Skip to Content
59:58 Webinar

AI & ML - Beyond the Buzzwords into Reality with Pure1

For March, host Andrew Miller invites Sandeep Phadke, Director of Product Management, to the Coffee Break. While Sandeep has a storage background, for years now he’s been focused on Predictive Analytics, Deep Neural Network Compression, and more.
This webinar first aired on March 21, 2023
The first 5 minute(s) of our recorded Webinars are open; however, if you are enjoying them, we’ll ask for a little information to finish watching.
Click to View Transcript
00:00
Hello, everyone and welcome to your this month's Coffee Break for March 2023. March 2021 or 2023. Depending on what time it is, depending on what time zone you're joining us from A I and ML Beyond the buzzwords into reality with pure one joined today by Sandeep Sandy. Thank you so much for making the time to join today and for the,
00:22
for the preparation and work, we'll uh we'll dive in a little bit later to Cindy's background as we go along here as always. Uh This is a series. So Coffee Break series, you can find the previous ones because we try and keep a solution focus means they age a little bit better. You can find the previous ones either at that bit link or as well at pure stores dot com
00:43
slash events. And you can even see previous ones there. You can tell when I took the screenshot it's been a month or two. Also, let me make sure I uh I pull down the music here. Um Also, I want to make sure to highlight for those if you are attending today. I realize it's partly for the coffee card. But it's also probably because you're somewhat
01:02
interested in pure. At least we try and keep it, you know, focused on industry background and details and then go into what pier does. But our conferences, our yearly conference is coming up June 14th to 16th in Resorts World, Las Vegas. And rather than me talk about it more than that. There's actually an amazing 42 2nd video.
01:17
Thanks to the amazing marketing team here. Let's roll. So I hope that you can join us. I'll be there. I'm not sure who else will be the virtual option from Jason. Good question. Actually, I might be actually um task some of the folks that are helping us with the background on the chat if they can answer that
02:13
a little bit of final housekeeping. As always, I recognize that to some degree uh folks join for the coffee card. That's cool. It, it's part of how it works. 1st 1000 attendees. As always, there are some people that we cannot send this to given who you are. But thank you.
02:27
If you're employing your employee, your partner, you're in government or state or sled. Thank you for joining us today. Finally, in the housekeeping. Um I'm not gonna go introduce myself as I often do don't as much each month. But when I think about data sets and analysis, uh for me,
02:42
at least that was supporting teams that did this stuff. But as often, sometimes I will both kind of cop presents. This is one where I'm definitely hosting a little bit more until we get to the final section. We'll try and work in some live demos and I'll do those, but I'm really glad to be joined by Sande, given his deep background in this space.
03:01
Uh Sande, do you mind uh Do you mind introducing yourself? Yeah. Um So I've been working in this uh field. I lo I loosely call internet infrastructure because if I go back and look at all the things I've done, I've started with storage, did a little bit of security. Uh Good few years of networking, uh some virtualization and now
03:27
back to storage. So, you know, this is, this is what makes the internet uh happens. And I think I'm, I'm clearly a person from the in, in internet infrastructure uh background. So, II I, you know, I'm an electrical engineer and uh spent almost 20 years building products. You know, first half of my career doing engineering and now the second half of my
03:51
career is essentially uh trying to tell engineering what to build. I love chatting about, we were chatting about timing for folks who, who, who go up, you look up on linkedin, there's some, some founding in there and even like thinking about like what's the right time to sell and how long do you hold on? I mean, you, you've lived through multiple
04:10
different places and waves in, in the value, which is, it was cool to hear. Um, as well. I'll say, uh, you know, I have something to say on that. I mean, you, you can do startups but, you know, you, you need to get the timing right on when to start and when to exit.
04:29
And, uh, I, I, I've done two of them. I sold the first one, in October 2000. I think that was a good time to sell anything pretty much because everything went down after that. Uh The second one I, I started and I sold to Cisco in 2010. And that's the one I feel like maybe I sold a little too early.
04:52
Sometimes the, the only other piece here, it was fun. Actually looking up on Amazon your undocumented Windows nt book for anyone who still needs to learn Windows. Nt it's a, it's a bargain and I think it was $94.53 for the coffee still on Amazon. So go, go look it up and, and also recognizing that that sometimes as I frankly spent more time in
05:11
Palo Alto area, I, I got to know more of what IIT means, Indian Institute of Technology and have a healthy respect for that as well as I grew up close to Notre Dame. So uh growing up in Maryland, so again, thank you for being here. So the last little bit of housekeeping, uh next month we'll be joined by Justin Emerson uh for those of you who are tracking,
05:29
we released a new product Flash Blade E this month and we're gonna have a little bit of fun. We try and be a little bit humorous, mildly snarky without being over the top right kind of thing. So like it says there, the disc dinosaurs never saw it coming. The meteor man. Now I'm having trouble saying the word meteor impact of Flash Blade.
05:47
E join us next month. This is also going to be actually a little bit of a continuation of the better science discussion with Brian Gold. Previously looking at some of Justin's history at a pure partner and why he's a pure kind of, you know, the being a chronic fixer and building and creating things and use cases actually that are unlocked there.
06:03
But into this month's topic, as always, we, we try, I often go with a little bit of a, a standard agenda style. Um It's not magic a IML, it's lots of math potentially or repetitive type stuff. So we'll start off with a little bit of just definition, you know what, what in the world really is a inml beyond the buzzwords, like the title said,
06:26
that's gonna go deeper in. Uh This is just a host of buzzwords actually, but there's meaning behind them or actually explore that and you're gonna hear some of Sandy's background mixed throughout all of this because he's been living this. Then we're gonna try and make it a little bit more real from an applications and industry standpoint. And then last, but not least, you know,
06:43
in action at pure how we're making your life better with this. A INML and specifically in this case, we'll illustrate it with pier one. And if the demo gods smile on me, we'll even do a live demo or two. As we, as we explore some of that, please don't hesitate to put questions in the chat throughout. We have um Emily Olivia Matt and Calvin here to
07:03
help with that who all know lots about this stuff as well. Uh So please don't hesitate and then we'll take questions at the end as we always do so to kick it off. Um Emily or Olivia, whoever can launch poll number one if you don't mind. So this is uh it's meant to be a little bit of maybe a,
07:19
a pallet cleanser here. We're just gonna have some fun with it. So if you could automate any of the tasks below tasks with a I, what would it be? And if you have other ones, you think like, I wish I could do that. Um Hey, feel free to put it in the chat and then I'm, I'm gonna take full credit more because I don't
07:34
think anyone else would want to take credit for it for the, the answers in question number two. Uh If you know what GP T stands for? Cool. Um If you just appreciate my sense of humor, you know, as, as good or not good as it is, feel free to pick some of the other answers.
07:47
But we'll leave that running up here for a second. So Sandy starting out here, this is a, I think a INML is almost about as buzz worthy as it gets in some cases. So, do you mind kind of starting helping out and maybe I'll just put this up and you're gonna talk around this for a while.
08:07
Um I know for you, it even starts with ML and then the others are related but just, just help me out, help us out with thinking about the categories and, and what they mean here. Yeah, absolutely. So, a I, you know, artificial intelligence, this uh you know, subject uh has fascinated us, you know, for decades, I mean,
08:27
you know, this is not a new word at all. It has been around for maybe 60 70 years now, maybe even longer. It's the whole field that uh relates to developing something that can act and think and do things like humans, whether it is computers, whether it is programs, whether it is robots, anything that can think like a human act,
08:51
like a human, pretty much imitate a human falls under, you know, the realm of artificial intelligence, you know. So uh that's, that's the broader definition, you know, pretty much everything that I feel we are doing today and, you know, you hear people talking about, you know, uh the N medias,
09:09
the Googles and so on. Uh To me is largely powered by what is in that second circle inside which is machine learning. Uh machine learning is actually a pathway to A I. So it's all about enabling the machine to learn and recognize patterns uh from data, everything starts with the data here. So uh as you have more data,
09:34
uh you know, there are techniques developed to recognize patterns from that data, whether it is trying to understand what is written, trying to read, what the text is trying to understand images. Uh You name it even trying to turn what is the sad word into? What is the written word? All of these things are, you know, patterns and trying to understand those patterns uh is,
10:00
is, you know, machine learning, you know, job number one, there's a lot of techniques to do that. So I'll, I'll, that's how I just differentiate between those two and, and what I found interesting when you and I were chatting was, was for you, it's almost like it's all machine learning or that's even though there,
10:15
it kind of shows concentric circles you machine learning is at the core of that. And then there's even a couple of different concepts there that then eventually wander into A I and, and ML you mind continuing. Yeah. So uh essentially, you know, the, the techniques that machine learning brings uh I would say broadly fall into three categories, you know, shown by those three medium sized circles.
10:37
One is the supervised learning. Yeah, that uh red or pink circle that you see there. And uh whenever there's lots of labor data, which means, you know, you sort of know, hey, this is the answer to this is a cat. This is a picture of a cat when you have a lot of label data. This is the technique you use where you train the model,
11:00
which is, which is, you know, the e effective learning of any machine learning uh endeavor is to build a model. So when you have lots of data and you use it to train that model uh that is called supervised learning. So it's kind of like, you know, you're supervising, making sure you know, it knows exactly what is what and then once it's
11:21
sufficiently trained, it's able to then, you know, sort of uh pick out the cats from all the pictures and so on, right. So that's supervised learning uh to be very contrasted with what's on the left hand side, the unsupervised learning where you know, you of just hand over the data uh to these machine learning technique, uh these algorithms and say what can you tell me about this data?
11:44
This is, this is often used uh when you are trying to segment groups of customers or you know groups of users uh into various clusters uh machines. There are techniques to easily differentiate, you don't even know how many groups there are. So, so this technique will tell you there's five groups and here they are and sort of cluster those uh you know, users or, you know, uh customers buy those.
12:10
So that's a very uh unsupervised technique where, you know, not a lot of human help is needed in this. And the third one is sort of half and half in some sense, you know, at the bottom there's reinforcement learning where you train the model and then you start giving it feedback saying, yeah, I think you got it right.
12:30
Thumbs up or hey, I think, you know, you got this part wrong and you go improve on it. So that's with the feedback loop, that third technique is reinforcement learning. Uh And this is, this is heavily used in robotics. So when, you know, you're training, training something, you sort of uh keep telling that uh machine or model or program saying, you know, it's doing good,
12:52
it's not doing good and you know, it, it improves based on that. So if you've heard of the popular uh you know, uh programs developed to, you know, play chess with humans uh or, you know, play, go with humans. And now they, they're doing extremely well better than any super, any human.
13:11
Uh those were all developed using reinforcement learning. And even I think you were saying uh chat GP T because probably everyone talk about that combination of two even there. Absolutely. Chat GP T uses a combination of uh supervised uh semi supervised. That's one more subdivision there. And reinforcement learning. Chat GP T has uh really uh you know,
13:35
used a lot of different techniques, not some, these basic ones, but some additional ones that have been developed over the last few years and uh with lots of data, uh it has, it has really trained a very complex model. But you know, for anybody looking to understand what is machine learning about, it's really a combination of one or more of these techniques uh or you know,
13:59
using the right kind of data and right kind of solution. So, and, and even even there, I was thinking of um this is way back, we didn't even talk about this. I think Amazon for a while has a service, maybe they still do called mechanical Turk where like people could actually identify things and provide the training of like is there a cat in this image and that kind of
14:16
thing? So this is uh for anyone who remembers that one. OK. So base definition. Thank you. Hope I didn't cut off anything. You were, you were gonna add in there more. I think we're in section two. I mean, there's more techniques that have come up, but I'm sure we'll talk about them in
14:29
future, you know, generative A I is one more technique where now, you know, the machine starts generating more data. So we'll, we'll talk about it when we, when we go for. So I'm gonna close out poll number one here and I will share it back to anybody everybody. So let's uh share the results here just because that, that's part of the fun.
14:48
You can kind of see how you, how you did so or not, how you did in this case, this is just like, you know, what do you think? So if you could automate any of the below tasks with A II, I think you're seeing the Sandy. We're actually pretty evenly split here. Uh It's a mix of do my job for me. Pay my bills, go to meetings for me and even what's for dinner.
15:08
So, you know, and that was a multi select, by the way. So people could choose multiple. Yeah, I didn't realize the big market for what's for dinner. That, that feels more like you're at the end of the day and it's just decision fatigue. I, I can't make any more decisions, just get something that I like, you know, kind of thing.
15:25
Um And, and do you mind helping out? So, so Sunday? But what does the GP T and chat GP T stand for? Uh Yes. So uh the poll has closed on that. Yes, it has. Yeah. Generalized prescriptive technology. No, actually, no. Yeah. Go ahead.
15:45
It's not that it's uh you know, generative pretrained transformer. That, that's, that's the long form of GP T and, uh you know, uh every word has, uh you know, a lot of meaning attached to it and I'm sure we'll, we'll talk about it in a minute. Uh But Chad GP T is really taking uh that very complex
16:08
uh language model and using it to interact with humans in a chat format. So, so, uh I'm sure a lot of you have heard this uh buzzword chat GP T by now. Uh A lot of you have probably tried it out. Uh a company called Open A I has announced it about four months ago. Uh And, you know, you can go get an account and you can ask it pretty much anything and it has
16:33
this fantastic ability to come up with a very elaborate and confident sounding response to your question and the questions uh can be in text, you could ask it to write a poem, you could ask it to write an essay, you could give it constraints, you can sort of um change uh the constraints based on the response you get.
16:56
And the program is able to do a fantastic job, much better than a human, probably in some cases, especially the confident I, I literally saw one this morning where someone uh trained on Steve Jobs voice with audio input and actually is getting back like, you know, how did COVID affect Apple spoken in Steve Job voice? So we're, we're starting to enter some pretty
17:16
crazy territory. Which actually that feels like a segway into number two. And for those who caught the typo. Good for you. Good. Catch. That's my fault, you know, typing up the whole questions. But going deeper there is a good bit underneath the covers here,
17:28
whether it's around image processing, text processing. We've got some acronyms in, in there and more. Um Olivia or Emily, if you don't mind launching the second poll and we will just go ahead and leave that up there. And um and actually I'll, I'm just gonna leave this up because curious of what everyone is
17:46
doing for A I will share these results back. But, but Sandeep, do you mind actually kind of taking through a little bit of, of some of these deeper layers and what's involved? Yes, absolutely. So uh going a little bit deeper, I mean, we talked about the three basic techniques, you know,
18:03
supervised unsupervised and uh you know, reinforcement learning techniques uh as you look to analyze more complex patterns, you know, newer techniques or rather more complex techniques are used. And uh we'll, we'll talk uh primarily about uh image recognition, text recognition and chat GP T a little bit.
18:27
So the way uh you start out with text, uh sorry, no, keep going, keep going. OK. Uh The way uh you know, the whole image recognition works is that, you know, by nature images, you know, sort of uh think of it like a, let's say 100 pixel by 100 pixel image uh and contrast uh you know,
18:51
it's, it's something that uh you know, is very easy to split up into smaller pieces and process uh as a result, you know, some of the models that are developed to process these images. Uh They, they are, you know, the one of the most common ones is called convolutional viral network. And the way it works is it, it sort of breaks the image into smaller chunks,
19:17
think of it like a four by four window and slides that across the 100 by 100. And it is able to extract little features from that image. So there might be a line, there might be a point, there might be a circle, there might be a track. Uh So if, if it's trying to analyze, let's say the picture of a cat, uh you know, if you give it sufficient images of a cat,
19:42
uh it will start realizing that, you know, cats have two ears and they're shaped in this fashion, they have two eyes and they're shaped in this fashion and they'll soon be able to, you know, contrast from images of dogs where you know, these features may be slightly different, maybe the years are longer or, you know, they're shaped differently and so on.
20:02
So by nature, uh image recognition or image processing uh is is very easy to paralyze and one can throw a lot of compute power at it and make it go faster. Uh The reason I'm, I'm talking about this paralyzed is this is, this was a huge short coming when it came through text analysis. Uh When, when you think of text, I mean, just, just think of if you open a one page of a book,
20:31
whereas, you know, if you have a one page with one large picture and it would take you probably a second to read that picture and understand what's going on. Uh You couldn't do that with a page of text, you would have to read it line by line to understand uh text is by nature, a very sequential, you have to read words in that sequence. Uh Every word has context based on the words
20:56
around it. Yeah, that is true for images as well. But uh by naturally, we are, we are very good at uh reading uh uh image. Uh you know, but maybe the way our uh you know, brain or maybe the visual cortex is built uh it does battle processing. But you know, in case of text, this is something that I think humans invented a language is invented by humans and it's very
21:19
sequential by nature. And what happened is, you know, the text learning algorithms, the natural language processing algorithms uh up until five years ago as a result uh were very sequential in nature. They needed to keep in uh remember, you know, what were the last 20 words in order to understand what was the meaning of a
21:39
sentence. And that made it very hard to train a model with large quantities of data. So if I if I needed a model that uh you know, sort of captured the gist of an entire book, it would take hours days to process that. Imagine if you wanted to, you know, sort of capture the sense of the entire Wikipedia.
22:04
It, it was nearly because of this very serial nature of the algorithms. And that's, that's because of, of, you know, text being very serial and sequential. So it was not possible to do some of these things. Uh what had changed about five or six years ago is uh a group of researchers at uh I think Google Brain, they came up with uh some new ideas around how to position,
22:31
encode the text so that its position can be remembered. And then they brought in some new concepts uh called, you know, self attention, which is really uh putting weight on words as to how they relate with other words in a sentence, other words in a paragraph and so on. So uh they essentially created this heat map of how does a given word
22:58
related to any other word? This made it paralyzing. So this made it suddenly possible to process very large sentences and also do it in a parallel fashion, which means you can throw more hardware, you could have more GP us uh process that then it would happen much faster. So, uh that's where chat GP T is today,
23:21
you know, chat G GP T I'm told is able to process, uh you know, firstly, it has been trained uh with all the text knowledge uh in the public internet to the Wikipedia, all the books, scientific papers, 95 some languages. So it really understands how human language works, how the, you know, words have meaning based on the context.
23:45
And it is able to generate new language using that, it's able to process 3000 words. And then I just heard, you know, last week they launched a newer version of GP D which can now, you know, uh essentially process 25,000 words. So if, if you ask it to write a book, probably it could. So I think what has happened is we, we started with, you know,
24:08
uh image processing which was naturally paralyzed. And then now we've made the text processing also paralyzed. And so with Chad GP T, I think this has really changed the world now, you can pretty much, you know, give this uh program, uh your bar exam or your sad and it's able to do better than 90% of people.
24:31
So I don't know how this changes education or, you know, uh getting admissions into universities and exams, but this is really uh going to impact the world very quickly. So, you know, that, that in a nutshell is you know how, how some of these processing are working. I mean,
24:50
we're still talking at a very, uh, you know, 20,000 ft level or a 50,000 ft level here. Uh, but I mean, that's what has happened, what has changed over the last five or six years is, you know, this whole, uh, you know, self attention and, you know, being able to process large quantities of data, uh very fast and law has helped, you know, computing has gotten faster and,
25:14
you know, it's cheaper and smaller and all of that, we're gonna have a kind of a uh frankly a promotion later at the end for a session that you're doing. Actually NVIDIA GTC, we thought about talking of some of the stages of building ML models if you're interested in hearing more about some of this, uh because we're gonna jump into the use cases. Now, make sure to look at the NVIDIA GTC
25:35
session that Sandy he recorded. And um and we actually have a link for that toward the end. But as a, as a teaser of you can go way deeper on this, which is part of why you're here, even though I appreciate you summarizing it, I'll just take the five more seconds to make one place on this previous thing.
25:50
You know, the chat GP T because it's such a buzzword these days, it's the model is so large that it has 1 75 billion parameters and there's, there's no way this can run on your phone or, you know, on a smaller computer, this requires massive compute power today. And, you know, it can only run in the cloud. So there's a whole concept around model
26:12
compression that that has come about. Uh And, you know, we, we'll hopefully get to talk about it later on. So for those of you who participated in the second poll, thank you. Uh there's almost a, almost 1000 folks so curious. I mean, this is all just fascinating to see and
26:27
hopefully it even gives you attending a sense of where you may be relative to your peers kind of thing. So how far along is your organization? The use of A I it looks like there's a more of I'm not sure and exploration, there's a healthy minority though, you know, 20% that are actually um more exploring it and, you know, some that are actually pretty far along and biggest challenges and actually the
26:50
top three seem pretty evenly split. Um any commentary there, Sandy before we keep going. No, I'm, I'm glad, you know, people are, uh looking to become more and more aware because the world is, uh you know, A I is gonna thrust itself into our lives very soon. So the sooner we do it the better. Um I'm glad to see a lot of people are already
27:10
adopting or looking to adopt it heavily. Section number three, let's start to make it a little bit real from a but from a general standpoint. So this is being used already and I think we all have some sense of this probably whether on social media or, or shopping or we go to the doctor, I mean, there or,
27:29
or in financial stuff and you know, various um advertising targeting. So we decided to kind of pick a couple industries here being real. Uh some of these industries fear, fear plays in and we help the companies do what they do. But I think we'll, we'll dive into here uh retail, social media and health care. And, and when Sandy and I were talking about this, it's because like all of these should be,
27:51
I'm pretty sure everybody goes shopping, everyone is on social media in some way or you're affected by it even if you're not and then we all are human and we have to go to the doctor periodically. So do you mind uh kind of walking through a little bit from a starting with retail about kind of the use cases and challenges there? Sure, sure. I've, I've worked with a, a high end retailer
28:12
that uh that had, you know, something like 4000 product categories, uh something like 35 million customers logging in buying their products. Um Lot of data, nearly 300 million transactions a year. One of the things that you know, these uh retailers end up doing and this is true for anything that's retail, I mean, even if you look at airlines,
28:37
you know, they are, they retail, they're selling to you as an individual. If you look at, uh, casinos, they're catering to clients. You know, you, you look at the, the retail when we, you know, go to, let's say the Walmart or Target or, uh, you know, uh, Macy's or so on, you know, these are, these are all big retailers,
28:54
they really need to understand you as a customer and you know, you are not the same, you are different from, you know, the guy sitting next to you and so on. And they, it's very important for them to really cluster these and segment these customers. They have these people who are, you know,
29:11
the people who just go and buy, they don't look at price, they don't look at uh you know, coupons. They, they're not looking to say there's a very small group of that, but there are people who will just go and buy, there are those who actually will go and only buy certain things in the stores, you know, so they, they really need to understand, you know,
29:30
what type of a customer are you. So in some sense, and you know, I always say this, you do five transactions with one retailer and they have slaughtered you, they have figured you out that category one of the clusters and guess what? As soon as they have done that now, you know, their behavior towards you is gonna be, you know, they have something to base it on. So you'll start getting coupons in the email if
29:56
you have that. Uh, you know, the customer who sort of, hey, I got a, uh, $20 off for spending $100 I'll go to the store to use that coupon. Then you, you're gonna get more of those. So they, they, they get very good at, you know, what kind of products do you buy? What time of the year you buy them? Uh What is your typical budget for that store
30:18
in a year? That's, that's what it's called customer lifetime value. So they actually build AC LV model for every customer they have, and let's say by halfway or 3/4 of the way through the year, you haven't spent that money, guess what? They're chasing you because they know you've got $1000 to spend that you haven't.
30:39
So typically around October or November, you're gonna start getting these promotions because, you know, it's Christmas time, they know you, you're gonna cough up that money. So they, they, they use a combination of customer segmentation slotting you and then, you know, trying to figure out what's your propensity, what's your annual budget? And, you know, they start making recommendations based on that.
31:02
So this is a very well thought through business and uh even how they contact you is a I driven, uh whether you get a, you know, flyer in the mail. What has worked with you in the past? They'll, they'll keep doing more of that. Is it a text? Is it a email? So the retail has really figured this out, out for quite some time.
31:23
Uh, we all know Amazon will recommend us, uh, other things to buy. Uh, you know, all of that is part of this whole personalized recommendations, uh, that the retail has mastered and this is coming into all industries, this is slowly but steadily coming to all industries. Yeah, I think, and there, I even think there was like a story that you can folks can look up
31:45
of target, contacting a, a father and a daughter and it was actually a little bit awkward. It was in the news a while back. But is that good sometimes? Um let's think about and just being can I I'm watching time here. So let's um I know there's challenges here around data volume and changing customer
32:00
behavior, any quick comments there before we jump to social media, I think the biggest challenge for these retailers is really the amount of transactions and you know, the the the frequency they they they follow the typical marketing principle. They look at the frequency, they look at the recency of your transaction, they look at the monetary value of their transactions.
32:21
Uh and those three are sort of key uh data points they collect about you and that's how they figure you out. They often use third party data about you. You know, they, they sort of try to figure out, hey, are you married? Do you own a home? They get all that data from third parties, mash it up and that's, that's how they, you know, figure out what should be the
32:41
recommendation for you and, and be real like I'm, we're all human so we can see this is like this can be used for great reasons actually to make sure there's inventory and supply and what you want. But also we can see the opposite flip side. These are all tools that can be used well and, and as well, that's the thing with A I I think, you know, if you don't use it irresponsibly,
33:02
uh they can be trouble. So here again, you know, I'll, I'll, I'll be quick, I think, you know, very similar. You know, all of us use, you know, some social media or the other, whether it is Facebook or Twitter or linkedin. Uh what have you, we are getting things in the news field that nobody else necessarily is
33:20
getting. And it's again, it's very personalized for you. And, you know, whenever I say the word personalized, it means a lot of data points have been picked up, you have been put in a cluster, uh the attributes of the cluster sort of determine what might be of interest to you. And you know, the, the challenge, I'll jump to the challenge and we all know this challenge very,
33:43
very well, which is, you know, the model bias, uh elections have been won and lost because of, you know, the political bias that some of these models have had. But it, it can suffer from a lot of different biases. So that's, that's one of the challenges when implementing A I in social media. Uh but, you know, I think it's heavily used, uh Facebook wouldn't be Facebook if,
34:07
if they did not contact uh suggestions and personalizing the ads and so on, this is not even at a little bit of a philosophical level. Um And if you have a friend and you say I like this thing, they'll recommend more of that to you. But then does that mean that you start to be in an echo chamber and not hearing other
34:24
perspectives, et cetera? So these, these are A I questions, we're kind of human friend questions and sometimes are we creating our own echo chambers? Now we're getting super philosophical and cultural, but it's, it's there, there is some of that, there is some of that. I mean, you know, there is the echo chamber effect.
34:39
Uh But, you know, there are also people having these guard rails built on top of these models to, you know, sort of not, you know, drink their own kool aid sort of and, you know, just, just go out of control there last but definitely not least and maybe actually the most important one because this cuts the very core of who we are at a a physical human level. Let's think about health care.
35:02
Yeah. So I, I worked very closely with a uh very large uh diagnostic, you know, health care diagnostic company. And I, I got to see firsthand, you know, where they were beginning to use A I, they were, they were in the early stages. But I think, you know, some of the cases I definitely saw uh that all of you,
35:20
all of us can uh you know, relate to medical imaging analysis. I think nowadays, you know, the doctors use uh A I as a co-pilot. Uh you know, it's, it's uh the doctors can easily read a x-ray and figure out what's going on. But it's always good if the A I sort of helps them and uh you know, points out areas they should pay attention to.
35:43
Uh I think this will only get better. Uh maybe it uh to a point that, you know, it may even get better than majority of the doctors. Uh But you know, this, this is a big field of investment uh looking for malignancy through, you know, various types of uh you know, imaging charts, whether it's x-rays or MRI or so on. Uh This is a big area of application um in addition, you know,
36:08
just like in retail or uh you know, in other places, there is tons of medical data, you know, uh what kind of treatment you receive? What what were your, you know, uh health parameters over the years, drawing insights from those analyzing them. And this is not necessarily very structured data in retail, everything is, you know, transactions in a database here you got charts,
36:32
you got images, uh things that came out of a printer, all of those, right? So uh it's, it's, it's a much more complex problem and uh surprisingly, a lot of text uh analysis gets used in, in the medical industry because doctors have written notes about you. And you know, there's this text mining uh that happens to figure out,
36:55
you know, uh which kinds of patients need, need more care uh based on, you know, the historical records. And finally, you know, the a really hot field of genomics where, you know, you know, we, we as humans, our DNA is extremely unique and trying to figure out the, you know, which uh sequence or which genome controls what and then,
37:20
you know, uh what are the exclusions for certain types of drugs? This is a very hot area. So every time they uh come up with a drug, they, they try to make sure it, it covers the majority of the, you know, uh user base for lack of patient base, I should say. And if there are exclusions, then they're able to sometimes detect them ahead of time.
37:43
If any of you have done 23 and me uh analysis of your DNA, uh just be just from the sheer data they have gathered, they're able to call out that, hey, you are more likely to have allergies, you are more likely to have asthma. You're more likely to be a cat person, you know, or whatever they can figure out, they'll, they'll figure out and pass on those insights.
38:06
But the, the point here is that DNA as a source of data is getting mine uh and, you know, match that with this whole uh process. Uh you know, this whole drugs and, you know, uh the impact of drugs, I think this is a huge area of investment uh for, for air. I, I am not even going to touch upon, you know, the, the application of A I and Pharma uh where,
38:35
you know, they have made great advances. There's a, there's a program or a project I should say called alpha four which has really helped uh you know, figure out how the proteins look like very rapidly. Thanks to Emily Olivia, if you don't mind going and launching the third poll. Thank you, Sandeep. I'm hoping this is one that is a little more uh
38:59
coffee breaks. That's a little more deeper educational than some we've done. That's thanks to what you're going through. I hope everyone appreciates that if, if you like it, you know, feel free to say so in the chat or, or not to, because we want to go both ways.
39:09
Um I'm just gonna leave up here. The poll of, you know, which industry or A I use case is most exciting to you. If you have other ones, please put it in the chat heads up. We will probably go a little bit past the 45 minute mark here. So if you're saying around for the uh for the drawing at the end,
39:24
which we will be doing for an Ember coffee meg, um ember coffee mug stick with us. But we wanted to make sure um this is a pure storage, you know, event, what a surprise. We want to keep it educational. But we also want to show you why someone like Sandeep is here like because and he's not the only one like you, you're, you're in a product management role.
39:42
We have engineers that are actually writing the code around this stuff. So thinking about Pier One, if you were with us about a year and a half ago, give or take a year and 34 months ago, we actually did a previous um coffee break on Pier one, a little bit of the origin of the history and then a live demo for 20 minutes with Kyle Keller. So we thought what we do today is kind of
40:01
continue the pier one theme and illustrate some of what Sandi has just talked about via examples in pure one. Talk a little bit out of the A I side of it for two of these, I'll see if I can pull off a little bit of a live demo. So diving right in a classic challenge. Uh If you're in the storage industry is thinking about what does it look like when I run out of something or how long until
40:25
something that's capacity, it's controllers, et cetera. So there's this ability around hardware simulation. I love the term. I, I think it was digital twin. You mentioned, you might kind of talk about the, the A I side of this. Yeah, essentially, you know, uh all hardware uh uh is,
40:42
is very complex in terms of its performance characteristic. And I wish it was simple that, you know, I I the true rate and the performance goes up, it, it's not linear and it's not even a two dimensional thing. It, it uh you know, our hardware uh just like any other hardware is it depends on so many different factors, you know, the io size, you know, the drr the compression rate.
41:05
What kind of data are you using? What app are you? So what we have done is we have modeled the behavior of hardware under varying conditions and we let you simulate the performance of that hardware for conditions that you haven't yet experienced, but you can simulate them and say, how will I do this? Christmas will, will this thing survive the surge in demand?
41:29
You know, I think those are the kind of questions you should be able to answer with this, but even to show that a little bit. Um So in this case, I'm in, I'm in pure one being real. This is a demo environment. It's a where there's lots of arrays that are not doing what they should do because it's a demo environment, internal lab environment.
41:47
So we've got one array here. We're in the analytic section underneath planning. You can actually see if we take an array, I wouldn't sort it on load. That's kind of a combined metric of how, how busy is an array because it's more than just CPU and you can actually go here and see, you know, it actually tells you projected load and you can either write here,
42:05
say simulate hardware or up at the top simulate hardware. And you can say, well, this is an X 90 today. What if I took that up non disruptively? That's part of evergreen. You can do it without downtime. What if I took that to say the, the biggest box that's out there?
42:18
And XL 1 70 takes a second, crunching a huge amount of data. Suddenly the load metric, you know that here was going to hit 100% within 90 days. We've actually got a reset level of that. It's still doing a lot of work. We look at it from a capacity standpoint, you know, we could even go in this case, I only changed the controller. But if I chose here to actually go and change
42:38
some of the data packs I could and we could even see how, what the capacity looks like. And that's taking into account what we see under the covers with in line and post process ded duplication and compression. Because as humans, that stuff is gets a little bit hard to track how reducible is your data. Hm. That's why we have a storage way to figure that
42:54
stuff out for us. So this is stuff that I used to actually as a partner se and as a customer, I would have to do uh NR data or auto supports. And then you do manual analysis and you kind of guess based on personal experience, you can do this as a customer and even if you choose not to use it, us E can use it to give that much better recommendations to you.
43:15
So the other piece there is and this is now a more hm I I hate how little time we're giving to some of these because they're so amazing. We also do this for workload simulation. If someone says man, I want to run a couple more copies of this application. What does that look like on the storage array capacity? Mm Data reduction makes that interesting to figure out.
43:33
But what actually about the load, you know, is it running at the same times or not? So back to you, Sandy for a little, little bit of work and you know, not just, you know, scaling your existing loads, but let's say you want to run a different, completely different application. Let's say you're article, but you wanna uh now repurpose this array for running SAP or VM ware,
43:52
we can simulate that for you. So even the loads that you don't have, we can simulate and you could say, I really want a large VM ware installation on this array. And we, we can tell you how the performance and capacity will behave. Well, even there, I, I'll just drag gross across here for a second. If I go and say uh on here, I can say in that same exact same screen,
44:14
I can say simulate new workload. It's right up here and either I can choose a workload on the array or I can go and choose AAA generic ish workload, you know, as you mentioned. So let's for instance, choose a oracle is kind of a big deal for most people that run it. So you want to simulate an oracle workload or even more valuably,
44:33
I can go and actually simulate a workload that matches one that's already on the array and that's super powerful. OK? I think there's also a piece there. We've taken it even further into this idea of uh intelligent recommendations back to you. Yeah. So you know, just, just like we talked about in
44:54
the retail, we can make recommendations. I think, you know, when I came to pure II I started pushing for this idea that, hey, look, we have all this customer data and it's not really your private data, but just the metadata which is, you know, data about your data just based on that we are able tell, hey, it looks like you, you're gonna run out of space in three months.
45:15
You are not gonna be able to handle this December this year. So why don't we do something about it? And you know what we're doing is now we are proactively sending you these recommendations. Uh you know, and this is not just based on the simple projection of your load curve. We also look, you know, how have you managed such highs in the past?
45:37
Uh you know, does this really uh impact your latency? If, if this is not really causing any problem for you, we, we're not gonna call you, we're not gonna recommend you. But if it looks like you, you are gonna be in some real trouble, we'll have, you know, these proactively generated recommendations shown here by the
45:53
light bulb. And then in addition, uh this third one on the ride, we also let our systems engineers who work closely with you who understand your business. Uh Let them make any custom recommendations for you. You know, all things, all the things considered that they know about you.
46:12
We are in the best position to recommend something for you. So, uh these will come to you and all you have to do is, yeah, it looks good. You know, can you give me a quote for it? Uh So I think, you know, this is, this is a new thing that we have started in the last one year. Uh It's, it's completely data driven sustainability, uh energy,
46:34
carbon, energy cost, there's all this stuff. It varies based on the geography. You're in the size of company. I think we were even talking about, about what your energy bills have done in the last year or two. Sandy. Um So I'll, I'll let you keep going here and then I'll pull up another demo.
46:49
Yeah, I mean, you know, uh look, the way this is going is, you know, II, I live in California and my energy company PGNE sends me a bill and tells me, hey, you are 100th percentile of your neighbors, which means I'm not doing, uh I'm not like the best energy user. You're not the worst, are not the worst. So I take some heart in that,
47:14
but there are months when, you know, I, I've, I've done badly because probably my kids have turned up the heater all the way and I have been slacking on my job to turn it down every time they turn it up. So, uh I think, you know, we, we're gonna do something similar and at least help our current customers understand how they are doing in terms of their energy usage. If they have a set of arrays,
47:37
we will identify the arrays where they are, you know, under utilizing something or over utilizing the power part of it. And uh uh soon also you, you'll see in other screens, you know, even call out recommendations. They can uh you know, things they can do to improve that.
47:54
As maybe as you see here uh swap out certain a flash modules, you know, in the way we have more power efficient modules available. Uh You know, maybe that's the reason why that particular is, you know, uh consuming a lot of power. So we, we have sort of a complete uh sustainability assessment view for customers that are very,
48:14
very uh you know, tuned to, you know, the sustainability aspect of it as well as you know, the power bill aspect of it. So, so I think this is a very useful screen uh for customers to look at how, how am I doing with respect to my power consumption, with respect to my peers, my neighbors service. If you haven't looked at this, make sure to, we actually just launched a 2.0 version of the
48:36
data protection assessment that you see here. If there's a time I won't click on it but, but go look at it yourself too. And sustainability. Th this is one thing sometimes I think about if I'm talking with folks that are, you know, admins or engineers, this may be something an area that's a board level initiative,
48:49
ac suite level initiative in focus and having knowledge and, and, and even data here can actually make you more relevant up to your chart in a good way for your career, you know, kind of thing. And maybe either just because it's an executive initiative, it may be due to very real costs that they're looking at. Ok. Last, but not least here is what if, what if
49:10
you say, you know, hey, that's all sounds interesting, but I'm not really in a Capex world anymore in a purchase. I want an on prem as a service consumption driven experience where in theory pure just takes care of a lot of this store for me. What do we have for those people? Yeah. So uh we have launched Evergreen one some time
49:30
back and this is, this is very much pay as you go service, which means you, you just sign up uh for a fixed amount of, you know, capacity license at a certain performance tier and whether we ship you, you know, one array or two arrays or three arrays to meet that capacity, you don't have to worry about it. You don't have to pay for the arrays themselves.
49:51
You just get a monthly bill and uh you know, through pure one, you can see what your uh you know, monthly usage. How is it tracking uh are you likely to go into on demand in the near future? As is shown here? On the screen, you know, that red line is sort of your, uh, you know, uh, reserve, uh, pre discounted amount,
50:11
right? So it looks like another couple of months, this particular customer is likely to cross that line and, you know, go into on demand. In which case, you know, he can simulate, hey, what if I just raise that by another 20 terabytes, would that cover me for the next few months? So we have this thing at the bottom sort of,
50:31
you know, how many miles to empty, uh, for this particular type of service tells you you're good for the next four months. And uh you know, you, you don't have to worry. So it helps you manage your minutes. So to say to use the cell phone analogy and, you know, sort of it, it helps you plan a little bit into the future and stay clear of,
50:53
of the on demand charges. We did a great session with Paul Ferraro uh last year on Evergreen one. It was called pure as a service at the time, even though, you know, the burst is good because you didn't pay ahead of time. You can burst up and down. But this is where this starts to feel a little
51:09
different than your regular storage company or your better than average way better than average storage company that the peer is right. This is now going into even more feeling like cloud type storage but delivered on prem with capabilities. Even we mentioned before this the workload simulation, the hardware simulation, the intelligent recommendations that even if you're purchasing a storage service from pure
51:29
consumption based, we use that stuff under the covers to deliver you a better service, right? It only makes sense. Yeah. So we're at the home stretch at the very end here. I'm gonna actually end and share this one poll here for folks to look at briefly and we'll get right to the drawing. So in case you were curious,
51:47
um it looks like actually uh it and security management, we've got to get great question, we'll answer on that in a second here. Um But there's, it's across the board here. There's no, there's no crazy outliers and then as well, um if we could go ahead and launch the fourth pole and as we uh close it out and do the drawing,
52:08
but if you could launch that La La Olivia, that'd be great. And this is just a, it's a little bit of a self serving question, but, you know, generally curious of, are, do you actually know about these capabilities are using them or not or maybe you're on a platform that doesn't have these. So I'll just let that sit for a minute.
52:25
Hopefully you feel like you learned a good bit today both at a general level from this space kind of what it really is beyond the buzzwords. Like we said, how it's being used some of the Real World. And even though we, we made the, made it made the, the, the summary version, we're really doing stuff in this space around pure, that actually helps you and makes your life better.
52:44
Sandeep. Thank you. As I know you would be for being such a, such a great guest. If you want to hear more, please go and look up NVIDIA GTC online. Actually, I think we're talking right before this. The sessions are now active. So you can actually hear uh a version of what
52:59
Sandy Covered, but more tuned to an audience that specifically focused on a INML frankly, it goes a little bit deeper as it should, you know, kind of thing. And if you were waiting for the drawing, let me make sure I get it right here. Brian S from Utah, you are the proud winner, hopefully happy winner of an ember mug retail value, 100 and $30 the kind you can control with your phone both because it's cool and
53:22
actually, it's kind of worthwhile if you want to keep your tea at a your dar jailing tea, maybe at a, at an even temperature for a very long time. And please join us next month where I'll be hosting Justin Emerson Flash Blade technology evangelist around a little bit of uh having a little bit of fun with the concept of disc and is a disc on its way out of the data center just maybe we'll talk about that and I think we
53:45
are now at questions and answers. So I will actually just to give a little bit of the relaxed feel here. Um, we will pull up the music a little bit, not too much and this is where we are still, we are officially done for folks that need to leave. You need to get a bio break before your next meeting. That's all cool.
54:00
But we're gonna hang out for a minute for uh for Q and A and I'll actually close this poll here and uh send the results back and poll and we will share results. So we had a solid 21% people that knew about these features and used them. That's awesome for 34% of you. Hey, you learn something that'll make your life better. That's cool.
54:23
And then there's some other people that uh you know, would maybe would want to be on debt and, and to those of you, actually, I'm genuinely curious if you're on a platform that does, does all of this already. Um If you even just reach out to me on Twitter linkedin, I'd be curious to hear what you are, what you're doing. We have competitors in the space that a lot of
54:39
them have a lot of, we have friends all over the place, but what we're doing here, best of my knowledge is going further than just about anybody else. Uh I just wanna mention there's a, there's a whole bunch of things we didn't talk about today, uh, that we're working on that are a IML driven and as soon as we're ready to, you know, talk about those, you'll probably see them on our website or maybe a conversation.
55:04
Um, I, I just wanted to drop the word anomaly detection. That's a space we are doing a lot of active work in and stay. Yeah. And that one's even a little bit public because there are, there are some patents filed and issued around that area on tier side. Uh My, I'm, I'm super fortunate that my name is on one of those,
55:24
one of those patents along with some other people. It wasn't just me. So uh yes, we're doing stuff there. It's, it's semipublic. If you go search the patent database, I think looking at questions, there was um there were a couple here, Matt and Calvin have been awesome about keeping up with the questions.
55:39
Thank you. Uh There was one I wanted to highlight just because it was an interesting food for thought when it's in the in the answered section here, but it's a little bit thought provoking. And you might be like, hey, are they gonna take that one of the live? Yeah. Yeah, we will.
55:53
So uh this is actually someone asking about, you know, could an attacker train a language model such as chat GP T generate code that exploits vulnerabilities in software applications you know, to carry out attacks. And this is, I think we were alluding to this a little bit earlier, Sande, I mean, this is a, this is a tool that can be used for good purposes or malevolent
56:12
purposes, whichever terms we want to use. So I have coworkers that already are using this thing. They think about security. They don't put pure identifiable information in but to actually do the basic framework of like powershell scripting to develop loops and if thens and that kind of stuff, um any other thoughts on hearing Sandeep.
56:29
Yeah, I mean, uh that is uh one of the bigger concerns that, you know, chat G PPP itself uh has been very aware of and they have uh you know, are included, you know, everything I've seen on the web uh for the next version they put in a serious guard rail uh to, to make sure that, you know, they, they are part of the advert area, uh you know, usage of uh chat GP D uh they have
56:57
put in several guard rails to prevent uh not only the, but also uh you know, kind of things that are socially unacceptable. So they, they're put in a bunch of guard rails to prevent that. Having said that, you know, uh is it? No, nobody is even playing, I mean, G BT is fool proof,
57:14
it is error and uh uh could be, we are error in their guard rail, all of those things are possible. But I think this is one of the top challenges for A I, which is to stay responsible and to prevent, uh, I think that that's gonna be a huge challenge going on and focus on that thought over the years. Any technology is an accelerator, it's a force
57:40
multiplier what we're multiplying toward or what we're accelerating toward. That's a human decision. Absolutely. The, um, there was one other one here that I wanted to pull up. Um, then maybe we'll make this the last one because it's just a little bit of a interesting thought experiment. So, uh can a, I totally phase out human labor?
58:05
Am I redundant? Are you redundant? You know, kind of thing? Um, I, I, I've thought about this actually, some, and even in the past, like the job that I do did not exist when I was a kid, um, some jobs did, my father was in finance, for instance,
58:19
you know, kind of thing. But even that's dramatically changed, like quantitative in the quants that didn't exist back when I was a kid, you know, kind of thing. So, but there, there's always opportunities there and I'd personally rather be the one leaning in to the changes than having the change, like, pushed on top of me kind of thing.
58:34
Your thoughts there, Cindy? No, I totally agree. I think, you know, if you look at the last 100 years every time new automation of any kind has come in, it has pushed out jobs and new ones have got created, right. Those who embraced automation, you know, um, where we're at the, you know, leading edge of finding those new job.
58:55
Uh So I think the cycle will continue. Um, it's sort of uh speeding up a little too quickly that, yeah, I think, you know, the cha changes are coming far more rapidly than they have in the past. But I guess that acceleration has always been there. So, you know, uh the pace of change is always
59:18
uh been evident reason and I just feel like it was faster than ever embrace it. I think with that we are right at time uh 59 minutes after the hour for me. Thank uh for those who are still with us, which is actually the vast majority of you want to make sure to do a final reminder, you can use the QR code or you can go to pure storage dot
59:40
com slash events to register for the coffee break next month. Hope to see you then Sundeep. Thank you again for being such a great guest. Hope to see you all next month. Thank you, everyone. Thanks. Thanks.
  • Artificial Intelligence
  • Log Analytics
  • Coffee Break
  • Modern Analytics
  • Pure1
  • FlashBlade

Andrew Miller

Lead Principal Technologist, Pure Storage

Sandeep Phadke

Director Product Management, Pure Storage

Who knew that the best coffee break conversations would end up happening online? Each month, Pure’s Coffee Break series invites experts in technology and business to chat about the themes driving today’s IT agenda - much more ‘podcast’ than ‘webinar’. This is no webinar or training session—it’s a freewheeling conversation that’s as fun as it is informative and the perfect way to break up your day. While we’ll wander into Pure technology, our goal is to educate and entertain rather than sell.

For March, host Andrew Miller invites Sandeep Phadke, Director of Product Management, to the Coffee Break. While Sandeep has a storage background, for years now he’s been focused on Predictive Analytics, Deep Neural Network Compression, and more.

During this webinar, we’ll explore:

  • Stories with Sandeep - how he has worked on AI, ML, Predictive Analytics, and more in various industries before joining Pure.
  • Industry Perspective - sometimes AI/ML projects are nothing more than a FOMO exercise….so what’s the reality?
  • Pure1 - helping drive a more consistent, proactive, AI-driven customer experience via So. Much. Data.

As always, we’ll keep it educational while exploring how Pure is using capabilities in this space to benefit you. The team will stay on after the webinar answering any questions for those that want to stay longer!

Test Drive FlashBlade

No hardware, no setup, no cost—no problem. Experience a self-service instance of Pure1® to manage FlashBlade, the industry's most advanced solution delivering native scale-out file and object storage.

Try Now
09/2024
Seven Key Storage Considerations for Digital Pathology
Explore 7 key considerations your organization should consider when choosing data storage for digital pathology.
White Paper
4 pages
Continue Watching
We hope you found this preview valuable. To continue watching this video please provide your information below.
This site is protected by reCAPTCHA and the Google Privacy Policy and Terms of Service apply.
Your Browser Is No Longer Supported!

Older browsers often represent security risks. In order to deliver the best possible experience when using our site, please update to any of these latest browsers.