00:00
All right, well, welcome everybody. Thanks for uh thanks for coming. Um, this is kind of the uh double click follow-up session for the small flash talk that I gave yesterday. So for those of you that are there yesterday, I see a couple of faces, uh, great. Um, if not, don't worry, we're gonna repeat ourselves because it is absolutely worth
00:18
repeating. Um, but what we really wanted to do is double click on not just the use cases around AI but specific to financial services, which is I'm assuming why you guys are in the room. Uh, so my name is Rob Blansman. I'm a global principal architect for financial services at Pure. Um, my esteemed colleague next to me is Non
00:37
Taraj. Um, he's from KX. Uh, for those of you in the financial services industry, I would imagine you would know KX, um, but, uh, please give a quick intro. Yeah, yeah, sure. So, uh, I'm not. I lead the AI group at KX Systems.
00:51
KX essentially builds, um, a database called KDB Plus, which is sort of um. Um, an extension of the Q language that is primarily used in all the hedge funds from trading desks, and so on, uh, very in a low latency, high performance language that's specifically built for high frequency trading, algo trading and stuff.
01:10
So, um, yeah, I lead the AI solutions. We have a lot of AI Gen AI work going on at the moment, um, highly optimized with GPUs and some very low level primitives and stuff, so we want to discuss that here. Yeah, and you know, for those of you that follow what we do, um, we have a long history with KX, um, certainly in the time series based and Quan
01:30
trading, um, publishing our our our results to stack, um, so we're, you know, this is the natural evolution of that partnership and leveraging and taking advantage of, um, all the things we're gonna talk about today. Uh, I will, uh, make an apology for Perboo, uh, from Nvidia. Unfortunately, he could not make it today. He had some personal things he needed to take
01:49
care of, um, but we will certainly channel him, um, you know, here today. So, for those of you that may not be familiar with Pure's pedigree and financial services, it's considerable. Um, I have been in and around Wall Street for almost 30 years now, and I've worked at Pure for just over 5, so I've seen what the industry has done in the
02:13
last 30 years. Um, and when I got to Pure. You know, I always like to look and create a slide of my own that kind of talks to the pedigree of what we do in financial services. This is what I made 5 years ago, and I'm just happy to say that all I do is increment numbers every year, you know, we, we were, we are 11 years solid,
02:31
right, in terms of our growth and, and market share in the um enterprise. But the really important one is the uh bottom rate for what we're talking about today. So we're in, in 9 of the 10 global investment banks, 7 of the 10 asset management companies, and 5 of the 10 insurance companies worldwide. Um, that's saying nothing about all the hedge funds and prop shops,
02:48
only the ones that we deal with, um, but, you know, this is, um, this is where we really like to focus our efforts from my selfish, you know, lens, which is fostering what the vision and strategy is going to be for Pure globally with my colleague Mike, who runs the vertical for Pure. And and uh so this is KX.
03:12
um it's been around actually for almost um I say actually almost 50 years. If you, if you go all the way back to the 70s when you had APL and KDBQ, the language is an, is an, is an evolution of the APL language, which was basically built by mathematicians, for mathematicians, very highly analytical and very terse and so on.
03:32
So, as it, as it says, you know, it's 1 almost 17, almost 19, more than 90% of all the stack benchmarks, um, real-time and historical uh data petabyte scale. So real time was one of the main, uh, advantages of KDB plus cause So on the trading floor, you know, when you're dealing with milliseconds, microsecs, and you need to make trading decisions at that scale, yes,
03:54
you do use low level, uh, technology, the CC+ + ASIC boards and stuff, but if you're doing analytics, it's very hard for other databases to keep pace at that level where you need to react in a millisecond or 2 milliseconds and say, OK, I need, I need to trigger this, there's a latency arbitrage going on. And um you know, it's not just in capital markets, so it's also there in aerospace and
04:18
defense, IoTs um so there's just, there's just an example of some of the clients and you can as you can see there's Lockheed Martin, there's Royal Air Force, uh, even healthcare firms. In fact, so I used to be at UBS in the FX group. In 2005, worked there for, uh, eight years on trading floor, got tired, moved into healthcare, build a startup using KDB Plus,
04:41
and then we got acquired. So for the last number of years, I've been actually using KDB Plus for 10 years in, in healthcare, and now, you know, back into finance. So it's a very versatile language, versatile platform. And wherever you need that, you know, edge to take you beyond what you're doing,
04:58
um, you know, that's where we use KDB. Cool, thanks. So, let's talk about quick, uh, kind of current state of the union. Um. You know, we, we, not to be dramatic about it, but we are definitely seeing an inflection point, um,
05:15
where the early adopters, um, really tried their best, as a matter of fact, I was just having a discussion after the, uh, panel, panel that we had earlier today where, um, one of our customers was lamenting about the fact that, you know, the alpha that was able to get obtained by the firm over the last 10 years has
05:34
been eroding rapidly from double digits down to single and They were an early adopter, you know, and they were struggling, and you think about all that we had to do 10 years ago to actually do something meaningful in the space, right? No one even was talking about AI, you know, then we have the ones that are coming.
05:54
You know, and just our cloud native to begin with and AI native to begin with. So there's this delta between what can be done, um, and dealing with, let's just say, decades of data that can be harvested, can be probably have never been looked at, you know, in a meaningful way and derive insight from it because it is valuable. Um, when, when we talk about benchmarking and, and speed and relevance to the,
06:19
to the industry. Um, speed is not going to help us much longer, right? Um, we'll, they'll always be carving a few nanoseconds more, but it's how fast can we derive insight from the data itself is where we are talking about today. Add anything to that.
06:36
Um, well, I assume you agree. Well, I partly agree because it still matters, you know, I just read about this hedge fund that spent like $300 million to save 3 milliseconds, and that's real because, you know, they're using holo core fibers from, you know, under the Atlantic and stuff, because those milliseconds really matter and um.
06:56
I mean, that's where, you know, storage, all the, the entire stack comes into play, you know, so yes, it's it's, it's going down more and more, but what's happening is that this, they were talking about millises, they went to microex, now it's down to nano sex, and who knows, peak of sex, you know.
07:12
Yeah, it's it, I believe I was being kind of strategious right that speed doesn't matter. It's more like we struggled so much to just carve a millisecond or two, right, in the last 20 years, because that was meaningful for the firms. Now, it's not a point of diminishing returns, but it's just not seeing what we used to in terms of the benefit of investing that much, you know,
07:33
into it. So they will always exist in parallel, um, but they, they, you know, what we're going to show you today is really how we can kind of elevate the strategy of the firms that you guys potentially work at, um, and supplement what we've been doing. You know, it's, it's something that we always get asked is,
07:56
and listen, all the firms talk to each other, right? You all know what you're doing, um, but we still get asked, um, what everybody else, uh, what, what state they're at, what, what level of maturity they're at. And I'm hoping that it's somewhat of a comforting, you know,
08:12
statement to say that most firms are really in the first, the first oval, right? The bottom left my dad, right? Most of them. If you look at the top of the bell curve, they're, they're. Just coming out of pilots, they've been at it for years and in dev and test,
08:27
but it's sure it's no surprise that crow Walk run, as we like to say, when the initial use cases are uh unleashed into production, um, the first thing that always comes up is how do we create, uh, and utilize an LLM inside the walls of the firm. That will prevent, you know, any leakage out into the public for doing research
08:55
or anything else that would be, uh, you know, the PMs or the quants require. It's unsurprising. That's true of any vertical, right? Um, I, I, I was talking to this other customers like, listen, we had to literally shut down all access to the internet because that's the only way we can do zero trust, right?
09:12
How can we prevent people from using this? Well, we certainly can't prevent them from doing it at home, but the data is resident inside the firm. But when we think about just the one bullet point that really is um striking. In terms of the current state, is the, the, the second one down, level one, which 100% of their decisions, traditionally are being made on 20% of the data.
09:36
That's, you know, kind of terrifying if you think about it. I'm sure it was a lot less years ago, but that is exactly what we're talking about here is it's not just the data that is resident inside the firm or sitting there and, uh, you know, even on backups or archives or kept for regulatory purposes only. How do we actually mine that for something useful and how do we actually augment that with
10:00
Things that are happening outside of the firm that would impact the decisions that are being made. So, yeah, so in, in real world practice, you know, it's AI sounds great, but then to actually get that benefit, you need to do a lot of experimentation. A lot of times, you know, we, when we speak to larger banks,
10:20
large banks and stuff like. One of the biggest challenges is where do we apply AI? And does it really drive a value? You might be better off just doing like averages on Excel sheet or something, right, than use AI for something. So AI has its space and, um, you know, I think basically what,
10:36
what, uh, my colleague is saying is that in capital markets and, you know, as you, as you, as you increase your level of sophistication, I think that is where you start seeing the benefits of AI more and more. If you go to Renaissance technologies, you know. You're going to see, um, they're optimizing at the level of,
10:55
I need to improve my accuracy to 50.27%. Whereas that 0.27% means a lot, you know, so you're not saying, I, I want to be accurate 80% of the time. It's just not gonna happen because you, you know, brilliant brains everywhere, but even that 0.27% alpha is, is a huge amount of money. So it's, it's a non-trivial problem, but you know, it is what it is.
11:16
Yeah, and, and the only thing I'll add to that is um in terms of the starting point. A lot of it is being used on results that can be checked manually, if you will, right? So if there's insights that are being derived, it's fairly easy to go through and see how that uh result was returned to the user. Um, that's table stakes because if there's anything that we fear more,
11:41
it's a risk, right, and making decisions that are uninformed, um, or, or even wrong, right? Yeah, the hallucinations. Yeah, so, uh, this is basically talking about some of the different use cases that we see in the markets, um. And, and, uh, well, you know, different kind of reason, the different reasons,
12:05
uh, KD plus is being used, so the hidden hidden market shifts, the subtle trading patterns, unpredictable spikes, uh, liquidity fragmentation, um. Generally and basically, you know, the, the partnership that's been highlighted here is the KXPR and Nvidia.
12:21
Uh, Nvidia is a key component here because um you know, today most, the state of the art in, in, in deep learning or even Gen AI is the Nvidia stack. That could change in the future, but today we're largely dependent on GPUs and even if we just take your current code and port it over or use the Nvidia stack on top of it, just a Python data frame and you switch it to QDF and
12:45
stuff, you're gonna see a huge benefit. But then the thing is, You, you need to put these, these, these technologies together. So if you just use Nvidia, it's, it's all well and good, but if your storage is not fast enough, that might be a problem. If your, if your database isn't fast enough, that might be a problem.
13:01
So you need to have the best of breed and Essentially it's you know KXPR Nvidia, there's an entire stack that's been built up that provides these sorts of advantages for these specific uh use cases and, and beyond this as well. Anything you wanna add? No, just, um, you know, this is all leading to all of the research and all of the data that we've been,
13:23
you know, privileged to speak to our customers about. Um, there's a lot of commonality raising to the top that would manifest in something that we wanted to do to help you guys get to, you know, making it meaningful inside the walls. To that end. So, I'm sure you've heard a lot of terminology this week, multimodal being one of them.
13:45
And this is an example of A subset of the data that PMs uh are using now to make decisions. And it's down the right and for those of you in the back, um, it's the usual suspects, ESG data, social media, things web crawls, uh, weather, right?
14:06
Vectors that probably were introduced and, you know, given some weight and consideration, but to aggregate all of these manually. And to go out and have, you know, a, a PM, you know, do an assessment and make assumptions based upon this. That's the way we used to do it in the past, right? Um, these are growing.
14:28
The magnitude of, of each of these for the most part is growing, and a lot of it is in various different forms, audio, video, text, PDFs, CD, I, I mean, you name it, it's in there, right? So when we think about How do we actually deal with ingesting that and making it meaningful from a training perspective, making it meaningful for returning,
14:54
uh, queries, um, you really start to have to double click on that, you know, and look at really what does that mean to the use cases and what do we have to check off as prerequisites to do it. Yeah, I think so. Traditionally, KDB plus has always been used for the structured data part,
15:12
right? And then once the Gen AI wave started, uh, so there's this big question that we should be supporting vector databases. The thing is in the vector database is just a term we were already managing huge arrays, numeric, you know, quantitative information KDB plus. So it was really just a question of, OK, we're going to,
15:30
we already have this, but now can we make it more optimized, you know, can we improvise on the rag aspect of things? Uh, can we improvise on some of our algorithms which does similarity searches and so on. So essentially what this, what this, what this slide is saying is that today's version of KDP. It doesn't only do the structured data, it also does, of course,
15:51
the unstructured data. And then, of course, it's time aware, you know, temporality is one of the main um main benefits of KDV plus because time, months, days, years, these are first-class primitives in KDB. So time is, you know, I, I know you're gonna have posits and all that stuff in, in regular other databases, but time is, is a fundamental dimension along which you part.
16:13
you do a lot of um types of interesting time series based joints like as of joints, window joints that are not that common in, in, in some of the other fields. And because we are also handling the unstructured data part now, you know, it, it, it's usable not just by the front desk but also by the wealth management, you know, portfolio advisors. It's used in retail banking and so on.
16:36
Uh, to the point that, you know, Rob was making earlier about all those different data sets that's out there. There's a lot of alt data that's also being loaded. It's not just the tick data, but what's happening in the news, you know, what's happening in the weather. So, we're trying to query that in the ordering of milliseconds.
16:54
We're looking at the price movement in milliseconds and basically we're combining the structured and the unstructured data analysis at scale at a very ultra low latency uh kind of paradigm. Yeah, yeah. Well, that's certainly where we come in as pure. Um, I like to say that uh that is exactly why we engineered.
17:17
The platforms that we have today, because we had to get out of the idea that storage had to be tuned for the workload, right? That that is not providing value to the firm other than maybe the one tuning exercise that took months to do. But how do you actually tune for something that is so erratic and so unpredictable and growing, right, at an unprecedented scale.
17:47
So if we look at, let's just say a fairly basic industry agreed data pipeline, right? And we look at all the components of each of those. Well, we did a lot of analysis in terms of what each of those stages looks like from a perspective of IO. And I won't take you through all of this.
18:05
Well, let's just say that is absolutely worst case scenario for us, right? And for you, if you don't have us. But you know, when we think about the nuances and what's required for each of these stages, um, there isn't anything missing here from the perspective of characteristics of an IO. What is missing though, is when you look at this, and you're like,
18:30
well, I can probably get something tuned, you know, around this and, you know, figuring out what each one of these is going to do and how, how, how heavy it gets hit and when. Yeah, but that's just one pipeline. We're talking about potentially deploying thousands of these simultaneously. So it's just, you know, compounded by compounded, uh,
18:50
in terms of what the, you know, what we need to do to ensure that the GPUs are as hot as possible, that there is no disruption in any part of the data pipeline that would effectively cause cause issue with each of these because it is serial, right? Um, oh, my bad. So, Words or something, visuals are better.
19:11
Uh, if you look at the characteristics of each of those, well, again, these are all happening simultaneously. So if I merge those, you effectively are looking at an IO pattern that, um, you know, we'd like to think we have purposely built our architecture. To ingest exactly that, which is not having you guys worry about the mechanics and the
19:35
nuances of tuning that or worrying out how to optimize it. That's our job. Your job is to figure out how to generate alpha and beta. One thing I will say though is depending on your philosophy in the firm, whether you decided that you were going to start net new and police tape off a section of the data center to do your AI implementation or you have
19:59
been at it for a while and have been integrating it, you know, successfully or struggling regardless where you're at. What we're seeing is the same exact thing that we've like to think we've solved for and are aware of in the industry is happening again inside the AI plants. You know, Data access being paramount and sprawl.
20:20
It's almost like reverse sprawl. I, I, I'm coming up with that in real time, uh, near real time. So what we see is if we think about all those data sources, and let's just assume that half of them are proprietary. Well, we know that for the most part, the larger the firm, the harder it is to get that data into a centralized place to be able to meaningfully be,
20:41
be operated on. Well, we, we actually call that a biological problem, right? It's not a technical problem, you know, it's easy to get from point A to point Z. It's how do we actually break down those barriers from lines of businesses to be able to train a model on all the data that's inside the firm.
20:56
So those are the type of things that we've been discussing and figuring out and sharing and, and under, you know, and, and cone of anonymity. Of what firms are doing to actually try that and like anything else you start small, you start with some, you know, data that isn't necessarily on the front lines, something that's potentially a little older, something that's maybe even not proprietary,
21:18
just public that you can do a what if scenario. Well, I won't take you through this, we've solved for all that, right? We know if you. This is what we do. Last thing I'll say that before we double click on what we're really talking about today is, is, is the,
21:34
you know, State of the Union. I, uh, hesitated to put this in here because I did it in my flash talk, but it was so well received I decided to do it. Um, who remembers that from back in the day, right? I'm sorry, I don't, you know, it's, it's late in Vegas for, you know, we're all waking up or going to sleep, but I was horrified to actually dig that one up
21:52
from the depths because the first time we saw that, We're just like, oh my God. Look at, I mean, how do I actually create a, you know, infrastructure and architecture that picks and chooses one of those? Well, I have unfortunate news for you that they released something very similar to this specific to AI. And no presentation is good without some kind
22:15
of an eye chart. Um, and when you look at that landscape, which was, you know, Couple of years ago, uh, excuse me, last year. How do you pick and choose that? How do you understand what is relevant to the firm, right? And that's where we really wanted to focus on today, is building a turnkey FSI
22:36
specific AI stack for Greg. That allows you to get to that initial point with curated, jointly engineered, jointly architected, um. Platform between ourselves, KX, Nvidia, Super Micro, AISA. When we think about, you know, if you were to have a magic wand and
23:07
say, how do I actually get started and what do I actually do to get started in a firm, everybody's all over the map, right? Some, some people have chosen their own components, that's great. What we decided to do was collectively share our expertise and our understanding of the market and what bubbled up to the top was our best and breed partners, um.
23:29
And in a architecture that would be able to demonstrate the most common use cases that we're seeing, so we can generate that what if. Yeah, I mean, basically what we're saying is that, you know, we have all these choices uh to go through, you know, it's, it's much, uh, it's, it's helpful to have a blueprint.
23:48
It's helpful if, if there are partners have already built a stack which they know works really well, which is highly optimized and essentially what this is is a collection of the best in class and you have Arista, you have PR, you have Nvidia, Super Micro, Port Works, KDB, uh Microservice, and so on. So, you know, I, I've done a lot of POCs with vendors and takes ages,
24:08
right? Sometimes you're going through benchmarks here, you're, you're comparing one against another. It's, it's always beneficial when you have this sort of a stack already done for you. And then, yeah, you can, you know, you can switch things out if you want to, but at least you have a starting point.
24:24
Um, and that's what, you know, we have, we have built this in, in the WWT labs, um, and, um, you know, we have some really good client testimonials and references. The entire system is, is, is, is highly performing, uh, works really well. Yeah, and I just pointed out the what we agreed upon, and I intentionally left the WWT to the
24:43
end because this is up and running, living GA live ready for you guys to take advantage of as a, in a demo environment in their um AI proving ground. Um, that is something that I can say, uh, was not trivial, right, to get done because of the previous slide I showed you, things are always changing, and we had to pick and choose what we were going to do,
25:08
put a stop gap and say this is, this is what we believe is the current state that's gonna be the most beneficial for you guys to see. When we think about what can we do in there, or we're gonna show you. Um, when we think about how can we help you if you engage us, well, leveraging WWE's WWT's expertise in the lab for just their deployment,
25:30
their understanding of how the, you know, the cooling if need be, and rack structure and power and cooling, everybody's got their part to play here. You know, ERISA and Super Micro, right, taking care of the connectivity and the compute, um, if you go and you look at all of the nuances that each of us bring to that stack, we're confident that when you see this and you see it in action,
25:54
it will just generate so much good conversation to have that, you know, it's bidirectional, right? We want to learn from you as much as, you know, you, you can learn from us. Yeah, and these are some of the use cases, uh, some of the, um, you know, projects that are underway within, within, within KX with our hedge funds,
26:14
with our institutional banks. So we have AIPod research assistant, relationship manager, portfolio construction, uh, Alpha Beta, deep learning, and so on, basically, wherever you need content comprehension, content reasoning. You know, that's where, uh, generative AI at least today really shines.
26:32
And, uh, those are the kinds of use cases you'll typically find. Alpha is slightly different, that's more on the quant side. Time series forecasting is also a bit more on the quant side. The, the, the good thing, the interesting thing with the 1st 3 though is because KDB Plus is already a very high performance structured database.
26:52
And you have KDV.AI, which is the vector database component. And because these are one and the same, you can, you can add, you can do both the rack part. There's a vector database part as well as the structured analysis, simultaneously, concurrently in database, in that ultra low latency fashion.
27:11
So, that is, that is the main difference here. You could be using a vector database as is, but it might not be the best structured, you know, database from a performance standpoint. But if you can correlate the price movements with the data that you have in your rack system in the a. Of mill, or whatever, you know, your time frame might be, basically very efficiently, that reduces, that,
27:34
that simplifies your architecture, reduces the overhead, you know, you have a much smoother ride than trying to manage multiple systems, multiple databases in different parts of the ark. Yeah. So we love numbers in this business. We'll give you a few. Um, we took a look at each one of those use
27:53
cases and applied the best estimate that we could in terms of what it takes a typical persona inside of well's firm to actually do something for each one of those. The one we're gonna show you today actually a demo of is the AI power research assistant, but, and this is I think being conservative. You know, the delta between what it would take for an analyst to ask a question,
28:17
like on the left-hand side. And actually to go and do the search in a traditional manner of what they're used to doing and what they historically have done versus being able to actually use an LLM. Ingestrag data to be able to uh answer that question and, you know, a best or magnitude less than than capable.
28:35
Hopefully, that is the type of, you know, delta that really It is, is intriguing, you know, and expected from the investments that you guys have made in a platform, and that will yield in, in, uh, in so many ways. Let me, uh, show you a quick double click on the archi architecture first and then we'll show you a quick demo of that. Um, just by starting I'll say.
29:02
Here's our, your friend of mine, our portfolio data manager, right? Coming from Disparate amount of data sources, market enterprise, um, and all of the other ones we're talking about multi multimodal. Well, looking at the flow of that through the NIMs, basically, you, we, we have a, again, a unpredictable amount of IO,
29:25
an unpredictable amount of uh scale and um from even a day to day basis. If there's an event that happens. Hurricane off the east coast or unfortunate, you know, events, uh, you know, from, from a perspective of world events, you know, that has obvious correlations and obvious impact in terms of how much data would
29:45
be ingested because the news sources erupt or the weather data erupts, right? So just trivial examples of how important it is to have that land on architecture that we've all developed, that's able to ingest that and still maintain a Um, a steady state, right? ingesting spikes, but also having predicted
30:08
performances, uh, to the front end. Yeah, I'll just mention that, yeah, there are lots of components here, the Nemo components and Nvidia components. It's, it's, it's, it looks very, it looks, might look a bit complicated, but in practice, you know, if you have someone who's already using Nvidia and stuff,
30:25
those are fairly standard stuff, you know, that you can just put together and they're just like. Basically packages that you connect with one another. And yes, I mean, you have the cakes, and data is the main source and data is the main stuff that you're going to use to make the decisions. So all of these relies on the data, the speed of the data, the, the accuracy, reliability of the system in order to perform
30:46
uh at scale. Uh, there's also, if you see on the left hand side, there are guardrails, there's fine tuned LLMs, those are essentially trying to make sure that the results you're getting are grounded in facts and not just, you know, so-called hallucinations. And there's, you know, I know people still talk about hallucinations,
31:04
but in the real world, in the enterprise setting, that's been already, that's been largely taken care of. You know, there are provisions in place. It's just a matter of becoming familiar with what those are. Yeah, and, and if it wasn't obvious when it comes to the pure part of this, um, and this was double clicked on uh during the uh panel
31:23
session that Mike had hosted, um, this is all containerized, right? So we have to account for the Kuberne's aspect of it and what Portworks is doing in this environment, right? So if you ever wanted an example of understanding. The criticality of something like Port Works in an environment like this, this is a great place to explore that,
31:43
um, because we've put, you know, great effort into demonstrating why, you know, once these start to scale, it just becomes untenable to, to manage. So let's, uh, let's show you a quick demo. Hey you can ski skip over here. You wanna start? Yeah, I'll start it up and.
32:07
Rob seems to like the cake stuff more than I do. I love cakes, you know, I talk about it all the time, twice a year at stack and um, you know, it's, it's uh. We, we, we are just enamored by the partnership because, you know, that if you want a wax park and a KX I will because when I first got to Pure.
32:29
And I, I was asked definitively, like, what do we need to do to make sure that your pedigree is cemented in financial services, right? I said, well, I mean we took a look at what everybody's doing and cats constantly came up to the top because I was familiar with it from my past lives. But I was like, well, you know, there are ways to prove definitively that we as few.
32:51
Can ingest this workload like nobody else can, and we did prove it right through stack, um, so once that happened, it was just a matter of a natural evolution of working together and frankly having similar philosophies in our portfolios. Um, yeah, I mean, the, you know, the, the speed of the, the IO speed is very much, very critical, especially when you're dealing with,
33:12
you know, terabytes, petabytes. Yes, the software can do a lot of optimization. You can have a sorted column and you're doing binary search and things like that, but then your IO always becomes a bottleneck. K2B plus is, is very tiny. It's like 400, 500 kilobytes.
33:27
That's the size of the file. It fits inside the L2 cache of the CPU, optimizes everything, you know, the communication between RAM, the, the CPU, and it looks at, you're going down to observing, you know, how much data are we passing through the PCI buses, you know. So you're going down to that level,
33:44
but then every time, you know, IO becomes kind of the, the, the constraints. So within KDB plus itself, in fact, um, and this has been there for like tens of years now, we would, whenever we would load, save the data, it would be distributed across multiple IO channels. KDB automatically does MapReduce,
34:05
which is what Hadoop, you know, essentially was known for. So given a query, it would distribute the query across multiple IOT. Channels in order to maximize the IO throughput. because it's going, it's trying to move at the speed of the CPU, the quartz crystals and stuff, you know.
34:21
So it doesn't want the IO to become that, that sort of constraint. And just anecdotally, you know, once I was working in in this, um, other firm, and there was not a financial firm, we, we, we install KD plus in the data centers on-premises data center. And I remember my first week in, they said, what did you guys do?
34:40
I said, what you said, OK, the, the, the rack sounds like a jet plane is taking off and the fire alarms went off yesterday because the system overheated, you know, so that stuff really, you know, actually happens when you work with certain types of technologies, not ours. So for those of you that I don't know if you guys can read that,
35:01
but what, what we're gonna show you here is um simulating. An interaction, um, with, you know, the, the what we've trained and basically asking a question and I'll read it just in case people can read at the back. What is the FY FY 2018 capital expenditure amount in US millions for 3M? You have a response to the question by relaying, relying, excuse me,
35:22
relaying the details shown in the cash flow statement. And forgive me for Going off from there. So we had, and by the way, I wanted to point out that was a prompt that we, that we pre-built, right? If you guys come to the lab, you can certainly use your own,
35:39
right? Um, a little bit more involved, this is the one we'll show you, um, but This is one about Apple Apple and 10K filings, right? I won't read the whole thing, but basically what we're doing is, um, there was some volatility, um, trying to estimate why in 14 days, uh, you know, compared to its trading value using a 30 day,
36:00
30 day VWOP and 60 days following disclosure. Um, so the question that there's a context, right, using an LLM. And the question is, I want to understand if these newly emphasized supply chain risks were gradually priced in or if there were sudden volatility spikes in the market digested in this information. Basically it's a complicated question.
36:18
What they're basically trying to do say is that, OK, I have this question from the portfolio adviser. He wants to know a lot of stuff. The data is in the, there's stick data in the structure database. There's all these, there's millions and millions of filings. Can you make it easy for me to find that information?
36:36
So the first thing that happens is refining the prompts. I pause and scroll that up. I'm just saying just the details, yeah. But the point is, when it comes to introducing this type of technology to a firm. I think I've said it before.
36:53
There's a great comfort in knowing how you can effectively prove that what you're getting back is correct and this is an example of showing you systematically how that was arrived at and also being able to go back and, you know, analyze it manually if need be to spot check it. Yeah, so basically what it's doing now is that it's breaking down the, the prompt into individual components like what are the symbols,
37:19
what are the filing types, what are the keywords, uh, you know, what documents do I need to retrieve, and so on. And um you can keep playing and uh you know, it's not like any problem that you just plug into chat GPT, right? That that's going to that's going to be that has breadth. This is where we're, we need depth.
37:40
So we're, we're going down into the financial use case and we have the appropriate tools and agents which actually, um, you know, parse that query in order to make sure you're getting data from the right sources. You know, I wanted to stop it there because just showing the documents that were retrieved, right? You had cited the statistic when you were
38:01
sitting here, what was it? Yeah, so that one was, I, even I wrote that one down. So, so basically, you know, um, there's, there's 3 million um documents in this system. It's done apparently in a matter of minutes or so. Um, the, when we were first, when the vector
38:22
databases thing first came out and KDB also came out with their version of vector databases, um, we looked at, we, we, we look at the benchmarks and how well do they, do these things perform and Generally, there's, there's sort of a uniformity across the industry and if you have any number of documents to give a particular query, how long does it take for you to retrieve the the right set of similar documents from your
38:47
rag database? So KDV.AI on the CPU it was all OK, but when, when we, so, let's, so in this particular use case, we had a, we had a portfolio advisor asking a query and we have 1 million documents through which we need to Do a very intense matrix multiplication, 1536 values times 1 million times
39:07
1536. So there's quadrillions of matrix, you know, multiplications that's happening. And uh we ran it in KDB.AI and it, it uses optimized liplas, GPUs and all that. And I was really stunned because they, they, they're showing us the numbers.
39:22
It took 7 microsecs, 7 microseconds to do all those multiplications, find the right documents, do the sorting, and return the results back. I mean, that basically means, you know, sometimes we, uh, people would tell me, if we don't need, we are not speed sensitive, especially in wealth management, we're not that speed sensitive,
39:40
I be the speed sensitive. But what that means is you can have just one machine doing the work of what you may otherwise need an entire cluster for. It, it simplifies your operations drastically, and net result, it, it reduces your total cost of operation as, uh, also drastically.
40:00
Yeah, so how many are here. 9, right? We're talking millions, right, upon millions, uh, and that's being generous, right, or, or conservative actually, um, you know, it really begins to paint a picture vividly in terms of what, uh, you know, what this looks like. I think we can, yeah, I'm gonna see that.
40:18
So what we wanted to show you is what is built currently in the WWC lab based on that architecture. What you're capable of coming in to do. We've obviously pre-populated this with our own data, as Naa just said, and we basically have given you the keys to go and explore this and simulate what this might look like inside the walls. What if,
40:41
right? You know, generating and showing you what the prompts look like and discussing what each of those components had, uh, to contribute to this. You know, that this is not something that You know, we built just to say, railroad you through, right? It's something that we want to be able to have
40:58
a meaningful dialogue with and share our expertise, so we can come to a, what does the firm need conclusion. Yeah, I think it's covered it well. Very good. Yeah, so. Please. Yeah, so I mean, we, you know, we, we showed you a lot of that we kind of tried to go
41:21
through it at, at, at a pretty fast pace. Um, You know, K KX has been there in this industry for many years now. Like I mentioned, 1020, 30 years and, and, and counting. And, uh, because, you know, it, it, it lives in that zone, which I guess becomes so niche that people don't hear about it that often,
41:42
or you think, oh, that's gonna be too difficult. You don't get into them, but the fact of the matter is, those who know what they, you know, those, those are kind of look, go exploring for these technologies and they find them. They, they, they, they discover a world that, that they've never seen before.
41:58
And I went into, from after 8.5 years on trading floor, I went into healthcare. Nobody knew KDV over there. I said, let's do KDB. They said, we'll put it in the lab. Within 2 years, our entire healthcare, our entire databases were moved into KDB Plus.
42:13
There's a highly conservative traditional. Pharmaceutical company, moving their entire workload because they're saying, you've never been able to query more than 2 years of our patient data. And suddenly now we're doing 20 years, and the responses are coming back in, in minutes. Earlier, we used to wait for offshore to send
42:31
it to us, and they would run a bad job on their, on their, you know, data warehouse. And that's just data that was already there and resident. Has nothing to do with what was going on that day, that year, right? That's the world we live in now, you know.
42:45
I lied. I'll show you one more I chart. Then I'll let you go, right? Without having to go through this, I think for those of you that know and love us, you know most of this, but I really wanted to point out a few things very specific to this environment, and that's true of all the partners that,
42:58
you know, worked on this with us. Power is paramount, right? Power and cooling, and you know, we often kind of dismissed as, oh well, we'll take care of you. You know what, when we think about the data centers that we've seen out there, most of them are pushing the walls, most of them have no way to expand.
43:17
And we, our philosophy is, well, how much can we actually squeeze in, you know, to a half of a rack. And how you know how much can we actually, you know, save you from the perspective of physical footprint versus power versus cooling what's required in those so all of those environmentals are things that we'll share with you when you come to us,
43:37
um, there's the fun stuff in terms of what it's actually doing, but then there's a pragmatic. Things like how you actually deploy this inside the walls, you know, so trust that when you come and explore this, uh, and by all means, I really hope you guys do, uh, you know, reach out to your, your contacts, uh, at Pure.
43:54
Uh, we will direct you to the right place so you can sign up and, uh, enjoy this with us. And, uh, we really love to have, you know, everybody come and explore this and, um, learn together, so.
