1
00:00:00,080 --> 00:00:04,280
In this episode, I want to walk 
you through how to build a truly

2
00:00:04,280 --> 00:00:09,400
AI first team of business. 
This month, Y Combinator, one of

3
00:00:09,400 --> 00:00:12,640
the most famous start up 
accelerators and venture capital

4
00:00:12,640 --> 00:00:17,960
firms in the entire world who 
have been the place that Airbnb,

5
00:00:18,280 --> 00:00:22,800
Coinbase, Dropbox, Reddit, 
Stripe, basically most of the 

6
00:00:22,800 --> 00:00:26,000
biggest businesses in the entire
world have come from. 

7
00:00:26,560 --> 00:00:29,840
And they share incredible 
insights every single quarter. 

8
00:00:30,240 --> 00:00:33,880
And this month they shared a 
playbook for building what they 

9
00:00:33,880 --> 00:00:37,360
see as an AI first company. 
And the contrast here is not 

10
00:00:37,360 --> 00:00:41,280
just somebody or a team or a 
business that uses AI, but that 

11
00:00:41,280 --> 00:00:44,720
is built on AI. 
They're the very heart of every 

12
00:00:44,720 --> 00:00:48,520
new idea that is emerging in the
world, based in San Francisco 

13
00:00:48,720 --> 00:00:52,240
around the most innovative 
people and ideas you can 

14
00:00:52,240 --> 00:00:55,360
possibly imagine. 
They showed 6 principles from 

15
00:00:55,360 --> 00:01:00,080
what they're seeing with all the
latest hot AI start-ups that 

16
00:01:00,080 --> 00:01:02,200
they're working with. 
And I'm going to talk about a 

17
00:01:02,200 --> 00:01:05,200
very publication of how this is 
actually playing out real time 

18
00:01:05,640 --> 00:01:09,880
by Jack Dorsey, the ex founder 
of Twitter, who has a new 

19
00:01:09,880 --> 00:01:12,920
company called Block. 
I'm going to explain the 

20
00:01:12,920 --> 00:01:14,800
playbook. 
I'm going to talk about what 

21
00:01:14,800 --> 00:01:17,760
that actually looks like when 
it's running properly and give 

22
00:01:17,760 --> 00:01:21,720
you some ideas to bring to your 
next team or business meeting. 

23
00:01:22,320 --> 00:01:23,720
This is in the loop with Jack 
Hot. 

24
00:01:23,720 --> 00:01:37,740
I hope you enjoy the show. 
So let's start with a bit of 

25
00:01:37,740 --> 00:01:41,540
background. 
So about 10 days ago, so about 

26
00:01:41,540 --> 00:01:46,540
24th of April, Diana Hugh, a 
general partner at YC, he 

27
00:01:46,540 --> 00:01:50,380
basically overseas their AI 
companies program published what

28
00:01:50,380 --> 00:01:53,140
was called the playbook for 
building an AI native company. 

29
00:01:53,900 --> 00:01:56,740
And it blew up, of course, soon 
as it was shared. 

30
00:01:57,660 --> 00:02:00,680
And the reason it spread so fast
is that a, it's named something 

31
00:02:00,680 --> 00:02:03,520
that a lot of people are 
watching happen without quite 

32
00:02:03,520 --> 00:02:05,880
having the language for or the 
playbook for. 

33
00:02:06,200 --> 00:02:09,120
And the fact that it covers team
or company going through this AI

34
00:02:09,120 --> 00:02:11,400
journey. 
It was described as AI, as your 

35
00:02:11,400 --> 00:02:16,000
company's operating system, open
versus closed loop systems and 

36
00:02:16,000 --> 00:02:20,680
companies and teams making a 
company or team fully queryable.

37
00:02:21,440 --> 00:02:24,680
And the rise of 1000 ex 
employee, team member and 

38
00:02:24,680 --> 00:02:26,640
engineer. 
And essentially why this means 

39
00:02:26,640 --> 00:02:29,480
that middle management 
disappears and why start-ups or 

40
00:02:29,480 --> 00:02:32,480
smaller companies have a huge 
opportunity to win right now 

41
00:02:32,960 --> 00:02:35,280
because of the natural 
challenges that big companies 

42
00:02:35,280 --> 00:02:38,000
face. 
And what's quite interesting is 

43
00:02:38,240 --> 00:02:41,920
a couple months earlier, Jack 
Dorsey, who has basically run a 

44
00:02:41,920 --> 00:02:44,400
lot of these principles real 
time. 

45
00:02:45,560 --> 00:02:49,080
So Block, Jack Dorsey's company,
for anyone not familiar is, is a

46
00:02:49,080 --> 00:02:51,840
fintech company. 
So you might have seen Square, 

47
00:02:51,840 --> 00:02:54,560
the payment provider systems 
when you go to buy a coffee or 

48
00:02:54,560 --> 00:02:56,440
whatever. 
And it's basically a 

49
00:02:56,440 --> 00:03:00,520
peer-to-peer money app used by 
about 57 million Americans and 

50
00:03:00,520 --> 00:03:03,760
companies. 
And in February of 2026, he 

51
00:03:03,760 --> 00:03:07,320
decided to cut 40% of the entire
block workforce in a single 

52
00:03:07,320 --> 00:03:09,440
letter, not because it was 
struggling. 

53
00:03:09,440 --> 00:03:13,360
The business actually profit was
up, you know, nearly 25% over 

54
00:03:13,360 --> 00:03:16,240
the last quarter, but actually 
because the architecture of the 

55
00:03:16,240 --> 00:03:19,680
company he believed no longer 
matched what the technology of 

56
00:03:19,680 --> 00:03:23,040
LLMS now makes possible. 
And so when she published this 

57
00:03:23,040 --> 00:03:26,200
playbook, you know, two months 
later, the block story was a 

58
00:03:26,200 --> 00:03:29,440
really fantastic case study of 
this playing out in real time. 

59
00:03:29,720 --> 00:03:32,080
And I'm not placing a value 
judgement or whether that is 

60
00:03:32,080 --> 00:03:34,960
good or bad right now. 
I'm just going to talk about 

61
00:03:34,960 --> 00:03:37,800
what an AI first company and 
team really looks like. 

62
00:03:55,340 --> 00:03:58,060
So let's get into the playbook. 
So there's 8 principles and and 

63
00:03:58,060 --> 00:04:01,300
principle number one is that AI 
is the operating system, not 

64
00:04:01,300 --> 00:04:03,700
just a tool. 
So most companies right now are 

65
00:04:04,020 --> 00:04:06,260
doing something quite similar. 
They're rolling out, you know, 

66
00:04:06,380 --> 00:04:09,460
enterprise licences, putting AI 
assistance on the sidebar of 

67
00:04:09,460 --> 00:04:12,060
every application. 
They've got internal prompt 

68
00:04:12,060 --> 00:04:15,220
engineering workshops. 
That's not AI as an operating 

69
00:04:15,220 --> 00:04:18,660
system that is making existing 
processes a little bit faster 

70
00:04:18,660 --> 00:04:19,899
without actually changing 
anything. 

71
00:04:20,399 --> 00:04:24,120
And Dorsey frame this as the 
turbocharger on a horse cart. 

72
00:04:24,400 --> 00:04:25,680
The horse is still doing the 
work. 

73
00:04:25,680 --> 00:04:28,280
You just made the horse a little
bit less tired. 

74
00:04:28,880 --> 00:04:31,120
And the distinction drawn in 
this principle here is that your

75
00:04:31,120 --> 00:04:34,600
company should not be able to 
function without AI in the same 

76
00:04:34,600 --> 00:04:37,240
way it can't function without 
Microsoft Teams or e-mail or 

77
00:04:37,240 --> 00:04:40,040
Slack. 
Not AI makes us work faster. 

78
00:04:40,840 --> 00:04:42,760
AI is actually how everything 
gets done. 

79
00:04:43,520 --> 00:04:46,240
And the test is quite simple. 
I think could your most 

80
00:04:46,240 --> 00:04:50,240
important daily team workflows 
run for a week if you switched 

81
00:04:50,240 --> 00:04:53,280
off all AI tools? 
If the answer is yes and the 

82
00:04:53,280 --> 00:04:56,880
work would carry on more or less
normal but slower, then you're 

83
00:04:56,880 --> 00:04:59,160
just using AI as a tool and 
you're not running it as an 

84
00:04:59,160 --> 00:05:01,160
operating system. 
I'll give you a couple of good 

85
00:05:01,160 --> 00:05:03,000
examples. 
We've quite literally built 

86
00:05:03,000 --> 00:05:06,840
internal tools for managing all 
systems that literally building 

87
00:05:06,840 --> 00:05:09,280
them using public APIs from 
Twitter, for example. 

88
00:05:09,960 --> 00:05:14,560
We have every process that is 
built around AI now and we 

89
00:05:14,560 --> 00:05:16,280
haven't got enough people to 
suddenly run all of those 

90
00:05:16,280 --> 00:05:18,760
anymore. 
We we would literally be ceasing

91
00:05:18,760 --> 00:05:22,240
to function as a team. 
For example, every task of 

92
00:05:22,240 --> 00:05:25,560
scheduled task on one system 
feeds into the next system to 

93
00:05:25,560 --> 00:05:28,600
run the next process. 
Every team takes those outputs 

94
00:05:28,600 --> 00:05:29,880
and puts it into the next 
process. 

95
00:05:29,880 --> 00:05:32,960
We we can't function anymore. 
Block shared a really 

96
00:05:32,960 --> 00:05:35,000
interesting example. 
Jack Dorsey's company. 

97
00:05:35,760 --> 00:05:37,200
Essentially they've got 
something called Manager Bot. 

98
00:05:37,200 --> 00:05:39,720
It sits inside Square and 
monitors each merchants 

99
00:05:40,160 --> 00:05:43,640
inventory levels, sales 
velocity, their staffing, and 

100
00:05:43,640 --> 00:05:46,000
other signals are upcoming 
events and weather patterns. 

101
00:05:46,560 --> 00:05:48,680
And it doesn't wait for a 
merchant to notice a problem or 

102
00:05:48,680 --> 00:05:51,560
ask for help. 
It literally surfaces an action 

103
00:05:51,560 --> 00:05:53,600
before the merchant even 
realizes there's a problem. 

104
00:05:54,320 --> 00:05:56,760
You can't just switch that off 
on a Friday and start again on a

105
00:05:56,760 --> 00:05:58,720
on a Monday. 
It's not a feature, it's now a 

106
00:05:58,720 --> 00:06:03,160
layer of their entire system. 
Principle number 2 is closed 

107
00:06:03,160 --> 00:06:06,680
loops, not open loops. 
So an open loop is basically 

108
00:06:06,680 --> 00:06:08,600
where a decision goes out, 
something happens, and the 

109
00:06:08,600 --> 00:06:10,000
result never really comes back 
in. 

110
00:06:10,520 --> 00:06:14,720
In AI, that is a pillar of 
success and effectiveness. 

111
00:06:15,440 --> 00:06:18,080
You know, most companies are 
running almost entirely on open 

112
00:06:18,080 --> 00:06:20,840
loops and they have no idea 
whether things are working or 

113
00:06:20,840 --> 00:06:24,280
not, or if the agents are making
the right decisions or they're 

114
00:06:24,280 --> 00:06:28,600
not able to scale these systems 
because that closed loop doesn't

115
00:06:28,600 --> 00:06:30,600
exist. 
And I've talked about This is 

116
00:06:30,600 --> 00:06:33,040
why, you know, MD file 
structures that we talked about 

117
00:06:33,040 --> 00:06:36,920
in the last episode exist where 
an agent logs its decision and, 

118
00:06:36,920 --> 00:06:38,600
and stuff like that. 
But that's just not scalable, as

119
00:06:38,600 --> 00:06:41,200
I've mentioned before, because 
these decisions need to be 

120
00:06:41,200 --> 00:06:43,480
logged at scale and each agent 
needs to pick up from that 

121
00:06:43,480 --> 00:06:46,240
decision. 
And most companies are basically

122
00:06:46,240 --> 00:06:49,560
running entirely on open leaps 
and have zero idea about what's 

123
00:06:49,560 --> 00:06:52,320
working as a result. 
And that's how companies have 

124
00:06:52,320 --> 00:06:54,360
always worked. 
They've relied on humans to be 

125
00:06:54,360 --> 00:06:57,320
the closed leap to retain that 
bit of information about whether

126
00:06:57,320 --> 00:07:00,240
something was successful, put in
a spreadsheet or a dashboard 

127
00:07:00,240 --> 00:07:03,440
maybe, that a human may check 
every few months or weeks. 

128
00:07:04,280 --> 00:07:06,560
A good example is you run a 
marketing campaign, you see the 

129
00:07:06,560 --> 00:07:10,160
total conversion numbers. 
You do not know really which 

130
00:07:10,160 --> 00:07:12,040
specific touch points drove that
conversion. 

131
00:07:12,360 --> 00:07:13,640
Some tools give you that 
insight. 

132
00:07:14,200 --> 00:07:17,200
You don't have an idea of maybe 
the sequence of interactions, 

133
00:07:17,440 --> 00:07:20,520
the variance of messaging that 
worked, and the information 

134
00:07:20,520 --> 00:07:23,680
always exists but doesn't root 
back into how you design that 

135
00:07:23,680 --> 00:07:25,560
next campaign in a systematic 
way. 

136
00:07:26,360 --> 00:07:28,720
And a lot of the time, that's 
because you take insight, you 

137
00:07:28,720 --> 00:07:31,360
take data and then you draw 
insights from it to shape your 

138
00:07:31,360 --> 00:07:34,680
next things or your next ideas. 
But for agents, that doesn't 

139
00:07:34,680 --> 00:07:37,760
work because an agent is not a 
human and you can't trust it in 

140
00:07:37,760 --> 00:07:39,360
the same way. 
It's fallible. 

141
00:07:39,800 --> 00:07:42,360
Whereas a closed loop system is 
a process where the outcome of 

142
00:07:42,400 --> 00:07:45,920
every single decision roots back
into a structured database 

143
00:07:45,920 --> 00:07:49,360
really, so that an agent can 
pick up that decision and make 

144
00:07:49,360 --> 00:07:50,960
subsequent decisions 
automatically. 

145
00:07:51,320 --> 00:07:52,920
Again, you're scaling 
intelligence. 

146
00:07:53,480 --> 00:07:56,040
A good example of this is what 
we're building with mindset, 

147
00:07:56,040 --> 00:08:00,160
which is called Memex. 
Essentially, you plan work in a 

148
00:08:00,200 --> 00:08:02,720
document. 
That document has all the why 

149
00:08:02,720 --> 00:08:05,560
and the how this is going to be 
delivered, which creates a set 

150
00:08:05,560 --> 00:08:08,000
of decisions to make. 
Humans and agents make those 

151
00:08:08,000 --> 00:08:10,520
decisions, and then suddenly all
of that is split into tasks. 

152
00:08:11,160 --> 00:08:13,360
You then assign that to all the 
agents and people and they go 

153
00:08:13,360 --> 00:08:17,520
off and deliver that work. 
Every single piece of task and 

154
00:08:17,520 --> 00:08:20,240
item within that task that's 
delivered by a human or an agent

155
00:08:20,240 --> 00:08:23,320
is automatically updated into 
that task. 

156
00:08:24,640 --> 00:08:26,920
So it's literally constantly 
scaling all the decisions that 

157
00:08:26,920 --> 00:08:30,080
were made and as a result, every
agent or person that picks up 

158
00:08:30,080 --> 00:08:33,480
from that next, say, task or 
project or whatever it might be 

159
00:08:33,480 --> 00:08:36,360
in the future has access to all 
decisions at scale. 

160
00:08:36,880 --> 00:08:38,320
You don't need to provide a 
context. 

161
00:08:38,679 --> 00:08:41,559
You don't need to worry if this 
is going to do the wrong thing 

162
00:08:41,559 --> 00:08:43,320
because it has access to those 
decisions. 

163
00:08:43,840 --> 00:08:48,000
As a result, you can suddenly 
run at scale AI systems that can

164
00:08:48,000 --> 00:08:50,880
take huge decisions and actions 
because it has such amazing 

165
00:08:50,880 --> 00:08:53,000
context. 
That's a closed loop system. 

166
00:08:53,640 --> 00:08:57,120
So as an action for you this 
week, pick a process and then 

167
00:08:57,120 --> 00:08:59,560
close it. 
What's the most repetitive task?

168
00:09:00,000 --> 00:09:02,840
What's the AI process or human 
process and where is that 

169
00:09:02,840 --> 00:09:05,600
getting locked? 
You could take an MD structure 

170
00:09:05,600 --> 00:09:07,960
format like we talked about last
week that can work. 

171
00:09:08,280 --> 00:09:11,920
That has some problems at scale.
You have what we call decision 

172
00:09:11,920 --> 00:09:14,880
drift, where not every decision 
gets logged in the proper way. 

173
00:09:14,880 --> 00:09:17,640
So the agent of the human misses
key context and decisions. 

174
00:09:18,200 --> 00:09:20,680
But it's a good start. 
If you have called Co work, 

175
00:09:20,680 --> 00:09:23,360
describe the process, ask it to 
design what a closed loop system

176
00:09:23,360 --> 00:09:26,160
would look like, where it gets 
measured, where the output goes,

177
00:09:26,160 --> 00:09:27,680
where the next agent picks up 
from. 

178
00:09:28,040 --> 00:09:31,160
So when you think about this, 
think about asking the agent to 

179
00:09:31,160 --> 00:09:35,160
create closed loop systems, and 
then every agent or system or 

180
00:09:35,160 --> 00:09:37,640
scheduled task that you create 
should then read from that 

181
00:09:37,640 --> 00:09:40,040
place. 
So principle #3 is making your 

182
00:09:40,040 --> 00:09:43,320
company fully queriable. 
So think about what happens in 

183
00:09:43,320 --> 00:09:46,480
most companies when, say, 
someone new joins at a senior 

184
00:09:46,480 --> 00:09:49,200
level. 
They spend months in meetings 

185
00:09:49,400 --> 00:09:51,120
talking to people, getting 
chucked into things. 

186
00:09:51,120 --> 00:09:52,280
So they get to grips with 
everything. 

187
00:09:52,280 --> 00:09:54,600
They have to understand what 
decisions have been made. 

188
00:09:54,600 --> 00:09:56,760
They have to figure out 
institutional knowledge. 

189
00:09:57,320 --> 00:10:00,400
It's not indexed anywhere. 
It cannot be queried easily. 

190
00:10:00,880 --> 00:10:02,240
They just have to ask people all
the time. 

191
00:10:03,000 --> 00:10:06,200
This is what Hugh calls an 
Inquiry Able Company, and it 

192
00:10:06,200 --> 00:10:09,320
describes most organizations 
that existed before 2024. 

193
00:10:09,880 --> 00:10:11,960
So what this means in practice, 
for example, is instead of 

194
00:10:11,960 --> 00:10:14,560
emailing A colleague to ask why 
a decision was made six months 

195
00:10:14,560 --> 00:10:17,880
ago or last week, you just open 
a search interface and ask it 

196
00:10:18,280 --> 00:10:20,640
and you get all the information,
the context, the emails, 

197
00:10:20,640 --> 00:10:24,040
everything instantly. 
A good example of this is we 

198
00:10:24,040 --> 00:10:28,840
have a system for managing every
project and plan and code that 

199
00:10:28,840 --> 00:10:32,240
was written at the moment. 
It was in an It was initially in

200
00:10:32,240 --> 00:10:35,200
an MD file structure that was 
very advanced, but we've now 

201
00:10:35,200 --> 00:10:37,880
moved out to Memex. 
Every decision, every 

202
00:10:37,880 --> 00:10:39,560
information, it's all stored in 
there. 

203
00:10:39,800 --> 00:10:43,040
So it's instantly queryable from
a product delivery perspective 

204
00:10:43,800 --> 00:10:46,880
and team members in the product 
process can query and out. 

205
00:10:47,400 --> 00:10:50,440
When a developer goes, OK, I 
want to deliver any project, 

206
00:10:51,000 --> 00:10:53,840
they don't have to ask anyone 
about architecture or decisions 

207
00:10:53,840 --> 00:10:56,360
or projects from a week, six 
weeks, a month ago. 

208
00:10:56,920 --> 00:10:59,080
They literally just go to that 
system and start making it from 

209
00:10:59,080 --> 00:11:00,760
there. 
Many companies have already, you

210
00:11:00,760 --> 00:11:02,880
know, started doing this. 
They've got AI note taking, 

211
00:11:03,600 --> 00:11:06,880
Notion for say documentation, 
Slack instead of e-mail. 

212
00:11:06,880 --> 00:11:10,800
But that's all very basic stuff.
And This is why I think one of 

213
00:11:10,800 --> 00:11:12,920
the most important spaces, and 
I'm not just saying this because

214
00:11:12,920 --> 00:11:16,520
this is what Memex does a lot of
the time as a company that's 

215
00:11:16,880 --> 00:11:19,280
I'll probably describe at the 
forefront of a lot of these new 

216
00:11:19,280 --> 00:11:22,360
changes. 
We spot things that are really, 

217
00:11:22,360 --> 00:11:26,000
really important quite early. 
And in the next 1218 months, the

218
00:11:26,000 --> 00:11:28,240
most important things is going 
to be a decision layer, A 

219
00:11:28,240 --> 00:11:31,840
decision register, not a meeting
summary that has loads of 

220
00:11:32,440 --> 00:11:36,240
unstructured text and 
information, An actual organized

221
00:11:36,240 --> 00:11:39,400
database of every decision that 
was made, connected to each 

222
00:11:39,400 --> 00:11:42,440
project and person that 
delivered it, and importantly, 

223
00:11:42,760 --> 00:11:45,320
every agent. 
Because again, the the new 

224
00:11:45,320 --> 00:11:49,320
paradigm here is that you no 
longer design processes and 

225
00:11:49,320 --> 00:11:51,360
systems for human to human 
collaboration. 

226
00:11:51,800 --> 00:11:54,720
It's all about agent to agent to
human collaboration. 

227
00:11:55,200 --> 00:11:58,160
Humans are a coordinator now, 
not just doers. 

228
00:11:58,720 --> 00:12:01,120
So as an action, think about 
where decisions are getting 

229
00:12:01,120 --> 00:12:03,440
logged. 
Think about how you can make 

230
00:12:03,440 --> 00:12:06,640
everything queryable so you can 
always say to somebody who asks 

231
00:12:06,640 --> 00:12:09,760
a question, have you asked X 
system? 

232
00:12:11,040 --> 00:12:15,120
So then the culture internally 
should be 1 of initiative. 

233
00:12:15,640 --> 00:12:17,880
Because once you remove those 
barriers, in my opinion, you 

234
00:12:17,880 --> 00:12:20,720
also remove excuses. 
So you encourage and enforce 

235
00:12:20,720 --> 00:12:23,840
people to take initiative and go
and make decisions. 

236
00:12:24,400 --> 00:12:28,600
So principle #4 is the idea of 
1000 X team member, engineer, 

237
00:12:28,600 --> 00:12:31,680
employee and what that means for
how you know, hire. 

238
00:12:32,480 --> 00:12:34,680
So you know, for software 
companies, we'll just take the 

239
00:12:34,680 --> 00:12:38,040
engineer as the first example. 
Software companies for 30 years 

240
00:12:38,040 --> 00:12:43,040
operated in a band of roughly 
200,000 to $400,000 of revenue 

241
00:12:43,120 --> 00:12:45,680
per employee. 
And that range is held 

242
00:12:45,680 --> 00:12:48,120
throughout most of the package 
software era. 

243
00:12:48,120 --> 00:12:51,480
So when you buy a perpetual 
license for 12 months all the 

244
00:12:51,480 --> 00:12:54,360
way through SAS through to 
mobile, it was a bit of a 

245
00:12:54,440 --> 00:12:58,480
structural ceiling on what a 
person plus a computer could 

246
00:12:58,480 --> 00:13:02,560
actually just produce. 
But that ceiling is now well and

247
00:13:02,560 --> 00:13:05,400
truly gone. 
You know, Cursor, the AI coding 

248
00:13:05,400 --> 00:13:11,240
tool, hit $2 billion of revenue 
and generates $6.1 million per 

249
00:13:11,240 --> 00:13:14,000
employee. 
So that historical benchmark is,

250
00:13:14,000 --> 00:13:17,760
is is gone. 
It's 20 times the traditional 

251
00:13:18,200 --> 00:13:20,480
revenue per employee ceiling 
that we used to have. 

252
00:13:21,200 --> 00:13:23,560
And these aren't productivity 
improvements, they are a 

253
00:13:23,560 --> 00:13:25,840
different level of output per 
person and function. 

254
00:13:26,640 --> 00:13:30,640
And that thousand X team member 
employee engineer concept 

255
00:13:30,640 --> 00:13:33,400
therefore actually means one 
person doing what used to take 

256
00:13:33,400 --> 00:13:35,840
an entire team. 
I'll give you a good example. 

257
00:13:36,920 --> 00:13:40,720
We for this product launch built
an internal Twitter tool, 

258
00:13:41,000 --> 00:13:44,240
created a research report that 
scanned over 20,000 public 

259
00:13:44,480 --> 00:13:46,760
GitHub repos. 
So loads of different public 

260
00:13:46,760 --> 00:13:50,800
repos for code to analyse the 
drift between the MD file 

261
00:13:50,800 --> 00:13:55,560
structure and what's actually 
being written in code to bring 

262
00:13:55,560 --> 00:13:58,000
to life what we see is this big 
problem of decisions getting 

263
00:13:58,000 --> 00:14:01,360
made by agents but not updating 
your rule set for agents. 

264
00:14:01,680 --> 00:14:04,400
Using the tools, we then create 
an entire strategy, a content 

265
00:14:04,400 --> 00:14:08,080
strategy based on real Twitter 
conversations that were 

266
00:14:08,080 --> 00:14:10,680
happening in vocabulary and 
language to make it super 

267
00:14:10,680 --> 00:14:13,880
applicable to people and really 
understand the problem set. 

268
00:14:14,400 --> 00:14:18,920
We then created and set up MCPS 
into Google, into Google 

269
00:14:18,920 --> 00:14:22,360
Analytics and Google Ads and to 
run experiments on our 

270
00:14:22,360 --> 00:14:24,560
proposition. 
So what is the best way of 

271
00:14:24,560 --> 00:14:27,280
describing this product and 
problem set that resonates to 

272
00:14:27,280 --> 00:14:29,720
people? 
And we did it all from just 

273
00:14:29,720 --> 00:14:31,920
talking to AI and setting up MCP
servers. 

274
00:14:32,320 --> 00:14:35,120
And then we pull in automatic 
reporting from different places 

275
00:14:35,280 --> 00:14:38,800
into reports and models, which 
create new insights, which 

276
00:14:38,800 --> 00:14:42,600
creates a new document as an MD 
doc, which we send straight to 

277
00:14:43,160 --> 00:14:46,560
our AI coding. 
So core code in in our MIDE 

278
00:14:46,800 --> 00:14:49,720
which spins up a new landing 
page to test new hypotheses. 

279
00:14:50,360 --> 00:14:54,920
All of this, all of this was 
done by literally probably the 

280
00:14:54,920 --> 00:14:57,560
equivalent of 1 1/2 persons 
working part time. 

281
00:14:58,040 --> 00:15:01,040
That would have been probably 
months of work. 

282
00:15:01,240 --> 00:15:04,640
And we literally were doing that
in literally over a bank holiday

283
00:15:04,640 --> 00:15:07,480
weekend. 
So the level of pace change and 

284
00:15:07,480 --> 00:15:09,200
things you can achieve is 
insane. 

285
00:15:09,960 --> 00:15:12,600
I think the hiring implication 
here is quite uncomfortable 

286
00:15:13,400 --> 00:15:16,560
because most hiring today still 
uses some version of challenges 

287
00:15:16,560 --> 00:15:19,200
or basic questions. 
If you take the engineers, 

288
00:15:19,640 --> 00:15:22,080
you're given a challenge. 
So write this function, debug 

289
00:15:22,080 --> 00:15:25,480
something, build this feature in
a few hours, and then you test 

290
00:15:25,480 --> 00:15:27,720
it. 
But in a world where agents do 

291
00:15:27,720 --> 00:15:30,800
all of this, do the execution, 
you're screening for the wrong 

292
00:15:30,800 --> 00:15:33,160
thing. 
The skill that matters now is AI

293
00:15:33,160 --> 00:15:37,480
management is being able to 
precisely describe what should 

294
00:15:37,480 --> 00:15:41,120
be done for you. 
Can this person define what good

295
00:15:41,120 --> 00:15:43,040
enough looks like? 
Can they have the taste and 

296
00:15:43,040 --> 00:15:46,520
judgement to make decisions? 
Do they have the initiative to 

297
00:15:46,520 --> 00:15:49,240
get an idea and execute and move
forward? 

298
00:15:49,600 --> 00:15:52,880
Can they take the output and 
verify whether that was done 

299
00:15:52,880 --> 00:15:56,320
properly and have the judgement 
to decide what to do next? 

300
00:15:57,000 --> 00:15:59,720
These skills are hard to test 
for, but the interview doesn't 

301
00:15:59,720 --> 00:16:02,200
have to be complicated. 
Instead of these challenges or 

302
00:16:02,200 --> 00:16:05,560
basic questions, give people a, 
let's say, a product brief and 

303
00:16:05,560 --> 00:16:07,120
ask them to write a 
specification. 

304
00:16:07,560 --> 00:16:10,520
Not to build anything, but to 
define what it should do, what 

305
00:16:10,520 --> 00:16:13,000
edge cases it needs to handle, 
the process that you're going to

306
00:16:13,000 --> 00:16:14,920
go through and then see what 
happens. 

307
00:16:15,440 --> 00:16:18,800
Just because someone's a really 
good person in their job doesn't

308
00:16:18,800 --> 00:16:21,400
mean they're going to be a 
fantastic candidate in this new 

309
00:16:21,480 --> 00:16:24,720
AI first world. 
And it's why education needs to 

310
00:16:24,720 --> 00:16:28,200
rapidly change. 
The way we teach people to think

311
00:16:28,200 --> 00:16:29,880
about the world needs to change 
fast. 

312
00:16:30,520 --> 00:16:34,040
So yeah, next time you go for 
hiring, think about how you can 

313
00:16:34,320 --> 00:16:38,480
attract the right person and 
actually learn whether the right

314
00:16:38,480 --> 00:16:42,360
candidate through a different 
type of process principle #5 is 

315
00:16:42,360 --> 00:16:45,800
token Max, not headcount Max. 
Now, again, I covered this in a 

316
00:16:45,800 --> 00:16:47,680
recent episode. 
I, I have mixed feelings with 

317
00:16:47,680 --> 00:16:49,080
this. 
I, I mentioned it a lot, so I'm 

318
00:16:49,080 --> 00:16:51,800
not going to go into it, but it 
was a really popular episode. 

319
00:16:51,800 --> 00:16:54,680
I recommend you go listen to it.
But the principle that they were

320
00:16:54,680 --> 00:16:57,080
talking about is that most 
companies are still treating 

321
00:16:57,080 --> 00:16:59,760
inference. 
So spending on tokens, which is 

322
00:16:59,760 --> 00:17:02,040
what happens when you produce 
words through a language model, 

323
00:17:03,000 --> 00:17:06,119
is essentially constrained. 
Like finance saying we don't 

324
00:17:06,119 --> 00:17:09,400
want to spend too much here. 
We have to approve tickets to 

325
00:17:09,920 --> 00:17:13,280
increase spend per person, which
basically means that if you want

326
00:17:13,280 --> 00:17:16,240
to run many agents across many 
larger data sets and bigger 

327
00:17:16,240 --> 00:17:18,880
tasks, you're going to burn 
through lots of tokens and spend

328
00:17:18,880 --> 00:17:20,760
a lot. 
And finance is constraining 

329
00:17:20,760 --> 00:17:22,920
that. 
And the principle here from this

330
00:17:22,920 --> 00:17:25,599
book, which as I said, I land in
a bit of a mixed feeling here, 

331
00:17:25,599 --> 00:17:29,160
is just to burn as many tokens 
as humanly possible, not get 

332
00:17:29,160 --> 00:17:32,000
more headcount. 
And if according to Y 

333
00:17:32,120 --> 00:17:34,640
Combinator, if that doesn't make
the CFOA little bit 

334
00:17:34,640 --> 00:17:36,960
uncomfortable, they're probably 
not pushing it far enough. 

335
00:17:37,560 --> 00:17:40,800
Now, I do think this encourages 
the wrong behaviours and you 

336
00:17:40,800 --> 00:17:43,760
need to be teaching people how 
to be efficient with token spend

337
00:17:44,080 --> 00:17:47,400
and context usage when using AI.
But if you're trying to get the 

338
00:17:47,400 --> 00:17:49,840
team to really start adopting 
this, it can be a good place to 

339
00:17:49,840 --> 00:17:53,680
start just to get people widely 
using it and then teach them how

340
00:17:53,680 --> 00:17:56,800
to be more efficient. 
So as an action, maybe whole the

341
00:17:56,800 --> 00:17:59,640
last three months of AI tool 
spend across all the different 

342
00:17:59,640 --> 00:18:03,400
subscriptions, use core Co work 
to paste it alongside your 

343
00:18:03,400 --> 00:18:06,440
headcount and ask it to 
calculate current spend per 

344
00:18:06,440 --> 00:18:09,360
person per month on AI and then 
judge whether that's high 

345
00:18:09,400 --> 00:18:11,480
enough. 
Take that into the next finance 

346
00:18:11,480 --> 00:18:15,280
conversation and reframe it as 
maybe not increasing overhead of

347
00:18:15,280 --> 00:18:17,400
new people, but actually 
increasing spend. 

348
00:18:17,800 --> 00:18:20,760
The goal really here is simple. 
It's not about cutting headcount

349
00:18:20,760 --> 00:18:23,840
here in my view, it's about 
getting people to stop asking 

350
00:18:23,840 --> 00:18:26,440
permission to use AI and just to
start using it. 

351
00:18:27,040 --> 00:18:29,880
Principle number six is the idea
of software factories. 

352
00:18:30,320 --> 00:18:32,920
So most teams today still work 
in the same way as they did in 

353
00:18:33,200 --> 00:18:36,640
2015. 
A ticket arrives, a person picks

354
00:18:36,640 --> 00:18:37,840
it up. 
So in your project management 

355
00:18:37,840 --> 00:18:41,240
tool or in Jira or whatever it 
might be, an engineer or a team 

356
00:18:41,240 --> 00:18:43,440
member picks it up. 
The write code or they do the 

357
00:18:43,440 --> 00:18:46,560
task, they put it in review 
state to wait for somebody to 

358
00:18:46,560 --> 00:18:50,240
look at it, and then it ships. 
AI is, you know, in a sidebar 

359
00:18:50,240 --> 00:18:54,160
helping with certain parts of 
that, but the process is broadly

360
00:18:54,160 --> 00:18:56,960
unchanged. 
Now, what Hugh described is 

361
00:18:57,120 --> 00:18:59,880
something very different. 
A-Team that does not write code 

362
00:18:59,880 --> 00:19:02,960
at all, that does not perform 
the tasks themselves at all. 

363
00:19:03,400 --> 00:19:05,920
Mindset's been like this for 
quite a long time. 

364
00:19:05,920 --> 00:19:08,280
Almost no code is really written
by humans anymore. 

365
00:19:08,840 --> 00:19:12,080
We really care about decision 
making though, and planning and 

366
00:19:12,080 --> 00:19:14,320
thinking about what the AI 
produces, because otherwise 

367
00:19:14,560 --> 00:19:18,080
you're just vibe coding crap. 
So it's not about not caring 

368
00:19:18,080 --> 00:19:22,680
about the input or the output, 
but it's trying to no longer 

369
00:19:22,680 --> 00:19:25,360
constrain people to having to do
the work themselves. 

370
00:19:25,840 --> 00:19:29,560
Now, Dan Shapiro built a 
taxonomy basically of AI 

371
00:19:29,560 --> 00:19:32,520
assistant coding, and the term 
comes from manufacturing. 

372
00:19:32,520 --> 00:19:35,160
So a facility where robots work 
in complete darkness because 

373
00:19:35,160 --> 00:19:37,120
they don't need to see. 
And a good example here is the 

374
00:19:37,120 --> 00:19:40,600
code compiles in the dark and 
humans just show up to maintain 

375
00:19:40,600 --> 00:19:43,720
the specification, verify the 
output and decide what comes 

376
00:19:43,720 --> 00:19:45,920
next. 
So if people think their job is 

377
00:19:45,920 --> 00:19:50,040
now writing the thing, coding 
the thing, this new thinking 

378
00:19:50,040 --> 00:19:53,880
tells them that their job 
definition is about telling 

379
00:19:53,880 --> 00:19:56,760
something what to do, making 
sure it makes the right 

380
00:19:56,760 --> 00:19:58,680
decisions, and verifying the 
output. 

381
00:19:59,200 --> 00:20:03,040
So precise specification is now 
a skill that's critical, and as 

382
00:20:03,040 --> 00:20:05,320
a result, therefore you need 
specifications that are 

383
00:20:05,320 --> 00:20:07,240
versioned and maintained like 
Coach used to be. 

384
00:20:07,680 --> 00:20:10,080
If the spec changes, you should 
be able to see the differential.

385
00:20:10,440 --> 00:20:12,640
And again, this is exactly what 
Memex is solving for. 

386
00:20:13,240 --> 00:20:15,640
If someone amends a test 
scenario, there should be a 

387
00:20:15,640 --> 00:20:18,160
record of why. 
If the agent made a decision 

388
00:20:18,160 --> 00:20:20,400
that impacts the plan, it should
have recorded why. 

389
00:20:20,840 --> 00:20:23,440
Principle 7. 
This is the intelligence layer 

390
00:20:23,440 --> 00:20:25,240
replaces the organizational 
chart. 

391
00:20:25,960 --> 00:20:28,120
So most people, when they hear 
middle management is going to 

392
00:20:28,120 --> 00:20:30,200
disappear, they think it's about
cutting costs. 

393
00:20:30,200 --> 00:20:34,160
Well, it it just isn't. 
This principle is actually about

394
00:20:34,400 --> 00:20:37,360
the specific job of a middle 
manager, which if you watch 

395
00:20:37,360 --> 00:20:41,720
carefully, is mostly rooting 
what a senior leader or leader 

396
00:20:41,720 --> 00:20:45,960
of a company wants to the right 
person, or aggregating statuses 

397
00:20:45,960 --> 00:20:49,920
from teams relaying contacts to,
you know, people above them. 

398
00:20:50,520 --> 00:20:53,640
Passing the decisions back down 
in the right way, making sure 

399
00:20:53,640 --> 00:20:55,480
the left hand knows what the 
right hand is doing. 

400
00:20:55,480 --> 00:20:59,080
Basically them having the right 
judgement of what to do or what 

401
00:20:59,080 --> 00:21:02,440
to build when things go wrong. 
Well, that is now a much smaller

402
00:21:02,440 --> 00:21:04,840
fraction of the role than the 
title actually implies. 

403
00:21:05,200 --> 00:21:08,600
Now what the intelligence layer 
replaces is the routing, not the

404
00:21:08,600 --> 00:21:11,680
judgement. 
When every person in the company

405
00:21:11,680 --> 00:21:14,320
can query the same set of 
information, see the same 

406
00:21:14,320 --> 00:21:18,040
decisions, understand the same 
context without a human having 

407
00:21:18,040 --> 00:21:22,200
to pass that information along 
well, therefore the layers that 

408
00:21:22,200 --> 00:21:26,920
exist only to pass information 
and just not important and 

409
00:21:26,920 --> 00:21:31,960
therefore the organizational 
chart now is no longer the same.

410
00:21:32,320 --> 00:21:34,920
It was previously a workaround 
for the fact that information 

411
00:21:34,920 --> 00:21:38,760
couldn't really travel at scale.
A good example here is now we 

412
00:21:38,760 --> 00:21:41,760
have MCP service for all of our 
primary systems, for sales 

413
00:21:41,760 --> 00:21:45,680
engineering, for product, for 
customer service, and every 

414
00:21:45,680 --> 00:21:48,240
single one of those accessed by 
every leader in the company. 

415
00:21:48,760 --> 00:21:51,280
You can get a scheduled daily 
report of everything that 

416
00:21:51,280 --> 00:21:54,080
happened automatically. 
So therefore, why would I go and

417
00:21:54,080 --> 00:21:56,560
ask a middle management about 
what happened, about progress, 

418
00:21:56,560 --> 00:21:58,480
about performance? 
I don't need that. 

419
00:21:59,320 --> 00:22:02,240
As a result, the people who are 
now middle management are about 

420
00:22:02,240 --> 00:22:05,760
deciding what the most important
actions are to take, aligning 

421
00:22:05,760 --> 00:22:09,800
the teams and their agents to 
take them, and making sure that 

422
00:22:09,800 --> 00:22:13,120
thing happens on time. 
So as a result, in many ways the

423
00:22:13,120 --> 00:22:16,440
org chart flattens and I think 
leadership becomes about 

424
00:22:16,440 --> 00:22:19,000
supporting people and getting 
them to think about the right 

425
00:22:19,000 --> 00:22:22,160
ideas and the right decisions to
make, not just passing 

426
00:22:22,160 --> 00:22:25,760
information along the process. 7
principles are extremely 

427
00:22:25,760 --> 00:22:29,840
concrete, really interesting by 
general partners who work with 

428
00:22:29,840 --> 00:22:32,880
all of the most exciting AI 
start-ups in the world. 

429
00:22:33,400 --> 00:22:36,600
And two months earlier, Jack 
Dorsey had run most of these at 

430
00:22:36,600 --> 00:22:39,800
his company block, and actually 
cut 40% of his workforce in the 

431
00:22:39,800 --> 00:22:57,600
process. 
So there are the seven 

432
00:22:57,600 --> 00:23:00,920
principles and I think it's it's
a good time to conclude and I'm 

433
00:23:00,920 --> 00:23:03,120
going to conclude with a few 
different actions I think you 

434
00:23:03,120 --> 00:23:05,000
should do. 
You don't need to implement them

435
00:23:05,000 --> 00:23:09,200
all at once, but I think you can
start by a running three 

436
00:23:09,200 --> 00:23:13,280
questions to test the team. 
What is the minimum number of 

437
00:23:13,280 --> 00:23:15,880
people to keep this service 
running at full capacity? 

438
00:23:16,160 --> 00:23:20,200
What is the minimum to stay 
compliant and secure and what is

439
00:23:20,200 --> 00:23:22,280
the minimum to grow and meet 
your commitments? 

440
00:23:22,720 --> 00:23:26,360
The first two give you the floor
and the third gives you the 

441
00:23:26,360 --> 00:23:28,800
ceiling. 
For this stage, most leaders 

442
00:23:28,800 --> 00:23:31,560
will know how many people work 
in their company or teams, but 

443
00:23:31,560 --> 00:23:35,320
very few calculate the floor. 
This isn't supposed to be a 

444
00:23:35,320 --> 00:23:38,360
redundancy exercise. 
It's supposed to be an audit of 

445
00:23:38,360 --> 00:23:41,320
the structure moving forward and
the number that comes back. 

446
00:23:41,320 --> 00:23:45,160
I think will tell you more about
how the organization runs today 

447
00:23:45,960 --> 00:23:48,000
more than anything else you can 
do in a single week. 

448
00:23:48,560 --> 00:23:52,480
Second, set a visible AI budget 
and just remove the approval 

449
00:23:52,480 --> 00:23:55,080
friction. 
Pull what people spend at the 

450
00:23:55,080 --> 00:23:56,920
moment per person per month on 
AI tools. 

451
00:23:57,320 --> 00:23:59,760
Open clock Co work. 
Paste it in alongside the 

452
00:23:59,760 --> 00:24:01,840
headcount and ask it to 
benchmark against what the 

453
00:24:01,840 --> 00:24:05,160
leading companies are spending. 
Set a team budget higher than 

454
00:24:05,160 --> 00:24:07,720
your current spend, but then 
work with them to make sure 

455
00:24:07,720 --> 00:24:11,120
they're efficiently spending. 
Make sure that people just don't

456
00:24:11,120 --> 00:24:16,360
ask permission #3 close a loop 
on one of the most important 

457
00:24:16,360 --> 00:24:19,040
repetitive tasks. 
Pick a task that someone on your

458
00:24:19,040 --> 00:24:23,040
team does all the time in the 
same way and use cloud Co work 

459
00:24:23,040 --> 00:24:26,080
to design what a closed loop 
system looks like for that task.

460
00:24:26,440 --> 00:24:28,960
What gets measured, where the 
output goes, what changes next 

461
00:24:28,960 --> 00:24:32,520
time, what the next agent 
process will read from. 

462
00:24:32,520 --> 00:24:34,640
So is it going to be connected 
to that same decision layer? 

463
00:24:35,320 --> 00:24:38,560
And then measure quality of 
outcomes from day one, not just,

464
00:24:38,800 --> 00:24:42,440
you know, how fast it runs and 
#4 run a software factory 

465
00:24:42,440 --> 00:24:44,960
pattern on at least one internal
process. 

466
00:24:45,720 --> 00:24:48,240
So something that touches no 
customers and no live system. 

467
00:24:48,800 --> 00:24:51,720
Just get a person to write the 
spec, One person writes the 

468
00:24:51,720 --> 00:24:54,820
criteria for success. 
And get core code or core Co 

469
00:24:54,820 --> 00:24:57,400
work or whatever it might be to 
just do the work. 

470
00:24:58,040 --> 00:25:00,720
Review the specification when 
the output is wrong, not just 

471
00:25:00,720 --> 00:25:03,760
the output itself. 
And then get it to the log its 

472
00:25:03,760 --> 00:25:06,200
decisions in a structured way, 
whether it's an MD file 

473
00:25:06,200 --> 00:25:08,520
structure, whether it's pushing 
it into a spreadsheet. 

474
00:25:09,440 --> 00:25:12,240
So yeah, there are the four 
things I think you need to start

475
00:25:12,480 --> 00:25:16,520
now. 
This is a 6-7 week project, you 

476
00:25:16,520 --> 00:25:19,360
can really transform your team 
and organization as a result. 

477
00:25:19,360 --> 00:25:24,040
And really I think many of these
principles, although quite scary

478
00:25:24,040 --> 00:25:27,040
for the future of employment 
quite frankly, and the fact that

479
00:25:27,040 --> 00:25:30,080
we really need to push for 
change for education and 

480
00:25:30,120 --> 00:25:33,280
opportunities, for 
apprenticeships and pre 

481
00:25:33,280 --> 00:25:35,720
employment training. 
But I think this is how 

482
00:25:35,720 --> 00:25:38,080
organisations are going to run 
in the future. 

483
00:25:38,560 --> 00:25:40,680
Anyway, I hope you find that 
interesting. 

484
00:25:40,680 --> 00:25:43,200
Thank you for listening and I'll
see you next week.