Episode Transcript
Transcripts are displayed as originally observed. Some content, including advertisements may have changed.
Use Ctrl + F to search
0:07
Welcome everyone to the AI
0:09
in Business podcast. I'm Matthew
0:12
D'Amelio, Senior Editor here at
0:14
Emerge Technology Research. Today's
0:16
guest is Steve Astorino, Director
0:18
of the Canada Lab and
0:20
VP of Development and Data
0:22
in AI at IBM. Steve
0:25
joins us on today's program to
0:27
talk about the biggest challenges for
0:29
enterprise leaders when it comes to
0:32
driving the infrastructure innovations necessary to
0:34
leverage new emerging AI use cases,
0:36
especially in new data-hungry, generative AI
0:39
tools. Without further ado, here's our
0:41
conversation. Steve,
0:49
thanks so much for being with us on the show today.
0:52
Thank you for having me. Especially, you
0:54
know, talking to someone who has
0:56
like a divisional control and we're
0:58
talking to someone from IBM, Canada.
1:01
I think this is a
1:03
particularly great perspective to have
1:05
on infrastructure challenges, especially
1:07
to know where they are in
1:10
the pipeline and what those specific
1:12
challenges are. So thank you so
1:14
much for being on the show.
1:16
We wanted to know what you
1:19
see as the biggest infrastructure challenges
1:21
when it comes to driving AI
1:23
capabilities, especially from a divisional
1:25
enterprise standpoint. Yeah,
1:27
sounds good. So we have, when
1:29
we look at where AI is
1:31
today, especially with, you know, foundation models,
1:34
large language models, the biggest
1:36
challenge that I see, and it's
1:38
kind of across the world, is
1:40
the availability of the GPUs that
1:42
are required, both for,
1:44
you know, faster training of these
1:46
models, as well as the
1:49
ability to actually execute and infer these
1:51
models when we're asking a question. I
1:53
think for most organizations, they want to
1:55
get their hands on it. That
1:57
is one of the biggest challenges that I see.
2:00
in the market in terms of
2:02
how do these clients get started and get
2:04
their hands on this infrastructure. There's
2:06
lots of different things that are trying to be done to
2:09
accelerate that, but the reality is we're probably going to be
2:11
in this situation for a little while. But
2:13
it's probably a good thing, I would say, as well,
2:16
because this technology has been moving
2:18
way too fast. I think everyone
2:21
can acknowledge that. And there are risks associated with
2:23
it, so we want to be careful. So
2:26
I would say GPUs, the availability of
2:28
that infrastructure needed for large language models,
2:30
is the number one, I would say,
2:32
challenge that we have. But
2:34
there are other things that are super critical,
2:37
in my opinion, to make sure that companies
2:40
and users of this technology are getting
2:43
into the right use of it. Having
2:45
the right tools is super important. What
2:48
I've seen is a lot of
2:50
businesses are overwhelmed, they're underprepared, they
2:52
want to get into this
2:55
AI space, but they don't know
2:57
how to do it exactly, they don't know how to
2:59
profit, but they're all having board-level
3:01
meetings that need to happen. The
3:03
business leaders are expecting this to happen.
3:05
So choosing the right
3:08
tools, I think, is also critical
3:10
in the success. And we've seen
3:12
a lot of mishaps in the market already, just
3:14
in the last six or 12 months,
3:16
I would say. Absolutely. It's
3:18
all moving so fast. It really feels like in
3:20
the last year, life has changed entirely in a
3:22
very much a paradigm shift. And
3:25
AI was here and in people's faces,
3:27
and a topic of conversation well before
3:29
that. But I think generative
3:32
AI on a cultural level is just way
3:34
bigger than even original AI for reasons we
3:36
can get to in another podcast. I really
3:38
want to pull apart what you were saying in
3:40
GPUs there for a moment. We did a big series
3:42
not too long ago on GPUs.
3:45
The audience is free to go back to
3:47
that episode, check it out for at least
3:49
the dynamic so we can get a little
3:51
deeper baseball here. Basically, as you describe it,
3:54
it's a good thing that on an entire
3:56
infrastructure level, there haven't been and there don't
3:58
seem to be what are called... step-level
4:00
improvements to the entire system. The
4:03
step-level improvement is basically like the
4:05
wheel, inventing the wheel. It makes a
4:08
lot of things a lot easier, you
4:10
know, or the cotton gin, if you
4:12
want a slightly better, slightly more recent
4:14
example. But it doesn't seem like there's
4:16
gonna be step improvements
4:18
for the recent future. Audience members who
4:20
tune into that series will remember our
4:22
guests talking about, well, when we do
4:24
have step-level improvements, we'll be able to
4:26
take data from every single cell phone
4:28
and do incredible things and yeah, taking
4:31
data from every single cell phone. Now,
4:33
I know we're already doing that on
4:35
some level, but even taking visual data,
4:37
that is a giant leap, not just
4:39
in the infrastructure and in the step-level
4:41
improvements, but also in the customers being
4:43
comfortable with their privacy. That, to your
4:45
credit and to your last answer, brings
4:47
up some ethical questions. So I'm
4:50
just wondering there, you know, also I'm curious
4:52
what makes you say, not that our big,
4:54
you know, message from that last series was
4:56
that there's all these step improvements coming. What
4:58
makes you say it's slowing down right
5:00
now? I get that sense too, but I'm
5:02
curious as to what makes you say. Do
5:05
you think it's just the cultural backlash of
5:07
AI? We all, everybody kind of agreeing, we
5:09
all need to slow down right now? Or
5:11
is there something a bit more in the
5:13
systems behind the scenes? Yeah, I think
5:15
there's a couple of things. I mean, so the
5:17
availability of the infrastructure is helping us slow it
5:20
down a bit or the inavailability of it. Then
5:22
I think, you know, it's almost
5:25
like in January everybody woke up and said,
5:27
oh AI is here. Well, to your
5:29
point, like you said, AI was here before. It's
5:31
just now we really went to
5:33
the next level to show its capability
5:36
to everyone rather than
5:38
just maybe data scientists, right? So
5:40
that has kind of, I
5:43
think everybody going through this year has
5:45
learned a lot more about it. That
5:47
has raised a lot of questions about
5:49
the reliability of it, about the security
5:51
of it, about all the wrong things
5:53
that can happen. There's a lot of
5:55
lawsuits happening right now and for
5:58
probably for the right reasons, right? We've kind
6:00
of dove in really, really fast. And
6:02
I think everyone is now to the point, okay, I
6:04
understand it a bit better. I know I need to
6:06
continue to get into it, but now I'm going to
6:09
step it back a little because I want
6:11
to make sure I do the right thing for my
6:13
company. Right. So I think that's the biggest thing. So
6:16
everyone is taking a little bit of a step back and
6:18
be more mature about the technology, if I can use those
6:20
words. Yes. Yes. Yeah. You can, you
6:22
can, you can bring maturity into the mix. No
6:24
one's no one's going to know. And from our
6:26
legal department is going to write in to me
6:28
with, with some angry messages. And I'll take that
6:30
hate mail. I'll take that hate mail. You also
6:33
mentioned the right tools. I love this conversation and
6:35
I want to put a finer point on this.
6:37
I'm seeing, and our audience
6:39
can go back in the last few episodes
6:41
that it's been with the, especially with the
6:44
explosion of generative AI beforehand,
6:47
the C-suite management, not everybody, of course,
6:49
but let's, let's call worst case offenders
6:51
or worst case scenarios, it was like,
6:53
you want to do what with AI?
6:56
You know, they hadn't heard of it yet. You know, what
6:58
is this machine learning? And then they have to
7:00
go find it with the explosion of generative
7:02
AI. It seems like, especially with, with the
7:04
media coverage and the notice, now it's like,
7:07
oh, you have C-suites, their kids are using
7:09
it, and now they're coming back down to
7:11
the data science and saying, why can't we
7:13
use this gen AI stuff to do whatever
7:15
we want? And what's being
7:17
lost in that mix is that the
7:19
first generation capabilities of AI, just the
7:21
plain old machine learning, the predictive analytics,
7:24
they're not even being leveraged well enough
7:26
for specific workflows where they can really
7:28
help in enterprises. I'm wondering
7:30
if that's what you mean there, but
7:32
by the right tools, maybe something a bit
7:35
deeper, but that's, that's what, what I'm hearing
7:37
for the conversations we have on the show
7:39
in terms of, you know, enterprises barking up
7:41
the wrong tree, not really thinking about what's
7:43
the right hammer to nail this nail. Yeah,
7:46
I think so. Yes. And what I
7:48
would say is what we know about
7:50
this technology, there's still a lot of
7:52
the answers are inaccurate. The, you know,
7:54
it's not scalable, it's not adaptable. There's
7:56
a lot of risk with it. Right.
8:00
There's a hallucination where you get the wrong
8:02
answer altogether. So I think
8:04
all of that has been helping
8:06
slow us down. But yeah, so I
8:08
think line of business, you
8:10
know, C-suite, they've been looking at this,
8:13
give me this fast, but I
8:15
think there's a learning curve that's been happening. And
8:18
I think it's the right thing for all of
8:20
us to understand what's there, what it
8:22
can do, what are the risks associated with
8:24
it. And I think we're
8:26
in the, probably I would say in the
8:28
right place, we should leverage this technology because
8:31
I believe it's going to transform kind
8:33
of everything that we do on
8:35
our daily lives in the good. And
8:38
we have to be careful for the bad because there
8:40
are a lot of risks with it. The
8:42
other thing I would say is the skills are
8:44
not there. I, you know, I applaud the market.
8:46
Everyone is becoming more and more knowledgeable
8:49
of the technology, trying to understand it, what it can
8:51
do, how it works. I think that's all goodness. And
8:53
by right now, there is a lack of skill and
8:56
we need to continue to build it. And
8:59
the other thing I would say, look,
9:01
at the end of the day, this is
9:03
all about data. And every conversation I have
9:05
with clients, no matter where we
9:07
start, we always end up with data. And
9:09
there's a lot of things associated with that
9:11
around data security, privacy, what's got done with
9:13
that data, the type of answer, who owns
9:15
it and all of that stuff. It's, there's
9:17
a lot of complexity around this,
9:20
where we need very strong regulations. But, you
9:22
know, there are a lot of things that
9:24
kind of get into this mix on how
9:26
we do this successfully. Very, very
9:28
interesting stuff. And I think you're bringing up a
9:30
lot of, a lot of really incredible points for
9:32
the solutions portion of our show. I'd like to
9:35
break these down because I think you're bringing up
9:37
very distinct problems. And even a few
9:39
of them, I just don't think that you can,
9:41
you can talk about them. Hey, how are, how
9:43
is everybody using solutions? Let me go to GPUs
9:45
for a second. It almost sounds like you actually
9:47
kind of think this, you know, this problem is
9:50
a good thing that it is slowing down. It
9:52
might not need a solution. Maybe the question then
9:54
for solutions is how should we go about talking
9:57
about being more
9:59
proactive? about when
10:01
that next step level improvement
10:03
comes, implementing it safely and
10:05
ethically? Yeah, I'm going to
10:08
tie it back to the comments I made before about
10:10
the right tools. So right
10:13
now, there's a lot of every
10:15
company probably in the world is coming up, oh, I
10:17
have an AI tool, I
10:19
have a Gen AI tool. But
10:21
the reality is that at
10:24
the enterprise level, our clients
10:26
expect the ability to, they either
10:28
want to build their own models, or they want to
10:31
fine tune models that exist out there, or they want
10:33
to just do prompt engineering or prompt tuning. But
10:35
that's just a kind of a small aspect of
10:37
it, right? Once you get past that, then
10:40
you really need to be able to want
10:42
foster innovation, but also in a secure way.
10:44
I think that's the biggest challenge right now. And
10:48
the problem is the guardrails don't exist,
10:50
both from legislation and regulations, as well
10:53
as kind of the tools themselves. So
10:55
it's very easy to go off a cliff
10:57
and get in trouble. And we're seeing that
11:00
we're seeing that in the market. So
11:02
to me, tying it
11:04
back to having the right tool is the number
11:06
one thing. I can share
11:09
what we're doing. And we're very proud of
11:11
what we're doing at IBM in terms of
11:13
how we have been taking our technology to market.
11:16
But it's really the ability to manage and
11:18
to end the entire lifecycle in a secure
11:21
risk-free and compliant way. I
11:24
said a lot in that one sentence, but at
11:26
the end of the day, we want to make sure our clients are
11:29
able to get the
11:31
innovation out, disrupt their own business in
11:33
a positive way, and then do it
11:35
in a safe way. Yeah, that
11:37
makes a lot of sense. And I think that also
11:40
narrowed down my question on the right
11:42
tools, like how do we think of
11:44
those solutions? I think we got a
11:46
good there. I want to talk about
11:48
what you just brought up in the
11:50
skills gap, especially what you find are
11:52
good approaches, at least maybe for hiring
11:54
or at least training in-house to develop
11:56
this know-how for enterprises. Yeah, we've
11:59
been working through this. challenge also with
12:01
data science and machine learning, right?
12:03
You go back five, six years,
12:05
most companies did not even have
12:07
data scientists. So there's
12:09
multiple approaches to this. We with an IBM
12:11
have done a ton of
12:13
an enablement and education and courses and
12:16
leverage our research team. Our research team has been
12:18
at the forefront of this technology
12:20
for many years. We didn't just wake up in
12:23
April or May and we said, oh, we
12:25
have Gen AI too. We've been working at
12:27
this for five, six years. And we
12:30
knew exactly what the capability was, also
12:32
what the dangers were. And that's important
12:34
to us. But on the skill side,
12:37
we've been leveraging that team to enable
12:39
and to train and we have dedicated
12:41
courses specifically for it. We're
12:43
doing a lot of hands-on challenges within
12:45
the company. Our brand is What's the
12:48
Next, as you can see on my
12:50
shirt here. But we are What's
12:53
the Next challenges where we're getting the
12:55
entire company to use the technology, to learn
12:57
about the technology, to see how we can innovate
12:59
with the technology. And this is a safe zone
13:01
where we can do that. And we can see
13:03
what the art of the possible that then we
13:05
can pass on to our clients. We've
13:07
been working with academia to be able to
13:10
strengthen these programs. And we've gone from machine
13:12
learning now to Gen AI as well. The
13:15
reality with academia is that they're a
13:17
little slower than probably everyone else in
13:19
the market. And it takes time to
13:22
kind of modify the courses that they have
13:24
and the programs. But we're working with them
13:26
to accelerate as much as we can. And
13:28
there are also a lot of other external
13:31
companies that are doing this education. So I
13:33
think it's an industry-wide task
13:36
that we all have to improve the skills,
13:38
increase the skills in the market. And I
13:40
think we all got to collaborate on that.
13:43
It doesn't matter which company is doing what,
13:45
but I think it's important for everybody. Absolutely.
13:47
Absolutely. I want to give some time to
13:49
the problem of data that you had brought
13:51
up before as well. But we could do
13:53
a whole episode on data, data
13:55
problems in the enterprise. We could do probably three
13:57
or four episodes on data problems in the enterprise.
14:00
enterprise. Let me know. Do you want to
14:02
know what is data mesh versus data fabric?
14:04
Hit me with that answer, but where I
14:06
was going to go with this is what
14:08
do you find is the biggest problem, biggest
14:10
conversation with data that you're having with enterprise
14:13
leaders? If you could narrow it down to
14:15
one. Data mesh in that difference, totally
14:17
on the table if you feel that's the case. I'll
14:20
talk about the biggest challenges really. There's
14:22
actually two. There are two that I
14:24
think are super critical. One is the
14:26
data needs to be clean, especially for
14:28
foundation models just because whatever answer you're
14:31
going to give is going to be
14:33
based on that data. Maybe it's
14:35
more than two, maybe it's three, but the ability
14:37
to access the data is still
14:39
a challenge in probably I
14:42
would say most of
14:44
not all organizations. The larger the
14:46
organization, the bigger the challenge is.
14:49
The way AI works is the more
14:51
data that is clean and accurate you
14:53
have, the better your model and
14:56
your answers that you get from those models
14:58
will be. Then the
15:00
other one is privacy around our data. It's
15:02
twofold. It's who can access the data, but
15:05
also if you look at the models, we
15:08
can put governance and controls around the data
15:10
on who can access it, but once you train
15:13
a model, then that model is a generic model
15:15
that users can access.
15:17
Let's say that you're a more senior person
15:20
in an organization and I'm a less senior
15:22
person and I want to be able to
15:24
ask a question, we're going to
15:26
get the same answer. The model doesn't actually
15:28
know that yet. When you think about where
15:30
we need to go in the future around
15:33
this type of capability, we should be able
15:35
to provide access control or role control granularity
15:37
so that I can get potentially a different
15:39
answer that you can based on what privileges
15:42
I should have. When we talk
15:44
about the data itself and privacy and
15:47
all of that, it's super critical that
15:49
we have full governance around that, not only
15:51
on the data itself, but also on the
15:53
models. When the models
15:56
get called, are we able to provide
15:58
a safe and right answer? are based on
16:00
not only the data itself, but also the
16:02
privileges of that data and that model. You're
16:05
making a lot of sense to me, and I think you're
16:07
making a lot of sense to the audience. I think
16:10
these are new ways of looking at the problem.
16:12
And very excited in our next episode to really
16:14
dive into AI ethics with you. So we'll have
16:16
to have the audience tune back in for that.
16:18
Steve, thank you so much for joining us today.
16:21
It was my pleasure being here, thank you. Before
16:34
we wrap up today's show, and
16:37
I'll give a full disclosure
16:39
here, while we often record
16:42
two-part episodes separately and
16:44
over a fair amount of week's time,
16:46
I'm actually recording this outroduction after we've
16:48
recorded our second episode with Steve. And
16:51
I would like to plug that episode, even though I'm
16:53
not quite sure when it's going to publish, and
16:56
just encourage our audience to stay tuned,
16:58
especially given Steve's last answer in this
17:00
first episode, because we're going to
17:03
focus a lot more on AI ethics
17:05
in our second episode. And
17:08
in many ways, especially for those of
17:10
you who remember my conversation with Scott
17:12
Zoldy, Chief Analytics Officer at FICO
17:14
from earlier this year, you know, and by
17:16
this year I mean 2023, you
17:19
know that many AI
17:21
ethics problems as they occur
17:23
at the end of the
17:26
pipeline are PR disasters. That
17:28
impugn the morality and
17:31
the characters of the leadership at
17:33
the brand in question. And
17:36
often these products begin as
17:38
simple, non-biased, no
17:41
social issues involved, but simple
17:44
data governance problems, just at
17:46
least where the problems
17:48
are originating. And
17:51
Steve goes into a lot of really great
17:53
depth on how that works, how the mistakes
17:55
are made, how these mistakes
17:57
happen, often with all the good intent.
18:00
the world or at least intend to cut costs and
18:02
make processes faster. But
18:05
as Steve emphasizes both in this
18:07
episode and in the coming episode,
18:10
there's many ways that just cutting
18:12
corners is not worth it. And
18:15
that is sort of the underlying
18:17
conversation underneath both conversations about data
18:19
governance and conversations about AI ethics,
18:22
which very often for that Venn
18:24
diagram, they're almost perfect circles. I
18:26
find the more we talk about
18:29
these topics throughout the show. On
18:32
behalf of Daniel and the entire team here in
18:34
Emerge, thanks so much for joining us today and
18:36
we'll catch you next time on the AI in
18:39
business podcast.
Podchaser is the ultimate destination for podcast data, search, and discovery. Learn More