WEBVTT
1
00:00:01.280 --> 00:00:07.040
Forbes Books presents the Startup Science Podcast
with Gregory Shepherd, brought to you by
2
00:00:07.080 --> 00:00:13.720
Startupscience dot Io. Greg is an
entrepreneur who has built and sold twelve businesses.
3
00:00:14.160 --> 00:00:18.760
He's a recipient of four Private Equity
awards and is a featured ten X
4
00:00:18.800 --> 00:00:24.559
speaker and Forbes author. Here's Gregory
Shepherd. My guest today is David Danks,
5
00:00:24.719 --> 00:00:29.120
a professor of data science and philosophy
at the University of California, San
6
00:00:29.199 --> 00:00:35.280
Diego. His research explores the intersection
of philosophy, cognitive science, and machine
7
00:00:35.359 --> 00:00:43.560
learning, tackling complex interdisciplinary problems.
He's examined the ethical, psychological, and
8
00:00:43.679 --> 00:00:50.159
policy implications of AI robotics in various
domains like transportation and healthcare. David,
9
00:00:50.159 --> 00:00:53.159
I'm excited to have you on the
podcast, and I'd love it if you
10
00:00:53.159 --> 00:00:57.520
could start by talking a little bit
about all of the incredible work you've done
11
00:00:57.880 --> 00:01:02.719
reaching these different fields if your brilliant
mind. Well, thank you for that
12
00:01:03.119 --> 00:01:08.200
very kind introduction. So I'm here, you see San Diego, and I
13
00:01:08.200 --> 00:01:14.719
do everything from actually build AI systems, primarily with biomedical data, through to
14
00:01:14.840 --> 00:01:19.239
thinking about the ethics and policy of
AI. On top of that, a
15
00:01:19.239 --> 00:01:23.599
lot of issues around philosophy more generally, So, what can we possibly know
16
00:01:23.719 --> 00:01:26.879
about the world, How do we
come to know about the world, what's
17
00:01:26.920 --> 00:01:32.040
the nature of science. I'm one
of these people that I've discovered I don't
18
00:01:32.079 --> 00:01:36.439
actually have any one big problem that
I'm trying to solve. Many academics do.
19
00:01:37.159 --> 00:01:41.519
Instead, I'm just kind of interested
in everything and lucky enough to be
20
00:01:41.640 --> 00:01:45.439
in a position that I get to
pursue all of these different ideas that might
21
00:01:45.480 --> 00:01:49.959
come to my mind. So when
it comes to AI and ethics and poopy,
22
00:01:49.000 --> 00:01:56.319
it seems like philosophy and this new
realm we're entering with AI, there
23
00:01:56.319 --> 00:02:00.519
would be a lot of benefits to
understanding philosophy. Right, how do you
24
00:02:00.599 --> 00:02:04.560
look at the way we're using AI, or what the future holds for AI,
25
00:02:04.640 --> 00:02:07.639
that sort of stuff. Do you
have any perspectives on that completely agree
26
00:02:07.680 --> 00:02:15.439
with you that philosophy is invaluable in
our current tying milieu. I think that's
27
00:02:15.479 --> 00:02:20.759
not always obvious to people, in
part because I think many people misunderstand the
28
00:02:20.840 --> 00:02:27.240
nature of philosophy. So I think
philosophy is sometimes thought of as it's either
29
00:02:27.639 --> 00:02:31.560
the navel gazing that occurs in a
dorm room at midnight or it's the province
30
00:02:31.639 --> 00:02:37.400
of a bunch of dead white guys
from Greece onward. But what philosophy is
31
00:02:37.439 --> 00:02:42.479
really about is it's about asking the
foundational questions. It's about asking those questions
32
00:02:42.960 --> 00:02:46.400
that in most of the parts of
our lives we don't have the time or
33
00:02:46.400 --> 00:02:51.639
even the opportunity in space to stop
and ask them what is the nature of
34
00:02:51.680 --> 00:02:54.879
knowledge? What is the good life? And so when we find ourselves in
35
00:02:54.919 --> 00:02:59.719
a time of significant disruption, I
think it's all the more important that we're
36
00:02:59.719 --> 00:03:04.520
able to go back to those foundations, to have a firm ground at least
37
00:03:05.159 --> 00:03:08.080
about our own values, what we
think is important, well, we're trying
38
00:03:08.120 --> 00:03:14.639
to achieve in our lives, and
philosophy is ultimately about giving people the skills
39
00:03:14.639 --> 00:03:19.280
and experience to be able to ask
those questions about their own foundations. Do
40
00:03:19.319 --> 00:03:23.639
you have or are you still working
on a philosophy of AI? Right?
41
00:03:23.719 --> 00:03:27.360
Like I think about it like this, it's moving so fast. It's been
42
00:03:27.400 --> 00:03:30.759
around for a long time, but
it's just hit this point where it's like
43
00:03:30.919 --> 00:03:35.159
really really moving fast and people are
starting to apply it in all these different
44
00:03:35.199 --> 00:03:38.879
areas. And when you think about
the philosophy of AI, are you thinking
45
00:03:38.960 --> 00:03:44.800
about it in terms of, you
know, it should be used for this
46
00:03:44.919 --> 00:03:49.159
and not used for that, or
we can expect this to happen in the
47
00:03:49.199 --> 00:03:52.199
future, you know, and that
sort of stuff. I'm just coming off
48
00:03:52.199 --> 00:03:57.240
of reading Sapiens, so I'd like
to look to hear about that. So
49
00:03:57.319 --> 00:04:02.240
a lot of my own work in
philosophy and has focused on understanding the ways
50
00:04:02.240 --> 00:04:08.479
in which our values get implemented and
bansed almost kind of expressed through the AI
51
00:04:08.560 --> 00:04:12.360
systems that we build. And I
think a lot of people think of AI
52
00:04:12.400 --> 00:04:15.479
as just code, it's just the
ones and zeros, or it's math.
53
00:04:15.040 --> 00:04:19.560
But I'm much more addrested in the
way that all of the different design choices,
54
00:04:19.680 --> 00:04:24.759
the choices made by the developers and
creators of AI systems, all the
55
00:04:24.759 --> 00:04:30.079
ways that those choices then lead to
the systems being designed for particular purposes,
56
00:04:30.279 --> 00:04:34.240
that they benefit some people and not
others, that they're good in some context
57
00:04:34.319 --> 00:04:40.000
and not others, And so really
trying to connect the values of the people
58
00:04:40.120 --> 00:04:46.399
creating the AI, the people using
the AI, connecting those values with the
59
00:04:46.560 --> 00:04:50.319
values that matter to us in our
lives. So I worry that a lot
60
00:04:50.319 --> 00:04:55.319
of the AI systems that are being
built these days, they are benefiting just
61
00:04:55.360 --> 00:04:59.120
a few people and not others.
So this means that I end up not
62
00:04:59.160 --> 00:05:01.920
thinking all that much of about the
definitional question what counts as AI, what
63
00:05:02.040 --> 00:05:10.120
doesn't count as AI, or thinking
about very abstract questions about what kinds of
64
00:05:10.120 --> 00:05:15.920
AIS might exist in thirty years.
And I'm much more trying to think about
65
00:05:15.360 --> 00:05:19.319
my neighbor who's a retired lawyer and
is trying to figure out should she be
66
00:05:19.399 --> 00:05:24.759
freaked out about AI, or the
person who's worried that AI is going to
67
00:05:24.759 --> 00:05:28.600
take their job. How do we
think about how to create AI that empowers
68
00:05:28.639 --> 00:05:33.959
people rather than just building things that
are seem cool to me but might end
69
00:05:34.040 --> 00:05:39.439
up having a lot of harmful side
effects. Yeah, so let me ask
70
00:05:39.480 --> 00:05:43.959
you about this. I was talking
to a pretty smart fellow in London a
71
00:05:44.040 --> 00:05:46.399
few weeks ago. What he said
is he said, listen, everybody was
72
00:05:46.439 --> 00:05:50.680
afraid of losing their jobs because of
robotics, and they did, but then
73
00:05:50.720 --> 00:05:55.879
they got reemployed. And AI,
on the other hand, is the other
74
00:05:56.000 --> 00:05:58.439
end of the spectrum. Right,
So you have the robotics, which is
75
00:05:58.879 --> 00:06:00.199
back to the blue collar work,
and then you have AI, which is
76
00:06:00.240 --> 00:06:04.240
going to affect the white collar workers, right, like doctors, lawyers.
77
00:06:04.680 --> 00:06:09.240
AI is just going to do a
better job than they could do in diagnosing
78
00:06:09.360 --> 00:06:13.759
or drafting or whatever it is.
When you think about that, do you
79
00:06:14.199 --> 00:06:17.639
feel like there is going to be
a point at which people aren't working as
80
00:06:17.720 --> 00:06:20.759
much? Right? So this guy
said, he said, listen, you
81
00:06:20.800 --> 00:06:23.879
know, it's not like people aren't
going to work. They're just going to
82
00:06:23.959 --> 00:06:28.240
work four hours instead of twelve hours
or whatever to do the same thing and
83
00:06:28.240 --> 00:06:30.360
get paid the same money. Do
you feel like that's a future that in
84
00:06:30.439 --> 00:06:34.519
terms of an optimistic outview, do
do you feel like that's a future that
85
00:06:34.800 --> 00:06:38.319
is possible or what are your thoughts
on that? So? I think that's
86
00:06:38.319 --> 00:06:43.480
a future that is possible, but
it's not going to happen without really concerted
87
00:06:43.519 --> 00:06:47.279
effort. Somewhat Famously, John Maynard
Kanes, the economist, thought that the
88
00:06:47.319 --> 00:06:53.120
steam engine and industrialization was going to
mean that by the early twentieth century,
89
00:06:53.160 --> 00:06:56.839
everybody would be working fifteen to twenty
hours. It'd have lives of leisure.
90
00:06:57.319 --> 00:07:01.160
Obviously that's not how it turned out. I think we're facing a similar sort
91
00:07:01.160 --> 00:07:05.879
of choice here with AI. Do
we take advantage of the massive productivity gains
92
00:07:05.920 --> 00:07:12.839
that we can achieve through AI to
reshape in certain ways quite frankly, our
93
00:07:12.920 --> 00:07:17.959
fundamental society. Do we have things
like universal basic income that enable people to
94
00:07:18.000 --> 00:07:23.319
step away from work and pursue their
passions, which for many people passion is
95
00:07:23.360 --> 00:07:26.680
not the same as work. That's
the part of which I start to be
96
00:07:26.800 --> 00:07:31.079
a little bit less optimistic, because
I think right now it's not clear to
97
00:07:31.120 --> 00:07:35.600
me that there is the political will, the political direction, the political coherence
98
00:07:36.240 --> 00:07:42.759
to be able to get us towards
that better future. I think it's possible,
99
00:07:43.199 --> 00:07:47.199
but unfortunately not probable. In fact, I worry in many ways that
100
00:07:47.240 --> 00:07:51.040
the risks are much more on the
downside than on the upside, in the
101
00:07:51.120 --> 00:07:59.240
sense that right now these past technological
changes, even robotics, but if we
102
00:07:59.360 --> 00:08:03.480
go back earlier to you know,
cars replacing courses and these kinds of things,
103
00:08:03.879 --> 00:08:07.560
as you said, people have lost
jobs, but they've always been able
104
00:08:07.600 --> 00:08:11.639
to get new jobs afterwards. And
so the challenge though is that there's a
105
00:08:11.680 --> 00:08:16.959
period of disruption in the middle,
and historically that's been ten plus years.
106
00:08:18.000 --> 00:08:22.519
It takes a long time for a
new technology to come into widespread use,
107
00:08:22.959 --> 00:08:26.920
in large part because technologies usually are
physical things, so just you have to
108
00:08:26.920 --> 00:08:31.879
build more of them, right,
So it takes time to convert all factories
109
00:08:31.920 --> 00:08:35.080
into being primarily run by robots because
we have to build a lot of robots
110
00:08:35.080 --> 00:08:41.000
to do that. AI costs almost
nothing to duplicate it, And so I
111
00:08:41.039 --> 00:08:46.559
think that the speed of adoption can
be much faster with AI, which also
112
00:08:46.600 --> 00:08:50.480
means the speed of disruption and dislocation
of workers is going to be much faster
113
00:08:50.559 --> 00:08:56.879
with AI. And I don't know
if our society can handle suddenly having twenty
114
00:08:56.919 --> 00:09:01.759
percent unemployment if it's not spread out
over ten years. I worry that there's
115
00:09:01.799 --> 00:09:11.759
actually a much bigger risk of significant
disruption socially rather than the positive upside possibility.
116
00:09:13.120 --> 00:09:16.600
So when you think about this,
so I'm going to sort of move
117
00:09:16.759 --> 00:09:22.200
my way into entrepreneurship and startups and
stuff like that. When you think about
118
00:09:22.200 --> 00:09:24.600
this, you think about industrial transfer. Right now, in the whole startup
119
00:09:24.600 --> 00:09:30.039
ecosystem, industrial transformation is like the
big thing, right everybody thinks AI is
120
00:09:30.039 --> 00:09:33.279
the big thing. The industrial transformation
is one hundred times bigger because it's basically
121
00:09:33.440 --> 00:09:39.000
taking the industrial Revolution and doing it
all over again with a big, giant
122
00:09:39.080 --> 00:09:43.840
upgrade. Would you put AI in
the category of industrial transformation or do you
123
00:09:43.840 --> 00:09:48.200
think it's inside and outside of that. I think that AI is a key
124
00:09:48.279 --> 00:09:54.080
part of the industrial transformations that are
occurring right now. I think it is
125
00:09:54.120 --> 00:09:58.919
also somewhat separate in some ways.
I think what AI does is it's fundamentally
126
00:10:00.080 --> 00:10:03.159
an optimization technique. I mean,
that's really almost all AI comes down to.
127
00:10:03.279 --> 00:10:09.240
We're optimizing a process, a mathematical
function code to do something, whether
128
00:10:09.279 --> 00:10:15.000
it's predict the next word or figure
out who has cancer. So we can
129
00:10:15.039 --> 00:10:20.919
apply those optimization techniques in industrial transformations, in optimizing factory workflows, in optimizing
130
00:10:22.000 --> 00:10:24.039
supply chains. You know, it's
one of the big areas that we're now
131
00:10:24.080 --> 00:10:28.919
seeing AI being used is in optimizing
logistics. It's been around for it for
132
00:10:28.000 --> 00:10:31.519
a while, and it just keeps
becoming more and more important. On the
133
00:10:31.600 --> 00:10:35.679
other hand, one of the things
that I think is important about AI is
134
00:10:35.720 --> 00:10:39.799
that it also can provide new capabilities
and a lot of industrial transformation. At
135
00:10:39.840 --> 00:10:46.039
least in my experiences with it,
is focused on how do we more efficiently
136
00:10:46.200 --> 00:10:50.480
do what we've done before. One
of the key changes with AI is that
137
00:10:50.519 --> 00:10:56.039
we start to expand our horizons of
the possible, at least in the good
138
00:10:56.120 --> 00:11:01.519
cases, so that we start to
really understand that we might not just be
139
00:11:01.600 --> 00:11:05.000
able to replace a human with a
machine, but rather get a human to
140
00:11:05.080 --> 00:11:09.919
do so much more than they could
before by being augmented by the machine.
141
00:11:11.399 --> 00:11:16.440
And so I think that there's opportunities
with AI to do much more than we're
142
00:11:16.480 --> 00:11:20.080
seeing in a lot of the kind
of standard industrial transformation approaches. But it's
143
00:11:20.120 --> 00:11:26.960
also a really important element of successful
transformation. Interesting. I love that.
144
00:11:28.039 --> 00:11:33.320
So it's almost like when computers came
out and we became one hundred thousand times
145
00:11:33.360 --> 00:11:37.159
more effective, and AI is like
another boost, right, Yeah, And
146
00:11:37.440 --> 00:11:41.720
I think, you know, the
introduction of computers, or let's be even
147
00:11:41.799 --> 00:11:46.200
more granular, the widespread use of
spell check. You know, spell check
148
00:11:46.240 --> 00:11:50.080
means that we all lost a little
bit of our spelling ability. I know
149
00:11:50.159 --> 00:11:52.559
I did. On the other hand, it means that we can focus on
150
00:11:52.639 --> 00:11:56.519
what really matters, which is the
expression of our ideas, and we're not
151
00:11:56.639 --> 00:12:01.399
bogged down necessarily by the details of
our we spell something. Word processing is
152
00:12:01.440 --> 00:12:05.519
similar. A lot of people who
are typists lost their jobs because of word
153
00:12:05.519 --> 00:12:09.799
processing programs. On the other hand, it makes it much easier to rapidly
154
00:12:09.840 --> 00:12:16.240
iterate, revise, and what we've
seen is large changes in the number of
155
00:12:16.279 --> 00:12:22.879
people who are able to express their
ideas more clearly, faster and to a
156
00:12:22.919 --> 00:12:26.600
broader audience. So, you know, computers don't just let us do all
157
00:12:26.639 --> 00:12:31.159
the things we did before faster.
They can lead to these real changes,
158
00:12:31.480 --> 00:12:39.559
almost phase changes in what we see
as possible in business contexts. That's really
159
00:12:39.559 --> 00:12:43.840
interesting. So let's move from there. So, if you are an entrepreneur
160
00:12:43.080 --> 00:12:48.480
right now and you're looking at AI
opportunities, the first thing you have to
161
00:12:48.480 --> 00:12:54.200
think about, right, is the
considerations for building trustworthy systems, right that
162
00:12:54.279 --> 00:12:58.200
align with the social values to your
point earlier, So, what are some
163
00:12:58.240 --> 00:13:05.799
of the key siderations an entrepreneur should
have when looking at their responsibility for this
164
00:13:05.919 --> 00:13:09.879
powerful tool, you know, to
prevent it from being weaponized, you know,
165
00:13:09.919 --> 00:13:13.200
as an example. Yeah, So
I think the first thing is,
166
00:13:13.480 --> 00:13:16.840
and I think this is probably good
advice for any entrepreneur, is know your
167
00:13:16.879 --> 00:13:20.440
audience. Know your customers, Know
your clients, and by that I mean
168
00:13:20.559 --> 00:13:24.240
know what matters to them, know
what they're trying to achieve, and know
169
00:13:24.320 --> 00:13:30.320
how what you're building will actually serve
their needs and interests, not what you
170
00:13:30.519 --> 00:13:35.240
think their needs and interests. Are. So there's a long history of people
171
00:13:35.320 --> 00:13:39.120
thinking that they know, oh,
here's what our customers want, when in
172
00:13:39.159 --> 00:13:43.279
fact the customers wanted something very different. Now, of course you can't ask
173
00:13:43.399 --> 00:13:46.240
people necessarily what they want. People
aren't always good at expressing it, as
174
00:13:46.480 --> 00:13:50.679
you know, Steve Jobs famously said, you know, people will know what
175
00:13:50.720 --> 00:13:54.480
they want when we give it to
them. They and I think that that's
176
00:13:54.799 --> 00:13:58.960
right. But that's all the more
reason to engage closely with your customer base,
177
00:14:00.080 --> 00:14:03.759
with your potential clients, so that
they can almost help co design what
178
00:14:03.840 --> 00:14:07.360
it is that you're building. You
can show them, hey, if we
179
00:14:07.399 --> 00:14:09.799
could do this, is that something
you would like? Do you think that
180
00:14:11.120 --> 00:14:13.639
AI could help with that? Because
that's a big problem, right, Like
181
00:14:13.799 --> 00:14:18.000
to your point, you know,
it is almost impossible to get customers to
182
00:14:18.039 --> 00:14:24.799
tell you what they want. It's
you have to decipher and translate what they
183
00:14:24.879 --> 00:14:26.759
tell you to turn it into something
that you could actually do. Do you
184
00:14:26.759 --> 00:14:31.039
think AI could help with that process, like conversationally or something. So,
185
00:14:31.120 --> 00:14:33.960
I think AI can help with that
in a couple of different ways. The
186
00:14:35.000 --> 00:14:41.000
first is that you could use conversational
generative AI systems to basically have free flowing
187
00:14:41.039 --> 00:14:46.360
conversations with your customers and your clients
without it requiring your time and energy.
188
00:14:46.639 --> 00:14:50.720
See I basically get a chatbot to
elicit a lot of these things and to
189
00:14:50.840 --> 00:14:54.879
explore scenarios, so you could design. You can have a chatbot that essentially,
190
00:14:54.960 --> 00:14:58.799
you know, an LLM for example, a large language model that essentially
191
00:14:58.919 --> 00:15:01.360
generates a whole bunch of of scenarios
and says, in this case, what
192
00:15:01.399 --> 00:15:03.440
would you want to do? In
this case, what would you want to
193
00:15:03.480 --> 00:15:07.399
do? The other place that I
think AI can really help is in rapid
194
00:15:07.440 --> 00:15:15.039
prototyping and exploring the possibility space.
So AI systems, especially generative ones,
195
00:15:15.120 --> 00:15:22.759
can very quickly create mockups of systems, can create sort of simulations of new
196
00:15:22.840 --> 00:15:28.279
capabilities, and so you can put
people into a more realistic, a more
197
00:15:28.360 --> 00:15:31.159
naturalistic setting and say, you know, would you use this in your workflows
198
00:15:31.279 --> 00:15:35.480
rather than just asking people, hey, is this something you would like?
199
00:15:35.519 --> 00:15:39.879
Which I think can often suffer when
people have failures of imagination. They might
200
00:15:39.919 --> 00:15:46.120
not realize what kind of a cool
thing they could create. So I wonder
201
00:15:46.200 --> 00:15:52.080
what regulations a founder should be cautious
of and how would they figure that out?
202
00:15:52.200 --> 00:15:56.639
You know, because they're passing regulation
right, So, if you're building
203
00:15:56.759 --> 00:16:00.000
a AI system, is there somewhere
you need to go to make sure you're
204
00:16:00.039 --> 00:16:07.120
not going to get yourself in trouble
for building something that is using potentially the
205
00:16:07.279 --> 00:16:11.440
entire Internet for its knowledge base.
Yeah, that's a great question. It
206
00:16:11.600 --> 00:16:15.759
is one that I found myself spending
far more time than I expected over the
207
00:16:15.840 --> 00:16:19.720
last couple of years thinking about and
talking with folks in places like DC about
208
00:16:19.759 --> 00:16:25.360
what's going on. So the first
thing that is really important for an entrepreneur
209
00:16:25.440 --> 00:16:33.039
to know is that right now we're
seeing basically the development of a patchwork approach
210
00:16:33.159 --> 00:16:38.200
to AI regulation that depends a lot
on where you're located. So the approach
211
00:16:38.240 --> 00:16:42.360
to regulation in the United States is
very different from the approach in Germany.
212
00:16:42.360 --> 00:16:45.919
It's very different from the approach in
China's very different from the approach in India,
213
00:16:47.799 --> 00:16:52.360
just to name four locations. One
of the first things is to recognize
214
00:16:52.600 --> 00:16:57.720
that just because you've checked all the
regulatory boxes to deploy a model in France
215
00:16:59.200 --> 00:17:00.919
does not mean that you going to
be on safe ground when it comes to
216
00:17:02.000 --> 00:17:04.839
the United States. And the reason
I wanted to make sure to call this
217
00:17:04.880 --> 00:17:11.240
out is obviously the idea that different
nations have different regulations, is not anything
218
00:17:11.319 --> 00:17:18.960
new, but AI and AI companies
are not bound by borders in the usual
219
00:17:18.000 --> 00:17:22.799
ways. Yeah, that's a really
good point, right, because it's like
220
00:17:22.920 --> 00:17:27.440
three exactly, and it's very easy
to build out an AI system and suddenly
221
00:17:27.480 --> 00:17:32.839
find people around the world using it, especially if you just throw it up
222
00:17:32.839 --> 00:17:36.680
on one of these standard cloud services, it can be accessed anywhere in the
223
00:17:36.720 --> 00:17:41.000
world. And so I think that
there is a much more complex regulatory environment
224
00:17:41.039 --> 00:17:45.319
that entrepreneurs are entering when it comes
to AI than many of them realize.
225
00:17:45.920 --> 00:17:51.480
And so I think that there's a
large amount of thought that entrepreneurs need to
226
00:17:51.480 --> 00:17:55.200
put into how are they going to
make sure that they don't step on any
227
00:17:55.240 --> 00:17:59.440
regulatory land lines. Now I mentioned
the different approaches, let me just briefly
228
00:17:59.519 --> 00:18:03.359
say, you know, for example, in the United States, the regulation
229
00:18:03.400 --> 00:18:07.440
of AI is happening in a very
sector specific way. So if I'm building
230
00:18:07.440 --> 00:18:10.920
an AI system in healthcare, I'm
going to have to know about HIPPA,
231
00:18:10.920 --> 00:18:15.440
the health privacy law that we have
in the United States. In contrast,
232
00:18:15.680 --> 00:18:18.200
if I'm building AI that's going to
operate on social media, well, in
233
00:18:18.240 --> 00:18:22.240
the United States, social media is
largely unregulated and so there isn't going to
234
00:18:22.240 --> 00:18:26.920
be a particular regulation on my AI
system. In contrasting the EU, there's
235
00:18:27.039 --> 00:18:30.599
what's called the AI Act, which
was just recently finalized. It's going to
236
00:18:30.640 --> 00:18:34.400
be going into effect in the next
couple of years, and the EUAI Act
237
00:18:34.440 --> 00:18:38.680
is much more focused on AI writ
Larch. It's a sector general kind of
238
00:18:38.720 --> 00:18:44.759
regulation that essentially says, if you're
going to be using AI systems in high
239
00:18:44.839 --> 00:18:48.839
risk ways or in ways that creates
systemic risks. So if you're building a
240
00:18:48.880 --> 00:18:52.599
generative AI system that's going to be
at the foundation of many different models,
241
00:18:52.559 --> 00:18:56.759
then there's a whole set of auditing
and testing that you have to do before
242
00:18:56.839 --> 00:19:02.519
you ever deployed in the first place. So if I'm an AI developer in
243
00:19:02.559 --> 00:19:06.400
Germany, my brain is thinking about, all right, what are the high
244
00:19:06.480 --> 00:19:10.440
risk use cases that this might be
put to, even if I don't intend
245
00:19:10.519 --> 00:19:12.319
it to be used in that way, Whereas in the US, what I'm
246
00:19:12.319 --> 00:19:15.839
thinking is, Okay, where are
the sectors I want to deploy and what
247
00:19:15.960 --> 00:19:21.799
are the regulations that are specific to
that sector. Now you also ask,
248
00:19:21.880 --> 00:19:25.160
I think a question which is very
near and dear to my heart because it's
249
00:19:25.200 --> 00:19:29.279
part of how I got into thinking
about issues around ethics and technology and ethics
250
00:19:29.279 --> 00:19:33.599
and AI in the first place,
which is how should we think about the
251
00:19:33.640 --> 00:19:38.119
ways that our algorithms and models might
end up being used for purposes that we
252
00:19:38.200 --> 00:19:41.920
never intended. Yeah, that's what
I'm worried about, and you're right to
253
00:19:41.960 --> 00:19:47.400
worry about it. That's part of
what drew me into ethics of AI is
254
00:19:47.440 --> 00:19:52.119
in part learning that an algorithm I
had helped to develop probably was used for
255
00:19:52.200 --> 00:19:56.279
purposes that I didn't necessarily agree with. So the first thing to say is
256
00:19:56.720 --> 00:20:00.680
that's less of a risk I think
actually for many entrepreneurs than it is for
257
00:20:00.799 --> 00:20:06.640
folks like me and academia, because
entrepreneurs are typically building products, which means
258
00:20:06.640 --> 00:20:10.960
that there's going to be a lot
of work that goes into translating the algorithm
259
00:20:10.960 --> 00:20:14.960
and model into the particular use cases
of your customers, your clients, whether
260
00:20:15.000 --> 00:20:19.079
those are the businesses or consumers.
In contrast, academics, we're much more
261
00:20:19.079 --> 00:20:23.400
interested in what is the general algorithm
that could be used in many different contexts.
262
00:20:23.640 --> 00:20:30.680
Oh, I see broad exactly.
AI is great because it's kind of
263
00:20:30.799 --> 00:20:36.079
domain agnostic. Right. In general
data science machine learning, we kind of
264
00:20:36.160 --> 00:20:40.119
don't care where the data come from. We're operating on the patterns that we
265
00:20:40.200 --> 00:20:45.839
find in patterns are independent of what
their patterns over. At the same time,
266
00:20:45.680 --> 00:20:51.680
obviously you're going to tune your algorithm
to the particular assumptions you can make
267
00:20:51.720 --> 00:20:56.880
the particular features of the problem,
and so as you get closer and closer
268
00:20:56.880 --> 00:21:00.920
to an actual product. I think
the domain generality, it's less and less
269
00:21:00.440 --> 00:21:04.359
even with the things that we think
of now as the most general purpose AI
270
00:21:04.480 --> 00:21:11.119
systems called foundation models or generative AI, the large language models, even those
271
00:21:11.200 --> 00:21:19.359
systems, we are increasingly seen that
very few actual applications in business or in
272
00:21:19.920 --> 00:21:26.119
society generally, very few of the
applications are based on the untuned foundation models.
273
00:21:26.680 --> 00:21:32.640
Most of what's happening even with something
like GPT four is that companies are
274
00:21:32.759 --> 00:21:37.160
fine tuning it in ways that make
it much more focused, specific and accurate
275
00:21:37.799 --> 00:21:42.559
for their particular needs. So I
think we're even though we talk a lot
276
00:21:42.599 --> 00:21:48.000
about general purpose AI, I don't
think that it happens as much in practice.
277
00:21:48.440 --> 00:21:52.480
And so as long as the entrepreneurs
retaining some measure of control over their
278
00:21:52.480 --> 00:21:59.480
intellectual property, they can help to
shape where the technology goes basically like commercial
279
00:21:59.519 --> 00:22:06.079
AI is relatively low risk, but
then broad AI is more high risk because
280
00:22:06.240 --> 00:22:11.039
one of the things that I notice
is that if I aim the AI at
281
00:22:11.160 --> 00:22:15.160
data, it uses that data to
return my answers. So if you aimed
282
00:22:15.279 --> 00:22:19.160
at the entire internet, we all
know there's a lot of trash on the
283
00:22:19.160 --> 00:22:23.119
Internet, right, Am I getting
the truth? Or am I getting somebody
284
00:22:23.160 --> 00:22:26.720
who has? You know, carpet
bombed the Internet with the way that they
285
00:22:26.759 --> 00:22:30.440
think, right, So we saw
in a couple of election cycles we saw
286
00:22:30.480 --> 00:22:37.039
this massive insurgence of flooding the internet, you know, carpet bombing the Internet
287
00:22:37.079 --> 00:22:41.079
with certain belief patterns that it seems
like if AOI was in place and somebody
288
00:22:41.119 --> 00:22:45.759
asked a question, that would be
the answer that they would get because there's
289
00:22:45.960 --> 00:22:49.680
so much data you know, reflecting
that answer, Like how do you think
290
00:22:49.720 --> 00:22:55.920
about that? I mean, that
seems like a pretty scary scenario and dividing
291
00:22:56.160 --> 00:23:00.720
you know, subs even more so, Yeah, it's a massive problem.
292
00:23:02.599 --> 00:23:07.559
Just recently, Google came out with
their AI summaries that's supposed to help with
293
00:23:07.599 --> 00:23:14.960
the search results, So it summarizes
the leading responses that were found through search
294
00:23:15.400 --> 00:23:18.039
so that you might not even have
to click through, and it's done.
295
00:23:18.079 --> 00:23:22.519
Things that have suggested, for example, that you should perhaps add glue to
296
00:23:22.599 --> 00:23:29.759
your cheese on your pizza so that
it is gooier and flows better, which,
297
00:23:30.200 --> 00:23:33.839
as I understand it, actually is
true of advertising. Advertisers often do
298
00:23:33.960 --> 00:23:37.599
things to the food in order to
make it look better. You don't want
299
00:23:37.599 --> 00:23:41.480
to eat it. Though. That's
a case where there's information on the Internet
300
00:23:41.599 --> 00:23:48.319
that is being misused or misapplied,
shall we say, by the search summarization
301
00:23:48.599 --> 00:23:52.640
system. Of course, it goes
beyond just the sort of garbage that's out
302
00:23:52.640 --> 00:23:57.240
there on the Internet in various places, to these systems actually completely. They're
303
00:23:57.279 --> 00:24:02.640
just trying to predict what the most
next few words are supposed to be,
304
00:24:02.759 --> 00:24:06.839
so they will make up things that
sound plausible. We call this hallucination.
305
00:24:07.000 --> 00:24:10.640
It's not the best term, but
that's okay. They'll essentially make things up.
306
00:24:10.920 --> 00:24:15.440
Rather infamously, many of us in
academia, when things like CHATCHBT first
307
00:24:15.519 --> 00:24:18.039
came out, we would say,
oh, what are the papers in my
308
00:24:18.200 --> 00:24:22.200
area that I should know about?
It would generate a lot of very plausible
309
00:24:22.200 --> 00:24:26.359
sounding paper names, none of which
existed, and in fact, recently one
310
00:24:26.359 --> 00:24:30.880
of the CEOs of the big tech
companies came out and just said, we're
311
00:24:30.960 --> 00:24:33.640
never going to solve these problems.
These systems, these generative AI systems,
312
00:24:33.680 --> 00:24:40.319
are always going to hallucinate. They're
always going to return problematic output in at
313
00:24:40.400 --> 00:24:45.480
least some cases. So it's a
huge problem and it's one that we're never
314
00:24:45.519 --> 00:24:49.759
going to get rid of. Having
said all of that, we can do
315
00:24:49.839 --> 00:24:53.640
so much better than we are right
now. There are a lot of techniques
316
00:24:53.720 --> 00:25:00.359
and ways to help people know for
these kinds of systems what we call provenence
317
00:25:00.559 --> 00:25:04.319
sort of the history of where the
information came from. And that's one of
318
00:25:04.319 --> 00:25:10.160
the big advances we've seen in the
last six to nine months is improved systems
319
00:25:10.200 --> 00:25:15.680
that can reveal the provenance of information
to people that are using these AI systems.
320
00:25:15.799 --> 00:25:19.200
Can these systems can trace back the
history of well, the reason I've
321
00:25:19.240 --> 00:25:22.480
given you this information is that it
came from this site. This site is
322
00:25:22.519 --> 00:25:27.920
reputable for these reasons. And we're
starting to see the emergence of technologies that
323
00:25:29.000 --> 00:25:33.599
can help people more critically evaluate the
information sources that are being put in front
324
00:25:33.640 --> 00:25:37.720
of them by transparency. Just given
this, because that's one of the things
325
00:25:37.759 --> 00:25:42.240
I worry about right if AI is
doing the thinking and we're not. Are
326
00:25:42.240 --> 00:25:45.039
we just going to get dumber,
just like we don't spell as well,
327
00:25:45.359 --> 00:25:49.720
so our cognition is going to change. I quite deliberately didn't use the term
328
00:25:49.799 --> 00:25:56.440
dumber a fair enough. There's there's
a phenomenon that's referred to as de skilly.
329
00:25:56.559 --> 00:26:00.519
It's the idea that if you have
a skill and then a machine or
330
00:26:00.519 --> 00:26:03.559
some automation starts to do that skill
for you, or even another human starts
331
00:26:03.599 --> 00:26:07.160
doing that skill for you, that
over time you'll lose that skill. You
332
00:26:07.200 --> 00:26:11.960
will d skill. That happens all
the time. When I was younger,
333
00:26:12.359 --> 00:26:18.680
much younger, I could drive a
manual transmission car pretty well. I still
334
00:26:18.759 --> 00:26:21.559
can kind of drive one, but
you wouldn't want to put me behind the
335
00:26:21.559 --> 00:26:25.960
wheel on a regular basis at the
moment because I've deskilled. I've lost that
336
00:26:26.000 --> 00:26:29.920
particular ability that one. Though.
Honestly, I'm okay with like. It
337
00:26:30.160 --> 00:26:33.920
doesn't bother me. I very rarely
need to drive a manual transmission car,
338
00:26:33.039 --> 00:26:37.960
so I'm okay with de skilling in
that way. Selective deskilling selective discilling exactly.
339
00:26:40.000 --> 00:26:44.119
So what that points to, though, is the importance that as we're
340
00:26:44.119 --> 00:26:48.279
bringing AI systems into our workflows,
into our companies, into our daily lives.
341
00:26:48.000 --> 00:26:55.200
We actually need to be intentional and
thoughtful about what skills are we likely
342
00:26:55.279 --> 00:27:00.720
to lose as a result, and
are we okay with that? So the
343
00:27:00.799 --> 00:27:04.279
three most powerful words when we're thinking
about technology adoption, in my mind,
344
00:27:04.359 --> 00:27:08.880
are are we sure? Just stopping
and asking are we sure that this is
345
00:27:08.880 --> 00:27:11.759
a good idea? Are we sure
it's okay that we lose that skill?
346
00:27:12.640 --> 00:27:17.960
I'm really glad that people thought about
that question when it came to the rollout
347
00:27:18.000 --> 00:27:22.839
of autopilots and airplanes. Autopilots can
actually land airplanes better than most commercial pilots
348
00:27:22.880 --> 00:27:29.359
would be able to, but we
don't allow commercial airlines to land on autopilot
349
00:27:29.559 --> 00:27:33.480
because we think it's really important that
the pilot's not d skill in terms of
350
00:27:33.599 --> 00:27:37.119
landing and airplane. That's the one
skill they need to make sure to never
351
00:27:37.279 --> 00:27:42.440
lose is to safely land an airplane. And I think that that's emblematic of
352
00:27:42.680 --> 00:27:48.559
a discussion and a debate that I
worry is not happening right now. AI
353
00:27:48.640 --> 00:27:51.839
is getting rolled out all over the
place. We are going to see some
354
00:27:51.880 --> 00:27:56.920
significant deskilling occurring. We're going to
see some significant new capabilities arising because of
355
00:27:56.960 --> 00:28:03.079
the AI systems being used widespread ways, and what we need to be having
356
00:28:03.200 --> 00:28:07.240
is an open, honest discussion of
which skills is it okay to lose,
357
00:28:07.680 --> 00:28:11.079
which capabilities do we want to make
sure to gain, and how we struck
358
00:28:11.200 --> 00:28:15.839
that balance appropriately, or are we
just throwing the technology out there seeing what
359
00:28:15.039 --> 00:28:21.960
sticks, chasing the latest bit of
market share in ways that are ultimately going
360
00:28:22.000 --> 00:28:23.720
to come back and we're going to
look up in ten years and say,
361
00:28:25.160 --> 00:28:27.039
oh God, we don't have anybody
who has those skills anymore. And I
362
00:28:27.039 --> 00:28:32.440
think where I feel this most acutely
right now is the rollout of code assistance,
363
00:28:32.640 --> 00:28:36.839
so systems that help people write computer
code. Yeah, I use those.
364
00:28:37.839 --> 00:28:41.480
They're great in some ways, and
they can really accelerate the writing of
365
00:28:41.519 --> 00:28:45.599
code. I've used them myself.
On the other hand, I worry if
366
00:28:45.599 --> 00:28:48.559
we end up in a world where
everybody ds skills in terms of writing code.
367
00:28:49.200 --> 00:28:52.960
I think that's problematic. And then
you get to the point where the
368
00:28:52.000 --> 00:28:57.279
code's writing the code, and then
you've got no transparency and nobody right code
369
00:28:57.359 --> 00:29:02.680
anymore. Exactly all right, David, I want to get into one of
370
00:29:02.680 --> 00:29:07.319
your mini areas of expertise, and
that's time series models, but we'll have
371
00:29:07.400 --> 00:29:11.400
to dig into that next time.
Also coming up, in the next episode
372
00:29:11.519 --> 00:29:15.799
of the Startup Science Podcast, David
Danks shares some valuable advice for founders.
373
00:29:17.119 --> 00:29:19.359
So we really are at a place
right now that if you want to be
374
00:29:19.400 --> 00:29:25.160
able to get ahead, the best
advice I would give an entrepreneur is figure
375
00:29:25.200 --> 00:29:29.000
out if you have access to data
that nobody else has figured out how to
376
00:29:29.119 --> 00:29:33.880
use. Get This Startup Science Podcast
with Gregory Shepherd is brought to you by
377
00:29:33.960 --> 00:29:41.039
startupscience dot Io. Founders can log
on too startupscience dot io and use code
378
00:29:41.279 --> 00:29:47.440
Forbes to get free access. To
connect and find out more about greg Go
379
00:29:47.519 --> 00:29:53.359
to Gregory Shepherd dot com. The
Startup Science Podcast is a production of Forbes Books.
1
00:00:01.280 --> 00:00:07.040
Forbes Books presents the Startup Science Podcast
with Gregory Shepherd, brought to you by
2
00:00:07.080 --> 00:00:13.720
Startupscience dot Io. Greg is an
entrepreneur who has built and sold twelve businesses.
3
00:00:14.160 --> 00:00:18.760
He's a recipient of four Private Equity
awards and is a featured ten X
4
00:00:18.800 --> 00:00:24.559
speaker and Forbes author. Here's Gregory
Shepherd. My guest today is David Danks,
5
00:00:24.719 --> 00:00:29.120
a professor of data science and philosophy
at the University of California, San
6
00:00:29.199 --> 00:00:35.280
Diego. His research explores the intersection
of philosophy, cognitive science, and machine
7
00:00:35.359 --> 00:00:43.560
learning, tackling complex interdisciplinary problems.
He's examined the ethical, psychological, and
8
00:00:43.679 --> 00:00:50.159
policy implications of AI robotics in various
domains like transportation and healthcare. David,
9
00:00:50.159 --> 00:00:53.159
I'm excited to have you on the
podcast, and I'd love it if you
10
00:00:53.159 --> 00:00:57.520
could start by talking a little bit
about all of the incredible work you've done
11
00:00:57.880 --> 00:01:02.719
reaching these different fields if your brilliant
mind. Well, thank you for that
12
00:01:03.119 --> 00:01:08.200
very kind introduction. So I'm here, you see San Diego, and I
13
00:01:08.200 --> 00:01:14.719
do everything from actually build AI systems, primarily with biomedical data, through to
14
00:01:14.840 --> 00:01:19.239
thinking about the ethics and policy of
AI. On top of that, a
15
00:01:19.239 --> 00:01:23.599
lot of issues around philosophy more generally, So, what can we possibly know
16
00:01:23.719 --> 00:01:26.879
about the world, How do we
come to know about the world, what's
17
00:01:26.920 --> 00:01:32.040
the nature of science. I'm one
of these people that I've discovered I don't
18
00:01:32.079 --> 00:01:36.439
actually have any one big problem that
I'm trying to solve. Many academics do.
19
00:01:37.159 --> 00:01:41.519
Instead, I'm just kind of interested
in everything and lucky enough to be
20
00:01:41.640 --> 00:01:45.439
in a position that I get to
pursue all of these different ideas that might
21
00:01:45.480 --> 00:01:49.959
come to my mind. So when
it comes to AI and ethics and poopy,
22
00:01:49.000 --> 00:01:56.319
it seems like philosophy and this new
realm we're entering with AI, there
23
00:01:56.319 --> 00:02:00.519
would be a lot of benefits to
understanding philosophy. Right, how do you
24
00:02:00.599 --> 00:02:04.560
look at the way we're using AI, or what the future holds for AI,
25
00:02:04.640 --> 00:02:07.639
that sort of stuff. Do you
have any perspectives on that completely agree
26
00:02:07.680 --> 00:02:15.439
with you that philosophy is invaluable in
our current tying milieu. I think that's
27
00:02:15.479 --> 00:02:20.759
not always obvious to people, in
part because I think many people misunderstand the
28
00:02:20.840 --> 00:02:27.240
nature of philosophy. So I think
philosophy is sometimes thought of as it's either
29
00:02:27.639 --> 00:02:31.560
the navel gazing that occurs in a
dorm room at midnight or it's the province
30
00:02:31.639 --> 00:02:37.400
of a bunch of dead white guys
from Greece onward. But what philosophy is
31
00:02:37.439 --> 00:02:42.479
really about is it's about asking the
foundational questions. It's about asking those questions
32
00:02:42.960 --> 00:02:46.400
that in most of the parts of
our lives we don't have the time or
33
00:02:46.400 --> 00:02:51.639
even the opportunity in space to stop
and ask them what is the nature of
34
00:02:51.680 --> 00:02:54.879
knowledge? What is the good life? And so when we find ourselves in
35
00:02:54.919 --> 00:02:59.719
a time of significant disruption, I
think it's all the more important that we're
36
00:02:59.719 --> 00:03:04.520
able to go back to those foundations, to have a firm ground at least
37
00:03:05.159 --> 00:03:08.080
about our own values, what we
think is important, well, we're trying
38
00:03:08.120 --> 00:03:14.639
to achieve in our lives, and
philosophy is ultimately about giving people the skills
39
00:03:14.639 --> 00:03:19.280
and experience to be able to ask
those questions about their own foundations. Do
40
00:03:19.319 --> 00:03:23.639
you have or are you still working
on a philosophy of AI? Right?
41
00:03:23.719 --> 00:03:27.360
Like I think about it like this, it's moving so fast. It's been
42
00:03:27.400 --> 00:03:30.759
around for a long time, but
it's just hit this point where it's like
43
00:03:30.919 --> 00:03:35.159
really really moving fast and people are
starting to apply it in all these different
44
00:03:35.199 --> 00:03:38.879
areas. And when you think about
the philosophy of AI, are you thinking
45
00:03:38.960 --> 00:03:44.800
about it in terms of, you
know, it should be used for this
46
00:03:44.919 --> 00:03:49.159
and not used for that, or
we can expect this to happen in the
47
00:03:49.199 --> 00:03:52.199
future, you know, and that
sort of stuff. I'm just coming off
48
00:03:52.199 --> 00:03:57.240
of reading Sapiens, so I'd like
to look to hear about that. So
49
00:03:57.319 --> 00:04:02.240
a lot of my own work in
philosophy and has focused on understanding the ways
50
00:04:02.240 --> 00:04:08.479
in which our values get implemented and
bansed almost kind of expressed through the AI
51
00:04:08.560 --> 00:04:12.360
systems that we build. And I
think a lot of people think of AI
52
00:04:12.400 --> 00:04:15.479
as just code, it's just the
ones and zeros, or it's math.
53
00:04:15.040 --> 00:04:19.560
But I'm much more addrested in the
way that all of the different design choices,
54
00:04:19.680 --> 00:04:24.759
the choices made by the developers and
creators of AI systems, all the
55
00:04:24.759 --> 00:04:30.079
ways that those choices then lead to
the systems being designed for particular purposes,
56
00:04:30.279 --> 00:04:34.240
that they benefit some people and not
others, that they're good in some context
57
00:04:34.319 --> 00:04:40.000
and not others, And so really
trying to connect the values of the people
58
00:04:40.120 --> 00:04:46.399
creating the AI, the people using
the AI, connecting those values with the
59
00:04:46.560 --> 00:04:50.319
values that matter to us in our
lives. So I worry that a lot
60
00:04:50.319 --> 00:04:55.319
of the AI systems that are being
built these days, they are benefiting just
61
00:04:55.360 --> 00:04:59.120
a few people and not others.
So this means that I end up not
62
00:04:59.160 --> 00:05:01.920
thinking all that much of about the
definitional question what counts as AI, what
63
00:05:02.040 --> 00:05:10.120
doesn't count as AI, or thinking
about very abstract questions about what kinds of
64
00:05:10.120 --> 00:05:15.920
AIS might exist in thirty years.
And I'm much more trying to think about
65
00:05:15.360 --> 00:05:19.319
my neighbor who's a retired lawyer and
is trying to figure out should she be
66
00:05:19.399 --> 00:05:24.759
freaked out about AI, or the
person who's worried that AI is going to
67
00:05:24.759 --> 00:05:28.600
take their job. How do we
think about how to create AI that empowers
68
00:05:28.639 --> 00:05:33.959
people rather than just building things that
are seem cool to me but might end
69
00:05:34.040 --> 00:05:39.439
up having a lot of harmful side
effects. Yeah, so let me ask
70
00:05:39.480 --> 00:05:43.959
you about this. I was talking
to a pretty smart fellow in London a
71
00:05:44.040 --> 00:05:46.399
few weeks ago. What he said
is he said, listen, everybody was
72
00:05:46.439 --> 00:05:50.680
afraid of losing their jobs because of
robotics, and they did, but then
73
00:05:50.720 --> 00:05:55.879
they got reemployed. And AI,
on the other hand, is the other
74
00:05:56.000 --> 00:05:58.439
end of the spectrum. Right,
So you have the robotics, which is
75
00:05:58.879 --> 00:06:00.199
back to the blue collar work,
and then you have AI, which is
76
00:06:00.240 --> 00:06:04.240
going to affect the white collar workers, right, like doctors, lawyers.
77
00:06:04.680 --> 00:06:09.240
AI is just going to do a
better job than they could do in diagnosing
78
00:06:09.360 --> 00:06:13.759
or drafting or whatever it is.
When you think about that, do you
79
00:06:14.199 --> 00:06:17.639
feel like there is going to be
a point at which people aren't working as
80
00:06:17.720 --> 00:06:20.759
much? Right? So this guy
said, he said, listen, you
81
00:06:20.800 --> 00:06:23.879
know, it's not like people aren't
going to work. They're just going to
82
00:06:23.959 --> 00:06:28.240
work four hours instead of twelve hours
or whatever to do the same thing and
83
00:06:28.240 --> 00:06:30.360
get paid the same money. Do
you feel like that's a future that in
84
00:06:30.439 --> 00:06:34.519
terms of an optimistic outview, do
do you feel like that's a future that
85
00:06:34.800 --> 00:06:38.319
is possible or what are your thoughts
on that? So? I think that's
86
00:06:38.319 --> 00:06:43.480
a future that is possible, but
it's not going to happen without really concerted
87
00:06:43.519 --> 00:06:47.279
effort. Somewhat Famously, John Maynard
Kanes, the economist, thought that the
88
00:06:47.319 --> 00:06:53.120
steam engine and industrialization was going to
mean that by the early twentieth century,
89
00:06:53.160 --> 00:06:56.839
everybody would be working fifteen to twenty
hours. It'd have lives of leisure.
90
00:06:57.319 --> 00:07:01.160
Obviously that's not how it turned out. I think we're facing a similar sort
91
00:07:01.160 --> 00:07:05.879
of choice here with AI. Do
we take advantage of the massive productivity gains
92
00:07:05.920 --> 00:07:12.839
that we can achieve through AI to
reshape in certain ways quite frankly, our
93
00:07:12.920 --> 00:07:17.959
fundamental society. Do we have things
like universal basic income that enable people to
94
00:07:18.000 --> 00:07:23.319
step away from work and pursue their
passions, which for many people passion is
95
00:07:23.360 --> 00:07:26.680
not the same as work. That's
the part of which I start to be
96
00:07:26.800 --> 00:07:31.079
a little bit less optimistic, because
I think right now it's not clear to
97
00:07:31.120 --> 00:07:35.600
me that there is the political will, the political direction, the political coherence
98
00:07:36.240 --> 00:07:42.759
to be able to get us towards
that better future. I think it's possible,
99
00:07:43.199 --> 00:07:47.199
but unfortunately not probable. In fact, I worry in many ways that
100
00:07:47.240 --> 00:07:51.040
the risks are much more on the
downside than on the upside, in the
101
00:07:51.120 --> 00:07:59.240
sense that right now these past technological
changes, even robotics, but if we
102
00:07:59.360 --> 00:08:03.480
go back earlier to you know,
cars replacing courses and these kinds of things,
103
00:08:03.879 --> 00:08:07.560
as you said, people have lost
jobs, but they've always been able
104
00:08:07.600 --> 00:08:11.639
to get new jobs afterwards. And
so the challenge though is that there's a
105
00:08:11.680 --> 00:08:16.959
period of disruption in the middle,
and historically that's been ten plus years.
106
00:08:18.000 --> 00:08:22.519
It takes a long time for a
new technology to come into widespread use,
107
00:08:22.959 --> 00:08:26.920
in large part because technologies usually are
physical things, so just you have to
108
00:08:26.920 --> 00:08:31.879
build more of them, right,
So it takes time to convert all factories
109
00:08:31.920 --> 00:08:35.080
into being primarily run by robots because
we have to build a lot of robots
110
00:08:35.080 --> 00:08:41.000
to do that. AI costs almost
nothing to duplicate it, And so I
111
00:08:41.039 --> 00:08:46.559
think that the speed of adoption can
be much faster with AI, which also
112
00:08:46.600 --> 00:08:50.480
means the speed of disruption and dislocation
of workers is going to be much faster
113
00:08:50.559 --> 00:08:56.879
with AI. And I don't know
if our society can handle suddenly having twenty
114
00:08:56.919 --> 00:09:01.759
percent unemployment if it's not spread out
over ten years. I worry that there's
115
00:09:01.799 --> 00:09:11.759
actually a much bigger risk of significant
disruption socially rather than the positive upside possibility.
116
00:09:13.120 --> 00:09:16.600
So when you think about this,
so I'm going to sort of move
117
00:09:16.759 --> 00:09:22.200
my way into entrepreneurship and startups and
stuff like that. When you think about
118
00:09:22.200 --> 00:09:24.600
this, you think about industrial transfer. Right now, in the whole startup
119
00:09:24.600 --> 00:09:30.039
ecosystem, industrial transformation is like the
big thing, right everybody thinks AI is
120
00:09:30.039 --> 00:09:33.279
the big thing. The industrial transformation
is one hundred times bigger because it's basically
121
00:09:33.440 --> 00:09:39.000
taking the industrial Revolution and doing it
all over again with a big, giant
122
00:09:39.080 --> 00:09:43.840
upgrade. Would you put AI in
the category of industrial transformation or do you
123
00:09:43.840 --> 00:09:48.200
think it's inside and outside of that. I think that AI is a key
124
00:09:48.279 --> 00:09:54.080
part of the industrial transformations that are
occurring right now. I think it is
125
00:09:54.120 --> 00:09:58.919
also somewhat separate in some ways.
I think what AI does is it's fundamentally
126
00:10:00.080 --> 00:10:03.159
an optimization technique. I mean,
that's really almost all AI comes down to.
127
00:10:03.279 --> 00:10:09.240
We're optimizing a process, a mathematical
function code to do something, whether
128
00:10:09.279 --> 00:10:15.000
it's predict the next word or figure
out who has cancer. So we can
129
00:10:15.039 --> 00:10:20.919
apply those optimization techniques in industrial transformations, in optimizing factory workflows, in optimizing
130
00:10:22.000 --> 00:10:24.039
supply chains. You know, it's
one of the big areas that we're now
131
00:10:24.080 --> 00:10:28.919
seeing AI being used is in optimizing
logistics. It's been around for it for
132
00:10:28.000 --> 00:10:31.519
a while, and it just keeps
becoming more and more important. On the
133
00:10:31.600 --> 00:10:35.679
other hand, one of the things
that I think is important about AI is
134
00:10:35.720 --> 00:10:39.799
that it also can provide new capabilities
and a lot of industrial transformation. At
135
00:10:39.840 --> 00:10:46.039
least in my experiences with it,
is focused on how do we more efficiently
136
00:10:46.200 --> 00:10:50.480
do what we've done before. One
of the key changes with AI is that
137
00:10:50.519 --> 00:10:56.039
we start to expand our horizons of
the possible, at least in the good
138
00:10:56.120 --> 00:11:01.519
cases, so that we start to
really understand that we might not just be
139
00:11:01.600 --> 00:11:05.000
able to replace a human with a
machine, but rather get a human to
140
00:11:05.080 --> 00:11:09.919
do so much more than they could
before by being augmented by the machine.
141
00:11:11.399 --> 00:11:16.440
And so I think that there's opportunities
with AI to do much more than we're
142
00:11:16.480 --> 00:11:20.080
seeing in a lot of the kind
of standard industrial transformation approaches. But it's
143
00:11:20.120 --> 00:11:26.960
also a really important element of successful
transformation. Interesting. I love that.
144
00:11:28.039 --> 00:11:33.320
So it's almost like when computers came
out and we became one hundred thousand times
145
00:11:33.360 --> 00:11:37.159
more effective, and AI is like
another boost, right, Yeah, And
146
00:11:37.440 --> 00:11:41.720
I think, you know, the
introduction of computers, or let's be even
147
00:11:41.799 --> 00:11:46.200
more granular, the widespread use of
spell check. You know, spell check
148
00:11:46.240 --> 00:11:50.080
means that we all lost a little
bit of our spelling ability. I know
149
00:11:50.159 --> 00:11:52.559
I did. On the other hand, it means that we can focus on
150
00:11:52.639 --> 00:11:56.519
what really matters, which is the
expression of our ideas, and we're not
151
00:11:56.639 --> 00:12:01.399
bogged down necessarily by the details of
our we spell something. Word processing is
152
00:12:01.440 --> 00:12:05.519
similar. A lot of people who
are typists lost their jobs because of word
153
00:12:05.519 --> 00:12:09.799
processing programs. On the other hand, it makes it much easier to rapidly
154
00:12:09.840 --> 00:12:16.240
iterate, revise, and what we've
seen is large changes in the number of
155
00:12:16.279 --> 00:12:22.879
people who are able to express their
ideas more clearly, faster and to a
156
00:12:22.919 --> 00:12:26.600
broader audience. So, you know, computers don't just let us do all
157
00:12:26.639 --> 00:12:31.159
the things we did before faster.
They can lead to these real changes,
158
00:12:31.480 --> 00:12:39.559
almost phase changes in what we see
as possible in business contexts. That's really
159
00:12:39.559 --> 00:12:43.840
interesting. So let's move from there. So, if you are an entrepreneur
160
00:12:43.080 --> 00:12:48.480
right now and you're looking at AI
opportunities, the first thing you have to
161
00:12:48.480 --> 00:12:54.200
think about, right, is the
considerations for building trustworthy systems, right that
162
00:12:54.279 --> 00:12:58.200
align with the social values to your
point earlier, So, what are some
163
00:12:58.240 --> 00:13:05.799
of the key siderations an entrepreneur should
have when looking at their responsibility for this
164
00:13:05.919 --> 00:13:09.879
powerful tool, you know, to
prevent it from being weaponized, you know,
165
00:13:09.919 --> 00:13:13.200
as an example. Yeah, So
I think the first thing is,
166
00:13:13.480 --> 00:13:16.840
and I think this is probably good
advice for any entrepreneur, is know your
167
00:13:16.879 --> 00:13:20.440
audience. Know your customers, Know
your clients, and by that I mean
168
00:13:20.559 --> 00:13:24.240
know what matters to them, know
what they're trying to achieve, and know
169
00:13:24.320 --> 00:13:30.320
how what you're building will actually serve
their needs and interests, not what you
170
00:13:30.519 --> 00:13:35.240
think their needs and interests. Are. So there's a long history of people
171
00:13:35.320 --> 00:13:39.120
thinking that they know, oh,
here's what our customers want, when in
172
00:13:39.159 --> 00:13:43.279
fact the customers wanted something very different. Now, of course you can't ask
173
00:13:43.399 --> 00:13:46.240
people necessarily what they want. People
aren't always good at expressing it, as
174
00:13:46.480 --> 00:13:50.679
you know, Steve Jobs famously said, you know, people will know what
175
00:13:50.720 --> 00:13:54.480
they want when we give it to
them. They and I think that that's
176
00:13:54.799 --> 00:13:58.960
right. But that's all the more
reason to engage closely with your customer base,
177
00:14:00.080 --> 00:14:03.759
with your potential clients, so that
they can almost help co design what
178
00:14:03.840 --> 00:14:07.360
it is that you're building. You
can show them, hey, if we
179
00:14:07.399 --> 00:14:09.799
could do this, is that something
you would like? Do you think that
180
00:14:11.120 --> 00:14:13.639
AI could help with that? Because
that's a big problem, right, Like
181
00:14:13.799 --> 00:14:18.000
to your point, you know,
it is almost impossible to get customers to
182
00:14:18.039 --> 00:14:24.799
tell you what they want. It's
you have to decipher and translate what they
183
00:14:24.879 --> 00:14:26.759
tell you to turn it into something
that you could actually do. Do you
184
00:14:26.759 --> 00:14:31.039
think AI could help with that process, like conversationally or something. So,
185
00:14:31.120 --> 00:14:33.960
I think AI can help with that
in a couple of different ways. The
186
00:14:35.000 --> 00:14:41.000
first is that you could use conversational
generative AI systems to basically have free flowing
187
00:14:41.039 --> 00:14:46.360
conversations with your customers and your clients
without it requiring your time and energy.
188
00:14:46.639 --> 00:14:50.720
See I basically get a chatbot to
elicit a lot of these things and to
189
00:14:50.840 --> 00:14:54.879
explore scenarios, so you could design. You can have a chatbot that essentially,
190
00:14:54.960 --> 00:14:58.799
you know, an LLM for example, a large language model that essentially
191
00:14:58.919 --> 00:15:01.360
generates a whole bunch of of scenarios
and says, in this case, what
192
00:15:01.399 --> 00:15:03.440
would you want to do? In
this case, what would you want to
193
00:15:03.480 --> 00:15:07.399
do? The other place that I
think AI can really help is in rapid
194
00:15:07.440 --> 00:15:15.039
prototyping and exploring the possibility space.
So AI systems, especially generative ones,
195
00:15:15.120 --> 00:15:22.759
can very quickly create mockups of systems, can create sort of simulations of new
196
00:15:22.840 --> 00:15:28.279
capabilities, and so you can put
people into a more realistic, a more
197
00:15:28.360 --> 00:15:31.159
naturalistic setting and say, you know, would you use this in your workflows
198
00:15:31.279 --> 00:15:35.480
rather than just asking people, hey, is this something you would like?
199
00:15:35.519 --> 00:15:39.879
Which I think can often suffer when
people have failures of imagination. They might
200
00:15:39.919 --> 00:15:46.120
not realize what kind of a cool
thing they could create. So I wonder
201
00:15:46.200 --> 00:15:52.080
what regulations a founder should be cautious
of and how would they figure that out?
202
00:15:52.200 --> 00:15:56.639
You know, because they're passing regulation
right, So, if you're building
203
00:15:56.759 --> 00:16:00.000
a AI system, is there somewhere
you need to go to make sure you're
204
00:16:00.039 --> 00:16:07.120
not going to get yourself in trouble
for building something that is using potentially the
205
00:16:07.279 --> 00:16:11.440
entire Internet for its knowledge base.
Yeah, that's a great question. It
206
00:16:11.600 --> 00:16:15.759
is one that I found myself spending
far more time than I expected over the
207
00:16:15.840 --> 00:16:19.720
last couple of years thinking about and
talking with folks in places like DC about
208
00:16:19.759 --> 00:16:25.360
what's going on. So the first
thing that is really important for an entrepreneur
209
00:16:25.440 --> 00:16:33.039
to know is that right now we're
seeing basically the development of a patchwork approach
210
00:16:33.159 --> 00:16:38.200
to AI regulation that depends a lot
on where you're located. So the approach
211
00:16:38.240 --> 00:16:42.360
to regulation in the United States is
very different from the approach in Germany.
212
00:16:42.360 --> 00:16:45.919
It's very different from the approach in
China's very different from the approach in India,
213
00:16:47.799 --> 00:16:52.360
just to name four locations. One
of the first things is to recognize
214
00:16:52.600 --> 00:16:57.720
that just because you've checked all the
regulatory boxes to deploy a model in France
215
00:16:59.200 --> 00:17:00.919
does not mean that you going to
be on safe ground when it comes to
216
00:17:02.000 --> 00:17:04.839
the United States. And the reason
I wanted to make sure to call this
217
00:17:04.880 --> 00:17:11.240
out is obviously the idea that different
nations have different regulations, is not anything
218
00:17:11.319 --> 00:17:18.960
new, but AI and AI companies
are not bound by borders in the usual
219
00:17:18.000 --> 00:17:22.799
ways. Yeah, that's a really
good point, right, because it's like
220
00:17:22.920 --> 00:17:27.440
three exactly, and it's very easy
to build out an AI system and suddenly
221
00:17:27.480 --> 00:17:32.839
find people around the world using it, especially if you just throw it up
222
00:17:32.839 --> 00:17:36.680
on one of these standard cloud services, it can be accessed anywhere in the
223
00:17:36.720 --> 00:17:41.000
world. And so I think that
there is a much more complex regulatory environment
224
00:17:41.039 --> 00:17:45.319
that entrepreneurs are entering when it comes
to AI than many of them realize.
225
00:17:45.920 --> 00:17:51.480
And so I think that there's a
large amount of thought that entrepreneurs need to
226
00:17:51.480 --> 00:17:55.200
put into how are they going to
make sure that they don't step on any
227
00:17:55.240 --> 00:17:59.440
regulatory land lines. Now I mentioned
the different approaches, let me just briefly
228
00:17:59.519 --> 00:18:03.359
say, you know, for example, in the United States, the regulation
229
00:18:03.400 --> 00:18:07.440
of AI is happening in a very
sector specific way. So if I'm building
230
00:18:07.440 --> 00:18:10.920
an AI system in healthcare, I'm
going to have to know about HIPPA,
231
00:18:10.920 --> 00:18:15.440
the health privacy law that we have
in the United States. In contrast,
232
00:18:15.680 --> 00:18:18.200
if I'm building AI that's going to
operate on social media, well, in
233
00:18:18.240 --> 00:18:22.240
the United States, social media is
largely unregulated and so there isn't going to
234
00:18:22.240 --> 00:18:26.920
be a particular regulation on my AI
system. In contrasting the EU, there's
235
00:18:27.039 --> 00:18:30.599
what's called the AI Act, which
was just recently finalized. It's going to
236
00:18:30.640 --> 00:18:34.400
be going into effect in the next
couple of years, and the EUAI Act
237
00:18:34.440 --> 00:18:38.680
is much more focused on AI writ
Larch. It's a sector general kind of
238
00:18:38.720 --> 00:18:44.759
regulation that essentially says, if you're
going to be using AI systems in high
239
00:18:44.839 --> 00:18:48.839
risk ways or in ways that creates
systemic risks. So if you're building a
240
00:18:48.880 --> 00:18:52.599
generative AI system that's going to be
at the foundation of many different models,
241
00:18:52.559 --> 00:18:56.759
then there's a whole set of auditing
and testing that you have to do before
242
00:18:56.839 --> 00:19:02.519
you ever deployed in the first place. So if I'm an AI developer in
243
00:19:02.559 --> 00:19:06.400
Germany, my brain is thinking about, all right, what are the high
244
00:19:06.480 --> 00:19:10.440
risk use cases that this might be
put to, even if I don't intend
245
00:19:10.519 --> 00:19:12.319
it to be used in that way, Whereas in the US, what I'm
246
00:19:12.319 --> 00:19:15.839
thinking is, Okay, where are
the sectors I want to deploy and what
247
00:19:15.960 --> 00:19:21.799
are the regulations that are specific to
that sector. Now you also ask,
248
00:19:21.880 --> 00:19:25.160
I think a question which is very
near and dear to my heart because it's
249
00:19:25.200 --> 00:19:29.279
part of how I got into thinking
about issues around ethics and technology and ethics
250
00:19:29.279 --> 00:19:33.599
and AI in the first place,
which is how should we think about the
251
00:19:33.640 --> 00:19:38.119
ways that our algorithms and models might
end up being used for purposes that we
252
00:19:38.200 --> 00:19:41.920
never intended. Yeah, that's what
I'm worried about, and you're right to
253
00:19:41.960 --> 00:19:47.400
worry about it. That's part of
what drew me into ethics of AI is
254
00:19:47.440 --> 00:19:52.119
in part learning that an algorithm I
had helped to develop probably was used for
255
00:19:52.200 --> 00:19:56.279
purposes that I didn't necessarily agree with. So the first thing to say is
256
00:19:56.720 --> 00:20:00.680
that's less of a risk I think
actually for many entrepreneurs than it is for
257
00:20:00.799 --> 00:20:06.640
folks like me and academia, because
entrepreneurs are typically building products, which means
258
00:20:06.640 --> 00:20:10.960
that there's going to be a lot
of work that goes into translating the algorithm
259
00:20:10.960 --> 00:20:14.960
and model into the particular use cases
of your customers, your clients, whether
260
00:20:15.000 --> 00:20:19.079
those are the businesses or consumers.
In contrast, academics, we're much more
261
00:20:19.079 --> 00:20:23.400
interested in what is the general algorithm
that could be used in many different contexts.
262
00:20:23.640 --> 00:20:30.680
Oh, I see broad exactly.
AI is great because it's kind of
263
00:20:30.799 --> 00:20:36.079
domain agnostic. Right. In general
data science machine learning, we kind of
264
00:20:36.160 --> 00:20:40.119
don't care where the data come from. We're operating on the patterns that we
265
00:20:40.200 --> 00:20:45.839
find in patterns are independent of what
their patterns over. At the same time,
266
00:20:45.680 --> 00:20:51.680
obviously you're going to tune your algorithm
to the particular assumptions you can make
267
00:20:51.720 --> 00:20:56.880
the particular features of the problem,
and so as you get closer and closer
268
00:20:56.880 --> 00:21:00.920
to an actual product. I think
the domain generality, it's less and less
269
00:21:00.440 --> 00:21:04.359
even with the things that we think
of now as the most general purpose AI
270
00:21:04.480 --> 00:21:11.119
systems called foundation models or generative AI, the large language models, even those
271
00:21:11.200 --> 00:21:19.359
systems, we are increasingly seen that
very few actual applications in business or in
272
00:21:19.920 --> 00:21:26.119
society generally, very few of the
applications are based on the untuned foundation models.
273
00:21:26.680 --> 00:21:32.640
Most of what's happening even with something
like GPT four is that companies are
274
00:21:32.759 --> 00:21:37.160
fine tuning it in ways that make
it much more focused, specific and accurate
275
00:21:37.799 --> 00:21:42.559
for their particular needs. So I
think we're even though we talk a lot
276
00:21:42.599 --> 00:21:48.000
about general purpose AI, I don't
think that it happens as much in practice.
277
00:21:48.440 --> 00:21:52.480
And so as long as the entrepreneurs
retaining some measure of control over their
278
00:21:52.480 --> 00:21:59.480
intellectual property, they can help to
shape where the technology goes basically like commercial
279
00:21:59.519 --> 00:22:06.079
AI is relatively low risk, but
then broad AI is more high risk because
280
00:22:06.240 --> 00:22:11.039
one of the things that I notice
is that if I aim the AI at
281
00:22:11.160 --> 00:22:15.160
data, it uses that data to
return my answers. So if you aimed
282
00:22:15.279 --> 00:22:19.160
at the entire internet, we all
know there's a lot of trash on the
283
00:22:19.160 --> 00:22:23.119
Internet, right, Am I getting
the truth? Or am I getting somebody
284
00:22:23.160 --> 00:22:26.720
who has? You know, carpet
bombed the Internet with the way that they
285
00:22:26.759 --> 00:22:30.440
think, right, So we saw
in a couple of election cycles we saw
286
00:22:30.480 --> 00:22:37.039
this massive insurgence of flooding the internet, you know, carpet bombing the Internet
287
00:22:37.079 --> 00:22:41.079
with certain belief patterns that it seems
like if AOI was in place and somebody
288
00:22:41.119 --> 00:22:45.759
asked a question, that would be
the answer that they would get because there's
289
00:22:45.960 --> 00:22:49.680
so much data you know, reflecting
that answer, Like how do you think
290
00:22:49.720 --> 00:22:55.920
about that? I mean, that
seems like a pretty scary scenario and dividing
291
00:22:56.160 --> 00:23:00.720
you know, subs even more so, Yeah, it's a massive problem.
292
00:23:02.599 --> 00:23:07.559
Just recently, Google came out with
their AI summaries that's supposed to help with
293
00:23:07.599 --> 00:23:14.960
the search results, So it summarizes
the leading responses that were found through search
294
00:23:15.400 --> 00:23:18.039
so that you might not even have
to click through, and it's done.
295
00:23:18.079 --> 00:23:22.519
Things that have suggested, for example, that you should perhaps add glue to
296
00:23:22.599 --> 00:23:29.759
your cheese on your pizza so that
it is gooier and flows better, which,
297
00:23:30.200 --> 00:23:33.839
as I understand it, actually is
true of advertising. Advertisers often do
298
00:23:33.960 --> 00:23:37.599
things to the food in order to
make it look better. You don't want
299
00:23:37.599 --> 00:23:41.480
to eat it. Though. That's
a case where there's information on the Internet
300
00:23:41.599 --> 00:23:48.319
that is being misused or misapplied,
shall we say, by the search summarization
301
00:23:48.599 --> 00:23:52.640
system. Of course, it goes
beyond just the sort of garbage that's out
302
00:23:52.640 --> 00:23:57.240
there on the Internet in various places, to these systems actually completely. They're
303
00:23:57.279 --> 00:24:02.640
just trying to predict what the most
next few words are supposed to be,
304
00:24:02.759 --> 00:24:06.839
so they will make up things that
sound plausible. We call this hallucination.
305
00:24:07.000 --> 00:24:10.640
It's not the best term, but
that's okay. They'll essentially make things up.
306
00:24:10.920 --> 00:24:15.440
Rather infamously, many of us in
academia, when things like CHATCHBT first
307
00:24:15.519 --> 00:24:18.039
came out, we would say,
oh, what are the papers in my
308
00:24:18.200 --> 00:24:22.200
area that I should know about?
It would generate a lot of very plausible
309
00:24:22.200 --> 00:24:26.359
sounding paper names, none of which
existed, and in fact, recently one
310
00:24:26.359 --> 00:24:30.880
of the CEOs of the big tech
companies came out and just said, we're
311
00:24:30.960 --> 00:24:33.640
never going to solve these problems.
These systems, these generative AI systems,
312
00:24:33.680 --> 00:24:40.319
are always going to hallucinate. They're
always going to return problematic output in at
313
00:24:40.400 --> 00:24:45.480
least some cases. So it's a
huge problem and it's one that we're never
314
00:24:45.519 --> 00:24:49.759
going to get rid of. Having
said all of that, we can do
315
00:24:49.839 --> 00:24:53.640
so much better than we are right
now. There are a lot of techniques
316
00:24:53.720 --> 00:25:00.359
and ways to help people know for
these kinds of systems what we call provenence
317
00:25:00.559 --> 00:25:04.319
sort of the history of where the
information came from. And that's one of
318
00:25:04.319 --> 00:25:10.160
the big advances we've seen in the
last six to nine months is improved systems
319
00:25:10.200 --> 00:25:15.680
that can reveal the provenance of information
to people that are using these AI systems.
320
00:25:15.799 --> 00:25:19.200
Can these systems can trace back the
history of well, the reason I've
321
00:25:19.240 --> 00:25:22.480
given you this information is that it
came from this site. This site is
322
00:25:22.519 --> 00:25:27.920
reputable for these reasons. And we're
starting to see the emergence of technologies that
323
00:25:29.000 --> 00:25:33.599
can help people more critically evaluate the
information sources that are being put in front
324
00:25:33.640 --> 00:25:37.720
of them by transparency. Just given
this, because that's one of the things
325
00:25:37.759 --> 00:25:42.240
I worry about right if AI is
doing the thinking and we're not. Are
326
00:25:42.240 --> 00:25:45.039
we just going to get dumber,
just like we don't spell as well,
327
00:25:45.359 --> 00:25:49.720
so our cognition is going to change. I quite deliberately didn't use the term
328
00:25:49.799 --> 00:25:56.440
dumber a fair enough. There's there's
a phenomenon that's referred to as de skilly.
329
00:25:56.559 --> 00:26:00.519
It's the idea that if you have
a skill and then a machine or
330
00:26:00.519 --> 00:26:03.559
some automation starts to do that skill
for you, or even another human starts
331
00:26:03.599 --> 00:26:07.160
doing that skill for you, that
over time you'll lose that skill. You
332
00:26:07.200 --> 00:26:11.960
will d skill. That happens all
the time. When I was younger,
333
00:26:12.359 --> 00:26:18.680
much younger, I could drive a
manual transmission car pretty well. I still
334
00:26:18.759 --> 00:26:21.559
can kind of drive one, but
you wouldn't want to put me behind the
335
00:26:21.559 --> 00:26:25.960
wheel on a regular basis at the
moment because I've deskilled. I've lost that
336
00:26:26.000 --> 00:26:29.920
particular ability that one. Though.
Honestly, I'm okay with like. It
337
00:26:30.160 --> 00:26:33.920
doesn't bother me. I very rarely
need to drive a manual transmission car,
338
00:26:33.039 --> 00:26:37.960
so I'm okay with de skilling in
that way. Selective deskilling selective discilling exactly.
339
00:26:40.000 --> 00:26:44.119
So what that points to, though, is the importance that as we're
340
00:26:44.119 --> 00:26:48.279
bringing AI systems into our workflows,
into our companies, into our daily lives.
341
00:26:48.000 --> 00:26:55.200
We actually need to be intentional and
thoughtful about what skills are we likely
342
00:26:55.279 --> 00:27:00.720
to lose as a result, and
are we okay with that? So the
343
00:27:00.799 --> 00:27:04.279
three most powerful words when we're thinking
about technology adoption, in my mind,
344
00:27:04.359 --> 00:27:08.880
are are we sure? Just stopping
and asking are we sure that this is
345
00:27:08.880 --> 00:27:11.759
a good idea? Are we sure
it's okay that we lose that skill?
346
00:27:12.640 --> 00:27:17.960
I'm really glad that people thought about
that question when it came to the rollout
347
00:27:18.000 --> 00:27:22.839
of autopilots and airplanes. Autopilots can
actually land airplanes better than most commercial pilots
348
00:27:22.880 --> 00:27:29.359
would be able to, but we
don't allow commercial airlines to land on autopilot
349
00:27:29.559 --> 00:27:33.480
because we think it's really important that
the pilot's not d skill in terms of
350
00:27:33.599 --> 00:27:37.119
landing and airplane. That's the one
skill they need to make sure to never
351
00:27:37.279 --> 00:27:42.440
lose is to safely land an airplane. And I think that that's emblematic of
352
00:27:42.680 --> 00:27:48.559
a discussion and a debate that I
worry is not happening right now. AI
353
00:27:48.640 --> 00:27:51.839
is getting rolled out all over the
place. We are going to see some
354
00:27:51.880 --> 00:27:56.920
significant deskilling occurring. We're going to
see some significant new capabilities arising because of
355
00:27:56.960 --> 00:28:03.079
the AI systems being used widespread ways, and what we need to be having
356
00:28:03.200 --> 00:28:07.240
is an open, honest discussion of
which skills is it okay to lose,
357
00:28:07.680 --> 00:28:11.079
which capabilities do we want to make
sure to gain, and how we struck
358
00:28:11.200 --> 00:28:15.839
that balance appropriately, or are we
just throwing the technology out there seeing what
359
00:28:15.039 --> 00:28:21.960
sticks, chasing the latest bit of
market share in ways that are ultimately going
360
00:28:22.000 --> 00:28:23.720
to come back and we're going to
look up in ten years and say,
361
00:28:25.160 --> 00:28:27.039
oh God, we don't have anybody
who has those skills anymore. And I
362
00:28:27.039 --> 00:28:32.440
think where I feel this most acutely
right now is the rollout of code assistance,
363
00:28:32.640 --> 00:28:36.839
so systems that help people write computer
code. Yeah, I use those.
364
00:28:37.839 --> 00:28:41.480
They're great in some ways, and
they can really accelerate the writing of
365
00:28:41.519 --> 00:28:45.599
code. I've used them myself.
On the other hand, I worry if
366
00:28:45.599 --> 00:28:48.559
we end up in a world where
everybody ds skills in terms of writing code.
367
00:28:49.200 --> 00:28:52.960
I think that's problematic. And then
you get to the point where the
368
00:28:52.000 --> 00:28:57.279
code's writing the code, and then
you've got no transparency and nobody right code
369
00:28:57.359 --> 00:29:02.680
anymore. Exactly all right, David, I want to get into one of
370
00:29:02.680 --> 00:29:07.319
your mini areas of expertise, and
that's time series models, but we'll have
371
00:29:07.400 --> 00:29:11.400
to dig into that next time.
Also coming up, in the next episode
372
00:29:11.519 --> 00:29:15.799
of the Startup Science Podcast, David
Danks shares some valuable advice for founders.
373
00:29:17.119 --> 00:29:19.359
So we really are at a place
right now that if you want to be
374
00:29:19.400 --> 00:29:25.160
able to get ahead, the best
advice I would give an entrepreneur is figure
375
00:29:25.200 --> 00:29:29.000
out if you have access to data
that nobody else has figured out how to
376
00:29:29.119 --> 00:29:33.880
use. Get This Startup Science Podcast
with Gregory Shepherd is brought to you by
377
00:29:33.960 --> 00:29:41.039
startupscience dot Io. Founders can log
on too startupscience dot io and use code
378
00:29:41.279 --> 00:29:47.440
Forbes to get free access. To
connect and find out more about greg Go
379
00:29:47.519 --> 00:29:53.359
to Gregory Shepherd dot com. The
Startup Science Podcast is a production of Forbes Books.
