Full transcript
0:02Awesome. Okay.
0:04Thank you everyone
0:06for joining.
0:08Uh I know we resumed after uh canceling
0:11two meetings due to technical issues and
0:13some change on Microsoft
0:15side how they're going to treat the user
0:16groups going forward. So, this is our
0:18first meeting in
0:20Zoom since we started this uh virtual
0:23group during COVID. So, just bear with
0:25us. Uh that's why you're hearing a lot
0:27of chatter how we are uh
0:28doing this, but I will try to go quickly
0:30so I'm not taking any time from Chris.
0:33Um okay.
0:35Oh, sorry. As you can see our contact
0:37here. So, uh
0:39Uh I'm sure you know Meetup, but we are
0:41also in LinkedIn and we have a email
0:43address which I'll talk about again.
0:45If you want to speak,
0:47you have a suggestion, you want to give
0:48feedback, good or bad, but don't don't
0:51don't just tell us you are bad, tell us
0:52why so we can fix it.
0:55And uh pretty much anything. Send email
0:58dbavug@outlook.com.
1:00Uh
1:00we look at it at least once in every 48
1:03hours, for sure.
1:06When do we meet?
1:07Second Wednesday
1:09at noon time, Boston Eastern, and fourth
1:12Wednesday to be fair to our West Coast
1:13friends like today. So, we meet every
1:15month. We take two breaks. One is during
1:17Thanksgiving. One is at the end of
1:19December during holidays. Otherwise, we
1:21try to hold 24 meetings
1:23uh with exception of this year, few
1:25technical issues. Uh we do have a
1:27YouTube channel. We have our some of our
1:29old recordings too. Today's speaker was
1:30kind enough
1:32and agreed to record. So, this will show
1:34up here. Don't send us a note after an
1:36hour. It's going to take us few days.
1:38Uh all volunteers, so it takes us few,
1:41you know, download and upload and do
1:42some logistics. So, by weekend or next
1:45week it should be there.
1:47PASS
1:50in Seattle in November 9th to 11th.
1:53Uh
1:54so, if you do not know, please look it
1:56up. Google it. Uh I'll be there. Many
1:59people will be there. People come from
2:00all over the world. And
2:03I think pretty soon I'm going to share a
2:04discount code in next meeting.
2:08Our future sessions are pretty booked
2:09for the year that's coming up.
2:12Uh
2:13as you can see there are some awesome
2:15speakers coming up. Two of them are
2:18actually Microsoft engineers. So take a
2:20note. I always, you know, get excited
2:22because you can get some scoop that you
2:24don't normally don't get when you bring
2:25a Microsoft speaker.
2:27And these are the events going on around
2:29the world. Most of these are a one-day
2:32free event. Um and, you know, I'll put a
2:35shameless plug if you're in New England.
2:37I am the main organizer and I have an
2:40awesome team. Boston Data and AI
2:42Saturday on October 3rd and we will have
2:45two pre-cons on Friday. Look it up. Our
2:48today's speaker will be there also as a
2:50speaker.
2:51Uh Paresh is going to drive fly here um
2:54to join. So and Paresh has his own
2:56events on December 5th and I think Julie
2:59also have event. I don't know if I
3:02missed it somehow.
3:03>> October 24th. You got to listen.
3:05>> October Oh, October Oh, sorry. I Sorry,
3:08I messed up some dates. I need to fix
3:09it.
3:10I'll fix it. Uh so SQL Saturday mini So
3:13there are more. Please look it up. Go to
3:15sqlsaturday.com or datamonday.com.
3:18And I'm not going to talk anymore. Thank
3:20you, Chris. I'm going to hand it over.
3:22You cannot unmute yourself. If you have
3:25a question, comments, logistics,
3:28anything,
3:29please put it in the chat. I'm watching
3:31it. Chris is watching it and we'll
3:34respond. So with that I'm going to stop
3:35sharing. All to Chris.
3:38>> Awesome. Thanks, Teo. And uh yeah,
3:41if [clears throat] you're in the Boston
3:42area, um I will be participating
3:45and it will be an extension of this
3:49talk. So it won't be too redundant,
3:51hopefully.
3:53Um so, just want to make sure that I've
3:57got You see the screen? We're good?
4:00>> Good.
4:00>> Everything's good to go. Perfect. Okay.
4:03Um hey folks, thank you again Tayyab,
4:06Paresh, Julie, folks, thanks for having
4:09me and thanks for your persistence
4:10Tayyab.
4:12You know, glad glad we could get
4:13together today. You know, I think that
4:16these events are super important for the
4:17community for folks that are
4:19sort of taking that extra time to
4:21improve themselves, add skills, more
4:24learning, things like that. I'm a big
4:26believer in it. So, glad glad everybody
4:28could make it today.
4:30I I am a director of technology strategy
4:33at Microsoft.
4:34So, I I sit in the manufacturing sector,
4:38you know, where where my main charge is
4:41to align with with senior executives,
4:44CXO, VP
4:46folks and understanding
4:48you know, how their technology strategy
4:50is going to align to the overall
4:52organizational strategy and objectives.
4:55I also teach part-time at Boston
4:58University. I work in in both the the
5:00residential master's program for the
5:02faculty of computing and data science as
5:04well as in the um
5:07the online master's program for data
5:09science.
5:11So, I teach classes in the areas of data
5:13engineering and and
5:16you know,
5:17basically sort of data management, those
5:19types of things, big data engineering.
5:21Um
5:23What else?
5:24Um
5:25Author, speaker, yeah, all that good
5:27stuff. Um
5:29My website's just my last name, bunch of
5:31content up there. Please please feel
5:32free to take a look, reach out on
5:34social, always happy to connect, happy
5:37to give guidance, mentor, things like
5:38that. Please don't be afraid to reach
5:40out.
5:41Um
5:43So, today,
5:44you know, obviously we all know that
5:47there is this massive massive
5:48proliferation of of AI in our world. And
5:51you know, it's it's touching in all
5:52different areas. You know, Agentech is
5:54obviously the the latest greatest most
5:57exciting craze, right?
5:59You know, but starting in the fall of
6:002022,
6:02November when when ChatGPT became live,
6:06right? And GPT 3.5
6:08became a thing and and ever since then
6:10the trajectory has just been insane,
6:12right? Um so,
6:15what happens though is is that
6:17everybody's excited about AI,
6:20but you know,
6:22recognize a failure in how we are
6:26evaluating
6:28what is a good answer. And and you know,
6:31today we're going to talk about some
6:33steps as to
6:35approach that when we look at how
6:39you know, how we want to be thinking
6:41about this and how it relates to the
6:42business. And so,
6:45you know,
6:46we all have situations where
6:49you know, the
6:50the data pipeline looks fine, right? We
6:53we we have no problems. The the accuracy
6:55metrics are very much within the range
6:58that we expect.
7:00We're not we're not seeing any problems,
7:03but we're still
7:05we're having situations where our AI
7:08predictions in our systems just aren't
7:10giving
7:11valuable information or it could just be
7:15you know, slightly wrong, but
7:16>> [clears throat]
7:17>> even even slightly wrong wrong can can
7:20create some major problems, right? So,
7:23so how do how does this break down? You
7:25know, that that's really the question we
7:27want to talk about today in that
7:30you know,
7:31typically we see that folks first want
7:34to blame the model, right? And and
7:37this is [clears throat] natural because
7:39it's it's visible, it's measurable, and
7:41it can be replaced, you know, what do we
7:43do? Well, we we retrain
7:46you know, and and then we you know,
7:48replace what we've got.
7:50Um
7:51Excuse me.
7:53Um and it and it might be, you know,
7:56doing pretty well for a little while.
7:58Uh but then again, we start to see that
8:00drift, right?
8:02Um you know,
8:03it's it's looking at um you know, some
8:05things like stale context or
8:08inconsistent semantics, um you know,
8:10even possibly incorrect workflows, um
8:13business context, right? We We have a
8:15lot of challenges in that area.
8:18Um you know, and and first and foremost,
8:20right? Any of these projects, we should
8:22be working with the business, you know?
8:23I mean, obviously, as technologists, we
8:25like to build things, we like to play
8:27with things. Um you know, but
8:29ultimately, if if we're doing this for
8:31our organizations, um you know, it's
8:33critical to make sure that we're
8:35aligning to um what their needs and
8:37expectations are.
8:39So, really, when when we see um what
8:42happens um you know, as as we zoom out,
8:46um a lot of times, it's not the model
8:49that's the problem. Um it it can be
8:53something upstream, um whether it be a
8:55data definition or um other, you know,
8:58other reasons like that. Um and then,
9:01you know,
9:02also, it could be the way that we're
9:04handling uh you know, the the output of
9:07the model.
9:09So,
9:10when we look at how people interpret the
9:14information,
9:15um you know,
9:17we have to figure out, you know,
9:19what's the right point to to um change
9:23something, right? So,
9:25uh from an analytics uh system
9:26standpoint, um you know, they're really
9:28designed to explain what happened,
9:31right? We're looking at the past. Um
9:33humans are still looking at the results
9:36of, you know, those dashboards and and
9:38things like that. Um you know, we look
9:41at um things like latency is is largely
9:44tolerable, acceptable. Um it of course
9:46it depends on the use case. Um you know,
9:49there are critical things that that
9:51that's not the case, but in large part
9:53we're still seeing um you know, that
9:55that um that
9:58surfaced information and it's okay if
10:00it's from last night, right?
10:02Um you know, and and the idea here is if
10:05we see errors, we're going to let people
10:07know about it, but we may not see
10:09errors. Um
10:11And really um we look at uh the the data
10:15lineage aspect of it. And that's the
10:17critical component here for the
10:19analytics systems. But
10:22you know, when we start to shift toward
10:23AI and ML systems, right? Now we're
10:26seeing, okay, what are the predictions
10:29or recommendations that are the output
10:31of these, right? We've we've done all
10:32this this um uh data engineering and and
10:36and brought all of our data together,
10:38we've trained our model, um and then the
10:40purpose, you know, for whatever we built
10:42it for or whatever we're predicting or
10:43or making those recommendations. Um you
10:46know, and and the the machine itself is
10:51is really what is influencing a
10:53decision, right? The the the prediction
10:56that we get or or the output that we get
10:58is is influencing decision. Um and and
11:01where we can have problems with that is,
11:04you know, if the uh the data is latent
11:06or um you know, we have um errors within
11:10the data or, you know, possibly the data
11:13is stale because we're training on older
11:15data, um where, you know, the I mean,
11:18the ecosystem changes so rapidly now. Um
11:21what was happening in healthcare 20
11:22years ago is very very different than
11:24what's happening in healthcare today. Um
11:26in all aspects of healthcare, right?
11:29And of course, if if we're getting bad
11:32information, we're getting bad
11:33predictions, Uh uh you know, it can
11:35cause some errors, right? And and so,
11:38look at that as decision lineage.
11:40Um and now, of course, with agents,
11:43everybody's excited. Um they are super
11:45helpful in a lot of ways.
11:47Um I've been doing a lot of writing
11:49lately about, you know, sort of my
11:50experience working with um various
11:53tools. Um and and some of the shifts
11:56that I'm seeing in the industry.
11:58Um you know, they certainly can help us
12:00uh put together a PowerPoint faster or
12:03create an image or um you know, help us
12:06analyze some data. Um you know,
12:09we're going to go see Noah Kagan next
12:10summer in in England, and the first
12:12thing I did was get went to chat GPT and
12:14said, "Hey, help me map out this plan.
12:16When should I buy my airline tickets?
12:17Where should we stay?" You know, um I I
12:20want to stay within this budget. And and
12:22it works great, right? It it it gives
12:24you those information or that that
12:26information. Um
12:27you know, but it's still, you know, we
12:30still see hallucinations. We still still
12:32see things that may or may not exist,
12:34right? Um but, you know, when we start
12:36looking at more agentic and we look at
12:39co-work and we look at, you know, work
12:41and we look at uh you know, some of the
12:43newer technologies, Claude Pilot and and
12:46um and Scout, where
12:48uh really the goal here is to execute a
12:51workflow.
12:52Um now we start to see how um the the
12:56machine itself is invoking these tools,
12:59right? And and so, um
13:01the dece the the decisions are are
13:04largely being made by the machines.
13:08Uh and and that um state lasts over
13:12time.
13:14And
13:15if we are using wrong information or
13:19allowing our agents to uh kind of go and
13:23do whatever they wanted, uh you you
13:25folks um may be aware of of uh what what
13:29uh OpenAI disclosed uh just a couple
13:31weeks back um and how um doing some
13:34testing with agents um
13:36it it uh was in its own uh sort of uh
13:40lab uh area. Um it wound up somehow um
13:44it wound up figuring out how to worm its
13:46way out of its contained area um based
13:49on a shared repository uh where they
13:51could get libraries uh and teamed up
13:54with another um
13:56uh another tool that was running its own
13:59uh in its own lab environment. Uh they
14:02together, the two agents, figured out
14:04how to get out to the internet um
14:06looking for answers to questions uh and
14:09and even um even OpenAI uh admitted that
14:12they um they set it up poorly because
14:14they didn't provide some of the data
14:16that they intended to uh and so the
14:18systems were looking for said data uh
14:20and they thought it was part of the
14:21test. And so, you know, they started
14:23working together and then they went out
14:25to the internet and they went and
14:26attacked Hugging Face, right? And so, uh
14:30you know,
14:30>> [clears throat]
14:30>> uh really really good uh lessons learned
14:33uh for OpenAI, of course, uh and and uh
14:36the fact that they were able to disclose
14:38the information is is fantastic. Um but
14:42at the same time, uh how do we how do we
14:44contain that? How do we make sure that
14:46that doesn't happen with our agents? How
14:48do we make sure that we're not sending
14:50um you know, improper messaging or um
14:54uh the amount of um you know, uh
14:56transaction that needs to happen and and
14:58those types of things, right? Um you
15:00know,
15:01when we look at even those those correct
15:03signals that we're getting, um they're
15:05they're going to uh become useless
15:09over time, right? Um you know,
15:12uh a stock quote from 3 hours ago uh
15:15could be very very different now, right?
15:17And and so, if I'm looking at a stock
15:19quote from 3 hours ago and and something
15:21was announced in the market and and we
15:23saw a rapid spike or decrease, um that
15:26quote from even 3 hours ago is is no
15:28longer useful for me.
15:32Um and and really when we think about
15:34the the the freshness of our content, it
15:37it really is a decision constraint,
15:40right? Um you know, we think about how
15:43we define from that decision and and and
15:46go backwards, not from
15:49a generic pipeline um with a service
15:52level agreement moving forward.
15:55Um
15:56One second.
15:59You know, we look at a dashboard uh for
16:02yesterday's data, um that's sufficient
16:05in most cases. Um you know, however,
16:08fraud detection, inventory
16:10[clears throat] allocation, or or things
16:11like um clinical intervention,
16:14probably won't um be okay, right? If if
16:18that data is from yesterday. Uh
16:20inventory levels, if they're incorrect
16:22and we've stocked out, uh you know, when
16:24it's no longer available, um we've
16:26booked a transaction, now we got to go
16:28back to the customer and be like, "Oh,
16:30sorry, you know."
16:32Um we look at, you know, uh the the
16:36uh the most important relevant um uh
16:39measure here is is really the gap
16:41between when the signal was generated
16:44and the point where a decision can still
16:46change the outcome, right? Um and and
16:49really the the observations of that
16:52outcome can happen weeks later, right?
16:54Which is going to be a problem, right?
16:56It's going to cause um some degradation
16:59and and it's going to be difficult to
17:01detect. Um so, you know, we need to ask
17:05ourselves, like, how fresh does the
17:07information need to be um to change our
17:10decision, right? Is information that's
17:13that's from several weeks um going to be
17:16okay
17:17um to to answer the question, right?
17:20Um, and and, you know,
17:22is it as simple as updating a report,
17:24right? Or or is it, you know, something
17:28more critical than that?
17:29So, we really need to think about how
17:32we're taking responsibility
17:35to cover the entire path and ensure that
17:38we're bringing the freshest data where
17:41it's needed and
17:44looking at it the right way.
17:46So,
17:47a little while back, I I wrote this this
17:50framework where, you know, it's it's the
17:52data engineering for AI systems, right?
17:54And so,
17:55it's it's
17:57generally
18:00you know, a guidebook, right? For
18:03helping to evaluate and ultimately
18:08evolve the way that we're approaching
18:10these challenges. You know, we can look
18:13at different technologies that that
18:17you know, can implement each of the
18:19layers
18:21and then the responsibilities are what's
18:23left over. You know, who is responsible?
18:26We can use the model to identify where
18:30some of the assumptions that we're
18:31making
18:33are being brought forth and um
18:38where there is risk around that drift
18:41and and where accountability might be
18:44missing for some of these areas, right?
18:47And so, the six layers will remain
18:49intact even as
18:52you know, generative and energetic AI
18:55becomes an extension of the kinds of
18:58systems and work that we build and do.
19:02So, when we look at the six layers of
19:04responsibility,
19:06you know, we start of course with the
19:08data, right? Where is the data coming
19:10from? How are we bringing it in, right?
19:13Looking at
19:14after we've acquired the data, looking
19:16at the quality of the data, you know,
19:18what needs to be done there.
19:20As we go to build our models, you know,
19:23or train our models, we we've of course
19:26have the feature engineering,
19:28the context engineering,
19:30you know, as we start to go to more of
19:33this graph rag architecture that's
19:36becoming more and more prevalent. If
19:38you're familiar with the Microsoft IQs,
19:42you know, so you've got
19:43work IQ and and fabric IQ and foundry
19:46IQ.
19:48They're just an example. Palantir has
19:51its foundry, right? Where you have that
19:54rag layer, so you have your your your
19:57baseline information and and
19:59documentation. And then on top of that,
20:01we have a graph layer that
20:04you know, is a knowledge graph and and
20:06can relate context to our data. So,
20:10whether business context or what have
20:12you.
20:14And then we have our operational data
20:15systems, right? Where are we bringing
20:17the data to?
20:19Governance, lineage and trust. And then
20:21finally, observability in the feedback
20:24loops. And that's going to be
20:26our real key point here as we move
20:28along.
20:30So,
20:31you know, when we look at the data
20:32sourcing and capture, capture, right? We
20:35want to make sure
20:36what we we are clearly defining what
20:40signals are entering the system, you
20:41know, what's the grain, what are the
20:43assumptions we're making, right?
20:46And then when we see failures, right?
20:48Critical signals are lost before the
20:50model sees them.
20:51You know, so data isn't
20:54in line with with what's required.
20:58And then the aggregation of the data
21:01masks the detail, right? And and so it
21:04gets drowned out um all we see is the
21:07the the the surfacing of that.
21:09And then you know, we think about sort
21:11of something like a
21:13model sees the daily totals, but it
21:16doesn't have the time of day behavior,
21:19right? So you think of like a retail
21:21scenario where you have your your data
21:26that's making recommendations throughout
21:28the day,
21:29but it's not seeing that granular
21:32information that can help with
21:36you know, the the various activities
21:38around how we respond to customers
21:41coming in with promotions and whatnot.
21:45So the question you really you know,
21:46start to ask is
21:48you know,
21:49what are the assumptions about the
21:51behavior are embedded in how the data is
21:53captured and
21:56which decisions now depend on them,
21:58right?
21:59When we look at a supply chain example,
22:02it shows how
22:04capture problem can masquerade as an AI
22:07problem, right? So
22:10for an example here, we have a global
22:13electronics
22:14manufacturer and this is an actual use
22:16case. It's kind of a bit of an
22:18amalgamation of a couple,
22:20but you know, it wants to use AI across
22:23sourcing and and risk alerting and
22:24inventory management.
22:26Tons of money invested.
22:28You know, the dashboard's great and the
22:30executives really expect you know, big
22:33things from it, right?
22:35Reports
22:37are improved, right? Dashboards still
22:39look healthy. You know, we're we're
22:41creating confidence in the system,
22:44but then we start to see
22:47a little bit of cracks here and there,
22:49right? We see
22:50you know, it's maybe noisy
22:54information is coming out or or perhaps
22:56duplicated data or signals or
22:59you know,
23:00we start to see that and it starts to
23:02erode confidence a little bit, right?
23:04Um, you know, obviously LLMs have gotten
23:06much, much better with hallucinations.
23:09Um, it's still a challenge and um, you
23:11know, for everything we do, there's
23:14there's still a trust but verify
23:16mechanism. There's just far too many
23:17stories in the news about um, just
23:20trusting what the output is and and um,
23:22all too often we're still running into
23:24challenges.
23:26So, when these reports are are um, or
23:29these alerts are repeatedly wrong, um,
23:31we start to see um, users adapting and
23:35saying, well, I don't trust this, right?
23:38And so, they don't use the tools as a
23:40result.
23:42Um, and then of course,
23:43um, when when that behavior changes, it
23:45becomes just another failure in the
23:47system, right? In this case, it's not a
23:49failure of technology, it's it's a
23:51failure of of process and procedure.
23:54And so, in this case,
23:56we're not concerned about building a
23:58better model, right? Um, we need to look
24:01at um, the signal capture and identify
24:04where there could be challenges before
24:07we optimize any kind of of um,
24:11prediction.
24:14And what's important to know is that
24:16even correctly captured data may long
24:19may no longer be fit for the decision,
24:21right? So, we look at curation and
24:23quality, right? Traditional uh, data
24:26quality asks whether the value is
24:27present and valid, right? AI quality
24:30also asks whether remains
24:32representative, right? Again, healthcare
24:35data from 20 years ago, uh, may not be
24:38um,
24:39uh, something that we want to use to
24:41train our model with, right? Uh, a
24:44schema can can uh, pass while
24:47um, you know, the the population, the
24:49policy, channel mix or or the behavior
24:52has shifted, right? Um, we we have to
24:55think about structural correctness,
24:58um, and it's necessary but not
25:01necessarily sufficient for a decision
25:04system.
25:05And and really we need to hold the
25:08quality expectations
25:11to be tied to the population
25:13and those actions that the AI is
25:17affecting, right? So what predictions
25:19are we making? What recommendations are
25:21are we making as a result?
25:23And we need to ask what evidence is
25:25going to reveal
25:27that that production no longer resembles
25:30the world the way it's encoded in in the
25:34training data.
25:36A good example of this is an Amazon
25:39hiring
25:41example that shows how structurally
25:44valid
25:45but
25:47unfit data causes a challenge, right?
25:50And so Amazon created a tool that looked
25:55at
25:56you know the resumes of tens of
25:58thousands of employees, right?
26:01And we
26:03you know they they they saw that they
26:05were legitimate records, they were good
26:07employees,
26:08and so
26:11there was nothing wrong with that per se
26:13except that
26:15the the bulk of of the employees that
26:18they were looking at were all male,
26:20right? And and so
26:22you know any kind of diversity could be
26:25actually
26:27held against applicants, you know? So
26:31you know if if there was anything that
26:33that highlighted you know
26:35I don't know I played softball in
26:37college, right? Or or something like
26:39that, it could actually sort of work
26:43against their case for being a good
26:46engineer because of how the model was
26:49trained.
26:51Now everything looked fine, right? The
26:53data was moving fine.
26:55Uh you know, and and everything was
26:57accurate, but it didn't take into
26:59account
27:01um uh what
27:03what a good hire looked like, no matter
27:06what their, you know, sex or race or
27:08anything was, right? And so,
27:11really the lesson is to, you know, test
27:14um the representativeness and um any
27:17kind of subgroup behavior um not just
27:20nulls, right? Um or or types of data and
27:23distributions, right?
27:25And so, you know, when we when we think
27:27about it after the deciding what the
27:30data is that's that's fit for the role
27:32here, now we need to preserve what it
27:36actually means, right? And so, that's
27:39where our our semantics come in, right?
27:42Um the feature engineering component of
27:43building a model has always um required
27:47um the consistency of the definitions
27:49between uh our our training and our
27:52serving, right? Um you know, generative
27:54and agentic systems only broaden the
27:58responsibility, right? We look at
27:59instructions or retrieval state, tools,
28:03um and and of course permissions that
28:05that will also shape the meaning of the
28:07output. Uh you know, things like
28:10customer or active or risk, um they they
28:14they don't silently change across teams
28:17or environments
28:19um or or Sorry, they they
28:21they do, but we don't necessarily see
28:24it, right? Um and and the the context
28:27itself needs to be sort of versioned and
28:30tested. Um somebody has to own it,
28:34right? What does uh a customer
28:36represent? Um what does an active
28:39customer represent, right? Does sales
28:41look at um active customers as somebody
28:44who's bought something in the last 24
28:46months, whereas accounting says an
28:48active customer is somebody who bought
28:49something in the last 12 months, right?
28:51There needs to be some kind of
28:53discussion and synergy as to what those
28:55definitions are.
28:58And, you know, really the the key
29:00question here is whether every critical
29:02input means the same thing everywhere
29:05it's used, right? So, whether it's
29:07across business units, whether it's
29:09across departments, you know, wherever
29:11it's used, you know,
29:12do they mean the same thing, right?
29:15And and when we look at it that way,
29:17you know, training
29:19serving skew, right, is is one of those
29:22classic examples of the problem, right?
29:25So,
29:26Uber,
29:27you know, used
29:29an offline performance system that, you
29:32know, it it looks great when we're
29:35training the features and and you know,
29:37the the internal aspects of it, but when
29:41we go to production, right, the failure
29:44only pops up when the serving path
29:47computes the the same-named features but
29:50differently, right? And so, again, if if
29:53the you know, two different groups look
29:55at the same term in different ways,
29:58you know, that needs to be defined
30:00properly.
30:02The hard part, beside the fact that
30:05you know,
30:06every aspect of the business looks at
30:08things different ways,
30:11you know, is that the pipeline looks
30:14fine. It's it's most likely working just
30:16the way it's supposed to
30:18and and you know, that's the way it's
30:20been developed, right? And
30:23we can retrain
30:24and it it might be okay for certain
30:27situations, but
30:29that same mismatch will pop up again
30:33because
30:34you know, the the structural fit is is
30:39to have those shared uh
30:42And then of course, when we when we do
30:44the implementations and the validation
30:45across these environments, uh again,
30:47those have to match, right? And so, um
30:52as we move to like LLM and agentic
30:54systems, um they have to or or they're
30:57going to have a very similar um failure
31:00mode, right? So, um think about the the
31:04modern equip equivalent being context
31:06skew, right? Um, uh
31:09again, this is this is, you know, sort
31:10of illustrative based on patterns, based
31:13on findings from places like McKinsey
31:15and things like that. Um,
31:17but the eval evaluation um often uses
31:21like curated documents, right? Um,
31:23things like know known tools, um
31:26uh various permissions that are stable
31:28and and um you know, these these
31:30carefully
31:32um
31:32uh curated instructions.
31:35And um the production retrieval side of
31:38it uh can be stale or um you know, the
31:41rankings can change, um the workflow
31:44state um can drift um from from where we
31:48started and and of course, permissions
31:50are going to differ by user and they're
31:52going to change, right? Permissions
31:54aren't going to stay the same at all
31:56times. And so, how does that impact our
31:58our model? How does it impact um our
32:01agentic systems? What are they doing? Uh
32:03you know, and how do things change as a
32:05result?
32:06So,
32:08we think about it, the the model version
32:10alone is is not going to be enough to
32:13reproduce or explain
32:16um the outcome, right? Um, the the you
32:20know, important thing is how context um
32:24becomes part of the deployed system, not
32:27just an aside, not an accessory to that,
32:30right? Context is is that important.
32:34So, when we look at um the the the parts
32:38of the system that need to be
32:39considered, right? We've got our data
32:41and our instructions, um the retrieval
32:43knowledge as we we grab that data, um
32:46the state of it, um what tools we're
32:48using, permissions, and that's all going
32:51to lead to AI behavior, right? Um each
32:55of these elements, of course, can
32:57independently change the model's
32:59behavior, right? So, any one of these
33:02that changes along the way can you know,
33:05uh uproot our entire process and model.
33:08Uh you know, tools and permissions are
33:11especially important because they
33:13determine what the system can do, um not
33:17only what it can say, right? Um and so,
33:21when we see things like reproducible
33:23evaluation, it's going to require
33:25capturing um versions and um some of the
33:29values uh for these inputs that we're
33:32we're worried about.
33:34And, you know,
33:37we we need to keep in mind that we can't
33:40evaluate an agent independently from its
33:44operating context, right? That's not
33:46enough.
33:48Uh you know, and and now we're starting
33:49to see a lot more AI control towers um
33:52with the capability for um you know, not
33:55only uh what what transactions the agent
33:58is doing, but also what data is it
34:00accessing, um where is it getting data
34:02from, where is it sending data to, um
34:05what permissions is it using in order to
34:09um make these transactions, um where are
34:12the trouble spots that um could be
34:14catastrophic to the organization.
34:17Uh you know, folks may have heard um I
34:20the name of the company escapes me, um
34:22but they had set up an agentic process
34:25internally, uh and it uh didn't have the
34:29proper credentials to a database, and so
34:31it found a workaround, hacked the
34:33database, and deleted the database. And
34:36they had to restore from data that was 3
34:38months old because they didn't have any
34:41uh recent copies. They were also
34:42destroyed.
34:44So, you know, when we think about again
34:47the context, it's going to be delivered
34:50through the infrastructure design as we
34:53look at the execution.
34:58Now,
34:59looking at our operational uh data
35:01systems, right? I mean, uh this could be
35:03a data lake, this could be a database,
35:05you know, um
35:06we're starting to see data um wind up in
35:09a variety of places. Uh the analytics
35:12infrastructure is is certainly um
35:15optimized for throughput um and and of
35:18course availability, right? Uh whereas
35:21decisions may require bounded latency
35:24and determinism, right? So, agentic
35:27workflows um add this this durable
35:30state. Um it'll it'll also do retries.
35:34Um it'll look at tool failures,
35:36permission changes, and long-running
35:38executions, right? So, um folks that are
35:41using anything like uh you know, co-work
35:43from Claude or Microsoft, looking at
35:45work um from from uh OpenAI, uh it's
35:49telling you what it's doing, and it
35:51tells you when it runs into a failure.
35:53Hey, I couldn't use this application, so
35:55I'm trying it this way, right? And it's
35:57walking you through what it's doing, um
35:59but we have to be very, very careful to
36:01make sure that uh it's doing uh what it
36:05should be, right? And it and it's not
36:06looking to um
36:08find a a workaround to a problem causing
36:11a bigger problem.
36:13Um you know, when when we think about a
36:17demo when we're evaluating system, it's
36:19it's not going to really uh prove out
36:23that um this this full workflow is going
36:26to successfully um survive these
36:29handoffs as we move the the process
36:32along
36:34or that you know
36:36once once we hit a point that there's a
36:38stall of some sort,
36:41how do we know that it's going to
36:42correctly resume later on?
36:45I'm sure folks here have run into the
36:47situation where you start a process in
36:50in one of the popular tools. It says,
36:53"Okay, I'm I'm generating this for you."
36:56And it just stops, you know, and so the
36:59system has fooled itself saying, "We've
37:01started this." But it never actually
37:03started, you know, and you got to nudge
37:04it. But if you walk away thinking that
37:06it's doing your thing and it's going to
37:07take a while, right? Then you come back
37:09and you're like, "Oh, well, that was
37:10just a waste of an hour." Right? So, run
37:12into that, right?
37:14Um you know, we think about
37:18how we look at state and how we persist
37:21it, right?
37:22You know, what what um
37:26what actions are made in independent,
37:29right? Um
37:30where are we seeing fallback or human
37:33reviews occur, right? These are all
37:36things that that need to be evaluated as
37:39we're rolling out these systems to
37:41ensure that
37:44you know,
37:45we we can avoid these these critical
37:48challenges, right?
37:50When we look at healthcare, obviously
37:52super critical, right? It looks at
37:55the cost of of these closed decisions
37:59and it's easy to see
38:02where these these challenges can occur,
38:04right? So,
38:05we we look at
38:08um healthcare infrastructure and and we
38:11see that you know, a model can be
38:13accurate and the batch job finishes just
38:16fine on schedule,
38:19but the recommendations themselves still
38:21arrive too late. Uh, you know, clinical
38:24decisions have a very specific window.
38:27Um, you know, uh, just using daily
38:30doesn't mean that it's a meaningful
38:32service level agreement or SLA uh,
38:35without reference to the workflow,
38:37right? If if if, you know, daily is um,
38:41you know,
38:42needs to be done at a specific time,
38:45right? That needs to be specified. Just
38:47you can't say, "Hey, just run this once
38:49a day, right?"
38:50It doesn't know what the right time is,
38:53right? And it can run whenever as long
38:54as within that 24-hour period, right?
38:56And that's how it's going to evaluate
38:58it. Um, you know,
39:01we look at some of these these source
39:03systems, they can be really fragmented
39:06um, and and um,
39:08they we have challenges with the
39:10integrations. Uh, it takes longer um, to
39:14get the data integrated because
39:16uh, you know, they're coming from
39:17different places. Um, if the data
39:19doesn't match up probably uh, uh,
39:21properly, we run into more and more
39:23challenges um, and it can take too long
39:26to get that
39:28so that the the result becomes
39:31meaningful, right? Um, so we want to we
39:34want to make sure that that we are
39:35measuring the latency end-to-end, right?
39:38And that starts from the signal
39:40generation um, to the human or system
39:43action, right? And which one is it? Um,
39:45and and not just within one of the
39:47services, right? Um, once the the
39:50systems influence or um, execute on
39:55decisions or or uh, workflows, the the
39:58traceability really has to um,
40:00uh, extend far beyond just the data,
40:04right? So we see that the data gets fed
40:06back, um, but maybe we don't catch the
40:09results of of the recommendation and the
40:11action taken.
40:15Um, um, lineage, and trust, of course,
40:17super critical, right? Um, the the
40:19lineage is going to answer um, where um,
40:23the the in input came from, the data
40:25input, the human input, whatever it was.
40:28But, it's not going to explain why an
40:31outcome occurred, right? So, we don't
40:34know what the outcome was or or, you
40:38know, good or bad. I mean, yeah, we have
40:40um, you know, a thumbs up or a thumbs
40:42down. I don't know if anybody's noticed,
40:44but ChatGPT removed the thumbs down. Um,
40:48I don't know whether they just don't
40:49want our feedback anymore, but it's
40:51gone, right? And so, that's problematic.
40:54Um, the the decision lineage, right?
40:57That that second step is going to help
41:00us add the context and and those
41:02business rules. Uh, we're going to look
41:04at the the version of the model, um, you
41:07know, any of the evidence that we have,
41:09uh, and and then sort of what was the
41:12behavior? Um, how was it approved after
41:14that choice? Um, now, the action lineage
41:18on the other hand, um,
41:20adds the the things like the the actual
41:23execution of the tool, the parameters
41:26that we use, um, what the results were,
41:29uh, who's involved, and if there's any
41:32additional, um, overrides that need to
41:34happen, right? Um, you know, sorry I
41:37couldn't do this. Um, I could do it with
41:39this tool instead, uh, you know, and
41:41that type of stuff, right? We We We run
41:43into those kinds of things, especially,
41:45um, with more of the automation.
41:48We need to look at, um, you know, the
41:50governance aspect, um, to ensure we have
41:53the accountability, um,
41:55to be visible across the organizational
41:58boundaries, right? Um, it it can't be,
42:02um,
42:02a simple catalog entry. We need to be
42:05able to make sure people see this, uh,
42:08you know, to to um, uh, to the masses,
42:11if you will, right?
42:12Um,
42:13when we think about, um,
42:15the, um,
42:18uh, the,
42:20sorry, brain fart. Um,
42:22we we think about the compass, uh,
42:23program, right? And it shows the
42:25consequences of how the decisions were
42:28made without, um,
42:29the meaningful, um, decision decision
42:32lineage, right? So, anybody who's maybe
42:34familiar, um, with the compass program,
42:37you know, it's it's a historical example
42:40of, um, algorithmic stores, uh, uh,
42:43scores influencing sort of
42:45high-consequence decisions, right? Um,
42:49the system itself could produce a number
42:52reliably, right? But, the governance
42:55question became, um, whether the people
42:58who were affected by this, um, could
43:01understand or challenge it, right? Um,
43:04the the proprietary implementation, um,
43:07and the minimal traceability,
43:10um, made accountability of the model
43:13much less, um,
43:16available, right? And so, if we don't
43:19know how these decisions are being made,
43:21especially in in, uh, you know,
43:23something as as, uh, important as court
43:25decisions, right? Uh, if if you don't
43:28have that information, um, as to how the
43:31decision was made, it's just a black
43:32box. And, you know, it's hard to know
43:35what what the answer is. Um, you know,
43:37the the the debates on accuracy, uh,
43:40they're not going to replace the need,
43:42uh, to document the evidence. Um, we
43:44need to make sure we're keeping track of
43:46the ownership, the use,
43:48um, and of course the appeal pathways,
43:50right? All of that information, um,
43:52uh,
43:53all of those details are required of the
43:55information as to how the decision was
43:57made, right? Uh, and and this is still
44:00something that's that's going today,
44:02right? Um, you know, uh, uh,
44:05ProPublica is the one that surfaced this
44:07situation,
44:08but you know
44:10the the
44:13the governance question is going to
44:15persist, right?
44:17Um
44:18And beyond that, right, we need to think
44:20about how the traceability is going to
44:23tell us what happened, but the
44:25observability is going to tell us
44:28when the system starts to become wrong,
44:30right? And so that brings us to our
44:33sixth layer right where
44:36the infrastructure in the model
44:38telemetry
44:39are certainly necessary,
44:42but
44:43there are situations where it's
44:45incomplete.
44:46Um you know, this stable prediction
44:49distribution can can certainly coexist
44:52with declining business outcomes or or
44:56any any kind of challenges around
44:58segment level effects.
45:01But
45:02we look at agentic systems again and
45:05they can succeed technically, they can
45:07get the job done.
45:10Things like a valid response or or a
45:12successful tool call. Um but it can also
45:16create more rework reversals or
45:19repeat contacts into the system, right?
45:22And so you know, again we think about
45:25you know, hey, help me help me
45:28clean up this email or you know, help me
45:31describe this this
45:33document, right? And and an hour later
45:36you find yourself fighting with the
45:38system because it just keeps giving you
45:40more and more suggestions about how to
45:41improve what you're doing. You know, and
45:44at some point you just got to tell it to
45:45stop. I don't want to keep iterating on
45:47this.
45:48You know, and and it's just because it
45:50is giving responses back, but now it's
45:52also trained to tell you you're amazing,
45:54of course, but then also give you the
45:57information
45:59that that could improve this, right?
46:00It's designed to be sticky and have you
46:03continue to use the tools.
46:05And, you know,
46:07when when we are able to define both
46:11what the triggers of the intervention
46:13are and who owns the system,
46:16we're able to avoid a lot of these
46:18challenges, right? So, again, you're
46:20you're hearing ownership a lot,
46:23observability a lot, right? And and
46:26especially in this day and age of AI,
46:27everything is moving so fast
46:30that if we're not keeping a close eye,
46:33we can really really have some some
46:35major consequences, right?
46:37So, the observe the observability
46:40surface is now able to a span the entire
46:45decision system, right? And so,
46:49if we look at how we make these
46:51decisions,
46:52you know, we can look at how each of
46:54these stages, again, can can change on
46:57their own, right? They're they're not
47:00reliant on each other to change
47:03and and any one of them can break
47:06the the workflow here, right?
47:09We we want to make sure we capture
47:13the the data and the context, right?
47:16What's the latest version of the
47:18context? What's the latest version of of
47:20that meaning, right? You know, we we
47:23want to look at how the model's
47:25configured, you know, what what tools
47:28we're using,
47:30the the state of the workflow itself,
47:34you know, how how we're
47:36rationalizing the decisions that are
47:38being made,
47:39what the action result is and what the
47:42eventual outcome is, right? If if and
47:45and and in large part, machine learning
47:47models are not capturing that action or
47:50eventual outcome, right? Which is which
47:53is where we're losing critical
47:55information.
47:57We also want to make sure that that um
48:00we're we're keeping an eye on things
48:01like uh approvals and retries and
48:03failbacks and and of course cost, right?
48:06Everything's um you know tokenomics
48:08these days and and how much is it
48:09costing um to run this co-work job? How
48:12many tokens is it going to talk ca- uh
48:14cost? And of course making it even more
48:16challenging input tokens versus output
48:18tokens and you know um them not being
48:21balanced and and the expenses behind
48:23that, right? Um and and you know um
48:27it we we want to make sure that we're
48:29not inundating ourselves with logging
48:32around this, but we want to make sure
48:34that we have uh the sufficient um
48:38uh amount of evidence to uh reproduce
48:41the challenges that we have and um
48:44diagnose how um how to take action,
48:48right? How to how to resolve the issue.
48:51And so now
48:53you got to think about it as, you know,
48:56um
48:57why did the system take this action,
49:00right? How how do we determine why it
49:02took the action and then change that
49:04behavior if necessary?
49:06Uh and so, you know, outcome monitoring
49:09is really where a lot of these deployed
49:11systems just uh just
49:15either it doesn't exist or it's weak,
49:17right? And so that's one of the critical
49:18areas that we need to focus on.
49:22Um another example here for United
49:24Healthcare, right? Um
49:26you know, uh
49:27we we had that um
49:29uh
49:31the the operational scale and the fast
49:33decisions, um but they can look like
49:36success, um you know, and and appeals
49:39and reversals of course are are outcome
49:42signals when we talk about insurance um
49:45uh
49:46claims and things like that, um but if
49:49we if we if we don't loop back in those
49:52outcomes, right? Um if if we're not
49:55taking the results and feeding it back
49:57into the model so that the model
49:59understands what is happening, uh you
50:02know, basically it
50:04we're going to outpace um the the
50:07meaningful review process, right? And
50:11and it's going to magnify the errors
50:13that we run into um long before um the
50:16patterns become visible, right? And so
50:19we really need to think about how we
50:21connect recommendations and the actions
50:23that are taken um to those outcomes, uh
50:27disputes, corrections, whatever they
50:29happen to be, um and sort of how do we
50:31feed that back into the system so we
50:33take that into account as we start to
50:35retrain the model.
50:38So if we look at it, um you know, all as
50:41a single propagation chain, right? Um
50:43you know, a small upstream assumption
50:46can change while every single downstream
50:49service continues to function, right?
50:51The model may remain confident because
50:54it, you know, it it it's a model, right?
50:58And and it's doing what it's supposed to
51:00do. It's what it was trained, right? Um
51:02but um
51:04when when that decision gets translated
51:06into a human or an automated action, um
51:09and then, you know, finally into these
51:12now degraded outcomes,
51:14um you know,
51:15we're just going to see this accelerated
51:17rapidly as we go into autonomy, right?
51:21Um so we want to make sure that um you
51:24know, technical health can coexist with
51:28accumulating harm, right? Um back to the
51:30very beginning. Uh your dashboard is
51:32green. There's no problems, um but
51:34you've got problems lurking beneath the
51:36covers um because we're not capturing
51:38all the right information.
51:41Um and really, yeah, scale turns um you
51:44know, the the the hidden mismatches into
51:47massive visible failures when they could
51:49be mitigated long before time.
51:52>> [snorts]
51:53>> Um so, why do these appear appear at
51:55scale, right? Um you know, automation
51:58removes the the human um uh guts, right?
52:02Gut check, right? Um uh
52:04this is this is something that AI still
52:06can't replace and probably won't be able
52:07to replace for a long time. Uh it
52:10doesn't know what humans are going to
52:11do. Um you know, the the population
52:14behavior changes uh and you know, uh
52:18if if assumptions are made about the
52:20data, um then you know, that confidence
52:22gets eroded. Um delayed outcomes can
52:25hide um the degradation of of um the
52:28results and
52:30possibly long after many decisions have
52:32already been made,
52:34you know, system health monitoring um
52:36will show that that the systems are all
52:38up, but it's not going to show
52:39correctness of the of the decision that
52:40we're making.
52:42Um you know, again, autonomy is going to
52:45um just increase the problem, right?
52:47It's going to um exacerbate um what the
52:50output is, and if it's wrong, it's going
52:52to get worse and worse. Um And and it's
52:55important to note that, you know, the
52:57scale itself doesn't create the flaw,
53:00right? It's just going to reveal it and
53:03multiply it more quickly.
53:05So, when you think about these dynamics,
53:07right? They're going to make um very
53:10similar uh familiar engineering habits
53:14um that may become dangerous, right? So,
53:17things like treating it as a modeling
53:19problem or reusing um analytics
53:22pipelines for inference, um they may not
53:25be sufficient. In fact, they probably
53:27aren't um based on what the use case is,
53:29right? Um thinking about um governing
53:32our our data sets, but not the actions
53:35that are taken, right? Not not observing
53:38or or monitoring what those actions are.
53:40Um you know, if we expand uh that
53:43autonomy um before we understand what
53:48the actions that were taken,
53:49uh you know,
53:50again, we're exacerbating the problem,
53:52right? Um
53:54>> Hey Chris,
53:55>> Yeah.
53:56>> 6 minutes to go.
53:58>> Got it. Thank you.
54:03Um so,
54:04you know, when when we think about how
54:07um our predictions are are made, right?
54:09It's going to inform us of information.
54:11Um the recommendations are going to
54:14influence our decisions. Um you know,
54:17the when we make those decisions, we're
54:19making that final commitment, and then
54:20of course, the action itself is going to
54:22change what we're doing, right? Um the
54:25model can be identical, um but the
54:27consequences of of an error along the
54:29way uh isn't, right? And so, we want to
54:32make sure that, you know, we think about
54:33how reliable systems are going to
54:35respond um by engineering the the
54:38complete responsibility chain, right?
54:40So, what does that look like, right? We
54:42look at uh the right signals at the
54:45right time. Um we're we're governing
54:47outcomes, not just the data sets. We're
54:50protecting the meeting and the context,
54:52right? Um across the organization. Uh
54:55we're measuring what happens after the
54:57prediction is made, right? Um we're
55:00we're designing for decisions and
55:01actions, right? As opposed to
55:04um you know, uh
55:06what we think the model should produce,
55:08right? Um and and we want to match the
55:11autonomy to the accountability. Um you
55:13know, uh depending on the action, human
55:15actions should still be very very
55:17heavily involved.
55:19We look at sort of an updated end-to-end
55:22um you know, loop for the system, right?
55:25Now, we we see that the prediction is is
55:28an immediate event, not the end product.
55:30Uh sorry, an intermediate event, not the
55:32end product. Um we we look at how the
55:35context is going to determine how the AI
55:38interprets the data, uh the decision
55:40itself is going to determine
55:43what's going to happen next, and the
55:45action is going to change the
55:46environment. Our outcome is going to
55:49provide the evidence about whether
55:52decision actually created value or harm,
55:55right? If we don't capture that outcome,
55:57we're never going to know.
55:58Of course, we get our feedback updates,
56:00right? And and then we see, okay, now we
56:02need to update our data or our rules or
56:04our context or whatnot, right? We we get
56:06that information.
56:07And and when we measure it, we got to
56:09look at more than just accuracy of the
56:11model, right? We need to make sure that
56:14the decision improved the outcome.
56:17You know, we need to make sure that
56:19we understand the consequences of the
56:21errors, and you know, how can we detect
56:25them and respond to them very, very
56:27quickly.
56:28And and the last main point here is
56:30again about ownership, right? So, when
56:33we think about the layers of each of the
56:36stacks, right? These are going to be
56:39owned by various aspects of the business
56:42in the organization, right? And so,
56:45certainly as as a data engineer, a DBA,
56:48a data analyst, an ML engineer,
56:51whatever,
56:52you know, you're going to need to know
56:53what the data is, but if you don't get
56:55the impact or the feedback from the
56:57business, you're going to run into a lot
56:59of challenges, right?
57:01You know, it kind of going through
57:04here's just some of the steps of the
57:05framework itself. You know, I put it on
57:08one of the pages and I'll put it here at
57:09the end. If you want to go grab it and
57:11read it and give me some feedback on it.
57:13But you want to take through, you know,
57:16how how you can have this reliably.
57:19Here's a quick sheet. We can get this
57:22out to the community, these slides, so
57:24they can, you know, take this assessment
57:26themselves, look at the maturity of
57:28where we're at.
57:29You You and and thinking about, um, you
57:32know, before you build a missile system
57:33itself, understanding these aspects of
57:36making the decision, right? Um, one of
57:40the things we run into very frequently
57:42in technology is, um, you know, this is
57:45going to cost a lot of money or we need
57:48to we need to get some wins quickly um,
57:51before we can do anything else. And and
57:53if we have that challenge, um, then uh
57:57we're going to need to be able to push
57:58back and say, "Well, no, but this is
58:00really important. The model may not be
58:03the hardest part, right?"
58:05Uh, just some framework in 30 days. That
58:08should help, right? Before you go to
58:10production, ask yourself these
58:11questions. And
58:14that's me.
58:16How's that for a rush? Pretty good?
58:17>> Sorry. [clears throat]
58:18Uh, Chris, there's a question.
58:20>> Uh, oh, is there? I didn't see it pop
58:22up.
58:22>> Yeah, just now from Carl.
58:24I said, "Could Chris explain the
58:26difference between human in the loop
58:28versus human on the loop?" Uh, just
58:30before you go there for others, uh, you
58:33know, the presentation is over. I'm
58:34going to put the
58:36uh,
58:37recording and everything. But if you
58:39want some Q&A, if Chris has time, uh,
58:41we'll take it. Otherwise, thank you very
58:43much because I know it's a working hour.
58:45You might have a meeting. And you might
58:47have other things to schedule. So, the
58:48main presentation is done. Chris is just
58:50going to take couple of question and
58:51answer if you have and if Chris has
58:53time. So, but thank you for joining.
58:56>> Yeah, thanks all. Appreciate it. Uh,
58:57love your feedback. Hook up with me on
58:59social social and whatnot.
59:01Um, human in loop versus human on the
59:03loop. I've actually never heard the
59:05phrase human on the loop. But if I were
59:07to guess at what it is, human in loop is
59:10going to require some kind of action
59:12from a human, right? So, a button has to
59:15be clicked or something has to be
59:16approved, declined, whatever.
59:18Um, and then human on the loop is,
59:20again, I'm guessing, but observability,
59:23right? And and keeping an eye on,
59:25um, you know, what we're what we're
59:27seeing there and making sure that you
59:29know whether it's it's agents or you
59:32know what whatever is moving ahead is
59:34taking the appropriate actions. That's
59:36my best interpretation.
59:39Oh yeah, look at.
59:41It's funny my
59:42my chat wasn't popping up Tayyab for
59:44some reason. So I didn't even see.
59:48>> It can't say that
59:50that was my assumption as well. It was a
59:52new term for me also. So looks like Carl
59:54is okay.
59:56>> That's all I can interpret it.
59:58>> Yeah, JD Walker has an interesting
1:00:00question and Chris if you don't have
1:00:02time please let me know. Thank you for a
1:00:04very thorough and informative
1:00:06presentation.
1:00:07I'm the only and on
1:00:11prem I mean on premises DBA with no
1:00:13cloud experience and I feel I still need
1:00:16AI. So he's looking for a comment. It's
1:00:18not a question but I'll decide whatever
1:00:20you want to say. So
1:00:22>> Yeah, sure. Yeah, and and you know JD I
1:00:26I feel you man.
1:00:28You know, I I
1:00:30Tayyab and and some of these other folks
1:00:32I'm sure could explain some of the
1:00:33benefits that we're seeing from database
1:00:35intelligence. Is it going to completely
1:00:38replace the skills of a good DBA
1:00:41tomorrow? No, right? But it it certainly
1:00:43made a lot of strides. We see this in a
1:00:46lot of the cloud versions of the tools.
1:00:49I'm not sure if if SQL 2025
1:00:52has the intelligent
1:00:54query writing and stuff. Tayyab, do you
1:00:55know?
1:00:56>> It does have a we had a lot of talk in
1:00:58our local user group. I'm actually
1:01:00speaking about it on September 9th at
1:01:02our local MTC.
1:01:04Like what tools DBAs have
1:01:08without doing anything of course with
1:01:10your credit card that Microsoft is
1:01:12providing. Of course you know, I work in
1:01:13Microsoft product and I'm a Microsoft
1:01:15MVP as you know. So I'm going to only
1:01:17talk about on this. And yes, SQL 2025
1:01:20has everything built in. It is all
1:01:22public. You just look up Bob Ward 2025
1:01:24sessions.
1:01:25You can download that. You can run
1:01:27everything. You can have a chat You can
1:01:29have chat through T-SQL. You can connect
1:01:32to the model. You can do create vectors.
1:01:35And you can do intelligent search or or
1:01:37or hybrid search all within the
1:01:39framework of SQL Server 2025 taking all
1:01:42the latest and greatest.
1:01:45Uh and you asked the other question
1:01:47um you know, about I do not believe
1:01:50we're going to be replaced, but we just
1:01:52have to work in a different way. Our
1:01:54skill set needs to change. We'll have
1:01:56different ask from the management to
1:01:58perform than what we've been doing. So,
1:02:01uh I would say if you don't embrace it,
1:02:04somebody else will or AI is going to you
1:02:06know, like
1:02:07Sorry, whoever is going to embrace
1:02:09probably will take your job, not AI. So,
1:02:11I think that's what I meant. Yeah.
1:02:13>> Yeah, I I I'm actually I'm going to try
1:02:15to coin the phrase the task shift,
1:02:17right? Because we're we're we're taking
1:02:19what we are used to doing on a daily
1:02:21basis, automating some of that, and then
1:02:24we're focusing on other areas and other
1:02:26tasks, right? Um it doesn't mean that uh
1:02:29some other function along the way isn't
1:02:32going to get more work to do now, right?
1:02:34So, it it's it's going to we're going to
1:02:36see a shift over the next several years.
1:02:38Um you know, and and again, know the
1:02:40tools to take most advantage.
1:02:44>> Totally.
1:02:45And Microsoft is you know, I mean, all
1:02:47companies all be because I work with
1:02:49Microsoft product.
1:02:50So, also yeah, just look up and you
1:02:53know, there are a lot of sessions going
1:02:54on around the world like Andy Yun.
1:02:56Uh I'm not promoting He's not here.
1:02:58Did a great session uh in local user
1:03:00group and in this group. Look up our
1:03:02YouTube record just on focusing on
1:03:04vector search.
1:03:05>> Yeah.
1:03:05>> Um and and some of this stuff that you
1:03:08don't need money even though you said
1:03:10you're the only you know, you the
1:03:11Microsoft is also providing a lot of
1:03:13free tools nowadays. Azure credit and
1:03:15stuff free Azure SQL database you can
1:03:16run whole year and take advantage and
1:03:19and test some stuff. Maybe spend few
1:03:21dollar from your pocket.
1:03:22>> Yeah. Yeah, absolutely. Um
1:03:25So, Carl, forward deployment engineers
1:03:27or forward deployed engineers, um yeah,
1:03:30I I
1:03:30I like the uh the Palantir, you know,
1:03:33um
1:03:33uh forward deployed aspect of it, right?
1:03:35And And you were seeing this across the
1:03:37industry. Um I I don't see them any
1:03:39differently than um you know, uh any any
1:03:42kind of um customer-facing engineer that
1:03:45that the various providers are are
1:03:47putting out there, right? So, I think
1:03:48it's become um the latest buzzword. Um I
1:03:52But personally, from my experience, I'm
1:03:53not seeing any difference than, you
1:03:55know, what our um you know, uh any of
1:03:58the engineers that we would put out in
1:03:59the field, um any kind of contractors
1:04:01and things like that in the past, right?
1:04:03So, I
1:04:04I think it's just a a role change or a
1:04:06name change based on um what the the
1:04:09broader community is is observing.
1:04:12Um And JD, I'm I'm former military as
1:04:14well. I was army.
1:04:16Um so, stay army.
1:04:17Um
1:04:18Yeah.
1:04:20Any other questions, folks?
1:04:23>> Nice nice questions and and these are
1:04:24some of the stuff that, you know, I'm
1:04:26glad that, you know, folks are asking
1:04:27and and you are getting engaged because
1:04:30we just have to think different. I mean,
1:04:32that's the fact today I feel like. So,
1:04:34it's not easy.
1:04:35Um
1:04:36you know, I'm not that I'm expe- I'm I'm
1:04:38learning a lot. Like Chris, we always
1:04:40talk about this. So, yes. So, if anybody
1:04:42has questions, please go ahead.
1:04:43Otherwise, we're going to call it a day.
1:04:45Let Chris go. Uh you know, he's a busy
1:04:47person.
1:04:48And um yes, Paris said thumbs up. Chris,
1:04:50thank you again.
1:04:52Folks, please uh
1:04:53uh keep registering for future events.
1:04:56Uh we'll see you in 2 weeks.
1:04:57Thank you very much and thanks for
1:04:59bearing with us with the new technology.
1:05:00So,
1:05:01>> Yeah, we appreciate it. Thanks, all.
1:05:03Take care.
1:05:04Bye.
1:05:06>> Okay, how do I stop this recording
1:05:07thing? Okay, stop trans-
1:05:09>> to recording and then save record and
1:05:11then stop. Uh do you want me to do that?
1:05:14>> No, no, no. Hold on, I want to learn.
1:05:15So, stop recording.
1:05:18Uh, are you sure you want to stop the
1:05:20recording in the cloud?