Latest / The Road to Accountable AI / Dominique Shelton Leipzig: Building Trust When Every Company is an AI Company

Transcript

Dominique Shelton Leipzig: Building Trust When Every Company is an AI Company · The Road to Accountable AI

0:01
Dominic Shelton -Laitzig, thank you so much for joining us on the road to
0:04
accountable AI.
0:06
Glad to be here.
0:07
Thank you for having me.
0:09
Tell us first about your background.
0:11
How did you get involved in technology law?
0:15
Well, it's been a really interesting road.
0:19
It would seem that it was all organized, but actually it's just a series of sort of
0:26
a confluence of different synergies.
0:29
So I've been practicing in the area of data and privacy, cyber ever since 1998
0:36
after studying at Brown University.
0:38
I studied international relations and French civilization.
0:42
So nothing to do with technology then.
0:44
And, but I was really interested in world affairs and policy.
0:48
So that's why after Brown, I went over to Georgetown University Law Center.
0:55
And my favorite course I remember was Privacy in American Law.
1:00
This predates the internet and everything.
1:04
This was back in 1990.
1:06
So fast forward eight years later, I started working for a firm that...
1:10
no longer exists, but a small boutique that did a lot of work for Yahoo and
1:15
PeopleSoft and some of the other technology companies.
1:18
It San Francisco -based firm.
1:21
I was the fifth lawyer in the LA office.
1:23
So from there, I just really stuck with the issues dealing with these tech
1:30
companies way back in 1998 and 99.
1:34
Their privacy issues articulated around ad tech and then later data breach.
1:40
in 2012.
1:41
So from there, just continued with it.
1:44
At this point in time, I've trained over 50 ,000 professionals on AI, privacy, and
1:50
cyber.
1:51
I spent really most of my time looking at legal trends in order to be able to
1:56
forecast for companies what's coming down the pike and how to be data leaders, not
2:02
just sort of skimming by with the minimum that they can do in the law, but really
2:06
excelling.
2:07
So my book, Trust, is
2:10
Trust, Responsible AI, Innovation, Privacy, and Data Leadership.
2:13
It's my fourth book.
2:15
And I really just try to focus there on forecasting those trends again.
2:21
One other thing I'll just mention in terms of how things can come full circle, as I
2:26
told you, I was really interested in international affairs.
2:28
That's why I was an international relations major at Brown University.
2:33
And how privacy and cyber and now AI have become global issues.
2:38
is really exciting over the past five years.
2:41
So I did an op -ed for the Financial Times in 2020 to deal with the stalemate of
2:49
crisis of data going back and forth between Europe and the US, which
2:54
jeopardized about 7 .1 trillion in commerce that was going on at the time.
3:00
So I wrote an op -ed that I wanted to...
3:02
to use in the financial times to influence policy.
3:05
And it was really rewarding to see President Biden adopt my recommendations
3:09
for an executive order to deal with the redress issues post -Shrimps.
3:15
So that's really what takes me here today.
3:19
No, a lot to cover.
3:21
Let me back up in what you said.
3:23
You mentioned that privacy has been around for a long time and cybersecurity issues
3:28
have been around for a long time.
3:29
How does that connect with AI and the issues that companies are facing today?
3:34
Well, yeah, so looking at the trustworthy AI legal frameworks, and there are 97
3:40
countries and six continents right now that have draft legislation or regulatory
3:46
frameworks that were issued by regulators and not laws.
3:50
But all of them that I've seen, I had an opportunity to canvas this in connection
3:54
with the research for the book and also advice I give to clients, all of them call
3:59
for, in addition to risk ranking the AI,
4:03
For high risk AI, they do call for continuous testing, monitoring, and
4:07
auditing of AI in some key areas.
4:10
And one of those, well, two of those areas include privacy and cyber, but the other
4:15
five, bias, IP, accuracy, health and safety are some of those.
4:20
So that's how it fits in.
4:23
It's going to be a, in Europe, we'll go first as far as the final laws, but in
4:28
Europe, it'll be a requirement to continuously test, monitor, and audit.
4:32
high -risk AI in the areas of privacy and cyber, along with five other areas.
4:37
You talk about in the book that every company now is a data company, regardless
4:41
of what industry they're in.
4:42
What exactly does that mean?
4:44
It means that data, although it doesn't appear on a balance sheet, unless you're
4:49
doing M &A and it's buried in an intangible asset, for the most part, data
4:54
does not appear on a balance sheet.
4:56
But as we've seen, Kevin, it was so critical during the pandemic, those
5:01
companies that did not have a data strategy and didn't have a good digital
5:06
game plan to pivot to are just not with us right now.
5:10
So that is just how critical data is to every single...
5:14
And when you think about how AI is going to fuel the insights that can be gained
5:19
from data, it's really going to be the bedrock of the future of successful
5:24
companies.
5:25
So would you say every company is now an AI company?
5:29
Uh, you know, every company is probably an AI company already right now between
5:35
whether they're using AI directly or they're relying on vendors that are using
5:40
AI.
5:40
But this is a, uh, we did almost close to 1 trillion in licensing deals, uh, last
5:47
year on AI.
5:48
So, uh, these projects are being deployed in pilot phase and enterprise, uh,
5:54
offering.
5:55
So, uh, this is going on all around us, all around the world.
5:59
How do you think about the topic of what I call accountable AI, but trustworthy AI,
6:05
responsible AI, AI governance, there are all these variants on the theme.
6:10
How would you define what that category means?
6:13
So prior to the legislative and regulatory frameworks, it was really kind of bespoke
6:22
and it depends upon what company you're in and what frameworks that you're following,
6:29
the NISC risk management framework, the White House AI Bill of Rights we were
6:33
talking about, one of the professors that helped draft that computer science large
6:39
language model expert.
6:41
So there are a lot of different.
6:43
and best practices floating around.
6:46
But now, where the legislation and the regulatory frameworks have stepped in, and
6:53
we've seen just an explosion of those, starting three years ago, but really
6:59
picking up steam with the release of ChatGPT4 last February and sort of the
7:07
explosion of this technology and the...
7:10
regular parlance of popular culture.
7:13
There's also been a rush to issue legislation and guidance.
7:19
And so it means very specific things for the legislators and regulators right now.
7:27
So that's an interesting point.
7:28
Is it a good thing that now it's no longer just companies figuring out on their own
7:33
what it means to be responsible, but now we've got all these specific regulatory
7:36
compliance requirements?
7:39
I think it is a good thing because it actually gives companies an opportunity to
7:43
benchmark against a standard rather than sort of be a little bit at sea and lost
7:51
about how to exercise governance.
7:53
The legislative frameworks and the regulatory guidance is actually very
7:57
didactic and it lays out what they all use the term that they're there to codify
8:04
trustworthy AI.
8:05
That's what the EU AI Act.
8:07
says right in the opening.
8:09
And that's what we're seeing.
8:11
And like I said, the 70 other countries that have similar frameworks on the table.
8:16
What they're looking for is first a risk ranking of the AI.
8:21
And they're very clear about what constitutes these three buckets,
8:26
prohibited AI.
8:28
So there's 17 areas where governments around the world just don't want private
8:33
companies involved in those activities.
8:36
things like continuously monitoring people in public or social scoring using, Kevin,
8:43
what you might post on LinkedIn or social as a basis for making a decision about
8:49
whether you should be a company's customer or not.
8:51
There's no governance around that activity because regulators around the world don't
8:55
want private companies engaged in that.
8:58
The other kind of risk ranking category, which I call the green light, it's
9:03
prohibited, maybe think of that as a stoplight.
9:05
tap the brakes, don't go there.
9:07
The green light is low risk AI, and that's things like chat bots for non -sensitive
9:13
use cases, going to a website to buy some jeans, and you're getting a suggestion
9:18
about what to purchase from the chat bot.
9:21
There's very little governance there because they consider it low risk.
9:25
And frankly, the view is most people have been interacting with an AI in that
9:31
context.
9:31
Think of Siri or your mapping app where you're
9:34
tells you, it gives you the voice to tell you which way to turn in directions to get
9:40
to a location.
9:41
Those were mostly powered by predictive AI.
9:46
So the view is, well, if a chat bot's powered by generative AI, the relative
9:52
risk of that is low.
9:54
So all they're asking there, these regulatory legal frameworks, all they're
10:00
asking there is that there be disclosure to...
10:03
end users that they're interacting with an AI.
10:06
The need of the governance, and we'll talk about that later, is in high risk.
10:10
And there's 141 examples.
10:13
Every company ought to know what those are and then know the governance steps that
10:17
the regulators think are important.
10:21
So this all sounds great for lawyers.
10:24
You're a lawyer, I'm a lawyer.
10:26
There's a lot of requirements now on companies.
10:28
But when you talk with boards, which I know you do a lot, how do you convince
10:34
them that this is not just something where they need to check a box because there's a
10:39
legal requirement they're going to have to meet, but maybe it's not even applicable
10:42
to them yet.
10:44
Maybe these laws are still on the drawing board.
10:46
How do you convince them to really invest both the resources and the time to be
10:52
serious about this kind of governance process?
10:56
Well, I talk about this in the book and you're right, when I'm talking to boards
11:00
and to CEOs about these concepts, I frame it in terms of unlocking innovation and
11:07
success with their brand, success with the services that they're offering with AI.
11:15
So let me give you some examples.
11:17
I try to give storytelling examples to help crystallize the importance of this.
11:22
So let's just even talk about low risk AI.
11:25
Just about a month ago a company a logistics company in the UK very similar
11:31
to our packaging delivery companies that deliver packages overnight and so forth
11:37
big company in the UK Had a chatbot that they were using to interact with customers
11:43
Letting them know where their when their packages would arrive and etc Suddenly the
11:50
CEO wakes up to a viral
11:53
thread going on on X, formerly Twitter, about the company and the chatbot where a
11:59
customer jumped onto X and posted the chatbot cursing out the customer and
12:05
criticizing the company as being the worst logistic company in the world, et cetera.
12:10
So the company disabled the chatbot.
12:14
Actually, they didn't need to disable the chatbot.
12:16
They just needed to insert guardrails that would have let the company know.
12:22
that the chatbot was cursing out customers.
12:25
So it's the type of thing that is very avoidable.
12:30
It doesn't need to happen at scale and you can protect your brand in the process.
12:35
So the point is that governance, putting the things in place that the legal
12:43
frameworks require, for example, continuous testing, monitoring and
12:47
auditing on things like accuracy, you know,
12:52
health and safety, privacy, cyber.
12:55
What that means is code needs to be entered into the application, the AI tool
13:01
itself, with the guardrails of what is acceptable to the company.
13:05
So in terms of accuracy, just think about, rewind the story I just told you and what
13:10
if accuracy meant that when talking to customers, it has to start with a please
13:15
and a thank you, and it's going to be inaccurate in their environment in any
13:19
language the company does business.
13:21
to curse at the customers, for example.
13:24
So rather than have tens of thousands of people experiencing this that are your
13:29
customer base, instead, the first time that the chat bot cursed at the first
13:34
customer, the company could be alerted, the governance, the human and the
13:38
governance team could be alerted.
13:40
Look at the logging data and metadata associated with that chat bot, see when it
13:46
started to drift into cursing activity, and just fix the model, right?
13:51
So it's nothing about waiting till two years until the law becomes final to
13:56
protect your brand.
13:57
You can jump in right away and do that.
13:59
And that's why it makes sense to do it.
14:02
It's a great example.
14:04
Is there any way that from the outside, I can tell if a company is really serious
14:12
about their responsible AI practices?
14:14
That might be as a customer, it might be a potential business partner that I'm trying
14:18
to decide whether to have a relationship with this company.
14:20
Is there any way to look and be able to, or what questions should one ask to make
14:24
that kind of assessment?
14:26
Yeah, so the questions to ask are, first of all, are you risk ranking the AI?
14:31
Where do we stand there?
14:34
Do you know if you're engaged in any high risk AI?
14:38
For high risk AI, are you making sure there's high quality data, accurate data
14:44
going in and accurate data coming out?
14:48
That's a key issue.
14:51
In terms of accuracy,
14:53
Are there IP rights, privacy rights to use the data that is being used to train the
14:57
model?
14:58
And then with respect to deployment and you're going to interact and engage a
15:04
vendor, for example, the question is, is there a mechanism for continuously
15:09
testing, monitoring and auditing this AI?
15:12
Why?
15:12
Because in the generative AI context,
15:15
the large language model experts, the folks at Wharton that you work with and
15:21
other data scientists will tell you that state of the technology is that models
15:26
move.
15:27
We want them to move because they're being informed by the whole wide internet that
15:31
is being added to, you know, we generate 2 .5 quintillion bytes, 18 zeros behind it
15:38
of data per day.
15:40
All of that roiling new information is getting into the model and making it.
15:45
smarter and having the ability to have greater insights.
15:49
But that means that what you tested it on today, it's going to move.
15:55
It won't be the same necessarily tomorrow or 30 days from now or 60 days from now.
16:02
So the only way for the company to know that its values are going to remain in the
16:07
model like it started, taking an example again that I mentioned when the
16:12
Chatbot was first working, it wasn't cursing out customers, but later over time
16:17
developed that language.
16:18
The only way the company's going to know about that is if there are guardrails that
16:22
it can be basically like a sensor on your window at home when the alarm goes off.
16:28
That's the only way you're going to know that a burglar has entered.
16:32
Similarly, you're not going to know that the model's drifting unless there's
16:36
guardrails to let the company know.
16:39
As you mentioned, the title of your new book is Trust.
16:44
Actually something I've written about too, in a different context or related context
16:48
having to do with blockchain technology as a new form of trust.
16:51
But I want to ask you about AI.
16:53
So why is it that you think trust is an important value or an important term to
16:58
use and how companies should think about AI?
17:01
Yeah, so that is just really plucked directly from the 97 regulatory and draft
17:08
legislative frameworks in 97 countries, six continents that do talk about
17:15
trustworthy AI.
17:16
And they say we're codifying trustworthy AI.
17:20
So given that that's the legal standard, I think it's the sooner we start adopting
17:27
that language, the better, because that is
17:30
That's what the standard is.
17:32
But more importantly, I think what I found interesting in my research and what has
17:37
been so helpful is to see that this time the legislation actually was built with
17:44
the large language model experts advice in mind.
17:49
So legislators around the world sat with large language model experts and data
17:54
scientists around the world and said, look, what if we were dealing with AI?
17:59
not about a chatbot that curses out customers, which is embarrassing, but
18:03
again, not really life -threatening.
18:08
But what about a situation where we have AI in our power grid or our hospital
18:12
systems or our water supply, and we have to deal with drift and we need to get an
18:20
AI model back on track where we really don't have room for error?
18:24
What would you need, data scientists, to be able to correct a model quickly?
18:28
And what they said was just what I told you, continuous testing, monitoring, and
18:32
auditing so that there is a baseline, technical logging data and metadata so
18:38
that when the model drifts, the data scientists have something to look at and
18:43
be able to track when did the model start to move to diagnose the issue and correct
18:47
the model and get it back on track as soon as possible.
18:50
So that's how we can trust that when AI enters our hospital systems, financial
18:55
networks, our...
18:58
power grid and electrical supply and also, let alone national security, that it's
19:05
trustworthy.
19:08
Is it more important though to trust the AI systems or to trust the organizations
19:13
that are deploying them?
19:16
Both, right?
19:19
The organizations, and let's just talk about companies, because that's what I
19:25
know best and I work with, they're in the business to, we talk about brand and
19:33
goodwill, that does appear on a balance sheet.
19:38
And a big bedrock of brand is trust, right?
19:43
If...
19:44
Customers do not trust a company.
19:46
We've seen this over and over again I was reading something in Forbes recently that
19:51
companies that are trusted are 400 % more successful The companies that have broken
19:57
the three trillion market cap club right now just a few of them there they put
20:03
trust front and center in their branding and their communications with customers
20:10
Trust is right there trust in privacy cyber
20:14
AI, people are entrusting their data with a company and they want to know that the
20:21
company takes that responsibility seriously.
20:24
And the only way to show that is to actually be aware of the legal
20:30
requirements and follow them.
20:34
So clearly you're right, every company wants to be trusted, but it's not
20:40
costless, right?
20:40
I'm sure you've dealt with companies and boards who say, sure, of course we want to
20:44
be responsible.
20:44
We don't want the example you gave with the chat bot or anything else, but it's
20:49
going to slow us down if we have to go through all of these compliance processes
20:54
and talk to our lawyers.
20:55
And the lawyers may say, don't implement this system that we think will have some
21:00
real business benefits.
21:01
and maybe our competitors are not going to be as careful, they're going to implement
21:04
it, how do you convince companies or how do you help them draw those lines most
21:09
effectively?
21:11
Yeah, so I talk about this in the book.
21:14
Rather than, I think the old mantra for Web 1 .0 and 2 .0 was move fast and break
21:20
things, what I've been encouraging is move fast with trust.
21:24
There's not a reason in the world that the steps that I mentioned are just six of
21:28
them.
21:29
High quality data, accurate data going in, accurate data going out is the first step.
21:36
Second step, continuous testing, monitoring, and auditing.
21:38
Third, technical documentation that should
21:41
be maintained from the continuous testing.
21:43
That just happens real time in the background.
21:46
And then a human that if the model starts to drift, a human can come in, look at the
21:51
logging data and metadata and correct the model.
21:53
And then if you can't get the model corrected, a failsafe to kill that AI use
21:57
case, that's six steps.
22:00
It doesn't take time.
22:01
It just takes attention.
22:03
It doesn't require a huge upheaval.
22:07
What requires more and what is more costly,
22:10
is the $1 .4 trillion, for example, that it cost our global economy in 2022 and
22:17
2023 when data privacy was missed, right?
22:21
Same issues, 161 countries have data protection laws.
22:25
And unfortunately, it wasn't something that was talked about in the boardroom
22:28
very often.
22:28
Cyber, yes.
22:30
But cyber, another issue, talked a lot about, but last year it cost our global
22:35
economy $6 .1 trillion in losses.
22:38
So if you take the losses in...
22:40
data privacy, 1 .4 trillion, plus the losses in the cost of cyber, 6 .1
22:46
trillion.
22:47
We're talking close to 8 trillion, which would, if that were GDP, if that were one
22:51
country, would be the third largest GDP in the world behind the US and China.
22:57
So we've got to get savvy economically here and keep that money to grow our
23:03
economies, grow our businesses, grow our customer base by paying attention to this
23:07
very straightforward issue.
23:10
We need to not be intimidated because it's tech.
23:14
You mentioned the European AI Act and some of the AI legislation going on.
23:18
There's this notion, kind of conventional wisdom I hear a lot, that the US is still
23:25
the Wild West, that in the US we have this attitude that companies should self
23:28
-regulate in general with tech, but with AI specifically.
23:32
Do you think that's right?
23:37
Yes and no.
23:38
So first of all, in terms of federal legislation, I think you're right.
23:44
The US is going to be behind Europe.
23:46
Europe has already reached a political agreement and they're on the verge of
23:51
adopting the final language for the final EU AI Act.
23:55
And that will go into effect 24 months after it's adopted formally, which should
24:00
be happening I think around May is the estimate.
24:04
So,
24:05
So May, 2026, we're probably looking at it at where enforcement will come in.
24:09
But we've got a lot going on here in the U .S.
24:12
The executive order, President Biden has issued an executive order on AI.
24:17
Prior to that, that executive order is sitting on top of and refers to the White
24:22
House AI Bill of Rights, as well as the NIST AI risk management framework.
24:28
So there's a lot of good guidance that's very synergistic with what we're seeing
24:33
with the EU AI Act.
24:34
There's these six steps that I've talked to you about exist in every place I just
24:39
mentioned in the US.
24:41
Also, we do have some laws in sectoral areas in the US already on the books.
24:48
So in New York, for example, there's an AI law as it relates to the use of AI for
24:55
employment and HR purposes and hiring purposes.
24:59
We also have laws in Illinois and et cetera.
25:03
related to biometrics that cover AI.
25:06
And then we have data protection laws in California, Colorado, Connecticut, that
25:12
talk of, that give consumers the right to opt out of automated decision making.
25:17
So we have things here, more of a patchwork than what we're seeing in
25:22
Europe, but frankly, all of our significant companies are doing business
25:28
in Europe.
25:28
So they're going to be highly influenced by what happens in Europe.
25:32
Isn't there some benefit of having a more distributed approach like the US where
25:39
laws can be passed more specifically focused on issues as opposed to having one
25:43
huge comprehensive code like they have in Europe?
25:47
Well, you know, as I said, the six steps that the Europeans have put forward, which
25:53
are mirror what is in the technical companion to the AI Bill of Rights that
25:58
came out of the White House, those are just good business, practical things.
26:03
There's just no reason for a company to be completely unaware if their AI is
26:10
denigrating the customers, for example.
26:13
Especially when there's tools.
26:14
at this day and age, 2024, to be able to identify that.
26:19
It's just all the way better for the brand.
26:21
But getting to things like healthcare and drug development, where it's really key
26:26
and critical that companies be aware if, say, for example, the AI is recommending
26:31
the wrong medicine to people, which has occurred, that's just going to hurt the
26:35
brand and hurt business down the road.
26:38
So I don't look at it as tethering to the...
26:43
waiting till the law goes into effect and holding out to the last second when there
26:48
are brand issues right on the table right now with the use of this technology.
26:54
And you wanna use the technology to stay competitive, of course.
26:57
I mean, AI has been used already.
27:00
A month after ChatGPT 4 was released, some doctors gave it, gave all their medical
27:06
data lake to an AI tool and it was able to identify.
27:11
an early treatment for liver cancer.
27:14
It's amazing.
27:15
LA right here at Cedars -Sinai Hospital, they're using AI right now, generative AI,
27:20
to be able to have early identification of pancreatic cancer.
27:24
So this is incredible innovation.
27:27
That's just in the healthcare context, but I mean everywhere, financial, education,
27:31
et cetera.
27:32
So we've got to use this technology, but we have to understand its limitations.
27:37
And one big one is that it drifts.
27:40
So...
27:41
What could be a great thing can be a terrible outcome for really no reason
27:47
other than just not having the guardrails necessary to manage the technology.
27:50
It so happens it's in the law as well, but it's just good business sense.
27:56
How much though will those kinds of obligations fall on the companies that are
28:01
not the ones building the technology?
28:03
So if you're Microsoft or OpenAI or Google, you have that technical expertise,
28:09
you probably have a lot of resources internally going to making the AI
28:13
responsible, doing that documentation, doing the testing and so forth.
28:16
But if you're a manufacturing company or agricultural company, one of these
28:21
companies to your point before that they're all data companies.
28:23
that is using this technology that they're implementing built somewhere else.
28:27
How are they going to have the requisite expertise either to comply with the law or
28:32
to just take those good business steps you're talking about?
28:36
So I'm excited.
28:37
We're going to talk about this at the Digital Trust Summit at Brown.
28:40
The obligations fall on both the providers of large language models or generative AI
28:48
if they have a high -risk use case.
28:52
And I want to put a big if there.
28:53
We'll put a pin, and I'll come back to that.
28:56
And they also fall on those that license those models for their own purposes.
29:02
So both the
29:03
LicenseOr, which is the tech company, and the Licensee, the business that is
29:09
licensing it and maybe training the models with their own proprietary data and
29:14
creating their own applications.
29:17
So the obligations in terms of the trustworthy legal AI frameworks fall on
29:22
both.
29:23
In terms of the ability to incorporate these things into their businesses that
29:30
are licensing the models to incorporate these six steps,
29:33
into their practices.
29:34
There are all kinds of vendors that can help with developing guardrails to put
29:39
into the AI tool.
29:41
Dropping some code into the tool is not particularly expensive exercise.
29:46
It's just knowing that you need to drop the code in to be able to do the testing
29:51
and then developing the guardrails.
29:53
And there are vendors that do this.
29:55
And we're going to talk about who they are.
29:57
And if anybody's listening to this podcast and wants to get those recommendations,
30:01
there are a lot of them.
30:02
that are now interoperable with all the major tech companies tech stack and it's
30:10
very easy to get these gargoyles in place.
30:13
So it's so great that the small and mid -sized companies don't have to go in blind
30:17
just because they're small and mid -sized.
30:20
Yeah, just picking up on what you said, what's the best place or some good places
30:25
you would recommend for someone to go to that just wants to be aware of either what
30:30
the requirements are or what are the best kinds of steps they could take.
30:34
Well, so a couple of places.
30:37
I have a free chapter of the book on my website.
30:41
It's just my name, dominiquesheltonleipzig .com.
30:44
And you can download that and see those six steps that I mentioned.
30:47
Also in the podcast notes, I'm happy to share an article about trustworthy AI
30:54
frameworks that was done by MIT based on a talk that I gave.
30:59
It's detailed also in the book, really a detail.
31:02
400 pages though, so it's a lot to read.
31:04
So if you're not interested in reading all that, the MIT article is about three pages
31:08
long and goes through all the backup.
31:12
Terrific, thank you.
31:13
So one more question.
31:15
One development that we should look to in the next, say, one to three years, either
31:20
something you think is coming with AI generally in terms of the technology of
31:23
the business or something in terms of the legal and responsibility aspects of AI.
31:29
There's one thing that we should be on the lookout for that you think is going to
31:33
happen in the near future.
31:34
What would you say?
31:36
I think we're probably two or three years away from getting large language models
31:44
coming out where the drift issue is taken care of, where they don't move as much.
31:52
So probably two to three years out from that, from what I understand from the
31:58
large language model experts, the computer and data science teams working on this,
32:02
but particularly in the area of...
32:05
AI used to apply to the genomic stream and find early identifications for generative
32:11
diseases like ALS or pancreatic cancer.
32:16
A lot of advanced work is going on by Nobel laureates and et cetera to see if
32:21
there's a way to tie trade on that and basically control the movement.
32:27
So we don't have to worry so much about the guardrails.
32:29
They'll be built into the models themselves.
32:31
We're probably about two, three years out from that, but I'm excited.
32:34
about that opportunity.
32:36
Wonderful.
32:36
Dominic Shelton -Lypsick, thank you so much for joining me on the road to
32:39
accountable AI.
32:41
Thank you for having me, Kevin.