The Digital Handyman: Claude Code vs. Codex, and the Rise of Bespoke Software

Ajay: 00:00:00

Hello, my name is Ajay Medury, and I'm a software engineer, and today I'm joined by

2

: 00:00:05

Ajay: my co-host.

3

: 00:00:07

Andrew: Hello, my name is Andrew Sierota, and I am a systems engineer.

4

: 00:00:11

Andrew: And we are hosting our podcast called Act of Intelligence.

5

: 00:00:15

Andrew: Which is about our experiences using Codex, Claude Code, Gemini, and other tools in

6

: 00:00:20

Andrew: the space.

7

: 00:00:21

Ajay: We are trying to answer the question of whether they are actually intelligent or

8

: 00:00:25

Ajay: just pretending.

9

: 00:00:27

Andrew: and we are a podcast for creatives creators and builders or anyone with an

10

: 00:00:32

Ajay: interest in the space nice there we go there we go I mean and now I think uh we

11

: 00:00:38

Ajay: just continue from here and I open up a different notes file and uh

12

: 00:00:43

Ajay: luckily luckily this time our notes are pretty

13

: 00:00:48

Ajay: straightforward yes and and I think uh we are here to have the conversation about

14

: 00:00:52

Ajay: the things we've learned um instead of uh follow the notes like I might have done a

15

: 00:00:58

Ajay: little too faithfully last time there we go a little more natural a little a little

16

: 00:01:03

Ajay: more yeah let it just let's just have the conversation and uh so talking

17

: 00:01:08

Ajay: about notes there's no good transition here so I'll just I'll just say the first

18

: 00:01:14

Ajay: thing on on our list is is there I'm doing it live we're doing it live so the the

19

: 00:01:19

Ajay: first thing on our list is something that I recently experienced because um maybe I

20

: 00:01:24

Ajay: will find actually some way to transition into it because in the last episode we

21

: 00:01:28

Ajay: talked a lot about software engineering life cycle software processes versus

22

: 00:01:34

Ajay: projects and I think in particular there's this whole thing of like hey what

23

: 00:01:40

Ajay: does a process look like with the new tools Claude Code

24

: 00:01:45

Ajay: Gemini.

25

: 00:01:47

Ajay: And in this case, a specific tool that I had in mind was an LLM wiki, which is not

26

: 00:01:53

Ajay: actually something that's vended by any particular company.

27

: 00:01:56

Ajay: The LLM wiki is more of a concept that Andre Karpathy, the gentleman, the vibe

28

: 00:02:02

Ajay: coder, the original vibe coder himself that coined the term vibe coding, had

29

: 00:02:08

Ajay: a GitHub gist talking about the concept of an LLM wiki and the kind of tools and

30

: 00:02:14

Ajay: things that would be useful uh to actually build one of your own which is

31

: 00:02:19

Ajay: interesting because the concept there if you just give that gist to Claude Code

32

: 00:02:24

Ajay: gemini codex whoever they'll actually intrinsically say okay I'll build the tools

33

: 00:02:29

Ajay: then and they'll go ahead and build them for you um the goal of the wiki itself

34

: 00:02:33

Ajay: though I think is where I want to get into it is like you have

35

: 00:02:38

Ajay: Codex, Claude Code, and all these tools, they're writing the code for you.

36

: 00:02:44

Andrew: Yeah.

37

: 00:02:44

Ajay: But at times, they may not exactly know all the details that you have in your head,

38

: 00:02:50

Ajay: you know, in your brain, in a sense.

39

: 00:02:53

Ajay: And the idea of the LLWiki is to kind of be almost a second brain.

40

: 00:02:58

Ajay: Basically contain all this information that you're thinking about, all the things

41

: 00:03:02

Ajay: that you researched ahead of time, Things that are not really code and probably

42

: 00:03:06

Ajay: maybe you will never put in code or maybe you will put it in your code repository.

43

: 00:03:10

Ajay: But the idea is that it is essentially a second brain for all the things that you've

44

: 00:03:14

Ajay: been researching, learning, building towards.

45

: 00:03:18

Ajay: And then and that's at least the way that I've used it is essentially a way to give

46

: 00:03:24

Ajay: a scratch pad with enough information that the coding tools can use prior to

47

: 00:03:30

Ajay: actually writing code.

48

: 00:03:32

Andrew: Like a source beyond the context window.

49

: 00:03:35

Ajay: Yes, exactly.

50

: 00:03:36

Ajay: Yeah, that's a good way of putting it.

51

: 00:03:38

Ajay: It's not memory because memory is also something that's directly attached to your

52

: 00:03:42

Ajay: coding assistant.

53

: 00:03:44

Ajay: It is more of a kind of a combination of markdown files that is

54

: 00:03:49

Ajay: representing a number of links that you've researched in the past.

55

: 00:03:55

Ajay: And they all come together in like an index file, which is one of the requirements

56

: 00:03:58

Ajay: of the LMWiki.

57

: 00:03:59

Ajay: they also have a log depicting all the changes that you've done all the new things

58

: 00:04:03

Ajay: that you've ingested you know that's again like the concept of ingesting is you have

59

: 00:04:07

Ajay: a raw set of information that you can give it you can ask your coding assistant

60

: 00:04:10

Ajay: because your coding assistant is also managing your alumbra key that is important

61

: 00:04:15

Ajay: part and it's creating the markdowns for you and the top process there is that it

62

: 00:04:20

Ajay: will actually focus on certain types of documents so create markdown files

63

: 00:04:25

Ajay: that are like concepts it will then create markdown files that are like synthesis so

64

: 00:04:31

Ajay: basically if you have a a concept could be very specifically like my

65

: 00:04:36

Ajay: home os which I'll talk a little more about is one concept may be the idea of a

66

: 00:04:42

Ajay: home finance assistant so that's a concept because that's something that I'm now

67

: 00:04:48

Ajay: building into its own thing but then the synthesis is more like all right what

68

: 00:04:53

Ajay: information sources did I actually use and what is the logic behind

69

: 00:04:59

Ajay: the business logic behind the concept so synthesis may be like oh I did a bunch of

70

: 00:05:04

Ajay: research now I'm going to write a document explaining uh what how to really go

71

: 00:05:09

Ajay: from concept to real product kind of idea so like what is the logic what are all the

72

: 00:05:15

Ajay: decisions that are important for this thing to just go from a concept to a real

73

: 00:05:19

Andrew: product with with the lm wiki like hold the spec sheets and stuff or would you build

74

: 00:05:24

Ajay: from that so so that's the that's the interesting thing is the spec sheets are a

75

: 00:05:29

Ajay: little more uh engineering focused the lm wiki is actually intended more conceptual

76

: 00:05:34

Andrew: yes yes I really like that characterization yes yeah okay okay it's it's kind of

77

: 00:05:39

Andrew: like place to put all your little bubbles that pop up in your head that is that oh I

78

: 00:05:44

Ajay: love that exactly yeah that's the intention of uh that is definitely a good way to

79

: 00:05:49

Ajay: think about it because there may be a lot of things that you will never formalize in

80

: 00:05:54

Ajay: my opinion uh that are just like high level thoughts and just high level like ideas

81

: 00:05:58

Ajay: uh the bubbles in your head yeah and it's funny in my projects right now that folder

82

: 00:06:03

Ajay: is called runoff all right I'll be that's that's an interesting name andrew

83

: 00:06:09

Andrew: I'm curious why well it's like uh it's something that it comes up once you

84

: 00:06:16

Andrew: want to notate it put something down and you know it's probably not going to be

85

: 00:06:20

Ajay: relevant beyond that okay I I definitely like the idea and one thing uh I would

86

: 00:06:25

Ajay: imagine that being a raw file that you would give the wiki but then the wiki would

87

: 00:06:31

Ajay: actually ingest it to take all the linked things with it and then link it up with

88

: 00:06:35

Ajay: everything else you've already talked about in the past yeah what's relevant right

89

: 00:06:39

Andrew: what's relevant out of the thought it's it's like some it's a place where where I

90

: 00:06:44

Andrew: you know I I midstream it's it's working like like Claude Code or codex it's

91

: 00:06:49

Andrew: coding yes and I think of something I'm like oh my gosh this would be a great

92

: 00:06:54

Andrew: optimization for this specific issue that we ran into I open up another agent and I

93

: 00:06:59

Andrew: say you know let's let's take note of this put into runoff that's just like a quick

94

: 00:07:04

Andrew: note taking place for me so that we can formalize it later and I think that's

95

: 00:07:09

Ajay: exactly the use case that the elm wiki is trying to provide is like hey I just came

96

: 00:07:13

Ajay: up with something I just observed something really quickly I read an article that

97

: 00:07:17

Ajay: actually was very relevant that I didn't even think would be relevant um let me just

98

: 00:07:21

Ajay: go ahead and add it to my wiki that is my concept wiki elm wiki is the idea of like

99

: 00:07:25

Ajay: various concepts at least the way that I use it I'm sure other users have different

100

: 00:07:30

Ajay: ideas as to how it should be used and how it is used um but yes I think the only

101

: 00:07:34

Ajay: difference is I'm taking the raw files now I'm following the methodology that andre

102

: 00:07:40

Ajay: karpati set out and creating a kind of a more formal but simplified representation

103

: 00:07:46

Ajay: of it okay and that is where you end up with like the concepts idea the synthesis

104

: 00:07:50

Ajay: idea um I think there was one more name that I'm forgetting right now but there's

105

: 00:07:54

Ajay: like a few of those uh top level uh high level ideas in fact if you want to get more

106

: 00:07:59

Ajay: low level the recommendation within the methodologies actually to create sub pages

107

: 00:08:04

Ajay: so always be high level so that anybody can read it but if you want to be really

108

: 00:08:09

Ajay: specific and get into real details you can do it you just got to link them as a like

109

: 00:08:13

Ajay: a hierarchical set of information saying that high level and then mid-level and then

110

: 00:08:18

Ajay: low level kind of and you don't do it the elem or the wiki manager assistant will do

111

: 00:08:23

Ajay: it for you and and in all these cases it's basically again I'm using clod you could

112

: 00:08:28

Andrew: probably use uh gemini or codex so what when you're like trying to build something

113

: 00:08:33

Andrew: from the concepts in the wiki so how how do you direct um codex Claude Code to like

114

: 00:08:39

Andrew: I'm just thinking in my head like the wiki is going to have a lot of files at some

115

: 00:08:44

Andrew: point yeah right it's it's not going to be able to ingest everything that would be

116

: 00:08:47

Andrew: inefficient you don't want to spend the tokens yeah like how would you normally

117

: 00:08:52

Andrew: direct it or kind of like let it hone in on what you need for your specific task

118

: 00:08:58

Ajay: That's a really great question because you're right.

119

: 00:09:00

Ajay: I'm already reached, I think, about 200 files, probably.

120

: 00:09:04

Ajay: That's a lot.

121

: 00:09:05

Ajay: Yeah, it's getting up there.

122

: 00:09:07

Ajay: And then there's a lot of links between them.

123

: 00:09:10

Ajay: They're not all interlinked, which at some point I think they will potentially get

124

: 00:09:13

Ajay: there.

125

: 00:09:14

Ajay: There's a concept within the methodology.

126

: 00:09:17

Ajay: There's a concept called query.

127

: 00:09:19

Ajay: So the query will use the wiki style links.

128

: 00:09:24

Ajay: So again, it's marked down.

129

: 00:09:25

Ajay: one more thing I forgot to mention is I typically use obsidian to host all my

130

: 00:09:28

Ajay: markdowns uh you could do it in any tool it doesn't have to be obsidian of course

131

: 00:09:33

Ajay: the idea is that when your index page gets created your index page essentially has

132

: 00:09:38

Ajay: um these links I'm forgetting what the wiki link style thing was like within my

133

: 00:09:43

Ajay: relative folder path uh this is the concept and this is the actual location of the

134

: 00:09:48

Ajay: actual markdown file within my relative folder path the um I benefit of that is

135

: 00:09:54

Ajay: saying that now I have progressive disclosure saying that now I have the super high

136

: 00:09:59

Ajay: level index giving me just a single line understanding of every single thing so my

137

: 00:10:03

Ajay: LLM can actually do a good job of just being like oh yeah that sounds like it's

138

: 00:10:07

Ajay: relevant let me actually now dig into it when I call query and I say this is my

139

: 00:10:11

Ajay: search string it'll actually be like what is a relevant thing in my index and then

140

: 00:10:15

Ajay: it'll slowly start digging deeper and deeper to find information about it so the

141

: 00:10:19

Ajay: obvious drawback here is if there's circular links somewhere in your wiki you may

142

: 00:10:24

Ajay: run into problems my coding assistants haven't run into this problem yet I think

143

: 00:10:28

Ajay: it's a combination of my wiki being relatively sane at the same time my coding

144

: 00:10:32

Ajay: assistant being intelligent enough not to like dig into itself like go get in an

145

: 00:10:37

Ajay: infinite loop but assuming that you don't run into that problem anytime soon and

146

: 00:10:43

Ajay: there's like enough ways that your coding assistant doesn't get into the trap um it

147

: 00:10:48

Ajay: should be able to reach the leaf node in the tree and and gather

148

: 00:10:54

Ajay: all the information from like high level all the way deep but progressively based on

149

: 00:10:58

Ajay: the query that you've given I see that's interesting so it's not exactly

150

: 00:11:03

Ajay: filtering uh but only looking up the information that it found relevant in every

151

: 00:11:09

Ajay: at every level of its search every branch has like a main trunk word to branch off

152

: 00:11:15

Andrew: to to other trunks right exactly and hopefully enough information maybe yeah it

153

: 00:11:20

Ajay: could be that's that's yes it could there could be roots it could be the branches

154

: 00:11:25

Ajay: but yes yeah it's all about uh giving because it's storing like a wiki the wiki

155

: 00:11:29

Ajay: every link has some metadata associated with it which was just like a small

156

: 00:11:33

Andrew: description could be one line could be more lines right yeah yeah yeah that's

157

: 00:11:37

Andrew: interesting um in my game design documents for Luxwander there is actually

158

: 00:11:43

Andrew: like wiki links and each page like I have an index um each

159

: 00:11:48

Andrew: page has like a little um appendix at the bottom where all the other concepts that

160

: 00:11:54

Andrew: touch on this concept are linked are linked yes exactly yes so that way you can see

161

: 00:12:00

Andrew: okay we talked about you know storing gold in the vault here well you're gonna

162

: 00:12:06

Andrew: need to know more about how the vault can get rate it that's yes right and so

163

: 00:12:12

Andrew: there will be a link to the war module yeah where there's like okay yeah well yeah

164

: 00:12:17

Andrew: it's a vault but it's not it's not impervious right and so and then that goes to

165

: 00:12:23

Andrew: well how do we stop the duplication of gold when people are desyncing while

166

: 00:12:29

Ajay: they're extracting the gold so it'll go to another place where it has more

167

: 00:12:33

Ajay: protections and security measures in place definitely I think that's a really

168

: 00:12:37

Ajay: interesting point that you're bringing up because it's if the comp based on the

169

: 00:12:40

Ajay: complexity of your query you could be loading up a bunch of stuff right which is 100

170

: 00:12:45

Ajay: like that could happen I think it certainly could happen and I've actually seen it

171

: 00:12:50

Ajay: happen at times where I'm like can you find all the relevant information for xyz it

172

: 00:12:55

Ajay: sometimes uh over complicates and unloads too much um so I have found

173

: 00:13:01

Ajay: that this if used in a particular way is definitely more beneficial but in the sense

174

: 00:13:06

Ajay: of saying that um I have I'm building one module right and that module may have

175

: 00:13:11

Ajay: relevant information uh that you need to be concerned about linkages to other

176

: 00:13:15

Ajay: modules so I think your exact point is mirrored there is like okay there is going to

177

: 00:13:19

Ajay: be linkage between modules but the kind of questions that comes out is like hey do I

178

: 00:13:24

Ajay: need to make changes in those modules or is the linkage more superficial so is it

179

: 00:13:30

Ajay: like does my current modules contract actually expect that if I'm feeding

180

: 00:13:36

Ajay: information to a different module and the contract going out changes, it is very

181

: 00:13:39

Ajay: likely, yes, I will need to go make changes to that other module.

182

: 00:13:43

Ajay: Assuming my prompt and the rest of the mechanisms around my coding assistant are

183

: 00:13:48

Ajay: actually doing a good job, it will be able to recognize when the contract change

184

: 00:13:53

Ajay: actually breaks the contract between the two modules and then looks up the

185

: 00:13:56

Ajay: information for both modules because then knows it has to change both bodies I think

186

: 00:14:00

Ajay: the linkages are the important part but also enough information in each one saying

187

: 00:14:06

Ajay: this is what the expected output is and this is what the expected that it is the

188

: 00:14:10

Ajay: expected input of the next module becomes relevant and codex will be like hmm oh I

189

: 00:14:15

Andrew: see this uh link to another document maybe I should take a look at that yeah that's

190

: 00:14:19

Andrew: exactly what it would say like in the little subtext like I should look at that

191

: 00:14:24

Ajay: that that is the hope and um I know I do feel like I got a little deep into this

192

: 00:14:30

Ajay: subject, but here's my take on it, though, right?

193

: 00:14:32

Ajay: Like, the shorter version of it is it's a scratch pad.

194

: 00:14:36

Ajay: Right, right, right.

195

: 00:14:37

Ajay: So before you start writing code, if the coding assistant does a good job and your

196

: 00:14:41

Ajay: wiki is set up in a reasonable way, which hopefully if you're using the methodology

197

: 00:14:45

Ajay: faithfully, it shouldn't be too crazy to do that.

198

: 00:14:50

Ajay: It will actually experiment ahead of time and be like, oh, somebody, he asked me to

199

: 00:14:53

Ajay: make this change.

200

: 00:14:54

Ajay: He asked me to do this.

201

: 00:14:55

Ajay: Or, you know, the coder or the builder asked me to do XYZ.

202

: 00:14:59

Ajay: Let me just make sure that I know enough visibility.

203

: 00:15:02

Ajay: And the way I do it, I'm like, I give it the folder location for now and say, hey,

204

: 00:15:07

Ajay: this is where the wiki lives.

205

: 00:15:08

Ajay: Go ahead and query the wiki.

206

: 00:15:10

Ajay: Then Claude can essentially go into that folder and understand how to query it, find

207

: 00:15:14

Ajay: all the linkages, get high-level information, not go too deep yet.

208

: 00:15:19

Ajay: And then it can be like, okay, this was a task.

209

: 00:15:22

Ajay: How many things do I need to change?

210

: 00:15:24

Ajay: And then it will hopefully come back to me and ask me questions like, hey, did you

211

: 00:15:27

Ajay: intend to change all of these?

212

: 00:15:29

Ajay: Should I do them now?

213

: 00:15:29

Ajay: Should I do them later?

214

: 00:15:31

Ajay: And that's the scratch pad part of it.

215

: 00:15:33

Ajay: It's like it's putting together, this is my notes.

216

: 00:15:35

Ajay: These are the modules that I believe I need to make changes to.

217

: 00:15:38

Ajay: Because when it gets into coding then, then it's like, okay, I know which ones I can

218

: 00:15:43

Ajay: do now or I need to do to make sure everything still works.

219

: 00:15:48

Ajay: Versus these are the ones I can do later to actually make everything really nice.

220

: 00:15:53

Ajay: So, this, I'm going to take a few, potentially make a jump here and mention.

221

: 00:15:58

Ajay: Do you remember we talked about the parallelization of work?

222

: 00:16:01

Ajay: I think so.

223

: 00:16:02

Ajay: So, this is kind of like the place where it helps me.

224

: 00:16:05

Ajay: It's like, now I know three different modules that are part of the same task.

225

: 00:16:09

Andrew: Right, okay.

226

: 00:16:10

Ajay: And all of a sudden, I can parallelize work across three.

227

: 00:16:14

Ajay: Because my scratch pad found them.

228

: 00:16:17

Ajay: Where, like, when I was doing my scratch work, I found them because of the linkages

229

: 00:16:22

Ajay: that were established.

230

: 00:16:23

Andrew: Yeah, all of these modules at this area in particular are interlinked.

231

: 00:16:27

Andrew: Yes, yes.

232

: 00:16:29

Andrew: And they're like, okay, if I can change something here, we'll update it there and

233

: 00:16:33

Andrew: update it here.

234

: 00:16:34

Ajay: Yes, yes.

235

: 00:16:35

Ajay: And I think a good example from my home OS without getting into too much depth is

236

: 00:16:39

Ajay: home OS is basically I have a VPN that allows me to make calls to a central

237

: 00:16:45

Ajay: machine that's running all my stuff, including the home assistant, some chore

238

: 00:16:49

Ajay: tracking stuff, as well as like a deal tracker that I'm trying to set up so now if I

239

: 00:16:55

Ajay: say that oh I want to make sure that the all of these become one website then it's

240

: 00:17:00

Ajay: actually saying okay there are three websites right now for all three of these

241

: 00:17:03

Ajay: running locally my album wiki should hopefully be able to say that all right that

242

: 00:17:07

Ajay: means that all three of them need to change in some way to centralize this website

243

: 00:17:10

Ajay: and that is where I'm hoping it'll recognize that ahead of time so I can just then

244

: 00:17:14

Ajay: say all right go ahead and make all the changes uh in parallel okay that makes sense

245

: 00:17:19

Andrew: I actually had before we move on to home os yes and uh you were just touching on it

246

: 00:17:24

Andrew: a little bit I had a I had a question actually so so something that I have an issue

247

: 00:17:29

Andrew: with in my projects is there is a moment in time where

248

: 00:17:35

Andrew: those concepts in the wiki right um my version of it I guess right

249

: 00:17:40

Andrew: they need to be updated like like I make a big design change right and

250

: 00:17:46

Andrew: I'm like okay now all of those references all those ideas need to to

251

: 00:17:52

Andrew: take into account that this change has occurred maybe it affects it could touch a

252

: 00:17:57

Andrew: lot you know blast radius could be huge so does you know Karpathy's

253

: 00:18:02

Andrew: llm wiki doesn't have uh you know a protocol in place to handle you know

254

: 00:18:08

Andrew: downstream to

255

: 00:18:10

Ajay: upstream changes that is a great use case and and I think my

256

: 00:18:15

Ajay: understanding I haven't had to make quite that change yet so this might be a

257

: 00:18:21

Ajay: little bit of a you know um a theoretical answer well hey

258

: 00:18:27

Andrew: well I'm glad I'm putting you there yeah yeah no you got what the questions are for

259

: 00:18:31

Ajay: absolutely you got it andrew you um the uh idea I would expect is um

260

: 00:18:37

Ajay: because it's ingesting things in a particular way it's like okay here's a concept

261

: 00:18:42

Ajay: here's a synthesis uh let's say you change the synthesis right and that's what it

262

: 00:18:45

Ajay: sounds like you're saying like you're saying that my business logic or something has

263

: 00:18:49

Ajay: changed.

264

: 00:18:49

Ajay: So technically my concept may go along with it.

265

: 00:18:51

Ajay: Um, I'm imagining in the case of like, by the way, I observed that this assumption I

266

: 00:18:57

Ajay: had in the synthesis was incorrect.

267

: 00:18:59

Ajay: Then now I need to update that synthesis and all the linked documents that rely on

268

: 00:19:04

Ajay: that particular assumption in the synthesis.

269

: 00:19:07

Ajay: Yes.

270

: 00:19:07

Ajay: Yes.

271

: 00:19:08

Ajay: My expectation is that, uh, there shouldn't be too much cross linkage between

272

: 00:19:13

Ajay: synthesis, but there would be linkages between concepts and potentially multiple

273

: 00:19:18

Ajay: synthesis hmm I'm not sure if that fully makes sense yeah I'm not sure

274

: 00:19:24

Ajay: that's not like hmm so so the concept is still saying that um again a

275

: 00:19:29

Ajay: concept could be like my home finance assistant could be a concept a synthesis

276

: 00:19:35

Ajay: would be something like okay now uh together with the vpn and the website

277

: 00:19:41

Ajay: how do I actually give my phone access to this website that's

278

: 00:19:46

Ajay: running on my machine so now I'm saying that oh I'm going to use tailscale as a vpn

279

: 00:19:52

Ajay: solution I can go log in and then I can also assign a magic dns to my home personal

280

: 00:19:57

Ajay: finance assistant essentially right uh the concept of personal finance assistant

281

: 00:20:03

Ajay: still exists but the synthesis that might change is like I want to replace Tailscale

282

: 00:20:07

Ajay: with something else.

283

: 00:20:08

Ajay: I see.

284

: 00:20:09

Ajay: Okay.

285

: 00:20:09

Ajay: So the delivery method now changes, which could also impact the other ones.

286

: 00:20:15

Ajay: So there might be the chore tracking is also basically connected in the same way.

287

: 00:20:20

Ajay: But the synthesis in this case would be saying that, how do I reach the home network

288

: 00:20:26

Ajay: rather than how do I access a website?

289

: 00:20:29

Ajay: There's a difference between saying that reaching the home network versus uh being

290

: 00:20:35

Ajay: able to access any website on the home network because if my synthesis contained

291

: 00:20:40

Ajay: information about uh accessing every website on the home network that means every

292

: 00:20:45

Ajay: single concept would now need to change I see because it's so the scope is pretty

293

: 00:20:51

Ajay: narrow intentionally it's basically saying that uh the synthesis is

294

: 00:20:57

Ajay: talking more about like accessing uh websites in general not a website in

295

: 00:21:02

Ajay: particular uh through a vpn solution from let's say a phone or any other device so

296

: 00:21:07

Andrew: then it's basically talking about the mechanism by which tail scale does that and

297

: 00:21:12

Andrew: it'll be like these are the pathways that match that pattern yes that I need to

298

: 00:21:16

Ajay: change precisely and and if it's isolated enough and that's what it should be with

299

: 00:21:21

Ajay: the wiki it should be a small wiki document it should be very focused on its

300

: 00:21:25

Ajay: requirements or maybe the the concept that it's basically representing or the

301

: 00:21:30

Ajay: synthesis that it's trying to answer questions about which is how to access things

302

: 00:21:34

Ajay: on the home network uh and ideally the linkages will exist with other things but now

303

: 00:21:39

Ajay: the entire aspect if I remove tail scale then ideally those links now just point to

304

: 00:21:43

Ajay: the new technology okay yeah yeah no I think

305

: 00:21:48

Ajay: that makes sense yeah I hope yeah maybe

306

: 00:21:54

Ajay: well for me it's it's more so um I

307

: 00:22:00

Ajay: guess to kind of put in the simpler terms maybe

308

: 00:22:05

Andrew: maybe it's too simple but it if you have your concepts and ideas

309

: 00:22:11

Andrew: modular enough yes and uh segregated and organized in a proper way

310

: 00:22:16

Andrew: yep when you want to change a piece of that you can just basically hot swap

311

: 00:22:22

Andrew: that module.

312

: 00:22:24

Ajay: Yes, yes.

313

: 00:22:25

Ajay: Okay, again, I really like the characterization there.

314

: 00:22:28

Ajay: Yes, there are small enough modules you just basically plug and play and you're

315

: 00:22:31

Ajay: like, hey, it still works with everything else.

316

: 00:22:33

Ajay: This one's responsibility is so small that I could just replace that and not impact

317

: 00:22:37

Ajay: everybody.

318

: 00:22:37

Andrew: And a great example of this actually just recently is with my CI tool for GitHub

319

: 00:22:42

Andrew: Actions.

320

: 00:22:43

Ajay: Oh, yeah.

321

: 00:22:44

Ajay: Right?

322

: 00:22:44

Andrew: So GitHub Actions, it was, you know, reviewing my stuff, code and stuff.

323

: 00:22:48

Andrew: but there's only one file you have to touch switch at the blacksmith yes okay I

324

: 00:22:53

Andrew: remember I was having this yeah you're right and so so that's a place kind of like

325

: 00:22:58

Andrew: maybe a little bit more precise than a wiki would be but you only got to change it

326

: 00:23:03

Andrew: in one place but it changes the entire flow of information it goes to a completely

327

: 00:23:08

Andrew: different tool completely different place but it doesn't break it doesn't break

328

: 00:23:12

Andrew: anything that's exactly plug and play hot swap absolutely and I think that's kind of

329

: 00:23:17

Ajay: the goal behind the wiki idea is there's links to many things, but then each one of

330

: 00:23:23

Ajay: them is so small or ideally small enough.

331

: 00:23:25

Ajay: And if the wiki is doing a good job, yeah, hot swap it out, put a new concept in,

332

: 00:23:29

Ajay: put a new synthesis and put in whatever you need.

333

: 00:23:32

Ajay: And then ideally, when you actually come back and you start writing code, it's also

334

: 00:23:35

Ajay: organizing the code in a similar-ish way so that you can also potentially do that

335

: 00:23:38

Ajay: with your code in the future.

336

: 00:23:41

Andrew: Yeah.

337

: 00:23:41

Andrew: You recognize patterns that you can repeat.

338

: 00:23:45

Ajay: And Wiki is hopefully helping you recognize them before they become many, many lines

339

: 00:23:49

Ajay: of code in the future.

340

: 00:23:50

Ajay: There you go.

341

: 00:23:51

Andrew: Yeah.

342

: 00:23:52

Andrew: Yeah.

343

: 00:23:52

Andrew: I'm going to have to take a look at this.

344

: 00:23:54

Andrew: Yeah.

345

: 00:23:56

Ajay: I have been talking to some folks in my work about, like, the main reason I think

346

: 00:24:01

Ajay: it's interesting to other people outside of this is because people do

347

: 00:24:05

Ajay: modularization.

348

: 00:24:07

Ajay: They break things up.

349

: 00:24:08

Ajay: They do these things intentionally and they repeat that process so that they can get

350

: 00:24:12

Ajay: modules that are small enough.

351

: 00:24:14

Ajay: In my past experience, one of the things I always used to hear is like, if you have

352

: 00:24:17

Ajay: a file that's over a thousand lines, you're probably doing too much in that file.

353

: 00:24:21

Ajay: Therefore, you should probably decompose it into smaller components and break it

354

: 00:24:24

Ajay: down into multiple files.

355

: 00:24:26

Ajay: That same idea is now essentially saying, before you even write code, that same idea

356

: 00:24:31

Ajay: is now being applied to Vail on Wiki.

357

: 00:24:32

Ajay: At least that's the way I use it.

358

: 00:24:34

Andrew: We're just making it more compact for the human context window.

359

: 00:24:37

Ajay: Yes, because then hopefully humans can deal with the markdown files and then the AI

360

: 00:24:42

Ajay: coding assistants can actually handle the coding files for there we go yeah yeah at

361

: 00:24:46

Ajay: least that's yeah that's what I'm thinking about so um yeah I I do I do question

362

: 00:24:51

Andrew: this because I think we talked a little bit of context windows and yes yes and I

363

: 00:24:56

Andrew: want to know a few I had a few questions for you andrew because I think you have uh

364

: 00:25:01

Ajay: you're making a transition for coding from one coding tool coding assistant to

365

: 00:25:06

Andrew: another that's true yes yes so so with you know we we've all heard um

366

: 00:25:12

Andrew: anyone who's using Claude Code right now is very very you know appraised on the

367

: 00:25:18

Andrew: limits in place you're running into limits yeah we're you know we just got fable

368

: 00:25:23

Andrew: back like a day or two ago you know that's going to chew through your your weekly

369

: 00:25:28

Andrew: limit like nothing if you give it a big enough task and just for relevance we're

370

: 00:25:32

Ajay: recording a couple of days before fourth of july here so

371

: 00:25:36

Ajay: just to kind of date it correctly but yes um my thing

372

: 00:25:42

Andrew: is like there's a trade-off between the two tools that I'm

373

: 00:25:48

Andrew: using so I I use both codex and Claude Code to develop my project right now

374

: 00:25:54

Andrew: and and my project is a a custom minecraft server um based about like nation

375

: 00:26:00

Andrew: combat um we can get into that more in detail later there's a lot there's a lot a

376

: 00:26:05

Andrew: non-trivial problem like a few different there there's a lot to it that meets the

377

: 00:26:10

Andrew: eye yes um but to get back to codex and claude code the the main thing that I'm

378

: 00:26:16

Andrew: running into is that there's just we were talking about budgets last

379

: 00:26:21

Andrew: episode yes and claude's budget for usage is pretty small actually

380

: 00:26:28

Andrew: it I can use it up 20x max it can use it up in like two or three days

381

: 00:26:34

Andrew: easy and that's a weekly limit we're talking about weekly limit yeah yeah and so I

382

: 00:26:38

Andrew: can only use it for two or three days it can code the whole time and then it's

383

: 00:26:45

Andrew: like well codex gives a lot more they're they're being a lot more

384

: 00:26:50

Andrew: generous in their uh you know what was it like limits I guess and yeah in

385

: 00:26:56

Ajay: their available capacity for individuals yes and and you also said that there was

386

: 00:27:01

Andrew: the concept of uh limit resets yes yes so for whatever reason open ai

387

: 00:27:07

Andrew: is being very generous with their you know usage and limit resets tokens they're

388

: 00:27:12

Andrew: practically handing them out like candy right now right now that's that's a relevant

389

: 00:27:16

Andrew: right now right now right now yes yes yes and and as you were talking about they're

390

: 00:27:22

Andrew: probably gearing up to try to uh be very competitive with anthropic right now yes

391

: 00:27:26

Ajay: and I think we were talking about it briefly before we start recording um it is

392

: 00:27:31

Ajay: interesting that in the last few months uh anthropics revenue has uh well reported

393

: 00:27:36

Ajay: revenue these are not public companies at this point so we don't know but the uh

394

: 00:27:41

Ajay: shared reported revenue in some way has exceeded that of OpenAI, which OpenAI had

395

: 00:27:47

Ajay: been winning this revenue battle for, oh, since they launched.

396

: 00:27:51

Ajay: So it's been many years and they've been well ahead of Anthropic.

397

: 00:27:55

Ajay: And Anthropic had been wagging behind for the longest time.

398

: 00:27:58

Ajay: And the revelation here really is that Anthropic is doing this increase in revenue,

399

: 00:28:03

Ajay: beating them in revenue with 5% essentially of the same user base.

400

: 00:28:07

Ajay: Exact numbers, as I said, I don't know.

401

: 00:28:09

Ajay: We don't know.

402

: 00:28:10

Ajay: This is just what is being reported uh but with a 20th of the user base they're

403

: 00:28:15

Andrew: exceeding that revenue number that is very impressive um I it's hard for me to

404

: 00:28:21

Andrew: conceive that you know maybe there's some hollywood accounting occurring here

405

: 00:28:27

Ajay: well the logic at least I heard which seemed reasonable at the time was uh we're

406

: 00:28:33

Ajay: talking about white collar workers uh using and enterprises using

407

: 00:28:38

Andrew: anthropic solutions oh yeah that's right I see so so more enterprise is using it

408

: 00:28:44

Ajay: compared to you know commercial yes and when enterprise uh conversations get into

409

: 00:28:50

Ajay: the picture you're talking about much higher revenue from api pricing and also for

410

: 00:28:54

Ajay: enterprise licensing each user could be hundreds of dollars per um per seat in the

411

: 00:29:00

Ajay: enterprise license essentially or more or it could be more yeah so it suddenly

412

: 00:29:06

Ajay: seems like not an unreasonable statement to make that by capturing an enterprise

413

: 00:29:11

Ajay: focused market anthropic might be winning the money battle the revenue battle so

414

: 00:29:15

Andrew: that makes sense and you you read online that all of the you know personal

415

: 00:29:21

Andrew: subscriptions are being subsidized by all these corporate um contracts and I

416

: 00:29:27

Ajay: wouldn't doubt it at all right and I think the use cases we talk a lot about cloud

417

: 00:29:31

Ajay: code for a reason uh they have built the agent harness which is essentially and

418

: 00:29:36

Ajay: again the concept I think uh I don't know andrew would you like to talk about it

419

: 00:29:39

Ajay: because I think harness is something we've talked a little bit about right you

420

: 00:29:42

Andrew: explained it in a really good way earlier um I don't remember exactly how I

421

: 00:29:45

Andrew: explained it but I can comment on that between at least codex in Claude Code I I very

422

: 00:29:50

Andrew: much like cloud codes harness compared to codex for

423

: 00:29:56

Andrew: example um at least in the latest versions here of Claude Code um you can see

424

: 00:30:01

Andrew: whenever it launches agents you can clearly navigate easily open the agent's

425

: 00:30:07

Andrew: own you know cli windows you can see what they're doing individually you can see how

426

: 00:30:13

Andrew: many shells they've opened up you can see how much tokens they're using how long

427

: 00:30:18

Andrew: it's been going it's actually not that verbose in codex oh you really

428

: 00:30:23

Andrew: don't get anything you you see codex open up an agent and that's

429

: 00:30:29

Andrew: it you don't really get to see anything else that's pretty that's pretty basic yeah

430

: 00:30:35

Andrew: and and you only get to see I believe the limit for just in in line for one cli

431

: 00:30:40

Andrew: session is six parallel agents wow yes it can do more if it

432

: 00:30:46

Andrew: uses like codec zek launching it in a shell you can do more if it launches in the

433

: 00:30:51

Andrew: shell but if built into the cli yeah right now as of today

434

: 00:30:57

Andrew: july 2nd I only up to six wow agents and that limitation is

435

: 00:31:03

Andrew: obviously not there in Claude Code you launch a work phone Claude Code you could

436

: 00:31:07

Andrew: launch a hundred agents and burn through your five hour limit within three seconds

437

: 00:31:12

Ajay: which you've done yeah which which I've done yes with fable when it was last time

438

: 00:31:16

Ajay: yes yes um and and I think so but I do want to go back to that six is so low

439

: 00:31:21

Ajay: especially if you're opening it up in a different execution shell you lose

440

: 00:31:25

Andrew: visibility over it yeah I mean it can do it it knows how to open up right like like

441

: 00:31:31

Andrew: so so codex exec is uh the same as uh claude dash p okay okay I and I

442

: 00:31:37

Ajay: believe that clod-p has existed for at least six months now so I think I was using

443

: 00:31:42

Ajay: it back in january it's essentially just headless clod right right which is um it's

444

: 00:31:47

Ajay: really interesting and and I'll maybe I'll also take a quick step back and mention

445

: 00:31:51

Ajay: like when we talk about the hardness we're essentially saying the place where the lm

446

: 00:31:55

Ajay: runs because we're saying that the clod model the codex model which is using chat

447

: 00:32:00

Ajay: gpd behind the scenes of some some flavor of chat gpd behind the scenes is

448

: 00:32:04

Ajay: essentially being wrapped around by a bunch of instructions last time I think we

449

: 00:32:09

Ajay: talked a little bit about the concept of a system prompt and a user prompt so the

450

: 00:32:12

Ajay: idea is thinking about like this CLI wrapper or maybe a you know like a desktop

451

: 00:32:18

Ajay: application wrapper whatever is being whatever is available I think Google provides

452

: 00:32:23

Ajay: the anti-gravity which is essentially their solution they build it up around Gemini

453

: 00:32:29

Ajay: the idea is this wrapper has a system prompt has many other features that they're

454

: 00:32:34

Ajay: building, like how do you implement skills and all these other fun things.

455

: 00:32:37

Ajay: Fun being a relative term here.

456

: 00:32:40

Ajay: But the intention is that the wrappers are building or the harnesses.

457

: 00:32:44

Andrew: And there is actually a lot of focus right now in the industry that I'm reading on

458

: 00:32:49

Andrew: that it's actually the wrappers that matter more than the model.

459

: 00:32:53

Andrew: You'll hear the Cursor CEO talk about that a lot.

460

: 00:32:56

Ajay: Exactly.

461

: 00:32:57

Ajay: Cursor is also wrappers and they are the harness.

462

: 00:33:00

Ajay: Yes.

463

: 00:33:00

Ajay: And it is becoming a very common tool to say that hey lm capabilities will

464

: 00:33:07

Ajay: go so far like the models themselves are getting really good but there is also a

465

: 00:33:12

Ajay: diminishing return aspect with the new models but the latest models may not be so

466

: 00:33:17

Ajay: much better compared to like the last ones so then essentially the wrappers that

467

: 00:33:22

Ajay: they're built on the harness that is built around them is the place where a lot of

468

: 00:33:26

Ajay: value can be added now because it also provides integrations and understanding of

469

: 00:33:31

Ajay: the surrounding environment the information that you get out of it I think as you

470

: 00:33:33

Ajay: mentioned just now uh clod code is going to give you so much information about like

471

: 00:33:37

Ajay: hey these are the tokens I sent these are the tokens I got back uh this is the

472

: 00:33:41

Ajay: prompt that I started with here all your sub-agents I've got a hundred of them like

473

: 00:33:45

Ajay: you can hierarchically track them uh recently that happened to me is I actually

474

: 00:33:49

Ajay: started I I started uh research similar to you um ended up with started with four

475

: 00:33:53

Ajay: agents oh no and the four agents each spun up four more and then I started going

476

: 00:33:58

Ajay: down the chain I think I was at like 25 or something yeah they can get it they can

477

: 00:34:02

Andrew: get away from you yeah yeah yeah yeah I I actually just realized that agents

478

: 00:34:08

Andrew: can spawn their own agents the sub agents can spawn more so can those agents spawn

479

: 00:34:13

Andrew: their own agents I don't know I went five level five levels down I I was like this

480

: 00:34:18

Ajay: is impressive I didn't even realize I said really five levels see yeah I don't check

481

: 00:34:23

Andrew: it enough to know like I know it goes one more down but five levels down Yes, and

482

: 00:34:28

Ajay: then you've got to keep pressing down.

483

: 00:34:29

Ajay: You've got to keep pressing down.

484

: 00:34:31

Ajay: Oh, okay.

485

: 00:34:34

Andrew: That's why it says there's 25 shells open when there's two agents.

486

: 00:34:38

Andrew: Yeah, yeah.

487

: 00:34:40

Ajay: You've just got to keep checking on that.

488

: 00:34:42

Andrew: Keep going, keep going.

489

: 00:34:44

Ajay: But the interesting thing here, I think, similar to what you mentioned,

490

: 00:34:48

Ajay: for Claude, I started recently using the actual desktop app, which is the first time

491

: 00:34:53

Ajay: I've done it.

492

: 00:34:54

Ajay: This week I started using it because I started writing code for a new project

493

: 00:34:58

Ajay: recently at work, and I think I was looking for some features as part of that new

494

: 00:35:03

Ajay: work.

495

: 00:35:04

Ajay: One of them was trying to get access to remote control, which that's a whole other

496

: 00:35:07

Ajay: story.

497

: 00:35:08

Ajay: I'm not going to go into that right now.

498

: 00:35:09

Ajay: But one thing that was really fascinating to me was I started, I gave it a prompt.

499

: 00:35:15

Ajay: I told it that, hey, this is the issue.

500

: 00:35:17

Ajay: Go ahead and solve this issue.

501

: 00:35:18

Ajay: We use linear at work, but I think GitHub issues would basically give you the same

502

: 00:35:22

Ajay: thing.

503

: 00:35:23

Ajay: notion probably would give you the same thing uh we're actually talking about a lot

504

: 00:35:26

Ajay: of interesting project management tools right now I do feel like we need to circle

505

: 00:35:30

Ajay: back on this at some point in the future because that's another part of the workflow

506

: 00:35:33

Ajay: that I really you know like find useful or my workflow that I find really useful for

507

: 00:35:38

Ajay: now the idea is I gave it an issue I give it three issues I'm like you can execute

508

: 00:35:42

Ajay: these in parallel because I did the whole element key thing I figured out what

509

: 00:35:47

Ajay: modules are related uh and then I said okay go ahead and execute them in parallel

510

: 00:35:50

Ajay: use subagents uh it did it used subagents everything worked it finished

511

: 00:35:56

Ajay: and then it's like oh by the way I want to spin up three more subagents I was like

512

: 00:36:00

Ajay: why for what I didn't tell you to do this and then it was like oh because I'm going

513

: 00:36:05

Ajay: to review the code I was like that has never happened to me in the claude cli but

514

: 00:36:10

Ajay: this case it actually did that but interestingly that was not the only thing.

515

: 00:36:16

Ajay: It also kept track of a new skill that it created.

516

: 00:36:18

Ajay: It added like 3,800 lines.

517

: 00:36:20

Andrew: So it was refining itself as it went along without your direction.

518

: 00:36:24

Andrew: Yes, yes.

519

: 00:36:25

Ajay: And it did a better job of picking up my local CLI settings.

520

: 00:36:28

Ajay: So things like the skills that I had previously installed, the plugins that I

521

: 00:36:32

Ajay: previously installed.

522

: 00:36:34

Ajay: So it was able to not only detect my local environment, it picked the right skills.

523

: 00:36:39

Ajay: It created a skill that was a gap in all the skills I already had and and this this

524

: 00:36:44

Ajay: new skill was only for this session and then it suggested that I

525

: 00:36:50

Ajay: review the code

526

: 00:36:52

Ajay: automatically so I was literally just pressing enter enter enter enter and I was

527

: 00:36:55

Ajay: like okay seemed reasonable some things that I would have had to go back in the past

528

: 00:36:59

Ajay: and specifically prompted for so this is another aspect of the harness which I'm

529

: 00:37:04

Andrew: like I actually don't know yet So I guess the CLI harness is different from the

530

: 00:37:08

Andrew: desktop harness.

531

: 00:37:10

Ajay: I think so.

532

: 00:37:12

Ajay: That's my, at least that's what it looks like to me.

533

: 00:37:16

Andrew: I haven't used it yet, but it does sound like it.

534

: 00:37:19

Andrew: I've never had Claude Code prompt me to review something on its own.

535

: 00:37:25

Andrew: But also, I've modified the memories and base instructions enough at this point

536

: 00:37:31

Andrew: that I don't think it would be a fair assessment on my part.

537

: 00:37:34

Ajay: I don't have memories enabled either, so that's interesting.

538

: 00:37:37

Ajay: I have a bunch of skills and stuff, so I do not have memories enabled for cloud

539

: 00:37:40

Ajay: code, at least for my work setup.

540

: 00:37:43

Andrew: That makes sense.

541

: 00:37:43

Ajay: So it's definitely an interesting point about Codex versus Claude Code.

542

: 00:37:50

Ajay: We've talked a lot about the models historically.

543

: 00:37:53

Ajay: I think the harness discussion is becoming very interesting because it does seem

544

: 00:37:58

Ajay: like that is where potentially these companies will add value in the near term, at

545

: 00:38:02

Ajay: least for maybe the next six months, in my opinion.

546

: 00:38:04

Ajay: um, and it could impact your workflow a reasonable amount.

547

: 00:38:08

Ajay: Um, I'm also curious how much of this they're getting feedback or how much of this

548

: 00:38:14

Ajay: they're seeing user metrics and doing all this stuff.

549

: 00:38:17

Ajay: Um, but I know, I don't know, would you, would you go back?

550

: 00:38:20

Ajay: Would you consider that a reason to take on the cost of, you know, Claude's

551

: 00:38:24

Ajay: non-trivially expensive models?

552

: 00:38:28

Andrew: I, you know, so, so it's interesting to me.

553

: 00:38:30

Andrew: I think what you're saying makes sense.

554

: 00:38:32

Andrew: I think what you're seeing in the desktop app like I would consider them to be

555

: 00:38:37

Andrew: additional guardrails so people who are using the desktop application versus cli at

556

: 00:38:42

Andrew: least at first right maybe they'll discover cli later but maybe their first um

557

: 00:38:48

Andrew: exposure to clog code would be obviously within the mobile app there are some parts

558

: 00:38:53

Andrew: of that that are on mobile now yeah and then on obviously the desktop app and so the

559

: 00:38:57

Andrew: people who are running into it on desktop probably don't have a proficiency or at

560

: 00:39:02

Andrew: least familiarity with cli or or it's more like you know they're not in the

561

: 00:39:08

Andrew: weeds they're more so I'm just gonna click stuff around and it's gonna work but that

562

: 00:39:13

Andrew: also means like we were talking about a few episodes ago they probably don't have

563

: 00:39:17

Andrew: you know ideas of like the software development life cycle or any formal software

564

: 00:39:22

Andrew: engineering practices yes but with a tuned harness

565

: 00:39:27

Andrew: yeah it can guide that person into doing that stuff just just on its own right

566

: 00:39:32

Andrew: because thereby actually enhancing their experience because it's forcing them

567

: 00:39:38

Andrew: and recommending to them that you know what maybe we should review your code instead

568

: 00:39:43

Andrew: of just go off vibes I I see what you do wow all right I

569

: 00:39:49

Ajay: think that actually opens up a whole nother thing in my mind is

570

: 00:39:55

Ajay: our question in my mind so um we talked about a little bit I don't know if you

571

: 00:40:00

Ajay: remember this I'm just gonna bring this up we talked about rag being dead because

572

: 00:40:05

Andrew: there's a whole aspect of that is vibe coding dead now if if claude is going

573

: 00:40:10

Andrew: to solve it oh that's interesting you know it I think if vibe

574

: 00:40:16

Andrew: coding is dead then the human in the loop has less value that's that's an

575

: 00:40:22

Ajay: interesting one that that is an interesting one yeah um I I will I will challenge

576

: 00:40:27

Ajay: that a little bit because I do believe that there is an aspect of uh judgment we

577

: 00:40:33

Ajay: talked about last time yes yes yes so that human in the loop skill in my opinion

578

: 00:40:38

Ajay: that is a I believe it's a skill um is still very relevant and no I fully agree with

579

: 00:40:44

Andrew: that and I guess what I'm trying to say is is the reason why people say vibe coding

580

: 00:40:50

Andrew: is like you can get away with it right right and so so if it's

581

: 00:40:56

Ajay: dead right that means that it's doing everything for you at

582

: 00:41:02

Ajay: this point that's that's that's interesting because um oh or maybe or

583

: 00:41:07

Andrew: is it taking vibe coding to the next level exactly the opposite maybe maybe I'm

584

: 00:41:12

Ajay: going about it the wrong way yeah yeah that's okay okay wow all right so are the

585

: 00:41:17

Andrew: guardrails promoting more vibe coding oh see that's that's actually where

586

: 00:41:22

Andrew: I'm turning it around like like I'm

587

: 00:41:25

Andrew: realizing that that like if if the desktop harness guides a

588

: 00:41:31

Andrew: newbie to follow sdlc to follow software engineering concepts wouldn't

589

: 00:41:37

Andrew: that promote them like I'm going like hypocritical against myself right now wouldn't

590

: 00:41:43

Andrew: that allow more vibe coding because people are going to get away with it because

591

: 00:41:49

Andrew: now you have people who are developing apps clod is forcing you to review it

592

: 00:41:55

Andrew: forcing you to follow some set of standards to their system prompt yes yes

593

: 00:42:01

Ajay: oh boy and then maybe vibe coding is back you know you know I I

594

: 00:42:07

Ajay: think then software engineering is becoming vibe coding is becoming software

595

: 00:42:12

Ajay: engineering because we started we started this a little bit ago we started this

596

: 00:42:16

Ajay: conversation one of the things in the first episode was um are the existing rituals

597

: 00:42:22

Ajay: existing software engineering process still relevant and I still think they

598

: 00:42:28

Andrew: are I I do too I do think I I think for any application that is built on

599

: 00:42:33

Andrew: non-cli Claude Code can live with the bare minimum that

600

: 00:42:39

Andrew: they're gonna throw in yeah right but anything that is you know has distributed

601

: 00:42:44

Andrew: architecture or has any sort of scale yeah it's going to take more than this

602

: 00:42:50

Ajay: code looks nice yeah this code is you know technically correct now let me actually I

603

: 00:42:56

Ajay: am changing my mind a little bit here and coming back to this question is uh but

604

: 00:43:00

Ajay: this is actually what software engineering is all about too is like self-improvement

605

: 00:43:05

Ajay: over time that's true if if Claude Code and tools coding assistants in general are

606

: 00:43:10

Ajay: actually now doing following the self-improvement loop to such an extent um where

607

: 00:43:15

Ajay: does it stop it stops when the human tells it to stop so now the human the loop is

608

: 00:43:21

Ajay: the liability no oh my god because per se you create the perfect

609

: 00:43:27

Andrew: harness yeah that will always create you know enterprise grade ready yeah

610

: 00:43:33

Andrew: production ready code yeah no matter what task you give it but it takes time to do

611

: 00:43:39

Andrew: that that is and and tokens yeah and tokens right so at some point the

612

: 00:43:45

Andrew: humans are going to say enough is enough I want the product now wow okay

613

: 00:43:51

Ajay: I'm actually going to go off the bar I'm going to go away from our show notes all

614

: 00:43:54

Ajay: together and pick up an interesting new vv I feel like we've uh hit a little bit of

615

: 00:43:58

Ajay: an interesting little vein of something here uh so I'm gonna I'm gonna go ahead and

616

: 00:44:02

Ajay: mention that I think there was a big consulting company that is out there doing

617

: 00:44:07

Ajay: consulting for software purposes.

618

: 00:44:09

Ajay: I'm forgetting the name, so I'm not going to bring it up right now.

619

: 00:44:13

Ajay: Though what I read in the news article about them was that they actually started

620

: 00:44:18

Ajay: doing consulting for VCs to see if the companies that the VCs wanted to invest

621

: 00:44:24

Ajay: in slash buy actually had a moat because of their software or because of something

622

: 00:44:30

Ajay: else their approach towards doing that was can I vibe code that company that that

623

: 00:44:36

Ajay: this vc wants to invest in can I vibe code their software and if I can vibe code

624

: 00:44:41

Ajay: their software then it's not a mode then their software is not a mode because I can

625

: 00:44:44

Andrew: just build it now with using ai that's a really interesting topic actually and I was

626

: 00:44:49

Andrew: thinking about this reading about it today even on reddit the limitation now is not

627

: 00:44:56

Ajay: building the application it is now the ecosystem you build around it

628

: 00:45:01

Ajay: that is yes okay or the data that you have to form it in order for it to become

629

: 00:45:07

Ajay: a moat and your data slash your existing brand is the moat or maybe you're building

630

: 00:45:13

Andrew: a new brand that becomes it's just like everyone who grabs Claude Code is now

631

: 00:45:18

Andrew: building like their own habit tracker yes right for whatever reason that's a

632

: 00:45:23

Andrew: big one that I see online and now there's billions of habit trackers on the app

633

: 00:45:29

Andrew: store vibe coded like we need another one of those things that no one ever has you

634

: 00:45:34

Ajay: know it's a bad habit making them at this point well please sir well but uh so

635

: 00:45:39

Ajay: you're telling me I shouldn't build my own homeless no I shouldn't build my

636

: 00:45:45

Andrew: own habit track maybe not the habit tracker all right all right maybe not I I think

637

: 00:45:51

Andrew: habits are more than an app and that's a whole nother discussion that's yes you're

638

: 00:45:57

Ajay: not going into that rabbit hole I agree I agree it is more

639

: 00:46:03

Ajay: than an app and um I think uh having the app to do certain things now that the

640

: 00:46:09

Ajay: cost of building apps is becoming so low uh well depending on the token limits

641

: 00:46:14

Ajay: depending on the model you use yeah I still do believe that if we get into specifics

642

: 00:46:17

Ajay: of models uh opus 47 is still giving you so much value and building using it with

643

: 00:46:24

Ajay: the right harness using it with actually the right methodology I think we've talked

644

: 00:46:27

Ajay: about gsd I'm not going to say the full name anymore I don't want to get flagged

645

: 00:46:31

Ajay: here um is actually going to give you so much value even without spending the uh you

646

: 00:46:35

Ajay: know 60 dollars for a million in and million out from fable um so that is actually

647

: 00:46:41

Ajay: where I I do want to talk a little bit about the homo s setup and like essentially

648

: 00:46:46

Ajay: saying that um does it now make sense for everybody to build their own uh personal

649

: 00:46:52

Ajay: finance app right like this used to be a whole thing like you I pay 15 an hour for

650

: 00:46:57

Ajay: monarch money today because it's a great finance app it has a bunch of great

651

: 00:47:00

Ajay: integrations uh the integrations are still broken all the time because the websites

652

: 00:47:06

Ajay: like the I I go to fidelity to look at my 401k a week later they changed something

653

: 00:47:10

Ajay: on the fidelity side and the integration from my finance tracker from Monarch to

654

: 00:47:15

Ajay: Fidelity fails and I have to re-sync.

655

: 00:47:18

Ajay: And it's a painful process because now I'm essentially hoping a server that Monarch

656

: 00:47:23

Ajay: is running on their cloud system is able to do the same steps that I'm supposed to

657

: 00:47:28

Ajay: actually be doing on my phone.

658

: 00:47:30

Andrew: So the integrations for Monarch money, let's talk about that real quick.

659

: 00:47:33

Andrew: So I know one of the integrations is Plaid.

660

: 00:47:36

Ajay: Yes.

661

: 00:47:37

Andrew: Right, that's the popular one that you use.

662

: 00:47:39

Andrew: Finicity is the other one.

663

: 00:47:40

Andrew: Yeah, Finicity.

664

: 00:47:41

Andrew: That's by MasterCard, right?

665

: 00:47:43

Ajay: Yes.

666

: 00:47:43

Andrew: So those integrations aren't from Monarch.

667

: 00:47:47

Andrew: That is true.

668

: 00:47:48

Andrew: So there's these big companies that are focused just on this product, and they can't

669

: 00:47:52

Andrew: get it right.

670

: 00:47:52

Ajay: They can't.

671

: 00:47:53

Ajay: No, they can't.

672

: 00:47:54

Andrew: And a part of that is the diversity of vendors, right?

673

: 00:48:00

Andrew: So Fidelity is kind of like your boomer, your boomer fintech, quote, quote,

674

: 00:48:06

Andrew: company, boomer fintech.

675

: 00:48:08

Ajay: maybe I'll say more uh traditional uh more dated traditional okay okay

676

: 00:48:14

Ajay: okay some friendly language I I I do understand where you're

677

: 00:48:20

Ajay: going with that you know I do understand that the institutions have been around

678

: 00:48:26

Ajay: for a certain amount of time yes um but sorry keep going keep going those

679

: 00:48:32

Andrew: standards that you know the the oldies like fidelity I think

680

: 00:48:38

Andrew: they are having trouble yeah being compatible with third-party products it's not

681

: 00:48:43

Andrew: necessarily monarch's fault or or it's not city or plant's fault right um it it can

682

: 00:48:49

Andrew: be the company and then I say that because I I do use monarchs there we go and I

683

: 00:48:54

Andrew: have tried to attach my fidelity account to it and you accept it like five

684

: 00:49:00

Andrew: times yep and it doesn't just quite work so so the interesting

685

: 00:49:06

Ajay: question I have there essentially is and and uh to provide a little bit of context

686

: 00:49:10

Ajay: before I go too deep here is my personal finance uh home project um part of

687

: 00:49:16

Ajay: my home os uh projects uh is intentionally there to follow the concept of a book

688

: 00:49:22

Ajay: uh I can get into the software development lifecycle how I did all this stuff in a

689

: 00:49:26

Ajay: minute uh but the problem it was trying to solve is the problem we're talking about

690

: 00:49:30

Ajay: uh the book was called pillars of wealth uh it's called pillars of wealth uh the

691

: 00:49:34

Ajay: concept is offense uh defense and investments so offense is you're earning money on

692

: 00:49:40

Ajay: a regular basis defense is how you're saving money investments is how you're

693

: 00:49:43

Ajay: investing the money that you're now saving um that is kind of like the super high

694

: 00:49:46

Ajay: level uh before you can do all these things you need visibility you need to be able

695

: 00:49:51

Ajay: to see uh how much money you're earning how much money you're saving how much money

696

: 00:49:55

Ajay: is going into existing investments like your 401k which is where we get into this

697

: 00:49:59

Ajay: whole conversation of like, can I see my 401k infidelity?

698

: 00:50:02

Ajay: And I have been, we've been trying to use Monarch for that purpose.

699

: 00:50:08

Ajay: I will not blame Monarch.

700

: 00:50:09

Ajay: I'm sorry, Monarch.

701

: 00:50:11

Ajay: The idea is that the integrations are too weak right now.

702

: 00:50:14

Ajay: They're just, it's just very finicky.

703

: 00:50:16

Ajay: My theory is that the integrations, as you said, these

704

: 00:50:22

Ajay: traditional companies have built them and user acceptance tested them

705

: 00:50:28

Ajay: for individuals accessing these systems on their phones, accessing them on their

706

: 00:50:33

Ajay: laptops or desktops or whatever other machines.

707

: 00:50:36

Ajay: I think the solution that Monarch's integrators like Plaid and Finicity are coming

708

: 00:50:42

Ajay: up with are saying that, hey, we will run these workflows to get you access in some

709

: 00:50:47

Ajay: cloud machine that is running in some network isolated environment and which has

710

: 00:50:53

Ajay: access to only certain things like the website directly.

711

: 00:50:56

Ajay: Those are traditionally things which get blocked by the banks, like get blocked by

712

: 00:51:01

Ajay: the WAFs that are hosted.

713

: 00:51:04

Ajay: Like basically Cloudflare is like a solution to avoid botnet attacks, right?

714

: 00:51:09

Ajay: Like denial of service attacks, like all of these bad actors.

715

: 00:51:12

Ajay: Like usually those websites treat these solutions almost like bad actors.

716

: 00:51:18

Ajay: So interestingly, the question comes to me is, well then doesn't it make it much

717

: 00:51:24

Ajay: doesn't it make a lot more sense for me to build something that is actually running

718

: 00:51:27

Ajay: on the desktop which is where I'm supposed to actually access the website from now

719

: 00:51:30

Ajay: it's running on my local machine my local machine has a lot more trust with those

720

: 00:51:35

Andrew: servers excuse me and you're also not hammering their services

721

: 00:51:40

Andrew: exactly yeah uh and that is the intention behind why I'm

722

: 00:51:46

Ajay: like oh can I spin up a home finance assistant running locally that can essentially

723

: 00:51:52

Ajay: do the integration for me using one of these coding assistants or something to

724

: 00:51:56

Ajay: basically make the connection work instead of relying on these cloud solutions to

725

: 00:52:02

Ajay: essentially collect all this.

726

: 00:52:04

Ajay: And then the cost of developing these applications is so relatively low, I was able

727

: 00:52:09

Ajay: to spin up the majority of the work fairly quickly.

728

: 00:52:13

Andrew: So an interesting concept just popped in my mind, actually, as you talked about

729

: 00:52:18

Andrew: this.

730

: 00:52:19

Andrew: it is almost like we are in the digital age of the handyman yeah yeah

731

: 00:52:25

Ajay: yeah I I guess now the question

732

: 00:52:30

Ajay: becomes is your handyman is be able to build so many things so

733

: 00:52:36

Ajay: quickly how are you actually going to build it and how is it actually going to

734

: 00:52:42

Ajay: get used is maybe a more interesting question because now oh yeah go ahead sorry

735

: 00:52:46

Andrew: yeah yeah no so so so I guess like you know if you think of you know like a handyman

736

: 00:52:51

Ajay: around the house or something like that every solution is bespoke oh yes

737

: 00:52:57

Ajay: yes I see okay I'm okay I had I was going down the wrong path I appreciate you

738

: 00:53:02

Ajay: correcting me that yeah 100 I agree like everything is custom built yeah because

739

: 00:53:06

Andrew: it's like there's no standardization for this house for this exact problem for this

740

: 00:53:11

Andrew: family for their requirements yes and that is the opportunity now for the digital

741

: 00:53:16

Andrew: handyman yeah for the everyday person to cater their requirements for whatever

742

: 00:53:22

Andrew: digital needs they have just like you're saying you want to make your own finance

743

: 00:53:27

Andrew: app for you yeah it's gonna fit your ecosystem yeah and I and I do say that online a

744

: 00:53:33

Andrew: lot that that is how people are typically using Claude Code and codex now is to build

745

: 00:53:39

Andrew: their own custom solutions not necessarily deploying it for revenue and I think

746

: 00:53:44

Andrew: maybe that's where vibe coding lives on that and I think that I genuinely do believe

747

: 00:53:49

Ajay: so and I feel like that is the part that I'm most excited about is um there is a

748

: 00:53:54

Ajay: bunch of solutions that are just nobody's going to build an app for that nobody's

749

: 00:53:58

Ajay: actually going to build a solution in fact even if they do it's probably not going

750

: 00:54:02

Ajay: to be good uh so it's it's like one of those cases where I'm like maybe I just vibe

751

: 00:54:05

Ajay: code it and if there's something that's less an ideal I just change it to meet my

752

: 00:54:09

Ajay: needs and uh you I really like the analogy again you mentioned you you're really

753

: 00:54:15

Ajay: hitting them out of the park today I mean as usual as always uh the the uh digital

754

: 00:54:20

Ajay: handyman is building a bespoke solution for me based on my needs and the cost of the

755

: 00:54:25

Ajay: digital handyman is so trivially low compared to what it would have been a few years

756

: 00:54:30

Ajay: ago uh that yeah I I would have had to hire a engineer like for thousands of dollars

757

: 00:54:34

Ajay: to do something similar now it's like in the you know it's the tens of dollars to

758

: 00:54:38

Ajay: probably do something relatively similar uh a few months of work now gets compressed

759

: 00:54:43

Ajay: into a few days or maybe a week I think as you mentioned your codex experience yes

760

: 00:54:48

Andrew: yes I am very thankful to open ai I love the usage for

761

: 00:54:53

Andrew: sets right now I I and and I think uh that is that is a

762

: 00:54:59

Ajay: thing now it becomes a matter of how many resets you can get and how many how much

763

: 00:55:03

Ajay: tokens how many tokens you can get and how well you can refine the workflow, right?

764

: 00:55:07

Ajay: So this unintentionally became a very philosophical episode for us.

765

: 00:55:13

Ajay: I don't know, Andrew, do you feel like we should actually, I could talk a little

766

: 00:55:16

Ajay: more about the SDLC hop-on building, the home OS solutions, but I think what we just

767

: 00:55:20

Ajay: mentioned is also very relevant of like, where does potentially vibe coding continue

768

: 00:55:25

Ajay: to live on and be very relevant versus where does it potentially change and need to

769

: 00:55:30

Ajay: become more like a software development lifecycle?

770

: 00:55:32

Ajay: I think we've kind of answered that.

771

: 00:55:34

Andrew: Yeah.

772

: 00:55:34

Andrew: I think within the household, for your own usage, when the risk is low,

773

: 00:55:40

Andrew: you'll probably accept the risk if you made it just for you,

774

: 00:55:45

Andrew: right?

775

: 00:55:46

Andrew: Yeah.

776

: 00:55:47

Andrew: Then vibe coding is a real thing.

777

: 00:55:49

Andrew: The everyday person, vibe code, whatever they need, if it fails, they'll just tell

778

: 00:55:54

Andrew: Cloud to fix it for that exact instance, and they'll move on with their life.

779

: 00:55:57

Ajay: Worst case, it doesn't work, and it wasn't very valuable, and you move on to the

780

: 00:56:00

Ajay: next project.

781

: 00:56:01

Ajay: You try and solve the next thing.

782

: 00:56:03

Ajay: So, yeah.

783

: 00:56:03

Ajay: So, yeah, I guess it really does depend.

784

: 00:56:06

Ajay: Yeah.

785

: 00:56:07

Ajay: I mean, I

786

: 00:56:08

Ajay: definitely agree.

787

: 00:56:10

Ajay: And I definitely think that this age of the digital handyman

788

: 00:56:15

Ajay: is just starting.

789

: 00:56:19

Ajay: And I feel like folks are just starting, at least I feel like I'm starting to just

790

: 00:56:23

Ajay: work on it now.

791

: 00:56:24

Ajay: Because one thing I was going to bring up is when I did traditional software, when I

792

: 00:56:28

Ajay: used to build systems and software.

793

: 00:56:32

Ajay: Usually, I'd be building it based on a set of requirements that I would get from

794

: 00:56:35

Ajay: somebody else.

795

: 00:56:37

Ajay: They'd have done all the research.

796

: 00:56:38

Ajay: They'd figured out who's going to buy it and who's going to essentially spend money

797

: 00:56:41

Ajay: on it.

798

: 00:56:42

Ajay: And if somebody was wrong, it would usually be like the product manager who would

799

: 00:56:45

Ajay: say that, hey, we made a mistake.

800

: 00:56:46

Ajay: Or they would say that we're going to try again in a different way.

801

: 00:56:48

Ajay: And they'd figure out how to get things right.

802

: 00:56:51

Ajay: When we talk about this particular use case, now I'm the product manager.

803

: 00:56:56

Ajay: So that becomes very interesting to me because I'm like, I need to know what my

804

: 00:57:01

Ajay: needs are even better.

805

: 00:57:02

Ajay: Or I need to know what my household needs are or the people that I'm building this

806

: 00:57:05

Ajay: for, their needs when I'm vibe coding.

807

: 00:57:08

Ajay: So at times I've actually built things and then at the end I get a product where I'm

808

: 00:57:11

Ajay: just like, oh, I actually don't want this.

809

: 00:57:15

Ajay: It's just like it is making me do more work.

810

: 00:57:18

Ajay: And that is one of the more interesting things I've learned and I've experienced.

811

: 00:57:22

Ajay: I don't know what you think about that.

812

: 00:57:24

Andrew: no no no I I think I think what you said it's uh it's very timely because just like

813

: 00:57:30

Andrew: I was telling you earlier I was like you know what I'm kind of annoyed with cursor

814

: 00:57:35

Andrew: and bs code it's overweight you know it's based off like chromium right electron

815

: 00:57:40

Andrew: like it's just so bloated with with all this stuff that they put onto it yeah and I

816

: 00:57:46

Andrew: have like six seven windows of cursor open right it's kind of heavy weight for my

817

: 00:57:52

Andrew: needs right I'm not using like the cursor agent I'm not using like composer or

818

: 00:57:56

Andrew: anything like that yeah um so I'm like maybe I could just make my own solution and

819

: 00:58:01

Andrew: just trim it down to exactly what I need right and and I looked

820

: 00:58:07

Andrew: into it and I actually did two different okay windows I actually was

821

: 00:58:13

Andrew: comparing fable versus opus I gave the same prompt to each one and I was comparing

822

: 00:58:18

Andrew: the output I was like I because I wanted to see I wanted to see you like what's

823

: 00:58:22

Andrew: fable going to tell me what's opus is going to tell me same exact question prompt

824

: 00:58:27

Andrew: and they actually did give me different answers okay yes yeah but it was both

825

: 00:58:34

Andrew: the same conclusion how many surveys did it spin up thankfully not all right good

826

: 00:58:39

Andrew: good good good I don't think any actually I I I was very specific that we were in

827

: 00:58:45

Ajay: the ideation no deep research yes no deeper yeah oh okay but but I'm

828

: 00:58:50

Ajay: curious sorry sorry yes yes so so but it both went around in different different

829

: 00:58:56

Andrew: ways it said different technologies different solutions there was a general tone

830

: 00:59:01

Andrew: between them that was the same it's obviously been trained on similar yes

831

: 00:59:07

Andrew: information yeah um but uh the solution at the end wasn't from either of them

832

: 00:59:13

Andrew: oh from from reading both of their responses I realized that this was a

833

: 00:59:19

Andrew: much bigger ordeal than I had imagined and I was just going to stick with

834

: 00:59:25

Andrew: vs code or cursor uh there there's a lot more to it than just

835

: 00:59:31

Ajay: fable fix this for me yeah that's that's interesting andrew because I think we

836

: 00:59:37

Ajay: were just having this discussion about you got you've had a session of codex running

837

: 00:59:42

Ajay: for how many days now over a week over a week and how many prompts have you given

838

: 00:59:48

Ajay: a total or how many approximately do you remember giving well of directive prompts

839

: 00:59:54

Ajay: right okay yeah one wow all right yeah all the other prompts are me just asking it

840

: 00:59:59

Andrew: how you've been doing so so like a trip like a

841

: 01:00:04

Andrew: traditional software engineering manager yeah yeah I was like hey hey you know

842

: 01:00:09

Andrew: like how's it going what you got for me today do you have

843

: 01:00:15

Andrew: anything you know every every hour I'm like so what now

844

: 01:00:21

Ajay: so now you understand so

845

: 01:00:25

Ajay: you're certainly learning every role of the software engineering life cycle

846

: 01:00:32

Andrew: right now you know you gotta have every hat these days you gotta learn

847

: 01:00:37

Ajay: every but okay okay well on a more interesting note I I think uh we

848

: 01:00:43

Ajay: wanted to talk a little bit about the different coding assistants yeah in this case

849

: 01:00:47

Ajay: you've clearly done a lot more work seven straight days of it apparently yeah um I

850

: 01:00:52

Ajay: guess I think you we mentioned uh some differences between the harnesses uh some

851

: 01:00:57

Ajay: some experience aspects um I think you also we were talking before we started

852

: 01:01:01

Ajay: recording about like the experience of using codecs versus the experience of

853

: 01:01:07

Ajay: using clod in particular um the the need to get it to continue its work or

854

: 01:01:13

Andrew: yes yes so so in my experience so Claude Code cloud cli um no matter

855

: 01:01:19

Andrew: what memories I put in no matter what you know directives I put in uh

856

: 01:01:25

Andrew: whatever I put in cloud md at some point claude's gonna

857

: 01:01:30

Andrew: stop and prompt you for something it's like it's built in that it's like at

858

: 01:01:36

Andrew: some point it could be 24 48 hours it's gonna stop and say hey what do you think of

859

: 01:01:42

Ajay: this some some guardrail maybe some some some issue that or not some issue but some

860

: 01:01:48

Andrew: intentional uh check-in point yep and and I've run into it so many times

861

: 01:01:53

Andrew: that I've put enough moat around this that I think for the most part it's it's

862

: 01:01:59

Andrew: become more rare for example one of the solutions I had is if you do have a solution

863

: 01:02:04

Andrew: or you do have a problem right that you're gonna ask me about yeah just file it

864

: 01:02:10

Andrew: away as a github issue for my review for my inbox yeah right and and you

865

: 01:02:16

Ajay: take what you think is the reasonable solution in the meanwhile all right

866

: 01:02:21

Ajay: because I trust you and then how do you feel now that uh you've had codex and you've

867

: 01:02:26

Ajay: had to follow up with it codex doesn't ask for follow-ups

868

: 01:02:32

Andrew: you you so you're the one asking yeah I'm like whoa hold on I'm like I actually oh

869

: 01:02:37

Andrew: so so it's funny I actually told codex to pause after it closed the

870

: 01:02:43

Andrew: module I came back several hours later it didn't pause it's like

871

: 01:02:48

Andrew: I'm not done with the module what are you talking well it did finish it it went on

872

: 01:02:54

Andrew: and it was finishing the other one oh my gosh because I said I said close like like

873

: 01:02:59

Andrew: follow the closed protocol it's all laid out after module six is done wow and pause

874

: 01:03:04

Andrew: I use that and pause and came back it was already done with module seven and

875

: 01:03:10

Ajay: starting module seven reviews so so it just it just heard the word pause it's like I

876

: 01:03:15

Ajay: did for two seconds for two milliseconds I mean I have no idea that's way out of

877

: 01:03:20

Ajay: context by now no okay okay I I am um definitely very curious

878

: 01:03:25

Ajay: about this because it almost sounds like um it would be something that as a

879

: 01:03:31

Ajay: person building software I'd be like all right this is doing a little too much

880

: 01:03:37

Andrew: and as a caveat there is one tool that I think I think I know what happened okay

881

: 01:03:42

Andrew: yeah yeah so for codex in

882

: 01:03:45

Andrew: particular there is a feature called the goal ah and I believe that overrides

883

: 01:03:51

Andrew: everything I see yeah so so you do slash goal you write in the goal so the goal

884

: 01:03:56

Andrew: funny enough I didn't write it in codex put the goal in after I told it what I

885

: 01:04:02

Andrew: wanted the first time and then that goal is like a persistent immutable

886

: 01:04:07

Andrew: command before the session I'm assuming for the yeah until you stop it until

887

: 01:04:13

Ajay: until you do control c oh my god until you kill it from stop it will keep going

888

: 01:04:19

Ajay: you know until you interrupt it that that's that's actually really no that's really

889

: 01:04:24

Ajay: a catch by you because actually I wouldn't be surprised if that was very intentional

890

: 01:04:28

Ajay: um and that actually almost gives me more insight into the way that software

891

: 01:04:32

Ajay: engineering has been occurring within the companies uh because I think well we hear

892

: 01:04:38

Ajay: all these things from uh you know like anthropic internally talk about how they use

893

: 01:04:43

Ajay: their tools and they talk about how much code is being generated by quad code um and

894

: 01:04:49

Ajay: then open AI also talks about these things and they also talk about how much code is

895

: 01:04:52

Ajay: being generated by their by codecs internally and because these are all internal

896

: 01:04:56

Ajay: tools that became external and became products that they wanted to sell now it

897

: 01:05:01

Ajay: definitely tells me that engineers at open AI are essentially building

898

: 01:05:07

Ajay: features end-to-end and giving goals and saying that hey go ahead and just spin this

899

: 01:05:11

Ajay: up and do everything whereas it almost seems like the engineers and the anthropic

900

: 01:05:16

Ajay: are having more regular check-ins and are doing more regular like points of like

901

: 01:05:20

Ajay: let's validate more often because that's also I think we talked a little bit about

902

: 01:05:25

Ajay: the user instructions for Claude Code last time I that I was reading and I found some

903

: 01:05:29

Ajay: information they always recommend do research then do planning and then do execution

904

: 01:05:34

Ajay: um it almost seems like that's less of a criticality in the codex workflow yes

905

: 01:05:40

Andrew: and and kind of to our first episode I was talking about trusting the

906

: 01:05:46

Andrew: model yeah and I think there is a lot of trust in the model on codex aside on

907

: 01:05:51

Andrew: codex aside okay all right more so than than claude because obviously claude will

908

: 01:05:56

Andrew: defer to your human judgment all the time or even after repeated

909

: 01:06:02

Andrew: instructions not to do that codex doesn't ask codex will keep going this

910

: 01:06:07

Andrew: keeps going as long as you have the money and they seem to be giving you more more

911

: 01:06:13

Ajay: yes the free resets so this is going to be the really interesting thing I I I wonder

912

: 01:06:19

Ajay: if you feel that it's uh a strategy by open ai to get you to use

913

: 01:06:24

Ajay: more tokens or do you feel it's an actual experience thing that they have in mind

914

: 01:06:30

Ajay: in terms of saying that hey uh somebody gave me a goal my ideal user is somebody

915

: 01:06:35

Ajay: who's non-technical let me just go execute the goal and on the anthropic side on the

916

: 01:06:41

Ajay: Claude Code side their user is a enterprise individual whose risk

917

: 01:06:47

Ajay: profile is much much more like uh you know I don't want to take so many risks I want

918

: 01:06:50

Ajay: regular check-ins I want to be able to take a look at things more regularly um their

919

: 01:06:55

Ajay: their tokens are very expensive so they don't worry about like maximizing token use

920

: 01:06:59

Ajay: you know uh token maxing they don't worry about that um but instead their focus is

921

: 01:07:04

Ajay: to give you good results at the check-in more frequently because the user profile

922

: 01:07:08

Ajay: that they're targeting is somebody who wants those results more often.

923

: 01:07:12

Andrew: That's really interesting.

924

: 01:07:13

Andrew: And I think that ties into what we were talking about earlier.

925

: 01:07:15

Andrew: It kind of sounds like OpenAI Code Extract GPT is more targeting the vibe coder.

926

: 01:07:21

Andrew: Yeah.

927

: 01:07:22

Andrew: And Anthropic Claude Code is targeting the engineer.

928

: 01:07:26

Ajay: So this is very fascinating to me because I've been building all my HomeOS projects

929

: 01:07:30

Ajay: with Claude Code because that is something that I use professionally.

930

: 01:07:33

Ajay: I'm almost wondering if I should actually not be doing that.

931

: 01:07:36

Ajay: I almost want wondering if I should be using codecs to actually do my vibe.

932

: 01:07:40

Ajay: Right.

933

: 01:07:41

Ajay: Right.

934

: 01:07:42

Ajay: Because it's maybe it's actually going to one problem I have run into when I'm been

935

: 01:07:47

Ajay: building, when I have been building software with quad code is because it asked me

936

: 01:07:50

Ajay: to check in so often at some point I will miss a check-in and then I'll come back a

937

: 01:07:55

Ajay: few days later.

938

: 01:07:56

Ajay: Oh, and, and then all of a sudden it's not, it's still stuck in that last spot,

939

: 01:08:00

Ajay: which becomes challenging because of my project stagnates.

940

: 01:08:05

Ajay: my my vibe coder project becomes a something more like a uh I need to check in more

941

: 01:08:10

Ajay: often which becomes like work and becomes less like something but and every time it

942

: 01:08:14

Ajay: asks you you become desensitized to the severity yes question exactly and I think I

943

: 01:08:19

Ajay: I do agree because at some point I'm just like the risk level of this project is not

944

: 01:08:23

Ajay: that high just go do what you need to do um so all of this to ask a simple question

945

: 01:08:29

Ajay: who is the better digital handyman codex or Claude Code yeah at least

946

: 01:08:35

Ajay: at least amongst these two based on what you have what you have experienced I I

947

: 01:08:38

Andrew: would have to say codex oh there we go wow the better the

948

: 01:08:44

Andrew: best handyman is hands off well that's a that's a good yes

949

: 01:08:49

Ajay: that's a good way I I I agree I agree so this has become an interest I I feel like

950

: 01:08:55

Ajay: we've actually gotten some very interesting perspective here because yeah I do feel

951

: 01:08:59

Ajay: like um you know we we got into maybe a little bit of software engineering life

952

: 01:09:04

Ajay: cycles to start talking about like the LM wiki um and trying to improve our software

953

: 01:09:08

Ajay: engineering life cycle but then the question really became by the end was like do we

954

: 01:09:12

Ajay: need the software engineering life cycle for things that we do at home uh probably

955

: 01:09:16

Ajay: not right probably not not at home because the risk is low the risk is low um and I

956

: 01:09:21

Ajay: think this is where I would eventually like to also talk about the hardnesses of

957

: 01:09:26

Ajay: other tools because I do feel like at some point I want to try an anti-gravity uh

958

: 01:09:31

Ajay: clf yes I haven't tried that yet yeah me neither and I think I'm really curious

959

: 01:09:36

Ajay: because um I had asked you last time and asked you this question last time is do you

960

: 01:09:40

Ajay: feel like it's it's a mirage but it's I well do I'll ask it to you again and maybe

961

: 01:09:46

Andrew: the answer will be more nuanced this time so that's very interesting I and and

962

: 01:09:52

Andrew: I I think most people who are using these tools aren't deploying their

963

: 01:09:57

Andrew: apps in a way that they can answer that question besides for their, you know, home

964

: 01:10:02

Andrew: usage.

965

: 01:10:02

Andrew: Um, and, and I think, although I haven't still have

966

: 01:10:08

Andrew: not reached practical testing, unfortunately, I, I feel like there is a scope at

967

: 01:10:14

Andrew: which I could prove that it works if I needed to.

968

: 01:10:19

Andrew: All right.

969

: 01:10:20

Andrew: Right.

970

: 01:10:20

Andrew: Yeah.

971

: 01:10:21

Andrew: Um, and, and the reason why I say that is um there's a level of confidence that

972

: 01:10:27

Andrew: you get when you're comparing the different models together right

973

: 01:10:33

Andrew: yes so for example any work that codex does yeah I will have opus

974

: 01:10:38

Andrew: review it any work that opus does I'll have codex review it right and you know and I

975

: 01:10:44

Andrew: already know that these models are being trusted at high levels in the industry

976

: 01:10:49

Ajay: right people much smarter much more capable than I am maybe and and I think

977

: 01:10:55

Ajay: the I I can give a little perspective and say I think the work

978

: 01:11:01

Ajay: that uh codex is executing it definitely feels as as we said

979

: 01:11:07

Ajay: a little more digital handyman so far uh the work that claude is executing is

980

: 01:11:12

Ajay: essentially plugging into

981

: 01:11:15

Ajay: workflows on like if we're software engineers yes so it's essentially meeting the uh

982

: 01:11:20

Andrew: you know like the knowledge worker where they are today I I I would actually I I

983

: 01:11:26

Andrew: want to actually say an

984

: 01:11:28

Andrew: observation I've seen between the two harnesses absolutely codex is a lot more

985

: 01:11:34

Andrew: direct okay by default all right yeah yeah cloud is a lot more friendly

986

: 01:11:40

Andrew: okay yeah like like it's more human interactable it's more human okay but codex

987

: 01:11:46

Andrew: is much more straight to the point and also when I ask codex to give instructions or

988

: 01:11:52

Andrew: um like like for example I'll be like I want you to give instructions to opus yeah

989

: 01:11:57

Andrew: to take over this task because you know what codex is at its limit or something

990

: 01:12:02

Andrew: codex will always give a very detailed briefing that always

991

: 01:12:08

Andrew: impresses me that it tells it will tell opus what files to look

992

: 01:12:14

Andrew: at what memories to check ah what modules and files it's

993

: 01:12:20

Andrew: touched what problems it's encountered wow there is a lot of detail like I didn't

994

: 01:12:26

Andrew: say to do any of that all I said was give detailed instructions to opus to continue

995

: 01:12:31

Andrew: this task and it will go through and and put in things that I didn't really

996

: 01:12:37

Andrew: think of that I have not seen opus do for the opposite and I

997

: 01:12:43

Andrew: think so what I'm thinking here is I guess what I'm getting at is I

998

: 01:12:48

Andrew: think in the system prompt for codex there is some software

999

: 01:12:54

Andrew: development lifecycle stuff baked in got it and that is what makes the vibe

: 1000

01:12:59,750 --> 01:13:05,340

Andrew: coding more it's an opinion it's an opinionated harness yes or or maybe model it

: 1001

01:13:05,370 --> 01:13:10,650

Ajay: might be an opinionated model maybe a little bit of both maybe a little bit of both

: 1002

01:13:10,660 --> 01:13:16,650

Andrew: yeah and so so those values yeah are incidentally being put into the work that

: 1003

01:13:16,890 --> 01:13:22,140

Ajay: codex does god which is okay that and I I really like the characterization there

: 1004

01:13:22,640 --> 01:13:28,420

Ajay: because interestingly it almost feels like clod code is or anthropic the way

: 1005

01:13:28,510 --> 01:13:34,090

Ajay: they've developed things is intended to meet uh the human operator

: 1006

01:13:34,370 --> 01:13:40,370

Ajay: where they are today yes whereas maybe codex and chat gpt um open ai in general

: 1007

01:13:40,390 --> 01:13:46,040

Ajay: are trying to provide an opinionated solution of where they want everybody to be in

: 1008

01:13:46,040 --> 01:13:50,560

Andrew: in like I think in this future I think that's a really good way to put it in the

: 1009

01:13:50,560 --> 01:13:54,820

Andrew: future that they're envisioning and you know what that kind of that kind of fits sam

: 1010

01:13:55,990 --> 01:14:01,780

Andrew: yes without

: 1011

01:14:02,100 --> 01:14:05,740

Andrew: getting without getting too deep or too much even there um we will leave that topic

: 1012

01:14:05,820 --> 01:14:11,600

Ajay: for a different day um I I will uh maybe close with a question and a statement

: 1013

01:14:11,690 --> 01:14:16,720

Ajay: and then a question okay um so I I think the mirage might be

: 1014

01:14:16,720 --> 01:14:22,470

Ajay: becoming more real right now um the question maybe it's looking like more of an

: 1015

01:14:22,560 --> 01:14:28,060

Ajay: oasis well look at looking like more of an oasis so so uh if I may ask ender do you

: 1016

01:14:28,060 --> 01:14:33,840

Ajay: feel like at least today um it doesn't feel like an act if it is it's

: 1017

01:14:33,840 --> 01:14:39,810

Ajay: scarily good wow wow all right I think I think

: 1018

01:14:39,810 --> 01:14:45,450

Ajay: I can't uh quite say anything more than that awesome

: 1019

01:14:45,500 --> 01:14:50,910

Ajay: so um yeah on a sober note I guess wow we

: 1020

01:14:51,220 --> 01:14:56,860

Ajay: will uh we will be back next time with more questions we'll be back next time

: 1021

01:14:56,860 --> 01:14:58,960

Ajay: awesome