Latest / The Road to Accountable AI / Navrina Singh: AI Alignment in Practice

Transcript

Navrina Singh: AI Alignment in Practice · The Road to Accountable AI

0:01
Navrina, it's a pleasure to have you on the podcast.
0:04
People have been talking about and studying questions about risks of AI and
0:10
bias and so forth for a long time, relatively speaking.
0:14
What gave you the idea in 2020 to set up a company to do the kind of work that you
0:19
do?
0:21
Kevin, thank you so much for having me.
0:23
So, you know, I've spent almost two decades in high tech building products
0:28
focused on mobile and artificial intelligence.
0:31
And I would say that, especially in the past decade, I saw not only the
0:36
proliferation of this very powerful technology impacting pretty much every
0:41
individual, every citizen you can potentially think about, but also I saw a
0:46
massive shift in the enterprises.
0:49
where the way we were building machine learning applications had a massive
0:53
oversight deficit.
0:55
So let me just unpack that for you what that means.
0:57
Most of the teams that I've managed in my career are product and data science teams.
1:03
And as you can imagine, the incentives for those teams are really to just create the
1:08
fastest performing models, put them out in the market, and get to the business
1:12
outcomes that we are looking for.
1:15
But given...
1:16
the impact of artificial intelligence and given that it is a socio -technical
1:21
technology, that means it has a far reaching impact than some of the
1:25
traditional software has.
1:27
What we found was that the way machine learning was getting built in organization
1:31
needed more compliance, needed more policy perspective, needed more understanding of
1:37
business objectives so that you could actually put out this machine learning
1:42
capabilities out in the market.
1:44
that was not only serving the users, but also the businesses.
1:47
And I would say that that's what I found that in the way ML Ops was showing up in
1:53
the market, there was an increasing deficit in terms of accountability, in
1:59
terms of how we were building these systems, designing these systems, testing
2:03
these systems, especially in production.
2:06
And so what the idea really was, is there a standardized and scalable way?
2:11
to bring that oversight and accountability to the socio -technical technology, which
2:18
is going to have such a far reaching impact in the world that we've not even
2:22
studied and grappled with with the past technological revolutions.
2:26
And that's what really led me to creation of Credo AI in 2020.
2:32
What do you say to companies that are still of the view that they just want to
2:37
do what you said, get the best business benefits out of their AI?
2:39
Of course, maybe they need to test it.
2:41
Maybe they'll do some bias evaluation, but they don't want to invest the resources or
2:48
take the time to do that kind of oversight.
2:52
You know, Kevin, I'm a firm believer that in this new age of human productivity
2:59
powered by artificial intelligence, the only organizations and enterprises that
3:05
are going to exist are the enterprises that are going to lead with trust, that
3:11
are going to be the enterprises that double down on governance and use that as
3:16
a competitive advantage, not as just a checkbox.
3:20
And so the thing that I would like to say to those enterprises is you won't exist
3:25
tomorrow if you don't invest in governance and oversight today so that you can have
3:31
accountable technology that is actually serving not only your business objectives,
3:35
but also consumer needs.
3:38
So maybe you could talk a little bit more specifically about what kinds of
3:42
functionality you provide at Credo AI and how does it actually promote that kind of
3:46
trust that you described.
3:48
Yeah.
3:48
So Kevin, as I mentioned, AI is a very socio -technical challenge.
3:54
And so one of the things that we are finding is defining what you're going to
4:00
measure to make sure that the AI systems actually works in your behalf and in your
4:04
favor is very challenging.
4:07
So for us at Credo AI, we sort of stepped back and really went back to first
4:12
principles and thought about...
4:14
what is really socio -technical mean and how do we actually figure out what to
4:20
measure?
4:20
And as you can imagine, this was a very interesting interplay, not just in
4:25
building the tech, but also humans who are involved in building that tech.
4:29
So the three things we do really well with our software solution is first, alignment.
4:36
As you can imagine, in AI, and especially because it is so context dependent,
4:42
we are seeing that with foundation models, which are multimodal and can, you know,
4:48
are dual purpose and can be adapted to any use case.
4:51
It is really important to understand the context of use.
4:55
And within that context of use, it is really important to understand what are
5:00
you going to be testing these systems against.
5:02
And this is where the alignment problem becomes really critical.
5:06
So the first step in Credo AI product that we've spent years,
5:11
building and pioneering is really alignment capabilities where we are
5:17
helping organizations define and understand and align on what does good
5:21
look like.
5:23
And within our platform, the way that we are making that happen is first and
5:27
foremost, really providing out of box policy intelligence where we have codified
5:33
what the world is saying what good looks like.
5:35
And this might come from regulations, it can come from standards like NIST, ISO.
5:40
It can come from industry best practices, whether it is defense innovation units,
5:45
responsible AI guidelines, or data interest alliances, requirements, or it
5:51
can come from the companies themselves.
5:53
As you can imagine, there is a need for the AI forward organizations to define
5:59
what they are willing to be held accountable to.
6:03
So how do you codify that within the product is one critical thing.
6:07
And I think for us at Credo AI, the way we are solving that alignment problem is
6:12
really bringing these stakeholders together around these codified set of
6:16
guardrails.
6:17
Once you've done that codification, Kevin, what we end up doing next is using those
6:22
codified policies as a set of controls to really test your entire stack.
6:29
And this goes back to the socio -technical nature because the entire stack does not
6:34
just mean your...
6:36
data sets, your models, your use cases, but it also means the processes within the
6:41
organizations.
6:42
It also means having an understanding of who is reviewing these systems, who is
6:47
getting impacted by these systems.
6:49
So in the second phase of Credo AI product, we basically do interrogation of
6:54
your tech stack as well as your process stack against those requirements.
6:59
And then lastly, as you can imagine, something that I'm deeply passionate about
7:04
is
7:05
How do you actually get those diverse voices actually engaging on understanding
7:10
is this AI working in our behalf?
7:12
So in the third step of Credo AI product, there is a massive translation engine,
7:18
which is basically taking all the outputs of this testing and interrogation against
7:23
a set of guardrails and translating it so that Kevin, the stakeholders were on both
7:28
sides, whether they are coming from technical backgrounds or they're coming
7:32
from business backgrounds.
7:33
can really understand what's happening within those systems.
7:37
And so the outputs could be things like audit reports, impact assessment reports,
7:42
model cards, governance dashboards, risk dashboards.
7:46
But we are really making a very, I would say, concerted and intentional effort so
7:51
that these stakeholders actually can have conversations at the right level, even
7:56
though they might not have the same level of expertise in AI, in risk, in policy, in
8:01
compliance.
8:02
but they are really coming together to solve a collective problem.
8:05
Is this AI working for the end users and making sure that they have the confidence
8:10
in deployment of these AI systems?
8:13
So what's really different in doing this in the AI context from what many
8:18
organizations have already been doing for data governance, compliance with GDPR and
8:23
so forth, or even just overall good IT governance?
8:27
Is it those particular regulatory obligations or is there something
8:32
fundamentally new in the AI context?
8:35
You know, that's a great question and we get asked this question quite a lot.
8:40
I think the way I would like to answer it is having foundation of technology
8:45
governance, corporate governance, data governance helps because that just
8:49
indicates that your organization has already thought about what does oversight
8:54
for those different functions look like.
8:56
And you have individuals who are incentivized to do that oversight and
9:01
governance in those different areas.
9:03
The reason AI governance is very different, and I would say that certainly
9:08
does build on that foundation, but needs a new kind of thinking, is primarily because
9:12
of three reasons.
9:15
The first reason is it really needs a very multi -stakeholder engagement, because
9:21
there are so many nuances to how these socio -technical systems are going to
9:26
impact users and businesses.
9:29
that it is not just one person who will hold the keys to understanding how this
9:34
very complicated technology works.
9:38
And so I would say the first big change that we are seeing is you really do need
9:42
perspective of data science, machine learning, risk, compliance, audit, you
9:47
name it, into this new governance methodology, which can be built on the
9:52
foundation of what you have.
9:54
The second thing that is critical, Kevin, is...
9:58
Unlike the previous technologies, which, you know, if you think about privacy and
10:03
cloud, yes, there were evolutions in their journey, but they weren't changing.
10:09
They weren't as dynamic.
10:11
With artificial intelligence, especially with these large frontier models and
10:15
foundation models, the dual nature of those actually is something that we should
10:21
be very concerned about, but also...
10:23
be aware that there's a lot of more unknown unknowns.
10:26
As a result of this, this technology is moving so fast, it's getting new
10:31
capabilities exponentially that the governance methodologies and frameworks
10:36
that you might have needs to be constantly evaluated and adapted to make sure that
10:41
they are making the right guardrails happen for AI.
10:45
So that's the second reason I think AI governance is different.
10:48
And I would say that the third reason AI governance is different,
10:52
is absolutely because of the kind of guardrails.
10:56
And as you can imagine, those guardrails first and foremost are still coming to
11:00
fruition, whether it is through regulations or standards.
11:04
But the second thing is because of the massive increase in capabilities, new kind
11:12
of guardrails really need to start surfacing very quickly.
11:15
And this is something we call as emergent properties of these models.
11:19
there are things that we don't even know about these systems that are going to
11:22
change very drastically.
11:23
And so how do you actually have a governance practice that can keep pace,
11:28
not only with the regulatory changes, but also capability changes is really
11:33
critical.
11:34
So I think just to summarize, Kevin, certainly if you have privacy, data
11:39
governance, cloud, you can build on those governance capabilities, but AI governance
11:45
does need.
11:46
a new accountability leader.
11:48
It does need new kind of thinking.
11:50
It does need this multi -stakeholder perspective to really make sure that you
11:54
can keep pace with this evolution of AI that's happening.
11:59
You've used the term sociotechnical a number of times, and I had another
12:03
conversation with Elhan Tabassi of NIST, and as you know, they very much have that
12:08
orientation as well in their AI risk management framework.
12:11
What does sociotechnical mean to you, and what's so important about that framing?
12:16
Yeah, you know, the way I define socio -technical is when you have a
12:19
technological change or capability set that are brought to the market in a way
12:26
that literally is going to change the way society, this world, our democracy
12:32
operates.
12:33
And the impact of that is really far reaching than just a set of small
12:38
stakeholders.
12:39
So for us, socio -technical really requires two things.
12:43
And we think about it as inputs and outputs.
12:45
But in terms of outputs, the impact of this technology, whether it is ML models
12:50
being used in hiring or whether they are used in climate change, it is really far
12:55
reaching.
12:56
And it really has a very important consequence that we need to, as
13:02
stakeholders think about, which goes beyond just getting the next customer.
13:06
So I would say the output part is really important to think about.
13:10
And then from the input part, this goes back to...
13:13
The development of this technology is dependent on so many different components.
13:18
As an example, where are you getting these data sets for training these systems from?
13:23
Are those data sets demographically diverse?
13:26
And as you can imagine, most of the data that is feeding even these large language
13:30
models is coming from internet.
13:33
And, you know, does it represent all parts of different demographics that exist on
13:38
this earth?
13:39
No.
13:40
So how do we actually make a very thoughtful decision in what these inputs
13:43
to these systems are?
13:44
So for us, social technical really is this input output understanding of where and
13:52
how these systems are getting built and who's building them, what is their
13:55
accountability, but also thinking about the very pervasive nature of impact that
14:00
these systems are gonna have on this world and on the businesses and on the impacted
14:05
users, which with the previous technological changes,
14:09
we didn't have to pay as much attention to.
14:13
Given that incredible importance though, is a compliance orientation enough?
14:19
Shouldn't companies really be thinking about ethics and values and how to raise
14:25
the level of what they're doing?
14:27
Isn't there a danger that this becomes just checking a box?
14:32
Absolutely.
14:32
And I think this is where, you know, we started this conversation with who's going
14:37
to win in this age of AI.
14:39
And if you are going to use AI governance and responsible AI as a checkbox
14:44
mechanism, you can already count yourself out.
14:47
I think organizations that are really going to start emerging as winners are
14:51
really taking a very comprehensive approach.
14:54
They are thinking through their not only structures, whether it is...
15:00
Who is actually on the table?
15:01
Who's looking and reviewing these systems?
15:03
But all the way from, are you sourcing the data from the right compliant individuals?
15:08
Are you making sure that the testing is done in a comprehensive way?
15:11
Are you making sure once these systems are put in production, that the reviews and
15:15
incident reporting is happening in a comprehensive way?
15:18
So I would say that if you're going to use it just to get to, MI -EUA Act compliance
15:24
checkbox, MI New York City law number 144 checkbox.
15:27
I think you're really missing out an opportunity to build that trust, which
15:31
actually can really power your business growth in ways that you can't even
15:36
imagine.
15:37
And we are already seeing classically, like in the companies that we work with at
15:42
Credo AI, that organizations who are taking that comprehensive approach, which
15:47
is not just a rubber stamping, are already establishing, I would say, core
15:52
competencies within their organizations that's going to really set them up for
15:56
success.
15:57
to really be able to innovate and iterate fast because that's what this new age of
16:02
AI will need so that you can respond to the technological advances very quickly
16:06
within the organization.
16:09
How does one measure that or evaluate that?
16:11
If you're looking at a company or someone else is looking at it, what should we look
16:16
to to evaluate if they've got that real competency?
16:20
Yeah, great question, Kevin.
16:21
And this is where at Credo AI, we've come up with a readiness index.
16:26
And that readiness index is really focused on a couple of core components, which not
16:31
only looks at the AI maturity of the organization, but also the organizational
16:35
commitment, leadership commitment, as well as accountability stakeholders within the
16:41
organization.
16:42
And the readiness index basically...
16:45
categorizes businesses where we've surveyed over thousands of businesses and
16:50
looked at some of these core vectors of what defines sort of like a solid AI
16:57
governance structured company.
16:59
I would say in that readiness, what we are finding is if you're at these early stages
17:05
where you are still exploring, the signals that we get from still exploring are.
17:12
those organizations are trying to baseline themselves with the rest of the world.
17:16
They are saying, they're looking and reading what the responsible AI principles
17:20
that might be emerging in the world are.
17:22
They are trying to do maybe some light form of audits.
17:25
They're trying to bring in some outside consultants and they're trying to bring in
17:30
some outside competencies to really do assessments of their systems.
17:34
Because there are these very early stages of exploration.
17:36
They don't have a stakeholder who's accountable for it.
17:39
And they're just trying to figure out,
17:41
who they want to be in this age of AI.
17:44
Now, when you go beyond this exploration stage, you start really aligning.
17:49
And this is when you're like, you know what?
17:52
AI governance is going to be critical for us to become leaders in this age of AI.
17:56
You at least had that understanding.
17:59
Now you start forming or formulating a set of responsible AI guidelines for yourself.
18:04
You actually maybe have created a committee and that committee is starting
18:08
to discuss.
18:09
What does good look like based on the baseline that you've done in the earlier
18:12
stages?
18:13
And this is, I would say the alignment stage is a very important stage because
18:18
that is where really you start to see signals from an organization.
18:22
Are they going to take actions to further their people processes tooling so that
18:29
they can start becoming the leader in AI governance?
18:32
And then as you can imagine, there are two other stages, but finally, when you get
18:36
to...
18:37
becoming an optimized stage or a leadership stage in AI governance, you
18:42
have actually an accountable stakeholder.
18:45
It could be a chief AI officer.
18:47
You actually have a budget, which is as high as your cybersecurity and privacy
18:51
budget, which is now going to be dedicated to AI governance.
18:54
You actually have a team of individuals who are responsible for that oversight and
18:59
accountability across your entire business.
19:02
You have not only your ops tooling,
19:05
in place for MLOps, LLM Ops.
19:08
But now you also have this pain of governance where many of the companies are
19:12
using Credo AI, but to really have that holistic tool chain that is providing that
19:17
continuous oversight.
19:19
And this is the time you actually start externally sharing assets with your
19:23
customers.
19:24
As an example, if you are a financial services organization selling a fraud
19:28
model to a bank, now along with that fraud model, you will also maybe share your
19:34
governance report.
19:35
and the governance report would have all the transparency around how did this fraud
19:38
model get built, how was it tested, who reviewed it, what kind of industry
19:42
standards is it compliant to, et cetera.
19:45
And that's when you start seeing the benefits because now the customer, if I am
19:50
an enterprise making a choice between vendor A and vendor B, and vendor A
19:55
provides me AI products and vendor B provides me AI products, but also
20:01
transparency reports and also governance reports.
20:04
I as the user, I'm going to choose vendor B because they are giving me more
20:08
confidence that there are these transparent practices to really make sure
20:13
that that AI is going to work for me at the end of the day.
20:16
So Kevin, like, you know, in this AI readiness journey, what we are finding is
20:20
there are different signals.
20:22
Some early stage organizations are posturing, they're writing blog posts
20:26
about responsible AI, but actually they start to move the needle when they start
20:31
to provide assets.
20:33
that can help them land more customers, that can help faster procurement cycle,
20:37
that can help retain customers longer.
20:39
They can actually adopt AI and Genitive AI faster to increase employee productivity.
20:45
Their employees actually have a very good sense of what they can and cannot use the
20:49
Gen.
20:49
AI systems for.
20:50
And I think that clarity exists.
20:52
So that journey is where we are seeing organizations need to get to rather than
20:57
just using compliance as a mechanism to do responsible AI governance.
21:04
In financial reporting, there's a whole infrastructure of external audit and that
21:10
gets regulated and filed with the SEC and so forth and standardized in addition to
21:16
the kinds of things that companies are doing internally.
21:19
Do you think that we need to get to something similar to that in AI?
21:24
Absolutely, Kevin.
21:25
I think they're going to see again a journey to getting to that level.
21:29
And the journey, in my view, has already started.
21:32
So as an example, we are seeing a very interesting ecosystem of auditors emerge.
21:38
And these auditors, as you can imagine, are primarily third -party assessment
21:43
capability providers.
21:46
Because right now, there are no standards against which you can audit an AI system.
21:50
So audit is a wrong word to use right now.
21:53
But think about them as a third party assessors who have already been brought in
22:00
to come and assess some of these models and AI use cases.
22:03
Classic example of that is New York City law number 144 that came into effect last
22:09
July.
22:10
And it requires a third party review, whether that third party is within your
22:14
organization or external of your ML -based systems, which are used for hiring.
22:21
Now, an evolution of that that's happening in the ecosystem is emergence of standards
22:25
like ISO 400001.
22:27
Now, ISO 400001 is the first, I would say, certifiable AI risk management standard
22:35
that has emerged in the market.
22:37
And there needs to be a second standard, which is coming out later this year, which
22:40
is ISO 400006, which actually will provide auditors the capability to certify against
22:46
400001.
22:47
That hasn't happened yet.
22:49
But we are already seeing a lot of auditors showing up in this market who are
22:54
like, I can get you to ISO 42001 compliance, which I think is a good
22:59
starting point.
23:00
But as you can imagine, it is very broad.
23:02
It does not have understanding of use case context.
23:06
It is really around sort of like a carry forward of, do you as an organization have
23:11
a risk management framework for AI in place?
23:14
And that's what ISO 42001 does.
23:18
The evolution beyond that you're going to see of auditors is when actually more and
23:22
more of these standards show up, that AI audits and audit auditors will actually
23:27
start to emerge.
23:28
And this is where you as an auditor can say, not only am I certified by a
23:33
certification body, but now I can actually provide you an independent review and
23:37
assessment and audit, similar to SOC 2 audits that have existed in that category.
23:44
And then lastly, the vision that I have for this world,
23:47
is especially if you think about public and private side.
23:51
On the public side, my hope is that there is responsible AI disclosures that are
23:57
mandated by SEC to really make sure that an organization when they are doing their,
24:05
you know, sort of like earnings reporting, they're also required to disclose how they
24:11
are building these AI systems.
24:13
How are they making sure that the responsibility is central and paramount to
24:16
it?
24:17
And I think those responsible AI disclosures are not very far out.
24:20
We are already activating the ecosystem so that on the public side, not only are the
24:26
investors holding companies accountable, but also bodies like SEC are holding these
24:32
public entities accountable.
24:34
And I would say a flip side that we've already started to see on the private
24:37
sector side is we actually are founding members of something called as Responsible
24:42
Innovation Labs, where we came up with our own set of commitments.
24:46
that the private investors will hold early stage startups accountable to as they are
24:52
building or buying AI capabilities.
24:55
And these responsible innovation labs, really the initial focus was how do we
25:00
have that contract between investors and founders who are starting companies in AI
25:05
to be responsible by design?
25:07
And if we do our jobs right and these companies become super successful, then
25:12
when they go IPO on the flip side, they will have these
25:16
requirements around public disclosures for responsible use, which should not be a new
25:21
requirement because you've done that from the beginning.
25:23
So I know long answer to your question around audit ecosystem, but I am thrilled
25:29
about the formation of the audit ecosystem.
25:31
It's just going to be a while till it comes to fruition, but there are bits and
25:35
pieces that we at Credo AI are already activating or we are working with the
25:39
ecosystem to make happen.
25:43
What's different for AI governance with the rise of generative AI?
25:47
Yeah, a lot.
25:50
So a couple of things again, the way I would like to tackle this is in terms of
25:57
companies that care about AI governance as a first vector.
26:01
What has been really interesting for us is when I started the company four and a half
26:05
years ago, we were focused on predictive ML governance and it was really primarily
26:10
focused on regulated sector.
26:13
12 months ago, everything changed.
26:16
And the reason everything changed is generative AI became truly a
26:20
democratization of artificial intelligence where anyone and everyone in any industry
26:26
could have access to these powerful systems and use it to their benefit and
26:31
mostly for their own productivity.
26:33
So I would say that what has shifted in AI governance because of Gen .AI first is we
26:39
are seeing organizations who are not even in regulated sector.
26:44
actually caring about AI governance because they recognize that with or
26:49
without them, their employees are going to be using generative AI systems and their
26:53
risk surface area has drastically exponentially increased.
26:57
And how do they not only manage risk, but is there a way that they can actually use
27:02
it to their advantage and build that trust that we've been talking about?
27:05
So that's one thing that has shifted.
27:07
The second thing that has shifted with AI governance is an acceleration in
27:13
what I like to call adaptive policy making.
27:16
We have never seen our government or global governments or global legislatures
27:21
move as fast as in the past 12 months because there truly is a requirement for
27:29
how do we not only rein in this technology but make sure that this technology can be
27:34
adopted very fast by businesses with the right intentional guardrails.
27:38
And so I would say that the second big movement that we have seen is,
27:42
in this rise of, let's make sure that there are some guardrails in place with an
27:47
understanding that these guardrails will change.
27:49
And so I've been a pretty vocal advocate for this adaptive policymaking where you
27:55
don't have to get everything right from day one.
27:57
You really have to work with some of the core AI experts like Credo AI and the
28:02
hyperscalers to really understand what does at least the beginning stages of good
28:06
policy look like.
28:07
And then we need to have...
28:09
sort of like new mechanisms in our government to be able to iterate very fast
28:13
on it.
28:13
And, you know, EUAI Act does, you know, I think they've done a fantastic job in
28:19
really thinking through what that might look like.
28:21
Obviously it has lots of gaps and challenges, but rather than focusing on
28:25
gaps and challenges, I would say that at least it is the first, you know, global
28:29
framework that organizations have started to pay attention to.
28:32
And then the third big thing that has shifted, I would say is,
28:37
It has put enterprises and stakeholders on alert that now they can't just rest on a
28:45
technology that's not going to change for another decade.
28:47
This technology literally is changing on monthly and quarterly basis.
28:51
So how do you actually embed AI literacy within your organization to be able to
28:57
respond to the speed of capabilities that are emerging in generative AI?
29:03
And I think, Kevin, that is really critical.
29:06
because what we are going to see is an evolution on what kind of skills and jobs
29:12
emerge in this ecosystem because individuals who are able to learn quickly,
29:17
try out new technologies, are creative problem solvers, are not just like holding
29:22
on to their expertise forever, but deploying their expertise in AI are the
29:26
ones who are going to actually start showing up and their jobs are not going to
29:29
be at risk.
29:30
So I think we are seeing a very new model of AI literacy emerge within
29:34
organizations.
29:35
to really respond to this very fast changing AI technology.
29:40
Just following up on one thing you said, you seem pretty optimistic that regulation
29:46
will actually promote better development.
29:50
Are you worried at all about some of the concerns about chilling innovation or
29:54
especially now in the US where we have so many states passing legislation that
29:59
there's going to be this thicket of laws that actually make it difficult,
30:03
especially for the smaller companies to comply?
30:06
You know, Kevin, the way I think about regulation, it's not that like I wake up
30:10
in the morning and say, my God, I'm all for regulation.
30:13
No, I think what regulation is really doing is bringing the focus on how
30:19
important it is for us to put guardrails around this technology.
30:23
So I see regulation as a way to wake up enterprises and stakeholders that you need
30:30
to take action because unfortunately as humans,
30:34
I always say this, we are motivated by two things.
30:36
We are motivated either by ambition or fear, right?
30:39
And the ambition here is, my God, Genitivia is going to bring in so much
30:44
prosperity for humans and for our world, and it's gonna make sure that this planet
30:50
continues to be a thriving planet.
30:52
And I truly, 100 % believe that.
30:54
However, on the flip side, there are individuals who take time to respond to
31:00
technological changes.
31:01
And the only way to get their attention is through fear.
31:04
And I think that's what unfortunately regulatory movements I use as a fear that
31:11
you need to be compliant otherwise there are these fines.
31:14
I think what I'm hoping to do is change that entire narrative.
31:18
The narrative really is think about who you are going to be as an individual in
31:23
this new age of AI, what characteristics and skills you need as an individual.
31:27
But secondly, what characteristics you need as an enterprise to actually be a
31:31
thriving business in this age of AI.
31:34
And for both of those, I would say there are certain core characteristics on
31:38
adaptability, being able to learn and being able to build trust to consistently
31:42
deliver are basically table stakes.
31:46
And I think for me, when I think about regulation, I think about as a wake up
31:52
call, not as a way that...
31:55
is the only way that we're going to adopt AI responsibly.
31:58
I don't think that's the only way.
32:01
So to answer your question, am I worried about regulation actually being a barrier
32:06
to innovation?
32:07
Yes, if it is not done thoughtfully.
32:10
But the good news right now in artificial intelligence is policymakers and
32:14
regulators have been very heartened with the private -public partnership that has
32:19
emerged where...
32:20
policymakers are recognizing they don't have the expertise.
32:23
They will need to lean on AI experts like us.
32:26
At the same time, we also understand we don't understand the nuances of actually
32:31
creating a good policy.
32:33
And this is where we need to lean on the policymakers and many a times sort of like
32:38
go against the bureaucracy that has existed and many a times go against the
32:43
power that has existed to make sure that this technology serves humanity.
32:48
So long story short,
32:49
I think regulation is a way to rally the troops around a common mission.
32:57
Now, can this be a barrier for smaller companies?
32:59
That's why Credo AI exists.
33:01
We want to make sure that not only are we holding big tech accountable through Credo
33:08
AI products, but we are reducing the cost of our technology and our product so that
33:15
all the emerging innovators, wherever they are on the globe,
33:19
can actually use governance as a competitive advantage and compete with
33:24
some of the big tech at global scale.
33:27
And that is actually a very important mission for us here at Credo AI, where we
33:32
don't want this to be a compliance burden, but as governance has an opportunity for
33:36
them to build their technical as well as capability modes.
33:41
Okay, last question.
33:43
What's the number one thing that someone should look to?
33:46
Biggest development or biggest area to focus on to understand the development of
33:51
responsible AI or AI governance in the next few years?
33:55
my God, that is a hard question, Kevin, because I would say certainly tracking
34:00
some of the core frontier model developments that are coming out of some
34:06
of the large companies is really, I would say, interesting to watch and track
34:13
because that is going to impact how we live, work, play.
34:17
I am a big fan of also really engaging with academia.
34:24
whether it is Stanford's Human Centered AI Institute or whether it is CMU's AI
34:30
Institute, I think they're really doing some phenomenal work, especially on the
34:35
accountability and transparency to really start thinking about what are the ethical
34:38
challenges we are going to run into and who needs to be on the table.
34:42
And then lastly, I would say from the capability perspective, I think we need to
34:51
be a little bit more
34:53
open to trying out all the new tools.
34:56
And this is something that I coach pretty much everybody, whether you're a nine year
35:00
old kid, like my daughter, who is like very well versed in some of these new
35:05
emerging tools, or whether you are a high school student or whether you are a grad
35:09
or you are an early in your employment is don't be afraid of artificial
35:14
intelligence.
35:14
This is the moment in time to try out all the tools that are showing up in your way,
35:19
because I'm a firm believer that
35:21
Again, we are going to emerge as winners if we know how to use these tools to our
35:27
benefit much more efficiently.
35:29
And so I think getting AI literate becomes really critical as we move forward.
35:35
Terrific, Navridha, thanks so much.
35:37
Really wonderful speaking with you.
35:39
Yeah, thank you so much for having me, Kevin.