Cognitive AI for Transformative Customer Service

WEBVTT

1
00:00:02.919 --> 00:00:07.400
You are listening to Leaders in Lending
from Upstart, a podcast dedicated to helping

2
00:00:07.400 --> 00:00:13.080
consumer lenders grow their programs and improve
their product offerings. Each week, here,

3
00:00:13.080 --> 00:00:17.160
decision makers in the finance industry offer
insights into the future of the lending

4
00:00:17.160 --> 00:00:22.600
industry, best practices around digital transformation, and more. Let's get into the

5
00:00:22.600 --> 00:00:26.839
show. Welcome to Leaders and Lending. I'm your host, Jeff Keltner.

6
00:00:27.239 --> 00:00:32.600
This week's episode features my conversation with
Peter Voss from AIGO, and Peter's actually

7
00:00:32.600 --> 00:00:35.560
the guy who back in the day
in the early thousands, point the term

8
00:00:35.640 --> 00:00:38.960
artificial general intelligence. So to be
a little outside of our normal thing,

9
00:00:38.960 --> 00:00:42.679
we're going to go deep on AI
in this conversation, and it was really

10
00:00:42.759 --> 00:00:47.520
fascinating for me because we talked about
the recent explosion of interest in generative AI,

11
00:00:48.000 --> 00:00:51.920
what it does well, what it
gets wrong, like hallucinations, and

12
00:00:51.960 --> 00:00:57.159
why Peter is both optimistic about our
ability to create artificial general intelligence, not

13
00:00:57.200 --> 00:01:02.240
just specific intelligence, but a very
generalizable, flexible intelligence, but also that

14
00:01:02.280 --> 00:01:07.040
the general technologies around statistical AI that
underpin most current approaches are really not the

15
00:01:07.079 --> 00:01:11.760
right way. And so he talks
about his approach to what he calls cognitive

16
00:01:11.760 --> 00:01:14.680
AI, what they're building on that
front, what he calls his chat pup

17
00:01:14.719 --> 00:01:19.680
with the brain, and what the
future looks like for generalized intelligence generalized artificial

18
00:01:19.760 --> 00:01:23.200
intelligence. So I think it's a
really fascinating conversation for anyone who's curious about

19
00:01:23.319 --> 00:01:26.439
what's going on in THEI, what's
next, and maybe talk about the edges

20
00:01:26.480 --> 00:01:29.200
of where some of these things are
not working as well and why. So

21
00:01:29.239 --> 00:01:34.280
please enjoy this conversation, Peter Boss
Peter, welcome to the podcast, and

22
00:01:34.319 --> 00:01:37.159
thanks for making the time to join
us. Yes, well, thank you.

23
00:01:37.560 --> 00:01:41.400
I'm looking forward to this conversation.
It's been a little off the beaten

24
00:01:41.439 --> 00:01:44.200
track compared to many of my conversations
with bankers. I don't think we'll go

25
00:01:44.239 --> 00:01:48.280
down any deep banking rabbit holes.
But you've got a little more of a

26
00:01:48.319 --> 00:01:49.680
background and AI, so give us
a little bit of your history to so

27
00:01:49.719 --> 00:01:53.560
people can have a little context of
where you're coming from and what you've been

28
00:01:53.599 --> 00:01:57.840
doing for the last number of years. Yeah, certainly so. I actually

29
00:01:57.879 --> 00:02:02.000
started up as an electronics engineer,
intarted my own company, and then I

30
00:02:02.000 --> 00:02:07.359
fell in love with software and my
company turned into a software company. I

31
00:02:07.480 --> 00:02:13.840
developed an ERP comprehensive ERP system for
multi medium sized businesses, and that company

32
00:02:13.960 --> 00:02:16.439
quite well. We went from the
garage to four hundred people and did an

33
00:02:16.479 --> 00:02:21.360
IPO so that that was very exciting. I'd love to do that again.

34
00:02:22.759 --> 00:02:25.680
And it's when I exited that that
company, I you know, I said,

35
00:02:25.719 --> 00:02:30.240
well, why do I want to
tackle next? And what occurred to

36
00:02:30.280 --> 00:02:35.520
me is that software really isn't inherently
very smart. You know, if the

37
00:02:35.840 --> 00:02:38.280
if the programer didn't think of something, then it'll just crash or give you

38
00:02:38.319 --> 00:02:44.439
an error message, or you know, do something stupid. So I really

39
00:02:44.520 --> 00:02:49.159
want to figure out how can we
build software that actually has common sense,

40
00:02:49.280 --> 00:02:53.280
that can learn, that can reason. So that's a journey I've been on

41
00:02:53.319 --> 00:02:55.759
for the last you know, twenty
plus years. And I took off five

42
00:02:55.840 --> 00:03:01.319
years to study intelligence, all different
aspects of intelligence to be able to build

43
00:03:01.400 --> 00:03:06.680
intelligent system. But I thought I
really needed to understand it, and so

44
00:03:07.039 --> 00:03:12.879
I studied how children learn, how
our intelligence differs from animals, what IQ

45
00:03:13.080 --> 00:03:15.800
tests measure, you know, all
the work that had been done in the

46
00:03:15.800 --> 00:03:20.400
field of AI, and that study
culminated and me coming up with a theory

47
00:03:20.400 --> 00:03:24.080
of how to actually build a thinking
machine and then two thousand and one I

48
00:03:24.120 --> 00:03:29.520
started in my first AI company,
and for several years we were just in

49
00:03:29.719 --> 00:03:34.159
R and D mode, sort of
turning my ideas into actual code prototypes,

50
00:03:34.199 --> 00:03:39.879
and eventually a platform that we then
commercialized and automated calls and core center automation,

51
00:03:40.439 --> 00:03:45.240
and doing that a lot more intelligently
than what had been done before,

52
00:03:45.280 --> 00:03:52.319
because we have a cognitive engine that
actually deeply understands what you're saying. Users

53
00:03:52.319 --> 00:03:55.240
context uses history of what you said
earlier in the conversation, so it has

54
00:03:55.280 --> 00:04:00.520
short term emy long term memory.
And that's really the journey I've been on.

55
00:04:00.719 --> 00:04:11.199
So my current company also commercializes conversational
AI, but we also are continuing

56
00:04:11.240 --> 00:04:14.520
to increase the IQ of a system. You know, we really want to

57
00:04:14.560 --> 00:04:20.879
get to human full human level intelligence. It's fascinating to me the discussion about

58
00:04:20.959 --> 00:04:26.040
human intelligence and how close or not, I mean, there seems to be

59
00:04:26.040 --> 00:04:29.759
this very wide range of opinions and
how close some of the you know,

60
00:04:29.839 --> 00:04:33.160
at least the consumer facing cutting edge
technology things like cloud or bard or chat.

61
00:04:33.199 --> 00:04:38.240
GPT is probably the most popular,
and that transformer architecture that seems to

62
00:04:38.319 --> 00:04:43.399
kind of revolutionize the quality of conversation
we're having and I hear everything. I

63
00:04:43.399 --> 00:04:46.920
think we're like one turn of the
crank on that architecture from having like a

64
00:04:46.959 --> 00:04:48.560
superhuman brain, and people who go, hey, we're just not even on

65
00:04:48.600 --> 00:04:51.439
the right path. You're kind of
doing this thing, and I'm kind of

66
00:04:51.439 --> 00:04:56.160
curious where you stand on that,
and what do you think true human like

67
00:04:56.240 --> 00:04:59.879
intelligence requires. And it's not just
like being able to hold a conversation at

68
00:05:00.079 --> 00:05:01.279
the touring test is to be out
of date in terms of how we think

69
00:05:01.279 --> 00:05:05.199
about what a truly intelligent machine needs
to look like. So what's your perspective

70
00:05:05.240 --> 00:05:09.879
on how far or close are those
kinds of technologies and what do you think,

71
00:05:10.199 --> 00:05:14.519
if anything, they're missing in terms
of how their approach generates or christ

72
00:05:14.560 --> 00:05:18.439
to mimic intelligence and humans. Yes, it's a very interesting and very important

73
00:05:18.480 --> 00:05:24.439
question at this point in time,
because, you know, check GPT became

74
00:05:24.519 --> 00:05:30.680
available. I think it really was
a major shock to many people and including

75
00:05:30.800 --> 00:05:36.759
pretty much all AI researchers. Virtually
nobody expected the system to be as good

76
00:05:36.800 --> 00:05:40.439
as it, you know, to
be able to do the hold the kind

77
00:05:40.439 --> 00:05:44.800
of conversation that it can and the
amount of knowledge that it has the you

78
00:05:44.839 --> 00:05:47.279
know, that's one of those fascinating
things to me is that people who built

79
00:05:47.279 --> 00:05:50.920
it really like it's kind I kind
of describe, feels like a really fancy

80
00:05:50.920 --> 00:05:56.000
autocomplete, and then how it got
to things like reasoning and understand like,

81
00:05:56.000 --> 00:06:00.439
they really don't understand how that emmerged
to really build a reasoning system. They

82
00:06:00.519 --> 00:06:03.399
built like an autocomplete, and somehow
it has some basic deductive you know,

83
00:06:03.439 --> 00:06:06.759
it's like they're really surprised by what
I can do. Yes, yes,

84
00:06:06.800 --> 00:06:11.879
absolutely, And you know, I
think I think there are ways, certainly

85
00:06:11.920 --> 00:06:16.600
ways of understanding it, but there
are as impressive as it is and as

86
00:06:16.680 --> 00:06:23.279
useful as it is, there are
inherent fundamental limitations to that. And it's

87
00:06:23.360 --> 00:06:27.920
difficult to look past the sort of
impressiveness of it. You know, it

88
00:06:28.079 --> 00:06:32.879
just seems so intelligent. It's really
good at faking intelligence, and of course

89
00:06:33.439 --> 00:06:41.680
it you know it, it can
do really really amazing things. But the

90
00:06:41.759 --> 00:06:46.959
problem is it's really not reliable.
It's not aware of what it's saying at

91
00:06:46.040 --> 00:06:49.920
all. So there is no metacognition, and so there are a couple of

92
00:06:49.959 --> 00:06:58.120
components that are really missing to truly
make it human like in human level intelligence.

93
00:06:58.800 --> 00:07:01.360
So and in fact, the name
GPT gives you a clue already.

94
00:07:01.839 --> 00:07:06.199
G stands for generative and which you
know another way of saying that it makes

95
00:07:06.279 --> 00:07:12.639
up stuff and it literally does.
I mean you can prompt it and and

96
00:07:12.839 --> 00:07:16.839
you know ask you know, I
can ask it, tell me about the

97
00:07:16.879 --> 00:07:23.920
white paper I posted on kidney revival
or something you know, and you know

98
00:07:24.000 --> 00:07:27.800
it'll it'll come up with something,
you know, complete with One of those

99
00:07:27.879 --> 00:07:31.040
lawyers recently who asked a question came
up with like three precedents and they submitted

100
00:07:31.040 --> 00:07:34.600
them in a brief and then judge
those are those are real cases. You

101
00:07:34.600 --> 00:07:38.959
can't just so that was the fine
I think a couple of thousand dollars for

102
00:07:39.040 --> 00:07:44.079
profression on this conduct. It's like
relied on it, right, that's exactly

103
00:07:44.160 --> 00:07:46.279
right. So yeah, I mean
that's the generative part. And that's inherent

104
00:07:46.360 --> 00:07:50.519
and generative. It's trained on trillions
of pieces of information, you know,

105
00:07:50.600 --> 00:07:55.199
good, bad, and ugly,
and you know what it comes up was

106
00:07:55.360 --> 00:08:00.160
is really just that it it's sort
of coherent in its statistical model, but

107
00:08:00.199 --> 00:08:05.000
it doesn't it doesn't need to bear
any kind of relationship to reality and so

108
00:08:05.000 --> 00:08:09.879
so, and that's inherent. It's
just inherent an inherent limitation of the system

109
00:08:11.199 --> 00:08:15.480
the way it's built. Now.
The second part is GP is a pre

110
00:08:15.560 --> 00:08:18.240
trained and it's really a read only
model. They build a model, it

111
00:08:18.279 --> 00:08:22.600
costs you know, one hundred million
dollars or whatever, you know, some

112
00:08:22.839 --> 00:08:26.399
enormous amount. Yeah, and it
takes you know, days or weeks to

113
00:08:26.439 --> 00:08:31.120
train the model. And but then
it's essentially after that it's largely read only.

114
00:08:31.519 --> 00:08:37.720
Now it cannot learn interactively. Now, there are tweaks that people you

115
00:08:37.759 --> 00:08:41.200
know, that that the system allows, that it has some kind of a

116
00:08:41.480 --> 00:08:45.799
short term buffer that it takes into
account. That's basically the PROMPTU you feed

117
00:08:45.840 --> 00:08:50.039
in. But that information isn't doesn't
update the model. So whatever you teach

118
00:08:50.080 --> 00:08:54.159
it, there is, you know, does not change the model, which

119
00:08:54.200 --> 00:08:58.399
is very different from the way humans
operate, you know. I mean we

120
00:08:58.440 --> 00:09:01.600
can hear, we can hear one
word or two words, and it can

121
00:09:01.679 --> 00:09:07.279
completely change the way we perceive things. And you know, like Trump one,

122
00:09:07.559 --> 00:09:09.279
you know, you know, for
better or for worse. I mean,

123
00:09:09.559 --> 00:09:13.000
it changes everything, you know,
if you didn't expect it. So

124
00:09:13.879 --> 00:09:18.480
it updates our model. Basically,
you know, certain things just one or

125
00:09:18.480 --> 00:09:22.200
two words can really update our model
significantly. And I think that's what intelligence

126
00:09:22.919 --> 00:09:28.000
requires. And then what I already
mentioned is a lack of metacognition. It

127
00:09:28.039 --> 00:09:33.120
doesn't know what it's what it's saying, It can't think about its own thinking.

128
00:09:33.399 --> 00:09:37.600
So these are requirements I think of
real intelligence. I mean, imagine

129
00:09:37.600 --> 00:09:41.039
not hiring an assistant, you know, human assistant for you and that come

130
00:09:41.120 --> 00:09:45.960
and they oh, yeah, I
can do Excel, and I can you

131
00:09:46.000 --> 00:09:50.279
know, do PowerPoint, and I
can you know, do whatever, you

132
00:09:50.320 --> 00:09:52.480
know, look up things on the
internet for you. And then you tell

133
00:09:52.519 --> 00:09:56.480
them, well, you know,
we introduce this new product, we shut

134
00:09:56.519 --> 00:09:58.720
down one of our branches, we
have this new relationship, and so on.

135
00:10:00.480 --> 00:10:03.480
And next day your assistant comes in
and they don't know any they don't

136
00:10:03.519 --> 00:10:07.039
remember any of that. You know. Yeah, my son had an article

137
00:10:07.159 --> 00:10:11.840
to respond to today, the middle
six, sixth grader, and he had

138
00:10:11.840 --> 00:10:13.559
to write a summary. And you
know, we were talking about this exact

139
00:10:13.559 --> 00:10:16.759
point because I said, imagine what
your history teacher would do if she gave

140
00:10:16.799 --> 00:10:20.120
you a lesson. And then you
said, hold on, give me a

141
00:10:20.120 --> 00:10:22.360
week and one hundred million dollars and
I'll come back and be able to pass

142
00:10:22.399 --> 00:10:24.039
the TESK. But between now and
then, I won't have any memory of

143
00:10:24.080 --> 00:10:28.639
what she said. He wouldn't past
sixth grade very well, let alone a

144
00:10:28.720 --> 00:10:31.240
job interview or do well in the
workforce. So it's it's kind of unbelievable

145
00:10:31.240 --> 00:10:35.840
when you think about how young a
person would have a better level of intelligence

146
00:10:35.879 --> 00:10:41.360
from that real time learning perspective,
it's really quite remarkable, right, And

147
00:10:41.639 --> 00:10:46.360
it's absolutely crucial for you know,
intelligence in the real world. You need

148
00:10:46.399 --> 00:10:50.240
to be able to learn things on
the fly and think through what the implications

149
00:10:50.240 --> 00:10:56.240
are of the new knowledge that you
have. And this technology you simply cannot

150
00:10:56.279 --> 00:10:58.840
do it. Transform a technology cannot
do it. So we need we need

151
00:10:58.840 --> 00:11:03.960
to different approach, and the approach
that I believe we need is something called

152
00:11:05.039 --> 00:11:11.919
cognitive AI as opposed to statistical or
generative and cognitive AI really starts with the

153
00:11:11.000 --> 00:11:16.960
question of what does intelligence require and
to build a system around that, where

154
00:11:16.000 --> 00:11:20.679
statistical AI started with, Hey,
we've got a lot of data. We've

155
00:11:20.720 --> 00:11:24.240
got a lot of computing power,
you know, the big companies like Google

156
00:11:24.320 --> 00:11:28.879
and Facebook. So what can we
do with a lot of data and a

157
00:11:28.879 --> 00:11:31.000
lot computing power? You know,
that's the hammer we've got. So what

158
00:11:31.080 --> 00:11:37.000
nails can we find? It's not
starting with what does intelligence require? And

159
00:11:37.320 --> 00:11:39.879
that's really the I think the approach
that's needed, and you know, that's

160
00:11:39.879 --> 00:11:45.720
the approach that we've taken. Do
you see any kind of recognition? I

161
00:11:45.720 --> 00:11:50.519
mean, it feels like there's so
much money in investment pouring into the generative

162
00:11:50.519 --> 00:11:52.399
AI space. Now, whether it's
you know, everybody trying to catch up

163
00:11:52.399 --> 00:11:58.159
with chat GPT in terms of anthropic
or Google investing in Claude and Bard,

164
00:11:58.919 --> 00:12:03.000
or whether it's vcs investing in startups, they're all using the same kind of

165
00:12:03.000 --> 00:12:07.039
statistical AI approach, right They're iterating
on the transformer or the diffusion models in

166
00:12:07.080 --> 00:12:13.600
the space of image image generation.
Do you see enough investment going into this

167
00:12:13.720 --> 00:12:16.080
kind of like new approaches because it
feels like we're all iterating on the same

168
00:12:16.200 --> 00:12:20.600
thing versus trying to build the new
thing. There's so much resource going there.

169
00:12:20.759 --> 00:12:24.120
I wonder what your thoughts are on
that versus seeing more investment into different

170
00:12:24.120 --> 00:12:28.200
approaches that more closely mirror what we
think is going on in human intelligence.

171
00:12:28.240 --> 00:12:33.639
It might be required for a more
generalized intelligence capability, right, Well,

172
00:12:33.799 --> 00:12:37.960
you know, obviously it has a
lot of momentum and really for I think

173
00:12:37.559 --> 00:12:43.720
since really the dot com boom.
People really only think in terms of momentum

174
00:12:43.720 --> 00:12:46.000
investing. You know, it's basically, we are things happening. Hey,

175
00:12:46.039 --> 00:12:50.039
let me pile in, let me
put more. And you know that that

176
00:12:50.399 --> 00:12:54.639
has this flywheel flywheel kind of effect
that the more people jump on, the

177
00:12:54.639 --> 00:12:58.039
more money comes in. You know. And you know, you can't even

178
00:12:58.120 --> 00:13:03.000
get a do a PhD computer science
if it doesn't involve machine learning, deep

179
00:13:03.039 --> 00:13:07.559
learning, statistical systems. These days, you know, you're probably not going

180
00:13:07.600 --> 00:13:11.120
to get funded on a VC if
you're not using that, so it you

181
00:13:11.159 --> 00:13:16.600
know, it hasn't effect. And
I think what's happening now is there is

182
00:13:16.679 --> 00:13:20.399
a realization of hang on, these
are real limitations that we have. We

183
00:13:20.480 --> 00:13:24.000
cannot trust these systems. They cannot
learn, they don't have metal cognition,

184
00:13:24.120 --> 00:13:31.039
you know, and and they're incredibly
expensive to run and to train. So

185
00:13:31.279 --> 00:13:35.039
people are starting to realize it.
But you know, the momentum has to

186
00:13:35.120 --> 00:13:39.360
kind of work its way out,
and I think a lot of vcs have

187
00:13:39.440 --> 00:13:43.279
to lose a lot of money,
which I think is going to happen soon.

188
00:13:43.639 --> 00:13:48.000
You know, as these these startups
that try to capitalize on it just

189
00:13:48.039 --> 00:13:54.279
find well, you know, what
are actually applications that are creating value.

190
00:13:54.679 --> 00:14:00.799
Now, there are a lot of
applications that create value, you know,

191
00:14:00.960 --> 00:14:05.480
such such as as a programming aid
and to help you brainstorm staff, or

192
00:14:05.519 --> 00:14:09.799
to generate images or logos or you
know, things where there's a human and

193
00:14:09.840 --> 00:14:13.840
loop using it as a tool.
But there's you know, there's a limited

194
00:14:13.919 --> 00:14:20.120
amount of value that creates. It
doesn't really replace human labor on a large

195
00:14:20.159 --> 00:14:26.559
scale, and that's really where the
real value would be created if it could

196
00:14:26.639 --> 00:14:31.159
do that. So I think we'll
see the tide turning. The problem is

197
00:14:31.080 --> 00:14:35.840
all of the experts, almost all
of the experts in the field, the

198
00:14:35.879 --> 00:14:41.799
only thing they know is statistical approaches. So you have a whole ecosystem where

199
00:14:41.840 --> 00:14:45.080
the AI experts in fact, you
know, if you talk to somebody now

200
00:14:45.320 --> 00:14:50.639
who graduated in the last ten years, they actually think that AI is statistical.

201
00:14:52.159 --> 00:14:54.600
That's it. You know, there's
no there's no other approach to AI.

202
00:14:56.000 --> 00:15:00.639
So you basically have all of all
of the experts in the field that

203
00:15:01.279 --> 00:15:07.679
their background is mathematics, statistics,
and logic, and those are kind of

204
00:15:07.679 --> 00:15:11.240
the only way they can think about
the problem. So it's going to be

205
00:15:11.480 --> 00:15:16.519
pretty difficult, you know, to
to have that whole ecosystem really change to

206
00:15:16.559 --> 00:15:22.639
say no, these approaches are not
going to get us to true human level

207
00:15:22.679 --> 00:15:28.120
AI. So I don't know how
long that will take, but certainly seeing

208
00:15:28.639 --> 00:15:33.279
an you know, an increasing realization
of what the inherent limitations are and you

209
00:15:33.679 --> 00:15:39.320
can't just brute force your way to
to human level intelligence. Yeah, it's

210
00:15:39.360 --> 00:15:43.080
interesting you talk about the evolution of
companies and how they invest in this or

211
00:15:43.120 --> 00:15:46.279
not. And it reminds me a
little bit of like the moment when Google

212
00:15:46.320 --> 00:15:50.440
was founded, because it feels so
inevitable looking back what Google became and what

213
00:15:50.559 --> 00:15:54.799
it is. But like search was
considered by many a solved problem and they're

214
00:15:54.799 --> 00:15:58.279
like, there's not room for other
search engine, there's licos, there's alt

215
00:15:58.320 --> 00:16:00.919
that this like because a starting a
search engine right now, it doesn't make

216
00:16:00.919 --> 00:16:04.440
any sense. Uh, And they
were clearly off the mark. And I

217
00:16:04.440 --> 00:16:10.679
think that feels like the same inevitability
that Google looks like in retrospect behind the

218
00:16:10.759 --> 00:16:14.879
guys you know open AI and chat
GPT, for instance. But it is

219
00:16:14.960 --> 00:16:18.559
often true that that's the moment when
the true innovation comes out and new things

220
00:16:18.600 --> 00:16:22.639
are founded that end up looking inevitable
in retrospects. But maybe foolhardy at the

221
00:16:22.639 --> 00:16:26.559
time, and maybe it's one of
those moments for a totally different approach despite

222
00:16:26.600 --> 00:16:30.480
the large, you know, influence
that some of these companies have right now.

223
00:16:30.720 --> 00:16:33.840
Yeah, now you make an excellent
point there, and I mean,

224
00:16:33.919 --> 00:16:40.639
yes, how could a little startup
Google possibly compete against the existing, you

225
00:16:40.679 --> 00:16:44.720
know, established search engines? You
take the same with Amazon. How could

226
00:16:44.840 --> 00:16:49.759
some you know, little mickey Mouse
online books books possibly take on Walmart?

227
00:16:49.799 --> 00:16:55.639
You know? And we see that
again and again. It's, you know,

228
00:16:55.720 --> 00:17:00.759
at a certain point, the establishment
is, you know, it is

229
00:17:00.159 --> 00:17:04.359
fixed on a particular approach and they
really can't break out of it. So

230
00:17:04.400 --> 00:17:07.079
I want to ask you a little
bit about what you guys do You describe

231
00:17:07.079 --> 00:17:11.720
it to me once at your company
as a chat bot with the brain one

232
00:17:11.720 --> 00:17:15.240
of the products you have, So
talk to me about how that plays out

233
00:17:15.279 --> 00:17:18.279
in a company and what you mean
when you say with a brain, Like

234
00:17:18.319 --> 00:17:22.319
that's an interesting phrase. I think
I'd love to understand where you're kind of

235
00:17:22.359 --> 00:17:26.960
coming at from with that. Right
Well, all the other chatbots out there,

236
00:17:26.119 --> 00:17:30.000
as far as I know, do
not have a brain. And what

237
00:17:30.039 --> 00:17:37.920
we mean by brain is basically a
cognitive engine that has deep understanding users context.

238
00:17:37.039 --> 00:17:41.319
Has short term memory, it remembers
what were said earlier in the conversation,

239
00:17:41.759 --> 00:17:45.759
has long term memory, it remembers
previous conversations with each individual. So

240
00:17:48.039 --> 00:17:55.440
it really has a comprehensive cognitive engine
or brain that manages the whole conversation,

241
00:17:56.039 --> 00:18:02.119
and other chatbots really use in a
approach that's been around for thirty years or

242
00:18:02.160 --> 00:18:07.720
so. So on the one hand, you have an intent classifier where you

243
00:18:07.759 --> 00:18:11.440
know, whatever the utterance is blah
blah blah, weather okay, it triggers

244
00:18:11.440 --> 00:18:18.440
the weather app. Or you say
I hate Uber, don't ever give me

245
00:18:18.559 --> 00:18:22.279
Uber again, Well, I'll probably
still trigger the Uber app because it latches

246
00:18:22.279 --> 00:18:26.839
onto that. So it's a fixed
a fixed list of options that can get

247
00:18:26.880 --> 00:18:30.720
triggered. And once they get triggered, there's basically little float charity type program

248
00:18:30.759 --> 00:18:34.440
that somebody writes, you know,
where do you want to go? How

249
00:18:34.440 --> 00:18:38.960
many people are going? And do
you want Uber X? So there you

250
00:18:38.960 --> 00:18:45.000
know. And then obviously people have
different levels of sophistication where they may hard

251
00:18:45.039 --> 00:18:49.599
code certain terms to be remembered,
so there may be some very limited specific

252
00:18:51.519 --> 00:18:53.920
short term memory or long term memory. They look up something in the database,

253
00:18:55.119 --> 00:18:59.839
but it's not an inherent part of
the architecture that there is deep understanding.

254
00:19:00.079 --> 00:19:04.480
Context is being used. The system
reasons about what it hears, and

255
00:19:04.839 --> 00:19:08.839
you know, and and so with
a brain you can have a much much

256
00:19:10.640 --> 00:19:15.640
more sophisticated conversation. So as an
example, one of one of our customers

257
00:19:15.720 --> 00:19:22.279
is one eight hundred Flowers, a
group of companies which includes Harry and David,

258
00:19:22.319 --> 00:19:26.680
popcorn factory and so on. So
they use our system very very successfully

259
00:19:26.079 --> 00:19:33.200
to provide a concierge type service to
their twenty million customers. And so our

260
00:19:33.279 --> 00:19:37.240
system remembers who you buy gifts for, for what occasion, what kind of

261
00:19:37.240 --> 00:19:41.359
gifts they're like, it's able to
do that because of the deep understanding and

262
00:19:41.759 --> 00:19:48.119
the memory. So you know,
this is where the brain really allows us

263
00:19:48.200 --> 00:19:55.160
to handle very complex conversations. And
the same approach you know, can be

264
00:19:55.200 --> 00:20:02.279
applied to you know, banking or
medical applicat as or whatever. I mean,

265
00:20:02.599 --> 00:20:07.960
we've all experienced how limited IVRs,
you know, voice interactions and chatbots

266
00:20:08.000 --> 00:20:14.119
are. I mean Bank of America
Erica, I think they have like what

267
00:20:14.319 --> 00:20:19.400
three percent repeat usage or something,
you know, very very low number in

268
00:20:19.440 --> 00:20:25.480
spite of hundreds of millions of dollars
being spent on that they don't have a

269
00:20:25.519 --> 00:20:29.240
brain. That's fascinating. I'm curious. How where do you think you guys

270
00:20:29.279 --> 00:20:33.359
are at in the taking the architecture
from what it's being used for today to

271
00:20:33.440 --> 00:20:37.000
more generalized tasks. You talk about
replacing human labor on a large scale,

272
00:20:37.319 --> 00:20:41.880
Where are your systems at and the
ability to move from a call center,

273
00:20:41.960 --> 00:20:48.720
a chap type interaction to you know, more generalized cognitive capabilities and types of

274
00:20:48.799 --> 00:20:52.880
labor. Yeah, So over the
last few years, our concentration has has

275
00:20:52.960 --> 00:20:56.559
been to really build out the infrastructure, the robustness, you know, and

276
00:20:56.920 --> 00:21:03.039
scalability and easily implementation and all of
that. So our focus has been very

277
00:21:03.119 --> 00:21:08.680
much on commercial aspect of our technology. But you know, we now have

278
00:21:08.759 --> 00:21:15.279
a very solid system and our focus
is actually shifting back again to cranking up

279
00:21:15.319 --> 00:21:18.160
the IQ of our system. You
know, there's still quite a ways to

280
00:21:18.240 --> 00:21:25.400
go to scale up the amount of
knowledge that the system has, the degree

281
00:21:25.440 --> 00:21:29.799
of complexity of what it can handle, how it can handle, For example,

282
00:21:30.559 --> 00:21:33.519
theory of mind. You know,
what is the other person thinking?

283
00:21:33.559 --> 00:21:37.480
What does the other person know?
So we need to scale up our brain

284
00:21:38.079 --> 00:21:45.720
in various areas to you know,
handle these more complex situations. Also,

285
00:21:45.759 --> 00:21:49.880
at the moment, there's still a
fair amount of work involved in actually teaching

286
00:21:49.920 --> 00:21:56.440
the system the specifics of each company
that we that use our system. And

287
00:21:56.160 --> 00:22:00.920
as our system becomes smarter, it
will be able to self configure to a

288
00:22:02.000 --> 00:22:06.240
much larger degree where you know,
it will be able to ask the question,

289
00:22:06.720 --> 00:22:08.440
you know, do you want me
to handle it this way or that

290
00:22:08.480 --> 00:22:14.039
way, or let let me read
let me read your training manuals. You

291
00:22:14.039 --> 00:22:17.759
know, at the moment, there's
still a human in the loop in many

292
00:22:17.799 --> 00:22:22.880
of us for many of these training
exercises. So as we crank up the

293
00:22:22.279 --> 00:22:30.680
IQ, the implementation implementations will become
much easier and it will be able to

294
00:22:30.759 --> 00:22:37.839
handle, you know, even more
complex situations and conversations. So we we

295
00:22:37.000 --> 00:22:47.440
just recently sort of rebooted our development
focus. We currently have about ten people

296
00:22:47.720 --> 00:22:52.279
working on that full time, but
we are actually in the process to raise

297
00:22:52.359 --> 00:22:56.720
money to increase it to one hundred. We want a hundred people on that

298
00:22:56.799 --> 00:23:00.000
project. We know what we need
to do to crank up IQ to get

299
00:23:00.079 --> 00:23:06.720
to human level intelligence, but were
currently looking for funding and so we're looking

300
00:23:06.759 --> 00:23:11.960
for the kind of visionary investors that
can look past the what if you know,

301
00:23:12.000 --> 00:23:18.160
the bandwagon that everybody is jumping jumping
on and say, well, what

302
00:23:18.240 --> 00:23:22.759
does intelligence really require and who has
who has potentially a solution for that.

303
00:23:22.359 --> 00:23:29.759
I'm fascinated by this idea that the
smarter version of your iteration of your program

304
00:23:29.839 --> 00:23:32.440
will start to kind of be able
to train itself. I mean, what

305
00:23:32.799 --> 00:23:36.319
are the things you think? What
are the real markers of as the system

306
00:23:36.359 --> 00:23:40.279
gets more and more intelligent, Like
the self configuration asking for the data it

307
00:23:40.359 --> 00:23:44.799
might need to be able to respond
more accurately is a really interesting one,

308
00:23:44.880 --> 00:23:48.079
kind of like self directed learning,
which is very different I think than in

309
00:23:48.119 --> 00:23:51.559
the way we might use those terms
today. But what are the markers you

310
00:23:51.599 --> 00:23:55.279
see of like how this evolves and
how you can tell that it's getting smarter

311
00:23:55.359 --> 00:23:56.759
and smarter? Like, what are
the you know, there's not like a

312
00:23:56.839 --> 00:24:00.279
ninth grade test to a high school
test you want to pass the way GPT

313
00:24:00.400 --> 00:24:03.240
I think passes the you know,
the bar exam or whatever. What are

314
00:24:03.240 --> 00:24:07.519
the markers for you as you get
farther down this line of real real cognition

315
00:24:08.279 --> 00:24:11.519
and you're ramping up the IQ of
the system, what does that look like?

316
00:24:11.319 --> 00:24:15.759
Yeah, so it's very different for
us. It's to us, it's

317
00:24:15.799 --> 00:24:19.200
this is not a black box.
So it's not that we kind of keep

318
00:24:19.359 --> 00:24:25.200
throwing more data and computing power at
it and then wow, you know what

319
00:24:25.240 --> 00:24:26.839
can it do? Now? Wow
that's cool? You know, Okay,

320
00:24:27.039 --> 00:24:32.160
can we tweak it a bit here
and there to to make sure it doesn't

321
00:24:32.160 --> 00:24:34.319
swear at us or say naughty things, you know, or whatever. That's

322
00:24:34.359 --> 00:24:38.079
not at all the approach I mean, uh, the with with the cognitive

323
00:24:38.160 --> 00:24:42.680
architecture that we have, it's it's
completely scrutable, so we can see what

324
00:24:42.839 --> 00:24:47.000
is doing and why it's doing what
is doing, and so we have a

325
00:24:47.039 --> 00:24:52.000
curriculum that we're going through to enable
the system to handle more and more complex

326
00:24:52.079 --> 00:24:56.279
tasks, to handle more and more
complex reasoning. So you know, I

327
00:24:56.279 --> 00:24:59.920
mean, they're little surprises along the
way we say, wow, this wor

328
00:25:00.119 --> 00:25:03.839
a lot better than we thought it
might, you know, or this is

329
00:25:03.880 --> 00:25:07.400
a lot harder than we thought,
and you know, it takes more time.

330
00:25:07.799 --> 00:25:12.359
But we really have a particular development
paths that we've mapped out to get

331
00:25:12.359 --> 00:25:17.200
from where we are to get to
full human level intelligence. And it's just

332
00:25:17.279 --> 00:25:22.599
an incremental incremental paths, and you
know, it requires a fair amount of

333
00:25:22.640 --> 00:25:27.799
effort, not billions or trillions of
dollars, but you know, certainly tens

334
00:25:27.799 --> 00:25:30.880
of millions, you know, I
mean for us to you know, we

335
00:25:32.200 --> 00:25:37.240
need you know, one hundred two
hundred people working on this to to really

336
00:25:37.279 --> 00:25:45.000
get to human level commercial product.
But we have a very specific road map

337
00:25:45.039 --> 00:25:48.839
to achieve that. What kinds of
things are on your curriculum? I'm curious.

338
00:25:48.839 --> 00:25:52.759
I always ask my kids teachers for
their you know, curriculum for the

339
00:25:52.839 --> 00:25:56.400
year. What are the kinds of
things just conceptually that are on the curriculum

340
00:25:56.480 --> 00:26:00.599
for your system? Yeah, So
with a new project that that we have

341
00:26:00.799 --> 00:26:06.960
now is one of the big changes
is our commercial product. The language ability

342
00:26:07.799 --> 00:26:11.759
was really put in by humans.
You know, we've fed it, we've

343
00:26:11.759 --> 00:26:15.119
feded the different language rules and so
on. So there was a lot of

344
00:26:15.240 --> 00:26:21.160
curation being done. And we've always
known, and we've known for you know,

345
00:26:21.200 --> 00:26:26.160
for the last twenty years that we've
been developing the technology and commercializing it.

346
00:26:26.480 --> 00:26:30.799
We've always known that ultimately the system
has to be able to learn language

347
00:26:30.000 --> 00:26:37.599
by itself, essentially with a teacher. So our current focus is on you

348
00:26:37.640 --> 00:26:45.279
know, finally replacing these human coded
rules with a knowledge that the system requires,

349
00:26:45.559 --> 00:26:49.559
and so that's the curriculum we're right
now doing, is to teach it

350
00:26:51.519 --> 00:27:00.799
the basic language skills interactively. And
you know that that curriculum, it has

351
00:27:00.000 --> 00:27:03.880
certainly has some overlap to the way
children learn. You know, there are

352
00:27:03.960 --> 00:27:08.759
layers of what is the most fundamental
knowledge that you need to have. You

353
00:27:08.799 --> 00:27:14.200
know about things. There are objects, there are activities, there are attributes

354
00:27:14.240 --> 00:27:19.240
that they are entities. You know, there are general categories and those things

355
00:27:19.279 --> 00:27:26.960
have to be kind of very deeply
embedded. So that that's our focus right

356
00:27:26.960 --> 00:27:30.640
now. Uh, the curriculum now
as we do that, we will we

357
00:27:30.680 --> 00:27:34.240
will then go back to the kinds
of things we've been doing. Is you

358
00:27:34.279 --> 00:27:40.160
know, how how do you take
an order or how do you interact with

359
00:27:40.279 --> 00:27:45.440
APIs to achieve different to achieve different
things, So that that will be the

360
00:27:45.480 --> 00:27:51.400
curriculum that that we execute as we
go along. That's a that's a fascinating

361
00:27:51.480 --> 00:27:55.519
journey. I can't wait to see
what comes out of it. It seems

362
00:27:55.559 --> 00:28:00.000
like a very interesting and different approach
and it'll be really interesting to see how

363
00:28:00.960 --> 00:28:06.839
hopefully your system doesn't have the same
kind of teenage angst and difficulty that humans

364
00:28:06.880 --> 00:28:12.000
do. But the deuteractive learning a
very interesting path. No, no,

365
00:28:12.000 --> 00:28:17.480
no, it will not. We're
not going to put injected with hormones or

366
00:28:17.839 --> 00:28:21.480
you know, so it's it's not
going to have to go through puberty or

367
00:28:21.480 --> 00:28:25.960
any any of that. So that's
thankfully not something we have to deal with

368
00:28:25.960 --> 00:28:30.839
in our AI. But you know, I mean, I am so excited

369
00:28:30.880 --> 00:28:33.319
about working on this project, and
I have been for as long as I

370
00:28:33.319 --> 00:28:37.480
have been working on it because of
the tremendous potential that it has. And

371
00:28:37.680 --> 00:28:42.319
you know, I mean I could
name just like three categories. The first

372
00:28:42.319 --> 00:28:51.200
one is imagine having a true PhD
level researchers AI researchers. You know,

373
00:28:51.279 --> 00:28:55.960
you get one that gets trained to
be a cancer researcher, for example,

374
00:28:56.039 --> 00:29:00.160
with all the background knowledge, then
you can make a million copies of that.

375
00:29:00.559 --> 00:29:04.960
You now have a million PhD level
cancer researchers chipping away at the problem.

376
00:29:06.279 --> 00:29:08.559
You know, I mean, we're
going to make much much more progress.

377
00:29:08.880 --> 00:29:12.039
You take that for other diseases.
You take it, you know,

378
00:29:12.519 --> 00:29:18.400
for energy for you know, battery
efficiency, for pollution, you know,

379
00:29:18.480 --> 00:29:26.200
climate change, all of you know, having these PhD level AIS chipping away

380
00:29:26.240 --> 00:29:30.559
at these these many problems that are
facing humanity. That's kind of the one

381
00:29:30.640 --> 00:29:38.000
area. The second area is obviously
dramatically reducing the cost of goods and services

382
00:29:38.039 --> 00:29:44.200
by having a lot more things automated. I mean, human labor is the

383
00:29:44.240 --> 00:29:49.480
bottleneck for affordability in in many things. And the third one that that is

384
00:29:49.519 --> 00:29:55.039
actually, in my mind, even
the most exciting one, is what we

385
00:29:55.160 --> 00:30:00.559
call a personal personal assistant, and
told really be called personal personal personal assistant

386
00:30:02.599 --> 00:30:07.400
because of three different aspects of personal. So the first personal is that you

387
00:30:07.480 --> 00:30:12.640
own it. It serves your agenda, not some mega corporation's agenda. The

388
00:30:12.680 --> 00:30:18.079
second personal is that it's hyper personalized
to you, so it learns your preferences,

389
00:30:18.119 --> 00:30:22.920
your history. You know your goals
and your likes, your dislikes and

390
00:30:22.960 --> 00:30:26.960
so on. And the third personal
isn't the one of privacy that you decide

391
00:30:27.680 --> 00:30:30.359
what you want to share with whom. So you know, certain things you

392
00:30:30.400 --> 00:30:34.279
share with your spouse, other things
you share with your coworkers, and yet

393
00:30:34.319 --> 00:30:38.880
other things you might share with Amazon. So you know, all of these

394
00:30:38.880 --> 00:30:44.440
things are going to be possible with
true human level AI, and I think

395
00:30:44.519 --> 00:30:49.240
it will be tremendous to improve the
quality of life, you know, and

396
00:30:49.440 --> 00:30:56.519
basically help us flourish. And your
second point is things often describe current generive

397
00:30:56.519 --> 00:31:00.720
AI technologies to some people and other
AI technologies is what would you do if

398
00:31:00.720 --> 00:31:03.920
you had unlimited free interns, like
not super talented people, but people who

399
00:31:03.920 --> 00:31:07.759
could image, process, write basic
copy, Like how would you leverage that

400
00:31:07.799 --> 00:31:11.599
in your business? And you're talking
about a world that you have maybe not

401
00:31:11.680 --> 00:31:15.720
unlimited, maybe not free, but
a plethora of highly trained experts in different

402
00:31:15.759 --> 00:31:18.000
areas. It's a fascinating thing to
think about what you could do with that

403
00:31:18.039 --> 00:31:23.640
capability in the world. Really yeah, and the important thing is the ability

404
00:31:23.720 --> 00:31:30.680
to learn in real time. You
know that they can learn these tasks by

405
00:31:30.720 --> 00:31:34.400
themselves the way human can. It's
not then having to go through a training

406
00:31:34.440 --> 00:31:38.880
process and whatever you train is that's
limited to what it can do what the

407
00:31:38.920 --> 00:31:45.359
training is. Now, this is
truly dynamic, ongoing learning, reasoning about

408
00:31:45.400 --> 00:31:49.960
it at a high level, and
you know that it's a completely different approach

409
00:31:51.039 --> 00:31:56.960
from from the statistical generative you know
AI, and you're not going to have

410
00:31:56.039 --> 00:32:02.079
these surprises of where it says some
with great confidence that is just completely utterly

411
00:32:02.119 --> 00:32:07.279
wrong. That may be a good
mirror of some humans intelligence. Plenty of

412
00:32:07.279 --> 00:32:10.440
people who also say things with high
degree of confidence that are completely wrong,

413
00:32:10.480 --> 00:32:15.920
but maybe not the one we want
a mirror in. RAI. Uh,

414
00:32:15.119 --> 00:32:19.640
well, Peter, this has been
a fascinating conversation. I really appreciate you

415
00:32:19.680 --> 00:32:22.319
making the time today. That's given
me a lot to think about, and

416
00:32:22.400 --> 00:32:24.160
I'm sure the audience as well.
I do have a couple of traditional closing

417
00:32:24.240 --> 00:32:30.880
questions I'd love to ask you before
we before you finish up here, and

418
00:32:30.920 --> 00:32:34.119
the first of them is what's the
best piece of career advice you've ever got?

419
00:32:34.680 --> 00:32:37.960
Well, for me personally, I
can kick myself that I only started

420
00:32:37.960 --> 00:32:43.960
my first business when I was twenty
five. I should have started when I

421
00:32:44.000 --> 00:32:49.480
was sixteen. Now this is not
going to be, you know, true

422
00:32:49.519 --> 00:32:52.440
for everyone, but for me,
it's until you actually if you know,

423
00:32:52.480 --> 00:32:58.279
if you aspire to run a company
or to be part of the you know,

424
00:32:58.319 --> 00:33:02.319
the core team, you really just
need to get the experience. You

425
00:33:02.359 --> 00:33:08.160
know, it's just completely different working
for a boss as an employee than having

426
00:33:08.240 --> 00:33:13.160
the responsibility of running a company.
So if that's something you want to do,

427
00:33:13.960 --> 00:33:15.279
the sooner you can start. I
mean, whether it's part time,

428
00:33:15.519 --> 00:33:20.680
you know, whether you do it
even for free, just to gain the

429
00:33:20.720 --> 00:33:24.480
experience, but to me that you
know, if I have to say one

430
00:33:24.559 --> 00:33:30.160
regret that I only started at twenty
five. Interesting reminds me of advice about

431
00:33:30.160 --> 00:33:34.440
parenting that you can never really understand
it until you do it. And if

432
00:33:34.440 --> 00:33:37.160
you're waiting for a better time,
like there isn't really a better time,

433
00:33:37.200 --> 00:33:38.960
you just got to go out there
unprepared as you are and learn on the

434
00:33:39.039 --> 00:33:43.559
job, because it's really the only
way to learn. My last question was

435
00:33:43.599 --> 00:33:45.240
maybe we've been talking about I mean, this may be easier one for you

436
00:33:45.279 --> 00:33:49.119
because we've kind of been going on
this all days, but it's always ask

437
00:33:49.200 --> 00:33:53.039
everybody what's one bold prediction for the
future that we are going to have AGI

438
00:33:53.720 --> 00:34:00.680
and I believe it will be fantastic
for humanity. It will really help us

439
00:34:01.160 --> 00:34:06.880
manage the many difficult challenges that are
facing us. Well. I love that

440
00:34:06.880 --> 00:34:09.960
that's a great place to leave it
later, because it's certainly I think possible

441
00:34:10.000 --> 00:34:15.280
and would be an amazing opportunity and
exciting development as we get there. So

442
00:34:15.639 --> 00:34:17.519
I hope you're right on this one, and we'll be watching carefully because it's

443
00:34:17.559 --> 00:34:21.199
a very exciting time and lots of
stuff going on. So thank you so

444
00:34:21.280 --> 00:34:22.880
much. For making the time to
join us on the podcast today. Yes,

445
00:34:22.920 --> 00:34:28.199
thanks for having me. It's fun. Upstart partners with banks and credit

446
00:34:28.280 --> 00:34:31.960
unions to help grow their consumer loan
portfolios and deliver a modern, all digital

447
00:34:32.039 --> 00:34:37.320
lending experience. As the average consumer
becomes more digitally savvy, it only makes

448
00:34:37.360 --> 00:34:43.960
sense that their bank does too.
Upstarts AI lending platform uses sophisticated machine learning

449
00:34:44.000 --> 00:34:49.679
models to more accurately identify risk and
approve more applicants than traditional credit models.

450
00:34:49.880 --> 00:34:55.119
With fraud rates near zero, Upstarts
All Digital experience reduces manual processing for banks

451
00:34:55.440 --> 00:35:00.360
and offers a simple and convenient experience
for consumers. Whether you're looking to grow

452
00:35:00.440 --> 00:35:06.039
and enhance your existing personal and auto
lending programs or you're just getting started,

453
00:35:06.400 --> 00:35:09.840
Upstart can help. Upstart offers an
end to end solution that can help you

454
00:35:09.880 --> 00:35:15.679
find more credit worthy borrowers within your
risk profile. With all digital underwriting,

455
00:35:15.880 --> 00:35:20.840
onboarding, loan closing, and servicing, It's all possible with Upstart in your

456
00:35:20.880 --> 00:35:25.199
corner. Learn more about finding new
borrowers, enhancing your credit decisioning process,

457
00:35:25.239 --> 00:35:30.760
and growing your business by visiting upstart
dot com Slash four dash Banks That's upstart

458
00:35:30.800 --> 00:35:36.760
dot Com slash four dash banks.
You've been listening to Leaders in Lending from

459
00:35:36.800 --> 00:35:40.119
upstart. Make sure you never miss
an episode. Subscribe to Leaders in Lending

460
00:35:40.199 --> 00:35:45.079
in your favorite podcast player using Apple
Podcasts. Leave us a quick rating by

461
00:35:45.119 --> 00:35:47.360
tapping the number of stars you think
the show deserves. Thanks for listening,

462
00:35:47.719 --> 00:35:49.440
until next time.