1
00:00:00,300 --> 00:00:05,300
This is epicenter episode, 471 
with guests, Harry grief, and 

2
00:00:05,300 --> 00:00:22,700
been Fielding from Jensen. 
Welcome to epicenter. 

3
00:00:22,800 --> 00:00:25,600
The show would shocks about the 
Technologies projects and people

4
00:00:25,600 --> 00:00:28,100
driving decentralisation and the
blockchain revolution. 

5
00:00:28,500 --> 00:00:31,400
I'm Frederick a handstand today.
I'm speaking with Harry grief 

6
00:00:31,400 --> 00:00:33,700
and been Fielding the founders 
of Jensen. 

7
00:00:34,200 --> 00:00:40,900
Jensen is an AI blockchain 
project, that is looking To 

8
00:00:40,900 --> 00:00:47,500
enable you to buy a i compute in
a decentralized manner and we 

9
00:00:47,500 --> 00:00:51,900
will get to that in just a 
second before that. 

10
00:00:51,900 --> 00:00:55,200
I will tell you about our 
sponsor this week, though, our 

11
00:00:55,200 --> 00:00:59,900
sponsors Tallyho and open source
wallet, redefining the wallet as

12
00:00:59,900 --> 00:01:02,400
a public. 
Good with Tallyho, you can 

13
00:01:02,400 --> 00:01:05,500
safely connect to defy and web 
three with everything you need 

14
00:01:05,600 --> 00:01:08,200
from metal mask. 
Plus a lot more, you can view 

15
00:01:08,200 --> 00:01:10,600
your nft Xin, what it across the
theorem polygon out. 

16
00:01:10,800 --> 00:01:13,800
Amazement a Bertram. 
There's also no need to manually

17
00:01:13,800 --> 00:01:16,500
at these networks. 
They already come plugged in 

18
00:01:17,200 --> 00:01:19,700
tally, ho has the best edge of 
support around built. 

19
00:01:19,700 --> 00:01:22,800
By a community of developers 
that listen to users swap 

20
00:01:22,800 --> 00:01:26,400
between Assets in a wallet at a 
fraction of the price and 

21
00:01:26,400 --> 00:01:29,300
conveniently view. 
All of your account balances 

22
00:01:29,300 --> 00:01:32,800
across multiple networks, with 
our new and improved portfolio 

23
00:01:32,800 --> 00:01:35,100
tab. 
Currently, they're running a 

24
00:01:35,100 --> 00:01:42,500
campaign called gyoji SSE, why 
da GC See a layer to Adventure 

25
00:01:42,600 --> 00:01:45,800
that rewards users for exploring
the arbitrary ecosystem with 

26
00:01:45,800 --> 00:01:48,300
Telly Hope from now until 
December 2nd. 

27
00:01:48,500 --> 00:01:51,400
So hurry up, rich fun. 
So Albert room with one of their

28
00:01:51,400 --> 00:01:55,500
participating Bridges and claim 
your trusty space Dock and ft+ 

29
00:01:55,900 --> 00:02:00,200
be entered into a giveaway for 
rare blueberry Club nft had to 

30
00:02:00,200 --> 00:02:04,000
their blog at block tortelli dot
cash or their Twitter, telecast 

31
00:02:04,000 --> 00:02:06,700
for more info. 
Tell you who isn't just building

32
00:02:06,700 --> 00:02:09,199
a wallet that work steady. 
Ho is building a wallet web 

33
00:02:09,199 --> 00:02:13,000
three can believe in Visit 
Tallyho to Iraq today and to 

34
00:02:13,000 --> 00:02:16,800
download the wallet and join 
Over. 150,000 people in signing 

35
00:02:16,800 --> 00:02:19,400
their Community pledge. 
Okay. 

36
00:02:19,400 --> 00:02:20,700
Fantastic guys. 
Hang on. 

37
00:02:20,700 --> 00:02:25,200
Let me let me order you a bit. 
Put myself in the middle here. 

38
00:02:27,500 --> 00:02:29,100
Fantastic. 
Harry and Ben. 

39
00:02:29,100 --> 00:02:30,500
Thank you so much for joining 
me. 

40
00:02:31,300 --> 00:02:33,400
Hey Federica! 
Thanks so much for having us. 

41
00:02:33,900 --> 00:02:36,300
I think we said this on Twitter,
but we both been longtime 

42
00:02:36,300 --> 00:02:38,000
listeners of the epicenter 
podcast. 

43
00:02:38,000 --> 00:02:41,700
So really pleased to be here. 
And I think that's so good to 

44
00:02:41,700 --> 00:02:43,100
hear. 
How are you? 

45
00:02:43,100 --> 00:02:46,400
And Ben, I'm tell me about 
yourselves. 

46
00:02:46,500 --> 00:02:48,600
What are your backgrounds? 
And what did you do before 

47
00:02:48,600 --> 00:02:50,500
Jensen? 
Sure. 

48
00:02:51,000 --> 00:02:54,600
So yeah, I guess my backgrounds 
in machine learning research. 

49
00:02:54,600 --> 00:02:58,100
Mainly. 
So I did a PhD in deep learning 

50
00:02:58,100 --> 00:03:00,900
focused on neural. 
Architecture search as a problem

51
00:03:00,900 --> 00:03:05,400
which is essentially searching 
the space of deep neural network

52
00:03:05,400 --> 00:03:08,300
structures to find one that kind
of most performant for a 

53
00:03:08,300 --> 00:03:11,700
specific task. 
So did a PhD in matte finish 

54
00:03:11,700 --> 00:03:15,100
that in 2019, and then moved up 
and moved into the startup world

55
00:03:15,600 --> 00:03:17,700
and co-founded a data, privacy 
startup. 

56
00:03:17,800 --> 00:03:21,200
So I've got quite a kind of 
strong interest in individual 

57
00:03:21,200 --> 00:03:23,600
data, privacy kind of data 
sovereignty and things like 

58
00:03:23,600 --> 00:03:26,000
that. 
Did that for a couple of years. 

59
00:03:26,000 --> 00:03:29,300
And then joined an accelerator 
program in London called the 

60
00:03:29,300 --> 00:03:32,000
entrepreneur entrepreneur, 
first, which is where I'm a 

61
00:03:32,000 --> 00:03:33,900
hurry. 
And we kind of went down the 

62
00:03:33,900 --> 00:03:36,100
rabbit hole of what we're 
building with Jensen. 

63
00:03:37,100 --> 00:03:40,300
Yeah, and I'm on my side, my 
backgrounds and applied 

64
00:03:40,300 --> 00:03:43,200
econometrics, so kind of a 
fusion of economics and 

65
00:03:43,200 --> 00:03:45,800
statistics. 
I was sort of introduced in 

66
00:03:45,800 --> 00:03:48,300
machine learning during my 
post-grad doing my Master's 

67
00:03:48,300 --> 00:03:52,600
Degree whilst while studying 
econometrics and fell in love 

68
00:03:52,600 --> 00:03:55,900
with it from there. 
I just thought it was so cool to

69
00:03:55,900 --> 00:03:58,800
be able to essentially quantify 
everything. 

70
00:04:00,000 --> 00:04:03,800
The kind of next step for me was
leading data research team and 

71
00:04:03,800 --> 00:04:06,700
they I start up in London so 
whereas been comes from over. 

72
00:04:06,900 --> 00:04:09,900
Of technical kind of academic, 
background lines from one and 

73
00:04:09,908 --> 00:04:13,400
the applied side. 
Commercially got to the point 

74
00:04:13,400 --> 00:04:15,800
where I really wanted to build 
something in the space. 

75
00:04:15,800 --> 00:04:20,300
I saw a lot of issues with 
respect to scaling and yeah, 

76
00:04:20,300 --> 00:04:23,800
join the entrepreneur to First 
accelerator met then for anyone 

77
00:04:23,800 --> 00:04:25,700
who doesn't know what the year 
for entrepreneur. 

78
00:04:25,700 --> 00:04:28,400
First accelerator is it's been 
described as love. 

79
00:04:28,400 --> 00:04:34,200
Island meets Shark Tank. 
So join as a individual and then

80
00:04:34,200 --> 00:04:36,700
you find a co-founder and then 
they kind of invest and you saw 

81
00:04:36,800 --> 00:04:41,000
sort of pre idea met then shared
a similar vision for the future 

82
00:04:41,000 --> 00:04:44,200
of a I shared a similar sense of
humor. 

83
00:04:44,800 --> 00:04:49,900
So yeah the rest is history. 
So it seems like you both come 

84
00:04:49,900 --> 00:04:54,200
from a fairly extensive a ID 
planning background. 

85
00:04:54,500 --> 00:04:57,700
What moved you to kind of marry 
this entire thing with 

86
00:04:57,700 --> 00:05:00,300
blockchain? 
Good question. 

87
00:05:00,800 --> 00:05:03,000
It wasn't a sort of instant 
thing. 

88
00:05:03,300 --> 00:05:06,500
It happens over a relatively 
long period of time, to be 

89
00:05:06,508 --> 00:05:09,500
honest and essentially, it was 
technology-driven. 

90
00:05:10,000 --> 00:05:15,600
So we knew we wanted to build 
massive scale, AI infrastructure

91
00:05:16,100 --> 00:05:18,700
and essentially as we were doing
the research to figure out how 

92
00:05:18,700 --> 00:05:21,100
we could make this, the absolute
maximum scale. 

93
00:05:21,300 --> 00:05:24,900
We realized that in order to do 
that, you need to have a 

94
00:05:24,900 --> 00:05:26,900
trustless layer. 
Essentially, you need to be able

95
00:05:26,900 --> 00:05:31,800
to unite compute without To do 
centralized onboarding of new 

96
00:05:31,800 --> 00:05:33,800
providers because at that point 
you end up with an 

97
00:05:33,800 --> 00:05:37,600
administrative kind of like 
scaling limit and we don't want 

98
00:05:37,600 --> 00:05:40,400
any limits. 
So we went down the kind of the 

99
00:05:40,400 --> 00:05:43,800
road of verifiable computation 
research until we hit that kind 

100
00:05:43,800 --> 00:05:47,800
of block of the always has to be
a trusted third party that has 

101
00:05:47,800 --> 00:05:51,200
to be this judge or Arbiter when
you're checking a computation 

102
00:05:51,200 --> 00:05:54,500
who makes a kind of consider a 
decision on whether something's 

103
00:05:54,500 --> 00:05:58,100
been done correctly, blockchain 
represents a way to kind of 

104
00:05:58,100 --> 00:06:00,600
break that and do it. 
Vai Vai by consensus 

105
00:06:00,600 --> 00:06:02,400
essentially. 
So a large group of people can 

106
00:06:02,400 --> 00:06:05,500
do it without having to nominate
a single person to make the 

107
00:06:05,500 --> 00:06:07,400
decision. 
And that was the light bulb 

108
00:06:07,400 --> 00:06:10,500
moment for us where we said, 
like this has to kind of be the 

109
00:06:10,500 --> 00:06:13,200
next step for AI to get the 
scale that we want. 

110
00:06:13,200 --> 00:06:16,200
Like planetary a i scale. 
The has to be this kind of 

111
00:06:16,200 --> 00:06:19,100
consensus later introduced and 
blockchains the way to do it 

112
00:06:19,700 --> 00:06:22,200
before that. 
Interestingly we were kind of 

113
00:06:22,600 --> 00:06:24,500
blockchain Skeptics to an 
extent. 

114
00:06:24,600 --> 00:06:27,400
We hadn't kind of dived into the
space before we'd sort of day, 

115
00:06:27,400 --> 00:06:30,900
taking the typical technical 
path of Of saying read only 

116
00:06:30,900 --> 00:06:33,600
database can do the same thing 
that for I Won't kind of dive 

117
00:06:33,600 --> 00:06:35,400
into it. 
But I know for me personally, 

118
00:06:35,700 --> 00:06:38,900
realizing that kind of trust 
layer was was an absolute light 

119
00:06:38,900 --> 00:06:40,600
bulb moment. 
It was when I realized the kind 

120
00:06:40,600 --> 00:06:44,100
of actual power behind it and 
got very into the space. 

121
00:06:45,500 --> 00:06:48,000
Yeah, interestingly that and I 
shared a lot of the kind of 

122
00:06:48,000 --> 00:06:50,900
ideals that you see Champions 
kind of in The Wider 

123
00:06:50,900 --> 00:06:54,300
decentralization scene. 
So we both were very like large,

124
00:06:54,300 --> 00:06:58,700
free speech maximalists. 
And we kind of a lot of the a 

125
00:06:58,700 --> 00:07:00,700
lot of the kind. 
Of censorship stuff that we saw 

126
00:07:00,700 --> 00:07:02,400
with Snowden and things like 
that. 

127
00:07:02,400 --> 00:07:05,300
We we bonded over prior to even 
talking about blockchain. 

128
00:07:05,300 --> 00:07:08,600
So it kind of felt almost like 
obviously we should have started

129
00:07:08,600 --> 00:07:10,900
in the blockchain space but we 
didn't interestingly right 

130
00:07:10,900 --> 00:07:13,900
before making the switch, we 
were trying to do Federated 

131
00:07:13,900 --> 00:07:16,700
learning which is an area of 
deep learning where you train 

132
00:07:16,700 --> 00:07:19,600
lots of models across 
distributed data and then 

133
00:07:19,600 --> 00:07:23,000
combine them to create a kind of
metal model that can learn from 

134
00:07:23,100 --> 00:07:26,600
all the data sources and we were
doing that with banks. 

135
00:07:27,500 --> 00:07:31,400
So the the kind of Of 
realization for other for me at 

136
00:07:31,400 --> 00:07:32,800
least. 
Well as that, there's a much 

137
00:07:32,800 --> 00:07:36,300
bigger problem with accessing 
computes or essentially just the

138
00:07:36,300 --> 00:07:39,500
processors on which the models 
could be trained and to do that,

139
00:07:39,500 --> 00:07:43,400
you need a decentralized kind of
method of trust and that's 

140
00:07:43,400 --> 00:07:45,500
basically a blockchain. 
Okay. 

141
00:07:45,500 --> 00:07:49,300
So basically it's kind of the 
platform and decentralized 

142
00:07:49,300 --> 00:07:53,200
incentive layer that kind of did
it for you in terms of that 

143
00:07:53,200 --> 00:07:53,900
form. 
Yeah. 

144
00:07:53,900 --> 00:08:00,200
And in terms of moving this to a
blockchain Maybe let's do many. 

145
00:08:00,200 --> 00:08:03,400
Let's talk about AI fast before 
we kind of go into what Jensen 

146
00:08:03,400 --> 00:08:08,600
exactly does as a as a 
blockchain protocol because most

147
00:08:08,600 --> 00:08:12,400
business of this podcast will be
familiar with blockchain to a 

148
00:08:12,400 --> 00:08:19,300
certain extent, but AI is not so
much our user cup of tea. 

149
00:08:19,300 --> 00:08:24,800
So let's talk about the state of
AI today as an outside. 

150
00:08:24,800 --> 00:08:28,400
It kind of seems like it's 
totally on fire. 

151
00:08:28,500 --> 00:08:31,800
Fire. 
I mean with GP T 3 and T PT 

152
00:08:31,800 --> 00:08:34,200
Force concurrent going to come 
out soon. 

153
00:08:34,200 --> 00:08:38,299
I think and then things like 
Dolly and mean it just it's 

154
00:08:38,299 --> 00:08:41,600
completely mind-blowing. 
And can you guys talk about the 

155
00:08:41,600 --> 00:08:44,600
advances in AI in the last 
couple of years? 

156
00:08:46,000 --> 00:08:48,900
Absolutely. 
Yeah, I think it's interesting 

157
00:08:48,900 --> 00:08:51,600
being in the kind of AI space 
and watching this explosion 

158
00:08:51,600 --> 00:08:54,800
happen because the sort of AI 
and machine learning space, over

159
00:08:54,800 --> 00:08:59,900
the past seven years I guess has
basically been a series of mini 

160
00:08:59,900 --> 00:09:02,000
explosions. 
So this one is just kind of a 

161
00:09:02,008 --> 00:09:06,100
next one in the in the sequence.
But I think to The Wider world, 

162
00:09:06,100 --> 00:09:09,200
it's one of the first times 
they've seen it actually create 

163
00:09:09,200 --> 00:09:12,400
real impact, and create 
applications that people see the

164
00:09:12,400 --> 00:09:16,200
value and essentially. 
But yeah, I think Deep learning 

165
00:09:16,200 --> 00:09:17,900
fundamentally has been the big 
change. 

166
00:09:17,900 --> 00:09:19,200
That's kind of enabled all of 
this. 

167
00:09:19,200 --> 00:09:23,400
It was when I first started my 
PhD, the Deep learning kind of 

168
00:09:23,408 --> 00:09:26,900
explosion was just happening. 
It just started it just kind of 

169
00:09:26,900 --> 00:09:29,900
Taken computer vision as an area
by storm. 

170
00:09:29,900 --> 00:09:32,700
They'd shown that essentially 
using a deep neural network, you

171
00:09:32,700 --> 00:09:36,700
could blow away all of the 
benchmarks set by sort of manual

172
00:09:37,100 --> 00:09:38,800
computer vision methods in the 
past. 

173
00:09:38,800 --> 00:09:41,900
So very, very I'll try it. 
Very, very briefly computer 

174
00:09:41,900 --> 00:09:44,900
vision before that used to be 
kind of manually defining sort 

175
00:09:44,900 --> 00:09:49,200
of Otters over images and then 
figuring out how to detect lines

176
00:09:49,200 --> 00:09:51,300
and things. 
And then you would have to 

177
00:09:51,300 --> 00:09:53,800
Define this filter to detect the
kind of line that you're looking

178
00:09:53,800 --> 00:09:56,300
for and textures that you're 
looking for and is a very manual

179
00:09:56,300 --> 00:09:58,300
process. 
Deep learning essentially just 

180
00:09:58,300 --> 00:10:00,900
came on the scene and said we 
can do all of this straight from

181
00:10:00,900 --> 00:10:03,800
the data and that was such a 
huge change. 

182
00:10:03,800 --> 00:10:06,300
It took away all of that kind of
expert knowledge, that was 

183
00:10:06,300 --> 00:10:09,400
required and just allowed 
somebody with enough compute to 

184
00:10:09,400 --> 00:10:12,700
design, a kind of relatively 
simple model, apply it to a very

185
00:10:12,700 --> 00:10:15,000
large amount of data and then 
just have the outcome that they 

186
00:10:15,000 --> 00:10:16,500
want. 
What we're seeing now is 

187
00:10:16,500 --> 00:10:19,900
essentially the kind of building
on top of that building models 

188
00:10:19,900 --> 00:10:23,000
that can do even more and then 
crucially getting them to the 

189
00:10:23,000 --> 00:10:26,100
consumer or to the developer who
doesn't necessarily know the 

190
00:10:26,100 --> 00:10:30,100
specific problem that's been 
going on for years, but image 

191
00:10:30,100 --> 00:10:33,900
and Ali G, PT 3, Etc, of really 
kind of fast track that. 

192
00:10:34,900 --> 00:10:37,500
I don't know if you want to 
speak to some of the deep 

193
00:10:37,500 --> 00:10:39,000
learning stuff as well. 
Hurry, yeah. 

194
00:10:39,000 --> 00:10:42,600
I think whenever we kind of 
talked to crypto crowds about it

195
00:10:43,000 --> 00:10:45,300
at conferences, we always do a 
kind of sharpener. 

196
00:10:45,400 --> 00:10:49,200
Iran, the distinction between 
three terms so the AI machine 

197
00:10:49,200 --> 00:10:51,400
learning and deep learning 
because you're used essentially 

198
00:10:51,400 --> 00:10:54,200
interchangeably but they're 
quite different and the best way

199
00:10:54,200 --> 00:10:56,900
to think about it. 
Our series of kind of like 

200
00:10:56,900 --> 00:10:59,900
circles which are like a 
matroska doll almost wear on the

201
00:10:59,900 --> 00:11:01,200
outside of the big on the 
outside. 

202
00:11:01,200 --> 00:11:03,500
You've got Ai. 
And Ai and bi the loosest 

203
00:11:03,500 --> 00:11:06,200
definition possible with many 
people will disagree with this, 

204
00:11:06,200 --> 00:11:09,400
but the losses definition is, 
it's just programming machine to

205
00:11:09,400 --> 00:11:11,900
do something. 
So, you know, kind of washing 

206
00:11:11,900 --> 00:11:15,100
machine, is in a sense, a narrow
version of artificial 

207
00:11:15,100 --> 00:11:16,700
intelligence. 
And you tell it to do something 

208
00:11:16,700 --> 00:11:20,400
and it kind of programmatically 
does it, or it works out how to 

209
00:11:20,400 --> 00:11:23,800
do it, machine learning kind of 
came into the scene, much more 

210
00:11:23,800 --> 00:11:27,000
prominently in the kind of 90s, 
and the kind of early 2000s, 

211
00:11:27,000 --> 00:11:30,200
where in you instead of having 
has been set expert systems for 

212
00:11:30,200 --> 00:11:34,500
you say, you know, if this then 
this you use data to advance. 

213
00:11:34,600 --> 00:11:37,500
So it's really work out, the 
kind of probability with which a

214
00:11:37,508 --> 00:11:40,800
certain decision will be made, 
deep learning takes that 

215
00:11:40,800 --> 00:11:44,800
concept, but allows different 
kind of Concepts, be modeled 

216
00:11:44,900 --> 00:11:47,700
much. 
More with much more kind of 

217
00:11:47,700 --> 00:11:50,200
fidelity. 
So it kind of has hierarchical 

218
00:11:50,200 --> 00:11:52,700
feature representation which 
means that the way that the 

219
00:11:53,100 --> 00:11:55,700
model Works, different parts, 
learn different things. 

220
00:11:56,600 --> 00:11:58,900
If, for example, if you're the 
classic example is if you want 

221
00:11:58,900 --> 00:12:02,900
to recognize handwritten letters
and your network typically 

222
00:12:02,900 --> 00:12:05,500
pushes the image through lots of
different layers. 

223
00:12:05,800 --> 00:12:09,100
Each layer will kind of pick up 
something like a kind of all 

224
00:12:09,100 --> 00:12:11,100
this. 
This kind of number has a has a 

225
00:12:11,108 --> 00:12:15,000
close look in it or as a stem 
and then we're over time and 

226
00:12:15,000 --> 00:12:18,500
over Some kind of computational 
cycles and lots of tweaking. 

227
00:12:18,600 --> 00:12:21,300
The model will be able to 
generalize any new image. 

228
00:12:21,300 --> 00:12:24,700
It sees to one of these kind of 
categories, you know, a number 

229
00:12:24,700 --> 00:12:28,300
between 0 and 9. 
So that's basically the 

230
00:12:28,300 --> 00:12:31,200
distinction between a IML deep 
learning deep learning is where 

231
00:12:31,200 --> 00:12:34,000
you see all the kind of big 
breakthroughs coming in. 

232
00:12:34,000 --> 00:12:36,700
So all the things you mentioned,
then gbd3, Thalia search for 

233
00:12:36,700 --> 00:12:40,600
stuff like stable, diffusion, 
all thats deep learning and the 

234
00:12:40,600 --> 00:12:44,100
story for deep learning over the
past, kind of, I guess, you 

235
00:12:44,100 --> 00:12:47,200
know, I guess it's like 20 
years. 16 2015 has been 

236
00:12:47,200 --> 00:12:50,000
transformed, our models, which 
are a specific type of deep 

237
00:12:50,000 --> 00:12:52,900
learning model that have been 
very useful for things like 

238
00:12:52,900 --> 00:12:55,900
large language modeling. 
I think what's crucial as well 

239
00:12:55,900 --> 00:12:57,800
as a kind of more social point 
is. 

240
00:12:58,600 --> 00:13:02,300
If you told people, you know, at
the beginning of the 2010's that

241
00:13:02,300 --> 00:13:05,700
they'd be able to essentially 
generate a comic book, which is 

242
00:13:05,700 --> 00:13:10,300
in really kind of convincing 
with really convincing art. 

243
00:13:10,700 --> 00:13:12,400
Just from a series of text 
prompts. 

244
00:13:12,800 --> 00:13:14,700
I honestly don't think most 
people would believe that's 

245
00:13:14,700 --> 00:13:16,200
possible. 
Particularly the kind of 

246
00:13:16,208 --> 00:13:19,000
consumer grade for like a normal
person, just to be able to type 

247
00:13:19,000 --> 00:13:22,500
text prompts to create a comic 
book in the next few years. 

248
00:13:22,900 --> 00:13:25,700
The kind of same order of 
magnitude jump is going to 

249
00:13:25,700 --> 00:13:30,100
happen. 
So in the 2020s, the ability to 

250
00:13:30,100 --> 00:13:33,700
sit down in front of say, you 
know, Netflix and instead of 

251
00:13:33,700 --> 00:13:37,100
picking a movie which has been, 
you know, pre-made you simply 

252
00:13:37,100 --> 00:13:40,800
enter a text prompt and you're 
like, you know, I want to see 

253
00:13:41,100 --> 00:13:44,400
free technologists talking a 
podcast for an hour, you know, 

254
00:13:44,400 --> 00:13:49,000
about about AI or something. 
And with other kind of prompts, 

255
00:13:49,000 --> 00:13:52,600
and maybe like a kind of set of 
initializations, you'll be able 

256
00:13:52,600 --> 00:13:55,900
to generate an entire movie, 
which you can then kind of steer

257
00:13:55,900 --> 00:13:59,000
or maybe a different place if 
you want or as a final point, 

258
00:13:59,100 --> 00:14:01,600
maybe you have the same story 
but you can change the genre of 

259
00:14:01,600 --> 00:14:03,800
the story. 
So you could turn something 

260
00:14:03,800 --> 00:14:07,400
like, you know, I don't know, 
Halloween into like a sci-fi 

261
00:14:07,400 --> 00:14:10,000
movie, where you can change 
Jurassic Park, it's a love story

262
00:14:10,000 --> 00:14:13,300
or something, all by changing 
the using same script that 

263
00:14:13,300 --> 00:14:16,600
changing the kind of rendering, 
Lots of exciting things coming. 

264
00:14:16,600 --> 00:14:21,300
In my opinion can't Kenny talked
about kind of the Paradigm Shift

265
00:14:21,300 --> 00:14:23,500
behind this. 
So basically I mean if you look 

266
00:14:23,500 --> 00:14:26,800
at like old-school programming 
it's a lot of deterministic if 

267
00:14:26,800 --> 00:14:30,700
this then that and so on. 
And and in my understanding and 

268
00:14:30,800 --> 00:14:34,500
admittedly the, this is a very 
layer understanding you kind of 

269
00:14:34,500 --> 00:14:39,900
you use like some sort of neural
network with, you know, like 

270
00:14:39,900 --> 00:14:43,700
complex connectivity. 
And where exactly is. 

271
00:14:43,900 --> 00:14:48,700
I mean, to Do people exactly 
understand how decisions in a 

272
00:14:48,700 --> 00:14:54,600
neural network actually reached?
Is this something that can could

273
00:14:54,600 --> 00:15:01,400
you kind of Transport this? 
But I mean, obviously, you don't

274
00:15:01,400 --> 00:15:05,300
use real neural networks, right?
So basically everything's in a 

275
00:15:05,300 --> 00:15:09,900
regular computer your, you don't
have to go to like the bio lab, 

276
00:15:10,700 --> 00:15:13,700
although, that would be not. 
Yeah, I'm not sure whether that 

277
00:15:13,700 --> 00:15:16,800
would be terrifying or fun. 
Um, but basically everything's 

278
00:15:16,800 --> 00:15:18,900
anyways in a computer. 
So basically kind of your 

279
00:15:18,900 --> 00:15:26,800
modeling like in a different 
system that's kind of more 

280
00:15:26,800 --> 00:15:30,000
interconnected and more flexible
and maybe you can, maybe you can

281
00:15:30,000 --> 00:15:34,000
kind of qualify how the system 
you're modeling with your 

282
00:15:34,000 --> 00:15:38,300
regular computers different from
just, you know, just giving the 

283
00:15:38,308 --> 00:15:42,100
computers prompts. 
I mean yeah sure. 

284
00:15:42,100 --> 00:15:46,000
I think the black box sort of 
nature of Learning models is 

285
00:15:46,000 --> 00:15:48,500
just down to the absolute size 
of them. 

286
00:15:49,100 --> 00:15:51,500
At the end of the day, you're 
still tracing a path through a 

287
00:15:51,500 --> 00:15:54,800
series of kind of decision 
points in the, in the network. 

288
00:15:54,900 --> 00:15:57,900
It's just that path is 
absolutely enormous and it's 

289
00:15:57,900 --> 00:16:02,200
hard to kind of Link the weights
or the parameters within that 

290
00:16:02,200 --> 00:16:05,500
model down to exactly why 
they're that sort of value. 

291
00:16:05,500 --> 00:16:09,500
Because they've come to that 
value after being fed millions 

292
00:16:09,500 --> 00:16:13,600
of samples and you can 
deterministically, you could do 

293
00:16:13,600 --> 00:16:17,100
that, you could track every 
single update, but the size of 

294
00:16:17,100 --> 00:16:19,200
data that you would end up 
generating, would be absolutely 

295
00:16:19,200 --> 00:16:23,000
enormous. 
I think it's the sort two things

296
00:16:23,000 --> 00:16:25,500
that I see happening. 
As we kind of go through. 

297
00:16:25,500 --> 00:16:29,300
This one, is the Black Box. 
Nature, is sort of falling away 

298
00:16:29,300 --> 00:16:31,200
a little bit. 
As we start to understand more 

299
00:16:31,200 --> 00:16:34,000
and more about the models that 
were building deep learning as a

300
00:16:34,008 --> 00:16:38,400
kind of research area is sort of
gone through a, an interesting 

301
00:16:38,400 --> 00:16:42,100
fast period where there's been a
lot of experimentation that 

302
00:16:42,100 --> 00:16:45,900
wasn't driven by the sort of Of 
the research. 

303
00:16:45,900 --> 00:16:49,000
It was more driven by seeing 
what we could get out of it. 

304
00:16:49,100 --> 00:16:52,200
So we throw more data at it. 
We try out new architectures and

305
00:16:52,200 --> 00:16:54,600
we just see what happens rather 
than starting from first 

306
00:16:54,600 --> 00:16:58,000
principles and designing this 
thing and knowing exactly how it

307
00:16:58,000 --> 00:16:59,400
works. 
So, it's been that kind of 

308
00:16:59,400 --> 00:17:02,300
exciting period where everything
has been very black box. 

309
00:17:02,400 --> 00:17:05,200
I think a lot of the gains that 
happened there or something to 

310
00:17:05,200 --> 00:17:08,400
sort of starting to slow down a 
little bit and we're seeing 

311
00:17:08,400 --> 00:17:10,800
people revisit those 
architectures and sort of check 

312
00:17:10,800 --> 00:17:14,099
and say, why does this work so 
well, let's dig into it and 

313
00:17:14,099 --> 00:17:15,500
that's kind of prove it out. 
Out. 

314
00:17:15,500 --> 00:17:17,800
So, in some ways that kind of 
curtain is lifting. 

315
00:17:18,300 --> 00:17:20,099
The other thing that's happening
which is a bit more 

316
00:17:20,099 --> 00:17:23,800
controversial, I guess, is the 
shift in people's perspectives 

317
00:17:23,800 --> 00:17:28,300
as to whether a kind of 
computational system needs to be

318
00:17:28,300 --> 00:17:30,800
fully deterministic or whether 
we can live in a probabilistic 

319
00:17:30,800 --> 00:17:32,700
world. 
We live in a probabilistic world

320
00:17:32,700 --> 00:17:35,400
as people that's kind of 
self-driving. 

321
00:17:35,400 --> 00:17:38,000
Cars example is probably the 
clearest where, when we're 

322
00:17:38,000 --> 00:17:40,600
driving around, we accept that 
there are kind of stochastic 

323
00:17:40,600 --> 00:17:44,300
events that happen and that the 
can be small accidents in the 

324
00:17:44,300 --> 00:17:47,400
can be issues that Button with a
self-driving car system. 

325
00:17:47,400 --> 00:17:50,100
We don't accept that at all and 
we say that this has to be a 

326
00:17:50,100 --> 00:17:52,400
fully completely deterministic 
process. 

327
00:17:52,700 --> 00:17:55,200
I think one of the challenges 
that the self-driving car 

328
00:17:55,200 --> 00:17:57,900
industry, as hard as been an 
assumption that people would 

329
00:17:57,900 --> 00:18:00,700
just accept that probabilistic 
mechanism, applied to 

330
00:18:00,700 --> 00:18:03,400
self-driving cars, and they 
haven't, but I think that will 

331
00:18:03,400 --> 00:18:04,600
change. 
And that's the probably the 

332
00:18:04,600 --> 00:18:07,500
controversial but as we as a 
society go towards actually 

333
00:18:07,500 --> 00:18:11,200
allowing kind of probabilistic 
computational systems to exist, 

334
00:18:11,200 --> 00:18:14,100
alongside us. 
Not sure if it'll be an easy 

335
00:18:14,100 --> 00:18:16,300
road but But I think it'll 
happen. 

336
00:18:17,500 --> 00:18:22,500
Yeah, thank you before we dive 
into the current landscape. 

337
00:18:22,500 --> 00:18:26,500
There's one term I have come 
across often kind of in 

338
00:18:26,500 --> 00:18:29,000
preparing for this episode. 
Also maybe that's a question for

339
00:18:29,000 --> 00:18:32,700
Harry because you already talked
about the different kind of 

340
00:18:32,900 --> 00:18:35,800
machine learning, deep, learning
artificial intelligence. 

341
00:18:35,800 --> 00:18:38,600
So basically there's this time 
of artificial general 

342
00:18:38,600 --> 00:18:43,200
intelligence is said different 
from the three terms you already

343
00:18:43,200 --> 00:18:45,000
talked about. 
Yes. 

344
00:18:45,700 --> 00:18:47,100
So it's a term which was 
popular. 

345
00:18:47,200 --> 00:18:50,800
Yes, I believe by at Bangor so 
who's on? 

346
00:18:50,800 --> 00:18:55,200
AI researcher and entrepreneur. 
The idea of AGI is similar to 

347
00:18:55,500 --> 00:19:00,100
also the singularity so it's the
idea that you get human level 

348
00:19:00,100 --> 00:19:04,500
intelligence from a machine. 
So you have right now, what you 

349
00:19:04,500 --> 00:19:07,500
might describe as a kind of like
artificial narrow intelligence 

350
00:19:07,500 --> 00:19:10,200
for by machines are good at 
doing certain tasks. 

351
00:19:10,200 --> 00:19:13,300
So for example, machines are 
very, very good at detecting 

352
00:19:13,300 --> 00:19:17,000
certain types of cancer from 
from medical scans. 

353
00:19:17,200 --> 00:19:20,200
Right, so pattern recognition, 
yes, yeah. 

354
00:19:20,300 --> 00:19:24,300
But kind of scaling that up to 
general intelligence for by a 

355
00:19:24,300 --> 00:19:28,900
machine, can be good at doing a 
task, which is kind of may be 

356
00:19:28,900 --> 00:19:31,500
simple to humans, but actually 
quite difficult to reflect in a 

357
00:19:31,508 --> 00:19:34,300
kind of computational. 
Like, can you give an example 

358
00:19:34,400 --> 00:19:36,500
prediction space? 
Yeah, good example, would be a 

359
00:19:36,508 --> 00:19:40,400
machine being able to walk 
through a crowded area in a 

360
00:19:40,400 --> 00:19:44,500
smooth way. 
Whilst being able to essentially

361
00:19:44,500 --> 00:19:47,900
make get a discrete assumptions 
about all the The inputs around 

362
00:19:47,900 --> 00:19:49,700
it. 
It's one of the reasons that I 

363
00:19:49,700 --> 00:19:52,900
can't believe I can't remember 
the level of driverless cars. 

364
00:19:52,900 --> 00:19:54,500
I think it's like maybe level 10
or something. 

365
00:19:54,500 --> 00:19:57,100
It's one of the reasons that 
driverless cars do really well 

366
00:19:57,100 --> 00:20:00,700
on the motorway because it's a 
very kind of, it's a kind of, 

367
00:20:00,800 --> 00:20:02,900
it's a problem was to humans 
might feel quite complex but 

368
00:20:02,900 --> 00:20:06,000
it's quite like a simple sort of
mathematical problem because 

369
00:20:06,000 --> 00:20:09,700
there's not much variation, but 
when you take that same car and 

370
00:20:09,700 --> 00:20:13,400
you put it kind of in a city 
street in Rome, you know, going 

371
00:20:13,400 --> 00:20:16,300
over cobbles Sakura squawking 
out in front of everything, it 

372
00:20:16,300 --> 00:20:18,000
becomes extremely I'm Lee 
difficult. 

373
00:20:18,700 --> 00:20:22,000
So it's kind of yeah, some of 
the stuff which we think is 

374
00:20:22,008 --> 00:20:24,700
really kind of difficult like 
being really good at chess. 

375
00:20:24,700 --> 00:20:26,900
It's actually quite easy for a 
machine but some of the stuff 

376
00:20:26,900 --> 00:20:29,200
that we think is really easy 
like being able to kind of walk 

377
00:20:29,200 --> 00:20:32,900
down the street or, you know, 
being able to like I guess 

378
00:20:33,700 --> 00:20:36,200
certain certain kind of things 
in conversation, like, you know,

379
00:20:36,200 --> 00:20:39,800
understanding looking at some of
the tire body language and 

380
00:20:39,900 --> 00:20:42,300
looking at everything of someone
saying, and being able to kind 

381
00:20:42,300 --> 00:20:45,800
of withdraw an emotion from 
that, there might be, she won't 

382
00:20:45,800 --> 00:20:48,300
be good at various things like 
Estimation, you know, how 

383
00:20:48,300 --> 00:20:51,000
someone sitting but combining 
that all together and making a 

384
00:20:51,000 --> 00:20:53,400
kind of decision, it's quite 
difficult. 

385
00:20:53,400 --> 00:20:56,600
So yeah. 
Artificial general intelligence 

386
00:20:56,600 --> 00:21:01,100
basically means a model for a 
set of models or a system which 

387
00:21:01,100 --> 00:21:04,600
is able to essentially be as 
good as humans at everyday 

388
00:21:04,600 --> 00:21:06,800
tasks. 
Critically. 

389
00:21:07,400 --> 00:21:12,500
The kind of Advent of AGI leads 
to artificial super intelligence

390
00:21:12,600 --> 00:21:15,000
because it follows that once a 
machine is kind of mastered 

391
00:21:15,000 --> 00:21:16,700
everything, a human would 
reasonably do. 

392
00:21:17,100 --> 00:21:20,900
Their rate of kind of marginal 
Mastery, over tasks moves a lot 

393
00:21:20,900 --> 00:21:24,400
faster than humans because of as
a kind of function of both the 

394
00:21:24,400 --> 00:21:26,700
kind of complexity of their 
model and the amount of computer

395
00:21:26,700 --> 00:21:28,900
available to them. 
So if we threw all the computer 

396
00:21:28,900 --> 00:21:32,400
in the world at a model which is
already a human level, it's got 

397
00:21:32,400 --> 00:21:34,900
much more energy than the kind 
of normal human does and it's 

398
00:21:34,900 --> 00:21:38,600
also got infinite lifespan and 
it's also got a perfect memory. 

399
00:21:38,900 --> 00:21:41,800
So horribly Your Perfect Memory.
So it kind of that's where you 

400
00:21:41,808 --> 00:21:44,000
kind of get into the realm of 
kind of Science, Fiction, horror

401
00:21:44,000 --> 00:21:48,700
movies but this is what eat on 
is afraid of Yes, you hear a lot

402
00:21:48,700 --> 00:21:53,200
and these kind of examples and 
you here also kind of what one 

403
00:21:53,200 --> 00:21:56,500
of the kind of Pathways that 
people are at least I estimate 

404
00:21:56,500 --> 00:22:00,400
will kind of take is, there is 
the kind of fusion, of humans 

405
00:22:00,400 --> 00:22:02,400
and machines. 
So, for example, if you have a 

406
00:22:02,400 --> 00:22:06,100
kind of brain-computer interface
or brain machine interface BMI, 

407
00:22:06,500 --> 00:22:10,900
and you're able to essentially 
augment your lived experience 

408
00:22:10,900 --> 00:22:15,800
with, you know, machine kind of 
inputs that machine, learns from

409
00:22:15,800 --> 00:22:18,800
all your kind of the way. 
Your brains working and firing 

410
00:22:18,800 --> 00:22:21,400
it learns patterns, you're 
helping it train is going to 

411
00:22:21,400 --> 00:22:25,100
helping you train your own brain
and that's going to help speed 

412
00:22:25,100 --> 00:22:28,500
up that process as well. 
It raises a kind of you know, 

413
00:22:28,700 --> 00:22:32,300
Treasure Trove of ethical could 
you know, kind of issues. 

414
00:22:33,300 --> 00:22:36,300
But the yeah that's basically 
definition of AGI and then 

415
00:22:36,300 --> 00:22:39,300
subsequently a SI artificial. 
Super intelligence. 

416
00:22:40,200 --> 00:22:44,100
Who is super interesting. 
So let's look at what the 

417
00:22:44,200 --> 00:22:46,000
landscape currently looks like, 
right? 

418
00:22:46,000 --> 00:22:48,500
So basically say I want to In a,
a martyr. 

419
00:22:49,400 --> 00:22:51,900
Where do I buy a? 
I compute? 

420
00:22:51,900 --> 00:22:55,500
So, I mean, I could just get an 
instance on AWS or I could run 

421
00:22:55,500 --> 00:22:58,400
it on my local machine. 
So kind of walk me through 

422
00:22:58,800 --> 00:23:00,600
through the options. 
Yeah. 

423
00:23:00,600 --> 00:23:03,700
So it really depends on the 
scale and model your training. 

424
00:23:03,900 --> 00:23:06,900
If you're a kind of student 
learning about AI, maybe you're 

425
00:23:06,900 --> 00:23:10,800
an undergrad, you typically just
use AWS or for small enough 

426
00:23:10,800 --> 00:23:12,500
models. 
Your local machine. 

427
00:23:12,500 --> 00:23:16,900
As you can say, the next level 
up, you might be a kind of 

428
00:23:17,200 --> 00:23:19,100
Startup. 
You've just burned through your 

429
00:23:19,100 --> 00:23:22,500
kind of hundred K of a and 
credits and you're kind of 

430
00:23:22,500 --> 00:23:25,300
looking at the kind of marginal 
cost of training models. 

431
00:23:26,300 --> 00:23:29,600
You might go for an on-demand. 
AWS instance, you might go for 

432
00:23:29,600 --> 00:23:32,300
something more kind of fixed. 
We're going to permanent which 

433
00:23:32,300 --> 00:23:35,300
is typically cheaper when you 
have booked them and in advance 

434
00:23:35,600 --> 00:23:38,900
but the reaches a certain point 
when you're training models that

435
00:23:38,900 --> 00:23:44,900
use a kind of are experiencing 
enormous cost in AWS or B, you 

436
00:23:44,900 --> 00:23:47,100
can actually achieve the scale 
required. 

437
00:23:47,300 --> 00:23:50,600
Terms of gpus. 
So you just get kind of limited 

438
00:23:50,600 --> 00:23:55,200
by AWS in terms of scale at that
point, you see companies go in 

439
00:23:55,200 --> 00:23:57,300
house. 
So in our, in our kind of 

440
00:23:57,300 --> 00:24:00,400
research prior to raising our 
last funding round bed, and I 

441
00:24:00,400 --> 00:24:03,100
spoke to about 150 machine, 
learning, researchers, and 

442
00:24:03,100 --> 00:24:06,400
Engineers, a variety of places 
from going to find companies to 

443
00:24:06,800 --> 00:24:10,900
startups to Academia and whilst 
a lot of academics at top 

444
00:24:10,900 --> 00:24:14,600
universities, have access to 
kind of clusters and large, can 

445
00:24:14,600 --> 00:24:17,300
high-performance compute. 
And people, it's a We have 

446
00:24:17,300 --> 00:24:21,000
access to the fear research lab 
supercluster there, which is 

447
00:24:21,000 --> 00:24:24,300
biggest a, I lost him. 
The world, most people in our 

448
00:24:24,300 --> 00:24:26,900
experience didn't manage to get 
the scale that the that they 

449
00:24:26,900 --> 00:24:29,700
wanted. 
And one of the ways that some of

450
00:24:29,700 --> 00:24:32,200
them kind of dealt with that 
would be, they buy gpus 

451
00:24:32,200 --> 00:24:35,000
themselves and they bring them 
in house and then we manage 

452
00:24:35,000 --> 00:24:36,000
them. 
And we heard all these horror 

453
00:24:36,000 --> 00:24:39,700
stories about people like in 
south of England, having a spare

454
00:24:39,700 --> 00:24:42,500
bedroom with a fan in it and 
loads of gpus, it's a good, I 

455
00:24:42,508 --> 00:24:45,700
like a bum, like a Bitcoin miner
upstairs, and like, also people 

456
00:24:45,700 --> 00:24:47,000
who would have them in their 
office. 

457
00:24:47,200 --> 00:24:50,200
Then it's a bit of a kind of 
fragmented Market. 

458
00:24:50,900 --> 00:24:54,500
However, basically, the bottom 
line is, if you buy the gpus 

459
00:24:54,500 --> 00:24:58,800
outright typically it costs less
marginally over the long term to

460
00:24:58,800 --> 00:25:00,700
run them. 
And that's a function of 

461
00:25:00,700 --> 00:25:05,400
basically not having to pay the 
sorts of 65% ish, premium for, 

462
00:25:05,900 --> 00:25:10,100
or should see margin for 
accessing Amazon ec2 instances. 

463
00:25:10,900 --> 00:25:14,100
So, those that's kind of cloud, 
local, or kind of get in your 

464
00:25:14,100 --> 00:25:15,800
own cluster. 
There's also high performance 

465
00:25:15,800 --> 00:25:18,300
compute if you're in Academia, 
You have access to that type of 

466
00:25:18,300 --> 00:25:20,700
compute but then again, it can 
be bottlenecks there. 

467
00:25:21,100 --> 00:25:24,900
There's other kind of options. 
So for example, if you're a kind

468
00:25:24,900 --> 00:25:28,500
of, I guess the neverland 
organization and you're wanting 

469
00:25:28,500 --> 00:25:32,600
to solve our highly 
parallelizable computer, science

470
00:25:32,600 --> 00:25:35,500
problem example, that would be 
like folding at home. 

471
00:25:36,200 --> 00:25:38,700
You can, you can access 
volunteer, computer networks 

472
00:25:38,700 --> 00:25:42,700
using things like boinc from 
from Berkeley originally. 

473
00:25:42,700 --> 00:25:44,500
It may be a lot listeners. 
Will remember things like seti 

474
00:25:44,500 --> 00:25:46,500
at home. 
Well, it's not, you know, it's 

475
00:25:46,500 --> 00:25:50,000
not It's not machine learning. 
It's kind of just on a 

476
00:25:50,000 --> 00:25:53,200
processing signals. 
It's a really good example of 

477
00:25:53,400 --> 00:25:56,300
grid Computing, reaching very 
large scale. 

478
00:25:57,100 --> 00:25:59,800
I think, right now that folding 
at home, which is kind of 

479
00:25:59,800 --> 00:26:04,000
successor has the largest kind 
of compute volume anywhere in 

480
00:26:04,000 --> 00:26:07,300
the world even greater mystery 
book and supercomputers like 

481
00:26:07,300 --> 00:26:12,500
'fuck a coup. 
So yeah, to summarize you have 

482
00:26:12,500 --> 00:26:17,100
you kind of go from your local 
machine onto the cloud maybe via

483
00:26:17,300 --> 00:26:19,400
High-performance cluster at 
University or and then 

484
00:26:19,600 --> 00:26:22,500
ultimately back off the clouds, 
taking it back on program. 

485
00:26:23,500 --> 00:26:26,900
The goal of Jensen, as a segue 
is to give everyone access to 

486
00:26:26,900 --> 00:26:30,500
the same kind of compute scale 
that the people who currently 

487
00:26:30,500 --> 00:26:34,000
have on Prime clusters can 
achieve and crucially to do. 

488
00:26:34,000 --> 00:26:39,400
So, in a way, which allows Fair 
access so kind of it's not can't

489
00:26:39,400 --> 00:26:41,900
be turned off by centralized 
entity. 

490
00:26:42,800 --> 00:26:46,000
There have been projects like 
this in the blockchain space 

491
00:26:46,100 --> 00:26:50,400
before one of Very, very old by 
blocking standards projects. 

492
00:26:50,400 --> 00:26:52,100
It's got em. 
I believe they actually did 

493
00:26:52,100 --> 00:26:57,200
their SEO and 2016, which is 
basically, like 50 years ago and

494
00:26:57,200 --> 00:26:58,000
block him. 
Yes. 

495
00:26:58,200 --> 00:27:02,600
So how does that Jensen compared
to Gollum? 

496
00:27:03,500 --> 00:27:06,100
Yes, great question. 
So, we've been given kind of two

497
00:27:06,100 --> 00:27:09,000
axes. 
The first one is the kind of 

498
00:27:09,300 --> 00:27:11,400
thinness of the protocol, so to 
speak. 

499
00:27:11,500 --> 00:27:15,200
So, Golems are General compute 
protocol, you can do lots of 

500
00:27:15,200 --> 00:27:19,100
things on it, and we are, Finn 
protocol more similar to like 

501
00:27:19,100 --> 00:27:22,100
render protocol. 
If you want the kind of analog 

502
00:27:22,100 --> 00:27:26,400
there where we do one thing and 
that's training, machine 

503
00:27:26,400 --> 00:27:30,400
learning models. 
The the second kind of point is 

504
00:27:30,400 --> 00:27:33,500
on the kind of scalability of 
the verification. 

505
00:27:33,900 --> 00:27:36,600
So what we see in a lot of the 
kind of earlier projects is a 

506
00:27:36,600 --> 00:27:41,100
tendency to use things like 
reputation or to use kind of 

507
00:27:41,100 --> 00:27:44,500
less Byzantine tolerant or fault
all, and I should say, methods 

508
00:27:44,500 --> 00:27:47,100
of replication when we looked at
those. 

509
00:27:48,500 --> 00:27:51,600
Those kind of those 
architectures for verification 

510
00:27:51,600 --> 00:27:53,800
systems, be just didn't work for
us. 

511
00:27:53,800 --> 00:27:56,000
As people who train machine 
learning models, we just 

512
00:27:56,000 --> 00:27:59,300
wouldn't, we wouldn't have 
enough faith in the results. 

513
00:27:59,300 --> 00:28:00,300
Doesn't mean that you don't 
work. 

514
00:28:00,300 --> 00:28:03,100
It just four of his comfort 
purely for kind of machine 

515
00:28:03,100 --> 00:28:04,900
learning. 
We just weren't such a villain 

516
00:28:04,900 --> 00:28:07,900
when we had conversations with 
kind of web to machine learning 

517
00:28:07,900 --> 00:28:11,800
people, they kind of agreed. 
So for us, the goal was to 

518
00:28:11,800 --> 00:28:14,900
basically take a lot of those 
initial learnings around, how do

519
00:28:14,900 --> 00:28:19,200
you kind of position? 
Our you protocol in this kind of

520
00:28:19,200 --> 00:28:23,200
world crypto world, but do it 
crystalline away which is only 

521
00:28:23,200 --> 00:28:26,000
for machine learning. 
So you can make super, you know,

522
00:28:26,000 --> 00:28:29,700
optimizations around the the 
kind of speed and that the cost 

523
00:28:29,700 --> 00:28:32,800
of the protocol number one. 
But number two, how do you can 

524
00:28:32,800 --> 00:28:36,400
reach a satisfactory level 
verification right now, that 

525
00:28:36,400 --> 00:28:39,500
verification and consensus 
pieces, really like the vast 

526
00:28:39,500 --> 00:28:44,200
majority of our time and energy.
You know it's it's the question 

527
00:28:44,200 --> 00:28:49,400
and we had a good initial stab 
at it with Our kind of inaugural

528
00:28:49,400 --> 00:28:53,000
light paper but we've expanded 
on it since since then I don't 

529
00:28:53,000 --> 00:28:54,500
know been if you'd add anything 
to that. 

530
00:28:55,400 --> 00:28:57,600
Yeah. 
Probably just to emphasize the 

531
00:28:57,700 --> 00:29:00,600
kind of general purpose approach
that most people before have 

532
00:29:00,600 --> 00:29:02,700
taken it's quite an attractive 
one. 

533
00:29:02,800 --> 00:29:05,400
You want to get the biggest 
Market you can possibly kind of 

534
00:29:05,700 --> 00:29:07,700
get to. 
So saying we do general-purpose 

535
00:29:07,700 --> 00:29:11,100
computation any scale, any kind 
of computational problem is 

536
00:29:11,100 --> 00:29:14,200
attractive at first, but you 
fall so quickly into the two 

537
00:29:14,200 --> 00:29:17,500
traps Harry mentioned the first 
drop is the very first Vacation 

538
00:29:17,500 --> 00:29:19,200
problem. 
It's very, very, very difficult.

539
00:29:19,200 --> 00:29:23,500
Our thesis is, you have to 
narrow and that we will have a 

540
00:29:23,500 --> 00:29:26,700
big sort of set of thin 
protocols at the bottom of the 

541
00:29:26,700 --> 00:29:29,700
kind of the decentralized 
infrastructure stack if you 

542
00:29:29,700 --> 00:29:33,400
think about AWS. 
But in web three, we think all 

543
00:29:33,400 --> 00:29:35,200
of the kind of functionality 
that exists. 

544
00:29:35,200 --> 00:29:37,600
There will be ported over and it
exists is this sort of 

545
00:29:37,608 --> 00:29:40,800
hierarchical stack of things 
getting closer and closer to the

546
00:29:40,800 --> 00:29:43,800
user as you go up and on the 
bottom is protocols, like 

547
00:29:43,800 --> 00:29:47,800
Jensen, protocols, like render 
token where you do want Specific

548
00:29:47,800 --> 00:29:51,000
type of computation really 
efficiently with really strong 

549
00:29:51,000 --> 00:29:52,900
verification. 
And then, on top of that, you 

550
00:29:52,900 --> 00:29:55,100
can have the kind of general 
purpose computer networks that 

551
00:29:55,100 --> 00:29:58,500
fall back onto that. 
So that's our kind of vision for

552
00:29:58,500 --> 00:30:00,800
the, the decentralized 
infrastructure. 

553
00:30:00,800 --> 00:30:03,800
I think, as part of that, when 
you launch is one of those then,

554
00:30:03,800 --> 00:30:06,700
protocols, you have a much 
easier job in initially 

555
00:30:06,700 --> 00:30:10,000
targeting your market. 
So, our Market isn't doing kind 

556
00:30:10,000 --> 00:30:12,100
of like chess simulations and 
things like that. 

557
00:30:12,100 --> 00:30:14,500
That just building machine 
learning models, that's it. 

558
00:30:14,800 --> 00:30:17,400
It can be really sort of 
attractive to say, I We could 

559
00:30:17,400 --> 00:30:19,600
just do this extra thing. 
We could do this extra thing, 

560
00:30:19,600 --> 00:30:22,100
maybe we could attach ourselves 
to an existing sort of thing 

561
00:30:22,100 --> 00:30:24,800
that's quite popular right now. 
Maybe we could generate ftes 

562
00:30:24,800 --> 00:30:27,400
things like that but I think 
when you do that you split the 

563
00:30:27,400 --> 00:30:30,000
mindshare massively in terms of 
product and people don't know 

564
00:30:30,000 --> 00:30:34,000
what you are for Jensen. 
We will always be very clear 

565
00:30:34,000 --> 00:30:36,700
that where machine learning 
compute if that's what you want.

566
00:30:36,700 --> 00:30:38,700
Then you come here, if you want 
something else, you go to a 

567
00:30:38,700 --> 00:30:41,500
different protocol, maybe it 
falls back onto Jensen at some 

568
00:30:41,500 --> 00:30:43,600
point. 
But fundamentally, that's all 

569
00:30:43,600 --> 00:30:45,700
kind of we are. 
And I think they'll very long 

570
00:30:45,700 --> 00:30:47,100
term of it is we're behind the 
scenes. 

571
00:30:47,200 --> 00:30:51,100
Since we're just like HTTP, but 
for machine learning compute to 

572
00:30:51,100 --> 00:30:53,500
an end user and a developer, you
won't even know that Jensen 

573
00:30:53,500 --> 00:30:55,600
exists. 
All, you know, is that the world

574
00:30:55,600 --> 00:30:57,000
has changed. 
And now, when you train a 

575
00:30:57,000 --> 00:30:59,800
machine learning model, it goes 
out somewhere and it gets 

576
00:30:59,800 --> 00:31:03,300
performed by someone in the 
world through a series of kind 

577
00:31:03,300 --> 00:31:06,700
of apps and apps, and things 
until it eventually sits on the 

578
00:31:06,708 --> 00:31:09,300
Jensen protocol. 
We think that's the kind of best

579
00:31:09,300 --> 00:31:13,000
way to provide this compute to 
the world is via that kind of 

580
00:31:13,000 --> 00:31:15,700
hierarchical infrastructure 
where we gradually go more and 

581
00:31:15,700 --> 00:31:19,700
more behind-the-scenes. 
Ali-A, dad had one final point 

582
00:31:19,700 --> 00:31:23,900
to that which is there's where 
we think about the kind of 

583
00:31:23,908 --> 00:31:26,700
properties that the network has 
to have it needs to be targeted 

584
00:31:26,700 --> 00:31:29,000
towards machine, learning 
engineers and researchers. 

585
00:31:29,000 --> 00:31:31,600
At least have the verification 
piece but crucially on the kind 

586
00:31:31,600 --> 00:31:34,700
of permissionless site in needs 
to have that level of sort of 

587
00:31:35,100 --> 00:31:40,000
censorship resistance but also 
kind of an agnostic relationship

588
00:31:40,000 --> 00:31:43,900
with Hardware. 
So in the kind of, I guess deep 

589
00:31:43,900 --> 00:31:47,900
learning Hardware space, you 
know dominated by Like Nvidia, 

590
00:31:48,300 --> 00:31:50,600
there's companies which are 
doing their own proprietary 

591
00:31:50,600 --> 00:31:54,500
Asics, like Google through TP. 
Use sensor processing units or a

592
00:31:54,508 --> 00:31:58,300
graph core and a good one with 
their IP use intelligent 

593
00:31:58,900 --> 00:32:04,100
conscious processing units. 
What kind of trap is I feel 

594
00:32:04,100 --> 00:32:06,500
which some protocols not even in
the kind of deep learning space 

595
00:32:06,500 --> 00:32:09,400
of gone done before is shipping 
like proprietary Hardware. 

596
00:32:09,700 --> 00:32:13,500
So, the idea that, you know, I 
think I may be a good example 

597
00:32:13,500 --> 00:32:15,900
for General compute. 
I've listened to that are the 

598
00:32:15,900 --> 00:32:20,000
epicenter session with ICP and 
Affinity where they have their 

599
00:32:20,000 --> 00:32:23,100
own boxes, basically, and 
they're sold by them. 

600
00:32:23,100 --> 00:32:26,100
That's actually very attractive 
to us. 

601
00:32:26,100 --> 00:32:29,400
The idea that you can 
essentially ship your own 

602
00:32:29,400 --> 00:32:32,400
Hardware because then you all 
the kind of issues you have with

603
00:32:32,400 --> 00:32:36,200
sort of, you know, rerunning 
proofs in a way, which is 

604
00:32:36,200 --> 00:32:39,400
deterministic or hashing Etc. 
Lots of our gets solved, but 

605
00:32:39,400 --> 00:32:41,600
crucially, it creates us a choke
point of centralization. 

606
00:32:41,600 --> 00:32:46,700
So one of the kind of rabbit 
holes, we've seen some other 

607
00:32:46,700 --> 00:32:50,700
kind of Computer articles, go 
down is the rely on certain kind

608
00:32:50,700 --> 00:32:55,100
of how would you say kind of 
secure enclaves? 

609
00:32:55,700 --> 00:32:58,800
So certain like secure Enclave 
like Intel sgx where they're 

610
00:32:58,800 --> 00:33:02,200
like, you know, we run we can 
run you computations for you in 

611
00:33:02,200 --> 00:33:05,300
a way which are kind of private.
But you only have to use the 

612
00:33:05,300 --> 00:33:08,300
specific chat which is 
manufactured by the specific 

613
00:33:08,300 --> 00:33:11,000
company. 
And you know, it's only rentable

614
00:33:11,000 --> 00:33:14,800
and these specific services and 
it just it doesn't hold true to 

615
00:33:14,800 --> 00:33:17,000
the decentralized, be false in 
our opinion. 

616
00:33:17,100 --> 00:33:20,600
It also doesn't scale. 
Well, currently at least. 

617
00:33:20,900 --> 00:33:23,600
Yeah. 
I mean, if you look at what 

618
00:33:23,600 --> 00:33:27,200
appears to me about Jensen's 
offering most aesthetic, kind of

619
00:33:27,400 --> 00:33:32,400
it can use resources that are 
currently lying fallow. 

620
00:33:32,400 --> 00:33:35,500
And I mean, this would not be 
the case of you actually had to 

621
00:33:35,500 --> 00:33:39,300
buy a dedicated piece of 
Hardware to kind of partake in 

622
00:33:39,300 --> 00:33:42,400
the network know. 
Yeah, exactly. 

623
00:33:43,200 --> 00:33:46,800
I think like I said it's really 
attractive to go that route from

624
00:33:46,800 --> 00:33:48,800
a A perspective because it's so 
easy. 

625
00:33:48,800 --> 00:33:51,900
But I think it intersects with 
one of the biggest things that 

626
00:33:51,900 --> 00:33:54,400
we think about when designing 
our verification system, which 

627
00:33:54,400 --> 00:33:58,100
is how what assumptions are we 
making, and how we can straining

628
00:33:58,100 --> 00:34:00,800
the system, because essentially 
we have to make some assumptions

629
00:34:00,800 --> 00:34:03,500
and we have to put some 
constraints in, but a constraint

630
00:34:03,500 --> 00:34:06,100
like that to us is massive. 
It's huge. 

631
00:34:06,100 --> 00:34:07,600
We don't want to do that. 
Unless we. 

632
00:34:07,600 --> 00:34:10,500
Absolutely, absolutely have to 
there's other things that we can

633
00:34:10,500 --> 00:34:13,800
do, we can sort of narrow, the 
space of devices in in a 

634
00:34:13,808 --> 00:34:17,500
temporary sense or in a 
permanent sense, we Can look at 

635
00:34:17,500 --> 00:34:20,500
certain manufacturers, we can 
look at certain libraries that 

636
00:34:20,500 --> 00:34:22,800
are provide determinism and 
things like that. 

637
00:34:22,808 --> 00:34:25,800
But every time that we make any 
decision like that, we make it 

638
00:34:25,800 --> 00:34:29,300
very deliberately and I think 
it's quite easy to jump over 

639
00:34:29,300 --> 00:34:32,199
those in the rush to ship 
something, but if you're going 

640
00:34:32,199 --> 00:34:35,300
to build the network that we 
want to build that kind of takes

641
00:34:35,300 --> 00:34:39,100
the entire world and turns it 
into an AI supercomputer, you 

642
00:34:39,100 --> 00:34:41,800
have to be very deliberate about
that and maybe it takes slightly

643
00:34:41,800 --> 00:34:44,900
longer, but you've made it 
generalizable and that's the 

644
00:34:44,900 --> 00:34:48,699
kind of Step change essentially.
Most 0 or 1, if you make those 

645
00:34:48,699 --> 00:34:51,100
assumptions, you won't reach 
that kind of end State. 

646
00:34:52,300 --> 00:34:55,199
It's I think it's sort of fits 
on three axes, this product 

647
00:34:55,199 --> 00:34:57,300
assumptions. 
There's research assumptions in 

648
00:34:57,300 --> 00:34:59,800
this technical assumptions and 
essentially, you have to balance

649
00:34:59,800 --> 00:35:03,400
all of those things, which makes
it, I think uniquely tricky. 

650
00:35:04,100 --> 00:35:08,100
You have to have kind of voices,
of each of them, equally kind of

651
00:35:08,300 --> 00:35:11,300
valid in the company and that's 
something that we've focused on 

652
00:35:11,300 --> 00:35:14,200
quite strongly with, with hiring
and things like that. 

653
00:35:14,400 --> 00:35:16,900
Just making sure that we don't 
accidentally overweight a sir, 

654
00:35:17,100 --> 00:35:20,000
Certain kind of area. 
I think the some protocols we've

655
00:35:20,000 --> 00:35:22,100
looked at before, who've fallen 
into traps. 

656
00:35:22,500 --> 00:35:25,900
There's some traps with research
where you can go down a, let's 

657
00:35:25,900 --> 00:35:28,500
make the most formally 
verifiable system, we possibly 

658
00:35:28,500 --> 00:35:29,900
can. 
And then you never ship 

659
00:35:29,900 --> 00:35:31,700
anything. 
And then you can go the other 

660
00:35:31,700 --> 00:35:34,700
route where you make their kind 
of flashiest thing that an end 

661
00:35:34,700 --> 00:35:37,000
user will like you ship 
something really quickly. 

662
00:35:37,200 --> 00:35:39,800
And in previous startup terms 
that will be fantastic ship. 

663
00:35:39,800 --> 00:35:43,500
It it breaks build it again in 
the web three world not quite as

664
00:35:43,500 --> 00:35:46,400
good as it breaks isn't just a 
little thing anymore. 

665
00:35:46,400 --> 00:35:50,100
It's a Problem. 
So I think it's sort of a unique

666
00:35:50,100 --> 00:35:52,300
area web 3 where you have to 
walk this. 

667
00:35:52,600 --> 00:35:55,300
I think of it like a ridge where
there's really attractive 

668
00:35:55,300 --> 00:35:58,200
looking paths that go down 
either side but they're not 

669
00:35:58,200 --> 00:36:01,300
attractive, they quite quickly, 
drop off the cliff and we're 

670
00:36:01,300 --> 00:36:03,600
being very careful to stay on 
that Ridge. 

671
00:36:05,200 --> 00:36:10,600
Quit before we dive into the ins
and outs of the protocol itself.

672
00:36:10,600 --> 00:36:16,300
So gentle its own layer, one 
block chain in principle. 

673
00:36:16,300 --> 00:36:21,000
It could have also been built as
a Dap on another chain. 

674
00:36:21,400 --> 00:36:24,000
Why did you go the layer one 
road? 

675
00:36:25,300 --> 00:36:28,200
Yeah, it was it was a big 
question for is at the start. 

676
00:36:28,200 --> 00:36:31,900
I think, like we said the the 
sort of blockchain world for us 

677
00:36:31,900 --> 00:36:34,900
was all about tack. 
So when we, when we entered it, 

678
00:36:34,900 --> 00:36:37,600
we were quite sort of deliberate
about it. 

679
00:36:37,600 --> 00:36:39,800
We looked at all of the 
potential ways. 

680
00:36:39,800 --> 00:36:42,400
We could build it. 
We made a massive lift list of 

681
00:36:42,400 --> 00:36:45,300
pros and cons and we kind of 
navigated through figuring out 

682
00:36:45,300 --> 00:36:48,000
what the again, I guess, the 
constraints and assumptions were

683
00:36:48,000 --> 00:36:51,400
for each each one, we quite 
quickly moved from layer to 

684
00:36:51,408 --> 00:36:54,900
layer 1 because we wanted the 
freedom to kind of change 

685
00:36:54,900 --> 00:36:56,800
things. 
Is on the layer, one side 

686
00:36:56,800 --> 00:36:58,800
essentially, the consensus 
mechanism. 

687
00:36:59,500 --> 00:37:02,700
We didn't want to be constrained
by certain smart contract 

688
00:37:02,700 --> 00:37:04,100
system. 
We wanted to be able to do as 

689
00:37:04,100 --> 00:37:06,500
much as we possibly could, 
because we knew this was going 

690
00:37:06,500 --> 00:37:08,800
to be a big sort of open-ended 
problem. 

691
00:37:09,700 --> 00:37:12,700
Essentially being a layer. 
One allows us to do a lot more 

692
00:37:12,700 --> 00:37:16,000
work on the Node side than we 
would otherwise be able to do. 

693
00:37:16,100 --> 00:37:18,700
I think if we'd built in the VM,
which you could absolutely do, 

694
00:37:18,700 --> 00:37:20,300
you could build what we're 
talking about there. 

695
00:37:20,500 --> 00:37:23,600
You'd be very, very constrained 
by what you can do in solidity 

696
00:37:23,600 --> 00:37:27,700
essentially whereas Building in 
Rust for us, we can do certain 

697
00:37:27,700 --> 00:37:30,400
things we can fall out and do 
some machine learning processes.

698
00:37:30,400 --> 00:37:32,600
Maybe we can do some tensor 
processing, things like that, 

699
00:37:32,600 --> 00:37:34,900
that just wouldn't be available 
to us within the evm. 

700
00:37:35,200 --> 00:37:38,300
It was a, I guess in a nutshell 
it was a future proofing thing 

701
00:37:38,300 --> 00:37:39,600
for us. 
We don't want to constrain 

702
00:37:39,600 --> 00:37:42,900
ourselves early when we don't 
understand fully why we're 

703
00:37:42,900 --> 00:37:44,900
making those constraints. 
So we kept it as open as 

704
00:37:44,908 --> 00:37:47,800
possible. 
And fundamental we also believe 

705
00:37:47,800 --> 00:37:50,900
in a multi-chain future. 
We think that the future is true

706
00:37:50,900 --> 00:37:53,500
multi-chain. 
It's not sort of ecosystems full

707
00:37:53,500 --> 00:37:56,400
of chains. 
Its Will change interacting with

708
00:37:56,400 --> 00:37:59,300
each other with a kind of 
generally, agreed messaging 

709
00:37:59,300 --> 00:38:01,300
protocol. 
I think we've seen some 

710
00:38:01,300 --> 00:38:04,700
movements through the ecosystem 
having their own messaging 

711
00:38:04,800 --> 00:38:07,400
message passing and now I'm 
moving back into kind of General

712
00:38:07,400 --> 00:38:09,800
message passing. 
And I think realistically, we're

713
00:38:09,800 --> 00:38:11,800
seeing the multi chain future, 
sort of play out. 

714
00:38:11,800 --> 00:38:14,800
So, we're quite pleased with 
that kind of that so far. 

715
00:38:15,800 --> 00:38:21,600
So you're looking at building 
this, as a parent chain, why 

716
00:38:21,900 --> 00:38:24,100
this on substrate in the 
polkadot ecosystem? 

717
00:38:25,300 --> 00:38:28,000
So we're not, we're not fully. 
Certain weather will be a pair 

718
00:38:28,000 --> 00:38:32,100
of chain or not. 
Yet, the substrate decision was 

719
00:38:32,100 --> 00:38:35,300
essentially the technology. 
So when we looked at everything,

720
00:38:35,300 --> 00:38:38,300
we looked at the sort of 
Frameworks that we could use in 

721
00:38:38,300 --> 00:38:41,100
the libraries that existed from 
Attack perspective just what was

722
00:38:41,100 --> 00:38:43,300
nice. 
What had sort of the best 

723
00:38:43,300 --> 00:38:45,900
technology built-in and 
substrate came out on top for 

724
00:38:45,900 --> 00:38:48,500
us. 
We weren't blockchain people. 

725
00:38:48,500 --> 00:38:51,500
We were machine learning people.
We came in knowing that we 

726
00:38:51,500 --> 00:38:54,100
wanted to like stand on the 
shoulders of giants, if you 

727
00:38:54,100 --> 00:38:55,500
will. 
We don't want to rebuild 

728
00:38:55,500 --> 00:38:57,800
consensus from scratch. 
We want to use whatever the best

729
00:38:57,800 --> 00:39:00,800
one is and then carry on with 
building the machine learning 

730
00:39:00,800 --> 00:39:03,600
stuff that we're focused on and 
substrate provided. 

731
00:39:03,600 --> 00:39:07,200
That tours, as a way, to very 
quickly, iterate build up the 

732
00:39:07,200 --> 00:39:11,300
chain and then get on with the 
off chain stuff with enough, 

733
00:39:11,300 --> 00:39:13,300
flexibility to change it when we
need to. 

734
00:39:13,300 --> 00:39:16,400
So the kind of frame sub system 
allows us to quickly get 

735
00:39:16,400 --> 00:39:18,900
something running but then if we
need to step in and completely 

736
00:39:18,900 --> 00:39:22,800
change it, which is really 
attractive is written in Rust 

737
00:39:22,800 --> 00:39:27,100
were fans of rust as a language.
Edge it just kind of made sense 

738
00:39:27,100 --> 00:39:28,500
from that perspective. 
It's interesting. 

739
00:39:28,500 --> 00:39:32,000
This was a year and a half ago 
and the kind of to that came out

740
00:39:32,000 --> 00:39:35,900
on top or Cosmos and substrate 
and essentially substrate one 

741
00:39:35,900 --> 00:39:38,000
because of the tack and the kind
of nice libraries in the 

742
00:39:38,000 --> 00:39:39,700
developer tooling and things 
like that. 

743
00:39:40,700 --> 00:39:43,200
But yeah, in the power of chain 
decision is one that we 

744
00:39:43,300 --> 00:39:46,100
essentially will make later as a
bit of a cheat answer. 

745
00:39:46,300 --> 00:39:49,100
We can be at para chain, we 
could not be a pair of chain, we

746
00:39:49,100 --> 00:39:54,300
don't need to decide right now. 
So essentially we don't if the 

747
00:39:54,400 --> 00:39:57,000
System starts to fill up with 
things that we can interact 

748
00:39:57,000 --> 00:39:58,300
with. 
So if there's like storage 

749
00:39:58,300 --> 00:40:01,300
layers in there, if there's sort
of, sovereign data layers, and 

750
00:40:01,300 --> 00:40:04,000
things like that, that we would 
want close ties with, then maybe

751
00:40:04,000 --> 00:40:06,200
it makes sense. 
If they Exist Elsewhere, then 

752
00:40:06,200 --> 00:40:09,900
maybe it makes sense to kind of 
Bolt IBC, on and exist in the 

753
00:40:09,900 --> 00:40:12,200
wider world. 
But yeah, yeah. 

754
00:40:12,200 --> 00:40:15,900
So you guys should look at 
Solutions like TC also. 

755
00:40:16,100 --> 00:40:22,800
So things have kind of allow you
to kind of have a legacy 

756
00:40:24,500 --> 00:40:28,200
Operating system, that kind of 
hooks into a blockchain for 

757
00:40:28,200 --> 00:40:32,100
provable compute. 
It's a super interesting. 

758
00:40:32,900 --> 00:40:34,500
As you can say, sounds 
interesting, I've not come 

759
00:40:34,500 --> 00:40:37,800
across it before, but yeah, 
we'll definitely check it out as

760
00:40:37,800 --> 00:40:41,900
you have an insulator. 
So let's dive into the protocol.

761
00:40:41,900 --> 00:40:45,300
So there's a couple of 
participants in the Jensen 

762
00:40:45,700 --> 00:40:48,500
economy. 
There are submitted service 

763
00:40:48,500 --> 00:40:51,900
verify as and whistleblowers. 
The submitters are the people 

764
00:40:52,100 --> 00:40:56,600
who actually wants work done. 
So let's say let's start in the 

765
00:40:56,600 --> 00:40:58,800
beginning. 
Let's say I'm a submitter. 

766
00:40:59,400 --> 00:41:03,800
What kind of AI problems can I 
submit a my constrained in any 

767
00:41:03,800 --> 00:41:07,800
way? 
Yes, currently you constrained 

768
00:41:08,200 --> 00:41:13,100
by your AI problem has to use 
gradient, based optimization at 

769
00:41:13,100 --> 00:41:16,500
some point in the in the 
computational process. 

770
00:41:16,500 --> 00:41:21,400
Basically, we use portions of 
the gradient calculations as 

771
00:41:21,400 --> 00:41:24,600
part of our proof system. 
That's not necessarily set in 

772
00:41:24,600 --> 00:41:26,000
stone. 
I think as a hurry mentioned 

773
00:41:26,000 --> 00:41:28,500
earlier, we've got our light 
paper, which is public. 

774
00:41:28,600 --> 00:41:31,500
We're iterating on that 
internally and there are lots of

775
00:41:31,500 --> 00:41:34,600
kind of things in play, 
essentially, but right now, it's

776
00:41:34,800 --> 00:41:36,100
great. 
Nice optimization. 

777
00:41:36,100 --> 00:41:38,700
We use the signals from that as 
part of the verification 

778
00:41:38,700 --> 00:41:42,100
mechanism, what does gradient 
based optimization means so to 

779
00:41:42,100 --> 00:41:45,100
me as an AI Noob. 
So how would I know whether a 

780
00:41:45,100 --> 00:41:49,100
problem falls into that category
or not sure. 

781
00:41:49,500 --> 00:41:52,000
Yeah. 
So I guess fundamentally, if we 

782
00:41:52,000 --> 00:41:56,800
think about a neural network, it
is a big set of players that 

783
00:41:56,800 --> 00:41:58,600
have parameters in them. 
And those parameters are 

784
00:41:58,600 --> 00:42:02,600
essentially just real numbers 
that could be millions billions 

785
00:42:02,600 --> 00:42:04,500
now trillions of those numbers 
in there. 

786
00:42:05,100 --> 00:42:08,100
But fundamentally, they are the 
kind of deciding factor in the 

787
00:42:08,100 --> 00:42:11,200
output of the network and the 
training of the network is 

788
00:42:11,200 --> 00:42:14,300
setting, those two realistic 
values that allow data to go 

789
00:42:14,300 --> 00:42:17,400
through and Trigger the kind of 
outputs that you like at the end

790
00:42:17,400 --> 00:42:19,400
of the network. 
So you go through lots of sort 

791
00:42:19,400 --> 00:42:22,000
of matrices, layers of these 
real numbers. 

792
00:42:22,000 --> 00:42:26,700
It sort of it changes the, the 
current input as its work going 

793
00:42:26,700 --> 00:42:28,000
through. 
And then you get the output that

794
00:42:28,000 --> 00:42:29,700
you want. 
By all of those changes that 

795
00:42:29,700 --> 00:42:32,600
have happened, you need to 
update those numbers to reflect 

796
00:42:32,600 --> 00:42:36,700
the output that you For a 
certain input. 

797
00:42:36,700 --> 00:42:39,300
And previously way way back in 
the day that would be done 

798
00:42:39,300 --> 00:42:41,300
manually. 
So you maybe not with a neural 

799
00:42:41,300 --> 00:42:44,600
network, but with kind of 
certain systems you would set 

800
00:42:44,600 --> 00:42:47,400
those using expert knowledge and
then you would know that when an

801
00:42:47,400 --> 00:42:49,500
input goes through, you would 
get the right output. 

802
00:42:49,800 --> 00:42:52,500
There's also different ways of 
setting them programmatically. 

803
00:42:52,500 --> 00:42:56,400
So you could imagine a super 
sort of naive way of just 

804
00:42:56,400 --> 00:42:59,900
randomly setting, all of the 
parameters running, a sample 

805
00:42:59,900 --> 00:43:03,600
through checking, how far away 
from the realistic sample, it is

806
00:43:03,800 --> 00:43:06,700
and then just doing random ones.
Again and then doing a random 

807
00:43:06,700 --> 00:43:10,100
search, essentially until you 
make a smaller error value at 

808
00:43:10,100 --> 00:43:12,700
the end and then you just keep 
decreasing that error value. 

809
00:43:13,300 --> 00:43:15,300
You can do other strategies 
where you do sort of more 

810
00:43:15,300 --> 00:43:17,800
targeted updates and there's 
lots of ways that you can do 

811
00:43:17,800 --> 00:43:19,700
that. 
Gradient based optimization 

812
00:43:19,700 --> 00:43:22,700
talks about essentially, what 
was the big change for neural 

813
00:43:22,700 --> 00:43:25,600
networks and deep learning which
was showing that you could 

814
00:43:26,100 --> 00:43:31,300
essentially use the gradient or 
differentiating, the parameters 

815
00:43:31,300 --> 00:43:34,100
of the layer with respect to the
arrow as you go through the 

816
00:43:34,100 --> 00:43:36,700
network and you can use the 
Chain rule to apply that all the

817
00:43:36,700 --> 00:43:39,000
way back through the 
hierarchical network that hurry 

818
00:43:39,000 --> 00:43:41,100
described. 
Essentially, in that way, you 

819
00:43:41,100 --> 00:43:46,500
get the position on the hill of 
loss, if that makes sense. 

820
00:43:46,600 --> 00:43:51,000
So, if you modeled the loss as a
in like euclidean space, you 

821
00:43:51,000 --> 00:43:55,000
would see it as this kind of 
really bumpy area, where 

822
00:43:55,200 --> 00:43:57,900
somewhere there's a big dip and 
at the bottom is where the loss 

823
00:43:57,900 --> 00:44:00,400
is really, really small. 
And you're trying to find that 

824
00:44:00,400 --> 00:44:03,500
dip getting the gradients for 
each layer of essentially shows 

825
00:44:03,500 --> 00:44:06,400
you for that layer. 
Where you It on that surface and

826
00:44:06,400 --> 00:44:09,800
what direction you should go. 
And so you use the gradient to 

827
00:44:09,800 --> 00:44:12,700
say, hey, we've got a massive 
like Drop here. 

828
00:44:12,700 --> 00:44:14,900
Let's go down it. 
So the direction that we want to

829
00:44:14,900 --> 00:44:17,900
update the parameters in is this
way and we want to update them 

830
00:44:17,900 --> 00:44:22,300
with this sort of size of Step 
because this is really steep or 

831
00:44:22,300 --> 00:44:24,300
it's not steep. 
So we want to make a big jump or

832
00:44:24,300 --> 00:44:26,500
a smaller jump and essentially 
that's it. 

833
00:44:26,500 --> 00:44:30,000
You're just navigating this huge
bumpy surface looking for a big 

834
00:44:30,000 --> 00:44:33,600
dip and the gradients give you 
sort of a position on that 

835
00:44:33,600 --> 00:44:36,200
surface so that you know which 
direction To go in and it was a 

836
00:44:36,200 --> 00:44:40,700
huge leap because that signal 
that direction is kind of really

837
00:44:40,700 --> 00:44:44,500
clearly useful rather than just 
taking random leaps, all of the 

838
00:44:44,500 --> 00:44:46,500
space and figuring out. 
Hey, I'm on the top of a hill 

839
00:44:46,500 --> 00:44:49,400
now, or hey, I'm at the bottom 
of a trench, you know where you 

840
00:44:49,400 --> 00:44:51,600
are, you know, that you're on 
the side of a trench or that 

841
00:44:51,600 --> 00:44:53,800
you're on this weird flat bit 
and you need to make a big jump 

842
00:44:53,800 --> 00:44:56,000
to try and get out of the flat 
bill, something like that. 

843
00:44:56,200 --> 00:44:59,000
How do you know this? 
How do you know there's only one

844
00:44:59,000 --> 00:45:00,700
Trend or how do you make sure 
you're on the right? 

845
00:45:00,700 --> 00:45:03,400
Train ch right because basically
if there's lots of trenches you 

846
00:45:03,400 --> 00:45:05,700
kind of you want to end up in 
the Deep First one. 

847
00:45:05,800 --> 00:45:07,700
You don't want to get stuck on a
mole here, right? 

848
00:45:07,700 --> 00:45:09,200
Yeah, I mean you want to go to 
Mount Everest. 

849
00:45:09,200 --> 00:45:12,300
So basically how do you make 
sure that basic? 

850
00:45:12,300 --> 00:45:16,000
How do you know how how you how 
low you can go? 

851
00:45:16,000 --> 00:45:20,400
I how how you can go with your 
mother very good question. 

852
00:45:20,400 --> 00:45:23,400
That's one of the big, big 
problems in deep learning 

853
00:45:23,400 --> 00:45:26,000
itself. 
Essentially little lots of 

854
00:45:26,000 --> 00:45:30,200
techniques for doing that. 
The very, very simple answer is 

855
00:45:30,200 --> 00:45:32,800
assume it's convex and then you 
don't have to think about the 

856
00:45:32,800 --> 00:45:34,600
being any other. 
Yeah. 

857
00:45:35,500 --> 00:45:38,000
Obviously in the real world, it 
doesn't work like that. 

858
00:45:38,800 --> 00:45:41,600
Essentially there's lots of sort
of regularization techniques 

859
00:45:41,600 --> 00:45:44,300
that happen in deep learning 
training and make it a really 

860
00:45:44,300 --> 00:45:46,500
complex thing. 
It makes it more of an art than 

861
00:45:46,500 --> 00:45:49,400
a science because a lot of 
people have their sort of little

862
00:45:49,400 --> 00:45:53,500
tricks that they do, there's 
things like within learning rate

863
00:45:53,500 --> 00:45:55,500
schedules. 
So you'll use a learning rate 

864
00:45:55,500 --> 00:45:58,300
set, the magnitude of the jump 
that you'll take in that 

865
00:45:58,300 --> 00:46:01,900
gradient space but your you can 
use certain schedules to sort of

866
00:46:01,908 --> 00:46:04,800
Decay The Learning rate, make it
smaller over time, which means 

867
00:46:05,100 --> 00:46:07,700
smaller and smaller jumps. 
So you don't accidentally jump 

868
00:46:07,700 --> 00:46:11,100
over kind of trench. 
But in the same case, you can 

869
00:46:11,100 --> 00:46:14,900
suddenly randomly introduced a 
huge jump, which just allows you

870
00:46:14,900 --> 00:46:16,900
to know that. 
Maybe if I am in a global 

871
00:46:16,900 --> 00:46:19,400
minimum, maybe I'm at, if I'm in
a tiny trench here and there's a

872
00:46:19,408 --> 00:46:22,600
massive one over here, I'll just
do a huge jump, I don't know 

873
00:46:22,600 --> 00:46:25,900
where I'll end up, but it 
should, it could be better if 

874
00:46:25,900 --> 00:46:28,100
not, I'll probably roll back to 
where I was before. 

875
00:46:28,400 --> 00:46:31,400
So there's lots of techniques 
like that, that sort of more 

876
00:46:31,400 --> 00:46:34,000
trial and error than they are 
sort of deliberate. 

877
00:46:34,600 --> 00:46:37,800
But Like I said before, they're 
becoming more deliberate over 

878
00:46:37,800 --> 00:46:39,000
time. 
So now that people have 

879
00:46:39,000 --> 00:46:43,000
introduced these regularization 
techniques, Dropout, norms and 

880
00:46:43,000 --> 00:46:44,700
things like that. 
Now, people are looking back at 

881
00:46:44,700 --> 00:46:47,300
them and saying, hey, did this 
work for the right reason? 

882
00:46:47,300 --> 00:46:51,100
Or was it just some weird random
Quirk of the model architecture 

883
00:46:51,100 --> 00:46:53,100
that made it work here? 
And can we kind of figure out 

884
00:46:53,100 --> 00:46:56,900
exactly why it works? 
But yeah, it is, it comes down 

885
00:46:56,900 --> 00:46:59,700
to an art more than a science, 
to be honest, it can be very 

886
00:46:59,700 --> 00:47:05,900
frustrating. 
So now, I understand that 

887
00:47:05,900 --> 00:47:10,100
gradient optimization problems 
are what I should submit in 

888
00:47:10,100 --> 00:47:15,100
terms of, I mean, can you can 
you can you talk about like real

889
00:47:15,100 --> 00:47:18,500
word problems, Robert problems 
and say which ones are gradient 

890
00:47:18,500 --> 00:47:22,900
optimizations and which one, 
which ones aren't just, so I can

891
00:47:22,900 --> 00:47:28,200
get like a feeling for what kind
of problems I should be able to 

892
00:47:28,200 --> 00:47:33,000
submit Yeah, I mean the simplest
way is thinking pretty much 

893
00:47:33,000 --> 00:47:36,100
every neural network is uses 
gradient based optimization. 

894
00:47:36,600 --> 00:47:39,300
There are other problems that 
use it as well, but within 

895
00:47:39,300 --> 00:47:43,400
neural networks, all of the kind
of big steps that we've seen all

896
00:47:43,400 --> 00:47:46,000
the big changes have been neural
networks recently, so it's a 

897
00:47:46,008 --> 00:47:49,400
logical place to focus for us 
whilst also allowing this big 

898
00:47:49,400 --> 00:47:52,000
space of other places. 
So any optimization problem you 

899
00:47:52,000 --> 00:47:54,800
could theoretically, as long as 
it's differentiable and we can 

900
00:47:54,800 --> 00:47:58,100
use the chain rule to flow back.
It could use gradient based 

901
00:47:58,100 --> 00:48:01,600
optimization and Use other 
optimization techniques with the

902
00:48:01,600 --> 00:48:04,000
grading as a signal in there. 
As long as you're calculating a 

903
00:48:04,000 --> 00:48:08,200
gradient, we can use, it's 
useful, but yeah, fundamentally 

904
00:48:08,200 --> 00:48:12,600
it's all neural networks. 
Every maybe two three years, 

905
00:48:12,600 --> 00:48:15,400
somebody comes out with a paper 
that says, hey, we're training 

906
00:48:15,400 --> 00:48:18,000
neural networks with 
evolutionary optimization that 

907
00:48:18,000 --> 00:48:19,500
doesn't use gradients, and it's 
better. 

908
00:48:19,500 --> 00:48:22,400
It's never better. 
It's better in a really 

909
00:48:22,400 --> 00:48:26,700
constrained system and it never 
takes, hold not to say. 

910
00:48:26,700 --> 00:48:30,000
It never will do. 
But so far gradients of Managed 

911
00:48:30,000 --> 00:48:32,800
to stay pretty kind of solidly 
at the top. 

912
00:48:33,100 --> 00:48:34,400
Okay. 
Then I have turn the question 

913
00:48:34,400 --> 00:48:37,000
around. 
What I've what kind of problems 

914
00:48:37,000 --> 00:48:41,400
can't I submit? 
So what what what problems are 

915
00:48:41,400 --> 00:48:44,000
not solvable or not? 
Where's Ava were with neural 

916
00:48:44,000 --> 00:48:45,900
networks? 
Hmm. 

917
00:48:45,900 --> 00:48:49,200
Good question, expanding at the 
question your own networks in 

918
00:48:49,200 --> 00:48:51,000
general neural. 
Networks are generally quite 

919
00:48:51,000 --> 00:48:55,700
data hungry algorithms. 
So if you have a have a problem 

920
00:48:55,800 --> 00:49:00,400
with very low data volume, Good 
example. 

921
00:49:00,400 --> 00:49:04,100
That would be like, I might be 
wrong about this, but some of 

922
00:49:04,100 --> 00:49:06,700
the kind of toy examples which 
are used to teach people to do 

923
00:49:06,700 --> 00:49:08,700
machine learning like the iris 
dataset and stuff. 

924
00:49:08,700 --> 00:49:10,600
Are you have like a very like 
you could actually have a 

925
00:49:10,600 --> 00:49:13,600
spreadsheet like a hundred rolls
and maybe if I think like seven 

926
00:49:13,600 --> 00:49:16,500
or eight features, I don't think
that intuitively. 

927
00:49:16,500 --> 00:49:19,200
They're like, well, suited to 
neural networks are typically 

928
00:49:19,200 --> 00:49:21,000
better handled by like 
statistical machine learning 

929
00:49:21,000 --> 00:49:25,300
techniques. 
So I think data volumes, one of 

930
00:49:25,308 --> 00:49:28,600
them. 
There's also just certain 

931
00:49:29,600 --> 00:49:31,900
Certain types of neural networks
which is very large. 

932
00:49:31,900 --> 00:49:35,500
So like fitting them on edge 
devices can be a challenge, but 

933
00:49:35,500 --> 00:49:37,900
in terms of like the actual, I 
guess, when you say problem, if 

934
00:49:37,900 --> 00:49:42,200
you think about the kind of, 
what's the type of thing, you're

935
00:49:42,200 --> 00:49:46,100
trying to predict in the world, 
there isn't something which 

936
00:49:46,100 --> 00:49:49,600
immediately comes to mind, that 
neural networks aren't are like 

937
00:49:49,600 --> 00:49:53,000
explicitly currently, and always
will be bad at. 

938
00:49:53,400 --> 00:49:55,500
I don't know if you've had any 
tuition there. 

939
00:49:55,500 --> 00:49:57,400
Been. 
I guess you can think of a 

940
00:49:57,400 --> 00:50:00,100
neural network as a universal. 
Action approximator. 

941
00:50:00,100 --> 00:50:03,700
So theoretically, it can do all 
of the things that you would do 

942
00:50:03,700 --> 00:50:05,900
with other methods. 
I think, like Harry said, the 

943
00:50:05,900 --> 00:50:08,200
reason you would not use a 
neural network would typically 

944
00:50:08,200 --> 00:50:11,000
be down to data volumes where 
with a statistical machine 

945
00:50:11,000 --> 00:50:14,400
learning mechanism you method, 
you could get a better result 

946
00:50:14,400 --> 00:50:17,100
essentially and then you would 
you wouldn't train that using 

947
00:50:17,900 --> 00:50:21,700
gradient based optimization. 
But fundamentally you could do 

948
00:50:21,700 --> 00:50:23,600
it within your own network if 
you wanted to. 

949
00:50:23,600 --> 00:50:27,400
It just might be a bit worse. 
Okay, good. 

950
00:50:27,400 --> 00:50:30,500
So I understand that. 
I can submit almost any 

951
00:50:30,500 --> 00:50:33,700
questions. 
So basically say do I why I 

952
00:50:33,700 --> 00:50:38,600
asked for an entire program or I
mean do I ask for like a dolly 

953
00:50:38,600 --> 00:50:43,400
kind of output can I say like I 
want a picture of accountants in

954
00:50:43,400 --> 00:50:46,400
hot air balloons over a 
waterfall and there should also 

955
00:50:46,400 --> 00:50:50,100
be a rainbow with scorpions on 
it and it'll do that for me. 

956
00:50:50,300 --> 00:50:54,800
Oh can I ask I'm betting? 
I'm reading this car and I need 

957
00:50:54,800 --> 00:50:57,900
an AI to drive that car. 
Can you deliver that AI? 

958
00:50:57,900 --> 00:51:01,800
Is that kind of the Both within 
the scope or do one of those 

959
00:51:01,800 --> 00:51:07,500
fall out of scope. 
So I guess it's more like you'd 

960
00:51:07,500 --> 00:51:09,800
use Jensen to train, the model 
itself. 

961
00:51:09,800 --> 00:51:11,900
So what you would do would be 
you'd think I want those things,

962
00:51:11,900 --> 00:51:14,700
I want to be able to create my 
scorpion rainbow kind of image 

963
00:51:14,700 --> 00:51:18,400
generator from the text prompt 
scorpion rainbow and which I 

964
00:51:18,408 --> 00:51:23,500
love and you'd you'd build a 
model which, you know, receives 

965
00:51:23,500 --> 00:51:26,700
a text prompt and then converts 
that text problems into into 

966
00:51:26,700 --> 00:51:29,100
images and then you would have 
the training data. 

967
00:51:29,200 --> 00:51:33,700
Which facilitated the kind of 
learning of that model and then 

968
00:51:33,700 --> 00:51:37,700
to magenta Network, you would 
submit model data and then some 

969
00:51:37,700 --> 00:51:40,600
hyperparameters which determine 
you know, like badminton League,

970
00:51:40,600 --> 00:51:42,300
The Learning rate, schedule 
things like that. 

971
00:51:43,400 --> 00:51:45,100
How do you how long you wanted 
to train for? 

972
00:51:45,100 --> 00:51:48,100
And then you're kind of the 
artifact you receive. 

973
00:51:48,100 --> 00:51:51,100
From that training process, the 
kind of product you get, is the 

974
00:51:51,100 --> 00:51:54,000
train model and then that model 
can, then you can then post that

975
00:51:54,000 --> 00:51:56,300
and then you can submit scorpion
rainbow. 

976
00:51:56,500 --> 00:51:59,000
Yeah, how do I decide, which 
untrained motto. 

977
00:51:59,100 --> 00:52:02,700
How to use? 
That's a brilliant question. 

978
00:52:02,700 --> 00:52:06,400
I think there's kind of two ways
of thinking about it, so it just

979
00:52:06,400 --> 00:52:10,100
kind of emerging and highly kind
of popular concept around 

980
00:52:10,100 --> 00:52:12,800
foundation models currently, 
which is, you know, you get like

981
00:52:12,800 --> 00:52:17,100
a big company like open the eye 
or something like mid-journey or

982
00:52:17,107 --> 00:52:19,500
something, and they build the 
base model. 

983
00:52:19,600 --> 00:52:22,400
And then you take the base model
with you, your training data, 

984
00:52:22,800 --> 00:52:24,600
which might have lots of 
rainbows scorpions in it and 

985
00:52:24,600 --> 00:52:28,800
then you train the model on that
and then the output of that. 

986
00:52:28,800 --> 00:52:30,500
And then I'm always very good 
at, you know, approximately 

987
00:52:30,600 --> 00:52:32,800
Amazing that those outputs. 
That's kind of option. 

988
00:52:32,800 --> 00:52:35,600
One and that's the most common 
for people who are quite compute

989
00:52:35,600 --> 00:52:38,700
restricted which is the kind of 
theme in the industry just now 

990
00:52:39,000 --> 00:52:41,800
second option is you build the 
model but they were be whatever 

991
00:52:41,800 --> 00:52:45,000
wise build from scratch and I 
don't know if you want to talk 

992
00:52:45,000 --> 00:52:50,300
about that then from our Yeah, I
suppose a lot of our thinking 

993
00:52:50,300 --> 00:52:52,500
comes down to the foundation 
models approach because we think

994
00:52:52,500 --> 00:52:54,300
it's the kind of the future of 
the space. 

995
00:52:55,100 --> 00:53:00,000
My research back in my PhD was 
specifically on essentially Auto

996
00:53:00,000 --> 00:53:03,000
ml techniques. 
So the idea of allowing somebody

997
00:53:03,000 --> 00:53:06,300
to optimize that that model 
structure and find the best 

998
00:53:06,300 --> 00:53:08,900
model structure without 
necessarily being an expert, 

999
00:53:08,900 --> 00:53:11,600
that's another way of doing it. 
And you've seen that sort of 

1000
00:53:11,600 --> 00:53:14,800
happen within like AWS Sage 
maker for example, where they 

1001
00:53:14,800 --> 00:53:17,600
and gcps. 
A few clouds as well, where they

1002
00:53:17,600 --> 00:53:20,900
build in some Auto ml techniques
to say to a developer. 

1003
00:53:21,100 --> 00:53:23,800
You don't need to know the 
specific machine learning 

1004
00:53:23,800 --> 00:53:25,800
architecture because 
essentially, we can just see 

1005
00:53:25,800 --> 00:53:27,500
that as something that's 
trainable as well. 

1006
00:53:27,700 --> 00:53:30,500
And we apply an optimization 
technique on top of that. 

1007
00:53:30,800 --> 00:53:34,000
Jensen is a protocol. 
Can have that if you wanted to, 

1008
00:53:34,000 --> 00:53:36,700
we would see that as something 
that you would build as adapt 

1009
00:53:36,800 --> 00:53:40,300
that would use Jensen. 
And that dap might Implement a 

1010
00:53:40,300 --> 00:53:42,400
evolutionary optimization 
technique or something like 

1011
00:53:42,400 --> 00:53:44,200
that. 
It would submit the sort of 

1012
00:53:44,200 --> 00:53:47,500
individual architectures, it 
wants to train and Best to the 

1013
00:53:47,500 --> 00:53:49,600
Jensen protocol, it would have 
them trained and then it would 

1014
00:53:49,600 --> 00:53:52,600
iterate on the structure and it 
could build up the kind of the 

1015
00:53:52,600 --> 00:53:55,300
model that you want. 
And that's a bit of a theme in 

1016
00:53:55,308 --> 00:53:58,900
the way that we think about 
Jensen as being purely machine, 

1017
00:53:58,900 --> 00:54:01,400
learning compute. 
All of these interesting things 

1018
00:54:01,400 --> 00:54:03,600
that exist around it. 
We would love to see build out 

1019
00:54:03,600 --> 00:54:06,300
as an ecosystem essentially. 
So all of the nice things that 

1020
00:54:06,300 --> 00:54:09,500
you see on stage maker and gcp, 
we see as being additional 

1021
00:54:09,500 --> 00:54:12,100
things on top, I think it could 
be very attractive to build them

1022
00:54:12,100 --> 00:54:16,400
yourselves but ultimately, it's 
a, it's a trap, but yeah. 

1023
00:54:16,500 --> 00:54:21,100
A on the foundation models we've
seen because of the compute 

1024
00:54:21,300 --> 00:54:23,500
problem like how he described, 
we've seen people take 

1025
00:54:23,500 --> 00:54:27,200
Foundation models, from very 
large research papers that have 

1026
00:54:27,200 --> 00:54:30,200
been spent maybe 10 million 
dollars in funding in order to 

1027
00:54:30,200 --> 00:54:33,000
kind of trial out all of these 
different architectures and then

1028
00:54:33,000 --> 00:54:35,900
they publish a new architecture 
and say, hey, this is the best 

1029
00:54:35,900 --> 00:54:39,300
in the world at doing these 
three computer vision tasks and 

1030
00:54:39,300 --> 00:54:42,100
then you take that, you use the 
vast majority of that 

1031
00:54:42,100 --> 00:54:44,700
pre-trained Network. 
That cost Millions to train, you

1032
00:54:44,707 --> 00:54:46,300
would add some layers on the 
end. 

1033
00:54:46,300 --> 00:54:49,300
You Some layers off and then you
train those layers on a smaller 

1034
00:54:49,300 --> 00:54:53,100
set of data and you have a kind 
of usable model for that and it 

1035
00:54:53,100 --> 00:54:55,700
generalized, loads of 
information from the first kind 

1036
00:54:55,700 --> 00:54:58,600
of training it did. 
So you call that pre-training 

1037
00:54:58,600 --> 00:55:01,200
and then you've got fine-tuning 
and that's very kind of classic.

1038
00:55:01,200 --> 00:55:05,400
And in the Deep learning space, 
one of the things that we find 

1039
00:55:05,400 --> 00:55:08,900
particularly difficult with that
is the bias that gets introduced

1040
00:55:08,900 --> 00:55:11,900
in that pre-training. 
So, one organization doing that 

1041
00:55:11,900 --> 00:55:15,400
on a proprietary data set or on 
a dataset that they haven't. 

1042
00:55:16,500 --> 00:55:20,000
Closed features about means that
when somebody else comes to use 

1043
00:55:20,000 --> 00:55:22,500
it, they don't know what's going
on because of those black box 

1044
00:55:23,100 --> 00:55:26,600
issues that you mentioned 
earlier, it can't kind of you 

1045
00:55:26,600 --> 00:55:28,800
can't go back in and say why did
it make this decision? 

1046
00:55:29,400 --> 00:55:32,200
The solution to that, in our 
minds isn't to go fully 

1047
00:55:32,200 --> 00:55:35,000
deterministic and kind of get 
rid of the black box. 

1048
00:55:35,100 --> 00:55:38,000
It's to open it up to everyone. 
And say, hey everyone train, 

1049
00:55:38,000 --> 00:55:41,000
this Foundation model, so design
it together, we train it 

1050
00:55:41,000 --> 00:55:43,800
together on an infrastructure 
that nobody owns and at the end 

1051
00:55:43,800 --> 00:55:46,100
of the day, we have a model that
we can all use. 

1052
00:55:46,100 --> 00:55:49,500
That's kind Global and hasn't 
necessarily been biased by a 

1053
00:55:49,500 --> 00:55:53,500
specific company's cash of data 
that they've kept back and they 

1054
00:55:53,500 --> 00:55:55,600
don't want to tell you what's in
it, and things like that. 

1055
00:55:55,700 --> 00:55:58,500
So, once we've got those kind of
global Foundation models, then 

1056
00:55:58,500 --> 00:56:01,800
anyone can come along and say, 
I'll take I'll find the hash of 

1057
00:56:01,800 --> 00:56:04,100
that model on the Chain. 
I know it's been trained. 

1058
00:56:04,100 --> 00:56:06,800
I'll pick that spot. 
And I'll continue training from 

1059
00:56:06,800 --> 00:56:10,500
there on my data set for my 
problem or task. 

1060
00:56:10,500 --> 00:56:14,200
And then I'll have a model that 
I know at least is as biased as 

1061
00:56:14,200 --> 00:56:17,500
the entire global population. 
Rather than Being as biased as a

1062
00:56:17,508 --> 00:56:21,500
company in California. 
Okay so basically until we have 

1063
00:56:21,500 --> 00:56:24,100
the global Foundation model 
maybe we'll talk. 

1064
00:56:24,100 --> 00:56:27,500
We can talk about how you plan 
on kind of delivering that 

1065
00:56:27,500 --> 00:56:29,300
later. 
But before we have that, I kind 

1066
00:56:29,300 --> 00:56:33,600
of have to decide on one of the 
commercially available ones and 

1067
00:56:34,300 --> 00:56:38,900
I have now submitted my my 
problem who gets to work on it. 

1068
00:56:39,500 --> 00:56:43,400
Do servers need to kind of prove
is some prerequisites. 

1069
00:56:43,900 --> 00:56:48,400
And basically is it one solve a 
Ma, can you Paradise this? 

1070
00:56:49,300 --> 00:56:53,800
That sure I would say it's at 
the task level. 

1071
00:56:53,900 --> 00:56:57,000
It's one solar per task but a 
model can break out into lots of

1072
00:56:57,000 --> 00:56:59,700
different tasks. 
So typically when large language

1073
00:56:59,700 --> 00:57:01,900
models are trained, it's 
interesting. 

1074
00:57:01,900 --> 00:57:05,300
They've kind of been built in a 
way which maxes out the current 

1075
00:57:05,300 --> 00:57:07,200
Hardware at their time of 
creation. 

1076
00:57:07,300 --> 00:57:10,500
So you know they are designed to
fit chunks on you know certain 

1077
00:57:10,500 --> 00:57:13,900
the video processors Etc. 
You imagine a similar thing 

1078
00:57:13,900 --> 00:57:16,400
happening across the network. 
It's complicated. 

1079
00:57:16,400 --> 00:57:18,500
Created by the fact that there's
heterogenous devices on the 

1080
00:57:18,508 --> 00:57:20,400
network. 
But essentially for any given 

1081
00:57:20,400 --> 00:57:24,000
task, you know, supplier of 
compute. 

1082
00:57:24,000 --> 00:57:27,700
So if I have a verifier or 
worker, they have their ability 

1083
00:57:27,700 --> 00:57:32,300
to basically say, I'll take that
from the mempool and then they 

1084
00:57:32,300 --> 00:57:34,700
are randomly chosen from the 
pool. 

1085
00:57:34,700 --> 00:57:37,900
People who say that they'd like 
to take that task. 

1086
00:57:38,500 --> 00:57:43,700
So everyone can do it if the 
model and the data can't fit on 

1087
00:57:43,700 --> 00:57:48,200
your device and you said that 
you know, it And then it follows

1088
00:57:48,200 --> 00:57:51,000
that there's deliver lightly be 
a penalty there because it's 

1089
00:57:51,000 --> 00:57:55,400
kind of clogging up the system 
but essentially if you're if a 

1090
00:57:55,408 --> 00:57:59,000
task and fits on your machine 
then yes your ability to run it 

1091
00:57:59,000 --> 00:58:01,800
is essentially just determined 
by a verifiably random function 

1092
00:58:01,800 --> 00:58:04,900
which is the likes you from a 
sub, a subset of the available 

1093
00:58:04,900 --> 00:58:11,400
minors or workers, I should say,
how do you verify the what kind 

1094
00:58:11,400 --> 00:58:16,000
of capacities the miners have? 
So busy, if I say I have like a 

1095
00:58:16,400 --> 00:58:22,700
16 core GPU and 400 gigabytes of
RAM. 

1096
00:58:23,600 --> 00:58:25,700
How do you verify that? 
Yeah. 

1097
00:58:25,700 --> 00:58:28,800
So it's essentially in the 
verification of the computation,

1098
00:58:29,200 --> 00:58:31,900
they won't be able to do the 
computation if they don't have 

1099
00:58:31,900 --> 00:58:36,000
that compute device essentially 
or that capacity. 

1100
00:58:36,500 --> 00:58:38,900
And when they come to submit 
their proof, when I gets 

1101
00:58:38,900 --> 00:58:41,300
checked, it will be found that 
they couldn't do the 

1102
00:58:41,300 --> 00:58:42,200
computation. 
Okay. 

1103
00:58:42,700 --> 00:58:45,000
There's a little bit of a sort 
of question there and how big 

1104
00:58:45,000 --> 00:58:48,800
you make a task, because You, if
you made a task in an enormous 

1105
00:58:48,800 --> 00:58:52,200
piece of compute, then that 
would kind of be an issue. 

1106
00:58:52,200 --> 00:58:54,700
Because you could quite easily 
Das the system by set grabbing 

1107
00:58:54,700 --> 00:58:57,300
lots of tasks and saying, I can 
do these never doing them 

1108
00:58:57,500 --> 00:58:59,900
wasting everybody's time and 
kind of money and things like 

1109
00:58:59,900 --> 00:59:02,200
that. 
So it really is a decision 

1110
00:59:02,200 --> 00:59:04,200
exactly. 
It feeds into that decision on 

1111
00:59:04,200 --> 00:59:06,600
the size of a task and that's 
there's lots of other things 

1112
00:59:06,600 --> 00:59:08,600
that feed into that. 
There's a parallelization that 

1113
00:59:08,600 --> 00:59:10,800
you mentioned as well. 
And how you split the tasks up 

1114
00:59:10,800 --> 00:59:14,300
into the most optimal structure.
At the end of the day, we're 

1115
00:59:14,300 --> 00:59:17,600
doing a lot of research on 
Figuring out what that should be

1116
00:59:17,600 --> 00:59:20,300
based on the constraints. 
When we launched our test, net 

1117
00:59:20,300 --> 00:59:23,300
will do it based on the kind of 
practical aspects as well. 

1118
00:59:23,300 --> 00:59:26,100
When we see how this actually 
works in the real world, we're 

1119
00:59:26,100 --> 00:59:29,400
very conscious that it's easy to
kind of divide Define this in 

1120
00:59:29,400 --> 00:59:32,400
the kind of perfect system and 
say, yes, this is the best size 

1121
00:59:32,400 --> 00:59:34,400
of task. 
And then you go out launch a 

1122
00:59:34,400 --> 00:59:36,900
test, that someone does 
something really weird and you 

1123
00:59:36,900 --> 00:59:38,900
realize that you have to 
completely change it again. 

1124
00:59:39,200 --> 00:59:43,300
So it's part research part. 
Let's just see how it functions 

1125
00:59:43,300 --> 00:59:44,700
when we get out there 
essentially. 

1126
00:59:45,300 --> 00:59:50,300
How do so basically if If I get 
a specified model and the 

1127
00:59:50,300 --> 00:59:54,300
training data, how can you make 
sure that I've actually done the

1128
00:59:54,300 --> 00:59:56,200
job right? 
Because it's very much not 

1129
00:59:56,200 --> 00:59:58,100
deterministic. 
So it's not like, you know, you 

1130
00:59:58,100 --> 01:00:01,700
can make me do a hash and then 
the hash will tell you whether 

1131
01:00:01,700 --> 01:00:05,500
I've done it or not. 
How do you build in checkpoints 

1132
01:00:05,500 --> 01:00:10,100
into this into this process? 
Because otherwise, I could, I 

1133
01:00:10,100 --> 01:00:13,100
could just, you know, pretend to
do the work and then kind of, 

1134
01:00:13,200 --> 01:00:16,200
you know, this was, this was a 
lazy moderate kind of didn't. 

1135
01:00:16,400 --> 01:00:18,600
Do the work, it kind of, maybe 
stupid. 

1136
01:00:18,600 --> 01:00:20,600
I don't know. 
I've done it but it just 

1137
01:00:20,600 --> 01:00:23,100
couldn't be taught. 
Essentially, that's the big 

1138
01:00:23,100 --> 01:00:24,800
challenge that verification 
system. 

1139
01:00:24,800 --> 01:00:30,500
It's a huge Challenge and I 
think the simplest most secure 

1140
01:00:30,500 --> 01:00:33,800
solution to say is a zero 
knowledge proof of the entire 

1141
01:00:33,800 --> 01:00:35,100
computation. 
Essentially. 

1142
01:00:35,400 --> 01:00:37,200
That's sort of what you think 
about. 

1143
01:00:37,200 --> 01:00:41,900
In X years time, we should be 
able to do any computation as 

1144
01:00:41,900 --> 01:00:44,600
part of his own knowledge proof.
And then we can you can check 

1145
01:00:44,600 --> 01:00:47,900
that proof to say. 
Definitively whether someone's 

1146
01:00:47,900 --> 01:00:51,600
done that computation or not? 
Don't you need a new second for 

1147
01:00:51,600 --> 01:00:54,500
each given computation? 
Yeah. 

1148
01:00:54,500 --> 01:00:56,600
So right now, that's that's the 
case. 

1149
01:00:56,600 --> 01:00:59,400
And it's horrible to try and do 
for machine learning work. 

1150
01:00:59,400 --> 01:01:03,400
The computations are massive. 
You need a DSL for defining 

1151
01:01:03,500 --> 01:01:07,200
circuit, with respect to a 
machine learning computation, 

1152
01:01:08,000 --> 01:01:10,800
it's horrible. 
Essentially, our approach is to 

1153
01:01:10,800 --> 01:01:14,000
have a hybrid between the that 
and a probabilistic mechanism. 

1154
01:01:14,300 --> 01:01:16,300
We sort of follow some 
principles. 

1155
01:01:16,400 --> 01:01:20,700
Suppose in work called proof of 
learning by Nicolas paper knots 

1156
01:01:20,700 --> 01:01:24,000
Group which is it was a paper 
within the machine learning 

1157
01:01:24,000 --> 01:01:27,500
World, essentially, showed that 
using the path through gradient 

1158
01:01:27,500 --> 01:01:31,300
space that we described before, 
you can sort of create this 

1159
01:01:31,300 --> 01:01:34,300
certificate proof using 
checkpoints and that space that 

1160
01:01:34,300 --> 01:01:39,800
theoretically it's just as hard 
to generate a realistic-looking 

1161
01:01:39,800 --> 01:01:42,000
path as it is to just do the 
work. 

1162
01:01:42,000 --> 01:01:45,300
And then using a kind of 
financial irrational assumption 

1163
01:01:45,600 --> 01:01:48,200
on all of the Ben's you can say 
they would just do the work. 

1164
01:01:48,200 --> 01:01:51,300
Essentially, there were issues 
with that paper and kind of 

1165
01:01:51,300 --> 01:01:54,000
flaws with the with the ways 
that check things but 

1166
01:01:54,000 --> 01:01:57,900
fundamentally what it showed was
using a stanch Ali, random 

1167
01:01:57,900 --> 01:02:02,000
auditing scheme on top of a 
path, through gradient space, 

1168
01:02:02,000 --> 01:02:05,100
you can build up a relatively 
robust check. 

1169
01:02:05,200 --> 01:02:07,900
And essentially, we take that 
one step further by introducing 

1170
01:02:07,900 --> 01:02:11,500
zero-knowledge proofs at certain
steps, and on top of the kind of

1171
01:02:11,500 --> 01:02:15,300
global loss of the model, just 
add another definitive kind of 

1172
01:02:15,300 --> 01:02:17,800
proof on top. 
And we package all of that up 

1173
01:02:17,800 --> 01:02:19,400
within a game theoretic 
mechanism. 

1174
01:02:19,400 --> 01:02:22,300
That looks quite a bit like true
bit from, from way back in the 

1175
01:02:22,300 --> 01:02:26,800
day with staking and slashing 
solving the verifies dilemma 

1176
01:02:26,800 --> 01:02:30,300
with like random jackpots 
essentially with whistleblowers,

1177
01:02:30,300 --> 01:02:32,600
and that's the full system. 
Although I'm aware. 

1178
01:02:32,700 --> 01:02:36,200
That is a, just a big word, 
vomit of things. 

1179
01:02:36,200 --> 01:02:39,500
So, happy to dig into bits of. 
Yeah, maybe that's, there's so 

1180
01:02:39,500 --> 01:02:42,800
much unpack here, so, maybe kind
of that's back up to true it. 

1181
01:02:42,800 --> 01:02:45,100
So I think lots of people kind 
of, remember prove it. 

1182
01:02:45,200 --> 01:02:49,600
It's kind of this. 
Basically, it lets you do large 

1183
01:02:49,600 --> 01:02:52,400
computations of chain and then 
base, you can prove it where the

1184
01:02:52,400 --> 01:02:56,000
binary search on Shane. 
If anything's is that, is that a

1185
01:02:56,000 --> 01:02:59,300
fair summary? 
Yeah, I think, essentially 

1186
01:02:59,300 --> 01:03:01,700
that's, that's exactly it. 
So, true bit prove that you 

1187
01:03:01,700 --> 01:03:05,100
could take a very large 
computation that wouldn't fit in

1188
01:03:05,100 --> 01:03:08,500
the evm or would be absolutely 
massive and really expensive, do

1189
01:03:08,500 --> 01:03:11,000
it off chain and using that 
challenge mechanism and that 

1190
01:03:11,008 --> 01:03:14,300
search that you described 
eventually, prove it on chain 

1191
01:03:14,300 --> 01:03:18,500
with the chain, doing a tiny 
Ian, we take that same 

1192
01:03:18,500 --> 01:03:22,000
principle, we apply it on top of
some of the sort of certificate 

1193
01:03:22,000 --> 01:03:23,600
proof stuff that we mentioned 
before. 

1194
01:03:23,800 --> 01:03:26,700
So if you applied that to a full
machine learning training job, 

1195
01:03:26,800 --> 01:03:29,100
you'd be searching forever like 
it's enormous. 

1196
01:03:29,300 --> 01:03:32,500
So we distill that down into a 
kind of a smaller proof that is 

1197
01:03:32,500 --> 01:03:36,700
still representative of the 
larger IE rather than doing the 

1198
01:03:36,700 --> 01:03:39,500
full thing. 
You do one in 100 checkpoints or

1199
01:03:39,500 --> 01:03:41,200
something like that. 
You've already reduced the size 

1200
01:03:41,200 --> 01:03:44,300
by 100, then you go into that 
challenge mechanism. 

1201
01:03:44,500 --> 01:03:47,000
There's also some work in the 
machine learning Space again 

1202
01:03:47,000 --> 01:03:50,600
which is applied the true B 
mechanism to neural networks. 

1203
01:03:50,700 --> 01:03:53,400
But rather than using virtual 
machine instructions for that 

1204
01:03:53,400 --> 01:03:57,300
search, use a graph and you 
Traverse kind of Merkle tree, 

1205
01:03:57,300 --> 01:04:01,500
graph of a neural network graph,
essentially of operations and 

1206
01:04:01,500 --> 01:04:03,200
you can do that at different 
granularities. 

1207
01:04:03,200 --> 01:04:05,000
So you can do it at Native 
operations. 

1208
01:04:05,000 --> 01:04:08,600
Like within Pi torture 
tensorflow on a convolution and 

1209
01:04:08,600 --> 01:04:11,100
then you can step into that 
convolution and do the Matrix 

1210
01:04:11,100 --> 01:04:13,800
multiplications that are 
involved and then you can step 

1211
01:04:13,800 --> 01:04:16,200
into that matrix multiplication 
and do the individual. 

1212
01:04:16,300 --> 01:04:19,100
And of floating Point operations
that are involved, it's quite 

1213
01:04:19,100 --> 01:04:21,000
large overhead. 
So it requires you to do that 

1214
01:04:21,000 --> 01:04:23,100
big reduction before you get to 
that stage. 

1215
01:04:23,100 --> 01:04:26,500
But once you do it provides, 
that crucial link, that goes 

1216
01:04:26,500 --> 01:04:30,200
from random off chain 
participant to full consensus of

1217
01:04:30,200 --> 01:04:33,100
the chain with the chain 
running, something and that, 

1218
01:04:33,100 --> 01:04:36,000
links back to what we said 
earlier about being a layer on 

1219
01:04:36,000 --> 01:04:39,000
versus a layer. 
To, as a layer one, we can also 

1220
01:04:39,000 --> 01:04:41,500
increase the size of computation
that the chain can do. 

1221
01:04:41,700 --> 01:04:44,600
So if we make the chain to a 
matrix multiplication and that's

1222
01:04:44,600 --> 01:04:48,100
okay, then we get to kind of 
skip that step in to a matrix 

1223
01:04:48,100 --> 01:04:50,500
multiplication doing following 
points and things which is quite

1224
01:04:50,500 --> 01:04:53,300
nice at the end of the day, it's
constraints and assumptions 

1225
01:04:53,300 --> 01:04:56,200
again, you increase the hardware
of the validators and things, so

1226
01:04:56,200 --> 01:04:59,100
you've got to be careful but we 
like the flexibility and being 

1227
01:04:59,100 --> 01:05:01,400
able to kind of change all those
levers and things. 

1228
01:05:02,700 --> 01:05:05,800
I totally get that. 
So, I think basically, kind of 

1229
01:05:06,200 --> 01:05:11,900
fixing the, the block gas limit 
and kind of maybe repricing some

1230
01:05:11,900 --> 01:05:14,000
opcodes. 
And I mean, obviously it gives 

1231
01:05:14,000 --> 01:05:19,800
you it goes a long way, right? 
Maybe that's talk about the how 

1232
01:05:19,800 --> 01:05:23,500
the blockchain itself Works in 
just a bit but there are two 

1233
01:05:23,500 --> 01:05:28,100
more parties in the process. 
So there's the verifier will 

1234
01:05:28,100 --> 01:05:31,600
actually make sure that the 
check points have been checked. 

1235
01:05:31,800 --> 01:05:35,600
T' and then The Whistleblower 
who makes who make sure that the

1236
01:05:35,607 --> 01:05:37,800
verifier actually operates 
correctly. 

1237
01:05:38,000 --> 01:05:41,100
Can you go through what their 
respective roles are? 

1238
01:05:41,800 --> 01:05:43,500
Yes. 
So the verifying The 

1239
01:05:43,500 --> 01:05:47,800
Whistleblower have a 
relationship similar to the 

1240
01:05:47,800 --> 01:05:51,900
verifier in the kind of worker 
in the tribute paper. 

1241
01:05:51,900 --> 01:05:56,300
So essentially The Whistleblower
solves the verifiers Dilemma 

1242
01:05:56,300 --> 01:06:01,100
problem which is the idea that 
you won't necessarily about you 

1243
01:06:01,100 --> 01:06:04,600
know, verify Mark, unless, you 
know, you can be simply expect 

1244
01:06:04,600 --> 01:06:08,200
there to be warkworth kind of, 
you know, catching is being 

1245
01:06:08,200 --> 01:06:12,000
wrong and being rewarded for. 
So the question blower 

1246
01:06:12,000 --> 01:06:15,200
essentially checks that the The 
verifiers Works being correct. 

1247
01:06:15,200 --> 01:06:18,900
There's also incentivized to do 
so by force Terrors from the 

1248
01:06:18,900 --> 01:06:21,600
verifier. 
So, the verifier wasn't from the

1249
01:06:21,600 --> 01:06:26,000
true bit paper right there. 
Yeah, it kind of like that the 

1250
01:06:26,000 --> 01:06:29,700
dogs at the baggage carousel, 
where we see if they don't find,

1251
01:06:29,700 --> 01:06:32,900
you know, any drugs in their 
hand Let's put, you know, like a

1252
01:06:32,900 --> 01:06:35,100
suitcase with drugs. 
So, you know, they don't end up 

1253
01:06:35,100 --> 01:06:38,600
depressed and you know, stuff 
working the dogs need dog 

1254
01:06:38,600 --> 01:06:43,600
treats, occasionally the that's 
basically the kind of the kind 

1255
01:06:43,600 --> 01:06:47,800
of thinking there, so it falls. 
The basically, the solver does 

1256
01:06:47,800 --> 01:06:51,900
the work. 
If the work is incorrect, the 

1257
01:06:52,000 --> 01:06:55,300
verifier shows it as being 
correct and the list of Laura, 

1258
01:06:55,300 --> 01:06:57,800
can they confirm? 
It's incorrect that then goes 

1259
01:06:57,800 --> 01:07:02,300
back onto the kind of chain 
which we can talk to in In terms

1260
01:07:02,300 --> 01:07:04,500
of, you know, being being 
verified on chain. 

1261
01:07:05,000 --> 01:07:10,100
But essentially, it periodically
and also kind of the rate at, 

1262
01:07:10,100 --> 01:07:13,100
which is kind of linked to the 
security of the system, I guess.

1263
01:07:13,300 --> 01:07:18,200
V, the verifier will show up an 
error on purpose to twist a 

1264
01:07:18,207 --> 01:07:21,600
blower, which keeps The 
Whistleblower wanting to be the 

1265
01:07:21,600 --> 01:07:24,100
kind of engaged with time, as 
well. 

1266
01:07:24,700 --> 01:07:28,100
If there was a blower, does find
a problem, the play a game 

1267
01:07:28,100 --> 01:07:30,400
pinpoint protocol, where we 
narrow down equal to down the 

1268
01:07:30,408 --> 01:07:34,100
computation to a single Kind of 
point in the, in the kind of I 

1269
01:07:34,100 --> 01:07:36,900
guess you could view it like the
miracle tree of computations for

1270
01:07:36,900 --> 01:07:39,800
that area of the neural network 
and then that goes to the the 

1271
01:07:39,800 --> 01:07:44,300
chain for, for arbitration. 
That's the kind of the the 

1272
01:07:44,300 --> 01:07:48,600
version of it it kind of plane 
way that we originally had has 

1273
01:07:48,600 --> 01:07:51,300
been mentioned earlier. 
We've Advanced on it in a couple

1274
01:07:51,300 --> 01:07:54,900
of areas after a basically 
closing or sit around and doing 

1275
01:07:54,900 --> 01:07:56,900
more research work. 
But yeah that's the verifying, 

1276
01:07:56,900 --> 01:08:01,000
it was Civil War. 
So tell me how that fits into 

1277
01:08:01,100 --> 01:08:04,500
the blockchain as a whole. 
So obviously someone has to 

1278
01:08:04,500 --> 01:08:08,400
build blocks, there has to be I 
assume this is some kind of 

1279
01:08:08,400 --> 01:08:12,900
shaking Network so has to be 
staking token and how does all 

1280
01:08:12,900 --> 01:08:17,300
of this fit in with the Jensen 
prodigal? 

1281
01:08:18,500 --> 01:08:22,399
Yeah. 
Essentially it's a vanilla to an

1282
01:08:22,399 --> 01:08:26,399
extent substrate blockchain, we 
use the proof of stake. 

1283
01:08:26,399 --> 01:08:28,500
Grandpa babe consensus 
mechanism. 

1284
01:08:28,600 --> 01:08:31,899
Validators. 
Just kind of doing things in the

1285
01:08:31,907 --> 01:08:36,200
normal way that they will, and 
all of the parts that Harry 

1286
01:08:36,200 --> 01:08:38,100
described. 
And I described earlier happen 

1287
01:08:38,100 --> 01:08:41,200
off Jane, they're all kind of 
off chain participant's doing 

1288
01:08:41,200 --> 01:08:44,500
portions of work and kind of 
being incentivized by the fact 

1289
01:08:44,500 --> 01:08:47,700
that they've staked through a 
kind of normal staking palette 

1290
01:08:47,700 --> 01:08:50,300
within substrate, but been a 
smart contract, it could just be

1291
01:08:50,300 --> 01:08:53,899
submitting, a certain amount of 
tokens and then there will be 

1292
01:08:53,899 --> 01:08:55,899
rewarded. 
When that work, ultimately gets 

1293
01:08:55,899 --> 01:08:58,100
checked. 
Now, the kind of game theoretic 

1294
01:08:58,600 --> 01:09:02,300
Faculty here is making sure that
all the staking potential 

1295
01:09:02,300 --> 01:09:05,899
slashing amounts and the reward 
amounts all add up. 

1296
01:09:05,899 --> 01:09:09,300
So that there isn't an incentive
somewhere for somebody to either

1297
01:09:09,300 --> 01:09:12,800
be lazy or to do something that 
is malicious essentially. 

1298
01:09:12,800 --> 01:09:15,600
So it gets complicated. 
When you add more participants 

1299
01:09:15,600 --> 01:09:17,500
in the kind of having the 
Whistleblower. 

1300
01:09:17,500 --> 01:09:20,800
There is an additional 
participant is annoying because 

1301
01:09:20,800 --> 01:09:24,000
it's over complicated but it's 
crucial for us given the size of

1302
01:09:24,000 --> 01:09:27,899
the computations to have it 
there to assure the honesty of 

1303
01:09:27,899 --> 01:09:31,200
the verifier. 
Essentially, it's not certain 

1304
01:09:31,200 --> 01:09:33,000
that we'll always have to have 
that we do. 

1305
01:09:33,000 --> 01:09:35,399
Keep thinking about ways that we
can potentially remove the 

1306
01:09:35,399 --> 01:09:37,300
Whistleblower. 
There's certain zero knowledge, 

1307
01:09:37,300 --> 01:09:39,500
proof techniques, that mean that
we potentially could. 

1308
01:09:39,800 --> 01:09:41,600
But we don't want to get ahead 
of ourselves essentially. 

1309
01:09:41,600 --> 01:09:44,000
So right now, it kind of looks 
like what's described in the 

1310
01:09:44,000 --> 01:09:47,700
light paper but we're chipping 
away at each bit of it to try 

1311
01:09:47,700 --> 01:09:50,600
and simplify it. 
We think, if you look at the way

1312
01:09:50,600 --> 01:09:53,399
that other protocols have gone 
in the past, there's a tendency 

1313
01:09:53,399 --> 01:09:56,600
to launch with a complicated 
system and then once you get it 

1314
01:09:56,600 --> 01:09:59,600
out there, realize that you can 
simplify buy it and we're 

1315
01:09:59,900 --> 01:10:01,500
expecting to go through that 
essentially. 

1316
01:10:01,500 --> 01:10:04,200
We kind of saw the same thing 
with polka dot on the fisherman 

1317
01:10:04,900 --> 01:10:08,000
mechanism that sort of got 
removed after the thing had 

1318
01:10:08,000 --> 01:10:11,100
launched and it was out in life.
I died. 

1319
01:10:11,100 --> 01:10:14,100
One other point. 
They're just on the kind of our 

1320
01:10:14,300 --> 01:10:17,000
augmentation of the vanilla kind
of substrate shown. 

1321
01:10:17,700 --> 01:10:20,200
There's an issue in the 
verification system as we 

1322
01:10:20,200 --> 01:10:23,400
originally proposed it and also 
as it currently looks for us 

1323
01:10:23,900 --> 01:10:29,100
state-of-the-art whereby if the 
the data which is being used To 

1324
01:10:29,200 --> 01:10:33,900
perform the initial kind of 
smart from the solver is removed

1325
01:10:33,900 --> 01:10:36,600
or made inaccessible. 
Halfway through the verification

1326
01:10:36,600 --> 01:10:39,100
process, you reach a kind of 
standoff. 

1327
01:10:39,100 --> 01:10:42,100
Because at the barrel fire, 
can't access the data. 

1328
01:10:42,100 --> 01:10:46,400
Then there's a follows that they
can't verify some kind of data 

1329
01:10:46,400 --> 01:10:49,900
availability solution that kind 
of plugs into it precisely. 

1330
01:10:49,900 --> 01:10:51,100
Yes. 
So we are. 

1331
01:10:51,300 --> 01:10:54,900
We built that in to on top of 
the old kind of substrate. 

1332
01:10:54,900 --> 01:10:57,000
So we have a proof of 
availability. 

1333
01:10:57,200 --> 01:11:00,100
P of a is this. 
Kind of what we've got adopted 

1334
01:11:00,100 --> 01:11:05,000
internally layer, which is 
Erasure encoded etcetera. 

1335
01:11:05,000 --> 01:11:09,800
And basically provides what we 
couldn't find in The Wider, kind

1336
01:11:09,800 --> 01:11:13,100
of storage layer market. 
And if anyone's listening to 

1337
01:11:13,100 --> 01:11:15,900
this huge building that space 
and this does exist, I would 

1338
01:11:16,200 --> 01:11:17,600
honestly be fascinated to see 
it. 

1339
01:11:17,600 --> 01:11:23,000
But essentially a layer where in
you can lock data for a period 

1340
01:11:23,000 --> 01:11:27,200
of time in a way, which is 
pinned on pinnable for that 

1341
01:11:27,200 --> 01:11:30,400
period of time. 
And verified on chain that the 

1342
01:11:30,400 --> 01:11:34,800
exist there and threw that on 
our we've it's too expensive and

1343
01:11:34,800 --> 01:11:38,000
are we've. 
So are we is the answer but the 

1344
01:11:38,000 --> 01:11:41,300
cost for if you think about, you
know a terabyte of training 

1345
01:11:41,300 --> 01:11:43,800
data, being stored Forever on 
our. 

1346
01:11:43,800 --> 01:11:47,800
We've it just doesn't work when 
the kind of alternative is like,

1347
01:11:47,800 --> 01:11:49,900
you know, storing it on S3. 
So, yeah. 

1348
01:11:49,900 --> 01:11:52,800
Should I should also caveat it 
has to be an expensive. 

1349
01:11:53,800 --> 01:11:59,500
Yeah, but yes, just on that with
our we've The reason we need it 

1350
01:11:59,500 --> 01:12:03,000
is is for the data, the training
data, but it's also for some of 

1351
01:12:03,000 --> 01:12:06,100
the intermediate, like proof 
data, and that doesn't need to 

1352
01:12:06,100 --> 01:12:09,300
be around for very long. 
It could just be 20 seconds. 

1353
01:12:09,300 --> 01:12:12,300
While we go through a certain 
number of block block, like, 

1354
01:12:12,300 --> 01:12:16,000
what releases or something with 
our we've, we don't need like 

1355
01:12:16,000 --> 01:12:18,800
200 years of storage if that 
thing, but that's what we're 

1356
01:12:18,800 --> 01:12:22,500
paying for essentially. 
So if somebody has really short 

1357
01:12:22,500 --> 01:12:26,100
term but with the guarantees of 
our we've so bringing the price 

1358
01:12:26,100 --> 01:12:28,400
down because it doesn't need to 
be 200 plus years. 

1359
01:12:29,000 --> 01:12:31,300
That's what we want. 
Basically, we just haven't seen 

1360
01:12:31,300 --> 01:12:32,900
it anywhere. 
You should go. 

1361
01:12:33,000 --> 01:12:36,500
You guys should talk to Ali's. 
I mean, I weave the storage 

1362
01:12:36,500 --> 01:12:39,700
rent. 
This would be, you know, this 

1363
01:12:39,700 --> 01:12:43,100
might be a thing. 
Yeah, it's like a sort of Perma 

1364
01:12:43,100 --> 01:12:47,700
web that are temporary web. 
The temperature at the Tempe 

1365
01:12:47,700 --> 01:12:49,100
Perma web. 
Right. 

1366
01:12:49,400 --> 01:12:54,300
So I assume that there's going 
to be agents in token somewhere 

1367
01:12:54,300 --> 01:12:56,500
in this eventually, tell me 
about that. 

1368
01:12:57,700 --> 01:12:59,100
Sure. 
So the Jensen token 

1369
01:12:59,100 --> 01:13:02,400
fundamentally is required for 
the attack essentially, 

1370
01:13:02,400 --> 01:13:05,100
everything that we've just 
described assumes that you have 

1371
01:13:05,700 --> 01:13:10,000
this token built-in that can be 
used to stake /, provide 

1372
01:13:10,000 --> 01:13:13,500
rewards, Etc. 
And also maintain the consensus 

1373
01:13:13,500 --> 01:13:15,400
of the system itself 
essentially. 

1374
01:13:15,400 --> 01:13:19,800
So use kind of a small inflation
amount to pay out validators, 

1375
01:13:19,800 --> 01:13:23,400
and then be used in that game 
theoretic mechanism to allow us 

1376
01:13:23,400 --> 01:13:26,000
to kind of guarantee that 
Financial irrational kind of 

1377
01:13:26,000 --> 01:13:28,000
assumption over the entire 
system. 

1378
01:13:28,200 --> 01:13:31,200
I think crucially for us, that's
what it's for. 

1379
01:13:31,300 --> 01:13:35,500
And that's like the only thing 
it's for we're very, very 

1380
01:13:35,500 --> 01:13:38,200
deliberate to say it's a 
technical thing that we need, 

1381
01:13:38,300 --> 01:13:41,300
but we will bring in when we 
technically need it. 

1382
01:13:41,300 --> 01:13:44,400
And not before we've seen what's
happened with with kind of 

1383
01:13:44,407 --> 01:13:46,900
utility tokens in the past where
people have kind of launched 

1384
01:13:46,900 --> 01:13:49,200
them too early. 
And then it's a distraction for 

1385
01:13:49,200 --> 01:13:51,000
the team as a distraction for 
everyone. 

1386
01:13:51,300 --> 01:13:54,100
People aren't buying it to use 
it in the system, they're buying

1387
01:13:54,100 --> 01:13:56,100
it for other reasons. 
We don't want any of that. 

1388
01:13:56,100 --> 01:13:59,800
Ideally, I mean, It's easy to 
say that it's hard to kind of 

1389
01:13:59,800 --> 01:14:01,100
see what will happen in 
practice. 

1390
01:14:01,100 --> 01:14:04,100
But our approach is to 
essentially delay it as much as 

1391
01:14:04,100 --> 01:14:07,300
possible and then quietly bring 
it in when it's required to 

1392
01:14:07,300 --> 01:14:11,000
maintain that consensus and pay 
out the participants with the 

1393
01:14:11,000 --> 01:14:15,100
game theoretic mechanism. 
Essentially, yeah, it's critical

1394
01:14:15,100 --> 01:14:19,600
to note as well that along with 
some other kind of early movers 

1395
01:14:19,600 --> 01:14:21,800
in the Deep learning crypto 
space. 

1396
01:14:22,300 --> 01:14:25,800
We're very much a minority with 
respect to the rest of a deep 

1397
01:14:25,800 --> 01:14:28,400
learning community. 
There's at least in our 

1398
01:14:28,400 --> 01:14:31,900
experience, quite a lot of 
skepticism about crypto more. 

1399
01:14:31,900 --> 01:14:36,100
Broadly, I mean, bed and eyes 
but histories kind of Testament 

1400
01:14:36,100 --> 01:14:39,600
today, you know, we were 
skeptical and we've obviously 

1401
01:14:39,600 --> 01:14:41,600
from a technological and kind of
ideological point. 

1402
01:14:41,600 --> 01:14:44,200
We like it, I think it's the 
right way. 

1403
01:14:44,300 --> 01:14:48,000
However, when we initially 
provide, when the network 

1404
01:14:48,000 --> 01:14:50,600
initially, launches we 
anticipate that the majority of 

1405
01:14:50,600 --> 01:14:53,900
the earning users will pay in 
Fiat and it will just simply be 

1406
01:14:53,900 --> 01:14:56,600
swapped into tokens, the 
solvers, and the people 

1407
01:14:56,600 --> 01:14:58,200
participating. 
On the supply side of the 

1408
01:14:58,208 --> 01:15:01,100
network will facilitate with 
with tokens. 

1409
01:15:01,600 --> 01:15:04,900
And we have huge interest from a
lot of the kind of old fiery M1 

1410
01:15:05,300 --> 01:15:08,100
miners who have lots of gpus who
want to attach them to 

1411
01:15:08,100 --> 01:15:12,000
something. 
But yeah, it's crucial that 

1412
01:15:12,000 --> 01:15:16,700
there's a kind of the kind of 
crypto the scary crypto words 

1413
01:15:16,700 --> 01:15:20,500
like Takin her kind of removed 
from the the end then deep 

1414
01:15:20,500 --> 01:15:23,900
learning and machine learning 
users, which is exciting because

1415
01:15:23,900 --> 01:15:27,200
they're ours, this is one of 
those use cases which really 

1416
01:15:27,600 --> 01:15:29,800
Bridges. 
The Two Worlds of web to and web

1417
01:15:29,800 --> 01:15:33,000
three, you know, there's our 
economic rationale for this 

1418
01:15:33,000 --> 01:15:36,300
existing. 
There's the technology now to 

1419
01:15:36,300 --> 01:15:39,700
enable it. 
Its existence, it's Dow. 

1420
01:15:39,700 --> 01:15:42,700
Almost like a kind of execution 
question to a large extent, you 

1421
01:15:42,700 --> 01:15:45,300
know, how do you just hold you 
get people comfortable with the 

1422
01:15:45,300 --> 01:15:49,500
idea that I'm not using Amazon? 
And, you know, there's a kind of

1423
01:15:49,600 --> 01:15:53,200
variable price kind of concert 
happening here with the talking 

1424
01:15:53,200 --> 01:15:56,400
and how the obfuscated that as 
much as possible and crucially 

1425
01:15:56,400 --> 01:15:58,800
how do you have to skate it? 
Wait, which is decentralized 

1426
01:15:58,800 --> 01:16:02,000
because if you very easy just to
stand out some centralized API 

1427
01:16:02,000 --> 01:16:05,200
front end which you know, just 
yeah automatically converts the 

1428
01:16:05,200 --> 01:16:07,700
tokens and a centralized 
exchange somewhere but that 

1429
01:16:07,700 --> 01:16:09,500
becomes that brings its own 
problems. 

1430
01:16:10,100 --> 01:16:12,300
So yeah, I just add that to the 
pot. 

1431
01:16:14,200 --> 01:16:17,500
So what is the roadmap look like
for you guys, there's going to 

1432
01:16:17,500 --> 01:16:20,900
be a test net early next year. 
Yep. 

1433
01:16:20,900 --> 01:16:22,500
Test net. 
Early next year. 

1434
01:16:23,400 --> 01:16:27,300
It won't be incentivized. 
Kind of torture. 

1435
01:16:27,400 --> 01:16:31,600
Earlier and it's more to pick up
the kind of do two things. 

1436
01:16:31,600 --> 01:16:35,500
Firstly to kind of battle test, 
some of the, some of the tech 

1437
01:16:35,500 --> 01:16:38,000
that we've been building 
internally and 22 going to get 

1438
01:16:38,000 --> 01:16:43,100
feedback on the usability of it.
Overall, that will be that 

1439
01:16:43,100 --> 01:16:47,600
precedes our kind of incentivize
test net which will be which, 

1440
01:16:47,800 --> 01:16:50,500
yeah, well, essentially, you'll 
be able to train models, kind 

1441
01:16:50,500 --> 01:16:56,500
of, you know, in anger on it. 
The the rate at which we move 

1442
01:16:56,500 --> 01:16:59,900
is, It's something that we 
talked about a lot, you know, we

1443
01:16:59,900 --> 01:17:03,800
could ship something very very 
soon, which doesn't really give 

1444
01:17:03,800 --> 01:17:06,100
us any meaningful feedback, but 
it kind of looks good because 

1445
01:17:06,100 --> 01:17:08,700
it's like all you kind of shit 
something and we don't want to 

1446
01:17:08,708 --> 01:17:13,000
fall into that trap because 
yeah, there's there's been lots 

1447
01:17:13,000 --> 01:17:16,400
of things which have kind of 
come and gone for us where there

1448
01:17:16,400 --> 01:17:18,500
have been like incentives to 
kind of ship something super 

1449
01:17:18,500 --> 01:17:20,500
early. 
So even kind of you know earlier

1450
01:17:20,500 --> 01:17:23,000
this year there was lots of kind
of hype around the idea of doing

1451
01:17:23,000 --> 01:17:26,600
generative nft arts and we could
have kind of provided like an 

1452
01:17:26,600 --> 01:17:28,600
inference. 
And for that, you know, really 

1453
01:17:28,600 --> 01:17:32,200
quickly, but we decided that it 
kind of, It kind of blows it 

1454
01:17:32,200 --> 01:17:34,700
out, step of her principles, you
know, it doesn't solve. 

1455
01:17:34,700 --> 01:17:36,700
The big problem is not really on
the way to solving the big 

1456
01:17:36,700 --> 01:17:38,300
problem. 
It had lots of kind of other 

1457
01:17:38,300 --> 01:17:40,900
things we have to build. 
So yeah, I guess I'm trying to 

1458
01:17:40,900 --> 01:17:45,100
say is we're not like in a kind 
of immediate rush to believe 

1459
01:17:45,100 --> 01:17:46,300
something. 
Tomorrow, we die were really 

1460
01:17:46,300 --> 01:17:48,700
something, which is Meaningful, 
which takes time giving a 

1461
01:17:48,700 --> 01:17:49,800
fundamental. 
Some of the stuff is, 

1462
01:17:49,800 --> 01:17:53,100
particularly the zero-knowledge 
stuff which is, you know, which 

1463
01:17:53,100 --> 01:17:55,100
is it was a pretty involved in 
terms of time. 

1464
01:17:55,900 --> 01:17:57,400
Cool. 
Thank you guys. 

1465
01:17:57,600 --> 01:17:59,900
Where can people go? 
To learn more about Jensen, 

1466
01:18:00,500 --> 01:18:01,200
yesil. 
Jensen. 

1467
01:18:01,200 --> 01:18:03,400
Do AI, is you're going to 
primary source? 

1468
01:18:04,000 --> 01:18:07,300
We have a discard, we don't have
a telegram group, but this cards

1469
01:18:07,300 --> 01:18:10,200
were a lot of chat happens. 
We're also hiring just now so if

1470
01:18:10,200 --> 01:18:13,100
anyone's listening to this and 
is interested in, you know, 

1471
01:18:13,100 --> 01:18:16,800
building a permissionless, deep 
learning compute protocol than 

1472
01:18:17,100 --> 01:18:22,200
we are, we're it and I guess. 
Moreover next year we should be 

1473
01:18:22,200 --> 01:18:24,600
hosting as your knowledge 
machine learning Summit. 

1474
01:18:24,700 --> 01:18:27,900
So if anyone's Particularly 
interested in that kind of 

1475
01:18:27,900 --> 01:18:30,600
crossover. 
There's that and maybe as a 

1476
01:18:30,600 --> 01:18:34,000
final point, maybe more for 
traditional kind of deep 

1477
01:18:34,000 --> 01:18:36,200
learning or machine learning 
people coming was listening or 

1478
01:18:36,200 --> 01:18:40,000
sponsoring the the New York's 
conference in New Orleans next 

1479
01:18:40,000 --> 01:18:41,900
week. 
So, we'll bet I will both be in 

1480
01:18:41,900 --> 01:18:45,200
Louisiana at the conference 
attending the talks flying the 

1481
01:18:45,208 --> 01:18:50,600
flag for a crypto and yeah, if 
anyone's there, we were more 

1482
01:18:50,600 --> 01:18:53,200
than happy to chat. 
Super good. 

1483
01:18:53,300 --> 01:18:56,600
Thank you both for coming on. 
This was Super interesting. 

1484
01:18:57,100 --> 01:19:00,200
We look forward to kind of 
seeing how this plays out with 

1485
01:19:00,200 --> 01:19:01,900
them. 
The test net and the mainland 

1486
01:19:02,600 --> 01:19:04,000
fantastic. 
Appreciate it. 

1487
01:19:04,000 --> 01:19:05,400
Thanks dropping songs been a 
pleasure. 

1488
01:19:06,000 --> 01:19:08,000
Yeah, thanks so much for having 
his own really enjoyed it. 

1489
01:19:08,000 --> 01:19:10,100
Great questions as well. 
Really interesting. 

1490
01:19:11,100 --> 01:19:14,700
Thank you. 
Thank you for joining us on this

1491
01:19:14,700 --> 01:19:17,100
week's episode. 
We release new episodes every 

1492
01:19:17,100 --> 01:19:19,100
week. 
You can find And subscribe to 

1493
01:19:19,100 --> 01:19:22,900
the show on iTunes Spotify, 
YouTube SoundCloud or wherever 

1494
01:19:22,900 --> 01:19:25,200
you listen to podcast. 
And if you have a Google home or

1495
01:19:25,200 --> 01:19:28,100
Alexa device, you can tell it to
listen to the latest episode of 

1496
01:19:28,100 --> 01:19:32,000
the epicenter podcast, go to 
epicenter, .t V /, subscribe for

1497
01:19:32,000 --> 01:19:34,600
a full list of places where you 
can watch and listen, while 

1498
01:19:34,600 --> 01:19:36,300
you're there, be sure to sign up
for the newsletter. 

1499
01:19:36,600 --> 01:19:39,200
You get new episodes in your 
inbox as they're released. 

1500
01:19:39,900 --> 01:19:42,300
If you want to interact with us 
guests or other podcast 

1501
01:19:42,300 --> 01:19:45,500
listeners, you can Us on Twitter
and please leave us a review on 

1502
01:19:45,500 --> 01:19:48,200
iTunes helps people find the 
show and we're always happy to 

1503
01:19:48,200 --> 01:19:51,400
read them but thanks so much and
we look forward to being back 

1504
01:19:51,400 --> 01:19:51,900
next week.