June 5, 2025 • 40 mins
Welcome to Episode 403 of the Microsoft Cloud IT Pro Podcast where Ben and Scott catch up on some of their favorite announcements and news from Microsoft Build 2025.
Your support makes this show possible! Please consider becoming a premium member for access to live shows and more. Check out our membership options.
Show Notes
Introducing Microsoft 365 Copilot Tuning
New Microsoft 365 Copilot Tuning | Create fine-tuned models to write like you do
Behind the Curtain: A white-collar bloodbath
Announcing Microsoft Entra Agent ID: Secure and manage your AI agents
Enterprise-grade controls for AI apps and agents built with Azure AI Foundry and Copilot Studio
Quickstart: View enterprise applications
BRK195: Inside Azure innovations with Mark Russinovich
Microsoft Build 2025 - BOOK OF NEWS
About the sponsors
Would you like to become the irreplaceable Microsoft 365 resource for your organization? Let us know!
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
(00:03):
Welcome to episode 403
of the Microsoft Cloud IT Pro podcast
recorded live 05/30/2025.
This is a show about Microsoft three sixty
five and Azure from the perspective of IT
pros and end users, where we discuss a
topic or recent news and how it relates
to you. Microsoft build has come and gone.
(00:24):
So today, we dive into some of the
announcements from the conference.
We'll catch you up on announcements around Microsoft
three sixty five Copilot tuning,
securing Copilot agents and enter ID, and a
can't miss build session with Mark Russinovich and
more. So let's dive into the show.
(00:44):
It's been a week, Scott, for both of
us. Time marches on. I I have not
told you about my latest adventure. Your latest
adventure? Not your latest toy, but your latest
adventure? My latest adventure. Scott story time. Scott
story time. So my latest adventure was
my oldest graduated high school Congratulations. In the
last couple weeks. Yeah. Thanks. Accomplishment. We will
see. We need to get him out of
(01:04):
the house still. As part of that, we
threw him a graduation party. So we had
people over to the house, and Yep. It's
hot outside, but you can still hang out
outside if you're in the shade.
So we bought these
pop up tents, like, 12 by 12 tents
without sides. They just had a roof on
them. Right. Like, the whole canopy thing, yeah,
you see them at all the sporting event.
Yeah. All that. Simple basic. We we we
(01:25):
found some at a a bargain bin store.
Like,
there's a store in Jacksonville where this stuff
just comes in on crates, and then they
just mark it down to clear it all
out. So it was whatever. They're they're normally
a hundred $50,
and so they were on sale for $60.70.
I don't know. My wife went and grabbed
a couple of them. I will say this
is one thing I like about Florida because
(01:47):
nobody has basements, nobody has places to store
anything, so everybody gets rid of stuff if
they don't need it. And I feel like
because of that, you can find good deals
and cheap stuff. Well, this was even brand
new because this is, like, a wholesaler.
So they're buying what's kinda like Amazon return
bin kinda thing. They're buying a big
40 foot shipping crate. They don't know what's
in it, and then they're just bringing it
into their store, and they're trying to clear
(02:07):
it out and sell it. So so they
must have had a pallet of these These
canopies. Ready to go. Yeah. These canopies. So
we bought a couple of them and stood
them up for the day of the party,
and it ended up raining.
And so
we ended up inside for a little bit,
in, out, in, out. But I left them
up after the party. So this was last
week, and then we went through this week,
(02:28):
and
we've had a couple good rainstorms
and squalls and things like that. So they're
still up? You left them up? I did.
I left them up. This wasn't, you know,
hurricane stuff. I wasn't scared of something, like,
blowing away and landing on a neighbor's house,
things like that. So day one, they went
up, party. It rained. It was it was
fine. Like, all the water shed off the
(02:50):
roof. Like, it was taut enough and and
tight enough that
all that stuff happened.
At some point, it must have just loosened
up or some of the some of the
Velcro must have slipped or something that was
holding that canopy onto
onto the legs. So I went out there,
one day after one of the rainstorms,
and one of them just had this big
(03:10):
bubble in a corner where it was all
just leaning down.
And we've got the lawn furniture and stuff
underneath it, like, the patio furniture, and it
was leaning down about to my head. So
this is maybe a tent that's about eight
feet high, and
I'm I'm five seven. So it's coming down,
like, two, two and a half feet and
just about touching my head with this big
bubble. And I'm like, oh, that's no good.
(03:31):
So I'm in there and pushing it up
and trying to get the water off it.
And I take a look, and I'm like,
oh, it came loose on on the side
because the Velcro slipped. So I kinda, like,
redid the Velcro. I went and pulled it
all back down and tightened it back up.
Fine. Great. Two more days go by. Rain.
Rain. Okay.
Yesterday,
we had some massive kinda squalls come through,
(03:54):
like sideways rain. I saw that. It was
a little nutty. It it came down for
basically
two hours straight sideways.
I was watching a little river form between
my house and and my neighbor's house because
there's kind of a natural divot there, and
and that leads down to the preserve behind
my house.
And so I'm I'm looking. I'm looking. So
(04:15):
I go outside, and I just I just
open the door. I'm not I'm not going
out in this weather, and I peek my
head over, and there's not just one bubble
coming down. Now there's two bubbles, and the
whole thing is, like, leaning, leaning, leaning, leaning.
And it's coming down, and it's almost touching
the patio furniture. Oh, man. Patio furniture's
only, what, three feet maybe? Right. Considering, like,
the top of the back of the chair,
three, four feet. So it's coming, like, just
(04:36):
down, down, down. I'm, like, I'm telling my
wife, I said, this might be it. Like,
it it it's it's it's probably gonna go.
Absolutely did. So it got so heavy with
the rain
on with those two bubbles on the one
side, it actually tore
the the entire top. Like, it just sheared
itself, like, right off. And then when it
sheared itself with the weight, it also bent
(04:57):
all all the metal and things because this
is like a big accordion. So now this
thing will not go back together. It's probably
the best $70 I've spent for, like, a
week of just being able to sit outside
and things like that. I was kinda telling
my wife. It's like, you know what? Worth
it. Maybe I should just go back and
see if we can buy a couple more.
Like, we'll we'll just keep them on the
side of the house in in stock and
(05:17):
and and ready to go for the next
one. So that that that's my next adventure
is after I'm off this call, I'm gonna
go grab a hacksaw, and I'm gonna go
out in the backyard. And now I gotta
cut all these metal spider legs off this
12 by 12 monstrosity thing because there's no
way it's going back together with how bent
it is and everything else. And then because
it sat there over my patio furniture so
(05:39):
normally in my patio furniture, it sits outside
and we leave the cushions on sometimes, and
it doesn't get soaked soaked. It's got, like,
Scotchgard or whatever on it. The water all
runs off. Uh-huh. Because this sat on it
for so long, it soaked through the material
on the tent, and it soaked into all
the patio cushions. So when I picked the
patio cushions up and I just turned them
from horizontal to vertical, they've just been leaking
(05:59):
out for the past hour. Like, yep. Still
watching them drip. I will not lie. I
thought for sure one of those tents was
gonna, like, take off and blow away. I
thought that's what was gonna happen. Because when
it starts raining sideways, I've seen I think
it was actually a hurricane. I've seen a
trampoline, like, blow from one house into the
neighbor's yard in some of the wind that
we've had. It had a great, like, vent
(06:20):
on the top that Okay. Would crosscut, so
it really wasn't gonna take off or go
up and down. And we had it all
tied down and everything. I I didn't use,
like, all the hurricane ties, but it it
was definitely, like, attached to the house and
and where it needed to be. So fun
fun little adventure for the week. Wow. Way
more adventurous than my week. I don't think
I have any fun adventure stories. So between
(06:41):
that and and my new terminal toy, which
we'll we'll talk about later. We'll we'll do
a toy episode and and annoy folks
sometime in the future. But, anyway, that's my
news for the week. Why don't we get
into the news from the last couple weeks
because we've had events like build
and all sorts of other stuff. Yeah. And
some of those events, they're definitely developer heavy.
(07:03):
Like, it's all different things. But there's
always news, I feel like, that relates to
IT pros because us as IT pros have
to manage our developers and keep them in
line and make sure they're behaving. And as
a result, we get different things, and let's
face it, there's Copilot stuff everywhere, so there's
always Copilot news to talk about as well.
And we haven't done much of a news
(07:24):
episode, so some of these I think everything
we kind of put together was the last
couple weeks. Some of it may have been
a little before build, all around the build
time frame. We really should do these more
often. Although, I don't feel like there's been
a ton of news either, or I've just
missed it. There's been quite a bit. I
think it depends It's kinda what you're looking
for. Where you sit and where you play
(07:44):
around. Yeah. Yeah. There's been a bunch of
new stuff come out across
M 365.
I was looking the other day
at some of the
latest announcements that have come out for
Azure Kubernetes service and and some of the
things going on in AKS land.
Tons going on over there.
(08:05):
The the trick is kinda sifting through the
AI, AI, AI to what's what's the what's
the real thing, or if AI was your
thing, then, hey, maybe there was something real
for you there too. Like, I definitely saw
some stuff with Copilot
that piqued my interest, and and maybe we
could start there. So
(08:26):
last week, we we talked to Do you
wanna start with that one? Yeah. So so
last week, we chatted with AEC
about declarative agents, and that was kind of
a follow-up to some stuff that we had
talked about previously with Copilot Studio and doing
kind of
well, for you and I, no code declarative
agents, just in Copilot Studio,
next, next, next in context of an m
(08:48):
$3.65
subscription, things like that.
They're introducing
fine tuning for Copilot,
which kinda takes this a whole another level
because some of the at at least from
my perspective and the things that I do
in my day to day, because what I
do today
in something like Copilot Studio with declarative agents
is I'm relying on RAG and and retrieval
(09:10):
augmented generation.
And this is really being able to take
a model that already exists and the weights
in that model that are already deployed out
there and exist, and then fine tune your
own model on top of it
or fine tune and refine that model so
that you can drive
specific business processes.
(09:30):
And the demo I saw around this really
got my head turning. So I'm a product
manager.
I live in a land where
I'm a I'm a remote employee.
Communication and particularly written communication is super important.
So we spend a lot of time on
product requirements documents, PRDs,
justifications,
specifications,
(09:51):
designs,
and we have formal templates for all this
stuff. One of my big problems with it
is
people who take the template and they massage
it into their own thing or they don't
know what a good one is. So they
they kinda look at the template that we
have today, and and they just get like
a deer in headlights and and scared about
it.
So I like, with this, I could take
(10:12):
all our existing
PRDs that we've done for, like, at least
in my time for the past five years
on my team.
I could train
and fine tune
the existing model on my PRDs,
and I could actually have it tune on
top of my templates as well.
So then I could make it super turnkey
(10:34):
to go in and turn these things out.
And that's a very simple scenario for me.
I saw a great demo on the Microsoft
Mechanics YouTube channel, and I'll I'll put a
link in the show notes, and I recommend
folks go out and watch that one. That
was generating legal documents,
and it actually did a pretty good job.
Like, I'm not a lawyer, all all that
jazz. Yep. It's just to couch all that
(10:56):
and and have that caveat.
But that said, I I think that one's
really cool and and really exciting, and it's
something I'll have access to in my tenant
and totally intend to go out and play
around with. So I don't know if you
had a chance to either play with this
in your tenant or watch any of the
watch any of the introduction videos or anything
like that that's out there, but we would
(11:18):
would love to get your thoughts on this
one as well. I did watch the Microsoft
Mechanics video while I was doing other stuff,
so I was kinda half watching it. This
is a really cool feature. I like how
you can fine tune it that way. I
would say the part that I was interested
in, kinda coming at this from
the security model is in my
head too when you see this announcement. You're
(11:39):
like, okay. You're now fine tuning
an LLM.
What does this mean for
data security? What does this mean for how
is my data now being used if I'm
fine tuning
an LLM? Is it going into the LLM?
Is it training the LLM? And they talked
about this in the Microsoft Mechanics video, and,
ironically, I don't I probably don't have the
(12:00):
business use cases for this yet, although I'm
starting to play with it more. But it's
more thinking about this from a client perspective
of questions they're gonna ask about their data,
their data security. Some of the stuff we
even talked about a couple weeks ago when
I was talking about how a lot of
clients are coming to me now and saying
we wanna prep for Copilot, we wanna make
sure data is secure.
(12:21):
Now what does fine tuning this mean? And
they did a pretty good job, I would
say, in the Microsoft Mechanics video from that
perspective of, like, you still have your LLM,
and it's not and I I could not
explain this from a deep technical level on
how it actually works, but they actually take
your data. And the way they illustrated it
is they, like, attach it to the LLM
for the fine tuning. It's not like your
(12:43):
data's gonna go into the LLM and they're
retraining the core LLM. It's like something between
RAG and actually having your own model, where
it's a model with your data attached to
it to do the fine tuning. And it
sounded like from that video that it's not
even always attached to the LLM. Like, it
attaches to the LLM to do the fine
tuning, and then it kinda goes and stands
(13:05):
separate from it after you've done the fine
tuning. It only reattaches when you do it,
and it still maintains, like, your data and
your own tenancy.
They're not taking your data to train the
core LLM.
All your access lists and everything, that is
all still maintained.
So it was that that part to me
was interesting to try to understand a little
(13:25):
bit more from that perspective
on how was it doing the fine tuning
of the model without actually
absorbing the data into the model? The way
you can think about it is
fine tuning,
and and that's really what we're talking about
here is is kinda
taking an existing model that's already been trained
Yep. Let's say chat GPT four o, and
(13:49):
then you're gonna take that, and that that
that's a stage in that model's life cycle,
and you can use it as it is
with the weights that exist and and and
all of that stuff. Now what you can
then do so you do all that, and
and that's kinda like training slash pre training.
And then you take that pre trained model
and you fine tune it. And fine tuning
(14:09):
is a refinement process. So you're taking this
very it doesn't have to be very small,
but you're taking a smaller
task specific
dataset. So, like, in my example,
PRDs and and justifications. Templates and yeah. You're
doing this to
optimize
performance
(14:29):
and
and and drive a business
process.
So, like, you you you're right. You're not
loading this stuff into the original
pretrained model. You're kind of creating your little
bit of, like, an your own offshoot of
a model, but you don't have to go
back and redo all the training and redo
all the weights for the original model. So
(14:51):
you've kind of augmented it with this set
of sidecar weights that can then be used
as well to
go ahead and make your responses better.
Now the the the purpose of this and
and the difference
versus,
say, something like rag
is
RAG has to go out and you you
(15:13):
have to chunk all these documents, you have
to put them in a vector database,
and you have to compute and and and
run all that stuff every time. This is
just having that set of
kinda fine tuned weights out there for you
and ready to go. So it's it's it's
a click stop if you think about it
as a series of stages. So pre training
is all about, let me get the model
(15:34):
out there, and that's
general representation
of whatever it was trained on. And then
fine tuning is let me take and
take this model to the next level by
really augmenting it in a way I shouldn't
say augment because of the whole rag thing,
by
adapting it. Yeah. Fine tuning it. We're fine
tuning it. Right?
(15:55):
We're fine tuning it. So you're going but
you're going from general representations
to specific tasks or Right. Knowledge domains.
Processes. Yeah. Processes. So you could do this
multiple times. So when you hear about maybe
a company that's working on
AI to solve cancer, well, guess what? They
did the same thing. Right? They they took
that general model that was out there, and
(16:18):
then they took a bunch of cancer research
that had already been done previously, and then
they fine tuned that model so that they
could have a purpose built model that they
could go and and work with
to
attack cancer, attack your
in your case, maybe your business processes
around statements of work and how you put
(16:38):
those together for clients. I should try one
of those. You've done that a whole bunch
throughout the years with things like document libraries
and document sets, and this is I I
I think you should give it a shot,
like, if you're licensed for it because it's
a little bit of a next level thing
for you. I should absolutely go try that.
I I didn't statements of work didn't click
in my head until we were just talking
about it. I'm like, yeah, master service agreements.
(17:00):
And maybe it's because I don't put a
ton in there. Statements of work, I put
a little bit more in there. Master service
agreements are a lot of addresses
and that. But I have had clients ask
can't even remember which client it was. Where
they did, they're like, well, if I have
a template of a document and I want
to use AI to fill out the template,
can you do it? And I'm like, well,
(17:21):
not really.
The way it was
two months ago even or before they talked
about this fine tuning, it's like, doesn't really
do a great job at that. And this
is where
now I think as clients get this, again,
licensed for it, willing to build it out,
it's a lot more gets you a lot
closer to being able to have templates that
you can auto fill out. Not auto fill
(17:42):
out, but you still as a user
can yeah. Yeah. So so you can generate
those
and you as a user providing inputs. So
the the way I think about it and
the way I see a lot of folks
use generative
AI today
is they kind of look at it and
and I've definitely seen,
(18:04):
particularly, like, my PRD example, I've seen some
PMs that go out there, and they just
say, hey. Write me the document.
And this isn't, hey. Write me the document.
This is write me help me write the
document so that it conforms to this process,
this knowledge domain,
any anything like that, you know, that exists
out there. So, yeah, it's it's a tool
(18:26):
in the toolbox
versus a do my job for me kinda
thing. So if you take an agent,
and I don't know that they talked about
this, and I don't know if you've seen
it, Could you potentially take like, let's say
you go do the co the fine tuning,
and, really, you're creating an agent. So that's
the other aspect about this. This isn't fine
tuning like your general
(18:47):
Microsoft three sixty five Copilot responses. You're creating
an agent that's fine tuned
to do
a specific task.
Can you combine
agents with actions yet where you could do
a fine tuned agent going back to the
proposals or statements of work where you're then
using an action
to prompt or to get certain information
(19:08):
to feed into a fine tuned agent to
generate these particular outputs? I haven't tried to
chain them together like that yet. I I
I think that's where all this stuff goes.
So if you think about
MCP servers and the whole model context protocol
thing,
that's where these agentic things are going. And
you see a little bit of this today
in
(19:29):
in maybe in the declarative agents that we
already talked about where there's a connector for
SAP, there's a connector for Workday,
things like that. Like, you have this agentic
system on the side that it can reach
out to and sidecar that knowledge in from
that other system
and bring it in. I always think it's
super funny, at least in today's world where
we're still
(19:50):
very early days in this stuff, when
you you see the funny two agents talking
to each other kinda thing. I don't I
don't know if you've seen the video. It
came out a couple months ago, but it
was like a robo calling agent called another
robo caller. And then they both figured out
that they were robo callers, so they just
basically started talking to themselves in it sounded
(20:10):
like an old, like, bot modem going off.
Like I think I did see that. Yes.
Back and forth because they knew there was
no human there and, hey. Like like, let's
go down a different path kinda thing. So
they're still a little wild and off the
rails. But, ultimately, I think that's where these
things do go. So if I think about
an
agentic chain to build a PRD, well, great.
(20:31):
So I've got the fine tuned
process where I've come in and said, hey.
Help me write this within these guard re
these guardrails, or help me generate this thing
within these guardrails and and these boundaries.
But part of my PRDs would be customer
stories.
Well, maybe I wanna go build a little
agent that does nothing other than reason over
(20:52):
my customer feedback.
And then there's another agent that I might
wanna go build that reasons over
the financial aspects of my business and and
the cogs of it and and how that
comes together. So all these little things have
their own
click stops along the way, and, eventually, I
think, yeah, that that's where you end up
is potentially a bunch of agents just chatting
(21:12):
back and forth with each other, and one
throws this over here, and then it throws
it back this, then it throws it into
a new one, and out the other end
comes something else. And I I think that's
the thing that we we all haven't, like,
really figured out as Right. I saw the
I think it was the CEO of Claude
this week, last week. He was doing an
interview, and he was talking about basically, like,
(21:34):
50%
of it was some crazy number, 20%, fifty
%. It was really high of kinda low
end knowledge worker jobs going away because of
AI and and Gen AI. So nobody knows
how this stuff's gonna go and where it
comes out. I'm trying not to be, like,
the doom and gloom person about it. Like,
I very much do look to this stuff
(21:55):
to augment my job today
and help me be faster? Like, I don't
know. IT Pro land, all we've ever been
doing is trying to automate ourselves out of
jobs. This is just another step in that
journey is the way I kinda look at
it. And then just go watch Mission Impossible.
Have you seen the latest Mission Impossible yet?
I haven't. Don't spoil it for me. I
will not spoil it for you, but all
(22:16):
I could think about was AI and Copilot
during Mission Impossible.
Do you feel overwhelmed by trying to manage
your Office three sixty five environment? Are you
facing unexpected issues that disrupt your company's productivity?
Intelligink is here to help. Much like you
take your car to the mechanic that has
(22:36):
specialized knowledge on how to best keep your
car running, Intelligink helps you with your Microsoft
cloud environment because that's their expertise. Intelligent keeps
up with the latest updates in the Microsoft
cloud to help keep your business running smoothly
and ahead of the curve. Whether you are
a small organization with just a few users
up to an organization of several thousand employees,
(22:57):
They want to partner with you to implement
and administer
your Microsoft cloud technology.
Visit them at inteliginc.com/podcast.
That's intelligink.com/podcast
for more information or to schedule a thirty
minute call to get started with them today.
(23:18):
Remember, Intelligink focuses on the Microsoft cloud so
you can focus on your business.
Alright. So Copilot Tuning, folks can go out
and sign up for that. I believe you
have to sign up for a preview for
that one or or kinda put yourself Yeah.
This one's still it's early access. About the
Copilot Tuning early access,
(23:38):
customers with more than so this is a
high barrier to entry, so I probably cannot
do this yet either. But it does say
early access, you have to have more than
5,000
Microsoft three 60 five Copilot seats to participate,
EAP candidates undergo
screening, make sure they have scenarios that align
with it. Like, Microsoft is at that point,
they're sounds like they really wanna learn from
(24:00):
this before they just throw it out there
to GA for everyone. This is the same
thing you saw with Copilot before. So
this shouldn't be a model that surprises folks
at this point. They have a couple
people that are already using it. Ernst and
Young, McCarthy,
I don't even know how to pronounce that
word, Tyrolt,
Land O'Lakes.
They have some big companies using it. So
(24:22):
if you're a large customer and wanna go
play with us, early access program, if you're
not, be patient. You'll get it eventually. Don't
let that stop you from going out and
watching some videos. Like, the mechanics video is,
like, ten, twelve minutes. Totally worth your time,
I think. This looks like a build video.
I haven't clicked on this one. This one's
probably, like, an hour long. Oh, no. This
one's ten minutes. There's an introduction video too
(24:43):
on the article, so definitely go out and
learn from it. Start building agents.
That's I've been playing more and more with
building some agents as well, which that leads
to another news article. This one's
one that caught my eye is
you now have
Microsoft
Entra Agent
ID, Secure and Manager AI agents. So this
(25:05):
one was, I think, right before
Build, and I saw Merrill posted something about
this. I saw that. I saw a few
blog articles about it. Alex Simmons has a
announcement on the Entra blog about it. But
this was interesting to me because, again, we've
been talking about we're building these agents. We're
giving them access to
(25:26):
SharePoint sites, to different data.
And this is starting to look at, so
how do we actually go in
and start securing those agents, securing what those
agents can do?
And now if
you're an admin and you're a tenant, you
can go out to Entra, You can actually
go out
to Entra
and go to your enterprise applications.
(25:48):
And if you look at application
types,
you now have an agent
ID
type in your
enterprise applications.
And this is this is cool
just from a pure
reporting perspective right now, and that it will
give you a list of every single agent
that has been created in your environment
(26:10):
by any users,
whether it's an agent created in Copilot Studio
or Azure AI Foundry that are registered a
part of your Microsoft three sixty five tenant.
This is one that there's some videos out
there. I went and did it. My tenant
saw all the agents there.
Great for reporting perspective
for right now. I would say from, like,
(26:31):
a security perspective,
there's not as much you can do, but
there's a lot of stuff coming. So they
have a whole part of this that's what
what's next with these EntraID agents now that
they're in EntraID, now that they're showing up
as enterprise applications.
It means that you can start going in
and looking at, like, eventually what Graph API
access they have or
(26:53):
even creating conditional access policies for your agent.
So now someone can interact with this agent.
Being able to go in and do
granular conditional access and all that detailed permissioning
that comes from that perspective
because,
I mean, I mean, we're getting to the
point with these things, like, going back a
little bit to what we talked about earlier
with
(27:13):
agentic workflows and these things effectively talking to
each other. Like, you almost are treating them
as individuals within your business. So it makes
sense from that perspective
to secure them in in the same way
that you would your users and other things
that are out there. So this will give
you additional operational controls
and, frankly, like, comfort food around locking some
(27:36):
of this stuff down when it comes to
what do they access, when do they access
it, how do they access it.
The the the I think the granular graph
stuff will be great. Like, hey, making sure
this thing only has
a read permission versus a read all, or
does it does it have the ability to
actually
generate content in that environment? Like, oh, no,
(27:57):
I wanna I wanna block rights for this
one because I don't actually want it to
go in and I don't know. Somebody goes
and builds the agent to rename titles in
the GAL.
You you you might not want that that
that thing
running that way. Yep. It's also good just
from a
visibility perspective for these things. Like, it's another
way to kind of track adoption in your
(28:19):
organization, maybe go talk to some users,
see where they're getting
ROI out of out of these kinds of
things,
and better manage the life cycle of them.
I still see a lot of places where
you go in and you look at their
their intro tenant, and it's just the graveyard
of lost stuff that that has not been
(28:40):
cleaned up over the years and hasn't been
attributed to a user. Hey. What's this old
enterprise app ID doing out here that has,
like, basically, god permissions in your environment? Well,
like, what did you do, and why was
this a deity? Oh, that was created way
back when five years ago when we spun
up our tenant. Great. Let's clean it up
today. Right. And I think some of those
(29:00):
two going back to life cycle,
above and beyond this announcement, permissions change too
with Graph API. Right? Like, there's old applications
out there that have full control over your
entire SharePoint environment
because
granular
site level API access in the graph wasn't
there when these apps were originally created. So
(29:20):
as things continue
to evolve, permissions continue to evolve over time,
being able to,
see that life cycle, see what they're doing,
see what they have access to, some of
the auditing and monitoring you'll be able to
do around
logs and visibilities into what these agents are
doing, what they're accessing.
(29:40):
You've even seen that some with Copilot and
some of the new stuff around,
sensitivity labels and being able to do DLP
and audit user interactions
with Copilot,
now being able to audit agent interactions with
stuff. Enterprises, this is I feel like it's
something you have to have when you start
getting into AI. If there's legal questions that
(30:02):
come up, questions that come up about how
something was generated, being able to dive into
these logs and see this does become super
important. And this is one that's out there
today. Like, folks can go play with this
one. Yes. You can see them. Like, they
have a whole list. Again, I could see
it in mine. It's labeled as preview, so
all the preview stuff, it's in preview.
I'd say don't use it in production, but
it's just a preview feature that shows up
(30:23):
in production.
So don't put production workloads with it, but
then a lot of this other stuff that
we talked about, the conditional access, the auditing,
life cycle management,
in this blog article from May 19, it
says,
this is all in the coming months. A
lot of these capabilities
around
zero trust security posture,
(30:45):
all of that is going to be rolling
out to these
agent IDs in Entra. So that's one I'm
looking forward to playing with, keeping an eye
on. What other one you wanna talk about?
We maybe have time. One more announcement. Why
don't we do the Russinovich session? So he
he did a
app build. Mark Russinovich did a session on
(31:07):
inside Azure innovations.
I don't I don't know if you had
a chance to go and watch that one.
I failed on my homework to watch that.
So
I am curious for you to fill me
in on that one. I'll put a link
in the show notes for everybody, but that's,
b r k
one ninety five.
Slides are out there, recordings out there,
(31:28):
all all that good kind of stuff, so
folks can go and take a look at
it. But, effectively, like, this talks about
the internals
of Microsoft and Azure infrastructure.
So what's the latest in boost and DPUs?
What's going on with FPGAs?
How are these things being
(31:49):
leveraged within the
within the Azure fabric.
One of the cool things
that was in there, or at least I
think it was cool more because, like,
some folks on my team were
were were involved in it,
is there's a demo in there where he
talks about blob storage and scaled accounts. And
(32:11):
so one of the things that we've been
doing in storage over the years is trying
to enable these large
AI training companies
who are doing these large pre training runs,
so OpenAI, so the world, things like that.
And
as these customers are bringing in petabytes
and petabytes and petabytes of data, it's not
like single digit, It's it's double to triple
(32:33):
digit
to exabyte scale because you got, like, the
world's knowledge in there. Like, downloading the New
York Times archive is a pretty big thing.
Do that across a bunch of
different sessions that exist out there, and
it it gets quite a bit bigger. So
we did a demo,
and we've been on this journey with scaled
accounts to ever increase the amount of throughput
(32:55):
that we can give them. So ingress, egress,
the amount of IOPS we can give them
in the form of TPS, all that. So
we did a demo where
we were able to, on a single storage
account,
provision
a workload
that ran
at 15 plus terabits a second of ingress,
25
plus terabits a second of egress, and this
(33:18):
is all random
random IO through
through an HPC benchmarking
suite called
IIOR. Okay. We were able to just run
this for hours.
We didn't do this in a special place,
like, just public fleet.
This this is stuff our our customers
can go in and do today.
(33:39):
So so I thought that was a super
cool one. And like I said, important to
me because while my team doesn't work on,
like, the perf and scale stuff, one of
the key components we work on is AI
and ML clients.
So I own our my team,
we we do our SDKs, client tools.
BlobFuse is one of those. So BlobFuse was
heavily used in that demo and ready to
(33:59):
go. So we've done a bunch of work
there just to be able to, like, spin
up workload
in BlobFuse and have it be able to
take, like, full advantage
of
a VM's NIC and soak it all the
way through. So I I would recommend everybody
go watch that session. Like, if you're interested
in the
either the internals of Azure or even a
little bit more about the internals of of
(34:21):
the way some of these workloads compose. So
there was some stuff in there about Linux
Guard and kind of how code integrity is
handled
within Linux, some cool demos for that one.
There was a bunch of stuff around
virtual machines
and kinda how we do host OS upgrades,
(34:42):
hot patch, driver swaps,
full reboots,
and and how those things go within the
within the infrastructure. There was some insights into
how we do RDMA
on the networking side, like, all sorts of
cool stuff. And because it's a Recenavit session,
it's also demo driven, so you actually get
to, like, see it in action. It's not
(35:02):
just a bunch of slideware. Yeah. I'm flipping
through
the slides here, and it looks like there
were, like, 10 or 12 demos in there.
And it is like, I always try to
go watch Russinovich's.
It has been a time constraint more than
lack of desire to watch his yet. But,
yeah, all this stuff, there's some stuff around
that. A confidential
compute,
a bunch of stuff around Azure confidential
(35:24):
GPU VMs.
Yeah. His are always,
like you said, super
demo driven, but I would also say super
nerdy.
Like, for
super nerdy or super technical. Maybe I should
be politically correct. Super technical, like, the amount
of detail that are in his sessions and
the amount of I don't know if he
just sits there and practices all these to
(35:46):
get all these technical details down or if
he just has this much technical knowledge floating
around us in his head. Either way, I
am always super impressed with the
level of technical depth that he's able to
go to in these discussions and these presentations.
So I always find them fascinating. They're great
to go watch.
Mark is
(36:07):
super deep in the platform
and and does understand a bunch of this
stuff. Like, he's not out there just
spouting it off based on a script that
I think somebody else wrote for him. So,
yeah, I I encourage everybody
to to to go and watch that one.
It's it's it's worth an hour if you're
into Azure in general, and you and you're
(36:28):
kinda looking for what's the latest in hardware
innovations there across really all the core services,
so compute, networking, storage. Yep. And if you're
only interested in one or two of those,
skip around the session. You can watch it
on YouTube, and it's got chapters and everything
in there for you. What is this? The
world's first analog optical computer
for accelerating
(36:48):
AI
inference.
Potentially a hundred times more energy efficient than
GPUs.
If that is at all intriguing to you
about light doing massive parallel computations
for AI, go watch the end of this
session. Right? Go watch the end of it.
So power consumption is important. Sustainability
is important. Yep. I saw and and
(37:08):
don't quote me exactly again on the percentage,
but it's something like three to 4%
of the world's power now is going to
data centers,
and it's been driven on this general uptick
around
more and more CPU and particularly more and
more GPU usage because these things are so
power hungry. So
I I think it is important that all
(37:29):
these companies that are working towards this stuff
continue to strive towards sustainability
and innovations
in
not only, like, capping power consumption, like, at
some point, it can't keep growing forever,
but also bringing it back down. So things
like like analog
optical computer AOC, not not the other AOC,
(37:50):
this AOC
fall into that category of let's make the
world a better place. Very cool. Well, those
are some fun announcements,
interesting news. We spent, like, the whole time
in, like, three news articles. You should see
the number of blog posts I have to
go through. I think we went through
a a couple hundred of them that are
out there. So the other thing that I'll
put in the show notes for folks,
(38:11):
the the book of news
is out and available as well for that,
and and that's publicly available. So if you're
interested in maybe the more
broad roundup of, hey, what were the bigger
announcements
at Microsoft Build, things like that, I would
encourage folks to go out and
take a look at the book of news.
(38:32):
Maybe
throw it into your favorite LLM and have
it summarize it for you. Absolutely. And if
you have any questions,
topics you want us to talk about,
news you want us to cover in future
episodes as well, let us know. I would
say either if you're part of the membership,
let us know in Discord. If you're not,
LinkedIn,
I'm quickly leaning towards as being the social
(38:54):
media social media probably social media, social network
that I am spending the most time on.
All the other ones
Yeah. I feel very hit or miss. LinkedIn
is probably the best place to get both
Scott and I and get our attention,
if you have topics or questions about the
show. You can tag the podcast there. You
can also go to the website and just
use our contact
(39:16):
contact
contact us form.
It's alright. It's Friday. It's time to
call it a day at 11:00 in the
morning and go relax for the rest of
the day or go to meetings. I'll probably
do the former. But as always, thanks, Ben.
Appreciate the conversation. Thank you, and enjoy your
weekend. We'll talk to you next time. Yep.
(39:36):
If you enjoyed the podcast, go leave us
a five star rating in iTunes. It helps
to get the word out so more IT
pros can learn about Office three sixty five
and Azure.
If you have any questions you want us
to address on the show, or feedback about
the show, feel free to reach out via
our website, Twitter, or Facebook.
Thanks again for listening, and have a great
(39:56):
day.
Welcome to episode 403
of the Microsoft Cloud IT Pro podcast
recorded live 05/30/2025.
This is a show about Microsoft three sixty
five and Azure from the perspective of IT
pros and end users, where we discuss a
topic or recent news and how it relates
to you. Microsoft build has come and gone.
(00:24):
So today, we dive into some of the
announcements from the conference.
We'll catch you up on announcements around Microsoft
three sixty five Copilot tuning,
securing Copilot agents and enter ID, and a
can't miss build session with Mark Russinovich and
more. So let's dive into the show.
(00:44):
It's been a week, Scott, for both of
us. Time marches on. I I have not
told you about my latest adventure. Your latest
adventure? Not your latest toy, but your latest
adventure? My latest adventure. Scott story time. Scott
story time. So my latest adventure was
my oldest graduated high school Congratulations. In the
last couple weeks. Yeah. Thanks. Accomplishment. We will
see. We need to get him out of
(01:04):
the house still. As part of that, we
threw him a graduation party. So we had
people over to the house, and Yep. It's
hot outside, but you can still hang out
outside if you're in the shade.
So we bought these
pop up tents, like, 12 by 12 tents
without sides. They just had a roof on
them. Right. Like, the whole canopy thing, yeah,
you see them at all the sporting event.
Yeah. All that. Simple basic. We we we
(01:25):
found some at a a bargain bin store.
Like,
there's a store in Jacksonville where this stuff
just comes in on crates, and then they
just mark it down to clear it all
out. So it was whatever. They're they're normally
a hundred $50,
and so they were on sale for $60.70.
I don't know. My wife went and grabbed
a couple of them. I will say this
is one thing I like about Florida because
(01:47):
nobody has basements, nobody has places to store
anything, so everybody gets rid of stuff if
they don't need it. And I feel like
because of that, you can find good deals
and cheap stuff. Well, this was even brand
new because this is, like, a wholesaler.
So they're buying what's kinda like Amazon return
bin kinda thing. They're buying a big
40 foot shipping crate. They don't know what's
in it, and then they're just bringing it
into their store, and they're trying to clear
(02:07):
it out and sell it. So so they
must have had a pallet of these These
canopies. Ready to go. Yeah. These canopies. So
we bought a couple of them and stood
them up for the day of the party,
and it ended up raining.
And so
we ended up inside for a little bit,
in, out, in, out. But I left them
up after the party. So this was last
week, and then we went through this week,
(02:28):
and
we've had a couple good rainstorms
and squalls and things like that. So they're
still up? You left them up? I did.
I left them up. This wasn't, you know,
hurricane stuff. I wasn't scared of something, like,
blowing away and landing on a neighbor's house,
things like that. So day one, they went
up, party. It rained. It was it was
fine. Like, all the water shed off the
(02:50):
roof. Like, it was taut enough and and
tight enough that
all that stuff happened.
At some point, it must have just loosened
up or some of the some of the
Velcro must have slipped or something that was
holding that canopy onto
onto the legs. So I went out there,
one day after one of the rainstorms,
and one of them just had this big
(03:10):
bubble in a corner where it was all
just leaning down.
And we've got the lawn furniture and stuff
underneath it, like, the patio furniture, and it
was leaning down about to my head. So
this is maybe a tent that's about eight
feet high, and
I'm I'm five seven. So it's coming down,
like, two, two and a half feet and
just about touching my head with this big
bubble. And I'm like, oh, that's no good.
(03:31):
So I'm in there and pushing it up
and trying to get the water off it.
And I take a look, and I'm like,
oh, it came loose on on the side
because the Velcro slipped. So I kinda, like,
redid the Velcro. I went and pulled it
all back down and tightened it back up.
Fine. Great. Two more days go by. Rain.
Rain. Okay.
Yesterday,
we had some massive kinda squalls come through,
(03:54):
like sideways rain. I saw that. It was
a little nutty. It it came down for
basically
two hours straight sideways.
I was watching a little river form between
my house and and my neighbor's house because
there's kind of a natural divot there, and
and that leads down to the preserve behind
my house.
And so I'm I'm looking. I'm looking. So
(04:15):
I go outside, and I just I just
open the door. I'm not I'm not going
out in this weather, and I peek my
head over, and there's not just one bubble
coming down. Now there's two bubbles, and the
whole thing is, like, leaning, leaning, leaning, leaning.
And it's coming down, and it's almost touching
the patio furniture. Oh, man. Patio furniture's
only, what, three feet maybe? Right. Considering, like,
the top of the back of the chair,
three, four feet. So it's coming, like, just
(04:36):
down, down, down. I'm, like, I'm telling my
wife, I said, this might be it. Like,
it it it's it's it's probably gonna go.
Absolutely did. So it got so heavy with
the rain
on with those two bubbles on the one
side, it actually tore
the the entire top. Like, it just sheared
itself, like, right off. And then when it
sheared itself with the weight, it also bent
(04:57):
all all the metal and things because this
is like a big accordion. So now this
thing will not go back together. It's probably
the best $70 I've spent for, like, a
week of just being able to sit outside
and things like that. I was kinda telling
my wife. It's like, you know what? Worth
it. Maybe I should just go back and
see if we can buy a couple more.
Like, we'll we'll just keep them on the
side of the house in in stock and
(05:17):
and and ready to go for the next
one. So that that that's my next adventure
is after I'm off this call, I'm gonna
go grab a hacksaw, and I'm gonna go
out in the backyard. And now I gotta
cut all these metal spider legs off this
12 by 12 monstrosity thing because there's no
way it's going back together with how bent
it is and everything else. And then because
it sat there over my patio furniture so
(05:39):
normally in my patio furniture, it sits outside
and we leave the cushions on sometimes, and
it doesn't get soaked soaked. It's got, like,
Scotchgard or whatever on it. The water all
runs off. Uh-huh. Because this sat on it
for so long, it soaked through the material
on the tent, and it soaked into all
the patio cushions. So when I picked the
patio cushions up and I just turned them
from horizontal to vertical, they've just been leaking
(05:59):
out for the past hour. Like, yep. Still
watching them drip. I will not lie. I
thought for sure one of those tents was
gonna, like, take off and blow away. I
thought that's what was gonna happen. Because when
it starts raining sideways, I've seen I think
it was actually a hurricane. I've seen a
trampoline, like, blow from one house into the
neighbor's yard in some of the wind that
we've had. It had a great, like, vent
(06:20):
on the top that Okay. Would crosscut, so
it really wasn't gonna take off or go
up and down. And we had it all
tied down and everything. I I didn't use,
like, all the hurricane ties, but it it
was definitely, like, attached to the house and
and where it needed to be. So fun
fun little adventure for the week. Wow. Way
more adventurous than my week. I don't think
I have any fun adventure stories. So between
(06:41):
that and and my new terminal toy, which
we'll we'll talk about later. We'll we'll do
a toy episode and and annoy folks
sometime in the future. But, anyway, that's my
news for the week. Why don't we get
into the news from the last couple weeks
because we've had events like build
and all sorts of other stuff. Yeah. And
some of those events, they're definitely developer heavy.
(07:03):
Like, it's all different things. But there's
always news, I feel like, that relates to
IT pros because us as IT pros have
to manage our developers and keep them in
line and make sure they're behaving. And as
a result, we get different things, and let's
face it, there's Copilot stuff everywhere, so there's
always Copilot news to talk about as well.
And we haven't done much of a news
(07:24):
episode, so some of these I think everything
we kind of put together was the last
couple weeks. Some of it may have been
a little before build, all around the build
time frame. We really should do these more
often. Although, I don't feel like there's been
a ton of news either, or I've just
missed it. There's been quite a bit. I
think it depends It's kinda what you're looking
for. Where you sit and where you play
(07:44):
around. Yeah. Yeah. There's been a bunch of
new stuff come out across
M 365.
I was looking the other day
at some of the
latest announcements that have come out for
Azure Kubernetes service and and some of the
things going on in AKS land.
Tons going on over there.
(08:05):
The the trick is kinda sifting through the
AI, AI, AI to what's what's the what's
the real thing, or if AI was your
thing, then, hey, maybe there was something real
for you there too. Like, I definitely saw
some stuff with Copilot
that piqued my interest, and and maybe we
could start there. So
(08:26):
last week, we we talked to Do you
wanna start with that one? Yeah. So so
last week, we chatted with AEC
about declarative agents, and that was kind of
a follow-up to some stuff that we had
talked about previously with Copilot Studio and doing
kind of
well, for you and I, no code declarative
agents, just in Copilot Studio,
next, next, next in context of an m
(08:48):
$3.65
subscription, things like that.
They're introducing
fine tuning for Copilot,
which kinda takes this a whole another level
because some of the at at least from
my perspective and the things that I do
in my day to day, because what I
do today
in something like Copilot Studio with declarative agents
is I'm relying on RAG and and retrieval
(09:10):
augmented generation.
And this is really being able to take
a model that already exists and the weights
in that model that are already deployed out
there and exist, and then fine tune your
own model on top of it
or fine tune and refine that model so
that you can drive
specific business processes.
(09:30):
And the demo I saw around this really
got my head turning. So I'm a product
manager.
I live in a land where
I'm a I'm a remote employee.
Communication and particularly written communication is super important.
So we spend a lot of time on
product requirements documents, PRDs,
justifications,
specifications,
(09:51):
designs,
and we have formal templates for all this
stuff. One of my big problems with it
is
people who take the template and they massage
it into their own thing or they don't
know what a good one is. So they
they kinda look at the template that we
have today, and and they just get like
a deer in headlights and and scared about
it.
So I like, with this, I could take
(10:12):
all our existing
PRDs that we've done for, like, at least
in my time for the past five years
on my team.
I could train
and fine tune
the existing model on my PRDs,
and I could actually have it tune on
top of my templates as well.
So then I could make it super turnkey
(10:34):
to go in and turn these things out.
And that's a very simple scenario for me.
I saw a great demo on the Microsoft
Mechanics YouTube channel, and I'll I'll put a
link in the show notes, and I recommend
folks go out and watch that one. That
was generating legal documents,
and it actually did a pretty good job.
Like, I'm not a lawyer, all all that
jazz. Yep. It's just to couch all that
(10:56):
and and have that caveat.
But that said, I I think that one's
really cool and and really exciting, and it's
something I'll have access to in my tenant
and totally intend to go out and play
around with. So I don't know if you
had a chance to either play with this
in your tenant or watch any of the
watch any of the introduction videos or anything
like that that's out there, but we would
(11:18):
would love to get your thoughts on this
one as well. I did watch the Microsoft
Mechanics video while I was doing other stuff,
so I was kinda half watching it. This
is a really cool feature. I like how
you can fine tune it that way. I
would say the part that I was interested
in, kinda coming at this from
the security model is in my
head too when you see this announcement. You're
(11:39):
like, okay. You're now fine tuning
an LLM.
What does this mean for
data security? What does this mean for how
is my data now being used if I'm
fine tuning
an LLM? Is it going into the LLM?
Is it training the LLM? And they talked
about this in the Microsoft Mechanics video, and,
ironically, I don't I probably don't have the
(12:00):
business use cases for this yet, although I'm
starting to play with it more. But it's
more thinking about this from a client perspective
of questions they're gonna ask about their data,
their data security. Some of the stuff we
even talked about a couple weeks ago when
I was talking about how a lot of
clients are coming to me now and saying
we wanna prep for Copilot, we wanna make
sure data is secure.
(12:21):
Now what does fine tuning this mean? And
they did a pretty good job, I would
say, in the Microsoft Mechanics video from that
perspective of, like, you still have your LLM,
and it's not and I I could not
explain this from a deep technical level on
how it actually works, but they actually take
your data. And the way they illustrated it
is they, like, attach it to the LLM
for the fine tuning. It's not like your
(12:43):
data's gonna go into the LLM and they're
retraining the core LLM. It's like something between
RAG and actually having your own model, where
it's a model with your data attached to
it to do the fine tuning. And it
sounded like from that video that it's not
even always attached to the LLM. Like, it
attaches to the LLM to do the fine
tuning, and then it kinda goes and stands
(13:05):
separate from it after you've done the fine
tuning. It only reattaches when you do it,
and it still maintains, like, your data and
your own tenancy.
They're not taking your data to train the
core LLM.
All your access lists and everything, that is
all still maintained.
So it was that that part to me
was interesting to try to understand a little
(13:25):
bit more from that perspective
on how was it doing the fine tuning
of the model without actually
absorbing the data into the model? The way
you can think about it is
fine tuning,
and and that's really what we're talking about
here is is kinda
taking an existing model that's already been trained
Yep. Let's say chat GPT four o, and
(13:49):
then you're gonna take that, and that that
that's a stage in that model's life cycle,
and you can use it as it is
with the weights that exist and and and
all of that stuff. Now what you can
then do so you do all that, and
and that's kinda like training slash pre training.
And then you take that pre trained model
and you fine tune it. And fine tuning
(14:09):
is a refinement process. So you're taking this
very it doesn't have to be very small,
but you're taking a smaller
task specific
dataset. So, like, in my example,
PRDs and and justifications. Templates and yeah. You're
doing this to
optimize
performance
(14:29):
and
and and drive a business
process.
So, like, you you you're right. You're not
loading this stuff into the original
pretrained model. You're kind of creating your little
bit of, like, an your own offshoot of
a model, but you don't have to go
back and redo all the training and redo
all the weights for the original model. So
(14:51):
you've kind of augmented it with this set
of sidecar weights that can then be used
as well to
go ahead and make your responses better.
Now the the the purpose of this and
and the difference
versus,
say, something like rag
is
RAG has to go out and you you
(15:13):
have to chunk all these documents, you have
to put them in a vector database,
and you have to compute and and and
run all that stuff every time. This is
just having that set of
kinda fine tuned weights out there for you
and ready to go. So it's it's it's
a click stop if you think about it
as a series of stages. So pre training
is all about, let me get the model
(15:34):
out there, and that's
general representation
of whatever it was trained on. And then
fine tuning is let me take and
take this model to the next level by
really augmenting it in a way I shouldn't
say augment because of the whole rag thing,
by
adapting it. Yeah. Fine tuning it. We're fine
tuning it. Right?
(15:55):
We're fine tuning it. So you're going but
you're going from general representations
to specific tasks or Right. Knowledge domains.
Processes. Yeah. Processes. So you could do this
multiple times. So when you hear about maybe
a company that's working on
AI to solve cancer, well, guess what? They
did the same thing. Right? They they took
that general model that was out there, and
(16:18):
then they took a bunch of cancer research
that had already been done previously, and then
they fine tuned that model so that they
could have a purpose built model that they
could go and and work with
to
attack cancer, attack your
in your case, maybe your business processes
around statements of work and how you put
(16:38):
those together for clients. I should try one
of those. You've done that a whole bunch
throughout the years with things like document libraries
and document sets, and this is I I
I think you should give it a shot,
like, if you're licensed for it because it's
a little bit of a next level thing
for you. I should absolutely go try that.
I I didn't statements of work didn't click
in my head until we were just talking
about it. I'm like, yeah, master service agreements.
(17:00):
And maybe it's because I don't put a
ton in there. Statements of work, I put
a little bit more in there. Master service
agreements are a lot of addresses
and that. But I have had clients ask
can't even remember which client it was. Where
they did, they're like, well, if I have
a template of a document and I want
to use AI to fill out the template,
can you do it? And I'm like, well,
(17:21):
not really.
The way it was
two months ago even or before they talked
about this fine tuning, it's like, doesn't really
do a great job at that. And this
is where
now I think as clients get this, again,
licensed for it, willing to build it out,
it's a lot more gets you a lot
closer to being able to have templates that
you can auto fill out. Not auto fill
(17:42):
out, but you still as a user
can yeah. Yeah. So so you can generate
those
and you as a user providing inputs. So
the the way I think about it and
the way I see a lot of folks
use generative
AI today
is they kind of look at it and
and I've definitely seen,
(18:04):
particularly, like, my PRD example, I've seen some
PMs that go out there, and they just
say, hey. Write me the document.
And this isn't, hey. Write me the document.
This is write me help me write the
document so that it conforms to this process,
this knowledge domain,
any anything like that, you know, that exists
out there. So, yeah, it's it's a tool
(18:26):
in the toolbox
versus a do my job for me kinda
thing. So if you take an agent,
and I don't know that they talked about
this, and I don't know if you've seen
it, Could you potentially take like, let's say
you go do the co the fine tuning,
and, really, you're creating an agent. So that's
the other aspect about this. This isn't fine
tuning like your general
(18:47):
Microsoft three sixty five Copilot responses. You're creating
an agent that's fine tuned
to do
a specific task.
Can you combine
agents with actions yet where you could do
a fine tuned agent going back to the
proposals or statements of work where you're then
using an action
to prompt or to get certain information
(19:08):
to feed into a fine tuned agent to
generate these particular outputs? I haven't tried to
chain them together like that yet. I I
I think that's where all this stuff goes.
So if you think about
MCP servers and the whole model context protocol
thing,
that's where these agentic things are going. And
you see a little bit of this today
in
(19:29):
in maybe in the declarative agents that we
already talked about where there's a connector for
SAP, there's a connector for Workday,
things like that. Like, you have this agentic
system on the side that it can reach
out to and sidecar that knowledge in from
that other system
and bring it in. I always think it's
super funny, at least in today's world where
we're still
(19:50):
very early days in this stuff, when
you you see the funny two agents talking
to each other kinda thing. I don't I
don't know if you've seen the video. It
came out a couple months ago, but it
was like a robo calling agent called another
robo caller. And then they both figured out
that they were robo callers, so they just
basically started talking to themselves in it sounded
(20:10):
like an old, like, bot modem going off.
Like I think I did see that. Yes.
Back and forth because they knew there was
no human there and, hey. Like like, let's
go down a different path kinda thing. So
they're still a little wild and off the
rails. But, ultimately, I think that's where these
things do go. So if I think about
an
agentic chain to build a PRD, well, great.
(20:31):
So I've got the fine tuned
process where I've come in and said, hey.
Help me write this within these guard re
these guardrails, or help me generate this thing
within these guardrails and and these boundaries.
But part of my PRDs would be customer
stories.
Well, maybe I wanna go build a little
agent that does nothing other than reason over
(20:52):
my customer feedback.
And then there's another agent that I might
wanna go build that reasons over
the financial aspects of my business and and
the cogs of it and and how that
comes together. So all these little things have
their own
click stops along the way, and, eventually, I
think, yeah, that that's where you end up
is potentially a bunch of agents just chatting
(21:12):
back and forth with each other, and one
throws this over here, and then it throws
it back this, then it throws it into
a new one, and out the other end
comes something else. And I I think that's
the thing that we we all haven't, like,
really figured out as Right. I saw the
I think it was the CEO of Claude
this week, last week. He was doing an
interview, and he was talking about basically, like,
(21:34):
50%
of it was some crazy number, 20%, fifty
%. It was really high of kinda low
end knowledge worker jobs going away because of
AI and and Gen AI. So nobody knows
how this stuff's gonna go and where it
comes out. I'm trying not to be, like,
the doom and gloom person about it. Like,
I very much do look to this stuff
(21:55):
to augment my job today
and help me be faster? Like, I don't
know. IT Pro land, all we've ever been
doing is trying to automate ourselves out of
jobs. This is just another step in that
journey is the way I kinda look at
it. And then just go watch Mission Impossible.
Have you seen the latest Mission Impossible yet?
I haven't. Don't spoil it for me. I
will not spoil it for you, but all
(22:16):
I could think about was AI and Copilot
during Mission Impossible.
Do you feel overwhelmed by trying to manage
your Office three sixty five environment? Are you
facing unexpected issues that disrupt your company's productivity?
Intelligink is here to help. Much like you
take your car to the mechanic that has
(22:36):
specialized knowledge on how to best keep your
car running, Intelligink helps you with your Microsoft
cloud environment because that's their expertise. Intelligent keeps
up with the latest updates in the Microsoft
cloud to help keep your business running smoothly
and ahead of the curve. Whether you are
a small organization with just a few users
up to an organization of several thousand employees,
(22:57):
They want to partner with you to implement
and administer
your Microsoft cloud technology.
Visit them at inteliginc.com/podcast.
That's intelligink.com/podcast
for more information or to schedule a thirty
minute call to get started with them today.
(23:18):
Remember, Intelligink focuses on the Microsoft cloud so
you can focus on your business.
Alright. So Copilot Tuning, folks can go out
and sign up for that. I believe you
have to sign up for a preview for
that one or or kinda put yourself Yeah.
This one's still it's early access. About the
Copilot Tuning early access,
(23:38):
customers with more than so this is a
high barrier to entry, so I probably cannot
do this yet either. But it does say
early access, you have to have more than
5,000
Microsoft three 60 five Copilot seats to participate,
EAP candidates undergo
screening, make sure they have scenarios that align
with it. Like, Microsoft is at that point,
they're sounds like they really wanna learn from
(24:00):
this before they just throw it out there
to GA for everyone. This is the same
thing you saw with Copilot before. So
this shouldn't be a model that surprises folks
at this point. They have a couple
people that are already using it. Ernst and
Young, McCarthy,
I don't even know how to pronounce that
word, Tyrolt,
Land O'Lakes.
They have some big companies using it. So
(24:22):
if you're a large customer and wanna go
play with us, early access program, if you're
not, be patient. You'll get it eventually. Don't
let that stop you from going out and
watching some videos. Like, the mechanics video is,
like, ten, twelve minutes. Totally worth your time,
I think. This looks like a build video.
I haven't clicked on this one. This one's
probably, like, an hour long. Oh, no. This
one's ten minutes. There's an introduction video too
(24:43):
on the article, so definitely go out and
learn from it. Start building agents.
That's I've been playing more and more with
building some agents as well, which that leads
to another news article. This one's
one that caught my eye is
you now have
Microsoft
Entra Agent
ID, Secure and Manager AI agents. So this
(25:05):
one was, I think, right before
Build, and I saw Merrill posted something about
this. I saw that. I saw a few
blog articles about it. Alex Simmons has a
announcement on the Entra blog about it. But
this was interesting to me because, again, we've
been talking about we're building these agents. We're
giving them access to
(25:26):
SharePoint sites, to different data.
And this is starting to look at, so
how do we actually go in
and start securing those agents, securing what those
agents can do?
And now if
you're an admin and you're a tenant, you
can go out to Entra, You can actually
go out
to Entra
and go to your enterprise applications.
(25:48):
And if you look at application
types,
you now have an agent
ID
type in your
enterprise applications.
And this is this is cool
just from a pure
reporting perspective right now, and that it will
give you a list of every single agent
that has been created in your environment
(26:10):
by any users,
whether it's an agent created in Copilot Studio
or Azure AI Foundry that are registered a
part of your Microsoft three sixty five tenant.
This is one that there's some videos out
there. I went and did it. My tenant
saw all the agents there.
Great for reporting perspective
for right now. I would say from, like,
(26:31):
a security perspective,
there's not as much you can do, but
there's a lot of stuff coming. So they
have a whole part of this that's what
what's next with these EntraID agents now that
they're in EntraID, now that they're showing up
as enterprise applications.
It means that you can start going in
and looking at, like, eventually what Graph API
access they have or
(26:53):
even creating conditional access policies for your agent.
So now someone can interact with this agent.
Being able to go in and do
granular conditional access and all that detailed permissioning
that comes from that perspective
because,
I mean, I mean, we're getting to the
point with these things, like, going back a
little bit to what we talked about earlier
with
(27:13):
agentic workflows and these things effectively talking to
each other. Like, you almost are treating them
as individuals within your business. So it makes
sense from that perspective
to secure them in in the same way
that you would your users and other things
that are out there. So this will give
you additional operational controls
and, frankly, like, comfort food around locking some
(27:36):
of this stuff down when it comes to
what do they access, when do they access
it, how do they access it.
The the the I think the granular graph
stuff will be great. Like, hey, making sure
this thing only has
a read permission versus a read all, or
does it does it have the ability to
actually
generate content in that environment? Like, oh, no,
(27:57):
I wanna I wanna block rights for this
one because I don't actually want it to
go in and I don't know. Somebody goes
and builds the agent to rename titles in
the GAL.
You you you might not want that that
that thing
running that way. Yep. It's also good just
from a
visibility perspective for these things. Like, it's another
way to kind of track adoption in your
(28:19):
organization, maybe go talk to some users,
see where they're getting
ROI out of out of these kinds of
things,
and better manage the life cycle of them.
I still see a lot of places where
you go in and you look at their
their intro tenant, and it's just the graveyard
of lost stuff that that has not been
(28:40):
cleaned up over the years and hasn't been
attributed to a user. Hey. What's this old
enterprise app ID doing out here that has,
like, basically, god permissions in your environment? Well,
like, what did you do, and why was
this a deity? Oh, that was created way
back when five years ago when we spun
up our tenant. Great. Let's clean it up
today. Right. And I think some of those
(29:00):
two going back to life cycle,
above and beyond this announcement, permissions change too
with Graph API. Right? Like, there's old applications
out there that have full control over your
entire SharePoint environment
because
granular
site level API access in the graph wasn't
there when these apps were originally created. So
(29:20):
as things continue
to evolve, permissions continue to evolve over time,
being able to,
see that life cycle, see what they're doing,
see what they have access to, some of
the auditing and monitoring you'll be able to
do around
logs and visibilities into what these agents are
doing, what they're accessing.
(29:40):
You've even seen that some with Copilot and
some of the new stuff around,
sensitivity labels and being able to do DLP
and audit user interactions
with Copilot,
now being able to audit agent interactions with
stuff. Enterprises, this is I feel like it's
something you have to have when you start
getting into AI. If there's legal questions that
(30:02):
come up, questions that come up about how
something was generated, being able to dive into
these logs and see this does become super
important. And this is one that's out there
today. Like, folks can go play with this
one. Yes. You can see them. Like, they
have a whole list. Again, I could see
it in mine. It's labeled as preview, so
all the preview stuff, it's in preview.
I'd say don't use it in production, but
it's just a preview feature that shows up
(30:23):
in production.
So don't put production workloads with it, but
then a lot of this other stuff that
we talked about, the conditional access, the auditing,
life cycle management,
in this blog article from May 19, it
says,
this is all in the coming months. A
lot of these capabilities
around
zero trust security posture,
(30:45):
all of that is going to be rolling
out to these
agent IDs in Entra. So that's one I'm
looking forward to playing with, keeping an eye
on. What other one you wanna talk about?
We maybe have time. One more announcement. Why
don't we do the Russinovich session? So he
he did a
app build. Mark Russinovich did a session on
(31:07):
inside Azure innovations.
I don't I don't know if you had
a chance to go and watch that one.
I failed on my homework to watch that.
So
I am curious for you to fill me
in on that one. I'll put a link
in the show notes for everybody, but that's,
b r k
one ninety five.
Slides are out there, recordings out there,
(31:28):
all all that good kind of stuff, so
folks can go and take a look at
it. But, effectively, like, this talks about
the internals
of Microsoft and Azure infrastructure.
So what's the latest in boost and DPUs?
What's going on with FPGAs?
How are these things being
(31:49):
leveraged within the
within the Azure fabric.
One of the cool things
that was in there, or at least I
think it was cool more because, like,
some folks on my team were
were were involved in it,
is there's a demo in there where he
talks about blob storage and scaled accounts. And
(32:11):
so one of the things that we've been
doing in storage over the years is trying
to enable these large
AI training companies
who are doing these large pre training runs,
so OpenAI, so the world, things like that.
And
as these customers are bringing in petabytes
and petabytes and petabytes of data, it's not
like single digit, It's it's double to triple
(32:33):
digit
to exabyte scale because you got, like, the
world's knowledge in there. Like, downloading the New
York Times archive is a pretty big thing.
Do that across a bunch of
different sessions that exist out there, and
it it gets quite a bit bigger. So
we did a demo,
and we've been on this journey with scaled
accounts to ever increase the amount of throughput
(32:55):
that we can give them. So ingress, egress,
the amount of IOPS we can give them
in the form of TPS, all that. So
we did a demo where
we were able to, on a single storage
account,
provision
a workload
that ran
at 15 plus terabits a second of ingress,
25
plus terabits a second of egress, and this
(33:18):
is all random
random IO through
through an HPC benchmarking
suite called
IIOR. Okay. We were able to just run
this for hours.
We didn't do this in a special place,
like, just public fleet.
This this is stuff our our customers
can go in and do today.
(33:39):
So so I thought that was a super
cool one. And like I said, important to
me because while my team doesn't work on,
like, the perf and scale stuff, one of
the key components we work on is AI
and ML clients.
So I own our my team,
we we do our SDKs, client tools.
BlobFuse is one of those. So BlobFuse was
heavily used in that demo and ready to
(33:59):
go. So we've done a bunch of work
there just to be able to, like, spin
up workload
in BlobFuse and have it be able to
take, like, full advantage
of
a VM's NIC and soak it all the
way through. So I I would recommend everybody
go watch that session. Like, if you're interested
in the
either the internals of Azure or even a
little bit more about the internals of of
(34:21):
the way some of these workloads compose. So
there was some stuff in there about Linux
Guard and kind of how code integrity is
handled
within Linux, some cool demos for that one.
There was a bunch of stuff around
virtual machines
and kinda how we do host OS upgrades,
(34:42):
hot patch, driver swaps,
full reboots,
and and how those things go within the
within the infrastructure. There was some insights into
how we do RDMA
on the networking side, like, all sorts of
cool stuff. And because it's a Recenavit session,
it's also demo driven, so you actually get
to, like, see it in action. It's not
(35:02):
just a bunch of slideware. Yeah. I'm flipping
through
the slides here, and it looks like there
were, like, 10 or 12 demos in there.
And it is like, I always try to
go watch Russinovich's.
It has been a time constraint more than
lack of desire to watch his yet. But,
yeah, all this stuff, there's some stuff around
that. A confidential
compute,
a bunch of stuff around Azure confidential
(35:24):
GPU VMs.
Yeah. His are always,
like you said, super
demo driven, but I would also say super
nerdy.
Like, for
super nerdy or super technical. Maybe I should
be politically correct. Super technical, like, the amount
of detail that are in his sessions and
the amount of I don't know if he
just sits there and practices all these to
(35:46):
get all these technical details down or if
he just has this much technical knowledge floating
around us in his head. Either way, I
am always super impressed with the
level of technical depth that he's able to
go to in these discussions and these presentations.
So I always find them fascinating. They're great
to go watch.
Mark is
(36:07):
super deep in the platform
and and does understand a bunch of this
stuff. Like, he's not out there just
spouting it off based on a script that
I think somebody else wrote for him. So,
yeah, I I encourage everybody
to to to go and watch that one.
It's it's it's worth an hour if you're
into Azure in general, and you and you're
(36:28):
kinda looking for what's the latest in hardware
innovations there across really all the core services,
so compute, networking, storage. Yep. And if you're
only interested in one or two of those,
skip around the session. You can watch it
on YouTube, and it's got chapters and everything
in there for you. What is this? The
world's first analog optical computer
for accelerating
(36:48):
AI
inference.
Potentially a hundred times more energy efficient than
GPUs.
If that is at all intriguing to you
about light doing massive parallel computations
for AI, go watch the end of this
session. Right? Go watch the end of it.
So power consumption is important. Sustainability
is important. Yep. I saw and and
(37:08):
don't quote me exactly again on the percentage,
but it's something like three to 4%
of the world's power now is going to
data centers,
and it's been driven on this general uptick
around
more and more CPU and particularly more and
more GPU usage because these things are so
power hungry. So
I I think it is important that all
(37:29):
these companies that are working towards this stuff
continue to strive towards sustainability
and innovations
in
not only, like, capping power consumption, like, at
some point, it can't keep growing forever,
but also bringing it back down. So things
like like analog
optical computer AOC, not not the other AOC,
(37:50):
this AOC
fall into that category of let's make the
world a better place. Very cool. Well, those
are some fun announcements,
interesting news. We spent, like, the whole time
in, like, three news articles. You should see
the number of blog posts I have to
go through. I think we went through
a a couple hundred of them that are
out there. So the other thing that I'll
put in the show notes for folks,
(38:11):
the the book of news
is out and available as well for that,
and and that's publicly available. So if you're
interested in maybe the more
broad roundup of, hey, what were the bigger
announcements
at Microsoft Build, things like that, I would
encourage folks to go out and
take a look at the book of news.
(38:32):
Maybe
throw it into your favorite LLM and have
it summarize it for you. Absolutely. And if
you have any questions,
topics you want us to talk about,
news you want us to cover in future
episodes as well, let us know. I would
say either if you're part of the membership,
let us know in Discord. If you're not,
LinkedIn,
I'm quickly leaning towards as being the social
(38:54):
media social media probably social media, social network
that I am spending the most time on.
All the other ones
Yeah. I feel very hit or miss. LinkedIn
is probably the best place to get both
Scott and I and get our attention,
if you have topics or questions about the
show. You can tag the podcast there. You
can also go to the website and just
use our contact
(39:16):
contact
contact us form.
It's alright. It's Friday. It's time to
call it a day at 11:00 in the
morning and go relax for the rest of
the day or go to meetings. I'll probably
do the former. But as always, thanks, Ben.
Appreciate the conversation. Thank you, and enjoy your
weekend. We'll talk to you next time. Yep.
(39:36):
If you enjoyed the podcast, go leave us
a five star rating in iTunes. It helps
to get the word out so more IT
pros can learn about Office three sixty five
and Azure.
If you have any questions you want us
to address on the show, or feedback about
the show, feel free to reach out via
our website, Twitter, or Facebook.
Thanks again for listening, and have a great
(39:56):
day.
Popular Podcasts
United States of Kennedy
United States of Kennedy is a podcast about our cultural fascination with the Kennedy dynasty. Every week, hosts Lyra Smith and George Civeris go into one aspect of the Kennedy story.
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.
Dateline NBC
Current and classic episodes, featuring compelling true-crime mysteries, powerful documentaries and in-depth investigations. Follow now to get the latest episodes of Dateline NBC completely free, or subscribe to Dateline Premium for ad-free listening and exclusive bonus content: DatelinePremium.com