YT YouTube2Text

YouTube2Text → Transcripts → Microsoft IQ Overview

Microsoft IQ Overview — Transcript

by John Savill's Technical Training · 4,036 words · 604 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00Hi everyone. I previously did an
0:03overview video about Microsoft IQ, but
0:06as things evolve, I thought I should
0:08probably update it. And honestly, I
0:11suspect it won't be the last time. Now,
0:14I think at this point, the idea that
0:17generative, this creative AI is hype.
0:20It's a nothing burger is behind us.
0:23Organizations understand and are looking
0:25to leverage AI to provide business value
0:28from employee productivity, reimaging
0:31operational processes, lighting up new
0:33customer experiences, accelerating
0:34innovation.
0:36Every organization wants to leverage AI
0:38in some ways. And if we think about as a
0:42human being what my job entails
0:46which really around a certain set of
0:48responsibilities
0:50and I accomplish those responsibilities
0:53through my skills, my knowledge, the use
0:56of tools, processes I follow. And sure
1:00with AI, I think those the skills I have
1:03are going to evolve. The tools I use are
1:05going to change. The processes will
1:06shift over time with AI. And as I look
1:10at using AI,
1:12well, the tasks I perform fall into a
1:16number of different buckets with how AI
1:19might apply. I think there's going to be
1:21things I can just hand over to AI. The
1:24more mundane, the repetitive, it's a
1:26very strict process that I can very
1:29easily define. A little bit of
1:30creativity. I can just let AI do that
1:32for me. Maybe I oversee the end result.
1:35There are things that I can partner with
1:37AI and AI can help augment what I do,
1:40accelerate what I do, make what I do
1:41better, then there's going to be things
1:43that AI can't help with. But where there
1:47are things that I can either hand off to
1:49AI or AI is going to work with me to
1:54perform best, it's going to need access
1:56to the same types of knowledge that I
2:00have, that I've learned, that I use, and
2:03likewise the same types of tools. So we
2:04need to understand the processes.
2:08So I have to think about well what do I
2:10use in a day and it obviously varies by
2:13the particular task but at a high level
2:15I can think about well there's there's
2:17productivity type things. So I can think
2:20yeah I email various things I have teams
2:25chats I have meetings
2:29and there's obviously different
2:32artifacts. So there's documents,
2:36there's spreadsheets, there's
2:39powerpoints,
2:40there's line of business applications.
2:45And for all of these things, there's
2:47also relationships between them. So for
2:50example, I know how certain people
2:54relate to each other. Oh, they're
2:56working on this project. I know how
2:59people relate to certain artifacts or
3:01they were in this meeting or hey they're
3:03working on this. Um I have a a mental
3:07graph the the connections the edges
3:12between the entities.
3:14I have the way I work how I write my
3:18processes. So there's this whole set of
3:21things about how I do my work.
3:26And then I can think about well then
3:28there's the state of the business. So I
3:30think of the state of the business I
3:32look at various reports I look at
3:34dashboards and for most organizations
3:38there's just a huge number of silos of
3:40data and a lot of that is historical
3:44that we have different systems that
3:46required us to have their own data
3:48stores their own interfaces their own
3:50way of using it. It's why many of us
3:52create PowerBI reports. So in my PowerBI
3:55report I create these models of the
3:57enterprise entities that map to the data
3:59so it's easier and more consistent to
4:01work with to understand
4:03and so I might hey define those various
4:06things and that hey maps to this data in
4:09this particular place great so I have
4:12the the state of my business
4:15and I can also think about well there's
4:17just a whole set of institutional
4:20knowledge so for example there's maybe
4:23contract tax,
4:25there are regulatory standards, there
4:27are policies,
4:30whatever these are, these authorative
4:32documents,
4:34I have to go and use these things. I
4:36have to go and check on these things.
4:38And then also quite a lot of the time
4:40these days, I suspect I'm not alone in
4:42this. I'm using that wonderful
4:44invention,
4:46the all knowing internet where it's not
4:51some company piece of knowledge, but I
4:55need to go out to the internet and look
4:57stuff up. Maybe it's a new standard,
5:00maybe it's information from a partner,
5:02maybe a competitor, whatever it is,
5:05the process I'm performing or I just as
5:07a human, I want to go to a web browser,
5:09I go and search for the stuff, I find
5:10what I need. and there's not maybe an
5:12API and I suspect we all go and use the
5:15internet.
5:17Fantastic. Okay. So now we talked about
5:20those different buckets where I can
5:22either hand it over to AI or I can work
5:24with AI. So I want to leverage AI. And
5:27remember when we think about artificial
5:29intelligence and these creative things,
5:31what we're really starting off with is
5:34the idea of this generative model. say
5:36hey I want to use AI
5:40and this newest thing is the idea that
5:43we have typically we talk about a large
5:45language model we have these neural
5:48networks
5:50that are trained they're trained on a
5:53vast corpus of knowledge books sites
5:57posts from Reddits you name it but it's
5:59all public knowledge so it knows nothing
6:02about your enterprise and there's very
6:04much a knowledge cut off because
6:06obviously it was trained on a certain
6:08date and then once it's trained it it's
6:10static. It doesn't continue learning. So
6:14there's a cut off date of what it knows
6:16about.
6:18And also what's interesting today is
6:20more and more we're seeing models that
6:22deliberately limit the knowledge and
6:26instead focus on a capability like
6:28reasoning. And the reason they limit the
6:30knowledge is because I can shrink the
6:32number of these parameters because these
6:34parameters all obviously have to carry
6:36values. So it the more parameters the
6:39bigger the model, the bigger the size,
6:41the bigger the hardware requirements,
6:43the bigger the cost to use it. And so if
6:46we can shrink these down, hey, it makes
6:48them more effective.
6:50And what we're going to see today is
6:52when I'm creating my solution.
6:55Sure, I might have these frontier
6:57state-of-the-art last language model,
7:00but we don't tend to use a model.
7:04What I want to use is probably multiple
7:07models. So, actually maybe there's also
7:09a smaller model I use with a much
7:11smaller number of parameters, but that
7:13does the job for what I need. Hey, I
7:15need a big powerful
7:17frontier model for some deep reasoning,
7:20huge context, lots of tools it's using.
7:23Fine. But then also sometimes it's just
7:26a I want a very low latency interactive
7:29simple interaction. I don't need to use
7:32that. Why use the additional computation
7:36and therefore the additional cost? I can
7:38use a smaller model. I don't want to
7:41just always use the highest common
7:43denominator.
7:44And so as I go and create my agent,
7:48the focus of this is yes, I'm going to
7:52use
7:55multiple different models.
7:58I want that model flexibility and models
8:01change so frequently. There is no best
8:02model. There's a best model for a
8:04certain type of activity today. Tomorrow
8:06it be something else. So I want to
8:08architect in such a way that it's mod
8:10model flexible. I can adapt. You never
8:12know when a model's going to be
8:13replaced. Um, as we've seen, models can
8:16become unavailable. So, I don't want to
8:18be tied to any particular model. So, I
8:20want to be able to talk to many
8:21different models based on the
8:22requirement, evolve as time goes on.
8:25It's why things like Foundry has a model
8:26router. So, we can do that for the
8:28agent. So, that's the whole goal here.
8:31So, my agent is going to use many models
8:34that have a finite knowledge. They have
8:37different sets of capabilities and I
8:39want it to work with me. I want it to do
8:41things for me. And I currently have all
8:43these different types of knowledge,
8:46intelligence, tooling that I leverage.
8:50So I want to be able to give this to my
8:53agent so it can tell the model. So it
8:56has the knowledge where it needs it. It
8:57has the knowledge. It can use certain
8:59tools. It has intelligence. And the key
9:02part is
9:04I want to be able to give it the very
9:08relevant pieces of data, the exact right
9:13amount. I don't want to just give it
9:15everything it could. There were times
9:17when we focused on the idea that hey, a
9:18model can have a million token context,
9:21so I can send it everything and it will
9:23decide what it wants to use. Now today
9:27we would laugh at that idea because now
9:29we we pay for tokens. The idea of just
9:32sending everything we possibly could
9:34from a computational from a money
9:36perspective and honestly even getting
9:37the highest quality outcome. I don't
9:39want to rely on the model to pick out of
9:41a million tokens what's the right stuff.
9:43I want to give it the most relevant
9:46data, the highest quality. So I optimize
9:49what I'm sending it. So I'm optimizing
9:50the computer. I'm optimizing the tokens.
9:52But not even optimize. I'm just going to
9:54get the highest quality answer. Like
9:56that is the focus here.
9:59And so how we do this? How do we bring
10:01the knowledge into this? Well, remember
10:04what is it we want? What I want to bring
10:08here is this idea of for my agent. I
10:12want to add in the right level of
10:15intelligence
10:19and tools.
10:22And so from a solution perspective,
10:25obviously this is the focus of the
10:26video. This is Microsoft IQ.
10:32That is the whole goal of what we want
10:36to do. Bringing that enterprise context
10:39and also the latest up-to-ate
10:42information from the internet to my AI
10:45capabilities. Now it's made up of
10:47different groupings of intelligence
10:51because for different agents I might
10:53decide I only need a subset of it.
10:55Remember we always think least
10:57privilege, least amount of access when
11:00it needs it. I'm not just going to give
11:01every agent everything.
11:04Like a human being, I only give what's
11:07required to do the particular task at
11:09hand. with an agent I only give it
11:12access to the intelligence the tooling
11:15for the task at hand. So whatever that
11:17agent capability needs to do it only
11:20gets enough. So when I think about this
11:22first grouping over here typically
11:26focused around M365 but that's obviously
11:28growing. Think about Dynamics 365 the
11:31line of business apps the messages the
11:32sites the exchange the shareepoint
11:36everything else.
11:37Well, this is work IQ. So, this whole
11:41segment here
11:45is when you hear that term, it's really
11:48focused on the idea of the how we work,
11:51the artifacts we use, how we do the
11:54jobs, the relationships between the
11:56things. It has a huge personalization.
11:59It has a memory. It remembers the
12:01interactions. It has a recency bias. So,
12:04it it it moves that memory window. So,
12:06it's going to pick the people, the
12:08documents, the things that I've done
12:10most recently are probably the things I
12:12want to do. But it does also learn on
12:14that long tail. It learns how we talk
12:16from the IMs, the meetings, the emails,
12:18the documents, how I structure my day,
12:20who I collaborate with, what their role
12:22is, what I care about, what my team
12:24does. It learns the objectives, it
12:26tracks them, it helps tee up meetings.
12:29And the whole goal here is it has tuned
12:32inferencing to give the best
12:33interactions across those data apps
12:36workflows. So I get the very rich set of
12:39context related to that collaboration.
12:43It can help predict and it has a full
12:46API. Now I can use agent to agents A to
12:49A model context protocol rest may not
12:52care about those things. Point is I can
12:54interact with this in huge numbers of
12:57different ways. So it's not just getting
12:59the intelligence the knowledge
13:02but agents can very simply call a
13:05certain set of tooling say create a new
13:07calendar appointment send an email
13:09upload a document and more. So it's not
13:12just reading it's enabling agents to do
13:14things as well.
13:17So we get into these idea of these silos
13:19of data and this sort of category is
13:21fabric IQ.
13:26Now the goal here is you're never going
13:29to be able to say hey we have this new
13:31corporate standard this is where all the
13:33data is going. Uh there's legacy reasons
13:38there's anchors to certain systems
13:40there's just a cost uh a migration
13:43challenge. So what we do is we set up
13:46this idea of sort of a data
13:47virtualization layer
13:51and yes I can store things directly in
13:53here as well. I can have structured,
13:55semistructured,
13:57unstructured. But then also what I do is
14:00most of the time with zero data copy
14:03wherever my data is and if it is if it's
14:07not an open format, sure I can do sort
14:10of mirroring capabilities into it. It
14:13all now surfaces via a single interface
14:16point. So all your analytical data,
14:18operational data, streaming data, native
14:21databases, time series, geospatial, you
14:24name it, it's in there. And then what we
14:27can do is because just as a human, you
14:30know, I I don't want to integrate with a
14:32million different tables. How do I know
14:33what's what? We would create the idea of
14:36sort of these semantic layers, these
14:38real enterprise entities.
14:41Well, I want to be able to do the same
14:42thing for humans, but also agentic
14:44capabilities. So, the agents can just
14:46understand the enterprise entities, the
14:48enterprise relationships, the
14:50constraints, the rules, the goals for
14:52them. So, I create these ontologies that
14:56map to the real things that can then its
14:58properties just point to the data
15:02in this virtualization layer. And again,
15:04the actual source can be anywhere. We do
15:06not care.
15:08So with what this provides with fabric
15:10IQ is now my agentic capabilities could
15:12go and talk to these ontologies the real
15:14enterprise entities and can understand
15:17everything that's going on. Sure, I can
15:19have native agents as well, but I can go
15:21and interact with this really, really
15:23easily. And and that's really the goal
15:25of this MCP APIs, whatever it is, the
15:28state of the business is available now
15:31to my agents.
15:34And then I can think about all all this
15:36curated sets of information
15:39that that part of it is foundry IQ.
15:44And the goal of this is a very strict
15:46created AI powered search where I only
15:49want specific sets of knowledge to be
15:51used by agent want to be very
15:52prescriptive about what data or sources
15:56the agent can see. And so these could be
15:58sitting in a blob SharePoint site they
16:02could be. So what I'm going to create
16:04these knowledge bases so I can point to
16:06different areas. I might point to stuff
16:08sitting in that virtualization layer
16:11unstructured semistructured. I might go
16:13and point to specific URLs
16:16on the internet. Again, this could be
16:18unstructured. It could be
16:19semi-structured from the lakes. I I can
16:22point to many different places, but I'm
16:23creating these knowledge bases of a set
16:26of intelligence that I want to be able
16:29to expose. And I'm exposing it in a very
16:31specific manner so that AI is leveraged
16:35to work out what's the right lookup to
16:38perform what's the right natural
16:40language or keyword based on this type
16:43of particular knowledge source because
16:45these are all knowledge sources and not
16:47just pass through the same request to
16:49all of them. I want to get highest
16:50quality data. I'm putting a restriction
16:53on so I get exactly what it needs to do
16:57and that's the whole goal around this.
17:00And then we do have this internet thing
17:03which is we all love. So then this part
17:06is web IQ
17:10and the whole goal here is obviously I
17:13can go and search for things. So it's
17:16using the Bing graph, that index that
17:19exists. So that massive breadth and
17:22depth that Bing has, but what it can do
17:25here is as humans, we we do a search and
17:27we get the page and we kind of scan
17:29through the page or we might kind of
17:30search for the bit that we actually care
17:32about. Well, it does that instead of
17:35returning the entire page,
17:37what I can do is just return the passage
17:42that's actually relevant to what I'm
17:43looking for. lowest latency of any
17:46provider, which is critical when you
17:47think about how agents work. They go and
17:49look for something. They get some data.
17:51It's like, oh, okay, now I'm going to go
17:53and look for this next thing. So, it's
17:54going to be multiple calls. I need a
17:55really low latency to not impact the
17:58experience I get here. So, with web IQ,
18:01it's super low latency, super broad,
18:04super deep actual set of knowledge. But
18:08I can say, hey, just return the passage.
18:09So, again, only give it the bit. to
18:13reducing the amount of tokens, reducing
18:14the computational work and the highest
18:17quality to the model.
18:20I'm just returning the bit it needs and
18:22it understands text, news, images,
18:25videos. I can even have a URL. Maybe my
18:28agent gets given a URL and instead of
18:31risking my agent having to go to the
18:33internet and get the entire site
18:36returned, which could be slow because
18:39it's the internet could have malicious
18:40content on it. whereby cube will
18:43actually use its index for that URL and
18:47return you a safe very low latency set
18:50of information
18:52and so that that is common across all of
18:54these things right so the goal of this
18:57is it's always getting
19:00the relevant
19:02pieces of intelligence you need I'm
19:05going to optimize the tokens but it's
19:07going to be the relevant data you need
19:09so I get the highest possible quality
19:11response And of course all of these I
19:13have a separate video if you actually
19:14wanted to go and dive into the detail.
19:17But if you think about what we just did
19:20here
19:22for me to do my responsibilities
19:26in most tasks I I go across all these
19:29things. Hey I had a meeting about
19:31something. There was uh a teams chat. I
19:34know this person's working here to
19:37contact them. I know this person's
19:38responsible for this. Hey, I have to go
19:40and look at this dashboard to see where
19:41we are. Hey, what is the contract? And
19:43oh, what? Okay, what's the current
19:45pricing or what's going on? I leverage
19:48all of these things
19:51to do my task for my responsibilities as
19:55I partner with AI to augment what I'm
19:57doing to maybe hand off certain things.
20:00I can provide that same level of
20:02intelligence
20:04to those agentic capabilities.
20:07And the whole goal of this when I I
20:09think about really what this is doing
20:11right here, this is providing your so
20:15you as an organization, this is your
20:18intelligence, your tooling.
20:22And what I want to do here is as I think
20:24about consuming these different things,
20:28especially as I move from an assistant,
20:30as an assistant probably runs as my
20:32identity, but I start moving to
20:34autonomous as a a digital colleague.
20:37What's really important here is this.
20:42It needs to be running with its own
20:44identity
20:47like that. That's critical. I do not
20:49want it running as me. I do not want it
20:51all running as the same
20:54identity for every agent because
20:55remember we want lease privilege. I want
20:57to be able to audit it. I want to be
20:58able to see what it's doing. And so the
21:00goal here is every agent has its own
21:04unique identity. So and this is again
21:06this is not just Microsoft stuff. This
21:08is could be any platform. Doesn't matter
21:09where the agent is. I can give it an
21:11identity. And so then I can have full
21:13role-based access control. It can only
21:15access things it's allowed to do. It
21:17would be subject to data classifications
21:20to policies and the same entry identity
21:23provider we're used to using for users
21:25and external users. We're just extending
21:27out to support the agents. So that
21:29existing level of trust and governance
21:32and security we're used to, I'm just
21:34extending out to my identities. Hey, I I
21:37can appmention the thing. So I and
21:39there's skills now for this. It's really
21:41easy to do. I can just enable it and I
21:44can app mention it in teams and it can
21:46go and start doing stuff in a comment in
21:49a word document. It can go and start
21:50doing stuff.
21:52And if we think about a lot of this I'm
21:54trying to map to as we we work with the
21:56agents as a human I want to be able to
21:58observe and protect we look at
22:01processes. So if I think about what I
22:03do,
22:05I want to be able to have full data
22:06governance, full data protection where
22:08data is used, I'm respecting those data
22:12labels and maintaining them. So if as
22:15part of an agentic usage, it uses
22:17something that's highly classified and
22:19it creates a new artifact or there's
22:21some data from here that's got data
22:23labeling, I need to make sure the
22:25artifact it creates is still highly
22:28confidential. That's critical or I start
22:30risking data leakage.
22:32I need threat protection for agents
22:34because there's new types of threat,
22:37jailbreaking, prompt injection,
22:38hallucinations, and a lot more. So, I
22:41think about what I I have to have here
22:43is this whole security.
22:47I need that governance. I need all of
22:49those things. And so, this is where a
22:51lot of the times
22:53you'll hear agent 365.
22:57So agent 365 is about bringing that full
23:00identity data threat protection registry
23:03of agents that can be through registry
23:05syncs. It can discover on uh clients
23:08through endpoint integrations. There's
23:09SDKbased registrations a whole bunch
23:12more. But the goal is data identity
23:17threat protection. All of those
23:19capabilities we bring to our agents with
23:21their own identity.
23:25And when I think about the agents
23:27running, many of us have probably maybe
23:29heard of DevOps. With DevOps, there's
23:32this whole cycle of hey, we we put
23:34something in, we observe, we have this
23:36idea of okay, then we continuously
23:38improve the process. So I want exactly
23:41the same thing here. I want this idea of
23:44observing
23:46everything that's going on because
23:51I want to keep improving.
23:54And that improvement could be tweaking
23:57prompts, tweaking tools. Hey, I need a
23:59different set of knowledge over here.
24:01Maybe it's even because again, the whole
24:03goal of what we're doing here, remember,
24:05is we have multimodel.
24:10Maybe we get to a point we say actually
24:13it would be worth creating a fine-tuned
24:15model because based on the interactions
24:18actually taking some of this information
24:21or or taking this behavior it has to be
24:23so prescriptive so deterministic I'm
24:26actually wanting it as part of the
24:28model's behavior as part of the model's
24:31knowledge so you might create a
24:32[clears throat] fine-tune model
24:34and again all of this this is not
24:37Microsoft hosted Microsoft hosted yes
24:39but it could be wherever those agents
24:41are. I want to be able to use these
24:42capabilities.
24:44So, I hope that helped. I mean, really,
24:46this is the whole goal of what is
24:49Microsoft IQ. It's about giving the
24:52right level of enterprise
24:55intelligence, knowledge, context, and
24:59up-to-date information from the web to
25:01your agents that is constantly evolving
25:04learning. And so I get this complete set
25:06of capability to make my agents as
25:09capable as they could possibly be. And
25:12through that capability
25:14helps me trust them. Especially with
25:16these agent 365 with this observability
25:20gives me the ability to actually run
25:22these agents in a way that I as an
25:23enterprise can trust to own those
25:26responsibilities. Hope that helps as
25:28always till next video. Take care.

About this transcript

This page contains the full transcript of Microsoft IQ Overview by John Savill's Technical Training, generated from the public captions YouTube serves with the video. The transcript has 4,036 words across 604 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.