1
00:00:00,240 --> 00:00:04,160
Even if the technology again 
plateaus, even with current 

2
00:00:04,160 --> 00:00:07,920
model shape, the way we're going
to walk in 10 years is probably 

3
00:00:07,920 --> 00:00:09,600
nothing compared to the way we 
work today. 

4
00:00:09,600 --> 00:00:12,760
And so there's many, many, many 
opportunities to build for for 

5
00:00:12,760 --> 00:00:21,640
everyone we. 
Should probably start with this 

6
00:00:21,640 --> 00:00:23,880
multiplayer versus single 
player. 

7
00:00:23,880 --> 00:00:28,320
AII want to mention to folks 
that you've got a decorated 

8
00:00:28,320 --> 00:00:32,600
background, you're at Stripe in 
the early days, you then you're 

9
00:00:32,600 --> 00:00:36,800
also at Open AI and now you're 
doing your own thing with dust. 

10
00:00:36,800 --> 00:00:39,560
So it's super cool to see this 
progression. 

11
00:00:40,040 --> 00:00:43,640
Are there things that in those 
early days of Stripe scaling up 

12
00:00:43,920 --> 00:00:49,760
or in the early days of Open AI 
scaling up that you have taken 

13
00:00:49,760 --> 00:00:54,040
away from that and you want to 
infuse into dust as you're 

14
00:00:54,040 --> 00:00:56,560
scaling up? 
Oh yeah, a ton. 

15
00:00:56,800 --> 00:00:59,320
Mostly Stripe over open eye 
first. 

16
00:00:59,680 --> 00:01:01,120
Yeah, I've, I've been very 
lucky. 

17
00:01:01,240 --> 00:01:04,959
I've been very lucky. 
So we joined Stripe through an 

18
00:01:04,959 --> 00:01:07,600
acquisition actually with my 
current Co founders. 

19
00:01:07,600 --> 00:01:09,040
It's not the first time we've 
started a company. 

20
00:01:10,000 --> 00:01:13,280
The big difference that it was 
15 years ago and we, we were 

21
00:01:13,280 --> 00:01:16,440
young, we had absolutely no clue
about what we were doing, but it

22
00:01:16,440 --> 00:01:19,080
was fun. 
But if we can chat about that 

23
00:01:19,080 --> 00:01:20,720
later, it's a whole subject of 
its own. 

24
00:01:21,560 --> 00:01:24,760
Stripe, we saw the 150 people to
3000. 

25
00:01:24,880 --> 00:01:30,120
It was absolutely incredible. 
And I think I owe very dear to 

26
00:01:30,120 --> 00:01:34,560
my heart the 150 to 500 people. 
I think it was pretty magical at

27
00:01:34,560 --> 00:01:40,520
Stripe because there was really 
very minimal management and it 

28
00:01:40,520 --> 00:01:43,680
still was working really well. 
And I'm trying, I've, I've been 

29
00:01:43,680 --> 00:01:46,760
trying since then to understand 
what was happening at the time 

30
00:01:46,760 --> 00:01:48,760
for that to work so well, do you
think? 

31
00:01:48,800 --> 00:01:53,040
That's the writing culture. 
So, yes, so I've got a, I've got

32
00:01:53,080 --> 00:01:57,040
a bit of a theory about that 
which we we actually turn into 

33
00:01:57,040 --> 00:01:59,840
operating principle at dust. 
But my my big theory about the 

34
00:01:59,840 --> 00:02:02,320
time it's slightly bullshit, but
I think it's interesting 

35
00:02:02,320 --> 00:02:05,000
nonetheless. 
My theory is that we were 

36
00:02:05,000 --> 00:02:07,480
actually implementing A flocking
algorithm. 

37
00:02:07,480 --> 00:02:09,479
So you know what the flocking 
algorithm is. 

38
00:02:09,639 --> 00:02:13,320
It's like when you so when you 
want to program a flock of birds

39
00:02:13,320 --> 00:02:17,160
or flock of whatnots in 
computers, there are kind of a 

40
00:02:17,200 --> 00:02:22,120
autonomous groups of things that
auto organize around obstacles. 

41
00:02:22,640 --> 00:02:24,960
There's actually only three 
things to there's only three 

42
00:02:24,960 --> 00:02:29,040
rules to make it work. 
It's local inspiration, so small

43
00:02:29,040 --> 00:02:35,760
files that push the the people 
away, distant attraction and 

44
00:02:35,760 --> 00:02:40,800
alignment and that gives you 
kind of some form of autonomous 

45
00:02:40,800 --> 00:02:44,160
behavior around obstacles. 
You can, you can look at 

46
00:02:44,160 --> 00:02:46,960
flocking algorithm online and 
you'll see the small videos. 

47
00:02:47,600 --> 00:02:50,760
And I think at Scribe, the 
writing culture was the distant 

48
00:02:50,760 --> 00:02:54,480
attraction. 
So Scribe was very early into 

49
00:02:54,480 --> 00:02:57,840
implementing that kind of a open
writing culture. 

50
00:02:57,880 --> 00:03:00,560
It was mostly through e-mail at 
the time, interestingly. 

51
00:03:01,120 --> 00:03:03,840
And what it was great. 
It was great. 

52
00:03:04,160 --> 00:03:07,640
Oh yeah. 
I don't like slag that much 

53
00:03:07,640 --> 00:03:10,440
personally, and I love e-mail, 
but that's another take that we 

54
00:03:10,440 --> 00:03:13,960
can spend a lot of time on. 
It's surprising me here. 

55
00:03:14,000 --> 00:03:16,880
All right. 
But anyway, so distant traction.

56
00:03:16,880 --> 00:03:18,800
Why? 
It's because we had those 

57
00:03:18,800 --> 00:03:21,120
mailing lists. 
And so there were people within 

58
00:03:21,120 --> 00:03:23,560
the org that were spending a lot
of time kind of sifting through 

59
00:03:23,560 --> 00:03:25,360
those main lists. 
And so that means that somebody 

60
00:03:25,360 --> 00:03:29,680
would see someone else talk 
about subject A over there and 

61
00:03:29,680 --> 00:03:32,440
see another person talk about 
the same subject over there in 

62
00:03:32,440 --> 00:03:34,600
the org and was able to tell 
them, hey, you're talking about 

63
00:03:34,600 --> 00:03:36,880
the same thing, you should chat.
And so that's kind of a distant 

64
00:03:36,880 --> 00:03:38,360
traction. 
False. 

65
00:03:39,200 --> 00:03:41,320
I think at Scribe, one thing 
that really struck me is the 

66
00:03:41,320 --> 00:03:46,080
amount of trusts between people 
at the time, it was just 

67
00:03:46,080 --> 00:03:48,040
ingrained in the culture in a 
very nice way. 

68
00:03:48,280 --> 00:03:51,280
And that's kind of a local 
separation force because you 

69
00:03:51,280 --> 00:03:53,760
don't look over the shoulder. 
If somebody says I'm I've got 

70
00:03:53,760 --> 00:03:56,440
this, they've got this and you 
stop caring about that. 

71
00:03:57,320 --> 00:04:00,240
And then finally, the last one 
was which is, which is almost 

72
00:04:00,240 --> 00:04:02,480
the, the the most bullshit one 
is alignment. 

73
00:04:02,480 --> 00:04:04,800
But I think at Scribe were we 
were and which is very 

74
00:04:04,800 --> 00:04:07,160
interesting because it's very 
hard to reproduce that in modern

75
00:04:07,160 --> 00:04:11,240
times with AI, but we had a very
strong alignment on the vision. 

76
00:04:11,240 --> 00:04:16,800
It was very easy to understand. 
We are a dev API for payments. 

77
00:04:17,160 --> 00:04:20,480
Basically there's a tensor of 
countries, payment methods, 

78
00:04:20,480 --> 00:04:23,880
payout methods, and you just 
just want to to fill it in. 

79
00:04:24,320 --> 00:04:26,480
It was very easy to understand. 
And so that means that you are 

80
00:04:26,480 --> 00:04:28,240
every is kind of aligned very 
naturally. 

81
00:04:28,920 --> 00:04:31,200
And so that gives you the three 
properties of looking, which 

82
00:04:31,200 --> 00:04:36,040
gives you kind of a autonomous 
behaviour in a sense, emergent 

83
00:04:36,040 --> 00:04:37,800
autonomous behaviour, which 
explains why. 

84
00:04:38,480 --> 00:04:40,640
But then surely we didn't need 
that that much management and it

85
00:04:40,680 --> 00:04:43,520
was so nice to work in that 
environment. 

86
00:04:44,280 --> 00:04:48,000
It feels like we could draw a 
parallel here and heads up 

87
00:04:48,000 --> 00:04:53,040
Segway into multiplayer AI, 
single player versus multiplayer

88
00:04:53,240 --> 00:04:56,600
with the Flock algorithm. 
And how have you thought about 

89
00:04:56,960 --> 00:05:01,360
humans working with agents in 
that Flock style? 

90
00:05:01,360 --> 00:05:03,240
Do you think that there's 
anything that you could look at 

91
00:05:03,240 --> 00:05:05,520
there? 
Yeah, it's probably a ton of 

92
00:05:05,520 --> 00:05:08,560
stuff, but it's very hard to 
anticipate because today we are 

93
00:05:08,560 --> 00:05:11,040
still very much stuck in the 
single player mode and. 

94
00:05:11,280 --> 00:05:14,200
What do you mean by that single 
player mode? 

95
00:05:14,440 --> 00:05:17,960
It's mostly that most of the 
agents interactions we have, we 

96
00:05:17,960 --> 00:05:21,480
just have them alone for two 
reasons. 

97
00:05:21,480 --> 00:05:26,720
I think the main 2 reasons I see
is first, the timer reason of 

98
00:05:26,720 --> 00:05:33,480
the tasks an agent can take is 
still pretty small according to 

99
00:05:33,480 --> 00:05:38,280
the meter benchmark we are 
topping at right now at a couple

100
00:05:38,280 --> 00:05:42,480
hours, let's say 1/2 a day and 
mostly in coding tasks by the 

101
00:05:42,480 --> 00:05:44,680
way. 
And you know half a day is still

102
00:05:44,680 --> 00:05:49,240
very much a single player thing.
You I can walk half a day on the

103
00:05:49,240 --> 00:05:50,960
task. 
You can walk half today on this 

104
00:05:50,960 --> 00:05:54,160
task and you walk alone and it's
fine because it's a kind of well

105
00:05:54,160 --> 00:05:57,000
shaped tasks. 
The other thing is that the the 

106
00:05:57,200 --> 00:05:59,400
they're still very imperfect in 
many ways. 

107
00:05:59,800 --> 00:06:02,120
And so you interact a lot with 
them. 

108
00:06:02,120 --> 00:06:05,800
It's messy really. 
When you look at the an agent 

109
00:06:05,800 --> 00:06:09,400
trace within a human and an 
agent, you look at the all the 

110
00:06:09,400 --> 00:06:12,880
sausages made and you're like, 
oh man, you did something funky 

111
00:06:12,880 --> 00:06:16,360
there. 
Or it's like, oh, you were so 

112
00:06:16,360 --> 00:06:20,120
close dude. 
A little bit over to the right 

113
00:06:20,120 --> 00:06:21,320
or left you would have been 
good. 

114
00:06:22,040 --> 00:06:24,760
And so on those time horizon 
and, and in terms of 

115
00:06:24,760 --> 00:06:28,400
capabilities, the models are 
super jagged, which means that 

116
00:06:28,800 --> 00:06:31,320
you on that you're looking in 
that direction. 

117
00:06:31,320 --> 00:06:37,400
The thing is clearly superhuman.
You move a few degrees off on a 

118
00:06:37,400 --> 00:06:40,080
slightly different task, you 
fall into a valley and the, the,

119
00:06:40,280 --> 00:06:43,160
the thing, that thing is the 
dumbest, dumbest thing on Earth.

120
00:06:43,760 --> 00:06:47,480
And so that's why it's messy. 
And so all of that makes it very

121
00:06:47,480 --> 00:06:51,160
kind of unnatural to collaborate
in an agent loop with multiple 

122
00:06:51,160 --> 00:06:55,040
humans to share your sessions 
with agents because they are 

123
00:06:55,040 --> 00:06:56,800
messy. 
You share the outputs, you don't

124
00:06:56,800 --> 00:07:00,320
share the session. 
And so that's that's why I think

125
00:07:00,320 --> 00:07:03,960
we're still very much single 
player AI in the way we we walk.

126
00:07:03,960 --> 00:07:06,640
We all spend a lot of day in, in
codecode or codecs. 

127
00:07:06,640 --> 00:07:10,520
We all spend a lot of day in our
favorite agent harness to walk 

128
00:07:10,520 --> 00:07:14,120
on presentation and stuff. 
And it's we pretty much do it 

129
00:07:15,120 --> 00:07:19,160
alone and we share in the in the
other platforms the I put and we

130
00:07:19,160 --> 00:07:21,920
collaborate there be it on 
GitHub, be it on Notion on 

131
00:07:21,920 --> 00:07:25,520
Google Drive, what not. 
I think the thing that the the 

132
00:07:25,520 --> 00:07:28,400
the thing that is supposed to 
is, is, is about to switch is 

133
00:07:28,400 --> 00:07:31,440
probably as the models are 
getting better and they seems to

134
00:07:31,440 --> 00:07:33,880
be getting better consistently. 
It's always an open question, 

135
00:07:33,880 --> 00:07:38,720
but when you look at the past 
few months, well, the the time 

136
00:07:38,720 --> 00:07:40,240
horizon of those tasks was 
increased. 

137
00:07:40,240 --> 00:07:43,760
And if you start considering a 
task like that represents a week

138
00:07:43,920 --> 00:07:51,560
of man walk for most job 
descriptions, those tasks are 

139
00:07:51,600 --> 00:07:55,120
inherently cross team, cross 
people involving multiple 

140
00:07:55,120 --> 00:07:56,560
person. 
Even if we if you think of 

141
00:07:56,760 --> 00:08:00,760
coding, there is very few coding
tasks that takes a week that 

142
00:08:00,760 --> 00:08:03,840
doesn't involve collaboration 
even with somebody external, be 

143
00:08:03,840 --> 00:08:08,160
it a product manager, be it a 
designer or maybe the infra 

144
00:08:08,160 --> 00:08:10,360
person because you have 
something finicky about the 

145
00:08:10,360 --> 00:08:14,280
infra and so. 
And if you look at something 

146
00:08:14,280 --> 00:08:17,160
that is outside of coding, then 
most tasks that takes a week, 

147
00:08:17,160 --> 00:08:18,880
they are extremely cross 
functional. 

148
00:08:19,600 --> 00:08:24,160
And so that will flip the 
picture where it's not going to 

149
00:08:24,160 --> 00:08:28,360
be possible anymore for just one
human to steer the agent in 

150
00:08:28,360 --> 00:08:31,560
those type of tasks, the agent, 
even the relationship between 

151
00:08:31,560 --> 00:08:33,480
the human and the agent will, 
will flip. 

152
00:08:33,720 --> 00:08:36,880
The agent will orchestrate a 
longer horizon task involving 

153
00:08:36,880 --> 00:08:38,400
humans and the humans will 
steer. 

154
00:08:38,400 --> 00:08:41,120
But they'll, they'll, I'm quite 
convinced there's going to be a 

155
00:08:41,120 --> 00:08:44,720
need for many humans to steer 
those agents beat an agent or 

156
00:08:44,720 --> 00:08:47,280
multiple agents. 
We don't really care for those 

157
00:08:47,280 --> 00:08:50,800
kind of long horizon tasks. 
They'll invoke the human agent 

158
00:08:51,040 --> 00:08:56,240
or the human tool when they need
to, so it's invoking A SRA tool.

159
00:08:57,960 --> 00:09:00,080
For that, there's going to be, 
there's going to be some of that

160
00:09:00,280 --> 00:09:01,400
for sure. 
Yeah. 

161
00:09:01,760 --> 00:09:06,600
And now don't you feel like this
is just an, it's a UX problem 

162
00:09:06,600 --> 00:09:09,840
more than anything? 
Because when you tell me about 

163
00:09:09,840 --> 00:09:15,720
the single player mode, I think,
OK, but what ways do I have to 

164
00:09:15,800 --> 00:09:22,400
share with my colleagues rather 
than exporting the chat or 

165
00:09:22,640 --> 00:09:26,400
putting it up with my PR? 
Things like that. 

166
00:09:26,560 --> 00:09:32,520
It doesn't really feel like I 
have the tools in order to get 

167
00:09:32,920 --> 00:09:36,240
to bring someone else along this
journey with me. 

168
00:09:37,040 --> 00:09:40,120
Yeah, I think, I think the, the,
the interaction for pretty much 

169
00:09:40,120 --> 00:09:43,000
remains kind of a, a you with 
the agent. 

170
00:09:43,000 --> 00:09:45,400
But it's just that the agents 
will create many of those 

171
00:09:45,400 --> 00:09:49,760
sessions in the to to achieve 
that task and have will kind of 

172
00:09:49,920 --> 00:09:55,400
collaborate with many humans. 
And so it is entirely a UX is 

173
00:09:55,400 --> 00:10:00,240
maybe slightly diminishing, but 
I do think it's not a, it's not 

174
00:10:00,240 --> 00:10:04,320
a, a machine learning problem. 
It's more of a product problem. 

175
00:10:05,800 --> 00:10:10,600
It means it just means that as 
an example, if we if we take Co 

176
00:10:10,600 --> 00:10:15,080
work, which is having kind of a 
real moments these days because 

177
00:10:15,080 --> 00:10:17,720
it's working locally on your 
computer, it is inherently a 

178
00:10:17,720 --> 00:10:19,880
single player thing. 
It's very hard to kind of 

179
00:10:20,840 --> 00:10:24,080
collaborate around something 
through Co work with multiple 

180
00:10:24,080 --> 00:10:25,680
humans. 
And that means that you need to 

181
00:10:25,680 --> 00:10:29,040
push that harness inside of a 
Austen environment. 

182
00:10:29,040 --> 00:10:33,120
This is the only way I think 
and, and there's a lot of things

183
00:10:33,120 --> 00:10:35,920
that we we are exploring around 
what is the right products of 

184
00:10:35,920 --> 00:10:38,960
face to enable that 
collaboration where you have 

185
00:10:38,960 --> 00:10:43,240
multiple humans, multiple 
sessions around a shared state 

186
00:10:44,440 --> 00:10:46,200
and how that what what it can 
look like. 

187
00:10:46,200 --> 00:10:49,440
I think this is the frontier in 
terms of products of interacting

188
00:10:49,440 --> 00:10:51,240
with agents, and that's the one 
we're very excited about. 

189
00:10:51,760 --> 00:10:54,720
Talk to me more about that. 
So one thing we're doing right 

190
00:10:54,720 --> 00:10:59,880
now is, is so we, we, we've, 
we've, we've released the first 

191
00:10:59,880 --> 00:11:04,760
version of that, which is called
really, we call it a pod. 

192
00:11:04,880 --> 00:11:11,000
And a pod is really a shared 
state with a number of sessions 

193
00:11:11,320 --> 00:11:14,680
between humans and agents. 
And I'll give you a very simple 

194
00:11:14,680 --> 00:11:16,680
example of where it plays really
nice. 

195
00:11:17,160 --> 00:11:22,240
So we have, as with any 
companies, we have a team weekly

196
00:11:22,840 --> 00:11:28,120
and the team weekly is mostly 
made of a set of slides that are

197
00:11:28,240 --> 00:11:31,360
somewhat always the same with 
small variations every week and 

198
00:11:31,360 --> 00:11:35,960
maybe some deep dive about 
subjects every other week. 

199
00:11:37,560 --> 00:11:40,400
And so in the past, the way it 
would work is that somebody 

200
00:11:40,400 --> 00:11:44,280
would stop doing a slide and and
ping the humans and have the 

201
00:11:44,280 --> 00:11:46,280
humans collaborate in those 
slides, etcetera. 

202
00:11:46,280 --> 00:11:50,000
And the kind of orchestration of
all of that would take a some 

203
00:11:50,000 --> 00:11:55,240
amount of time to that human. 
No comes agents. 

204
00:11:55,240 --> 00:11:58,920
Maybe you can the the initial 
organization building a slide. 

205
00:11:58,920 --> 00:12:01,960
You can, you can, you can 
offload to an agent, but you 

206
00:12:01,960 --> 00:12:05,840
still have to ping the humans 
and organize the, the kind of 

207
00:12:05,960 --> 00:12:09,120
organize the collaboration so 
that we get to a shared result. 

208
00:12:10,240 --> 00:12:15,520
And so with a pod, which is a 
group of conversations, group of

209
00:12:15,520 --> 00:12:19,120
humans and a shared state. 
No, our team weekly is much more

210
00:12:19,120 --> 00:12:21,240
driven by an agent than driven 
by a human. 

211
00:12:21,240 --> 00:12:25,360
So basically on our teamwork is 
on Tuesdays, Monday morning, 

212
00:12:25,680 --> 00:12:28,840
somebody triggers an agent with 
a skilled team weekly, the team,

213
00:12:28,920 --> 00:12:34,720
the agent creates a pod, it 
creates one session pair slide 

214
00:12:35,480 --> 00:12:39,400
with the owner of each of those 
areas and maybe one session for 

215
00:12:39,400 --> 00:12:42,920
the people that are doing the 
deep dive that week. 

216
00:12:43,600 --> 00:12:45,640
And it will will pre walk the 
slides. 

217
00:12:45,640 --> 00:12:48,720
So going into meta base, pulling
the data, keeping up to date 

218
00:12:48,720 --> 00:12:53,040
revenue customers, going on 
GitHub for the the product walk 

219
00:12:53,040 --> 00:12:55,480
and then going on whatever 
platform for the support state, 

220
00:12:55,480 --> 00:12:57,440
whatever, you know, prebuild the
slides. 

221
00:12:57,440 --> 00:13:00,240
And then it'll ping the the the 
the person responsible for that 

222
00:13:00,240 --> 00:13:02,360
slide and say, Hey, what do you 
want to talk this week about? 

223
00:13:02,360 --> 00:13:04,880
And let me let's collaborate on 
on this slides. 

224
00:13:04,880 --> 00:13:06,640
What do you want to had? 
Is there anything on the 

225
00:13:06,640 --> 00:13:08,400
highlight? 
And so that collaboration 

226
00:13:08,400 --> 00:13:12,160
happens in one of that session. 
Each humans that are responsible

227
00:13:12,160 --> 00:13:14,080
for each of those subjects are 
collaborating in different 

228
00:13:14,080 --> 00:13:16,880
sessions. 
And on Tuesday, let's say noon, 

229
00:13:17,080 --> 00:13:20,960
the agents, the agents wakes up 
and like, looks at everything, 

230
00:13:21,440 --> 00:13:24,840
probably Ding on the humans that
haven't finished their work, 

231
00:13:26,880 --> 00:13:29,160
wakes up again in two hours, 
like 2 hours before the two 

232
00:13:29,160 --> 00:13:31,880
weekly look at all the, the 
things that have been built and 

233
00:13:31,880 --> 00:13:34,280
coalesce them into one shed 
presentation. 

234
00:13:34,280 --> 00:13:35,920
So we, we don't use any 
presentation anymore. 

235
00:13:35,920 --> 00:13:39,480
We use, we have a frame which is
the equivalent of canvas or 

236
00:13:39,480 --> 00:13:43,000
artifacts or really kind of a 
culturated UIS. 

237
00:13:43,400 --> 00:13:47,960
And so it it takes all of those 
frames and and and pull them in 

238
00:13:47,960 --> 00:13:50,240
that in that short frame and 
that thing is ready. 

239
00:13:50,640 --> 00:13:53,920
And really the the weekly, the 
team weekly, which is kind of a 

240
00:13:54,320 --> 00:13:59,000
a multi days collaborative task 
is today mostly on auto mode, 

241
00:13:59,000 --> 00:14:00,800
right. 
And I think that's that's the 

242
00:14:00,800 --> 00:14:02,720
kind of a product surface that 
are interesting. 

243
00:14:02,920 --> 00:14:05,240
If you don't have that pod, 
which is that group with that 

244
00:14:05,240 --> 00:14:08,480
shut state feels kind of weird 
because you have a bunch of 

245
00:14:08,480 --> 00:14:11,760
conversations over the place. 
You don't have a shuts place to 

246
00:14:11,760 --> 00:14:13,800
see them. 
All the agents doesn't have a 

247
00:14:13,800 --> 00:14:16,120
place to organize all of that 
work with multiple humans. 

248
00:14:16,480 --> 00:14:19,120
And so I think that's that's the
type of things we're exploring 

249
00:14:19,120 --> 00:14:23,400
and which is very exciting for, 
for what's to come up for more 

250
00:14:23,400 --> 00:14:26,120
than for what's where we're at. 
Can you talk to me more about 

251
00:14:26,120 --> 00:14:27,640
this pod? 
Yeah. 

252
00:14:27,880 --> 00:14:32,760
So yeah, the the pod is really 
like, technically it's really a 

253
00:14:32,760 --> 00:14:38,000
group of humans, a group of 
agents, a group of sessions, 

254
00:14:38,000 --> 00:14:39,840
which is the interaction between
the human and agents. 

255
00:14:40,480 --> 00:14:44,080
And these are sandboxed. 
Each of the sessions has a 

256
00:14:44,080 --> 00:14:47,680
sandbox. 
Each of the session has a, a 

257
00:14:47,680 --> 00:14:51,600
state. 
So a state, let's go to file 

258
00:14:51,600 --> 00:14:54,080
system. 
The file system is visible by 

259
00:14:54,080 --> 00:14:57,040
the humans but also mounted on 
the sandbox basically. 

260
00:14:57,040 --> 00:14:59,880
To make that easy, we're diving 
into the technical details here.

261
00:14:59,880 --> 00:15:02,880
We use GCS FUSE. 
So we back that with GCS. 

262
00:15:02,880 --> 00:15:04,360
So it's very easy to show it in 
the UI. 

263
00:15:04,360 --> 00:15:07,920
It's very easy to show it as a 
mounted as a file system to the 

264
00:15:07,920 --> 00:15:09,600
agent. 
So the home directory of the 

265
00:15:09,600 --> 00:15:12,880
agent in the sandbox is is GCS 
back basically. 

266
00:15:13,960 --> 00:15:15,840
And so that's the state in one 
session. 

267
00:15:15,840 --> 00:15:19,600
And then inside of the pod we 
also have another file system 

268
00:15:20,040 --> 00:15:22,200
which is shared across the 
sessions. 

269
00:15:22,200 --> 00:15:24,880
And the agents are free to move 
files between the two file 

270
00:15:24,880 --> 00:15:26,720
systems. 
So basically they see slash 

271
00:15:26,720 --> 00:15:29,840
session slash blah, blah, blah. 
And in there they have all the 

272
00:15:30,160 --> 00:15:32,760
all the file system related to 
the current session. 

273
00:15:33,320 --> 00:15:37,840
And then they have slash pod and
the under slash pod. 

274
00:15:37,840 --> 00:15:40,280
That's the shared file system 
where they can move some of 

275
00:15:40,280 --> 00:15:44,000
those files up or down as they 
see fit, and so they can really 

276
00:15:44,000 --> 00:15:46,360
collaborate across sessions. 
You can have collaboration 

277
00:15:46,360 --> 00:15:48,440
across sections through that 
shut file system at the pod 

278
00:15:48,440 --> 00:15:51,320
level. 
And when I'm interacting with my

279
00:15:51,320 --> 00:15:55,880
agent, I can drop into somebody 
else's slides, presumably also 

280
00:15:56,080 --> 00:16:00,160
it's not just my slide. 
So if yeah, if the if the slides

281
00:16:00,160 --> 00:16:02,920
of that also, so somebody has a 
session with an agent inside of 

282
00:16:02,920 --> 00:16:06,000
the pod is working on a slide, 
it leaves naturally as a 

283
00:16:06,080 --> 00:16:08,440
starting point on the file 
system of the session. 

284
00:16:08,440 --> 00:16:11,280
So it's mostly visible to that 
session and then that person 

285
00:16:11,280 --> 00:16:13,480
will say, OK, the slides are 
ready, move them to the move 

286
00:16:13,480 --> 00:16:16,720
them to the to the pod, the 
slides get moved and then 

287
00:16:16,720 --> 00:16:20,920
somebody else will be able to 
see them or even modify them, 

288
00:16:21,240 --> 00:16:23,320
which is risky in a sense, but 
it's fine. 

289
00:16:23,320 --> 00:16:25,800
I mean, that's that's in Google 
Drive. 

290
00:16:25,800 --> 00:16:29,400
We can modify whatever else has 
been doing in your company as 

291
00:16:29,400 --> 00:16:30,360
well. 
So I think it's OK. 

292
00:16:31,560 --> 00:16:36,960
Do you have the possibility to 
look at the history of the 

293
00:16:36,960 --> 00:16:39,520
modifications in case you're 
like oh control Z? 

294
00:16:39,920 --> 00:16:44,400
Yeah, no, not right now, 
technically. 

295
00:16:44,400 --> 00:16:45,960
That's why it's risky then, 
yeah. 

296
00:16:45,960 --> 00:16:49,200
Technically, technically we have
it because GCS is version. 

297
00:16:49,200 --> 00:16:52,360
So technically we do have it and
that stuff is something that we 

298
00:16:52,360 --> 00:16:55,360
would want. 
Yes, you can. 

299
00:16:55,600 --> 00:16:58,280
You can. 
The the though the reality is 

300
00:16:58,280 --> 00:17:00,760
that you do something with an 
agent, it will. 

301
00:17:01,080 --> 00:17:03,680
What will happen is it's going 
to read the file and then it's 

302
00:17:03,680 --> 00:17:08,079
going to emit and edit. 
And so generally they're able to

303
00:17:08,079 --> 00:17:11,440
pretty easily backtrack their 
change because they still see 

304
00:17:11,440 --> 00:17:14,079
them in context. 
The only thing that the agent 

305
00:17:14,079 --> 00:17:18,800
can do that would be bad is like
like deleting the whole pod 

306
00:17:18,800 --> 00:17:22,599
sheds file system. 
That's possible. 

307
00:17:22,599 --> 00:17:26,200
Technically, I don't think it 
happens much. 

308
00:17:26,839 --> 00:17:29,960
So that's that's kind of fine. 
And also I imagine you have some

309
00:17:29,960 --> 00:17:32,840
kind of redundancy, maybe some 
backup somewhere, just in case 

310
00:17:32,840 --> 00:17:33,040
you. 
Yeah. 

311
00:17:33,040 --> 00:17:34,880
So again and just yes, 
everything is version. 

312
00:17:34,880 --> 00:17:38,360
So we could, we could totally 
restore it and and and yeah. 

313
00:17:40,440 --> 00:17:47,720
So OK, now the shared state and 
the ability for almost like this

314
00:17:47,720 --> 00:17:52,760
tiered level of I'm working, I'm
bringing it down, not 

315
00:17:52,760 --> 00:17:57,120
necessarily locally, but I'm 
having my one player experience 

316
00:17:58,280 --> 00:18:01,560
and then I'm pushing to the 
multiplayer so everyone can 

317
00:18:01,560 --> 00:18:05,920
share and look at it. 
How are you envisioning that in 

318
00:18:06,000 --> 00:18:12,720
other scenarios? 
There's, there's so many of 

319
00:18:12,720 --> 00:18:16,440
them. 
It's, it's one of the, by the 

320
00:18:16,440 --> 00:18:21,160
way, it's one of the very hard 
difficulty of building an 

321
00:18:21,160 --> 00:18:23,760
Orbiton platform. 
There's a lot of value into 

322
00:18:23,760 --> 00:18:25,320
building in the northern 
platform. 

323
00:18:25,360 --> 00:18:27,880
There's a lot of pain in 
building a northern platform. 

324
00:18:28,200 --> 00:18:32,360
The pain is mostly this one is 
that the, there is so many use 

325
00:18:32,360 --> 00:18:36,920
cases that's that you, it's, you
cannot, I mean, I would love to 

326
00:18:36,920 --> 00:18:39,920
be able to tell you dust is 
awesome for that, but that means

327
00:18:39,920 --> 00:18:42,280
we would be verticalized. 
If we were to be verticalized, 

328
00:18:42,280 --> 00:18:45,000
we wouldn't have the right to 
really try to equip the entire 

329
00:18:45,000 --> 00:18:47,120
company. 
And so that's why we accept that

330
00:18:47,200 --> 00:18:50,480
that tension. 
Many other ways to use a pod, 

331
00:18:50,480 --> 00:18:54,120
You can, you can use a pod as a,
as a, as a. 

332
00:18:54,240 --> 00:18:57,600
So basically we also have in 
that pod, we also have a list of

333
00:18:57,800 --> 00:19:01,080
tasks, which is a way to try to 
organize, walk. 

334
00:19:01,440 --> 00:19:03,400
And so you can create a pod for 
any project you do. 

335
00:19:04,040 --> 00:19:07,600
And let's say you're on Slack 
and you're like, oh, I need to 

336
00:19:07,600 --> 00:19:09,560
do, we need to do this, this and
that. 

337
00:19:09,600 --> 00:19:13,880
And you call Dust from Slack and
you ask it to add those tasks to

338
00:19:13,880 --> 00:19:15,920
the pod you're working on May, 
let's say you're working on the 

339
00:19:16,160 --> 00:19:18,600
initiative sandbox. 
You're working on adding sandbox

340
00:19:18,720 --> 00:19:22,680
to Dust and say we should do 
this, this and that the agent is

341
00:19:22,680 --> 00:19:26,080
capable of interacting with the 
pod and add tasks itself. 

342
00:19:26,360 --> 00:19:30,400
And then as a human, you can you
can the task are kind of a nice 

343
00:19:30,400 --> 00:19:36,160
way to organize the organize the
walk and trigger sessions to 

344
00:19:36,160 --> 00:19:40,680
start solving the task with an 
agent and a human working 

345
00:19:40,680 --> 00:19:44,560
together to solve that task. 
So you have that kind of a task 

346
00:19:44,680 --> 00:19:47,800
kind of trying to explore the 
task management part of working 

347
00:19:48,040 --> 00:19:52,920
with agents around the project. 
There's obviously another usage 

348
00:19:52,920 --> 00:19:57,800
of POD which is which works well
is around internal kind of 

349
00:19:58,160 --> 00:20:02,160
internal supports like. 
So obviously the kind of 

350
00:20:02,160 --> 00:20:05,520
external support, the tier one 
support will, will is best 

351
00:20:05,560 --> 00:20:07,960
tackled in platform that are 
specialized for that. 

352
00:20:08,320 --> 00:20:11,080
But when the questions is 
finicky and it reaches your 

353
00:20:11,080 --> 00:20:15,560
internal support team, a pod is 
a really great place to share 

354
00:20:15,640 --> 00:20:20,080
the way you solve those 
questions for people to, to, to,

355
00:20:20,160 --> 00:20:22,520
to look at and, and, and 
explore. 

356
00:20:22,960 --> 00:20:26,560
And so I think there's many, 
many use cases around that we're

357
00:20:26,560 --> 00:20:30,160
really in the business of 
creating Lego bricks that work 

358
00:20:30,160 --> 00:20:33,640
well together and and make sure 
that they are somewhat 

359
00:20:33,640 --> 00:20:38,520
universal. 
And so we see the very varied 

360
00:20:38,680 --> 00:20:40,520
usages. 
One principle that I think is 

361
00:20:40,520 --> 00:20:45,240
interesting that we try to 
enforce is what we call the BI 

362
00:20:45,240 --> 00:20:50,280
directional harness, meaning 
that we enforce us. 

363
00:20:50,360 --> 00:20:53,400
A very simple rule is that any 
feature we build, it should be 

364
00:20:53,400 --> 00:20:55,280
equally accessible by agents and
humans. 

365
00:20:56,720 --> 00:20:58,600
OK. 
There shouldn't be anything that

366
00:20:58,600 --> 00:21:01,280
a human can do in dust that an 
agent cannot do. 

367
00:21:01,640 --> 00:21:05,360
And this vessel? 
And how does that play out in 

368
00:21:05,480 --> 00:21:07,560
the time when you're actually 
building? 

369
00:21:08,800 --> 00:21:11,680
So it, it, it kind of creates a 
constraint of the things you can

370
00:21:12,040 --> 00:21:15,880
build. 
As an example, I mentioned the 

371
00:21:15,880 --> 00:21:19,120
tasks around the pod. 
This is a product surface that 

372
00:21:19,120 --> 00:21:24,040
is kind of very human centric. 
Well, it's, it's, it's, it's, 

373
00:21:24,400 --> 00:21:27,520
it's very important that those 
tasks are, can be listed by 

374
00:21:27,520 --> 00:21:30,920
agents, can be edited by agents,
can be created by agents, can be

375
00:21:31,360 --> 00:21:33,120
the what can be triggered. 
Another conversation can be 

376
00:21:33,120 --> 00:21:34,200
triggered. 
The other session can be 

377
00:21:34,200 --> 00:21:36,080
triggered from the stats by 
agents as well. 

378
00:21:36,560 --> 00:21:39,680
And so it's really kind of a 
great principle to make sure 

379
00:21:39,680 --> 00:21:43,720
that we, we don't overfit on any
of those two users because we're

380
00:21:43,720 --> 00:21:46,360
really building a product for 
both those users. 

381
00:21:47,280 --> 00:21:51,400
And we really care about that 
because we see it's the only way

382
00:21:51,400 --> 00:21:55,240
to start meshing the work of 
humans and agents together. 

383
00:21:55,280 --> 00:21:57,840
In our opinion. 
Only if you have that kind of 

384
00:21:57,840 --> 00:22:04,280
symmetry will you progressively 
stop making it fuzzy whether the

385
00:22:04,280 --> 00:22:07,280
work has been done by humans or 
an agents, which we think is the

386
00:22:07,280 --> 00:22:08,840
future and is exciting in many 
ways. 

387
00:22:10,080 --> 00:22:13,320
Other things that there is so 
much, so much places where you 

388
00:22:13,320 --> 00:22:15,480
really want to build something 
for humans. 

389
00:22:15,480 --> 00:22:17,800
And when you try to apply and 
believe strongly in that 

390
00:22:17,800 --> 00:22:20,720
principle, it really makes you 
rethink what it should look 

391
00:22:20,720 --> 00:22:23,680
like. 
And so it's a useful principle 

392
00:22:23,680 --> 00:22:25,960
to to apply. 
But I think most AI products 

393
00:22:25,960 --> 00:22:28,080
today are kind of building for 
2/2 users. 

394
00:22:28,080 --> 00:22:29,960
And so it's important to build 
for them both. 

395
00:22:30,440 --> 00:22:35,040
Yeah, it's a It reminds me of 
when someone told me you can 

396
00:22:35,040 --> 00:22:40,400
create documentation that humans
can read, but agents can't. 

397
00:22:41,160 --> 00:22:44,600
But you can't do it vice versa. 
If you create documentation for 

398
00:22:44,600 --> 00:22:46,880
agents, a human is going to 
understand it. 

399
00:22:46,880 --> 00:22:51,400
So you should default to agents 
being able to understand your 

400
00:22:51,400 --> 00:22:56,360
documentation first and 
foremost, and humans will by 

401
00:22:56,360 --> 00:23:00,440
default also understand. 
Now, when you were talking about

402
00:23:00,720 --> 00:23:07,720
the idea of the pods and also 
how work gets done over longer 

403
00:23:07,720 --> 00:23:13,640
time horizons, it made me 
realize, and I had the vision of

404
00:23:13,800 --> 00:23:19,200
data moving through like a DAG 
and how a finished piece of work

405
00:23:19,840 --> 00:23:25,760
is not dissimilar to that. 
You have different nodes where 

406
00:23:25,760 --> 00:23:28,600
work gets done and then it goes 
to the next node. 

407
00:23:28,600 --> 00:23:33,760
And whether that node and that 
piece of work is being done by a

408
00:23:33,760 --> 00:23:37,200
human or it's being done by an 
agent, it doesn't always have to

409
00:23:37,200 --> 00:23:39,720
be the same human that is doing 
it. 

410
00:23:39,720 --> 00:23:41,920
And I think that's the huge 
unlock that I'm understanding 

411
00:23:41,920 --> 00:23:49,160
with you is that a human is an 
expert in different parts of 

412
00:23:49,160 --> 00:23:52,920
that process. 
So it shouldn't be the same 

413
00:23:53,320 --> 00:23:55,400
human that owns it all the way 
through. 

414
00:23:55,400 --> 00:23:58,840
We hear a lot of talk of how 
well now PMS and designers are 

415
00:23:58,840 --> 00:24:06,240
shipping code, and that's great.
But they have their unique skill

416
00:24:06,240 --> 00:24:08,480
set where they're very good at 
things. 

417
00:24:08,960 --> 00:24:14,760
And isn't it almost more 
efficient if they can do what 

418
00:24:14,760 --> 00:24:19,880
they're best at in that note of 
the DAG and then it gets shipped

419
00:24:20,000 --> 00:24:22,680
to someone who can make it 
battle hard. 

420
00:24:22,680 --> 00:24:25,160
Maybe it's the dev OPS person 
and the security person. 

421
00:24:25,360 --> 00:24:27,560
They can harden it and then it 
goes to production. 

422
00:24:27,560 --> 00:24:30,760
And then you have that feedback 
loop where it will go back 

423
00:24:30,760 --> 00:24:33,280
around and you're constantly 
iterating and updating. 

424
00:24:33,680 --> 00:24:36,760
But those longer time horizon 
events. 

425
00:24:37,080 --> 00:24:40,600
Now when you talk about 
multiplayer, I'm understanding 

426
00:24:40,600 --> 00:24:44,760
what you're saying there. 
Yeah, I think, I think that's 

427
00:24:45,080 --> 00:24:52,320
that's exactly that. 
And and and the grail is the 

428
00:24:52,320 --> 00:24:55,360
grail is really try to create a 
place where it's really easy to 

429
00:24:55,480 --> 00:24:58,440
end off. 
And so to really create an edge 

430
00:24:58,520 --> 00:25:02,520
and an A next node and end off 
to the the the next processing 

431
00:25:02,520 --> 00:25:03,840
unit. 
And the next processing might be

432
00:25:03,840 --> 00:25:06,600
humans, might be an agents. 
And it should be really easy to 

433
00:25:06,600 --> 00:25:11,240
root towards one or the other 
and and and and and and, and, 

434
00:25:11,400 --> 00:25:14,560
and that's, that's the thing. 
And, and to make that a reality,

435
00:25:15,280 --> 00:25:19,400
the, the complex, the, the, the,
The Dirty bulk that needs to 

436
00:25:19,400 --> 00:25:23,560
happen is that those agents has 
to be very contextful. 

437
00:25:23,720 --> 00:25:29,560
And like it's super interesting 
into trying to think what is the

438
00:25:29,560 --> 00:25:34,040
right context for those agents 
to be useful agents of work. 

439
00:25:34,720 --> 00:25:39,520
Is it just, I mean, today the 
basic answer is a long agency 

440
00:25:39,520 --> 00:25:43,560
group and the gazillion MCP 
calls to rebuild the context in 

441
00:25:43,560 --> 00:25:49,360
every session, the kind of a 
second stage and Sir is is turn 

442
00:25:49,360 --> 00:25:52,400
everything into a skill and 
eventually the context will be 

443
00:25:52,400 --> 00:25:55,400
rebuilt. 
Is that part automated? 

444
00:25:55,400 --> 00:25:58,040
Is that part on autopilot? 
Is that part manual? 

445
00:25:58,040 --> 00:26:00,520
These are all the kind of a 
question at the frontier of what

446
00:26:00,520 --> 00:26:03,920
it means to create that that 
right context for agents, which 

447
00:26:03,920 --> 00:26:08,280
is quite interesting. 
Yeah, how are you seeing success

448
00:26:08,280 --> 00:26:11,240
with that and breaking down 
those barriers? 

449
00:26:11,240 --> 00:26:14,760
So it is more fuzzy when you 
have the hand off and you can 

450
00:26:14,760 --> 00:26:17,760
still. 
I am also understand there's 

451
00:26:17,760 --> 00:26:22,120
moments where there's two 
different people, humans that 

452
00:26:22,120 --> 00:26:24,760
are working on the same thing 
and then the agent in the 

453
00:26:24,760 --> 00:26:26,440
middle. 
And so you almost have like this

454
00:26:26,440 --> 00:26:29,840
Venn diagram of you have the 
product manager, you have the 

455
00:26:29,840 --> 00:26:34,240
engineer, you have the agents 
all looking at one node, if 

456
00:26:34,240 --> 00:26:37,240
we're going to continue the 
metaphor of the DAG. 

457
00:26:40,200 --> 00:26:44,120
I, I, it's, I think we are 
still, I don't have a crisp 

458
00:26:44,120 --> 00:26:48,040
necessary answer to that. 
I think we are, again, I think 

459
00:26:48,040 --> 00:26:51,840
we are still very early on, on 
this one because as we said it 

460
00:26:52,200 --> 00:26:54,840
as, as we said at the very 
beginning, the capabilities of 

461
00:26:54,840 --> 00:26:58,280
the agents are still kind of 
jagged. 

462
00:26:58,360 --> 00:27:02,280
And so that makes that whole 
process not super smooth at all 

463
00:27:02,280 --> 00:27:04,600
time, but we start seeing light 
at the end of the tunnel. 

464
00:27:04,960 --> 00:27:07,680
And so our goal is really to 
discover that with our user to 

465
00:27:07,680 --> 00:27:11,280
give them the right Lego bricks.
That means that the, the, the, 

466
00:27:11,520 --> 00:27:14,320
the within, within our user 
base, there's always going to be

467
00:27:14,320 --> 00:27:19,320
kind of a AI operators, 
tinkerers that will see the 

468
00:27:19,320 --> 00:27:22,160
bricks and will want to assemble
them in kind of new ways. 

469
00:27:22,520 --> 00:27:25,880
And it's so it's super exciting 
to see what people do with the 

470
00:27:25,880 --> 00:27:28,000
bricks. 
You give them sometime they come

471
00:27:28,000 --> 00:27:30,960
up with stuff that are just mind
blowing. 

472
00:27:30,960 --> 00:27:34,080
They just rebuild entire 
systems, all of that. 

473
00:27:34,080 --> 00:27:36,400
And it's, it's exciting to see, 
But I think it's a, it's a Co 

474
00:27:36,400 --> 00:27:38,360
discovery. 
It's a Co discovery thing. 

475
00:27:38,360 --> 00:27:41,200
So it's a would be very, would 
be very. 

476
00:27:41,200 --> 00:27:45,600
I mean, on top of that, it's so 
dependent on the capability of 

477
00:27:45,600 --> 00:27:48,880
models which is moving so fast 
and and which is still kind of 

478
00:27:48,920 --> 00:27:50,640
somewhat unpredictable in many 
ways. 

479
00:27:51,280 --> 00:27:56,120
So it would be very adventurous 
of of me of trying to to have a 

480
00:27:56,120 --> 00:27:59,840
very crisp answer on this one. 
Slant a flag right now and then 

481
00:28:00,240 --> 00:28:02,800
it's like by the time this 
podcast airs. 

482
00:28:03,360 --> 00:28:09,760
Exactly, exactly. 
Going back to the alignment, I 

483
00:28:09,760 --> 00:28:12,920
kind of seated that like at the 
very beginning. 

484
00:28:13,360 --> 00:28:15,680
Going back to the alignment, 
this is one of the major 

485
00:28:15,680 --> 00:28:19,320
challenge I see into trying to 
reproduce the kind of conditions

486
00:28:19,320 --> 00:28:22,240
of Stripe early on. 
I think it's the alignment part 

487
00:28:22,240 --> 00:28:25,720
because when you're operating in
that space, the technological 

488
00:28:25,720 --> 00:28:29,640
substrate is moving under your 
feet in a way that wasn't true 

489
00:28:29,640 --> 00:28:31,080
before. 
For the past 20 years when we 

490
00:28:31,080 --> 00:28:34,480
are building tech companies, you
are building on JavaScript and 

491
00:28:34,480 --> 00:28:39,920
Postgres was really stable. 
Like there's nothing more 

492
00:28:39,920 --> 00:28:41,920
stable. 
I mean, those things evolve, but

493
00:28:41,920 --> 00:28:45,160
it's like stable, stable, 
concrete type stable. 

494
00:28:46,000 --> 00:28:48,760
And today we're building on the 
on models that are changing 

495
00:28:48,760 --> 00:28:51,840
every weeks. 
And so every time we try to 

496
00:28:51,840 --> 00:28:55,640
build a very crisp picture of, 
of what we could be constructing

497
00:28:55,640 --> 00:28:59,600
the year, we'd got it completely
wrong because within that time 

498
00:28:59,600 --> 00:29:03,280
frame, the models shifted the 
place we wanted to build that 

499
00:29:03,280 --> 00:29:06,920
nice city of yours kind of went 
undersea and the big mountain 

500
00:29:06,920 --> 00:29:09,400
appeared right next to it. 
And so now you want to be on top

501
00:29:09,400 --> 00:29:13,960
of the mountain, obviously. 
And so that creates the that 

502
00:29:13,960 --> 00:29:17,600
creates, that creates a real 
challenge for creating a crisp 

503
00:29:17,640 --> 00:29:20,960
alignment of the team because 
you, you, you, you're subject to

504
00:29:20,960 --> 00:29:24,520
the fog of AI. 
And it's very hard to see past 

505
00:29:24,520 --> 00:29:29,480
six months, let's say. 
I like this term, the fog of AI.

506
00:29:29,520 --> 00:29:32,120
The fog of AI, yeah, that's a, 
that's a, that's a pretty, 

507
00:29:32,160 --> 00:29:34,440
that's one of an interesting 
challenge of building a building

508
00:29:34,720 --> 00:29:39,720
an AI company, I guess. 
And now just going back to dust 

509
00:29:39,720 --> 00:29:45,480
and how you're building. 
Do you allow folks to interact 

510
00:29:45,480 --> 00:29:50,080
and be multiplayer in all 
different types of scenarios 

511
00:29:50,080 --> 00:29:55,720
like Slack, like in their web 
browser or in GitHub or they 

512
00:29:55,720 --> 00:29:59,400
throw like a JIRA or linear 
issue at it? 

513
00:29:59,400 --> 00:30:01,960
Is that kind of the vision where
you're saying we have all these 

514
00:30:01,960 --> 00:30:04,680
tentacles? 
Yeah, we try to build as much 

515
00:30:04,680 --> 00:30:07,600
tentacles as as possible. 
The big one is obviously. 

516
00:30:07,600 --> 00:30:10,640
So there's always A and there's 
always an interesting tension is

517
00:30:10,640 --> 00:30:14,240
that when you build one of those
connection like Slack is one of 

518
00:30:14,240 --> 00:30:16,720
the one of the big one. 
It's the very natural one. 

519
00:30:16,720 --> 00:30:20,240
It's a chat system. 
And so that's the place where 

520
00:30:20,240 --> 00:30:22,920
you might want to trigger work 
with agents. 

521
00:30:24,840 --> 00:30:28,240
It is also an acquisition engine
internally within work within 

522
00:30:28,720 --> 00:30:32,280
companies because people are on 
Slack and they see other people 

523
00:30:32,280 --> 00:30:35,080
using a system, so they discover
it and they might start using 

524
00:30:35,080 --> 00:30:38,000
it. 
Obviously the, the tension is 

525
00:30:38,000 --> 00:30:41,760
that the, you're constrained by 
the interface of the platform 

526
00:30:41,760 --> 00:30:44,600
you're connected to. 
And so the experience, 

527
00:30:44,600 --> 00:30:47,800
everything we just said is that 
is very hard to do in Slag 

528
00:30:47,800 --> 00:30:50,600
because you, you're bound by the
interface. 

529
00:30:50,600 --> 00:30:52,800
And Slag has been doing a lot of
work there and, and they 

530
00:30:53,080 --> 00:30:55,640
actually you can create pretty 
smooth interface today in Slag. 

531
00:30:55,640 --> 00:30:57,280
But let's say GitHub as an 
example. 

532
00:30:58,360 --> 00:31:03,640
It's just a text, text, I think.
You don't see the syncing 

533
00:31:03,640 --> 00:31:06,440
tokens, you don't see the tools 
being used, all of that. 

534
00:31:06,440 --> 00:31:12,720
And so I, we really see those 
places as entry points more than

535
00:31:12,720 --> 00:31:16,040
place where the collaborations 
or the collaboration with agents

536
00:31:16,040 --> 00:31:20,040
really happens because you 
always want to, you always try 

537
00:31:20,040 --> 00:31:23,960
to drive the user to go open the
session inside of those because 

538
00:31:23,960 --> 00:31:25,840
that's where all the richness 
happens. 

539
00:31:26,840 --> 00:31:29,680
I like that. 
So it's the gateway into it and 

540
00:31:29,680 --> 00:31:34,760
you're pushing them to go and 
look and inspect on dust so that

541
00:31:35,000 --> 00:31:36,840
you can get that richer 
experience. 

542
00:31:36,840 --> 00:31:40,640
Yeah, some actions that agents 
take for the for obvious reasons

543
00:31:40,680 --> 00:31:44,040
of some admins mark as high 
stakes. 

544
00:31:44,040 --> 00:31:47,600
So it means it requires a user 
confirmation before the action 

545
00:31:47,600 --> 00:31:49,320
is taken. 
And a good example is sending an

546
00:31:49,320 --> 00:31:54,320
e-mail and. 
And so in Slack you can actually

547
00:31:54,400 --> 00:31:57,960
build a pretty good tool 
approval experience. 

548
00:31:58,400 --> 00:32:02,200
But in GitHub you cannot. 
You can't send the next message 

549
00:32:02,200 --> 00:32:04,280
on the issue. 
Say you need to click here to 

550
00:32:04,280 --> 00:32:06,000
approve the tool. 
That wouldn't make sense. 

551
00:32:06,520 --> 00:32:11,040
Well, I'm also thinking about 
even things like MCP apps that 

552
00:32:11,040 --> 00:32:13,920
make the experience so much 
richer. 

553
00:32:14,040 --> 00:32:17,960
You don't have that ability if 
I'm not mistaken. 

554
00:32:17,960 --> 00:32:22,120
You could you could, you could 
imagine that Slack rebuilds 

555
00:32:22,240 --> 00:32:25,960
supports MCP app and you managed
to forward the MCP app payload 

556
00:32:25,960 --> 00:32:28,080
to Slack. 
That could be, but again, GitHub

557
00:32:28,080 --> 00:32:30,280
wouldn't work, Zendesk, all all 
of that. 

558
00:32:30,440 --> 00:32:34,320
Pretty complicated. 
Yeah, we're there's so many 

559
00:32:34,320 --> 00:32:39,480
areas or surfaces that you touch
that do not support that. 

560
00:32:39,560 --> 00:32:43,240
And it would be very nice to 
have that richer data 

561
00:32:43,240 --> 00:32:48,160
experience, like even at JIRA or
Linear, you might want to see 

562
00:32:48,160 --> 00:32:52,120
some of these data points or 
data exploration that happens 

563
00:32:52,120 --> 00:32:57,600
with the agent before you commit
to a quote UN quote Sprint, 

564
00:32:57,600 --> 00:33:00,400
because I don't know what 
happens in sprints these days. 

565
00:33:00,400 --> 00:33:04,040
It's an absolute mess. 
Maybe that's a thread we can 

566
00:33:04,040 --> 00:33:08,600
pull on, on how sprints work 
now, but don't continuing on 

567
00:33:08,600 --> 00:33:12,760
this MCP apps. 
You almost want an area where 

568
00:33:12,760 --> 00:33:18,920
you can use all of the data that
you're having happen in these 

569
00:33:19,200 --> 00:33:23,560
different spots, like the slacks
or the Jiras or the linears. 

570
00:33:23,560 --> 00:33:30,240
And then put them in something 
to allow folks to have that 

571
00:33:30,240 --> 00:33:33,640
richer experience. 
So I can see that being a huge 

572
00:33:33,640 --> 00:33:34,600
value prop. 
Yeah. 

573
00:33:35,040 --> 00:33:37,880
But we, we really see those 
platforms as I mean we try to 

574
00:33:37,920 --> 00:33:39,640
insert ourselves in a discos 
system. 

575
00:33:39,640 --> 00:33:42,080
We don't know, we don't want 
everything to happen on dust. 

576
00:33:42,080 --> 00:33:47,120
We have walked some some amount 
of walk happen on dust, but we 

577
00:33:47,240 --> 00:33:50,520
sometimes that walk starts from 
Slack, continues on dust and 

578
00:33:50,520 --> 00:33:52,920
ends up in GitHub as a as an 
issue. 

579
00:33:54,040 --> 00:33:56,240
And that's really fine. 
It's really important to be able

580
00:33:56,240 --> 00:33:57,880
to read from them. 
It's really important to be able

581
00:33:57,880 --> 00:34:01,440
to write back to them. 
I think that's the that's the at

582
00:34:01,440 --> 00:34:03,000
this opposition on this one. 
Yeah. 

583
00:34:03,000 --> 00:34:06,160
And it goes back to the whole 
fuzziness of the handoffs and 

584
00:34:06,160 --> 00:34:11,120
how those work right now and how
that actually is is being done. 

585
00:34:11,120 --> 00:34:13,440
Because you've got some folks 
that are working in Slack or 

586
00:34:13,440 --> 00:34:18,080
you've got decisions being made 
in Slack and then an agent will 

587
00:34:18,080 --> 00:34:21,199
execute on that decision that's 
being made, which then becomes a

588
00:34:21,199 --> 00:34:26,840
GitHub issue or APR. 
And then somebody else has to 

589
00:34:26,840 --> 00:34:30,600
review it, but they are using an
agent to review it. 

590
00:34:30,600 --> 00:34:37,080
So you've, you've got all these 
different ways that it could be 

591
00:34:37,080 --> 00:34:41,480
siloed or it probably is being 
siloed right now. 

592
00:34:41,719 --> 00:34:45,760
And I know when we had our lunch
and learns, one of the big 

593
00:34:45,760 --> 00:34:49,920
things that folks asked for was 
how are teams working with 

594
00:34:50,120 --> 00:34:53,880
coding agents? 
How are you figuring out which 

595
00:34:53,880 --> 00:34:57,520
skills should be part of the 
whole company? 

596
00:34:57,520 --> 00:35:02,040
Or should it be team wide skill?
How are you making sure that 

597
00:35:02,040 --> 00:35:04,720
sessions can be reproduced or 
shared? 

598
00:35:04,720 --> 00:35:10,520
Or are you committing your whole
chat history to the PR? 

599
00:35:10,520 --> 00:35:13,800
Like there's these questions 
that are coming up and we don't 

600
00:35:13,800 --> 00:35:15,480
have the tools for them right 
now. 

601
00:35:15,880 --> 00:35:19,240
No, at least in the in the 
coding agent space, we don't 

602
00:35:19,240 --> 00:35:22,000
have them yet. 
But that seems a I mean it seems

603
00:35:22,000 --> 00:35:28,560
pretty obvious that at some 
point your codecs dash dash 

604
00:35:28,560 --> 00:35:31,600
session and that ID or cloud 
dash dash session and that ID 

605
00:35:31,600 --> 00:35:34,960
will be hosted in the cloud and 
you'll be able to tie it to the 

606
00:35:34,960 --> 00:35:37,920
PR and somebody will want to re 
restart from there. 

607
00:35:38,480 --> 00:35:41,200
That seem pretty obvious. 
For. 

608
00:35:41,320 --> 00:35:46,480
Them. 
Now I don't want to Passover 

609
00:35:46,480 --> 00:35:53,240
this very big important thought 
of finding alignment is very 

610
00:35:53,240 --> 00:35:57,160
tricky in the age of AI because 
you have the fog of AI, right? 

611
00:35:57,240 --> 00:35:59,720
Yep. 
Are there things that you have 

612
00:36:00,240 --> 00:36:05,840
done or seen work, or just like 
a principle that you're 

613
00:36:05,840 --> 00:36:10,800
operating on in dust so that you
can not necessarily predict the 

614
00:36:10,800 --> 00:36:13,680
future but iterate quickly when 
it does happen? 

615
00:36:13,960 --> 00:36:16,280
Yeah, exactly. 
I think, I think the the, the 

616
00:36:16,280 --> 00:36:22,240
failure mode of that is to 
refuse to paint a picture. 

617
00:36:22,880 --> 00:36:27,280
I think you have to operate with
some form of conviction within 

618
00:36:27,440 --> 00:36:30,840
uncertainty. 
It has to be clear to everyone 

619
00:36:30,840 --> 00:36:34,200
that we are operating in a very 
uncertain environment and that 

620
00:36:34,200 --> 00:36:37,960
whatever we say will be subject 
to change because the 

621
00:36:37,960 --> 00:36:42,400
environment change, but that we 
create conviction on the actual 

622
00:36:42,400 --> 00:36:48,000
current direction we're taking. 
And so that means people, I 

623
00:36:48,000 --> 00:36:51,560
mean, everybody wants to rally 
on a vision or rally on a 

624
00:36:51,560 --> 00:36:54,440
direction. 
And so if you cannot give a one 

625
00:36:54,440 --> 00:36:58,200
year and two year direction, you
have to give a six months 

626
00:36:58,200 --> 00:37:00,960
direction, that's fine. 
And whenever you give a six 

627
00:37:00,960 --> 00:37:03,680
months direction, by month 3, 
it's, it's going to be updated. 

628
00:37:03,720 --> 00:37:06,440
And so that means that you, you,
you need to have that kind of 

629
00:37:06,440 --> 00:37:10,560
muscle of updating it and being,
providing a lot of conviction 

630
00:37:10,840 --> 00:37:14,880
in, in, in that process, but 
really being very explicit each 

631
00:37:14,880 --> 00:37:17,160
time in doing it because it's 
wise. 

632
00:37:17,400 --> 00:37:21,840
You might fall in the trap of, 
of not not creating that, that, 

633
00:37:21,840 --> 00:37:26,880
that shared at least speed 
vector that you're trying to, to

634
00:37:26,880 --> 00:37:29,960
have for the group. 
And that's the worst, that's 

635
00:37:29,960 --> 00:37:32,800
the, the, the, the worst part. 
Basically the longer the speed 

636
00:37:32,800 --> 00:37:37,600
vector, the best it is, the 
shorter it's, it's worse. 

637
00:37:38,800 --> 00:37:43,200
Three months is so that's we're 
mixing time with speed here, but

638
00:37:43,200 --> 00:37:45,760
it's fine. 
Three months, what would be what

639
00:37:45,760 --> 00:37:48,080
would be anyway, three months 
definitely swing starts. 

640
00:37:48,560 --> 00:37:51,120
Three months is better than 0 
because then you don't know the 

641
00:37:51,120 --> 00:37:52,520
direction. 
Yeah. 

642
00:37:52,800 --> 00:37:55,360
And I imagine how it plays out 
in practices. 

643
00:37:55,360 --> 00:37:59,160
Every day folks are showing you 
new things saying hey, should we

644
00:37:59,160 --> 00:38:03,480
maybe go in this direction and 
you have to make those hard 

645
00:38:03,480 --> 00:38:08,760
decisions of is it worth course 
correcting for this? 

646
00:38:08,760 --> 00:38:12,280
Have we seen something that 
fundamentally changes where we 

647
00:38:12,280 --> 00:38:14,520
want to go? 
Yep, yeah, this. 

648
00:38:15,120 --> 00:38:18,440
So what we're trying to do is 
we're trying to really give a 

649
00:38:18,440 --> 00:38:21,800
good picture of what we want to 
achieve within three to six 

650
00:38:21,800 --> 00:38:24,840
months. 
But then we also, so that's, 

651
00:38:24,840 --> 00:38:30,640
that's kind of a, a photo like 
it's a, it's a snapshots and we 

652
00:38:30,640 --> 00:38:34,040
try to update it, but it's kind 
of at the horizon of big O of 

653
00:38:34,040 --> 00:38:37,640
months. 
And then we, we have what we 

654
00:38:37,640 --> 00:38:40,200
actually doing. 
And so we call it our stack 

655
00:38:40,200 --> 00:38:41,000
ranks. 
That's pretty easy. 

656
00:38:41,000 --> 00:38:44,640
It's a list of projects that we 
actively working on and we try 

657
00:38:44,640 --> 00:38:47,440
to create it very clearly so 
that the rest of everybody 

658
00:38:47,440 --> 00:38:49,640
knows. 
So engineering generally knows 

659
00:38:49,640 --> 00:38:52,760
what they're working on, not 
always true, but most of the 

660
00:38:52,760 --> 00:38:56,880
time it's better if true and 
enables everybody else within 

661
00:38:56,880 --> 00:39:00,000
the company to know what we are 
changing on the product right 

662
00:39:00,000 --> 00:39:02,880
now. 
And and then that's, that's 

663
00:39:02,880 --> 00:39:05,440
stack rank. 
We really try to keep it very 

664
00:39:05,440 --> 00:39:08,200
dynamic. 
So that means that we, we have a

665
00:39:08,200 --> 00:39:11,480
process which we call stack rank
update that is open to anybody 

666
00:39:11,480 --> 00:39:14,280
in the company to update the 
stack rank. 

667
00:39:14,840 --> 00:39:17,720
And so obviously it gets windy 
at the top. 

668
00:39:17,880 --> 00:39:21,320
So it's not an easy process. 
It's not a, that hard process 

669
00:39:21,320 --> 00:39:25,120
either, but basically a button 
you can push and both founders 

670
00:39:25,120 --> 00:39:27,840
appear in the room and we chat 
about how we should update the 

671
00:39:27,840 --> 00:39:30,360
stack rank And everybody was in 
the company has access to that, 

672
00:39:30,440 --> 00:39:33,360
to that. 
But generally we try to, we 

673
00:39:33,360 --> 00:39:36,600
expect, we, we have a high 
expectation for the input to 

674
00:39:36,600 --> 00:39:41,080
that process depending on the on
the size of the, the kind of 

675
00:39:41,240 --> 00:39:43,760
appetite that is traded with the
updates. 

676
00:39:44,520 --> 00:39:46,880
But but I think it's very nice 
because it gives a policy to 

677
00:39:46,880 --> 00:39:49,920
everybody to be a bit of a 
product person within the 

678
00:39:49,920 --> 00:39:53,640
company and we're trying to push
everybody to be a bit of product

679
00:39:53,640 --> 00:39:57,320
person inside of a company. 
Let's change gears and talk 

680
00:39:57,320 --> 00:39:59,320
about something. 
Before we hit record, we were 

681
00:39:59,320 --> 00:40:04,000
mentioning tokenomics and I 
think that is. 

682
00:40:04,480 --> 00:40:08,160
Very hot topic these days 
because the general narrative 

683
00:40:08,160 --> 00:40:13,360
you hear from the Internet is, 
oh, the era of subsidies is over

684
00:40:13,680 --> 00:40:15,520
now folks are trying to go 
public. 

685
00:40:15,520 --> 00:40:17,240
We're not going to get these 
cheap tokens. 

686
00:40:17,560 --> 00:40:20,200
I'm not sure that I fully buy 
into that. 

687
00:40:20,360 --> 00:40:24,520
Whenever I hear something from 
the masses, I that automatically

688
00:40:24,520 --> 00:40:31,640
gives me like red flag. 
However, I do see that the costs

689
00:40:31,640 --> 00:40:38,640
are going up a lot and it is 
really easy to spend a lot of 

690
00:40:38,640 --> 00:40:41,000
money and not create anything of
value. 

691
00:40:41,000 --> 00:40:43,760
So there's like that. 
Is. 

692
00:40:45,320 --> 00:40:47,240
I can burn a whole hell of a lot
of tokens. 

693
00:40:47,240 --> 00:40:50,520
That doesn't necessarily mean 
I'm good at my job, right? 

694
00:40:51,000 --> 00:40:54,920
And so I do understand that you 
were mentioning there's the cost

695
00:40:54,920 --> 00:40:57,240
of tokens, but then there's also
the cost of inference. 

696
00:40:57,240 --> 00:41:02,320
And so let's maybe center the 
next 5-10 minutes around that. 

697
00:41:03,600 --> 00:41:07,080
Well, I think the the way the 
way I we think, or at least I 

698
00:41:07,080 --> 00:41:11,760
think about that is we know the 
end states this we are subject 

699
00:41:11,760 --> 00:41:17,240
to gravity, this technology as a
sailing, it will converge and 

700
00:41:17,240 --> 00:41:20,040
when it does, it will be 
commoditized and tokens will be 

701
00:41:20,040 --> 00:41:22,120
cheap and plentiful. 
That's the end state. 

702
00:41:23,320 --> 00:41:27,520
The question is, and maybe we'll
have a side by them, but 

703
00:41:27,640 --> 00:41:30,400
whether, whatever is the state 
where it reached, it will 

704
00:41:30,400 --> 00:41:34,000
plateau at some point because 
it, and there's a limited amount

705
00:41:34,000 --> 00:41:37,920
of energy and stuff, you know, 
and, and I think it when, when 

706
00:41:37,920 --> 00:41:41,520
it plateaus, it'll get 
communitized and it'll be pretty

707
00:41:41,520 --> 00:41:44,480
cheap, very close to the price 
of, of, of power. 

708
00:41:44,600 --> 00:41:47,120
Basically. 
The question is when that 

709
00:41:47,120 --> 00:41:50,320
happens, maybe it opens in a in 
a year, maybe it happens in five

710
00:41:50,320 --> 00:41:53,600
years, maybe in 20 years, maybe 
it happens in 100 years, who 

711
00:41:53,600 --> 00:41:56,520
knows. 
And so the question is, is 

712
00:41:56,520 --> 00:42:00,240
trying to think about what's the
transition state between now and

713
00:42:00,240 --> 00:42:02,280
then. 
But I guess the end state is 

714
00:42:02,360 --> 00:42:05,760
plentifulness, cheap 
intelligence from from 

715
00:42:05,760 --> 00:42:10,960
electricity. 
And then, well, the one thing 

716
00:42:10,960 --> 00:42:13,720
that I've kind of was a new 
learning for me. 

717
00:42:13,720 --> 00:42:16,960
So I'll, I'll share it because I
think it was a learning for me 

718
00:42:16,960 --> 00:42:21,560
in the past couple weeks months 
is that I've always supposed 

719
00:42:21,560 --> 00:42:26,280
that if models performance 
plateau, we would enter a very 

720
00:42:26,280 --> 00:42:29,920
rapid communication phase. 
And I think I've been coming 

721
00:42:29,920 --> 00:42:36,200
back from that, that assumption 
because the the demand has 

722
00:42:36,200 --> 00:42:39,800
increased so much. 
That's the pressure on 

723
00:42:39,800 --> 00:42:42,080
inference. 
So just the the fact of serving 

724
00:42:42,080 --> 00:42:49,480
the tokens is so high that the 
actors that have pre bought that

725
00:42:49,480 --> 00:42:54,680
inference capacity will be able 
to maintain high margin due to 

726
00:42:54,680 --> 00:42:58,400
that pressure. 
Which means that at least my 

727
00:42:58,400 --> 00:43:01,000
learning is that even if I 
suppose that today is the best 

728
00:43:01,000 --> 00:43:03,080
we'll have a get in terms of 
models, which is probably not 

729
00:43:03,080 --> 00:43:04,840
true. 
But even if I make that that 

730
00:43:04,840 --> 00:43:07,440
sort of experiments, that 
doesn't mean that in six months 

731
00:43:07,440 --> 00:43:11,240
the tokens will be close to 0. 
Because there is so much build 

732
00:43:11,240 --> 00:43:14,440
up demands And some of and some 
of that demand is might not be 

733
00:43:14,440 --> 00:43:18,120
the best demand because as you 
said, it's easy to to be token 

734
00:43:18,120 --> 00:43:21,480
maxing on some pure AI slot 
bullshit. 

735
00:43:21,480 --> 00:43:26,280
But but this the demand is still
there and people are still ready

736
00:43:26,280 --> 00:43:28,280
to pay for it. 
And so there's a there's kind of

737
00:43:28,280 --> 00:43:31,400
a nozzle pressure in the market 
for around in France that will 

738
00:43:31,400 --> 00:43:33,200
maintain the price of token 
high. 

739
00:43:33,640 --> 00:43:38,080
But but anyway, it's very hard 
to to know. 

740
00:43:38,080 --> 00:43:43,000
There's a there's also the other
thing that you can save that is 

741
00:43:43,000 --> 00:43:46,680
very easy to say that makes 
sense is we're getting in the 

742
00:43:46,680 --> 00:43:50,160
world where does the those 
models on some tasks are PhD 

743
00:43:50,160 --> 00:43:55,080
level, if not better. 
We need the PhD level to fill in

744
00:43:55,560 --> 00:43:57,880
Salesforce. 
No one. 

745
00:43:59,160 --> 00:44:03,720
But the counter argument to that
there is 2 is we need the 

746
00:44:03,720 --> 00:44:09,080
MacBook M2 to do spreadsheets. 
No one but everybody wants 1 

747
00:44:11,400 --> 00:44:15,200
because it's the best thing. 
And the most serious arguments 

748
00:44:15,200 --> 00:44:19,920
is even if the model is APFC 
level, if you apply to task like

749
00:44:19,920 --> 00:44:23,520
feeling self force, you might 
gain 19 of of reliability. 

750
00:44:24,200 --> 00:44:27,400
And whenever you gain 19 of 
reliability, some funky stuff, 

751
00:44:27,400 --> 00:44:31,880
interesting stuff happen. 
You go from I look up the I look

752
00:44:31,880 --> 00:44:36,240
up what's going on. 
I'm checking the work to to I'm 

753
00:44:36,240 --> 00:44:39,320
checking every to This thing is 
fully entirely automated and I 

754
00:44:39,320 --> 00:44:44,240
will never ever in the story of 
humankind feeling a Salesforce 

755
00:44:45,320 --> 00:44:48,000
card ever. 
And so I think that. 

756
00:44:50,360 --> 00:44:52,840
And so I think there's there's 
still an argument for using the 

757
00:44:52,840 --> 00:44:56,720
best frontier model even for 
tasks that seems pretty not so 

758
00:44:56,800 --> 00:45:00,440
PhD level esque because of that 
kind of added nine of 

759
00:45:00,440 --> 00:45:06,960
reliability on the task. 
Well, if the demand right now is

760
00:45:06,960 --> 00:45:11,720
the main constraint on what's 
making this plentiful, like you 

761
00:45:11,720 --> 00:45:18,440
were saying, because 4.8 dropped
recently and I've played around 

762
00:45:18,440 --> 00:45:21,760
with it, I haven't seen that 
explosion of wow, this is so 

763
00:45:21,760 --> 00:45:25,560
amazing and I need to use it all
the time. 

764
00:45:26,320 --> 00:45:29,560
There's a lot of talk about how 
Methos is going to drop and it's

765
00:45:29,560 --> 00:45:32,240
going to be a game changer, 
etcetera, etcetera. 

766
00:45:32,560 --> 00:45:35,880
We'll see you when that happens.
But assuming in this thought 

767
00:45:35,880 --> 00:45:40,440
experiment that the models are 
what they are and we have this 

768
00:45:40,440 --> 00:45:46,640
pent up demand, the question in 
my mind is how long does this 

769
00:45:46,640 --> 00:45:50,800
pent up demand stick around? 
Does it grow over time? 

770
00:45:50,800 --> 00:45:54,960
Does it diminish over time? 
Is it something that stays 

771
00:45:54,960 --> 00:45:57,200
stable? 
And so we just need to keep 

772
00:45:57,200 --> 00:46:02,280
bringing more energy, more 
inference online to service that

773
00:46:02,280 --> 00:46:05,520
demand. 
Well, I think there's a, there's

774
00:46:05,800 --> 00:46:08,840
a everybody, I get presumed 
everybody along the, the, the 

775
00:46:08,920 --> 00:46:14,080
chain, the, the value chain will
be happy to provide more offer, 

776
00:46:14,080 --> 00:46:18,200
right? 
The, the ecstasy companies in 

777
00:46:18,200 --> 00:46:21,880
every countries will be happy to
build more nuclear plants. 

778
00:46:21,920 --> 00:46:26,040
And then the, the cluster 
building companies will be happy

779
00:46:26,040 --> 00:46:30,080
to build new cluster in the 
place of old industrial places 

780
00:46:30,080 --> 00:46:35,600
that are not used anymore. 
And the GPU providers mostly 1 

781
00:46:35,600 --> 00:46:38,520
today, but probably more 
competition tomorrow, we'll be 

782
00:46:38,520 --> 00:46:42,360
happy to provide the GPUs there.
And so I mean, it will have 

783
00:46:42,400 --> 00:46:46,280
economically, we'll have to 
accelerate and, and we know 

784
00:46:46,280 --> 00:46:51,840
where the I mean, it's with 
those type of technology, it's 

785
00:46:51,840 --> 00:46:55,560
always, it always equilibrates 
to towards commensuration. 

786
00:46:55,560 --> 00:47:00,680
It's like your, it's like your, 
your, your mobile phone, your 

787
00:47:00,680 --> 00:47:06,200
mobile phone service service. 
You, you pay 10 bucks, 20 bucks 

788
00:47:06,200 --> 00:47:09,240
a month or maybe a bit more in 
the USI don't know. 

789
00:47:09,240 --> 00:47:11,960
But it's, it's so cheap compared
to the value it provides you. 

790
00:47:13,240 --> 00:47:15,440
It's just because it's in fully 
commuted eyes and it's 

791
00:47:15,440 --> 00:47:17,000
equilibrates to cost plus 
margin. 

792
00:47:20,600 --> 00:47:22,440
It is fascinating to think about
that. 

793
00:47:22,440 --> 00:47:28,120
And right now though, where 
we're at living today, I imagine

794
00:47:28,120 --> 00:47:31,440
that's another piece that makes 
the ground underneath you very 

795
00:47:31,440 --> 00:47:35,880
shaky. 
And so trying to build in going 

796
00:47:35,880 --> 00:47:40,520
back to this metaphor of the fog
of AI, knowing that the costs 

797
00:47:40,520 --> 00:47:43,680
aren't really stable, locked in,
reliable. 

798
00:47:44,480 --> 00:47:49,240
And so then you're going out 
there and talking to users and 

799
00:47:49,240 --> 00:47:53,040
you have your own product 
pricing models and all of that 

800
00:47:53,040 --> 00:47:56,440
fun stuff. 
And then the users are already 

801
00:47:56,600 --> 00:48:00,080
having their own thoughts about 
the tokens that they're spending

802
00:48:00,080 --> 00:48:04,280
and how they're spending it. 
That makes for some interesting 

803
00:48:04,280 --> 00:48:06,600
conversations. 
Yeah, definitely. 

804
00:48:06,600 --> 00:48:10,160
I think these are it's the fog 
of pricing. 

805
00:48:10,160 --> 00:48:16,360
I guess this time around we 
interestingly add dust. 

806
00:48:16,360 --> 00:48:22,760
We so we started with the with 
the flat price and the, the 

807
00:48:22,760 --> 00:48:27,080
thinking was, was the following.
The thinking was you want a flat

808
00:48:27,080 --> 00:48:29,360
price because you want to 
encourage usage and value 

809
00:48:29,360 --> 00:48:31,440
creation. 
So you want kind of a Chinese 

810
00:48:31,440 --> 00:48:36,000
Buffet type of pricing because 
they you pay once, can use as 

811
00:48:36,000 --> 00:48:40,520
much as you want and let's go 
for usage, let's go for for for 

812
00:48:40,520 --> 00:48:44,840
value creation. 
And it was also built on the 

813
00:48:44,840 --> 00:48:48,840
assumption that eventually, even
if the eventually was very 

814
00:48:48,840 --> 00:48:52,280
loosely defined, eventually cost
would go down. 

815
00:48:52,720 --> 00:48:58,200
And as it happens with the 
emergence of the models, I mean 

816
00:48:58,200 --> 00:49:04,480
with the awakening of the models
at the end of last year, it is 

817
00:49:04,480 --> 00:49:07,440
not sustainable anymore for us. 
And so we have to refactor our 

818
00:49:07,440 --> 00:49:10,480
entire pricing to move to what 
is the industry send out today. 

819
00:49:10,480 --> 00:49:13,480
And I think it's the only 
sustainable pricing unless you 

820
00:49:13,480 --> 00:49:18,760
are willing to go extremely 
negative margins. 

821
00:49:19,440 --> 00:49:23,360
You, you, you, you have to move 
to a crate based pricing because

822
00:49:23,360 --> 00:49:25,360
you don't control those costs. 
We don't know what's going to be

823
00:49:25,360 --> 00:49:28,760
made of tomorrow. 
We don't know what's maybe mitos

824
00:49:28,760 --> 00:49:33,680
will be really a banger and it 
will be maybe really expensive 

825
00:49:33,680 --> 00:49:35,720
at the same time. 
And that's great. 

826
00:49:35,800 --> 00:49:38,360
That's great if that's the case.
That's great if you are in a 

827
00:49:38,360 --> 00:49:41,880
crate based pricing wall because
you're like people just go for 

828
00:49:41,880 --> 00:49:43,040
it. 
Use meadows if you want. 

829
00:49:43,040 --> 00:49:45,640
It's an awesome model. 
It's going to cost you a lot. 

830
00:49:45,920 --> 00:49:50,480
But we I mean, it's fine if you 
are on a kind of a flat price 

831
00:49:50,960 --> 00:49:56,680
like a typical SAS pricing you 
just you just like if a model 

832
00:49:56,680 --> 00:50:00,040
drops, awesome, very powerful, 
but super expensive, just can't 

833
00:50:00,040 --> 00:50:02,440
serve it. 
Just can't serve it because it's

834
00:50:02,800 --> 00:50:04,320
the the the economics don't work
anymore. 

835
00:50:05,280 --> 00:50:07,560
So. 
Oh, sorry, go. 

836
00:50:07,560 --> 00:50:10,440
Ahead, yeah, no, sorry. 
So I think the uncertainty and 

837
00:50:10,440 --> 00:50:15,720
kind of the time horizon for the
evolution of pricing makes it a 

838
00:50:15,760 --> 00:50:18,400
pretty much an unstable position
to not have a crate based 

839
00:50:18,400 --> 00:50:19,640
pricing in the market today I 
think. 

840
00:50:21,200 --> 00:50:25,360
Yeah, it is very dangerous to go
per seat. 

841
00:50:25,360 --> 00:50:30,120
And then especially if you're 
abstracting away the cost of the

842
00:50:30,120 --> 00:50:33,400
providers, there's some creative
things you can do like you were 

843
00:50:33,400 --> 00:50:38,360
talking about earlier of using 
smaller models, but that at the 

844
00:50:38,360 --> 00:50:44,160
end of the day is not going to 
work, especially if folks are 

845
00:50:44,200 --> 00:50:46,600
wanting these the best. 
Yeah, exactly. 

846
00:50:46,600 --> 00:50:51,200
Models all the time. 
For, I mean, maybe there is a 

847
00:50:51,200 --> 00:50:55,080
set of task where I'm happy to 
use this model though, but for 

848
00:50:55,080 --> 00:50:58,680
most of the stuff I do with 
agents I I just want the best, 

849
00:50:58,960 --> 00:51:00,960
yeah? 
Well, and you don't. 

850
00:51:01,520 --> 00:51:05,360
It almost takes more time to 
figure out where to insert the 

851
00:51:05,360 --> 00:51:10,200
small model, which slows you 
down and you just, it's like, 

852
00:51:10,200 --> 00:51:15,240
well, I know I could save a few 
cents here, but let's try and 

853
00:51:15,680 --> 00:51:21,280
use it because it's going to be 
a lot easier for me then if you 

854
00:51:21,280 --> 00:51:24,760
set up and, and that's one thing
that has changed a lot recently 

855
00:51:24,760 --> 00:51:30,320
too, is how we don't necessarily
set up by hand each step of this

856
00:51:30,360 --> 00:51:32,960
graph. 
We let the agents, we kind of 

857
00:51:32,960 --> 00:51:35,560
throw the problem at the agents 
and let them figure it out. 

858
00:51:36,000 --> 00:51:41,880
And so if the agents have the 
ability to choose the smaller 

859
00:51:41,880 --> 00:51:44,960
models or kick off tasks with 
smaller models, that's great. 

860
00:51:45,520 --> 00:51:49,640
I'm not going to go in there and
like use a smaller model for 

861
00:51:49,640 --> 00:51:52,920
these specific tasks when you 
are going and doing them in the 

862
00:51:52,920 --> 00:51:55,840
plan. 
You know, maybe there are people

863
00:51:55,840 --> 00:52:00,760
that do and those folks, wow, 
hats off to you for now. 

864
00:52:00,760 --> 00:52:04,160
I imagine you could. 
Optimize micro micromanagement 

865
00:52:04,280 --> 00:52:07,200
of and. 
You probably could get more lift

866
00:52:07,200 --> 00:52:11,480
or more cost savings just by 
optimizing certain context 

867
00:52:11,480 --> 00:52:18,080
window tricks or the ability to 
make sure that your agents 

868
00:52:18,080 --> 00:52:21,160
aren't constantly ingesting 
files. 

869
00:52:21,400 --> 00:52:24,360
Maybe it's like cashing tricks 
or things like that that are 

870
00:52:24,360 --> 00:52:29,640
going to be much, I'm not going 
to say like much cheaper, but 

871
00:52:29,640 --> 00:52:32,480
they are going to potentially 
give you more bang for your 

872
00:52:32,480 --> 00:52:36,360
buck. 
So there's there's all that fun 

873
00:52:36,360 --> 00:52:39,240
stuff. 
Now I have one thing that has 

874
00:52:39,240 --> 00:52:41,320
been going through my head that 
I've been wanting to ask you, 

875
00:52:41,320 --> 00:52:45,080
and I'm sure you hear it every 
day that you talk to customers 

876
00:52:45,800 --> 00:52:50,680
and it's well, yeah, cool dust. 
I see this vision, but can't I 

877
00:52:50,680 --> 00:52:53,000
just do that with Co work? 
As we were talking about 

878
00:52:53,000 --> 00:52:54,760
earlier, Co work is having its 
moment. 

879
00:52:55,480 --> 00:52:59,400
What is the unlock on this? 
Like single player to 

880
00:52:59,400 --> 00:53:02,880
multiplayer? 
I can share my slides with 

881
00:53:02,880 --> 00:53:06,740
people or I can share my 
sessions with people from Co 

882
00:53:06,740 --> 00:53:09,840
work. 
So where do I go with dust? 

883
00:53:10,600 --> 00:53:13,160
Yeah, I think there's, so 
there's, there's three pillars 

884
00:53:13,160 --> 00:53:14,720
to our different situation with 
cohort. 

885
00:53:15,600 --> 00:53:18,560
It's so first on the on, on that
multiplayer, there is some use 

886
00:53:18,560 --> 00:53:21,560
cases that you can explain for, 
you know, as I did for the team 

887
00:53:21,560 --> 00:53:24,680
weekly where I think you can 
paint a picture with you. 

888
00:53:24,720 --> 00:53:27,880
You it's very hard to do with 
cohort that kind of a 

889
00:53:28,160 --> 00:53:32,280
collaboration around a unit of 
work that needs multiple people.

890
00:53:32,840 --> 00:53:36,360
The, I mean, you can obviously 
do it, but it creates more, more

891
00:53:36,360 --> 00:53:39,040
kind of manual work and it's 
more, it's less integrated. 

892
00:53:39,720 --> 00:53:43,160
I think there's in terms of 
enterprise, there's the 

893
00:53:43,880 --> 00:53:46,240
governance part, which is very 
important. 

894
00:53:47,200 --> 00:53:52,040
Having the ability to, we, we 
try to do a better job than code

895
00:53:52,440 --> 00:53:55,680
on that. 
Then be able to distribute in a 

896
00:53:55,680 --> 00:53:59,600
control way the different MCP 
servers, different tools, the 

897
00:53:59,600 --> 00:54:03,680
different skills limits the 
access to some data to some 

898
00:54:03,680 --> 00:54:06,760
users have that things 
transitively managed so that's 

899
00:54:06,760 --> 00:54:09,560
people can discover the the 
agents and the skills they can 

900
00:54:09,560 --> 00:54:11,680
use. 
But in a way that's is aligned 

901
00:54:11,680 --> 00:54:13,720
with other what are the 
restrictions of the workspace? 

902
00:54:14,200 --> 00:54:18,040
I think cohort with all the 
things turning locally, the 

903
00:54:18,040 --> 00:54:22,080
skills being shared through kind
of out of out of bands to some 

904
00:54:22,080 --> 00:54:27,240
extent as of today is a bit a 
bit more of a of a far West when

905
00:54:27,240 --> 00:54:31,400
you want to have kind of a 
slightly more governed 

906
00:54:31,480 --> 00:54:33,320
deployment of AI with a new 
company. 

907
00:54:33,680 --> 00:54:35,960
So that will talk to different 
type of company obviously. 

908
00:54:36,400 --> 00:54:38,720
And finally, the third pillar, 
which is a pretty obvious 1 is 

909
00:54:38,720 --> 00:54:41,840
like you don't want to get stuck
in with one model provider. 

910
00:54:43,640 --> 00:54:45,440
Especially as we're talking 
about tokenomics. 

911
00:54:45,880 --> 00:54:47,600
Exactly. 
You don't want to be stuck with 

912
00:54:47,600 --> 00:54:50,680
one token provider because the 
next best model you don't know 

913
00:54:50,680 --> 00:54:52,880
where it's going to know where 
it's going from and you surely 

914
00:54:52,880 --> 00:54:56,600
want to be have access to it. 
So I think it's kind of a mix of

915
00:54:56,600 --> 00:54:59,160
those three. 
We try to be very innovative on 

916
00:54:59,160 --> 00:55:04,480
the products in terms of 
multiplayer AI and that's a 

917
00:55:04,480 --> 00:55:07,480
differentiation. 
That's something that we hope 

918
00:55:07,480 --> 00:55:10,200
will awaken and become stronger 
and stronger as we go. 

919
00:55:11,040 --> 00:55:14,640
We want to be better at the 
governance and kind of the 

920
00:55:14,760 --> 00:55:19,080
enterprise readiness as much as 
possible in at the product 

921
00:55:19,080 --> 00:55:21,800
layer. 
Obviously we're not talking 

922
00:55:21,800 --> 00:55:25,040
about we're not talking about FD
ES trying to implement stuff in 

923
00:55:25,040 --> 00:55:27,360
your company, but really as a 
product, be more enterprise 

924
00:55:27,360 --> 00:55:29,360
ready. 
And then the sub pillar is that 

925
00:55:29,360 --> 00:55:33,440
kind of a, is that kind of a, is
that kind of a a freedom of of 

926
00:55:33,440 --> 00:55:37,280
freedom of choosing your model 
and being able to test open 

927
00:55:37,280 --> 00:55:40,440
source models and keep your 
finger on that and have all of 

928
00:55:40,440 --> 00:55:42,760
that in one central place. 
So I think it's really that's 

929
00:55:42,760 --> 00:55:46,800
kind of a those three factors 
that makes us still a, a 

930
00:55:46,960 --> 00:55:51,400
compelling option in the time 
where everybody, everybody talks

931
00:55:51,400 --> 00:55:53,680
about cohort. 
But I'm I mean, we've lived 

932
00:55:53,680 --> 00:55:56,280
through other hype micro hype 
cycles. 

933
00:55:56,960 --> 00:55:59,800
We've had obviously tragedy. 
We've had we've rolled enough. 

934
00:55:59,800 --> 00:56:02,400
We've we've rolled enough for 
for having lived as a company 

935
00:56:02,400 --> 00:56:07,560
the emergence of tragedy Imagine
and we've lived through the 

936
00:56:07,560 --> 00:56:11,160
glean moments like it was all 
about glean 2 years ago. 

937
00:56:11,160 --> 00:56:12,800
Glean what? 
Happened Glean. 

938
00:56:12,800 --> 00:56:15,440
That's true, Glean. 
Well, I think it's still a very 

939
00:56:15,440 --> 00:56:19,880
powerful enterprise search and 
AI answer products and I think 

940
00:56:20,000 --> 00:56:21,640
they're doing actually really 
well. 

941
00:56:21,640 --> 00:56:24,160
I think they're just they're 
just I've seen, I don't know if 

942
00:56:24,160 --> 00:56:26,920
it's official or not, but I've 
seen on on newspapers, I don't 

943
00:56:26,920 --> 00:56:29,240
know if it's true number, but 
they've reached 200 me in 

944
00:56:29,720 --> 00:56:32,000
revenue. 
It's it's a very healthy growing

945
00:56:32,000 --> 00:56:35,400
company at that. 
At that time, it was the IT was 

946
00:56:35,400 --> 00:56:36,240
really darling. 
Yeah. 

947
00:56:36,240 --> 00:56:40,120
The be all to end all. 
I remember those days in 2023 or

948
00:56:40,120 --> 00:56:42,040
2024. 
Exactly, exactly. 

949
00:56:42,920 --> 00:56:45,440
So there'll be there'll be other
hype cycles, that's fine. 

950
00:56:45,440 --> 00:56:48,520
I think we what we want to 
provide our users is really a 

951
00:56:48,520 --> 00:56:53,600
platform where there is there is
really building and and helping 

952
00:56:53,600 --> 00:56:56,720
them being at the forefront of 
what can be done with with 

953
00:56:56,720 --> 00:57:01,240
agents. 
So at the risk of jumping all 

954
00:57:01,240 --> 00:57:08,720
over the place, Yep, file 
systems, Yep, you, we got into 

955
00:57:08,800 --> 00:57:12,760
this earlier. 
Are you in a place where you 

956
00:57:12,760 --> 00:57:18,000
feel like file systems can be 
updated for more agentic work 

957
00:57:19,040 --> 00:57:25,080
And or, or add to this longer 
question, cause the way the 

958
00:57:25,080 --> 00:57:28,920
agents work and the way that 
they will almost fan out and do 

959
00:57:28,920 --> 00:57:29,800
things. 
Yep. 

960
00:57:30,000 --> 00:57:34,600
And do things very compute 
intensive things and then just 

961
00:57:34,600 --> 00:57:38,480
hang out for a minute and be 
like, hey, a human, I want I 

962
00:57:38,480 --> 00:57:41,120
want your feedback. 
And maybe it's a minute, maybe 

963
00:57:41,120 --> 00:57:43,680
it is an eternity in computer 
time. 

964
00:57:44,560 --> 00:57:48,440
Are there other areas where you 
feel like the current 

965
00:57:48,440 --> 00:57:55,880
infrastructure that you have to 
build dust, you need new tools? 

966
00:57:57,720 --> 00:58:02,400
Well, on the file system, it's 
true that we can predict a 

967
00:58:02,400 --> 00:58:06,120
moment where. 
So right now we discuss about 

968
00:58:06,120 --> 00:58:07,480
it. 
We have a file system mounted on

969
00:58:07,480 --> 00:58:09,400
the pod, we have a file system 
mounted on the sandbox. 

970
00:58:09,840 --> 00:58:11,360
Some of the file system in the 
pod are shared. 

971
00:58:11,360 --> 00:58:13,520
You can have multiple agents 
working with that file system. 

972
00:58:13,560 --> 00:58:17,760
As of today, it's mostly still 
human, collaborative driven. 

973
00:58:17,760 --> 00:58:21,920
And so the kind of concurrency 
on that file system remains 

974
00:58:22,080 --> 00:58:25,960
within typical human bounds. 
Yeah, but you could totally 

975
00:58:25,960 --> 00:58:29,040
imagine like an army of agents 
cooperating on a project. 

976
00:58:29,040 --> 00:58:32,880
And then the file system becomes
the wrong abstraction. 

977
00:58:34,240 --> 00:58:36,200
Oh, I mean, it's not necessarily
the wrong abstraction, but 

978
00:58:36,200 --> 00:58:38,720
there's going to be there's 
going to be a concurrency 

979
00:58:38,720 --> 00:58:40,960
pressure on it. 
And it doesn't have any 

980
00:58:40,960 --> 00:58:44,560
primitive for that. 
I mean it does file locking, but

981
00:58:46,440 --> 00:58:49,880
who uses file looking? 
Agents will. 

982
00:58:50,240 --> 00:58:56,680
Agents will, agents might, but I
think but to some extent this, 

983
00:58:56,920 --> 00:59:00,280
to me this has been sold as 
there's a couple of companies 

984
00:59:00,520 --> 00:59:03,480
that are building kind of a git 
infrastructure. 

985
00:59:03,480 --> 00:59:08,160
Because if you want to scale, 
providing GitHub repository to 

986
00:59:08,200 --> 00:59:11,760
agents using GitHub is not very 
convenient because there are 

987
00:59:11,760 --> 00:59:14,480
rate limits that are meant for 
humans on the platform. 

988
00:59:14,880 --> 00:59:16,920
And so there's a bunch of 
companies, they're pretty 

989
00:59:16,920 --> 00:59:19,120
exciting that that builds Git 
infra. 

990
00:59:20,120 --> 00:59:25,040
And Git infra seems to be a 
fair, fair solution to the to 

991
00:59:25,040 --> 00:59:29,120
the problem because at least you
have that resolution first 

992
00:59:29,120 --> 00:59:31,400
resolution step when you merge 
back to Maine. 

993
00:59:31,400 --> 00:59:36,320
That makes kind of some sense. 
But at the same time, yeah, the 

994
00:59:37,240 --> 00:59:40,920
but I think the as was as was 
chatting. 

995
00:59:41,480 --> 00:59:45,320
I think one, one step before 
that, that excites me a lot is 

996
00:59:45,320 --> 00:59:49,760
the, is the kind of a stateful 
sandbox, the sandbox that stays 

997
00:59:49,760 --> 00:59:54,600
live because I'm super excited 
about kind of a hyper 

998
00:59:54,600 --> 00:59:58,800
specialized business apps. 
Many of our users are already 

999
00:59:58,800 --> 01:00:01,400
doing that with those frames 
that we mentioned. 

1000
01:00:01,400 --> 01:00:04,280
But those frames, they're not 
backed, they're not backed by a 

1001
01:00:04,280 --> 01:00:07,440
particular sandbox. 
So they, they, they, they kind 

1002
01:00:07,440 --> 01:00:11,520
of, they kind of they're 
snapshots, which means that if 

1003
01:00:11,520 --> 01:00:14,640
you want to update it, you need 
to ask the agent to go fetch the

1004
01:00:14,680 --> 01:00:16,760
data again and go update the 
frame. 

1005
01:00:16,760 --> 01:00:19,160
So what you really want is you 
want those frames to be backed 

1006
01:00:19,160 --> 01:00:22,800
by a sandbox that stays around. 
Because then you can have inside

1007
01:00:22,800 --> 01:00:25,120
of a sandbox, you can have a 
SQLite database and you can 

1008
01:00:25,120 --> 01:00:28,560
rebuild the CRM and ERP. 
You can rebuild the whatever 

1009
01:00:29,440 --> 01:00:34,440
things you need to do your job. 
And the frame is the UI for the 

1010
01:00:34,440 --> 01:00:37,000
human. 
The sandbox is accessible by 

1011
01:00:37,000 --> 01:00:40,640
agents and so you have an API 
for you have a way to walk with 

1012
01:00:40,640 --> 01:00:43,800
agents or directly as a human. 
Every human can create a 

1013
01:00:43,800 --> 01:00:47,600
different view of the the app. 
If let's say we are talking 

1014
01:00:47,600 --> 01:00:53,720
about feedback system for the 
product where you automatically 

1015
01:00:53,720 --> 01:00:56,600
routes to different owners. 
And that's something that you 

1016
01:00:56,600 --> 01:00:59,040
cannot manage to do well in 
GitHub or linear. 

1017
01:00:59,040 --> 01:01:02,080
And so you build your own thing.
Maybe the manager will want a 

1018
01:01:02,080 --> 01:01:04,680
different view than the IC, 
which will want a different view

1019
01:01:04,680 --> 01:01:06,960
than the person that report the 
feedback. 

1020
01:01:07,400 --> 01:01:10,760
And that's having that habit is 
extremely exciting. 

1021
01:01:11,040 --> 01:01:15,000
And in that world, the kind of a
de facto way to build that I 

1022
01:01:15,000 --> 01:01:19,320
think that is very easy is to is
to do a shared nothing type of 

1023
01:01:19,320 --> 01:01:23,800
approach where the sandbox is 
the state and you just, you just

1024
01:01:23,800 --> 01:01:27,440
let the agents do whatever they 
want, use whatever technology 

1025
01:01:27,440 --> 01:01:29,400
they want and you just Snapchat 
the sandbox. 

1026
01:01:29,400 --> 01:01:33,040
We can go back in time easily. 
And there is nothing more to it.

1027
01:01:33,040 --> 01:01:35,760
There's no external database, 
there's no external Git 

1028
01:01:35,760 --> 01:01:39,280
repository. 
I think that will that's kind of

1029
01:01:39,280 --> 01:01:42,080
a very nice convenient way to 
package those kind of things for

1030
01:01:42,080 --> 01:01:44,520
agents. 
The only problem is like if you 

1031
01:01:44,520 --> 01:01:47,520
imagine that you have hundreds 
of agents collaborating on the 

1032
01:01:47,520 --> 01:01:49,400
shut sandbox, then it starts 
failing. 

1033
01:01:49,400 --> 01:01:53,480
And maybe then you need the 
codes to be managed by GitHub. 

1034
01:01:53,480 --> 01:01:56,640
Maybe you need the the data to 
be managed by your hosted 

1035
01:01:56,640 --> 01:02:01,120
database, which so we'll see. 
So it works until you get to a 

1036
01:02:01,120 --> 01:02:05,640
certain scale, which it doesn't 
feel like a lot of these 

1037
01:02:05,640 --> 01:02:11,480
workloads are at yet. 
Oh yeah, I mean it's so the 

1038
01:02:11,480 --> 01:02:18,000
moment we're so imagine you are 
you are a small SMB in whatever 

1039
01:02:18,000 --> 01:02:22,920
countryside of of Germany and 
you and you do cookies and 

1040
01:02:22,920 --> 01:02:27,960
you're paying your you're paying
your ERP20K or 30 KA year to use

1041
01:02:27,960 --> 01:02:31,080
three features of an ERP that 
has 10,000 features. 

1042
01:02:31,800 --> 01:02:33,560
That's that's the state of SAS 
today. 

1043
01:02:33,840 --> 01:02:38,560
And and so you may want to 
rebuild that, that that that 

1044
01:02:38,640 --> 01:02:42,040
process, which is get me the 
dog, get me the missing. 

1045
01:02:42,040 --> 01:02:43,880
And then here are the the 
comments. 

1046
01:02:43,880 --> 01:02:47,880
And it's pretty easy one inside 
of one of those kind of a 

1047
01:02:48,160 --> 01:02:53,320
specialized app. 
And so there indeed, there is no

1048
01:02:53,640 --> 01:02:56,840
that's going to be a few people 
interacting with it. 

1049
01:02:56,840 --> 01:02:59,160
Maybe it's for agents, maybe 
directly through UIS. 

1050
01:02:59,160 --> 01:03:01,840
And I think having that that's 
that that shared nothing 

1051
01:03:01,840 --> 01:03:04,440
approach really works. 
The moment where it would fail 

1052
01:03:04,440 --> 01:03:08,960
is where you start growing and 
you have thousands of humans. 

1053
01:03:08,960 --> 01:03:11,440
If they interact through the UI 
that would the model would still

1054
01:03:11,440 --> 01:03:13,800
work fine. 
The model would fail is like, 

1055
01:03:13,800 --> 01:03:18,760
you know, 10,000 people and and 
although you're 10,000, you're 

1056
01:03:18,760 --> 01:03:22,640
10 times bigger. 
And no, you need to make that 

1057
01:03:22,640 --> 01:03:25,720
system evolve and you have 
agents working with it and 

1058
01:03:25,720 --> 01:03:28,880
changing the codes at the speed 
of hundreds of agents at the 

1059
01:03:28,880 --> 01:03:31,680
same time. 
We can imagine that, Paul, but 

1060
01:03:31,680 --> 01:03:32,800
that's that's where it's going 
to fail. 

1061
01:03:32,800 --> 01:03:36,840
But it's so far away that's it 
seems like a fine to not over in

1062
01:03:36,840 --> 01:03:41,560
sure for it. 
And even the current state of a 

1063
01:03:41,880 --> 01:03:45,600
large company, you don't 
necessarily have the majority of

1064
01:03:45,600 --> 01:03:51,400
your projects being hit hundreds
of times a minute by agents. 

1065
01:03:52,360 --> 01:03:57,560
So it's almost like I see a 
208020 principle here where you 

1066
01:03:57,560 --> 01:04:02,280
could have that working for a 
lot of your different projects, 

1067
01:04:02,880 --> 01:04:09,800
but then those projects that 
really need scale and they need 

1068
01:04:09,800 --> 01:04:17,200
that type of limiting or 
unlimitedness to it, you would 

1069
01:04:17,200 --> 01:04:20,560
take those or you would 
architect those in a different 

1070
01:04:20,560 --> 01:04:28,360
way. 
Exactly, so that is stateful 

1071
01:04:28,360 --> 01:04:32,520
sandboxes. 
Are there any other hang ups 

1072
01:04:32,520 --> 01:04:35,240
that we might encounter if we 
tried to do them? 

1073
01:04:35,240 --> 01:04:37,080
It feels like it could get 
expensive. 

1074
01:04:37,080 --> 01:04:41,080
If I constantly have a sandbox 
running and I only use it once a

1075
01:04:41,080 --> 01:04:45,040
week, it could be wasteful. 
Oh, that, that, that, that we 

1076
01:04:45,040 --> 01:04:47,800
have really great companies that
gives you the nice privacy 

1077
01:04:47,800 --> 01:04:49,760
there. 
You you just shut them down. 

1078
01:04:49,880 --> 01:04:54,000
The time is boot up time is 100 
milliseconds. 

1079
01:04:54,600 --> 01:04:56,040
You don't have if, if it's sure 
nothing. 

1080
01:04:56,040 --> 01:04:57,960
The beauty of that is that you 
don't have to care about where 

1081
01:04:57,960 --> 01:04:59,880
is the DB. 
Do you have to shut down the DB?

1082
01:04:59,880 --> 01:05:02,120
No, everything is in the sandbox
SQLA database. 

1083
01:05:02,120 --> 01:05:04,640
It's a file. 
It's 1 megabytes to do Most of 

1084
01:05:04,640 --> 01:05:07,080
any companies would do with 
anything serious. 

1085
01:05:07,640 --> 01:05:09,160
It just works. 
I think it's what I really see 

1086
01:05:09,160 --> 01:05:11,640
as the prolongation of the Excel
spreadsheets. 

1087
01:05:12,560 --> 01:05:16,240
Today we have the Excel 
spreadsheets, No man's Lands and

1088
01:05:16,240 --> 01:05:19,600
SAS. 
And it's kind of the, it's kind 

1089
01:05:19,600 --> 01:05:23,240
of a way to it's, it's, it's, 
it's a capture of a lot more 

1090
01:05:23,240 --> 01:05:26,200
than it can do with a 
spreadsheet that you were forced

1091
01:05:26,200 --> 01:05:28,640
to do with SAS. 
And so that's funny because if 

1092
01:05:28,640 --> 01:05:31,720
you really believe in that 
vision, it feels like Wall 

1093
01:05:31,720 --> 01:05:38,160
Street was right way in advance 
for dangling the SAS businesses.

1094
01:05:39,240 --> 01:05:41,880
And that's, it's pretty rare to 
think that Wall Street was 

1095
01:05:41,920 --> 01:05:46,320
really, really right in advance.
But but that's yeah, that's the,

1096
01:05:47,560 --> 01:05:50,120
the fun, the fun thoughts. 
But yeah, I think there's a 

1097
01:05:50,160 --> 01:05:54,160
massive opportunity to have 
anybody create the, the app they

1098
01:05:54,160 --> 01:05:58,040
need to do their job that is 
equally accessible. 

1099
01:05:58,040 --> 01:06:00,560
The state is accessible by 
agents and the, the UI is 

1100
01:06:00,560 --> 01:06:02,920
actually can be create. 
They can be customized to every 

1101
01:06:02,920 --> 01:06:05,160
human basically. 
That's that's something very 

1102
01:06:05,160 --> 01:06:09,200
exciting over there. 
Yeah, especially if you're 

1103
01:06:09,240 --> 01:06:16,120
interacting with this through 
the chat ability or I, I guess 

1104
01:06:16,120 --> 01:06:19,400
where my mind gets caught up and
where I've never fully bought 

1105
01:06:19,400 --> 01:06:23,440
into this generative UI 
principle, although I'm coming 

1106
01:06:23,440 --> 01:06:26,200
around to it a little more. 
The more and more that I use 

1107
01:06:27,000 --> 01:06:31,520
things like the HTML that the 
agents will create on the fly. 

1108
01:06:32,640 --> 01:06:35,920
It's like this is, this is 
actually kind of that idea of 

1109
01:06:35,920 --> 01:06:41,360
generative UI. 
But what I would get hung up on 

1110
01:06:41,360 --> 01:06:46,600
is that a lot of times we have 
learned and we are comfortable 

1111
01:06:46,600 --> 01:06:50,440
with certain tools because we 
know those five things that we 

1112
01:06:50,440 --> 01:06:52,400
need to do. 
We know how to do them and where

1113
01:06:52,400 --> 01:06:55,640
to do them. 
If we now are being asked to 

1114
01:06:56,000 --> 01:07:00,640
create our own tool that does 
those five things and we have to

1115
01:07:00,640 --> 01:07:04,080
spend the time creating it, even
if it is going to give us a lot 

1116
01:07:04,080 --> 01:07:07,320
of success on the tail end 
because it's going to be much 

1117
01:07:07,320 --> 01:07:09,000
faster. 
It's going to be more custom. 

1118
01:07:09,000 --> 01:07:10,480
We don't have to spend as much 
money. 

1119
01:07:10,960 --> 01:07:15,160
It's a pretty hard ask to get 
get it set up. 

1120
01:07:16,440 --> 01:07:19,960
Yeah, but it's going to be as 
with any content economy ever, 

1121
01:07:20,000 --> 01:07:25,720
it's a 1.1% producer, 99% 
consumer, which is true of of of

1122
01:07:25,800 --> 01:07:30,120
spreadsheets as well. 
There are people that knows that

1123
01:07:30,120 --> 01:07:32,240
know how to do a great 
spreadsheets and they're not 

1124
01:07:32,320 --> 01:07:35,400
that many of them and they are 
the ones building a 

1125
01:07:35,400 --> 01:07:37,200
spreadsheets. 
And then there are the rest of 

1126
01:07:37,200 --> 01:07:40,720
us consuming, and there's the 
rest of us consuming consuming 

1127
01:07:40,720 --> 01:07:45,800
them. 
So you're thinking there will be

1128
01:07:45,800 --> 01:07:49,560
the builders of these. 
It's just not. 

1129
01:07:49,600 --> 01:07:51,960
We're not all going to be 
expected to build it. 

1130
01:07:52,800 --> 01:07:55,680
There's going to be the specific
builders and they're going to, I

1131
01:07:55,680 --> 01:07:58,480
like that idea of extending the 
spreadsheet. 

1132
01:07:58,920 --> 01:08:03,400
You're now going to get this 
micro SAS that is very specific 

1133
01:08:03,400 --> 01:08:07,520
to your business. 
Yep, and anybody can build if 

1134
01:08:07,520 --> 01:08:10,120
they want. 
It's just that we know how it 

1135
01:08:10,200 --> 01:08:12,760
wet lens. 
It's like some will build, some 

1136
01:08:12,760 --> 01:08:14,960
will consume, and. 
Yeah. 

1137
01:08:15,280 --> 01:08:21,920
The the one place that I think 
that is for sure having it's 

1138
01:08:22,760 --> 01:08:28,880
questions is the no code low 
code type of tools because that 

1139
01:08:28,880 --> 01:08:33,080
is 100% in the crosshairs of 
what we're talking about. 

1140
01:08:33,520 --> 01:08:39,680
However, I did just see a tweet 
of somebody laughing, saying I 

1141
01:08:39,680 --> 01:08:46,880
am constantly surprised at how 
folks will set up with their 

1142
01:08:46,920 --> 01:08:51,760
agents a whole workflow to 
automate something that costs 

1143
01:08:52,680 --> 01:08:57,760
like $0.90 to run, when with 
Zapier you could do it for five 

1144
01:08:57,760 --> 01:08:58,760
cents, yeah. 
Yeah. 

1145
01:09:00,040 --> 01:09:03,279
So it goes back to that thing of
like what it costs versus 

1146
01:09:03,279 --> 01:09:06,479
convenience type of thing. 
And it is much more convenient 

1147
01:09:06,560 --> 01:09:09,319
to explain what we want, 
especially if you're using some 

1148
01:09:09,359 --> 01:09:13,680
dictation tool, you just mind 
dump the agent organizes it and 

1149
01:09:13,680 --> 01:09:16,359
says, all right, let's try it. 
You test it, it works, it looks 

1150
01:09:16,359 --> 01:09:17,160
good. 
Ship it. 

1151
01:09:17,920 --> 01:09:21,680
Exactly, we'll see. 
It's hard to it's hard to 

1152
01:09:21,680 --> 01:09:24,399
predict where it's going to 
land, but it's surely there's 

1153
01:09:24,399 --> 01:09:27,720
truly a ton of ton of exciting 
stuff to explore. 

1154
01:09:27,720 --> 01:09:32,240
Agreed that the kind of on the 
fly dynamic UI, it's hard to see

1155
01:09:32,240 --> 01:09:34,720
it fly right now just because of
the latency. 

1156
01:09:34,720 --> 01:09:37,120
But here it's kind of slightly 
more. 

1157
01:09:37,120 --> 01:09:41,120
I think it's like you build ones
use many times, but you can 

1158
01:09:41,120 --> 01:09:45,760
build very custom and agreed. 
That's all of that is doable 

1159
01:09:45,760 --> 01:09:50,840
with there was many no code 
platforms out there and it's 

1160
01:09:51,200 --> 01:09:56,240
it's very hard to understand why
they why they didn't conquer the

1161
01:09:56,240 --> 01:09:57,400
world. 
Yeah. 

1162
01:09:58,880 --> 01:10:03,920
Well, there there's also the 
other idea of just spitballing 

1163
01:10:03,920 --> 01:10:09,120
with you on product here because
you're taking the approach of 1 

1164
01:10:09,120 --> 01:10:14,720
sandbox that will have different
views that people can drop into 

1165
01:10:14,720 --> 01:10:21,080
and they can look at. 
However, what I wonder and, and 

1166
01:10:21,080 --> 01:10:25,080
I see it as the use case in my 
mind is you have a lot of 

1167
01:10:25,080 --> 01:10:27,080
customer calls. 
Like I'm sure you're on calls 

1168
01:10:27,160 --> 01:10:29,680
all day long talking with 
customers or talking with 

1169
01:10:29,680 --> 01:10:34,400
potential customers. 
And some of the time of that 

1170
01:10:34,400 --> 01:10:38,280
call or the same call can have 
value for a salesperson, it can 

1171
01:10:38,280 --> 01:10:41,360
have value for APM, it can have 
value for your engineering team.

1172
01:10:42,320 --> 01:10:44,960
And so you want to take that 
transcript and you want to 

1173
01:10:44,960 --> 01:10:49,720
display the valuable parts to 
each party that's looking at it.

1174
01:10:50,600 --> 01:10:54,520
And so going back to this 
analogy of, all right, we've got

1175
01:10:54,520 --> 01:10:57,280
data and we want to display it 
to different stakeholders in 

1176
01:10:57,280 --> 01:11:01,680
different ways. 
Wouldn't it be, or is there a 

1177
01:11:01,680 --> 01:11:07,400
world where I can as APM have my
sandbox and everything gets 

1178
01:11:07,400 --> 01:11:12,160
piped into that one place? 
Or maybe that is just like you 

1179
01:11:12,160 --> 01:11:15,880
don't need to do it at the 
sandbox level because you are 

1180
01:11:16,440 --> 01:11:19,240
doing it in all these different 
ways and you have something 

1181
01:11:19,240 --> 01:11:24,040
that's on top of all the 
sandboxes that's coagulating all

1182
01:11:24,040 --> 01:11:26,200
of that data. 
Yeah. 

1183
01:11:26,280 --> 01:11:30,720
I mean some use cases will be 
leverable by just having the 

1184
01:11:30,720 --> 01:11:34,520
data and having agent just just 
process the data. 

1185
01:11:34,760 --> 01:11:36,400
There's a lot of that happening 
in dust today. 

1186
01:11:36,400 --> 01:11:40,680
Like we, we have a, it's a gun 
connector. 

1187
01:11:40,680 --> 01:11:43,560
And so companies that have 
connected to gun, they have all 

1188
01:11:43,560 --> 01:11:46,880
those gun calls that gets 
transcribed and recorded. 

1189
01:11:47,280 --> 01:11:50,160
And that enables interesting use
case where indeed product teams 

1190
01:11:50,160 --> 01:11:54,320
can have launch an agent, do a 
deep dive on all of the customer

1191
01:11:54,320 --> 01:11:57,080
calls for the past months. 
Tell me everything that they 

1192
01:11:57,080 --> 01:11:59,160
said about that feature or that 
thing. 

1193
01:11:59,160 --> 01:12:01,760
And it's also completely 
possible today. 

1194
01:12:04,000 --> 01:12:06,520
That is an exciting that that's 
something that's possible today.

1195
01:12:07,120 --> 01:12:15,040
The kind of a micro SAS view of 
that would be you have a call, 

1196
01:12:15,040 --> 01:12:17,920
you dump this transcript. 
There's a it triggered in in a, 

1197
01:12:18,120 --> 01:12:20,600
in a, in AUI. 
It triggers an agent that will 

1198
01:12:21,480 --> 01:12:24,840
create stuff for the sales team,
create stuff for a proc team, 

1199
01:12:25,360 --> 01:12:29,440
auto assign tags for the product
surfaces that are concerned, 

1200
01:12:29,440 --> 01:12:31,360
auto assign tags for the 
geography. 

1201
01:12:32,040 --> 01:12:36,280
And maybe you have a way to say,
OK, looks good and submit and 

1202
01:12:36,280 --> 01:12:38,480
then it gets stored in different
places. 

1203
01:12:39,040 --> 01:12:42,120
I mean there are two ways of 
tackling the same problems. 

1204
01:12:42,720 --> 01:12:50,480
But I do think that like there 
is many cases, many use cases 

1205
01:12:50,480 --> 01:12:57,880
for kind of ADB backed apps that
will that that could be 

1206
01:12:57,880 --> 01:12:59,960
customized for special way of 
doing work. 

1207
01:13:00,400 --> 01:13:04,800
I have I have many of my own add
dust on that could change a 

1208
01:13:04,800 --> 01:13:06,600
little bit. 
The way we operate, we kind of, 

1209
01:13:06,760 --> 01:13:11,000
we can always kind of, you know,
you, you bend the way you want 

1210
01:13:11,000 --> 01:13:15,880
to operate to the tools that, 
that, that you use, the way you 

1211
01:13:15,880 --> 01:13:21,040
share goals for your week or 
your days at, at dust happens in

1212
01:13:21,040 --> 01:13:24,320
Slack. 
So it has a certain fun factor 

1213
01:13:24,320 --> 01:13:29,080
that is banded compared to the 
ideal state, the way we make 

1214
01:13:29,080 --> 01:13:31,760
decisions. 
We do that on GitHub, which is 

1215
01:13:31,760 --> 01:13:34,640
interestingly a fun place to do 
because it helps kind of a 

1216
01:13:34,640 --> 01:13:37,680
asynchronous collaboration. 
And sometimes it's done on Slack

1217
01:13:37,680 --> 01:13:39,840
and it's bended in very 
different ways depending on the 

1218
01:13:39,840 --> 01:13:43,320
platform you use. 
And so there's so many processes

1219
01:13:43,320 --> 01:13:46,760
and ways to ways to work that we
that are shaped by our, by our 

1220
01:13:46,760 --> 01:13:49,720
tools and the other way around. 
And so that's what I find that's

1221
01:13:49,760 --> 01:13:53,040
pretty exciting. 
So then you would have the 

1222
01:13:53,040 --> 01:13:58,680
creative control to do what you 
want and not be beholden to what

1223
01:13:58,680 --> 01:14:08,320
the tools forced on you. 
I could see that and and the my 

1224
01:14:08,320 --> 01:14:12,040
imagination is going wild as I 
think through what you could do 

1225
01:14:12,360 --> 01:14:17,480
in those instances. 
I like this like database backed

1226
01:14:17,480 --> 01:14:22,760
apps or just a sandbox with a 
database that's stateful is a 

1227
01:14:22,760 --> 01:14:27,920
very fun thing to ponder. 
Have you felt like the way that 

1228
01:14:27,920 --> 01:14:35,200
agents use databases, we need a 
new type of database or the way 

1229
01:14:35,200 --> 01:14:40,040
that agents are constantly going
around and trying to collect 

1230
01:14:40,040 --> 01:14:42,040
data? 
And the reason I bring this up 

1231
01:14:42,040 --> 01:14:47,480
is because I was just reading 
that quote UN quote, old paper, 

1232
01:14:47,520 --> 01:14:49,960
I think it was from December 
2025. 

1233
01:14:50,960 --> 01:14:53,680
It was like supporting our AI 
overlords. 

1234
01:14:54,040 --> 01:14:58,080
And it's written by, you know, 
some of the greats, I think 

1235
01:14:58,080 --> 01:15:03,760
Matteis Arya and Jan Stoika and 
Shreya Shankar. 

1236
01:15:03,760 --> 01:15:08,440
And so they, they mentioned how 
you have this, these agents and 

1237
01:15:08,440 --> 01:15:11,000
what they called it was like 
agent speculation. 

1238
01:15:11,920 --> 01:15:15,000
So the agents will go and they 
kind of have a theory and then 

1239
01:15:15,000 --> 01:15:18,280
they'll try and find the data to
support that theory and execute 

1240
01:15:18,280 --> 01:15:22,520
on what they need to do. 
But a lot of times if they're 

1241
01:15:23,120 --> 01:15:27,160
trying to get data from a 
database, it's not that 

1242
01:15:27,160 --> 01:15:33,240
efficient to pull the whole row 
or create this join from these 

1243
01:15:33,240 --> 01:15:39,560
gigantic tables or whatever it 
may be that you're like, can we 

1244
01:15:39,560 --> 01:15:44,480
have agents interact with data 
in a more efficient way? 

1245
01:15:45,480 --> 01:15:47,960
Yeah. 
I mean there surely is many 

1246
01:15:48,080 --> 01:15:52,360
things to invent there, but I'm 
also, I'm a bit of a pragmatic 

1247
01:15:52,360 --> 01:15:55,240
system, those questions. 
And at the end of the day, when 

1248
01:15:55,240 --> 01:15:57,280
you think about that technology,
it's trained on the Internet, 

1249
01:15:57,280 --> 01:16:01,760
it's trained from reinforcement 
learning, from human feedback. 

1250
01:16:01,760 --> 01:16:06,360
It's trained on, on, on human 
traces being built by many labs 

1251
01:16:06,360 --> 01:16:08,560
extra. 
And so my point is that it's a 

1252
01:16:08,560 --> 01:16:11,480
technology that is extremely, 
extremely, extremely 

1253
01:16:11,480 --> 01:16:16,880
anthropomorphic. 
And so it and it feels it, it 

1254
01:16:16,880 --> 01:16:21,360
actually we had that weird 
moment where we needed rag. 

1255
01:16:21,800 --> 01:16:24,680
Yeah. 
And it was kind of a new way of 

1256
01:16:24,680 --> 01:16:27,960
presenting the data made for 
agents, made for the constraint 

1257
01:16:27,960 --> 01:16:29,880
of the context size that was 
rather small. 

1258
01:16:31,040 --> 01:16:34,120
And with the context size 
augmenting and the the very 

1259
01:16:34,120 --> 01:16:35,840
entropometric nature of that 
technology. 

1260
01:16:35,840 --> 01:16:40,120
The what's the best way to do 
search in company data today is 

1261
01:16:40,120 --> 01:16:43,880
just let's do what people call 
agentic search, which is 

1262
01:16:43,880 --> 01:16:46,240
basically just let's use the 
tool that the human use. 

1263
01:16:46,240 --> 01:16:47,680
If we can do it, they can do it 
as well. 

1264
01:16:48,440 --> 01:16:52,760
And we're ready to pay the tax 
of the latency shaded because 

1265
01:16:52,760 --> 01:16:55,600
it's just simpler to just give 
them the tool and let them let 

1266
01:16:55,600 --> 01:16:59,800
them do the work. 
And so I'm, I think I'm, it 

1267
01:16:59,800 --> 01:17:02,880
feels like as the context is 
getting longer, the, the, the, 

1268
01:17:02,960 --> 01:17:06,200
the, the, the way those, those, 
those agents use tools is 

1269
01:17:06,200 --> 01:17:09,360
looking much more. 
I mean, it's looking very much 

1270
01:17:09,360 --> 01:17:13,240
like we use tools. 
And so if we don't have the need

1271
01:17:13,240 --> 01:17:15,880
in the past, why would they? 
Interesting question. 

1272
01:17:16,120 --> 01:17:20,920
And they've been RLD on what the
humans will expect of it. 

1273
01:17:20,920 --> 01:17:25,320
So if you, and they've been 
railed on all these traditional 

1274
01:17:25,320 --> 01:17:29,480
tools also. 
So if you're now creating a new 

1275
01:17:29,480 --> 01:17:35,400
tool that you are expecting the 
agents to use, well, it may take

1276
01:17:35,400 --> 01:17:38,040
a few cycles. 
I I'm fascinated by that space 

1277
01:17:38,040 --> 01:17:43,240
just because I feel like you've 
got this whole new way of 

1278
01:17:43,320 --> 01:17:47,240
interacting with machines 
through agents and the 

1279
01:17:47,240 --> 01:17:51,400
infrastructure that we have. 
It's fun to think about. 

1280
01:17:51,960 --> 01:17:54,640
Is it good enough? 
Can we do like you were saying, 

1281
01:17:54,640 --> 01:17:59,920
can we explore and not be 
beholden to what tech currently 

1282
01:18:00,960 --> 01:18:04,600
constrains us to? 
Or is there, like the other 

1283
01:18:04,600 --> 01:18:07,240
piece, the other side of the 
coin that you're saying, be more

1284
01:18:07,240 --> 01:18:13,360
of a pragmatist and say, look, I
think if we can use it this way,

1285
01:18:13,360 --> 01:18:18,440
the agents can use it this way? 
It's it's obviously going to be 

1286
01:18:18,560 --> 01:18:21,800
always a bit of a bit of a bit 
of a bit of both. 

1287
01:18:21,880 --> 01:18:26,040
But even in the way we use tool,
we use them very differently. 

1288
01:18:26,160 --> 01:18:32,960
It's like we use UIS, they use 
AP is turned MCP, turned CLI and

1289
01:18:33,880 --> 01:18:37,120
eventually it's the CLI. 
And so I would love to use every

1290
01:18:37,120 --> 01:18:39,120
SAS in the world with the CLI 
myself. 

1291
01:18:39,120 --> 01:18:42,760
So in the sense at the end of 
the day we merge back into the 

1292
01:18:43,000 --> 01:18:47,040
most efficient place. 
I might use, I might use the the

1293
01:18:47,040 --> 01:18:50,400
CLI that agents have been that 
have been built for agents for 

1294
01:18:50,400 --> 01:18:54,120
using using the SAS tool myself.
I'm a big of a fan of the 

1295
01:18:54,240 --> 01:18:57,440
terminal, but. 
I just saw a tool, I can't 

1296
01:18:57,440 --> 01:19:00,360
remember what it was. 
It was like turn any SAS into 

1297
01:19:00,480 --> 01:19:02,400
ACLI. 
That was their whole tool. 

1298
01:19:02,560 --> 01:19:06,360
I can't remember what it was, 
but that is. 

1299
01:19:07,520 --> 01:19:14,360
It's exciting to me, yeah. 
It's yeah, it's a different 

1300
01:19:14,360 --> 01:19:18,040
level that we're playing at. 
Well, man, this has been great. 

1301
01:19:18,040 --> 01:19:20,920
Is there anything that you want 
to mention before we jump? 

1302
01:19:21,280 --> 01:19:23,080
Is there anything that I didn't 
ask you about? 

1303
01:19:23,480 --> 01:19:26,400
No, I think this was great. 
We covered, we covered the the 

1304
01:19:26,400 --> 01:19:28,960
main subject. 
I think we on the verge of going

1305
01:19:28,960 --> 01:19:31,880
towards multiplayer AI. 
We don't quite know yet what it 

1306
01:19:31,880 --> 01:19:34,720
look like, but surely won't look
like what we're doing today. 

1307
01:19:34,720 --> 01:19:38,520
And so that's, that's why I, I 
think it's in the, in that wall 

1308
01:19:38,520 --> 01:19:43,480
where there are two black holes 
being created in the market. 

1309
01:19:43,480 --> 01:19:47,320
I think it's always an 
interesting questions of are 

1310
01:19:47,320 --> 01:19:48,960
they not going to going to take 
it all? 

1311
01:19:50,000 --> 01:19:54,040
We, we, we, we really attached 
to the, to the, to the thinking 

1312
01:19:54,040 --> 01:19:58,560
that even if the technology 
again, plateaus, even with 

1313
01:19:58,560 --> 01:20:02,040
current model shape, the way 
we're going to walk in 10 years 

1314
01:20:02,040 --> 01:20:04,440
is probably nothing compared to 
the way we work today. 

1315
01:20:04,440 --> 01:20:07,560
And so there's many, many, many 
opportunities to build for for 

1316
01:20:07,560 --> 01:20:12,400
everyone. 
I will also finish with dust is 

1317
01:20:12,400 --> 01:20:16,840
known throughout the world for 
having the absolute best in 

1318
01:20:16,840 --> 01:20:19,520
person event. 
So if you see a dust event 

1319
01:20:19,720 --> 01:20:22,920
happening near you, I highly 
recommend go for it. 

1320
01:20:23,360 --> 01:20:27,600
Yeah that is for sure. 
They are awesome events. 

1321
01:20:27,600 --> 01:20:30,480
Anybody that has been to one 
will tell you the same. 

1322
01:20:30,840 --> 01:20:32,600
So thanks for doing this dude. 
I really thank. 

1323
01:20:32,600 --> 01:20:33,400
You very much. 
It was great.