1
00:00:00,040 --> 00:00:02,520
Hey everybody, and welcome to 
another episode of NF Twins. 

2
00:00:02,720 --> 00:00:04,960
Today we are joined by Ganesh 
Krishnan. 

3
00:00:05,280 --> 00:00:08,360
He's the founder of AI Hello, 
which he scaled to over 36 

4
00:00:08,360 --> 00:00:11,680
people as well as Hell 0, from 
which he has now taken on a 

5
00:00:11,680 --> 00:00:13,360
different bet on the future of 
AI. 

6
00:00:14,040 --> 00:00:17,480
While everybody's chasing big 
LLMS, Ganesh believes the real 

7
00:00:17,480 --> 00:00:20,640
path to super intelligence might
be something entirely different.

8
00:00:20,960 --> 00:00:22,800
I personally really enjoyed this
episode. 

9
00:00:22,840 --> 00:00:25,480
I'm sure you will learn a lot as
well, so let's just jump right 

10
00:00:25,480 --> 00:00:28,320
in. 
Ganesh, I'm home from Australia 

11
00:00:28,320 --> 00:00:30,920
to Canada because I want my own 
AI startup. 

12
00:00:30,920 --> 00:00:32,840
And I think, you know, it's 
better to be in North America. 

13
00:00:33,760 --> 00:00:36,320
And we've been quite successful.
You know, we moved on. 

14
00:00:36,320 --> 00:00:38,560
We have around 36 employees now 
who start. 

15
00:00:38,800 --> 00:00:43,640
And because of this coming craze
of AI, we decided to branch off 

16
00:00:43,640 --> 00:00:50,880
into hallucination free models, 
AI models and reading about all 

17
00:00:50,880 --> 00:00:52,720
of it. 
You know how the CEO's of 

18
00:00:52,720 --> 00:00:57,280
Anthropic and Open AI are they 
all talking about, you know, ASI

19
00:00:57,280 --> 00:01:00,000
is around the corner, but we've 
been, you know, talking about it

20
00:01:00,000 --> 00:01:02,920
for the last 40 years and they 
all have the vested interest in 

21
00:01:02,920 --> 00:01:06,600
their own startup. 
So my point is I I think more in

22
00:01:06,600 --> 00:01:11,040
terms of yad lacun that LLMS is 
not where the Super intelligence

23
00:01:11,040 --> 00:01:12,640
lies. 
It's in a different kind of 

24
00:01:12,760 --> 00:01:14,960
model. 
And we are taking a stab at it 

25
00:01:14,960 --> 00:01:18,440
and we hope we know we make it 
into getting some kind of a 

26
00:01:18,440 --> 00:01:21,640
super intelligence that is as 
intelligent as we once if not. 

27
00:01:23,240 --> 00:01:28,400
So, Ganesha, what makes this 
model less likely to 

28
00:01:28,400 --> 00:01:33,520
hallucinate? 
It is interesting to talk about 

29
00:01:33,520 --> 00:01:36,720
it. 
Most of these LLM companies, 

30
00:01:36,720 --> 00:01:39,520
they move towards a weight model
where all the intelligence is 

31
00:01:39,520 --> 00:01:44,440
innovate and it's like numbers 
0101023 and they use some kind 

32
00:01:44,440 --> 00:01:46,880
of a pattern matching to look 
back and talk about it. 

33
00:01:47,920 --> 00:01:52,160
And because of the way that the 
data is all it's just that, you 

34
00:01:52,160 --> 00:01:54,440
know, weight model, they always 
will. 

35
00:01:54,440 --> 00:01:57,920
Hellesmet. 
So in my opinion, elements are 

36
00:01:57,920 --> 00:02:00,240
another way that we can create 
super intelligence. 

37
00:02:00,720 --> 00:02:05,480
And the way that we can know is 
we ask the simple question, will

38
00:02:05,480 --> 00:02:08,000
this LLMS ever criticize the 
training data? 

39
00:02:08,759 --> 00:02:10,919
Can it ever say that the 
training data that I've been 

40
00:02:10,919 --> 00:02:17,200
trained on is wrong? 
If you take a model from scratch

41
00:02:17,880 --> 00:02:21,280
and we train it that the sun is 
always green, we feed it text 

42
00:02:21,280 --> 00:02:24,160
that the sun is always green, it
will never question that. 

43
00:02:26,320 --> 00:02:28,880
Yeah, I mean, it's really hard, 
if you think about it, if you've

44
00:02:28,880 --> 00:02:33,360
been told something is true the 
entire life and to question that

45
00:02:33,360 --> 00:02:35,480
it's, I mean, even with humans 
we notice, right? 

46
00:02:35,480 --> 00:02:39,480
If you have a core belief that 
you truly hold dear, it is so 

47
00:02:39,480 --> 00:02:42,680
hard to question that. 
Yeah, my. 

48
00:02:42,680 --> 00:02:45,480
Question is then how can you 
make sure that this is not the 

49
00:02:45,480 --> 00:02:48,800
case? 
With, with elements, or with 

50
00:02:48,800 --> 00:02:51,120
humans? 
Well go with humans, you can 

51
00:02:51,120 --> 00:02:53,360
teach them right? 
Is I guess hard? 

52
00:02:53,680 --> 00:02:56,520
Wouldn't it make LLM very 
similar to humans in that case? 

53
00:02:56,520 --> 00:03:00,840
That if you teach them long 
enough, they will start 

54
00:03:00,840 --> 00:03:05,920
believing a certain aspect. 
Just But, you know, that's a 

55
00:03:05,920 --> 00:03:08,800
critical point of humans. 
We all bought, you know, with 

56
00:03:08,800 --> 00:03:11,360
our own nationality, with our 
own religion, with our own 

57
00:03:11,360 --> 00:03:13,560
beliefs. 
But at some point of a time, 

58
00:03:13,560 --> 00:03:15,560
something clicks. 
You know, it's just one event 

59
00:03:16,000 --> 00:03:17,360
that clicks something, we change
it. 

60
00:03:18,280 --> 00:03:21,320
But with LLM, if you have been 
trained for a million times, you

61
00:03:21,320 --> 00:03:24,960
know something is true and then 
a million and one time we say 

62
00:03:24,960 --> 00:03:28,720
that is not true because it has 
been trained so many times on 

63
00:03:28,720 --> 00:03:30,440
the previous data. 
It's just the number of 

64
00:03:30,440 --> 00:03:33,720
reputations it takes it. 
Well, that is a certain 

65
00:03:33,720 --> 00:03:35,440
similarity between humans and 
LLM. 

66
00:03:35,640 --> 00:03:38,920
At some point we can reject our 
data and we measure it now we 

67
00:03:38,920 --> 00:03:42,360
observe it on real life data and
then, you know, we can make a 

68
00:03:42,480 --> 00:03:44,480
change. 
And kids grew out of it, right? 

69
00:03:44,480 --> 00:03:46,480
By the time they're 9 or 10, 
they start questioning the 

70
00:03:46,480 --> 00:03:49,000
parents. 
They're not white like what I 

71
00:03:49,000 --> 00:03:50,760
say is white. 
So. 

72
00:03:50,760 --> 00:03:54,280
Basically it's that we humans 
get a frame breaker so to speak,

73
00:03:54,280 --> 00:03:57,880
or something that resets our 
current belief. 

74
00:03:57,960 --> 00:04:02,520
That enables us then to learn 
something that is new, which as 

75
00:04:02,520 --> 00:04:08,280
LMS have a model that was 
trained a million times and then

76
00:04:08,280 --> 00:04:11,080
the reinforced learning is not 
strong enough because they don't

77
00:04:11,080 --> 00:04:13,800
get a reset. 
Yeah, for sure. 

78
00:04:14,000 --> 00:04:16,720
And it's also, you know, one 
interesting thing is, is the 

79
00:04:16,720 --> 00:04:20,399
quality of data. 
So if you, if your teacher tells

80
00:04:20,399 --> 00:04:24,640
you know something the Earth is 
round, or if a guy on the street

81
00:04:24,640 --> 00:04:28,120
you know while waiting for the 
surfaces the earth is flat, who 

82
00:04:28,120 --> 00:04:31,520
do you rather trust? 
Depends on the guy on the street

83
00:04:31,520 --> 00:04:37,160
is. 
You you will trust less the guy 

84
00:04:37,160 --> 00:04:38,400
on the street than on the 
teacher. 

85
00:04:38,400 --> 00:04:41,120
So we have a quality. 
Of it, yeah, yeah. 

86
00:04:42,440 --> 00:04:45,360
They tell them so tap they trust
all the data blindly, right? 

87
00:04:45,360 --> 00:04:46,720
You feed them thousand 
documents. 

88
00:04:47,040 --> 00:04:48,800
They have equal weight as to all
the documents. 

89
00:04:51,120 --> 00:04:55,240
So and basically you believing 
more in something called a world

90
00:04:55,320 --> 00:05:00,320
model or OK for the for people 
who are unfamiliar with the 

91
00:05:00,320 --> 00:05:04,320
concept, could you explain what 
actually a world model means and

92
00:05:04,320 --> 00:05:08,400
how it's derived? 
So the world model is, is the 

93
00:05:08,400 --> 00:05:11,760
observation. 
We observe things and then we 

94
00:05:11,760 --> 00:05:15,080
make an inference and then we 
say, OK, what I learned before 

95
00:05:15,080 --> 00:05:17,280
was wrong because what I'm 
seeing is different. 

96
00:05:17,880 --> 00:05:20,840
So that means, you know, we need
to change our beliefs. 

97
00:05:21,120 --> 00:05:23,800
So we keep on updating our 
beliefs depending on the model. 

98
00:05:23,800 --> 00:05:27,440
And it's funny, you know, you 
talk about it because we 2 days 

99
00:05:27,440 --> 00:05:30,400
ago we talked about Leon Lacuna 
and you know, his eating salon 

100
00:05:30,400 --> 00:05:32,840
is different. 
And yesterday he published out a

101
00:05:32,840 --> 00:05:34,840
tweet that he got funded a 
billion dollars. 

102
00:05:34,840 --> 00:05:36,280
Did you see that tweet by? 
Him. 

103
00:05:36,280 --> 00:05:37,800
I did. 
I didn't see the tweet. 

104
00:05:37,800 --> 00:05:40,880
I saw it in the rundown in the 
newsletter that I followed. 

105
00:05:41,840 --> 00:05:44,240
Yeah, yeah. 
He got funded a billion dollars 

106
00:05:44,240 --> 00:05:46,760
for this world model thing and 
he's probably not. 

107
00:05:46,760 --> 00:05:48,360
He's taking a stab at it as 
well. 

108
00:05:48,360 --> 00:05:51,480
And he's working from a real 
life model and we are working in

109
00:05:51,480 --> 00:05:54,080
a different kind of a model. 
We take the data and then we say

110
00:05:54,080 --> 00:05:57,320
what quality, what kind of 
quality is better, what kind of 

111
00:05:57,320 --> 00:06:00,280
quality is not better. 
And we teach the LLM to critique

112
00:06:00,280 --> 00:06:02,640
his own training data as well. 
So we work slightly differently 

113
00:06:02,640 --> 00:06:05,520
from Yan Lacun, but I think Yan 
Lacun has a higher chance of 

114
00:06:05,520 --> 00:06:07,640
success, probably because he's 
more intelligent, he's more 

115
00:06:07,640 --> 00:06:11,640
funded, he's more connected. 
I mean billion dollars is quite 

116
00:06:11,640 --> 00:06:14,880
a lot of money. 
Also makes it hard for the next 

117
00:06:14,880 --> 00:06:17,040
round. 
A quick question. 

118
00:06:17,040 --> 00:06:19,360
So you wait. 
So basically you put still 

119
00:06:19,360 --> 00:06:22,320
weights towards the quality of 
the data as well? 

120
00:06:22,560 --> 00:06:27,640
You say hey if this is from the 
physics review then the data 

121
00:06:27,640 --> 00:06:30,120
should trust it versus A blog 
post. 

122
00:06:31,080 --> 00:06:34,920
That is correct, yeah. 
So the domain, the page, the 

123
00:06:34,960 --> 00:06:36,600
author, you know, like depends 
on that. 

124
00:06:37,800 --> 00:06:42,000
Google used to use something 
like page bank how authority web

125
00:06:42,080 --> 00:06:44,120
pages. 
So a bit based on that, we're 

126
00:06:44,120 --> 00:06:46,000
trying to, you know, put a 
qualitative. 

127
00:06:47,800 --> 00:06:50,080
Data, that's super interesting. 
You know what I also thought was

128
00:06:50,080 --> 00:06:52,560
super interesting? 
That John Lekoun actually went 

129
00:06:52,560 --> 00:06:57,120
to Europe for his his business. 
Paris, right? 

130
00:06:57,400 --> 00:07:00,320
Yeah. 
That is interesting because you 

131
00:07:00,320 --> 00:07:03,920
wouldn't think Europe would have
a lot of funding. 

132
00:07:03,920 --> 00:07:06,040
No 100%. 
Now that's I was I was thinking 

133
00:07:06,040 --> 00:07:10,360
like, because I don't know, 
everyone says always look, the 

134
00:07:10,360 --> 00:07:14,160
US is a great with the AI and in
general technological advances 

135
00:07:14,160 --> 00:07:18,080
and Europe is good in 
regulation, a lot of regulation.

136
00:07:18,880 --> 00:07:21,840
And it's just super interesting 
that he chose then to move to 

137
00:07:21,840 --> 00:07:26,760
France for his LM also. 
Well, maybe given the recent 

138
00:07:26,760 --> 00:07:30,440
news that in the US, what 
they're doing with AI companies 

139
00:07:30,440 --> 00:07:34,240
in terms of defensive, like they
want to have defense contracts 

140
00:07:34,240 --> 00:07:37,000
and whatever the regulations 
might come that you need to 

141
00:07:37,000 --> 00:07:41,320
have, I know a back door and all
the, I don't know, user 

142
00:07:41,320 --> 00:07:44,840
funneling the information. 
So it's super interesting how 

143
00:07:44,840 --> 00:07:47,760
how that was going to play out 
that Europe actually might have 

144
00:07:47,840 --> 00:07:50,440
a fundamental model coming out 
of Europe, which would be 

145
00:07:50,520 --> 00:07:53,000
insane. 
Make Europe great again. 

146
00:07:55,080 --> 00:08:01,080
I'm on Team Euro. 
My question is if you do you 

147
00:08:01,080 --> 00:08:04,040
think scaling LNMS will 
eventually hit a ceiling? 

148
00:08:04,040 --> 00:08:07,960
So for example, we've seen like 
some models that not like when 

149
00:08:07,960 --> 00:08:11,560
they bring out a new model is 
not getting so much better is 

150
00:08:11,560 --> 00:08:15,120
because we're getting closer to 
a ceiling or is just a question 

151
00:08:15,120 --> 00:08:18,040
of we didn't put enough data in 
there or high quality data? 

152
00:08:19,840 --> 00:08:23,080
I think, you know, like my 
opinion is it's a ceiling and 

153
00:08:23,400 --> 00:08:26,240
they're trying to bypass a 
ceiling using chain of thoughts,

154
00:08:27,280 --> 00:08:30,600
making it smarter. 
But I imagine a human that was 

155
00:08:30,600 --> 00:08:34,039
born blind and the only way the 
human would learn is by, you 

156
00:08:34,039 --> 00:08:37,240
know, repeating the data. 
So how much intelligent would 

157
00:08:37,240 --> 00:08:38,520
be? 
I think someone need to do an 

158
00:08:38,520 --> 00:08:42,400
experiment to find out. 
But humans do objectively, you 

159
00:08:42,400 --> 00:08:45,320
know, question the data as well.
You told me it is true. 

160
00:08:45,320 --> 00:08:50,120
Is it true everything but LLM 
blindly accepts the training 

161
00:08:50,120 --> 00:08:52,280
data? 
But in my opinion, to answer 

162
00:08:52,280 --> 00:08:56,480
your question, I definitely 
think there is a ceiling and you

163
00:08:56,480 --> 00:08:58,200
can buy a possibly chain of 
thoughts. 

164
00:08:58,360 --> 00:09:02,240
It doesn't solve the problem. 
We need to completely change the

165
00:09:02,240 --> 00:09:05,040
direction in which we're going 
if we want to have ASR. 

166
00:09:07,800 --> 00:09:11,880
Do you think also that has 
there's a connection then to how

167
00:09:11,880 --> 00:09:15,960
much data we can absorb or let's
say the LLM can absorb? 

168
00:09:15,960 --> 00:09:20,640
Because like you mentioned in 
our call last time, Ian Likun, I

169
00:09:20,640 --> 00:09:23,800
did some research. 
I mean he said, look, LLMS are 

170
00:09:23,800 --> 00:09:26,200
currently trained on data that's
written in text. 

171
00:09:26,200 --> 00:09:29,600
If you would read it, it would 
take you 100,000 years, but it's

172
00:09:29,600 --> 00:09:34,040
the same amount of data in terms
of bytes as a kid from the age 

173
00:09:34,040 --> 00:09:39,240
of 0 to 4 takes in visually, so 
that visual is just a much more 

174
00:09:39,240 --> 00:09:44,520
efficient way to absorb data. 
Could be as long as the model 

175
00:09:44,520 --> 00:09:46,160
can critique the visual take, 
right? 

176
00:09:46,320 --> 00:09:49,000
As long as it observes its 
different object, how they're 

177
00:09:49,000 --> 00:09:52,680
interacting observation and you 
know changing the behaviour is 

178
00:09:53,480 --> 00:09:57,960
fundamental of the universe. 
My question, so if what if you 

179
00:09:57,960 --> 00:10:01,360
believe that what models are the
future and LMS are currently 

180
00:10:01,360 --> 00:10:04,960
going like it's like a one way 
street or it's just like at 

181
00:10:04,960 --> 00:10:08,800
least hitting the ceiling soon? 
Well, how we AI for AIG like 

182
00:10:09,240 --> 00:10:10,800
like super intelligence. 
OK. 

183
00:10:11,800 --> 00:10:14,480
But would you believe then then 
basically that right now with 

184
00:10:14,480 --> 00:10:17,800
the world model that we're right
now at starting at a lower point

185
00:10:17,800 --> 00:10:21,360
and currently LLMS like the 
current outstanding models are, 

186
00:10:21,640 --> 00:10:26,960
but then they have a higher 
curve of exponential curve right

187
00:10:26,960 --> 00:10:29,000
off the ceiling. 
So basically, I don't know, for 

188
00:10:29,000 --> 00:10:33,160
example, back in the day with 
Henry Ford with horses, right? 

189
00:10:33,160 --> 00:10:35,240
So basically back in the day, 
horses were so much more 

190
00:10:35,240 --> 00:10:39,120
reliable than the first car. 
But then of course, the ceiling,

191
00:10:39,120 --> 00:10:42,760
a horse can only, I don't know, 
run 30 miles an hour maximum, 

192
00:10:43,000 --> 00:10:47,120
whereas then the car is now 
what, 250 miles or faster? 

193
00:10:47,120 --> 00:10:49,840
So it's like a completely 
different ceiling, but it takes 

194
00:10:49,840 --> 00:10:52,600
time to mature. 
So it could be the same with 

195
00:10:52,600 --> 00:10:57,880
work models. 
You know, it's our reverse 

196
00:10:57,880 --> 00:11:01,200
question to ask is do humans 
have a ceiling on intelligence? 

197
00:11:01,200 --> 00:11:03,800
You know, when your two year old
or intelligence is pretty low 

198
00:11:04,240 --> 00:11:06,120
and then you keep growing 
intelligent and intelligent. 

199
00:11:06,120 --> 00:11:10,520
But at what point do we have a 
ceiling on our intelligence 

200
00:11:10,560 --> 00:11:12,200
using visual model? 
Is it 18? 

201
00:11:12,200 --> 00:11:13,640
Is it 21? 
Do we stop growing an 

202
00:11:13,640 --> 00:11:16,680
intelligence after 21? 
It's a very hard answer to, you 

203
00:11:16,680 --> 00:11:19,160
know, to know what is the limit 
to intelligence, to anything. 

204
00:11:21,160 --> 00:11:24,160
You know, that actually makes 
also I wonder myself sometimes 

205
00:11:24,160 --> 00:11:29,000
this have humans had the same IQ
over the last thousand years or 

206
00:11:29,000 --> 00:11:31,720
not? 
Is it just that we have like 

207
00:11:31,960 --> 00:11:36,120
written down all the learning so
we absorb it earlier the data 

208
00:11:36,120 --> 00:11:39,480
and that's why our IQ is the way
it is or be using the IQ? 

209
00:11:39,480 --> 00:11:41,160
Better. 
Yeah. 

210
00:11:41,560 --> 00:11:43,600
Yeah, it's a good question. 
You know, like we don't know the

211
00:11:43,600 --> 00:11:45,640
answer. 
Are we getting smarter? 

212
00:11:45,920 --> 00:11:49,200
Are we getting more data? 
Yeah, exactly. 

213
00:11:49,200 --> 00:11:51,760
A better dater. 
Also a better dater because like

214
00:11:52,000 --> 00:11:53,920
before I mean. 
Back in the day, just remember 

215
00:11:53,920 --> 00:11:56,040
when we were in in school, 
right, you always had to believe

216
00:11:56,040 --> 00:11:59,120
what the teacher said, but he 
was also just hopefully reading 

217
00:11:59,120 --> 00:12:01,920
it in one point in time. 
Sometimes just saying things to 

218
00:12:02,560 --> 00:12:05,600
like and you were no, he was 
like that didn't make sense. 

219
00:12:05,600 --> 00:12:07,880
Like we got us so often in 
trouble that like you're saying,

220
00:12:07,960 --> 00:12:11,720
but I'm sure that's true 
nowadays you can just ask HBT 

221
00:12:11,720 --> 00:12:14,240
and directly. 
No, it just reminds me of the I 

222
00:12:14,280 --> 00:12:17,680
don't know in your education, 
but during our education in. 

223
00:12:17,960 --> 00:12:22,480
Harmony In physics class we 
learned the how electrons work 

224
00:12:22,840 --> 00:12:26,240
and the atoms and the nucleus 
and all that stuff differently 

225
00:12:26,240 --> 00:12:27,720
than they actually work in real 
life. 

226
00:12:28,280 --> 00:12:31,240
It's super super interesting 
that because it was too complex 

227
00:12:31,240 --> 00:12:34,240
or something like but why are 
you teaching the kids then the 

228
00:12:34,240 --> 00:12:37,000
wrong one first if you know why 
you're not teaching the right 

229
00:12:37,000 --> 00:12:39,320
one when you're supposed to do 
it. 

230
00:12:39,920 --> 00:12:43,520
I was like, that's weird. 
Yeah, it depends on the teacher 

231
00:12:43,520 --> 00:12:45,680
too. 
Yeah, 100%. 

232
00:12:45,680 --> 00:12:47,800
I mean well, but this? 
Was this was everyone that had 

233
00:12:47,800 --> 00:12:49,200
to learn it this way, right? 
It was in the. 

234
00:12:49,200 --> 00:12:53,480
Books, Yeah, yeah, yeah, 
unfortunately at that time, you 

235
00:12:53,480 --> 00:12:57,000
know, like it's the method of 
teaching was so well, but if you

236
00:12:57,000 --> 00:13:00,240
look at the YouTube videos now, 
so interesting. 

237
00:13:00,840 --> 00:13:03,320
It's too much better. 
Yeah, I mean, there are. 

238
00:13:03,400 --> 00:13:07,160
The kids learning. 
That's the thing, but they have 

239
00:13:07,160 --> 00:13:09,160
the option at least have the 
opportunity. 

240
00:13:09,400 --> 00:13:11,960
Like there's there's something 
is a huge difference because 

241
00:13:12,720 --> 00:13:14,720
when we're still in high school,
like most of the videos were 

242
00:13:14,720 --> 00:13:17,760
just random stuff like on 
teaching. 

243
00:13:17,760 --> 00:13:20,400
And because there's more and 
more competition that gets more 

244
00:13:20,400 --> 00:13:23,640
and more professional. 
And now you can see like really 

245
00:13:24,040 --> 00:13:27,800
probably better education on 
YouTube, which is free than most

246
00:13:27,800 --> 00:13:31,480
courses that are being sold to 
people like oh, join my 

247
00:13:31,480 --> 00:13:34,520
whatever, scaling course, 
business course or whatever. 

248
00:13:34,520 --> 00:13:36,720
And these like just to make 
money. 

249
00:13:36,720 --> 00:13:39,800
Worse either offer it for free. 
I mean, I can't talk about kids,

250
00:13:39,800 --> 00:13:43,560
but myself, I know I learned a 
shit ton of YouTube the last 

251
00:13:43,560 --> 00:13:47,080
three years just learning how to
code, etcetera and all that. 

252
00:13:47,080 --> 00:13:49,600
You just so many tutorials are 
for free and it's incredible. 

253
00:13:49,840 --> 00:13:53,680
You can just directly copy or 
learn like step by step, follow 

254
00:13:53,680 --> 00:13:56,560
them along. 
Which again, like my IQ hasn't 

255
00:13:56,560 --> 00:13:59,640
increased, but I have the, the 
ability to absorb this 

256
00:13:59,640 --> 00:14:02,520
information because someone 
already did it, which is super, 

257
00:14:02,520 --> 00:14:06,480
super interesting. 
And I think I just do you know 

258
00:14:06,480 --> 00:14:09,160
the, do you follow the podcast 
from first principles? 

259
00:14:10,560 --> 00:14:13,440
Now what is it? 
So it's basically these two 

260
00:14:14,000 --> 00:14:17,760
physics nerds. 
They talk about new developments

261
00:14:17,760 --> 00:14:23,720
in science and AI, whatever is 
in basically physics based. 

262
00:14:24,040 --> 00:14:26,200
And they were just talking about
the one that episode I was 

263
00:14:26,200 --> 00:14:29,160
listening to about string 
theory. 

264
00:14:30,040 --> 00:14:33,400
And string theory has been 
apparently the the end of the 

265
00:14:33,480 --> 00:14:36,520
like the gets joked around like 
mess around a lot because 

266
00:14:36,760 --> 00:14:38,080
there's a real application to 
it. 

267
00:14:38,400 --> 00:14:41,200
But they just recently found out
that you can use string theory 

268
00:14:41,200 --> 00:14:47,360
to predict how tissue tissue 
grows properly. 

269
00:14:47,360 --> 00:14:50,640
And what's the most efficient 
way to have the surface area 

270
00:14:50,880 --> 00:14:53,920
always the smallest. 
So this could be used for skin 

271
00:14:53,920 --> 00:14:58,920
grafts, for predicting how 
fungus actually grow brain 

272
00:14:58,960 --> 00:15:02,720
neurons, etcetera. 
So that this model was basically

273
00:15:02,760 --> 00:15:06,600
this described everything else 
somewhere else can now be used 

274
00:15:06,760 --> 00:15:10,760
for something that wasn't even 
related to it or didn't think 

275
00:15:10,760 --> 00:15:13,600
about before. 
And I think these kind of steps 

276
00:15:13,600 --> 00:15:16,560
are super interesting. 
And I'm wondering, will AI with 

277
00:15:16,560 --> 00:15:20,200
LLMS be able to do this or do we
need a world model? 

278
00:15:20,200 --> 00:15:22,280
Then they're like look you know 
what, I'm going to try this on 

279
00:15:22,280 --> 00:15:26,520
this now and it works. 
LLMS will never be able to do 

280
00:15:26,520 --> 00:15:29,520
this. 
I mean, it's because it's not in

281
00:15:29,520 --> 00:15:32,560
the training data. 
It won't predict the future, 

282
00:15:33,040 --> 00:15:36,240
right? 
That's my opinion. 

283
00:15:36,240 --> 00:15:38,840
Although I've been very 
interested with how LLMS can do 

284
00:15:38,840 --> 00:15:40,320
some of the prediction in our 
data. 

285
00:15:40,320 --> 00:15:43,360
Yeah, using our data. 
But that is a kind of a ceiling 

286
00:15:43,360 --> 00:15:44,480
to it. 
You know it can look at the 

287
00:15:44,480 --> 00:15:46,040
past, apply something. 
Exactly. 

288
00:15:46,040 --> 00:15:48,240
That's, but that's my question 
because there's the expression, 

289
00:15:48,240 --> 00:15:49,800
right? 
Yeah, if you know the past, you 

290
00:15:49,800 --> 00:15:51,960
can predict the future because 
there are a lot of patterns 

291
00:15:51,960 --> 00:15:54,600
matching. 
So basically you see like we 

292
00:15:54,600 --> 00:15:57,160
always end cycles. 
So that's like for me. 

293
00:15:57,160 --> 00:15:58,920
Interesting. 
You see of course for maybe 

294
00:15:58,920 --> 00:16:04,920
discovery of new, I guess new 
medicines etcetera, it might be 

295
00:16:04,920 --> 00:16:07,360
harder. 
Oh, same with business ideas, 

296
00:16:07,360 --> 00:16:09,480
right. 
The problem is when you use it 

297
00:16:09,480 --> 00:16:14,160
or generate your new business 
ideas, it just you try it 

298
00:16:14,160 --> 00:16:15,920
ourselves. 
It just like sometimes just 

299
00:16:15,920 --> 00:16:20,800
generates random ideas, but it's
great at reproducing. 

300
00:16:20,800 --> 00:16:26,040
For example, maybe a help me 
draft a contract, a contract, 

301
00:16:26,360 --> 00:16:29,280
etcetera, help me review this 
contract, point out all the red 

302
00:16:29,280 --> 00:16:32,000
flags, etcetera. 
This is like fantastic. 

303
00:16:32,000 --> 00:16:35,680
It has so much data on it, which
makes it fantastic for 

304
00:16:36,160 --> 00:16:41,040
reproduction, but not great at, 
I guess the creation of or 

305
00:16:41,040 --> 00:16:42,160
invention of new things. 
Yeah. 

306
00:16:43,000 --> 00:16:45,760
Yeah, I guess we'll have to test
it out. 

307
00:16:45,760 --> 00:16:50,520
You know, the question to ask is
what has an LLM discovered or 

308
00:16:50,560 --> 00:16:56,640
invented in the last two years 
that we know we have ChatGPT? 

309
00:16:56,760 --> 00:16:58,280
What has it distorted our 
inventor? 

310
00:16:58,280 --> 00:17:01,840
You know that is a question. 
Do we know the answer to that? 

311
00:17:03,520 --> 00:17:05,200
How about we ask? 
Right now on the first, yeah. 

312
00:17:05,359 --> 00:17:08,839
Actually right now checking out,
I mean, I'm curious to see what 

313
00:17:08,839 --> 00:17:12,800
what actually would say. 
Nothing. 

314
00:17:16,079 --> 00:17:19,680
Think about open claws. 
Well, open claw went really, 

315
00:17:19,680 --> 00:17:22,599
really viral. 
Every you know, like grandmother

316
00:17:22,599 --> 00:17:26,079
was an open claw. 
And the question is who has 

317
00:17:26,079 --> 00:17:28,560
created a billion dollar 
industry out of open claw? 

318
00:17:28,560 --> 00:17:31,920
Who has launched a successful 
SAS using, you know, everyone is

319
00:17:31,920 --> 00:17:34,240
like, you know, like all the SAS
companies are gone. 

320
00:17:34,240 --> 00:17:36,840
SAS is outdated. 
You know, junior devils are out 

321
00:17:36,840 --> 00:17:38,760
of business. 
And I'm I'm sick of this 

322
00:17:38,760 --> 00:17:40,720
nonsense. 
You know, it's like who has 

323
00:17:40,720 --> 00:17:42,520
launched a successful startup 
using Open Claw. 

324
00:17:43,480 --> 00:17:46,400
But that's the thing who the 
only people who benefit from 

325
00:17:46,400 --> 00:17:49,280
these are those who are just 
content creators, because it's 

326
00:17:49,280 --> 00:17:52,200
clicky, right? 
If I say I replace these people,

327
00:17:52,200 --> 00:17:54,960
everybody clicks on it, 
everybody gets FOMO, tries it 

328
00:17:54,960 --> 00:17:58,360
out, might get hacked because 
they don't install it correctly,

329
00:17:58,360 --> 00:18:01,560
not properly. 
Now they have a security flaw, 

330
00:18:02,160 --> 00:18:04,040
etcetera. 
I mean, I've seen on Instagram, 

331
00:18:04,040 --> 00:18:06,640
I mean, I've watched a few 
videos myself and I was like, 

332
00:18:06,640 --> 00:18:09,920
damn, yeah, we implement this. 
And they say that's like, Oh, 

333
00:18:09,920 --> 00:18:11,880
no. 
Security risk is like also 

334
00:18:11,880 --> 00:18:14,680
another thing is just like, I 
think it instills a lot of fear 

335
00:18:14,680 --> 00:18:17,160
in people. 
For example, our roommate, he 

336
00:18:17,160 --> 00:18:20,480
has also business like Doc man, 
in 3-4 years, we're all going to

337
00:18:20,480 --> 00:18:23,400
be out of jobs because AI is 
going to take over all this 

338
00:18:23,400 --> 00:18:25,880
stuff like ecommerce. 
They're going to do everything. 

339
00:18:25,880 --> 00:18:29,000
I'm like man, and honestly I 
think AI is not capable of doing

340
00:18:29,000 --> 00:18:31,680
that first of all. 
And second of all agents like 

341
00:18:31,840 --> 00:18:34,240
agents are not that good like if
you not. 

342
00:18:34,240 --> 00:18:35,440
Yet. 
Not yet, yeah. 

343
00:18:35,480 --> 00:18:38,080
But they're not good yet. 
It's like open Claw is cool and 

344
00:18:38,080 --> 00:18:41,320
the concept is great, but the 
real implication is like, oh, 

345
00:18:41,320 --> 00:18:42,360
what are you doing? 
Is just. 

346
00:18:42,520 --> 00:18:44,760
It's not there yet. 
It's just will. 

347
00:18:44,760 --> 00:18:47,720
Take some time. 
Can they drive a car? 

348
00:18:47,720 --> 00:18:50,200
You know you? 
I think the 60 IQ is a minimum 

349
00:18:50,200 --> 00:18:50,960
to drive a car. 
Yeah. 

350
00:18:52,000 --> 00:18:56,840
And they can drive better than 
any of these AI latest AI models

351
00:18:57,000 --> 00:19:00,600
on Tesla. 
So the question to ask is if AI 

352
00:19:00,600 --> 00:19:03,160
is going to take over the world,
can it serve a coffee? 

353
00:19:03,360 --> 00:19:05,200
Can it drive a car, you know, 
better than a human? 

354
00:19:05,920 --> 00:19:08,880
And yeah, they're stuck on 
stupid things, you know, like 

355
00:19:08,880 --> 00:19:11,320
the two memo cars are coming 
down the street. 

356
00:19:11,920 --> 00:19:13,880
They're stuck with both, with 
each other for, you know, 

357
00:19:14,200 --> 00:19:16,120
minutes and minutes. 
They can't even think of moving 

358
00:19:16,440 --> 00:19:18,360
right. 
I heard that, I heard that 

359
00:19:18,360 --> 00:19:20,880
people are taking advantage of 
the situation with Vamos. 

360
00:19:20,880 --> 00:19:23,120
When they see a Vamo, they drive
more aggressive because they 

361
00:19:23,120 --> 00:19:25,520
know this Vamos gonna stop. 
It's gonna behave. 

362
00:19:25,520 --> 00:19:28,760
So they're gonna just like pass 
it and just like cut corners or 

363
00:19:28,760 --> 00:19:30,880
something because they're like, 
yeah, I have to stop. 

364
00:19:31,080 --> 00:19:32,000
So. 
Yeah. 

365
00:19:32,440 --> 00:19:35,680
And if they're off the training 
data, you know this driving self

366
00:19:35,680 --> 00:19:38,320
driving cars, they're not in the
driving, you know, within the 

367
00:19:38,320 --> 00:19:40,000
training data, they get stuck 
permanently. 

368
00:19:40,600 --> 00:19:42,440
We can't move out of it. 
But as a human would think 

369
00:19:42,440 --> 00:19:44,800
instinctively out. 
This is another wrong way how to

370
00:19:44,800 --> 00:19:47,360
make AU turn, how to go. 
You know, we plan and we 

371
00:19:47,400 --> 00:19:50,200
logically get out, even if it's 
not in our training data. 

372
00:19:50,440 --> 00:19:51,960
That's why we can drive in new 
cities. 

373
00:19:51,960 --> 00:19:56,120
You know, we move to a new city,
we instinctively pick up based 

374
00:19:56,120 --> 00:19:59,640
on this, you know, how to drive.
But a self driving car is like 

375
00:19:59,640 --> 00:20:01,160
that. 
We want to launch it this in 

376
00:20:01,160 --> 00:20:03,200
this city because we have 
trained it out this city. 

377
00:20:03,400 --> 00:20:08,160
You cannot go to the next city. 
I wonder how the conscious works

378
00:20:08,160 --> 00:20:11,840
with different layers and in 
terms of how we process data, 

379
00:20:11,840 --> 00:20:15,040
what's the data that we rely on?
For example, when you just said 

380
00:20:15,040 --> 00:20:18,520
that, I was thinking of if we 
went, we went to Australia and 

381
00:20:18,560 --> 00:20:21,480
you guys drive while they drive 
on the other side of the road, 

382
00:20:21,480 --> 00:20:24,400
you're like, yeah, yo, I need to
rethink this for a second. 

383
00:20:24,400 --> 00:20:28,000
It's like you need some time, 
like an hour or something or so 

384
00:20:28,080 --> 00:20:30,000
you're accustomed to like, OK, 
yeah. 

385
00:20:30,920 --> 00:20:32,920
This is odd. 
Also walking. 

386
00:20:32,920 --> 00:20:34,120
You walk on the other side 
right? 

387
00:20:34,120 --> 00:20:36,840
Because normally you're always 
on the right side, but then you 

388
00:20:36,840 --> 00:20:39,440
walk on the left side, so 
everything is reversed. 

389
00:20:39,440 --> 00:20:41,640
So it's like. 
But you get this used to it, 

390
00:20:41,640 --> 00:20:43,600
which is super interesting and 
of course because you go with 

391
00:20:43,600 --> 00:20:44,320
the flow. 
Right. 

392
00:20:44,320 --> 00:20:47,520
You're just like, if they're on 
a pedestrian walk, everybody 

393
00:20:47,520 --> 00:20:51,600
walks on the left side. 
So yes, you just follow the herd

394
00:20:51,600 --> 00:20:55,520
and just accept the rules that 
are put on like that are just 

395
00:20:55,520 --> 00:20:58,080
like displayed. 
Yeah, But the question is, like,

396
00:20:58,080 --> 00:21:00,400
for me, that's the conscious put
in a new container. 

397
00:21:00,400 --> 00:21:02,560
Like, look, you have all this 
information, I suppose this is 

398
00:21:02,680 --> 00:21:05,240
how everything works. 
And it's like, OK, this is how 

399
00:21:05,240 --> 00:21:07,040
it works in Europe. 
And then it puts a smaller 

400
00:21:07,040 --> 00:21:09,600
container. 
I'm currently in Australia, so 

401
00:21:09,600 --> 00:21:12,040
I'm need to rethink this. 
And this is a temporary 

402
00:21:12,040 --> 00:21:14,560
container where I'm in. 
And then when you go back, 

403
00:21:14,560 --> 00:21:16,480
you're like, OK, can I release 
the container? 

404
00:21:16,480 --> 00:21:18,840
And then you're like OK, and a 
container is gone again. 

405
00:21:19,680 --> 00:21:22,360
Yeah, that's interesting. 
You know like it's you have to 

406
00:21:22,360 --> 00:21:25,360
switch your context to the other
side of the road and we do it. 

407
00:21:25,360 --> 00:21:26,760
But can a self driving car do 
it? 

408
00:21:27,880 --> 00:21:30,400
I know I know way too little 
about this, but it's for me 

409
00:21:30,400 --> 00:21:33,080
interesting to see like just 
because like in the end, I 

410
00:21:33,080 --> 00:21:36,960
really believe that AIO the self
driving cars are out right now 

411
00:21:37,480 --> 00:21:44,120
are driving safer than most 
people assisted assisted like if

412
00:21:44,120 --> 00:21:47,760
I look at it, just man, people, 
humans, especially if in their 

413
00:21:47,760 --> 00:21:51,560
car, they especially like and 
you have to drive for every lane

414
00:21:51,560 --> 00:21:53,200
you can like. 
Distract every. 

415
00:21:53,200 --> 00:21:56,520
Lane so and then distracted you 
like you're not you're late to 

416
00:21:56,520 --> 00:21:59,280
work put up makeup or you're 
looking on the phone. 

417
00:21:59,280 --> 00:22:02,200
So you're very distracted so 
you're not very focused and that

418
00:22:02,240 --> 00:22:06,760
all together just combining man.
I just saw another video of a 

419
00:22:07,120 --> 00:22:13,120
Mercedes basically a car was he 
was driving 250 on the autobahn 

420
00:22:13,400 --> 00:22:18,600
and the guy was on the middle 
lane with 120 and going to the 

421
00:22:18,600 --> 00:22:20,560
left lane and was already far 
away. 

422
00:22:20,840 --> 00:22:26,360
But just in one second the car 
already passes 30 to 40 meters 

423
00:22:26,360 --> 00:22:28,400
in that time so you can't even 
react. 

424
00:22:28,400 --> 00:22:31,560
So you basically before you even
see he's pulling over you 

425
00:22:31,560 --> 00:22:33,600
already moved 30 or 40 meters 
ahead. 

426
00:22:33,840 --> 00:22:38,320
So like he he could barely like 
like pass or like dodge it. 

427
00:22:38,480 --> 00:22:40,640
Yeah, dodge it. 
But at the same time, just like 

428
00:22:40,640 --> 00:22:44,000
this human, the human eye is not
trained for like the speeds and 

429
00:22:44,000 --> 00:22:48,880
reaction time where AI or at 
least self driving cars can do 

430
00:22:48,880 --> 00:22:50,720
this. 
So I'm like, yes, that's a very 

431
00:22:50,720 --> 00:22:53,080
valid point for. 
You but this assist it's just 

432
00:22:53,080 --> 00:22:54,920
like a break assist right it's 
just physics. 

433
00:22:54,920 --> 00:22:57,680
It just does the math and like 
OK, Oh yeah not going to happen 

434
00:22:58,920 --> 00:23:00,920
but Ganesh let's let's switch 
gears again. 

435
00:23:00,920 --> 00:23:03,840
I want to hear more about Hell 0
AI, what you currently have, 

436
00:23:03,840 --> 00:23:08,560
what the vision is going to be, 
the way you going with this and 

437
00:23:08,880 --> 00:23:11,040
give us a lowdown please. 
Sure. 

438
00:23:11,040 --> 00:23:14,120
OK, so just before we wrap up 
the self driving thing, you know

439
00:23:14,120 --> 00:23:17,440
one interesting thing to talk 
about is 2 interesting things is

440
00:23:18,280 --> 00:23:21,560
humans have accident because 
they lose distraction, they lose

441
00:23:21,560 --> 00:23:23,280
focus. 
Now we are on the phone so we 

442
00:23:23,280 --> 00:23:26,560
stop thinking and machines never
stop thinking. 

443
00:23:26,960 --> 00:23:29,000
So that's the important 
distinct, you know to have. 

444
00:23:29,000 --> 00:23:31,520
It's not that, you know, 
machines are smarter, it's just 

445
00:23:31,520 --> 00:23:33,960
that we become stupider and 
that's why we have accident. 

446
00:23:34,400 --> 00:23:37,960
100% believe it. 
Second thing is when we see 

447
00:23:37,960 --> 00:23:42,400
Tesla's video, you know, like 
some of the what it thinks. 

448
00:23:42,880 --> 00:23:46,640
So one time there was a truck 
with a stop sign carrying it on 

449
00:23:46,640 --> 00:23:51,360
the back because it was a the 
fixing truck. 

450
00:23:52,200 --> 00:23:55,040
And Tesla kept on thinking 
there's a stop sign ahead of it.

451
00:23:55,040 --> 00:23:57,520
And then they kept on stopping 
because there was a stop sign at

452
00:23:57,520 --> 00:24:00,480
the back of the truck. 
You know, it, it didn't take, 

453
00:24:00,480 --> 00:24:03,280
you know, like even a 5 year old
could look at it and see, you 

454
00:24:03,280 --> 00:24:05,160
know, there's a stop sign on the
back of a truck so it doesn't 

455
00:24:05,160 --> 00:24:07,720
need to stop. 
But the Tesla but confused. 

456
00:24:07,720 --> 00:24:09,640
And the second thing was in a 
Crispus, you know, long 

457
00:24:09,640 --> 00:24:13,200
decoration on the trees, there 
were some red lanterns. 

458
00:24:13,800 --> 00:24:17,000
And Tesla kept on thinking it 
was a, you know, a signal. 

459
00:24:19,720 --> 00:24:21,840
Turn signal, yeah. 
And then you're like, yeah. 

460
00:24:22,880 --> 00:24:26,280
Yeah, so even a 5 year old would
know there's a Lantern on a 

461
00:24:26,280 --> 00:24:28,800
tree, you know, like why would a
traffic sign be on a? 

462
00:24:29,000 --> 00:24:31,800
So these are the common, you 
know, like stupid mistakes that 

463
00:24:31,800 --> 00:24:33,240
AI would do. 
But if it is really, really 

464
00:24:33,240 --> 00:24:37,840
intrigent on 120 IQ, which all 
you know the hype it up, it's 

465
00:24:37,840 --> 00:24:40,760
not behaving like that. 
Do you think then this is 

466
00:24:41,080 --> 00:24:43,840
because they're, I don't know 
sure if they're using LMS yet. 

467
00:24:43,840 --> 00:24:46,880
They have some machine learning,
but because Tesla is using 

468
00:24:47,040 --> 00:24:52,760
visual just camera wise as an 
input versus other companies are

469
00:24:52,760 --> 00:24:57,600
using Lidar, which has a 
different kind of I guess 

470
00:24:57,600 --> 00:25:01,840
information broadband that they 
can use and that they would 

471
00:25:01,840 --> 00:25:05,760
obviously if they combine these 
two would be much more efficient

472
00:25:05,760 --> 00:25:09,120
I guess or smarter. 
It to it, but that's just the 

473
00:25:09,120 --> 00:25:11,400
input data. 
What it thinks is, you know, is 

474
00:25:11,400 --> 00:25:13,360
different. 
So even with the input data, you

475
00:25:13,360 --> 00:25:16,520
know, if you have radar, if you 
have visual, which is again 

476
00:25:16,520 --> 00:25:19,440
debatable, I think even with 
visual, you should be able to do

477
00:25:19,440 --> 00:25:21,200
it. 
But the fact that it takes the 

478
00:25:21,200 --> 00:25:24,520
data and it thinks that there 
is, you know, a turn signal on 

479
00:25:24,520 --> 00:25:28,280
top of a tree all along the way 
is, you know, even if you had 

480
00:25:28,280 --> 00:25:30,680
radar, it's it's going to see a 
stop sign, it's going to see 

481
00:25:30,680 --> 00:25:31,880
something and it's going to 
take. 

482
00:25:32,320 --> 00:25:33,960
So the question is what does it 
think? 

483
00:25:34,360 --> 00:25:35,880
It's thinking something really 
absurd. 

484
00:25:37,600 --> 00:25:42,480
Makes sense? 
It doesn't question what it sees

485
00:25:42,720 --> 00:25:45,560
and it doesn't criticize what it
sees, and it doesn't, you know? 

486
00:25:47,440 --> 00:25:50,200
Do you think it would be quick 
enough to question? 

487
00:25:52,000 --> 00:25:54,240
It has to be quick enough. 
That's how intelligence can say,

488
00:25:54,240 --> 00:25:58,200
you know, when you criticize, it
breaks to a important 

489
00:25:58,200 --> 00:26:01,760
philosophical point. 
You know, like, to me, the mark 

490
00:26:01,760 --> 00:26:05,080
of intelligence is if you can 
question what you've been 

491
00:26:05,080 --> 00:26:06,920
taught, then you're truly 
intelligent. 

492
00:26:09,720 --> 00:26:13,040
That's why I mean, to be honest 
then most people are not 

493
00:26:13,040 --> 00:26:14,720
intelligent because. 
They're not. 

494
00:26:15,560 --> 00:26:18,360
The the majority don't question 
most of the things that are 

495
00:26:18,360 --> 00:26:20,560
being taught and it just 
accepted or they don't even want

496
00:26:20,560 --> 00:26:23,880
to think about it because in the
end it's, it would cost energy 

497
00:26:23,880 --> 00:26:27,080
to think about the decision or 
energy and the knowledge and 

498
00:26:27,080 --> 00:26:30,400
they just accept it. 
So I mean, that's the herd, 

499
00:26:30,400 --> 00:26:31,760
right? 
The herd just goes with 

500
00:26:31,760 --> 00:26:33,120
everything. 
They're not question it. 

501
00:26:33,320 --> 00:26:37,680
So I mean, we've seen it in the 
past, like history is the best 

502
00:26:37,680 --> 00:26:39,920
teacher. 
It just shows people just like 

503
00:26:39,920 --> 00:26:41,880
going with the flow because 
they're like, Nah, I mean. 

504
00:26:42,760 --> 00:26:45,280
He said it, so it must be right.
Exactly. 

505
00:26:45,280 --> 00:26:47,320
Yep, the government said it. 
It must be right. 

506
00:26:47,800 --> 00:26:51,280
I mean, yeah. 
Sorry, go ahead. 

507
00:26:52,080 --> 00:26:54,400
No, I was just saying basically 
you saw that also with the Max 

508
00:26:54,400 --> 00:26:57,800
vaccination for COVID. 
That's like a great example 

509
00:26:58,280 --> 00:27:01,200
because the government said, 
yes, it should be done, it's 

510
00:27:01,200 --> 00:27:03,320
better for everybody, even 
though there were not real 

511
00:27:03,320 --> 00:27:05,760
studies about it. 
So it's like a lot of these 

512
00:27:05,760 --> 00:27:08,400
things were just like you 
accepted because it's basically 

513
00:27:08,400 --> 00:27:12,640
somebody in power or at least 
with authority says it and then 

514
00:27:12,640 --> 00:27:16,160
at the the rest will just follow
instead of questioning it. 

515
00:27:16,160 --> 00:27:20,400
And those who are questioned 
tend to be tend to be either. 

516
00:27:20,520 --> 00:27:23,200
Well, luckily nowadays you have 
like more more ways to 

517
00:27:23,200 --> 00:27:27,160
distribute your opinions or like
social media or podcasts, 

518
00:27:27,160 --> 00:27:29,600
etcetera. 
But back in the day, there was 

519
00:27:29,600 --> 00:27:32,760
no option for them to actually 
question it out loud. 

520
00:27:33,640 --> 00:27:37,200
Yeah, and even if you do 
question it, you ought to paint 

521
00:27:37,200 --> 00:27:41,040
it as a crazy one. 
Yeah, exercise. 

522
00:27:41,040 --> 00:27:43,880
But The thing is, nowadays you 
have better ways to. 

523
00:27:43,880 --> 00:27:49,360
For example, thanks to Elamas 
buying Twitter, now you and have

524
00:27:49,360 --> 00:27:52,280
your opinion is there, right, 
You can you can debate whether 

525
00:27:52,280 --> 00:27:54,520
they should be restricted or 
not, but at least you can see 

526
00:27:54,520 --> 00:27:56,800
always both sides, which is like
a great part. 

527
00:27:57,000 --> 00:28:03,560
So like people have always a way
to say or explain themselves and

528
00:28:03,560 --> 00:28:07,200
then the entire public can judge
it whether that's correct or 

529
00:28:07,200 --> 00:28:08,480
not. 
Yeah. 

530
00:28:08,480 --> 00:28:10,600
Which I think is very important 
because otherwise you always 

531
00:28:10,600 --> 00:28:15,120
have the narrative of whoever is
winning and, or who has more 

532
00:28:15,120 --> 00:28:17,040
money or power. 
And it's, I think a very, very 

533
00:28:17,040 --> 00:28:19,320
great one. 
I wonder in general, how much of

534
00:28:19,320 --> 00:28:24,280
this is actually taken as 
training data for AI? 

535
00:28:25,960 --> 00:28:27,640
Because I would see a pattern, 
right? 

536
00:28:27,640 --> 00:28:29,240
It would see a. 
It would see a. 

537
00:28:34,880 --> 00:28:37,840
I'm curious to hear also what 
before we. 

538
00:28:38,760 --> 00:28:42,520
I know we got distracted here, 
but what's your opinion on 

539
00:28:44,640 --> 00:28:47,680
everyone having the, the ability
to say whatever they want 

540
00:28:48,000 --> 00:28:51,160
without repercussions or well, 
the freedom of speech end of the

541
00:28:51,160 --> 00:28:52,880
day? 
Because in recent years, 

542
00:28:52,880 --> 00:28:57,040
especially in Canada or in other
countries in the UK as well, we 

543
00:28:57,040 --> 00:29:01,800
have seen that this is not true 
anymore. 

544
00:29:01,800 --> 00:29:05,760
That whatever you can say, what 
you want to say is there is real

545
00:29:05,760 --> 00:29:08,960
repercussions if you don't 
follow the narrative. 

546
00:29:10,200 --> 00:29:12,160
For sure there are two cameras 
to this. 

547
00:29:12,160 --> 00:29:15,720
You know, 1 is the data that is 
generated could be good for the 

548
00:29:15,720 --> 00:29:18,040
AI because it gets an 
alternative viewpoint. 

549
00:29:19,320 --> 00:29:22,360
The second viewpoint is that we 
humans tend to have diverse 

550
00:29:22,360 --> 00:29:25,600
opinion on different things and 
we might be crazy in some aspect

551
00:29:25,600 --> 00:29:27,320
of the area. 
Like some, you know, we could be

552
00:29:27,720 --> 00:29:30,440
crazy when it comes to 
nationalism or religion and we 

553
00:29:30,440 --> 00:29:34,040
could be really, really saying, 
you know, when it comes to AI or

554
00:29:34,040 --> 00:29:36,280
physics or something like that. 
So we could be right in 

555
00:29:36,360 --> 00:29:38,000
different fields. 
We could be wrong or different 

556
00:29:38,000 --> 00:29:39,400
fields. 
We could be crazy in different 

557
00:29:39,400 --> 00:29:41,120
fields. 
There's a good thing about being

558
00:29:41,120 --> 00:29:43,960
humans, right? 
We're not universally uniform, 

559
00:29:44,440 --> 00:29:49,720
no talented in everything. 
So the question is like what? 

560
00:29:49,720 --> 00:29:52,720
What happens when we take this 
data and we train an AI on it? 

561
00:29:52,720 --> 00:29:55,440
You know, because diverse 
opinion on all of it, but it has

562
00:29:55,440 --> 00:29:58,560
to learn to critique which is 
right and what is wrong. 

563
00:29:59,320 --> 00:30:03,000
Yeah, I mean, that's one of the 
things I always see when I often

564
00:30:03,000 --> 00:30:05,040
see when I ask something touch 
routine. 

565
00:30:05,040 --> 00:30:06,680
I just. 
And it's like, oh, this is a 

566
00:30:06,680 --> 00:30:09,000
great idea. 
And I'm like, why you're not 

567
00:30:09,000 --> 00:30:13,120
criticizing at all? 
I mean, can you criticize her? 

568
00:30:13,120 --> 00:30:17,120
You're doing a great job. 
And I was like, please point out

569
00:30:17,120 --> 00:30:19,360
all the flaws I need. 
I don't need to hear like how 

570
00:30:19,360 --> 00:30:22,080
great I am because if it's in 
real, if real life doesn't work,

571
00:30:22,320 --> 00:30:25,640
this will not help me. 
I need you to be that. 

572
00:30:25,640 --> 00:30:29,640
My question is, though, did it 
learn this from human behavior? 

573
00:30:29,640 --> 00:30:33,440
Because most humans rather say, 
oh, it's hard for them to 

574
00:30:33,440 --> 00:30:36,960
criticize. 
Well, criticize friends, right? 

575
00:30:37,000 --> 00:30:40,400
It's like like, for example, 
when you a kid does something 

576
00:30:40,400 --> 00:30:42,440
great, you say, oh, good job, 
whatever. 

577
00:30:42,720 --> 00:30:46,920
Oh, even though it was nothing 
great or it didn't do a good 

578
00:30:46,920 --> 00:30:49,440
job. 
So I wonder whether this is a 

579
00:30:49,440 --> 00:30:52,320
training question. 
Like it basically learns so much

580
00:30:52,320 --> 00:30:56,400
on oh, you should not be giving 
direct feedback, etcetera. 

581
00:30:57,000 --> 00:30:58,800
Just a question I have for 
myself. 

582
00:31:01,280 --> 00:31:04,760
I think kids, you know, even 
when they're young, can revolt 

583
00:31:04,760 --> 00:31:08,640
against what we said, right, is 
tell them, you know, like this 

584
00:31:08,640 --> 00:31:10,760
is not good for you. 
They'll say I no, no, no, this 

585
00:31:10,760 --> 00:31:13,560
is good for, you know, if you 
don't want space, no, I want 

586
00:31:13,560 --> 00:31:15,240
space. 
You don't want, you know, screen

587
00:31:15,240 --> 00:31:16,360
time. 
No, I want screen time. 

588
00:31:16,360 --> 00:31:19,680
So they'll revolt against it. 
This revolt is kind of a markup 

589
00:31:19,680 --> 00:31:22,160
in intelligence. 
And I think we all humans have 

590
00:31:22,160 --> 00:31:26,000
it, which AI doesn't have. 
If you feed it something, it'll 

591
00:31:26,000 --> 00:31:28,280
always take it as a gospel for 
training data. 

592
00:31:28,440 --> 00:31:31,640
And because it has been fed so 
many times, it becomes the truth

593
00:31:31,640 --> 00:31:35,760
for it. 
Makes sense? 

594
00:31:36,800 --> 00:31:41,440
We have to keep the ability for 
a LLM or AI model to question 

595
00:31:41,440 --> 00:31:44,480
the training data that is given 
to it and assign a probability 

596
00:31:44,480 --> 00:31:51,320
that it is not quite are. 
You doing that already with the 

597
00:31:51,320 --> 00:31:54,680
Hell 0? 
Yes, and that is we know we what

598
00:31:54,680 --> 00:31:56,680
we wanted to talk about. 
Sorry you asked this question in

599
00:31:56,680 --> 00:31:57,640
a few minutes. 
Good. 

600
00:32:01,840 --> 00:32:05,160
When it comes to trading the Hal
0, there are two which I like 

601
00:32:05,160 --> 00:32:07,720
Yang Lukan's model mode, but it 
requires a huge amount of 

602
00:32:07,720 --> 00:32:09,920
investment. 
You know, it requires a physical

603
00:32:09,920 --> 00:32:11,520
world model. 
We might go into this, you know,

604
00:32:11,520 --> 00:32:15,360
once we have more bit of 
attraction and we discovered how

605
00:32:15,360 --> 00:32:19,440
we can do it with just text. 
We're building out a kind of a 

606
00:32:19,440 --> 00:32:22,640
proof of concept where, you 
know, this AI model can take a 

607
00:32:22,640 --> 00:32:27,120
text, can critique it, can, you 
know, absorb that knowledge, can

608
00:32:27,120 --> 00:32:30,840
criticize it, can go against it.
And then it has 0 hallucination 

609
00:32:30,840 --> 00:32:32,520
because we build it on top of 
the data. 

610
00:32:32,520 --> 00:32:35,880
It's just not the weights. 
That is the weights, but 

611
00:32:35,880 --> 00:32:39,040
underneath is the data. 
So it uses the weight to look at

612
00:32:39,040 --> 00:32:42,120
the data, pull it out, and then 
give the information. 

613
00:32:42,120 --> 00:32:45,080
In some ways it's like RAG, 
except this is on a latent 

614
00:32:45,240 --> 00:32:47,920
space. 
My question there, because I'm 

615
00:32:47,920 --> 00:32:52,440
thinking about it, can this be 
used also as a new training 

616
00:32:52,440 --> 00:32:56,160
method for the current LLMS? 
Or would that actually go 

617
00:32:56,160 --> 00:32:58,840
against it? 
Because all everything they have

618
00:32:58,840 --> 00:33:02,160
is basically in weights, right? 
Because for example, the 

619
00:33:02,160 --> 00:33:05,640
critique and the critiquing it 
would be super helpful. 

620
00:33:05,640 --> 00:33:07,520
And why are they currently not 
doing it? 

621
00:33:07,760 --> 00:33:10,560
Because I'm sure they're aware 
of this flaw. 

622
00:33:12,240 --> 00:33:15,360
Yep, we can't use LLM for it 
because that's how the 

623
00:33:15,360 --> 00:33:19,080
transformer models are, unless 
you put one more layer before it

624
00:33:19,120 --> 00:33:20,280
that you know kind of predicts 
it. 

625
00:33:20,280 --> 00:33:23,160
But the LLM, you know, the 
transformer model is not 

626
00:33:23,160 --> 00:33:25,880
inherently made for kind of a 
critique. 

627
00:33:25,880 --> 00:33:28,600
You know, it's just a vector in,
vector out. 

628
00:33:29,640 --> 00:33:32,840
And again, it comes back to the 
point that we need to radically 

629
00:33:32,840 --> 00:33:36,200
change that algorithm if we want
to have some kind of a 

630
00:33:36,400 --> 00:33:38,480
intelligence that can predict 
the input data. 

631
00:33:38,880 --> 00:33:42,280
And also when it looks up the 
data, it needs to verify if that

632
00:33:42,280 --> 00:33:46,320
is right or wrong. 
So you're not using Transformers

633
00:33:46,320 --> 00:33:47,760
then? 
What are you using? 

634
00:33:48,000 --> 00:33:50,120
No, we're using a different kind
of model. 

635
00:33:50,120 --> 00:33:53,360
We're using reinforcement 
learning and we're using a kind 

636
00:33:53,360 --> 00:33:57,800
of a diffusion model. 
Explain that to the to our 

637
00:33:57,800 --> 00:34:00,400
listeners please. 
What's the difference there and 

638
00:34:00,600 --> 00:34:03,080
how does that impact the 
outcome? 

639
00:34:04,400 --> 00:34:07,760
So diffusion model for text is 
where it has some kind of a 

640
00:34:07,760 --> 00:34:11,040
rough data on what it wants to 
talk about. 

641
00:34:12,080 --> 00:34:14,679
If I wait, for example, if I ask
you a question, you know 

642
00:34:14,679 --> 00:34:17,199
something that you're interested
to talk about like philosophy, 

643
00:34:17,800 --> 00:34:19,719
right? 
What do you think is the human 

644
00:34:19,719 --> 00:34:23,159
philosophy of who is successful?
You don't know the whole 

645
00:34:23,159 --> 00:34:26,199
sentence, but you have rough 
idea on what is right and what 

646
00:34:26,199 --> 00:34:28,440
is wrong. 
So you know a few topics, you 

647
00:34:28,440 --> 00:34:32,440
know that observation. 
We are the observer, we are part

648
00:34:32,440 --> 00:34:35,280
of the one universe. 
We have some key concept and 

649
00:34:35,280 --> 00:34:38,760
based on that you take the data 
and you build out your entire 

650
00:34:38,760 --> 00:34:40,239
answer. 
Right. 

651
00:34:40,760 --> 00:34:43,880
The adjectives, the nouns, the 
pronouns, they're all you know 

652
00:34:43,880 --> 00:34:47,480
built out later after you have 
this knowledge, and that's what 

653
00:34:47,480 --> 00:34:52,320
we're trying to build it up. 
And diffuses are commonly used 

654
00:34:52,320 --> 00:34:55,080
normally for image generation, 
right to use the pixels and then

655
00:34:55,080 --> 00:34:57,400
it diffuses that way. 
That's right. 

656
00:34:57,440 --> 00:35:00,400
Yeah. 
So the way we do it with text is

657
00:35:00,400 --> 00:35:03,440
we get the important data first.
We get what we need to talk 

658
00:35:03,440 --> 00:35:04,720
about. 
We need to go the latent 

659
00:35:04,720 --> 00:35:07,760
direction we need to talk about.
But those are just abstract 

660
00:35:07,760 --> 00:35:09,680
data. 
It cannot express itself. 

661
00:35:10,120 --> 00:35:12,800
And then we express ourselves 
using diffusion. 

662
00:35:12,800 --> 00:35:15,040
So we build the words that 
pronounce and everything around 

663
00:35:15,040 --> 00:35:18,640
it. 
So this is OK. 

664
00:35:18,640 --> 00:35:21,680
So the training data does 
obviously understand language 

665
00:35:21,680 --> 00:35:25,040
then. 
So it has some something already

666
00:35:25,040 --> 00:35:29,280
baked into it that are, I don't 
know language criterias for 

667
00:35:29,520 --> 00:35:35,520
English, Spanish, whatever it is
and or is it some form of a base

668
00:35:35,520 --> 00:35:39,120
model and then you retrain it? 
What's how does it work? 

669
00:35:40,160 --> 00:35:42,640
So it's a vector model. 
So what we have is a vector 

670
00:35:42,640 --> 00:35:46,800
model of different text, and if 
it's different languages, they 

671
00:35:46,840 --> 00:35:50,240
all combine to the same vector. 
So the advantage over LLM is an 

672
00:35:50,800 --> 00:35:54,720
LLM trained on English does not 
have the same intelligence. 

673
00:35:54,720 --> 00:35:56,320
That is an LLM trained on 
German. 

674
00:35:57,040 --> 00:36:00,280
If it has the same question in 
German and if it has the same 

675
00:36:00,280 --> 00:36:02,520
question in English, you could 
get probably two different 

676
00:36:02,520 --> 00:36:04,720
answer because it has two 
different knowledges. 

677
00:36:06,440 --> 00:36:09,320
OK, and for the vector base it 
is the same regardless of the 

678
00:36:09,320 --> 00:36:10,520
language. 
Yeah, yeah. 

679
00:36:11,440 --> 00:36:13,680
Because when you think of, you 
know, the lion is the king of 

680
00:36:13,680 --> 00:36:17,160
the the jungle, whether you 
express it in, you know, in 

681
00:36:17,160 --> 00:36:20,400
German or whether you express it
in English, it's the same 

682
00:36:20,400 --> 00:36:22,000
meaning. 
Yeah. 

683
00:36:22,080 --> 00:36:25,120
But if LLM has never been taught
that the lion is the king of the

684
00:36:25,120 --> 00:36:27,400
jungle in German, I don't know 
what you guys, you know, do it 

685
00:36:27,600 --> 00:36:29,760
in German. 
Who's the king of the jungle? 

686
00:36:30,360 --> 00:36:32,800
Let's assume in if German is 
giraffe is the king of the 

687
00:36:32,800 --> 00:36:38,080
jungle and then you ask it in 
German, you know, like who's the

688
00:36:38,080 --> 00:36:39,720
king of the jungle, then it'll 
say giraffe. 

689
00:36:39,720 --> 00:36:42,600
But if you offer to ask it in 
English, so the king, you know, 

690
00:36:43,280 --> 00:36:45,200
lion is the king of the jungle. 
So two different, you know, 

691
00:36:45,880 --> 00:36:46,960
that's. 
Why they didn't know this? 

692
00:36:46,960 --> 00:36:50,360
Yeah. 
So it's basically it's the best 

693
00:36:50,360 --> 00:36:54,160
in English, probably because 
most of the context is English 

694
00:36:54,360 --> 00:36:57,800
trained on. 
How do you guys talk to charity?

695
00:36:58,360 --> 00:37:04,000
English, Well, well, yes and no.
So mostly English, but I also 

696
00:37:04,000 --> 00:37:09,520
have German, but only German for
German Texas because I want, for

697
00:37:09,520 --> 00:37:11,880
example, if I have to do for a 
German entity that we still 

698
00:37:11,880 --> 00:37:15,040
have, I if I do do anything 
related to it because it already

699
00:37:15,040 --> 00:37:18,600
has everything there in German. 
But otherwise it's English or 

700
00:37:18,600 --> 00:37:20,720
Spanish text. 
Actually, if I do certain things

701
00:37:20,720 --> 00:37:25,640
in Spanish, but it's also like 
contextual, but mostly like 99% 

702
00:37:25,640 --> 00:37:28,440
of the time 95 is English 
because it's just like that's 

703
00:37:28,440 --> 00:37:31,840
how we like, it's everything 
work related, just easier. 

704
00:37:32,640 --> 00:37:35,000
And I was like, most of my 
knowledge is acquired in 

705
00:37:35,000 --> 00:37:37,480
English, not in German or 
Spanish. 

706
00:37:37,480 --> 00:37:39,880
So it's basically the similar to
an LLM. 

707
00:37:39,880 --> 00:37:41,840
I guess it's like mostly trade 
on English data. 

708
00:37:44,840 --> 00:37:46,440
It's. 
Probably how you express 

709
00:37:46,440 --> 00:37:49,200
yourself, you know, like, but 
your knowledge is carried over 

710
00:37:49,200 --> 00:37:52,520
to German or what you learned in
English, Yeah, it's also carried

711
00:37:52,520 --> 00:37:54,200
over to German. 
But with LLMS is not. 

712
00:37:54,200 --> 00:37:56,640
You know, what is learned in one
language is not carried over to 

713
00:37:56,640 --> 00:37:59,600
the other language. 
It does have a similarity 

714
00:37:59,600 --> 00:38:03,920
though, in the context of, I 
mean, people who speak multiple 

715
00:38:03,920 --> 00:38:05,680
language have noticed that that 
they have different 

716
00:38:05,680 --> 00:38:07,360
personalities and different 
languages. 

717
00:38:07,840 --> 00:38:11,800
And I don't wonder if LMS is 
that's also not personality 

718
00:38:11,800 --> 00:38:13,440
wise. 
They're different, but just like

719
00:38:13,440 --> 00:38:17,880
because they're trained in 
different, the languages are 

720
00:38:17,880 --> 00:38:20,640
this, the text that they're 
trained on is different. 

721
00:38:20,920 --> 00:38:25,640
They also act different the the 
way they I would say predict 

722
00:38:26,720 --> 00:38:28,360
because I don't know, you 
probably noticed it as well, 

723
00:38:28,360 --> 00:38:31,920
like a different language, 
different personality, the same 

724
00:38:31,920 --> 00:38:36,160
guy, same girl. 
I think it probably depends on 

725
00:38:36,280 --> 00:38:39,640
you know who you are surrounding
is because if you speak 

726
00:38:39,640 --> 00:38:41,920
different language, you are with
different people, so your 

727
00:38:41,920 --> 00:38:44,680
personality changes. 
I don't think it's because I 

728
00:38:44,680 --> 00:38:47,440
speak around 5 languages. 
I think 5 or 6, I don't even 

729
00:38:47,440 --> 00:38:51,520
remember. 
So I know inherently I'm saying,

730
00:38:51,720 --> 00:38:55,080
but the person that I'm talking 
to, the surroundings change. 

731
00:38:55,080 --> 00:38:58,160
So my personality changes. 
It's not because I'm speaking a 

732
00:38:58,160 --> 00:39:00,920
different languages, because the
people around me are different. 

733
00:39:01,800 --> 00:39:04,520
So you're adjusting to it. 
And I notice also the 

734
00:39:04,520 --> 00:39:08,680
expressions are different, the 
way I express, because the way I

735
00:39:08,680 --> 00:39:11,640
learn, for example in Spanish is
more expression like more 

736
00:39:11,640 --> 00:39:16,240
expressive than other languages 
than English, for example, that 

737
00:39:16,800 --> 00:39:20,640
you change modes because the way
you absorbed and then you're 

738
00:39:20,640 --> 00:39:23,440
obviously we see monkey see, 
monkey do. 

739
00:39:23,800 --> 00:39:28,560
And yeah. 
We reuse it, yeah. 

740
00:39:28,560 --> 00:39:32,480
So I actually, I actually have, 
I just Google it. 

741
00:39:32,480 --> 00:39:35,080
Why we behave differently or at 
least we believe it's different 

742
00:39:35,440 --> 00:39:38,840
is as you mentioned, we are put 
in a different cultural frame 

743
00:39:39,080 --> 00:39:42,440
because we are seeing. 
So maybe Spanish might be warmer

744
00:39:42,440 --> 00:39:47,120
or more touchy, etcetera. 
Then also emotional distance. 

745
00:39:47,120 --> 00:39:49,400
For example, your first language
you learned when you're a child,

746
00:39:49,560 --> 00:39:53,040
so you have childhood memories. 
So it's also impacts this while 

747
00:39:53,040 --> 00:39:56,040
as if you learn a different 
language later on, you're just 

748
00:39:56,040 --> 00:39:59,160
like having different memories 
which are related only in that 

749
00:39:59,160 --> 00:40:00,480
language. 
So that has. 

750
00:40:00,480 --> 00:40:04,000
So it's both. 
It's a mix of the environment. 

751
00:40:04,000 --> 00:40:06,680
So the cultures or people. 
If you talk with people, for 

752
00:40:06,680 --> 00:40:10,800
example, in Spanish who are 
warmer, then you will also 

753
00:40:10,800 --> 00:40:13,840
behave warmer because you're 
just like adjusting, right? 

754
00:40:13,840 --> 00:40:17,800
You're adjusting to that culture
as well as the memory that I 

755
00:40:17,800 --> 00:40:20,440
actually made in that language. 
Reinforce this, right? 

756
00:40:20,440 --> 00:40:22,960
It's also reinforcement learning
because in the end, when you 

757
00:40:22,960 --> 00:40:28,640
only talk with people who are. 
Warmer or more touchy etcetera 

758
00:40:28,640 --> 00:40:31,240
because of the culture you will 
directly associate with the 

759
00:40:31,240 --> 00:40:34,360
language and also behave in that
that way, which is reinforcing. 

760
00:40:35,400 --> 00:40:37,640
Yeah, interesting. 
I just thought it was 

761
00:40:37,840 --> 00:40:41,040
interesting. 
But it's funny how we become 

762
00:40:41,040 --> 00:40:44,520
different languages and how we 
switch personality, but 

763
00:40:44,800 --> 00:40:46,520
underneath the knowledge is 
still the same. 

764
00:40:46,520 --> 00:40:49,080
If you're noted in English, you 
can carry it over into the 

765
00:40:49,280 --> 00:40:50,200
other. 
Language, yeah, that's the 

766
00:40:50,200 --> 00:40:52,240
difference. 
I mean, that's why we are still 

767
00:40:52,920 --> 00:40:56,760
the apex on this world. 
We can, we can transfer all this

768
00:40:56,760 --> 00:41:00,920
information, but coming back to 
hell 0, what's the go to market 

769
00:41:00,920 --> 00:41:02,360
then? 
Because what you mentioned so 

770
00:41:02,360 --> 00:41:05,440
far is a lot of learning that 
needs to be done and data 

771
00:41:05,440 --> 00:41:09,480
modeling and and all that stuff.
So how's what's go to market? 

772
00:41:09,480 --> 00:41:12,640
And you said if it's going well 
later on, you can still expand 

773
00:41:12,640 --> 00:41:16,040
with the more the young, the kun
mod like work model style. 

774
00:41:16,240 --> 00:41:19,000
So what's the goal to market? 
When do you know you can switch 

775
00:41:19,000 --> 00:41:24,840
gears or the next steps? 
It's when we when we are 

776
00:41:24,840 --> 00:41:28,040
bootstrapping, it's a completely
different aspect from how you 

777
00:41:28,040 --> 00:41:29,600
know, if you raise a billion 
dollars in euro. 

778
00:41:30,880 --> 00:41:35,080
Especially like think about it 
bootstrapping, building a base 

779
00:41:35,080 --> 00:41:39,120
model basically. 
Yeah, it's, it's, it's a hard 

780
00:41:39,120 --> 00:41:41,640
job. 
And the number one priority, 

781
00:41:41,640 --> 00:41:44,560
whether we like it or not, is to
make sure we can monetize this 

782
00:41:44,560 --> 00:41:46,680
as soon as possible. 
So every step has to be 

783
00:41:46,680 --> 00:41:48,920
monetized. 
Are customers paying for it? 

784
00:41:48,920 --> 00:41:52,800
If they're not paying for it, it
doesn't make sense and I don't 

785
00:41:52,800 --> 00:41:56,040
have the energy or the knowledge
or the efficiency to go and 

786
00:41:56,040 --> 00:41:57,320
then, you know, raise money for 
it. 

787
00:41:58,720 --> 00:42:00,760
I don't even have the skills or 
the expertise for it. 

788
00:42:01,120 --> 00:42:03,560
So I have the skills and 
expertise to, you know, to 

789
00:42:03,560 --> 00:42:06,520
bootstrap it to, you know, to 
market it and to make money. 

790
00:42:06,880 --> 00:42:09,000
So why not do it, you know, and 
have fun along the way. 

791
00:42:09,400 --> 00:42:11,760
So to answer the question 
directly, I think the first 

792
00:42:11,760 --> 00:42:15,600
thing is we, we have their hello
where we do a lot of, you know, 

793
00:42:15,720 --> 00:42:17,360
we have our customers, we're 
growing well. 

794
00:42:17,760 --> 00:42:20,600
We plan to use this 0 
hallucination within our startup

795
00:42:20,600 --> 00:42:24,600
that is successful. 
Can you explain the user what AI

796
00:42:24,600 --> 00:42:26,960
Hello actually does? 
So because we didn't talk about 

797
00:42:26,960 --> 00:42:29,920
it at all yet and because I 
think it's a very, very cool 

798
00:42:29,920 --> 00:42:33,080
product and I'm sure all this 
was also loved hearing about it.

799
00:42:34,320 --> 00:42:38,200
So AI, hello, is AE commerce 
advertising platform and 

800
00:42:38,720 --> 00:42:40,720
specifically targeted towards 
Amazon. 

801
00:42:41,000 --> 00:42:44,040
We do a lot of bid optimization.
We do a lot of keyword research.

802
00:42:44,480 --> 00:42:48,680
There's departing, so there are 
multiple modules inside it where

803
00:42:48,680 --> 00:42:50,880
we automate it using our own AI 
model. 

804
00:42:50,880 --> 00:42:53,120
They are not LLM, they are the 
old style, you know, neural 

805
00:42:53,120 --> 00:42:56,600
networks, reinforcement 
learning, standard regression 

806
00:42:56,600 --> 00:42:59,400
statistics and those kind of AI 
models. 

807
00:42:59,840 --> 00:43:03,920
We have we have been growing, 
you know, around 20% year on. 

808
00:43:04,040 --> 00:43:06,680
This is a nice bootstrap is 
profitable. 

809
00:43:06,680 --> 00:43:08,520
We're around 36 employees right 
now. 

810
00:43:09,200 --> 00:43:13,040
The most of my time goes into a 
hello and we have a software 

811
00:43:13,040 --> 00:43:14,960
based as well. 
People can just sign up, you 

812
00:43:14,960 --> 00:43:17,480
know, set up the platform and do
automation. 

813
00:43:17,480 --> 00:43:20,560
And we also have, you know, 
guided where we have specialists

814
00:43:20,560 --> 00:43:23,360
that help them out. 
This is, you know, in progress. 

815
00:43:23,360 --> 00:43:27,480
This is this has been set in 
motion doesn't require much of 

816
00:43:27,480 --> 00:43:29,800
my input. 
I have senior people handling 

817
00:43:29,800 --> 00:43:32,040
it. 
So this gives me go to the next 

818
00:43:32,040 --> 00:43:36,560
stage, which is the AI and the, 
you know, e-mail and whatever 

819
00:43:36,560 --> 00:43:39,120
meeting we have, it collects all
of it. 

820
00:43:39,720 --> 00:43:42,800
And then with 0 Halo solution, 
it can give you context for each

821
00:43:42,800 --> 00:43:45,560
e-mail. 
So Hal 0 is going to take over 

822
00:43:45,560 --> 00:43:47,640
the e-mail, is going to take 
over the meeting notes. 

823
00:43:47,640 --> 00:43:50,440
It's going to take over what you
chat about and it's going to 

824
00:43:50,440 --> 00:43:53,680
bring the whole context into it.
And it can go and ask it a 

825
00:43:53,680 --> 00:43:56,920
question like, hey, I spoke to 
Mark, what do I talk about next?

826
00:43:57,520 --> 00:43:59,120
Then it can sit on top of the 
data. 

827
00:43:59,120 --> 00:44:00,840
It can give you a complete 
overview. 

828
00:44:02,120 --> 00:44:05,160
How do you ensure? 
I mean this sounds magical, I 

829
00:44:05,160 --> 00:44:08,280
love that already because that 
will make everything so much 

830
00:44:08,280 --> 00:44:11,720
faster and they can just give me
a TLDR afterwards so it would be

831
00:44:11,720 --> 00:44:14,840
fucking amazing. 
But how do you ensure that first

832
00:44:14,840 --> 00:44:19,560
of all my privacy is private, 
and the second of all that this 

833
00:44:19,560 --> 00:44:22,680
is in general safe, that no one 
else can hack it? 

834
00:44:23,880 --> 00:44:26,520
Yeah. 
So we built it on mail zero. 

835
00:44:26,520 --> 00:44:29,800
So we took a existing mail 
application that is really 

836
00:44:29,800 --> 00:44:32,680
secure and we loaded it on our 
server. 

837
00:44:32,680 --> 00:44:36,360
I think the big aspect is that 
when you use other note taker 

838
00:44:36,520 --> 00:44:40,400
tools, they invariably call up 
in our ChatGPT, they call up in 

839
00:44:40,400 --> 00:44:43,640
where the Gemini, they make API 
calls, summarize it. 

840
00:44:44,040 --> 00:44:46,560
We have our own internal model 
loaded on our GPU. 

841
00:44:46,560 --> 00:44:49,920
So your data never leaves us 
once we release it. 

842
00:44:49,920 --> 00:44:52,160
We're just using it internally. 
Once we release it, we'll be 

843
00:44:52,160 --> 00:44:55,240
writing on the privacy laws for 
it because we don't depend on 

844
00:44:55,240 --> 00:44:58,080
ads, we depend on you paying the
money for the subscription. 

845
00:44:58,440 --> 00:45:00,960
Make sure the data is just 
close, no one else. 

846
00:45:02,560 --> 00:45:05,160
Austin, do you already know the 
the plan pricing tiers? 

847
00:45:06,440 --> 00:45:08,800
We have a rough idea. 
We're going to release it to 

848
00:45:08,800 --> 00:45:12,280
some users and we're going to 
see this push back on the cost 

849
00:45:12,280 --> 00:45:14,160
of it and. 
Then I would love to try it 

850
00:45:14,440 --> 00:45:16,960
whenever you you're ready, 
please let us know. 

851
00:45:17,200 --> 00:45:21,120
Also, another question regarding
this is regarding the emails. 

852
00:45:21,120 --> 00:45:23,400
Could be then could would it be 
possible then? 

853
00:45:23,400 --> 00:45:27,080
For example, if everybody has AG
Suite, for example, would you be

854
00:45:27,080 --> 00:45:31,440
able to integrate that or would 
you need to create new accounts,

855
00:45:31,440 --> 00:45:33,560
new e-mail accounts which make 
it would make it harder? 

856
00:45:34,840 --> 00:45:36,840
We right now integrate just with
Google. 

857
00:45:36,840 --> 00:45:40,080
So everything is yeah, perfect. 
Yeah, yeah. 

858
00:45:40,320 --> 00:45:43,280
So it makes it easier, right? 
It's you don't have to create 

859
00:45:43,280 --> 00:45:45,240
new e-mail for it. 
You just click login by Google 

860
00:45:45,360 --> 00:45:48,040
and you connect to calendar. 
Everything is handled 30 

861
00:45:48,040 --> 00:45:51,120
seconds. 
Is it also the website or do you

862
00:45:51,120 --> 00:45:52,720
guys have an app? 
Is it also on the phone? 

863
00:45:52,720 --> 00:45:57,760
Is it a mobile app as well or? 
It is just a new website right 

864
00:45:57,760 --> 00:46:00,400
now. 
If we don't have plans to make a

865
00:46:00,400 --> 00:46:03,360
mobile app, I think once I get 
10 users, then I might have the 

866
00:46:03,360 --> 00:46:05,240
second path. 
The 1st is to get 10 paying 

867
00:46:05,240 --> 00:46:07,200
users and then go for the next 
step. 

868
00:46:07,200 --> 00:46:09,400
Then you know, 50 users and then
go for the next step. 

869
00:46:09,400 --> 00:46:11,320
Everything is as I said, 
bootstrap. 

870
00:46:11,320 --> 00:46:13,080
Everything is how many people 
are paying but. 

871
00:46:14,600 --> 00:46:17,640
To be honest, that's I think one
of the most important parts, 

872
00:46:17,640 --> 00:46:22,840
especially if you're going now 
towards more, I guess not an 

873
00:46:22,840 --> 00:46:24,960
economy boom, but more like a 
recession. 

874
00:46:24,960 --> 00:46:29,560
That's so important to think 
about bootstrapping, getting 

875
00:46:29,560 --> 00:46:32,680
money for some clients because 
like there's always when there's

876
00:46:32,680 --> 00:46:36,840
a lot of cash out there and a 
lot of investment start-ups just

877
00:46:36,840 --> 00:46:40,560
like blow through money. 
But then when it's going down 

878
00:46:40,560 --> 00:46:44,680
again, a normal cycle, right 
then people are here 

879
00:46:44,680 --> 00:46:47,240
profitability first. 
And I believe really believe you

880
00:46:47,240 --> 00:46:53,200
should from day one think about 
not building something like a 

881
00:46:54,560 --> 00:46:57,320
something where you need a lot 
of investment, right And you 

882
00:46:57,320 --> 00:47:00,360
have a longer time horizon. 
Otherwise I would like 

883
00:47:00,440 --> 00:47:04,280
bootstrapping works. 
Yeah, yeah. 

884
00:47:04,440 --> 00:47:05,880
Default alive. 
You're making money. 

885
00:47:05,880 --> 00:47:07,840
Are you going to? 
Yeah. 

886
00:47:08,360 --> 00:47:10,960
And take a stab and tell 0 the 
main time, yeah? 

887
00:47:11,240 --> 00:47:13,800
Hell yeah. 
No, I agree. 

888
00:47:13,800 --> 00:47:16,160
Agree. 
OK, you already have 

889
00:47:16,160 --> 00:47:18,760
bootstrapped hell zero and hey 
AI hello. 

890
00:47:19,040 --> 00:47:21,120
The question is for me. 
They already have a big heck. 

891
00:47:21,120 --> 00:47:23,880
Well, like I think it's a big 
head count 30 something 

892
00:47:24,440 --> 00:47:26,320
employees. 
Especially with AI, because now 

893
00:47:26,320 --> 00:47:31,040
they're with that kind of this 
amount of employees, you can 

894
00:47:32,000 --> 00:47:36,040
have a lot of revenue because 
like with AI, if they're really 

895
00:47:36,040 --> 00:47:39,600
like supporting. 
So my my question is how do you 

896
00:47:39,720 --> 00:47:43,800
see strategically a go to market
strategy in terms of what are 

897
00:47:43,800 --> 00:47:46,360
the channels that you focus on 
because I think that everyone 

898
00:47:46,360 --> 00:47:49,280
has their own expertise and what
do you bank on? 

899
00:47:49,520 --> 00:47:53,200
Are you also still, how do you 
feel about influencer marketing?

900
00:47:54,000 --> 00:47:57,600
What are these takes like B to B
like in general, like how how do

901
00:47:57,600 --> 00:48:00,720
you see the whole function the 
stack right now, especially for 

902
00:48:00,720 --> 00:48:02,280
your different, different kind 
of products? 

903
00:48:02,280 --> 00:48:06,720
I guess the e-mail is direct 
directly to consumers versus low

904
00:48:06,720 --> 00:48:09,520
is to businesses. 
How do you see that for from the

905
00:48:09,520 --> 00:48:12,040
B to B and detect consumer? 
What's your stack? 

906
00:48:13,280 --> 00:48:14,440
Right. 
So as I mentioned, you know, 

907
00:48:14,440 --> 00:48:19,400
like raising money and bootstrap
is completely different and in 

908
00:48:19,400 --> 00:48:21,960
the same way B to B is 
completely different from B to 

909
00:48:21,960 --> 00:48:23,600
C. 
So air hello is B to B. 

910
00:48:23,720 --> 00:48:27,080
The e-mail thing would be B to C
radically different approach. 

911
00:48:27,680 --> 00:48:29,560
Influenza marketing is really 
exciting. 

912
00:48:29,560 --> 00:48:33,320
I never managed to get it 
working though, but I'm, I'm, 

913
00:48:33,720 --> 00:48:35,440
I'm confident, you know, it does
work. 

914
00:48:36,200 --> 00:48:38,360
I haven't managed to make it 
work in the meantime, you know, 

915
00:48:38,360 --> 00:48:41,840
running ads, marketing, social 
media marketing is the one that 

916
00:48:41,840 --> 00:48:44,360
it works. 
But for B2C, you know what I do 

917
00:48:44,360 --> 00:48:47,480
is I go to LinkedIn and I reach 
out to people and say, hey, I'm 

918
00:48:47,480 --> 00:48:50,760
testing out a free tool. 
I'm giving it away for free just

919
00:48:50,760 --> 00:48:53,080
for the 10, first 10 users. 
After that I'll charge. 

920
00:48:53,440 --> 00:48:55,840
Would you be willing to test it 
out, get the feedback from them,

921
00:48:56,160 --> 00:48:58,760
improve on it, and then go to 
the next 10 users? 

922
00:48:58,760 --> 00:49:02,600
I'm giving away at 90% discount.
We want to test it out, then 80%

923
00:49:02,600 --> 00:49:03,840
discount. 
That's 90%. 

924
00:49:03,840 --> 00:49:06,440
It's one way. 
Just reach out to your contacts.

925
00:49:07,320 --> 00:49:09,200
That is so cool never thought 
about this. 

926
00:49:09,200 --> 00:49:13,840
Always like giving a temp like 
9% discount, 10% going down. 

927
00:49:14,280 --> 00:49:16,240
I really really liked it. 
Never thought about this. 

928
00:49:16,360 --> 00:49:18,680
So it's the classical do things 
that don't scale. 

929
00:49:18,680 --> 00:49:22,360
It's manual, reach out. 
Yes, the beginning, yeah. 

930
00:49:23,360 --> 00:49:26,480
Oh, you get the feedback right 
if it is working or not working.

931
00:49:27,400 --> 00:49:29,680
Yeah, that's true. 
Like because we're we're with 

932
00:49:29,680 --> 00:49:33,920
our new project also we are, I 
mean we have Bootstrap as well. 

933
00:49:34,480 --> 00:49:37,680
So we did. 
Reach out to our contacts, try 

934
00:49:37,680 --> 00:49:41,000
it out. 
And we also running some ads 

935
00:49:41,000 --> 00:49:42,400
right now. 
I mean we're doing also things 

936
00:49:42,400 --> 00:49:46,320
that not scaling literally our 
plans are like starting AT150250

937
00:49:46,320 --> 00:49:51,680
and €450 per month. 
And for example, for 250 plan, I

938
00:49:51,800 --> 00:49:55,120
just worked with the lady for, I
was on the call with her for 

939
00:49:55,120 --> 00:49:58,520
over 7 hours. 
So it's like, think about it for

940
00:49:58,520 --> 00:50:03,480
250 plan, which is should be all
they can serve themselves. 

941
00:50:03,840 --> 00:50:07,120
But like the onboarding, but 
it's a lot of feedback. 

942
00:50:07,120 --> 00:50:09,080
I think these are things that I 
don't scale. 

943
00:50:09,080 --> 00:50:12,440
But in the end, you learn so 
much during the phone call what 

944
00:50:12,440 --> 00:50:15,480
is not working or what can be 
approved because you need to 

945
00:50:15,520 --> 00:50:19,520
shut yourself up to like try, 
see them fail like in the 

946
00:50:19,520 --> 00:50:23,120
process because you are so an 
expert of your own product 

947
00:50:23,120 --> 00:50:24,680
because you have used it so 
much. 

948
00:50:24,800 --> 00:50:28,520
So you know all the kinks and 
just know how to ignore them. 

949
00:50:28,760 --> 00:50:31,280
When you see somebody else 
struggle for the first time, you

950
00:50:31,280 --> 00:50:33,720
learn so much about it. 
So it's at the beginning. 

951
00:50:33,720 --> 00:50:37,040
It is worth doing things that 
don't scale. 

952
00:50:37,800 --> 00:50:40,520
Yeah, yeah, It's the one way to 
do it actually. 

953
00:50:40,520 --> 00:50:42,920
You can't just put on ads and 
then forget about it. 

954
00:50:43,280 --> 00:50:45,120
You don't get the feedback. 
Yeah. 

955
00:50:46,240 --> 00:50:48,240
I agree. 
That's that's I think super 

956
00:50:48,240 --> 00:50:50,360
important. 
I think this is nowadays because

957
00:50:50,360 --> 00:50:53,400
everything is so easy, just 
generating ads or just popping 

958
00:50:53,400 --> 00:50:55,560
something out there. 
It's like, hey, check this out 

959
00:50:55,640 --> 00:50:57,800
and then you get feedback or 
don't get feedback. 

960
00:50:57,800 --> 00:51:00,080
It's just like overwhelming. 
It's like if you go on tech 

961
00:51:00,080 --> 00:51:02,800
Twitter, you see so many 
releases and there's like no one

962
00:51:02,800 --> 00:51:06,320
trying it out or it's just it's 
overwhelming or what is it? 

963
00:51:06,560 --> 00:51:08,280
What's the other one called? 
The launching 1? 

964
00:51:09,200 --> 00:51:11,600
What Product hunt. 
Product hunt, Yeah. 

965
00:51:11,800 --> 00:51:13,600
Exactly. 
I mean, there's there's a whole 

966
00:51:13,600 --> 00:51:14,880
new. 
Have you ever tried it? 

967
00:51:16,200 --> 00:51:17,600
Yes, I used to be on Product 
Hunt. 

968
00:51:17,600 --> 00:51:19,320
We did try to launch it in 
Product Hunt. 

969
00:51:19,320 --> 00:51:21,000
It didn't work out. 
I think it depends on who 

970
00:51:21,000 --> 00:51:22,720
launches it for you. 
Yeah, I'd be. 

971
00:51:22,720 --> 00:51:25,360
We also tried it back in the day
like when we started our first 

972
00:51:25,360 --> 00:51:30,560
gaming startup Man we we just 
failed at we did not the. 

973
00:51:30,560 --> 00:51:32,920
Strategy either like it's like a
whole strategy. 

974
00:51:32,920 --> 00:51:35,720
You have like specific hunters, 
you need to have the right 

975
00:51:35,720 --> 00:51:38,440
hunters for you and then you 
need to incentivize and probably

976
00:51:38,520 --> 00:51:41,080
it's a very complex game. 
I mean, for our last business, 

977
00:51:41,080 --> 00:51:43,120
we for example, we ran only on 
affiliates. 

978
00:51:43,400 --> 00:51:46,320
We just had affiliates that did 
it for us and was great. 

979
00:51:46,600 --> 00:51:50,280
That was that was a good thing. 
But otherwise it's just you got 

980
00:51:50,280 --> 00:51:52,560
to find something that works for
you and then go with it. 

981
00:51:53,400 --> 00:51:55,800
Yeah, like it's just sad. 
You got to reach out and just 

982
00:51:55,800 --> 00:51:58,640
like, I think, I mean, I noticed
myself as well. 

983
00:51:58,640 --> 00:52:00,440
I'm getting lazy, man. 
They're reaching out. 

984
00:52:00,440 --> 00:52:02,760
I'm like, I don't want to do it.
But you're right. 

985
00:52:02,760 --> 00:52:05,400
It's the best thing to do it. 
Reach out to your network first.

986
00:52:06,160 --> 00:52:08,880
Yeah, initial stages at least. 
After that, you know. 

987
00:52:08,880 --> 00:52:11,800
You 100% because later you need 
to automate it, you need to find

988
00:52:11,800 --> 00:52:14,200
ways how you can scale it 
because at the beginning, like 

989
00:52:14,240 --> 00:52:16,280
once you have the kinks out, the
product is good. 

990
00:52:16,280 --> 00:52:19,360
I mean, there's also one of the 
differences between a simple 

991
00:52:19,360 --> 00:52:23,240
product is right? 
Like, for example, super thing, 

992
00:52:23,240 --> 00:52:27,560
you have a you you sell candy, 
OK, you sell candy, you don't 

993
00:52:27,840 --> 00:52:30,640
once you have to flavor out, you
don't need to customer feedback 

994
00:52:30,640 --> 00:52:32,160
anymore. 
Like you can just run ass and 

995
00:52:32,160 --> 00:52:35,040
they buy it and if they like 
they will purchase it after you 

996
00:52:35,040 --> 00:52:37,680
have some. 
Well, as you have a product, 

997
00:52:37,680 --> 00:52:41,160
like a system, like a software 
product, which is more complex, 

998
00:52:41,800 --> 00:52:43,960
it takes way more time to get it
right. 

999
00:52:44,480 --> 00:52:49,920
Like, yeah, like so many moving 
parts and and I think there's 

1000
00:52:49,920 --> 00:52:54,120
like makes a huge difference. 
Yeah, It's easy to create 

1001
00:52:54,120 --> 00:52:55,960
software now. 
It's harder to market it. 

1002
00:52:56,040 --> 00:52:58,760
Earlier it was, you know, easier
to market and hard to pay. 

1003
00:52:59,520 --> 00:53:02,400
Yeah. 
We've been saying for the last, 

1004
00:53:02,400 --> 00:53:05,080
I guess probably almost two 
years now, that distribution is 

1005
00:53:05,080 --> 00:53:08,640
king nowadays because yes, 
creation is so simple. 

1006
00:53:08,640 --> 00:53:11,720
Everybody can just have a 
problem to create an entire 

1007
00:53:11,720 --> 00:53:14,160
website. 
But The thing is the 

1008
00:53:14,160 --> 00:53:17,840
distribution just it's so 
important. 

1009
00:53:18,120 --> 00:53:22,280
That's why nowadays you see more
and more influencer who become 

1010
00:53:22,280 --> 00:53:27,040
part owners of start-ups, maybe 
tech start-ups or product, 

1011
00:53:27,400 --> 00:53:31,200
because they have the 
distribution and they can for 

1012
00:53:31,200 --> 00:53:36,480
example, for example, MQPHD, 
he's like a huge YouTube on 

1013
00:53:36,480 --> 00:53:38,760
tech. 
I think he has like 15,000,000 

1014
00:53:38,760 --> 00:53:42,800
subscribers and he became part 
owner of which wallets? 

1015
00:53:43,480 --> 00:53:46,720
Like. 
So which is like crazy to see 

1016
00:53:46,720 --> 00:53:49,000
that so late. 
They already made like hundreds 

1017
00:53:49,000 --> 00:53:50,920
of millions of dollars, but it's
so late. 

1018
00:53:50,920 --> 00:53:54,280
He they still got him on board 
because of the distribution, 

1019
00:53:54,520 --> 00:53:57,720
because it makes sense for him 
to have him there because his 

1020
00:53:58,480 --> 00:54:01,760
people who watch it have so much
buying power and they can 

1021
00:54:01,760 --> 00:54:03,400
directly promote their products 
there. 

1022
00:54:04,200 --> 00:54:06,360
Yeah, that's something on here 
to learn as well. 

1023
00:54:07,320 --> 00:54:08,800
Same. 
I mean, that's The thing is that

1024
00:54:08,800 --> 00:54:10,840
that's something, but it's good 
to know. 

1025
00:54:10,840 --> 00:54:13,720
That's why I also like talking 
to different people, so other 

1026
00:54:13,720 --> 00:54:16,320
entrepreneurs like yourself, 
because you're seeing different 

1027
00:54:16,320 --> 00:54:21,640
angles and it opens you up to 
explore in that direction, which

1028
00:54:21,640 --> 00:54:24,440
are really, really like because 
The thing is at the gate right 

1029
00:54:24,440 --> 00:54:29,800
now, it being a journalist can 
be quite helpful because thanks 

1030
00:54:29,800 --> 00:54:33,480
to AI, you can be better in that
certain point as long as you 

1031
00:54:33,480 --> 00:54:37,160
know it exists. 
Yeah, for sure. 

1032
00:54:37,280 --> 00:54:39,040
Yeah. 
It's all about, you know, 

1033
00:54:39,440 --> 00:54:42,240
reaching out and distribution. 
And marketing 100%. 

1034
00:54:43,080 --> 00:54:47,320
Do you think now that we are in,
I mean it was already the last 

1035
00:54:47,320 --> 00:54:51,480
ticket, the era of social media 
and that this is even becoming 

1036
00:54:51,480 --> 00:54:53,960
more and more important because 
it's one of the biggest 

1037
00:54:54,400 --> 00:54:57,360
distribution factor right now in
terms of ads. 

1038
00:54:57,360 --> 00:54:59,400
Do you think it's going to scale
further? 

1039
00:55:00,360 --> 00:55:03,040
You know what, I've 
interestingly seen that a lot of

1040
00:55:03,040 --> 00:55:07,280
the social media is now become 
an SCO graveyard with LLM taking

1041
00:55:07,280 --> 00:55:09,480
over. 
If you go to Twitter and I click

1042
00:55:09,480 --> 00:55:13,200
on an interesting topic, almost 
all the replies, I can smell it 

1043
00:55:13,200 --> 00:55:14,840
out. 
That is chatty video response. 

1044
00:55:15,080 --> 00:55:19,800
It's oh wow. 
So counting to do. 

1045
00:55:19,880 --> 00:55:21,240
I think the future is going to 
be here. 

1046
00:55:21,240 --> 00:55:24,480
Social media where there's only 
humans allowed and no, no robots

1047
00:55:24,480 --> 00:55:28,160
allowed that if someone builds 
it out and you get who is just 

1048
00:55:28,160 --> 00:55:31,600
human, it's. 
Probably just very hard, which 

1049
00:55:31,600 --> 00:55:33,800
is like very hard to prove. 
That's one of the things, 

1050
00:55:33,800 --> 00:55:37,400
especially with automation, 
which I'm also the. 

1051
00:55:37,600 --> 00:55:40,840
Yeah, it's. 
Very I agree. 

1052
00:55:41,000 --> 00:55:44,040
I mean this also what the Trump,
sorry, you know, Musk said right

1053
00:55:44,040 --> 00:55:46,480
when he went like took over 
tours, like there's so many 

1054
00:55:46,480 --> 00:55:48,480
freaking bots on here. 
What's going on? 

1055
00:55:48,800 --> 00:55:52,520
Get rid of the bots. 
It's it's not the things. 

1056
00:55:52,520 --> 00:55:56,840
It's a big problem for sure. 
And yeah. 

1057
00:55:58,120 --> 00:56:01,920
One way to resolve this would be
you can onely send 160 

1058
00:56:01,920 --> 00:56:04,280
characters, but you can onely 
send one tweet at it. 

1059
00:56:04,680 --> 00:56:07,400
So think about it. 
Or one tweet or one. 

1060
00:56:08,160 --> 00:56:09,880
But then they create multiple 
accounts, right? 

1061
00:56:09,880 --> 00:56:12,920
So they have like a million 
accounts that only tweet once 

1062
00:56:12,920 --> 00:56:16,440
per day, which makes it. 
And also at the same time, it's 

1063
00:56:16,440 --> 00:56:19,120
interesting for me to see where 
this all is going. 

1064
00:56:19,320 --> 00:56:24,560
But in in general, especially 
with now glasses being basically

1065
00:56:25,320 --> 00:56:27,560
like smart glasses. 
So basically you will have your 

1066
00:56:27,560 --> 00:56:29,640
screen directly always in front 
of your eyes. 

1067
00:56:29,960 --> 00:56:34,640
And what it also means how 
content is being consumed. 

1068
00:56:34,640 --> 00:56:35,840
Yeah. 
Yeah. 

1069
00:56:36,840 --> 00:56:39,600
Makes it completely different. 
Is audio the next thing? 

1070
00:56:39,760 --> 00:56:41,160
Who knows? 
That's something that's for me 

1071
00:56:41,160 --> 00:56:45,760
very, very interesting to see. 
Yeah, it's video probably. 

1072
00:56:45,840 --> 00:56:48,160
And 24 by 7 video with smart 
glasses. 

1073
00:56:48,280 --> 00:56:50,200
Yeah. 
Yeah, that's crazy. 

1074
00:56:50,200 --> 00:56:51,600
It's going to be a new 
interface. 

1075
00:56:51,680 --> 00:56:54,280
It's it's it's great. 
But at the end of the day, smart

1076
00:56:54,280 --> 00:56:56,680
glasses the same as a phone, 
just a different spot. 

1077
00:56:57,000 --> 00:56:57,840
You know what I mean? 
Yes. 

1078
00:56:58,200 --> 00:57:02,560
It's nothing and nothing crazy 
new That's also I didn't 

1079
00:57:02,560 --> 00:57:06,360
understand the complete hype 
about the the VR glasses AR 

1080
00:57:06,360 --> 00:57:10,760
glasses just like it's the same.
It's just they the well, the 

1081
00:57:10,760 --> 00:57:15,640
phone is already so fucking cool
and optimized for so many things

1082
00:57:16,040 --> 00:57:18,400
that it's a. 
Future you won't. 

1083
00:57:18,400 --> 00:57:20,280
You won't even have to take it 
in your hands, which is like 

1084
00:57:20,400 --> 00:57:24,040
honesty 5 to 10 years. 
But it's dangerous because then 

1085
00:57:24,040 --> 00:57:26,920
you can't even put it away. 
Oh no, 100% that's the thing, 

1086
00:57:27,120 --> 00:57:29,880
but that's the thing. 
That's where that's why you see 

1087
00:57:29,880 --> 00:57:32,720
these social media giants 
building these because in the 

1088
00:57:32,720 --> 00:57:36,160
end you want to spend more time 
on the platform and that's also 

1089
00:57:36,160 --> 00:57:37,480
why you don't want to be late in
it. 

1090
00:57:37,480 --> 00:57:39,560
You saw Snapchat, one of the 
first one building smack 

1091
00:57:39,560 --> 00:57:43,640
glasses, Google Google glasses. 
All these failed now meta gets 

1092
00:57:43,640 --> 00:57:47,440
it Well, got it partially right 
with the Ray Ben with just 

1093
00:57:47,440 --> 00:57:50,240
recording now they're brew and 
bringing out the new glasses, 

1094
00:57:50,480 --> 00:57:54,760
which have a small display 
integrated. 

1095
00:57:54,960 --> 00:57:58,240
So it's like it's interesting to
see how it goes because like 

1096
00:57:58,240 --> 00:58:01,000
everybody, it's basically back 
in the day when the smartphone 

1097
00:58:01,000 --> 00:58:04,400
was coming out and you were 
still using desktop like 

1098
00:58:04,400 --> 00:58:07,560
basically main day desktop yes, 
then you had to go to mobile 

1099
00:58:07,560 --> 00:58:10,720
Facebook, the right job 
optimize, like stop developing 

1100
00:58:10,720 --> 00:58:15,520
new new features and went full 
head on mobile development 

1101
00:58:15,520 --> 00:58:18,800
because mobile was the future 
and thinks that they grew even 

1102
00:58:18,800 --> 00:58:21,560
faster. 
The next might be the next case 

1103
00:58:21,560 --> 00:58:26,160
once the glasses are mass 
adopted that we're seeing also 

1104
00:58:26,320 --> 00:58:29,440
big push towards the these 
glasses because it's just easier

1105
00:58:29,440 --> 00:58:34,000
to take videos, video calls. 
You can have your hands free 

1106
00:58:34,400 --> 00:58:36,560
while still the person seeing 
you, right? 

1107
00:58:36,560 --> 00:58:39,200
Because there you've seen these 
things where the other person 

1108
00:58:39,200 --> 00:58:42,520
actually still sees your face, 
which is just an animated 

1109
00:58:42,960 --> 00:58:44,880
version of yourself talking 
Avatar. 

1110
00:58:45,080 --> 00:58:47,880
Because you've scanned it once 
and now it just like you see it 

1111
00:58:47,880 --> 00:58:50,440
on the other side, which is also
like an interesting thing. 

1112
00:58:51,400 --> 00:58:57,560
Yeah, have you tried VR? 
VR only I think I put the 

1113
00:58:57,560 --> 00:59:01,160
glasses on but like not too much
to be honest. 

1114
00:59:02,120 --> 00:59:04,480
VR is something that blew my 
mind actually. 

1115
00:59:04,560 --> 00:59:11,080
Yeah, seriously. 
I mean, we played two or three 

1116
00:59:11,080 --> 00:59:15,160
times VR shooters like literally
you, because in in Germany there

1117
00:59:15,160 --> 00:59:19,440
are basically you go there with 
five people and then you have 

1118
00:59:19,440 --> 00:59:23,960
like Star Trek, we think we 
played last time and before we 

1119
00:59:23,960 --> 00:59:26,720
played a different game. 
It is cool, which was cool 

1120
00:59:26,720 --> 00:59:30,520
because you're in a room and 
they're like limitations, which 

1121
00:59:30,560 --> 00:59:33,480
makes it very cool. 
But still some of these games 

1122
00:59:33,480 --> 00:59:39,520
were laggy, leggy, but but it 
was, I was still mind blowing, 

1123
00:59:39,520 --> 00:59:43,560
especially like sometimes, 
especially when do your eyes 

1124
00:59:43,560 --> 00:59:46,600
perceive that you're moving, 
then you're also like feeling 

1125
00:59:46,800 --> 00:59:49,600
shaky, which is like for me, 
it's like very, very 

1126
00:59:49,640 --> 00:59:52,720
interesting. 
It's like your body is 2 feet on

1127
00:59:52,720 --> 00:59:55,480
the ground. 
You should not have any problems

1128
00:59:55,480 --> 00:59:58,280
with. 
It's crazy how much we rely on 

1129
00:59:58,280 --> 01:00:01,520
our eyesight. 
That's what's the scary part is 

1130
01:00:01,640 --> 01:00:04,520
it's just like you get. 
If you close your eyes, you're 

1131
01:00:04,520 --> 01:00:06,240
like, oh, all of a sudden you're
like, OK, all right. 

1132
01:00:06,560 --> 01:00:09,840
But how we perceive it through 
these reality classes, you think

1133
01:00:09,840 --> 01:00:13,080
you're somewhere else and 
something really is like flying 

1134
01:00:13,080 --> 01:00:17,840
towards you and you're like, 
shit, it's my it is crazy. 

1135
01:00:18,440 --> 01:00:19,800
Even though the graphics are not
that great. 

1136
01:00:20,040 --> 01:00:23,880
No, not yet. 
But I mean, that's that's why I 

1137
01:00:24,160 --> 01:00:26,600
also believe for AI will be very
helpful, especially when 

1138
01:00:26,680 --> 01:00:30,360
building these worlds, right? 
Because you can basically let 

1139
01:00:30,360 --> 01:00:34,480
them build it for you. 
If you have like let them create

1140
01:00:34,480 --> 01:00:37,560
all these stuff and then it's 
easy because otherwise it takes 

1141
01:00:37,560 --> 01:00:43,200
forever, man, building, building
all of these, right? 

1142
01:00:44,360 --> 01:00:46,800
I think we got a little bit off 
topic, which are also very 

1143
01:00:46,800 --> 01:00:50,240
really like. 
But can I? 

1144
01:00:50,240 --> 01:00:52,360
Do you still have something that
you want to talk to us about? 

1145
01:00:52,440 --> 01:00:56,160
Hell Zero. 
And hello before we wrap this 

1146
01:00:56,160 --> 01:00:58,800
up. 
I think Hell 0 is very 

1147
01:00:58,800 --> 01:01:00,880
interesting. 
I'm very, very interested to 

1148
01:01:00,880 --> 01:01:05,760
move into that aspect, primarily
because, you know, like I would 

1149
01:01:05,760 --> 01:01:08,800
like to model it on how human, 
human intelligence came through.

1150
01:01:09,400 --> 01:01:12,360
And if you can, you know, model 
the human intelligence and make 

1151
01:01:12,360 --> 01:01:16,520
it as smart as human. 
That's all we need, right, As 

1152
01:01:16,520 --> 01:01:18,840
smart as human. 
And then we can try to think of 

1153
01:01:18,840 --> 01:01:20,200
making something smarter than 
human. 

1154
01:01:20,200 --> 01:01:22,600
It's it's a philosophical 
question, you know, is there is 

1155
01:01:22,600 --> 01:01:25,200
a limit to the intelligence when
we do the visual model that the 

1156
01:01:25,200 --> 01:01:27,400
question you asked. 
It's a very philosophical 

1157
01:01:27,400 --> 01:01:30,200
question, but I'm hoping I 
managed to, you know, set up 

1158
01:01:30,200 --> 01:01:34,200
some kind of infrastructure so 
we can this build up on that and

1159
01:01:34,200 --> 01:01:36,000
have some kind of ASI in the 
future. 

1160
01:01:38,120 --> 01:01:41,240
I'm sure you both do great. 
Especially I like the concept 

1161
01:01:41,240 --> 01:01:44,000
that you already have a cash cow
or like a started that actually 

1162
01:01:44,000 --> 01:01:47,880
is working. 
And then you can pour the money 

1163
01:01:47,880 --> 01:01:52,280
of this into the new one, which 
makes it then like basically you

1164
01:01:52,280 --> 01:01:54,400
can use the technology that 
you're learning there to 

1165
01:01:54,560 --> 01:01:57,400
optimize the other one again, 
which is really, really cool. 

1166
01:01:58,280 --> 01:01:59,840
Yeah, Ganesh, where can people 
find you? 

1167
01:02:00,080 --> 01:02:02,240
You can find me on 
ganesh@halo.com. 

1168
01:02:02,560 --> 01:02:06,720
That is my primary e-mail and 
I'll respond immediately to send

1169
01:02:06,720 --> 01:02:07,280
a message. 
Terrific. 

1170
01:02:07,320 --> 01:02:10,920
Any Twitter or anything else 
that you don't want the users to

1171
01:02:10,920 --> 01:02:12,120
know? 
About the Twitter, the 

1172
01:02:12,120 --> 01:02:15,000
Instagram, the YouTube all air 
hello. 

1173
01:02:15,000 --> 01:02:18,000
You just find air hello over 
there and message over there. 

1174
01:02:18,000 --> 01:02:19,720
It's usually handled by our 
social media team. 

1175
01:02:20,080 --> 01:02:22,160
But if we send a message to my 
e-mail and I'll, I'm the one you

1176
01:02:22,160 --> 01:02:23,960
want handling that. 
Awesome, awesome. 

1177
01:02:23,960 --> 01:02:26,120
Thank you, Ganesh. 
Ganesh, it was a pleasure. 

1178
01:02:26,120 --> 01:02:30,240
Talking to you. 
So yeah, I really, I'm excited 

1179
01:02:30,240 --> 01:02:33,280
to see where this is going, 
especially seeing in like 6 

1180
01:02:33,280 --> 01:02:36,760
months or eight months where all
these leads.