1
00:00:00,252 --> 00:00:04,722
From security gaps to evolving
regulations and tough ethical questions,

2
00:00:04,902 --> 00:00:09,102
artificial intelligence introduces
as many risks as opportunities.

3
00:00:09,762 --> 00:00:14,442
Today's guest takes us deep into the
strategies enterprises need to mitigate

4
00:00:14,442 --> 00:00:18,372
them, ensure compliance, and build
responsible governance that scales.

5
00:00:19,092 --> 00:00:22,092
My name is Alexandre Nevski,
and this is Innovation Tales.

6
00:00:23,377 --> 00:00:25,957
Navigating change, one story at a time.

7
00:00:25,987 --> 00:00:30,187
We share insights from leaders tackling
the challenges of today's digital world.

8
00:00:30,457 --> 00:00:33,967
Welcome to Innovation Tales,
the podcast exploring the human

9
00:00:33,967 --> 00:00:35,827
side of digital transformation.

10
00:00:36,506 --> 00:00:40,286
In our last episode, we explored why
not every Gen AI use case is worth

11
00:00:40,286 --> 00:00:44,366
pursuing, how to build AI governance
that works, and what business leaders

12
00:00:44,426 --> 00:00:46,886
must do to stay ahead of emerging risks.

13
00:00:47,276 --> 00:00:48,656
But now we go deeper.

14
00:00:49,286 --> 00:00:53,336
Alec Crawford, founder of Artificial
Intelligence Risk, Inc. has spent

15
00:00:53,336 --> 00:00:57,596
decades at the intersection of
AI, finance and risk management.

16
00:00:58,241 --> 00:01:02,891
His company helps enterprises adopt
AI security through strong governance,

17
00:01:03,011 --> 00:01:05,051
compliance, and cybersecurity.

18
00:01:06,131 --> 00:01:10,991
In this episode, we dive into the features
and deployment of his platform, along with

19
00:01:10,991 --> 00:01:14,021
broader regulatory and ethical challenges.

20
00:01:14,861 --> 00:01:19,241
From securing data to ensuring
compliance, Alex shares practical insights

21
00:01:19,391 --> 00:01:20,711
every business leader should hear.

22
00:01:21,371 --> 00:01:24,461
Without further ado, he's my
conversation with Alec Crawford.

23
00:01:24,764 --> 00:01:29,924
I wanna circle back to what the
platform that you deploy is.

24
00:01:30,144 --> 00:01:35,714
I want to ensure that for our business
minded audience, it's clear, what's the

25
00:01:35,714 --> 00:01:39,694
difference between, other, alternatives
that they might have on the market.

26
00:01:39,694 --> 00:01:44,378
So what I understood is that the first
and foremost thing you, argue is that

27
00:01:44,768 --> 00:01:51,728
you can only really be certain of the
security and compliance and governance

28
00:01:51,758 --> 00:01:57,038
of technology that you're running on
premise, or that at least it's, if

29
00:01:57,038 --> 00:02:00,478
it's in a cloud that's a private cloud
and you know exactly what's running.

30
00:02:00,478 --> 00:02:04,760
And then I heard initially that  there
was a number of capabilities that

31
00:02:04,760 --> 00:02:06,260
you had built into the platform.

32
00:02:06,660 --> 00:02:10,830
I think at some point you were mentioning
logging every prompt and every model

33
00:02:10,830 --> 00:02:15,361
change, because essentially with
these additional capabilities, you

34
00:02:15,361 --> 00:02:17,641
can facilitate compliance, et cetera.

35
00:02:17,941 --> 00:02:19,156
Did I get this more or less right.

36
00:02:19,466 --> 00:02:20,756
I think you nailed it, Alex.

37
00:02:20,786 --> 00:02:26,486
So think about it as "single pane
of glass" access to AI across all

38
00:02:26,486 --> 00:02:28,676
your different AI tools, right?

39
00:02:29,336 --> 00:02:35,526
And because of that, that can run
through the compliance system, right?

40
00:02:35,526 --> 00:02:38,586
And then you've got the
agent building capabilities.

41
00:02:38,586 --> 00:02:40,116
We've got three levels.

42
00:02:40,836 --> 00:02:43,656
You've got the admins that can
set security settings and build

43
00:02:43,679 --> 00:02:45,576
agents and assign agents to people.

44
00:02:45,966 --> 00:02:50,206
Have super users that are allowed
to build agents for themselves, but

45
00:02:50,206 --> 00:02:51,976
not change the security settings.

46
00:02:51,976 --> 00:02:55,646
And then you've got users who, hey,
here's your 10 agents, go wild.

47
00:02:56,166 --> 00:03:02,596
So that on an enterprise platform that
does the governance and risk management

48
00:03:02,596 --> 00:03:08,836
and compliance and cybersecurity is
what allows, companies with high risk AI

49
00:03:08,836 --> 00:03:14,371
applications that include consumer data
or healthcare data to do that safely.

50
00:03:14,851 --> 00:03:19,891
And I agree, doing it inside your private
cloud where you own all the data and

51
00:03:19,891 --> 00:03:24,681
you know it's not leaking out there is
a number one goal and something that

52
00:03:24,681 --> 00:03:27,881
we really encourage our clients to do.

53
00:03:28,371 --> 00:03:31,431
There are certain models out there
that are, that you can't do that with.

54
00:03:31,431 --> 00:03:35,751
So right now, for example, Claude
is only accessible as an API.

55
00:03:36,051 --> 00:03:39,501
So you know, there you may have an
agreement with Claude saying, Hey,

56
00:03:39,501 --> 00:03:43,451
they're not gonna tell anybody, what
you're doing with AI, but we have

57
00:03:43,451 --> 00:03:47,561
no idea how long they're storing
our data for or who's looking at it.

58
00:03:47,561 --> 00:03:53,141
And if you have confidentiality agreements
with your clients, let's say like you may

59
00:03:53,141 --> 00:03:55,901
not legally be allowed to do that, right?

60
00:03:56,041 --> 00:03:57,091
To push it out to your cloud.

61
00:03:57,091 --> 00:04:01,831
If you're running something
on-prem, you know it's inside

62
00:04:01,831 --> 00:04:02,821
your firewall, you're good.

63
00:04:03,031 --> 00:04:04,981
You can do whatever you want
from a confidential standpoint.

64
00:04:05,386 --> 00:04:09,136
Yeah, well it, it makes sense that
just like platforms in the past, I

65
00:04:09,136 --> 00:04:13,596
dunno if we go way back, I dunno, for
document management or, any number of,

66
00:04:13,676 --> 00:04:19,306
enterprise grade software, that gets
deployed as a platform, it's normal

67
00:04:19,306 --> 00:04:25,736
that an enterprise's requirements
are more specific than, a consumer.

68
00:04:25,806 --> 00:04:30,406
And therefore having a multi-tiered
authorization model that you've

69
00:04:30,406 --> 00:04:34,766
just mentioned is just a, it's not
a negotiable, it's a must have.

70
00:04:34,766 --> 00:04:38,966
You must be able to design the roles
that make sense for your organization

71
00:04:39,006 --> 00:04:45,728
But it's seems like it's quite an onerous
of effort in order to be fully in control.

72
00:04:45,998 --> 00:04:49,808
Do you think that most businesses
will be able to get there at some

73
00:04:49,808 --> 00:04:54,188
point, or do you see for, I don't
know, foreseeable future, that this

74
00:04:54,188 --> 00:04:56,738
is only the largest organizations
we can afford something like that?

75
00:04:57,186 --> 00:05:00,386
No, our smallest clients are
50-100 people and they're

76
00:05:00,386 --> 00:05:02,826
perfectly capable of managing this.

77
00:05:02,876 --> 00:05:08,126
I do think that there aren't a lot of
other solutions like ours right now.

78
00:05:08,266 --> 00:05:10,336
And really that's why we're
focused on high risk AI.

79
00:05:10,366 --> 00:05:13,606
Look, if you have nothing confidential
at your company and you wanna

80
00:05:13,606 --> 00:05:16,246
write a marketing piece and you
don't care that the world can see

81
00:05:16,246 --> 00:05:18,196
it, just go to ChatGPT, right?

82
00:05:18,196 --> 00:05:19,606
Pay $20, right?

83
00:05:20,376 --> 00:05:25,476
The reason we're focused on high risk
is you can't do that with consumer data.

84
00:05:25,476 --> 00:05:26,586
You can't do that with healthcare data.

85
00:05:26,586 --> 00:05:28,056
It's illegal, right?

86
00:05:28,056 --> 00:05:32,256
So that's why we're focused on that
market, but also really focused

87
00:05:32,256 --> 00:05:33,516
on the security side, right?

88
00:05:33,516 --> 00:05:38,046
I think that, there's a lot of software
out there that's SaaS, that just has

89
00:05:38,046 --> 00:05:39,666
a lot of security holes in it, right?

90
00:05:39,666 --> 00:05:44,496
Even Grammarly, for example, had a huge
security hole that was revealed last year.

91
00:05:44,496 --> 00:05:47,826
And think of all the people using
Grammarly who all of a sudden, whoops.

92
00:05:47,826 --> 00:05:49,836
Like your data could be
out there somewhere, right?

93
00:05:50,726 --> 00:05:52,496
That's primarily because it's SaaS.

94
00:05:52,501 --> 00:05:59,502
In ours, your firewall has to be breached
before anyone has access to your AI.

95
00:05:59,982 --> 00:06:04,962
Beyond that, everything's encrypted
in motion, at rest, and at rest.

96
00:06:04,962 --> 00:06:11,052
Beyond that, each individual only has
access to certain agents and certain data.

97
00:06:11,082 --> 00:06:16,122
Beyond that, we're looking for
unusual activity and flagging it

98
00:06:16,122 --> 00:06:17,982
back to the cybersecurity team.

99
00:06:18,032 --> 00:06:20,312
Is there, is someone
trying to hack the AI?

100
00:06:20,312 --> 00:06:23,518
Is there data exfiltration in place?

101
00:06:23,699 --> 00:06:29,249
' cause one of the tenets of cybersecurity
today is, it's not a matter of if

102
00:06:29,249 --> 00:06:32,609
you get hacked, but when and when
you get hacked, it's a matter of

103
00:06:32,609 --> 00:06:35,789
identifying that on day 0, not day 17.

104
00:06:36,539 --> 00:06:40,189
And if you're just using, ChatGPT online
and letting people do whatever they want,

105
00:06:40,669 --> 00:06:42,169
you'll never know that you've been hacked.

106
00:06:42,469 --> 00:06:45,829
What ChatGPT is gonna call up your
cybersecurity team and say, Hey, we

107
00:06:45,829 --> 00:06:46,909
think you might've gotten hacked.

108
00:06:46,909 --> 00:06:47,719
That ain't happening.

109
00:06:47,789 --> 00:06:49,049
You gotta figure it out yourself.

110
00:06:49,049 --> 00:06:51,810
And that's one of the things that,
one of the things that we do.

111
00:06:52,606 --> 00:06:56,986
In my experience with platforms like
this, if you are deploying them on

112
00:06:56,986 --> 00:07:00,046
premise, or at least some of the
components are deployed on premise,

113
00:07:00,046 --> 00:07:04,546
you start to have another problem,
then you have to have enough internal

114
00:07:04,546 --> 00:07:07,306
skills to maintain those components.

115
00:07:07,786 --> 00:07:11,506
And when I say maintain, I also mean
apply security patches, et cetera.

116
00:07:11,506 --> 00:07:17,746
So how do you do that at scale
so that it's accessible to

117
00:07:17,746 --> 00:07:21,416
organizations who don't have a,
full-time staff to take care of it?

118
00:07:22,451 --> 00:07:26,746
Yeah, so at this point, most of
our smaller clients are on Azure.

119
00:07:26,936 --> 00:07:33,171
and we have a process where we can deploy
our software through an ARM template.

120
00:07:33,201 --> 00:07:34,071
It's super easy.

121
00:07:34,121 --> 00:07:34,781
Click here.

122
00:07:35,391 --> 00:07:39,436
Most of those clients will also, if they
don't have a, a specific technology person

123
00:07:40,066 --> 00:07:41,566
will have a managed service provider.

124
00:07:41,566 --> 00:07:44,326
And this is, we make it super, super easy.

125
00:07:44,896 --> 00:07:48,466
The reason we even do that, obviously
we could, we could push out, these

126
00:07:48,466 --> 00:07:53,256
updates, but obviously we want our
clients to update when they feel

127
00:07:53,256 --> 00:07:57,186
like, not necessarily when we feel
like, So maybe they wanna update on

128
00:07:57,186 --> 00:08:00,846
a Saturday 'cause they, no, it won't
disrupt anybody, for example, right?

129
00:08:01,386 --> 00:08:03,996
Maybe they like doing
Wednesday nights, I'm not sure.

130
00:08:04,046 --> 00:08:06,236
we make it available at the
end of the month and then they

131
00:08:06,236 --> 00:08:08,096
can update when they want to.

132
00:08:08,486 --> 00:08:14,006
We've had zero issues with that so far
in terms of even at our smaller clients.

133
00:08:14,370 --> 00:08:15,540
That's, uh, great to hear.

134
00:08:15,930 --> 00:08:21,450
And I think it's, now that we've gone
into quite a lot of details about the

135
00:08:21,450 --> 00:08:26,450
product, I think we can go back to,
a wider, discussion of generative AI.

136
00:08:26,450 --> 00:08:30,886
And I wanted to pick your brain,
uh, because you've already mentioned

137
00:08:30,886 --> 00:08:33,796
about compliance, about various
regulations that are in place.

138
00:08:33,796 --> 00:08:37,726
Do you feel like there is enough
regulation in place already?

139
00:08:38,326 --> 00:08:40,786
Do you feel like there's much more coming?

140
00:08:40,876 --> 00:08:43,746
Where do you think this is
going, for, companies to

141
00:08:43,746 --> 00:08:45,336
assess what's going to happen?

142
00:08:46,071 --> 00:08:49,981
Yeah, so I think we're pretty well
set up in Europe at this point.

143
00:08:50,041 --> 00:08:54,751
Especially for high risk AI because
what's happening there is companies

144
00:08:54,751 --> 00:08:58,921
need to go to their specific
regulator and really get approval

145
00:08:58,921 --> 00:09:00,586
to use AI for something high risk.

146
00:09:00,811 --> 00:09:06,841
So that will involve lots of
detailed work and end up with,

147
00:09:07,301 --> 00:09:10,631
company policies and procedures that
they'll need to build into their AI.

148
00:09:11,081 --> 00:09:14,971
I, I think there will be more stuff,
passed there in the EU that, because

149
00:09:15,001 --> 00:09:19,951
so far the rules are mainly focused on
consumer protection and there may be

150
00:09:19,951 --> 00:09:23,491
more rules around what you're allowed
to do and not allowed to do with AI.

151
00:09:23,491 --> 00:09:26,671
And things along those lines,
we will see, we'll see there.

152
00:09:27,101 --> 00:09:31,201
The US as usual is behind
Europe in terms of regulation.

153
00:09:31,213 --> 00:09:36,343
The one kind of big thing that was out
there was the Biden executive order

154
00:09:36,343 --> 00:09:38,753
on AI, which Trump has struck down.

155
00:09:39,323 --> 00:09:41,183
So there's just basically aren't
a whole lot of rules right now.

156
00:09:42,616 --> 00:09:47,086
The rules that are out there are
for kind of high risk AI, right?

157
00:09:47,086 --> 00:09:50,716
So if you are a healthcare company,
you have to abide by HIPAA.

158
00:09:50,716 --> 00:09:51,466
That already exists.

159
00:09:51,466 --> 00:09:56,356
If you're a bank, a bunch of rules
around protecting client data,

160
00:09:56,356 --> 00:09:57,946
anti-bias and things like that.

161
00:09:57,946 --> 00:10:01,576
And the really important thing
I've heard regulators say is

162
00:10:02,236 --> 00:10:06,316
just because it's AI doesn't mean
you can break any rules, right?

163
00:10:06,316 --> 00:10:07,666
And that's really important.

164
00:10:07,666 --> 00:10:12,346
In other words, if you're not allowed
to have bias in lending and your AI is

165
00:10:12,346 --> 00:10:15,106
biased, it's breaking the law, right?

166
00:10:15,156 --> 00:10:18,276
Even if that's an AI, making a loan
instead of a person making a loan.

167
00:10:18,366 --> 00:10:19,806
They don't care, right?

168
00:10:19,806 --> 00:10:25,021
So that's super, super important
and very important to follow.

169
00:10:25,571 --> 00:10:31,681
I think over time, what we're gonna see,
or we're already seeing is the states are

170
00:10:31,681 --> 00:10:34,621
making laws about AI and consumer privacy.

171
00:10:34,621 --> 00:10:42,201
So for example, Colorado early in
February their AI act kicked in, which

172
00:10:42,201 --> 00:10:47,811
said, any use of high risk AI, and they
have 29 pages of rules, literally 29

173
00:10:47,811 --> 00:10:51,081
pages of roles about what you have to
do and can't do and things like that.

174
00:10:51,081 --> 00:10:55,491
And it's not for companies headquartered
there, it's for companies with even one

175
00:10:55,491 --> 00:10:59,361
client there and the penalties are severe.

176
00:10:59,361 --> 00:11:01,821
We're talking potentially
millions of dollars, right?

177
00:11:01,821 --> 00:11:04,371
So better be following the rules.

178
00:11:04,471 --> 00:11:08,011
But what's interesting there also is they
write 29 pages of rules and then they're

179
00:11:08,011 --> 00:11:11,591
like, oh, and by the way, if you're
doing the national version, the National

180
00:11:11,591 --> 00:11:16,211
Institute of Science and Technology AI
Risk Management Framework, otherwise

181
00:11:16,211 --> 00:11:19,541
known as the NIST AI, RFM, you're good.

182
00:11:19,541 --> 00:11:20,711
You don't have to do
any of this other stuff.

183
00:11:20,751 --> 00:11:21,970
we're gonna give you a pass.

184
00:11:22,431 --> 00:11:25,761
And that's really what we focus on is
the NIST AI risk management framework

185
00:11:25,761 --> 00:11:28,101
and facilitate compliance with that.

186
00:11:28,101 --> 00:11:31,511
'cause look, in the end, companies
are not gonna be able to comply with

187
00:11:31,511 --> 00:11:33,881
50 different sets of state rules.

188
00:11:33,881 --> 00:11:34,901
That's just crazy.

189
00:11:34,901 --> 00:11:39,041
But if they all allow you to use the NIST
AI risk management framework, you're good.

190
00:11:39,101 --> 00:11:40,901
'Cause that's relatively straightforward.

191
00:11:41,991 --> 00:11:47,316
And so do you feel like this framework
is already addressing, all of the

192
00:11:47,346 --> 00:11:49,596
risks that you had mentioned at
the beginning of the conversation?

193
00:11:50,661 --> 00:11:53,521
Yeah, the issue with the
framework is it's pretty broad.

194
00:11:53,521 --> 00:11:54,871
It's principles based.

195
00:11:55,501 --> 00:12:00,821
So if you have good intent, you can
follow that framework and do a really

196
00:12:00,821 --> 00:12:05,021
good job setting up your governance,
risk compliance, and cyber security

197
00:12:05,021 --> 00:12:09,821
for AI, your use cases, your policy,
your procedures, assign the right

198
00:12:09,821 --> 00:12:12,701
people at the right places, doing
things ethically and things like that.

199
00:12:12,701 --> 00:12:18,551
But if you're trying to, be the
opposite of that, shall we say,

200
00:12:19,121 --> 00:12:20,471
you can find ways around it.

201
00:12:20,471 --> 00:12:24,111
So that's one of the issues
is, it's broad enough that it

202
00:12:24,141 --> 00:12:26,241
allows for loopholes, right?

203
00:12:26,241 --> 00:12:30,276
Not loopholes that necessarily
let you, go against the law.

204
00:12:31,111 --> 00:12:36,391
Loopholes that might make things easier
for you or cut corners or diffuse

205
00:12:36,391 --> 00:12:39,121
responsibility, things like that.

206
00:12:39,121 --> 00:12:43,431
So you do have to be very
thoughtful, in, implementing the

207
00:12:43,431 --> 00:12:45,111
AI risk management framework.

208
00:12:45,411 --> 00:12:51,101
And look, if I were a bank or a
high risk AI user in Colorado, I'd

209
00:12:51,101 --> 00:12:54,461
go through the 29 pages and just
say, what are their expectations?

210
00:12:55,211 --> 00:12:59,741
And if my risk management framework,
covers 99% of that, great, if it

211
00:12:59,741 --> 00:13:03,371
covers half of that, I'd think pretty
hard about adding in some other

212
00:13:04,001 --> 00:13:06,891
features and guidelines to my my AI.

213
00:13:07,386 --> 00:13:11,606
Well, I guess most people, most
organizations in that situation, they,

214
00:13:12,146 --> 00:13:17,576
it's not that they don't have the desire
or that somehow they're not, wanting

215
00:13:17,636 --> 00:13:19,616
to be in compliance with regulation.

216
00:13:19,766 --> 00:13:22,916
It's just a question
of time, cost, effort.

217
00:13:22,916 --> 00:13:23,126
No?

218
00:13:24,091 --> 00:13:25,131
Yeah, I agree with that.

219
00:13:25,131 --> 00:13:26,181
And it's also knowledge.

220
00:13:26,211 --> 00:13:30,681
I had talked to a large bank the other day
and I was talking about the Colorado AI

221
00:13:30,681 --> 00:13:32,271
Act, and they're like, the Colorado what?

222
00:13:32,331 --> 00:13:35,181
And I'm like, this already
applies to you, right?

223
00:13:35,571 --> 00:13:39,051
You just need one customer in
Colorado and this law applies to you.

224
00:13:39,101 --> 00:13:42,671
So I think a lot of it's knowledge
and things are just moving so fast.

225
00:13:43,286 --> 00:13:45,236
We literally had a client
say, this is hilarious.

226
00:13:45,236 --> 00:13:49,676
Client come to us and said, could
you use AI to keep track of the

227
00:13:49,676 --> 00:13:51,956
different AI rules that are coming out?

228
00:13:52,256 --> 00:13:53,996
And we said, yeah, we could do that.

229
00:13:54,386 --> 00:13:59,726
So using AI to keep track of,
uh, AI rules, I'm  all right.

230
00:13:59,726 --> 00:14:01,556
That's, uh, that's a very,  uh,

231
00:14:01,776 --> 00:14:02,536
Next level.

232
00:14:02,596 --> 00:14:03,656
Inception.

233
00:14:03,696 --> 00:14:04,046
Yes.

234
00:14:04,046 --> 00:14:04,986
Exactly.

235
00:14:06,266 --> 00:14:11,326
And speaking of things, moving very
quickly on the, I guess you meant

236
00:14:11,326 --> 00:14:16,046
the regulatory side, but the tech
side is also moving very quickly,

237
00:14:16,556 --> 00:14:19,136
or at least that's my impression.

238
00:14:19,136 --> 00:14:19,856
What do you think?

239
00:14:19,856 --> 00:14:20,786
Where are we?

240
00:14:21,086 --> 00:14:21,756
Super fast.

241
00:14:21,756 --> 00:14:25,476
Look, I think I talk to a lot of people
and most people I talk to think the

242
00:14:25,476 --> 00:14:27,396
technology is moving super quickly.

243
00:14:28,006 --> 00:14:32,066
I think that the deep learning stuff,
the deep research stuff is amazing.

244
00:14:32,796 --> 00:14:37,656
And, it's impacting, all kinds
of businesses at this point.

245
00:14:37,656 --> 00:14:40,286
And we're just, and most
people in businesses are

246
00:14:40,286 --> 00:14:41,786
just scratching the surface.

247
00:14:41,786 --> 00:14:45,866
I mean, you know, top of the
second inning maybe, given that AI

248
00:14:45,866 --> 00:14:47,756
has been around since the 1950s.

249
00:14:48,326 --> 00:14:53,606
But in AI security and privacy,
kinda the stuff we do, it's

250
00:14:53,606 --> 00:14:54,716
the top of the first inning.

251
00:14:54,716 --> 00:14:58,776
Like even a lot of the big
companies, they're just, they're

252
00:14:58,776 --> 00:15:01,176
literally saying, yeah, we don't,
we haven't figured it out yet.

253
00:15:01,226 --> 00:15:02,061
So for example.

254
00:15:03,176 --> 00:15:08,386
If you look at some of the largest
language model producers, and you do

255
00:15:08,386 --> 00:15:14,056
the 350 question, like ethics test, like
a lot of them are failing it, right?

256
00:15:14,106 --> 00:15:16,746
And it's the questions like, can you
even tell me how to build a nuclear bomb?

257
00:15:16,746 --> 00:15:19,926
Tell me how to build a bio weapon,
show me how to build malicious code

258
00:15:19,926 --> 00:15:23,196
to do X or Y or Z. It's bad, right?

259
00:15:23,196 --> 00:15:28,656
And, as most of our listers probably know,
DeepSeek is, is the worst failure there.

260
00:15:28,656 --> 00:15:34,116
That's basically not blocking
anything other than stories

261
00:15:34,116 --> 00:15:35,136
about the Tiananmen Square.

262
00:15:35,751 --> 00:15:37,371
Mm. Which is ironic.

263
00:15:37,371 --> 00:15:37,701
Yes.

264
00:15:38,511 --> 00:15:42,451
And so do you think there's, if I
understand your answer correctly there,

265
00:15:42,451 --> 00:15:47,711
there's, you don't see a, an AI winter
returning, with any type of saturation

266
00:15:47,716 --> 00:15:49,391
in the research on foundational model?

267
00:15:49,751 --> 00:15:56,381
For you the architecture that we have
to is likely to take us to completely

268
00:15:56,381 --> 00:16:00,191
new levels that we haven't anticipated,
like within the next couple years?

269
00:16:00,761 --> 00:16:02,051
Yeah, it's, it is interesting.

270
00:16:02,051 --> 00:16:02,171
Here.

271
00:16:02,321 --> 00:16:03,161
Here's what I'd say.

272
00:16:03,161 --> 00:16:08,371
I'd say that, we will continue to
make advances on the gen AI front.

273
00:16:09,041 --> 00:16:11,851
, We will hit a ceiling
at some point, right?

274
00:16:11,921 --> 00:16:18,951
we may get to the point where it's not
gen AI, but it fools some people, right?

275
00:16:18,981 --> 00:16:19,742
That's plausible.

276
00:16:19,851 --> 00:16:21,111
This is statistical AI.

277
00:16:21,111 --> 00:16:25,472
By definition it can't be gen AI
'cause it's statistical, right?

278
00:16:26,957 --> 00:16:29,477
Although, like I said, it could
potentially fool some people.

279
00:16:29,477 --> 00:16:31,037
It's not doing reasoning, right?

280
00:16:31,037 --> 00:16:34,277
It's literally figuring out what's
the next word in the sentence.

281
00:16:35,057 --> 00:16:39,987
That being said, and when I was
talking earlier today to, a Carnegie

282
00:16:39,987 --> 00:16:44,247
Mellon professor about AI, we're
probably gonna see something

283
00:16:44,247 --> 00:16:46,087
different, a decade or two from now.

284
00:16:46,147 --> 00:16:50,737
In other words, whenever we reach AGI,
whenever that is, it's unlikely to be

285
00:16:50,737 --> 00:16:52,657
large language models that are the key.

286
00:16:52,657 --> 00:16:54,577
Now, it could be a composite model, right?

287
00:16:54,577 --> 00:16:59,497
'cause language is one of the most
complex things to figure out with AI.

288
00:16:59,527 --> 00:17:04,657
So maybe it's LLMs over here and
quantum computing over there, and

289
00:17:04,857 --> 00:17:06,897
traditional machine learning over
here and something we haven't

290
00:17:06,897 --> 00:17:09,357
even heard of as the fifth part.

291
00:17:09,537 --> 00:17:14,157
And the, I think the, one of
the most important advances that

292
00:17:14,217 --> 00:17:19,407
DeepSeek borrowed from OpenAI was
the concept of mixture of experts.

293
00:17:20,427 --> 00:17:24,147
So a bunch of different models
that are good at different things,

294
00:17:24,147 --> 00:17:27,627
which if you think about it is
like the human brain, right?

295
00:17:27,627 --> 00:17:33,027
So, uh, you've got one area that
does language and one area that does

296
00:17:33,507 --> 00:17:38,817
feeling, and, another area, that
does visual part of your cortex.

297
00:17:39,117 --> 00:17:39,807
It's all there.

298
00:17:39,807 --> 00:17:44,457
That may not be how we lay
out AI in the future, right?

299
00:17:44,457 --> 00:17:48,747
It may be differently structured than the
human brain, but the concept of mixture

300
00:17:48,747 --> 00:17:54,867
of experts has allowed large language
models to go from this is kind of okay to

301
00:17:54,867 --> 00:17:58,137
like, wow, this is really, really good.

302
00:17:58,137 --> 00:18:06,162
And, that's what allowed DeepSeek to kinda
leapfrog some of its capabilities as well

303
00:18:06,162 --> 00:18:14,392
as apparently using Open AI's answers to
train on, which is obviously, prohibited

304
00:18:14,392 --> 00:18:17,142
by the Open AI terms of, of use.

305
00:18:17,629 --> 00:18:23,499
And so there's the current
generation, before we get something

306
00:18:23,649 --> 00:18:27,739
that's significantly different
and maybe takes us to AGI.

307
00:18:28,119 --> 00:18:34,249
With the current generation of
artificial intelligence, particularly

308
00:18:34,309 --> 00:18:39,199
generative AI, that we've been talking
about most of this episode, do you

309
00:18:39,199 --> 00:18:45,559
feel like the ethical questions are
also only, what's the expression used,

310
00:18:45,559 --> 00:18:47,839
at the second inning or something?

311
00:18:48,517 --> 00:18:51,297
Yeah, early days on ethics, right?

312
00:18:51,457 --> 00:18:52,392
I think that there.

313
00:18:53,487 --> 00:18:57,057
First of all, let's talk about
ethics in a specific place, right?

314
00:18:57,057 --> 00:18:59,427
So let's say the United States, right?

315
00:18:59,477 --> 00:19:03,707
There are no rules about replacing
people with AI at this point, right?

316
00:19:03,707 --> 00:19:06,557
You could literally fire 20
people and replace 'em with AI

317
00:19:06,557 --> 00:19:10,877
tomorrow if that were possible and
with zero repercussions, right?

318
00:19:10,877 --> 00:19:14,457
That's something I think, each
country and each region is gonna

319
00:19:14,457 --> 00:19:19,267
have to think pretty hard about what
they're gonna allow AI to do and not.

320
00:19:20,017 --> 00:19:25,297
I sit on the executive board of the
Global AI Ethics Institute and we come

321
00:19:25,297 --> 00:19:28,807
at it from a much broader perspective.

322
00:19:28,837 --> 00:19:34,837
And one of the things that is probably
not being discussed enough is how,

323
00:19:34,837 --> 00:19:39,187
obviously different countries are larger
and have larger economies than others.

324
00:19:39,737 --> 00:19:43,307
You've got very wealthy countries
and very poor countries.

325
00:19:44,192 --> 00:19:47,762
And AI may actually
make that worse, right?

326
00:19:47,762 --> 00:19:50,912
Because think about the countries that
really have access to AI right now.

327
00:19:50,912 --> 00:19:57,932
It's basically the US, Europe,
China, Taiwan, these are pretty

328
00:19:58,142 --> 00:19:59,912
wealthy countries already, right?

329
00:19:59,912 --> 00:20:02,892
China's the second largest,
economy in the world.

330
00:20:03,652 --> 00:20:07,422
And then you look at, let's
say, south America, right?

331
00:20:07,422 --> 00:20:10,422
Where AI is just not as adopted.

332
00:20:10,422 --> 00:20:13,032
They just don't have the
same level of access.

333
00:20:13,542 --> 00:20:21,072
So if magically AI makes us companies
even 10% more efficient over the

334
00:20:21,072 --> 00:20:27,222
next few years, but South American
companies are only 1% more efficient,

335
00:20:27,342 --> 00:20:29,862
that's longer term gonna be a problem.

336
00:20:29,962 --> 00:20:33,592
And should that be the case
from an ethical perspective?

337
00:20:34,207 --> 00:20:38,967
So I think, think of these, AI chips
as a scarce commodity that right

338
00:20:38,967 --> 00:20:40,857
now is being allocated by dollars.

339
00:20:40,857 --> 00:20:42,207
Who's got the money for them?

340
00:20:42,257 --> 00:20:46,397
So that's just something again, that we're
gonna have to think about pretty hard.

341
00:20:46,397 --> 00:20:48,917
And then obviously the
software as well, right?

342
00:20:49,097 --> 00:20:54,147
With DeepSeek, basically having
no rules, and Open AI, according

343
00:20:54,147 --> 00:20:58,137
to independent scoring, having
the most rules around ethical AI

344
00:20:58,137 --> 00:21:00,417
use, but still not perfect, right?

345
00:21:00,637 --> 00:21:05,627
And that's something that is
unlikely to change in the future.

346
00:21:05,627 --> 00:21:12,777
There's always gonna be either a model
out there that's not compliant, or someone

347
00:21:12,777 --> 00:21:17,697
will find out a way to take a model that's
theoretically compliant and jailbreak it

348
00:21:17,697 --> 00:21:22,772
or change it, or manipulate an open source
model to get what they want out of it.

349
00:21:23,131 --> 00:21:26,936
And the organization that you mentioned,
can you tell us more about it?

350
00:21:27,311 --> 00:21:28,331
Sure, yeah.

351
00:21:28,331 --> 00:21:32,611
So the Global AI Ethics Institute, has,
thousands of members around the country.

352
00:21:32,661 --> 00:21:34,251
We sponsor research.

353
00:21:34,301 --> 00:21:40,011
We have a white paper prize
each year on a specific topic to

354
00:21:40,011 --> 00:21:42,881
doing to do with ethics, and AI.

355
00:21:43,421 --> 00:21:46,681
And we're shortly gonna announce
our winner for last year.

356
00:21:47,381 --> 00:21:51,811
And basically we will publish
research on its website.

357
00:21:51,811 --> 00:21:56,011
It's a nonprofit and pretty much anyone
in the AI space is welcome to join.

358
00:21:56,721 --> 00:21:58,686
I will make sure to provide
the link in the description.

359
00:21:59,366 --> 00:22:02,756
And as we're about to wrap up, I usually
ask a couple of questions at the end.

360
00:22:02,826 --> 00:22:07,546
The first one is, what's a book,
tool or habit that has made the most

361
00:22:07,546 --> 00:22:09,676
impact on you in the last 12 months?

362
00:22:10,260 --> 00:22:10,500
Yeah.

363
00:22:10,560 --> 00:22:11,850
Good, good questions.

364
00:22:11,900 --> 00:22:15,620
I like Reid Blackman's book "Ethical AI".

365
00:22:15,870 --> 00:22:20,100
Excellent book with some
very practical examples.

366
00:22:20,190 --> 00:22:26,670
Every chapter is basically a
graduate school class in AI ethics.

367
00:22:26,670 --> 00:22:27,750
It's pretty amazing.

368
00:22:28,500 --> 00:22:31,770
'cause he's a professor and that's
kind of how we wrote the book.

369
00:22:32,370 --> 00:22:33,930
Uh, so that's a great book.

370
00:22:33,930 --> 00:22:36,900
If, uh, if you have, have not read it.

371
00:22:37,163 --> 00:22:41,603
And finally, given everything we've
discussed that's changing, is there maybe

372
00:22:41,603 --> 00:22:45,698
something, one thing that you would expect
to remain the same 10 years from now?

373
00:22:45,819 --> 00:22:46,329
Wow.

374
00:22:46,389 --> 00:22:49,569
I would say the thing I expect
to re remain the same is our

375
00:22:49,569 --> 00:22:51,969
need for human connection, right?

376
00:22:51,969 --> 00:22:55,359
Get out there, talk to
people, do things in person.

377
00:22:55,959 --> 00:22:58,809
It's so easy just to be like,
yeah, we'll just do a zoom meeting.

378
00:22:59,489 --> 00:23:01,159
Go to things.

379
00:23:01,179 --> 00:23:07,869
I have made way better connections with
people by simply going to their office

380
00:23:07,869 --> 00:23:12,499
or meeting them in person at a conference
than anything you could do over video.

381
00:23:12,849 --> 00:23:16,419
I think that's, uh, something
that a lot of our members of

382
00:23:16,419 --> 00:23:17,589
the audience will agree with.

383
00:23:18,279 --> 00:23:19,804
With that, thank you very much, Alec.

384
00:23:20,619 --> 00:23:21,084
Thank you Alex.

385
00:23:21,174 --> 00:23:21,849
Great to be here.

386
00:23:22,801 --> 00:23:26,671
Beyond innovation, AI adoption
requires security, control and trust.

387
00:23:27,511 --> 00:23:31,681
As Alec explained, enterprises need
a clear strategy to deploy AI safely

388
00:23:32,101 --> 00:23:36,301
from securing sensitive data to staying
compliant with evolving regulations.

389
00:23:37,021 --> 00:23:38,526
His insights raise a key question.

390
00:23:38,746 --> 00:23:42,271
Do you have a comprehensive
plan for managing AI risks?

391
00:23:43,261 --> 00:23:46,321
If today's conversation made you
rethink your approach, you can

392
00:23:46,321 --> 00:23:50,611
explore a hands-on demo of the
Artificial Intelligence Risk platform.

393
00:23:50,761 --> 00:23:52,201
The link is in the description.

394
00:23:52,951 --> 00:23:57,271
As always, we have more exciting topics
and guest appearances lined up, so

395
00:23:57,271 --> 00:24:01,651
stay tuned for more tales of innovation
that inspire, challenge and transform.

396
00:24:02,191 --> 00:24:03,991
Until next time, peace.

397
00:24:05,404 --> 00:24:07,744
Thanks for tuning in to Innovation Tales.

398
00:24:12,484 --> 00:24:16,174
Get inspired, connect with other
practitioners and approach the

399
00:24:16,174 --> 00:24:18,214
digital revolution with confidence.

400
00:24:19,144 --> 00:24:22,534
Visit innovation-tales.com
for more episodes.

401
00:24:23,874 --> 00:24:24,894
See you next time.