Empowering Development Teams

1

00:00:00,320 --> 00:00:06,320

AI is triggering a fundamental shift in the software development process. Instead of being in

2

00:00:06,320 --> 00:00:12,080

the IDE and actually looking directly at the code yourself, you are talking to an agent or sometimes

3

00:00:12,080 --> 00:00:19,040

multiple agents doing the coding for you. What is the next generation of software teams look like? I

4

00:00:19,040 --> 00:00:25,320

talk with Camille, VP of Engineering at CoreWeave, to get her insights. There's a whole pipeline from

5

00:00:25,320 --> 00:00:31,800

idea to code to validation and testing to release, to support. Yeah.

6

00:00:31,840 --> 00:00:37,520

And I think we've all really gone deep into the generating code part of that. The rest of that

7

00:00:37,520 --> 00:00:42,480

pipeline, we're just going to hit all the bottlenecks everywhere else. Whether you had a CIO,

8

00:00:42,520 --> 00:00:48,079

VP of software development or a developer, you will find this exploration insightful and

9

00:00:48,080 --> 00:00:49,039

enlightening.

10

00:00:58,160 --> 00:01:04,329

Hello, everyone. Welcome to season two of the AI Cloud Essentials, a podcast series brought to you

11

00:01:04,330 --> 00:01:10,529

by CoreWeave. Today we are talking about the autonomous engineering frontier and I'm super

12

00:01:10,529 --> 00:01:15,969

excited to be joined by Camille. So Camille, welcome. Thank you. You know, in the past couple of

13

00:01:15,970 --> 00:01:21,849

years, we have seen. A lot of advancements. Into how the developers role and the developer's life is

14

00:01:21,849 --> 00:01:27,930

completely transformed. I would love to get your take on your personal journey and what you're

15

00:01:27,930 --> 00:01:33,410

seeing in the industry. Yeah, sure. So I mean, it's been a very interesting few years, as you

16

00:01:33,410 --> 00:01:40,169

said. You know, I think that two years ago, two and a half years ago, we were sort of

17

00:01:40,209 --> 00:01:47,050

starting to see a little bit of traction with copilot and, you know, some of the tools

18

00:01:47,050 --> 00:01:53,609

where you would be in your, in your IDE or, you know, and it would sort of suggest code for you. You

19

00:01:53,609 --> 00:01:58,569

could, you know, you could get a little bit more than your normal autocomplete kind of activity

20

00:01:58,609 --> 00:02:03,809

from the AI. And that was definitely a benefit. And lots of teams use that. Lots of teams have

21

00:02:03,809 --> 00:02:10,809

been using that. But things started to really change about, you know, last November ish

22

00:02:10,809 --> 00:02:15,609

and certainly even really early this year. Right. Just very quickly, as more and more people have

23

00:02:15,609 --> 00:02:22,329

realized the power of that sort of a genetic coding where instead of being in the IDE and

24

00:02:22,369 --> 00:02:27,969

actually looking directly at the code yourself, you are talking to an agent or sometimes multiple

25

00:02:27,969 --> 00:02:33,609

agents doing the coding for you, and you are giving them various types of instructions in

26

00:02:33,609 --> 00:02:40,569

order to actually generate the code without you really necessarily looking at the code in context

27

00:02:40,570 --> 00:02:47,289

yourself. And that is a huge change that I think is really rocking the tech industry right now. Um,

28

00:02:47,289 --> 00:02:52,009

you know, as people try to figure out how in the world do I adapt to this? What does this mean? What

29

00:02:52,009 --> 00:02:57,369

does this mean for me as a developer? Uh, you know, what does this mean for me as a product manager?

30

00:02:57,369 --> 00:03:01,289

What does this mean for me as a marketing person? You know, here, here at CoreWeave, we're actually

31

00:03:01,289 --> 00:03:08,179

doing a hack week this week. Um, and it is amazing to see. I so I, I

32

00:03:08,220 --> 00:03:13,059

happen to, you know, lead the developer experience team here. So I have sort of access to the stats

33

00:03:13,059 --> 00:03:18,579

of what kind, you know, people using these AI coding tools. And it is remarkable how many people

34

00:03:18,580 --> 00:03:24,459

not in engineering are using them to play around, to do things for, you know, for their job in

35

00:03:24,460 --> 00:03:31,100

marketing or revenue or, you know, HR recruiting, right? People are really, you know,

36

00:03:31,139 --> 00:03:35,739

non coders are definitely getting into the game. And it's definitely a very interesting time. Yeah,

37

00:03:35,779 --> 00:03:40,659

I can completely relate with it because I started my career as a developer and I remember, you know,

38

00:03:40,659 --> 00:03:46,859

all the boilerplate text and all the conundrums that we had to do. So GitHub Copilot itself was

39

00:03:46,860 --> 00:03:51,939

interesting, but now, you know, looking into all the agents and now started my own company and I have

40

00:03:51,939 --> 00:03:56,099

moved, gotten back to my engineering roots and doing the development. And you are so spot on. You

41

00:03:56,099 --> 00:04:01,820

know, like in terms of talking about how you could actually see an agent do the things for you, and

42

00:04:01,820 --> 00:04:07,100

I'm experiencing it firsthand. So I'm super excited to talk to you further on this. How do you

43

00:04:07,100 --> 00:04:13,179

see the use cases and scenarios that are being awarded? Right? As you said, people were using it

44

00:04:13,179 --> 00:04:19,539

first just for code completion, but now helping it reason and plan and write the entire code. So

45

00:04:19,820 --> 00:04:24,299

could you talk about a couple of use cases in scenarios where the internal or external that

46

00:04:24,299 --> 00:04:30,940

you're seeing? Yeah, I mean, the most I would say that still in, you know, most

47

00:04:30,940 --> 00:04:34,659

companies that maybe aren't like, I don't know, Anthropic where they say they're not writing any

48

00:04:34,659 --> 00:04:40,099

code at all by hand, which, you know, may very well be true for them, but I think most companies that

49

00:04:40,099 --> 00:04:46,499

are still trying to figure out how to adopt this. Uh, people definitely see a huge amount of value

50

00:04:46,499 --> 00:04:52,179

in using it, both for like, kind of new things. So Greenfield, as we call it. Right. And engineering. So

51

00:04:52,179 --> 00:04:58,179

you, you have a new idea of a new brand new project starting from that and letting the agent

52

00:04:58,179 --> 00:05:04,859

really do it, do all of the all of the, like, work itself from the jump, or in a lot of cases, people

53

00:05:04,859 --> 00:05:11,510

are also using it for kind of like the side work that you might not normally have bothered to

54

00:05:11,549 --> 00:05:16,790

do, but that's so easy to get generated from the agents that now it's like actually like, you know,

55

00:05:16,869 --> 00:05:23,829

I can make this like weird sort of testing setup that, you know, because Claude can just kind of

56

00:05:23,870 --> 00:05:27,989

write it for me. Or Cursor can kind of generate it for me or whatever, you know, whatever I'm using.

57

00:05:28,269 --> 00:05:33,349

Um, and I don't have to remember how to use that particular library and this other weird, weird

58

00:05:33,390 --> 00:05:38,630

tool. Uh, and it's okay if it isn't perfect, because it's not like something I'm putting into

59

00:05:38,630 --> 00:05:43,829

production. It's just a a tool to help me do the parts of my job faster. Yeah. So I feel like we see

60

00:05:43,829 --> 00:05:50,069

a lot of that adoption. Um, and then a lot of greenfield adoption. And slowly people are trying

61

00:05:50,070 --> 00:05:55,070

to work it into how do I actually use this for my preexisting code base? So that's still, I think the

62

00:05:55,070 --> 00:06:00,189

area where everyone's doing a lot more work to figure it out. I see a lot of people doing some

63

00:06:00,190 --> 00:06:06,030

migration of their legacy code into the new code. I know cursor is one of your customers. I've been

64

00:06:06,030 --> 00:06:11,589

reading about them. And I also seen that people are using it, as you said, for the greenfield

65

00:06:11,589 --> 00:06:17,510

opportunities. Like, I love to use it for doing some mock ups, right? If I have to do some mock ups

66

00:06:17,510 --> 00:06:22,709

to get some inspiration, and then I might write the whole code myself. So internally, how are you

67

00:06:22,709 --> 00:06:27,789

folks using it at code V? How is your own developers and your own internal customers? It is

68

00:06:27,829 --> 00:06:32,590

a, you know, a breadth of of usages, right? So definitely lots of greenfield. Right? We have a lot

69

00:06:32,590 --> 00:06:39,589

of new initiatives that, you know, have kicked off very, very quickly because, you know, the people,

70

00:06:39,630 --> 00:06:42,750

you know, the engineers working on those initiatives are very comfortable working in this

71

00:06:42,750 --> 00:06:46,950

multi-agent environment and throwing a bunch of things at it. One of the most interesting things

72

00:06:46,950 --> 00:06:53,829

here, so our CTO is the biggest spender on AI tools here. Yeah. Um, which and I actually

73

00:06:53,829 --> 00:06:57,429

discovered this yesterday and I asked him, hey, Peter, like, what are you what are you doing? And

74

00:06:57,429 --> 00:07:04,149

he's like, well, you know, I am. He's using a huge amount for operational automation,

75

00:07:04,190 --> 00:07:09,439

understanding what's going on, making better and better dashboards and correlations. You know,

76

00:07:09,480 --> 00:07:15,119

because a lot of what we do here at CoreWeave, we've obviously is we support these massively complex

77

00:07:15,160 --> 00:07:21,479

AI data centers. And, you know, this very complex AI cloud, right. And a lot of that work is, you know,

78

00:07:21,519 --> 00:07:27,520

having really deep awareness of what's going on and sort of why something is failing and in

79

00:07:27,520 --> 00:07:32,159

detecting it before our customers detected so that we can get on it and fix it and make sure it

80

00:07:32,160 --> 00:07:35,679

doesn't, you know, doesn't affect their training workloads or their their inference workloads.

81

00:07:35,680 --> 00:07:42,479

Right. Um, and, you know, Peter has always been incredibly hands on when it comes to, you know,

82

00:07:42,520 --> 00:07:46,879

detecting and fixing problems. Frankly, throughout our customers. I mean, anybody who's a

83

00:07:46,880 --> 00:07:53,358

customer, of course, knows knows Peter. And so he's just making incredibly valuable use of these

84

00:07:53,359 --> 00:08:00,079

tools to help him do that part of his job and, you know, to get more sleep. But, you know, to to help

85

00:08:00,079 --> 00:08:04,479

all the rest of us who do those kinds of activities from time to time or all the time. Uh,

86

00:08:04,480 --> 00:08:08,999

you know, also answer those questions. And so I think that's a really, you know, that is really

87

00:08:09,040 --> 00:08:12,679

interesting to me is where I'm seeing it, not for the thing that everybody thinks about, which is

88

00:08:12,679 --> 00:08:17,639

like writing new code and building new features, but actually, like, you know, just like thinking

89

00:08:17,640 --> 00:08:23,840

about your job and how you gather information and how you answer questions and how these tools can

90

00:08:23,879 --> 00:08:29,159

build automation and build dashboards and build support for you to do that. Even better. Yeah, it's.

91

00:08:29,159 --> 00:08:33,679

Such an interesting example, as you said, right? Because if you start from what your pain points

92

00:08:33,679 --> 00:08:38,479

are currently, not just the greenfield, which is exciting and how you could do it better and how

93

00:08:38,479 --> 00:08:44,039

you could actually enhance it, uh, where do you see where we are heading in terms in the future? I

94

00:08:44,039 --> 00:08:50,479

mean, the pace of innovation has been out, you know, outright insane, right? I mean, so where do you see

95

00:08:50,479 --> 00:08:55,599

what do you predict for 2026? The later half of 2026? You know, people love to ask me for

96

00:08:55,599 --> 00:09:00,918

predictions. And I always tell them I'm like, I'm not a futurist. The pace of change has

97

00:09:00,919 --> 00:09:07,679

changed is is very rapid. And I think one of the reasons that the overall world as a whole has

98

00:09:07,680 --> 00:09:12,329

still sort of struggled to make the most use of AI is because it keeps changing so quickly, and

99

00:09:12,330 --> 00:09:18,450

it's really hard to figure out how to adopt something that a month later, you have to

100

00:09:18,490 --> 00:09:24,368

completely change the way you're using it to get the most value out of it. And so I think that we

101

00:09:24,369 --> 00:09:30,089

are not yet done with that particular treadmill. Um, but I also look, I definitely think no matter

102

00:09:30,090 --> 00:09:36,769

what, everybody is seeing the value in these coding agents, it's just it's undeniable. Not I'm

103

00:09:36,769 --> 00:09:41,770

not saying that it solves every coding problem. I know plenty of people who regularly apply it to

104

00:09:41,770 --> 00:09:46,649

their code bases, and they're like, it's just not ready for this complexity, this gigantic, super

105

00:09:46,650 --> 00:09:52,569

complex, odd language code base, what have you. Right. But, you know, for a lot of things that

106

00:09:52,570 --> 00:09:56,408

people do, they get they are, you know, in the engineering space, I think they are figuring out

107

00:09:56,409 --> 00:10:02,530

how to get value from it. Um, now, as a person who, among other things, runs a developer experience

108

00:10:02,530 --> 00:10:07,288

team, one of the challenges is, of course, writing code is not the only thing that we do, but there's

109

00:10:07,289 --> 00:10:13,699

a whole pipeline from idea to code to validation and testing to, you know,

110

00:10:13,739 --> 00:10:20,259

release to support. And I think we've all really gone deep into the generating code part of that.

111

00:10:20,380 --> 00:10:25,299

The rest of that pipeline, you know, we're just going to hit all the bottlenecks everywhere else,

112

00:10:25,340 --> 00:10:29,539

right? Um, you know, a lot of people are talking about code review is a big bottleneck right now,

113

00:10:29,580 --> 00:10:35,820

right? So when you've got, you know, when you spend, you know, so I was talking to a friend who almost

114

00:10:35,820 --> 00:10:40,179

entirely just writes code through agents. And I was like, you know, what is your okay. So if you

115

00:10:40,179 --> 00:10:44,659

never, barely ever touch the code yourself, like, what does it look like? He's like, well, you know, we

116

00:10:44,659 --> 00:10:51,579

spend whatever like a month, like banging away on this design and like, you know, sort of proposal

117

00:10:51,580 --> 00:10:55,499

spec that we want to do. So then me and other engineers are really talking about what do we

118

00:10:55,499 --> 00:11:00,820

want to do? Okay. Then we give it to the agent and whatever, you know, breakdown they have. And the

119

00:11:00,820 --> 00:11:06,339

agent spends like 30 hours or however long chugging away at it. And then, you know, I was like,

120

00:11:06,379 --> 00:11:12,139

and now I've got a stack of pull requests and I'm babysitting those through kind of the review and

121

00:11:12,140 --> 00:11:17,419

validation process. And, you know, on the one hand, he's like, and it's like, this is an this is

122

00:11:17,419 --> 00:11:21,900

interesting for me because like, it's like the as an engineer, some of the fun part of being an

123

00:11:21,900 --> 00:11:25,780

engineer is actually writing the code. And there's none of that in there. Right? It's all the kind of

124

00:11:25,820 --> 00:11:31,099

like, uh, you know, dealing with your coworkers, you know, trying to figure out the design and then

125

00:11:31,139 --> 00:11:35,538

dealing with your coworkers, trying to get the pull requests approved. But, you know, it is

126

00:11:35,539 --> 00:11:40,779

interesting to see, like, okay, if if he's in the situation where he's really managed to get away

127

00:11:40,780 --> 00:11:46,378

from that writing of the code entirely now it's like, okay, well, we've got the rest of these steps.

128

00:11:46,419 --> 00:11:52,140

What does that look like? You know, and this is just a first pass of optimization. So okay, you

129

00:11:52,140 --> 00:11:57,459

know, now that now that we've got, you know, the the agent writing all these stack pull requests, you

130

00:11:57,460 --> 00:12:04,178

know, can we use the agent to actually help with the pull requests. What are we reviewing really.

131

00:12:04,259 --> 00:12:10,139

Are we really expecting the engineers to really review the code that they didn't write, or are we

132

00:12:10,140 --> 00:12:16,069

expecting them to review the spec and review the testing and review the, you know, the support plan.

133

00:12:16,109 --> 00:12:21,909

Right? I think these are all questions that we are going to be answering and trying and failing at

134

00:12:21,909 --> 00:12:26,709

and trying and succeeding at in both ways throughout. Certainly the rest of the year. I would

135

00:12:26,710 --> 00:12:30,829

guess the next several years, to be honest with you, as things change, we're going to continuously

136

00:12:30,830 --> 00:12:36,589

have to be changing our workflows and really identifying those bottlenecks and optimizing for

137

00:12:36,590 --> 00:12:42,309

them. What are the early lessons learned based on all the code that has been written by an agent?

138

00:12:42,390 --> 00:12:47,469

You know, the the concerns and of course, people, you know, when you're using it in your own

139

00:12:47,510 --> 00:12:52,788

internal setup, you might be able to kind of provide it to your existing code, your existing

140

00:12:52,789 --> 00:12:58,548

style of writing and all. So what are the lessons and takeaways from that aspect? And as you said,

141

00:12:58,590 --> 00:13:03,869

you know, code review the code that has not been written by them. But at the same time, if it has

142

00:13:03,870 --> 00:13:09,589

picked up some of their nuances and best practices and how we can actually do it better, I

143

00:13:09,590 --> 00:13:14,309

do. I think that's, you know, so much of the concerns and challenges and issues right now are

144

00:13:14,309 --> 00:13:18,430

really just about like the humans and how they interact with one another with regards to this

145

00:13:18,430 --> 00:13:25,030

stuff. So like, yeah, definitely everyone I know has had the experience of someone on the team went a

146

00:13:25,030 --> 00:13:30,989

little too wild, generated a whole bunch of code that they themselves didn't really

147

00:13:31,349 --> 00:13:36,869

review very carefully, and then threw it over the fence to their colleagues to say, review this. And

148

00:13:36,869 --> 00:13:42,749

their colleagues are like, I'm looking at this and it's clearly sloppy or it's not. You know, it seems

149

00:13:42,750 --> 00:13:46,589

like it makes sense. But then I have to like, peel it apart. And I'm just like, wait, what? What does

150

00:13:46,590 --> 00:13:52,549

this do? What does that do? And I feel like you're giving me a lot of work, uh, because you didn't

151

00:13:52,549 --> 00:13:57,189

really think about what was produced that carefully. You were just like, I produced a bunch,

152

00:13:57,229 --> 00:14:03,629

and you guys go now. Have to have to read it. I think that is like, that is a

153

00:14:03,630 --> 00:14:09,790

lesson that it's interesting. I think we've all learned it, but not all of us know how to approach

154

00:14:09,790 --> 00:14:13,989

it. You know, I have friends that are talking about, you know, they have, uh, early career developers

155

00:14:14,030 --> 00:14:19,559

right on their teams. and they like, look. Some of them are very, you know, are very careful and

156

00:14:19,560 --> 00:14:23,840

they're really trying to do the work and they're not even always generating. They're actually

157

00:14:23,840 --> 00:14:27,798

sometimes writing it by hand to really make sure they understand. And that's that's going fine

158

00:14:27,840 --> 00:14:31,839

right there. They're doing great. They're learning. They're very, very productive. And then some of

159

00:14:31,840 --> 00:14:36,879

them like are very clearly using a lot of AI. And that's fine, except that when we come in and give

160

00:14:36,879 --> 00:14:41,359

them feedback about like this isn't that good and you need to redo it, they're like, why should I do

161

00:14:41,359 --> 00:14:46,120

that? Like the AI is just going to support it. The AI is just going to write the next one. And you

162

00:14:46,120 --> 00:14:51,359

know, nobody's quite comfortable with that yet. Is my is my feeling right? Like, you know, I think that,

163

00:14:51,399 --> 00:14:55,879

you know, in reality, especially when it comes to more early career engineers who are trying to

164

00:14:55,879 --> 00:15:00,640

make the most use out of this, as they should. Yeah. Um, I think, you know, people who don't want to

165

00:15:00,640 --> 00:15:06,040

listen to their colleagues saying, hey, like, no, we can't. We especially with certain types of systems.

166

00:15:06,040 --> 00:15:09,439

And I'm working on a database, I'm working on an operating system, working on something that's

167

00:15:09,440 --> 00:15:15,399

really foundational to the way a lot of people work. You know, we actually have to have higher

168

00:15:15,399 --> 00:15:20,199

standards for what good looks like. This isn't a toy web app that you know three people are going

169

00:15:20,199 --> 00:15:25,199

to use, and it doesn't really matter, right? Um, so I think that a lot of it really is just still the

170

00:15:25,199 --> 00:15:31,599

interpersonal. It's the how do teams work together once these content is generated and, you know, make

171

00:15:31,640 --> 00:15:37,879

sure that they're not setting themselves up for later challenges and failures. Great, great. I think,

172

00:15:37,879 --> 00:15:43,879

uh, as you were talking about this whole, you know, context, I was thinking about it. What we've seen

173

00:15:43,879 --> 00:15:48,559

so far, in my humble opinion, is that it's very good in sometimes, you know, creating some one

174

00:15:48,559 --> 00:15:53,919

piece of module or something, but it's not very good in a holistic, systematic design or a whole

175

00:15:53,920 --> 00:15:58,879

system design. So I think a lot of advancements need to be there as well. But this is fantastic

176

00:15:58,880 --> 00:16:04,480

conversation, Camila. I am sure our viewers learned a lot. What will be your parting advice to anyone

177

00:16:04,480 --> 00:16:11,479

who's listening to us today? My parting advice is always, you know, keep, keep learning

178

00:16:11,479 --> 00:16:18,209

and stay open minded and accept that you are going to do some not quite

179

00:16:18,209 --> 00:16:23,928

throw away work, but you're going to do some like iterations right now, right? If you you can't just

180

00:16:23,969 --> 00:16:28,729

ignore what's going on and wait for it to settle, because it's going to take a while to settle. So

181

00:16:28,729 --> 00:16:33,889

you do need to be keeping up and learning and trying to figure it out and just be comfortable

182

00:16:33,889 --> 00:16:38,449

with the fact that, like the way you approach getting value from these tools today, it may

183

00:16:38,449 --> 00:16:42,369

change in a month and that will be a little bit annoying because you're going to have to like,

184

00:16:42,409 --> 00:16:48,850

change it up and to meet to meet the new, you know, the new styles. But you know, this is also kind of

185

00:16:48,890 --> 00:16:55,609

exciting, right? And I think the people, everyone I know who is, um, you know, open

186

00:16:55,609 --> 00:17:02,129

minded, you don't have to be open minded and, and like, totally like, but in that it is the way the

187

00:17:02,129 --> 00:17:07,368

world will be for all things. Right. You can be open minded and skeptical, but open mindedness is

188

00:17:07,369 --> 00:17:13,728

so important right now when it comes to learning how to use these tools, making the best, getting

189

00:17:13,729 --> 00:17:18,529

the best out of them, you know, take the best and leave the rest. But you've got to keep looking

190

00:17:18,530 --> 00:17:23,409

regularly because they change so quickly that something that they couldn't do three months ago,

191

00:17:23,449 --> 00:17:30,289

they may be able to do for you today. And you know that hopefully over time will allow you

192

00:17:30,290 --> 00:17:35,449

to focus more on the things that are fun and get rid of some of the grunt work and tedious work

193

00:17:35,449 --> 00:17:41,369

for yourself. And you know, I am cautiously optimistic that that is the direction we can we

194

00:17:41,369 --> 00:17:46,889

can make this all go if we if we really put our minds to it. Awesome. Very nice. Nice. Couldn't agree

195

00:17:46,890 --> 00:17:53,009

more. One cannot sit on the fence, but I have to go with the flow. But also be cautious and be, you

196

00:17:53,009 --> 00:17:57,969

know, pragmatic about what really works and what doesn't work. Thank you so much, Camille. It was

197

00:17:57,969 --> 00:18:03,009

such a pleasure to have you with us in this episode. And thank you, everyone, all our viewers

198

00:18:03,010 --> 00:18:08,208

who are listening to us, and stay tuned. We are coming up with more episodes. Thank you.

‍