it
14:38
for ourselves? When do we use the prompts?
14:40
When do we create rules? When do we spin up multiple
14:43
agents? How do we look at like creating
14:47
a lot of parallel agents for the same task to try and pass the
14:50
test, etcetera, etcetera. There's so many questions that
14:53
are involved in how you operate now as a team.
14:57
It's a moving goal post, honestly, and no one has the
15:01
answer here. I, I think so much of what we're
15:04
what we're, we're discovering a lot.
15:06
I think as a community, everyone has like a hypothesis, but, but
15:10
it's funny, even again, I keep bringing up, you know, but this
15:13
was mentioned also in the conversation.
15:14
That was what, you know, what's becoming relevant here is like
15:16
cybernetics, like this concept of like, how do you, what does
15:18
it mean to have like feedback loops in, in, in systems?
15:21
And how do you, how do you oversee those feedback loops so
15:25
the system can use to improve? And so there's like a, there's
15:27
a, there's a broad literature around what that looks like in
15:29
practice. Again, I don't think any of that
15:31
was designed necessarily with, with the intention that one day
15:33
we were going to have software systems that that sort of do
15:36
what they're doing now. But you.
15:38
Know that's all to say that like it is kind of.
15:40
This for the first time we're, we're, we're maybe putting some
15:42
of these, these theories a little bit into practice.
15:44
So, you know, theory suggests that we don't yet know how all
15:48
the answers, but in practice, what I've seen and it different
15:51
players of questions. I mean, I think one of those
15:53
important aspects here is just teams need to disseminate
15:56
information. They need to share their
15:57
learnings, they need to share their failures and they need to
15:59
have the channels to do that. I think what I, what I've seen
16:01
with some, some teams is they are, they've gotten really good
16:05
at, hey, everyone has quad code access, everyone has cursor
16:07
access. Everyone is now has as much
16:10
tokens, as many tokens as they need to to kick off any tasks
16:13
they want. No limitations there.
16:15
And they start seeing the gains as they normally should.
16:17
And then, but they, we would expect that they would, and
16:18
they're hitting like that 20 to 30% mark or something.
16:21
Now, now in practice, what is maybe not as clear is like, hey,
16:24
how do you like centralized information now to, to, to
16:28
everyone so that we have a shared set of practices and
16:30
everyone, like everyone is as good at this.
16:33
A top 1% user in your team, you know, 'cause there's always even
16:36
if everyone has access to cloud code, there's people that
16:39
prompted fairy vanilla and there's people that are have the
16:41
most intricate setups. I'm not saying intricacy
16:43
suggests that it's the most that it's the most helpful, but it at
16:47
least suggest more thoughtfulness maybe here.
16:49
And so how do you bring like, how do you raise the bar on, on
16:52
everyone having that, that knowledge?
16:53
And so I think a lot of this is just like human process a little
16:56
bit. You go and you have the lunch
16:58
and learns, you go and you, you put, you have a shared Slack
17:00
channel. You go, when you create a, a
17:03
maybe centralized team that's responsible for instituting best
17:06
practices. I think in a lot of ways it, it
17:09
mimics what happened with, you know, Cici or the fact that it's
17:13
a lot of teams have internal dev dev success or dev UX teams that
17:18
with our sole responsibility is to make sure that CI and the
17:20
build system is well instituted and, and any developer can join
17:24
the team and they can be productive from day one, right?
17:26
There's literally teams dedicated to just that aspect,
17:28
like internal developer infrastructure.
17:30
I think the same that this will probably also fall under their
17:32
purview is like, hey, now, should we use an agents MD which
17:35
go into our agents MD? What's what suggestions you have
17:37
to a new team for agents MD. So a lot of these these like
17:40
follow on questions that I think actually may still exist with
17:43
the under the under the umbrella of responsibilities for for
17:47
those existing teams actually. Also a bit higher up the
17:50
questions are how do we enable the teams to have what they
17:56
need? But then how do we not spend out
18:00
the ass but. You know, I, I don't think this
18:02
the future is always going to be like this where I don't even
18:05
know that in the long term, this token based pricing model is
18:08
going to be exactly what what the ways these companies are
18:11
going to charge. I think that there's, there's
18:12
conceivably a reality where because like you don't, because
18:16
the tokens is, is like an implementation detail, right?
18:18
Like the fact that there's a, a token that's being generated, A
18:21
token that's being input, a token that's being output, some
18:23
maybe tokens that are cached in the middle, maybe some thinking
18:26
or reasoning tokens that are happening.
18:27
And in some of these reasoning models, that is sort of not
18:30
really, it's not made it visible to you, but it's happening.
18:33
And sometimes those reasoning model, those reasoning tokens
18:35
can actually take up quite a bit of the context because they're
18:36
doing a lot of thinking and, and sort of trying this, trying that
18:39
and kind of unwinding and trying again.
18:41
So, so, but, but at the end of the day, you, you don't really
18:45
care too much about the trajectories.
18:47
A tokens are just a proxy for the trajectory that the agent
18:49
took. And, and to an to an extent, at
18:51
the end of the day, you just care about the task being
18:52
completed. And so I think that there, there
18:54
is definitely a world where maybe these, these companies
18:58
actually start pricing based on like some, some flavor task
19:00
completion more so than just number of tokens, right?
19:04
Because you can imagine a system where you literally just while
19:06
loop something and you keep trying.
19:09
And and I mean, that's super like the Ralph loop.
19:10
I'm sure that's exactly what it's doing.
19:12
It's just it's just wild truing until it completes something
19:15
very token and efficient in a lot of ways.
19:17
And I don't think that's that people aren't going to be happy
19:20
with that for for long. But if you if you but what you
19:23
really care about is that the end success of the thing you're
19:25
doing. So maybe just charge people for
19:28
almost like a task rabbit, like you pay them to to, you know,
19:30
human, human context. You pay them to to assemble your
19:33
furniture, you pay them to do your laundry.
19:35
And that's you don't care if they like went and sat for like
19:37
an hour and, you know, huffed and puffed in the last five
19:40
minutes. Got it done in some sense,
19:41
right. So so maybe something's true
19:43
here, right? We just, we like want to see, we
19:45
have like a good system. And of course there's big
19:47
questions here like how do you define that task success?
19:49
But you define the task, you define some success criteria and
19:52
then go agent, go do something. I was thinking more, how do you,
19:55
as a company that is not a model company, have that be a
20:00
possibility? Because right now, if you don't
20:02
own the model, yeah, you're just going to be paying out the ass
20:05
to Anthropic or opening up. Yeah, that's, that's tricky
20:08
that, that I don't have a great answer.
20:09
And I know you have a lot of friends that, that run these
20:13
companies and, and they're involved in these companies.
20:14
And so, yeah, it's, it's a huge tension for them because they're
20:17
both simultaneously some of the biggest users of these, of these
20:21
of the underlying model and at the same time maybe competitors.
20:25
And they're, they're going and you know, they're, they're going
20:27
to buy it to dinners with the, with the providers, but they're
20:30
also when they try and sell the providers like, and actually we
20:33
were building our own version of that.
20:34
And then, you know, there's interesting dynamics around.
20:38
Yeah, they see that, right? Like they can, they because so
20:41
much of it is getting proxy to them in terms of the tokens and
20:43
the Proms. They also have the the the
20:45
underlying models can have a clear sense for where's the,
20:49
where are people's needs, right? In a way that that that allows
20:53
them to basically repurpose and build offerings on top of that,
20:56
right. You can think of it almost as
20:59
like, yeah, I, I mean, at least what people claimed that Amazon
21:02
did are like, oh, yeah, hey, we, we, we, you're like, you're,
21:06
you're all these third parties are in the marketplace.
21:08
But you know, hey, like, and I don't know whether this is
21:10
actually true, but but you know, then people would at least
21:13
nominally claim that they were saying, Oh, well, that implies
21:16
that now you can just RIP off based on the data and make your
21:18
own, like Amazon Basics, you know?
21:20
Which I mean, they did it with products on their marketplace,
21:23
they did it with products in AWS.
21:26
They we're pretty aggressive in that strategy.
21:29
So it makes sense. There's always going to be the
21:32
argument for open source models. I guess that's what we have to
21:38
like cross our fingers for that the open source models are going
21:41
to get good enough to play ball and be frontier enough or not
21:47
far enough behind to where that would work.
21:50
So in that world I could see the outcome based pricing.
21:54
Right, right, Yeah, right, right.
21:55
That's an interesting one. And and and that feels like one
21:58
of those. It's almost like once people's
22:01
once people's wallets are getting hit, then they're
22:03
actually fine with some of the deceptive practices that at
22:06
least now many people have have accused the Chinese companies of
22:09
doing right. The distillation of anthropic
22:11
models and things like that where where fundamentally still
22:14
completely deceptive and definitely cheating in in its
22:17
own way. But then.
22:18
They're like, yeah, but if I get like.
22:19
A Kinney model that's like now Opus 4-5 level and I don't
22:23
really. Care too much?
22:25
Yeah, then all of a sudden. Principles go out the window a
22:28
little bit like. I guess that's a chance to
22:30
happen in in general with economics is like everyone has
22:33
an opinion until everyone has is sitting on their moral high
22:35
ground. Until until they look at their
22:37
PNA. Until they.
22:38
Yeah, I don't know about that. That one.
22:40
She. I think that that principle is a
22:42
little too expensive. For me to hold, for me to.
22:45
Hold. Yeah, exactly.
22:47
Which when I list out my principles, which ones are like,
22:50
like they're good enough, but nice to have, nice to have, must
22:54
have indefinitely, like wow, we just dropped our revenue by 50%
22:57
because of that principle. Oh, man.
23:00
And you're funding your competitor.
23:03
That's the craziest part. That's about it, is that all
23:06
these folks are funding the competitors.
23:08
And So what are you going to do about that?
23:10
Like, yeah, yeah, it's tricky. Claude Code, did you see the
23:13
revenue for Claude Code? Is that like 14 billion now?
23:17
That's wild. And a year ago, is that 1
23:19
billion? I mean, it's only a year old.
23:21
Of a product. Exactly dude, that is.
23:24
It was 0 actually a year ago. I mean, like if you for like a
23:26
year, like very technically speaking, it was nothing, right?
23:29
That's. So maybe somebody will have to
23:32
Fact Check me on that. Maybe it's all of Anthropics
23:35
revenue is that 14 billion and Cloud Code makes up whatever
23:37
like 80% of that. I mean, if it's a billion,
23:40
that's a lot, right? I mean, and I don't know the
23:41
exact numbers, but, but, but I suspect probably in the handfuls
23:45
of billions at this point. I mean, yeah, their, their,
23:48
their growth has been, has been astronomical.
23:52
How do you feel about just the ever changing nature of how
23:57
folks? This week it's cloud code.
24:00
Last week it was cursor, next week it's something else and
24:05
you're jumping around to all of them and just like the tides and
24:10
the winds of change are coming. Yeah.
24:13
And you think more from the what is that experience as a
24:15
developer or what, how do I, how would I, what do you think about
24:18
that as, as the one who has to service the developers?
24:21
I think how do you look at what you need to be testing versus
24:26
not? How do you think about what
24:29
actually is valid here? Should I listen when people are
24:32
saying of a cloud code now because it's not that good?
24:35
I'm using codecs and. Yeah.
24:38
So OK, a few few thoughts here. 1 is, I think there is something
24:43
to be said about too much optimization and people micro
24:48
optimizing the hell out of their setups and, and maybe almost
24:50
enjoying the setup optimization, glorifying it more so than the
24:54
actual thing that they're trying to do right.
24:56
Like I, I love development, but I also, I think I'm compared to
25:00
many developers. I'm a lot more of a pragmatic
25:01
developer where like the most important thing to me is that I
25:03
have a toolkit that enables me to solve problems.
25:05
That's like how I've always thought about the discipline of,
25:07
of, of engineering, at least for me personally, versus some
25:10
people are like, they just love geeking out about, you know,
25:13
their vim plug insurance Once Upon a time.
25:14
That's I don't like irrelevant anymore, but Once Upon a time
25:16
that was people cared. And maybe there's still some
25:18
developers that, that are really devout in that respect, But I
25:22
think people do tend to micro optimize and, and I think it
25:24
would be, given how much FOMO there is in the industry, people
25:26
are very there. They'll switch because they hear
25:29
this is like a few percent better and some benchmark comes
25:31
out and they've won and, you know, beat the benchmark by a
25:33
few percentage points. So maybe I should have this
25:34
again. I, I think there's actually a
25:36
lot my, my, my prescription. I genuinely believe for most
25:40
developers who are feeling that FOMO like there, there's a lot
25:43
more value in just picking one of the top three and sticking
25:47
with it. And just really more exercising
25:50
the muscle around how to think about building systems with this
25:52
tool rather than, hey, does this one, how is it?
25:56
How does this one do on sweet bench versus how does and how it
25:59
is it an ID? And then the you know, so
26:01
there's like there's like a base on set of almost a commoditized
26:04
features that they all have. Pick one of the top three and
26:06
then just really spend the time to get good at it more so then,
26:10
yeah, do I have to like at the moment, 4-5 Opus 47 comes out
26:14
versus Codex 54 and you know, then I you swap, swap it around.
26:17
I don't, I don't think we're not seeing like tremendously massive
26:20
gains yet that I think it's going to meaningfully change the
26:24
scope of things you can do. Though the one caveat that I
26:27
would say here is where I, you know, so, so I'll commit to a
26:31
set of tools in my own sort of work flows, but where I think
26:35
there is flexibility that at least I'm comfortable with
26:37
switching to another one is. And you see this often is the
26:40
one there's downtime, right? So the fact that these, these
26:42
these services are moving so quickly, they're building so
26:44
quickly and as they should, they're trying to aggressively
26:46
capture market share. That means probably the systems
26:49
are like the most stable in the world.
26:50
And we see anthropic go down. We see, we see, you know, other
26:53
systems go down. Not not, you know, Anthropic is
26:55
great and that's the those are usually the moments where you
26:59
see on Reddit and subreddit, the people are are ho and topic is
27:04
down. Go to codex.
27:04
Wow, codex is actually the best thing since sliced bread.
27:06
And then codex goes down and they're like, oh, actually,
27:08
like, you know, cursor is the best thing since sliced bread.
27:11
And so that's where you see the switching and it's comical to
27:14
see this thing play out in in real time.
27:15
But it's also it's it's it's like having the fall back like
27:19
you, it's good to have some fall backs and that you're
27:21
comfortable with. And so like, I think I'm very
27:23
comfortable with Claude. I'm also quite comfortable
27:25
cursor. I'm also quite comfortable with
27:26
with Devin, you know what I mean?
27:27
And and if one of them goes down I I can go to the other one and
27:30
and my my work flows. It won't be hindered in any way.
27:33
How do you create more parallel agents?
27:38
Because this is another one that I try and do super often but
27:42
then by the time I set it up agent 1 is already done and I'm
27:46
like damn I can't have more parallel agents because I I'm
27:50
too slow to like have them fan out and do their shit.
27:53
Yeah, I think people are, people are, are are romanticizing this
27:57
a little bit too much. My in my opinion, and I think
27:59
part of it is, is well. It's because your boy Boris all
28:02
over Twitter. Saying he's like 10 of these,
28:04
Yeah. And it's like, do one of his 10
28:07
commandments is like run more agents in parallel?
28:10
And so no, whenever I talk to people, it's like, Oh yeah, we
28:13
got to be running them in parallel.
28:15
Yeah, versus awesome. And and this is super, super
28:19
cool guy and then super smart. I think that can be true.
28:23
And for some people that works. I think it can also be true that
28:26
that I don't know, when you see these these Twitter demos, these
28:29
Twitter videos of people where that, you know, they'll show the
28:32
video of someone like panning across their three screens.
28:34
There's like 15 tessellated, tiled, clawed code instances and
28:39
they're just like going and it's become like a meme at this point
28:41
how you can just do this circular view of yourself of
28:44
like 700 agents like you're not doing like.
28:47
That's just what are they working on.
28:49
Yeah, what are they doing? It's like those long running
28:52
tasks, like I can't get anything to do those beefy tasks.
28:55
How do you have 50 different agents?
28:57
Yeah. No, I don't.
28:58
I don't buy it. And quite frankly, like I think
29:01
the most productive developers that I've seen do these things.
29:04
I you know you have what I would consider.
29:09
Maybe on the order of two like really needy tasks, right?
29:12
Ones that you have to be like very heavily involved with as a
29:14
human to like really carefully review spec docs, feature
29:19
requirements, all these things. And you know, that's where
29:22
you're like really injecting your, your own taste and your
29:24
own design principles and your own understanding architecture.
29:28
And then I think the stuff that they can be like kicked off more
29:30
is the ones where you are, they're not such meaty tasks
29:33
actually by design. You have to change a
29:35
configuration file, you have to make a small refactor, you have
29:37
to add a test here or something like that.
29:39
You know, that's if you can, like, you don't have to be like
29:41
overly detailed and prompting to just a kick it off, right.
29:43
And just like do these things. And so, you know, in, in
29:45
practice, what I, what I've seen and I think where you know, so
29:49
that's how I think you can achieve some of that scale.
29:51
And there's a few notes I want to make about that is in order
29:54
for that for to confidently kick off sub agents, you know,
29:57
multiple agents, you do have to go back to the thing we were
30:00
discussing, which is like the validations.
30:02
They like to have have a meaningful harness in place
30:05
that, that the agent can do work and then check that it's still
30:07
OK and it hasn't broken something and that it's added
30:10
enough coverage of tests and then test will pass.
30:12
That still has to exist, right? I think what I honestly
30:15
recommend for more for people is start with the one, get really
30:19
good at the ones, then add two if you feel like you've gotten
30:22
like start exercising that context switching muscle.
30:24
And then once you have one and two, you're like, well, like I'm
30:27
doing good here. And you know, there's kind of
30:28
running some tasks that that requires a lot of test, test
30:31
checking. And therefore it's taking a
30:33
little bit longer. And then you're like, well, I
30:34
have a bit of downtime and rather than doom scroll, let me
30:36
go and let me go and fix, fix another thing.
30:39
And I think that does require that way of thinking does
30:42
require more planning upfront. And that's simply the way I
30:44
think about the tasks a lot is you, you start the day by, hey,
30:47
here's like the three meaty hard tasks that I do.
30:50
And here's the five not so hard tasks that I need to do.
30:52
And then you at least always have this like running list of
30:54
things you can you can kick off to an agent when the time is
30:56
right. I can see that.
31:02
Are you always writing your own tests?
31:05
I mean, not manually like I, I'm, I'm at least what I'm
31:08
always doing is checking that the things I want the test to
31:11
check are valid and and are, are in the the and I'm thinking
31:15
thoughtfully about what are those edge cases that I want to
31:18
make? Like what are the things that I
31:19
want to start to be true? That part I am absolutely still
31:22
involved in. But yeah, when it when it comes
31:23
to actually, if I'm writing Python, if you know, making the
31:26
Pi test lines of code, absolutely not.
31:28
I'm not doing that. But I am making sure that those
31:31
those assertions and those validations are existing in some
31:33
files, some spec that I've given to the to the agent, yeah.
31:38
There there was something else on the.
31:40
I'll put this down for a SEC. Just another thing that might be
31:43
interesting to talk about is just like, you know, I don't
31:45
know what what the what the byline for this episode is going
31:47
to be, but the kind of the future like a native teams.
31:49
I think there's there's there's in the coding agent specific
31:52
things and there's also just what I think is interesting and
31:55
interesting. The size of the team or.
31:57
The well, it says the team, but also like how, how
31:59
interdisciplinary the the roles are getting yeah, like there's
32:04
8, you know, engineers getting, you know, and there's.
32:07
And I fucking product people are taking down websites all over
32:12
the place. Yeah, yeah, yeah.
32:14
Well, I don't, I don't endorse that, but but I but it was.
32:17
You said it, not me, shit. And it's not the product people,
32:21
it's the designers. It's they're not even.
32:23
No, it's. Still not saying it.
32:26
I'm just, I'm not going to, I'm not going to incriminate myself
32:28
here. That's what I'm saying.
32:29
Like you're trying to, you're trying to goad me into something
32:31
that's going to be caught on camera and then and then held
32:34
against me in a court of law and I.
32:35
Refuse to give you that that firepower.
32:38
We do love the designers and the product folks.
32:40
Actually, that's one of the I've seen and I've been reaching out
32:43
to a ton of product folks for the conference that we're doing
32:46
because they stand to benefit and they potentially are the
32:50
ones that are benefiting from this whole revolution more than
32:52
anyone. Yeah.
32:53
But I do think that like you're saying, the product engineer and
32:57
designer, that role is kind of merging into one and you're
33:01
expecting that you have depth in one of those.
33:05
But you can get by pretty easy with letting the agents to those
33:11
other parts. Yeah, yeah.
33:13
And I think it's a welcome expansion of responsibilities
33:17
across every single role, quite frankly.
33:20
And you can, I don't know if it's converging to one role per
33:22
SE, but but there's definitely some, you know, people on the
33:25
right coming to the left, people on the left come to the right.
33:26
I mean, directions really matter here.
33:28
But but the engineer, you know, we as engineers are getting
33:31
faster, faster at, at generating code, building features and
33:33
things like that. As a consequence, we are in a
33:37
good way, I think, being tasked down where we're thinking
33:39
through like product considerations.
33:41
So we also have to start asking ourselves a little bit more
33:43
like, Hey, my own, my own responsibility is not finger on
33:45
keyboard, write this function. It's it's designing, it's
33:48
architecting. And, and beyond that, it's, it's
33:50
actually like understanding more intimately, like why does this
33:53
feature acquired? So, so we're putting on more of
33:55
the product app product people, I think are also, you know, they
33:59
were always living in the, in the user stories and, and the,
34:02
and talking to the users and, and developing like a very
34:04
intuitive and intimate understanding of a feature
34:07
requirements. They also, but, but you know,
34:10
their attention has always been, I don't want to wait for the
34:12
engineer to implement this. That's always been the the like
34:15
one of the gripes of most modern engineering organizations.
34:17
How many story points? Yeah, exactly.
34:20
Like how can I get 3 resources to do this two cycles for this
34:23
particular feature? And and they, they, and that's
34:25
always been obviously a, a big source of tension, I think in
34:28
teams and it's annoying people. But now, I mean, we do this in
34:32
in my company and, and, and it's, it's, it's awesome.
34:34
Like I encourage it all the time.
34:36
We're we're, you know, our product people can put a TRS,
34:38
they still have to get a review, they still have to go through
34:40
our checks. It's not Willy nilly push to
34:42
prod, but they're meaningfully kicking off little, you know,
34:46
few, few, few line changes through factors like things that
34:50
are actually touching parts of the code base.
34:51
And and I think that's awesome, right, Because the alternative
34:55
is I have to contact switch into that, right?
34:56
I have to go in and make like a relatively not super complex
34:59
change just to unblock them and they can unblock themselves.
35:03
And so those two roles are like kind of coming toward each other
35:06
in a good way. I think the designers, to your
35:07
point also, so I think we're even seeing one thing we're even
35:11
encouraging is a lot of GTM people and you know, these,
35:14
these, these folks that are effectively salespeople are,
35:16
are, are in calls and, and they're, they're selling a
35:19
product and they need to have a little more clarity on, Hey,
35:23
what feature was shipped in last week and what features now
35:26
present? And you can actually, you know,
35:28
and, and in yesteryear that would be, hey, now go message
35:32
someone on slack and say, Hey, engineer, what's the latest?
35:36
Like, what did, what did we ship?
35:37
What's, what are the new, the things that are in the road map
35:39
or what not? It's a way easier now.
35:41
And we encourage it and have it given the tools to our own teams
35:44
to ask Claude and, and have Claude, of course, it's
35:47
connected to our code base, get the most up to date source of
35:50
truth on these things. And I think that's awesome.
35:53
So even they, you know, they're maybe not shipping code, but
35:55
they're at least using it to understand from the source of
35:58
truth the code base, what is already there and what maybe you
36:02
know, was not there a week ago. And because you can now look at
36:04
the get logs or something like that.
36:06
So it's all a bit of a convergence, which I think is
36:09
just just everyone is up leveling.
36:11
Do you buy into the idea that you're now going to have smaller
36:15
teams? I, I would, I mean, here's the
36:19
thing. I think that that is, I think
36:21
that would be good. I think that that would be like
36:25
a welcome thing for new teams, right?
36:27
Where I think it is absolutely the case where if you're
36:31
designing an engineering org or a team from scratch from day
36:33
one, then you should be very thoughtful about what was
36:39
typically the answer to the question, which is you're more
36:41
people enables you to move faster.
36:43
We always said that, but it was never true that if you doubled
36:46
an engineering team, you doubled output.
36:48
That was never the case. Why?
36:49
Why is that the case? Because fundamentally, more
36:52
people means there's now communication that has to be
36:55
things that have to be communicated across.
36:57
Law for that isn't there. I think there's a law that they
37:00
call it where it's like, yeah, N yeah, now you have N
37:03
communication or n ^2 or whatever the hell it is.
37:06
I can't remember exactly but I've seen it in memes.
37:09
I think I know you're talking about I also the name eludes me,
37:11
but it's never it's never linear in number of people, right?
37:14
That that's never the case. And that's just a human process,
37:17
human limitation problem, human communication problem.
37:19
And so I, I, it is the case that the more people you, you throw
37:23
into the mix and something more, the context gets distributed and
37:26
more the context that needs to get communicated when you're
37:27
moving from one thing to the other.
37:29
And so if, if that, if you have fewer people, more context is
37:32
concentrated. And that's, I think a net
37:34
positive because then you, if the goal is speed, then yes,
37:38
that is the way to have more speed, right?
37:39
Because now people are able to have more context if the goal
37:43
is, you know, now if the goal is team bonding and team dynamics,
37:48
which is which is like it's it's this is nothing with
37:50
productivity per se, but it has it has all the human happiness
37:53
is like then that's not the answer, right.
37:55
It people do still want to go in like imagine the future one
37:59
person billion dollar startup and imagine their off site.
38:01
It's just you by. Yourself at a dinner and having.
38:03
Dinner. You know what I mean?
38:05
Hey, you know, and they're like, hey, what like let's listen.
38:07
Every day is an offsite. Every day is an offsite.
38:10
And they're like, hey, let's let's get a rundown from our
38:11
CFO. You go and put on like a
38:12
mustache, like, Hey, what is what is our marketing person
38:15
think you want to put on a hat? You're just like, you know, this
38:19
is that's very sad. It's like that that Pablo
38:21
Escobar meme from Narcos. You know, he's like waiting,
38:25
like it's a little bit of that. So I, I don't, I don't again,
38:30
that's, that's a different effect, but not a not an
38:32
insignificant one. It's just it depends what you're
38:34
otimizing for. But if you're otimizing for
38:36
reliability. Yeah, I I think like in the
38:39
case. Of.
38:40
There I would argue that fewer is also better just because
38:44
again, on in the context is more concentrated.
38:47
You know, I've certainly run into issues in the past where
38:51
you know, you, you're some, some something goes down or
38:54
something, and in some of the code base and you just like it
38:56
takes your time to understand what is, how this was set up,
38:59
who's responsible for this? And usually even then you're not
39:02
that comfortable making the change because you're not the
39:04
one who built it and you don't know what the, what the
39:06
downstream effects are making that change.
39:08
And so I do also think that they're it, it, it does
39:11
introduce attention does like make things a bit slower versus
39:14
if I built the whole thing, I'd be, I like, I know how this is
39:17
set up. I know what, what's going to
39:18
break if I hit this thing, if I issue this command.
39:20
And so I'm a little bit more comfortable doing that.
39:23
So that's the, that's the teams. Anything else on the new
39:27
paradigm of just like the skills as a person that you need or the
39:34
skills and the team structure and the team?
39:37
Yeah, I think maybe that's the last thing I can a set of
39:40
thoughts I can offer. There is you know when we talk
39:42
about multi agent, I think one of the the set of skills that
39:44
has become increasingly more important is like is the
39:47
planning and the delegation. Planning and delegation, I think
39:49
are still undervalued and yet super crucial if you want to
39:54
institute, If you wanted to kick off multiple agents and have
39:57
them do meaningful work, then you have to be the one at least
39:59
delegating and being technical lead on their work.
40:01
And so I think what we're going to see is people earlier on in
40:05
their career developing manager skill sets, basically people
40:09
that that, you know, came on as a junior and sort of indirectly.
40:14
They are effectively managers the moment they start
40:16
instituting multiple agents, right?
40:17
The moment they're a couple of agents, they are effectively
40:20
managers. And so they've gotten there way
40:22
before maybe the typical career trajectory where you spend one
40:26
year as a software dev one, then two years of software dev 2,
40:29
then three as a senior range. And then that year six, you're
40:32
like ready to be in a manager if you start from like fresh out of
40:35
school grad. Well, and if you want to go down
40:38
that path. And if you want to go, that's
40:40
right, that's right, you could just keep going down the ICIC
40:42
entirely. Yeah.
40:43
Now there's no IC choice. Yeah, exactly.
40:46
Everyone has to be manager. As an IC you are still a
40:49
manager. Good luck, but at least you
40:51
don't have to deal with people because that's thinking you're
40:53
messy. That's and I wonder if that's I
40:55
wonder if that's the piece that is.
40:58
I mean, that's the piece that both draws people to the
41:00
management route and turns in the way from the management
41:03
route is like the the people. Like I like to be the the the
41:07
figurehead and and and have my inspire exactly lead.
41:10
But the student. What is it?
41:11
The servant leader, right? Be be the one who who who guides
41:14
it to the path forward. And then there's also be like, I
41:17
also never want to talk to human being in my life at any part of
41:19
my day. So why?
41:20
How can I not do that? And put me in a room and just
41:23
with my my terminal and my IDE. And so I think it does.
41:26
It also bifurcates a little bit there.
41:29
But yes, by design, I think now more people are getting.
41:31
I noticed with myself and the skill that I really appreciate
41:36
and want to work more on is my ability to articulate and.
41:41
Not good at that. Yeah, no.
41:45
Dude, words are hard. All right, give.
41:47
Me a break. So the, I think if I, you know,
41:52
as I'm raising my daughters, that's one thing that if we're
41:56
going to lean in on one skill, it's how do you articulate what
42:00
you're needing and what you want and what good enough is and what
42:04
success looks like, all of that stuff.
42:06
Because if you start talking to agents and you can't properly
42:12
articulate it and you can't give it that context, then it makes
42:16
it really hard. And I've noticed in my own life
42:20
when I deal with others, working with others humans, it can be
42:25
very difficult. If somebody comes and they ask
42:28
you something and it's like, hey, can you take care of this?
42:31
It's like, well, what is this? And that ambiguity is so hard to
42:36
deal with. And it causes, OK, now we got to
42:38
go back and forth. And so if you can properly
42:42
articulate things, give the right context, that is a
42:46
superpower and it's only going to get rewarded more as time
42:50
goes on. I think that's a fantastic
42:52
point, honestly. Yeah, I think that that's super,
42:54
super on the nose. It it it is a forcing function
42:56
for us to get better at words and get better at.
42:59
At yeah, I mean, so I'm fucked basically.
43:03
Well, we'll start practicing exercise that muscle.
43:05
Yeah, exactly. Duolingo.
43:08
Have you heard of it, Grammarly? No, no, I think that that is
43:12
absolutely right and and it's a good forcing function.
43:15
It's a good up level. Of us as as just a species as as
43:18
like a skill set to have in in your yeah you you you don't
43:22
speak clearly you're going to be misconstrued with people and
43:26
certainly with agents and. So you're just going to waste
43:28
time, right? And it happens already.
43:31
But I think because we're dealing with another human, we
43:35
almost take it for granted because they normally can come
43:39
back to us with the right questions to ask to understand
43:43
it. I guess with agents, they can do
43:45
like I have the, I love this skill, the brainstorming skill
43:50
that was from like the superpowers pack, and that is
43:55
fairly good at interrogating me before kicking off any tasks.
44:00
That's right. That's right.
44:01
Yeah. So yeah.
44:06
I mean, and maybe it's just that that you create a skill that
44:11
will make sure to get everything, you know, and get
44:14
all the right context. And so we don't need to be good
44:17
at articulating, but it is time. Like if you and I, if you're
44:21
good at articulating and I'm not, I need that skill to ask me
44:25
things. I'm going to spend an extra 5
44:27
minutes for every time that I wanted to get something done.
44:36
I'm a fucking masochist. We're trying to organize
44:39
conferences, dude. They suck.
44:42
Except you do them and you have that moment of bliss afterwards
44:46
and you're like, it wasn't that bad.
44:48
And then leading up to it, it's like, why did I ever think that
44:53
I wanted to do this? I've heard that.
44:55
I mean, I don't, I can't speak for this, but I've heard this
44:58
through my friends to be true of childbirth as well.
45:01
For for women at least. Again, I don't, you know, take
45:03
everything with a huge grain of salt and I'm not trying to
45:05
mansplain. I don't know any about anything
45:06
about this. Don't dudes talk about.
45:09
I, I, I only know this through, through, through my fiance and
45:13
and her friends who, who, who like, you know, sometimes will
45:16
have like really bad childbirth, right?
45:19
Like it's just, it's a. There's a hormone that gets
45:21
released. Yeah, yeah, exactly like, which
45:22
is fascinating to me. It's like, it's like this thing
45:24
that is, I mean, you know, the bleak and the bleak as it's true
45:28
is like Once Upon a time, like 1/3 of women died in childbirth
45:30
or something like that, like some, some really stark
45:31
statistic, which is crazy, right?
45:33
And and yet modern medicine has made that no longer true.
45:36
And in a lot of ways, reducing that rate is, is indicative of
45:39
like a very forward culture, like a society, which is, which
45:42
is, which is great. But but yeah, like you have, you
45:46
have friends that go through like really bad, really bad
45:49
childbirth. And it's like, you know.
45:52
C sections are crazy. I don't want, you know, again,
45:54
it's wild stuff. And then and then they're like,
45:56
hey, like like 3 months later they're like, yeah, how about
45:59
Round 2? And I like if, if I got cut open
46:02
for, I mean, I've got, I've had many surgeries and you know,
46:05
even like some that are more evasive than others.
46:07
I mean, I had a knee surgery a few months ago and even just a
46:09
knee surgery where all they did was like poke a thing in my knee
46:11
with a little camera and like twiddled around a little bit.
46:14
I'm like, dude, that I would never want, I don't want to ever
46:17
do that again. Or, or every like consider doing
46:20
that again. And even the fact that it was
46:21
just a little tiny hole in my knee was was enough for doing
46:24
like, you know. Yeah.
46:26
So basically me organizing this conference.
46:29
It's your. It's your brainchild.
46:31
Is what I'm saying. My birth.
46:36
Well, that that yes, I think that they, by the extension of
46:38
the analogy, you are in fact birthing A brainchild, which is
46:41
this conference, which is a good brainchild as all children are
46:45
generally good children. Yeah, well, yeah.
46:47
I mean, except for Dimitri. I know some shitty kids too, but
46:52
I think it's great. Like to bring it back.
46:55
That was an incredibly long analogy that I hope is not
46:57
considered as mansplaining because I don't know anything
47:02
about childbirth at all. But and we'll say that 1000
47:06
times, but but what I will say is when it comes to AI coding
47:09
and AI software development, which I have an opinion or two.
47:12
About a little more. Which I feel more qualified to
47:15
talk about. What I'll say is this that that.
47:20
There is. There's so much changing so
47:23
quickly. FOMO is a huge, huge thing.
47:26
Maybe this is an obvious take away, but but there is genuinely
47:30
a, a, a need and desire to, to formalize what's happening a
47:33
little bit and, and, and mean this can take so many different
47:36
forms from yeah, what agent to use, what model to use, what
47:39
harness to use Devon, you know, cognition cursor, Claude,
47:43
etcetera. But but also just as it's skills
47:46
is it is it, is it MCP? Is it tools, it's API, is it, is
47:50
it sub commands? I mean, sub agents like all
47:52
these things and people. It's so easy when you combine an
47:57
incredibly rapidly moving industry with a with a, a lot of
48:03
a lot of social conversation suggesting that if you're not on
48:08
it, you are going to be left behind.
48:10
You're you're cooked. Like the end of the technical
48:12
term. Yes, yes, you are cooked in the
48:13
in the strictest sense, that's a, that's a, that's a that's
48:17
basically a recipe for a lot of anxious engineers and, and I
48:20
think it's it's good to. Bring it all back a little bit.
48:23
I mean, I, I, I, I think, you know, when I teach classes and
48:27
things like this for me and I don't, I don't know the people
48:31
necessarily value this. I, I, I try to make it more of a
48:33
point And because it's not very tactical, it's a little bit
48:36
abstract. But like one of the things I
48:38
genuinely, I'm trying to get people to appreciate is like
48:41
just a breathe, you know, like breathe a little bit like, like
48:43
in a very meta sense, just like breathe like it's, I understand
48:46
things are changing. I understand things are
48:48
incredibly quickly and they will continue to move quickly, except
48:50
that is part of the process, right?
48:51
Like that, that is experimentation is going to be
48:53
part of the process. And we are developers and we're
48:55
engineers and we love building and we will, we will need to
48:58
experiment And, and you know, you're not going to lose your
49:00
job immediately and, or maybe if at all, right, if you, if you
49:03
learn how to like up level and, and up skill in the right ways.
49:06
And so it, but, but before we dive into the anxious, you know,
49:10
I think in, in moments of anxiety, you're just kind of
49:12
grasping for security, however you can.
49:15
And that might mean you, you like immediately look at a tweet
49:17
and say, should I do that? And then you try that thing and
49:19
then you know. Dude.
49:21
But you know what I love about this time is that we're all our
49:25
own little mini R&D departments. Yeah, R and DI guess is
49:30
departments is a little repetitive in that sentence.
49:33
Research and department departments and it's research
49:35
and development departments, right?
49:36
Oh, right, you're good. What research?
49:38
Department Departments. What did you think of DUS?
49:43
As I said, words are hard. We've.
49:46
Already established. That, I think that's the name of
49:48
this podcast, is Dimitra's discoveries.
49:50
Words are hard. Yeah, that's the thumbnail
49:54
words, me looking all dumb on the thumbnail.
49:58
But basically we're all out there exploring and we are
50:03
reporting back. And that's the cool thing.
50:05
That's why I wanted to have this conference is because we get to
50:08
report back. There's some people that are
50:11
getting more exposure to it because of the companies that
50:14
they run and they have seen hundreds or thousands of these
50:21
little R&D departments and they get, they're on calls everyday
50:24
where people are talking about what's working, what's not
50:26
working. And there's people like you that
50:27
are out there on the forefront. And so you get to showcase, hey,
50:33
this is a design pattern that's happening now.
50:37
And if we can bring all of those people together into a room, I
50:41
think there's potential for at least, at the very least, you
50:47
miss out on having to bang your head against the wall because
50:53
you're thinking, oh, this is the new one, or that's the new one,
50:55
or I'm trying it like this and it's not working.
50:59
But now you get to hear from other practitioners that, yeah,
51:02
well, I tried these five things. Here's the one that really was
51:04
important. Yeah, love that.
51:06
Yeah, I think that that's that's absolutely right.
51:08
I think that's why there's so much value in these kinds of
51:10
efforts and, and, you know, masochism aside, putting
51:13
yourself through this experience and it was a ton of work.
51:15
But but genuinely, I think there's so much value in these
51:17
conferences where you know, you're getting the best of
51:20
you're getting the best of the insights, you're getting the,
51:22
the, the, the summary on the summaries.
51:24
You know, you're getting the, the best of the Twitter feed in
51:26
one tweet. You know, I mean in, in the, in
51:29
the most very practical way. And so I think there's a there's
51:33
genuinely a lot of value. People should invest more in
51:35
these things And, and, and then you, you, you build your
51:37
connections as well, right where then you, the people that you
51:40
trust, that you develop more rapport, people, people that you
51:42
can keep bouncing ideas off of. I think the, this is the, the
51:46
part of like us being social creatures that I think is, is
51:50
one of those beautiful things, right?
51:51
Is that actually we, we do share information and especially
51:54
engineers, they love, like, I mean, that's why open source is
51:57
such a, such a core pillar of, of what, what we do is we love
52:00
sharing those things. We love proudly sharing what
52:03
we've built and we love, we love sharing of, of our own ideas
52:06
around these things. And so giving forums for that to
52:08
happen organically is, is, I think healthy.
52:11
And it's also definitely a a way to combat the inherent anxiety
52:16
that might happen when you're just on Twitter doom scrolling
52:19
and you don't know any of these people.
52:21
And but they're all the strangers in the grand scheme of
52:24
your life. And you will never talk to them.
52:25
But somehow they have such a huge influence on on what you're
52:28
thinking and feeling at any given moment, which I think is
52:29
just a funny, it's brutal, very funny dynamic.
52:32
Yeah, it is. Hard on the mental health.
52:33
Totally, totally. Even just coming to San
52:35
Francisco, it's hard on my fucking mental health.
52:37
I don't know how you. Live here dude.
52:39
Seeing all the. Billboards.
52:40
Ever a lot of Academy. Yeah.
52:44
You see this? This Burning Man is up on the
52:46
4th floor. No, it's a.
52:47
Burning Man is how people resolve these things.
52:48
No, it's. True.
52:49
Kind of a little bit. Right pressure valve that you
52:51
escape. Yeah, I mean, I don't think
52:54
that's the topic of this conversation, but like, it's
52:55
just because we are in a building where I discovered that
52:59
Burning Man has a floor. Turns out they are renting out a
53:02
floor of a San Francisco building.
53:04
Turns out that is a thing. People at home, so much burden
53:07
for beating the man. It is so novel that they have
53:10
that right. Like.
53:11
Yeah, it's so. It's they're paying rent
53:13
somewhere. They're paying rent.
53:14
They're, they're, they're paying taxes if they're doing it
53:17
correctly, which is very much against.
53:20
Anyway, the only reason I bring that up, outside of the fact
53:22
that it's a very funny anecdote, is is the fact.
53:24
That, yeah, I think that that is a little.
53:25
Bit how people deal with this is is you, you play you play
53:29
serious adult person most of your life 99.9% of the year and
53:34
then you go into a desert and and then don't ask you what
53:36
happened there.