July 16, 2025 • 32 mins
Cloud‑native AI is rewriting the enterprise playbook; in this episode, guest host Robb Boyd convenes data strategist Ina Poecher and cloud architect Zaid Elkhateeb to unpack why 2025 belongs to firms that treat hyperscale cloud as their default R&D lab. Reporting from WWT’s AI Proving Ground, they reveal how on‑demand GPUs, modular services and policy‑as‑code governance collapse GenAI timelines from quarters to days —without riling security auditors. Tune in for a four‑step blueprint — clean data, outcome‑driven architecture, elastic cloud services and continuous guardrails — that turns experiments into production‑grade advantage.
More about this week's guests:
Ina Poecher is a Technical Solutions Architect at World Wide Technology (WWT), and collaborates with customers and internal teams to design and validate innovative technology solutions. Working within WWT’s Advanced Technology Center, she leverages extensive experience across IT infrastructure, cloud, networking, and automation to develop and test complex architectures that drive business outcomes and support strategic initiatives.
Zaid Elkhateeb is a Technical Solutions Architect at World Wide Technology (WWT), specializing in Google Cloud Platform. He brings deep expertise in cloud engineering, data architecture, and AI/ML services, joining WWT's GS&A team from another leading Google partner. Zaid helps enterprise clients design and implement scalable cloud solutions that power innovation and drive real business outcomes.
Learn more about WWT's AI Proving Ground.
The AI Proving Ground is a composable lab environment that features the latest high-performance infrastructure and reference architectures from the world's leading AI companies, such as NVIDIA, Cisco, Dell, F5, AMD, Intel and others.
Developed within our Advanced Technology Center (ATC), this one-of-a-kind lab environment empowers IT teams to evaluate and test AI infrastructure, software and solutions for efficacy, scalability and flexibility — all under one roof. The AI Proving Ground provides visibility into data flows across the entire development pipeline, enabling more informed decision-making while safeguarding production environments.
The AI Proving Ground Podcast leverages the deep AI technical and business expertise from within World Wide Technology's one-of-a-kind AI Proving Ground, which provides unrivaled access to the world's leading AI technologies. This unique lab environment accelerates your ability to learn about, test, train and implement AI solutions.
Learn more about WWT's AI Proving Ground.
The AI Proving Ground is a composable lab environment that features the latest high-performance infrastructure and reference architectures from the world's leading AI companies, such as NVIDIA, Cisco, Dell, F5, AMD, Intel and others.
Developed within our Advanced Technology Center (ATC), this one-of-a-kind lab environment empowers IT teams to evaluate and test AI infrastructure, software and solutions for efficacy, scalability and flexibility — all under one roof. The AI Proving Ground provides visibility into data flows across the entire development pipeline, enabling more informed decision-making while safeguarding production environments.
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Speaker 1 (00:00):
Every week, another organization discovers their
million-dollar AI initiative wasbuilt on a lie the lie that you
can skip the boring stuff andjump straight to the
transformative part, that dataquality is someone else's
problem or security can bebolted on later, or that cloud
infrastructure just somehowworks.
You know where this is going,probably.
(00:22):
By some estimates, more than80% of AI projects fail twice
the rate of failure forinformation technology projects
that don't involve AI.
Organizations are hemorrhagingmoney trying to build
breakthrough AI solutions, whiletheir cloud costs spiral out of
control, their data sits insilos across disconnected
systems and their securitypolicies haven't caught up to
(00:43):
their ambitions.
Its systems and their securitypolicies haven't caught up to
their ambitions.
Today, on the AI Proving Groundpodcast, I'm throwing it over to
my colleague, rob Boyd, whorecently talked with Ina Posher,
a data scientist who bridgesthe gap between what data
scientists want to build andwhat actually works in
production, and Zaid El-Khatib,a technical solutions architect
at WWT, who helps organizationsescape the AI corners they've
(01:04):
backed themselves into.
You're about to discover whymost AI failures happen before
the first model runs and theroadmap that separates AI
winners from casualties.
So, without further ado, let'sthrow it over to Rob.
Speaker 3 (01:38):
What have you learned about this gap between what
data scientists which, as Iunderstand, that's really your
focus here what data scientistswant to build versus what
actually works when it comes toreal production.
I would say the biggest gapthat exists is that there's so
many high-level ideas and somany things that I could solve
so many use cases that have highvalue and would have high
return, and then the data isjust not there or there's not
enough data there, and when thedata is not there as that
(01:59):
foundation, then there's notmuch you can do for your use
case and you're just stucktrying to create something out
of nothing.
So that's the biggest gap ondelivering a solution and really
making it worthwhile to investin.
Is your data there and is itready to go?
Speaker 2 (02:18):
And sometimes I guess that results as a mismatch
maybe between what theorganization is where they're
setting their expectationsversus you know what can
actually be done.
So it comes into.
You know, pick and we'll talkabout this further but picking
your use case especially earlyon and kind of being narrow
about that and understandingwhat all feeds into it, it
sounds like.
Speaker 3 (02:35):
Yeah, and prioritizing it based on the
data you have available, thedata that you can easily access
and the tools you have availableto solve it.
So it's not just what willprovide the most value to the
company, it's a return oninvestment.
So there has to be you have toinvest something and the return
comes from what's available andwe can't just invest, invest,
(02:57):
invest and not return anything.
Speaker 2 (02:59):
All right.
Well, obviously we're going totalk about that one a little bit
further in detail about how weactually pull that off.
But, zaid, let's get you inthere here.
You had told me you specializein Google Cloud and specifically
, you kind of help account teamsand customers.
However, it happens, they getthemselves back into a corner,
perhaps with cloud, with AI,with some combination of the two
(03:19):
, and you specialize in helpingbring them out.
Are you seeing any kind ofpatterns in terms of common
mistakes that are made?
Any patterns.
Speaker 4 (03:28):
So, yeah, definitely, I think what we see a lot of is
that you know, because,admittedly, right in our field
there are politics that areinvolved, and you know, I don't
mean government in the state ofthe world or anything like that,
but I mean in how you know howa solution is given to a
customer or a product is.
And so a lot of the times I'llsee a customer who, for one
(03:50):
reason or another, has beentalking a certain product or a
certain tool for a very longtime and they've purchased it
and they're trying to use it andget the most value out of it,
but they only really qualifiedthe product itself to see how
sound it is, you know, in abubble, and they don't often
look at what the outcome they'retrying to achieve is, and so
(04:13):
they'll purchase a product wherethey'll start using a
technology and it doesn'tnecessarily fit the outcome that
they are looking for, or it'snot the only thing that's going
to get them to the outcome thatthey're looking for.
And so sometimes I'll come in,because cloud is so wide and
there's so many technologies andthere's so many tools that are
hosted on the cloud cloud native, et cetera so I'll come in and
(04:34):
further qualify the tool thatthey're using and see if we can
mess that with another tool, orsee if they're not using it
correctly and see how we can getthem out of it.
Speaker 2 (04:43):
Yeah, and if you see that happening on a repetitive
basis, are there common elementsthat are just not taken into
consideration when approachingthese things?
Is that it, or does it comefrom many different directions?
Speaker 4 (04:55):
You know it comes from a lot of different
directions.
I brought up politics a littlebit because a lot of the times,
right, you have C-levelexecutives who are being told
really know, told really goodthings by people that they're
familiar with in a companyalready, and so, honestly, I see
a lot of I would say a lot ofissue come from those kinds of
conversations.
But then a lot of the timesit's that you know, perhaps they
(05:19):
are have a generative AIinitiative that is a little too
lofty and so they're trying tohit that you know initiative.
Or they're trying to hit thatis a little too lofty and so
they're trying to hit that youknow initiative, or they're
trying to hit that goal a littletoo quickly.
Or you know, maybe their youknow outcome and the tool that
they're using to get to thatoutcome is, is solid and it's
qualified really well, but youknow their data isn't quite
ready for it yet and so we haveto backtrack and so they didn't
(05:40):
look at the entire roadmap.
They, they didn't look at theentire roadmap, they were just
looking at the end of thedestination.
Speaker 2 (05:45):
Well, there's nothing more fun than finding those out
when you're mid-deployment.
Oh yeah, and then it comes up.
Those are not goodconversations, yeah.
Speaker 4 (06:01):
Then timelines, money everything just starts
expanding.
Yeah, that's when the customersstart bleeding money, For sure.
Speaker 2 (06:05):
Yeah, Well, and we're here to help hopefully prevent
the bleeding.
So there's two priorities thatwe're kind of blending together
into this conversation and wantto acknowledge that, because we
kind of bounce back and forthbetween some of them, but I
actually think this works welltogether.
In the research paper there'spriority two, which is about
data security and governance,and then there's priority three,
which is harnessing cloudinfrastructure for AI.
(06:26):
We've had elements of thesethings talked about on the show
and other ways before, but I'mcurious we talk about.
It's interesting to me always.
I always love seeing wheresecurity gets mentioned.
Is it before or after certainelements, because I feel like
it's always put in as kind of ohyeah, don't forget security.
Of course, now we have troublewith regulatory and compliance
and things with that, and I'mprobably just as guilty as
(06:48):
anyone else at doing that, aswe'll see here in a minute.
But, ina, I'm curious about whywe can't just skip ahead to the
AI part.
Why would it be prioritized?
Why would we need to thinkabout security?
Speaker 3 (07:02):
There are a lot of reasons.
Speaker 4 (07:05):
There's a lot of opportunity to give your data
away without even knowing it.
Speaker 3 (07:08):
There's a lot of opportunity to give your data
away without even knowing it andleaking important information
out into the ether or havingavailable to be taken.
So incorporating security fromthe get-go allows for your
product or your model, yourwhatever to be developed in a
secure fashion so that at everygiven point in time you are
(07:32):
doing that check and balance andmaking sure that what you're
developing is sound and will bevery robust once it gets to
deployment.
When you throw it on like alittle extra layer at the end,
it's not baked all the waythrough.
There are opportunities forthere to be vulnerabilities
throughout your data collectionwhere somebody could have
(07:55):
injected malicious data.
There's opportunities withinyour pipelines where somebody
could be changing values or noteven maliciously, but something
be going wrong when you're notcatching it changing values or
not even maliciously, butsomething be going wrong when
you're not catching it.
And if you had security baked insome checks there, then you
would have caught that earlieron.
Or if you're letting your, ifyou're not securing your model
(08:16):
weights files and somebody isable to get them and they're
able to just basically stealyour model for lack of a better
term Then all of a sudden allthat time and effort that you
put in is just it's gone, um,not gone, but uh, available to
somebody else without youknowing it.
And if you work hand in hand,um then then you're avoiding
that and you're building it froma strong foundation rather than
(08:37):
just putting like a littlewhipped cream on top of your
sundae security as well is notjust something that you, it's
not a conversation you have atthe beginning, and it's
certainly not a conversation youhave at just the end.
Speaker 4 (08:49):
It's a conversation you have throughout the entire
process.
Every work stream, every youknow if you're doing a
deployment, every sprint andeverything like that needs to
have a security conversation andneeds to have a security expert
involved in it, becauseeverything and I'm thinking
about the cloud right now everysingle thing you're going to be
doing needs to be secure.
If you're working in a data set, you need to let's get down to
(09:10):
the specifics of even role levelsecurity in your data set.
These little things really domatter, and so it needs to be a
conversation throughout theentire process.
Speaker 3 (09:19):
I also know that data scientists don't have that
inherently built in.
Speaking from my ownperspective, you don't really
learn a lot about cybersecurityin your data science curriculum.
It's not something that'splaced as high value, but as the
ecosystem that we're currentlyoperating within has moved
forward at such a rapid pace,you need to incorporate that
(09:40):
security and it's kind of beingleft behind right now.
So we're playing the game ofcatch up to make sure that we
have cross-functional teams,that we're including those
security experts and teachingthem the world of AI, because
maybe not always all expertsjust like we're not experts on
cybersecurity and trying to findsome common ground in terms of
verbiage can be really difficultand is where some time can be
(10:01):
spent to really help the endgoal.
Speaker 2 (10:07):
The two things I think of is, when I think of
Zaid I think you'd mentionedwhen we talked earlier about how
ideal cloud can be forprototyping, ai understanding
the caveats that are going tocome along with that or need to
be aware of.
But also, in my experience, atleast initially, a lot of
organizations may have overassumed that security is kind of
taken care of by using cloud.
(10:28):
It's almost like well, theytake care of the infrastructure,
so now they're also taking careof the security, don't they?
And obviously there's a sharedresponsibility model We'll talk
a little bit more about and theunderstanding of where that
begins and ends, because Iassume the cloud is probably a
good place to do security, butit doesn't mean you don't have
to think about it and bake it in, as you're saying.
Do you see organizations thatare baking it in, as you said at
(10:51):
the beginning and all the waythrough?
I assume there's got to be somethat are at least attempting to
do that.
Oh yeah, oh yeah definitely.
Speaker 4 (10:58):
I mean, I think this is more about maturity, right.
And so a lot of customers,right.
When cloud really started tofind its feet or its ground, a
lot of people were justmigrating their workloads from
whatever on-prem data centerthey had to the cloud.
Some people did it well, somepeople didn't do it well.
(11:18):
People who didn't do it well aretrying to catch up, but there
are definitely a lot of peoplewho are doing it well and a lot
of the times, what I've seen inmy experience is that they are
from industries that havehistorically always done it well
, right.
Like, if we're talking aboutglobal fire health care, there's
really strict complianceregulations that they have to
follow at every, at every stageof an application development
(11:39):
cycle.
Or just literally like thenurses and the doctors who are
badging into the building, right.
So these are industries who,historically, have done security
very, very well, right.
So these are industries whohistorically, have done security
very, very well, and so you see, a lot of times, those
(12:06):
industries are also excelling attheir security in cloud because
they're the ones putting itfirst versus you know, a retail
company who maybe historicallyhasn't been the best at security
, because they have a need toyou know, consider cybersecurity
for their small business thatis now rapidly growing.
Speaker 3 (12:10):
So specifically for AI, cybersecurity for AI is just
something, especially worldwide, that we are just pushing at
the forefront because it's sonecessary for those industries
that don't have compliance thatthey have to make or adhere to.
Speaker 2 (12:22):
It was always struggled with the fact that we
want security to be moreproactive, but by its very
nature it is reactive.
It is very difficult to getmoney and get time and attention
paid to something that hasn'thappened yet, and so there is a
chicken and egg challenge here,I think, because I don't think
anybody would disagree that it'simportant, but it is probably
harder to establish up front.
You're like I just want toprototype, make sure it works
(12:42):
first, then we'll worry aboutsecuring it.
And you're like well, you'reprobably going to pass up some
opportunities to really do itcorrectly if you don't do that
early on.
Oh yeah, definitely, definitely.
Well, I want to get into actionsteps.
I took the liberty of kind ofcombining these into four areas
(13:03):
and somewhat of a practicalroadmap.
And it's foundationarchitecture, strategy and then
security.
Notice how I put security atthe end.
We'll just roll with it andpretend that he knows better, he
knows.
I want to go through these realquick, just so I can bring up
on screen, because number one isestablish your data foundation.
(13:24):
Two is build with the rightarchitecture.
Three is align strategy withexecution.
And then four implementsecurity and governance.
I love the actionable nature ofthese actions, of these action
steps, that's not redundant Ina.
So, number one establish yourdata foundation.
What does it mean, especiallyfrom your perspective as a data
scientist, about having gooddata?
Speaker 3 (13:44):
When it comes to use cases, making sure that you have
your use case well-defined andthat there's data there.
It's like step one ofeverything.
And then, once you've made itthat far, making sure that we
have access to all that data,making sure it's readily
available when we go to pullfrom it and that it is actually
(14:07):
full of rich information and notjust full of empty values.
Speaker 2 (14:12):
That's kind of the whole point of building your own
models right, which is it needsto be in something that is
going to be expressly anduniquely yours, because that's
going to become part of yourextending your IP.
I guess, so to speak, step twobuild with the right
architecture.
Zaid, I'm curious about when itcomes about securing access to
high-performance architecture.
What does that mean?
When you talk about securingaccess to that architecture for
(14:33):
an organization, there's theability to actually use it,
secure your ability to use it.
Speaker 4 (14:37):
and then there's actually securing the
infrastructure itself, whetherthat's on-prem or in the cloud.
So, you know, being able toactually leverage the
(14:57):
infrastructure itself is oneconversation because, you know,
famously, especially for a lotof large organizations like, for
example, nvidia GPUs are reallyhard to come by.
And then also there's, likeCapEx versus OpEx, in terms of
actually purchasing theinfrastructure that you're going
to run models or do training on.
So securing, you know, in termsof getting the infrastructure,
is different for every companyand it's different depending on
your use case.
But I would say that it's asignificant problem that people
are facing, especially if theywant to build their own data
(15:18):
center.
But then it's also asignificant issue for people who
are running in the cloud.
Because if you are running inthe cloud and you're doing, you
know, high performance compute,you're probably it's probably
going to be pretty expensive.
So when I hear securing accessto high performance architecture
, there's the issue of actuallyacquiring the infrastructure or
the ability right to computenecessary.
(15:42):
But then there's also the skillissue right, the skill and the
education that's required whenit comes to building, with the
skill issue right, the skill andthe education that's required.
Speaker 2 (15:49):
When it comes to building with the right
architecture.
I know what lessons could bedrawn with that in mind in terms
of understanding architecturaldifferences between something
that's maybe proof of conceptversus scalable production,
ready Right.
Speaker 3 (16:03):
Yeah, I mean when Zed and I were working together on
building out the demo that webuilt out.
It was really a proof ofconference Whoa.
Speaker 4 (16:14):
It was a proof of conference at a conference.
It was a proof of concept for aconference.
Speaker 3 (16:19):
I swear we were going to have gone, but the idea was
like okay, we have a reallyshort amount of time, I need to
make sure I have access to someGPUs that go fast, because it
was going to be using a largelanguage model and typically
when you need to use those, youneed a little bit more support.
I needed access to some sort offoundational model, so some
(16:40):
sort of LLM that we could usefor our use case, spin it up in
a short amount of time, and allit had to do was run on any
computer where we pulled it up,uh, at any given point in time
in front of an audience which,uh, is is demo nightmare, um,
but it is what it was.
(17:01):
It we weren't selling it toanybody.
We were doing it to show whatthe art of the possible really
was and what you can do withwhat exists out there right now.
Specifically, we were using umgcp and they'd was was
supporting on the architectureside and was able to show hey,
we don't, we don't need to tospin up a whole new environment
within the air proving ground.
(17:22):
I can click a couple buttoningsand within 10 minutes we can
have an environment for you.
You know to go put some data inand start developing, which for
me is fantastic and definitelynot always the case.
Sometimes it takes a lot longerto get the environment for
myself to work within spun upand ready to go with all the
hardware different libraries andthen models that I would like
(17:47):
to build upon available to me orand have enough storage for the
data in a somewhat securelocation.
Speaker 4 (17:54):
Part of what we are trying to paint with the art as
possible there is that you know,the time to actual valuable
insight is definitely sped up byAI.
Right, there's generative AI iskind of is kind of integrated
into every piece of what we did,for example, you know, was
using Google collab.
Google Colab to actually writethe scripts that she was working
with, and there's like Codasystem there, and so that's
(18:15):
something cool to help youfinish what you're writing.
But then you know, I don't wantto diminish the fact that we
got to results, meaningfulresults, very quickly, but then
to kind of answer the originalquestion, it was just a POC,
right?
There's so many other thingsthat we need to consider.
I mean, like we were usingsample datasets that we, I think
(18:35):
that we used like ChagPT orGemini to create for us.
Speaker 3 (18:40):
Yeah.
Speaker 4 (18:41):
But what?
Speaker 3 (18:42):
the value there was was to show hey, even with fake
data, we can draw insights.
If you have a use case that'ssimilar to this, this is
something that can be solved ina similar manner to which we did
, and the proof of conceptreally sparks ideas in people
and then they start thinkingabout well, I have a similar
(19:02):
problem, but it's slightlytweaked.
Can you address this in asimilar fashion?
Yes, we can.
Well, I have this way.
Way, I would like to do it theopposite way around.
Could I?
Could I invert it?
Yep, models can also handlethat.
So it's the architecture and thepoc.
The value there is doing itquick, because when you're
talking with customers, usuallyyou only have so much momentum
(19:23):
for so long, and so when you canspin it up and you can spin it
up fast then you can show whatthe value is.
They keep their excitement,momentum.
Everybody's happy.
If you need to spin up anentire lab space and take about
a week and a half, two weeks, todo so because you want more
robust architecture for along-term engagement, you might
(19:44):
lose that momentum.
And then great, you have thissuper robust environment but
nobody, no customer, to fill it.
Speaker 2 (19:51):
So it's a give and a take, and when you're poc things
out and doing quickly, like zayand I do um on a weekly basis,
then then the short and dirtyway uh is is very nice saying
there's a it's important to setexpectations correctly about
this does not equate to yes, ifyour team's not doing it at this
(20:13):
speed in production, there'ssomething wrong with them.
You're not saying that at all.
You're saying this is a proofof concept, or even more work
was done to make it for a lab,but the lab is not production.
Correct and it shouldn't berelied on as much.
Speaker 3 (20:25):
Yes.
Speaker 2 (20:29):
Because understand our point we're making and the
points we're not making, to makesure that you don't conflate
the two.
Speaker 4 (20:31):
Yeah, yeah, and I think that speaks to just how
far AI has come right, the factthat we were able to spin up a
demo like that so quickly,honestly and I mean, this was
our second job, right, we stillhad all of our main priorities
that we're working on on a dailybasis, and then we kind of
hodgepodge this together in likea week and a half or two weeks,
and so that's just.
It's awesome that we were ableto do that, and I think that's
(20:53):
also speaks to the flexibilitythat cloud gives you, because we
were able to run it on thecloud If we tried to build this
in the ATC or anything like that, which is an amazing, you know,
an amazing thing that we weredoing, and so spinning up a
really quick instance in cloudwas really useful.
Speaker 2 (21:14):
Let's double down on the paper, and then we talked
about this a little bit.
Talks about the importance ofreal world use cases, and then
step three, where we talk aboutaligned strategy with execution.
How do you help customers focuson actual or pick use cases
that matter to them?
What are some things you mayhelp them keep in mind?
Speaker 4 (21:31):
So I think every customer is different, and I
think part of the great thingabout my role is that, you know,
I honestly don't get as deep asIna does into you know, she
gets super deep into datascience.
I get to be needy about a lot ofdifferent things, so I get to
work with a lot of differentindustries, which is awesome,
and so what I've noticed acrossthe industries is that I mean,
that's why we have industrycommittees, right Is that
(21:52):
they're all.
All their use cases are verydifferent, and so I think what
helps is that when you firstspeak to somebody, if you're
speaking to whatever stakeholder, it is, focus on the outcome.
Focus on the outcome ratherthan the tool, rather than the
process.
Honestly, before I even talkabout their current environment
and what struggles they'refacing, in that minute I talk
(22:14):
about what their ideal futurestate is going to look like, and
when I get a really goodunderstanding of what outcome
they want to achieve, that tellsme about their current
environment.
That tells me about the painpoints that they're seeing,
without even having to have thatconversation.
Of course, we'll dive intothose conversations, but
focusing on the outcome thatthey want to see and what their
ideal future state looks likereally gives me a good ground to
(22:38):
understand where we should go.
Speaker 2 (22:41):
How often do you guys run into a situation where you
then have to?
Then you know, okay, I see whatyou're trying to achieve.
I think we need to narrow itdown to ensure.
Speaker 4 (22:51):
Everybody wants to boil the ocean.
Everybody wants to boil theocean.
As a trusted partner, I get tosee people level set and be
honest, like maybe this toolthat you're using isn't great,
or maybe you need to come downnot just one step or two steps,
but like 10 steps, so that wecan start having a real
conversation, talking about areal use case.
Speaker 2 (23:09):
Do you get a lot of pushback often on those things,
or do they listen to whatthey're paying you to do?
Speaker 3 (23:13):
People always want more than you can deliver.
That's my two cents from thedata science perspective.
Speaker 2 (23:20):
Do you like to see them maybe run full circle on
something simple that can showresults before they start maybe
tackling more complex?
Things Like even with the wayWorldwide rolled out Atom I
don't know what it's calledtoday things Like even with the
way Worldwide rolled out ATOM Idon't know what it's called
today.
It may still be somewhatsimilar, but I remember it was
very constrained in terms ofwhen it was being rolled out, as
to who it affected, what datait was, resourcing and things
(23:40):
like this, and they waited untilthey got comfort level with
that before expanding access aswell as expanding the training
models, I believe.
Speaker 3 (23:46):
Yeah.
So I'm going to talk about datascience and how we approach it
model-wise.
But we really like to iterateand you can grow when you're in
this iteration phase.
So, rather than going fullcircle, think more spiral, where
you're spiraling up in a spring.
I guess you achieve that firstuse case or you get to a point
where you're comfortable with itand then you can build upon it.
(24:08):
So with Atom, we started justwith the basic LLM, evaluated it
yeah, basic LLM and then gaveit access to information on
worldwidetechnologycom or wwwcom, the platform.
Okay, great, everybody can haveaccess to that.
All right, what was the nextlevel?
Next level was okay, make surethat the model is responding
intelligently.
It's not hallucinating so much.
(24:29):
Okay, improve upon those pieces, all right.
Next step more data sources.
Let's feed in more data sources.
Now we have five that areavailable to us Awesome.
Next step above that well,let's limit who can access what.
I'm not necessarily an accountmanager.
I don't need access to the sametype of data that they do.
And then, finally, now we'vegotten to the point where we're
(24:50):
deploying agents, so differentmodels that have different
responsibilities and arefine-tuned to a specific task.
So when you start small andyou're able to build that solid
foundation that you can thenreuse and build upon.
Then you have the ability to dothis exponential growth, which
is what we've seen with Atom.
Exponential growth, which iswhat we've seen with Adam.
(25:11):
It was a very slow build at thebeginning and I think that was
something a lot of people whoweren't scientists at this point
in time that's the feedbacklist.
We've been working on WWT, gptfor so long, so long, so long,
so long, so long.
And just recently, have youseen that exponential growth of
like oh, we all have access, oh,there's new data services, oh,
now there's agents, and all ofthat in a short number of months
(25:32):
?
And you see the time that ittakes to add on those extra
features and build upon what youhave go down significantly,
because you started small andthen iterated upon that initial
solution.
Speaker 2 (25:45):
So step four is around implement security and
governance From a cloudperspective.
From a cloud perspective, zayd.
What specific skills doorganizations need for
governance when it comes tocloud?
What expectations should theyhave about what cloud's doing
for them versus where they needto take more responsibility when
it comes to security andgovernance?
Speaker 4 (26:08):
Governance isn't just a specific tool.
There are tools in cloud thathelp you with governance so that
you are following right toright permissions access and
you're labeling and tagging yourdata correctly, et cetera.
But governance, to me, I think,starts first and foremost with
education.
Speaker 2 (26:22):
Yeah, and I think and that's a big issue because, as
you guys both had mentionedearlier you know, one of the
challenges we're dealing with isthat we're all we're learning
what AI can do for us.
At the same time we're tryingto learn how to operate it, and
usually historically, thosethings have been really
separated.
We didn't have technology rollout so fast that we're using it,
while it's still changing asrapid as it is.
Speaker 3 (26:44):
My two cents on governance is get in early,
understand what is important toyour organization, whether it's
within the cloud, legal data,science, network space.
How are you running yourgovernance?
And then have every solutionadhere to that, Because you can
(27:06):
build models and products andsolutions in any type of way,
shape or form, but when giventhose guidelines, you are much
more likely to get a productthat adheres to it rather than
at the end being like oh, by theway, this needs to adhere to
this one policy that we didn'ttell you about ahead of time and
that might just be the exactkey piece of information that
(27:27):
the model is using to make itspredictions.
Speaker 4 (27:30):
Once you're educated and you've established what your
framework's going to be and howyou're going to implement
governance, then let's talkabout tools, because there's a
lot of tools inside.
You know, of course I'mspeaking from the cloud
perspective, but there's a lotof really cool tools to
implement governance, whetherit's at the data level for
labeling and tagging, likeDataplex in GCP or like
(27:51):
something like Assured Workloadsand again, I speak from the GCP
perspective.
But if you have a compliancestandard you need to hit or you
need to meet excuse me you canimplement that across your
entire cloud environment inGoogle Cloud by using something
called Assured Workloads, and sowhen you try to use a tool or
you try to access data orsomething like that that you
shouldn't be able to, itliterally flags you and stops
(28:12):
you.
So if you have a solidfoundation of how you want to
implement governance and whatstandards you need to meet,
there's a likelihood that thetool is there to help you follow
that.
Speaker 2 (28:29):
Do you think?
Is it logical to assume maybethis is too easy to answer but
the notion that I think we'regoing to have more regulations
and oversight that we can'tpredict now, and so if you say,
well, I'm safe doing thesethings now because no one's
watching or cares or something,and so planning it out and
making sure that you'readdressing this not based on
what the law or anyone else saysyou should do, but what you
(28:51):
know should be done, might saveyou a few steps in the future
and your ability to pivot whenthose things inevitably change,
perhaps, oh yeah, 100%.
Speaker 3 (28:59):
There's also a bunch of voluntary frameworks out
there right now, or even likeheads up like OWASP and NIST.
They put out a top 10vulnerabilities of LLMs.
So if you look at what thecurrent vulnerabilities are and
what the industry trends are,you can see where those
guidelines are then pointing.
(29:19):
At this point, within theUnited States, there is no AI
regulation.
That's very clear.
So you have to look a littlebit deeper into what the
industry is recommending andthat is not easy because there's
a lot of voices.
But paying attention to thoselike OWASP and NIST that have
(29:42):
been very strong in the securityrealm for a long time and are
now looking to include AI aspart of that.
And I think Cloud's a littlefurther along.
So tell me if I'm wrong, butit's still following.
It's still behind traditionalsecurity.
Speaker 4 (29:56):
Yeah, yeah, there are trends and publications that
come out all the time with youknow, recommendations about how
you can continue to secure yourworkloads and, specifically, how
you should be prepping for AI.
Speaker 2 (30:10):
Yeah, and then we generally see a lot of those get
adopted, because that becomesthe easier route, especially
because a lot of forwardthinkers are engaging with that
and they're going.
This one doesn't make sense,even though they're all
voluntary.
If we engage enough, then wecome out with something that is
reasonable and allows us all tostill hit our objectives without
being too onerous.
Yeah, yeah, you're right,they're probably going to come
(30:32):
from those pre-existingvoluntary frameworks.
Okay, so final questions.
What's the one thing that youwish that they knew more about
in terms of building AI-readyfoundations early on?
Who wants to take that first?
Speaker 4 (30:45):
Tina does I do I do.
Speaker 3 (30:48):
80% of your problem is a data problem.
Every AI problem 80% of itsdata.
So just expect that, becausethe first thing I'm going to
tell you is that you need tohandle your data, and being
surprised every single time whenyou have a new use case and
you're like but my data was fine, last night, we fixed it.
(31:09):
It's like we fixed it for thatuse case.
Every single AI problem is 80%a data problem.
Speaker 1 (31:16):
Okay, that was a great conversation and I want to
thank Rob, ina and Zaid fortheir insight on this important
issue facing any organizationlooking to scale their AI
initiatives.
A few things stood out to meabout the conversation.
First, your AI initiative willrise or fall on the quality and
accessibility of your data.
Prioritize projects for whichthe needed data already exists
(31:39):
and is well governed.
Second, bake security andgovernance into every step,
never as an afterthought.
Integrating security controls,role-based access and compliance
frameworks, early prevents dataleakage, model theft and costly
rework.
And third, start small, iteratefast and align architecture to
outcomes, not politics or hype.
(32:01):
Success comes from a clearoutcome, the right
high-performance infrastructureand a spiral, iterative roadmap
that expands only after smallwins are reached.
Bottom line.
The companies dominating withAI in 2025 won't be the ones who
move first.
They'll be the ones who movedright, who understood that
boring data governance andinfrastructure work is what
(32:22):
makes breakthrough AI possible.
That's it for this episode ofthe AI Proving Ground podcast.
A special thanks again to RobBoyd for hosting, brian Flavin
and Amy Riddle for their contentsupport.
If you're interested in moreepisodes of the AI Proving
Ground podcast, please considersubscribing on your favorite
podcast channel or check us outon WWTcom and we'll see you next
(32:44):
week.
Every week, another organization discovers their
million-dollar AI initiative wasbuilt on a lie the lie that you
can skip the boring stuff andjump straight to the
transformative part, that dataquality is someone else's
problem or security can bebolted on later, or that cloud
infrastructure just somehowworks.
You know where this is going,probably.
(00:22):
By some estimates, more than80% of AI projects fail twice
the rate of failure forinformation technology projects
that don't involve AI.
Organizations are hemorrhagingmoney trying to build
breakthrough AI solutions, whiletheir cloud costs spiral out of
control, their data sits insilos across disconnected
systems and their securitypolicies haven't caught up to
(00:43):
their ambitions.
Its systems and their securitypolicies haven't caught up to
their ambitions.
Today, on the AI Proving Groundpodcast, I'm throwing it over to
my colleague, rob Boyd, whorecently talked with Ina Posher,
a data scientist who bridgesthe gap between what data
scientists want to build andwhat actually works in
production, and Zaid El-Khatib,a technical solutions architect
at WWT, who helps organizationsescape the AI corners they've
(01:04):
backed themselves into.
You're about to discover whymost AI failures happen before
the first model runs and theroadmap that separates AI
winners from casualties.
So, without further ado, let'sthrow it over to Rob.
Speaker 3 (01:38):
What have you learned about this gap between what
data scientists which, as Iunderstand, that's really your
focus here what data scientistswant to build versus what
actually works when it comes toreal production.
I would say the biggest gapthat exists is that there's so
many high-level ideas and somany things that I could solve
so many use cases that have highvalue and would have high
return, and then the data isjust not there or there's not
enough data there, and when thedata is not there as that
(01:59):
foundation, then there's notmuch you can do for your use
case and you're just stucktrying to create something out
of nothing.
So that's the biggest gap ondelivering a solution and really
making it worthwhile to investin.
Is your data there and is itready to go?
Speaker 2 (02:18):
And sometimes I guess that results as a mismatch
maybe between what theorganization is where they're
setting their expectationsversus you know what can
actually be done.
So it comes into.
You know, pick and we'll talkabout this further but picking
your use case especially earlyon and kind of being narrow
about that and understandingwhat all feeds into it, it
sounds like.
Speaker 3 (02:35):
Yeah, and prioritizing it based on the
data you have available, thedata that you can easily access
and the tools you have availableto solve it.
So it's not just what willprovide the most value to the
company, it's a return oninvestment.
So there has to be you have toinvest something and the return
comes from what's available andwe can't just invest, invest,
(02:57):
invest and not return anything.
Speaker 2 (02:59):
All right.
Well, obviously we're going totalk about that one a little bit
further in detail about how weactually pull that off.
But, zaid, let's get you inthere here.
You had told me you specializein Google Cloud and specifically
, you kind of help account teamsand customers.
However, it happens, they getthemselves back into a corner,
perhaps with cloud, with AI,with some combination of the two
(03:19):
, and you specialize in helpingbring them out.
Are you seeing any kind ofpatterns in terms of common
mistakes that are made?
Any patterns.
Speaker 4 (03:28):
So, yeah, definitely, I think what we see a lot of is
that you know, because,admittedly, right in our field
there are politics that areinvolved, and you know, I don't
mean government in the state ofthe world or anything like that,
but I mean in how you know howa solution is given to a
customer or a product is.
And so a lot of the times I'llsee a customer who, for one
(03:50):
reason or another, has beentalking a certain product or a
certain tool for a very longtime and they've purchased it
and they're trying to use it andget the most value out of it,
but they only really qualifiedthe product itself to see how
sound it is, you know, in abubble, and they don't often
look at what the outcome they'retrying to achieve is, and so
(04:13):
they'll purchase a product wherethey'll start using a
technology and it doesn'tnecessarily fit the outcome that
they are looking for, or it'snot the only thing that's going
to get them to the outcome thatthey're looking for.
And so sometimes I'll come in,because cloud is so wide and
there's so many technologies andthere's so many tools that are
hosted on the cloud cloud native, et cetera so I'll come in and
(04:34):
further qualify the tool thatthey're using and see if we can
mess that with another tool, orsee if they're not using it
correctly and see how we can getthem out of it.
Speaker 2 (04:43):
Yeah, and if you see that happening on a repetitive
basis, are there common elementsthat are just not taken into
consideration when approachingthese things?
Is that it, or does it comefrom many different directions?
Speaker 4 (04:55):
You know it comes from a lot of different
directions.
I brought up politics a littlebit because a lot of the times,
right, you have C-levelexecutives who are being told
really know, told really goodthings by people that they're
familiar with in a companyalready, and so, honestly, I see
a lot of I would say a lot ofissue come from those kinds of
conversations.
But then a lot of the timesit's that you know, perhaps they
(05:19):
are have a generative AIinitiative that is a little too
lofty and so they're trying tohit that you know initiative.
Or they're trying to hit thatis a little too lofty and so
they're trying to hit that youknow initiative, or they're
trying to hit that goal a littletoo quickly.
Or you know, maybe their youknow outcome and the tool that
they're using to get to thatoutcome is, is solid and it's
qualified really well, but youknow their data isn't quite
ready for it yet and so we haveto backtrack and so they didn't
(05:40):
look at the entire roadmap.
They, they didn't look at theentire roadmap, they were just
looking at the end of thedestination.
Speaker 2 (05:45):
Well, there's nothing more fun than finding those out
when you're mid-deployment.
Oh yeah, and then it comes up.
Those are not goodconversations, yeah.
Speaker 4 (06:01):
Then timelines, money everything just starts
expanding.
Yeah, that's when the customersstart bleeding money, For sure.
Speaker 2 (06:05):
Yeah, Well, and we're here to help hopefully prevent
the bleeding.
So there's two priorities thatwe're kind of blending together
into this conversation and wantto acknowledge that, because we
kind of bounce back and forthbetween some of them, but I
actually think this works welltogether.
In the research paper there'spriority two, which is about
data security and governance,and then there's priority three,
which is harnessing cloudinfrastructure for AI.
(06:26):
We've had elements of thesethings talked about on the show
and other ways before, but I'mcurious we talk about.
It's interesting to me always.
I always love seeing wheresecurity gets mentioned.
Is it before or after certainelements, because I feel like
it's always put in as kind of ohyeah, don't forget security.
Of course, now we have troublewith regulatory and compliance
and things with that, and I'mprobably just as guilty as
(06:48):
anyone else at doing that, aswe'll see here in a minute.
But, ina, I'm curious about whywe can't just skip ahead to the
AI part.
Why would it be prioritized?
Why would we need to thinkabout security?
Speaker 3 (07:02):
There are a lot of reasons.
Speaker 4 (07:05):
There's a lot of opportunity to give your data
away without even knowing it.
Speaker 3 (07:08):
There's a lot of opportunity to give your data
away without even knowing it andleaking important information
out into the ether or havingavailable to be taken.
So incorporating security fromthe get-go allows for your
product or your model, yourwhatever to be developed in a
secure fashion so that at everygiven point in time you are
(07:32):
doing that check and balance andmaking sure that what you're
developing is sound and will bevery robust once it gets to
deployment.
When you throw it on like alittle extra layer at the end,
it's not baked all the waythrough.
There are opportunities forthere to be vulnerabilities
throughout your data collectionwhere somebody could have
(07:55):
injected malicious data.
There's opportunities withinyour pipelines where somebody
could be changing values or noteven maliciously, but something
be going wrong when you're notcatching it changing values or
not even maliciously, butsomething be going wrong when
you're not catching it.
And if you had security baked insome checks there, then you
would have caught that earlieron.
Or if you're letting your, ifyou're not securing your model
(08:16):
weights files and somebody isable to get them and they're
able to just basically stealyour model for lack of a better
term Then all of a sudden allthat time and effort that you
put in is just it's gone, um,not gone, but uh, available to
somebody else without youknowing it.
And if you work hand in hand,um then then you're avoiding
that and you're building it froma strong foundation rather than
(08:37):
just putting like a littlewhipped cream on top of your
sundae security as well is notjust something that you, it's
not a conversation you have atthe beginning, and it's
certainly not a conversation youhave at just the end.
Speaker 4 (08:49):
It's a conversation you have throughout the entire
process.
Every work stream, every youknow if you're doing a
deployment, every sprint andeverything like that needs to
have a security conversation andneeds to have a security expert
involved in it, becauseeverything and I'm thinking
about the cloud right now everysingle thing you're going to be
doing needs to be secure.
If you're working in a data set, you need to let's get down to
(09:10):
the specifics of even role levelsecurity in your data set.
These little things really domatter, and so it needs to be a
conversation throughout theentire process.
Speaker 3 (09:19):
I also know that data scientists don't have that
inherently built in.
Speaking from my ownperspective, you don't really
learn a lot about cybersecurityin your data science curriculum.
It's not something that'splaced as high value, but as the
ecosystem that we're currentlyoperating within has moved
forward at such a rapid pace,you need to incorporate that
(09:40):
security and it's kind of beingleft behind right now.
So we're playing the game ofcatch up to make sure that we
have cross-functional teams,that we're including those
security experts and teachingthem the world of AI, because
maybe not always all expertsjust like we're not experts on
cybersecurity and trying to findsome common ground in terms of
verbiage can be really difficultand is where some time can be
(10:01):
spent to really help the endgoal.
Speaker 2 (10:07):
The two things I think of is, when I think of
Zaid I think you'd mentionedwhen we talked earlier about how
ideal cloud can be forprototyping, ai understanding
the caveats that are going tocome along with that or need to
be aware of.
But also, in my experience, atleast initially, a lot of
organizations may have overassumed that security is kind of
taken care of by using cloud.
(10:28):
It's almost like well, theytake care of the infrastructure,
so now they're also taking careof the security, don't they?
And obviously there's a sharedresponsibility model We'll talk
a little bit more about and theunderstanding of where that
begins and ends, because Iassume the cloud is probably a
good place to do security, butit doesn't mean you don't have
to think about it and bake it in, as you're saying.
Do you see organizations thatare baking it in, as you said at
(10:51):
the beginning and all the waythrough?
I assume there's got to be somethat are at least attempting to
do that.
Oh yeah, oh yeah definitely.
Speaker 4 (10:58):
I mean, I think this is more about maturity, right.
And so a lot of customers,right.
When cloud really started tofind its feet or its ground, a
lot of people were justmigrating their workloads from
whatever on-prem data centerthey had to the cloud.
Some people did it well, somepeople didn't do it well.
(11:18):
People who didn't do it well aretrying to catch up, but there
are definitely a lot of peoplewho are doing it well and a lot
of the times, what I've seen inmy experience is that they are
from industries that havehistorically always done it well
, right.
Like, if we're talking aboutglobal fire health care, there's
really strict complianceregulations that they have to
follow at every, at every stageof an application development
(11:39):
cycle.
Or just literally like thenurses and the doctors who are
badging into the building, right.
So these are industries who,historically, have done security
very, very well, right.
So these are industries whohistorically, have done security
very, very well, and so you see, a lot of times, those
(12:06):
industries are also excelling attheir security in cloud because
they're the ones putting itfirst versus you know, a retail
company who maybe historicallyhasn't been the best at security
, because they have a need toyou know, consider cybersecurity
for their small business thatis now rapidly growing.
Speaker 3 (12:10):
So specifically for AI, cybersecurity for AI is just
something, especially worldwide, that we are just pushing at
the forefront because it's sonecessary for those industries
that don't have compliance thatthey have to make or adhere to.
Speaker 2 (12:22):
It was always struggled with the fact that we
want security to be moreproactive, but by its very
nature it is reactive.
It is very difficult to getmoney and get time and attention
paid to something that hasn'thappened yet, and so there is a
chicken and egg challenge here,I think, because I don't think
anybody would disagree that it'simportant, but it is probably
harder to establish up front.
You're like I just want toprototype, make sure it works
(12:42):
first, then we'll worry aboutsecuring it.
And you're like well, you'reprobably going to pass up some
opportunities to really do itcorrectly if you don't do that
early on.
Oh yeah, definitely, definitely.
Well, I want to get into actionsteps.
I took the liberty of kind ofcombining these into four areas
(13:03):
and somewhat of a practicalroadmap.
And it's foundationarchitecture, strategy and then
security.
Notice how I put security atthe end.
We'll just roll with it andpretend that he knows better, he
knows.
I want to go through these realquick, just so I can bring up
on screen, because number one isestablish your data foundation.
(13:24):
Two is build with the rightarchitecture.
Three is align strategy withexecution.
And then four implementsecurity and governance.
I love the actionable nature ofthese actions, of these action
steps, that's not redundant Ina.
So, number one establish yourdata foundation.
What does it mean, especiallyfrom your perspective as a data
scientist, about having gooddata?
Speaker 3 (13:44):
When it comes to use cases, making sure that you have
your use case well-defined andthat there's data there.
It's like step one ofeverything.
And then, once you've made itthat far, making sure that we
have access to all that data,making sure it's readily
available when we go to pullfrom it and that it is actually
(14:07):
full of rich information and notjust full of empty values.
Speaker 2 (14:12):
That's kind of the whole point of building your own
models right, which is it needsto be in something that is
going to be expressly anduniquely yours, because that's
going to become part of yourextending your IP.
I guess, so to speak, step twobuild with the right
architecture.
Zaid, I'm curious about when itcomes about securing access to
high-performance architecture.
What does that mean?
When you talk about securingaccess to that architecture for
(14:33):
an organization, there's theability to actually use it,
secure your ability to use it.
Speaker 4 (14:37):
and then there's actually securing the
infrastructure itself, whetherthat's on-prem or in the cloud.
So, you know, being able toactually leverage the
(14:57):
infrastructure itself is oneconversation because, you know,
famously, especially for a lotof large organizations like, for
example, nvidia GPUs are reallyhard to come by.
And then also there's, likeCapEx versus OpEx, in terms of
actually purchasing theinfrastructure that you're going
to run models or do training on.
So securing, you know, in termsof getting the infrastructure,
is different for every companyand it's different depending on
your use case.
But I would say that it's asignificant problem that people
are facing, especially if theywant to build their own data
(15:18):
center.
But then it's also asignificant issue for people who
are running in the cloud.
Because if you are running inthe cloud and you're doing, you
know, high performance compute,you're probably it's probably
going to be pretty expensive.
So when I hear securing accessto high performance architecture
, there's the issue of actuallyacquiring the infrastructure or
the ability right to computenecessary.
(15:42):
But then there's also the skillissue right, the skill and the
education that's required whenit comes to building, with the
skill issue right, the skill andthe education that's required.
Speaker 2 (15:49):
When it comes to building with the right
architecture.
I know what lessons could bedrawn with that in mind in terms
of understanding architecturaldifferences between something
that's maybe proof of conceptversus scalable production,
ready Right.
Speaker 3 (16:03):
Yeah, I mean when Zed and I were working together on
building out the demo that webuilt out.
It was really a proof ofconference Whoa.
Speaker 4 (16:14):
It was a proof of conference at a conference.
It was a proof of concept for aconference.
Speaker 3 (16:19):
I swear we were going to have gone, but the idea was
like okay, we have a reallyshort amount of time, I need to
make sure I have access to someGPUs that go fast, because it
was going to be using a largelanguage model and typically
when you need to use those, youneed a little bit more support.
I needed access to some sort offoundational model, so some
(16:40):
sort of LLM that we could usefor our use case, spin it up in
a short amount of time, and allit had to do was run on any
computer where we pulled it up,uh, at any given point in time
in front of an audience which,uh, is is demo nightmare, um,
but it is what it was.
(17:01):
It we weren't selling it toanybody.
We were doing it to show whatthe art of the possible really
was and what you can do withwhat exists out there right now.
Specifically, we were using umgcp and they'd was was
supporting on the architectureside and was able to show hey,
we don't, we don't need to tospin up a whole new environment
within the air proving ground.
(17:22):
I can click a couple buttoningsand within 10 minutes we can
have an environment for you.
You know to go put some data inand start developing, which for
me is fantastic and definitelynot always the case.
Sometimes it takes a lot longerto get the environment for
myself to work within spun upand ready to go with all the
hardware different libraries andthen models that I would like
(17:47):
to build upon available to me orand have enough storage for the
data in a somewhat securelocation.
Speaker 4 (17:54):
Part of what we are trying to paint with the art as
possible there is that you know,the time to actual valuable
insight is definitely sped up byAI.
Right, there's generative AI iskind of is kind of integrated
into every piece of what we did,for example, you know, was
using Google collab.
Google Colab to actually writethe scripts that she was working
with, and there's like Codasystem there, and so that's
(18:15):
something cool to help youfinish what you're writing.
But then you know, I don't wantto diminish the fact that we
got to results, meaningfulresults, very quickly, but then
to kind of answer the originalquestion, it was just a POC,
right?
There's so many other thingsthat we need to consider.
I mean, like we were usingsample datasets that we, I think
(18:35):
that we used like ChagPT orGemini to create for us.
Speaker 3 (18:40):
Yeah.
Speaker 4 (18:41):
But what?
Speaker 3 (18:42):
the value there was was to show hey, even with fake
data, we can draw insights.
If you have a use case that'ssimilar to this, this is
something that can be solved ina similar manner to which we did
, and the proof of conceptreally sparks ideas in people
and then they start thinkingabout well, I have a similar
(19:02):
problem, but it's slightlytweaked.
Can you address this in asimilar fashion?
Yes, we can.
Well, I have this way.
Way, I would like to do it theopposite way around.
Could I?
Could I invert it?
Yep, models can also handlethat.
So it's the architecture and thepoc.
The value there is doing itquick, because when you're
talking with customers, usuallyyou only have so much momentum
(19:23):
for so long, and so when you canspin it up and you can spin it
up fast then you can show whatthe value is.
They keep their excitement,momentum.
Everybody's happy.
If you need to spin up anentire lab space and take about
a week and a half, two weeks, todo so because you want more
robust architecture for along-term engagement, you might
(19:44):
lose that momentum.
And then great, you have thissuper robust environment but
nobody, no customer, to fill it.
Speaker 2 (19:51):
So it's a give and a take, and when you're poc things
out and doing quickly, like zayand I do um on a weekly basis,
then then the short and dirtyway uh is is very nice saying
there's a it's important to setexpectations correctly about
this does not equate to yes, ifyour team's not doing it at this
(20:13):
speed in production, there'ssomething wrong with them.
You're not saying that at all.
You're saying this is a proofof concept, or even more work
was done to make it for a lab,but the lab is not production.
Correct and it shouldn't berelied on as much.
Speaker 3 (20:25):
Yes.
Speaker 2 (20:29):
Because understand our point we're making and the
points we're not making, to makesure that you don't conflate
the two.
Speaker 4 (20:31):
Yeah, yeah, and I think that speaks to just how
far AI has come right, the factthat we were able to spin up a
demo like that so quickly,honestly and I mean, this was
our second job, right, we stillhad all of our main priorities
that we're working on on a dailybasis, and then we kind of
hodgepodge this together in likea week and a half or two weeks,
and so that's just.
It's awesome that we were ableto do that, and I think that's
(20:53):
also speaks to the flexibilitythat cloud gives you, because we
were able to run it on thecloud If we tried to build this
in the ATC or anything like that, which is an amazing, you know,
an amazing thing that we weredoing, and so spinning up a
really quick instance in cloudwas really useful.
Speaker 2 (21:14):
Let's double down on the paper, and then we talked
about this a little bit.
Talks about the importance ofreal world use cases, and then
step three, where we talk aboutaligned strategy with execution.
How do you help customers focuson actual or pick use cases
that matter to them?
What are some things you mayhelp them keep in mind?
Speaker 4 (21:31):
So I think every customer is different, and I
think part of the great thingabout my role is that, you know,
I honestly don't get as deep asIna does into you know, she
gets super deep into datascience.
I get to be needy about a lot ofdifferent things, so I get to
work with a lot of differentindustries, which is awesome,
and so what I've noticed acrossthe industries is that I mean,
that's why we have industrycommittees, right Is that
(21:52):
they're all.
All their use cases are verydifferent, and so I think what
helps is that when you firstspeak to somebody, if you're
speaking to whatever stakeholder, it is, focus on the outcome.
Focus on the outcome ratherthan the tool, rather than the
process.
Honestly, before I even talkabout their current environment
and what struggles they'refacing, in that minute I talk
(22:14):
about what their ideal futurestate is going to look like, and
when I get a really goodunderstanding of what outcome
they want to achieve, that tellsme about their current
environment.
That tells me about the painpoints that they're seeing,
without even having to have thatconversation.
Of course, we'll dive intothose conversations, but
focusing on the outcome thatthey want to see and what their
ideal future state looks likereally gives me a good ground to
(22:38):
understand where we should go.
Speaker 2 (22:41):
How often do you guys run into a situation where you
then have to?
Then you know, okay, I see whatyou're trying to achieve.
I think we need to narrow itdown to ensure.
Speaker 4 (22:51):
Everybody wants to boil the ocean.
Everybody wants to boil theocean.
As a trusted partner, I get tosee people level set and be
honest, like maybe this toolthat you're using isn't great,
or maybe you need to come downnot just one step or two steps,
but like 10 steps, so that wecan start having a real
conversation, talking about areal use case.
Speaker 2 (23:09):
Do you get a lot of pushback often on those things,
or do they listen to whatthey're paying you to do?
Speaker 3 (23:13):
People always want more than you can deliver.
That's my two cents from thedata science perspective.
Speaker 2 (23:20):
Do you like to see them maybe run full circle on
something simple that can showresults before they start maybe
tackling more complex?
Things Like even with the wayWorldwide rolled out Atom I
don't know what it's calledtoday things Like even with the
way Worldwide rolled out ATOM Idon't know what it's called
today.
It may still be somewhatsimilar, but I remember it was
very constrained in terms ofwhen it was being rolled out, as
to who it affected, what datait was, resourcing and things
(23:40):
like this, and they waited untilthey got comfort level with
that before expanding access aswell as expanding the training
models, I believe.
Speaker 3 (23:46):
Yeah.
So I'm going to talk about datascience and how we approach it
model-wise.
But we really like to iterateand you can grow when you're in
this iteration phase.
So, rather than going fullcircle, think more spiral, where
you're spiraling up in a spring.
I guess you achieve that firstuse case or you get to a point
where you're comfortable with itand then you can build upon it.
(24:08):
So with Atom, we started justwith the basic LLM, evaluated it
yeah, basic LLM and then gaveit access to information on
worldwidetechnologycom or wwwcom, the platform.
Okay, great, everybody can haveaccess to that.
All right, what was the nextlevel?
Next level was okay, make surethat the model is responding
intelligently.
It's not hallucinating so much.
(24:29):
Okay, improve upon those pieces, all right.
Next step more data sources.
Let's feed in more data sources.
Now we have five that areavailable to us Awesome.
Next step above that well,let's limit who can access what.
I'm not necessarily an accountmanager.
I don't need access to the sametype of data that they do.
And then, finally, now we'vegotten to the point where we're
(24:50):
deploying agents, so differentmodels that have different
responsibilities and arefine-tuned to a specific task.
So when you start small andyou're able to build that solid
foundation that you can thenreuse and build upon.
Then you have the ability to dothis exponential growth, which
is what we've seen with Atom.
Exponential growth, which iswhat we've seen with Adam.
(25:11):
It was a very slow build at thebeginning and I think that was
something a lot of people whoweren't scientists at this point
in time that's the feedbacklist.
We've been working on WWT, gptfor so long, so long, so long,
so long, so long.
And just recently, have youseen that exponential growth of
like oh, we all have access, oh,there's new data services, oh,
now there's agents, and all ofthat in a short number of months
(25:32):
?
And you see the time that ittakes to add on those extra
features and build upon what youhave go down significantly,
because you started small andthen iterated upon that initial
solution.
Speaker 2 (25:45):
So step four is around implement security and
governance From a cloudperspective.
From a cloud perspective, zayd.
What specific skills doorganizations need for
governance when it comes tocloud?
What expectations should theyhave about what cloud's doing
for them versus where they needto take more responsibility when
it comes to security andgovernance?
Speaker 4 (26:08):
Governance isn't just a specific tool.
There are tools in cloud thathelp you with governance so that
you are following right toright permissions access and
you're labeling and tagging yourdata correctly, et cetera.
But governance, to me, I think,starts first and foremost with
education.
Speaker 2 (26:22):
Yeah, and I think and that's a big issue because, as
you guys both had mentionedearlier you know, one of the
challenges we're dealing with isthat we're all we're learning
what AI can do for us.
At the same time we're tryingto learn how to operate it, and
usually historically, thosethings have been really
separated.
We didn't have technology rollout so fast that we're using it,
while it's still changing asrapid as it is.
Speaker 3 (26:44):
My two cents on governance is get in early,
understand what is important toyour organization, whether it's
within the cloud, legal data,science, network space.
How are you running yourgovernance?
And then have every solutionadhere to that, Because you can
(27:06):
build models and products andsolutions in any type of way,
shape or form, but when giventhose guidelines, you are much
more likely to get a productthat adheres to it rather than
at the end being like oh, by theway, this needs to adhere to
this one policy that we didn'ttell you about ahead of time and
that might just be the exactkey piece of information that
(27:27):
the model is using to make itspredictions.
Speaker 4 (27:30):
Once you're educated and you've established what your
framework's going to be and howyou're going to implement
governance, then let's talkabout tools, because there's a
lot of tools inside.
You know, of course I'mspeaking from the cloud
perspective, but there's a lotof really cool tools to
implement governance, whetherit's at the data level for
labeling and tagging, likeDataplex in GCP or like
(27:51):
something like Assured Workloadsand again, I speak from the GCP
perspective.
But if you have a compliancestandard you need to hit or you
need to meet excuse me you canimplement that across your
entire cloud environment inGoogle Cloud by using something
called Assured Workloads, and sowhen you try to use a tool or
you try to access data orsomething like that that you
shouldn't be able to, itliterally flags you and stops
(28:12):
you.
So if you have a solidfoundation of how you want to
implement governance and whatstandards you need to meet,
there's a likelihood that thetool is there to help you follow
that.
Speaker 2 (28:29):
Do you think?
Is it logical to assume maybethis is too easy to answer but
the notion that I think we'regoing to have more regulations
and oversight that we can'tpredict now, and so if you say,
well, I'm safe doing thesethings now because no one's
watching or cares or something,and so planning it out and
making sure that you'readdressing this not based on
what the law or anyone else saysyou should do, but what you
(28:51):
know should be done, might saveyou a few steps in the future
and your ability to pivot whenthose things inevitably change,
perhaps, oh yeah, 100%.
Speaker 3 (28:59):
There's also a bunch of voluntary frameworks out
there right now, or even likeheads up like OWASP and NIST.
They put out a top 10vulnerabilities of LLMs.
So if you look at what thecurrent vulnerabilities are and
what the industry trends are,you can see where those
guidelines are then pointing.
(29:19):
At this point, within theUnited States, there is no AI
regulation.
That's very clear.
So you have to look a littlebit deeper into what the
industry is recommending andthat is not easy because there's
a lot of voices.
But paying attention to thoselike OWASP and NIST that have
(29:42):
been very strong in the securityrealm for a long time and are
now looking to include AI aspart of that.
And I think Cloud's a littlefurther along.
So tell me if I'm wrong, butit's still following.
It's still behind traditionalsecurity.
Speaker 4 (29:56):
Yeah, yeah, there are trends and publications that
come out all the time with youknow, recommendations about how
you can continue to secure yourworkloads and, specifically, how
you should be prepping for AI.
Speaker 2 (30:10):
Yeah, and then we generally see a lot of those get
adopted, because that becomesthe easier route, especially
because a lot of forwardthinkers are engaging with that
and they're going.
This one doesn't make sense,even though they're all
voluntary.
If we engage enough, then wecome out with something that is
reasonable and allows us all tostill hit our objectives without
being too onerous.
Yeah, yeah, you're right,they're probably going to come
(30:32):
from those pre-existingvoluntary frameworks.
Okay, so final questions.
What's the one thing that youwish that they knew more about
in terms of building AI-readyfoundations early on?
Who wants to take that first?
Speaker 4 (30:45):
Tina does I do I do.
Speaker 3 (30:48):
80% of your problem is a data problem.
Every AI problem 80% of itsdata.
So just expect that, becausethe first thing I'm going to
tell you is that you need tohandle your data, and being
surprised every single time whenyou have a new use case and
you're like but my data was fine, last night, we fixed it.
(31:09):
It's like we fixed it for thatuse case.
Every single AI problem is 80%a data problem.
Speaker 1 (31:16):
Okay, that was a great conversation and I want to
thank Rob, ina and Zaid fortheir insight on this important
issue facing any organizationlooking to scale their AI
initiatives.
A few things stood out to meabout the conversation.
First, your AI initiative willrise or fall on the quality and
accessibility of your data.
Prioritize projects for whichthe needed data already exists
(31:39):
and is well governed.
Second, bake security andgovernance into every step,
never as an afterthought.
Integrating security controls,role-based access and compliance
frameworks, early prevents dataleakage, model theft and costly
rework.
And third, start small, iteratefast and align architecture to
outcomes, not politics or hype.
(32:01):
Success comes from a clearoutcome, the right
high-performance infrastructureand a spiral, iterative roadmap
that expands only after smallwins are reached.
Bottom line.
The companies dominating withAI in 2025 won't be the ones who
move first.
They'll be the ones who movedright, who understood that
boring data governance andinfrastructure work is what
(32:22):
makes breakthrough AI possible.
That's it for this episode ofthe AI Proving Ground podcast.
A special thanks again to RobBoyd for hosting, brian Flavin
and Amy Riddle for their contentsupport.
If you're interested in moreepisodes of the AI Proving
Ground podcast, please considersubscribing on your favorite
podcast channel or check us outon WWTcom and we'll see you next
(32:44):
week.
Popular Podcasts
Crime Junkie
Does hearing about a true crime case always leave you scouring the internet for the truth behind the story? Dive into your next mystery with Crime Junkie. Every Monday, join your host Ashley Flowers as she unravels all the details of infamous and underreported true crime cases with her best friend Brit Prawat. From cold cases to missing persons and heroes in our community who seek justice, Crime Junkie is your destination for theories and stories you won’t hear anywhere else. Whether you're a seasoned true crime enthusiast or new to the genre, you'll find yourself on the edge of your seat awaiting a new episode every Monday. If you can never get enough true crime... Congratulations, you’ve found your people. Follow to join a community of Crime Junkies! Crime Junkie is presented by audiochuck Media Company.
24/7 News: The Latest
The latest news in 4 minutes updated every hour, every day.
Stuff You Should Know
If you've ever wanted to know about champagne, satanism, the Stonewall Uprising, chaos theory, LSD, El Nino, true crime and Rosa Parks, then look no further. Josh and Chuck have you covered.