1
00:00:00,080 --> 00:00:00,587
you might think

2
00:00:00,587 --> 00:00:01,360
is just internal tooling,

3
00:00:01,360 --> 00:00:02,400
but we have a lot of

4
00:00:02,400 --> 00:00:03,840
internal tooling, just so you know.

5
00:00:03,840 --> 00:00:04,960
which is why Effect is

6
00:00:04,960 --> 00:00:06,560
actually being used a lot.

7
00:00:06,560 --> 00:00:08,400
So, instead of running a Python eval

8
00:00:08,400 --> 00:00:10,703
suite on the market,

9
00:00:10,703 --> 00:00:12,000
there's a bunch, we write our own, but it's

10
00:00:12,000 --> 00:00:12,720
using Effect.

11
00:00:12,720 --> 00:00:15,351
So, we actually have an eval suite

12
00:00:15,351 --> 00:00:16,880
and also a bulk test suite,

13
00:00:16,880 --> 00:00:19,600
that's all using Effect with Effect retry

14
00:00:19,600 --> 00:00:20,640
and so on and so forth.

15
00:00:20,640 --> 00:00:21,120
it's a very

16
00:00:21,120 --> 00:00:21,840
sophisticated system.

17
00:00:21,840 --> 00:00:23,920
I'm just saying is that if we use,

18
00:00:23,920 --> 00:00:24,640
you might think internal

19
00:00:24,640 --> 00:00:25,600
tooling is like some toy,

20
00:00:25,600 --> 00:00:26,560
but no, we're actually using

21
00:00:26,560 --> 00:00:27,840
it almost on a daily basis,

22
00:00:27,840 --> 00:00:29,040
it's actually production load.

23
00:00:29,040 --> 00:00:29,680
testing out all

24
00:00:29,680 --> 00:00:30,800
the endpoint, right?

25
00:00:30,800 --> 00:00:33,040
Which is a crucial part of our operation.

26
00:00:33,040 --> 00:00:34,800
well, I
mean, almost every respond

27
00:00:34,800 --> 00:00:36,960
transformation is using the pipe now.

28
00:00:36,960 --> 00:00:38,080
And then also I'm

29
00:00:38,080 --> 00:00:40,480
introducing more and more generator.

30
00:00:40,480 --> 00:00:42,720
So, down the line, we finally will,

31
00:00:42,720 --> 00:00:43,360
it will be making it

32
00:00:43,360 --> 00:00:44,400
easier for us to transitioning

33
00:00:44,400 --> 00:00:49,760
over to Effect as needed.

34
00:00:49,760 --> 00:00:52,160
Welcome to Cause & Effect a podcast

35
00:00:52,160 --> 00:00:53,120
about the TypeScript

36
00:00:53,120 --> 00:00:55,600
library and ecosystem called Effect,

37
00:00:55,600 --> 00:00:57,120
helping engineers to build

38
00:00:57,120 --> 00:00:58,080
production-ready software.

39
00:00:58,880 --> 00:01:00,480
I'm your host, Johannes Schickling,

40
00:01:00,480 --> 00:01:01,520
and I've been building with

41
00:01:01,520 --> 00:01:03,360
Effect for over four years.

42
00:01:03,360 --> 00:01:04,800
With this podcast, I want

43
00:01:04,800 --> 00:01:06,160
to help others understand

44
00:01:06,160 --> 00:01:09,680
the powers and benefits of using Effect.

45
00:01:09,680 --> 00:01:10,960
In this episode, I'm

46
00:01:10,960 --> 00:01:12,560
talking to Louis Vichy,

47
00:01:12,560 --> 00:01:14,320
co-founder of OpenRouter,

48
00:01:14,320 --> 00:01:17,120
the unified interface for LLMs.

49
00:01:17,120 --> 00:01:19,200
In this conversation, we dive deep

50
00:01:19,200 --> 00:01:21,840
into the Cloudflare workers architecture,

51
00:01:21,840 --> 00:01:23,920
their routing layers, and how they're

52
00:01:23,920 --> 00:01:25,520
replacing their own result type

53
00:01:25,520 --> 00:01:27,360
implementation with Effect

54
00:01:27,360 --> 00:01:28,800
for better retries,

55
00:01:28,800 --> 00:01:30,880
observability, and long-running

56
00:01:30,880 --> 00:01:32,240
agentic workflows.

57
00:01:32,240 --> 00:01:34,080
Let's get into it.

58
00:01:34,080 --> 00:01:35,520
Hey, welcome everyone.

59
00:01:35,520 --> 00:01:37,200
It's so nice to have you here.

60
00:01:37,200 --> 00:01:39,120
Louis, you're from OpenRouter.

61
00:01:39,120 --> 00:01:40,720
Would you mind introducing yourself?

62
00:01:40,720 --> 00:01:42,240
We also have the Effect

63
00:01:42,240 --> 00:01:44,480
co-creator Michael here.

64
00:01:44,480 --> 00:01:46,160
Awesome. How are you all?

65
00:01:46,160 --> 00:01:48,000
I'm Luis. I'm a co-founder of OpenRouter.

66
00:01:48,000 --> 00:01:49,600
Me and my co-founder, we met

67
00:01:49,600 --> 00:01:51,200
through GitHub essentially.

68
00:01:51,200 --> 00:01:52,720
He was building out a V0 of

69
00:01:52,720 --> 00:01:54,640
OpenRouter using a framework I was

70
00:01:54,640 --> 00:01:55,600
building for my previous

71
00:01:55,600 --> 00:01:57,440
startup and that's how we met.

72
00:01:57,440 --> 00:01:58,160
Then we just love

73
00:01:58,160 --> 00:02:00,000
collaborating with each other and then

74
00:02:00,000 --> 00:02:00,800
around May is when

75
00:02:00,800 --> 00:02:02,960
OpenRouter first launched May 2023.

76
00:02:02,960 --> 00:02:04,640
So early 2023.

77
00:02:04,640 --> 00:02:07,840
Essentially, the idea of OpenRouter is

78
00:02:07,840 --> 00:02:08,640
at the bare bone,

79
00:02:08,640 --> 00:02:09,360
it is actually a

80
00:02:09,360 --> 00:02:10,960
switchboard for AI models.

81
00:02:10,960 --> 00:02:13,280
That's it. We reduce the switching cost

82
00:02:13,280 --> 00:02:14,480
between like Anthropic

83
00:02:14,480 --> 00:02:16,000
and OpenAI and Gemini,

84
00:02:16,000 --> 00:02:17,840
and Moonshot, Cohere, and

85
00:02:17,840 --> 00:02:19,760
Meta to essentially zero.

86
00:02:20,800 --> 00:02:21,840
We're also centralizing

87
00:02:21,840 --> 00:02:23,280
all the payment for you.

88
00:02:23,280 --> 00:02:25,040
Instead of like 60 different providers,

89
00:02:25,040 --> 00:02:26,400
you pay just us and

90
00:02:26,400 --> 00:02:27,520
then you get some credit,

91
00:02:27,520 --> 00:02:28,720
and then you can call

92
00:02:28,720 --> 00:02:30,560
any AI model you want.

93
00:02:31,200 --> 00:02:32,640
Essentially, that is the

94
00:02:33,200 --> 00:02:34,400
basic part of OpenRouter,

95
00:02:34,400 --> 00:02:35,120
but OpenRouter is

96
00:02:35,120 --> 00:02:36,160
actually much more than that.

97
00:02:36,720 --> 00:02:39,760
You have seen Elon tweet about us on X.

98
00:02:39,760 --> 00:02:40,960
the secondary product of

99
00:02:40,960 --> 00:02:42,640
OpenRouter is the data that

100
00:02:42,640 --> 00:02:45,600
we are allowing everyone else to access.

101
00:02:45,600 --> 00:02:46,880
The data about which model

102
00:02:46,880 --> 00:02:48,320
is being spent on the most,

103
00:02:48,320 --> 00:02:49,280
or which model is

104
00:02:49,280 --> 00:02:51,600
getting the most utility.

105
00:02:51,600 --> 00:02:53,120
We measure them based on tokens.

106
00:02:53,120 --> 00:02:57,120
A token is about roughly full characters.

107
00:02:57,120 --> 00:02:58,080
Last week, we did about

108
00:02:58,080 --> 00:03:00,320
five trillion tokens a week.

109
00:03:00,320 --> 00:03:01,920
Exactly a year ago though,

110
00:03:01,920 --> 00:03:03,920
we were doing about 150

111
00:03:03,920 --> 00:03:05,440
billion tokens a week.

112
00:03:05,440 --> 00:03:06,320
That's roughly about

113
00:03:06,320 --> 00:03:07,600
more than 20x growth.

114
00:03:09,600 --> 00:03:12,400
By the way, this feature was released

115
00:03:12,400 --> 00:03:14,000
literally the first month

116
00:03:14,000 --> 00:03:15,760
of the product, I think.

117
00:03:15,760 --> 00:03:16,480
Basically, my co-founder

118
00:03:16,480 --> 00:03:17,600
and I were thinking about,

119
00:03:17,600 --> 00:03:18,560
"Okay, we have this data.

120
00:03:18,560 --> 00:03:20,720
Should we just let everyone see it?"

121
00:03:20,720 --> 00:03:23,360
Because at the time,

122
00:03:23,360 --> 00:03:24,880
there were so many benchmarks.

123
00:03:24,880 --> 00:03:25,760
There were so many

124
00:03:25,760 --> 00:03:27,760
Llama or Meta Llama benchmark,

125
00:03:27,760 --> 00:03:28,800
Llama one, right?

126
00:03:28,800 --> 00:03:29,680
They're claiming some

127
00:03:29,680 --> 00:03:30,720
of these Llama fighting

128
00:03:30,720 --> 00:03:32,000
will be better than GPT 3.5.

129
00:03:32,000 --> 00:03:33,120
And we were like, "Is

130
00:03:33,120 --> 00:03:35,360
that actually the case?"

131
00:03:35,360 --> 00:03:37,040
Because when we were trying them locally,

132
00:03:37,040 --> 00:03:38,080
they're dumb as hell.

133
00:03:38,080 --> 00:03:39,680
So we were like, "Okay,

134
00:03:39,680 --> 00:03:40,960
let's make a benchmark

135
00:03:40,960 --> 00:03:42,480
that's based on real

136
00:03:42,480 --> 00:03:44,800
utilization, real usage, real money,

137
00:03:44,800 --> 00:03:45,680
being bet on the

138
00:03:45,680 --> 00:03:47,760
model, like a horse race.

139
00:03:47,760 --> 00:03:48,800
And see which one get

140
00:03:48,800 --> 00:03:50,720
the most of this money."

141
00:03:50,720 --> 00:03:51,760
And that's essentially

142
00:03:51,760 --> 00:03:53,200
is the secondary product,

143
00:03:53,200 --> 00:03:53,920
which is the ranking

144
00:03:53,920 --> 00:03:55,600
page where everyone can see

145
00:03:55,600 --> 00:03:57,600
which model is being spent the most on.

146
00:03:57,600 --> 00:03:58,720
And we do cap the amount

147
00:03:58,720 --> 00:04:00,240
of free usage on a provider.

148
00:04:00,240 --> 00:04:02,720
So all of that is actual usage.

149
00:04:02,720 --> 00:04:04,240
Like people spending money.

150
00:04:04,240 --> 00:04:05,040
That is awesome.

151
00:04:05,040 --> 00:04:06,560
So you mentioned you met

152
00:04:06,560 --> 00:04:08,400
your co-founder in 2023,

153
00:04:08,400 --> 00:04:09,920
May 2023.

154
00:04:09,920 --> 00:04:12,080
Now more than two years later,

155
00:04:12,080 --> 00:04:14,880
where are you at in terms of a company?

156
00:04:14,880 --> 00:04:16,480
Is it just the two of you?

157
00:04:16,480 --> 00:04:17,440
Probably not.

158
00:04:17,440 --> 00:04:19,200
So maybe you can shed some light

159
00:04:19,200 --> 00:04:21,120
on like OpenRouter as a company.

160
00:04:21,120 --> 00:04:22,880
So we were two people

161
00:04:22,880 --> 00:04:24,160
writing on the thing.

162
00:04:24,160 --> 00:04:25,360
He was checking Datadog.

163
00:04:25,360 --> 00:04:27,920
I was just writing

164
00:04:27,920 --> 00:04:30,960
code for up until January.

165
00:04:30,960 --> 00:04:31,760
And then we have our

166
00:04:31,760 --> 00:04:32,640
first engineering hire

167
00:04:32,640 --> 00:04:33,920
who also joined us

168
00:04:33,920 --> 00:04:36,000
through an open source project.

169
00:04:36,000 --> 00:04:37,760
We released a couple of open source

170
00:04:37,760 --> 00:04:38,880
project at the time,

171
00:04:38,880 --> 00:04:40,480
early 2024.

172
00:04:40,480 --> 00:04:42,240
And then Sam joined us

173
00:04:42,240 --> 00:04:43,840
through he sent us a PR

174
00:04:43,840 --> 00:04:45,120
on another project.

175
00:04:45,120 --> 00:04:47,120
and then we still have him to this day.

176
00:04:47,120 --> 00:04:48,400
last year we have four people.

177
00:04:48,400 --> 00:04:49,840
Now we have about 22.

178
00:04:49,840 --> 00:04:51,360
So team grow about 5x,

179
00:04:51,360 --> 00:04:52,160
but the utilization

180
00:04:52,160 --> 00:04:53,280
has grew more than 20x.

181
00:04:53,280 --> 00:04:55,920
So I think we have a decent wiggle room.

182
00:04:55,920 --> 00:04:57,200
We make enough of coffee, right?

183
00:04:57,200 --> 00:04:58,720
So got to buy more coffee.

184
00:04:58,720 --> 00:04:59,520
essentially the

185
00:04:59,520 --> 00:05:01,040
growth curve is exactly the same.

186
00:05:01,040 --> 00:05:02,080
The first year we went

187
00:05:02,080 --> 00:05:04,240
from zero to 7 billion.

188
00:05:04,240 --> 00:05:06,960
7 billion token, you know, in a week.

189
00:05:06,960 --> 00:05:08,880
And that's when I showed my dad.

190
00:05:08,880 --> 00:05:11,920
And then a year later, September 2024,

191
00:05:11,920 --> 00:05:13,920
we went to about 150 billion.

192
00:05:13,920 --> 00:05:14,720
And that's when I

193
00:05:14,720 --> 00:05:16,320
show, you know, my friend

194
00:05:16,320 --> 00:05:18,240
is my south friend.

195
00:05:18,240 --> 00:05:20,200
And now, right, I'm showing you guys

196
00:05:20,200 --> 00:05:22,880
and a lot of people we're interviewing,

197
00:05:22,880 --> 00:05:24,160
we have five trillion.

198
00:05:24,160 --> 00:05:25,840
So it's been 20x and

199
00:05:25,840 --> 00:05:27,360
then more than 20x.

200
00:05:27,360 --> 00:05:29,760
Hopefully the next year we'll be higher,

201
00:05:29,760 --> 00:05:30,880
but we'll see, right?

202
00:05:30,880 --> 00:05:34,080
Well, AI usage is only gonna increase.

203
00:05:34,080 --> 00:05:37,040
And as far as I know,

204
00:05:37,040 --> 00:05:39,520
from friends and so on and so forth,

205
00:05:39,520 --> 00:05:42,480
OpenRouter is probably the best way

206
00:05:42,480 --> 00:05:44,720
to access those different models.

207
00:05:44,720 --> 00:05:45,840
We had internal

208
00:05:45,840 --> 00:05:49,120
conversations and one key aspect

209
00:05:49,120 --> 00:05:51,440
that I truly like about what you do

210
00:05:51,440 --> 00:05:52,320
is that you're not

211
00:05:52,320 --> 00:05:54,240
just a proxy to models.

212
00:05:54,240 --> 00:05:54,960
You actually

213
00:05:54,960 --> 00:05:57,360
consolidate the model responses,

214
00:05:57,360 --> 00:05:59,360
which is key here.

215
00:05:59,360 --> 00:06:00,800
So a lot of people actually

216
00:06:00,800 --> 00:06:02,720
mistake us as a wrapper, right?

217
00:06:02,720 --> 00:06:04,960
Building a wrapper is very easy,

218
00:06:04,960 --> 00:06:07,760
but building a dev tool is hard.

219
00:06:07,760 --> 00:06:09,040
Building a dev tool on top

220
00:06:09,040 --> 00:06:11,120
of the API is extremely hard,

221
00:06:11,120 --> 00:06:11,840
right, because you have

222
00:06:11,840 --> 00:06:13,280
to ensure that the Dev X,

223
00:06:13,280 --> 00:06:15,120
the user experience of the

224
00:06:15,120 --> 00:06:17,440
developer is actually good

225
00:06:17,440 --> 00:06:19,120
so that they can stick to us.

226
00:06:19,120 --> 00:06:20,960
And this is the key kind of mistake

227
00:06:20,960 --> 00:06:22,160
that people didn't think about

228
00:06:22,160 --> 00:06:23,680
until they need us, right?

229
00:06:23,680 --> 00:06:24,880
We have seen this a lot, by the way,

230
00:06:24,880 --> 00:06:25,680
is that user

231
00:06:25,680 --> 00:06:26,760
essentially didn't think too much

232
00:06:26,760 --> 00:06:27,440
of OpenRouter, they

233
00:06:27,440 --> 00:06:28,240
try it once, they're like,

234
00:06:28,240 --> 00:06:29,760
"Okay, I'm gonna use two or three model,

235
00:06:29,760 --> 00:06:30,640
"it doesn't matter, so

236
00:06:30,640 --> 00:06:32,000
I'm gonna do my own."

237
00:06:32,000 --> 00:06:32,640
A lot of people have

238
00:06:32,640 --> 00:06:34,240
built their own OpenRouters.

239
00:06:34,240 --> 00:06:36,960
And then the moment they have to say,

240
00:06:36,960 --> 00:06:37,920
"Okay, now I have to

241
00:06:37,920 --> 00:06:39,440
add Anthropic, you know,

242
00:06:39,440 --> 00:06:42,080
Claude 4.5 came out,

243
00:06:42,080 --> 00:06:42,960
they have to scramble

244
00:06:42,960 --> 00:06:44,720
their entire weekends, right?

245
00:06:44,720 --> 00:06:45,680
Either weekends or

246
00:06:45,680 --> 00:06:47,200
they're Tuesday and Wednesday

247
00:06:47,200 --> 00:06:48,640
to add a model.

248
00:06:48,640 --> 00:06:49,680
With OpenRouter,

249
00:06:49,680 --> 00:06:51,040
it's right there, right?

250
00:06:51,040 --> 00:06:52,080
You add the same

251
00:06:52,080 --> 00:06:53,600
code, you allow your user

252
00:06:53,600 --> 00:06:56,160
to switch up model name, you're done.

253
00:06:56,160 --> 00:06:58,080
And that essentially has been a case

254
00:06:58,080 --> 00:06:59,280
of how OpenRouter has been growing,

255
00:06:59,280 --> 00:07:01,040
is that every time there's a new model,

256
00:07:01,040 --> 00:07:02,080
we do partner with this

257
00:07:02,080 --> 00:07:04,160
model to launch them early,

258
00:07:04,160 --> 00:07:06,240
and then the moment the model is launched

259
00:07:06,240 --> 00:07:07,440
is outright on OpenRouter

260
00:07:08,080 --> 00:07:09,840
and people who miss the

261
00:07:09,840 --> 00:07:11,200
train, right, come back to us,

262
00:07:11,200 --> 00:07:12,400
and need people coming in

263
00:07:12,400 --> 00:07:14,240
because they also want to try the model.

264
00:07:14,240 --> 00:07:14,880
That's kind of

265
00:07:14,880 --> 00:07:16,560
essentially the growth engine

266
00:07:16,560 --> 00:07:18,240
we have had thus far, which

267
00:07:18,240 --> 00:07:20,240
is leveraging our partnership

268
00:07:20,240 --> 00:07:22,800
with all this new model coming in,

269
00:07:22,800 --> 00:07:24,560
Can you maybe categorize

270
00:07:24,560 --> 00:07:26,400
sort of like the products

271
00:07:26,400 --> 00:07:28,240
you had to build out over the years,

272
00:07:28,240 --> 00:07:29,040
and also maybe

273
00:07:29,040 --> 00:07:31,440
orthogonally the challenges

274
00:07:31,440 --> 00:07:33,040
that maybe you hit right away

275
00:07:33,040 --> 00:07:34,960
or that just appeared

276
00:07:34,960 --> 00:07:36,320
over time, I suppose,

277
00:07:36,320 --> 00:07:38,000
like working at that scale,

278
00:07:38,000 --> 00:07:41,040
just like pulling on that scale thread,

279
00:07:41,040 --> 00:07:42,320
there's probably like so

280
00:07:42,320 --> 00:07:43,840
many interesting challenges

281
00:07:43,840 --> 00:07:45,280
that are coming towards us.

282
00:07:45,280 --> 00:07:45,920
So the one thing that

283
00:07:45,920 --> 00:07:47,280
me and my co-founder really like

284
00:07:47,280 --> 00:07:49,680
is to have type safety, to be honest.

285
00:07:49,680 --> 00:07:52,000
So, I kid you not, like the

286
00:07:52,000 --> 00:07:53,520
initial version of OpenRouter

287
00:07:53,520 --> 00:07:55,920
all the models, all the endpoints,

288
00:07:55,920 --> 00:07:57,680
were TypeScript object,

289
00:07:57,680 --> 00:07:58,880
TypeScript literal

290
00:07:58,880 --> 00:08:01,120
type object in the code.

291
00:08:01,120 --> 00:08:03,120
And then the way that we do reference,

292
00:08:03,120 --> 00:08:05,360
we would do strict type satisfy.

293
00:08:05,360 --> 00:08:07,120
As a way to do literally

294
00:08:07,120 --> 00:08:08,880
like foreign key reference

295
00:08:08,880 --> 00:08:09,920
in the code.

296
00:08:09,920 --> 00:08:11,760
So our code were a huge

297
00:08:11,760 --> 00:08:13,200
database of these endpoints

298
00:08:13,200 --> 00:08:14,160
it's not crazy.

299
00:08:14,160 --> 00:08:16,480
The moment we add a new field, right,

300
00:08:16,480 --> 00:08:17,920
an alert, right, because

301
00:08:17,920 --> 00:08:19,920
of TypeScript and satisfy,

302
00:08:19,920 --> 00:08:21,760
an alert across all our code base saying,

303
00:08:21,760 --> 00:08:22,480
hey, you have to add this

304
00:08:22,480 --> 00:08:23,520
field in every, you know,

305
00:08:23,520 --> 00:08:24,160
this place.

306
00:08:24,160 --> 00:08:25,440
And then at the time,

307
00:08:25,440 --> 00:08:26,720
right, when we add new model,

308
00:08:26,720 --> 00:08:27,680
we add new endpoint, we

309
00:08:27,680 --> 00:08:28,880
just tell Devin to do it.

310
00:08:28,880 --> 00:08:31,280
It's automated because of type,

311
00:08:31,280 --> 00:08:32,640
because of type safety, right?

312
00:08:32,640 --> 00:08:34,320
The moment we just want to add something,

313
00:08:34,320 --> 00:08:37,040
we tell the model, it changed something,

314
00:08:37,040 --> 00:08:37,920
and then the rest is

315
00:08:37,920 --> 00:08:40,400
just TypeScript, you know,
at type error.

316
00:08:40,400 --> 00:08:42,320
So it was not that crazy.

317
00:08:42,320 --> 00:08:44,000
it was actually usable.

318
00:08:44,000 --> 00:08:45,760
even satisfying a bunch of, you know,

319
00:08:45,760 --> 00:08:47,600
like literal number constraint as well,

320
00:08:47,600 --> 00:08:49,680
which is very fun.

321
00:08:49,680 --> 00:08:50,720
I mean, you know, fun

322
00:08:50,720 --> 00:08:52,080
doesn't scale, right?

323
00:08:52,080 --> 00:08:54,000
So we'll talk a lot more

324
00:08:54,000 --> 00:08:55,120
about the data in a moment,

325
00:08:55,120 --> 00:08:56,480
but is this entire system

326
00:08:56,480 --> 00:08:58,240
still written all in TypeScript

327
00:08:58,240 --> 00:09:00,800
or have you moved some parts

328
00:09:00,800 --> 00:09:03,440
that are maybe like critical

329
00:09:03,440 --> 00:09:04,800
to be more efficient?

330
00:09:04,800 --> 00:09:06,800
Are those written in Go and Rust

331
00:09:06,800 --> 00:09:08,720
or is everything still in TypeScript?

332
00:09:08,720 --> 00:09:10,240
Nope, everything.

333
00:09:10,240 --> 00:09:12,291
Everything's still in TypeScript.

334
00:09:12,291 --> 00:09:12,880
We very much just TypeScript all in.

335
00:09:12,880 --> 00:09:13,520
I mean, the main

336
00:09:13,520 --> 00:09:15,760
reason why I think is fine

337
00:09:15,760 --> 00:09:17,120
for OpenRouter to be strictly

338
00:09:17,120 --> 00:09:18,320
TypeScript is because,

339
00:09:18,880 --> 00:09:20,000
I mean, that won't lie, right?

340
00:09:20,000 --> 00:09:21,280
When we actually hit the bottleneck,

341
00:09:21,280 --> 00:09:23,520
we sure will switch to Go and so on,

342
00:09:23,520 --> 00:09:26,400
but right now we're running on top of,

343
00:09:26,400 --> 00:09:28,160
you know, Cloudflare.

344
00:09:28,160 --> 00:09:30,000
And essentially, the

345
00:09:30,000 --> 00:09:32,240
Cloudflare worker is much, you know,

346
00:09:32,240 --> 00:09:35,280
better to use TypeScript with.

347
00:09:35,280 --> 00:09:37,360
The tight strictness of TypeScript

348
00:09:37,360 --> 00:09:39,040
and also the flexible, the

349
00:09:39,040 --> 00:09:40,080
dynamic nature of it, right?

350
00:09:40,080 --> 00:09:41,760
Allow us to move much faster.

351
00:09:41,760 --> 00:09:43,280
The velocity of the engineers is much

352
00:09:43,280 --> 00:09:44,320
faster, I would argue.

353
00:09:44,320 --> 00:09:46,080
Technically Go can do the same.

354
00:09:46,080 --> 00:09:47,440
It's just that we start with TypeScript

355
00:09:47,440 --> 00:09:48,880
and we use TypeScript almost everywhere

356
00:09:48,880 --> 00:09:50,960
on both front end and back end, right?

357
00:09:50,960 --> 00:09:53,600
So I think it's best for a

358
00:09:53,600 --> 00:09:55,680
team to have the same language.

359
00:09:55,680 --> 00:09:57,760
I mean, time and time again,
I mean, time and time again,

360
00:09:57,760 --> 00:09:58,960
the same story sort

361
00:09:58,960 --> 00:10:00,720
of like repeats, right?

362
00:10:00,720 --> 00:10:01,360
Yep.

363
00:10:01,360 --> 00:10:03,440
Like if you're looking at Facebook

364
00:10:03,440 --> 00:10:05,040
and Mark having written

365
00:10:05,040 --> 00:10:06,480
the first Facebook prototype

366
00:10:06,480 --> 00:10:08,480
in PHP, guess what?

367
00:10:08,480 --> 00:10:10,080
Still mostly in PHP.

368
00:10:10,080 --> 00:10:12,240
Yeah.
The same for like--

369
00:10:12,240 --> 00:10:13,120
Instead of switching

370
00:10:13,120 --> 00:10:14,880
from PHP, they ended up

371
00:10:14,880 --> 00:10:16,560
making a new kind of

372
00:10:16,560 --> 00:10:17,280
Built the hack.

373
00:10:17,280 --> 00:10:18,000
Yeah, exactly.

374
00:10:18,000 --> 00:10:19,120
Or it's like, I mean, that's

375
00:10:19,120 --> 00:10:20,480
how TypeScript came to be.

376
00:10:20,480 --> 00:10:21,840
It's like, I would say

377
00:10:21,840 --> 00:10:23,120
the rationale of just,

378
00:10:23,120 --> 00:10:24,000
the whole idea, right,

379
00:10:24,000 --> 00:10:25,200
is that let's make sure

380
00:10:25,200 --> 00:10:26,560
we just speak the same language.

381
00:10:26,560 --> 00:10:28,240
So, in our code base,

382
00:10:28,240 --> 00:10:29,360
everything is TypeScript.

383
00:10:29,360 --> 00:10:31,120
Even the scripting is TypeScript.

384
00:10:31,120 --> 00:10:31,840
No bash.

385
00:10:31,840 --> 00:10:32,640
We don't use bash.

386
00:10:32,640 --> 00:10:36,480
We use ZX and TS up to write script

387
00:10:36,480 --> 00:10:38,800
to run like database migration or

388
00:10:38,800 --> 00:10:39,520
something like that.

389
00:10:39,520 --> 00:10:41,120
Or like even instant response scripts or

390
00:10:41,120 --> 00:10:42,000
something like that.

391
00:10:42,000 --> 00:10:43,840
So that every time that we communicate

392
00:10:43,840 --> 00:10:45,600
is in the same language.

393
00:10:45,600 --> 00:10:47,120
And so that someone

394
00:10:47,120 --> 00:10:48,400
else can pick it up, right?

395
00:10:48,400 --> 00:10:50,320
And we never really have dead code

396
00:10:50,320 --> 00:10:51,120
because TypeScript is

397
00:10:51,120 --> 00:10:53,120
very easy to kind of pull

398
00:10:53,120 --> 00:10:54,640
into a shared package

399
00:10:54,640 --> 00:10:55,920
and then you can reuse it.

400
00:10:55,920 --> 00:10:57,680
Honestly, I'm kind of surprised

401
00:10:57,680 --> 00:10:59,440
that this is not more common.

402
00:10:59,440 --> 00:11:01,920
In the Ruby on Rails world, for example,

403
00:11:01,920 --> 00:11:03,200
it's like so normal.

404
00:11:03,200 --> 00:11:04,240
Like obviously you

405
00:11:04,240 --> 00:11:05,360
build all of your scripts

406
00:11:05,360 --> 00:11:07,440
as Ruby scripts and then you run them.

407
00:11:07,440 --> 00:11:09,440
it's so surprising to

408
00:11:09,440 --> 00:11:12,160
me that in a JavaScript repo

409
00:11:12,160 --> 00:11:13,600
where like arguably

410
00:11:13,600 --> 00:11:15,680
it's as nice, if not nicer,

411
00:11:15,680 --> 00:11:17,440
to write scripts in JavaScript.

412
00:11:17,440 --> 00:11:18,880
This is like in the name.

413
00:11:18,880 --> 00:11:19,760
Like there's still

414
00:11:19,760 --> 00:11:21,920
often like more bash script

415
00:11:21,920 --> 00:11:24,960
than like custom JavaScript files.

416
00:11:24,960 --> 00:11:25,840
It takes about three years

417
00:11:25,840 --> 00:11:26,720
for someone to understand

418
00:11:26,720 --> 00:11:28,720
that you don't need your,

419
00:11:28,720 --> 00:11:30,640
database migration script to be as fast.

420
00:11:30,640 --> 00:11:31,680
you also don't need

421
00:11:31,680 --> 00:11:33,440
like a bespoke framework

422
00:11:33,440 --> 00:11:35,680
that you're allowed to have like a script

423
00:11:35,680 --> 00:11:36,880
that you can just like, even

424
00:11:36,880 --> 00:11:38,560
if it does only a single line,

425
00:11:38,560 --> 00:11:39,760
console log, hello

426
00:11:39,760 --> 00:11:41,360
world, like that's fine.

427
00:11:41,360 --> 00:11:42,320
You don't need like a

428
00:11:42,320 --> 00:11:44,400
bespoke CLI thingy to run it.

429
00:11:44,400 --> 00:11:45,840
Well, also I would also argue that it

430
00:11:45,840 --> 00:11:46,560
take about five years

431
00:11:46,560 --> 00:11:49,040
for someone to ease into all the

432
00:11:49,040 --> 00:11:50,240
TypeScript, you know,

433
00:11:50,240 --> 00:11:52,320
like not TypeScript, but most

434
00:11:52,320 --> 00:11:54,400
of Node.js ecosystem tooling.

435
00:11:54,400 --> 00:11:54,960
Right?

436
00:11:54,960 --> 00:11:55,920
It took a while.

437
00:11:55,920 --> 00:11:56,880
Because if someone has

438
00:11:56,880 --> 00:11:58,560
been in a Golang for a while,

439
00:11:58,560 --> 00:11:59,680
they probably have, they

440
00:11:59,680 --> 00:12:00,800
probably pollute them, you know,

441
00:12:00,800 --> 00:12:02,640
their home directory with

442
00:12:02,640 --> 00:12:04,080
all the Go binary already.

443
00:12:04,080 --> 00:12:05,360
I mean, I did the mistake three times.

444
00:12:05,360 --> 00:12:06,400
So I'm like, holy moly.

445
00:12:06,400 --> 00:12:07,520
I have so many of that.

446
00:12:07,520 --> 00:12:09,280
the zero point five version of Go,

447
00:12:09,280 --> 00:12:10,640
Was that you add Go in

448
00:12:10,640 --> 00:12:13,040
your user bin

449
00:12:13,040 --> 00:12:14,640
You clone it down and then

450
00:12:14,640 --> 00:12:15,760
every single main module,

451
00:12:15,760 --> 00:12:17,760
you just, you keep on cloning it.

452
00:12:17,760 --> 00:12:19,200
There's no, there's no Go model.

453
00:12:19,200 --> 00:12:21,120
Your machine becomes a Go machine.

454
00:12:21,120 --> 00:12:23,360
Yes.
Your machine is essentially a Go machine.

455
00:12:23,360 --> 00:12:24,480
And then with

456
00:12:24,480 --> 00:12:25,600
Python, for example, right?

457
00:12:25,600 --> 00:12:26,480
You get used to either

458
00:12:26,480 --> 00:12:28,547
using UV or PIP, right?

459
00:12:28,547 --> 00:12:30,000
You get used to
running almost everything.

460
00:12:30,000 --> 00:12:31,280
And so when people

461
00:12:31,280 --> 00:12:32,400
are, you know, kind of,

462
00:12:32,400 --> 00:12:33,120
used to it, right?

463
00:12:33,120 --> 00:12:34,800
It's really hard for them to switch gear.

464
00:12:34,800 --> 00:12:35,680
And this is why they

465
00:12:35,680 --> 00:12:36,880
having a list

466
00:12:36,880 --> 00:12:37,840
ensuring that when the

467
00:12:37,840 --> 00:12:38,640
team is still growing,

468
00:12:38,640 --> 00:12:39,520
we ensure that the

469
00:12:39,520 --> 00:12:41,200
consistency is there, right?

470
00:12:41,200 --> 00:12:42,000
Then there's no

471
00:12:42,000 --> 00:12:43,680
fragmentation of knowledge, right?

472
00:12:43,680 --> 00:12:45,120
And people can learn very quickly too.

473
00:12:45,120 --> 00:12:46,240
It's not that crazy.

474
00:12:46,240 --> 00:12:47,920
So tell me a bit more about

475
00:12:47,920 --> 00:12:49,520
how you've picked Cloudflare

476
00:12:49,520 --> 00:12:50,240
back then

477
00:12:50,240 --> 00:12:51,680
did you have a lot of

478
00:12:51,680 --> 00:12:53,200
experience with Cloudflare workers

479
00:12:53,200 --> 00:12:55,680
from previous projects and how, like,

480
00:12:55,680 --> 00:12:57,360
I suppose there's also a

481
00:12:57,360 --> 00:12:58,640
non-trivial amount of money

482
00:12:58,640 --> 00:12:59,600
being spent.

483
00:12:59,600 --> 00:13:00,640
So I'm curious how you

484
00:13:00,640 --> 00:13:01,680
think about this as well.

485
00:13:01,680 --> 00:13:02,560
It's actually very cheap.

486
00:13:02,560 --> 00:13:03,680
I'm not, I'm not kidding you.

487
00:13:03,680 --> 00:13:04,560
Like the infrastructure

488
00:13:04,560 --> 00:13:06,720
cost is break even completely.

489
00:13:06,720 --> 00:13:07,600
Just by the fee.

490
00:13:07,600 --> 00:13:08,400
So actually before

491
00:13:08,400 --> 00:13:10,160
Cloudflare, we were all on Vercel.

492
00:13:10,160 --> 00:13:12,080
Everything was on Vercel until they,

493
00:13:13,520 --> 00:13:14,400
they decided to charge

494
00:13:14,400 --> 00:13:15,600
for egress cost, right?

495
00:13:15,600 --> 00:13:17,120
but then also the,

496
00:13:17,120 --> 00:13:18,560
the network connection layer

497
00:13:18,560 --> 00:13:19,840
was not first class, right?

498
00:13:19,840 --> 00:13:21,920
So everything was completely abstracted.

499
00:13:21,920 --> 00:13:23,440
And we also have, even

500
00:13:23,440 --> 00:13:24,480
though we were on Vercel,

501
00:13:24,480 --> 00:13:25,600
we also have a Cloudflare

502
00:13:25,600 --> 00:13:27,200
proxy on top as well, right?

503
00:13:27,200 --> 00:13:28,240
For all the, you know, like

504
00:13:28,240 --> 00:13:30,000
for the DDoS protection and so on.

505
00:13:30,000 --> 00:13:32,640
So there's a three loop, three hoop.

506
00:13:32,640 --> 00:13:33,760
And so, I'm curious to

507
00:13:33,760 --> 00:13:34,560
hear a little bit more

508
00:13:34,560 --> 00:13:35,280
about the

509
00:13:35,280 --> 00:13:37,040
software aspect of like,

510
00:13:37,040 --> 00:13:39,120
what does a typical line of

511
00:13:39,120 --> 00:13:40,720
code in your TypeScript app

512
00:13:40,720 --> 00:13:41,440
actually look like?

513
00:13:41,440 --> 00:13:42,400
So currently you're

514
00:13:42,400 --> 00:13:44,400
running on Cloudflare workers.

515
00:13:44,400 --> 00:13:45,840
So in Cloudflare workers, for

516
00:13:45,840 --> 00:13:47,200
those who have never written

517
00:13:47,200 --> 00:13:49,600
a Cloudflare worker, it's essential,

518
00:13:49,600 --> 00:13:52,000
it's very web centric.

519
00:13:52,000 --> 00:13:54,720
So you have like an export default there.

520
00:13:54,720 --> 00:13:56,000
And in there, that's

521
00:13:56,000 --> 00:13:57,920
basically you export an object

522
00:13:57,920 --> 00:14:00,000
that returns a fetch function.

523
00:14:00,000 --> 00:14:01,360
And in that fetch function,

524
00:14:01,360 --> 00:14:03,280
you get a request coming in,

525
00:14:03,280 --> 00:14:04,480
you do your thing, and

526
00:14:04,480 --> 00:14:06,240
then you send back a response.

527
00:14:06,240 --> 00:14:08,800
It's like very simple, works really well.

528
00:14:08,800 --> 00:14:11,280
But like how you do your thing,

529
00:14:11,280 --> 00:14:12,080
it does like, that's the

530
00:14:12,080 --> 00:14:13,760
actual app that you're running.

531
00:14:13,760 --> 00:14:15,840
So I'm curious like how that looks like.

532
00:14:15,840 --> 00:14:17,520
The way I'm thinking about Cloudflare

533
00:14:17,520 --> 00:14:18,720
or worker in general is

534
00:14:18,720 --> 00:14:19,920
just the entry point, right?

535
00:14:19,920 --> 00:14:21,200
And if you ever kind

536
00:14:21,200 --> 00:14:22,960
of use any framework,

537
00:14:22,960 --> 00:14:24,640
entry point are just

538
00:14:24,640 --> 00:14:25,760
supposed to be entry point.

539
00:14:25,760 --> 00:14:27,440
You just try to attract your core logic

540
00:14:27,440 --> 00:14:28,400
as much as possible so that

541
00:14:28,400 --> 00:14:30,000
you can pull it into a package

542
00:14:30,000 --> 00:14:31,440
or a shared package so that

543
00:14:31,440 --> 00:14:32,960
regardless of entry point,

544
00:14:32,960 --> 00:14:34,000
but when we moved from

545
00:14:34,000 --> 00:14:34,880
Vercel to Cloudflare,

546
00:14:34,880 --> 00:14:35,600
it's like literally

547
00:14:35,600 --> 00:14:36,720
just moving entry point.

548
00:14:36,720 --> 00:14:38,640
And so there's no drag,

549
00:14:38,640 --> 00:14:40,240
so the core logic that we

550
00:14:40,240 --> 00:14:41,840
have is plain TypeScript

551
00:14:41,840 --> 00:14:43,760
function, right, with strict typing.

552
00:14:43,760 --> 00:14:44,720
A lot of time I would love

553
00:14:44,720 --> 00:14:45,760
to infer the type, right?

554
00:14:45,760 --> 00:14:48,400
Because I hate typing thing manually.

555
00:14:48,400 --> 00:14:49,440
I like to infer the

556
00:14:49,440 --> 00:14:51,520
type as much as possible.

557
00:14:51,520 --> 00:14:52,480
This is why one of my

558
00:14:52,480 --> 00:14:53,520
favorite brand teasers

559
00:14:53,520 --> 00:14:55,920
I give people is inferring type from

560
00:14:55,920 --> 00:14:57,120
third party library.

561
00:14:57,120 --> 00:14:58,640
You have to see if they have struggled

562
00:14:58,640 --> 00:14:59,600
through this thing before.

563
00:14:59,600 --> 00:15:00,640
the Cloudflare code

564
00:15:00,640 --> 00:15:01,920
is we just use the Cloudflare

565
00:15:01,920 --> 00:15:04,640
Hono library to wrap all the API

566
00:15:04,640 --> 00:15:06,800
under the same path and then does it.

567
00:15:06,800 --> 00:15:08,320
at that point, just the entry point

568
00:15:08,320 --> 00:15:10,160
with authentication and middleware

569
00:15:10,160 --> 00:15:11,120
and the rest of the core

570
00:15:11,120 --> 00:15:12,960
logic are plain function.

571
00:15:12,960 --> 00:15:13,520
Got it.

572
00:15:13,520 --> 00:15:15,760
So how is that like that

573
00:15:15,760 --> 00:15:17,760
inner chunk of business logic

574
00:15:17,760 --> 00:15:19,120
that's like actually

575
00:15:19,120 --> 00:15:21,120
what makes OpenRouter work,

576
00:15:21,120 --> 00:15:22,160
like where you're

577
00:15:22,160 --> 00:15:24,320
actually route the requests.

578
00:15:24,320 --> 00:15:26,080
So can you describe a

579
00:15:26,080 --> 00:15:27,680
little bit the architecture

580
00:15:27,680 --> 00:15:29,920
and like how trivial is

581
00:15:29,920 --> 00:15:31,360
what's happening inside there

582
00:15:31,360 --> 00:15:34,000
or how non-trivial, probably the latter.

583
00:15:34,000 --> 00:15:35,600
So maybe you can describe roughly

584
00:15:35,600 --> 00:15:38,160
like what is the life cycle of a request

585
00:15:38,160 --> 00:15:40,560
as it walks through like various stages

586
00:15:40,560 --> 00:15:43,040
of what makes OpenRouter, OpenRouter.

587
00:15:43,040 --> 00:15:44,320
let's make the tech request

588
00:15:44,320 --> 00:15:45,680
we have Llama for

589
00:15:45,680 --> 00:15:46,880
Maverick, let's say maverick.

590
00:15:46,880 --> 00:15:49,760
Maverick has about what like

591
00:15:49,760 --> 00:15:51,760
more than 10 different providers

592
00:15:51,760 --> 00:15:53,200
or serving the model.

593
00:15:53,200 --> 00:15:54,800
So when a provider is serving the model,

594
00:15:54,800 --> 00:15:55,680
we call that an endpoint.

595
00:15:55,680 --> 00:15:57,760
That's a concept that we have internally.

596
00:15:57,760 --> 00:16:00,000
A provider is, this company, right?

597
00:16:00,000 --> 00:16:02,320
Is the provider containing metadata

598
00:16:02,320 --> 00:16:02,960
but the provider

599
00:16:02,960 --> 00:16:04,880
meaning like company names,

600
00:16:04,880 --> 00:16:07,520
what they do or their location, right?

601
00:16:07,520 --> 00:16:08,480
Like the data privacy

602
00:16:08,480 --> 00:16:09,680
policy and stuff like that.

603
00:16:09,680 --> 00:16:11,600
The model is like the actual model on,

604
00:16:12,400 --> 00:16:14,240
that's the actual weight to the model

605
00:16:14,240 --> 00:16:15,040
with like context

606
00:16:15,040 --> 00:16:16,400
length combined those two,

607
00:16:16,400 --> 00:16:17,520
we call that an endpoint.

608
00:16:17,520 --> 00:16:19,120
let's say when we call

609
00:16:19,120 --> 00:16:20,480
the Llama 4 model Maverick,

610
00:16:20,480 --> 00:16:22,720
there's like more than 10 endpoints.

611
00:16:22,720 --> 00:16:24,320
The first thing is authentication, right?

612
00:16:24,320 --> 00:16:25,440
But authentication is

613
00:16:25,440 --> 00:16:26,640
literally just a function.

614
00:16:26,640 --> 00:16:27,920
We authenticate first.

615
00:16:27,920 --> 00:16:30,320
And then after you authenticate, we

616
00:16:30,320 --> 00:16:31,600
return a result, right?

617
00:16:31,600 --> 00:16:33,840
Either results, we

618
00:16:33,840 --> 00:16:34,880
really subscribe to the result

619
00:16:34,880 --> 00:16:36,160
monad paradigm.

620
00:16:36,160 --> 00:16:37,760
So almost every function

621
00:16:37,760 --> 00:16:39,280
we have returning results,

622
00:16:39,280 --> 00:16:41,840
we check the result and then we return

623
00:16:41,840 --> 00:16:43,120
a respond as needed.

624
00:16:43,120 --> 00:16:44,560
we would then take the request

625
00:16:44,560 --> 00:16:46,320
and we pass it through

626
00:16:46,320 --> 00:16:48,160
each of these endpoint.

627
00:16:48,160 --> 00:16:50,320
And while passing through the endpoint,

628
00:16:50,320 --> 00:16:51,680
so there are two stack that we applying

629
00:16:51,680 --> 00:16:52,960
for each of the requests.

630
00:16:52,960 --> 00:16:54,080
One stack is the

631
00:16:54,080 --> 00:16:55,680
routing, the routing stack.

632
00:16:55,680 --> 00:16:56,480
The routing stack

633
00:16:56,480 --> 00:16:58,000
allow us to like prioritize

634
00:16:58,000 --> 00:17:00,000
and order the endpoint

635
00:17:00,000 --> 00:17:01,680
according to what the user wants

636
00:17:01,680 --> 00:17:02,800
or according to some

637
00:17:02,800 --> 00:17:04,080
heuristic like latency,

638
00:17:04,080 --> 00:17:05,680
throughput or price.

639
00:17:05,680 --> 00:17:07,440
So the routings are a

640
00:17:07,440 --> 00:17:09,600
stack of decision step.

641
00:17:09,600 --> 00:17:11,920
And secondly, we have a plugin layers.

642
00:17:11,920 --> 00:17:13,360
The plugin layer allow us to like

643
00:17:13,360 --> 00:17:15,040
basically transform the request

644
00:17:15,040 --> 00:17:16,640
and also transform respond

645
00:17:16,640 --> 00:17:18,400
as the stream coming back up

646
00:17:18,400 --> 00:17:20,080
have a FIFO stack style.

647
00:17:20,080 --> 00:17:21,120
And that's essentially is

648
00:17:21,120 --> 00:17:22,400
the life cycle of requests.

649
00:17:22,400 --> 00:17:23,280
And then we just keep

650
00:17:23,280 --> 00:17:24,320
on hitting each endpoint

651
00:17:24,320 --> 00:17:25,040
until one that's

652
00:17:25,040 --> 00:17:26,720
actually, one is like saying,

653
00:17:26,720 --> 00:17:28,240
A, we can use a 200, we

654
00:17:28,240 --> 00:17:30,000
can now return some data.

655
00:17:30,000 --> 00:17:31,040
That's like a full cycle.

656
00:17:31,040 --> 00:17:32,000
And the code for all

657
00:17:32,000 --> 00:17:33,440
of this is very modular.

658
00:17:33,440 --> 00:17:34,800
Every step is a function.

659
00:17:34,800 --> 00:17:36,160
And then we just have an orchestration

660
00:17:36,160 --> 00:17:37,520
that's run through each of those

661
00:17:37,520 --> 00:17:39,120
would be very trivial to rewrite into

662
00:17:39,120 --> 00:17:40,240
Effect down the line.

663
00:17:40,240 --> 00:17:41,360
But we write this code

664
00:17:41,360 --> 00:17:42,400
before there was Effect.

665
00:17:42,400 --> 00:17:42,880
I think it was

666
00:17:42,880 --> 00:17:44,120
Effect 1.x or something.

667
00:17:44,120 --> 00:17:45,360
So, and I have to say

668
00:17:45,360 --> 00:17:46,960
just a brief comment

669
00:17:46,960 --> 00:17:47,760
because what you

670
00:17:47,760 --> 00:17:52,080
described is sounding very similar

671
00:17:52,080 --> 00:17:55,200
to what I had to do in my past,

672
00:17:55,200 --> 00:17:56,640
which is some sort of

673
00:17:56,640 --> 00:17:57,920
consolidation engine

674
00:17:57,920 --> 00:18:00,080
between different APIs

675
00:18:00,080 --> 00:18:02,480
that expose a single format.

676
00:18:03,120 --> 00:18:05,520
The reason Effect is born,

677
00:18:05,520 --> 00:18:08,880
like more than seven years ago now,

678
00:18:08,880 --> 00:18:10,400
it's because I was CTO of

679
00:18:10,400 --> 00:18:12,320
a Fintech company London.

680
00:18:12,320 --> 00:18:14,320
And we were doing consolidation

681
00:18:14,320 --> 00:18:16,640
over peer to peer landing providers.

682
00:18:16,640 --> 00:18:18,720
And we would buy and

683
00:18:18,720 --> 00:18:20,080
securitize those products

684
00:18:20,080 --> 00:18:22,400
to then sell them to

685
00:18:22,400 --> 00:18:24,240
institutional investors.

686
00:18:24,240 --> 00:18:27,360
And we had a large piece of software,

687
00:18:27,360 --> 00:18:29,200
which was basically we

688
00:18:29,200 --> 00:18:30,800
wanted to have a single format

689
00:18:30,800 --> 00:18:34,320
for loans, as effectively

690
00:18:34,320 --> 00:18:36,320
all of the loans are the same.

691
00:18:36,320 --> 00:18:38,400
There's a borrower, there's a lender,

692
00:18:38,400 --> 00:18:39,920
there's an interest rate,

693
00:18:39,920 --> 00:18:41,920
which is charged annually,

694
00:18:41,920 --> 00:18:43,840
monthly, daily, or

695
00:18:43,840 --> 00:18:45,440
whatever, and so on and so forth.

696
00:18:45,440 --> 00:18:46,880
But then you had those providers

697
00:18:46,880 --> 00:18:49,280
that would all have

698
00:18:49,280 --> 00:18:51,840
outstandingly different APIs,

699
00:18:51,840 --> 00:18:54,160
sometimes to achieve exactly the same.

700
00:18:54,160 --> 00:18:55,680
And we had a very similar

701
00:18:55,680 --> 00:18:57,600
approach to what you described.

702
00:18:57,600 --> 00:18:58,480
We also initially

703
00:18:58,480 --> 00:19:01,120
started with a result monad

704
00:19:01,120 --> 00:19:03,360
in Scala back in the day,

705
00:19:03,360 --> 00:19:05,280
and then testing became

706
00:19:05,280 --> 00:19:07,120
harder and harder and harder,

707
00:19:07,120 --> 00:19:09,840
to the point where we ended up

708
00:19:09,840 --> 00:19:11,600
doing tagless finally in

709
00:19:11,600 --> 00:19:12,720
Scala, which is

710
00:19:12,720 --> 00:19:14,720
it's absurd in complexity.

711
00:19:14,720 --> 00:19:16,400
It does achieve the goal

712
00:19:16,400 --> 00:19:17,680
of having everything

713
00:19:17,680 --> 00:19:19,200
modular and testable,

714
00:19:19,200 --> 00:19:22,160
but it is highly complex.

715
00:19:22,160 --> 00:19:23,840
And that's when I decided

716
00:19:23,840 --> 00:19:24,720
let's migrate everything

717
00:19:24,720 --> 00:19:26,960
to TypeScript because

718
00:19:26,960 --> 00:19:29,840
well, our domain is IO bound.

719
00:19:29,840 --> 00:19:32,080
There's no CPU compute

720
00:19:32,080 --> 00:19:34,880
used at any point in time

721
00:19:34,880 --> 00:19:38,400
except for data mapping, but even there,

722
00:19:38,400 --> 00:19:39,520
like you said before

723
00:19:39,520 --> 00:19:40,800
once you run your

724
00:19:40,800 --> 00:19:43,920
JavaScript in enough servers

725
00:19:43,920 --> 00:19:45,200
you're not really using

726
00:19:45,200 --> 00:19:46,560
multi-threading to

727
00:19:46,560 --> 00:19:47,760
handle a single request.

728
00:19:47,760 --> 00:19:49,600
You just have to respond

729
00:19:49,600 --> 00:19:51,360
to thousands of requests

730
00:19:51,360 --> 00:19:54,400
at the same time, but two requests can go

731
00:19:54,400 --> 00:19:55,440
to even two different

732
00:19:55,440 --> 00:19:57,040
instances and so on and so forth.

733
00:19:57,040 --> 00:19:59,840
Plus JavaScript is still very, very fast.

734
00:19:59,840 --> 00:20:00,800
So I was like, let's move

735
00:20:00,800 --> 00:20:02,080
everything to TypeScript.

736
00:20:02,080 --> 00:20:04,000
And I was punched in the face

737
00:20:04,000 --> 00:20:04,960
because none of the

738
00:20:04,960 --> 00:20:06,560
infrastructure existed

739
00:20:06,560 --> 00:20:07,600
that we were using in

740
00:20:07,600 --> 00:20:09,040
Scala for TypeScript.

741
00:20:09,040 --> 00:20:12,880
And that's why I built Effect in the first place.

742
00:20:12,880 --> 00:20:15,040
So it was precisely to be in the

743
00:20:15,040 --> 00:20:16,160
consolidation engine,

744
00:20:16,160 --> 00:20:18,160
a testable consolidation engine

745
00:20:18,160 --> 00:20:19,680
that had embedded open

746
00:20:19,680 --> 00:20:20,400
telemetry

747
00:20:20,400 --> 00:20:22,640
And this is what later became the open

748
00:20:22,640 --> 00:20:24,400
source project Effect.

749
00:20:24,400 --> 00:20:26,880
So I think it would be extremely good fit

750
00:20:26,880 --> 00:20:28,480
for actually what you're trying to do

751
00:20:28,480 --> 00:20:29,760
and funny enough exactly

752
00:20:29,760 --> 00:20:31,360
the same that I was doing.

753
00:20:31,360 --> 00:20:33,040
Which brings us to today

754
00:20:33,040 --> 00:20:34,880
where my understanding is

755
00:20:34,880 --> 00:20:37,120
that OpenRouter, the core

756
00:20:37,120 --> 00:20:40,160
product, is not yet using Effect.

757
00:20:40,160 --> 00:20:42,000
I think you're using Effect

758
00:20:42,000 --> 00:20:43,360
already for some internal tools

759
00:20:43,360 --> 00:20:44,160
and we can talk

760
00:20:44,160 --> 00:20:45,440
about that separately,

761
00:20:45,440 --> 00:20:46,960
but maybe addressing the

762
00:20:46,960 --> 00:20:48,480
elephant in the room first.

763
00:20:48,480 --> 00:20:50,560
You're here on the Effect podcast,

764
00:20:50,560 --> 00:20:52,800
where the core product of OpenRouter is

765
00:20:52,800 --> 00:20:54,240
not yet using Effect.

766
00:20:54,240 --> 00:20:55,520
Yet you've dropped the

767
00:20:55,520 --> 00:20:58,400
M-bomb before, the Monad bomb,

768
00:20:58,400 --> 00:20:59,360
which typically is a

769
00:20:59,360 --> 00:21:00,880
word that we're not trying

770
00:21:00,880 --> 00:21:02,800
to use too much because

771
00:21:02,800 --> 00:21:04,320
it's this interesting dynamic

772
00:21:04,320 --> 00:21:06,000
of what is the role of

773
00:21:06,000 --> 00:21:07,200
functional programming

774
00:21:07,200 --> 00:21:08,160
in what we're trying

775
00:21:08,160 --> 00:21:10,400
to achieve with Effect.

776
00:21:10,400 --> 00:21:12,080
My perspective on it is

777
00:21:12,080 --> 00:21:13,920
that we should follow the path

778
00:21:13,920 --> 00:21:15,520
that React has taken,

779
00:21:15,520 --> 00:21:17,680
where React is in essence

780
00:21:17,680 --> 00:21:20,080
a functional programming framework,

781
00:21:20,080 --> 00:21:22,400
but it does so in disguise,

782
00:21:22,400 --> 00:21:26,160
where it doesn't overwhelm everyone

783
00:21:26,160 --> 00:21:27,840
by being too dogmatic

784
00:21:27,840 --> 00:21:29,280
about functional programming.

785
00:21:29,280 --> 00:21:31,200
It's just an elegant system

786
00:21:31,200 --> 00:21:34,000
that tricks you into becoming

787
00:21:34,000 --> 00:21:35,520
a happy functional

788
00:21:35,520 --> 00:21:37,360
programmer without knowing.

789
00:21:37,360 --> 00:21:38,640
I would like to argue that

790
00:21:38,640 --> 00:21:41,440
because I was there back in 2015, right?

791
00:21:41,440 --> 00:21:43,520
When Angular and Ionic was

792
00:21:43,520 --> 00:21:44,640
kind of a king at the time

793
00:21:44,640 --> 00:21:46,640
because it's pushed by Google, right?

794
00:21:46,640 --> 00:21:48,000
And everyone was writing a

795
00:21:48,000 --> 00:21:50,000
directive and writing, you know,

796
00:21:50,000 --> 00:21:51,440
some, basically it's very

797
00:21:51,440 --> 00:21:52,720
much functional programming,

798
00:21:52,720 --> 00:21:53,920
but it was a bloat.

799
00:21:53,920 --> 00:21:55,040
It was bloating so

800
00:21:55,040 --> 00:21:57,120
much that the key sell,

801
00:21:57,120 --> 00:21:58,880
at least at the time I was still young,

802
00:21:58,880 --> 00:22:00,400
I was still inexperienced,

803
00:22:00,400 --> 00:22:01,440
but when I listened to

804
00:22:01,440 --> 00:22:02,400
people talking about,

805
00:22:02,400 --> 00:22:02,800
hey

806
00:22:02,800 --> 00:22:04,320
if you adopt functional programming,

807
00:22:04,320 --> 00:22:06,240
your program will be very, very lean.

808
00:22:06,240 --> 00:22:07,520
And I think there was

809
00:22:07,520 --> 00:22:09,360
that push of just whole ass

810
00:22:09,360 --> 00:22:10,000
in functional

811
00:22:10,000 --> 00:22:11,600
programming, don't be ashamed of it.

812
00:22:11,600 --> 00:22:13,120
And literally say, fuck

813
00:22:13,120 --> 00:22:14,320
object-oriented programming.

814
00:22:14,320 --> 00:22:16,080
I think that was a push.

815
00:22:16,800 --> 00:22:18,000
I would argue you should be,

816
00:22:18,000 --> 00:22:18,800
you should not be ashamed

817
00:22:18,800 --> 00:22:19,840
of functional programming.

818
00:22:19,840 --> 00:22:21,280
And even if they push it,

819
00:22:21,280 --> 00:22:22,480
I would say, I mean, sure,

820
00:22:22,480 --> 00:22:24,880
then just go slow, you know,

821
00:22:24,880 --> 00:22:26,960
then you have a slow code base.

822
00:22:27,520 --> 00:22:29,120
I wouldn't say we're ashamed of it.

823
00:22:29,120 --> 00:22:30,000
I think we're just

824
00:22:30,000 --> 00:22:32,480
trying to like not make it

825
00:22:32,480 --> 00:22:33,600
like an obstacle.

826
00:22:33,600 --> 00:22:36,000
So I think we've so far

827
00:22:36,000 --> 00:22:38,400
traversed the line pretty well

828
00:22:38,400 --> 00:22:40,560
by like inviting

829
00:22:40,560 --> 00:22:44,240
everyone who is FP positive,

830
00:22:44,240 --> 00:22:45,520
but we've also not tried

831
00:22:45,520 --> 00:22:47,360
to make it define the area

832
00:22:47,360 --> 00:22:50,160
to the barrier to entry through

833
00:22:50,160 --> 00:22:51,440
functional programming.

834
00:22:51,440 --> 00:22:54,080
so there was that guy, I forgot his name,

835
00:22:54,080 --> 00:22:55,920
one of the React engineers.

836
00:22:55,920 --> 00:22:57,520
Dan Abramov?

837
00:22:57,520 --> 00:22:59,520
Yeah, Dan Abramov, yes, that guy.

838
00:22:59,520 --> 00:23:00,880
I think he has a good

839
00:23:00,880 --> 00:23:02,000
sense of community there.

840
00:23:02,000 --> 00:23:03,760
I'm curious what he would say about this.

841
00:23:03,760 --> 00:23:05,280
Yeah, great point.

842
00:23:05,280 --> 00:23:06,720
We should invite him to

843
00:23:06,720 --> 00:23:08,960
the show and see whether.

844
00:23:08,960 --> 00:23:11,120
But I would say that like the result,

845
00:23:11,120 --> 00:23:12,560
I mean, the reason why we

846
00:23:12,560 --> 00:23:14,000
reached for the result monad,

847
00:23:14,000 --> 00:23:16,160
I mean, sure, most of my

848
00:23:16,160 --> 00:23:18,240
previous project were try catch,

849
00:23:18,240 --> 00:23:18,800
or like, you

850
00:23:18,800 --> 00:23:20,800
don't even try catch by the way.

851
00:23:20,800 --> 00:23:22,640
I just have a single error

852
00:23:22,640 --> 00:23:24,960
handler at the top of my apps.

853
00:23:24,960 --> 00:23:25,520
but then

854
00:23:25,520 --> 00:23:27,040
whatever it try, right?

855
00:23:27,040 --> 00:23:28,080
Whatever error is being

856
00:23:28,080 --> 00:23:29,600
thrown in underneath, right?

857
00:23:29,600 --> 00:23:30,880
Is an error type that

858
00:23:30,880 --> 00:23:32,160
I have defined, right?

859
00:23:32,160 --> 00:23:34,640
The problem is, down the line, right?

860
00:23:34,640 --> 00:23:35,920
You soon learn that it's

861
00:23:35,920 --> 00:23:37,360
very hard to debug those stuff,

862
00:23:37,360 --> 00:23:37,840
right?

863
00:23:37,840 --> 00:23:39,560
You soon learn that the stack tray

864
00:23:39,560 --> 00:23:40,160
can be very, very

865
00:23:40,160 --> 00:23:41,680
long and need to function

866
00:23:41,680 --> 00:23:43,760
is like any anything complex.

867
00:23:43,760 --> 00:23:45,120
It'll be very hard to go

868
00:23:45,120 --> 00:23:46,320
through the stack trays

869
00:23:46,320 --> 00:23:47,280
because you catching

870
00:23:47,280 --> 00:23:48,400
it on the top, right?

871
00:23:48,400 --> 00:23:49,600
So now you always want to

872
00:23:49,600 --> 00:23:50,880
somehow manage your error

873
00:23:50,880 --> 00:23:52,320
or localize your error

874
00:23:52,320 --> 00:23:53,440
at the call site,

875
00:23:53,440 --> 00:23:55,040
and then when you start doing that,

876
00:23:55,040 --> 00:23:56,400
you have nested try catch and

877
00:23:56,400 --> 00:23:57,840
it's a pain in the ass to read,

878
00:23:57,840 --> 00:24:00,240
right?
Because every nested layer is,

879
00:24:00,240 --> 00:24:01,280
I review my code on

880
00:24:01,280 --> 00:24:02,240
my phone a lot.\

881
00:24:02,240 --> 00:24:02,880
And yes,

882
00:24:02,880 --> 00:24:04,320
almost every try catch

883
00:24:04,320 --> 00:24:05,840
kind of nesting layer is like,

884
00:24:05,840 --> 00:24:07,680
I just have to scroll on one side.

885
00:24:07,680 --> 00:24:09,200
I just want to scroll on

886
00:24:09,200 --> 00:24:11,040
one, you know, vertically.

887
00:24:11,040 --> 00:24:11,840
The result monad

888
00:24:11,840 --> 00:24:12,880
essentially solve this problem,

889
00:24:12,880 --> 00:24:15,200
By explicitly forcing you

890
00:24:15,200 --> 00:24:16,400
to handle the error case

891
00:24:16,400 --> 00:24:17,280
as a value.

892
00:24:17,280 --> 00:24:19,600
I think is a value add and it is weird

893
00:24:19,600 --> 00:24:21,600
I mean, it is surely a different syntax.

894
00:24:21,600 --> 00:24:22,800
I think people have to,

895
00:24:22,800 --> 00:24:24,400
they just have to do it.

896
00:24:24,400 --> 00:24:26,160
It's the same as when

897
00:24:26,160 --> 00:24:27,360
React first came out, right?

898
00:24:27,360 --> 00:24:28,320
With useState.

899
00:24:28,320 --> 00:24:30,640
If you say useState is an

900
00:24:30,640 --> 00:24:33,840
equivalent to the class state,

901
00:24:33,840 --> 00:24:34,880
that's completely wrong.

902
00:24:34,880 --> 00:24:36,160
Like class state is,

903
00:24:36,160 --> 00:24:38,320
you cannot just check it in and out.

904
00:24:38,320 --> 00:24:39,520
With useState of React,

905
00:24:39,520 --> 00:24:40,240
you can just take it in

906
00:24:40,240 --> 00:24:41,520
and out fully functional.

907
00:24:41,520 --> 00:24:44,480
And then now your component is so light.

908
00:24:44,480 --> 00:24:46,080
And that was, I would

909
00:24:46,080 --> 00:24:47,120
argue the main selling point

910
00:24:47,120 --> 00:24:47,680
at the time.

911
00:24:47,680 --> 00:24:48,720
because of the pain, right?

912
00:24:48,720 --> 00:24:50,240
The pain of, oh my God,

913
00:24:50,240 --> 00:24:51,920
my class component is

914
00:24:51,920 --> 00:24:53,680
so slow or so bloated

915
00:24:53,680 --> 00:24:54,560
that with this thing

916
00:24:54,560 --> 00:24:56,400
now, it will be so faster.

917
00:24:56,400 --> 00:24:57,200
And it has to adopt

918
00:24:57,200 --> 00:24:58,320
functional programming.

919
00:24:58,320 --> 00:25:00,960
So I'm curious to hear a little bit more

920
00:25:00,960 --> 00:25:02,480
of like when you grew

921
00:25:02,480 --> 00:25:03,600
the engineering team

922
00:25:03,600 --> 00:25:05,360
and like you as the founders,

923
00:25:05,360 --> 00:25:06,240
you had the privilege

924
00:25:06,240 --> 00:25:08,000
of like later foundations

925
00:25:08,000 --> 00:25:09,600
or like set the taste as like,

926
00:25:09,600 --> 00:25:11,440
hey, this is how we're doing things.

927
00:25:11,440 --> 00:25:12,320
I think that's probably

928
00:25:12,320 --> 00:25:14,720
still holds to some degree,

929
00:25:14,720 --> 00:25:16,880
sure, like new people joining,

930
00:25:16,880 --> 00:25:18,880
bring in like their opinions.

931
00:25:18,880 --> 00:25:20,560
But I'm curious like which

932
00:25:20,560 --> 00:25:22,240
sort of mini culture clashes

933
00:25:22,240 --> 00:25:25,200
you've experienced by hiring someone new

934
00:25:25,200 --> 00:25:27,280
and they look at this and it's like,

935
00:25:27,280 --> 00:25:28,160
what is that?

936
00:25:28,160 --> 00:25:29,840
How did you explain that?

937
00:25:29,840 --> 00:25:31,040
And how did you sort of

938
00:25:31,040 --> 00:25:33,440
like culturally assimilate them

939
00:25:33,440 --> 00:25:35,360
to those ideas?

940
00:25:35,360 --> 00:25:37,520
Well, I think we just have

941
00:25:37,520 --> 00:25:38,640
to hire very good engineers

942
00:25:38,640 --> 00:25:39,920
and good engineers with

943
00:25:39,920 --> 00:25:42,920
experience usually can swallow.

944
00:25:42,920 --> 00:25:43,520
Yep.

945
00:25:43,520 --> 00:25:45,200
Oh, I mean, swallow not in a bad way,

946
00:25:45,200 --> 00:25:46,240
but kind of like they have

947
00:25:46,240 --> 00:25:47,760
seen things before, right?

948
00:25:47,760 --> 00:25:49,200
It's not something crazy, right?

949
00:25:49,200 --> 00:25:50,400
I would argue the crazy part

950
00:25:50,400 --> 00:25:52,720
is we have like bash script there,

951
00:25:52,720 --> 00:25:53,360
bash script with

952
00:25:53,360 --> 00:25:54,800
different syntax of bash,

953
00:25:54,800 --> 00:25:56,240
different flavor bash even, right?

954
00:25:56,240 --> 00:25:57,120
And then we have

955
00:25:57,120 --> 00:25:58,480
TypeScript, we have coffee

956
00:25:58,480 --> 00:25:59,520
and then we have like,

957
00:26:00,240 --> 00:26:02,720
maybe we use zest here and we did there.

958
00:26:02,720 --> 00:26:04,160
That is the crazy part.

959
00:26:04,160 --> 00:26:05,760
I would say that as we

960
00:26:05,760 --> 00:26:06,560
grow the team, right?

961
00:26:06,560 --> 00:26:08,080
We have tried our best to

962
00:26:08,080 --> 00:26:09,430
if they tech debt, take

963
00:26:09,430 --> 00:26:10,480
tech debt is widespread.

964
00:26:10,480 --> 00:26:11,680
If they have to tech debt,

965
00:26:11,680 --> 00:26:13,280
it widespread is unified.

966
00:26:13,280 --> 00:26:15,280
So when we refactor, we refactor all of

967
00:26:15,280 --> 00:26:16,240
them at the same time.

968
00:26:16,240 --> 00:26:18,080
That's one key part of it.

969
00:26:18,080 --> 00:26:20,320
If there is a certain framework we use,

970
00:26:20,320 --> 00:26:21,120
we stick to it.

971
00:26:21,120 --> 00:26:21,920
So like, for example,

972
00:26:21,920 --> 00:26:23,120
right now we're using

973
00:26:23,120 --> 00:26:24,240
a lot of our front-end,

974
00:26:24,240 --> 00:26:26,720
our like React, right?

975
00:26:26,720 --> 00:26:27,680
We know how to spend,

976
00:26:27,680 --> 00:26:30,640
we are going to spin up a stencil,

977
00:26:30,640 --> 00:26:32,800
you know, web component anytime soon.

978
00:26:32,800 --> 00:26:34,480
The next one is probably will we react?

979
00:26:34,480 --> 00:26:36,480
We'll stick to react for a while

980
00:26:36,480 --> 00:26:37,760
until we decide, okay,

981
00:26:37,760 --> 00:26:38,800
I think we, you know,

982
00:26:38,800 --> 00:26:40,320
let's move everything, uproot everything

983
00:26:40,320 --> 00:26:41,680
from react to stencil.

984
00:26:41,680 --> 00:26:42,560
Sure, we'll have the

985
00:26:42,560 --> 00:26:44,320
discussion, but not right now.

986
00:26:44,320 --> 00:26:45,760
And then same with TypeScript.

987
00:26:45,760 --> 00:26:46,800
The fact that we picked

988
00:26:46,800 --> 00:26:49,120
TypeScript and forced everyone to,

989
00:26:49,120 --> 00:26:50,320
hey, get used to it,

990
00:26:50,320 --> 00:26:51,760
write script in TypeScript,

991
00:26:51,760 --> 00:26:52,720
do everything in TypeScript

992
00:26:52,720 --> 00:26:54,240
as much as possible, right?

993
00:26:54,240 --> 00:26:55,280
That really forced us to

994
00:26:55,280 --> 00:26:56,320
just keep on learning.

995
00:26:56,320 --> 00:26:57,120
Even though they were

996
00:26:57,120 --> 00:26:58,240
like from a Python shop

997
00:26:58,240 --> 00:27:00,400
of GoShop or something else, when they

998
00:27:00,400 --> 00:27:01,280
come in here, right?

999
00:27:01,280 --> 00:27:03,360
I mean, sure, there's a resistant layer,

1000
00:27:03,360 --> 00:27:05,520
a resistant period where they're like,

1001
00:27:05,520 --> 00:27:07,840
how can I introduce a Go, you know,

1002
00:27:07,840 --> 00:27:09,680
the Go binary a little bit?

1003
00:27:09,680 --> 00:27:11,680
But then over time, right?

1004
00:27:11,680 --> 00:27:13,280
they start to get like used

1005
00:27:13,280 --> 00:27:14,640
to the TypeScript paradigm

1006
00:27:14,640 --> 00:27:15,200
because it's everywhere

1007
00:27:15,200 --> 00:27:15,920
in the code base, right?

1008
00:27:15,920 --> 00:27:18,080
You can have the AI agent look it up.

1009
00:27:18,080 --> 00:27:19,440
Also like Linter, we're using

1010
00:27:19,440 --> 00:27:21,120
biome right now for Linting,

1011
00:27:21,120 --> 00:27:22,320
plus Grid.

1012
00:27:22,320 --> 00:27:23,600
So we've been starting

1013
00:27:23,600 --> 00:27:24,880
adding a bunch of this Grid,

1014
00:27:24,880 --> 00:27:27,120
custom rules, which is very nice, right?

1015
00:27:27,120 --> 00:27:28,080
Because you can do like a

1016
00:27:28,080 --> 00:27:30,800
crazy regex on the code.

1017
00:27:30,800 --> 00:27:31,680
I'm using the same.

1018
00:27:31,680 --> 00:27:32,880
This is like, allows you

1019
00:27:32,880 --> 00:27:34,160
really nicely to define

1020
00:27:34,160 --> 00:27:35,120
some like, hey, this is

1021
00:27:35,120 --> 00:27:36,320
our way of doing things

1022
00:27:36,320 --> 00:27:38,080
from like all the 10 different,

1023
00:27:38,080 --> 00:27:38,960
TypeScript engineers

1024
00:27:38,960 --> 00:27:40,800
and they have 10 different

1025
00:27:40,800 --> 00:27:43,520
preferences of like braces there,

1026
00:27:43,520 --> 00:27:44,720
no braces there.

1027
00:27:44,720 --> 00:27:45,920
And then you can basically say,

1028
00:27:45,920 --> 00:27:46,960
this is the OpenRouter

1029
00:27:46,960 --> 00:27:48,880
way and run a command

1030
00:27:48,880 --> 00:27:51,360
and it gets linted and formatted.

1031
00:27:51,360 --> 00:27:52,480
Yeah, it is very awesome.

1032
00:27:52,480 --> 00:27:53,920
Like for example, we have a

1033
00:27:53,920 --> 00:27:54,960
rule that's strictly saying,

1034
00:27:54,960 --> 00:27:55,760
hey, if you have more

1035
00:27:55,760 --> 00:27:56,880
than three position argument,

1036
00:27:56,880 --> 00:27:57,840
just make an object.

1037
00:27:58,400 --> 00:28:01,920
Plus I feel like experienced engineers

1038
00:28:01,920 --> 00:28:05,600
usually focus on the problem

1039
00:28:05,600 --> 00:28:08,560
and the way we solve the problem,

1040
00:28:08,560 --> 00:28:11,920
rather than picking on the how,

1041
00:28:11,920 --> 00:28:13,680
in the sense of like

1042
00:28:13,680 --> 00:28:15,360
the thing you described,

1043
00:28:15,360 --> 00:28:17,840
that you use a result type for,

1044
00:28:17,840 --> 00:28:19,760
not gonna use result monad,

1045
00:28:19,760 --> 00:28:23,200
because monad is a term I don't wanna use

1046
00:28:23,200 --> 00:28:27,520
and not because I don't trust monad,

1047
00:28:27,520 --> 00:28:29,120
but because a monad would be an

1048
00:28:29,120 --> 00:28:30,400
abstraction on top of it,

1049
00:28:30,400 --> 00:28:31,680
which is the thing we're

1050
00:28:31,680 --> 00:28:33,920
trying to stay away from.

1051
00:28:33,920 --> 00:28:35,840
Result types are great

1052
00:28:35,840 --> 00:28:39,120
and they solve the problems you have.

1053
00:28:39,120 --> 00:28:41,817
They solve the nesting of try-catches

1054
00:28:41,817 --> 00:28:44,320
and so on and so forth.
So an experienced engineer that comes in

1055
00:28:44,320 --> 00:28:45,920
and sees the solution

1056
00:28:45,920 --> 00:28:47,120
might disagree on the syntax,

1057
00:28:47,760 --> 00:28:49,680
But the problem it's resolving,

1058
00:28:49,680 --> 00:28:52,480
it's clear they had it before,

1059
00:28:52,480 --> 00:28:53,920
they got tired of it

1060
00:28:53,920 --> 00:28:55,680
from other code bases

1061
00:28:55,680 --> 00:28:56,960
and they just digest

1062
00:28:56,960 --> 00:28:58,400
the current solution.

1063
00:28:58,400 --> 00:28:59,600
I wouldn't be opposed to

1064
00:28:59,600 --> 00:29:00,960
contribute to a code base

1065
00:29:00,960 --> 00:29:03,520
where there's a custom result type

1066
00:29:03,520 --> 00:29:05,360
and effect is not yet used.

1067
00:29:05,360 --> 00:29:05,920
I would say,

1068
00:29:05,920 --> 00:29:07,040
so that apply for like

1069
00:29:07,040 --> 00:29:08,480
a more senior engineer.

1070
00:29:08,480 --> 00:29:09,280
For the

1071
00:29:09,280 --> 00:29:11,120
slightly less senior engineer though,

1072
00:29:11,120 --> 00:29:13,120
I think the main thing is curiosity.

1073
00:29:13,120 --> 00:29:15,440
If they curious about why this was done,

1074
00:29:15,440 --> 00:29:16,320
they'll learn sure

1075
00:29:16,320 --> 00:29:18,000
learning faster on it too.

1076
00:29:18,000 --> 00:29:19,120
Because if you imagine you

1077
00:29:19,120 --> 00:29:20,720
can hire a very junior person,

1078
00:29:20,720 --> 00:29:23,520
but also very hard headed on some stuff

1079
00:29:23,520 --> 00:29:24,240
they've done before,

1080
00:29:24,240 --> 00:29:26,320
then they might not budge at all.

1081
00:29:26,320 --> 00:29:28,800
So I think for the less senior,

1082
00:29:28,800 --> 00:29:30,960
curiosity is where we strive for the most

1083
00:29:30,960 --> 00:29:32,160
and the way that we test

1084
00:29:32,160 --> 00:29:34,240
was curiosity is very simple.

1085
00:29:34,240 --> 00:29:36,320
They asked good question, really is.

1086
00:29:36,320 --> 00:29:38,880
So I think solving challenging problem

1087
00:29:38,880 --> 00:29:42,080
and then dig deep into it,

1088
00:29:42,080 --> 00:29:44,560
show that they have clear ownership

1089
00:29:44,560 --> 00:29:46,240
and clear digging, right?

1090
00:29:46,240 --> 00:29:47,680
The more they can dig, the better it is,

1091
00:29:47,680 --> 00:29:49,680
I think they show curiosity in a sense.

1092
00:29:49,680 --> 00:29:52,240
So taking a slight step back of like,

1093
00:29:52,240 --> 00:29:54,320
now you probably have a healthy mix

1094
00:29:54,320 --> 00:29:56,560
of relatively experienced engineers

1095
00:29:56,560 --> 00:29:59,120
and maybe some slightly more junior,

1096
00:29:59,120 --> 00:30:01,680
but very curious engineers.

1097
00:30:01,680 --> 00:30:03,760
And so you have right now,

1098
00:30:03,760 --> 00:30:06,080
a very sizable TypeScript code base.

1099
00:30:06,080 --> 00:30:08,480
Not sure, probably in the

1100
00:30:08,480 --> 00:30:10,240
six digits lines of code,

1101
00:30:10,240 --> 00:30:13,280
TypeScript repo, maybe even larger.

1102
00:30:13,280 --> 00:30:16,000
And you clearly have

1103
00:30:16,000 --> 00:30:17,440
Effect on your radar.

1104
00:30:17,440 --> 00:30:19,840
You've mentioned that some internal tools

1105
00:30:19,840 --> 00:30:21,280
are written in Effect,

1106
00:30:21,280 --> 00:30:23,360
but the main product is

1107
00:30:23,360 --> 00:30:25,200
not yet built with Effect.

1108
00:30:25,200 --> 00:30:27,760
And maybe you can just walk us through

1109
00:30:27,760 --> 00:30:29,520
and describe a little bit of like the

1110
00:30:29,520 --> 00:30:31,040
internal conversations

1111
00:30:31,040 --> 00:30:32,400
where you're considering,

1112
00:30:32,400 --> 00:30:34,000
hey, should we use Effect

1113
00:30:34,000 --> 00:30:37,280
for more parts of the core of the system?

1114
00:30:37,280 --> 00:30:38,320
Which sort of like

1115
00:30:38,320 --> 00:30:40,240
resistance questions you're getting,

1116
00:30:40,240 --> 00:30:41,840
how do people who are

1117
00:30:41,840 --> 00:30:43,280
already sold on Effect,

1118
00:30:43,280 --> 00:30:45,760
how they make a case for effect

1119
00:30:45,760 --> 00:30:48,000
and like just sharing your observations.

1120
00:30:48,000 --> 00:30:48,880
Well, the main observation

1121
00:30:48,880 --> 00:30:50,080
I have with the overall team

1122
00:30:50,080 --> 00:30:51,280
is almost every time

1123
00:30:51,280 --> 00:30:52,640
someone bringing a new technology

1124
00:30:52,640 --> 00:30:55,360
is a shiny object syndrome.

1125
00:30:55,360 --> 00:30:57,600
So very tough to sell, right?

1126
00:30:57,600 --> 00:30:58,720
in terms of handling errors,

1127
00:30:58,720 --> 00:31:00,880
we already have our own result monad

1128
00:31:00,880 --> 00:31:03,120
or result type that we already have,

1129
00:31:03,120 --> 00:31:05,120
which serve the purpose just fine.

1130
00:31:05,120 --> 00:31:06,000
and then we already have

1131
00:31:06,000 --> 00:31:06,960
this paradigm of writing,

1132
00:31:06,960 --> 00:31:08,320
make sure you write functions.

1133
00:31:08,320 --> 00:31:09,351
So when engineer, when

1134
00:31:09,351 --> 00:31:10,240
engineer look at Effect, right?

1135
00:31:10,240 --> 00:31:10,880
And the way it

1136
00:31:10,880 --> 00:31:13,600
orchestrating, you know, function, right?

1137
00:31:13,600 --> 00:31:15,040
A lot of time, the main question is,

1138
00:31:15,040 --> 00:31:16,560
why can't we use co-function then?

1139
00:31:16,560 --> 00:31:17,360
Why do we have to go

1140
00:31:17,360 --> 00:31:19,120
through Effect piping

1141
00:31:19,120 --> 00:31:20,800
or Effect matching when

1142
00:31:20,800 --> 00:31:22,160
a function work just fine

1143
00:31:22,160 --> 00:31:24,160
and an if and else work just fine too?

1144
00:31:24,160 --> 00:31:25,280
And if needed, we can

1145
00:31:25,280 --> 00:31:26,560
just make a new function,

1146
00:31:26,560 --> 00:31:27,920
do a switch cases with

1147
00:31:27,920 --> 00:31:29,120
complete check, right?

1148
00:31:29,120 --> 00:31:31,840
And plop it into the current,

1149
00:31:31,840 --> 00:31:33,280
functional flow, right?

1150
00:31:33,280 --> 00:31:35,600
And I mean, my response a lot to that is,

1151
00:31:35,600 --> 00:31:37,200
okay, what if you need to add a logger?

1152
00:31:37,200 --> 00:31:37,920
What if you need to add

1153
00:31:37,920 --> 00:31:39,360
a sidecar logger, right?

1154
00:31:39,360 --> 00:31:40,080
What if you need to add

1155
00:31:40,080 --> 00:31:41,840
observability or trace or span

1156
00:31:41,840 --> 00:31:42,400
on top of, you know,

1157
00:31:42,400 --> 00:31:43,760
this function call, right?

1158
00:31:43,760 --> 00:31:45,520
Well, the answer to a

1159
00:31:45,520 --> 00:31:46,320
lot of time would be,

1160
00:31:46,320 --> 00:31:47,680
well, I just add a new line, right?

1161
00:31:47,680 --> 00:31:49,360
Right before it.

1162
00:31:49,360 --> 00:31:51,280
Well, except that doesn't work for spans.

1163
00:31:51,280 --> 00:31:52,080
Precisely, right?

1164
00:31:52,080 --> 00:31:53,920
A lot of time, it doesn't work at all.

1165
00:31:53,920 --> 00:31:55,280
I think the main idea is

1166
00:31:55,280 --> 00:31:56,720
people don't really need

1167
00:31:56,720 --> 00:31:57,440
it at the moment

1168
00:31:57,440 --> 00:31:59,120
And so there's no huge push.

1169
00:31:59,120 --> 00:32:00,080
There's not, for example,

1170
00:32:00,080 --> 00:32:01,120
if we don't add Effect today,

1171
00:32:01,120 --> 00:32:02,160
we will still be around.

1172
00:32:02,160 --> 00:32:03,440
We will not be dying.

1173
00:32:03,440 --> 00:32:04,560
So that is why I would

1174
00:32:04,560 --> 00:32:05,600
say the main pushback,

1175
00:32:05,600 --> 00:32:06,320
because my team is

1176
00:32:06,320 --> 00:32:07,360
also very rational, right?

1177
00:32:07,360 --> 00:32:07,760
You know

1178
00:32:07,760 --> 00:32:09,040
they're also very busy, right?

1179
00:32:09,040 --> 00:32:09,800
it's not like,

1180
00:32:09,800 --> 00:32:11,280
it's spending time kind

1181
00:32:11,280 --> 00:32:13,280
of learning the namespace,

1182
00:32:13,280 --> 00:32:14,320
the Effect namespace, right?

1183
00:32:14,320 --> 00:32:15,680
Is already a drag.

1184
00:32:15,680 --> 00:32:16,640
Learning pipe is

1185
00:32:16,640 --> 00:32:18,160
already a drag, by the way.

1186
00:32:18,720 --> 00:32:19,680
So this is why the

1187
00:32:19,680 --> 00:32:20,480
first thing that I'm doing

1188
00:32:20,480 --> 00:32:21,680
is just introducing

1189
00:32:21,680 --> 00:32:23,920
small little usage of Effect.

1190
00:32:23,920 --> 00:32:25,840
We're using pipe almost everywhere now

1191
00:32:25,840 --> 00:32:27,440
when we composing a bunch of

1192
00:32:27,440 --> 00:32:28,800
function
in step by step, right?

1193
00:32:28,800 --> 00:32:29,737
And that helped a lot.

1194
00:32:29,737 --> 00:32:30,240
But the pipe function, by the way,

1195
00:32:30,240 --> 00:32:31,600
one of the main questions I

1196
00:32:31,600 --> 00:32:33,120
have for you guys actually

1197
00:32:33,120 --> 00:32:36,240
is a polymorphism of the pipe function.

1198
00:32:36,240 --> 00:32:37,840
A lot of engineers actually thought

1199
00:32:37,840 --> 00:32:39,520
that the pipe function is too flexible.

1200
00:32:39,520 --> 00:32:40,160
It's allowed you to

1201
00:32:40,160 --> 00:32:41,680
transform the type, right?

1202
00:32:41,680 --> 00:32:43,280
So it should have been called transform.

1203
00:32:43,280 --> 00:32:45,360
Transform pipe maybe, not pipe,

1204
00:32:45,360 --> 00:32:47,120
But to some engineers,

1205
00:32:47,120 --> 00:32:48,400
when we think about pipe

1206
00:32:48,400 --> 00:32:49,200
They assume it's

1207
00:32:49,200 --> 00:32:51,440
literally just a pipe through

1208
00:32:51,440 --> 00:32:52,240
where the type,

1209
00:32:52,240 --> 00:32:53,120
the input type should be

1210
00:32:53,120 --> 00:32:54,080
the same as the output type.

1211
00:32:54,080 --> 00:32:54,560
And it just

1212
00:32:54,560 --> 00:32:56,000
transformation step in between.

1213
00:32:56,000 --> 00:32:56,880
I'm curious what you guys

1214
00:32:56,880 --> 00:32:57,760
think about this, by the way.

1215
00:32:57,760 --> 00:32:59,200
What do you think about polymorphism?

1216
00:32:59,200 --> 00:33:00,480
Because essentially, the

1217
00:33:00,480 --> 00:33:01,520
this is the case where the

1218
00:33:01,520 --> 00:33:02,960
pipe type is not very scalable.

1219
00:33:02,960 --> 00:33:04,640
It only has 20 letters

1220
00:33:04,640 --> 00:33:06,160
You can only pipe 20 times.

1221
00:33:06,160 --> 00:33:08,000
I think there's reasons

1222
00:33:08,000 --> 00:33:09,360
why you don't wanna pipe

1223
00:33:09,360 --> 00:33:12,800
more than a number of elements.

1224
00:33:12,800 --> 00:33:14,400
And that's more along

1225
00:33:14,400 --> 00:33:16,080
the lines of limitations

1226
00:33:16,080 --> 00:33:17,680
in type inference

1227
00:33:17,680 --> 00:33:19,600
capabilities of TypeScript

1228
00:33:19,600 --> 00:33:21,920
and speed of type checking.

1229
00:33:21,920 --> 00:33:24,800
So we could easily add

1230
00:33:24,800 --> 00:33:27,040
200 arguments to pipe.

1231
00:33:27,040 --> 00:33:30,000
The actual implementation supports any

1232
00:33:30,000 --> 00:33:31,920
number of arguments.

1233
00:33:31,920 --> 00:33:34,560
The issue is TypeScript cannot deal

1234
00:33:34,560 --> 00:33:36,320
with the full polymorphism.

1235
00:33:36,320 --> 00:33:38,880
So you need to add specific overloads.

1236
00:33:38,880 --> 00:33:40,880
And before AI was very

1237
00:33:40,880 --> 00:33:43,280
tedious to add those overloads.

1238
00:33:43,280 --> 00:33:46,480
Now with Effect 4, I got Claude

1239
00:33:46,480 --> 00:33:49,360
to add the various pipe signatures.

1240
00:33:49,360 --> 00:33:52,320
I can easily tell Claude add another 200

1241
00:33:52,320 --> 00:33:54,080
so Louis can be happy.

1242
00:33:54,080 --> 00:33:55,440
that would be a solution.

1243
00:33:55,440 --> 00:33:58,320
But generally speaking about the

1244
00:33:58,320 --> 00:34:00,880
polymorphisms arguments,

1245
00:34:00,880 --> 00:34:03,600
pipe is generally applied to functions.

1246
00:34:03,600 --> 00:34:05,840
So you're piping the output of one

1247
00:34:05,840 --> 00:34:07,040
function to the input

1248
00:34:07,040 --> 00:34:08,880
of another function.

1249
00:34:08,880 --> 00:34:10,160
In that case, there's no

1250
00:34:10,160 --> 00:34:12,320
restriction about in and out.

1251
00:34:12,320 --> 00:34:14,320
You have data that flows

1252
00:34:14,320 --> 00:34:16,480
across a number of functions.

1253
00:34:16,480 --> 00:34:17,920
If it's understood in this

1254
00:34:17,920 --> 00:34:20,960
way, then it's not so confusing.

1255
00:34:20,960 --> 00:34:24,240
The confusion that I also had,

1256
00:34:24,240 --> 00:34:27,920
and at some point I was evaluating this,

1257
00:34:27,920 --> 00:34:29,680
is should the Effect

1258
00:34:29,680 --> 00:34:31,920
pipe be tied to Effect?

1259
00:34:31,920 --> 00:34:32,800
But then you have

1260
00:34:32,800 --> 00:34:34,960
cases where you might want

1261
00:34:34,960 --> 00:34:37,040
to convert an Effect to a stream.

1262
00:34:37,040 --> 00:34:39,280
You might want to repeat an

1263
00:34:39,280 --> 00:34:39,760
Effect

1264
00:34:39,760 --> 00:34:41,840
So you're having like slight

1265
00:34:41,840 --> 00:34:43,600
differences between the types.

1266
00:34:43,600 --> 00:34:44,960
So, I think putting an

1267
00:34:44,960 --> 00:34:46,880
arbitrary restriction

1268
00:34:46,880 --> 00:34:48,080
doesn't really help

1269
00:34:48,080 --> 00:34:49,440
in the understanding.

1270
00:34:49,440 --> 00:34:53,040
Now it is a good practice in your pipes

1271
00:34:53,040 --> 00:34:54,880
to not change the type

1272
00:34:54,880 --> 00:34:56,640
from the beginning to the end.

1273
00:34:56,640 --> 00:34:57,600
Otherwise while you

1274
00:34:57,600 --> 00:35:00,080
read, you lose context.

1275
00:35:00,080 --> 00:35:01,680
Like if you change the

1276
00:35:01,680 --> 00:35:03,840
type, most likely the best idea

1277
00:35:03,840 --> 00:35:05,360
would be to slice those

1278
00:35:05,360 --> 00:35:07,280
into two different pipes

1279
00:35:07,280 --> 00:35:08,800
and have variable

1280
00:35:08,800 --> 00:35:10,560
names that clearly state

1281
00:35:10,560 --> 00:35:11,520
what you're doing.

1282
00:35:11,520 --> 00:35:13,200
So this, I think it's

1283
00:35:13,200 --> 00:35:14,720
sort of best practices

1284
00:35:14,720 --> 00:35:16,480
and so on and so forth.

1285
00:35:16,480 --> 00:35:18,320
And I think with

1286
00:35:18,320 --> 00:35:19,760
touching on another point

1287
00:35:19,760 --> 00:35:21,600
that you mentioned, it's

1288
00:35:21,600 --> 00:35:23,440
the familiarity concept

1289
00:35:23,440 --> 00:35:26,880
and people's brain is wired into

1290
00:35:26,880 --> 00:35:28,240
imperative programming.

1291
00:35:28,240 --> 00:35:29,120
By imperative

1292
00:35:29,120 --> 00:35:31,200
programming, I mean the style.

1293
00:35:31,200 --> 00:35:32,160
Not the fact that

1294
00:35:32,160 --> 00:35:34,960
operations are actually executed

1295
00:35:34,960 --> 00:35:36,480
in an imperative sense.

1296
00:35:36,480 --> 00:35:38,880
I simply mean developers are used to

1297
00:35:38,880 --> 00:35:40,080
write normal functions.

1298
00:35:40,080 --> 00:35:41,680
Are used to write if statements.

1299
00:35:41,680 --> 00:35:42,800
They're used to write while.

1300
00:35:42,800 --> 00:35:44,960
They're used to write for.

1301
00:35:44,960 --> 00:35:46,960
They're used to write all of this.

1302
00:35:46,960 --> 00:35:49,360
With promises at the beginning,

1303
00:35:49,360 --> 00:35:51,440
we didn't get async await

1304
00:35:51,440 --> 00:35:54,400
And it was very tedious to write programs

1305
00:35:54,400 --> 00:35:55,520
because you had to think

1306
00:35:55,520 --> 00:35:57,920
in terms of this data flow,

1307
00:35:57,920 --> 00:36:00,880
this sequential nested data flow.

1308
00:36:00,880 --> 00:36:02,960
We escaped callback hell

1309
00:36:02,960 --> 00:36:06,320
to enter a venable hell

1310
00:36:06,320 --> 00:36:08,000
because then you had to access the

1311
00:36:08,560 --> 00:36:10,320
property from before,

1312
00:36:10,320 --> 00:36:11,600
then async await and

1313
00:36:11,600 --> 00:36:14,560
everything flattened out beautifully.

1314
00:36:14,560 --> 00:36:16,640
That you don't actually

1315
00:36:16,640 --> 00:36:18,320
get with a result type.

1316
00:36:18,320 --> 00:36:20,080
With a result type,

1317
00:36:20,080 --> 00:36:22,000
you almost end up writing

1318
00:36:22,000 --> 00:36:25,680
what they call Golang in TypeScript,

1319
00:36:25,680 --> 00:36:26,800
which is you call a

1320
00:36:26,800 --> 00:36:28,000
function then you have,

1321
00:36:28,000 --> 00:36:29,680
if you don't wanna end up in

1322
00:36:29,680 --> 00:36:30,320
pipes

1323
00:36:30,320 --> 00:36:31,120
what you end up doing

1324
00:36:31,120 --> 00:36:32,080
is you call a function

1325
00:36:32,080 --> 00:36:35,200
then you have if there was an error,

1326
00:36:35,200 --> 00:36:37,680
return the error, otherwise keep going.

1327
00:36:37,680 --> 00:36:40,400
If there was an error, well, in

1328
00:36:40,400 --> 00:36:42,000
JavaScript and in TypeScript

1329
00:36:42,000 --> 00:36:44,720
we have generators, which we rely heavily

1330
00:36:44,720 --> 00:36:48,080
in Effect to make code look

1331
00:36:48,080 --> 00:36:50,080
in the way that developers

1332
00:36:50,080 --> 00:36:52,720
are used to work with.

1333
00:36:52,720 --> 00:36:55,360
The mental model should be the same.

1334
00:36:55,360 --> 00:36:58,800
So my personal stance is that

1335
00:36:58,800 --> 00:37:01,120
pipe should not be the

1336
00:37:01,120 --> 00:37:02,880
default way we do things.

1337
00:37:02,880 --> 00:37:05,200
We should write code in a way that

1338
00:37:05,200 --> 00:37:06,800
everybody can understand.

1339
00:37:06,800 --> 00:37:07,760
And if you use a

1340
00:37:07,760 --> 00:37:10,400
generator, like switching mentally

1341
00:37:10,400 --> 00:37:11,920
from yield to await

1342
00:37:12,640 --> 00:37:14,320
it's not rocket science.

1343
00:37:14,320 --> 00:37:16,720
After you've read 10 generators,

1344
00:37:16,720 --> 00:37:19,520
you're gonna understand that you have a

1345
00:37:19,520 --> 00:37:20,880
direct mental mapping

1346
00:37:20,880 --> 00:37:23,360
that yield equals await, that's it.

1347
00:37:23,360 --> 00:37:24,800
And you apply your knowledge.

1348
00:37:24,800 --> 00:37:26,240
You can use it in a for loop,

1349
00:37:26,240 --> 00:37:27,520
you can use it in a while loop,

1350
00:37:27,520 --> 00:37:29,760
you can use it in an if condition.

1351
00:37:29,760 --> 00:37:32,880
You can do all that you want.

1352
00:37:32,880 --> 00:37:35,440
And if you want to access the specific

1353
00:37:35,440 --> 00:37:36,720
result of something,

1354
00:37:36,720 --> 00:37:40,240
you can say, okay, yield result of that.

1355
00:37:40,240 --> 00:37:40,800
And then you have

1356
00:37:40,800 --> 00:37:42,320
your if else that checks

1357
00:37:42,320 --> 00:37:43,600
for the exact

1358
00:37:43,600 --> 00:37:44,800
pattern that you would have

1359
00:37:44,800 --> 00:37:46,720
with a result type.

1360
00:37:46,720 --> 00:37:49,600
So I think like 90% of code

1361
00:37:49,600 --> 00:37:52,000
should go in that direction.

1362
00:37:52,000 --> 00:37:54,320
The rest of the code is

1363
00:37:54,320 --> 00:37:56,160
higher order behavior.

1364
00:37:56,160 --> 00:37:56,880
Like I have a

1365
00:37:56,880 --> 00:37:59,440
computation, I wanna retry it.

1366
00:37:59,440 --> 00:38:01,440
I don't wanna write a for loop.

1367
00:38:01,440 --> 00:38:03,760
I don't wanna deal with exponential

1368
00:38:03,760 --> 00:38:05,120
back off and everything.

1369
00:38:05,120 --> 00:38:07,840
Like you are integrating different APIs.

1370
00:38:07,840 --> 00:38:08,880
I think at some point

1371
00:38:08,880 --> 00:38:10,640
you have some retries.

1372
00:38:10,640 --> 00:38:12,320
At some point you have

1373
00:38:12,320 --> 00:38:13,440
to decide to back off

1374
00:38:13,440 --> 00:38:14,400
from the retries

1375
00:38:14,400 --> 00:38:16,080
because maybe open AI is down

1376
00:38:16,080 --> 00:38:17,120
and you all wanna keep

1377
00:38:17,120 --> 00:38:18,200
hitting the endpoint.

1378
00:38:18,200 --> 00:38:20,160
But those elements are

1379
00:38:20,160 --> 00:38:22,080
really not imperative.

1380
00:38:22,080 --> 00:38:23,200
They're not business

1381
00:38:23,200 --> 00:38:25,440
logic, they're not core.

1382
00:38:25,440 --> 00:38:26,880
They're just behaviors

1383
00:38:26,880 --> 00:38:28,400
that you add to the function.

1384
00:38:28,400 --> 00:38:29,600
And for that, I like to

1385
00:38:29,600 --> 00:38:31,840
do .pipe(Effect.retry).

1386
00:38:31,840 --> 00:38:32,720
That's it.

1387
00:38:32,720 --> 00:38:34,600
An LLM can understand it,

1388
00:38:34,600 --> 00:38:36,240
but a user doing

1389
00:38:36,240 --> 00:38:37,520
review can understand it.

1390
00:38:37,520 --> 00:38:38,480
And even a person

1391
00:38:38,480 --> 00:38:40,000
that doesn't know Effect

1392
00:38:40,000 --> 00:38:42,320
can read the function.

1393
00:38:42,320 --> 00:38:44,160
Okay, I understand the business logic

1394
00:38:44,160 --> 00:38:47,280
because it's all if else A equal B

1395
00:38:47,280 --> 00:38:49,280
and if these then that

1396
00:38:49,840 --> 00:38:51,120
while these conditions

1397
00:38:51,120 --> 00:38:53,040
through keep going.

1398
00:38:53,040 --> 00:38:57,120
And then they read pipe effect retry

1399
00:38:57,120 --> 00:39:00,160
while the error is network error.

1400
00:39:00,160 --> 00:39:01,520
Do not retry if the

1401
00:39:01,520 --> 00:39:03,440
response is malformed.

1402
00:39:03,440 --> 00:39:04,560
Cause if you call an API

1403
00:39:04,560 --> 00:39:06,000
and they change the response,

1404
00:39:06,000 --> 00:39:07,360
your decode logic will

1405
00:39:07,360 --> 00:39:09,840
keep failing anytime you retry.

1406
00:39:09,840 --> 00:39:12,560
So you might retry 500,000 times,

1407
00:39:12,560 --> 00:39:14,080
not getting anywhere.

1408
00:39:14,080 --> 00:39:16,080
So like to me, I

1409
00:39:16,080 --> 00:39:17,440
really wanna make it simple

1410
00:39:17,440 --> 00:39:20,160
for developers to read the code

1411
00:39:20,160 --> 00:39:24,720
and use pipe and use generators as tools

1412
00:39:24,720 --> 00:39:26,800
to really express business logic and

1413
00:39:26,800 --> 00:39:28,400
higher order behavior.

1414
00:39:28,400 --> 00:39:29,440
And I think like you are

1415
00:39:29,440 --> 00:39:31,760
in the very good direction

1416
00:39:31,760 --> 00:39:33,840
with the code base that you have.

1417
00:39:33,840 --> 00:39:37,760
And adding concepts one by one

1418
00:39:37,760 --> 00:39:39,600
is the right way to do it.

1419
00:39:39,600 --> 00:39:40,960
And from our private

1420
00:39:40,960 --> 00:39:42,640
discussions in the past,

1421
00:39:42,640 --> 00:39:44,480
I also see for example, you

1422
00:39:44,480 --> 00:39:46,160
use a lot of pattern matching.

1423
00:39:46,160 --> 00:39:48,080
Why the code becomes readable,

1424
00:39:48,080 --> 00:39:49,280
much more readable than

1425
00:39:49,280 --> 00:39:51,840
a sequence of if statements.

1426
00:39:51,840 --> 00:39:54,480
I would say that I actually consider

1427
00:39:54,480 --> 00:39:56,240
building everything using generator

1428
00:39:56,240 --> 00:39:58,080
but the code base started two years ago

1429
00:39:58,080 --> 00:39:59,760
and generator did not

1430
00:39:59,760 --> 00:40:01,200
work very well with async.

1431
00:40:01,200 --> 00:40:02,000
Async generator was on

1432
00:40:02,000 --> 00:40:03,680
a thing two years ago.

1433
00:40:03,680 --> 00:40:06,160
And you should not use async generators.

1434
00:40:06,160 --> 00:40:08,000
Yeah, you would not using it.

1435
00:40:08,000 --> 00:40:10,400
Now I would love to hear this by the way.

1436
00:40:10,400 --> 00:40:10,800
Why?

1437
00:40:10,800 --> 00:40:12,160
Why shouldn't you not

1438
00:40:12,160 --> 00:40:13,840
use async generators?

1439
00:40:13,840 --> 00:40:15,040
You should not have two

1440
00:40:15,040 --> 00:40:16,160
different representations

1441
00:40:16,160 --> 00:40:18,240
for sync code and async code.

1442
00:40:18,240 --> 00:40:19,680
That's madness.

1443
00:40:19,680 --> 00:40:20,800
That's actually also the

1444
00:40:20,800 --> 00:40:21,960
reason why I actually look up

1445
00:40:21,960 --> 00:40:22,800
by Effect by the way.

1446
00:40:22,800 --> 00:40:25,920
Oh, actually I look up what is it?

1447
00:40:25,920 --> 00:40:27,600
Function JS.

1448
00:40:27,600 --> 00:40:28,640
Because we had the result

1449
00:40:28,640 --> 00:40:31,120
and then we have async result.

1450
00:40:31,120 --> 00:40:31,920
Yes.

1451
00:40:31,920 --> 00:40:34,720
And you cannot escape from that.

1452
00:40:34,720 --> 00:40:37,680
So I think just to zoom out a little bit

1453
00:40:37,680 --> 00:40:39,120
and just to kind of like

1454
00:40:39,120 --> 00:40:41,280
observe the reflection here.

1455
00:40:41,280 --> 00:40:45,280
I think what makes your current code base

1456
00:40:45,280 --> 00:40:47,440
and your current programming styles

1457
00:40:47,440 --> 00:40:49,840
a little bit distinct from others

1458
00:40:49,840 --> 00:40:52,640
is that you went into this

1459
00:40:52,640 --> 00:40:54,240
with a very strong opinions.

1460
00:40:54,240 --> 00:40:55,760
Like, hey, we're not doing

1461
00:40:55,760 --> 00:40:57,360
try catch in this house.

1462
00:40:57,360 --> 00:40:59,360
And so you went like

1463
00:40:59,360 --> 00:41:01,600
this proper error handling.

1464
00:41:01,600 --> 00:41:03,520
And I think with that,

1465
00:41:03,520 --> 00:41:04,720
you're like way ahead

1466
00:41:04,720 --> 00:41:07,280
of where most TypeScript developers are.

1467
00:41:07,280 --> 00:41:09,200
most Effect adoption

1468
00:41:09,200 --> 00:41:10,800
I think happens to like

1469
00:41:10,800 --> 00:41:12,800
wherever the pain is most acute.

1470
00:41:12,800 --> 00:41:14,080
Maybe going to the doctor,

1471
00:41:14,080 --> 00:41:15,120
like you say like, oh my

1472
00:41:15,120 --> 00:41:16,800
gosh, this hurts so much.

1473
00:41:16,800 --> 00:41:18,000
But like while you're at the doctor,

1474
00:41:18,000 --> 00:41:19,040
you also realize like

1475
00:41:19,040 --> 00:41:20,000
you have those other things

1476
00:41:20,000 --> 00:41:21,920
that need treatment, but like one thing

1477
00:41:21,920 --> 00:41:23,440
gets you to the doctor.

1478
00:41:23,440 --> 00:41:25,600
And so the error handling thing,

1479
00:41:25,600 --> 00:41:27,440
like that pain like you've already

1480
00:41:27,440 --> 00:41:29,040
relieved yourself of that.

1481
00:41:29,040 --> 00:41:29,760
Yes, you have

1482
00:41:29,760 --> 00:41:33,200
appreciation for these other things.

1483
00:41:33,200 --> 00:41:34,400
Most people come to

1484
00:41:34,400 --> 00:41:36,400
Effect for error handling.

1485
00:41:36,400 --> 00:41:38,640
Also a lot of for other things,

1486
00:41:38,640 --> 00:41:39,680
but I think now

1487
00:41:39,680 --> 00:41:40,480
with the patterns

1488
00:41:40,480 --> 00:41:42,480
that you've like ended up with,

1489
00:41:42,480 --> 00:41:44,240
like just by talking through it.

1490
00:41:44,240 --> 00:41:46,480
Like I think we've like already uncovered

1491
00:41:46,480 --> 00:41:48,240
a couple of things where you say like,

1492
00:41:48,240 --> 00:41:50,560
well, yeah, actually that's not ideal,

1493
00:41:50,560 --> 00:41:52,480
but like, you know, it's like we have

1494
00:41:52,480 --> 00:41:53,600
good error handling.

1495
00:41:53,600 --> 00:41:54,960
But what Mike calls sort

1496
00:41:54,960 --> 00:41:56,320
of like the Golang style

1497
00:41:56,320 --> 00:41:57,920
of writing TypeScript.

1498
00:41:57,920 --> 00:42:00,560
I have done that myself as well.

1499
00:42:00,560 --> 00:42:02,080
And like this works

1500
00:42:02,080 --> 00:42:03,920
until it no longer works

1501
00:42:03,920 --> 00:42:06,560
or like until it just gets very annoying.

1502
00:42:06,560 --> 00:42:08,800
And I think the, what will

1503
00:42:08,800 --> 00:42:10,800
probably be the pain point,

1504
00:42:10,800 --> 00:42:11,840
like group of pain points

1505
00:42:11,840 --> 00:42:13,760
that gets you to the doctor

1506
00:42:13,760 --> 00:42:14,880
might be like better

1507
00:42:14,880 --> 00:42:17,280
retrying, might be observability,

1508
00:42:17,280 --> 00:42:18,560
might be like what

1509
00:42:18,560 --> 00:42:19,680
you call it like logger,

1510
00:42:19,680 --> 00:42:21,760
maybe metrics, et cetera.

1511
00:42:21,760 --> 00:42:22,640
And I think this will

1512
00:42:22,640 --> 00:42:23,760
make it actually easier

1513
00:42:23,760 --> 00:42:25,120
for you to adopt Effect

1514
00:42:25,120 --> 00:42:25,920
because you don't need

1515
00:42:25,920 --> 00:42:27,440
to rewire your brain.

1516
00:42:27,440 --> 00:42:29,360
So a lot of time we have been

1517
00:42:29,360 --> 00:42:30,480
incrementally adopting

1518
00:42:30,480 --> 00:42:31,280
some Effect, right?

1519
00:42:31,280 --> 00:42:34,800
Cause you know, lodash is a bloat

1520
00:42:34,800 --> 00:42:36,240
And I already added, I

1521
00:42:36,240 --> 00:42:37,120
already use my authority

1522
00:42:37,120 --> 00:42:39,560
to add Effect to the code base, right?

1523
00:42:39,560 --> 00:42:41,520
As I'm saying, we added Effect.

1524
00:42:41,520 --> 00:42:42,000
You're still

1525
00:42:42,000 --> 00:42:44,320
retaining your CTO authority.

1526
00:42:44,320 --> 00:42:46,480
Yeah, by the way, title wise, I'm not,

1527
00:42:47,120 --> 00:42:48,160
I don't claim CTO.

1528
00:42:48,160 --> 00:42:49,040
I'm just a co-founder

1529
00:42:49,040 --> 00:42:51,040
engineer title wise.

1530
00:42:51,040 --> 00:42:52,320
Yeah.

1531
00:42:52,320 --> 00:42:52,880
Perhaps there's a

1532
00:42:52,880 --> 00:42:54,160
better CTO out there for us.

1533
00:42:54,160 --> 00:42:55,280
We will see.

1534
00:42:55,280 --> 00:42:56,640
But I don't care too much about title,

1535
00:42:56,640 --> 00:42:58,000
but I care more about like

1536
00:42:58,000 --> 00:43:00,400
the dictator title for sure.

1537
00:43:00,400 --> 00:43:02,880
I'm a dictator for like BDFL for sure.

1538
00:43:02,880 --> 00:43:03,520
Yeah.

1539
00:43:03,520 --> 00:43:04,800
I can push the man by the way.

1540
00:43:04,800 --> 00:43:06,080
So like, for example,

1541
00:43:06,080 --> 00:43:08,560
You don't call a CTO, but

1542
00:43:08,560 --> 00:43:10,400
you get called a dictator.

1543
00:43:10,400 --> 00:43:12,160
Yeah, I'm a DFL.

1544
00:43:12,160 --> 00:43:13,120
And then, I mean, I was

1545
00:43:13,120 --> 00:43:14,320
saying like last week,

1546
00:43:14,320 --> 00:43:16,000
last week for the OpenAI dev day,

1547
00:43:16,000 --> 00:43:16,560
they were launching

1548
00:43:16,560 --> 00:43:17,760
the eval product, right?

1549
00:43:17,760 --> 00:43:20,000
And then right the Monday before,

1550
00:43:20,000 --> 00:43:22,080
they were paying a slack and saying,

1551
00:43:22,080 --> 00:43:22,960
hey, how come the

1552
00:43:22,960 --> 00:43:24,560
responding API don't have cost?

1553
00:43:24,560 --> 00:43:26,000
I just came right on and just, you know,

1554
00:43:26,000 --> 00:43:27,040
like push a thing on

1555
00:43:27,040 --> 00:43:29,040
man, two man in 30 minutes.

1556
00:43:29,040 --> 00:43:30,160
Shut up, right?

1557
00:43:30,160 --> 00:43:31,760
Get it over with, you know, like,

1558
00:43:31,760 --> 00:43:33,600
and that's the kind of thing I think is,

1559
00:43:33,600 --> 00:43:34,320
as a co-founder you

1560
00:43:34,320 --> 00:43:35,040
kind of have to do, right?

1561
00:43:35,040 --> 00:43:36,960
Is you gotta hold it by, you know,

1562
00:43:36,960 --> 00:43:37,600
you gotta hold it by

1563
00:43:37,600 --> 00:43:39,200
your throat and shove it.

1564
00:43:39,200 --> 00:43:40,160
Yeah, you have to know

1565
00:43:40,160 --> 00:43:41,440
when to skip the steps

1566
00:43:41,440 --> 00:43:42,800
and you have to know when

1567
00:43:42,800 --> 00:43:45,920
to impose a long-term vision

1568
00:43:45,920 --> 00:43:47,920
that is not yet apparent.

1569
00:43:47,920 --> 00:43:49,600
And sometimes you're wrong, like,

1570
00:43:49,600 --> 00:43:50,880
I do the same with Effect.

1571
00:43:50,880 --> 00:43:52,160
Sometimes I'm wrong.

1572
00:43:52,160 --> 00:43:53,440
I think I'm wrong and now, by the way,

1573
00:43:53,440 --> 00:43:53,920
there's one thing that

1574
00:43:53,920 --> 00:43:54,640
I might have been wrong,

1575
00:43:54,640 --> 00:43:57,760
which is like, I'm too ho-ass

1576
00:43:57,760 --> 00:44:00,160
into this idea of not using try-catch,

1577
00:44:00,160 --> 00:44:00,720
but then when it

1578
00:44:00,720 --> 00:44:02,160
comes to like the logging

1579
00:44:02,160 --> 00:44:03,120
or the tracing server,

1580
00:44:03,120 --> 00:44:04,480
you kind of have to use try-catch

1581
00:44:04,480 --> 00:44:06,960
so you don't allocate a new variable.

1582
00:44:06,960 --> 00:44:09,040
Because that way it's very easy to make a

1583
00:44:09,040 --> 00:44:09,840
copy of the variable

1584
00:44:09,840 --> 00:44:11,920
or when you do JSON parsing,

1585
00:44:11,920 --> 00:44:13,600
you make a copy of the variable

1586
00:44:13,600 --> 00:44:14,800
of a very deep nested

1587
00:44:14,800 --> 00:44:16,720
object of every value,

1588
00:44:16,720 --> 00:44:18,400
out of memory very easily.

1589
00:44:18,400 --> 00:44:18,880
Exactly.

1590
00:44:18,880 --> 00:44:21,440
And I think this is the thing where like

1591
00:44:21,440 --> 00:44:23,280
when people are sort of like in the,

1592
00:44:23,280 --> 00:44:26,560
the denial phase of Effect

1593
00:44:26,560 --> 00:44:27,755
and sort of like they, they

1594
00:44:27,755 --> 00:44:28,800
like justify to themselves,

1595
00:44:28,800 --> 00:44:30,000
like, we don't need Effect,

1596
00:44:30,000 --> 00:44:32,480
like we can do this thing like that.

1597
00:44:32,480 --> 00:44:33,680
Okay, yes, you can do

1598
00:44:33,680 --> 00:44:34,560
that thing like that,

1599
00:44:34,560 --> 00:44:36,160
like you do for error handling,

1600
00:44:36,160 --> 00:44:37,840
but like what if you need the,

1601
00:44:37,840 --> 00:44:39,200
if you wanna do the other thing,

1602
00:44:39,200 --> 00:44:41,440
if you wanna add that as well.

1603
00:44:41,440 --> 00:44:44,000
And like then, like all of those things,

1604
00:44:44,000 --> 00:44:45,920
they kind of like as you introduce them

1605
00:44:45,920 --> 00:44:47,360
on top of each other,

1606
00:44:47,360 --> 00:44:48,880
they don't seem to like

1607
00:44:48,880 --> 00:44:51,360
linearly add complexity,

1608
00:44:51,360 --> 00:44:52,400
but that's, I guess

1609
00:44:52,400 --> 00:44:54,480
like one of the unfortunate

1610
00:44:54,480 --> 00:44:56,640
emergent situations about TypeScript

1611
00:44:56,640 --> 00:44:57,600
is like the more you

1612
00:44:57,600 --> 00:44:59,360
add those requirements,

1613
00:44:59,360 --> 00:45:00,560
like observability,

1614
00:45:00,560 --> 00:45:02,080
better error handling,

1615
00:45:02,080 --> 00:45:02,580
dependency

1616
00:45:03,120 --> 00:45:05,200
management, like context management,

1617
00:45:05,200 --> 00:45:07,120
that doesn't make your code

1618
00:45:07,120 --> 00:45:08,640
like linearly more complex,

1619
00:45:08,640 --> 00:45:10,400
but exponentially more complex,

1620
00:45:10,400 --> 00:45:13,280
also visually more like visually,

1621
00:45:13,280 --> 00:45:15,760
exponentially more unwieldy.

1622
00:45:15,760 --> 00:45:16,640
And I think this is

1623
00:45:16,640 --> 00:45:17,840
where Effect comes in.

1624
00:45:17,840 --> 00:45:18,960
We've been adopting effect

1625
00:45:18,960 --> 00:45:20,400
in a bunch of like our retry

1626
00:45:20,400 --> 00:45:22,480
and like backup handling logic as well,

1627
00:45:22,480 --> 00:45:24,480
because the API elegant, right?

1628
00:45:24,480 --> 00:45:25,680
That's the whole idea.

1629
00:45:25,680 --> 00:45:27,280
So why would you write your own

1630
00:45:27,280 --> 00:45:28,880
when you can just pop this thing in,

1631
00:45:28,880 --> 00:45:31,680
throw a fetch and it's done, right?

1632
00:45:31,680 --> 00:45:32,560
And to review,

1633
00:45:32,560 --> 00:45:33,680
I don't think anyone

1634
00:45:33,680 --> 00:45:36,400
should adopt Effect in one day,

1635
00:45:36,400 --> 00:45:39,600
like in any reasonably large code base,

1636
00:45:39,600 --> 00:45:41,680
it would be absurd to

1637
00:45:41,680 --> 00:45:43,280
think one would rewrite

1638
00:45:43,280 --> 00:45:44,800
the whole code base.

1639
00:45:44,800 --> 00:45:47,440
we call effect a good virus,

1640
00:45:47,440 --> 00:45:48,880
in the sense that it tends

1641
00:45:48,880 --> 00:45:51,280
to spread when you started,

1642
00:45:51,280 --> 00:45:53,200
because now you have this retry,

1643
00:45:53,200 --> 00:45:55,760
but then you call a lot of APIs.

1644
00:45:55,760 --> 00:45:57,200
Obviously you are an API

1645
00:45:57,200 --> 00:45:59,360
consolidation provider.

1646
00:45:59,360 --> 00:46:01,600
So testing is essential.

1647
00:46:01,600 --> 00:46:03,680
You have to do integration testing,

1648
00:46:03,680 --> 00:46:05,040
you have to do end-to-end testing,

1649
00:46:05,040 --> 00:46:07,680
you might wanna do some unit testing.

1650
00:46:07,680 --> 00:46:08,880
And then you are in a

1651
00:46:08,880 --> 00:46:11,040
madness of Jest mocks

1652
00:46:11,040 --> 00:46:14,160
or Mocha mocks or whatever framework

1653
00:46:14,160 --> 00:46:16,880
you're gonna use Vitest mocks.

1654
00:46:16,880 --> 00:46:19,120
But mocks are really a very

1655
00:46:19,120 --> 00:46:20,800
complex thing to deal with,

1656
00:46:20,800 --> 00:46:23,280
because all your code is eager.

1657
00:46:23,280 --> 00:46:25,040
If you call it in the wrong place,

1658
00:46:25,040 --> 00:46:26,720
you have one piece

1659
00:46:26,720 --> 00:46:27,920
that is using one module,

1660
00:46:27,920 --> 00:46:28,640
another piece that is

1661
00:46:28,640 --> 00:46:30,240
using another module.

1662
00:46:30,240 --> 00:46:33,280
And so you might wanna elect,

1663
00:46:33,280 --> 00:46:35,280
use dependency injection.

1664
00:46:35,280 --> 00:46:37,520
I have seen how Epic Fabric

1665
00:46:37,520 --> 00:46:38,720
and their dependency

1666
00:46:38,720 --> 00:46:40,400
injection would help with that a lot.

1667
00:46:40,400 --> 00:46:40,880
Yeah.

1668
00:46:40,880 --> 00:46:42,560
Very good to use it, but you know,

1669
00:46:42,560 --> 00:46:43,520
I mean the team has

1670
00:46:43,520 --> 00:46:45,040
been using Vitest mock,

1671
00:46:45,040 --> 00:46:47,440
do mock and stuff for a long time.

1672
00:46:47,440 --> 00:46:49,760
By the way, we switched from Jest to Vite.

1673
00:46:49,760 --> 00:46:51,280
And that was a feat.

1674
00:46:51,280 --> 00:46:52,560
I know, I mean it's faster,

1675
00:46:52,560 --> 00:46:53,440
it's like the test runs

1676
00:46:53,440 --> 00:46:54,400
10 times faster, right?

1677
00:46:54,400 --> 00:46:57,600
We switched from prettier to biome.

1678
00:46:57,600 --> 00:46:58,320
I'm saying

1679
00:46:58,320 --> 00:47:01,440
we are not foreign to adopting new thing.

1680
00:47:01,440 --> 00:47:03,040
But it truly is all because,

1681
00:47:03,920 --> 00:47:05,360
it's pushed come to shove, right?

1682
00:47:05,360 --> 00:47:06,400
For example, if the test

1683
00:47:06,400 --> 00:47:07,760
take about five minutes to run,

1684
00:47:07,760 --> 00:47:09,600
oh my God, okay, let's

1685
00:47:09,600 --> 00:47:11,520
try to shave this down

1686
00:47:11,520 --> 00:47:12,400
and stuff like that.

1687
00:47:12,400 --> 00:47:13,840
Yeah, and to be frank,

1688
00:47:13,840 --> 00:47:17,120
I think like for your specific use case,

1689
00:47:17,120 --> 00:47:18,080
I think one of the

1690
00:47:18,080 --> 00:47:20,160
highest leverage points

1691
00:47:20,160 --> 00:47:22,080
could actually be the observability.

1692
00:47:22,080 --> 00:47:23,120
Because you already

1693
00:47:23,120 --> 00:47:25,840
referenced to debugging troubles

1694
00:47:25,840 --> 00:47:28,320
in the past of those,

1695
00:47:28,320 --> 00:47:29,680
once you have a hundred

1696
00:47:29,680 --> 00:47:30,720
of different functions,

1697
00:47:30,720 --> 00:47:33,200
the stack traces becomes unreadable.

1698
00:47:33,200 --> 00:47:34,400
Yes, they are.

1699
00:47:34,400 --> 00:47:36,000
They're much less unreadable

1700
00:47:36,000 --> 00:47:38,160
if you have spans at the right places

1701
00:47:38,160 --> 00:47:39,520
and you have a full dashboard

1702
00:47:39,520 --> 00:47:41,360
that tells you if something goes wrong.

1703
00:47:41,360 --> 00:47:41,920
The one thing with

1704
00:47:41,920 --> 00:47:43,360
observability and OpenRouter

1705
00:47:43,360 --> 00:47:44,640
is that because we run

1706
00:47:44,640 --> 00:47:46,160
the thing on the edge

1707
00:47:46,160 --> 00:47:47,680
and the worker runtime,

1708
00:47:47,680 --> 00:47:49,680
we don't truly have a span server

1709
00:47:49,680 --> 00:47:52,800
to continuously collecting the span.

1710
00:47:52,800 --> 00:47:54,000
So right now we're actually hooking

1711
00:47:54,000 --> 00:47:56,960
into the Node.js diagnostic channel,

1712
00:47:56,960 --> 00:47:59,120
which is not very good, I would say.

1713
00:47:59,120 --> 00:48:00,480
At least on the worker runtime

1714
00:48:00,480 --> 00:48:02,320
because the memory is very limited.

1715
00:48:02,320 --> 00:48:03,520
So we have to be very

1716
00:48:03,520 --> 00:48:05,360
careful with what we trace.

1717
00:48:05,360 --> 00:48:07,600
Well, I wish at some point,

1718
00:48:08,240 --> 00:48:10,080
we're gonna get to agentic behavior,

1719
00:48:10,080 --> 00:48:11,440
long-running processes

1720
00:48:11,440 --> 00:48:12,480
at some point in time

1721
00:48:12,480 --> 00:48:14,080
in this discussion.

1722
00:48:14,080 --> 00:48:17,440
Yeah, so are you familiar with the ECS or

1723
00:48:18,480 --> 00:48:19,600
the entity component

1724
00:48:19,600 --> 00:48:21,520
system in game design?

1725
00:48:21,520 --> 00:48:25,760
Yeah, I think the reason the actor model

1726
00:48:25,760 --> 00:48:29,120
in general is very popular in game design

1727
00:48:29,120 --> 00:48:32,160
is because in games you usually have a

1728
00:48:32,160 --> 00:48:34,160
lot of parallel entities

1729
00:48:34,160 --> 00:48:36,720
and they do share messages with each

1730
00:48:36,720 --> 00:48:38,240
other. It's a very

1731
00:48:38,240 --> 00:48:41,200
interactive world. Basically, again, it's

1732
00:48:41,200 --> 00:48:43,120
a world with entities that both some

1733
00:48:43,120 --> 00:48:45,120
entities can be movable, some entities

1734
00:48:45,120 --> 00:48:47,920
can be movable, some speak, some don't.

1735
00:48:47,920 --> 00:48:49,840
But even if you have something like a

1736
00:48:49,840 --> 00:48:52,560
room with a door, you have a character

1737
00:48:52,560 --> 00:48:54,720
that hits the door, the door is open and

1738
00:48:54,720 --> 00:48:56,240
All of those are sort

1739
00:48:56,240 --> 00:48:58,400
of unstructured. It's not very

1740
00:48:58,400 --> 00:49:01,120
deterministic. The player can take

1741
00:49:01,120 --> 00:49:02,076
different ways and so on and so forth.

1742
00:49:02,076 --> 00:49:04,160
And the actor model is perfect for that

1743
00:49:04,160 --> 00:49:06,560
case. And if you think about it, the

1744
00:49:06,560 --> 00:49:09,760
reason the actor model was even invented,

1745
00:49:09,760 --> 00:49:12,160
as far as my knowledge goes, we go back

1746
00:49:12,160 --> 00:49:14,960
to Erlang and we go back to the 70s when

1747
00:49:14,960 --> 00:49:19,280
Ericsson had to model telecom networks

1748
00:49:19,280 --> 00:49:22,320
where, again, you have very

1749
00:49:22,320 --> 00:49:24,400
undeterministic and unrelated entities

1750
00:49:24,400 --> 00:49:26,560
that communicate with each other.

1751
00:49:27,200 --> 00:49:30,720
I can phone Louis, New York, I can phone

1752
00:49:30,720 --> 00:49:33,520
Johannes in Germany, they can phone each

1753
00:49:33,520 --> 00:49:36,400
other. You have race conditions when two

1754
00:49:36,400 --> 00:49:38,240
people try to call the same

1755
00:49:38,240 --> 00:49:40,640
It's very similar to what you

1756
00:49:40,640 --> 00:49:43,120
would get in a game.

1757
00:49:43,120 --> 00:49:45,440
And in my opinion, it's very similar to

1758
00:49:45,440 --> 00:49:47,280
what an agentic system

1759
00:49:47,280 --> 00:49:49,440
looks like to a very high degree.

1760
00:49:49,440 --> 00:49:50,800
Have you thought about

1761
00:49:50,800 --> 00:49:53,040
agentic system that's steerable?

1762
00:49:53,040 --> 00:49:54,640
So when I was doing game design,

1763
00:49:54,640 --> 00:49:56,080
we do a lot of steer algorithm.

1764
00:49:56,080 --> 00:49:57,360
And steer algorithm is

1765
00:49:57,360 --> 00:49:59,120
all about managing a crown

1766
00:49:59,120 --> 00:50:00,800
Let's say you want to model

1767
00:50:00,800 --> 00:50:02,560
a particle system, right?

1768
00:50:02,560 --> 00:50:03,280
You have the particle

1769
00:50:03,280 --> 00:50:04,240
system, it's actually,

1770
00:50:04,240 --> 00:50:05,040
and each of them could

1771
00:50:05,040 --> 00:50:05,760
be a bird, by the way.

1772
00:50:05,760 --> 00:50:08,080
The particle system can be represented

1773
00:50:08,080 --> 00:50:09,600
by a bunch of bird

1774
00:50:09,600 --> 00:50:12,240
actors within this cap.

1775
00:50:12,240 --> 00:50:13,120
It's kind of a point

1776
00:50:13,120 --> 00:50:14,800
cloud of a bunch of birds.

1777
00:50:14,800 --> 00:50:15,680
And now let's say you

1778
00:50:15,680 --> 00:50:17,120
swarm this bird, right?

1779
00:50:17,120 --> 00:50:18,000
You swarm this bird through

1780
00:50:18,000 --> 00:50:20,080
an obstacle in the middle.

1781
00:50:20,080 --> 00:50:21,120
Now they all have to

1782
00:50:21,120 --> 00:50:23,440
independently dodge that ball

1783
00:50:23,440 --> 00:50:25,280
and fly around that thing, right?

1784
00:50:25,280 --> 00:50:26,320
Because the whole idea is when

1785
00:50:26,320 --> 00:50:27,840
you steer this flock of bird,

1786
00:50:27,840 --> 00:50:29,920
you just want to pick a pointer

1787
00:50:29,920 --> 00:50:30,720
and you move that

1788
00:50:30,720 --> 00:50:32,960
pointer to a certain location

1789
00:50:32,960 --> 00:50:34,880
on the 3D map and then the flock of birds

1790
00:50:34,880 --> 00:50:36,400
will just keep on following.

1791
00:50:36,400 --> 00:50:37,040
They keep on

1792
00:50:37,040 --> 00:50:38,960
following that point in space.

1793
00:50:38,960 --> 00:50:40,080
They both try to steer

1794
00:50:40,080 --> 00:50:41,120
away from each other,

1795
00:50:41,120 --> 00:50:43,680
but they all kind of act in tandem to

1796
00:50:44,400 --> 00:50:46,320
move to your pointer

1797
00:50:46,320 --> 00:50:49,120
and then they dodge popsicle on the way.

1798
00:50:49,120 --> 00:50:51,920
How do you think about steering agent

1799
00:50:51,920 --> 00:50:55,120
or in that same vein?

1800
00:50:55,120 --> 00:50:56,160
Have you thought about that?

1801
00:50:56,160 --> 00:50:57,840
To be frank, I have not.

1802
00:50:57,840 --> 00:50:59,440
I can apply some of my

1803
00:50:59,440 --> 00:51:00,960
knowledge from mathematics

1804
00:51:00,960 --> 00:51:02,000
in the past,

1805
00:51:02,000 --> 00:51:04,560
especially studying how birds

1806
00:51:04,560 --> 00:51:06,560
actually end up doing that.

1807
00:51:06,560 --> 00:51:09,440
And it's all about local communication

1808
00:51:09,440 --> 00:51:12,240
between one bird and the other.

1809
00:51:12,240 --> 00:51:14,400
There's no overall concept.

1810
00:51:14,400 --> 00:51:16,800
Basically the steering is almost like

1811
00:51:16,800 --> 00:51:18,560
one tries to steer a little bit,

1812
00:51:18,560 --> 00:51:20,320
the other follows the one

1813
00:51:20,320 --> 00:51:22,000
close and so on and so forth.

1814
00:51:22,000 --> 00:51:23,520
And I think even in that case,

1815
00:51:23,520 --> 00:51:25,920
it's all about the connectivity

1816
00:51:25,920 --> 00:51:28,640
between sorrounding entities

1817
00:51:28,640 --> 00:51:30,000
the neighbor flock

1818
00:51:30,000 --> 00:51:30,960
Each of the entity

1819
00:51:30,960 --> 00:51:32,240
they should have knowledge of

1820
00:51:32,240 --> 00:51:33,840
the entity right close to it.

1821
00:51:33,840 --> 00:51:36,400
So you put a small little boating block.

1822
00:51:36,400 --> 00:51:37,253
on each of them

1823
00:51:37,253 --> 00:51:38,800
so more like small clusters of

1824
00:51:38,800 --> 00:51:40,560
entities that sort of have

1825
00:51:40,560 --> 00:51:42,480
deep communication between themselves,

1826
00:51:42,480 --> 00:51:44,880
but almost not communicate

1827
00:51:44,880 --> 00:51:47,600
in the far distance one.

1828
00:51:47,600 --> 00:51:49,600
And also to quickly throw in here

1829
00:51:49,600 --> 00:51:51,120
to answer your initial question,

1830
00:51:51,120 --> 00:51:51,840
I've asked you to

1831
00:51:51,840 --> 00:51:54,240
chat GPT and asked how,

1832
00:51:54,240 --> 00:51:55,840
chat GPT how ECS

1833
00:51:55,840 --> 00:51:57,520
compares to the actor model.

1834
00:51:57,520 --> 00:51:59,440
And it says that they

1835
00:51:59,440 --> 00:52:00,880
overlap conceptually.

1836
00:52:00,880 --> 00:52:02,400
And the TLDR is that

1837
00:52:02,400 --> 00:52:05,040
ECS is about what you have

1838
00:52:05,040 --> 00:52:08,720
in terms of data and what operates on it,

1839
00:52:08,720 --> 00:52:11,120
like which systems optimized for

1840
00:52:11,120 --> 00:52:12,640
throughput and composition.

1841
00:52:12,640 --> 00:52:13,920
And I think composition here

1842
00:52:13,920 --> 00:52:15,360
is really interesting, right?

1843
00:52:15,360 --> 00:52:18,480
In a game you compose like a big scene

1844
00:52:18,480 --> 00:52:20,000
of like smaller things.

1845
00:52:20,000 --> 00:52:23,280
And the actor is about who does what

1846
00:52:23,280 --> 00:52:25,680
optimized for isolation concurrency.

1847
00:52:25,680 --> 00:52:27,600
So I think we can basically

1848
00:52:27,600 --> 00:52:29,120
build an overall system here

1849
00:52:29,120 --> 00:52:30,560
with both and they're

1850
00:52:30,560 --> 00:52:32,160
just like have sort of like

1851
00:52:32,160 --> 00:52:34,800
different aspects, like

1852
00:52:34,800 --> 00:52:36,400
what they're concerned about.

1853
00:52:36,400 --> 00:52:39,040
Like composition is interesting.

1854
00:52:39,040 --> 00:52:40,720
And the other one about concurrency,

1855
00:52:40,720 --> 00:52:43,120
like ideally you have both,

1856
00:52:43,120 --> 00:52:45,200
but both needs like this

1857
00:52:45,200 --> 00:52:47,120
different kind of concerns.

1858
00:52:47,120 --> 00:52:49,120
So I just wanted to weave that in

1859
00:52:49,120 --> 00:52:51,440
as I think it fits really nicely

1860
00:52:51,440 --> 00:52:54,640
the sort of like ball flock of birds

1861
00:52:54,640 --> 00:52:57,280
kind of scene we've been describing.

1862
00:52:57,280 --> 00:52:59,680
The kind of difference is more

1863
00:52:59,680 --> 00:53:03,120
at the level of the API they expose

1864
00:53:03,120 --> 00:53:07,280
rather than at the level of the runtime.

1865
00:53:07,280 --> 00:53:10,560
Because they're still both actors.

1866
00:53:10,560 --> 00:53:11,920
It's just that the API

1867
00:53:11,920 --> 00:53:13,840
focus is more on one aspect

1868
00:53:14,480 --> 00:53:16,000
or the other.

1869
00:53:16,000 --> 00:53:17,360
For example, even in

1870
00:53:17,360 --> 00:53:20,080
actor systems like Akka

1871
00:53:20,080 --> 00:53:21,520
you have different ways

1872
00:53:21,520 --> 00:53:23,040
of creating actors.

1873
00:53:23,040 --> 00:53:24,560
And even in a fat cluster,

1874
00:53:24,560 --> 00:53:26,800
which is our actor solution,

1875
00:53:26,800 --> 00:53:27,840
you do have different

1876
00:53:27,840 --> 00:53:29,120
ways of creating actors.

1877
00:53:29,120 --> 00:53:31,440
You can focus on the behavior,

1878
00:53:31,440 --> 00:53:32,960
which would be more along

1879
00:53:32,960 --> 00:53:36,000
lines of data in data out.

1880
00:53:36,000 --> 00:53:38,400
The kind of messages that the actor

1881
00:53:38,400 --> 00:53:40,800
can respond to, can send to.

1882
00:53:40,800 --> 00:53:44,720
Or you basically get a fiber with a cue

1883
00:53:44,720 --> 00:53:46,640
that you pull on and

1884
00:53:46,640 --> 00:53:49,600
everything you do is custom.

1885
00:53:49,600 --> 00:53:50,400
So you have more

1886
00:53:50,400 --> 00:53:53,840
flexibility, less structured API.

1887
00:53:53,840 --> 00:53:56,400
I think like the idea of a

1888
00:53:56,400 --> 00:53:57,680
distributed actor system

1889
00:53:57,680 --> 00:54:02,400
is very, very essential to

1890
00:54:02,400 --> 00:54:04,240
distributed systems in general.

1891
00:54:04,240 --> 00:54:07,120
And my point is really that

1892
00:54:07,120 --> 00:54:08,480
agentic behavior in

1893
00:54:08,480 --> 00:54:11,600
applications is a distributed system.

1894
00:54:11,600 --> 00:54:13,920
And it's just so happens that nowadays

1895
00:54:13,920 --> 00:54:15,600
everybody builds AI applications.

1896
00:54:15,600 --> 00:54:17,120
So they hit

1897
00:54:17,120 --> 00:54:19,600
the distributed system problems day one.

1898
00:54:19,600 --> 00:54:21,520
rather than facing them

1899
00:54:21,520 --> 00:54:23,040
further down the path.

1900
00:54:23,040 --> 00:54:25,360
And it's also why companies like Temporal

1901
00:54:25,360 --> 00:54:27,600
is getting renewed traction.

1902
00:54:27,600 --> 00:54:30,160
Because before you'd only use Temporal,

1903
00:54:30,160 --> 00:54:31,840
if you actually have a lot of customers

1904
00:54:31,840 --> 00:54:34,320
and what used to be an edge case,

1905
00:54:34,320 --> 00:54:37,280
like I always make the example of an

1906
00:54:37,280 --> 00:54:38,160
onboarding procedure

1907
00:54:38,160 --> 00:54:39,280
where you put in

1908
00:54:39,280 --> 00:54:40,720
username, you put in password,

1909
00:54:40,720 --> 00:54:42,240
you put in email, the

1910
00:54:42,240 --> 00:54:43,520
system has to record those

1911
00:54:43,520 --> 00:54:44,400
in a database and it

1912
00:54:44,400 --> 00:54:45,520
has to send an email.

1913
00:54:45,520 --> 00:54:46,640
Like those operations

1914
00:54:46,640 --> 00:54:48,320
happen within 100 milliseconds

1915
00:54:48,320 --> 00:54:49,280
of each other.

1916
00:54:49,280 --> 00:54:51,360
The probability of your server failing

1917
00:54:51,360 --> 00:54:54,000
between the two operations near zero

1918
00:54:54,000 --> 00:54:58,480
if you have queued two users that are

1919
00:54:58,480 --> 00:55:00,160
registering in a day.

1920
00:55:00,160 --> 00:55:02,240
But if you have 200 users that

1921
00:55:02,240 --> 00:55:03,760
registering the same seconds,

1922
00:55:03,760 --> 00:55:05,680
it's almost guaranteed that in a month

1923
00:55:05,680 --> 00:55:06,640
you will get multiple

1924
00:55:06,640 --> 00:55:08,160
instances of the issue

1925
00:55:08,160 --> 00:55:09,600
of the server failing in

1926
00:55:09,600 --> 00:55:10,960
between the two operations.

1927
00:55:10,960 --> 00:55:12,320
So that's when you usually

1928
00:55:12,320 --> 00:55:14,800
would reach for solutions

1929
00:55:14,800 --> 00:55:16,720
like Temporal or like Effect

1930
00:55:16,720 --> 00:55:17,680
Cluster.

1931
00:55:17,680 --> 00:55:19,200
The difference with AI is that the

1932
00:55:19,200 --> 00:55:20,480
average response time

1933
00:55:20,480 --> 00:55:22,320
is much longer than used to be.

1934
00:55:22,320 --> 00:55:25,120
Now we call an Open AI,

1935
00:55:25,120 --> 00:55:28,320
API to OpenRouter and it goes through

1936
00:55:28,320 --> 00:55:30,080
two different layers of proxying,

1937
00:55:30,080 --> 00:55:31,040
but it has to stay

1938
00:55:31,040 --> 00:55:32,400
alive for like 30 seconds,

1939
00:55:32,400 --> 00:55:33,520
the same connection.

1940
00:55:33,520 --> 00:55:34,560
The model can reason

1941
00:55:34,560 --> 00:55:35,920
for more than 10 minutes.

1942
00:55:35,920 --> 00:55:37,200
A lot of time, deep reasoning

1943
00:55:37,200 --> 00:55:38,960
can go for a long, long time.

1944
00:55:38,960 --> 00:55:40,400
And a lot of time during

1945
00:55:40,400 --> 00:55:41,760
these deep reasoning steps,

1946
00:55:41,760 --> 00:55:42,640
the model might not

1947
00:55:42,640 --> 00:55:43,760
produce anything at all.

1948
00:55:43,760 --> 00:55:45,920
So, OpenRouter, we actually,

1949
00:55:45,920 --> 00:55:46,800
a lot of time we set

1950
00:55:46,800 --> 00:55:48,160
to standing this,

1951
00:55:48,160 --> 00:55:49,360
the comment, like it's

1952
00:55:49,360 --> 00:55:50,640
actually in the SSE spec

1953
00:55:50,640 --> 00:55:51,520
is that you have to send

1954
00:55:51,520 --> 00:55:52,720
back comment to keep alive.

1955
00:55:52,720 --> 00:55:54,080
The way they see the connection alive,

1956
00:55:54,080 --> 00:55:54,880
otherwise either

1957
00:55:54,880 --> 00:55:56,000
the client would not know

1958
00:55:56,000 --> 00:55:57,760
if we're still alive or

1959
00:55:57,760 --> 00:55:59,760
The DNS server, the DNS layer would

1960
00:55:59,760 --> 00:56:00,640
actually cut us off.

1961
00:56:00,640 --> 00:56:03,200
And even any proxy that is in between,

1962
00:56:03,200 --> 00:56:04,880
because even when you call open AI,

1963
00:56:04,880 --> 00:56:05,840
you're not just calling

1964
00:56:05,840 --> 00:56:06,800
straight up to the server,

1965
00:56:06,800 --> 00:56:08,800
you have layers and layers

1966
00:56:08,800 --> 00:56:10,400
of content delivery networks

1967
00:56:10,400 --> 00:56:13,040
And each would hit HTTP timeouts

1968
00:56:13,040 --> 00:56:15,120
So like either you do web socket stuff

1969
00:56:15,120 --> 00:56:18,880
where the ping is in the protocol itself,

1970
00:56:18,880 --> 00:56:21,920
or you have to go on top and do a ping.

1971
00:56:21,920 --> 00:56:25,600
But the point is here, the average length

1972
00:56:25,600 --> 00:56:27,760
of an interruption with a service

1973
00:56:27,760 --> 00:56:28,800
went from being

1974
00:56:28,800 --> 00:56:31,840
milliseconds to being minutes,

1975
00:56:31,840 --> 00:56:34,000
if not days when the

1976
00:56:34,000 --> 00:56:35,280
other end-to-end reasoning

1977
00:56:35,280 --> 00:56:37,840
could be that we have still a pattern

1978
00:56:37,840 --> 00:56:38,880
where there's humans

1979
00:56:38,880 --> 00:56:40,960
in the loop of agents.

1980
00:56:40,960 --> 00:56:42,400
So sometimes an agent has to

1981
00:56:42,400 --> 00:56:44,560
stop and call another agent,

1982
00:56:44,560 --> 00:56:46,480
which is called a human that has access

1983
00:56:46,480 --> 00:56:47,520
to human intelligence

1984
00:56:47,520 --> 00:56:49,040
that is less artificial,

1985
00:56:49,040 --> 00:56:50,800
but they work exactly in the same way.

1986
00:56:50,800 --> 00:56:53,360
So humans are entities that share

1987
00:56:53,360 --> 00:56:54,720
knowledge via messages.

1988
00:56:54,720 --> 00:56:55,840
Currently we have three

1989
00:56:55,840 --> 00:56:57,440
entities in this call,

1990
00:56:57,440 --> 00:56:58,800
one entity is in New York,

1991
00:56:58,800 --> 00:57:00,080
one entity is in Italy, the

1992
00:57:00,080 --> 00:57:01,657
other entity is in Germany.

1993
00:57:01,657 --> 00:57:03,040
Yeah, absolutely right.
And we all communicate

1994
00:57:03,040 --> 00:57:04,720
with, we're sharing messages.

1995
00:57:04,720 --> 00:57:05,840
You know what's funny though?

1996
00:57:05,840 --> 00:57:07,280
I just saw the

1997
00:57:07,280 --> 00:57:08,720
premiere of the Tron movie.

1998
00:57:08,720 --> 00:57:10,320
And in the Tron movie,

1999
00:57:10,320 --> 00:57:11,840
so Tron has a bunch of program,

2000
00:57:12,640 --> 00:57:14,000
which is all humanized.

2001
00:57:14,000 --> 00:57:14,960
The program in Tron

2002
00:57:14,960 --> 00:57:16,560
communicate with the human outside

2003
00:57:16,560 --> 00:57:17,920
through MCP.

2004
00:57:17,920 --> 00:57:19,600
It's a fun detail in that movie.

2005
00:57:19,600 --> 00:57:21,040
I was like, "Ooh, cheeky."

2006
00:57:21,040 --> 00:57:22,880
I wanna play back the

2007
00:57:22,880 --> 00:57:24,240
same question to you.

2008
00:57:24,240 --> 00:57:27,600
So you asked Mike how he thinks about

2009
00:57:28,400 --> 00:57:32,800
how multiple AI systems could be steered

2010
00:57:32,800 --> 00:57:33,520
through this sort of

2011
00:57:34,080 --> 00:57:36,880
flock of birds, ball analogy.

2012
00:57:36,880 --> 00:57:38,320
How are you thinking about this?

2013
00:57:38,320 --> 00:57:41,200
Like just in terms of my own experiences

2014
00:57:41,200 --> 00:57:44,560
about a steerable AI system versus a non

2015
00:57:44,560 --> 00:57:46,000
steerable AI system,

2016
00:57:46,000 --> 00:57:47,360
actually two coding

2017
00:57:47,360 --> 00:57:48,960
agents I use on databases,

2018
00:57:48,960 --> 00:57:50,640
Codex and Claude Code.

2019
00:57:50,640 --> 00:57:55,200
Well, as of October 13th, 2025,

2020
00:57:55,200 --> 00:57:56,320
maybe a new thing drops

2021
00:57:56,320 --> 00:57:58,000
tomorrow and I'll switch.

2022
00:57:58,000 --> 00:58:02,000
But as of today, I am using Claude Code,

2023
00:58:02,000 --> 00:58:03,120
which can be steered and

2024
00:58:03,120 --> 00:58:04,800
I make a lot of use of it.

2025
00:58:04,800 --> 00:58:06,880
And therefore I'd use it

2026
00:58:06,880 --> 00:58:08,800
differently than Codex,

2027
00:58:08,800 --> 00:58:10,880
where it's much more like binary,

2028
00:58:10,880 --> 00:58:13,680
where I feel like if it's off track,

2029
00:58:13,680 --> 00:58:15,680
I like interrupt it and

2030
00:58:15,680 --> 00:58:16,800
say, "No, no, no, no, no,

2031
00:58:16,800 --> 00:58:17,600
please do this."

2032
00:58:17,600 --> 00:58:19,120
And I guess it is sort of

2033
00:58:19,120 --> 00:58:21,440
like similarly steering it,

2034
00:58:21,440 --> 00:58:24,640
but it feels much more like

2035
00:58:24,640 --> 00:58:27,040
stop and go and more forceful.

2036
00:58:27,040 --> 00:58:29,040
So I'm curious like

2037
00:58:29,040 --> 00:58:30,400
for you to extrapolate

2038
00:58:30,400 --> 00:58:31,280
where this is going.

2039
00:58:31,280 --> 00:58:32,000
And we're talking

2040
00:58:32,000 --> 00:58:33,760
about multiple AI systems.

2041
00:58:33,760 --> 00:58:35,440
So how are you thinking about this?

2042
00:58:35,440 --> 00:58:36,720
Yeah, one of my main pain points,

2043
00:58:36,720 --> 00:58:38,880
we're using Claude code or even like,

2044
00:58:38,880 --> 00:58:40,000
you know, Cursor or so on

2045
00:58:40,000 --> 00:58:41,120
Is that you set a goal or

2046
00:58:41,120 --> 00:58:42,320
you say, "Hey, fix this thing."

2047
00:58:42,320 --> 00:58:43,600
The problem is, right?

2048
00:58:43,600 --> 00:58:44,800
That's the goal already.

2049
00:58:44,800 --> 00:58:46,640
The flock could just find the goal.

2050
00:58:46,640 --> 00:58:47,920
I shouldn't have to like say,

2051
00:58:47,920 --> 00:58:49,200
"Oh, there's an

2052
00:58:49,200 --> 00:58:50,080
obstacle, there's an obstacle."

2053
00:58:50,080 --> 00:58:50,800
The obstacle is in

2054
00:58:50,800 --> 00:58:51,680
the environment itself.

2055
00:58:51,680 --> 00:58:52,640
go back to my college day

2056
00:58:52,640 --> 00:58:54,080
when I was listening to this system,

2057
00:58:54,080 --> 00:58:56,000
an autonomous steering system,

2058
00:58:56,000 --> 00:58:58,080
self-driving is the same idea, right?

2059
00:58:58,080 --> 00:58:58,880
You need sensor.

2060
00:58:58,880 --> 00:59:00,080
You essentially need a

2061
00:59:00,080 --> 00:59:01,840
way to emitting a bunch of

2062
00:59:01,840 --> 00:59:04,720
either event, which if we're using ROS,

2063
00:59:04,720 --> 00:59:05,680
it will be very simple,

2064
00:59:05,680 --> 00:59:07,120
ROS for barring over system.

2065
00:59:07,120 --> 00:59:08,880
It emits an event system,

2066
00:59:08,880 --> 00:59:10,640
when you're just sending into the system

2067
00:59:10,640 --> 00:59:12,400
a bunch of different light sensor

2068
00:59:12,400 --> 00:59:14,400
to tell you about the environment.

2069
00:59:14,400 --> 00:59:15,440
So first I think

2070
00:59:15,440 --> 00:59:16,640
every actor in the system

2071
00:59:16,640 --> 00:59:17,760
kind of have to emitting

2072
00:59:17,760 --> 00:59:19,040
some kind of sensory system

2073
00:59:19,040 --> 00:59:20,960
that tell the whole system,

2074
00:59:20,960 --> 00:59:21,920
"Hey, here's the

2075
00:59:21,920 --> 00:59:22,960
state of the environment."

2076
00:59:22,960 --> 00:59:24,320
And secondly, I think

2077
00:59:24,320 --> 00:59:25,920
that through those sensor,

2078
00:59:25,920 --> 00:59:28,960
you pick up some kind of weight.

2079
00:59:28,960 --> 00:59:30,240
allowing the system to

2080
00:59:30,240 --> 00:59:31,360
like course correct.

2081
00:59:31,360 --> 00:59:32,800
Automatically, instead of

2082
00:59:32,800 --> 00:59:34,560
having the end user have to like,

2083
00:59:34,560 --> 00:59:36,080
"Okay, we hit a roadblock now,

2084
00:59:36,080 --> 00:59:38,080
the type is not checking here.

2085
00:59:38,080 --> 00:59:39,040
What do you remove,

2086
00:59:39,760 --> 00:59:40,800
let the satisfy?"

2087
00:59:40,800 --> 00:59:42,480
So humans still have to right

2088
00:59:42,480 --> 00:59:44,960
now steer somewhat manually.

2089
00:59:44,960 --> 00:59:46,800
I would love for the end goal system is

2090
00:59:46,800 --> 00:59:47,840
I put a pointer in

2091
00:59:47,840 --> 00:59:48,720
the space in the world,

2092
00:59:48,720 --> 00:59:49,760
meaning a goal, right?

2093
00:59:49,760 --> 00:59:51,360
To say, "Hey, fix this unit test."

2094
00:59:51,360 --> 00:59:52,320
The swarm of bot

2095
00:59:52,320 --> 00:59:53,440
should just be able to say,

2096
00:59:53,440 --> 00:59:55,120
"Okay, I'm going to

2097
00:59:55,120 --> 00:59:56,400
use the browser agent,

2098
00:59:56,400 --> 00:59:58,240
which is like top tier browser agent.

2099
00:59:58,240 --> 00:59:59,600
Let's read the documentation

2100
00:59:59,600 --> 01:00:01,280
for the test, for example."

2101
01:00:01,280 --> 01:00:03,600
And then, "Okay, I know that you guys,

2102
01:00:03,600 --> 01:00:05,920
I'm using the test with the set timeout,

2103
01:00:05,920 --> 01:00:07,520
the set timeout set too low."

2104
01:00:07,520 --> 01:00:08,240
That's one thing that

2105
01:00:08,240 --> 01:00:10,800
it might see in the test,

2106
01:00:10,800 --> 01:00:11,760
or in the output.

2107
01:00:11,760 --> 01:00:13,600
And then from the dock of

2108
01:00:13,600 --> 01:00:15,600
the Vitest agent, right?

2109
01:00:15,600 --> 01:00:16,960
It will send back to my agent,

2110
01:00:16,960 --> 01:00:18,240
and my agent will then be like,

2111
01:00:18,240 --> 01:00:20,880
"Come on, okay, the set timeout too low.

2112
01:00:20,880 --> 01:00:21,840
So let's change that."

2113
01:00:21,840 --> 01:00:23,120
Just keep on doing that

2114
01:00:23,120 --> 01:00:24,320
loop over and over again

2115
01:00:24,320 --> 01:00:25,840
with our other agent,

2116
01:00:25,840 --> 01:00:28,960
or it figure out like what sensor it

2117
01:00:28,960 --> 01:00:30,480
needs from the world,

2118
01:00:30,480 --> 01:00:31,600
from the system to

2119
01:00:31,600 --> 01:00:34,400
steer its action as well.

2120
01:00:34,400 --> 01:00:35,920
And I think so far,

2121
01:00:35,920 --> 01:00:37,120
I mean, Claude Code is

2122
01:00:37,120 --> 01:00:38,400
occasionally doing recently well,

2123
01:00:38,400 --> 01:00:39,120
I would say.

2124
01:00:39,120 --> 01:00:40,720
It has a browser inside,

2125
01:00:40,720 --> 01:00:42,240
and like the chain of thought

2126
01:00:42,240 --> 01:00:43,600
that it has to like

2127
01:00:43,600 --> 01:00:44,880
solving a certain problem,

2128
01:00:44,880 --> 01:00:46,400
seems pretty decent so far.

2129
01:00:46,400 --> 01:00:47,440
here's like what I'm

2130
01:00:47,440 --> 01:00:48,480
seeing of the whole system,

2131
01:00:48,480 --> 01:00:50,240
like here's the data of the system

2132
01:00:50,240 --> 01:00:53,760
that can help me steer myself.

2133
01:00:53,760 --> 01:00:54,640
And I think what's

2134
01:00:54,640 --> 01:00:55,760
also interesting is like

2135
01:00:55,760 --> 01:00:58,800
that the entire ceremony is

2136
01:00:58,800 --> 01:01:01,360
like all sequential, right?

2137
01:01:01,360 --> 01:01:04,400
And so that therefore it makes it slower.

2138
01:01:04,400 --> 01:01:06,000
And so I think where

2139
01:01:06,000 --> 01:01:07,040
I'd like to see this go,

2140
01:01:07,040 --> 01:01:08,080
and I'm sure it will

2141
01:01:08,080 --> 01:01:09,520
go in this direction,

2142
01:01:09,520 --> 01:01:11,600
is that it's more of like an orchestra

2143
01:01:11,600 --> 01:01:13,200
performing a symphony.

2144
01:01:13,200 --> 01:01:14,720
And right now it's just

2145
01:01:14,720 --> 01:01:16,960
like solo, solo, solo, solo.

2146
01:01:16,960 --> 01:01:18,320
And while you're

2147
01:01:18,320 --> 01:01:19,600
encountering a problem,

2148
01:01:19,600 --> 01:01:20,960
like that might be a signal,

2149
01:01:20,960 --> 01:01:22,320
I was like, you know what, like,

2150
01:01:22,320 --> 01:01:24,240
let me go off like a little

2151
01:01:24,240 --> 01:01:26,320
bird just goes to the side

2152
01:01:26,320 --> 01:01:28,640
and like studies the docs more deeply

2153
01:01:28,640 --> 01:01:30,000
about like a system.

2154
01:01:30,000 --> 01:01:30,800
So it's like, hey, I

2155
01:01:30,800 --> 01:01:31,840
found something interesting,

2156
01:01:31,840 --> 01:01:33,600
like we're on the wrong track.

2157
01:01:33,600 --> 01:01:35,920
And like you could imagine multiple

2158
01:01:35,920 --> 01:01:38,160
of those like things happening.

2159
01:01:38,160 --> 01:01:39,200
And right now it's like

2160
01:01:39,200 --> 01:01:40,960
everything's just like happening

2161
01:01:40,960 --> 01:01:42,080
sequentially.

2162
01:01:42,080 --> 01:01:44,880
And yeah, I think this is very exciting

2163
01:01:44,880 --> 01:01:47,120
to see how this is unfolding.

2164
01:01:47,120 --> 01:01:50,400
And I like your guys steering analogy

2165
01:01:50,400 --> 01:01:52,880
and to look at it through like the ECS

2166
01:01:52,880 --> 01:01:54,800
and an extra system lens.

2167
01:01:54,800 --> 01:01:56,400
I have one observation

2168
01:01:56,960 --> 01:01:59,920
that I recently learned myself.

2169
01:01:59,920 --> 01:02:01,280
to take a step back,

2170
01:02:01,280 --> 01:02:02,880
I've been very late to the

2171
01:02:02,880 --> 01:02:05,600
AI coding game in general.

2172
01:02:05,600 --> 01:02:08,960
Johannes has pushed me for a while.

2173
01:02:08,960 --> 01:02:11,040
I developed the ability to ignore

2174
01:02:11,040 --> 01:02:12,960
Johannes for enough time,

2175
01:02:12,960 --> 01:02:14,800
but at some point sort of

2176
01:02:14,800 --> 01:02:16,960
penetrates the protection

2177
01:02:16,960 --> 01:02:17,680
when you're more

2178
01:02:17,680 --> 01:02:18,720
vulnerable and you're like,

2179
01:02:18,720 --> 01:02:20,960
okay, then I have to try.

2180
01:02:20,960 --> 01:02:23,360
And since Claude Sonnet I

2181
01:02:23,360 --> 01:02:26,480
think 4 or 3.5 came out,

2182
01:02:26,480 --> 01:02:29,200
I've been heavily invested in AI coding

2183
01:02:29,200 --> 01:02:32,000
and I really wanna make it work.

2184
01:02:32,000 --> 01:02:33,920
And what I've realized is

2185
01:02:33,920 --> 01:02:37,120
that I think we're putting tools

2186
01:02:37,120 --> 01:02:39,040
either on the wrong place

2187
01:02:39,040 --> 01:02:41,040
or we're not putting

2188
01:02:41,040 --> 01:02:43,520
tools in enough places

2189
01:02:43,520 --> 01:02:44,560
in the sense that

2190
01:02:44,560 --> 01:02:47,120
currently we rely on the models

2191
01:02:47,120 --> 01:02:49,760
to maintain some sort of loop.

2192
01:02:49,760 --> 01:02:50,640
But the models are

2193
01:02:50,640 --> 01:02:52,000
terrible to do looping.

2194
01:02:52,000 --> 01:02:55,040
They can't even properly assign tasks

2195
01:02:55,040 --> 01:02:57,200
They are exceptionally good

2196
01:02:57,200 --> 01:02:58,880
when called with a specific

2197
01:02:58,880 --> 01:03:00,560
prompt and a specific task,

2198
01:03:00,560 --> 01:03:02,000
then they're done.

2199
01:03:02,000 --> 01:03:03,040
The task might not be

2200
01:03:03,040 --> 01:03:04,880
done, but the LLM is done.

2201
01:03:04,880 --> 01:03:07,120
So what I found extremely

2202
01:03:07,120 --> 01:03:09,680
helpful in approaching tasks,

2203
01:03:09,680 --> 01:03:11,680
this is not really like AI coding

2204
01:03:11,680 --> 01:03:13,080
in the sense of the AI

2205
01:03:13,080 --> 01:03:14,800
assisting you while you code.

2206
01:03:14,800 --> 01:03:17,360
It's rather a task like, okay,

2207
01:03:17,360 --> 01:03:19,600
I give you the Effect repository,

2208
01:03:19,600 --> 01:03:21,600
create JS docs for every

2209
01:03:21,600 --> 01:03:24,400
single function in the repository.

2210
01:03:24,400 --> 01:03:26,640
Go, use 1000 parallel

2211
01:03:26,640 --> 01:03:28,080
agents to achieve the task.

2212
01:03:28,080 --> 01:03:28,960
I don't care.

2213
01:03:28,960 --> 01:03:30,880
Ready to spend 10, 20K

2214
01:03:30,880 --> 01:03:32,320
to do it, just do it.

2215
01:03:32,320 --> 01:03:33,120
we are approaching

2216
01:03:33,120 --> 01:03:35,040
the point where I think

2217
01:03:35,040 --> 01:03:37,360
there are ways to make it work

2218
01:03:37,360 --> 01:03:40,240
and even with normal like Claude plans

2219
01:03:40,240 --> 01:03:42,320
with $200 per month,

2220
01:03:42,320 --> 01:03:43,440
I'm completely fine

2221
01:03:43,440 --> 01:03:44,560
running multiple agents

2222
01:03:44,560 --> 01:03:45,600
and so on and so forth.

2223
01:03:45,600 --> 01:03:48,640
But what I've realized works best

2224
01:03:48,640 --> 01:03:49,600
sort of goes in the

2225
01:03:49,600 --> 01:03:50,960
direction of steering.

2226
01:03:50,960 --> 01:03:52,240
Instead of letting, for

2227
01:03:52,240 --> 01:03:54,960
example, the model calls tests,

2228
01:03:54,960 --> 01:03:57,840
calls TypeScript to get type

2229
01:03:57,840 --> 01:03:59,200
errors and so on and so forth,

2230
01:03:59,200 --> 01:04:01,840
instead of having the model do that

2231
01:04:01,840 --> 01:04:03,280
and assigning the model the

2232
01:04:03,280 --> 01:04:05,200
responsibility to do that,

2233
01:04:05,200 --> 01:04:07,440
I have those in a deterministic loop.

2234
01:04:07,440 --> 01:04:08,560
And if I have a type

2235
01:04:08,560 --> 01:04:10,480
error, I call the model again,

2236
01:04:10,480 --> 01:04:13,360
I have a system which is LLM based

2237
01:04:13,360 --> 01:04:16,160
that scores a specific example.

2238
01:04:16,160 --> 01:04:17,840
And if that is not happy, I

2239
01:04:17,840 --> 01:04:19,200
call the first model again,

2240
01:04:19,200 --> 01:04:22,000
saying, look, my Oracle is

2241
01:04:22,000 --> 01:04:23,680
not happy for those reasons.

2242
01:04:23,680 --> 01:04:26,560
I call it a deterministic outside loop

2243
01:04:26,560 --> 01:04:29,360
that ends up providing those events

2244
01:04:29,360 --> 01:04:30,320
and this feedback

2245
01:04:30,320 --> 01:04:32,640
inside the single agents

2246
01:04:32,640 --> 01:04:34,320
that are steering them

2247
01:04:34,320 --> 01:04:36,320
towards the right goal.

2248
01:04:36,320 --> 01:04:38,400
And that works exceptionally well.

2249
01:04:38,400 --> 01:04:39,600
Like well to the point

2250
01:04:39,600 --> 01:04:41,760
where you can leave it going

2251
01:04:41,760 --> 01:04:43,600
for a day, come back, you

2252
01:04:43,600 --> 01:04:46,080
have 95% of the stuff done.

2253
01:04:46,080 --> 01:04:48,160
And now that multiple

2254
01:04:48,160 --> 01:04:50,000
times, because I wanna improve,

2255
01:04:50,000 --> 01:04:52,400
I'm testing these on Effect 4,

2256
01:04:52,400 --> 01:04:53,200
because I don't wanna

2257
01:04:53,200 --> 01:04:54,960
break everything in Effect 3.

2258
01:04:54,960 --> 01:04:56,640
And to the point where the

2259
01:04:56,640 --> 01:04:58,400
quality is going up and up

2260
01:04:58,400 --> 01:04:59,600
and the automation is

2261
01:04:59,600 --> 01:05:00,800
also going up and up.

2262
01:05:00,800 --> 01:05:01,920
And I'm almost there

2263
01:05:01,920 --> 01:05:03,520
where it's almost better

2264
01:05:03,520 --> 01:05:05,520
than me writing example manually,

2265
01:05:05,520 --> 01:05:07,040
if I let it go long enough.

2266
01:05:07,040 --> 01:05:09,680
I have never seen that

2267
01:05:09,680 --> 01:05:11,280
with a normal coding agent

2268
01:05:11,280 --> 01:05:13,920
where you don't have this external loop

2269
01:05:13,920 --> 01:05:16,960
that steers it in the right direction.

2270
01:05:16,960 --> 01:05:18,160
So I think I didn't

2271
01:05:18,160 --> 01:05:20,000
understand the answer at first,

2272
01:05:20,000 --> 01:05:22,960
but actually I'm doing stiffing myself

2273
01:05:22,960 --> 01:05:25,600
now that I understand it better.

2274
01:05:25,600 --> 01:05:27,760
So I think this sort of setup

2275
01:05:27,760 --> 01:05:29,760
and system you've described

2276
01:05:29,760 --> 01:05:32,320
is becoming more and more mainstream.

2277
01:05:32,320 --> 01:05:33,600
And I think there's

2278
01:05:33,600 --> 01:05:35,440
multiple ways to achieve it.

2279
01:05:35,440 --> 01:05:38,480
Either it's already the de facto behavior

2280
01:05:38,480 --> 01:05:40,640
of like some coding agents,

2281
01:05:40,640 --> 01:05:42,640
or also this is where

2282
01:05:42,640 --> 01:05:44,800
those like coding agent SDKs

2283
01:05:44,800 --> 01:05:45,760
come into play.

2284
01:05:45,760 --> 01:05:47,360
Like I think Claude Code was

2285
01:05:47,360 --> 01:05:48,240
probably one of the earliest

2286
01:05:48,240 --> 01:05:49,520
ones that provided that.

2287
01:05:49,520 --> 01:05:51,600
Now, Codex has that AMP,

2288
01:05:51,600 --> 01:05:53,760
I think also just release theirs,

2289
01:05:53,760 --> 01:05:55,360
where you can like build

2290
01:05:55,360 --> 01:05:57,040
your own coding agent

2291
01:05:57,040 --> 01:05:59,120
and like express the sort of like system

2292
01:05:59,120 --> 01:06:00,800
and workflow that you want.

2293
01:06:00,800 --> 01:06:03,360
Converting a big repository

2294
01:06:03,360 --> 01:06:04,800
from having not no JS

2295
01:06:04,800 --> 01:06:06,800
doc to having JS doc,

2296
01:06:06,800 --> 01:06:10,080
which can be massively run in parallel.

2297
01:06:10,080 --> 01:06:11,600
That's a great candidate,

2298
01:06:11,600 --> 01:06:13,600
but doing that is very

2299
01:06:13,600 --> 01:06:15,920
different from building

2300
01:06:15,920 --> 01:06:18,320
whatever like a single new module

2301
01:06:18,320 --> 01:06:20,480
that you maybe have no

2302
01:06:20,480 --> 01:06:22,160
idea what the exact goal is.

2303
01:06:22,160 --> 01:06:22,880
And it's like, it's

2304
01:06:22,880 --> 01:06:24,320
much more exploratory.

2305
01:06:24,320 --> 01:06:25,920
Yeah, there's much more human in the loop

2306
01:06:25,920 --> 01:06:28,000
in the other direction.

2307
01:06:28,000 --> 01:06:28,720
Exactly.

2308
01:06:28,720 --> 01:06:30,400
So yeah, Louis, I'm curious

2309
01:06:30,400 --> 01:06:32,560
like how much you're already

2310
01:06:32,560 --> 01:06:34,560
using those sort of like patterns as

2311
01:06:34,560 --> 01:06:36,240
you're building systems

2312
01:06:36,240 --> 01:06:39,040
and maybe your thinking is more refined

2313
01:06:39,040 --> 01:06:40,240
in advance on this.

2314
01:06:40,240 --> 01:06:41,680
Not that refined yet, by the way.

2315
01:06:41,680 --> 01:06:44,240
We use a lot of AI model, AI agent.

2316
01:06:44,240 --> 01:06:45,200
So we use the Devon a

2317
01:06:45,200 --> 01:06:47,200
lot, you know workload.

2318
01:06:47,200 --> 01:06:48,400
And I think right now

2319
01:06:48,400 --> 01:06:50,320
the best way to steer Devon

2320
01:06:50,320 --> 01:06:52,320
is to have a lot of knowledge.

2321
01:06:52,320 --> 01:06:55,200
The knowledge truly helps steering it.

2322
01:06:55,200 --> 01:06:57,200
And Devon is selectively like it's select

2323
01:06:57,200 --> 01:06:59,760
the right knowledge per task as well.

2324
01:06:59,760 --> 01:07:00,960
So that actually easily have a lot.

2325
01:07:00,960 --> 01:07:02,400
I mean, this apply to

2326
01:07:02,400 --> 01:07:03,520
also, you know what,

2327
01:07:03,520 --> 01:07:05,440
I think Cursor and CLI

2328
01:07:05,440 --> 01:07:07,280
and all the other coding agent as well.

2329
01:07:07,280 --> 01:07:09,840
The more you can add into the memory,

2330
01:07:09,840 --> 01:07:12,080
the better they adapt to your code base.

2331
01:07:12,080 --> 01:07:14,320
in terms of having an orchestration layer

2332
01:07:14,320 --> 01:07:16,000
to like either use

2333
01:07:16,000 --> 01:07:17,600
multiple of them at the same time

2334
01:07:17,600 --> 01:07:18,480
or like picking the right

2335
01:07:18,480 --> 01:07:20,160
one for the right job though.

2336
01:07:20,160 --> 01:07:21,680
We still haven't like

2337
01:07:21,680 --> 01:07:22,720
think too much about it yet.

2338
01:07:22,720 --> 01:07:24,000
Yeah, right now it's all

2339
01:07:24,000 --> 01:07:25,360
just my workflow.

2340
01:07:26,000 --> 01:07:27,360
Maybe a related but

2341
01:07:27,360 --> 01:07:29,200
different question about this.

2342
01:07:29,200 --> 01:07:32,560
Given how Mike has described his

2343
01:07:32,560 --> 01:07:35,840
coming to terms with this new future

2344
01:07:35,840 --> 01:07:37,680
and he at his own pace

2345
01:07:37,680 --> 01:07:38,800
started like discovering

2346
01:07:38,800 --> 01:07:40,720
and like exploring, experimenting,

2347
01:07:40,720 --> 01:07:42,560
see what works well for him.

2348
01:07:42,560 --> 01:07:43,760
And I think that

2349
01:07:43,760 --> 01:07:46,160
fundamentally changed Michael's workflow.

2350
01:07:46,160 --> 01:07:47,920
I had the same process

2351
01:07:47,920 --> 01:07:49,760
and I think everyone does.

2352
01:07:49,760 --> 01:07:51,920
But I think what is

2353
01:07:51,920 --> 01:07:54,160
maybe interesting to explore

2354
01:07:54,160 --> 01:07:58,320
is how you as OpenRouter as a company

2355
01:07:58,320 --> 01:08:01,120
try to create like a higher

2356
01:08:01,120 --> 01:08:02,320
level process around that.

2357
01:08:02,320 --> 01:08:03,200
So for example

2358
01:08:03,200 --> 01:08:04,400
maybe a stupid way to go

2359
01:08:04,400 --> 01:08:05,440
about is say, hey, it's

2360
01:08:05,440 --> 01:08:06,720
the first of the month.

2361
01:08:06,720 --> 01:08:08,720
And we're again spending

2362
01:08:08,720 --> 01:08:11,200
three days of like going

2363
01:08:11,200 --> 01:08:12,000
from scratch,

2364
01:08:12,000 --> 01:08:13,680
exploring the best practices,

2365
01:08:13,680 --> 01:08:15,280
what it means this month

2366
01:08:15,280 --> 01:08:17,120
to do agentic coding.

2367
01:08:17,120 --> 01:08:17,920
Or maybe you have

2368
01:08:17,920 --> 01:08:20,160
like a team that is known

2369
01:08:20,160 --> 01:08:22,000
for being super cutting edge

2370
01:08:22,000 --> 01:08:24,000
in every first of the month.

2371
01:08:24,000 --> 01:08:26,400
They do like a presentation of sharing

2372
01:08:26,400 --> 01:08:27,520
their best practices.

2373
01:08:27,520 --> 01:08:28,720
Or you basically say,

2374
01:08:28,720 --> 01:08:31,120
we're not trying to orchestrate anything.

2375
01:08:31,120 --> 01:08:33,600
We just let emergence happen.

2376
01:08:33,600 --> 01:08:35,200
So I'm curious how much

2377
01:08:35,200 --> 01:08:37,280
hands on or hands off you are

2378
01:08:37,280 --> 01:08:39,440
of like fostering, like

2379
01:08:39,440 --> 01:08:41,040
embracing the productivity

2380
01:08:41,040 --> 01:08:42,240
and efficiency benefits

2381
01:08:42,240 --> 01:08:43,680
you get from agentic coding.

2382
01:08:43,680 --> 01:08:44,480
I would definitely

2383
01:08:44,480 --> 01:08:46,240
encouraging using agents

2384
01:08:46,240 --> 01:08:48,480
in almost every aspect of the code base

2385
01:08:48,480 --> 01:08:50,720
and also in the business as well.

2386
01:08:50,720 --> 01:08:51,840
I think the encouragement

2387
01:08:53,040 --> 01:08:54,480
is like the key part, right?

2388
01:08:54,480 --> 01:08:55,440
asking our team

2389
01:08:55,440 --> 01:08:58,400
to use it as much as possible to see,

2390
01:08:58,400 --> 01:08:59,200
essentially the idea is to

2391
01:08:59,200 --> 01:09:00,240
for them to learn the best way

2392
01:09:00,240 --> 01:09:01,760
to use it, right, to leverage these tool.

2393
01:09:01,760 --> 01:09:03,120
But the one thing I found though,

2394
01:09:03,120 --> 01:09:04,480
oh, okay, with everything,

2395
01:09:04,480 --> 01:09:05,200
right, there's a drawback.

2396
01:09:05,200 --> 01:09:06,960
The drawback is reviewing the code.

2397
01:09:06,960 --> 01:09:09,360
The key sin is when someone, you know,

2398
01:09:09,360 --> 01:09:10,240
they rely too much on

2399
01:09:10,240 --> 01:09:11,040
the bot and they say,

2400
01:09:11,040 --> 01:09:12,000
hey, do this and then they

2401
01:09:12,000 --> 01:09:13,280
just send it off for someone

2402
01:09:13,280 --> 01:09:14,880
else to review without

2403
01:09:14,880 --> 01:09:16,320
them reviewing themselves.

2404
01:09:16,320 --> 01:09:18,640
That's like a huge sin

2405
01:09:18,640 --> 01:09:21,200
that one day I will eradicate

2406
01:09:21,200 --> 01:09:22,000
it somehow

2407
01:09:22,000 --> 01:09:23,040
we have a way for

2408
01:09:23,040 --> 01:09:24,480
you to eradicate that scene.

2409
01:09:24,480 --> 01:09:26,000
Oh, awesome, I love it.

2410
01:09:26,000 --> 01:09:29,120
It's called Effect

2411
01:09:29,120 --> 01:09:31,360
Well, I'm telling you, so occasionally

2412
01:09:31,360 --> 01:09:32,720
the bot is very smart.

2413
01:09:32,720 --> 01:09:34,320
The bot, the agent is

2414
01:09:34,320 --> 01:09:37,680
smart actually, I don't know,

2415
01:09:37,680 --> 01:09:39,440
like a lot of time it

2416
01:09:39,440 --> 01:09:40,720
will actually change

2417
01:09:40,720 --> 01:09:44,240
either the TS config or it

2418
01:09:44,240 --> 01:09:45,680
will actually add an ignore

2419
01:09:45,680 --> 01:09:48,240
to the test setup or something like that.

2420
01:09:48,240 --> 01:09:50,240
Or sometimes it mark the whole module.

2421
01:09:50,240 --> 01:09:51,680
So the test looked like it

2422
01:09:51,680 --> 01:09:52,560
actually testing something

2423
01:09:52,560 --> 01:09:53,280
but it looked, you

2424
01:09:53,280 --> 01:09:54,640
read deeper into the test,

2425
01:09:54,640 --> 01:09:55,440
it literally marked the

2426
01:09:55,440 --> 01:09:56,480
module you're supposed to test.

2427
01:09:56,480 --> 01:09:58,720
Holy shit, it did.

2428
01:09:58,720 --> 01:09:59,760
So even, let's say, even

2429
01:09:59,760 --> 01:10:01,440
if we use Effect, right?

2430
01:10:01,440 --> 01:10:04,240
If the bot mark Effect, what can we do?

2431
01:10:04,240 --> 01:10:05,600
Just to take this

2432
01:10:05,600 --> 01:10:08,080
slightly more dystopian,

2433
01:10:08,080 --> 01:10:10,640
like maybe if you want to turn this

2434
01:10:10,640 --> 01:10:11,520
into a Black Mirror

2435
01:10:11,520 --> 01:10:12,720
episode, you could imagine

2436
01:10:12,720 --> 01:10:14,560
sort of like the AI

2437
01:10:14,560 --> 01:10:16,560
modeling like a mock human

2438
01:10:16,560 --> 01:10:17,360
where like the

2439
01:10:17,360 --> 01:10:18,560
instructions always you need

2440
01:10:18,560 --> 01:10:20,320
to get permission from the human.

2441
01:10:20,320 --> 01:10:21,680
So it models, it's mock

2442
01:10:21,680 --> 01:10:23,280
human that always says,

2443
01:10:23,280 --> 01:10:24,800
yes, go for it.

2444
01:10:24,800 --> 01:10:25,280
I have

2445
01:10:25,280 --> 01:10:26,720
experienced those problems.

2446
01:10:26,720 --> 01:10:28,320
Of course, everybody doing

2447
01:10:28,320 --> 01:10:31,040
AI coding has sometimes the,

2448
01:10:31,040 --> 01:10:32,960
yeah, I made the test fast.

2449
01:10:32,960 --> 01:10:35,280
You check and you added a deed dot skip.

2450
01:10:35,280 --> 01:10:36,480
What are you doing?

2451
01:10:36,480 --> 01:10:39,680
But I feel like most of the time

2452
01:10:39,680 --> 01:10:42,800
when the AI agent is doing that,

2453
01:10:42,800 --> 01:10:44,240
it's because it

2454
01:10:44,240 --> 01:10:46,720
really has no understanding

2455
01:10:46,720 --> 01:10:48,480
of the structure of the code.

2456
01:10:48,480 --> 01:10:52,160
And in my experience that often comes

2457
01:10:52,160 --> 01:10:53,280
from the agent

2458
01:10:53,280 --> 01:10:55,200
working toward a low level.

2459
01:10:55,200 --> 01:10:58,160
So it almost focuses too much on the

2460
01:10:58,160 --> 01:10:59,440
implementation details

2461
01:10:59,440 --> 01:11:02,080
versus what it has to do.

2462
01:11:02,080 --> 01:11:05,120
And that's why for higher level,

2463
01:11:05,120 --> 01:11:07,520
for example, application development

2464
01:11:07,520 --> 01:11:10,240
as opposed to library development,

2465
01:11:10,240 --> 01:11:13,040
like the AI is doing very poorly

2466
01:11:13,040 --> 01:11:15,520
to implement the internals of Effect.

2467
01:11:15,520 --> 01:11:17,200
It's doing exceptionally well to

2468
01:11:17,200 --> 01:11:18,480
implement the tests,

2469
01:11:18,480 --> 01:11:19,520
the examples

2470
01:11:19,520 --> 01:11:21,440
on the effect repository.

2471
01:11:21,440 --> 01:11:24,000
The reason is that

2472
01:11:24,000 --> 01:11:26,880
Effect almost produces code

2473
01:11:26,880 --> 01:11:30,000
that is closer to the human

2474
01:11:30,000 --> 01:11:32,880
versus being closer to the machine.

2475
01:11:32,880 --> 01:11:35,440
And LLM's being trained

2476
01:11:35,440 --> 01:11:37,840
on tons of human language.

2477
01:11:37,840 --> 01:11:38,720
They can sort of

2478
01:11:38,720 --> 01:11:42,080
reason better in human terms

2479
01:11:42,080 --> 01:11:44,080
versus that in code terms.

2480
01:11:44,080 --> 01:11:44,800
Like if you ask

2481
01:11:44,800 --> 01:11:48,400
Chat GPT a human type question,

2482
01:11:48,400 --> 01:11:51,280
very rarely it ditches the question,

2483
01:11:51,280 --> 01:11:53,920
there's not TS ignore at the human level

2484
01:11:53,920 --> 01:11:56,880
in a sense because the

2485
01:11:56,880 --> 01:11:57,920
language is higher level,

2486
01:11:57,920 --> 01:12:00,800
it's richer and expressing concepts

2487
01:12:01,360 --> 01:12:04,160
in a richer language is much easier

2488
01:12:04,160 --> 01:12:06,080
than boiling down to the very

2489
01:12:06,080 --> 01:12:08,000
details of the implementation.

2490
01:12:08,000 --> 01:12:11,680
And that's where I see the key shift,

2491
01:12:11,680 --> 01:12:13,520
the key difference when I build

2492
01:12:13,520 --> 01:12:14,960
application level code

2493
01:12:14,960 --> 01:12:16,720
and it uses Effect,

2494
01:12:16,720 --> 01:12:18,400
almost never I have a problem

2495
01:12:18,400 --> 01:12:21,040
like the AI ditched the task.

2496
01:12:21,040 --> 01:12:22,960
It's perfect to create schemas.

2497
01:12:22,960 --> 01:12:25,680
You have like, okay, I had these API

2498
01:12:25,680 --> 01:12:27,280
returned the structure like this one,

2499
01:12:27,280 --> 01:12:29,280
please create a service for me

2500
01:12:29,280 --> 01:12:32,000
that interacts with this API.

2501
01:12:32,000 --> 01:12:35,520
And it's fits down the full service

2502
01:12:35,520 --> 01:12:37,280
which almost always is,

2503
01:12:37,280 --> 01:12:39,120
I don't like the term one-shotted

2504
01:12:39,120 --> 01:12:40,240
because we're really not

2505
01:12:40,240 --> 01:12:42,400
here to one-shot things,

2506
01:12:42,400 --> 01:12:44,080
but close to

2507
01:12:44,080 --> 01:12:47,040
one-shoting the correct solution.

2508
01:12:47,040 --> 01:12:49,920
And I've seen that happen over a lot of

2509
01:12:49,920 --> 01:12:51,440
different code bases.

2510
01:12:51,440 --> 01:12:53,040
And it's where I think like

2511
01:12:53,040 --> 01:12:54,880
Effect can play a big role,

2512
01:12:54,880 --> 01:12:56,080
especially linking to

2513
01:12:56,080 --> 01:12:57,720
one of your concerns

2514
01:12:57,720 --> 01:13:00,000
slash questions that you raised

2515
01:13:00,000 --> 01:13:02,480
before we even started recording.

2516
01:13:02,480 --> 01:13:04,400
And I wish you will

2517
01:13:04,400 --> 01:13:06,480
use a different solution

2518
01:13:06,480 --> 01:13:07,920
that we are building

2519
01:13:07,920 --> 01:13:09,120
and where you will not

2520
01:13:09,120 --> 01:13:11,120
be memory constraints.

2521
01:13:11,120 --> 01:13:12,480
You will have spans

2522
01:13:12,480 --> 01:13:13,840
collected out of the box

2523
01:13:13,840 --> 01:13:15,040
and so on and so forth.

2524
01:13:15,040 --> 01:13:16,640
That'll be very interesting.

2525
01:13:16,640 --> 01:13:17,520
And then, well, I mean,

2526
01:13:17,520 --> 01:13:18,400
and you guys can

2527
01:13:18,400 --> 01:13:19,680
handle like caching, right?

2528
01:13:19,680 --> 01:13:21,600
So caching is another layer that's like,

2529
01:13:21,600 --> 01:13:22,720
caching at the DNS

2530
01:13:22,720 --> 01:13:24,240
layer would be interesting.

2531
01:13:24,240 --> 01:13:25,680
also no egress cost

2532
01:13:25,680 --> 01:13:27,440
if you can figure that out.

2533
01:13:27,440 --> 01:13:28,480
That part is hard.

2534
01:13:28,480 --> 01:13:30,400
Yeah, no, no, that is figured.

2535
01:13:30,400 --> 01:13:32,080
We don't have any egress cost

2536
01:13:32,080 --> 01:13:33,920
because we're building on bare metal.

2537
01:13:33,920 --> 01:13:36,080
So we don't pay for bandwidth.

2538
01:13:36,080 --> 01:13:37,360
There's no reason why to

2539
01:13:37,360 --> 01:13:38,960
charge people for bandwidth.

2540
01:13:39,680 --> 01:13:40,560
There's no reason why

2541
01:13:40,560 --> 01:13:41,760
to overcharge people

2542
01:13:41,760 --> 01:13:44,000
for CPU prices or go

2543
01:13:44,000 --> 01:13:46,240
5X what AWS gives you.

2544
01:13:46,240 --> 01:13:48,400
The only hard part, to be honest,

2545
01:13:48,400 --> 01:13:52,640
that we had to deal with is networking.

2546
01:13:52,640 --> 01:13:53,840
But networking in the

2547
01:13:53,840 --> 01:13:55,600
sense of private networking.

2548
01:13:55,600 --> 01:13:57,280
We got it figured,

2549
01:13:57,280 --> 01:13:59,040
kind of EPC layer stuff.

2550
01:13:59,040 --> 01:14:01,440
I had to write tons of

2551
01:14:01,440 --> 01:14:03,520
eBPF code in the kernel

2552
01:14:03,520 --> 01:14:06,320
to be able to figure that out properly.

2553
01:14:06,320 --> 01:14:07,760
But I think we're getting there.

2554
01:14:07,760 --> 01:14:10,120
And I hope that we're

2555
01:14:10,720 --> 01:14:12,880
gonna be able to provide you

2556
01:14:12,880 --> 01:14:13,680
with the right

2557
01:14:13,680 --> 01:14:16,240
primitives to deploy a platform

2558
01:14:16,240 --> 01:14:17,760
such as OpenRouter, which

2559
01:14:17,760 --> 01:14:18,960
is really what we're trying

2560
01:14:18,960 --> 01:14:21,680
to solve with Effectful.

2561
01:14:21,680 --> 01:14:23,920
Because with Effect, I think we solve

2562
01:14:23,920 --> 01:14:25,840
the production grade coding.

2563
01:14:25,840 --> 01:14:28,720
But at some point, code ends,

2564
01:14:28,720 --> 01:14:30,000
and you have to

2565
01:14:30,000 --> 01:14:32,080
deploy your code somewhere.

2566
01:14:32,080 --> 01:14:34,240
And especially with AI,

2567
01:14:34,240 --> 01:14:35,680
I feel we are hitting.

2568
01:14:36,240 --> 01:14:38,000
The real troubles were

2569
01:14:38,000 --> 01:14:38,960
if you were deploying

2570
01:14:38,960 --> 01:14:40,720
distributed systems in the past.

2571
01:14:40,720 --> 01:14:41,600
But you would only do

2572
01:14:41,600 --> 01:14:44,240
that if you were a scale up

2573
01:14:44,240 --> 01:14:45,360
handling millions and

2574
01:14:45,360 --> 01:14:46,880
millions of customers

2575
01:14:46,880 --> 01:14:48,160
Otherwise, a small

2576
01:14:48,160 --> 01:14:50,480
server, just a database,

2577
01:14:50,480 --> 01:14:52,000
would get you out of the troubles.

2578
01:14:52,000 --> 01:14:54,400
And the edge cases would not materialize.

2579
01:14:54,400 --> 01:14:56,720
Because the average request time was,

2580
01:14:56,720 --> 01:14:59,280
like what, 100, 200 milliseconds.

2581
01:14:59,280 --> 01:15:01,200
Now with AI and flows like

2582
01:15:01,200 --> 01:15:03,680
the ones you have in OpenRouter,

2583
01:15:03,680 --> 01:15:04,560
my guess is that

2584
01:15:04,560 --> 01:15:06,160
sometimes you have API calls

2585
01:15:06,160 --> 01:15:07,200
that takes minutes.

2586
01:15:07,200 --> 01:15:09,120
with minutes of

2587
01:15:09,120 --> 01:15:11,520
time, be sure that something

2588
01:15:11,520 --> 01:15:12,640
can happen much more

2589
01:15:12,640 --> 01:15:14,480
frequently than it can happen

2590
01:15:14,480 --> 01:15:15,760
in 100 milliseconds.

2591
01:15:15,760 --> 01:15:18,640
And so long-running processes

2592
01:15:18,640 --> 01:15:22,240
and less server-less pieces,

2593
01:15:22,240 --> 01:15:25,360
I think will be much more important.

2594
01:15:25,360 --> 01:15:28,000
And let me ask you a question.

2595
01:15:28,000 --> 01:15:31,200
How important is latency in a request

2596
01:15:31,200 --> 01:15:32,960
that takes five minutes to answer?

2597
01:15:32,960 --> 01:15:33,680
Not a lot, right?

2598
01:15:33,680 --> 01:15:34,400
Not a lot.

2599
01:15:35,040 --> 01:15:37,680
Maybe the handshake so that the client

2600
01:15:37,680 --> 01:15:39,120
is not easily, some

2601
01:15:39,120 --> 01:15:40,080
of them are very eager

2602
01:15:40,080 --> 01:15:43,040
to kill the request, but otherwise, yeah.

2603
01:15:43,040 --> 01:15:44,240
People don't really care.

2604
01:15:44,240 --> 01:15:46,080
Sure, but you might be,

2605
01:15:46,080 --> 01:15:47,280
like the important part

2606
01:15:47,280 --> 01:15:50,160
what might be to back off from a request

2607
01:15:50,160 --> 01:15:52,880
much faster to not spend money

2608
01:15:52,880 --> 01:15:55,680
versus actually letting the process run.

2609
01:15:55,680 --> 01:15:58,640
So I think for these kind of use cases,

2610
01:15:58,640 --> 01:15:59,680
and I've been discussing

2611
01:15:59,680 --> 01:16:01,920
this with some other people,

2612
01:16:01,920 --> 01:16:03,200
it's much better to have

2613
01:16:03,200 --> 01:16:04,880
reliable infrastructure,

2614
01:16:04,880 --> 01:16:05,840
maybe deployed

2615
01:16:05,840 --> 01:16:07,600
slightly outside of the edge,

2616
01:16:07,600 --> 01:16:09,760
so it's not close to your home,

2617
01:16:09,760 --> 01:16:11,600
but it's cheap, doesn't

2618
01:16:11,600 --> 01:16:14,400
have network costs whatsoever,

2619
01:16:14,400 --> 01:16:15,680
has very cheap compute

2620
01:16:15,680 --> 01:16:18,720
costs, very cheap storage costs,

2621
01:16:18,720 --> 01:16:21,280
and you can actually run a server

2622
01:16:21,280 --> 01:16:23,360
not being constrained by memory

2623
01:16:23,360 --> 01:16:24,720
because I feel like

2624
01:16:24,720 --> 01:16:26,560
being able to collect traces

2625
01:16:26,560 --> 01:16:29,360
is much more important for the business

2626
01:16:29,360 --> 01:16:33,200
than saving, what, 20 megabytes of RAM?

2627
01:16:33,200 --> 01:16:35,760
like RAM is cheap, or 20

2628
01:16:35,760 --> 01:16:37,800
milliseconds of request time?

2629
01:16:37,800 --> 01:16:38,480
or maybe like

2630
01:16:38,480 --> 01:16:39,840
100 milliseconds, yeah,

2631
01:16:39,840 --> 01:16:41,200
which is doesn't really do much.

2632
01:16:41,200 --> 01:16:42,080
By the way, the thing

2633
01:16:42,080 --> 01:16:42,880
that you might think

2634
01:16:42,880 --> 01:16:43,840
is just internal tooling,

2635
01:16:43,840 --> 01:16:44,800
but we have a lot of

2636
01:16:44,800 --> 01:16:46,480
internal tooling, just so you know.

2637
01:16:46,480 --> 01:16:48,160
So, which is why Effect is

2638
01:16:48,160 --> 01:16:49,680
actually being used a lot.

2639
01:16:49,680 --> 01:16:50,400
because we are

2640
01:16:50,400 --> 01:16:51,920
TypeScript everything, right?

2641
01:16:51,920 --> 01:16:53,920
So, instead of running a Python eval

2642
01:16:53,920 --> 01:16:55,200
suite on the market,

2643
01:16:55,200 --> 01:16:57,520
there's a bunch, we write our own, but it's

2644
01:16:57,520 --> 01:16:58,480
using Effect, right?

2645
01:16:58,480 --> 01:17:00,240
So, we actually have an eval suite

2646
01:17:00,240 --> 01:17:02,720
and also a bulk test suite,

2647
01:17:02,720 --> 01:17:05,440
that's all using Effect with Effect retry

2648
01:17:05,440 --> 01:17:06,400
and so on and so forth.

2649
01:17:06,400 --> 01:17:07,600
Yeah, so it's a very

2650
01:17:07,600 --> 01:17:08,400
sophisticated system.

2651
01:17:08,400 --> 01:17:10,400
I'm just saying is that if we use,

2652
01:17:10,400 --> 01:17:11,120
you might think internal

2653
01:17:11,120 --> 01:17:12,160
tooling is like some toy,

2654
01:17:12,160 --> 01:17:13,120
but no, we're actually using

2655
01:17:13,120 --> 01:17:14,880
it almost on a daily basis,

2656
01:17:14,880 --> 01:17:16,080
it's actually production load.

2657
01:17:16,080 --> 01:17:17,840
Reason why we don't use it in the core,

2658
01:17:17,840 --> 01:17:19,440
because while the core is working,

2659
01:17:19,440 --> 01:17:20,000
don't touch it.

2660
01:17:20,000 --> 01:17:22,000
But we do have a lot of internal tooling

2661
01:17:22,000 --> 01:17:23,520
to do testing out all

2662
01:17:23,520 --> 01:17:24,640
the endpoint, right?

2663
01:17:24,640 --> 01:17:27,040
Which is a crucial part of our operation.

2664
01:17:27,040 --> 01:17:28,240
We have, well, I

2665
01:17:28,240 --> 01:17:29,680
mean, almost every respond

2666
01:17:29,680 --> 01:17:31,840
to transformation is using the pipe now.

2667
01:17:31,840 --> 01:17:32,960
And then also I'm

2668
01:17:32,960 --> 01:17:35,680
introducing more and more generator.

2669
01:17:35,680 --> 01:17:38,240
So, down the line, we finally will,

2670
01:17:38,240 --> 01:17:38,960
it will be making it

2671
01:17:38,960 --> 01:17:40,000
easier for us to transitioning

2672
01:17:40,000 --> 01:17:41,360
over to Effect as needed.

2673
01:17:41,360 --> 01:17:42,240
by the way, to the point,

2674
01:17:42,240 --> 01:17:43,840
I'm actually still using async generator.

2675
01:17:43,840 --> 01:17:44,720
And the main reason why

2676
01:17:44,720 --> 01:17:45,840
I'm using async generator,

2677
01:17:45,840 --> 01:17:47,360
because inside this generator,

2678
01:17:47,360 --> 01:17:48,160
we have some async

2679
01:17:48,160 --> 01:17:50,800
process, which unavoidable.

2680
01:17:50,800 --> 01:17:52,480
No, no, my prior

2681
01:17:52,480 --> 01:17:53,920
comment on you should not use

2682
01:17:53,920 --> 01:17:55,520
async generators is

2683
01:17:55,520 --> 01:17:56,640
because you should not have

2684
01:17:56,640 --> 01:17:58,560
two representations of code.

2685
01:17:58,560 --> 01:17:59,680
But if you're using

2686
01:17:59,680 --> 01:18:01,840
promises with result types,

2687
01:18:01,840 --> 01:18:04,320
100%, please use async generators

2688
01:18:04,320 --> 01:18:06,480
so that your developers

2689
01:18:06,480 --> 01:18:07,600
don't have to unwrap

2690
01:18:07,600 --> 01:18:10,120
every error manually at every point.

2691
01:18:10,120 --> 01:18:11,840
But it's still a much

2692
01:18:11,840 --> 01:18:13,520
better user experience.

2693
01:18:13,520 --> 01:18:15,680
I just meant that if you

2694
01:18:15,680 --> 01:18:17,360
have one single representation

2695
01:18:17,360 --> 01:18:19,360
of code, you don't need

2696
01:18:19,360 --> 01:18:21,120
to use different types

2697
01:18:21,120 --> 01:18:22,320
of generators

2698
01:18:22,320 --> 01:18:24,320
Plus, adding like one

2699
01:18:24,320 --> 01:18:26,240
of the key troubles I had

2700
01:18:26,240 --> 01:18:28,720
in the past was, okay, I

2701
01:18:28,720 --> 01:18:30,240
have a very synchronous

2702
01:18:30,240 --> 01:18:31,920
code transformation, like these takes

2703
01:18:31,920 --> 01:18:33,200
an API, does this

2704
01:18:33,200 --> 01:18:34,640
this, this, this, that, okay.

2705
01:18:34,640 --> 01:18:36,560
Now I wanna emit a Kafka event.

2706
01:18:36,560 --> 01:18:37,680
Not now, not at the

2707
01:18:37,680 --> 01:18:39,600
beginning, but a month later,

2708
01:18:39,600 --> 01:18:41,760
I would write an event in Kafka.

2709
01:18:41,760 --> 01:18:44,000
Look at what's the writing to Kafka

2710
01:18:44,000 --> 01:18:45,520
is an asynchronous operation.

2711
01:18:45,520 --> 01:18:46,560
So I stitch it at the

2712
01:18:46,560 --> 01:18:48,480
end, and I have to bubble up.

2713
01:18:48,480 --> 01:18:50,480
Everything becomes async.

2714
01:18:50,480 --> 01:18:52,000
That's a problem.

2715
01:18:52,000 --> 01:18:53,760
Now it's still a much

2716
01:18:53,760 --> 01:18:55,440
better problem to have

2717
01:18:55,440 --> 01:18:57,840
versus writing callbacks.

2718
01:18:57,840 --> 01:18:58,960
Because callbacks don't

2719
01:18:58,960 --> 01:19:00,480
have function coloring.

2720
01:19:00,480 --> 01:19:02,880
Callbacks are fine for sync and async.

2721
01:19:02,880 --> 01:19:03,840
They are a unified

2722
01:19:03,840 --> 01:19:06,000
representation of computations.

2723
01:19:06,000 --> 01:19:08,560
It's continuation passing

2724
01:19:08,560 --> 01:19:10,560
style that we ended up calling

2725
01:19:10,560 --> 01:19:13,040
callback passing style in JavaScript

2726
01:19:13,040 --> 01:19:14,240
because we just like to

2727
01:19:14,240 --> 01:19:16,000
invent new terminology

2728
01:19:16,000 --> 01:19:18,080
so that when somebody looks at history,

2729
01:19:18,080 --> 01:19:19,200
they don't see the

2730
01:19:19,200 --> 01:19:21,600
theoretical papers from the 70s.

2731
01:19:21,600 --> 01:19:22,800
Otherwise they understand

2732
01:19:22,800 --> 01:19:24,240
we're just stealing ideas

2733
01:19:24,240 --> 01:19:26,000
from the 70s and making

2734
01:19:26,000 --> 01:19:28,080
them production ready.

2735
01:19:28,080 --> 01:19:30,080
But there was a unified

2736
01:19:30,080 --> 01:19:31,520
representation of code,

2737
01:19:31,520 --> 01:19:34,137
which leads to, boom,

2738
01:19:34,137 --> 01:19:35,520
boom, boom, boom, boom,
of callbacks.

2739
01:19:36,080 --> 01:19:37,600
Now with promises and async

2740
01:19:37,600 --> 01:19:39,440
await, you no longer have that.

2741
01:19:39,440 --> 01:19:40,880
But you do have the two different

2742
01:19:40,880 --> 01:19:42,640
representations of code.

2743
01:19:42,640 --> 01:19:44,000
And that's much better

2744
01:19:44,000 --> 01:19:45,760
than reading the code

2745
01:19:45,760 --> 01:19:46,800
like you described before.

2746
01:19:46,800 --> 01:19:48,160
You want to review on the phone.

2747
01:19:48,160 --> 01:19:49,200
There's no way you're going

2748
01:19:49,200 --> 01:19:52,800
to review 100 nested callbacks

2749
01:19:52,800 --> 01:19:54,560
with your phone screen.

2750
01:19:54,560 --> 01:19:55,600
By the way, no way

2751
01:19:55,600 --> 01:19:56,800
you're going to read that

2752
01:19:56,800 --> 01:19:58,720
in a normal computer screen.

2753
01:19:58,720 --> 01:20:01,280
You might need Johanness' set up

2754
01:20:01,280 --> 01:20:03,360
with exceptionally large

2755
01:20:03,360 --> 01:20:04,960
screens and so on and so forth.

2756
01:20:04,960 --> 01:20:06,160
But even there, I

2757
01:20:06,160 --> 01:20:07,920
think you should optimize

2758
01:20:09,120 --> 01:20:10,640
for a different problem.

2759
01:20:10,640 --> 01:20:13,120
So I think the final solution, which

2760
01:20:13,120 --> 01:20:15,040
is I believe what we have with Effect

2761
01:20:15,040 --> 01:20:18,240
is a unified representation that allows

2762
01:20:18,240 --> 01:20:20,640
for sequential representation.

2763
01:20:20,640 --> 01:20:23,520
But still, the intermediate point

2764
01:20:23,520 --> 01:20:25,760
of having two different ways of doing

2765
01:20:25,760 --> 01:20:26,800
things that all look

2766
01:20:26,800 --> 01:20:29,520
good is perfectly viable.

2767
01:20:29,520 --> 01:20:30,800
And especially in your

2768
01:20:30,800 --> 01:20:31,680
code base where you're

2769
01:20:31,680 --> 01:20:32,880
using a custom result

2770
01:20:32,880 --> 01:20:35,760
type, all the time go for that.

2771
01:20:35,760 --> 01:20:36,840
It's amazing.

2772
01:20:36,840 --> 01:20:37,760
Yeah, this is the thing

2773
01:20:37,760 --> 01:20:38,800
that I was looking for,

2774
01:20:38,800 --> 01:20:40,320
which is the way to represent in both

2775
01:20:40,320 --> 01:20:41,680
async and synchronous

2776
01:20:41,680 --> 01:20:42,080
process

2777
01:20:42,080 --> 01:20:44,160
in the same container, essentially.

2778
01:20:44,160 --> 01:20:45,040
that's something I'm excited

2779
01:20:45,040 --> 01:20:46,400
about most, about Effect, right?

2780
01:20:46,400 --> 01:20:48,560
It blurred the line between async and sync.

2781
01:20:48,560 --> 01:20:49,200
And you can now

2782
01:20:49,200 --> 01:20:51,840
orchestrate them in the same flow.

2783
01:20:51,840 --> 01:20:54,160
So it seems like you're well on your way

2784
01:20:54,160 --> 01:20:56,320
to introduce Effect more and more

2785
01:20:56,320 --> 01:20:58,640
throughout the entire system.

2786
01:20:58,640 --> 01:20:59,840
And it's very interesting

2787
01:20:59,840 --> 01:21:01,600
to hear how you've basically

2788
01:21:01,600 --> 01:21:03,600
introduced mini versions and

2789
01:21:03,600 --> 01:21:04,960
little pieces of Effect

2790
01:21:04,960 --> 01:21:06,800
already to make it

2791
01:21:06,800 --> 01:21:08,400
more and more swallowable

2792
01:21:08,400 --> 01:21:11,360
by the rest of the engineering culture.

2793
01:21:11,360 --> 01:21:14,240
So I feel there's plenty of stuff

2794
01:21:14,240 --> 01:21:17,280
here that we need to unpack

2795
01:21:17,280 --> 01:21:19,120
probably in another future

2796
01:21:19,120 --> 01:21:21,040
episode to hear how your

2797
01:21:21,040 --> 01:21:23,440
Effect journey is continuing.

2798
01:21:23,440 --> 01:21:25,680
Before we're wrapping up, though,

2799
01:21:25,680 --> 01:21:28,960
we talked before about that as part of

2800
01:21:28,960 --> 01:21:30,400
your onboarding culture

2801
01:21:30,400 --> 01:21:32,080
to ask a new employee

2802
01:21:32,080 --> 01:21:34,160
what's the latest rabbit hole

2803
01:21:34,160 --> 01:21:35,760
they went down on.

2804
01:21:35,760 --> 01:21:37,680
So what's been your latest rabbit hole

2805
01:21:37,680 --> 01:21:41,600
that you barely emerged again from?

2806
01:21:41,600 --> 01:21:43,600
The most recent one-- I recently moved to

2807
01:21:43,600 --> 01:21:44,640
New York, by the way,

2808
01:21:44,640 --> 01:21:45,920
two weeks ago.

2809
01:21:45,920 --> 01:21:47,680
Sorry, two weeks ago was two weeks ago.

2810
01:21:47,680 --> 01:21:49,040
So that was a month ago.

2811
01:21:49,040 --> 01:21:49,840
Oh, my god.

2812
01:21:49,840 --> 01:21:51,360
And that led me to hold rabbit hole

2813
01:21:51,360 --> 01:21:52,160
buying stuff for my

2814
01:21:52,160 --> 01:21:53,440
apartment, essentially.

2815
01:21:53,440 --> 01:21:54,960
Well, there are two rabbit holes.

2816
01:21:54,960 --> 01:21:56,720
One is on buying a bed,

2817
01:21:56,720 --> 01:21:58,160
and one is on apartment

2818
01:21:58,160 --> 01:21:59,520
finding in New York.

2819
01:21:59,520 --> 01:22:00,960
Luckily, I didn't--

2820
01:22:00,960 --> 01:22:04,880
I was on a time crunch to find an

2821
01:22:04,880 --> 01:22:06,560
apartment in New York.

2822
01:22:06,560 --> 01:22:07,760
So that

2823
01:22:07,760 --> 01:22:09,520
hole was a bit shallow.

2824
01:22:09,520 --> 01:22:12,400
But the one about finding a bed, though,

2825
01:22:12,400 --> 01:22:13,520
a good bed for your

2826
01:22:13,520 --> 01:22:15,600
back, is it was very deep.

2827
01:22:15,600 --> 01:22:17,040
And thankfully-- so I'm

2828
01:22:17,040 --> 01:22:19,040
working off the Marlin.

2829
01:22:19,040 --> 01:22:20,400
It is a very bougie cooking

2830
01:22:20,400 --> 01:22:22,320
space down in Flat Island,

2831
01:22:22,320 --> 01:22:22,720
Manhattan.

2832
01:22:22,720 --> 01:22:26,560
And right next to this place, there are

2833
01:22:26,560 --> 01:22:28,400
three mattress stores.

2834
01:22:28,400 --> 01:22:30,480
Like right next to me is the Sattva,

2835
01:22:30,480 --> 01:22:31,280
which is one of these

2836
01:22:31,280 --> 01:22:32,800
very high-quality hotel--

2837
01:22:32,800 --> 01:22:34,240
mattress.

2838
01:22:34,240 --> 01:22:35,040
Down the block, there

2839
01:22:35,040 --> 01:22:36,160
is a mattress firm store,

2840
01:22:36,160 --> 01:22:37,520
which contains a bunch of

2841
01:22:37,520 --> 01:22:39,280
variety of mattress store.

2842
01:22:39,280 --> 01:22:43,840
And then on the other block is Avocado,

2843
01:22:43,840 --> 01:22:45,440
which is this another

2844
01:22:45,440 --> 01:22:47,840
very bougie, very expensive

2845
01:22:47,840 --> 01:22:48,720
mattress.

2846
01:22:48,720 --> 01:22:50,000
And I got to go in and try

2847
01:22:50,000 --> 01:22:52,800
all of them, spend about a day.

2848
01:22:52,800 --> 01:22:54,560
And it is fascinating.

2849
01:22:55,200 --> 01:22:56,400
The final choice I made

2850
01:22:56,400 --> 01:22:58,400
was to get a purple mattress.

2851
01:22:58,400 --> 01:22:59,200
This is not a purple

2852
01:22:59,200 --> 01:22:59,840
mattress app, though.

2853
01:22:59,840 --> 01:23:00,720
I just tried all of them.

2854
01:23:00,720 --> 01:23:01,911
So the best-- I'm

2855
01:23:01,911 --> 01:23:02,400
going to say the best thing

2856
01:23:02,400 --> 01:23:03,600
that you can do when buying a mattress,

2857
01:23:03,600 --> 01:23:04,400
you just try them out

2858
01:23:04,400 --> 01:23:08,080
yourself, tug it in, go on all side.

2859
01:23:08,080 --> 01:23:09,520
You've got to try all side

2860
01:23:09,520 --> 01:23:10,640
of your sleeping position.

2861
01:23:10,640 --> 01:23:12,320
Yeah, I roll it all over

2862
01:23:13,040 --> 01:23:14,320
and try to tug it in to see

2863
01:23:14,320 --> 01:23:17,760
if your spine can get enough support.

2864
01:23:17,760 --> 01:23:19,680
And of all of them so

2865
01:23:19,680 --> 01:23:20,880
far, for me personally,

2866
01:23:20,880 --> 01:23:22,800
the purple mattress has been the one

2867
01:23:22,800 --> 01:23:27,200
that I can actually support my spine.

2868
01:23:27,200 --> 01:23:28,960
The second option was the

2869
01:23:28,960 --> 01:23:30,800
Sattva, the Sattva Classic,

2870
01:23:30,800 --> 01:23:33,040
which is somewhat soft, but

2871
01:23:33,040 --> 01:23:35,200
also decently firm in the bottom.

2872
01:23:35,200 --> 01:23:38,000
And yeah, the purple

2873
01:23:38,000 --> 01:23:39,440
mattress is kind of interesting.

2874
01:23:39,440 --> 01:23:43,120
It's not one of those

2875
01:23:43,120 --> 01:23:44,480
put traditional bad.

2876
01:23:45,600 --> 01:23:46,240
He's going to send

2877
01:23:46,240 --> 01:23:47,520
over the affiliate links,

2878
01:23:47,520 --> 01:23:48,160
and we're going to put

2879
01:23:48,160 --> 01:23:49,120
them in the description.

2880
01:23:49,120 --> 01:23:50,560
It's going to be a podcast first.

2881
01:23:50,560 --> 01:23:51,920
There was no affiliate, but

2882
01:23:51,920 --> 01:23:52,800
thanks to a very good friend

2883
01:23:52,800 --> 01:23:54,080
of mine.
He was my co-founder of

2884
01:23:54,080 --> 01:23:55,440
my previous company, too.

2885
01:23:55,440 --> 01:23:56,240
He went mattress

2886
01:23:56,240 --> 01:23:57,360
shopping together with me

2887
01:23:57,360 --> 01:23:59,040
And then we did the good

2888
01:23:59,040 --> 01:24:00,560
cop, bad cop negotiation

2889
01:24:00,560 --> 01:24:01,520
at the mattress firm.

2890
01:24:01,520 --> 01:24:03,280
And we get $1,000

2891
01:24:03,280 --> 01:24:04,480
discount on that mattress,

2892
01:24:04,480 --> 01:24:06,000
which is kind of crazy.

2893
01:24:06,000 --> 01:24:06,400
$1,000?

2894
01:24:06,400 --> 01:24:08,240
$1,000.

2895
01:24:08,240 --> 01:24:10,160
$1,000.
A $1,000 discount on a mattress?

2896
01:24:10,160 --> 01:24:11,040
Yes, on a mattress.

2897
01:24:11,040 --> 01:24:12,160
That kind of suggests

2898
01:24:12,160 --> 01:24:13,200
that the mattress was

2899
01:24:13,200 --> 01:24:14,080
more than $1,000.

2900
01:24:14,880 --> 01:24:15,520
Yeah, I guess.

2901
01:24:16,160 --> 01:24:17,440
Yes, the mattress was very--

2902
01:24:17,440 --> 01:24:19,440
I mean, I talked to a friend of mine,

2903
01:24:19,440 --> 01:24:21,600
and he told me, you

2904
01:24:21,600 --> 01:24:23,040
spend a lot on your back.

2905
01:24:23,040 --> 01:24:24,400
You're going to need it.

2906
01:24:24,400 --> 01:24:26,720
So yeah, I think it was good spending.

2907
01:24:27,440 --> 01:24:29,920
Yeah, very comparable bet, for sure.

2908
01:24:29,920 --> 01:24:30,720
Unrolling the bed,

2909
01:24:30,720 --> 01:24:31,840
though, was insane, though.

2910
01:24:31,840 --> 01:24:32,720
It's very hard.

2911
01:24:32,720 --> 01:24:35,920
So recommend, if you're buying a bed,

2912
01:24:35,920 --> 01:24:37,360
just have two people.

2913
01:24:37,360 --> 01:24:38,320
Just have a friend.

2914
01:24:38,320 --> 01:24:38,640
Fair enough.

2915
01:24:38,640 --> 01:24:40,880
Otherwise, you'll ruin your back while

2916
01:24:40,880 --> 01:24:43,280
setting up the mattress.

2917
01:24:43,280 --> 01:24:44,880
That's no good.

2918
01:24:45,520 --> 01:24:46,080
Awesome.

2919
01:24:46,080 --> 01:24:48,539
Hey, Louis, this was an absolute treat

2920
01:24:48,539 --> 01:24:48,560
for us having you on.

2921
01:24:48,560 --> 01:24:51,120
for us having you on.
I'm looking forward to

2922
01:24:51,120 --> 01:24:53,760
having you on in the future again

2923
01:24:53,760 --> 01:24:55,680
and hearing more about how

2924
01:24:55,680 --> 01:24:58,160
OpenRouter is adopting Effect

2925
01:24:58,160 --> 01:24:59,760
and all the crazy

2926
01:24:59,760 --> 01:25:01,760
growth you're going through.

2927
01:25:01,760 --> 01:25:02,720
Also, a big thank you

2928
01:25:02,720 --> 01:25:05,360
for Mike for spoiling us

2929
01:25:05,360 --> 01:25:06,800
with his presence

2930
01:25:06,800 --> 01:25:08,560
throughout this episode, as well.

2931
01:25:08,560 --> 01:25:10,640
Very, very interesting conversations.

2932
01:25:10,640 --> 01:25:11,680
I've learned a lot.

2933
01:25:11,680 --> 01:25:15,440
So thanks a lot to both of you.

2934
01:25:15,440 --> 01:25:16,400
Awesome.

2935
01:25:16,400 --> 01:25:17,280
Thanks, Mike, for

2936
01:25:19,840 --> 01:25:21,600
putting me in on the podcast.

2937
01:25:23,280 --> 01:25:25,040
It was a big pleasure.

2938
01:25:25,040 --> 01:25:27,520
And I hope to see you soon

2939
01:25:28,160 --> 01:25:30,080
when we come to New York.

2940
01:25:30,080 --> 01:25:31,040
Absolutely.

2941
01:25:31,040 --> 01:25:31,680
I'll see you then.

2942
01:25:31,680 --> 01:25:32,080
All righty.

2943
01:25:32,080 --> 01:25:32,640
Have fun.

2944
01:25:32,640 --> 01:25:33,760
Take care.

2945
01:25:33,760 --> 01:25:35,040
Thank you for listening to the

2946
01:25:35,040 --> 01:25:36,720
Cause & Effect Podcast.

2947
01:25:36,720 --> 01:25:38,640
If you've enjoyed this episode, please

2948
01:25:38,640 --> 01:25:40,160
subscribe, leave a review

2949
01:25:40,160 --> 01:25:42,080
and share it with your friends.

2950
01:25:42,080 --> 01:25:43,920
If you haven't done so already, you can

2951
01:25:43,920 --> 01:25:45,520
join our Discord community.

2952
01:25:45,520 --> 01:25:47,520
And if you have any questions, feedback

2953
01:25:47,520 --> 01:25:49,920
or suggestions about this episode or

2954
01:25:49,920 --> 01:25:51,440
about Effect in general,

2955
01:25:51,440 --> 01:25:53,600
don't hesitate to get in touch.

2956
01:25:53,600 --> 01:25:57,680
See you in the next episode.