1
00:00:00,200 --> 00:00:03,040
That's exactly what's going on 
right now, which is basically 

2
00:00:03,360 --> 00:00:08,240
that service based architecture 
SOA, you know, we're we've 

3
00:00:08,240 --> 00:00:13,280
switched from this idea that we 
had to go ahead and originally 

4
00:00:13,280 --> 00:00:16,480
it was 1 central command to 
understand what the budget was. 

5
00:00:16,800 --> 00:00:19,800
And now that got pushed down to 
developers because the 

6
00:00:19,800 --> 00:00:23,440
developers could actually then 
care, understand and help 

7
00:00:23,440 --> 00:00:26,120
control that token. 
Novice is the exact same thing 

8
00:00:26,120 --> 00:00:34,640
100%. 
I will have you know I'm still 

9
00:00:34,640 --> 00:00:37,880
waiting for the coffee from 
Taiwan that you promised you 

10
00:00:37,880 --> 00:00:39,840
were going to send me after that
blog post you. 

11
00:00:39,840 --> 00:00:42,320
Wrote and I'm and I'm glad 
you're recording it to basically

12
00:00:42,320 --> 00:00:45,280
make sure that I'm held. 
Accountability one O 1 Yeah, 

13
00:00:45,280 --> 00:00:47,440
this is. 
No, no, you're right, I'm wrong.

14
00:00:47,960 --> 00:00:50,000
The problem is that I'm actually
not going this summer, 

15
00:00:50,680 --> 00:00:55,400
unfortunately. 
But if when I do, yes, there is 

16
00:00:55,400 --> 00:00:58,960
a specific set of farmers and a 
specific set of places that I 

17
00:00:58,960 --> 00:01:00,760
want to go to. 
And it's actually interesting 

18
00:01:00,760 --> 00:01:03,320
enough related to agents and we 
will get into that when you're 

19
00:01:03,320 --> 00:01:05,280
ready to go. 
But it's actually related to 

20
00:01:05,280 --> 00:01:07,760
agents, my friend. 
Are are there some of your 

21
00:01:08,040 --> 00:01:10,880
agentic workflows that you've 
got going on finding different 

22
00:01:10,880 --> 00:01:16,040
micro fincas around the world? 
Not just that, but also which 

23
00:01:16,040 --> 00:01:19,720
regions in Taiwan versus which 
regions of Japan or Korea 

24
00:01:19,840 --> 00:01:21,720
specifically for macho and shit 
like that. 

25
00:01:22,080 --> 00:01:23,840
What? 
How? 

26
00:01:25,760 --> 00:01:28,480
How are you getting agents to do
this research? 

27
00:01:28,480 --> 00:01:31,400
Is it just like exploring the? 
Yeah, it's exploring. 

28
00:01:31,400 --> 00:01:34,760
So so basically, OK, so we're 
probably jumping ahead of her, 

29
00:01:34,760 --> 00:01:37,200
but like I'm using omnigents, 
Polly. 

30
00:01:37,920 --> 00:01:40,000
Sorry, Debbie, excuse me. 
That's the one where basically 

31
00:01:40,000 --> 00:01:43,200
they debate each other, right? 
And so this case is basically a 

32
00:01:43,560 --> 00:01:48,120
my ChatGPT, you know, 55 versus 
my Claude Opus 47, right? 

33
00:01:48,480 --> 00:01:49,760
And they basically debate each 
other. 

34
00:01:49,960 --> 00:01:54,640
And so basically I start, I, I 
already have enough like agentic

35
00:01:54,640 --> 00:01:58,360
memory and skills, MD's and all 
this other shit that provide 

36
00:01:58,360 --> 00:02:01,800
like a lot of huge background on
what I'm specifically looking 

37
00:02:01,800 --> 00:02:04,560
for anyways, right? 
So if you go to my blog either 

38
00:02:04,560 --> 00:02:07,400
like the daily jeely.com blog, 
the personal one, you'll 

39
00:02:07,400 --> 00:02:11,039
remember that there's this one 
asinine long Taiwanese blog 

40
00:02:11,039 --> 00:02:14,680
about Taiwanese coffee, right? 
It was supposed to be Yeah, 

41
00:02:14,680 --> 00:02:15,400
Yeah, exactly. 
Yeah. 

42
00:02:15,800 --> 00:02:18,320
It's supposed to be small. 
And then it turned out to this 

43
00:02:18,320 --> 00:02:21,440
basically be this long diatribe 
that went from like, here, all 

44
00:02:21,440 --> 00:02:24,160
the great coffee shops in 
Taiwan, to like, let me go visit

45
00:02:24,160 --> 00:02:27,400
a farmer in Taiwan, in Nanto 
Province, in Nanto County, 

46
00:02:27,560 --> 00:02:29,720
specifically to go ahead and 
understand how he's going to 

47
00:02:29,720 --> 00:02:32,160
beans and how like. 
Yeah, like, and how used to be 

48
00:02:32,160 --> 00:02:35,320
Bidlon leaves, right? 
And then how the Taiwanese 

49
00:02:35,320 --> 00:02:38,080
government was switching over to
convince people not to do grow 

50
00:02:38,080 --> 00:02:41,400
bin Laden leaves because bin 
Laden is a it's like a like 

51
00:02:41,400 --> 00:02:43,400
cigarettes, people get super 
addicted to it. 

52
00:02:43,400 --> 00:02:47,200
But also the leaves are huge, so
the water actually degrades the 

53
00:02:47,200 --> 00:02:48,960
land underneath it. 
So the Taiwanese government's 

54
00:02:48,960 --> 00:02:50,560
trying to convince people to 
grow coffee. 

55
00:02:50,920 --> 00:02:55,480
Well the this is already too 
long, but the TLDR version of 

56
00:02:55,480 --> 00:02:58,440
this is that Nanto County is 
actually famous for growing tea 

57
00:02:59,480 --> 00:03:03,760
and 1st before growing coffee. 
So I've been actually using a 

58
00:03:03,760 --> 00:03:09,000
homogene to find me specific 
regions in Nanto to grow matcha.

59
00:03:12,480 --> 00:03:17,040
Wait with the intention of you 
going there and then growing 

60
00:03:17,040 --> 00:03:18,840
matcha or? 
Partnering or working with the 

61
00:03:18,840 --> 00:03:21,520
farmers, working with the 
farmers in that area to grow the

62
00:03:21,520 --> 00:03:24,320
matcha. 
And so some interesting tidbits 

63
00:03:24,320 --> 00:03:29,000
is that the oxidase when you, 
when you working, when you pick 

64
00:03:29,000 --> 00:03:32,160
the leaves, right it, it could 
oxidize really fast. 

65
00:03:32,160 --> 00:03:34,680
So you actually have to steam 
and dry them immediately. 

66
00:03:34,680 --> 00:03:38,320
So, so the problem like Japan 
and Korea, they both have the 

67
00:03:38,320 --> 00:03:42,280
infrastructure set up that next 
to where the leaf picking is 

68
00:03:42,280 --> 00:03:45,480
actually happening, they can 
immediately deal with the 

69
00:03:45,480 --> 00:03:48,640
steaming and drying OK and of 
the leaves. 

70
00:03:49,800 --> 00:03:53,640
For example, this is a good 
setup for coffee because coffee 

71
00:03:53,720 --> 00:03:57,040
in the coffee regions, they know
there's like bourbon style, 

72
00:03:57,040 --> 00:03:59,200
honey style, whatever else, but 
it's all about drying the coffee

73
00:03:59,200 --> 00:04:01,280
beans after you pick them and 
how do you do it correctly, 

74
00:04:01,440 --> 00:04:05,440
right? 
And so the for oolong and other 

75
00:04:05,440 --> 00:04:06,920
types of teas in Taiwan, they're
all set. 

76
00:04:06,920 --> 00:04:10,760
That's nonto's famous for that, 
for that producing amazing like 

77
00:04:10,760 --> 00:04:13,320
for example, Juan tea, like 
they're, they're great at that 

78
00:04:13,320 --> 00:04:15,920
stuff. 
But for teas that oxidize fast, 

79
00:04:16,120 --> 00:04:21,079
like they, they don't have the 
infrastructure yet. 

80
00:04:21,160 --> 00:04:24,280
And so the idea is to actually 
ensure that they have the, the 

81
00:04:24,280 --> 00:04:27,880
steaming and the drying 
technology there right away #1 

82
00:04:28,240 --> 00:04:31,080
but then is there an area in 
Taiwan that actually can 

83
00:04:31,080 --> 00:04:36,320
actually handle the how to deal 
with oxidized tea? 

84
00:04:36,440 --> 00:04:38,840
Basically, like once you've 
tried, how do you process it? 

85
00:04:39,200 --> 00:04:42,360
There's only one area in Taiwan 
that actually knows how to do 

86
00:04:42,360 --> 00:04:43,680
it. 
It's a it's a, it's a region 

87
00:04:43,680 --> 00:04:47,360
called Sung Chao, which is near 
Banchao. 

88
00:04:47,360 --> 00:04:49,320
Banchao is like a, a major city 
in Taiwan. 

89
00:04:49,400 --> 00:04:52,280
Yeah, OK. 
And so ironically, that's an 

90
00:04:52,280 --> 00:04:54,320
area that I've been trying to 
convince my family to move to 

91
00:04:54,800 --> 00:04:57,440
like in other like to stay in 
Taiwan just because I love the 

92
00:04:57,440 --> 00:05:00,080
area a lot. 
It's got, it's got a lot of 

93
00:05:00,080 --> 00:05:02,240
great cycling, dude. 
Like the, the cycling there is 

94
00:05:02,240 --> 00:05:04,240
gorgeous. 
There's a lot of really, really 

95
00:05:04,240 --> 00:05:06,760
painfully painful hills. 
It's amazing. 

96
00:05:06,760 --> 00:05:10,560
So, so, so I'm basically getting
the agents to debate with each 

97
00:05:10,560 --> 00:05:13,760
other to get me to dive in. 
And so it's showing me which 

98
00:05:13,760 --> 00:05:18,000
areas of social I can work with 
specifically, which with 

99
00:05:18,000 --> 00:05:20,680
farmers, which small 
distributors I can work with to 

100
00:05:20,680 --> 00:05:24,480
basically figure out how to go 
ahead and have them process the 

101
00:05:24,480 --> 00:05:28,720
oxidized tea as well. 
So it's like, if I'm actually 

102
00:05:28,720 --> 00:05:31,840
ever decide that I want to go 
into that business, then the 

103
00:05:31,840 --> 00:05:34,360
idea is like, yeah, like 
obviously it's a side business, 

104
00:05:34,360 --> 00:05:37,880
but as you can tell, I'm just a 
tiny bit OCD stuff. 

105
00:05:37,880 --> 00:05:40,640
Like that? 
Just a little bit, yeah, yeah, 

106
00:05:40,640 --> 00:05:43,000
just just a little bit. 
So so because of that, then the 

107
00:05:43,000 --> 00:05:45,680
idea is like the one of my fun 
projects when if I'm not cycling

108
00:05:45,680 --> 00:05:49,040
would be like, OK, let's figure 
out how to actually like grow 

109
00:05:49,080 --> 00:05:52,320
high quality matcha. 
So obviously Japan has the lead.

110
00:05:52,480 --> 00:05:55,600
Korea's infrastructure is 
amazing in typical fashion, the 

111
00:05:55,600 --> 00:05:57,520
Korea does. 
And so for me, I'm like, huh, 

112
00:05:58,280 --> 00:06:00,720
can I do this in Taiwan? 
Because the the final tip is 

113
00:06:01,000 --> 00:06:06,480
Taiwan Nanto elevation is 
typically isn't high enough to 

114
00:06:06,480 --> 00:06:10,000
grow some of the very 
interesting coffee beans. 

115
00:06:10,880 --> 00:06:15,840
Yeah, it does. 
So for some reason it works even

116
00:06:15,840 --> 00:06:18,840
though it shouldn't exactly. 
You have the theory that that 

117
00:06:18,840 --> 00:06:22,280
also can translate to matcha, 
matcha. 

118
00:06:23,080 --> 00:06:27,440
And now with these agents 
debating each other, that's just

119
00:06:27,440 --> 00:06:29,840
giving you all the context that 
you would need. 

120
00:06:29,840 --> 00:06:32,920
If you want to pull the trigger 
and say, OK, I'm going to go do 

121
00:06:32,920 --> 00:06:37,320
it now, I've got it, all the 
information I need, I can point 

122
00:06:37,320 --> 00:06:41,840
another agent at it and say read
through this context and then 

123
00:06:41,880 --> 00:06:46,600
give me what I need. 
Yeah, so I save obviously all 

124
00:06:46,760 --> 00:06:49,880
the all the like, it's the old 
like build skills or build an 

125
00:06:49,880 --> 00:06:51,200
agentic memory to store all the 
stuff. 

126
00:06:51,200 --> 00:06:53,440
So I have a reference point 
because there's only so much 

127
00:06:53,440 --> 00:06:56,360
context you can compress before 
the thing blows up, right? 

128
00:06:56,520 --> 00:06:59,240
But don't forget it's like 
things like, OK, cool, now I 

129
00:06:59,240 --> 00:07:01,600
have a general idea of what's 
going on. 

130
00:07:01,600 --> 00:07:03,520
But then how do I for seek 
argument? 

131
00:07:04,040 --> 00:07:06,120
My mandarin's horrible. 
OK, let's be very clear. 

132
00:07:06,120 --> 00:07:12,240
It is disgustingly bad. 
OK, my at the time 5 year old 

133
00:07:12,360 --> 00:07:15,480
when we're in Taiwan makes fun 
of my Chinese accent and she's 

134
00:07:15,560 --> 00:07:17,040
it's fair. 
It's fair game. 

135
00:07:17,040 --> 00:07:19,200
It's not even like no, that's 
true, right? 

136
00:07:19,560 --> 00:07:22,680
Because my my Mandarin sounds 
like I'm speaking like 

137
00:07:22,680 --> 00:07:25,440
Cantonese, like I have a 
Cantonese accent Mandarin. 

138
00:07:25,800 --> 00:07:28,280
And mainly because I've loaded 
most of my Mandarin by watching 

139
00:07:28,280 --> 00:07:30,040
Jackie Chan movies that were 
dumped into Mandarin. 

140
00:07:30,440 --> 00:07:32,720
And so they had a Cantonese 
speaker speaking Mandarin. 

141
00:07:32,960 --> 00:07:34,680
And that's how I actually 
learned a lot of my Mandarin. 

142
00:07:35,120 --> 00:07:37,520
I wish I was joking. 
This is completely freaking 

143
00:07:37,520 --> 00:07:41,960
true. 
OK, so, so in other words, no, 

144
00:07:41,960 --> 00:07:43,960
I'm, I'm bad. 
OK. 

145
00:07:44,320 --> 00:07:48,040
And so now I need to go ahead 
and mess around with agents to 

146
00:07:48,040 --> 00:07:50,600
go ahead and read Mandarin 
properly. 

147
00:07:51,040 --> 00:07:54,080
And a lot of it's handwritten, 
not digitized, because we're 

148
00:07:54,080 --> 00:07:56,640
talking small farmers, small 
distributors, right? 

149
00:07:56,880 --> 00:08:00,080
So this is when I start 
switching over to the the 

150
00:08:00,080 --> 00:08:04,760
DeepSeek models, except now here
we get into this, the DeepSeek 

151
00:08:04,760 --> 00:08:07,080
models and all the Chinese open 
source models are great, but 

152
00:08:07,080 --> 00:08:10,600
they're for simplified Chinese, 
not for traditional Chinese and 

153
00:08:10,600 --> 00:08:13,800
traditional Chinese because 
there's actually there's 

154
00:08:13,800 --> 00:08:17,320
actually more context. 
So you actually need more tokens

155
00:08:17,480 --> 00:08:21,560
in order to process the same 
amount of Chinese characters for

156
00:08:21,680 --> 00:08:23,800
traditional Chinese versus 
simplified Chinese. 

157
00:08:23,880 --> 00:08:26,040
OK, I forget the name of it. 
Those are really good. 

158
00:08:26,040 --> 00:08:30,520
Cool researcher, she had done 
all this research about creating

159
00:08:30,960 --> 00:08:35,600
Taiwanese specific models. 
All right, LMS and how that it 

160
00:08:35,600 --> 00:08:38,919
was actually required more 
context required more tokens to 

161
00:08:38,919 --> 00:08:41,200
process this in essence the same
amount of words. 

162
00:08:41,600 --> 00:08:44,320
And so there was some, I forgot 
who did it, but there was some 

163
00:08:44,320 --> 00:08:47,480
person was actually trying to do
like, no, no, we'll translate 

164
00:08:47,480 --> 00:08:52,840
the, the traditional Chinese to 
simplify Chinese run against the

165
00:08:52,840 --> 00:08:55,120
LMS that way and then translate 
it back. 

166
00:08:55,400 --> 00:08:58,240
And because, because then 
they'll be less tokens required.

167
00:08:58,480 --> 00:09:01,720
And the translation actually is 
relatively straightforward. 

168
00:09:01,880 --> 00:09:05,040
It's just a direct mapping in a 
lot of cases you don't like need

169
00:09:05,040 --> 00:09:06,640
thinking. 
In that case, it's like, no, 

170
00:09:07,160 --> 00:09:09,200
here's a simple vibe version. 
Here's the traditional version. 

171
00:09:09,200 --> 00:09:11,400
So I'm actually still testing 
that out. 

172
00:09:11,400 --> 00:09:14,640
I, I screwed up my selections of
those models. 

173
00:09:14,880 --> 00:09:16,400
Why? 
Because my Mandarin sucks. 

174
00:09:16,480 --> 00:09:19,880
I think that's the thing here. 
But the idea is that I'm, I'm 

175
00:09:19,880 --> 00:09:21,960
trying out different things like
that because then that's the 

176
00:09:21,960 --> 00:09:27,680
whole point. 
Like how do I then work with the

177
00:09:27,680 --> 00:09:31,480
traditional farmer, the 
traditional distributor, with 

178
00:09:31,800 --> 00:09:37,200
the ones that actually have real
longevity and real knowledge of 

179
00:09:37,200 --> 00:09:42,800
the topic, right. 
And so, so part of the reason 

180
00:09:42,800 --> 00:09:44,360
I'm bringing this up, by the 
way, is just because a lot of 

181
00:09:44,360 --> 00:09:48,160
people talk about how AI agents 
are removing jobs, right? 

182
00:09:48,160 --> 00:09:51,400
And, and there's obviously 
validity of that statement by 

183
00:09:51,400 --> 00:09:53,560
every stretch of imagination. 
But I keep trying to remind 

184
00:09:53,560 --> 00:09:55,920
people like, yeah, but it's like
every other technology we have, 

185
00:09:55,920 --> 00:09:59,360
it also introduces new jobs. 
It also introduces a way for us 

186
00:09:59,360 --> 00:10:03,200
to rediscover things that we 
forgot using this small example 

187
00:10:03,200 --> 00:10:06,400
of like this, my obsession with 
matcha or my obsession with tea 

188
00:10:06,400 --> 00:10:08,840
or coffee in general, right? 
No, just any obsession I have, 

189
00:10:09,080 --> 00:10:11,200
right? 
Now you can just ask an agent to

190
00:10:11,200 --> 00:10:14,240
go do these things that you 
probably would have never gotten

191
00:10:14,240 --> 00:10:16,800
around to doing. 
No, because think about how much

192
00:10:16,800 --> 00:10:19,000
research you can think of how 
horrible my Mandarin is, right? 

193
00:10:19,360 --> 00:10:22,320
And so instead I'm having fun 
being a geek again, But then I'm

194
00:10:22,320 --> 00:10:25,400
also connecting with old 
farmers. 

195
00:10:26,560 --> 00:10:27,840
I'm connecting with old 
distributors. 

196
00:10:27,880 --> 00:10:32,040
Like this is fucking sweet. 
Like, so in other words, I get 

197
00:10:32,040 --> 00:10:35,040
to go ahead and work with people
that normally I wouldn't even 

198
00:10:35,040 --> 00:10:38,360
have a clue to where to even 
start now, right? 

199
00:10:38,360 --> 00:10:40,040
Which is. 
Which is to say that's. 

200
00:10:40,160 --> 00:10:43,240
That's cool. 
Yeah, that is very cool that. 

201
00:10:43,360 --> 00:10:49,040
And so you've got these you 
mentioned there's the setup 

202
00:10:49,280 --> 00:10:52,360
where the agents will debate 
each other and they debate each 

203
00:10:52,360 --> 00:10:57,120
other just because you're 
saying, hey, look, give me a 

204
00:10:57,120 --> 00:10:59,640
reason not to do this, like try 
and. 

205
00:10:59,720 --> 00:11:00,560
Find. 
No, it's not even that. 

206
00:11:00,560 --> 00:11:03,160
No, no, no, it's it's actually 
better than that. 

207
00:11:03,160 --> 00:11:06,240
It's not like, because most of 
the time, at least with the 

208
00:11:06,240 --> 00:11:09,280
prompts I use or the context I'm
using, it's not about doing or 

209
00:11:09,280 --> 00:11:11,680
not doing. 
No, you, you've known me. 

210
00:11:11,680 --> 00:11:15,960
Not unknown Demetrius, I'm 
always going to go do it right, 

211
00:11:17,120 --> 00:11:18,760
right. 
But my problem isn't am I going 

212
00:11:18,760 --> 00:11:21,040
to do it or not. 
My problem is like, am I going 

213
00:11:21,040 --> 00:11:24,520
to basically slam into a wall or
how thick that wall is when I 

214
00:11:24,520 --> 00:11:26,600
slam into it right? 
That that's usually my problem. 

215
00:11:26,800 --> 00:11:31,560
So so you can tell like basic 
how I look at the tech is 

216
00:11:31,560 --> 00:11:34,120
actually you can still look at 
it from an optimistic 

217
00:11:34,120 --> 00:11:37,000
perspective, right? 
And so for me, it's not about 

218
00:11:37,400 --> 00:11:38,800
like, should I do it or not 
should I do it? 

219
00:11:38,800 --> 00:11:42,200
It's more like what's the right 
way to do it? 

220
00:11:42,320 --> 00:11:44,600
Like. 
And so in other words, as you 

221
00:11:44,600 --> 00:11:46,760
can probably hear or tell from 
the way I'm describing 

222
00:11:46,760 --> 00:11:51,720
everything, my obsession is high
quality going back to the roots 

223
00:11:51,720 --> 00:11:54,640
of actual like tea making, 
right? 

224
00:11:54,800 --> 00:11:59,840
I'm not an expert in that at 
all, but I respect the quality. 

225
00:11:59,840 --> 00:12:02,360
I respect the art. 
I respect the cultural 

226
00:12:02,360 --> 00:12:04,720
significance of of it. 
That that I can say. 

227
00:12:05,160 --> 00:12:07,000
Yeah, I can vibe with that. 
Right. 

228
00:12:07,120 --> 00:12:09,880
And so because of that, then the
idea is that what I'm asking the

229
00:12:09,880 --> 00:12:13,880
agents is basically look, I'm 
giving the agents that context. 

230
00:12:14,600 --> 00:12:19,200
So then why they're debating 
with each other, like, oh, well 

231
00:12:19,200 --> 00:12:22,120
then what region of Taiwan am I 
looking at? 

232
00:12:22,280 --> 00:12:26,560
What's the elevation, what's the
pH, what's the soil quality? 

233
00:12:26,720 --> 00:12:31,880
Am I dealing with lots of, for 
example, like the theory is if 

234
00:12:31,880 --> 00:12:35,280
we were to grow matcha in 
Taiwan, it would be more floral 

235
00:12:35,280 --> 00:12:37,640
based, especially in Nanto 
County. 

236
00:12:37,640 --> 00:12:40,600
Why? 
Because that's what Nanto's 

237
00:12:40,840 --> 00:12:42,960
famous for. 
Like their teas have a lot of 

238
00:12:42,960 --> 00:12:45,960
floral, lot of lychee like. 
So for example, like coffee, 

239
00:12:45,960 --> 00:12:48,080
that coffee block that I, I sent
you, right? 

240
00:12:48,640 --> 00:12:52,880
That one, his coffee beans were 
notorious, were famous, 

241
00:12:53,200 --> 00:12:59,480
notorious, the wrong word famous
partially because there was very

242
00:12:59,480 --> 00:13:02,200
strong, very, very strong hints 
of lychee in the coffee beans 

243
00:13:02,200 --> 00:13:03,320
themselves. 
Right. 

244
00:13:03,840 --> 00:13:06,360
And the reason why is because 
literally the lychees were 

245
00:13:06,360 --> 00:13:11,680
growing wild in his coffee farm.
So literally like no joke, we 

246
00:13:11,680 --> 00:13:15,040
when we got there to try out the
coffee, the grandma was 

247
00:13:15,040 --> 00:13:17,000
literally just giving us lychees
for free. 

248
00:13:17,000 --> 00:13:19,080
And just like, don't you want us
to pay for them? 

249
00:13:19,280 --> 00:13:22,520
No, no, they grow everywhere. 
They're literally a weed at this

250
00:13:22,520 --> 00:13:24,760
point. 
Like I'm like, I've never heard 

251
00:13:24,920 --> 00:13:28,800
lychees referred to as weeds, 
but OK, like right, but that's 

252
00:13:28,800 --> 00:13:31,560
literally the context. 
It's like 05 be able to find 

253
00:13:31,560 --> 00:13:35,640
cool things like that. 
And so that's the that's the 

254
00:13:35,640 --> 00:13:37,760
attempt to discover. 
That's the attempt to like try 

255
00:13:37,760 --> 00:13:42,200
to like somebody like myself who
has the obsession of it or the 

256
00:13:42,200 --> 00:13:45,360
OCD Ness, but not necessarily 
the knowledge. 

257
00:13:45,360 --> 00:13:49,640
Letting the two systems debate 
with each other provide me the 

258
00:13:49,640 --> 00:13:52,840
additional information to say to
know what other questions to 

259
00:13:52,840 --> 00:13:55,480
ask. 
Yeah, because when you read 

260
00:13:55,480 --> 00:13:58,200
through that debate, then you 
go, oh, I didn't even think 

261
00:13:58,200 --> 00:14:03,560
about that. 
That is such a great trick that 

262
00:14:03,600 --> 00:14:07,680
I recognize and you know what? 
I one of the actually toys that 

263
00:14:07,680 --> 00:14:13,080
I've been making, and I don't 
know why this isn't a common UX 

264
00:14:13,400 --> 00:14:21,680
for dealing with agents is I now
have created a little widget 

265
00:14:21,840 --> 00:14:26,800
that you can highlight any piece
of text, whether you're inside 

266
00:14:26,800 --> 00:14:30,960
of the chat window with the 
agents or you're on a website. 

267
00:14:31,440 --> 00:14:35,120
And it will pop up a little 
prompt box and it will read the 

268
00:14:35,120 --> 00:14:38,920
context of whatever the either 
the paragraph or the whole page 

269
00:14:38,920 --> 00:14:42,320
that you're on. 
And you can prompt whatever ask 

270
00:14:42,320 --> 00:14:45,320
a question to it just about. 
Because a lot of times what I 

271
00:14:45,320 --> 00:14:49,880
recognize with myself is that I 
don't want to continue this 

272
00:14:50,160 --> 00:14:53,320
conversation in the actual chat 
that I'm having. 

273
00:14:53,320 --> 00:14:57,400
I want to shoot off and do this 
on a different thread because 

274
00:14:57,400 --> 00:15:02,200
it's just like something that 
isn't necessarily needed to 

275
00:15:02,200 --> 00:15:05,440
pollute this conversation with 
the agent that I'm having. 

276
00:15:05,440 --> 00:15:08,480
But I still want to know like, 
hey, that you see, you reference

277
00:15:08,480 --> 00:15:11,080
this thing here. 
What does that mean? 

278
00:15:11,080 --> 00:15:13,400
Go deeper. 
Give me more context on this 

279
00:15:13,400 --> 00:15:16,800
specific thing, but not in the 
main chat. 

280
00:15:16,800 --> 00:15:21,360
So it's like little sidebar 
conversations and I've got this.

281
00:15:22,400 --> 00:15:25,120
I, I use this tool called 
clearance to read like markdown 

282
00:15:25,120 --> 00:15:27,880
files. 
And so now I created a little 

283
00:15:27,880 --> 00:15:32,600
right hand sidebar where you can
have a side chat with an agent. 

284
00:15:33,200 --> 00:15:36,680
OK, I I know this is definitely 
wasn't done on purpose for 

285
00:15:36,680 --> 00:15:39,960
everybody that's listening to us
right now, but you you wonderful

286
00:15:39,960 --> 00:15:45,280
Segway to Omniton by the way, 
because OK, so for example, I 

287
00:15:45,280 --> 00:15:48,560
did this already. 
So I when I first started asking

288
00:15:48,560 --> 00:15:51,560
the question about Machos and 
things like that, like what 

289
00:15:51,560 --> 00:15:55,880
regions, I learned a lot about 
more about the regions of Japan 

290
00:15:56,520 --> 00:15:58,520
for starters, and then the 
regions of Korea. 

291
00:15:58,520 --> 00:16:03,880
All right. 
But then where my end goal was 

292
00:16:03,880 --> 00:16:06,440
to figure out if I could do it 
in Nonto County in Taiwan. 

293
00:16:06,560 --> 00:16:10,880
All right. 
But of course, I, as I'm to your

294
00:16:10,880 --> 00:16:13,480
point, I'm learning more about 
these reasons to Japan and these

295
00:16:13,480 --> 00:16:16,640
reasons in Korea. 
I'm like oh this is really cool 

296
00:16:17,040 --> 00:16:20,320
so I want to know more but now. 
So I really wish there was a way

297
00:16:20,320 --> 00:16:23,240
to fork my conversation. 
Yes, that's it. 

298
00:16:24,160 --> 00:16:26,960
So homage, it has the ability 
for me to fork. 

299
00:16:28,280 --> 00:16:30,240
So that's what I did. 
I basically had the 

300
00:16:30,240 --> 00:16:33,000
conversation, I forked it off 
that I allowed myself to 

301
00:16:33,000 --> 00:16:36,160
continue one conversation 
specifically about the Japan and

302
00:16:36,160 --> 00:16:41,440
Korea areas, while the other one
the the main line, I just 

303
00:16:41,520 --> 00:16:44,680
redirected the Nanto County. 
And so they both had the 

304
00:16:44,680 --> 00:16:47,560
original context by which we we 
had the conversation with. 

305
00:16:47,760 --> 00:16:50,440
But now I can have two very 
separate conversations so I can 

306
00:16:50,600 --> 00:16:51,760
learn more. 
Why? 

307
00:16:51,880 --> 00:16:55,120
Because I'm fortunate enough for
it, because I live in Seattle. 

308
00:16:56,080 --> 00:17:00,040
No, I'm not being paid for this,
but there's direct on Alaska 

309
00:17:00,040 --> 00:17:05,400
Airlines from Seattle to Tokyo, 
Seattle to Seoul now. 

310
00:17:05,599 --> 00:17:08,720
So I'm going. 
So yeah, so I hopefully maybe 

311
00:17:08,720 --> 00:17:09,880
they'll give me something for 
that. 

312
00:17:09,880 --> 00:17:13,960
But in that the reason I called 
that I was like. 

313
00:17:14,280 --> 00:17:16,880
Sponsored by Sponsored. 
Right. 

314
00:17:16,880 --> 00:17:22,240
But no, it's not I wish I'm not 
that we're we're both not big. 

315
00:17:22,440 --> 00:17:23,720
We're both not big enough for 
that right. 

316
00:17:23,839 --> 00:17:24,839
But. 
Exactly. 

317
00:17:24,839 --> 00:17:27,200
We need to become the true 
travel bloggers. 

318
00:17:27,200 --> 00:17:28,359
Exactly. 
Yeah, we're. 

319
00:17:28,359 --> 00:17:30,240
Just we're not we're not travel 
influencers. 

320
00:17:30,240 --> 00:17:31,920
So unfortunately that's not 
going to work. 

321
00:17:31,960 --> 00:17:33,360
But not yet. 
There you go. 

322
00:17:33,360 --> 00:17:34,320
True. 
Point taken. 

323
00:17:34,320 --> 00:17:38,400
Yeah, we we have goals, right? 
But nevertheless, the idea being

324
00:17:38,600 --> 00:17:43,680
now I'll go directly to Korea, 
I'll go directly to Japan so I 

325
00:17:43,680 --> 00:17:46,480
can explore the regions myself. 
Yeah. 

326
00:17:46,480 --> 00:17:47,800
So that's why I started going 
down. 

327
00:17:48,040 --> 00:17:50,480
So I forked the conversation 
specifically so I could go do 

328
00:17:50,480 --> 00:17:52,280
that, yeah. 
But I don't understand why that 

329
00:17:52,280 --> 00:17:55,200
isn't a common UX pattern with 
these agents. 

330
00:17:55,200 --> 00:18:00,480
And specifically for me it's 
like I have this one sentence or

331
00:18:00,480 --> 00:18:04,080
these 5 words that appear 
together in my context and I 

332
00:18:04,080 --> 00:18:08,680
want to fork them out and have a
separate thread on that specific

333
00:18:08,680 --> 00:18:12,680
piece. 
Like explain more but I haven't 

334
00:18:12,680 --> 00:18:20,240
seen any of the labs create this
as a integral part of how we 

335
00:18:20,240 --> 00:18:25,800
interact with agents and so it 
it makes me think like maybe I'm

336
00:18:25,800 --> 00:18:28,680
the only one who cares about. 
That no, no, no, no, no, no, no,

337
00:18:28,680 --> 00:18:31,840
no, no, no, no. 
You're, you're look, when we 

338
00:18:31,840 --> 00:18:34,880
started all this, this entire AI
community, right? 

339
00:18:35,720 --> 00:18:40,160
Think about how chat windows are
still the even after all this 

340
00:18:40,160 --> 00:18:44,240
time, chat windows are still the
way we work with things even our

341
00:18:44,240 --> 00:18:47,640
sticking when we decide to use 
CLI and terminals, it's still 

342
00:18:47,640 --> 00:18:51,960
basically ACLI chat window 
That's basically all this for 

343
00:18:51,960 --> 00:18:53,840
all intensive purposes, right? 
Just that we might have some 

344
00:18:53,840 --> 00:18:56,360
Jason stuff shoved in there. 
Now that's that's the extent of 

345
00:18:56,360 --> 00:19:02,200
the difference, right? 
The reason why is because, and 

346
00:19:02,200 --> 00:19:05,120
you actually see harnesses like 
being created right now, like a 

347
00:19:05,240 --> 00:19:07,440
bunch of start-ups, a bunch, 
even the major lab, they all 

348
00:19:07,440 --> 00:19:08,720
create harnesses themselves, 
right? 

349
00:19:09,200 --> 00:19:13,640
And the reason why is because 
it's like look up to this point.

350
00:19:13,720 --> 00:19:16,400
And this is like our definitely 
our learnings within database 

351
00:19:16,400 --> 00:19:20,440
itself, right? 
We all had like 5 terminal 

352
00:19:20,440 --> 00:19:23,320
windows open, 6 browser windows 
open. 

353
00:19:23,320 --> 00:19:26,320
Yeah, we're all going back and 
forth, back and forth. 

354
00:19:27,560 --> 00:19:30,640
You deep dive on a particular 
topic. 

355
00:19:30,640 --> 00:19:32,920
I'm coding something for Delta 
Rust, right? 

356
00:19:33,360 --> 00:19:35,560
Like so I can get off of T for a
second, right. 

357
00:19:36,480 --> 00:19:38,680
So I'm, I'm coding something for
Delta Rust. 

358
00:19:38,760 --> 00:19:45,000
I'm trying to understand the 
latest version of arrow or and 

359
00:19:45,000 --> 00:19:47,120
data fusion because they keep 
changing. 

360
00:19:48,240 --> 00:19:49,840
Hats off to the convenience. 
This is not a complaint. 

361
00:19:49,840 --> 00:19:52,080
This is just, but they do they 
do change things. 

362
00:19:52,080 --> 00:19:54,160
So we're trying to keep Delta 
Rust up to date, right? 

363
00:19:54,680 --> 00:19:57,080
And so when we're trying to make
those changes, then I need to 

364
00:19:57,680 --> 00:20:02,280
run some terminal windows to I 
need to figure out. 

365
00:20:02,960 --> 00:20:07,200
What's the current version? 
Another terminal set is creating

366
00:20:07,440 --> 00:20:10,960
the tests harnesses that I want 
to go create another browser 

367
00:20:10,960 --> 00:20:14,480
window is giving me oh, what's 
the latest security contacts? 

368
00:20:14,480 --> 00:20:18,840
What's the latest like, you 
know, security alerts associated

369
00:20:18,960 --> 00:20:22,760
to each and everyone, the 20 
different dependencies we have 

370
00:20:23,320 --> 00:20:24,520
that, that we have to go 
through, right? 

371
00:20:25,080 --> 00:20:28,240
And, and what are you doing? 
You're opening multiple windows 

372
00:20:28,240 --> 00:20:31,680
and you're losing context, no 
pun intended, right? 

373
00:20:32,960 --> 00:20:39,360
And you're, you're losing all of
the connections because you've 

374
00:20:39,360 --> 00:20:41,560
got so many windows, so many 
terminal windows open, right? 

375
00:20:41,800 --> 00:20:48,160
So you know how Amazon made the 
idea of like the each team 

376
00:20:48,160 --> 00:20:51,240
should only have like 2 beats of
boxes, right? 

377
00:20:51,240 --> 00:20:56,440
All right, the whole reason that
was Crave was because it was the

378
00:20:56,440 --> 00:20:59,640
idea that each team should only 
have five to seven people 

379
00:20:59,640 --> 00:21:02,000
reporting to it. 
OK, so each manager should only 

380
00:21:02,000 --> 00:21:04,200
have 5:00 to 7:00 because 
anything greater than that, they

381
00:21:04,200 --> 00:21:06,920
basically start losing context 
and again, pun intended. 

382
00:21:07,240 --> 00:21:11,840
OK, well, this was determined, 
not Amazon might have made it 

383
00:21:11,840 --> 00:21:13,560
famous, but it was actually a 
military thing. 

384
00:21:13,720 --> 00:21:17,360
Like military, military command.
In terms of units, teams 

385
00:21:17,560 --> 00:21:23,560
understood that each commander 
could only logistically handle 

386
00:21:23,560 --> 00:21:26,760
five to seven people reporting 
to them in order to be 

387
00:21:26,760 --> 00:21:27,920
effective. 
Right. 

388
00:21:27,920 --> 00:21:31,040
And so I'm not military, but I I
am. 

389
00:21:31,280 --> 00:21:39,360
I used to be King County OPS, 
the search and rescue OPS. 

390
00:21:39,360 --> 00:21:40,560
Excuse me. 
I'm sorry. 

391
00:21:40,640 --> 00:21:43,920
I'm trying to remember SNR look,
this is how tired I am. 

392
00:21:43,920 --> 00:21:46,080
OK, dude. 
Search and rescue operations. 

393
00:21:46,080 --> 00:21:50,600
So we learned that from so SNR 
OPS learned that from obviously 

394
00:21:50,600 --> 00:21:51,920
military. 
And then the idea is that this 

395
00:21:51,920 --> 00:21:56,320
is a a time value tradition, 
like an understanding and so so 

396
00:21:56,320 --> 00:21:59,120
in the tech world, everybody 
thought about two pizza boxes, 

397
00:21:59,120 --> 00:22:03,480
but basically this is a well 
understood logistical structure 

398
00:22:04,200 --> 00:22:05,480
within the military. 
OK. 

399
00:22:05,520 --> 00:22:09,080
And so cool. 
The reason I bring that up is 

400
00:22:09,080 --> 00:22:14,280
because, well, then in a lot of 
ways, if you consider an agent, 

401
00:22:14,960 --> 00:22:16,720
I'm almost like a junior 
developer that you're managing 

402
00:22:20,440 --> 00:22:22,640
once you go beyond 5. 
Yeah, yeah. 

403
00:22:22,880 --> 00:22:26,240
Once you go with seven, you're 
losing track of everything right

404
00:22:26,800 --> 00:22:31,520
And so more than anything else, 
the things like to your point 

405
00:22:31,520 --> 00:22:36,160
exactly, but I want to go off 
and do that it's actually going 

406
00:22:36,160 --> 00:22:39,320
to make you lose context even 
more so right. 

407
00:22:39,960 --> 00:22:44,440
Part of the reason why we we've 
had basically built this idea of

408
00:22:44,440 --> 00:22:47,840
forking was actually not 
originally that idea you talked 

409
00:22:47,840 --> 00:22:49,760
about, by the way, like to be 
very clear, like it was more 

410
00:22:49,760 --> 00:22:53,560
about literally we're coding. 
So you fork. 

411
00:22:53,560 --> 00:22:56,520
So that way you can literally go
ahead and try out different 

412
00:22:56,520 --> 00:22:59,520
ideas, but had the initial same 
context. 

413
00:23:00,000 --> 00:23:03,440
But to your point exactly, 
coding and actually having a 

414
00:23:03,440 --> 00:23:05,480
real conversation isn't actually
that much different. 

415
00:23:05,480 --> 00:23:08,360
Coding happens to be just around
code, but it's still an 

416
00:23:08,360 --> 00:23:10,920
abstraction layer, right? 
And so conversations are just 

417
00:23:10,920 --> 00:23:12,720
another abstraction layer. 
Like everyone's like, well, 

418
00:23:12,720 --> 00:23:14,800
Cody, no, that's real code. 
I'm like, no, are you writing 

419
00:23:14,800 --> 00:23:16,440
this thing in binary or 
assembly? 

420
00:23:16,440 --> 00:23:18,080
No. 
OK, well then stop talking to 

421
00:23:18,080 --> 00:23:19,600
me. 
It's an abstraction layer, OK. 

422
00:23:20,040 --> 00:23:24,760
And so the idea is that if 
you've got all these different 

423
00:23:24,760 --> 00:23:27,880
things floating around right, 
you will lose context. 

424
00:23:27,880 --> 00:23:31,320
You will not remember why you 
asked in the first place. 

425
00:23:31,960 --> 00:23:35,600
So there's so there's a lot of 
research being done by the labs 

426
00:23:35,600 --> 00:23:38,560
and to an extent even a lot of 
the harnesses, which is like how

427
00:23:38,560 --> 00:23:44,040
do you basically have now an 
agent fleet and the fleet manage

428
00:23:44,040 --> 00:23:45,360
itself. 
Right. 

429
00:23:45,920 --> 00:23:49,840
And there's various like by the 
time this podcast comes out, 

430
00:23:49,840 --> 00:23:52,040
which is only a week or two, I'm
sure everything will change. 

431
00:23:52,040 --> 00:23:54,360
So I'm not going to bother with 
the pretense of knowing what's 

432
00:23:54,360 --> 00:23:57,080
going to happen with fleets. 
But irrelevant, whether you you 

433
00:23:57,080 --> 00:24:00,400
call the fleet or you talk about
an agent or there's a, some 

434
00:24:00,400 --> 00:24:04,160
other layer, the the whole 
reason we built what we built 

435
00:24:04,160 --> 00:24:06,720
was because we understood that 
this was going to be a problem 

436
00:24:06,720 --> 00:24:11,080
either way. 
And just as importantly, or if 

437
00:24:11,080 --> 00:24:14,800
not in many cases more 
importantly, the models keep 

438
00:24:14,800 --> 00:24:17,160
changing. 
The hardest keep changing. 

439
00:24:18,320 --> 00:24:21,120
I want to be able to go ahead 
and move all that context, all 

440
00:24:21,120 --> 00:24:23,080
the agentic memory, all that 
configuration. 

441
00:24:23,720 --> 00:24:27,160
I don't want to rebuild it so 
now I can use the latest open AI

442
00:24:27,160 --> 00:24:30,360
model or the latest DeepSeek 
model or the latest anthropic 

443
00:24:30,360 --> 00:24:31,760
model or whatever it is I'm 
using. 

444
00:24:32,280 --> 00:24:35,720
I want this one layer that says 
whatever harness you're using, 

445
00:24:36,000 --> 00:24:40,080
whatever model you're using, 
it'll flow through irrelevant. 

446
00:24:41,560 --> 00:24:45,080
Where I get a little bit 
confused there is the harnesses 

447
00:24:45,080 --> 00:24:48,200
are optimized for certain 
models, right? 

448
00:24:48,200 --> 00:24:52,120
I guess if you're building a 
harness, you're building it for 

449
00:24:52,240 --> 00:24:55,720
a model and there's certain 
things that the labs are going 

450
00:24:55,720 --> 00:24:59,400
to be doing on their own that's 
like the inner harness and then 

451
00:24:59,400 --> 00:25:01,720
you have the outer harness and 
the inner harness. 

452
00:25:02,160 --> 00:25:07,480
They're RL ING that model so 
that it works really well with 

453
00:25:07,720 --> 00:25:12,240
their harness. 100%. 
How do you like envision these 

454
00:25:12,240 --> 00:25:15,880
hot swaps in a world where you 
know that? 

455
00:25:16,640 --> 00:25:20,320
Oh, because that that's actually
why we value Omnigen as an open 

456
00:25:20,320 --> 00:25:21,680
source project to begin with, 
right? 

457
00:25:21,680 --> 00:25:24,560
Because the idea is that like 
for example, even just building 

458
00:25:24,560 --> 00:25:30,040
it, we notice for example, what 
deems to be simple but actually 

459
00:25:30,080 --> 00:25:34,520
changes people's workflows is 
like approvals, like the 

460
00:25:34,520 --> 00:25:37,680
approval process. 
I'm not going to get this quite 

461
00:25:37,680 --> 00:25:38,640
right. 
So I'm going to stay vague. 

462
00:25:38,640 --> 00:25:41,680
So that way I can, I can not 
make it sound like 1's better 

463
00:25:41,680 --> 00:25:42,960
than the other because that's 
not true at all. 

464
00:25:42,960 --> 00:25:48,040
OK, But the process is slightly 
different between harness A and 

465
00:25:48,040 --> 00:25:51,880
harness B and the process is 
slightly different between model

466
00:25:51,880 --> 00:25:53,720
A and model B, just to be very 
vague. 

467
00:25:53,720 --> 00:25:56,320
So that way, again, I'm not 
calling anybody out, right? 

468
00:25:56,640 --> 00:25:58,880
And again, it's not even calling
because like I said, it's not 

469
00:25:58,880 --> 00:26:01,000
like 1's better or one's worse. 
It's just purely they're 

470
00:26:01,000 --> 00:26:02,840
different. 
That's all this is OK. 

471
00:26:03,480 --> 00:26:07,400
So if if you look in the Omni 
code base, that's we actually 

472
00:26:07,400 --> 00:26:10,960
reflect that, right? 
And so then that way we're 

473
00:26:10,960 --> 00:26:17,840
actually leveraging what model A
or harness A is really good at 

474
00:26:18,640 --> 00:26:22,760
because like the creation object
isn't about going and saying 

475
00:26:22,760 --> 00:26:24,920
we're a better harness than 
everybody else. 

476
00:26:25,000 --> 00:26:27,800
We're not. 
No, that's why we call ourselves

477
00:26:27,800 --> 00:26:30,360
a meta harness because we're 
saying no, no, we're on top of 

478
00:26:30,360 --> 00:26:33,960
this. 
So that way you, the individual,

479
00:26:34,680 --> 00:26:38,920
the person that's using these 
things, all that configuration, 

480
00:26:38,920 --> 00:26:45,200
all that context, all that agent
of memory that is portable, that

481
00:26:45,200 --> 00:26:48,040
you have control over, that you 
can flow, right? 

482
00:26:48,320 --> 00:26:52,600
You can compose these different 
ideas where basically I want to 

483
00:26:52,600 --> 00:26:55,920
go ahead and run like just be a 
little bit more specific now, 

484
00:26:56,680 --> 00:27:02,200
forsake argument. 
I was using Opus 4.8 and 

485
00:27:02,200 --> 00:27:05,480
actually 4.7 of them implicate a
real example to generate. 

486
00:27:05,480 --> 00:27:08,160
The code Owen better to be 
truthful. 

487
00:27:08,240 --> 00:27:12,320
Also, I'm glad I glad I have a 
kindred spirit here. 

488
00:27:13,400 --> 00:27:15,440
Right. 
Way more. 4.8 is not. 

489
00:27:15,520 --> 00:27:16,120
I'm not. 
Stoked. 

490
00:27:16,120 --> 00:27:18,480
Yeah, yeah, yeah, yeah, yeah. 
That's a whole other thing. 

491
00:27:18,480 --> 00:27:21,240
But let's not debate that, OK? 
I don't want to throw a kick in 

492
00:27:21,240 --> 00:27:23,040
my ass, OK, Right. 
So right. 

493
00:27:23,160 --> 00:27:24,400
But I'm using global force of 
it. 

494
00:27:24,400 --> 00:27:29,800
And then, but then I was using 
GBD 5.4 at the time to do the 

495
00:27:29,800 --> 00:27:35,320
test cases and fuck yeah, like 
it was wonderful because then I 

496
00:27:35,320 --> 00:27:38,840
got the 2 using Poly to debate 
each other in terms of high 

497
00:27:38,840 --> 00:27:40,920
quality coding. 
And then I and then I got one to

498
00:27:40,920 --> 00:27:43,720
actually go ahead and and 
generate the test cases for me 

499
00:27:43,720 --> 00:27:45,600
as well, which is amazing, 
right? 

500
00:27:46,080 --> 00:27:48,200
And then maybe a week later I 
would have switched over to say,

501
00:27:48,560 --> 00:27:52,200
you know what chat GVD 5.5. 
I like their code base a little 

502
00:27:52,200 --> 00:27:53,600
better. 
So I'm using that now. 

503
00:27:53,720 --> 00:27:59,400
I'll stick to Alpos 47 for the 
of the debate it I'll use maybe 

504
00:27:59,400 --> 00:28:02,640
sonnet to to go ahead and do the
the test case because those are 

505
00:28:02,640 --> 00:28:05,120
relatively easier to once you 
put the original template. 

506
00:28:05,120 --> 00:28:07,680
I don't actually need something 
as powerful to generate the test

507
00:28:07,680 --> 00:28:09,160
cases. 
I just need to sort of update 

508
00:28:09,160 --> 00:28:10,800
them. 
So this is perfectly fine. 

509
00:28:11,960 --> 00:28:14,480
And then when you know when 
Fable and all that stuff came 

510
00:28:14,480 --> 00:28:17,680
out that rewrote everything and 
then Fable got removed and then 

511
00:28:17,680 --> 00:28:20,840
we wrote everything right. 
The whole point is that I'm just

512
00:28:20,840 --> 00:28:21,960
using Omni jets. 
I didn't care. 

513
00:28:22,440 --> 00:28:28,120
I just redirected it and like 
hell, I'm using π for a bunch of

514
00:28:28,120 --> 00:28:31,080
my test cases now because it's 
cheap, like really dirt cheap. 

515
00:28:31,080 --> 00:28:33,080
So I'm like, why not? 
It's really good. 

516
00:28:33,480 --> 00:28:35,160
So just use that in some cases, 
right? 

517
00:28:35,320 --> 00:28:38,680
So that's most the point. 
Like it just gives you that 

518
00:28:38,680 --> 00:28:42,000
flexibility to say like, OK, you
start off with everybody telling

519
00:28:42,000 --> 00:28:44,520
you you're supposed to use the 
most advanced 1. 

520
00:28:44,800 --> 00:28:48,040
And I get it that you're tempted
to do that. 

521
00:28:48,040 --> 00:28:50,800
And if you if you're allowed 
toka maxing, you will do that, 

522
00:28:50,800 --> 00:28:52,560
right? 
But the reality is this stuff 

523
00:28:52,560 --> 00:28:54,600
starting to get expensive. 
Right. 

524
00:28:55,080 --> 00:28:58,120
Prices are going up or discounts
are going down or both, right. 

525
00:28:58,560 --> 00:29:02,800
And so my take on the whole 
thing's like, OK, well then try 

526
00:29:02,800 --> 00:29:05,960
to find the right model or the 
right harness for the job. 

527
00:29:06,000 --> 00:29:09,720
And so maybe even if you are 
initially using the more 

528
00:29:09,720 --> 00:29:12,440
advanced models to build it, a 
lot of the maintenance doesn't 

529
00:29:12,440 --> 00:29:14,000
require the most advanced model 
at all. 

530
00:29:14,360 --> 00:29:17,000
In fact, the most advanced 
models more times than not 

531
00:29:17,000 --> 00:29:19,760
actually make things literally 
more complicated. 

532
00:29:19,920 --> 00:29:22,320
Like you're going, I only needed
simple test cases. 

533
00:29:22,640 --> 00:29:26,200
I didn't need this gargantuan 
thing that is impossible for me 

534
00:29:26,200 --> 00:29:26,840
to read. 
Right. 

535
00:29:26,840 --> 00:29:29,520
That's a feature, not a bug. 
That's why Anthropic's worth 

536
00:29:29,520 --> 00:29:33,800
it's whatever trillion dollars 
they are. 

537
00:29:33,800 --> 00:29:35,960
Very no, but see, that's what 
it's like in the end, you get my

538
00:29:35,960 --> 00:29:37,800
trip, right? 
Like it's, it's about going 

539
00:29:37,800 --> 00:29:39,680
ahead. 
And then because after all, like

540
00:29:40,400 --> 00:29:42,400
I'm definitely not mean to get 
in the debate just right now, 

541
00:29:42,400 --> 00:29:45,800
but like during, for example, 
DNA summit during the open lake 

542
00:29:45,800 --> 00:29:48,360
house contributors meeting, 
right? 

543
00:29:49,560 --> 00:29:52,560
One of the sessions that I 
actually end up helping to lead 

544
00:29:52,560 --> 00:29:55,960
was the deal with coding AI 
slump, right? 

545
00:29:56,200 --> 00:30:01,640
And one of the call outs we had,
because the way we run open lake

546
00:30:01,640 --> 00:30:04,520
house was very much in the thing
of like, yeah, I might have led 

547
00:30:04,520 --> 00:30:06,800
the discussion, but it was meant
for everybody to chat. 

548
00:30:06,800 --> 00:30:09,480
I'm not interested in like me 
telling people what to do. 

549
00:30:09,480 --> 00:30:13,040
That's not interesting at all. 
I'm we were getting all these 

550
00:30:13,040 --> 00:30:15,640
people in the, the open lake 
house and the AI community 

551
00:30:15,640 --> 00:30:19,200
specifically to go to debate. 
Like, like we're marking on 

552
00:30:19,200 --> 00:30:20,560
competing projects, whatever, I 
don't care. 

553
00:30:20,560 --> 00:30:23,040
But like, we all are suffering 
from same version of, of 

554
00:30:23,040 --> 00:30:26,800
insanity in one form or another.
So can we help each other out on

555
00:30:26,800 --> 00:30:31,000
that part at least, right. 
And so for example, there was a 

556
00:30:31,000 --> 00:30:36,200
whole debate like on the on 
various distribution lists, 

557
00:30:36,200 --> 00:30:38,680
including like Apache Iceberg on
the delta list, whatever I was 

558
00:30:38,960 --> 00:30:40,520
like created by versus assisted 
by. 

559
00:30:40,520 --> 00:30:43,760
So created by is the person who 
did it assisted by is like it's 

560
00:30:43,760 --> 00:30:46,280
almost 47 or Sonnet or whatever 
it is that you were using at the

561
00:30:46,280 --> 00:30:48,880
time, right? 
And the call out more than 

562
00:30:48,880 --> 00:30:54,160
anything else unequivocally was 
that you needed to. 

563
00:30:55,560 --> 00:30:58,200
In the end, everybody agrees 
with the idea that it is still 

564
00:30:58,200 --> 00:31:03,160
the individual, the user, that 
is the one that submitted the 

565
00:31:03,160 --> 00:31:07,840
code that there still is from a 
TLDR perspective, who's the neck

566
00:31:07,840 --> 00:31:09,640
you're going to choke? 
It's not, I'm not choking the 

567
00:31:09,640 --> 00:31:11,280
neck of an agent. 
I'm not choking the neck of 

568
00:31:11,280 --> 00:31:12,640
Anthropic. 
I'm choking the neck of the 

569
00:31:12,640 --> 00:31:14,600
person that's contributing the 
code, right. 

570
00:31:15,120 --> 00:31:19,640
And so ultimately, like whatever
we're phrasing as in terms of 

571
00:31:19,640 --> 00:31:23,560
creative biases by or anything 
else, the point is that me, the 

572
00:31:23,560 --> 00:31:27,360
developer, whether it's 
developing code or developing a 

573
00:31:27,360 --> 00:31:31,360
business project or whatever it 
is, you, it's still me, it's 

574
00:31:31,360 --> 00:31:33,880
still you, it's still the person
that's making that decision that

575
00:31:33,880 --> 00:31:36,560
how ultimately succeeds or 
fails. 

576
00:31:36,560 --> 00:31:40,040
So whether it's the the code for
Delta rust or whether it's, you 

577
00:31:40,040 --> 00:31:44,600
know, figuring out which region 
of non telecom county that I'm I

578
00:31:44,600 --> 00:31:48,520
want to grow macho leave, right.
Ultimately it's like it's me, 

579
00:31:48,560 --> 00:31:51,440
right. 
The onus has to be on me alone, 

580
00:31:51,680 --> 00:31:56,400
right. 
And so knowing how to follow a 

581
00:31:56,400 --> 00:31:59,280
thread, knowing what these 
harness can do, knowing that you

582
00:31:59,280 --> 00:32:03,760
can move your context, your 
thought processes across the 

583
00:32:03,760 --> 00:32:06,400
different models so that way you
yourself can succeed. 

584
00:32:07,040 --> 00:32:09,840
Ultimately, that's it. 
And and I think you, Dimitra, 

585
00:32:09,840 --> 00:32:11,720
you and I have actually had this
conversation on the side. 

586
00:32:11,720 --> 00:32:14,960
You've also did this as part of 
your own set of podcasts where 

587
00:32:15,400 --> 00:32:19,600
ultimately the main reason 
people fail when they're using 

588
00:32:19,600 --> 00:32:24,440
agents is because they they zero
shot it and they have no context

589
00:32:24,440 --> 00:32:26,800
applied at all. 
They there's no free thinking. 

590
00:32:26,800 --> 00:32:29,080
So people are like, well, of 
course it's garbage that came 

591
00:32:29,080 --> 00:32:30,800
out of it. 
Like, yeah, so that means these 

592
00:32:30,800 --> 00:32:33,400
ancients suck. 
I'm like, well, it's the old 

593
00:32:33,400 --> 00:32:35,840
garbage in, garbage out. 
Like if you, if you're not 

594
00:32:35,840 --> 00:32:38,360
putting any effort into this, 
it's it's going to be garbage 

595
00:32:38,360 --> 00:32:39,840
coming out. 
It doesn't matter how good the 

596
00:32:39,840 --> 00:32:41,440
agent is. 
It doesn't matter how good the 

597
00:32:41,440 --> 00:32:46,040
model is. 
Have you been playing with the 

598
00:32:46,040 --> 00:32:48,040
different? 
Because basically what I 

599
00:32:48,040 --> 00:32:51,880
understand is, hey, you bring 
your own model and you can bring

600
00:32:51,880 --> 00:32:58,160
your own harness and then you 
have this meta harness above it 

601
00:32:58,720 --> 00:33:02,600
that is unopinionated. 
And it's like, cool, you want to

602
00:33:02,880 --> 00:33:05,440
come with your open claw over 
here and then your Hermes. 

603
00:33:05,440 --> 00:33:06,640
You want to come with your 
Hermes? 

604
00:33:06,640 --> 00:33:08,920
You want to toss an integration 
with a llama? 

605
00:33:09,160 --> 00:33:10,960
Have fun at it by all means, 
yeah? 

606
00:33:11,400 --> 00:33:15,720
How have you and how have you 
been seeing that working in 

607
00:33:15,720 --> 00:33:18,760
like? 
Is there a world where you're 

608
00:33:18,760 --> 00:33:22,640
recognizing that now? 
What is important is 

609
00:33:22,640 --> 00:33:27,240
understanding what each one of 
these is best at. 

610
00:33:28,400 --> 00:33:31,000
We built homogene in about two 
weeks by the way, to give you 

611
00:33:31,000 --> 00:33:33,120
some context. 
All right, again, pun intended. 

612
00:33:35,640 --> 00:33:39,280
We literally used homogene to 
help build homogene, OK, like so

613
00:33:39,640 --> 00:33:42,400
everything you see, we we 
basically used homogene itself 

614
00:33:42,400 --> 00:33:45,720
to basically build it, improve 
it the whole kin Kaboodle. 

615
00:33:46,520 --> 00:33:50,640
The reason I bring that up is 
because look we have a ton of 

616
00:33:50,640 --> 00:33:56,040
benchmarks right but before the 
two weeks we data bricks 

617
00:33:56,160 --> 00:33:59,640
engineering product we we were 
we were all suffering from the 

618
00:33:59,640 --> 00:34:02,480
same thing right We were 
suffering from like which wall 

619
00:34:02,480 --> 00:34:04,600
should we use this week? 
Should I use Kimi now? 

620
00:34:04,600 --> 00:34:08,560
Should I use right? 
Should I be looking more into 

621
00:34:08,560 --> 00:34:12,639
the open source ones right 
should like so if you like, if 

622
00:34:12,639 --> 00:34:14,159
you look at data bricks itself 
right, you'll notice that 

623
00:34:14,159 --> 00:34:16,520
there's a whole bunch of 
different models like that all 

624
00:34:16,520 --> 00:34:19,520
natively run within data bricks.
They're all governed that's 

625
00:34:19,520 --> 00:34:21,679
great. 
I love that fact. 

626
00:34:21,679 --> 00:34:25,320
But then exactly to your point, 
I don't know which one's better.

627
00:34:25,719 --> 00:34:28,239
And so, yeah, if you even go to 
the database website, we 

628
00:34:28,239 --> 00:34:30,560
actually have this thing where 
we update monthly, forget the 

629
00:34:30,560 --> 00:34:33,840
exact URL, but basically say 
like, here's the best ones based

630
00:34:33,840 --> 00:34:35,960
on what we see this month, 
right? 

631
00:34:35,960 --> 00:34:40,760
And then we'll obviously for 
documentation, for debate, for 

632
00:34:40,880 --> 00:34:43,120
financial, like whatever, we 
just, we actually write down all

633
00:34:43,120 --> 00:34:45,920
the steps because we're like, 
this is what we think in real 

634
00:34:45,920 --> 00:34:47,960
world scenarios. 
And you'll notice that the real 

635
00:34:47,960 --> 00:34:52,920
world scenarios often do not 
match benchmarks, right? 

636
00:34:53,320 --> 00:34:57,680
And so that's what got us to 
build Omni in the 1st place 

637
00:34:57,680 --> 00:35:00,600
because we were basically, we 
actually had a version of Omni 

638
00:35:00,600 --> 00:35:02,640
internally. 
We obviously can't share that 

639
00:35:02,640 --> 00:35:05,240
one. 
But the basically use that as 

640
00:35:05,240 --> 00:35:08,520
our starting point, mainly 
because we were recognizing that

641
00:35:09,160 --> 00:35:12,000
every single engineer needed the
ability to have that flexibility

642
00:35:12,000 --> 00:35:17,080
to switch, right? 
Because they realized that maybe

643
00:35:17,080 --> 00:35:20,440
that particular Co base that 
they were working on was better 

644
00:35:20,440 --> 00:35:26,680
with X while different team, 
same org, hell, same team, like 

645
00:35:26,680 --> 00:35:30,360
same division, everything 
realizing Oh no, but I'm doing 

646
00:35:30,600 --> 00:35:33,280
why? 
And it's better with this. 

647
00:35:34,200 --> 00:35:38,760
And so the idea of us trying to 
dictate you're supposed to only 

648
00:35:38,760 --> 00:35:41,920
use one like we were never, 
we're going to do that in the 

649
00:35:41,920 --> 00:35:43,520
1st place just because of the 
nature of data bricks. 

650
00:35:43,520 --> 00:35:49,280
But like it just the whole thing
fell on his face almost 

651
00:35:49,280 --> 00:35:51,320
immediately. 
And So what we did. 

652
00:35:51,920 --> 00:35:55,320
The model provider could be 
down, which we see all the time.

653
00:35:55,520 --> 00:35:58,520
Oh my gosh, yeah, yeah, yeah. 
Claude was just down yesterday. 

654
00:35:58,560 --> 00:36:00,560
So I was when I was just trying 
to run my stuff. 

655
00:36:00,560 --> 00:36:03,880
So fortunately, not a big deal. 
Flip it automatically flipped 

656
00:36:03,880 --> 00:36:05,720
over to codecs and all right, 
I'm back up and running. 

657
00:36:05,720 --> 00:36:08,280
Yippee, KA, barely, barely a 
blip of the whole thing, right. 

658
00:36:08,560 --> 00:36:13,120
But so ultimately it was more 
like, OK, do I have the 

659
00:36:13,120 --> 00:36:16,400
flexibility to change? 
Do I have the flexibility to, to

660
00:36:16,920 --> 00:36:19,720
all, but also not to change but 
control my costs? 

661
00:36:20,280 --> 00:36:23,080
So like with the knowledge that 
we actually have policies put 

662
00:36:23,080 --> 00:36:26,200
right in there, would you say OK
policies as simple as like, OK, 

663
00:36:26,200 --> 00:36:29,000
can I run this command? 
OK, that's one of them. 

664
00:36:29,320 --> 00:36:31,520
But like a like a shell command,
right? 

665
00:36:31,680 --> 00:36:33,960
For example, you and I were 
joking about how just starting 

666
00:36:33,960 --> 00:36:38,160
off, you had run a skill that 
basically shut off your redirect

667
00:36:38,160 --> 00:36:40,240
your browser, right? 
Well, guess what? 

668
00:36:40,480 --> 00:36:42,600
Like Aldrin has it built in 
where you have a policy that 

669
00:36:42,600 --> 00:36:44,360
says, no, you actually have to 
prove a such thing. 

670
00:36:44,360 --> 00:36:47,680
So it can't do that, right? 
But also even budgets, which is 

671
00:36:47,680 --> 00:36:50,520
like, OK, if this thing's going 
to cost you more than X dollars 

672
00:36:50,520 --> 00:36:53,160
or X cents or whatever it is, 
guess what you have to prove 

673
00:36:53,440 --> 00:36:55,240
before you're allowed running 
it, right? 

674
00:36:55,560 --> 00:36:58,480
So just things like that, which 
allow you have control over the 

675
00:36:58,480 --> 00:37:00,760
situation. 
And then the final little tidbit

676
00:37:00,760 --> 00:37:03,760
is that inside of all the 
conversations you and I've just 

677
00:37:03,760 --> 00:37:06,320
had, right? 
The one thing is that it's a, 

678
00:37:07,080 --> 00:37:11,280
I'm still very like the OR a 
developer centric, right? 

679
00:37:11,960 --> 00:37:14,880
But in order to be successful 
more times than not, what we 

680
00:37:14,880 --> 00:37:17,520
have to do is we have to work 
with your team. 

681
00:37:18,000 --> 00:37:19,880
You want to share your learnings
for your team. 

682
00:37:20,160 --> 00:37:23,040
And so that's the the coup de 
grace for on that, which is 

683
00:37:23,040 --> 00:37:26,600
like, OK, it's like a Google Doc
type interface where basically 

684
00:37:27,000 --> 00:37:31,160
now, hey, I ran this agent or 
had this Polly do this 

685
00:37:31,160 --> 00:37:34,880
particular debate. 
Why don't you take a look, you 

686
00:37:34,880 --> 00:37:38,880
know, team leader X. 
And then not only can they see 

687
00:37:39,000 --> 00:37:44,160
what's being generated, they can
then themselves run on that 

688
00:37:44,160 --> 00:37:46,800
session themselves because you 
shared it with them and they can

689
00:37:46,800 --> 00:37:49,560
run the session, say, wait, can 
you run this particular test? 

690
00:37:49,560 --> 00:37:51,760
Did you take an account of X? 
Did you take an account of like 

691
00:37:51,960 --> 00:37:55,000
the new updated data fusion or 
whatever else that you need to 

692
00:37:55,000 --> 00:37:55,800
do? 
Right. 

693
00:37:55,800 --> 00:37:59,120
And then basically now that 
process continues on. 

694
00:37:59,320 --> 00:38:01,560
So you're actually sharing the 
workload together so the teams 

695
00:38:01,560 --> 00:38:02,920
can actually work together 
really well. 

696
00:38:03,640 --> 00:38:05,440
Right. 
Yeah, that is so cool. 

697
00:38:05,560 --> 00:38:09,760
And actually funny thing that 
you were mentioning, another one

698
00:38:09,760 --> 00:38:12,760
of my little side projects, toy 
projects that I'm creating for 

699
00:38:12,760 --> 00:38:18,040
myself is a widget that goes up 
into the corner of your screen 

700
00:38:18,520 --> 00:38:21,280
and it just calculates how much 
money you've spent. 

701
00:38:21,760 --> 00:38:24,040
There you go. 
By day or by week. 

702
00:38:24,120 --> 00:38:28,520
And every time you hit another 
$100 increment, it goes cha 

703
00:38:28,520 --> 00:38:31,240
ching and confetti goes 
everywhere. 

704
00:38:31,320 --> 00:38:35,280
And this. 
Dude you should so do APR into 

705
00:38:35,280 --> 00:38:37,640
Omnagent to do that. 
Seriously because we already 

706
00:38:37,640 --> 00:38:42,120
have the basis for calculation. 
Do APR baby like I will use use 

707
00:38:42,280 --> 00:38:44,880
Omnagent to create the PR for 
you because that's all I've 

708
00:38:44,880 --> 00:38:47,600
done. 
Yeah, you know what else I did 

709
00:38:47,600 --> 00:38:50,880
though? 
I put a little GIF so all of a 

710
00:38:50,880 --> 00:38:55,440
sudden a gift pops up of like 
someone making it rain. 

711
00:38:56,120 --> 00:38:58,840
Make it so it's like 5 different
random gifts that you can get 

712
00:38:58,840 --> 00:39:02,560
and it's like making it rain. 
And so you recognize, and I'm 

713
00:39:02,560 --> 00:39:07,880
created this with because I was 
at Fin OPS X which is this whole

714
00:39:07,880 --> 00:39:11,760
conference around budgets and 
all they care about are 

715
00:39:11,760 --> 00:39:14,120
tokenomics. 
That's the whole thing now. 

716
00:39:14,120 --> 00:39:16,400
Nobody cares about dashboards 
for your clouds. 

717
00:39:16,560 --> 00:39:19,240
Oh God, yeah, yeah, 100. 
Percent, dude, we got that 

718
00:39:19,240 --> 00:39:21,120
figured out right. 
All they care about are 

719
00:39:21,120 --> 00:39:24,360
tokenomics and how people are 
spending their tokens. 

720
00:39:24,800 --> 00:39:28,720
And what I realized is that a 
lot of folks are creating 

721
00:39:28,720 --> 00:39:33,200
dashboards so that their CF OS 
or whoever in the finance 

722
00:39:33,200 --> 00:39:36,560
department can see how much the 
engineers themselves are 

723
00:39:36,560 --> 00:39:39,520
spending. 
But what I didn't see a lot of 

724
00:39:39,560 --> 00:39:43,760
is pushing that visibility onto 
the end users. 

725
00:39:44,080 --> 00:39:48,760
And I'm like, if I knew that I 
was spending 1000 bucks a day, I

726
00:39:48,760 --> 00:39:53,840
might be a little bit more 
conscientious about the cost. 

727
00:39:54,320 --> 00:39:59,240
So wonderful call out like so I 
still remember because I'm a 

728
00:39:59,240 --> 00:40:01,400
long time Microsoft guy, OK, 
before all this. 

729
00:40:01,400 --> 00:40:06,600
And then when we switched to the
cloud, the whole premise is that

730
00:40:06,600 --> 00:40:10,080
like this idea of CapEx to OpEx,
right, like we're changing cable

731
00:40:10,080 --> 00:40:13,360
expenditures topics. 
In that process of OpEx, it was 

732
00:40:13,360 --> 00:40:17,400
the into DevOps, meaning that 
the developers actually had to 

733
00:40:17,400 --> 00:40:19,680
understand how much money they 
were spending on the 

734
00:40:19,680 --> 00:40:20,880
infrastructure that they're 
building. 

735
00:40:21,840 --> 00:40:23,560
Right. 
So if there weren't like 

736
00:40:23,560 --> 00:40:27,400
needlessly running a small query
and spinning up a A50 node 

737
00:40:27,400 --> 00:40:31,560
Hadoop cluster when their laptop
could have solved the problem, 

738
00:40:31,560 --> 00:40:35,200
right? 
So that's exactly what's going 

739
00:40:35,200 --> 00:40:39,680
on right now, which is basically
that service based architecture,

740
00:40:39,680 --> 00:40:44,040
SOA, you know where we've 
switched from this idea that we 

741
00:40:44,040 --> 00:40:48,440
had to go ahead and originally 
it was 1 central command to 

742
00:40:48,440 --> 00:40:52,320
understand what the budget was. 
And now that got pushed down to 

743
00:40:52,320 --> 00:40:54,960
developers because the 
developers could actually then 

744
00:40:55,240 --> 00:40:58,880
care, understand and help 
control that token. 

745
00:40:58,880 --> 00:41:00,480
Novice is the exact same thing 
100. 

746
00:41:00,480 --> 00:41:04,560
Percent that's so true it's the 
same motion and you're just 

747
00:41:04,800 --> 00:41:08,720
we're reliving it, but it's a 
little bit different now because

748
00:41:08,720 --> 00:41:12,640
what is funny and I think one of
the hardest questions that I saw

749
00:41:12,640 --> 00:41:17,800
people struggling with at the 
event was how do you know what 

750
00:41:18,680 --> 00:41:22,120
token spend is valuable versus 
wasteful? 

751
00:41:22,600 --> 00:41:27,920
Yeah and so there's all kinds of
tricks that you can do on these 

752
00:41:27,920 --> 00:41:32,280
like these details and one of 
them being hey, choose a 

753
00:41:32,280 --> 00:41:34,680
different model. 
If it's not that big of or 

754
00:41:34,680 --> 00:41:38,680
complex of a problem, just use a
better model, use an open source

755
00:41:38,680 --> 00:41:41,400
model, use whatever. 
But still you have certain 

756
00:41:41,400 --> 00:41:48,840
engineers that are burning a ton
of tokens and for some sub set 

757
00:41:48,840 --> 00:41:52,760
of those engineers that burn is 
worth it. 

758
00:41:53,240 --> 00:41:58,200
And then for another subset of 
those engineers that burn is not

759
00:41:58,200 --> 00:42:00,480
worth it. 
And it's like what are they 

760
00:42:00,480 --> 00:42:01,480
getting done? 
No. 

761
00:42:01,640 --> 00:42:03,920
No, it's 100%. 
So this goes back to the idea 

762
00:42:03,920 --> 00:42:06,880
where it's actually just like 
with the switch from CapEx 

763
00:42:06,880 --> 00:42:09,200
topics, right? 
You actually need both. 

764
00:42:09,440 --> 00:42:12,680
You need to give the developers 
the tools to understand what 

765
00:42:12,680 --> 00:42:16,040
they're doing and what the spend
is, but you also need to have 

766
00:42:16,040 --> 00:42:19,840
some central governance story so
you can understand what's going 

767
00:42:19,840 --> 00:42:22,440
on right? 
Like so like you've, I've come 

768
00:42:22,440 --> 00:42:24,720
to you, we've actually had this 
conversations like why do we 

769
00:42:24,720 --> 00:42:28,640
have, you know, for example, uni
catalog and why did we database 

770
00:42:28,640 --> 00:42:31,960
introduce uni a gateway. 
That's the whole point. 

771
00:42:32,520 --> 00:42:35,160
It's two sides of the same coin.
Like homogen is there for the 

772
00:42:35,160 --> 00:42:37,920
developer so that they 
themselves understand they 

773
00:42:37,960 --> 00:42:41,200
themselves can apply policies. 
But then we have Uni a gateway 

774
00:42:41,200 --> 00:42:45,120
to to for centralized form, 
which says a look, yeah, you can

775
00:42:45,120 --> 00:42:50,560
use Opus 4748, whichever one 
you're doing 47, right? 

776
00:42:51,400 --> 00:42:55,400
But wait, like for example, if 
you were to look at my 

777
00:42:55,400 --> 00:42:57,640
questions, especially in terms 
of when I was still beginning 

778
00:42:57,640 --> 00:43:00,720
with Ongen, when I was testing 
it out, there will be an absurd 

779
00:43:00,720 --> 00:43:03,720
number of which Bagel is better,
Montreal versus New York, right.

780
00:43:04,160 --> 00:43:07,240
And so if I was a central 
person, they'd be like, no, this

781
00:43:07,240 --> 00:43:09,560
Denny guy, we we got. 
A token. 

782
00:43:10,720 --> 00:43:13,240
This is a ridiculous amount of 
tokens being spent on this 

783
00:43:13,240 --> 00:43:14,440
particular question. 
OK? 

784
00:43:14,800 --> 00:43:16,400
Especially because everybody 
knows Montreal bagels are 

785
00:43:16,400 --> 00:43:17,440
better. 
Yes, I see it. 

786
00:43:17,880 --> 00:43:20,720
All right, bye. 
But you get my drift. 

787
00:43:20,720 --> 00:43:23,280
Like, obviously there was a 
reason for what I did it, but if

788
00:43:23,280 --> 00:43:25,080
I still do it today, oh, my 
goodness. 

789
00:43:25,080 --> 00:43:27,080
Yeah. 
They would just lock into Uniaga

790
00:43:27,080 --> 00:43:30,120
anyway. 
No, we feel like no, you're not 

791
00:43:30,120 --> 00:43:32,600
allowed doing this. 
Shifting budget over. 

792
00:43:32,600 --> 00:43:33,880
Yeah, yeah. 
They're shifting budgets. 

793
00:43:33,880 --> 00:43:37,680
They're like, OK, like, no, 
daddy, you're allowed using pie.

794
00:43:37,800 --> 00:43:40,760
You're allowed using like you're
allowed. 

795
00:43:40,960 --> 00:43:44,760
Maybe if we're, if, if they're, 
if you're lucky, we'll let you 

796
00:43:44,760 --> 00:43:46,520
use Kimi just for the sake of 
doing it. 

797
00:43:46,520 --> 00:43:48,960
But not even a later version of 
Kimi, like an older version, 

798
00:43:49,200 --> 00:43:51,800
right? 
But like Opus, Oh hell no, we're

799
00:43:51,800 --> 00:43:53,360
never going to let you use that,
right? 

800
00:43:53,360 --> 00:43:56,640
So. 
You know what, on this idea of 

801
00:43:56,640 --> 00:44:00,040
choosing the models and choosing
the harnesses and all of that, 

802
00:44:00,040 --> 00:44:07,000
where I see the biggest friction
is when I have to manually make 

803
00:44:07,000 --> 00:44:12,360
the decision because I am a 
little bit lazy and I just kind 

804
00:44:12,360 --> 00:44:15,320
of go with whatever I've got on,
you know? 

805
00:44:15,400 --> 00:44:19,200
And so I'm not going to, 
especially if it's my company's 

806
00:44:19,200 --> 00:44:22,800
money in a way, I'm going to be 
like, whatever, let's see if it.

807
00:44:22,800 --> 00:44:24,360
Works. 
I can just blow the money and 

808
00:44:24,360 --> 00:44:25,560
ask the bagel. 
Question. 

809
00:44:25,600 --> 00:44:28,080
It's going to be like let's use 
the smallest model. 

810
00:44:28,080 --> 00:44:30,640
It doesn't work and then I'm 
going to go back and try and go 

811
00:44:30,640 --> 00:44:33,760
to the next smallest and 
cheapest and then it still 

812
00:44:33,760 --> 00:44:35,920
doesn't work. 
And then I've wasted time, which

813
00:44:35,920 --> 00:44:39,320
is the whole reason I thought I 
was using these coding agents 

814
00:44:39,320 --> 00:44:44,280
was to save myself time. 
And so it's like I see that I 

815
00:44:44,280 --> 00:44:50,280
will overcompensate with bigger 
models just for the fear of a 

816
00:44:50,320 --> 00:44:52,400
smaller model. 
It's probably not going to be 

817
00:44:52,400 --> 00:44:54,600
able to get this, even though it
totally will get it. 

818
00:44:54,840 --> 00:44:57,320
Like it's like, oh, you just 
changed some names in the code 

819
00:44:57,320 --> 00:45:01,600
base and it went through and you
know, like control FT, you did 

820
00:45:01,600 --> 00:45:04,960
not need a 4.7 for that. 
Right. 

821
00:45:04,960 --> 00:45:08,480
So I don't think that's 
completely solved yet because 

822
00:45:09,160 --> 00:45:12,240
this is sort of like the same 
idea like the journey from CapEx

823
00:45:12,240 --> 00:45:16,720
topics like we don't know when, 
when one should be applied where

824
00:45:16,720 --> 00:45:19,560
like like in other words, 
preventing you from using it in 

825
00:45:19,560 --> 00:45:20,800
the first place is very 
limiting. 

826
00:45:20,800 --> 00:45:23,480
We don't want to do that, but we
also don't want to get to a 

827
00:45:23,480 --> 00:45:29,240
point where you've blown 
1,000,000 tokens for which color

828
00:45:29,240 --> 00:45:36,600
of mauve is the best, OK, right.
And so we're at a point where 

829
00:45:36,920 --> 00:45:38,920
especially with meta harnesses 
and especially with things like 

830
00:45:38,920 --> 00:45:42,920
in the AI Gateway, that now we 
at least have visibility and 

831
00:45:43,120 --> 00:45:46,560
some form of control. 
One of the things that we've 

832
00:45:46,560 --> 00:45:48,840
been experimenting on, and 
you're going to see some of it 

833
00:45:48,840 --> 00:45:52,080
show up actually in object 
itself because we've been 

834
00:45:52,080 --> 00:45:54,680
experimenting internally first 
because we want to see how many 

835
00:45:54,680 --> 00:45:55,840
engineers are going to kill us, 
right? 

836
00:45:57,280 --> 00:46:02,640
Is the idea of where do we auto 
choose for you based on the 

837
00:46:02,640 --> 00:46:04,200
questions that's being asked 
right? 

838
00:46:04,360 --> 00:46:07,640
But that's itself is hard and 
that itself may require us to 

839
00:46:07,640 --> 00:46:10,920
use an LM to make the choice. 
So do we just shift the token 

840
00:46:10,920 --> 00:46:11,880
usage that way? 
Right? 

841
00:46:12,640 --> 00:46:16,680
But this is where like a good 
old fashioned machine learning 

842
00:46:16,680 --> 00:46:19,520
may kick in for us. 
You know, like you know how. 

843
00:46:19,520 --> 00:46:21,840
RL is back in vogue again. 
Which is it? 

844
00:46:21,840 --> 00:46:23,240
Never should have not been 
invoked, by the way. 

845
00:46:23,240 --> 00:46:25,400
But you get my trip, right? 
RL is back in vogue. 

846
00:46:25,400 --> 00:46:29,040
It's like, oh, that's cheaper, 
That's faster, that's better. 

847
00:46:29,960 --> 00:46:32,520
Yeah, we've been saying this 
pretty much the entire time 

848
00:46:32,520 --> 00:46:34,760
guys. 
But yeah, so goes back to the 

849
00:46:34,760 --> 00:46:35,960
same idea. 
Machine learning is back in 

850
00:46:35,960 --> 00:46:37,200
vocal. 
I'm like, no, a lot of these 

851
00:46:37,200 --> 00:46:39,640
types of questions though, like 
in terms of the complexity of 

852
00:46:39,640 --> 00:46:42,360
the question, you can literally 
use good old fashioned machine 

853
00:46:42,360 --> 00:46:44,040
learning to make the choices. 
So that's what I'm saying. 

854
00:46:44,040 --> 00:46:45,840
We can. 
We're definitely seeing trends 

855
00:46:45,840 --> 00:46:47,440
like that to help with that 
process. 

856
00:46:47,920 --> 00:46:52,120
And ultimately, however it's 
done, the key thing is to your 

857
00:46:52,120 --> 00:46:55,840
point is like is that for 
example, with homogene there 

858
00:46:55,840 --> 00:46:58,880
isn't changing the code base to 
choose. 

859
00:46:59,120 --> 00:47:03,600
It's literally click and it 
automatically chooses for and 

860
00:47:03,640 --> 00:47:06,720
you have the the drop down list 
of which one to go with in the 

861
00:47:06,720 --> 00:47:10,160
1st place for starters. 
And So what we're hoping for by 

862
00:47:10,160 --> 00:47:14,160
the time we're done with 
everything is that once we get 

863
00:47:14,160 --> 00:47:18,120
into some good patterns around 
that, we actually tell you 

864
00:47:18,120 --> 00:47:20,840
upfront we're going to switch to
Model X. 

865
00:47:21,240 --> 00:47:24,200
Here's the reason why and let it
run. 

866
00:47:25,080 --> 00:47:29,880
As long as it if it fails it 
will retry with a better model 

867
00:47:30,280 --> 00:47:33,320
then I'm stoked because I don't 
have to spend my time being like

868
00:47:33,760 --> 00:47:35,120
this. 
Doesn't really look that good. 

869
00:47:35,320 --> 00:47:37,400
Right, right, right, right. 
But then it goes back to like, 

870
00:47:37,560 --> 00:47:39,200
This is why are we like Polly so
much? 

871
00:47:39,200 --> 00:47:41,880
And and Debbie? 
Because there's different 

872
00:47:41,880 --> 00:47:45,040
versions of what it means to say
it doesn't work, if you get my 

873
00:47:45,080 --> 00:47:46,240
drift. 
Right. 

874
00:47:46,240 --> 00:47:50,000
So that's why we like the idea 
of the the agents debating each 

875
00:47:50,000 --> 00:47:54,680
other, because then you can walk
away, let the two agents fight 

876
00:47:54,680 --> 00:47:58,840
each other out. 
And basically, yeah, because 

877
00:47:58,840 --> 00:48:01,640
they have enough context, they 
have enough memory, they can 

878
00:48:01,640 --> 00:48:05,680
actually figure out, OK, there's
there's a reasonable assertion 

879
00:48:05,680 --> 00:48:08,960
here that actually because 
they're debating each other, you

880
00:48:08,960 --> 00:48:10,480
will get a result that actually 
works. 

881
00:48:10,560 --> 00:48:14,000
Or at least even if it doesn't 
work, at least now you have the 

882
00:48:14,000 --> 00:48:17,560
basis of why it doesn't work. 
So it left because in the end, 

883
00:48:17,720 --> 00:48:20,040
especially with how powerful are
these open source models are 

884
00:48:20,040 --> 00:48:22,120
these days? 
For me, a lot of time it's. 

885
00:48:22,280 --> 00:48:26,120
It's less about which model I'm 
using nowadays, and it's a lot 

886
00:48:26,120 --> 00:48:30,920
more about did I provide the 
data, the context, the 

887
00:48:30,920 --> 00:48:34,640
background, the configuration, 
the memory. 

888
00:48:34,640 --> 00:48:37,240
The big thing is, did I set up 
the correct loops? 

889
00:48:37,560 --> 00:48:42,400
Did I give it what it needed to 
just be able to run for an hour 

890
00:48:42,680 --> 00:48:44,280
so that I don't have to come 
back? 

891
00:48:44,320 --> 00:48:46,920
And that is one of them. 
I think that's a great point is 

892
00:48:47,280 --> 00:48:50,640
being able to debate and 
recognize what good looks like 

893
00:48:51,160 --> 00:48:55,840
so that you can trust, OK, it's 
a smaller, cheaper model, but I 

894
00:48:55,840 --> 00:49:01,080
know that the debate is going to
happen and if it doesn't pass 

895
00:49:01,080 --> 00:49:04,480
the sniff test, then it will go 
back and it will loop again. 

896
00:49:04,960 --> 00:49:07,400
Right. 
And then this goes and why do we

897
00:49:07,400 --> 00:49:11,440
see yet another resurgence of 
evals all over again for. 

898
00:49:11,440 --> 00:49:13,480
Precisely that we're still 
talking about evals. 

899
00:49:13,720 --> 00:49:17,440
We have to, because in the end 
there has to be some numerical, 

900
00:49:17,760 --> 00:49:21,720
There has to be a number that 
allows us to state if what was 

901
00:49:21,720 --> 00:49:25,480
created actually per evaluation 
actually is. 

902
00:49:25,880 --> 00:49:28,480
It's the old definition, what is
good, it goes back to that 

903
00:49:28,480 --> 00:49:29,840
definition over and over again, 
right? 

904
00:49:30,280 --> 00:49:33,680
And so in the end, like I said, 
it's, it's becoming more and 

905
00:49:33,680 --> 00:49:35,680
more of an obvious that it's 
less about the models. 

906
00:49:35,680 --> 00:49:37,120
This is not a knock on any of 
the models. 

907
00:49:37,160 --> 00:49:39,040
They're all really good. 
Now at the point that's what I'm

908
00:49:39,040 --> 00:49:41,840
just trying to get AT. 
And it's much more about, OK, 

909
00:49:42,120 --> 00:49:46,160
now that I know the models good,
what else is missing? 

910
00:49:46,200 --> 00:49:48,920
And then the evals will actually
help us figure that out. 

911
00:49:49,360 --> 00:49:54,040
Yeah, yeah, yeah, that. 
How can I make this harness top 

912
00:49:54,040 --> 00:49:58,560
notch exactly, 100%. 
I think the auto model choosing 

913
00:49:58,560 --> 00:50:02,840
is the way to go because if it 
is left on to the developer, 

914
00:50:03,440 --> 00:50:05,320
they're never going to choose a 
shitty model. 

915
00:50:05,440 --> 00:50:07,720
Or no, No, they're always going 
to choose the best model, yeah. 

916
00:50:08,680 --> 00:50:13,120
It's a psychological thing. 
You're like, I need to do this 

917
00:50:13,120 --> 00:50:18,360
because what if it encounter 
something in this run that it 

918
00:50:18,360 --> 00:50:20,960
needs the power? 
Let's just let's do it just in 

919
00:50:20,960 --> 00:50:24,520
case it's a few extra cents or 
dollars. 

920
00:50:24,680 --> 00:50:27,000
Depending on what, yeah, I mean,
if a few extra cents, we're 

921
00:50:27,000 --> 00:50:29,520
probably OK with that. 
It's the few extra dollars $100.

922
00:50:30,120 --> 00:50:31,840
And unequivocally, no, you're 
not wrong. 

923
00:50:31,840 --> 00:50:36,400
You're absolutely right. 
And I think like this is part of

924
00:50:36,400 --> 00:50:38,560
that story for us. 
Like it's, that's part of the 

925
00:50:38,560 --> 00:50:41,920
story why we made it open source
because we're going, look, we're

926
00:50:41,920 --> 00:50:43,200
not pretending we have all the 
answers. 

927
00:50:43,200 --> 00:50:45,280
We're just pretending. 
We're just stating the fact that

928
00:50:45,560 --> 00:50:47,600
this is the problem we faced 
internally. 

929
00:50:48,160 --> 00:50:49,920
We want to share this with the 
world. 

930
00:50:49,920 --> 00:50:54,920
So that way y'all can like we 
can all work together to get to 

931
00:50:54,920 --> 00:50:56,680
that solution that you're 
talking about, which is like, 

932
00:50:56,960 --> 00:50:59,960
how do we do a proper auto 
model, right? 

933
00:51:00,440 --> 00:51:02,280
How do we, how do we do a proper
thing? 

934
00:51:02,440 --> 00:51:05,760
So that is absolutely 100% 
vendor agnostic. 

935
00:51:06,920 --> 00:51:10,240
How you know what I kind of 
question with because certain 

936
00:51:10,240 --> 00:51:14,360
harnesses have certain ways of 
expressing their harness, right?

937
00:51:14,360 --> 00:51:18,480
And I'm thinking specifically 
like in Hermes and Open Claw, 

938
00:51:18,480 --> 00:51:22,720
you have the soul, but then in 
Cloud code you don't have that. 

939
00:51:23,240 --> 00:51:28,600
How does this meta harness of 
Omni agent act when you're 

940
00:51:28,600 --> 00:51:32,480
running these different shapes? 
Of harnesses. 

941
00:51:32,480 --> 00:51:36,560
So the the TLDR is the root you,
you ruined the DS Π project 

942
00:51:36,840 --> 00:51:38,280
right? 
Yeah, right. 

943
00:51:38,280 --> 00:51:41,200
The root of a lot of our 
learnings is basic from DSPI in 

944
00:51:41,200 --> 00:51:44,280
the 1st place, which is the idea
of how do you, how do I go ahead

945
00:51:44,280 --> 00:51:48,160
and allow myself to functionally
run these queries against all 

946
00:51:48,160 --> 00:51:50,760
these different models that 
actually instinct that actually 

947
00:51:50,760 --> 00:51:53,120
have different ways of running 
in the 1st place, right? 

948
00:51:53,200 --> 00:51:55,080
Different parameters, different 
weights, different everything, 

949
00:51:55,080 --> 00:52:01,040
right. 
And so with that, the TLDR is we

950
00:52:01,040 --> 00:52:04,920
started looking at the systems 
that we're integrating with and 

951
00:52:04,920 --> 00:52:09,560
we tried our best to basically 
take an account of each one. 

952
00:52:10,560 --> 00:52:13,680
And so that way any integration 
we have basically takes an 

953
00:52:13,680 --> 00:52:17,200
account of like what each one is
strong at at the code level, 

954
00:52:17,560 --> 00:52:21,760
right at the conflict level. 
But I would be remiss with if I 

955
00:52:21,760 --> 00:52:23,160
didn't also basically mention 
the facts. 

956
00:52:23,160 --> 00:52:26,080
Like that's actually why we're 
reaching out to a lot of these, 

957
00:52:26,240 --> 00:52:28,760
both the model makers and also 
the harness of makers, because 

958
00:52:28,760 --> 00:52:30,920
we're going to look, we're not 
pretending we got it right. 

959
00:52:30,920 --> 00:52:32,600
We're pretending we got a good 
start. 

960
00:52:32,640 --> 00:52:35,720
That's all we're saying, right? 
We'd love to work with you to 

961
00:52:35,720 --> 00:52:41,480
make sure we have configured it 
the best to represent, you know,

962
00:52:41,600 --> 00:52:44,360
your particular model, your 
particular harness the best. 

963
00:52:44,720 --> 00:52:47,200
Because, yeah, we're, we we're 
not pretending we understand at 

964
00:52:47,200 --> 00:52:49,840
all. 
We're just simply saying that as

965
00:52:49,840 --> 00:52:51,600
you can probably guess, we've 
used a lot of them. 

966
00:52:52,280 --> 00:52:54,880
Yeah. 
So we we've incorporated as much

967
00:52:54,880 --> 00:52:56,480
of that knowledge as we possibly
can. 

968
00:52:56,480 --> 00:52:58,720
But that still doesn't replace 
the fact that there are people 

969
00:52:58,920 --> 00:53:01,480
in these other who crave the 
harnesses themselves or crave 

970
00:53:01,480 --> 00:53:03,600
the models themselves. 
They would know better than us. 

971
00:53:03,920 --> 00:53:06,280
Fair. 
And again, why is an open source

972
00:53:06,480 --> 00:53:08,880
project? 
So you get to look at the code, 

973
00:53:08,960 --> 00:53:12,040
tell us what we screwed up on. 
We want to go ahead and work 

974
00:53:12,040 --> 00:53:13,440
with you to figure out how to 
make it better. 

975
00:53:13,680 --> 00:53:17,400
Simple as that. 
You know what gives me this like

976
00:53:17,400 --> 00:53:23,480
visual in my head of almost like
you have your different 

977
00:53:23,480 --> 00:53:26,640
harnesses. 
And so the outer harness we 

978
00:53:26,640 --> 00:53:30,120
could consider like the MCP 
servers that I set up, maybe I 

979
00:53:30,120 --> 00:53:32,560
do memory in a certain way, 
etcetera, etcetera. 

980
00:53:33,320 --> 00:53:38,880
And then you've got the meta 
harness, which is the it's 

981
00:53:39,000 --> 00:53:43,400
outside of all these outer 
harnesses. 

982
00:53:43,480 --> 00:53:48,600
And that is where you have this 
way to almost like be an 

983
00:53:48,640 --> 00:53:55,040
overlord of all your agents. 
So like if you're looking at a a

984
00:53:55,160 --> 00:53:57,600
circular thing like the inner 
harness and the outer. 

985
00:53:57,760 --> 00:53:59,160
Right. 
Basically the outside layer of 

986
00:53:59,160 --> 00:54:00,200
the onion. 
Yeah, yeah, yeah. 

987
00:54:00,360 --> 00:54:03,160
Yeah, exactly. 
It's it's the onion or it's the 

988
00:54:03,160 --> 00:54:06,200
earth with the different like 
rocks. 

989
00:54:06,440 --> 00:54:10,560
That core mantle shelf. 
Yeah, the magma, all of that. 

990
00:54:10,560 --> 00:54:13,360
So that's kind of how I I 
visualize it. 

991
00:54:13,360 --> 00:54:18,800
Is that the way that you would? 
Is that a good way of thinking 

992
00:54:18,800 --> 00:54:19,280
of it? 
Oh. 

993
00:54:19,440 --> 00:54:23,240
Absolutely. 
I mean, look, everything we do 

994
00:54:25,280 --> 00:54:29,520
business wise, development wise,
whatever else, we're always 

995
00:54:29,520 --> 00:54:32,640
trying to figure out ways to 
abstract it, right? 

996
00:54:32,680 --> 00:54:35,240
The reason for the abstraction 
is because it gives us 

997
00:54:35,680 --> 00:54:37,600
flexibility. 
It gives U.S. 

998
00:54:37,600 --> 00:54:42,720
Open standards, it gives us the 
ability to go ahead and like 

999
00:54:43,320 --> 00:54:45,440
simplify said business logic, 
right? 

1000
00:54:45,880 --> 00:54:50,920
So that way you're able to focus
on the real problem, not on it 

1001
00:54:50,920 --> 00:54:52,360
was a little problem. 
I mean, if if you and I were 

1002
00:54:52,360 --> 00:54:56,200
having discussion five years ago
when ChatGPT just came out 

1003
00:54:56,520 --> 00:55:02,000
right, you and I would have 
definitely been debating on 

1004
00:55:02,240 --> 00:55:04,480
things like what were the 
weights? 

1005
00:55:05,360 --> 00:55:06,720
What were the hyperparameters 
used? 

1006
00:55:07,360 --> 00:55:09,520
What's the? 
Prompt right, right. 

1007
00:55:10,280 --> 00:55:11,800
Are we still having those 
debates now? 

1008
00:55:12,440 --> 00:55:15,160
Some damn good AI researchers, 
of course they are, because 

1009
00:55:15,160 --> 00:55:16,760
they're trying to improve the 
models, right. 

1010
00:55:17,080 --> 00:55:21,960
But you and I, as users of these
systems, we're not having that 

1011
00:55:21,960 --> 00:55:23,880
debate anymore. 
We were barely qualified to have

1012
00:55:23,880 --> 00:55:25,960
that debate then. 
We're certainly not qualified to

1013
00:55:25,960 --> 00:55:28,960
have that debate now, right? 
And so, but what, what can we 

1014
00:55:28,960 --> 00:55:31,000
do? 
We can talk about like how to 

1015
00:55:31,080 --> 00:55:36,040
effectively use agents, right? 
Well, remember what I just said,

1016
00:55:36,040 --> 00:55:38,840
effectively use agents, not 
effectively use LMS, right? 

1017
00:55:38,840 --> 00:55:41,800
Because we've already accepted 
the fact that like whether it 

1018
00:55:41,800 --> 00:55:44,520
was called compound AI systems 
before or whatever else 

1019
00:55:44,520 --> 00:55:47,640
afterwards, the reality is like 
the agent is basically not just 

1020
00:55:47,640 --> 00:55:50,440
the LM, it's actually a lot of 
the things that are around it as

1021
00:55:50,440 --> 00:55:52,640
well, right? 
And then us saying harnesses, 

1022
00:55:52,680 --> 00:55:55,320
yes, exactly to your point, like
whether it's an MCP server, it's

1023
00:55:55,360 --> 00:55:59,520
ACLI, it's reading adjacent from
somewhere and XML from 

1024
00:55:59,520 --> 00:56:02,720
somewhere, hitting a database, 
whatever, whatever it's doing 

1025
00:56:02,720 --> 00:56:05,920
doesn't matter, right? 
The point is that we're trying 

1026
00:56:05,920 --> 00:56:09,520
to figure out how to make these 
things portable and stateful. 

1027
00:56:09,840 --> 00:56:12,480
Well, guess what? 
That is an old fashioned 

1028
00:56:13,120 --> 00:56:17,560
business database problem that 
we've been doing since the 60s, 

1029
00:56:18,200 --> 00:56:22,120
right? 
Like, why are databases back? 

1030
00:56:22,160 --> 00:56:24,320
Like why? 
Why do we talk about like 

1031
00:56:24,320 --> 00:56:26,440
Serverless Postcrest? 
Why did Databricks talk about 

1032
00:56:26,440 --> 00:56:27,920
Lake Base and Neon? 
Why? 

1033
00:56:28,160 --> 00:56:33,080
Because literally we have to 
have cheap stateful memory for 

1034
00:56:33,080 --> 00:56:35,800
these agents to run, right? 
As simple as that, right? 

1035
00:56:35,800 --> 00:56:38,240
Well then if I have cheap 
stateful, then I'm like, OK, 

1036
00:56:38,760 --> 00:56:41,560
what's stateful? 
The best thing is a database. 

1037
00:56:41,720 --> 00:56:43,360
Cool. 
Why is poster sequel? 

1038
00:56:43,360 --> 00:56:45,880
Why is there resurgence? 
I would argue there never was a 

1039
00:56:45,880 --> 00:56:48,080
drop in the 1st place. 
That's what I would argue 

1040
00:56:48,080 --> 00:56:50,600
actually. 
But irrelevant of like you want 

1041
00:56:50,600 --> 00:56:53,880
to call resurgence or not, the 
premise is that I need something

1042
00:56:53,880 --> 00:56:55,520
that's trustworthy, that can 
keep state. 

1043
00:56:55,640 --> 00:56:58,120
That's it, that, that. 
And so that's never left us. 

1044
00:56:58,840 --> 00:57:01,520
And I mean, like, remember I was
like former sequel server. 

1045
00:57:01,520 --> 00:57:03,920
So like I'm like, yeah, yeah, 
like the database is still 

1046
00:57:03,920 --> 00:57:05,800
there. 
It's no matter what we've done, 

1047
00:57:05,800 --> 00:57:09,080
no matter how much we've talked 
about lake houses, which 

1048
00:57:09,080 --> 00:57:11,920
obviously I'm a big fan of, and 
we're talking about agents, 

1049
00:57:12,080 --> 00:57:16,160
which is the to the shore these 
days, we still need something 

1050
00:57:16,160 --> 00:57:18,880
called state. 
Well, that state is databases. 

1051
00:57:18,880 --> 00:57:21,120
And then guess what? 
That's if you can make it 

1052
00:57:21,120 --> 00:57:25,840
cheaper, you're more successful.
Well, that's how these agents 

1053
00:57:25,840 --> 00:57:27,560
have run. 
That's how all these systems 

1054
00:57:27,560 --> 00:57:31,640
work where we're just extracting
the way and then storing what 

1055
00:57:31,640 --> 00:57:35,520
needs to be stored at IE state. 
So that way you can be 

1056
00:57:35,520 --> 00:57:38,200
effective. 
And so and everything when it 

1057
00:57:38,200 --> 00:57:44,440
comes to like how we started 
with databases before B i.e., TL

1058
00:57:44,440 --> 00:57:48,320
pipelines, it's the same action 
plan, right? 

1059
00:57:48,320 --> 00:57:52,280
With agents right now, which is 
like, how do I modularize this? 

1060
00:57:53,320 --> 00:57:57,520
How do I make it portable? 
And then for a bunch of us, how 

1061
00:57:57,520 --> 00:58:02,040
do I make it open so that way 
it's easy for you to move from 

1062
00:58:02,040 --> 00:58:02,680
one to the other?