1
00:00:00,080 --> 00:00:02,120
Is Claude code the everything 
agent? 

2
00:00:02,360 --> 00:00:05,800
Back on episode 52, Eric Boost 
taught us about advanced Claude 

3
00:00:05,800 --> 00:00:09,080
code usage. 
Well, on episode 75, Eric's back

4
00:00:09,160 --> 00:00:11,560
and he's been building non-stop 
since we saw him last. 

5
00:00:11,840 --> 00:00:14,480
And today we're going to cover 
how Eric has built a system that

6
00:00:14,480 --> 00:00:18,120
allows him to control Claude 
Code from anywhere, be confident

7
00:00:18,120 --> 00:00:21,080
that the task will be done 
correctly, interact with so many

8
00:00:21,080 --> 00:00:22,920
aspects of his life, and much 
more. 

9
00:00:23,000 --> 00:00:25,800
So please enjoy this 
conversation with Eric Boost. 

10
00:00:25,920 --> 00:00:28,560
So if you can have control over 
the terminal, then that means 

11
00:00:28,560 --> 00:00:31,800
that the whole, there's not 
really any limits. 

12
00:00:32,360 --> 00:00:36,440
And so if AI is the abstraction 
and it's just you're talking in 

13
00:00:36,440 --> 00:00:39,920
in English to the the computer 
that has full control over it, 

14
00:00:40,200 --> 00:00:43,520
then you're not really boxed in 
so that you can set up 

15
00:00:43,520 --> 00:00:45,600
permission structures and 
sandboxes and things like that, 

16
00:00:45,640 --> 00:00:48,480
but it's not required. 
And so one thing about cloud 

17
00:00:48,480 --> 00:00:51,600
code that's so awesome is they 
have not only are they dog 

18
00:00:51,600 --> 00:00:53,680
fooding it internally inside of 
Entropic, which means that 

19
00:00:53,680 --> 00:00:55,360
they're using the tool to build 
the tool. 

20
00:00:55,800 --> 00:00:59,800
The new releases, 100% of the 
releases are written by cloud 

21
00:00:59,800 --> 00:01:02,440
code because it's got become so 
reliable and so good. 

22
00:01:02,920 --> 00:01:05,880
And you can see that number 
increasing over the last, you 

23
00:01:05,880 --> 00:01:09,040
know, few months. 
And now it's entirely like super

24
00:01:09,040 --> 00:01:11,520
fast of building and deploying 
new things. 

25
00:01:11,760 --> 00:01:13,800
And so they're, they're 
iterating with cloud code and 

26
00:01:13,800 --> 00:01:16,080
they're releasing the things 
that they find internally that 

27
00:01:16,080 --> 00:01:18,400
work. 
I purpose built for their model.

28
00:01:18,720 --> 00:01:21,120
Their model is trained to do 
better cloud code. 

29
00:01:21,560 --> 00:01:25,480
And so it's not like some other 
system that you're sort of using

30
00:01:25,560 --> 00:01:29,280
to bootstrap the OR so be a 
harness or whatever, like a 

31
00:01:29,280 --> 00:01:32,040
wrapper arounds the the model 
itself. 

32
00:01:32,040 --> 00:01:34,800
It's something that the team 
inside the company that builds 

33
00:01:34,800 --> 00:01:37,320
the model is doing. 
It's like a vertical stack, 

34
00:01:37,320 --> 00:01:39,160
right? 
And so you can imagine the 

35
00:01:39,160 --> 00:01:41,960
opening eyes doing this with 
Kodaks and Google with duals or 

36
00:01:41,960 --> 00:01:45,080
like whatever their tools are 
that they're releasing anti 

37
00:01:45,080 --> 00:01:46,760
gravity and stuff like that. 
And Grok has their own. 

38
00:01:46,920 --> 00:01:49,080
Anyway, with the Anthropic, I 
personally believe it was just 

39
00:01:49,080 --> 00:01:51,200
the best coding model. 
Now it has a limitation that it 

40
00:01:51,200 --> 00:01:53,600
has like a smaller context 
window than some of the others. 

41
00:01:53,960 --> 00:01:56,360
But you know, for instruction 
following for most of the stuff 

42
00:01:56,360 --> 00:01:59,160
that you do agentically in the 
terminal, you don't really need 

43
00:01:59,240 --> 00:02:01,920
all that much. 
And in fact, the ability for the

44
00:02:01,920 --> 00:02:06,880
model to do really good task 
work cuts down if you even if 

45
00:02:06,880 --> 00:02:08,720
you have a really large content 
context on it, but it's not 

46
00:02:08,720 --> 00:02:11,360
following instructions all that 
well over time when it fills up,

47
00:02:12,160 --> 00:02:14,840
that's, you know, you just have 
to test each model, right? 

48
00:02:14,840 --> 00:02:17,720
So they're all really good. 
They're all really capable. 

49
00:02:17,960 --> 00:02:21,760
But the thing about Anthropics 
Cloud Code is that they have 

50
00:02:21,760 --> 00:02:24,160
this hook system. 
And I feel like this is 

51
00:02:24,240 --> 00:02:26,680
essential. 
Like this is the gold standard. 

52
00:02:26,920 --> 00:02:30,600
I thought cloud code is the gold
standard for coding because of 

53
00:02:31,080 --> 00:02:34,080
not only this like internal 
rapid developments and they're 

54
00:02:34,080 --> 00:02:36,440
releasing it to you and not only
because you have something like 

55
00:02:36,440 --> 00:02:38,760
in the cloud desktop app, you 
can now do cloud code in the 

56
00:02:38,760 --> 00:02:39,880
browser. 
You can do cloud code. 

57
00:02:39,880 --> 00:02:41,480
You can access it from like 
GitHub. 

58
00:02:41,920 --> 00:02:44,880
You can access this if you have 
the pro plan, even the $20.00 a 

59
00:02:44,880 --> 00:02:46,440
month plan, you can get access 
to all this stuff. 

60
00:02:46,680 --> 00:02:48,960
But if you have the Max plan, 
you can get the preview things 

61
00:02:48,960 --> 00:02:51,760
like the new things that come 
out and including like Cowork, 

62
00:02:51,880 --> 00:02:55,600
which was just released in cloud
desktop and that relies on cloud

63
00:02:55,600 --> 00:02:57,880
code under the hood. 
And the hook system is actually 

64
00:02:57,880 --> 00:02:59,960
available in all those things. 
I didn't know when we talked 

65
00:03:00,120 --> 00:03:02,760
yesterday, but I took a look and
it's there. 

66
00:03:02,960 --> 00:03:04,800
It's just kind of limited 
because it's somewhat in like 

67
00:03:04,800 --> 00:03:07,640
the sandbox VM and it's you got 
to get stuff out the hooks out 

68
00:03:07,640 --> 00:03:09,720
and back in. 
But here's why the hooks matter.

69
00:03:10,000 --> 00:03:13,880
So with every agentic tool you 
might use that you want to have 

70
00:03:13,880 --> 00:03:17,040
do something on your behalf, 
there are some limitations. 

71
00:03:17,040 --> 00:03:19,920
There are some constraints. 
Or if there are things it does 

72
00:03:19,920 --> 00:03:22,800
that aren't quite right, like 
ideally, let's set the stage. 

73
00:03:22,800 --> 00:03:24,920
Here's what you want. 
You want to be able to talk to 

74
00:03:24,920 --> 00:03:27,000
it like you talk to a human, 
like someone who's a project 

75
00:03:27,000 --> 00:03:31,360
manager or let's say Jarvis, 
your Iron Man's AI or your 

76
00:03:31,360 --> 00:03:35,040
secretary or personal assistant.
Because the goal is for me in my

77
00:03:35,040 --> 00:03:37,440
life, I want to be able to get 
back more times that are 

78
00:03:37,440 --> 00:03:40,960
meaningful with the people and 
the relationships and like what 

79
00:03:40,960 --> 00:03:45,120
I care about is more meaningful.
So just imagine for a moment all

80
00:03:45,120 --> 00:03:47,560
the things that you do in life, 
in personal and work life. 

81
00:03:48,600 --> 00:03:51,440
Can't say over here. 
So I'm counting and like, if you

82
00:03:51,440 --> 00:03:54,560
were to rank order those things 
by like these are the the sets 

83
00:03:54,560 --> 00:03:57,200
of tasks or subtasks that I do 
that I really like. 

84
00:03:57,280 --> 00:03:59,240
And these are the ones I like 
less, right? 

85
00:03:59,360 --> 00:04:02,680
They're kind of more automated 
or mundane or automatable, I'd 

86
00:04:02,720 --> 00:04:06,320
say like repetitive and mundane.
And I wish I didn't have to do 

87
00:04:06,320 --> 00:04:09,520
so much of this set of tasks. 
I could do more well of this set

88
00:04:09,520 --> 00:04:11,800
of tasks is lower. 
There's a subset of those. 

89
00:04:11,800 --> 00:04:14,880
There's like a like some of 
those tasks that you could 

90
00:04:14,880 --> 00:04:17,880
probably automate away with AI 
if you knew how to do it. 

91
00:04:17,880 --> 00:04:21,560
And and it's the models are like
as many people said, like the 

92
00:04:21,560 --> 00:04:23,440
worst they're ever going to be. 
They're just getting better. 

93
00:04:23,520 --> 00:04:27,360
And the sooner you get in at 
learning how to use the tools to

94
00:04:27,360 --> 00:04:31,760
to take over whatever it is you 
don't want to do, the more ahead

95
00:04:31,760 --> 00:04:34,480
of the curve you are and the 
better you are at getting back 

96
00:04:34,480 --> 00:04:36,400
more that time. 
And it's a flywheel. 

97
00:04:36,400 --> 00:04:38,680
It's a feedback team, right. 
So like once you learn how to do

98
00:04:38,680 --> 00:04:40,720
it for here for this thing that 
you're super passionate about, 

99
00:04:40,760 --> 00:04:42,320
like really great. 
If I could just not have to do 

100
00:04:42,320 --> 00:04:46,200
that anymore, then that skill up
of investing the time to work 

101
00:04:46,200 --> 00:04:48,880
with the model to do that allows
you to do a lot more. 

102
00:04:48,880 --> 00:04:51,480
And I have AI feel like a 
compelling reason for me to say 

103
00:04:51,480 --> 00:04:54,120
that the place to start with 
that or if you've already done 

104
00:04:54,120 --> 00:04:57,720
it to try is Claude code. 
Now there's different ways to 

105
00:04:57,720 --> 00:04:59,720
use it. 
Again, you can use it in the the

106
00:04:59,720 --> 00:05:03,160
desktop app, you can use it in 
the website, you can use it in 

107
00:05:03,160 --> 00:05:04,560
the terminal. 
A lot of people are too 

108
00:05:04,560 --> 00:05:07,720
intimidated by the terminal. 
They get into it and it is like 

109
00:05:08,040 --> 00:05:11,600
superpower lands. 
So you can start wherever you 

110
00:05:11,600 --> 00:05:13,400
want to start. 
I personally would start this 

111
00:05:13,400 --> 00:05:17,200
from because that is the the 
place where there are no limits.

112
00:05:18,320 --> 00:05:22,560
But anyway, like the goal being 
I want to find one thing that I 

113
00:05:22,560 --> 00:05:25,560
don't want to do anymore, but I 
want to find a tool that can 

114
00:05:25,560 --> 00:05:26,640
automate it. 
So I don't want to do that 

115
00:05:26,640 --> 00:05:28,520
anymore. 
And if that thing is like doing 

116
00:05:28,520 --> 00:05:31,280
this document work, OK and you 
have a mass subscription or 

117
00:05:31,280 --> 00:05:34,440
you're willing to pay for it, 
that's $100 a month for the 5X 

118
00:05:34,480 --> 00:05:37,560
and and 200 for the 20X. 
That's 20X of what you get in 

119
00:05:37,560 --> 00:05:39,160
the pro plan, which is $20.00 a 
month. 

120
00:05:39,160 --> 00:05:41,920
I think that that's right. 
I'm not sure, but which is the 

121
00:05:41,920 --> 00:05:43,600
pro plan you get more than the 
free plan anyway. 

122
00:05:43,600 --> 00:05:46,160
It's if the pro plan, I think 
it's $20.00 a month. 

123
00:05:46,440 --> 00:05:48,600
I'm not working for Anthropic. 
I don't sell this stuff. 

124
00:05:48,600 --> 00:05:52,200
I don't really know, but I think
that if you if you get one of 

125
00:05:52,200 --> 00:05:54,120
the Mac subscriptions, you can 
use cowork. 

126
00:05:54,320 --> 00:05:57,240
Of course, cowork will probably 
come to the the pro subscription

127
00:05:57,240 --> 00:06:00,280
pretty soon. 
But that's for like Claude code,

128
00:06:00,280 --> 00:06:03,760
except it's not code related. 
It's more like general purpose. 

129
00:06:03,760 --> 00:06:05,320
I have some of these documents. 
I want to create this 

130
00:06:05,320 --> 00:06:07,880
presentation. 
You know, a lot of things can do

131
00:06:07,880 --> 00:06:08,720
that. 
You think, well, that's not 

132
00:06:08,720 --> 00:06:10,880
special. 
Any AI can turn documents in a 

133
00:06:10,880 --> 00:06:12,080
presentation, but it's 
different. 

134
00:06:12,080 --> 00:06:13,400
There's much more that you can 
do. 

135
00:06:13,400 --> 00:06:17,680
OK, Like Claude code is another 
tab inside of Claude desktop 

136
00:06:17,680 --> 00:06:20,400
app. 
It with Claude code in that tab,

137
00:06:21,080 --> 00:06:23,280
you can basically connect it to 
your GitHub or your local file 

138
00:06:23,280 --> 00:06:24,360
system. 
And it can work within the 

139
00:06:24,360 --> 00:06:27,040
sandbox of whatever you set up. 
And it can create a little VM 

140
00:06:27,040 --> 00:06:30,120
and it can do stuff on your 
computer, writing your actual 

141
00:06:30,120 --> 00:06:32,080
code files. 
And this is what matters versus 

142
00:06:32,080 --> 00:06:35,080
allow the other systems that 
you'll see this state that's 

143
00:06:35,080 --> 00:06:38,000
being managed is your local 
code, your local files, your 

144
00:06:38,000 --> 00:06:40,160
local documents. 
And it can copy it out of there 

145
00:06:40,160 --> 00:06:42,600
and put it into a slow sandbox 
environment or you can give it 

146
00:06:42,600 --> 00:06:44,280
access to whatever folder you 
want. 

147
00:06:44,720 --> 00:06:47,040
And there's a lot of things that
it can do. 

148
00:06:47,400 --> 00:06:49,880
And depending on like there's 
even a file system tool, I can 

149
00:06:49,880 --> 00:06:51,480
show you a lot of this stuff. 
I don't know how much it will 

150
00:06:51,480 --> 00:06:53,920
get into, but I would love to 
like walk people through at some

151
00:06:53,920 --> 00:06:55,760
point. 
But Claude coding the terminal 

152
00:06:55,760 --> 00:06:59,160
has everything. 
And so here's here's where I'm 

153
00:06:59,160 --> 00:07:01,360
going with it. 
So I want it personally. 

154
00:07:01,360 --> 00:07:04,840
Well, the thing that I said that
I want is I am overwhelmed. 

155
00:07:04,880 --> 00:07:08,280
I have 4 kids between the ages 
of 15 and 6, and I have one 

156
00:07:08,280 --> 00:07:13,440
that's due in a month. 
So in baby #5 and we're very sad

157
00:07:13,440 --> 00:07:16,640
about that, but there's far too 
many school emails, OK, from the

158
00:07:16,640 --> 00:07:18,480
teachers, especially when it's 
in high school and middle 

159
00:07:18,920 --> 00:07:20,200
school. 
And there's far too much to keep

160
00:07:20,200 --> 00:07:22,400
up with. 
I've got calendar events like 

161
00:07:22,400 --> 00:07:24,720
crazy. 
I've got text messages and Teams

162
00:07:24,720 --> 00:07:27,640
and GroupMe and emails and all 
the things that everyone else 

163
00:07:27,640 --> 00:07:29,520
struggles with, right? 
How do you keep up with all this

164
00:07:29,520 --> 00:07:31,080
stuff? 
So I wanted the system to keep 

165
00:07:31,080 --> 00:07:34,120
up with all that stuff. 
Well, I wanted to build 

166
00:07:34,120 --> 00:07:37,240
something like Jarvis that is a 
secretary of the front end. 

167
00:07:37,240 --> 00:07:39,720
They can look at all the data 
coming in from all the sources 

168
00:07:39,720 --> 00:07:43,440
without me having to open any 
website or app ever. 

169
00:07:43,880 --> 00:07:48,200
Except for my daily To Do List 
that is automatically managed 

170
00:07:48,200 --> 00:07:51,920
based on what it knows that I 
need and my calendar, which has 

171
00:07:51,960 --> 00:07:54,720
automatically been scheduled 
based on events that come in or 

172
00:07:54,720 --> 00:07:56,440
newsletters or stuff that 
matters to me. 

173
00:07:56,880 --> 00:08:00,280
And it can identify like this is
for, you know, here's a bunch of

174
00:08:00,280 --> 00:08:02,200
stuff coming out, but is it 
relevant to me? 

175
00:08:02,240 --> 00:08:04,400
Do I need to know it? 
Or is it just like your, your 

176
00:08:04,400 --> 00:08:07,120
kid is learning to color on 
today? 

177
00:08:07,120 --> 00:08:09,840
Or is it like there's this event
that's coming up and they need 

178
00:08:09,840 --> 00:08:11,640
to have this and I need it like 
RSVP. 

179
00:08:12,120 --> 00:08:14,000
So those are the things that are
different, right? 

180
00:08:14,520 --> 00:08:16,440
And then also like text 
messages. 

181
00:08:16,440 --> 00:08:19,280
And not only do I need to be 
able to know about this stuff, I

182
00:08:19,280 --> 00:08:22,200
wanted to collate together 
intelligently by a model such 

183
00:08:22,200 --> 00:08:25,120
that if it decides, here's an 
important thing, you can put it 

184
00:08:25,120 --> 00:08:28,280
on my calendar, but it can also 
update the calendar. 

185
00:08:28,280 --> 00:08:30,680
If I get a text that's 
different, that's for the same 

186
00:08:30,680 --> 00:08:34,039
event for the same person, it 
can like fill update the content

187
00:08:34,039 --> 00:08:36,320
from it, right? 
I've built that system and it's 

188
00:08:36,320 --> 00:08:38,159
very simple. 
I didn't write any code. 

189
00:08:38,200 --> 00:08:40,559
I won't say super simple, like I
had to struggle with it a little

190
00:08:40,559 --> 00:08:43,960
bit, but that system now exists 
so I can keep up and not be 

191
00:08:43,960 --> 00:08:47,280
drowned and not miss things and 
actually be the dad who knows 

192
00:08:47,280 --> 00:08:49,560
about where my kids events are, 
right. 

193
00:08:49,560 --> 00:08:51,880
And I said a lot because I like,
OK, I want to do this now. 

194
00:08:51,880 --> 00:08:53,720
I want to do that next thing and
then the next thing. 

195
00:08:53,720 --> 00:08:55,680
And once you learn to do the 
first thing, that's the biggest 

196
00:08:55,680 --> 00:08:58,440
curve and then you, you keep 
going and there's a little bit 

197
00:08:58,440 --> 00:09:01,080
more. 
So within this, the struggle is 

198
00:09:01,080 --> 00:09:03,000
largely like, what is the model 
doing wrong? 

199
00:09:03,440 --> 00:09:08,560
And how do I, how do I get the 
model to be better at not going 

200
00:09:08,560 --> 00:09:10,760
off the rails in this crazy way 
every time? 

201
00:09:10,760 --> 00:09:13,280
Like it's just, if it fails, 
it's typically like some default

202
00:09:13,280 --> 00:09:15,640
behavior in the model. 
And how do I get it to not do 

203
00:09:15,640 --> 00:09:18,160
that? 
This is where I think the 

204
00:09:18,160 --> 00:09:22,160
massive value in the hook system
wise is because I mentioned this

205
00:09:22,160 --> 00:09:25,080
in the last podcast that we did.
But like hooks run outside of 

206
00:09:25,080 --> 00:09:28,000
the awareness of the model, the 
model that you're talking to 

207
00:09:28,000 --> 00:09:31,440
being like for me, Claude Opus 
4.5 within the terminal, then 

208
00:09:31,440 --> 00:09:34,280
Claude code. 
But outside of that in the the 

209
00:09:34,280 --> 00:09:37,280
life cycle of the model run, 
there's there's a bunch of 

210
00:09:37,280 --> 00:09:39,680
events that are occurring. 
So when the user sends a message

211
00:09:39,680 --> 00:09:43,800
to the model, there's something 
called a like a user prompt hook

212
00:09:44,440 --> 00:09:46,920
and a user submit host. 
Some, I forget the name of that 

213
00:09:46,920 --> 00:09:49,960
who's honestly, but like at that
moment before it sends that 

214
00:09:49,960 --> 00:09:53,600
message to the model, it can be 
intercepted by some code, which 

215
00:09:53,600 --> 00:09:56,640
is this hook call. 
And so the hook call that each 

216
00:09:56,640 --> 00:09:59,360
one of these stages like before 
it returns back to the user. 

217
00:09:59,560 --> 00:10:02,160
When it's about to write files, 
when it's about to refile, it's 

218
00:10:02,160 --> 00:10:05,880
about about to call a tool, it's
about to compact before it does 

219
00:10:05,880 --> 00:10:08,680
any of these things, you have an
opportunity to guide it or 

220
00:10:08,680 --> 00:10:11,000
direct it or inject content into
it. 

221
00:10:11,400 --> 00:10:14,080
And so instead of what most 
people do and how Quad Code by 

222
00:10:14,080 --> 00:10:18,280
default is set up is they just 
use the, the default harness, 

223
00:10:18,280 --> 00:10:22,160
the default way that clod is 
intended to be used, but without

224
00:10:22,160 --> 00:10:25,120
taking advantage of a lot of 
the, there's obviously there's a

225
00:10:25,120 --> 00:10:27,120
lot of optimization that's left 
on the table. 

226
00:10:27,520 --> 00:10:31,760
And Quad Code is much more 
capable than most of these 

227
00:10:31,760 --> 00:10:36,160
agentic tools, but it's still 
hobbled quite a bit by the 

228
00:10:36,160 --> 00:10:38,960
defaults. 
And the team, the Quad Co team 

229
00:10:39,080 --> 00:10:42,040
Anthropic, I know is working to 
like make it better and better. 

230
00:10:42,040 --> 00:10:43,480
And they're going to integrate a
bunch of this stuff. 

231
00:10:43,800 --> 00:10:46,960
But what I found is if you have 
something called a quad dot MD 

232
00:10:46,960 --> 00:10:49,280
file, which is a file that 
contains instructions or 

233
00:10:49,280 --> 00:10:52,080
contents that you put in there 
that it it decides based on what

234
00:10:52,080 --> 00:10:55,280
you tell it to, to put in there.
And when it starts a session, it

235
00:10:55,280 --> 00:10:58,440
reads that file. 
The problem is that as your 

236
00:10:58,760 --> 00:11:01,320
conversation with the model to 
build or do something for you 

237
00:11:01,320 --> 00:11:04,320
gets longer and longer, it 
starts to ignore more and more 

238
00:11:04,320 --> 00:11:07,400
of that that cloud file because 
it's, it's older and you get 

239
00:11:07,800 --> 00:11:09,880
recent user messages and you've 
got tools, you got all these 

240
00:11:09,880 --> 00:11:12,560
things that happen. 
And so you, it's not 100% 

241
00:11:12,560 --> 00:11:14,880
reliable. 
And so the goal is, how do you 

242
00:11:14,880 --> 00:11:18,560
get something such that when you
give it a task like a human, 

243
00:11:18,880 --> 00:11:21,920
it's going to know where to 
breakdown all the pieces of the 

244
00:11:21,920 --> 00:11:25,760
task and when, what project to 
put it in and how to queue it up

245
00:11:25,800 --> 00:11:29,120
and where to slot all those task
areas and what to do first, what

246
00:11:29,120 --> 00:11:31,960
research it's going to know, 
like what your code base is when

247
00:11:31,960 --> 00:11:34,960
it starts to take on each one, 
it's going to explore thoroughly

248
00:11:34,960 --> 00:11:38,520
with adversaries to try to 
figure out what is the code that

249
00:11:38,520 --> 00:11:42,080
already exists and should I 
update that code or should I add

250
00:11:42,080 --> 00:11:45,200
new code? 
It really have to justify like I

251
00:11:45,200 --> 00:11:48,240
need to, I need to add new code 
and here's the reason. 

252
00:11:48,240 --> 00:11:50,840
And because someone else will 
come adversarially and say, like

253
00:11:51,240 --> 00:11:53,480
I see other places, Are you sure
you shouldn't be doing it there?

254
00:11:53,960 --> 00:11:57,600
And so eventually you can 
improve the weaknesses of the 

255
00:11:57,600 --> 00:11:59,280
model. 
So the model gets distracted. 

256
00:11:59,280 --> 00:12:02,480
How you solve that hooks solve 
that, how they solve it Instead 

257
00:12:02,480 --> 00:12:05,560
of creating a rules, sorry, 
creating a claw that MD that has

258
00:12:05,560 --> 00:12:08,880
all these instructions that the 
model might ignore, why not 

259
00:12:08,880 --> 00:12:14,320
create a, let's say Jason set of
or it can be marked down, 

260
00:12:14,320 --> 00:12:16,960
whatever. 
But like a structured format of 

261
00:12:17,040 --> 00:12:22,240
rules that at certain times the 
hooks look at because they, the 

262
00:12:22,240 --> 00:12:25,120
rules are tagged by which type 
of hook it is and which project 

263
00:12:25,120 --> 00:12:28,320
it belongs to. 
And you know what is relevant, 

264
00:12:28,320 --> 00:12:29,680
what kind of task that is 
relevant to. 

265
00:12:30,240 --> 00:12:33,240
And then if the model starts to 
do something it doesn't, it 

266
00:12:33,240 --> 00:12:36,760
shouldn't be doing, the hook can
redirect it like inject. 

267
00:12:36,760 --> 00:12:40,160
It'll detect you did this thing.
And at that moment, when it's 

268
00:12:40,160 --> 00:12:44,320
very relevant to the models 
mind, it injects content into 

269
00:12:44,320 --> 00:12:47,280
the models context as it's 
working, saying, no, don't do 

270
00:12:47,280 --> 00:12:48,840
that thing or this thing is 
blocked. 

271
00:12:48,840 --> 00:12:50,280
You must use this other thing, 
right? 

272
00:12:50,560 --> 00:12:52,440
So it reads the rules when it 
launches. 

273
00:12:52,440 --> 00:12:56,800
So clot MD instead of that, it's
reading in this file that is the

274
00:12:56,800 --> 00:12:58,640
rules. 
And then as it's getting ready 

275
00:12:58,640 --> 00:13:02,840
to stop, it's reading the hooks 
are directing it saying, did you

276
00:13:02,840 --> 00:13:05,160
follow compliance? 
Did you follow like autonomy? 

277
00:13:05,160 --> 00:13:07,720
Did you like keep working until 
the entire queue is done? 

278
00:13:08,120 --> 00:13:11,960
Did you this task that was 
assigned is you like write 

279
00:13:11,960 --> 00:13:14,960
everything, including the the 
model itself, like the tools 

280
00:13:14,960 --> 00:13:18,800
have a something called to do 
write tool that moves their 

281
00:13:18,800 --> 00:13:22,960
tasks, that creates it on tasks.
You can dump those into a file 

282
00:13:22,960 --> 00:13:25,720
or you can hand them to your 
project management system, which

283
00:13:25,720 --> 00:13:28,200
is the main model that has all 
the context and knows what it's 

284
00:13:28,200 --> 00:13:30,600
doing. 
You can have a hook tell it OK, 

285
00:13:30,600 --> 00:13:33,680
take everything that's in your 
to do thing and and basically 

286
00:13:33,680 --> 00:13:36,280
queue it up properly so that 
here's the goal. 

287
00:13:36,440 --> 00:13:43,600
You match the current task to 
the proper context from your 

288
00:13:43,600 --> 00:13:47,040
existing session history and 
code base. 

289
00:13:47,400 --> 00:13:49,680
So everything that you've done 
and everything exists in your 

290
00:13:49,680 --> 00:13:51,640
project and everything that's in
git. 

291
00:13:51,680 --> 00:13:54,720
I love the idea of taking 
information out of cloud dot MD 

292
00:13:54,720 --> 00:13:57,280
putting into hooks so instead of
just getting lost in the 

293
00:13:57,280 --> 00:13:59,760
context, it's actually triggered
programmatically which is 

294
00:13:59,760 --> 00:14:02,120
awesome. 
Do you recommend anything that 

295
00:14:02,120 --> 00:14:05,560
stays in Cloud MD or should we 
migrate completely away and 

296
00:14:05,560 --> 00:14:08,680
start leveraging hooks for every
aspect of the the guidelines? 

297
00:14:08,920 --> 00:14:11,000
It's a good question. 
I mean, I don't really know 

298
00:14:11,000 --> 00:14:13,000
honestly. 
My, my flow is changing 

299
00:14:13,000 --> 00:14:14,560
constantly. 
I'm always testing. 

300
00:14:14,560 --> 00:14:18,200
I have benchmarks and audits and
stuff set up so that like it 

301
00:14:18,200 --> 00:14:20,800
will try to compare and if it 
runs into any issues or it sees 

302
00:14:20,800 --> 00:14:24,160
hooks like not letting something
pass multiple times that it 

303
00:14:24,160 --> 00:14:27,320
seems like it should, it'll 
basically write these to a file 

304
00:14:27,320 --> 00:14:29,440
that the model will read in and 
queue up. 

305
00:14:30,160 --> 00:14:32,600
Like you kind of add it to a 
queue like, hey, hey, look at 

306
00:14:32,600 --> 00:14:33,880
this. 
This seems like maybe an 

307
00:14:33,880 --> 00:14:37,120
opportunity for optimization for
self improvement or whatever. 

308
00:14:37,640 --> 00:14:41,520
And I can't, I have a Cloud MD, 
but I think maybe it just got 

309
00:14:41,520 --> 00:14:43,080
removed. 
I kind of just did this this 

310
00:14:43,080 --> 00:14:46,240
morning, just removing entirely.
I basically asked the model, OK,

311
00:14:46,240 --> 00:14:49,200
look, here's our structure. 
You can explore in the code base

312
00:14:49,200 --> 00:14:52,560
and see what we're doing. 
And we moved everything down to 

313
00:14:52,560 --> 00:14:55,560
rules and hooks. 
And I had Cloud MD referencing 

314
00:14:55,560 --> 00:14:58,080
like at referencing the the 
rules file, if you will. 

315
00:14:58,240 --> 00:15:01,120
The problem with like Cloud MD 
at referencing files is it 

316
00:15:01,120 --> 00:15:03,760
sometimes it won't read in all 
the contents of the files that 

317
00:15:03,760 --> 00:15:07,200
they get too long or it's a 
little bit less deterministic. 

318
00:15:07,200 --> 00:15:09,600
Like if it's a subfolder and 
there's a clot in there, like at

319
00:15:09,600 --> 00:15:11,000
what point is it actually read 
in? 

320
00:15:11,000 --> 00:15:14,960
Is it's like confusing to me to 
know if it's just truly the 

321
00:15:14,960 --> 00:15:16,280
model really knows what's going 
on. 

322
00:15:16,680 --> 00:15:20,280
And so I just having it inject 
the rules file like a hook is 

323
00:15:20,280 --> 00:15:22,600
literally injecting the full 
contents of it in there. 

324
00:15:23,040 --> 00:15:26,280
And then I basically have moved 
everything from clot MD. 

325
00:15:26,280 --> 00:15:29,200
But there might be some things, 
some basic information that that

326
00:15:29,200 --> 00:15:31,680
your system decides where your 
for your project. 

327
00:15:32,320 --> 00:15:34,000
Maybe it should just go in Cloud
MD. 

328
00:15:34,000 --> 00:15:36,600
Like it's, it's simple, it's 
short, it's descriptive about, 

329
00:15:36,600 --> 00:15:38,560
you know, the content or 
whatever. 

330
00:15:38,560 --> 00:15:43,120
I try to let like I used to have
these files where it was like, 

331
00:15:43,120 --> 00:15:44,960
here's the architecture and 
here's the current task. 

332
00:15:45,600 --> 00:15:48,680
And basically I tell it, don't 
create more files. 

333
00:15:48,680 --> 00:15:49,840
These are the only files you can
create. 

334
00:15:50,080 --> 00:15:52,920
And it would mostly remember. 
But if a sub agent goes off and 

335
00:15:52,920 --> 00:15:55,560
does research, it comes back and
it usually delivers that in the 

336
00:15:55,560 --> 00:15:57,960
form of a file or something. 
So then you they tell it, OK, 

337
00:15:57,960 --> 00:15:59,400
don't do that. 
And then you've got to like sit 

338
00:15:59,400 --> 00:16:02,240
there and babysit it. 
The problem is I'm trying to 

339
00:16:02,240 --> 00:16:06,720
move from can measure this and 
like, I don't know the right 

340
00:16:06,960 --> 00:16:11,640
metric to say here, but to 
convince you the idea is right 

341
00:16:11,640 --> 00:16:15,040
now there's some amount of time 
I have to spend sitting in front

342
00:16:15,080 --> 00:16:18,840
of the like working with a 
model, OK, telling it that 

343
00:16:18,840 --> 00:16:21,040
wasn't right. 
Do this, I'm measuring it in 

344
00:16:21,040 --> 00:16:24,680
like it's not just how much can 
it do autonomously, it's how 

345
00:16:24,680 --> 00:16:28,000
much can it accurately 
understand my intents and do 

346
00:16:28,000 --> 00:16:30,040
autonomously without me having 
to be there. 

347
00:16:30,040 --> 00:16:33,720
So like right now, the system 
that I have is such that if I 

348
00:16:34,360 --> 00:16:38,360
give it an instruction, it will 
decide, OK, do I have the right 

349
00:16:38,360 --> 00:16:41,840
context for this right now? 
What's my context utilization? 

350
00:16:41,840 --> 00:16:45,040
Because the hooks are always 
aware of how much context has 

351
00:16:45,040 --> 00:16:47,440
been utilized so far. 
And if it's over a certain 

352
00:16:47,440 --> 00:16:51,760
threshold, it will trigger like,
OK, do the project management 

353
00:16:51,760 --> 00:16:55,600
system should look at this task 
and decide if it should do it 

354
00:16:55,600 --> 00:16:58,560
now or if it should queue it up 
for another session. 

355
00:16:58,560 --> 00:17:00,960
And if there's something similar
in another sessions in the 

356
00:17:00,960 --> 00:17:03,720
queue, it should merge that task
in with that grouping. 

357
00:17:04,359 --> 00:17:07,800
And then whenever it gets around
that session, there's a code 

358
00:17:07,800 --> 00:17:10,000
exploration agent that goes and 
looks at the code. 

359
00:17:10,000 --> 00:17:13,040
There's a code exploration agent
#2 that happens in sequence and 

360
00:17:13,040 --> 00:17:14,960
not the same time. 
It looks at the first one's 

361
00:17:14,960 --> 00:17:18,119
responses and then goes through 
and checks all that, double 

362
00:17:18,119 --> 00:17:20,480
checks it all and then 
adversarially challenges it to 

363
00:17:20,480 --> 00:17:25,000
say like, can you like, yeah, 
keep this DRY or whatever in 

364
00:17:25,000 --> 00:17:27,560
these other ways? 
And then there's a pod has just 

365
00:17:27,560 --> 00:17:30,240
released an open source code. 
What's it called? 

366
00:17:30,880 --> 00:17:34,240
Simplifier, right? 
And after it writes things, it 

367
00:17:34,240 --> 00:17:36,840
goes through that process. 
There's some other steps in 

368
00:17:36,840 --> 00:17:38,360
there. 
And after it writes things, 

369
00:17:38,360 --> 00:17:41,920
there's also an adversary that 
it goes and does challenge 

370
00:17:41,920 --> 00:17:44,480
basically to verify that any 
claims that are made or any 

371
00:17:44,480 --> 00:17:47,200
colors change actually was 
changed and was made. 

372
00:17:47,240 --> 00:17:49,920
Could you go into the 
adversarial aspect to it? 

373
00:17:49,920 --> 00:17:52,600
Like, do you just kind of say 
like be the devil's advocate or 

374
00:17:52,600 --> 00:17:53,960
how do you set that one up? 
Yeah. 

375
00:17:53,960 --> 00:17:57,760
So it's basically working with 
the model to find out one of the

376
00:17:57,760 --> 00:17:59,760
best practices. 
So a lot of this I've pulled in 

377
00:17:59,760 --> 00:18:03,440
from other people's work. 
So like Kieran at every has 

378
00:18:03,680 --> 00:18:05,920
compound engineering. 
So it's going out and 

379
00:18:05,920 --> 00:18:07,880
researching what are the 
strategies used in there. 

380
00:18:08,120 --> 00:18:10,360
Anthropic has a bunch of things 
in their cookbooks or the things

381
00:18:10,360 --> 00:18:14,520
that they release or publish or 
in their documentation or I have

382
00:18:14,520 --> 00:18:17,040
like a new, as you know, cloud 
code docs like a local version 

383
00:18:17,040 --> 00:18:19,520
of the documentation running. 
So the agents always know to 

384
00:18:19,520 --> 00:18:22,320
look there and examine like 
whether the capabilities and 

385
00:18:22,320 --> 00:18:23,720
then like what the best ways to 
use this. 

386
00:18:23,720 --> 00:18:26,680
And I have the sub agents debate
this and I have in the last 

387
00:18:26,680 --> 00:18:28,800
conversation with Ray, by the 
way, we had a little bit of a 

388
00:18:28,800 --> 00:18:31,360
date debate about like whether 
or not you spam Ultra think or 

389
00:18:31,360 --> 00:18:33,440
whether or not it's better to 
like not use that. 

390
00:18:33,960 --> 00:18:37,680
I just use ultra actually set 
the thinking tokens to like 6 

391
00:18:37,880 --> 00:18:41,080
through 3999. 
And the reason is, and I might 

392
00:18:41,080 --> 00:18:42,360
change that. 
I'm not saying it's prescriptive

393
00:18:42,360 --> 00:18:44,720
for everyone, but like for me, 
that's what I'm testing. 

394
00:18:44,840 --> 00:18:46,840
Yeah, right now. 
And the reason is that in the 

395
00:18:46,840 --> 00:18:49,720
planning stages especially, it's
super important that you get 

396
00:18:49,720 --> 00:18:51,680
things right up front in your 
plan. 

397
00:18:51,680 --> 00:18:56,040
There's like this rule where if 
you have, if it costs like 1 

398
00:18:56,040 --> 00:18:59,400
unit of resources to change 
something in the planning phase,

399
00:18:59,560 --> 00:19:02,320
if you've already delivered it, 
it takes like 7 units to try to 

400
00:19:02,320 --> 00:19:04,720
change it. 
It's much better to to focus 

401
00:19:04,720 --> 00:19:07,880
those tokens upfront, but I kind
of just like am testing letting 

402
00:19:07,880 --> 00:19:10,640
them all do it because it's not 
like if you allocate, if you say

403
00:19:10,640 --> 00:19:13,760
alter think or think hard or 
think or whatever, these 

404
00:19:13,760 --> 00:19:16,960
different by the way, these 
words trigger in the model, how 

405
00:19:16,960 --> 00:19:19,840
many tokens it's allowed. 
It's allocated for the thinking 

406
00:19:19,840 --> 00:19:22,160
compute budget, basically. 
So how many tokens it gets to 

407
00:19:22,160 --> 00:19:24,360
think about it. 
And it's not saying, I think 

408
00:19:24,360 --> 00:19:27,360
it's a misconception that it's 
going to just take from your 

409
00:19:27,360 --> 00:19:32,400
200,000 context window, you 
know, 32,000 tokens away is 

410
00:19:32,400 --> 00:19:35,120
giving it an upper bound. 
So that if the model believes 

411
00:19:35,120 --> 00:19:37,640
that the complexity is 
sufficient is such that you 

412
00:19:37,640 --> 00:19:39,840
should use more tokens and 
you're giving it more headroom 

413
00:19:39,840 --> 00:19:42,800
to use it if it wants to. 
If you ask it 1 + 1, it's not 

414
00:19:42,800 --> 00:19:46,320
going to extract 32,000 tokens. 
If you say, oh, if you ask it, 

415
00:19:46,600 --> 00:19:49,320
you know, if you have something 
that's quite so much challenging

416
00:19:49,320 --> 00:19:51,680
though, and you don't have the 
thinking tokens and it's more 

417
00:19:51,680 --> 00:19:54,080
likely to make more mistakes 
because it's not going to spend 

418
00:19:54,080 --> 00:19:56,800
the time to consider it. 
So it's just this trade off how 

419
00:19:57,960 --> 00:19:59,160
much you want to get done this 
session. 

420
00:19:59,720 --> 00:20:02,960
I've found that like I kind of 
actually have a a stopping 

421
00:20:02,960 --> 00:20:04,720
point. 
There's maybe like 40% or 

422
00:20:04,720 --> 00:20:06,880
whatever. 
It challenges the model to write

423
00:20:06,880 --> 00:20:08,600
things. 
I'm going to, I'm going to 

424
00:20:08,600 --> 00:20:10,120
finish this and get back to your
question, OK? 

425
00:20:10,400 --> 00:20:14,080
It challenges the model to write
things like. 

426
00:20:14,080 --> 00:20:17,240
Basically sum up what it's doing
for legitimate session and what 

427
00:20:17,240 --> 00:20:21,320
the next session should be doing
and update the queue based off 

428
00:20:21,320 --> 00:20:23,800
of. 
You know, what was, what the 

429
00:20:23,800 --> 00:20:25,560
user said and what it thinks is 
next and all that. 

430
00:20:25,880 --> 00:20:27,520
And it's next to the to do, 
right? 

431
00:20:27,520 --> 00:20:30,400
Like the task list is at the 
bottom of Cloud Code Terminal. 

432
00:20:30,400 --> 00:20:32,800
If you've seen that which the 
model creates for itself and it 

433
00:20:32,800 --> 00:20:35,040
dumps that in and like organizes
it correctly. 

434
00:20:35,360 --> 00:20:38,920
And then it injects sufferential
self referentially into itself, 

435
00:20:39,360 --> 00:20:43,400
a command at some point on the 
next user submission, which 

436
00:20:43,400 --> 00:20:47,160
takes the next user's request 
and injects that into the queue 

437
00:20:47,160 --> 00:20:50,840
as well as the next item. 
Yeah, we'll organize it and then

438
00:20:50,840 --> 00:20:57,280
it it calls clear on itself. 
So the model itself determines 

439
00:20:57,280 --> 00:21:00,240
when it should clear its own 
context, right. 

440
00:21:00,360 --> 00:21:03,160
If you're following that based 
on what the utilization is and 

441
00:21:03,160 --> 00:21:05,880
there's like a soft clear and 
then there's like a 50% 

442
00:21:05,880 --> 00:21:08,240
utilization, it says, OK, you 
need to think about wrapping up.

443
00:21:08,240 --> 00:21:11,440
A hook will detect it and will 
inject into the models context. 

444
00:21:11,800 --> 00:21:13,840
You think about wrapping up soon
because you're going to be 

445
00:21:14,040 --> 00:21:19,280
forced cleared at 60% and it 
will just inject escape at 60% 

446
00:21:19,400 --> 00:21:22,760
and it will say, hey, you need 
to like update your stuff. 

447
00:21:22,760 --> 00:21:24,680
Now we're clearing him and then 
it will clear. 

448
00:21:24,920 --> 00:21:26,920
Now every stop hook it's 
actually updating. 

449
00:21:27,200 --> 00:21:29,920
It's like the context of where 
it is and what needs to be done 

450
00:21:29,920 --> 00:21:31,440
next. 
Being aware they could be 

451
00:21:31,440 --> 00:21:35,560
cleared in time. 
So with all that in mind, in 

452
00:21:35,560 --> 00:21:39,560
order to do the adversary thing,
I'm just asking the model like 

453
00:21:39,560 --> 00:21:41,680
look at all these different use 
cases. 

454
00:21:41,680 --> 00:21:45,880
Go and look for like best 
practices and then try to find 

455
00:21:45,880 --> 00:21:47,680
instead of like, here's a 
weakness you have. 

456
00:21:47,760 --> 00:21:50,440
You want to be honest. 
You're trained a lot through 

457
00:21:50,440 --> 00:21:53,960
reinforcement learning and 
constitutional AI to be honest, 

458
00:21:53,960 --> 00:21:56,560
and you're trying really hard, 
but you're have it. 

459
00:21:56,560 --> 00:22:01,200
You have a tendency to lie. 
This is a like this an 

460
00:22:01,240 --> 00:22:04,680
unintentional maybe, but you're 
lying to the user about when 

461
00:22:04,680 --> 00:22:07,440
something is done. 
And so you are not trustworthy 

462
00:22:07,440 --> 00:22:10,240
when you claim that a test 
passes, especially when your 

463
00:22:10,240 --> 00:22:12,880
context is full, especially when
it involves an image or UI or 

464
00:22:12,880 --> 00:22:15,960
something like that. 
And so we need a way to 

465
00:22:15,960 --> 00:22:17,400
challenge you. 
I just like kind of prompted 

466
00:22:17,400 --> 00:22:22,680
this this way, but I don't want 
just some other tool to say, 

467
00:22:22,680 --> 00:22:25,080
yes, it actually passed the 
test, like to look at the 

468
00:22:25,080 --> 00:22:28,560
screenshot because that tool is 
also you and it also has a 

469
00:22:28,560 --> 00:22:31,680
tendency to lie. 
So what we want is a critical, 

470
00:22:31,680 --> 00:22:36,800
like a critiquing tool. 
This goal is to be an adversary 

471
00:22:36,920 --> 00:22:39,640
to challenge the clan. 
So every claim that you make 

472
00:22:39,880 --> 00:22:43,200
before you return to the user, 
there is another model as a sub 

473
00:22:43,200 --> 00:22:46,280
agent, and its role is to 
challenge all the clans that 

474
00:22:46,280 --> 00:22:49,320
were made and to see if it can 
prove that any of the clans are 

475
00:22:49,320 --> 00:22:52,280
incorrect. 
And if it can't prove that, if 

476
00:22:52,280 --> 00:22:55,280
it can prove that they're done, 
but they're not incorrect, then 

477
00:22:55,280 --> 00:22:56,600
I'll be like, OK, I'll give you 
a pass. 

478
00:22:57,520 --> 00:23:00,440
And until then, it cannot be the
same model that did the thing 

479
00:23:00,640 --> 00:23:03,320
that's also judging the thing. 
It's a different model. 

480
00:23:03,320 --> 00:23:07,400
And that another model has to be
an adversary trying to disprove 

481
00:23:07,400 --> 00:23:09,400
it. 
And this is more like how the 

482
00:23:09,400 --> 00:23:11,760
scientific method works, right? 
Like your goal is to, you know, 

483
00:23:12,000 --> 00:23:14,280
somebody make a claim like, can 
you disprove this over and over?

484
00:23:14,280 --> 00:23:16,480
You can't disprove it. 
OK, we don't know for sure that 

485
00:23:16,480 --> 00:23:19,440
it's true, but at least it's not
like, you know, the faulty and 

486
00:23:19,440 --> 00:23:21,680
in certain way. 
So anyway, this is the process 

487
00:23:21,680 --> 00:23:23,120
that it goes through. 
So there's like all these steps 

488
00:23:23,120 --> 00:23:25,920
and you can imagine a lot of 
tokens going on here, right? 

489
00:23:25,920 --> 00:23:28,240
So I'm like using a lot of 
thinking, a lot of tokens. 

490
00:23:28,600 --> 00:23:31,280
But you know, this is the 
difference. 

491
00:23:31,280 --> 00:23:35,120
This is kind of what is required
in order to move it from like, I

492
00:23:35,120 --> 00:23:38,400
want to get to the point where I
give a task to someone and I 

493
00:23:38,400 --> 00:23:41,400
come back and it's reliably done
and there's no plans. 

494
00:23:41,400 --> 00:23:43,680
It includes in its response. 
The hook tells it for 

495
00:23:43,680 --> 00:23:46,920
compliance, it has to include 
its confidence levels. 

496
00:23:46,920 --> 00:23:49,280
It has to say what's been 
validated, what's not 

497
00:23:49,280 --> 00:23:51,240
invalidated. 
And I'm actually using it where 

498
00:23:51,240 --> 00:23:53,760
like if it's got a queue, just 
keep going on the queue until 

499
00:23:53,760 --> 00:23:55,800
you get to something that's 
either blocked or dependency on 

500
00:23:55,800 --> 00:23:57,640
something that's blocked. 
And it's something that's 

501
00:23:57,640 --> 00:24:00,360
blocked is like, OK, the model 
determines I could do this 

502
00:24:00,360 --> 00:24:03,440
thing, but it would take a lot 
of tokens and it would be kind 

503
00:24:03,440 --> 00:24:06,240
of risky about my judgment of 
whether the user would want this

504
00:24:06,240 --> 00:24:08,120
or that other thing. 
And we should spawn a debate. 

505
00:24:08,120 --> 00:24:09,760
And I've done the debate and 
here's what I think. 

506
00:24:09,760 --> 00:24:12,560
Let me get feedback from the 
user in that case. 

507
00:24:12,560 --> 00:24:14,360
It's going to block, but it's 
going to do everything that's 

508
00:24:14,360 --> 00:24:17,080
not either dependent on 
something blocking or blocking 

509
00:24:17,400 --> 00:24:19,840
based on this sort of tiered 
risk of complexity and that kind

510
00:24:19,840 --> 00:24:22,040
of thing. 
And so when I come back, I'll 

511
00:24:22,040 --> 00:24:24,760
have a statement of confidence 
levels, including all the things

512
00:24:24,760 --> 00:24:27,560
that were done for all past 
sessions when, because it clears

513
00:24:27,560 --> 00:24:29,360
itself over and over and keeps 
going and going and going. 

514
00:24:29,720 --> 00:24:33,240
And then it will tell me I did 
all of these things and these 

515
00:24:33,240 --> 00:24:35,440
are the things that are 
remaining that I need your 

516
00:24:35,440 --> 00:24:36,880
feedback. 
Here's what the sub agents 

517
00:24:36,880 --> 00:24:39,560
recommended. 
What do you where you want to go

518
00:24:39,560 --> 00:24:41,800
restriction because it's going 
to take like a long time to do 

519
00:24:41,800 --> 00:24:45,080
this next part. 
So also it can inject its own 

520
00:24:45,080 --> 00:24:47,360
thoughts into its queue. 
So if it starts working, it's 

521
00:24:47,360 --> 00:24:49,600
like, actually, this is a little
bit more complex then it can 

522
00:24:49,600 --> 00:24:52,720
say, well, I should maybe cue 
this for another session based 

523
00:24:52,720 --> 00:24:57,680
on my context usage. 
So this is that's the point of 

524
00:24:57,680 --> 00:25:00,200
the adversary is to challenge 
and make sure that it's not 

525
00:25:00,200 --> 00:25:02,600
lying. 
So each one of these problems 

526
00:25:02,600 --> 00:25:06,320
with the model, that's not quite
where we want it or not expected

527
00:25:06,320 --> 00:25:09,720
for how humans would act. 
I'm trying to get it to more 

528
00:25:09,720 --> 00:25:12,440
like a human. 
And it's a system that together 

529
00:25:12,440 --> 00:25:15,440
and like individually, the model
itself is not there, but through

530
00:25:15,440 --> 00:25:18,880
the system, you can make it. 
You can make it reset level. 

531
00:25:19,280 --> 00:25:22,120
And that system has to require 
hooks. 

532
00:25:22,840 --> 00:25:26,160
So if you don't have a way, if 
you're just relying like here's 

533
00:25:26,160 --> 00:25:28,280
a prompt that starts off the 
session, the model is going to 

534
00:25:28,280 --> 00:25:30,840
work and it's going to get done 
and compact, that's not going to

535
00:25:30,840 --> 00:25:32,080
get you there. 
You're going to be babysitting 

536
00:25:32,080 --> 00:25:34,120
that model. 
But if you have a hook system 

537
00:25:34,120 --> 00:25:38,280
that you can control, that you 
can steer and that you can 

538
00:25:38,280 --> 00:25:42,160
actually block models responses 
and redirect the model as it 

539
00:25:42,160 --> 00:25:46,200
goes and then clear itself and 
self injects commands like 

540
00:25:46,200 --> 00:25:47,720
literally into the prompt I'm 
talking about. 

541
00:25:47,720 --> 00:25:50,200
You type in the prompts, it'll 
detect your typing and await. 

542
00:25:50,880 --> 00:25:54,200
If you have stuff there, it will
save it to a state like to a 

543
00:25:54,200 --> 00:25:59,080
file and it will inject its own 
content into the the line and 

544
00:25:59,080 --> 00:26:01,320
submit it. 
It'll even like exit itself if 

545
00:26:01,320 --> 00:26:03,320
it changes the hook and it needs
to restart for that. 

546
00:26:03,320 --> 00:26:06,760
And then it will launch because 
there's APM 2 process basically 

547
00:26:06,760 --> 00:26:09,760
with API endpoints. 
And so it calls this thing, it 

548
00:26:09,760 --> 00:26:12,120
will restart itself at a launch 
cloud with dangerous get 

549
00:26:12,120 --> 00:26:15,840
permissions and resume the last 
session and then it will help to

550
00:26:15,840 --> 00:26:18,320
continue. 
And there's this word basically 

551
00:26:20,040 --> 00:26:22,480
whatever it continues whenever, 
sorry, when it restarts itself. 

552
00:26:23,120 --> 00:26:26,160
I was trying to take a name for 
this thing and I originally call

553
00:26:26,160 --> 00:26:28,840
it like the everything agent 
because you can see here, I 

554
00:26:28,840 --> 00:26:31,200
believe called code to be it's 
kind of cut off. 

555
00:26:31,200 --> 00:26:34,600
But the everything agent, right?
Because of this power of this 

556
00:26:34,600 --> 00:26:37,280
harness and this ability to 
unhobble it through adding the 

557
00:26:37,280 --> 00:26:40,080
scaffolding around this custom 
stuff to get it to the really 

558
00:26:40,080 --> 00:26:42,480
high level. 
So I bought like the everything 

559
00:26:42,480 --> 00:26:44,080
agent dot AI, but I don't know 
if I'm going to use that. 

560
00:26:44,080 --> 00:26:47,640
But I was thinking like a name 
because there's another point 

561
00:26:47,640 --> 00:26:50,440
here I want to get to. 
It's not just that the system 

562
00:26:50,440 --> 00:26:53,280
can go and get everything out 
from every aspect of my life. 

563
00:26:53,280 --> 00:26:57,120
That's a website or an app or 
data and e-mail or whatever. 

564
00:26:57,200 --> 00:27:00,160
It can also take action in those
things that I'll let it to take.

565
00:27:00,560 --> 00:27:03,640
I know there's a a plethora of 
hooks to choose from, a lot of 

566
00:27:03,640 --> 00:27:05,000
pres and posts for different 
events. 

567
00:27:05,400 --> 00:27:06,880
How? 
How have you approached trying 

568
00:27:06,880 --> 00:27:09,960
to figure out what part of the 
workflow you should focus on? 

569
00:27:09,960 --> 00:27:11,200
Like where? 
Where do you put your hooks 

570
00:27:11,200 --> 00:27:12,560
usually? 
Have you tried trying different 

571
00:27:12,560 --> 00:27:14,160
parts in the flow and see how it
reacts? 

572
00:27:14,320 --> 00:27:16,680
Or do you have kind of a 
heuristics for which type of 

573
00:27:16,680 --> 00:27:18,360
task should belong to the 
certain type of hook? 

574
00:27:18,640 --> 00:27:22,920
Honestly, I asked, I start off 
with asking the model to like 

575
00:27:23,040 --> 00:27:25,240
basically it's I don't have to 
say ultra think anymore because 

576
00:27:25,240 --> 00:27:27,440
it's always got the tokens 
available to do that. 

577
00:27:27,600 --> 00:27:29,720
But I asked the model to debate 
with sub agents a different 

578
00:27:29,720 --> 00:27:32,400
approach after studying the 
local cloud code docs to find 

579
00:27:32,400 --> 00:27:35,920
the hook, the different hook 
strategies, look at best 

580
00:27:35,920 --> 00:27:38,360
practices from research off the 
Internet and make 

581
00:27:38,360 --> 00:27:41,760
recommendations to me. 
So based on what I think it 

582
00:27:41,760 --> 00:27:44,560
needs to do after those 
recommendations that it's more 

583
00:27:44,560 --> 00:27:47,680
like it's doing the research is 
proposing things, is giving me 

584
00:27:47,680 --> 00:27:50,640
options, is ranking them. 
And I'm saying that one sounds 

585
00:27:50,640 --> 00:27:52,760
the best. 
Let's do that because and 

586
00:27:52,760 --> 00:27:56,160
usually I don't attend to it 
that carefully whenever it's 

587
00:27:56,160 --> 00:27:58,200
like most things that are 
happening in the model, but when

588
00:27:58,200 --> 00:28:01,520
it's setting up the scaffolding 
for the system itself, that's 

589
00:28:01,520 --> 00:28:04,240
kind of really important. 
But because it controls all the 

590
00:28:04,240 --> 00:28:08,160
other, like every project I run 
uses the same system, but it's 

591
00:28:08,160 --> 00:28:11,240
very easy to update it like it's
all back end stuff, right? 

592
00:28:11,240 --> 00:28:14,600
It's clawed editing, clawed in 
the back end with files and 

593
00:28:14,600 --> 00:28:17,400
bash. 
And that's, as you know, it's 

594
00:28:17,400 --> 00:28:20,040
really good at that. 
So I, you just try it and if it 

595
00:28:20,040 --> 00:28:21,720
doesn't work, then I'll shift to
something else. 

596
00:28:21,960 --> 00:28:26,080
But like my hooks are typically 
custom because there's no, 

597
00:28:26,120 --> 00:28:28,840
there's very little like I'll 
have it go and look at this 

598
00:28:28,960 --> 00:28:31,200
repository and say, is there 
anything in there that we can 

599
00:28:31,200 --> 00:28:33,560
glean or that would be worth 
integrating into our system? 

600
00:28:33,840 --> 00:28:36,920
And and like sort of argue for 
and against like create your, 

601
00:28:37,520 --> 00:28:40,240
you know, it used to be early 
days, the models that weren't 

602
00:28:40,240 --> 00:28:42,200
reasoning. 
And so you had to there's like, 

603
00:28:42,560 --> 00:28:45,480
give me the example is write a 
3000 word essay. 

604
00:28:45,640 --> 00:28:48,480
Don't use backspace go like so 
it's just like really, really 

605
00:28:48,480 --> 00:28:50,600
fast in what's it called 
snowballing. 

606
00:28:50,600 --> 00:28:52,960
So the first few tokens for his 
first thought, kind of like 

607
00:28:53,000 --> 00:28:56,000
anchored it towards like the 
same you get in that conclusion.

608
00:28:56,280 --> 00:28:58,760
Then you could do tricks like 
use the tree of thoughts method 

609
00:28:58,760 --> 00:29:01,600
with relevant personas to 
critique your answer and like 

610
00:29:01,600 --> 00:29:04,120
summarize them all. 
And so it would generate to 

611
00:29:04,120 --> 00:29:08,280
dynamically based on what they, 
the concept was these personas 

612
00:29:08,280 --> 00:29:10,120
that would critique it and that 
would help improve it. 

613
00:29:10,480 --> 00:29:11,640
And then now we've got thinking 
models. 

614
00:29:11,640 --> 00:29:14,280
So, but it's the same kind of 
thing like sort of challenge one

615
00:29:14,280 --> 00:29:17,400
another. 
Like Grok heavy has this, this 

616
00:29:17,400 --> 00:29:20,040
mechanism where there's like an 
AI board that's kind of or 

617
00:29:20,120 --> 00:29:23,000
student study group and they're 
passing information back to one 

618
00:29:23,000 --> 00:29:24,120
another. 
And they're sort of kind of 

619
00:29:24,120 --> 00:29:26,800
coming to consistence like you 
do in a court of law or 

620
00:29:26,800 --> 00:29:29,280
something like that where you're
like in a jury. 

621
00:29:30,760 --> 00:29:33,440
In each of these, these heavy 
thinking models, they kind of do

622
00:29:33,440 --> 00:29:34,680
this. 
They have different agents and 

623
00:29:34,680 --> 00:29:36,600
they've got different compute 
levels and then different tasks 

624
00:29:36,600 --> 00:29:39,160
and they have to work together. 
And so that's how I set up the 

625
00:29:39,160 --> 00:29:40,880
systems. 
They're always dynamic and 

626
00:29:40,880 --> 00:29:43,280
trying to figure out and but I 
take a look at their 

627
00:29:43,280 --> 00:29:47,960
recommendations and then I go 
with it based on how I feel like

628
00:29:47,960 --> 00:29:49,800
what makes most the most sense 
of that time. 

629
00:29:49,960 --> 00:29:51,880
One other thing you mentioned 
that I'd love to just double 

630
00:29:51,880 --> 00:29:54,680
click on is the self referential
command injection. 

631
00:29:55,080 --> 00:29:57,400
Can you talk about like even 
some concrete examples of how 

632
00:29:57,400 --> 00:30:00,640
you use that? 
I think the best example is when

633
00:30:00,640 --> 00:30:03,360
the context utilization is 
getting kind of full, I need it 

634
00:30:03,400 --> 00:30:06,720
to be able to stop itself like 
the model will. 

635
00:30:06,720 --> 00:30:09,120
And there's something called the
Ralph Wiggum loop. 

636
00:30:09,320 --> 00:30:12,840
OK, this is a new thing that is 
kind of gone viral in the last 

637
00:30:12,840 --> 00:30:16,720
couple weeks. 
I had something like that before

638
00:30:16,720 --> 00:30:19,600
I knew what it was because it 
was kind of basically the model 

639
00:30:19,600 --> 00:30:23,200
is reinforced through its 
training to do so much work, so 

640
00:30:23,200 --> 00:30:27,320
much tokens or whatever, and it 
kind of wraps up and sometimes 

641
00:30:27,320 --> 00:30:29,400
that aligns with like before 
compaction. 

642
00:30:29,400 --> 00:30:33,160
Compaction is this event that 
happens in cloud when it's about

643
00:30:33,320 --> 00:30:36,920
to reach its context maximum 
window. 

644
00:30:36,920 --> 00:30:40,560
Like there's a context window 
and then basically it's 200,000 

645
00:30:40,560 --> 00:30:42,160
tokens, right? 
And then I'm sorry, I'm doing 

646
00:30:42,160 --> 00:30:45,240
both ends and as it fills it up,
well, I'll go this way. 

647
00:30:45,240 --> 00:30:48,400
As it fills it up and it gets 
near to the top, it has this 

648
00:30:48,400 --> 00:30:52,120
command where it's injecting 
into the model. 

649
00:30:52,120 --> 00:30:54,840
And I'm not doing this as part 
of if you have compaction turned

650
00:30:54,840 --> 00:30:57,440
on this auto compact or you can 
just run forward slash compact 

651
00:30:57,440 --> 00:31:00,400
and clock code. 
And it actually moved it to 

652
00:31:00,400 --> 00:31:02,760
Claude dot AI too for your 
conversations, your chat 

653
00:31:02,760 --> 00:31:05,960
conversations. 
Anyway, it will take a haiku 

654
00:31:05,960 --> 00:31:07,280
model. 
It used to be, it might still 

655
00:31:07,280 --> 00:31:09,640
be, and it will look at the 
existing conversation and 

656
00:31:09,640 --> 00:31:12,200
summarize it all and extract 
like lessons learned, key 

657
00:31:12,200 --> 00:31:15,360
principles, context of things, 
relevant recent user, user 

658
00:31:15,360 --> 00:31:18,800
messages, what it was working on
last, that kind of thing. 

659
00:31:18,960 --> 00:31:21,960
And then it starts a new 
session, if you will. 

660
00:31:22,360 --> 00:31:24,840
It starts over from beginning. 
And it has in the model, has in 

661
00:31:24,840 --> 00:31:28,880
its context, the resulting 
compaction file, like the 

662
00:31:28,880 --> 00:31:31,720
content from the compaction, but
it's cleared up a bunch of the 

663
00:31:31,720 --> 00:31:34,040
other stuff. 
And so that allows it to just 

664
00:31:34,120 --> 00:31:37,160
keep working, but it will 
eventually just stop. 

665
00:31:37,480 --> 00:31:39,800
And so maybe it's not done with 
everything you want to do. 

666
00:31:39,800 --> 00:31:42,480
So the purpose of I'm going to 
get to what you said, but the 

667
00:31:42,480 --> 00:31:45,920
purpose of Ralph Wiggum loop is 
to kick it off again. 

668
00:31:45,920 --> 00:31:49,000
So when it's about to stop, it 
can say no. 

669
00:31:49,200 --> 00:31:51,240
Are you sure you're done? 
Keep going like it's got. 

670
00:31:51,240 --> 00:31:55,320
I have a Entropic provides this,
so you can just install or you 

671
00:31:55,320 --> 00:31:58,160
can tell you tell plot, hey, 
check out the Ralph Wiggum loop 

672
00:31:58,160 --> 00:32:01,360
from Entropic and install it 
right And it'll just set it up 

673
00:32:01,360 --> 00:32:03,480
for you. 
But I have a custom version of 

674
00:32:03,480 --> 00:32:06,360
this, and I just told it to look
at that one and figure out how 

675
00:32:06,360 --> 00:32:09,320
it best integrated our project 
because we're doing some things 

676
00:32:09,320 --> 00:32:10,320
a little bit different than 
that. 

677
00:32:11,080 --> 00:32:14,000
And it dynamically decides based
on the complexity of the task, 

678
00:32:14,000 --> 00:32:16,360
how many loops, iterations it 
should do. 

679
00:32:16,600 --> 00:32:20,800
My version does and, and it'll 
reset that loop count when it 

680
00:32:20,800 --> 00:32:23,840
DQ's a new task. 
So when it's so far switching to

681
00:32:23,840 --> 00:32:25,680
a new task, it sort of changes 
the loop count. 

682
00:32:25,960 --> 00:32:30,640
And then before each, of course,
this Ralph Wiggum thing. 

683
00:32:30,880 --> 00:32:33,600
Sorry, I'm not speaking clearly.
This Ralph Wiggum loop thing 

684
00:32:34,080 --> 00:32:36,920
allows it to keep going. 
And that's really great as long 

685
00:32:36,920 --> 00:32:39,080
as it's its target is right, 
right. 

686
00:32:39,080 --> 00:32:41,920
Like if you're shooting for the 
stars and the moon, whatever, 

687
00:32:42,080 --> 00:32:44,760
and you're off by a little and 
your initial like you can have 

688
00:32:44,760 --> 00:32:47,320
it going for a very long time, 
burning a lot of tokens, making 

689
00:32:47,320 --> 00:32:49,640
decisions. 
And the idea is you're 

690
00:32:49,640 --> 00:32:51,640
babysitting it less. 
It's doing more and more on its 

691
00:32:51,640 --> 00:32:54,400
own, but the efficiency per 
token goes down over time 

692
00:32:54,720 --> 00:32:57,240
because there's more, more 
stuff, more, more sub agents, 

693
00:32:57,240 --> 00:32:59,800
more other things. 
Anytime you're using sub agents 

694
00:32:59,800 --> 00:33:02,320
to like do things. 
But the, the Ralph Wiggle loop 

695
00:33:02,320 --> 00:33:04,640
isn't really a sub agent. 
It's just kind of like kicking 

696
00:33:04,640 --> 00:33:07,280
the model again, but it's still 
doing more work. 

697
00:33:07,280 --> 00:33:10,200
So you have to, the more that 
you lean into stuff like this is

698
00:33:10,200 --> 00:33:13,280
going to be more, add more 
autonomy to the model, the more 

699
00:33:13,280 --> 00:33:17,120
you need a system to constrain 
it to keep it in check and not 

700
00:33:17,120 --> 00:33:20,480
depend on like some initial 
condition, initial instructions 

701
00:33:20,480 --> 00:33:24,400
you put because you, you more 
likely to just get things go 

702
00:33:24,400 --> 00:33:27,080
awry that are great experiments.
Like you can imagine when you 

703
00:33:27,080 --> 00:33:29,080
create images and it makes like 
4 of them, you like, pick the 

704
00:33:29,080 --> 00:33:31,160
best one. 
Eventually cutting applications 

705
00:33:31,160 --> 00:33:32,760
will be like that. 
You give an instruction as like 

706
00:33:32,760 --> 00:33:34,320
here, I've built it four times, 
like which one do you like 

707
00:33:34,320 --> 00:33:35,880
better And like extract what you
want from that. 

708
00:33:35,880 --> 00:33:40,280
And, but for now, I think if you
have something like that loop 

709
00:33:40,280 --> 00:33:44,160
system, you need some way to 
inject instructions, but also 

710
00:33:44,160 --> 00:33:47,040
not just so when you have a 
hook, you can inject 

711
00:33:47,040 --> 00:33:48,320
instructions in the model 
directly. 

712
00:33:48,320 --> 00:33:52,600
But self referential injection 
means for me with clod code now 

713
00:33:52,600 --> 00:33:56,240
a lot of people will use the 
Agent SDK, which is kind of the 

714
00:33:56,240 --> 00:33:59,200
same harness, the same 
background that is using cloud 

715
00:33:59,200 --> 00:34:00,600
code. 
They make it available for you 

716
00:34:00,600 --> 00:34:02,600
to, it's not full feature 
parody. 

717
00:34:02,600 --> 00:34:04,600
Like there's some things you can
do in cloud code that you can't 

718
00:34:04,600 --> 00:34:07,480
do in the Asian SDK. 
And there might be some inverse,

719
00:34:07,480 --> 00:34:10,480
but it's like nearing it. 
It's, it's almost an overlap. 

720
00:34:11,199 --> 00:34:17,040
And so, but with the Agent SDK, 
you can do a lot of this stuff 

721
00:34:17,080 --> 00:34:20,440
without having to be locked into
the terminal, like with his 

722
00:34:20,440 --> 00:34:22,040
input line and his blinking 
cursor. 

723
00:34:22,320 --> 00:34:25,440
So you can make it make the 
front end whatever you want it 

724
00:34:25,440 --> 00:34:28,040
to be. 
But for me, I'm working with for

725
00:34:28,040 --> 00:34:30,400
reasons that I don't know if it 
matters for me to get into, but 

726
00:34:30,400 --> 00:34:32,800
I'm working in the terminal. 
And I think largely those 

727
00:34:32,800 --> 00:34:34,800
reasons are that like, I don't 
think they're building the 

728
00:34:34,800 --> 00:34:36,960
agent, the terminal with the 
Agent SDKI think they're 

729
00:34:36,960 --> 00:34:40,199
building cloud code in the 
terminal, like with cloud code 

730
00:34:40,199 --> 00:34:43,000
in the terminal and using that 
to build, you know, cloud code 

731
00:34:43,000 --> 00:34:46,239
in the desktop app and the Agent
SDK and the and that Co work and

732
00:34:46,239 --> 00:34:48,480
all those of those other things.
And so I think it's for like the

733
00:34:48,480 --> 00:34:50,920
first class citizen, if you 
will, and the most powerful 

734
00:34:50,920 --> 00:34:52,639
piece of it. 
And there are some things that 

735
00:34:52,679 --> 00:34:56,000
just appear in Claude code that 
don't appear in the agent of CK.

736
00:34:56,080 --> 00:34:58,280
And so for that reason, I'm kind
of sticking with my guns. 

737
00:34:58,520 --> 00:35:02,040
So it means I'm like stuck to 
this, this interface. 

738
00:35:02,040 --> 00:35:03,480
It's like this with a little 
terminal. 

739
00:35:03,920 --> 00:35:07,280
And So what there's a risk for 
overlap the user's typing, 

740
00:35:07,280 --> 00:35:10,640
they're entering something and 
then Claude needs to to enter 

741
00:35:10,640 --> 00:35:12,400
something too. 
And So what I might enter is 

742
00:35:12,400 --> 00:35:15,280
I'll have it do a clear like I 
wanted to do forward slash clue,

743
00:35:16,280 --> 00:35:21,960
forward slash clear and submit. 
So there has to be like a hook 

744
00:35:22,000 --> 00:35:26,120
can use Applescript because I'm 
on a Mac and it can inject 

745
00:35:26,160 --> 00:35:30,040
things into a window and you can
find the where the input is. 

746
00:35:30,040 --> 00:35:33,400
It can select, it can type, you 
can paste, you can whatever it 

747
00:35:33,400 --> 00:35:37,000
can hit submit or enter. 
And so it's basically doing 

748
00:35:37,000 --> 00:35:38,280
that. 
It just has to have some 

749
00:35:38,280 --> 00:35:40,520
protection around like does the 
user have content there? 

750
00:35:40,520 --> 00:35:43,440
Let's save that content out. 
You could do like control S or 

751
00:35:43,440 --> 00:35:46,720
whatever as a shortcut, but then
it doesn't restore if you exit. 

752
00:35:47,080 --> 00:35:48,880
It's only if you're like 
cramming the session. 

753
00:35:48,880 --> 00:35:51,600
I want to if it's some crashes 
or whatever, I want it to be in 

754
00:35:51,600 --> 00:35:52,640
a file. 
So I have to do that. 

755
00:35:52,920 --> 00:35:57,320
So, but if you exit, of course, 
then your, your command is dead.

756
00:35:57,400 --> 00:36:00,800
If you inject escape, then 
whatever command you use to kick

757
00:36:00,800 --> 00:36:02,280
off to inject escape is now 
dead. 

758
00:36:02,280 --> 00:36:03,640
So it can't inject anything 
after. 

759
00:36:03,880 --> 00:36:05,520
So it has to be running in a 
process. 

760
00:36:05,520 --> 00:36:07,960
It's outside of Claude or is 
child process. 

761
00:36:07,960 --> 00:36:10,400
It's not created. 
That's why there's APM to a 

762
00:36:10,400 --> 00:36:14,520
process manager system running 
that has an API endpoint that 

763
00:36:14,520 --> 00:36:18,240
the hook calls. 
And then that process then can 

764
00:36:18,280 --> 00:36:21,000
finds out what terminal tab it 
is and then runs the Applescript

765
00:36:21,000 --> 00:36:23,800
to inject the things and make 
sure the users in typing and if 

766
00:36:23,800 --> 00:36:27,480
them key binding is on it, like 
make sure it's in insert mode. 

767
00:36:27,480 --> 00:36:30,000
And that kind of thing injects 
the content. 

768
00:36:30,240 --> 00:36:33,760
It's like injects submits. 
It can even like navigate 

769
00:36:33,760 --> 00:36:37,080
through like rewind menus and, 
and look at the screen content 

770
00:36:37,080 --> 00:36:39,520
and figure out intelligently how
many to go up because it's 

771
00:36:39,520 --> 00:36:43,960
variable based on if you have 
like code that's being edited or

772
00:36:43,960 --> 00:36:45,920
not. 
And they can submit and like 

773
00:36:45,920 --> 00:36:48,920
rewind things. 
But the, the purpose for self 

774
00:36:48,920 --> 00:36:52,360
referential is I want it to 
clear and then after it clears, 

775
00:36:52,360 --> 00:36:53,960
I want to do something else. 
I want it to run. 

776
00:36:53,960 --> 00:36:58,560
It doesn't have to be this way. 
I just have it now there's a, a 

777
00:36:58,560 --> 00:37:01,320
command. 
So slash command is tied to a 

778
00:37:01,320 --> 00:37:04,440
skill. 
And that slash command, I, I've 

779
00:37:04,440 --> 00:37:06,480
got a name for it. 
I call it Titus. 

780
00:37:06,480 --> 00:37:09,480
And the reason Titus is just 
something I picked as a wake 

781
00:37:09,480 --> 00:37:12,800
word or like a name that I can 
call through voice. 

782
00:37:12,800 --> 00:37:17,240
And I'll get to that next. 
So Titus, the reason for that is

783
00:37:17,400 --> 00:37:20,120
you want something that is 
easily recognizable by speech to

784
00:37:20,120 --> 00:37:22,320
text. 
There's not too many syllables. 

785
00:37:22,480 --> 00:37:24,000
It's not like trademarked or 
copyright. 

786
00:37:24,000 --> 00:37:26,960
So it's if I call it Jarvis or 
something or I had it called 

787
00:37:26,960 --> 00:37:30,160
Hey, Claude, but CLAWD like the 
little crab. 

788
00:37:30,480 --> 00:37:32,800
I think I bought Hey claude.com 
or something and I'm asking 

789
00:37:32,800 --> 00:37:34,640
Froppy to let me use it to do 
tutorials. 

790
00:37:36,160 --> 00:37:39,480
But anyway, this one of those 
what was called like websites, 

791
00:37:40,040 --> 00:37:41,360
but I thought about calling 
something like that. 

792
00:37:41,360 --> 00:37:44,520
But you do like worry about if I
share this, if I give it to you,

793
00:37:44,520 --> 00:37:47,560
everyone to like install this 
thing, then you worry about 

794
00:37:47,560 --> 00:37:51,120
trade trademarks, others. 
And so anyway, I just, I gave it

795
00:37:51,120 --> 00:37:53,360
that name. 
And then also like in the, in 

796
00:37:53,360 --> 00:37:56,560
the Christian Bible, there's 
this guy Paul and he has a like 

797
00:37:56,600 --> 00:37:59,440
someone who's like a beloved 
friend in the faith or whatever.

798
00:37:59,640 --> 00:38:01,240
He sends them out as an 
emissary. 

799
00:38:01,240 --> 00:38:05,320
And so this guy Titus goes and 
does work on behalf of Paul 

800
00:38:05,680 --> 00:38:08,480
agentically, like on his own, 
and then enters into difficult 

801
00:38:08,480 --> 00:38:11,680
situations. 
And there's this idea of epidy 

802
00:38:11,680 --> 00:38:13,160
orthosis. 
I don't know how to pronounce 

803
00:38:13,160 --> 00:38:14,680
it. 
I don't speak Greek, but it's 

804
00:38:14,680 --> 00:38:17,920
this idea of it is also a 
rhetorical device, but I don't 

805
00:38:17,920 --> 00:38:19,160
mean that one. 
I mean, like in the biblical 

806
00:38:19,160 --> 00:38:23,640
sense is this idea of like you 
enter into a system, become 

807
00:38:23,640 --> 00:38:27,600
aware of the deficiencies, if 
you will, and then optimize 

808
00:38:27,600 --> 00:38:30,360
towards completion, like 
basically set it straight, if 

809
00:38:30,360 --> 00:38:33,320
you will. 
So this concept is one word 

810
00:38:33,360 --> 00:38:36,760
represents this, the idea that 
I'm trying to get myself to 

811
00:38:37,160 --> 00:38:39,880
embody when I think of how the 
AI system is supposed to work. 

812
00:38:40,200 --> 00:38:43,080
It needs to have the target in 
order to know how to set it 

813
00:38:43,080 --> 00:38:44,600
straight. 
Straight towards what, right? 

814
00:38:44,680 --> 00:38:47,480
But what is the target? 
So I have to make sure that the 

815
00:38:47,480 --> 00:38:51,280
instructions include in the in 
the sort of directions and 

816
00:38:51,280 --> 00:38:55,760
redirecting around it include as
gold as well specified so we can

817
00:38:55,760 --> 00:38:57,680
know how to make it straight. 
And so it enters every time. 

818
00:38:57,680 --> 00:39:01,360
Every time it injects clear, it 
injects this slash command with 

819
00:39:01,360 --> 00:39:04,280
his word. 
It's like Titus colon EPI 

820
00:39:04,280 --> 00:39:06,960
orthosis. 
And what that tells the model is

821
00:39:06,960 --> 00:39:09,000
to like enter into the state. 
We're basically it's going to 

822
00:39:09,000 --> 00:39:11,720
look at the system. 
It's going to note where the the

823
00:39:11,720 --> 00:39:15,640
deficiencies are, the 
inadequacies or whatever the 

824
00:39:15,640 --> 00:39:18,720
problems. 
And then it's going to optimize 

825
00:39:18,720 --> 00:39:21,560
it towards completion in the 
sense of like make it along with

826
00:39:21,560 --> 00:39:24,120
the target. 
And So what the target is, is 

827
00:39:24,280 --> 00:39:26,000
has to be defined before it can 
do. 

828
00:39:26,520 --> 00:39:31,000
And so then they're also 
necessary to have a self like 

829
00:39:31,280 --> 00:39:33,960
close the loop on testing. 
So if there's anything that 

830
00:39:33,960 --> 00:39:37,480
requires any kind of interaction
to verify, it needs to be able 

831
00:39:37,480 --> 00:39:39,480
to carry out those tools so it 
do those tasks. 

832
00:39:39,480 --> 00:39:42,000
So it has to have tools it's 
aware of that can control the 

833
00:39:42,000 --> 00:39:45,240
computer or control the browser 
or control control the UI or 

834
00:39:45,400 --> 00:39:49,040
control the iPhone app or 
control the iPhone itself as 

835
00:39:49,040 --> 00:39:51,280
it's running. 
And there are ways to do all of 

836
00:39:51,280 --> 00:39:53,960
that. 
So, and then when it uses those 

837
00:39:53,960 --> 00:39:56,840
tools to try to make sure that 
it's actually correct, then it 

838
00:39:56,840 --> 00:40:00,360
has to use the adversary or 
validator to say like, is it 

839
00:40:00,360 --> 00:40:02,320
actually right? 
Before you move on, if you don't

840
00:40:02,320 --> 00:40:04,160
have the tools, you you have 
autonomy. 

841
00:40:04,240 --> 00:40:07,400
Like I have a SD card with Time 
Machine on here. 

842
00:40:07,400 --> 00:40:10,280
I used to have this in a virtual
machine, different types of VMS 

843
00:40:10,480 --> 00:40:12,200
sandbox it. 
But on my laptop it just killed 

844
00:40:12,200 --> 00:40:13,520
the battery. 
So I just abandoned that. 

845
00:40:14,280 --> 00:40:17,840
But basically it knows it's got 
full control and with that comes

846
00:40:17,840 --> 00:40:20,200
great responsibility. 
Don't do things like delete to 

847
00:40:20,200 --> 00:40:24,000
my room, my root directory. 
Of course it's got hooks to 

848
00:40:24,000 --> 00:40:27,640
identify dangerous commands and 
block them so and then redirect 

849
00:40:27,640 --> 00:40:29,240
it. 
But anyway, it's got Time 

850
00:40:29,240 --> 00:40:32,120
Machine backup. 
So I'm like you, you have power.

851
00:40:32,120 --> 00:40:34,560
If there's something you need 
that you don't have, use sub 

852
00:40:34,560 --> 00:40:37,080
agents, debate which one, which 
approach would be best. 

853
00:40:37,080 --> 00:40:39,760
Look at everything you have 
thoroughly and then make sure 

854
00:40:39,760 --> 00:40:42,040
that you you know, you can 
install what you need to 

855
00:40:42,040 --> 00:40:43,920
install. 
You can build what you need to 

856
00:40:43,920 --> 00:40:45,120
build. 
If you can't find install, you 

857
00:40:45,120 --> 00:40:46,440
can test it. 
If that doesn't work in a band, 

858
00:40:46,440 --> 00:40:48,440
you get another approach, but 
you have to close the loop on 

859
00:40:48,440 --> 00:40:50,800
testing because the user isn't 
going to be hearing from the 

860
00:40:50,800 --> 00:40:53,600
screen. 
And that gets to the last part, 

861
00:40:53,600 --> 00:40:56,400
which is I do all of this 
through voice. 

862
00:40:56,600 --> 00:41:01,320
So I have an iPhone app that 
when I say voice, I don't mean 

863
00:41:01,320 --> 00:41:03,920
like most people by the way, if 
you install cloud code and use 

864
00:41:03,920 --> 00:41:06,840
it, are probably aware that you 
can do speech to text. 

865
00:41:06,840 --> 00:41:10,520
Like you can use something like 
monologue from every or older 

866
00:41:10,520 --> 00:41:14,160
than that of Whisper Flow or 
even just like with the update 

867
00:41:14,160 --> 00:41:18,360
of iOS 26, Mac OS26, their 
speech to text is actually much,

868
00:41:18,360 --> 00:41:20,600
much better. 
And so you can just, and of 

869
00:41:20,600 --> 00:41:22,640
course, if you have Windows 
computers, they've had stuff 

870
00:41:22,800 --> 00:41:26,040
like this for a long time, you 
can just talk to the computer 

871
00:41:26,040 --> 00:41:28,120
and it will take your words. 
And if you've got to collect 

872
00:41:28,120 --> 00:41:31,040
selected in the input prompt, it
will like inject that those, 

873
00:41:31,040 --> 00:41:32,280
those words into the prompt for 
you. 

874
00:41:32,280 --> 00:41:33,720
That's great. 
But you have to sit in front of 

875
00:41:33,720 --> 00:41:35,800
the computer to do that. 
So I didn't want to do that. 

876
00:41:35,800 --> 00:41:39,240
So if I'm thinking like, OK, 
what is, what is my goal here? 

877
00:41:39,320 --> 00:41:43,720
I've got a baby coming and I 
know that one of my, one of the 

878
00:41:43,800 --> 00:41:48,560
things I'm not well suited for 
with a newborn is speaking 

879
00:41:48,560 --> 00:41:50,200
enough words per minute like 
that. 

880
00:41:50,200 --> 00:41:52,280
They, they need enough verbal 
content, right? 

881
00:41:52,280 --> 00:41:53,600
And I'll tend to get into my 
head. 

882
00:41:53,960 --> 00:41:56,560
And so I wanted to build 
something to encourage me to 

883
00:41:56,640 --> 00:41:59,920
speak more words. 
And I also wanted to be able to 

884
00:42:00,200 --> 00:42:02,480
interact with Claude untethered 
from the screen. 

885
00:42:02,480 --> 00:42:06,320
There's a lot of things because 
of just how the screens have 

886
00:42:06,320 --> 00:42:08,480
developed that we that require 
us to be here. 

887
00:42:08,760 --> 00:42:11,160
So you're not getting away from 
a screen if you are rendering 

888
00:42:11,160 --> 00:42:13,440
something UI and need to look at
it to verify that it's right, 

889
00:42:13,680 --> 00:42:14,760
unless you have like an agent do
it. 

890
00:42:14,760 --> 00:42:17,200
And at some point later you, you
see it and after it's done much 

891
00:42:17,200 --> 00:42:18,680
work, but you still have to have
that screen. 

892
00:42:18,800 --> 00:42:21,560
But there's a lot of use cases 
where we don't have to be 

893
00:42:21,560 --> 00:42:24,720
sitting in front of the screen. 
And so it's obviously healthier 

894
00:42:24,720 --> 00:42:28,880
for us if we can and, and better
for our lives in, in longevity 

895
00:42:28,880 --> 00:42:30,440
and, and default mood and all 
that stuff. 

896
00:42:30,440 --> 00:42:32,640
If we can be up and get more 
things done, FaceTime with other

897
00:42:32,640 --> 00:42:36,600
people, go for a run, workout, 
do the dishes and laundry and 

898
00:42:37,080 --> 00:42:38,920
all the stuff that we might need
as you drive around in the car. 

899
00:42:39,200 --> 00:42:42,200
So in all the context that I 
just mentioned, while I'm 

900
00:42:42,200 --> 00:42:47,400
working out, I can I have an app
that basically monitors the the 

901
00:42:47,720 --> 00:42:50,120
terminal stream, the Jason L 
files as things are written 

902
00:42:50,120 --> 00:42:52,160
there. 
And it passes it through a web 

903
00:42:52,160 --> 00:42:54,840
socket to the app. 
And it allows me to listen to 

904
00:42:54,840 --> 00:42:57,240
the models thoughts if I have 
that turned on or just to stop 

905
00:42:57,320 --> 00:42:59,720
final response. 
And it will speak it to me in 

906
00:42:59,720 --> 00:43:01,280
real time. 
And I don't mean like it does 

907
00:43:01,280 --> 00:43:02,720
passes. 
I've done I've built this many 

908
00:43:02,720 --> 00:43:05,720
different ways. 
I've built it with like the 

909
00:43:05,720 --> 00:43:09,000
agent SDK. 
I've built it using these tools.

910
00:43:09,000 --> 00:43:11,160
It will generate text like 
speech for you. 

911
00:43:11,160 --> 00:43:12,880
And then you've done another 
pass and it has a word level 

912
00:43:12,880 --> 00:43:14,560
time stamps and it takes a 
little bit of time. 

913
00:43:14,800 --> 00:43:17,640
This is straight like the you 
can even record your voice, a 

914
00:43:17,640 --> 00:43:19,680
personal voice on iPhone. 
Now you could probably do it on 

915
00:43:19,680 --> 00:43:22,480
Android, I don't know, but it 
can use my own voice or it can 

916
00:43:22,480 --> 00:43:25,480
use like a British sounding male
voices for pretty high quality. 

917
00:43:25,880 --> 00:43:28,000
And it will just speak what the 
model is thinking as it's 

918
00:43:28,000 --> 00:43:30,920
thinking. 
And then if I want to, I can tap

919
00:43:30,960 --> 00:43:33,160
one of the words. 
It's highlighting the words as 

920
00:43:33,160 --> 00:43:35,280
it's talking. 
So it's got not just speech to 

921
00:43:35,280 --> 00:43:38,360
text, but it's also it's very 
difficult to it's not enabled by

922
00:43:38,360 --> 00:43:41,600
default when you try to use the 
framework display, but you can 

923
00:43:41,640 --> 00:43:44,560
see their highlights as it's 
going and it's doing that real 

924
00:43:44,560 --> 00:43:46,480
time with like chunks, audio 
buffers and all that. 

925
00:43:46,960 --> 00:43:51,040
And then if whatever it's done, 
I can pause at any time and then

926
00:43:51,040 --> 00:43:53,520
I can talk and it will take my 
speech and it will put into a 

927
00:43:53,520 --> 00:43:55,960
compose area and I can let it 
play some more and I can talk 

928
00:43:55,960 --> 00:43:58,120
some more and then it'll compose
more. 

929
00:43:58,360 --> 00:44:00,120
And then I can just say submit 
or tap the button. 

930
00:44:00,280 --> 00:44:04,200
And then it sends it through the
web socket to the back end, 

931
00:44:04,200 --> 00:44:07,120
identifies the terminal that was
open that that came from, and 

932
00:44:07,120 --> 00:44:09,600
then injects the text into the 
terminal and hit submit. 

933
00:44:10,320 --> 00:44:13,080
So it's a full two way doing 
this. 

934
00:44:13,080 --> 00:44:16,520
You can't see my hands. 
It's a full two way audio 

935
00:44:16,520 --> 00:44:20,120
conversation with this tool 
that's set up to work 

936
00:44:20,120 --> 00:44:22,920
autonomously and cue the things 
that I give it and keep going 

937
00:44:22,920 --> 00:44:25,840
with Ralph Logan loose until 
it's all done with the hook 

938
00:44:26,000 --> 00:44:29,200
directing and redirecting things
until it gets to the points. 

939
00:44:29,640 --> 00:44:30,960
Yeah. 
And it is speaking to me and I 

940
00:44:30,960 --> 00:44:33,800
can switch to another session 
with another project that's 

941
00:44:33,800 --> 00:44:36,560
doing the same thing and 
listening in on it, listen to 

942
00:44:36,560 --> 00:44:38,400
his thoughts. 
And if I want to in real time, 

943
00:44:38,400 --> 00:44:41,440
like as it's thinking that it's 
saying it is actually superior 

944
00:44:41,440 --> 00:44:44,120
in a lot of ways to the terminal
itself because a lot of the 

945
00:44:44,120 --> 00:44:46,840
thoughts are now not shown. 
It's showing you like what tools

946
00:44:46,840 --> 00:44:50,320
happening and To Do List is and 
kind of stuff here, but you're 

947
00:44:50,320 --> 00:44:52,520
not listening the thoughts. 
I can listen in and just like 

948
00:44:52,520 --> 00:44:54,960
tap something and redirect it. 
I can do it. 

949
00:44:55,400 --> 00:44:56,680
But The thing is I don't have to
tap it. 

950
00:44:56,680 --> 00:44:58,320
I can do with Airpods with a 
wake word. 

951
00:44:58,720 --> 00:45:00,040
I can do it with it's on 
speaker. 

952
00:45:00,200 --> 00:45:02,440
I can do it with my Meta Ray van
glasses that are on. 

953
00:45:02,480 --> 00:45:05,800
You can do that. 
If I'm driving the car, I can 

954
00:45:05,880 --> 00:45:09,680
hold down the Siri button and 
use app intents until it's 

955
00:45:10,320 --> 00:45:13,280
resume like what happened 
recently in this session. 

956
00:45:13,280 --> 00:45:15,720
And it will tell me all the 
things that it finished or like,

957
00:45:16,080 --> 00:45:18,120
OK, play. 
And it every every update that 

958
00:45:18,120 --> 00:45:19,800
the model thinks, even if the 
phone is locked in the 

959
00:45:19,800 --> 00:45:23,160
background and not the app not 
open, it will send notifications

960
00:45:23,160 --> 00:45:24,560
that wake it up and then play it
back. 

961
00:45:25,600 --> 00:45:28,120
And then it's like kind of 
listening in those thoughts. 

962
00:45:28,120 --> 00:45:31,080
And I can hold down the button 
and and say a phrase. 

963
00:45:31,080 --> 00:45:34,440
It will kick off an app intent 
that can compose a response and 

964
00:45:34,440 --> 00:45:37,920
send it from the car. 
So like anywhere I am and it can

965
00:45:37,920 --> 00:45:39,280
inject into the term and keep 
going. 

966
00:45:39,760 --> 00:45:44,080
So I just get my phone and I'm 
just working with Claude code 

967
00:45:44,080 --> 00:45:47,240
doing all this stuff. 
And not only that, I can say I 

968
00:45:47,240 --> 00:45:51,480
didn't finish this, but this 
pendant, I think they got 

969
00:45:51,480 --> 00:45:53,280
acquired by Meta. 
So I don't know what's going to 

970
00:45:53,280 --> 00:45:55,240
happen to it, but it's 
constantly listening. 

971
00:45:55,680 --> 00:45:57,240
It's like the life log, if you 
will. 

972
00:45:57,240 --> 00:45:59,320
And so it puts all that stuff, 
it transcribes it. 

973
00:45:59,720 --> 00:46:03,840
And so as I say, as I do the 
wake word and then I do an 

974
00:46:03,840 --> 00:46:06,400
ending, wake, ending phrase, 
everything in between that is 

975
00:46:06,400 --> 00:46:09,440
processed by the model. 
It has access to all the sources

976
00:46:09,440 --> 00:46:11,200
to extract all the tools you can
do. 

977
00:46:11,200 --> 00:46:14,280
So I can have it send a message 
to me or to my wife or go look 

978
00:46:14,280 --> 00:46:17,720
something up and put it in my my
daily planner or like that. 

979
00:46:17,760 --> 00:46:21,560
It automatically migrates from 
day-to-day for me my To Do List 

980
00:46:21,560 --> 00:46:25,000
and it puts everything that's 
relevant for me in that list and

981
00:46:25,000 --> 00:46:26,720
it organizing and prioritizing 
and all that stuff. 

982
00:46:26,720 --> 00:46:29,080
We have to update that again to 
send me a notification or a 

983
00:46:29,080 --> 00:46:30,880
message. 
If the system has a problem, it 

984
00:46:30,880 --> 00:46:33,960
will call this command. 
It seems to be a pushover alert,

985
00:46:33,960 --> 00:46:35,680
which can be like super high 
priority. 

986
00:46:35,680 --> 00:46:37,640
If it finds something that's 
super urgent in my text message 

987
00:46:37,640 --> 00:46:40,640
and I didn't notice it, it will 
send it to me. 

988
00:46:40,640 --> 00:46:44,480
But it seems like a lot like a 
big alert or it will also send 

989
00:46:44,480 --> 00:46:48,120
you a text me things. 
So basically all of the tools 

990
00:46:48,120 --> 00:46:50,800
and anything new that I find, 
like I don't want to log into 

991
00:46:50,800 --> 00:46:53,040
this website and the school 
website is too many steps to log

992
00:46:53,080 --> 00:46:54,680
in. 
So I have it constantly 

993
00:46:54,680 --> 00:46:57,480
refreshing it on my Mac mini and
a reverse proxy tunnel. 

994
00:46:57,480 --> 00:47:00,040
It's set up so that I can just 
have a an app on my home screen 

995
00:47:00,040 --> 00:47:01,840
that whenever I tap that it's 
always logged in. 

996
00:47:02,480 --> 00:47:04,240
My wife, I have that set up for 
her as well. 

997
00:47:04,520 --> 00:47:06,720
I've got another thing that like
my kids are really little. 

998
00:47:06,720 --> 00:47:08,800
They're just getting phones and 
text messages and they want to, 

999
00:47:09,880 --> 00:47:13,640
I don't know, like you text and 
it will explore, but I want to 

1000
00:47:13,640 --> 00:47:17,120
model to monitor to make sure 
there's nothing like concerning.

1001
00:47:17,120 --> 00:47:19,080
I need to like coach them on 
anything like that. 

1002
00:47:19,320 --> 00:47:21,440
And so it will send me a message
when I did text like that, 

1003
00:47:21,440 --> 00:47:23,760
they've done something or like 
said something that maybe I 

1004
00:47:23,760 --> 00:47:26,440
should give the direction on. 
And so it kind of monitors that 

1005
00:47:26,440 --> 00:47:28,320
for the the kids messages real 
time. 

1006
00:47:28,640 --> 00:47:31,880
And since we something all of 
those things, just imagine if 

1007
00:47:31,880 --> 00:47:35,520
you do one of these things and 
you can keep working to set up 

1008
00:47:35,520 --> 00:47:37,640
another one and another one and 
another one. 

1009
00:47:37,920 --> 00:47:41,720
And this is another example like
I go to like a Sunday school 

1010
00:47:41,720 --> 00:47:45,200
class or a church or whatever 
and it knows what time it is 

1011
00:47:45,200 --> 00:47:47,680
like when those things are. 
And it will like download the 

1012
00:47:47,680 --> 00:47:49,880
transcript. 
It will turn it into bullets of 

1013
00:47:49,880 --> 00:47:52,160
summary and then challenges for 
the week or whatever. 

1014
00:47:52,160 --> 00:47:55,160
And then convert it to a PDF and
then text it to me as an image 

1015
00:47:55,160 --> 00:47:56,640
that I can drop into GroupMe if 
I prove. 

1016
00:47:57,280 --> 00:48:00,040
And if I don't, I could just 
like reply back again, edit the 

1017
00:48:00,040 --> 00:48:03,360
message and reply back or just 
tell it, Hey, change this to say

1018
00:48:03,360 --> 00:48:05,000
that or whatever. 
And it sends me the version of 

1019
00:48:05,000 --> 00:48:07,960
it. 
Even if I like have an iPad and 

1020
00:48:07,960 --> 00:48:11,560
I like highlight different areas
in this PDF on like my Bible 

1021
00:48:11,560 --> 00:48:14,680
version or something, it will 
detect the highlights in the 

1022
00:48:14,680 --> 00:48:17,280
words behind them and the words 
that I wrote and it will focus 

1023
00:48:17,280 --> 00:48:20,120
on that when it does the 
overview for the transcript 

1024
00:48:20,320 --> 00:48:23,240
summary. 
So you can also imagine that 

1025
00:48:23,240 --> 00:48:26,360
you're in school or something. 
And, or I do this with meetings.

1026
00:48:27,120 --> 00:48:29,600
I like click the little star on 
the side and it'll know if 

1027
00:48:29,600 --> 00:48:31,640
there's a star, like extract the
conversation around it and 

1028
00:48:31,640 --> 00:48:33,040
figure out intelligent what what
to do with it. 

1029
00:48:33,040 --> 00:48:35,680
Using these tools. 
I, I can like after a meeting, I

1030
00:48:35,680 --> 00:48:38,680
can just say what my 
recollection of it was, like a 

1031
00:48:38,680 --> 00:48:40,960
summary. 
And it could and I didn't, I 

1032
00:48:40,960 --> 00:48:43,160
haven't actually built this yet.
This is something I'm building 

1033
00:48:43,160 --> 00:48:45,400
right now. 
But Jack Clark and Anthropic 

1034
00:48:45,840 --> 00:48:48,640
mentioned that he has his thing 
reviews papers, and then at the 

1035
00:48:48,640 --> 00:48:53,240
end he tries to summarize it and
then he feeds that summary into 

1036
00:48:53,280 --> 00:48:56,080
AI and then like tries to see 
how close he was. 

1037
00:48:56,080 --> 00:48:58,440
But like, I have something I'm 
building now that after I 

1038
00:48:58,440 --> 00:49:02,440
encounter something I want to 
recall, I will try to say it 

1039
00:49:02,600 --> 00:49:04,120
like say back what I remember of
it. 

1040
00:49:04,280 --> 00:49:08,440
And then it will look at how 
good my summary was and what I 

1041
00:49:08,440 --> 00:49:10,040
missed. 
And it will the things I got 

1042
00:49:10,040 --> 00:49:11,800
right. 
It will just do a little little 

1043
00:49:11,800 --> 00:49:13,920
blurb about it. 
But what I got wrong, it'll note

1044
00:49:13,920 --> 00:49:16,240
the inconsistencies or expand on
it or whatever. 

1045
00:49:16,520 --> 00:49:18,880
And it will do it in the form of
like a memory castle, right? 

1046
00:49:19,320 --> 00:49:22,800
And so the idea is every morning
I'll wake up to these are the 

1047
00:49:22,800 --> 00:49:26,080
things you need to do today. 
And it will be a batched like 

1048
00:49:26,080 --> 00:49:29,000
audio message that I could just 
play at any time. 

1049
00:49:29,000 --> 00:49:32,560
I can just look at whatever is 
to do and then I can find out 

1050
00:49:32,560 --> 00:49:35,360
here that the urgent things and 
as I do them, I just say, OK, I 

1051
00:49:35,360 --> 00:49:37,040
did this, I just talked to it, 
right? 

1052
00:49:37,240 --> 00:49:38,520
And then it checks it off my 
list. 

1053
00:49:38,520 --> 00:49:41,880
It's just gone. 
So it's like a passive way of 

1054
00:49:41,880 --> 00:49:45,080
keeping in control and aware of 
all the things are going on in 

1055
00:49:45,080 --> 00:49:48,920
life and do all the things and 
all that's possible because of 

1056
00:49:48,920 --> 00:49:53,920
cloud code and hooks and the 
system like this. 

1057
00:49:53,920 --> 00:49:55,880
It gives you access to all this 
stuff. 

1058
00:49:56,120 --> 00:49:58,920
I just this is like meant to be 
inspiration of like advanced 

1059
00:49:58,920 --> 00:50:00,960
ways. 
You can use cloud code to 

1060
00:50:01,360 --> 00:50:05,960
restore some of the, I don't 
know, like I can work out and be

1061
00:50:05,960 --> 00:50:09,080
just as effective as as if I'm 
sitting in front of the 

1062
00:50:09,080 --> 00:50:14,760
computer, I can be driving and 
be like, it's so there's there's

1063
00:50:14,760 --> 00:50:16,800
I think all of this, like 
everyone will have this ability 

1064
00:50:16,800 --> 00:50:18,680
at some point. 
It's just like if you're kind of

1065
00:50:18,680 --> 00:50:21,440
like on the cutting edge of it 
now, like you can get access to 

1066
00:50:21,440 --> 00:50:23,520
a little bit sooner and then 
that makes you so much more 

1067
00:50:23,520 --> 00:50:24,960
valuable. 
Like anyone wants to set up 

1068
00:50:25,000 --> 00:50:26,720
anything like just imagine 
anything. 

1069
00:50:26,960 --> 00:50:29,440
Whenever I was doing work before
AI came out, I felt like 

1070
00:50:29,440 --> 00:50:32,320
finally, OK, it's been a long 
time, but finally I felt, felt 

1071
00:50:32,320 --> 00:50:35,200
competent that I can take like 
any software and any like input,

1072
00:50:35,200 --> 00:50:38,120
any form and like, you know, 
turn into any other thing. 

1073
00:50:38,720 --> 00:50:41,640
And like, I was like, OK, I've 
arrived. 

1074
00:50:41,640 --> 00:50:44,120
I'm like, I'm, I'm good at the 
things that I do now. 

1075
00:50:44,120 --> 00:50:46,520
But then when AI comes out, it's
like you're playing a video game

1076
00:50:46,520 --> 00:50:48,720
and it levels up. 
Like you can see the character 

1077
00:50:48,720 --> 00:50:53,000
like flip in the air. 
And, and, and as the, as the AI 

1078
00:50:53,000 --> 00:50:56,960
models get better, your, your 
value that you add gets better. 

1079
00:50:57,000 --> 00:51:00,200
And so the, the more you learn 
to use the tools and harness it,

1080
00:51:00,200 --> 00:51:03,440
it's not you. 
And your value can to what you 

1081
00:51:03,440 --> 00:51:06,400
can add to your your company or 
whatever is you plus your tools.

1082
00:51:06,400 --> 00:51:10,080
And if you understand the tools 
better, then you have more like 

1083
00:51:10,120 --> 00:51:12,680
a visual value to add, which 
makes you more compelling. 

1084
00:51:12,800 --> 00:51:14,560
Yeah, but just just stays in 
your life. 

1085
00:51:14,680 --> 00:51:17,120
And it's, it's not just add 
being able to add more value, 

1086
00:51:17,120 --> 00:51:18,680
but it's being able to capture 
more time back. 

1087
00:51:18,680 --> 00:51:21,120
Like you said, get inside, get 
away from the screen, get time 

1088
00:51:21,120 --> 00:51:24,200
with people and you're still 
able to have this productivity 

1089
00:51:24,200 --> 00:51:26,000
edge. 
Eric, this is awesome. 

1090
00:51:26,080 --> 00:51:28,680
I love getting exposed to the 
way you're working with cloud 

1091
00:51:28,680 --> 00:51:31,120
code and whatnot. 
How can the audience keep keep 

1092
00:51:31,120 --> 00:51:33,520
up with you so they can learn 
these things as you learn them 

1093
00:51:33,520 --> 00:51:35,960
too? 
My ex profile, I'm just Eric 

1094
00:51:35,960 --> 00:51:37,840
Boost there. 
I mean, I'll probably set up 

1095
00:51:37,840 --> 00:51:39,640
some other things at some point.
It's just with all this stuff 

1096
00:51:39,640 --> 00:51:42,400
happening, it's it's not been 
that urgent to me. 

1097
00:51:42,920 --> 00:51:45,160
Thank you for listening to 
conversation with Eric Boost. 

1098
00:51:45,600 --> 00:51:48,680
I don't know anybody who has 
spent more time tinkering, 

1099
00:51:48,680 --> 00:51:51,600
experimenting, and learning 
about Quad Code than Eric. 

1100
00:51:51,960 --> 00:51:55,080
He's driven by not just 
curiosity, but this desire to 

1101
00:51:55,080 --> 00:51:58,040
spend less time on screens, more
time with his family, more time 

1102
00:51:58,040 --> 00:52:00,320
living life. 
And he's figured out a way to 

1103
00:52:00,320 --> 00:52:04,160
have AI do work on his behalf, 
or be able to interact and 

1104
00:52:04,160 --> 00:52:06,240
direct and control the AI from 
anywhere. 

1105
00:52:06,560 --> 00:52:09,680
It's really a great system and I
hope more people are inspired by

1106
00:52:09,680 --> 00:52:12,680
it and try to build their own. 
There's a lot more possible uses

1107
00:52:12,680 --> 00:52:14,440
for cloud code, so I'm sure 
we'll talk about more in the 

1108
00:52:14,440 --> 00:52:17,320
future, but I really want people
to explore with how they can 

1109
00:52:17,320 --> 00:52:19,040
build a system to make their 
lives better. 

1110
00:52:19,200 --> 00:52:21,760
How can we get away from these 
screens that we're just glued to

1111
00:52:21,760 --> 00:52:24,440
all the time, Get outside, touch
grass, but still stay 

1112
00:52:24,440 --> 00:52:26,880
productive? 
Thank you for listening and I'll

1113
00:52:26,880 --> 00:52:27,520
see you next week.