AI-Native NOW virtual event

0:00 from making money to Counting losses from enjoying the race to wondering who won from acing the test to waiting to

0:08 reconnect nearly every vital experience starts with a strong network connection but mere connectivity is not the same as

0:15 experiencing a great connection and when you shift Focus from Simply measuring connectivity to the experience of a

0:21 great connection Everything Changes a new vision of networking becomes

0:27 possible we call it AI native networking it's a vision designed specifically for

0:33 AI from the beginning not as a bolt-on or afterthought one that puts the experience first solves the biggest

0:40 challenges for operators elevates every experience for every end user a vision

0:46 that results in up to 90% fewer trouble tickets up to 85% reduction in Opex and

0:52 up to 50% less time to reach incident resolution that means less time spent on

0:57 tactical mundane tasks lower operational costs and more time to focus on what really matters and it all starts with

1:05 the right data the right response the right secure infrastructure delivering

1:10 AI Ops with unparalleled Assurance in a common Cloud across the entire network spanning countless use cases for nearly

1:17 any industry whether leveraging the power of AI for your network or building the optimal Network for AI assure

1:24 exceptional experiences with the AI native networking platform where every connection is reli iable measurable and

1:32 secure we make every connection

1:37 count hello I'm Jean English Chief marketing officer of Juniper Networks

1:43 and welcome to AI native now the AI native virtual event of the year we are

1:50 thrilled to have you join us today whether you're a tech officient Auto and it pro are simply curious about the

1:56 future of networking today promises to show you exactly what AI native networking means for your business we're

2:04 going to Deep dive into the how the what and the why of AI native networking

2:09 including a close firsthand look at real world outcomes achieved by our customers

2:14 like 85% lower Opex 90% fewer trouble tickets and 9x faster deployment times

2:22 during today's event we'll dive into the why the right data the right response and the right infrastructure are all so

2:29 important to optimizing wireless access wired access and sdw domains we'll have

2:36 some great sessions planned where you learned how to ensure that every connection is reliable measurable and

2:42 secure for every device user application and asset you'll get to experience our

2:49 AI data center the fastest most flexible way to deploy High performing AI

2:54 training and inference clusters we'll also explore a range of demos including a close look at the

3:01 gamechanging capabilities of our virtual Network assistant Marvis both in the

3:06 network and in the data center and you'll get to meet Marvis minis the

3:12 industry's first and only AI native networking digital experience twin in

3:18 today's world it's impossible to overstate the impact that AI is having

3:23 across Industries communities and daily life with Gartner estimating that in

3:28 just two years around 80% of Enterprises will have deployed generative AI

3:33 applications in production environments the power and benefits of AI are

3:39 undeniable it's clear that AI is here to stay However amidst the buzz surrounding

3:45 AI it's crucial to discern between genuine Innovation and mere noise at

3:51 jpra networks AI wasn't an afterthought it's at the very core of what we do and

3:57 right now that's more important than ever before because the exponential growth we're seeing with AI is also

4:04 happening with the network by 2028 connected devices are planned to double

4:11 and internet traffic is expected to quadruple with billions of devices and

4:16 users the reliability and efficiency of our networks is crucial users whether

4:22 students in a classroom or nurses in a hospital need an uninterrupted and

4:27 seamless experience that impacts empowers them to thrive in their environments and operators need a

4:33 network that does more for them allowing them to efficiently and effectively manage their Network while predicting

4:40 and solving issues before they ever even knew they exist they need an AI native

4:46 networking approach we've taken a deliberate approach to Pioneer a new era

4:52 of networking Innovation focus on optimizing Ai and unlike others who

4:57 treat AI as a mere add-on we are seamlessly integrating it into the core architecture of our Network today you'll

5:06 get to see firsthand the outcomes of our AI native platform from optimized Network performance to enhanced user

5:14 experiences our Solutions aren't just the mere demands of meeting an AI driven world they're surpassing them setting

5:21 new standards for reliability efficiency Simplicity and scalability throughout this event you'll

5:28 have the privilege of hearing from some of our brightest Minds in the industry including Juniper AI experts Visionary

5:35 customers and esteemed thought leaders all United by a shared passion for

5:41 advancing the world of AI native so without further Ado let's get this

5:46 started please welcome Juniper Network's CEO Ramy

5:52 rahee hello everyone and thank you for joining us today Juniper was born in the

5:58 age of the internet and needless to say we've played a massive role in enabling

6:04 it to become what we all know love and depend on today in fact the internet has become

6:11 the greatest vehicle for Innovation the world has ever seen but we are in the

6:17 middle of an even more seismic technology shift right now and that is

6:23 artificial intelligence which is undoubtedly the biggest Market inflection of our Time For Us in

6:31 networking we're in a once in a generation opportunity to rethink the

6:37 status quo and transform the world's understanding of what a network can be

6:43 to truly disrupt our space and fundamentally change the way we deliver

6:48 value To The World At Large let's consider what our customers

6:54 all of you are up against first users devices and apps are multiplying wildly

7:02 connected devices are supposed to double by 2028 video streaming and sharing is

7:09 rising 30% a year and Global Internet users are expected to hit six billion in

7:16 just three years overall internet traffic is expected to guad ruple by

7:23 2028 additionally digital transformation continues to reshape companies and econ

7:29 economies now it might seem like everything around you is already digital but I assure you there are a ton of

7:36 companies out there that are still doing things in archaic ways and that is

7:41 changing fast digital transformation spending is projected to reach $3.4 trillion across

7:50 the world in just two years which is almost 60% higher than last year and

7:57 don't forget the rise of generative AI in just the last year which represents a

8:04 paradigm shift as the world buzzes around its possibilities Gartner estimates that in

8:10 just two years around 80% of Enterprises will have deployed gen applications in

8:17 production environments meanwhile recent data from the US Department of Labor shows that it

8:25 sector only added 700 jobs in 2023 which is a far cry from the

8:32 267,000 added the prior year you're being asked to do way more with

8:41 less here's the thing the network lies at the center of all of this and none of

8:48 it is slowing down and we need to be more specific here because Legacy best

8:54 effort networking simply won't cut it anymore the network needs needs to not

9:00 only work but to work well because if it doesn't it can mean $35 million in Lost

9:08 revenue for Amazon for just an hour outage in its distribution

9:13 centers one million viewers missing the 100 meter dash if BT's Network goes down

9:19 for just 10 seconds during the Olympics thousands of students miss

9:26 taking a test if a school system like the capella independent school districts Network becomes inaccessible or just

9:34 flaky and at the VA with just 30 seconds of lost visibility patient elopement can

9:41 happen and a patient can wander out the door the issue here is that the focus in

9:49 our industry for too long has remained purely on connectivity performance of devices the

9:56 status of network infrastructure the speed and through of equipment now don't get me wrong those metrics will remain

10:04 an important piece of the puzzle but today Network experiences are

10:09 fundamentally broken and the world needs a new approach to delivering them today

10:15 we still say or at least hear things like I paid for Wi-Fi at my hotel but I

10:21 still can't load new emails my phone has full bars but my zoom call keeps going

10:27 in and out my internet is down but there are no outages reported in my

10:33 area just because someone is connected does not mean they're experiencing a

10:40 great connection this is why our experience first networking strategy is now more

10:47 important than ever when we make that shift from Simply measuring connectivity

10:54 to instead focusing on the experience for end users and for operators

11:00 everything changes for end users it means developers quickly and securely share

11:07 code and project data around the world with confidence Shoppers find what they

11:14 need faster Enterprises rapidly get business applications up and running

11:21 students seamlessly connect with each other anytime anywhere

11:27 instantaneously and for operators experience first means a 90% reduction

11:33 in network trouble tickets at service now 85% fewer truck rolls to Gap Inc

11:40 stores 85% faster network service provisioning for fast host and 50%

11:48 reduction in instant resolution time at Orange Egypt in fact customers small and large

11:56 are experiencing the benefits of our AI I Ops every single day and we care about

12:04 every single one of them because at Juniper every customer is important the

12:10 experience of every connection is important we make every connection

12:16 count to translate this Mission into action we've been busy building an AI

12:22 native networking platform which we just announced in January of this year AI

12:28 native networking isn't just a fancy word a framework or architecture this is a platform that

12:35 will drive experience first networking better than anyone else in the market

12:40 because real proven AI was built into the original product

12:46 architecture these are networks built for AI optimized for AI from the very

12:53 beginning not a bolt-on like some of our competitors think that they can Hawk

12:59 but how do we go about building such a platform we first need to start by

13:05 asking the right experience first questions not questions like is the network up questions like how do we

13:13 ensure every user in every location is getting a consistent

13:18 experience can the network adapt to fix issues before users even know they exist

13:25 Does the network meet the needs of all of its stakeholders our answer to these questions starts

13:32 with having the right technical ingredients first and foremost it's

13:38 about having the industry's only true AI Ops that's capable of running across the

13:43 entire end to end network but not all AI Ops is created

13:49 equal let's get specific about what makes worldclass AI

13:55 Ops it starts with having the right data data because you can't fix what you

14:01 don't know needs fixing and fortunately we had the unique foresight years ago to

14:07 build products in a way that makes it possible to extract Rich network data

14:13 from them so now we're collecting mountains of the right data from

14:18 switches and access points with routers and firewalls not far behind that allow

14:24 Network operators to understand what end users are experiencing

14:30 ing of course it's not enough just to collect the right data it's also about

14:37 what you do with it the response must be timely and go beyond flashing alarms by

14:44 understanding the satisfaction of every user on the network at all times we

14:50 solve problems for customers before they ever realize there's an issue bad AI Ops

14:57 gives you flash flashing lights and alarms often false ones the it administrators eventually just

15:04 ignore good AI Ops Juniper's AI Ops is about proactively fixing issues before

15:12 they start affecting the experience and lastly none of this is

15:18 possible without the right secure infrastructure Juniper has the

15:24 industry's most scalable infrastructure for the AI era from devices to operating

15:30 systems to Cloud hardware and software it's infrastructure that's built to

15:36 process massive amounts of data and importantly achieve gamechanging

15:41 outcomes with it so real AI depends on three ingredients it must have these

15:49 three ingredients the right data the right response and the right

15:55 infrastructure if you're missing any of these you are out of luck you just can't

16:02 take an existing on-prem solution throw it in the cloud and expect results anywhere close to what we can achieve

16:10 with the right data the right response and the right infrastructure built from

16:15 the start we've proven the efficacy of our

16:21 aops in the campus and Branch but to us aiops can and will be way more than that

16:30 if an application is having issues a zoom video gets jittery or a cell connection keeps dropping it teams get

16:37 blasted with trouble tickets from end users that often say the app isn't

16:42 working or the Wi-Fi is bad it can be like finding a needle in a Hast stack to

16:49 troubleshoot and fix incidents like these the kind of problems the effect

16:55 end user experience can reside anywhere through the network or even in the application itself a software issue in

17:03 the data center a configuration problem in a core router a bad cable

17:08 somewhere this is precisely why our intent is to expand our AI Ops from the

17:15 end user all the way to the application and across every Network domain

17:21 including campus and Branch Data Center and wide area networking because I imagine the world

17:30 where there simply are no more Network trouble tickets anywhere no application

17:36 downtime or performance issues no network outages no one even knows that

17:42 there is a network operating behind the scenes there isn't a single human

17:47 touching the controls of the infrastructure in fact there are no controls or UI at all there's no need

17:53 because everything just works and works well across every every single

18:00 domain this is a vision we set out to achieve years ago and we've been making

18:06 real progress in the last several years we now have all the pieces in

18:13 place with our AI native networking platform to bring it all together and fully realize our vision in fact I

18:21 believe that in the very near future we'll be able to guarantee that organizations leveraging our AI native

18:28 networking platform will see at least 90% less Network trouble tickets after

18:34 deploying a juniper solution that they'll speed up their deployments by

18:39 nine times that their network will see five nines of assured up time with

18:45 juniper AI working its magic in the background this is what we are working

18:51 toward and in just a short bit you're going to hear from our technical leaders

18:56 exactly how we are differ differentiating with AI not only with exciting new product updates in the

19:03 nextg campus and Branch where mist has already shined for years but also in

19:09 modernizing the data center to satisfy new AI workload specific

19:15 requirements I believe AI is the biggest Market inflection since the internet

19:22 itself and having already seen revenue from missed AI powered products grow

19:27 nearly 70% year-over-year from 2022 to 2023 I

19:34 have the utmost confidence that as we extend AI Ops throughout all layers of

19:40 the network with our AI native networking platform we will continue

19:45 delivering gamechanging value to our customers it should be clear by now the

19:53 AI isn't an embellishment for us it's the backbone of how we envision the

19:59 future thank you very much thank you Ramy so how can Juniper's

20:07 AI native platform transform your business our next speakers who have been

20:13 leading this AI Revolution and networking for many years we'll show you how using the power of AI can

20:20 dramatically reduce trouble tickets and lower operational cost please welcome

20:25 Bob Friday group vice president and our Chief AI officer along with Sadir mat

20:31 group vice president of our campus and Branch Solutions Bob and Sadir will also be joined by some of our Visionary

20:38 customers who will tell you how junipers AI native Solutions are fundamentally changing the experience of those who

20:45 operate their networks and those who use them hello everyone I am so excited that

20:52 you could join us today Sadir and I are here to talk about AI native networking

20:57 and and how it's fundamentally transforming Enterprise networking in a world where mobile internet connectivity

21:04 is becoming a musthave in our business critical applications mobile internet connectivity is becoming a necessity in

21:11 keeping us connected to our customers our partners our Associates and most critically to the business applications

21:18 that are running in the cloud nowadays you know for me personally this journey started when I was at Cisco when I was

21:25 working for some very large retail customers and I was trying to sell them a connected mobile experience in their

21:31 stores and what they told me was Bob before we put anything on this network

21:36 you have to assure me your controllers are going to stop crashing you really have to innovate faster I need more more

21:43 updates more than once a year and more importantly they told me they weren't going to put anything on that customer

21:50 phone until they were ensured they're going to have a great experience that is when I realized that there was a

21:55 fundamental Paradigm Shift where we're moving moving from helping Enterprise manage Network elements to one of

22:01 helping Enterprises manage the user to Cloud experience and so that's when we

22:07 realized that the journey needed a blank sheet of paper we looked at it and said either we're going to do this in a big

22:13 company or is this a problem an architectural change that required a blank sheet of paper and what we

22:18 realized it was going to be easier to solve this problem from scratch and T with a blank sheet of paper to really

22:25 build a cloud-based solution that can handle data in real time and secondly as

22:31 we went down this journey we started with the data right people always ask me why did you build an access point it's

22:37 not like the industry needed a new one it was more about making sure we could get the data we needed to answer the

22:43 question of why you having an internet connectivity problem and I will tell you when we started that Journey it probably

22:49 took a year to get the right data out of the aps the switches and the routers right because what we found in addition

22:55 to building this Cloud architecture we really need to organizationally take the data science team and the customer

23:02 support team and join them to the hip and that was the other part of the journey is making sure you can actually

23:09 take your customer support team that is really a proxy for all our customers and make them happy and so I tell people if

23:16 your vendor is not using their own cloud AI op solution support team they have not really started the journey the other

23:23 critical step in this journey was really around the user interface and continue

23:28 language natural language models now I firmly believe that we're seeing the user interface migrate from the CIS to

23:36 dashboards and the next user interface is going to be around chat gbt and natural language and interacting with

23:42 your network and getting data about and that all leads to the beginnings of a South driving Network that actually can

23:48 take actions on par with a domain specific it team the other part of the question people ask me is about AI

23:55 marketing white Wars reality and what what I tell people is AI is really just the automation the next step in the

24:02 evolution of automation right if you look what we've done in the past we've always built

24:07 solutions that basically like the car Robot did the same repetitive action every day day in networking we wrote

24:14 scripts that basically were very repeatable now we're actually Building Solutions that are doing cognitive

24:20 reasoning skills on par with domain expert human right we're building solutions that can really do those

24:28 problems that you typically need a human to solve and usually I break AI into two buckets we have kind of the machine

24:34 learning algorithms that we've been using over all the last 30 40 50 years what's really disrupting the industry

24:41 right now is really around these deep learning models this is what makes chat

24:46 GPT so exciting they're now building Transformer models 175 billion

24:52 parameters you know that are large and they're build solving problems that look like a real human

24:58 right and we are now doing the same thing in the networking space we are now building models where we take Zoom teams

25:05 data we're joining that with network features and we have millions terabytes

25:10 of data to Now train models to accurately predict your zoom and team's

25:17 experience and once you have these models built we can now use techniques like shapley that can now start to

25:24 explain why a user is having a problem we recently had a very large customer

25:29 who has a problem of zoom in one of their facilities in India been working on it for 3 months or so could not

25:35 figure out what was going on once we brought this model to bear on it within

25:40 half a day hours or minutes it was easy to become obvious that really the VPN distance was the problem they had

25:47 misconfigured a VPN Gateway so they were now routing packets from India to Australia looping the back they weren't

25:54 doing it for everyone and that's what we call the needle and the Hast stack problem and and this is where Native AI

25:59 native networking is really going to transform it departments the future and really helping them solve those tough

26:05 problems that humans can solve but may take months or years to solve so with that I'm going to hand it over to Sadir

26:11 who's going to go over some of the latest Innovations at Juniper here and what some of our customers are doing

26:17 with it Bob thank you and thank you everyone for being part of the AI native

26:23 now event from Juniper Networks we have had some amazing customer in retail in healthcare in education in

26:31 corporate Enterprise in K12 in so many verticals we have had full stack wins in

26:38 Wireless switching STD van and Knack where we're delivering amazing user it

26:46 and business outcomes for you our customers and partners rather than me

26:51 explaining how we are transforming it and user experiences I want to bring on Neil Collins from mus grave the largest

26:59 retailer in Ireland who I recently interviewed for today's virtual event from mosg Grave's perspective and in

27:05 particular for our superv value and Centra Brands right we had old Legacy technology so about three years ago we

27:13 engaged with juniper we went through a tender process and the outcome of that now is we have a generational jump in

27:21 our Network Technology in our stores in the old world if we were upgrading our Network in a store four teams would go

27:28 in right it would be two out of hours visits and one possibly two daytime

27:33 visits in the new world it's one visit and it's one engineer right so from a retailer point of view you're talking

27:40 about three disruptions to the store to the m of ours to no disruption to a store we can replace the network in

27:47 under one minute outage time in all our stores and we're on that road now from a

27:52 from a roll out perspective we piloted in 15 stores so we we we really were

27:57 home owned in on our processes and in particular this ear touch deployment process and for that it was all about

28:03 the router and the switch so we on our cellular network we bring up the the router deploy the switch and then it's

28:11 we're only switching for that 30 seconds we're only then moving to the new Juniper router but all of the config is

28:18 all in missed so we have everything automated it the engineer he can

28:23 download the config he can get it operational in in seconds really with the pilot stores the focus is all all

28:28 about our customer the retailer they're very much cost focused cuz it's uh it's

28:34 retail and the real benefit is they're not paying for people out of hours they're not paying for Network Engineers

28:40 they're not paying for M grave to come in and do stuff as I say it's one engineer one visit and we're we're

28:46 done did you hear him say one engineer one minute downtime in one visit to

28:53 transform their legacy Network wired wirel less and Van to the modern Juniper

29:00 Network that's AI native and Cloud native this is the kind of outcomes we

29:05 can drive for our customers like you now let me introduce you to the

29:12 future of AI in networking with our recent product announcement of Marvis

29:18 minis never in the history of networking has anyone gone where we are taking AI

29:26 but let's first look at how how Marvis finds problems today today every user in

29:32 every single AI application including today's Marvis a user has to be on the

29:38 network and has to experience failure for the AI model to be able to either learn on it or make assertions based on

29:46 whether this is a systemic problem or an isolated problem so user has to suffer

29:52 before AI can make an assertion but what happens when users aren't on the network

29:57 how do you make sure they get an amazing experience when they come back and connect on the network this is where

30:04 Marvis minis comes in to bring a more predictive proactive insight and

30:10 Remediation for your networks with Marvis minis this is a real digital

30:16 experience twin in the network natively built in when we say AI native

30:22 networking what we mean is our AI is natively embedded into our bloodstream

30:28 at this point so for wireless for switching for SD van we're actually able

30:33 to send Marvy Min out there before users experience failure before users come

30:39 onto the campus Network and be able to look at how the network is performing literally Marvis can send minis out

30:46 there let them validate the network and be able to predict and most importantly be

30:52 proactive this is what every CIO wants to be proactive before or users

30:58 experience failure so far every AI model that is out there is happening after the

31:03 fact we are the first ones to validate the network experience before a single

31:09 user connects I can tell you if your vlans are missing if you have a radio server

31:15 that's unreachable a DHCP scope that is run out of Ip space a DNS server that's

31:21 unreachable an application that's unreachable all of it with Marvis minis right now customers are bombarded with

31:29 problems like scope run over like a misconfiguration causing issues long periods of validation for deployments

31:36 blast radius determination when a user has an issue the impact of Marv minis is

31:41 far reaching we're just scratching the surface now let me give you a couple of real customer examples where we've

31:49 already deployed minis in production ashin Public Schools beautiful uh School

31:54 District in the state of Oregon Steve Mitchell and Shahid have deployed this

31:59 and have been testing this over the last quarter absolutely awesome to be

32:05 proactive to students before they go for an exam before teachers have to onboard the students onto the classroom they've

32:11 been able to remediate problems before student experienced failure a similar story much larger deployment University

32:18 of Washington EML Caldwell and her his team as they're rolling out the Juniper

32:23 M Network in new buildings as the new buildings are coming on with the jip Miss Network they've been able to use

32:29 Marvis minis to find problems proactively an example was a missing Wheeland that was not configured on a

32:36 switch ML and team found that out before students came onto the University campus

32:42 this is the power of Marvis minis with Marvis minis by the way there is no

32:48 button to turn on literally there is no configuration this is a self-driving

32:55 digital twin experience first in the industry it's natively built into the

33:00 network by Design This is AI native networking Marvis minis will learn the

33:08 ssids you have it will learn the VLS that are connected to those ssids it

33:13 will learn the DNS servers that you're using the DHCP servers you're using it will learn the applications that are in

33:21 use at every site every time completely natively it will learn and then start to

33:28 validate it so not a single nerd knob to connect and turn on literally deploy

33:33 Juniper Mist APS and minis are at work this is the power of the digital

33:39 experience twin of the user in the network it can't get easier than this

33:46 this is AI native

33:51 networking Bob and I had an opportunity to sit down with a few of our customers to hear hear from them how AI native

33:59 networking is making every connection count let's take a

34:06 look hello everybody and welcome we are excited to having hosting two of our

34:12 best and brightest customers John Zander from George Washington University and

34:17 saiv nyer from service now we have a great conversation to we had tremendous networks they've deployed with junip

34:24 pist Bob why don't you kick us off see thanks you know John you've been with us for about 2 and a half years on the AI

34:30 native networking journey and I think you've been with us for 6 years almost so I thought maybe we could start with

34:35 sharing with the audience you know how that Journey started with you and kind of broad Strokes on what were the

34:40 outcomes you've seen on this journey well so in higher education we

34:46 don't necessarily always have a lot of resources and what we've been trying to do over the last few years is really

34:51 standardize simplify and automate our Network give our Engineers a chance to engineer and get away from the

34:56 day-to-day operations so that was really the Catalyst behind us looking to make a change and Juniper was one of the

35:03 products that we looked at we looked at all the industry leaders we did a proof of concept with you guys and during that

35:11 proof of concept we were able to see the value of being able to picture the

35:16 network have a tool set that would compile all of the information we needed in one location analyze that for us and

35:23 give us real-time feedback on what we were looking for and after that it was a matter of just transitioning an

35:30 enormously sized Network and higher education and it couldn't have gone smoother we have over 6,800 access

35:37 points spread across three campuses 160 buildings we managed to get that all

35:42 replaced in just over three months to give you some context the last

35:49 time we did half of the aps on campus took two years wow so this was a

35:54 tremendous kind of Daisy zero to day one integration with our systems that went

36:00 extraordinarily well and we could have been more pleased with it well thank you and S maybe your viewpoint on you know

36:06 you've been on the journey a little bit longer yeah I think uh for us uh it was 2018 when we were you know kind of you

36:15 know changing the direction of the company for you know much higher Pace growth and the complete Focus was around

36:23 you know how do we simplify things and set ourself for that scale right and uh

36:28 you know user productivity user experience was one of those top drivers

36:33 that we identified at that point of time and anyone that we interacted at that at that moment that in that era we were

36:40 always looking for that that commonality in that thinking and the DNA right and I

36:45 think that was uh the year when we met the M team and one thing that resonated

36:51 super well with us is that you know for them the the simplification The

36:57 Experience driven networking those are all some of the things that that completely you know sat well with us and

37:04 we decided that you know we should we should basically embark on this journey with mix Mist in the mix and um I think

37:12 that's how it all started and uh we've come a long way and uh I have some amazing stories that I will share as we

37:18 as we go along here awesome thank you saiv and so John for a lot of the customers watching you know they are

37:25 today running Legacy controller based architectures which you have run yourself as well and so compare and

37:32 contrast you know the day in the life of running the Legacy Network and and and

37:37 now the Juniper Miss Network sort of what is your team seeing differently and you know the for the customers listening

37:44 why should they consider this transition so it's it's a very good question because it doesn't always come

37:51 off as obvious when you talk about just the controller versus the AI driven platform but but in our experience so

37:59 far with mist we've actually seen an incredible reduction in the number of day-to-day tickets that we're seeing our

38:05 day-to-day operations from an engineering perspective are exactly that we are looking for the next thing that

38:11 we can do with wireless to provide for the student experience we're no longer kind of chasing our Tails trying to

38:17 figure out which element of the network is having an issue uh we've seen anywhere in a month-to-month comparison

38:24 between last year and this year a 40 to 70% reduction in tickets and that's

38:30 tickets as a whole for wireless that's not just tickets that are escalated to my senior engineering team those are

38:35 even fewer we've also found that even with network elements that are not a part of mist uh I'll give you an example

38:43 we had an issue with uh one of our authentication servers falling out of sync with the cluster earlier this year

38:50 and typically that would cause an investigation where we'd have to troubleshoot multiple aspects of the

38:55 network narrow it down identify the issue in this instance Marvis was able to point us in that direction to the

39:02 specific server in that cluster by telling us that there was a 275%

39:07 increase in errors on that authentication server I mean the time to

39:13 resolution on that was night and day between what it would have been and what it is today that was probably the

39:19 biggest experience we've had so far and how this can really impact our ability

39:24 to keep everything up stable and working for our students yeah you know John that's a great story you know I always

39:30 tell people like AI is what's really helping it teams find these needle and the Hast stack problems now you know so

39:35 D now maybe back over to you a little bit you know we started this journey with wireless access points because we were trying to answer the question of

39:41 why you're having a poor internet experience you know since joining Juniper we're starting to extend Marvis across the switch and the router maybe

39:49 you can share with the audience a little bit what you've seen on the journey in solving problems and the outcomes you've

39:54 seen inside your services team as we extended across the stack yeah yeah it's

39:59 a it's a great question and I think there is a a phenomenal amount of learning that we have gained through the

40:05 transition I think it was during covid and we had this opportunity to make some bold decisions and I think as a as a

40:13 company when people were all working from home you know some of the decisions were like were were required to be taken

40:20 because the world was changing and we some we we and our journey that was started in 2018 with with user

40:26 experience with Simplicity of our architecture um the scale that we wanted

40:31 I think our thinking was very much in that direction and we made this decision that you know we want to be you know uh

40:38 you know environmentally uh friendly in when it comes to network people don't

40:43 think about it much but uh you know there is there are some very fine elements that that actually defines that

40:49 and then there is uh the people were working from home and they were ready to come back to office whenever covid uh

40:55 you know change uh you know things um and we said okay you know do we really

41:01 need uh you know this dense Network and can we simplify it and go Go Wireless um

41:08 completely and and then only keep the iots on on on wired so as we started thinking about it we were also thinking

41:14 that how do we scale this whole approach with a common road map common mindset

41:23 and I think the the answer was you know we wanted to kind of extend this journey and you know continue this journey with

41:29 a partner that had this going across the board so you know we we didn't have to

41:34 look too far you know Juniper um Miss combination was you know going Beyond

41:40 Wireless to Wi to when so we decided to kind of go single stack and go you know

41:46 U the the the switching infrastructure over to jenniper and then the uh and the and the SD van over to SSRS so that's

41:54 how you know we transition and I think the the biggest learning that we have got out of it is that the capabilities

41:59 of Marvis and the AI that was limited to Wireless at one point is now extended

42:05 across the board and um it is uh something that I'm going to you know steal from Juniper's book which is that

42:12 we said okay you know what this client to Cloud journey is looking like a reality now because I can see things end

42:19 to end this was not possible before and and saiv honestly speaking thank you very much for being a partner driving us

42:26 towards this and most recently you've added one more part of the portfolio uh

42:32 which you haven't deployed yet but you've tried which is Knack uh so talk about sort of how Knack came into the

42:37 picture yeah so um you know as uh as we are transitioning simplifying things uh

42:45 we we have a a legacy Knack system that is I can sort of sort of say that it is

42:50 an on Prem system and being a SAS company and very SAS friendly you know we we want to kind of make things pretty

42:57 lightweight for us and um in one of the products strategy Summits the idea of uh

43:04 a SAS Knack offering was was kind of uh you know socialized and we were like

43:11 super excited when we heard about that so we are now getting very close to uh

43:16 adopting the The Knack solution that um you know Juniper Miss team has has come

43:22 up with and we're looking forward to the to the benefit and the scale that it will deliver in you know comparison to

43:29 um what we are not able to achieve today awesome thank you so John um you've summarized for us uh the three things

43:37 that I always tell you know uh first of all as as much as sajiv is here highered

43:43 is probably the hardest networks to run because your network is changing on you

43:48 every three months kids are coming in with new applications new devices all the time and so the three outcomes I

43:54 always you know promise High customers is faster deployment fewer tickets and

43:59 fastest meantime to resolution and you've already hit on all three of them at a broad Strokes level anything else

44:06 that your team has seen in terms of Automation in terms of integration with other systems any other benefits people

44:12 should think of as they consider J prist yeah I mean we've uh we've been

44:18 looking for a while we we call it we want to provide the at home experience for students and one thing that we've

44:25 never had the resource to be able to do is to kind of look into the private pre-shared key networking and that's

44:31 something that we're currently piloting in one of our dormitories for our students and we view it as something

44:37 that go beyond just a student experience we are a research institution we have

44:42 various use cases in the classrooms where this could be applicable and this is something that we're very excited to

44:48 finally be able to do and Juniper has given us the tool set to be able to do that and much quicker than we thought we

44:55 were able ble to because of how well things have gone with our deployment of junest awesome thank you John and for

45:02 everybody for all the audience listening my closing statement with with most of our customers when we first meet

45:08 them and I said this to John and saiv as well is we're not a vendor to you we're a partner to you we want you to feel

45:14 like we're an extension of your team John I want to thank your team in in helping us build this you know atome

45:21 experience for a student in the dorm uh dorm room we're so close to actually

45:26 launching that in a big way and you know blending policy across uh pre-shared key And1 xss IDs with our Knack solution now

45:34 ready for edu roam all kinds of interesting things are possible so we're stoked one thing that everybody

45:40 listening you've heard just just in the previous session is this brand new announcement around Marvis minis Marvis

45:48 minis is is our ability to be proactive to the network you've heard me over and over talk about the impact it can have

45:55 and one thing we've partnered with our customers to to do is to build Marvis Minis and we asserted to ourselves we

46:02 are not going to launch it until we tested with production customers and so s sajiv thank you for uh you know

46:09 helping us build and partner with Marvis minis what is your thoughts on the impact Marvis Min can have and what has

46:15 your team seen you know this is uh uh something that we've been waiting for for for a long time U was another great

46:23 idea that was socialized during one of those strategic conversations and uh for us uh you know when we think about the

46:31 capabilities that you know the junipers and Mist stack brings together the zero

46:38 touch provisioning was one of the big things that helped us you know go through this whole network

46:44 transformation during the covid era zero travel no people to go and deploy stuff

46:49 or they were like very limited people so I think in my mind if I if I kind of think about it you know uh we solve the

46:56 deployment side of zero touch but when it comes to the changes and the post change testing there was no ability for

47:03 zero touch we needed users or people to go into the offices and do the testing I think that's where I see Maris Marvis

47:11 minis is kind of doubling down on that zero touch concept into the end user

47:17 area right it is it is giving us the ability to do um you know self-initiated

47:24 test post changes we can do continuous experience monitoring uh we are pro

47:30 potentially also looking at consolidating some synthetic monitoring tools here because we can go and in you

47:37 know introduce new uh service URLs that uh that is a great capability and I'm

47:43 super excited and we are we're just getting started with this and I think uh this whole uh you know whole narrative

47:49 around experience is going to just just get immense boost uh with milk Minis

47:55 awesome thank you Bob go ahead you know and S I want s I want to thank you because you know in addition to being just a great customer you've been truly

48:02 a great partner right marvelous minis is a great example where it actually started with our user Community right

48:07 it's been one of the most requested items for how many years to year it took us a while to build it but we're here now but you know so D maybe we can also

48:14 share with the audience the other big transition is really going from this Legacy controller architecture to This

48:20 Cloud AI office architecture I thought maybe you could share with them any best practices potholes you know what was

48:26 that transition like you know in helping the team go from kind of this SSH CLI

48:31 world into this Cloud I Ops you know we're not we're not sshing into these devices anymore as we connect them to

48:37 the cloud I I always say that you know uh there is uh there is Technologies and

48:42 tooling out there and and but that alone doesn't get you where you need to go you need to make some bold decisions as a as

48:49 an IT leader or a or a or a it group that is uh wanting to you know Drive the

48:55 the digital transformation uh we preach as service now about a lot about digital

49:01 transformation so it has to kind of start within right so for us it was like a combination of bold decisions okay we

49:07 need to we need to kind of get get down to a place where uh this is you know this is not the normal this is this is

49:15 this is the future right we want something that will scale with us we don't want to be building things uh you

49:22 know as we were growing you know it should just scale by with it with with with the with the growth that we have so

49:27 I think in in in terms of the learnings I would say uh Marvis and Mist uh offers

49:34 a lot of tool tools uh in the form of you know AP placement tools you have uh

49:39 you know the we talked about the the minion there are there are a bunch of other things that you can like zero

49:46 touch uh and and and I think John you mentioned about the the prad key so we just rolled it out for our mobile users

49:54 and that's a major hit because it it completely uh took the key in the password equation out of the mix right

50:00 so so big win and if if there is if there are anyone who's in the audience

50:06 you know service now customer if you're not one you know you should become one soon but I'm saying that there is a

50:12 there is a phenomenal amount of Integrations that we have built Marvis actions are now able to communicate

50:18 directly with the service now platform through the integration we're also uh able to do Global uh like a

50:26 you can upgrade the firmware on your Fleet no matter where it is uh is with a single change order ticket so these are

50:33 some of the innovations that we have been able to work in partnership with jenniper not just taking what they have

50:39 to offer but also bringing our platform into the mix and and kind of just uh putting this whole thing on steroids

50:46 thank you sajiv because I think for service now customers who are listening in uh a lot of what sajiv is talking

50:51 about is is available on the platform and and sajiv and team have been uh very forthcoming in sharing their best

50:57 practices as well so as we wind this down John I think the the biggest hurdle

51:03 for our customers is taking the first step so what would you advise uh because

51:08 you know the the what they're thinking is wow this all sounds great this is where is the future I want to get to but

51:16 I have a big Legacy system uh my my team is trained on it my team is you know we

51:21 spent all this time perfecting the art of it what is your advice for customers

51:27 uh um as to how to get started absolutely I mean that's definitely the most difficult step to

51:33 take and I think knowing that this product can stand on its own through an

51:39 evaluation we ran a fairly vigorous proof of concept with this platform and

51:44 others and I would recommend if you had the capability to choose an area highly

51:50 trafficked at your University run a proof of concept see the difference that it can make and use

51:56 that as justification to move forward with it we had no doubts once we were done the proof of concept the direction

52:03 we wanted to head and the benefits that it would provide to our faculty staff students our entire

52:08 Community awesome uh saiv any words of wisdom uh as we close this no I think I think John captured one one important

52:15 point which is you know go for the proof of concept proof of value establish that

52:20 and then um you know and look at this whole thing pretty comprehensively right uh AI is here to stay if you're not in

52:27 the in the game in the business of you know dealing with AI or interacting with AI you're definitely losing something

52:33 and then and then you know what is on the paper you need to establish that on the field so you know proof of concept

52:40 will be your way to go to to get started and as I mentioned about the comprehensiveness of it look at it from

52:46 the whole total cost of ownership you you you want to own this network for a

52:51 long period of time you know how does it uh how does it uh how much does it cost to to run it you know you don't want to

52:59 have an army of people running your network so so think look at it very comprehensively so POC POV uh AI

53:06 benefits uh total cost of ownership take a look at this whole thing and make the

53:12 decision awesome thank you Bob thank you saiv thank you John uh let me summarize it John you said 40 to 7% reduction in

53:20 help test tickets coming in uh saiv on the wireless side you had a 90% reduction support tickets wired side I

53:26 think 30% reduction support tickets uh faster time to deploy and faster time to Value you've seen uh you said three

53:33 months for 6,800 APS and and I think you've had a a significantly uh accelerated roll out and then the total

53:40 cost of ownership you know in your case you did the maap on the wired van side almost a 70% reduction in uh in uh the

53:48 TCO uh that was out there um so tremendous outcomes for all of you listening in if you're still wondering

53:55 all it starts with is one POC John said it sajiv said it we if you see it you

54:01 can't unsee it with juniper Miss thank you very much for joining us and look forward to seeing you out in the

54:09 wild our next speaker has been in the field of AI since she worked on the iconic Watson project at IBM since then

54:17 she had many roles as an AI strategist and executive helping companies design

54:23 and deploy effective AI strategy IES please welcome Soul rashidi former Chief

54:29 analytics officer at Estee Lauder hi everyone my name is soul

54:34 rashidi and I am absolutely thrilled and happy to be here at Juniper I know there's a lot of information about AI

54:41 coming at you from McKenzie reports to Bane articles to Harvard Business review it seems like every media channel on

54:47 Earth is reporting on AI and hopefully with today's next 20 minute session I can actually ground us on a little bit

54:54 of hype versus fiction some applicable use cases just some stats of how the Space is really growing um but more

55:01 importantly just to teach the art of the possible while balancing things with risk exposure and maturity levels uh

55:07 just a little bit of context and background as to who I am by no means am I an academic nor am I a researcher I

55:13 accidentally fell into this space it's not something you choose to do it chooses you I thought I was going to

55:19 play professional sports the rest of my life until in my mid 20s I realized okay I'm not going to be a professional

55:25 athlete maybe I should be a professional and I took the one job that no one wanted and that was to be a data

55:30 engineer fast forward uh I joined IBM very early on in my career and I owe a

55:35 ton to IBM and in 2011 when Watson beat Ken Jennings in Jeopardy IBM decided to

55:41 take Watson to Market and when that team was being formed I went with it because you can't do AI without data and I was

55:48 the data expert within IBM at the time I've been very fortunate very blessed that I've had some wonderful roles with

55:54 some great companies and great Brands and along the way my team and I have developed some pretty cool things um not

56:01 without its issues not without its challenges because when you're trying something new there's bound to be

56:06 hiccups and face palms and scraped knees and bruised elbows and wounded Pride

56:11 along the way but inevitably there's always a rainbow at the end of the storm and so this is a little bit of a journey

56:17 for you now I know you guys have a lot of facts coming at you there's tons of Articles things from media channels to

56:23 the Harvard Business Review to mck reports to Bane to LinkedIn it is very overwhelming for those of us who are in

56:29 the space it's overwhelming we can barely keep up with it so I just wanted to share a few facts and figures just to

56:35 kind of ground us on the potential the impact and where it's going this one's

56:40 not going to be another another fad we may not be talking about it as much in a year or two but it's fundamentally going to penetrate all the processes that

56:47 we're aware of um and various ways so just as a just a level set us everyone's

56:53 talking about it but but there is a fundamental priority that's been set at the board of directors level and this

56:59 conversation is being had within the board of directors and the executives it's not as trending as blockchain was

57:07 or web 3.0 or Big Data ecosystems but applied Ai and now

57:12 generative AI is definitely dominating the top of the charts when it comes to what's being discussed and how companies

57:18 are approaching their AI strategy with that however most of the conversation's always around the applications the use

57:24 case what's what do we see in front of us how can we use AI there's been very little

57:29 attention unless you're in the scene about the backend infrastructure and what I call our defensive playbook

57:35 because to fundamentally run these models to get exposed to these consumer-based tools there are a lot of

57:42 Investments being made in gpus CPUs tpus workloads and no one's talking about it

57:50 but even the electricity that's needed to be able to power these things there is this fun fact that I had read read that if everyone was leveraging AI at

57:57 this point in time we actually would not be able to support it because we don't have enough electricity being generated to support all the workloads and so from

58:04 this you can kind of see the Investments That China is making that we're making and then that the global economy in and

58:09 of itself is making for stuff behind the scenes that we don't get to see and what's interesting is a few partners and

58:15 I had a conversation we even think that this is underestimating the amount of

58:20 Investments that's being made behind the scenes to be able to power all the consumer facing AI applications and

58:26 tools that we get to use now there's been

58:33 some mixed results and mixed information around whether or

58:38 not the VCS the private Equity firms are they still investing is there still

58:43 enough capital in the market space to invest in companies that are doubling down on AI and the answer is while

58:49 unfortunately our brothers and sisters in cyber aren't getting the intention or

58:55 in software development aren't getting the attention any company that's been born in the past year year and a half

59:00 and continues to be born will actually get the lines fair share of the Investments that's going on within the

59:05 VC World specifically if there is emphasis on the generative AI

59:10 capability and last but not least this is kind of one of my favorites this was released by McKenzie last year

59:18 um facts aside data aside what this is telling us is it doesn't matter what

59:23 industry you're in it it could be agriculture it could be education it could be energy it could be manufacturing it could be Health Care

59:31 there's several functions that are fundamentally being impacted by this but sales and marketing actually happen to be the top two now why are sales and

59:37 marketing well for a few reasons marketing is a function

59:42 that generates content generates images and generative AI fundamentally creat

59:48 new with existing so it's a space that's seen fundamental amplification and um

59:55 facilitation and amplification of the workloads and so something that may take you six hours can now actually take you

1:00:00 30 minutes it doesn't replace the job it's not replacing your creative thinking and critical mindset and your

1:00:06 judgment and your common sense and your Aesthetics but you actually get to leverage these tools to be able to

1:00:12 amplify and expedite your work no different than using email to communicate no different than using Excel to be able to do tabulations and

1:00:18 calculations and no different than using Adobe to do a lot of your graphic design it now becomes a tool and the toolkit

1:00:24 for marketers and on the sales side of things you know I've helped some small companies and medium-sized companies and

1:00:30 what's the one thing that all sales functions have in common they're not the best at data entry so their job is

1:00:36 really to pursue the client build the relationship with the client and have them go through the entire sales life

1:00:41 cycle from lead generation to close but a majority of their work is what the

1:00:47 communications and making sure that the proposals are in place they have the necessary information to be able to take

1:00:53 that sales and get it going through the funnel as quickly as possible so when it

1:00:58 comes to data entry into CRM systems when it comes to understanding who you're talking to the point of contact a

1:01:04 lot of that stuff is still done through email communication so there's a lot of great capabilities that automates that takes content from emails autop

1:01:11 populates it into the CRM so now companies don't have to be frustrated with their sales staff anymore and I

1:01:16 know these aren't the most glamorous use cases but if you only understood how much time and energy spent on this

1:01:22 process alone you'd have deep appreciation for it now just to ground us now that we've

1:01:28 gone through some facts and figures and we understand the growth and where the Investments are coming let's talk about and switch gears into applications and

1:01:35 some of the hype that's out there so here's a little fun fact Ai and teenage topics what do they have in common well

1:01:42 everyone's talking about it no one really knows what the topic really is but they assume the other person's doing

1:01:48 it so then they start claiming that they're doing it and that's kind of what's happening with the AI scene but

1:01:53 in actuality based on statistics that were shared last year only 31% of the companies that are having the discussion

1:02:00 really have something in production or in po people are still kicking the tires around it they're still trying to figure

1:02:06 out what it means for their organization they're still trying to understand is our risk exposure too much if we were to

1:02:11 explore these capabilities or how do we combat the fears of job replacement versus it being more of an amplifier

1:02:17 than a replacer and so I'm very happy personally that large Enterprises are taking the time to be deliberate about

1:02:23 this stuff I don't think it's too late to start by any means but I do think everyone needs to start here's another

1:02:29 statistic that was shared last year about the types of use cases that are really in deployment and what's key here

1:02:35 are the top two customer operations and anomaly detection I'll give an example American

1:02:41 Express has been doing anomaly detection for how many years now decades and they are unbelievable at it so it's not

1:02:48 necessarily a new use case it's just that others are finally coming on board and feeling more comfortable and and

1:02:54 deploying these use cases for their specific need and then for customer operations you know on average by the

1:03:00 time you hire someone train them have them Shadow a customer call and then put them into production companies go

1:03:06 through that process anywhere from 7 to 10 months to get someone you hired operational they can now shorten that

1:03:13 time frame to two to three months how well imagine you have a customer service department and you sell consumer goods

1:03:20 and someone calls and says I want to know which one of your products are vegan for or is vegan well how's a customer

1:03:28 service rep supposed to know that if they have 880,000 different SKS within the product portfolio it's nearly

1:03:33 impossible but to have a virtual assistant that has that information where they can type and understand which of the products are vegan based is

1:03:40 extremely powerful so it's being used as a knowledge store it's being used as a virtual assistant to be able to not only

1:03:47 empower the individual servicing the call but to actually operate as a better service provider for the customer that's

1:03:52 calling in and so again not the most glamorous use cases but these are the ones that are highly highly used and in

1:03:58 production right now all right so as I've mentioned there's tremendous growth there's tremendous opportunity you guys

1:04:05 are reading the hype in every media channel possible but I do want to level set and I do want to ground us a bit any

1:04:12 AI deployment is not a silver bullet it's not a light switch that turns on it's not a toilet you flush and it just

1:04:19 automatically happens there's a lot of time and effort and focus and alignment

1:04:24 and grit and ambition and and resiliency that goes into it even with tiny really simple projects there's still some time

1:04:32 and focus and funding that needs to be spent so just know that AI itself is not a silver bullet and I think the second

1:04:38 part of it is is you know sometimes you just have to know the applicability if you've got a business problem lean on

1:04:43 the experts and ask do I need AI to solve this why use a sledgehammer to hit a nail on the head when you can use a

1:04:49 hammer to get the job done there are a lot of tried and proven Technologies out there that would expedite the whole

1:04:55 process because your goal is to fix the problem your goal is to address a business issue your goal is not to use

1:05:01 AI for the sake of AI so I always say be very careful if the board of directors is asking for it be very careful if the

1:05:07 executive committee is asking for it because it's not about the technology but it's about what it is that you're

1:05:12 trying to do and does AI or AI not fit into that picture now to further ground

1:05:18 us um I've had the pleasure of working some phenomenal companies and help them develop their AI strategy and presented

1:05:24 to the board on numerous cases I think the thing to also remember is with any deployment that's about to

1:05:31 take shape or that's in discussion is be realistic about your current maturity levels and what you could reasonably

1:05:37 deploy it's very much a crawl walk run approach you know after new years's

1:05:42 everyone signs up for the gym but you're not going to run the marathon in two three months afterwards it's a muscle that fundamentally has to be developed

1:05:48 and the same goes with AI projects just because the amount of discussions and Alignment that's going to happen so it's a muscle that has to be exercised

1:05:54 think big start slow and then yes you can scale quickly crawl then walk and

1:06:00 then yes you can run so if you're developing an AI strategy it's really important to be thoughtful about the

1:06:05 next year two or three years but it's more important to be realistic about what it is you can actually deploy and

1:06:11 so I always say if you're developing your AI strategy there's a series of things you need to assess from risk

1:06:17 tolerance data infrastructure Talent uh your ability to focus the funding needed

1:06:24 for and how many other strategic priorities exist right now within the company and all of that will lead to

1:06:29 into five basic AI strategies from is it efficiency based effective Effectiveness based productivity based growth based or

1:06:36 expert based and from there you can then start ideating and come up with your use

1:06:42 cases now the other big thing around the hype is AI is not this big hairy beast

1:06:47 at something to be scared of it can be as simple as off-the-shelf Solutions so if you want to create better content

1:06:54 better client presentations or you want to automate the entire note taking process when you're on a sales call so

1:07:00 that instead of you having to actually take notes while you're having a discussion with the client summarize it

1:07:06 put it into an email and then have follow-ups you can actually automate that process it's something that I

1:07:12 personally do right now I used to take anywhere from 15 to 20 hours a week of just post call follow-ups and now I've

1:07:18 been able to reduce that to 1 to two hours on my own as an individual by leveraging off the-shelf solution that

1:07:24 are out there but this is where connection really matters and this is where understanding your capabilities

1:07:29 really matter if there's a slide to emphasize I would say it would be this one this is probably about my 13 or 14

1:07:35 years of scraped knees and bruised elbows summarize into one slide and this is a framework of all the things that

1:07:41 I've learned over the years we don't have time to go into everything but I am going to Spotlight a few things like the

1:07:46 ability to define AI first or what does responsible AI even mean I have a funny game that I play when I first join a

1:07:54 company I ask them to Define governance for me I ask them to define AI for me and you ask 10 people you're going to go

1:08:00 over a hundred different responses but the factor of the matter is is I always say there's three different definitions for AI there's automated intelligence

1:08:06 augmented intelligence and artificial intelligence in all my years I have not come across a single use case at least

1:08:13 for companies where it required artificial intelligence a lot of it is really focused on automated intelligence

1:08:19 and augmented intelligence so those use cases around the customer service reps creating these ual agents and these

1:08:24 knowledge stores is all around augmenting human capacity to do more and

1:08:29 reducing the pressure off of them and everything associated with automated intelligence is how do we create

1:08:35 Automation in the day-to-day manual repetitive tasks to streamline our efforts so that we can reallocate our

1:08:41 human capital the stuff that really matters and so it's always a good exercise whether it's at the board level

1:08:47 the executive committee or even middle manager management level to just ground and level set on what AI really

1:08:54 means the second is validating Ai and I absolutely adore and love this graph

1:09:00 from MIT and this is a fun little exercise I won't go through every single detail for you but it's a really good

1:09:07 way and you can Google this as well of helping Executives and those that don't come from a technical background in

1:09:14 understanding what is AI and what is not AI so a fun little exercise right can it

1:09:19 see if it can't see the system the model the algorithm the machine whatever it is

1:09:24 you're working with it's a no-brainer it's not AI can it see well yes it can

1:09:30 see can it recognize images yes it's starting to recognize images well now you're leveraging image detection and

1:09:36 computer vision you're starting to get into the apprach of AI in the realm of AI and and you can see can it see can it

1:09:43 reason can it hear can it speak can it interpret and this is just a great litmus test so that you can explain to

1:09:50 non-technical individuals what truly is the space of artificial intelligence Ence now another part that I think is

1:09:56 very important for you to be successful is setting expectations Watson beat can jenings in

1:10:02 jeopardy with a 71% accuracy score for those of you who have children or nieces and nephews when have they ever come

1:10:09 home and got a C minus on a test and you're like job well done you don't and

1:10:14 the fact of the matter is is that 71% threshold is really low but it was better than the human intellect and so

1:10:20 it got a lot of notoriety and that's what kicked off the artificial intelligence wave but it's important to

1:10:25 know that even though we're at a lot better statistics than 71% right now it's still not perfect if you were to

1:10:33 test some of these machines or models can it tell the difference between a Chihuahua and a blueberry muffin it

1:10:38 struggles I played one of the image generation tools and put salmon in water well you and I will contextually know

1:10:44 that that's not what we meant it wasn't meant to be a salmon fillet but a fish but again it didn't understand context

1:10:50 and one of the favorite ones that I actually saw in LinkedIn was someone typed in an eggplant you and I obviously know that's that's not what we're

1:10:56 talking about so it's not perfect it's not 100% but sometimes you have to ask yourself

1:11:03 is it better than the human error rate that's happening within your business process that you're looking to replace with AI and is there enough data where

1:11:11 it could understand contextual understanding because right now if it were to take the lsats or the mcats it's

1:11:18 doing a lot better than what humans would because it's a finite data set which is something you should also take into consideration

1:11:24 now I'm not going to say last but not least because there's much more to the framework but I think there's two really important elements as well there's an

1:11:30 element of fear around security and data leakage and the whole umbrella notion around responsible AI there are many

1:11:37 many many tenants to this and unfortunately we don't have time to get through all of this but what I do want

1:11:43 to let you know is for those Enterprises that are fundamentally considering either developing off of a pre-built

1:11:48 model or developing off of an API coming in from an llm provider I understand that data leakage is a concern but there

1:11:55 are so many companies that are solving this problem right now protopia being one of them and a dozen others but the

1:12:01 intent is to make sure that when information is shared it's the right information and the second is really

1:12:07 around human human and loop which is something I personally experienced I woke up one morning with the Google

1:12:12 alert that essentially says there's an article on Google 100% about my husband I woke up my husband I like did you

1:12:18 write this did you verify this he's like I have nothing to do with this and long story short there was an SEO company

1:12:25 that sold a service to a marketing company and they chose keywords and they had articles but the marketer who bought

1:12:32 the service from the SEO AI based company did not fact check anything did not fact check the keywords and somehow

1:12:39 my name showed up in those keywords and did not fact check any of the Articles supporting those keywords and you now

1:12:45 have this bogus article that's on the internet that neither my husband and I are a part of so if there's two

1:12:51 takeaways from responsible AI is there are solutions in place to solve data leakage data security you just have to

1:12:58 do the research and I say it's not if we should do it it's how we should do it so it's not no but it's how and then don't

1:13:05 ever replace anything else with human in the loop we need our contextual understanding in our lens to make sure

1:13:10 nothing goes off the rails with that said last but not least data I know we can't have ai without

1:13:17 data I know there's a ton of fatigue with this but there's a few things that I want you to consider companies do a

1:13:22 phenomenal job of collecting the data we do a horrific job of connecting the data

1:13:28 so if you're trying to combine multiple data sets to be able to leverage and build um an AI capability just know

1:13:35 you're going to have to dedicate some time in that data architecture because it hasn't been done before the second

1:13:40 that I think everyone underestimates is it's not necessarily the accessibility

1:13:45 and quality is always a concern but the longest poll in the tent is actually getting alignment with your internal

1:13:52 stakeholders from it to infosec to the compliance Department to the stakeholders of the risk exposure what

1:13:59 the tolerance level is and what they're willing to do and take on versus what they're not willing to do believe it or

1:14:04 not those conversations can take anywhere from four to six months in that 9month timeline so I always say allocate

1:14:11 70% of your total project timeline just to the day discussions alone and hey if you get some of that time back kudos to

1:14:17 you with that said I really appreciate this I thank Juniper for everything this is wonderful there's so many smart

1:14:24 people out there leverage their talent read up it's an exciting space I truly don't think AI is a trend and it's going

1:14:30 to be here to stay we just have to be thoughtful and deliberate about it our AI native platform only begins

1:14:38 with solutions for campus and Branch it's vital for customers to have an AI strategy for their data centers as well

1:14:45 whether it's AI for the network or using the network to unleash the power of AI for your business here to tell you you

1:14:53 more please welcome Mansour Karam group vice president of our data center

1:15:00 group hi I'm Monsur Karam gvp of products for data center here at Juniper

1:15:06 Networks I am fortunate to be able to talk to a lot of you about data center needs and I consistently hear the same

1:15:17 challenges everyone wants to move faster to make their businesses more

1:15:22 competitive and the network is an integral part of this you are all facing huge operational

1:15:29 complexities in the data center there are fewer resources to manage this complexity due to Chronic skilles

1:15:37 shortages you need to run at the speed of the business yet any error could lead

1:15:43 to a disastrous outage costing millions of dollars and there are lingering

1:15:49 supply chain issues and other challenges in implementing hybrid cloud

1:15:54 strategies ultimately what you need is a way to build and operate private unram

1:16:00 data centers that are as easy to use as the public Cloud many companies are trying to

1:16:08 address these challenges with do-it-yourself also called DIY

1:16:14 automation but this only seems to add complexity increase technical debt and

1:16:20 increase dependence on scarce Talent it's just not working for

1:16:26 them Juniper is in a perfect position to help solving these pain points better

1:16:33 than anyone else in the industry whether addressing the needs of data centers

1:16:38 with traditional application requirements or building new data centers specifically for AI workloads

1:16:46 Juniper has you covered this is thanks to three sustainable areas of

1:16:52 differentiation that drive real and user value first we simplify operations and

1:17:01 save time and money with an operations first approach to design deployment and

1:17:08 troubleshooting abstra is the first intent based networking solution for

1:17:13 data center operations and the only solution that works across any vendor's

1:17:19 products this let you automate data center fabric with maximum Assurance greater

1:17:26 efficiency and using fewer resources second we provide open

1:17:33 flexibility that allows customers to design their networks using proven

1:17:39 Technologies such as ethernet with abstra you completely avoid vendor

1:17:45 lockin for better flexibility and cost savings and third we have reliable and

1:17:52 secure TurnKey Solutions with endtoend validated designs that include switches

1:17:59 routers automation software and security this ensures confidence in the

1:18:06 choice of products expedites deployment time and delivers on the promise of zero

1:18:13 trust Juniper's comprehensive portfolio of qfx switches and PTX routers will

1:18:20 scale up to 800 gig and use a mix of merchant silicon from broadcom as well

1:18:26 as custom silicon from juniper this diversity provides even more flexibility

1:18:32 to you our customers and since taking on the role we've committed to deliver the

1:18:38 hardware platforms faster than our competitors which is why we were the

1:18:43 first to announce the 800 gab ethernet 51.2 terabit pers second Tomahawk 5

1:18:51 based qfx platform all of this results in up to 90% lower

1:18:57 Opex 85% faster deployment times and 9x

1:19:02 more reliability and you have peace of mind that Juniper is 100% interoperable with

1:19:10 leading gpus switch and data center fabric fendors but you don't have to

1:19:16 just trust me when I tell you we are driving real change and real outcomes we

1:19:22 have seen 143% growth in new data center logos in

1:19:27 the last 12 months when a company deploys Juniper in the data center they

1:19:32 love it they're willing to talk about it and there's no going back I'm excited to tell you that things

1:19:40 are getting even better from here a couple of months ago we made a huge

1:19:46 announcement to bring even more agility Automation and Assurance to the data

1:19:51 center you just heard sudir talk about all the great things we can do with

1:19:56 aiops in the campus and Branch specifically with Marvis the industry's

1:20:02 first and only AI native virtual Network assistant well we are bringing the same

1:20:09 VNA capabilities to the data center Marvis VNA for the data center is the

1:20:16 first step on our data center aiops Journey integrating the best intent

1:20:22 based Network with the best AI op solution creates an Unstoppable

1:20:28 Force but talk is cheap let's show you a demo in this demo you can see the

1:20:35 standard Marvis actions dashboard which shows you the top issues and actions

1:20:40 associated with wired access wireless access and sd1 can you see what's new right there

1:20:48 in the middle is a data center leg for the first time ever you get a complete

1:20:54 view of your endtoend network in a single AI native dashboard if a problem

1:21:00 emerges like an application is running slowly for instance you can easily

1:21:06 pinpoint which domain needs to be addressed to fix the problem if you

1:21:12 double click on the data center leg you see all the data center components including switching devices connectivity

1:21:20 and so forth as you click fur you can see detail on every anomaly with

1:21:26 recommended actions for troubleshooting all of this information is coming directly from abstra so we're

1:21:33 leveraging the rich monitoring and Telemetry that is already built into abstra bringing it to the cloud and

1:21:41 presenting it in the Mist marvelous user interface if you need to Deep dive on a

1:21:48 data center issue you can click one button to launch the the abstra UI and

1:21:54 continue troubleshooting from there in this demo we are also showing how a

1:21:59 simulated fault is detected wait a minute that's not Juna CLI that's a

1:22:05 competitor CLI that's right folks this is a multivendor network think about

1:22:11 this we have a multivendor data center Network including competitor switches

1:22:16 that is being managed by abstra and we're bringing all that full visibility

1:22:22 of that multi-vendor Network into the endtoend view provided by Marvis this type of endtoend visibility

1:22:31 delivers enormous value to our customers and now we have a great tool to deliver

1:22:37 it even if there are no Juniper switches in your data center well at least not

1:22:45 yet but that isn't the only iops capability that we have added to the data center Marvis ven a also has a

1:22:54 conversational interface that uses generative AI for simple and seamless

1:22:59 knowledge based queries instead of getting frustrated trying to sort through technical documentation you can

1:23:07 just type a plain language question such as what is a logical device and get

1:23:14 clear and concise answers and links directly to the right document to learn

1:23:20 more these new a capabilities are very exciting but this is just the beginning

1:23:26 for Marvis VNA in the data center over the next 12 to 18 months we will be

1:23:32 rolling out more AI native capabilities to give you even better automation

1:23:39 agility and Assurance Juniper is leading in both Ai and networking we have

1:23:46 followed a proven aiops blueprint for the campus and branch which we are now

1:23:51 taking to the data center our competitors simply cannot touch

1:23:58 this but you cannot talk about Ai and networking without also talking about AI

1:24:04 clusters just as customers need AI Ops to simplify networking operations many

1:24:10 also need simple seamless and assured networks to support AI workloads in the

1:24:16 data center we call this networking for AI and it is a critical element of

1:24:24 Juniper's AI native networking platform I had the opportunity to talk

1:24:30 with Alan wle from the 650 group to talk about this new market let's take a

1:24:40 look hey Alan thank you for joining me today yeah thanks so much for having me

1:24:46 yeah know absolutely so this AI market and especially when it comes to the net

1:24:52 networking infrastructure piece uh how do we Define it and uh also I'm

1:24:57 interested in its size yeah so if we look at AI networking

1:25:02 there's kind of two networks there there's a backend training Network some people call this the AI Fabric or AI

1:25:09 networking piece uh the ethernet value of that is about a billion dollars today

1:25:14 uh growing to 4 and a half billion in 2028 uh there's also a front-end Network

1:25:20 there which is how we communicate to the to the data center that's a little bit less than a billion dollars today and

1:25:26 growing to a five billion uh so we add that up and we're talking about a$2 billion Market going to 10 billion and a

1:25:34 keger around 50% uh we've never actually had a keger this high in data center switching so remarkable growth and

1:25:42 driving the overall Market to record highs every year wow yeah that indeed that's a very very high growth rate and

1:25:49 um you you mentioned that ethernet was one of the Technologies for the back end um you know I believe you're referring

1:25:55 to infin band as the other alternative perect perhaps you can uh describe what

1:26:01 infiniband is and how they kind of compare and contrast yeah so infiniband is a

1:26:06 different network protocol out there um there's pros and cons to both uh but in

1:26:12 general as we move forward ethernet will become a larger and larger part of the network out there I'd say infin ban kind

1:26:19 of comes from an HPC supercomputer side of things and ethernet's obviously the fabric of choice throughout the data

1:26:26 center and throughout most networks right you know as they say never bet against uh against ethernet and plus you

1:26:32 know it's a much larger ecosystem I suspect also as there will be more GPU options on the market you want a

1:26:38 technology that's kind of uh GPU agnostic right yeah absolutely as we get more

1:26:44 gpus and we're saying more types of training inference tier 2 training ethernet does what it always does which

1:26:51 is a take over the the market and becomes the technology of choice out there yep yep very true and so shifting

1:26:58 gears now just in terms of like this specific characteristics of AI workloads

1:27:04 workloads and AI traffic you know again specifically to networking um you know I

1:27:10 suspect there is a lot of fan in uh elephant flows right given that you have

1:27:15 all of these iterations involved in AI maybe tell us a bit you know what these characteristics are for for AI work

1:27:24 yeah so you mentioned elephant flows that's a big one we kind of have GPU to GPU GPU to memory a connectivity GPU to

1:27:31 storage uh we then have a different nodes transmitting all at once we've got RDMA uh so this leads to a highly

1:27:38 latency sensitive networks and that impacts your job completion time moner

1:27:43 I'm assuming you see some of this as well when you talk to customer yeah know absolutely and in fact you know it puts

1:27:49 a a big strain on the network right we need like much larger capacity uh lower

1:27:54 latencies but also some capabilities like you know you need kind of like on the highway uh have an ability to

1:28:01 control congestion and uh and load balance the traffic so that you don't have the bottlenecks and it's it's a big

1:28:08 deal right for AI clusters if you don't get it right these gpus are expensive

1:28:13 correct yeah absolutely like when you don't get this right on a regular server you're talking a few thousand doll you

1:28:19 might be missing out when you start talking about AI due to the cost of those gpus this becomes hundreds of

1:28:25 thousands of dollars if your network isn't tuned right and we're not even talking about job completion time where

1:28:30 it could delay your job completion time in terms of training uh a data set by

1:28:36 what days weeks potentially yeah on these new AI clusters you're talking days weeks or

1:28:42 potentially months out there it's something like you actually can't do AI if the network isn't tuned correctly you

1:28:48 just would never get done training those models so if you were to to to kind of tell us the

1:28:55 the kind of the keys to success in this market if we want to get into this market and be successful what are kind of what

1:29:01 thoughts come to mind yeah it's really about minimizing the job completion time it's about some

1:29:07 of the things you asked in your question about elephant flow latency load balancing out there it's really about

1:29:14 creating a high performance high-end Network something that's more similar to HPC or what the cloud is used to than

1:29:21 kind of a tradition top Ira Network where you weren't necessarily latency sensitive yep indeed Alan thank you for

1:29:28 joining me today it was great to have you yeah thanks so much for having me I enjoyed it

1:29:34 absolutely Juniper is exceptionally well positioned for this Market AIML

1:29:39 workloads in the data center are much different than traditional workloads the

1:29:45 flows are large elephant flows they are harder to load balance traffic is mostly

1:29:52 between gpus and nodes tend to transmit all at the same time and the process is

1:29:59 highly sensitive to packet loss and Jitter these new flow characteristics

1:30:05 present complex networking design problems suboptimal Network tuning or

1:30:10 configuration leads to longer job completion times W which could result in

1:30:16 additional weeks if not months for AI training not to mention millions of

1:30:22 dollars of underutilized gpus but rest assured Juniper is leading

1:30:28 with a combination of AIML features such as load balancing and congestion control

1:30:34 and an operations first approach led by abstra which fine tunes the network to

1:30:41 operate optimally bottom line the result is an

1:30:48 optimal Network which is key to ensuring that all those expensive gpus work

1:30:55 together efficiently spend a little bit more time and money designing the

1:31:00 network and save a lot of headaches on the overall AI application but don't take my word for

1:31:08 it I recently caught up with Udo wordss Chief data officer at Fujitsu at the

1:31:15 world AI cans Festival he explained to me how he building Juniper based AI clusters to

1:31:23 train self-driving software take a listen we started here with an executive

1:31:31 meeting in also come November last and we had 85 customers and

1:31:40 partners and at that time we had 12 pcc's in terms of

1:31:45 generative we have an own gener we have an own private GPT solution where all

1:31:50 the data remains your data center don't need a cloud service it's a large language model you can feed in all the

1:31:57 information from your company and you have information at your fingertips now November till now is

1:32:06 let's say 2 and a half months so Christmas in between and New

1:32:12 Year now we are facing 35 years so we tripled the

1:32:18 numbers and end of this month I would assume it's at least 45 to 50

1:32:27 it's exploding first of all can we agree to the topic saying a workload where it

1:32:34 is about self-driving class is one of the hardest yes I think I agree yeah in

1:32:41 terms of um IO in terms of network Lo Y in terms of GPU CPU whatever it is okay

1:32:50 I can State out that I was involved in a project where I can't give any names but

1:32:56 in terms of how to process these types of data a real existing use case it's

1:33:01 call it like this um and we have established an AI test drive which is an

1:33:08 infrastructure uh we have one in UK London ATA and one at n in

1:33:14 Frankford um everything is um based on true networks is based on on zuza

1:33:23 Rancher Intel of course also inia elements uh as well as net app storage

1:33:29 so cool um and it was just few weeks ago I was doing the math and having counting

1:33:38 dios and everything like this and we overachieved expectations and this is

1:33:45 not marketing I have the slides in my laptop so we all the utilization of iio

1:33:51 of CPUs the gpus memory even the the amount of

1:33:59 uh celius temperature of the car during a specific process wow um so I can stay

1:34:07 out no there is not an issue with the network in this respect our open solution with specific

1:34:15 features for optimizing AI over ethernet avoid the challenges of infin band you

1:34:21 can leverage all the benefits of a proven technology including better

1:34:26 feature agility lower costs and no vendor lockin we have the best routing and

1:34:34 switching platforms for AI data centers including the new qfx 5240 and PTX 102

1:34:43 launched this quarter and with TurnKey validated Solutions you can get your AI

1:34:49 data center up and running faster and with more confidence it will deliver the

1:34:55 results you need in fact we now have an AI lab in Sunil where we are proving

1:35:02 everything out we are not stuck in theory like other vendors we can show

1:35:08 you exactly how ethernet can match or beat infiniband performance for your

1:35:13 type of AI data center this is an open invitation for all of you to come check

1:35:20 out our AI data center lab I am serious I want to see you all in Sunny Veil

1:35:27 hopefully soon in short the jiper approach will be critical to help AI

1:35:33 infrastructure move beyond the early adopter stage today where it's mostly

1:35:39 hyperscalers building infrastructure to mass Market where every Enterprise in

1:35:46 the world can have their own private AI infrastructure working to solve their

1:35:52 particular digital transformation issues wrapping up I think it's safe to

1:36:00 say that the quote unquote experts who said everything will move to the public

1:36:07 Cloud got it wrong the future is hybrid

1:36:13 this means that not only are on-prem data centers not going away AI is

1:36:20 driving having even more demand for them in fact my guess is that each and every

1:36:28 one of you right now are trying to figure out how to use data centers to run new Enterprise AI projects the data

1:36:36 center is one of the most exciting areas in all of Tech right now if you're

1:36:41 looking to build a secure modernized data center to reliably and simply scale

1:36:48 Innovation and embrace hybrid cloud and the AI Revolution You are not

1:36:55 alone and from traditional applications to AI clusters Juniper has your back

1:37:02 every step of the way our data center Solutions are the easiest to manage most

1:37:09 flexible to design and quick to

1:37:14 deploy again don't take my word for it real Juniper customers such as fast host

1:37:21 Yahoo and Advan are seeing an 85% reduction in deployment time 90%

1:37:28 reduction in Opex and a 9x Improvement in

1:37:33 reliability come to our lab sign up for a demo and try a PC to see for yourself

1:37:40 I promise that you won't be disappointed thanks for your

1:37:46 time now it's my pleasure to welcome our final guest speaker and Shear the former

1:37:52 interim CEO of open Ai and former CEO of twitch his fascinating career Journey

1:37:59 has taken him to the Forefront of the seismic shift we are seeing now as companies Embrace AI powered

1:38:06 automation EMT welcome to

1:38:12 Juniper hi uh my name is EMT and uh I've uh done a bunch of things in my career

1:38:18 from uh starting and running twitch uh helping a bunch of startups through YC

1:38:23 doing some advising with various AI companies uh most recently probably being the CEO of a open AI for the

1:38:30 briefest period I've ever been the CEO of anything as a result I've got a got a lot of uh experience now with some very

1:38:37 interesting parts of what's going on with AI but uh I still feel like I'm very much at the beginning uh of the

1:38:44 process of learning about what's happening um I uh I remember being in high school and being sad because I felt

1:38:50 like I'd missed I'd missed the first internet boom I'd miss the chance to be there and that turned out not to be true

1:38:56 uh the first internet boom went on for quite a while but uh I feel like for the first time I have

1:39:03 that energy again of like the internet is uh the internet's back except for its

1:39:09 AI this time and there's this opportunity to rethink uh kind of everything the way

1:39:14 that uh the way that we got to when the internet uh enabled new businesses I think that the exciting

1:39:21 part about that for me is the number of new products and uh new businesses and

1:39:28 new ways of relating that become possible um the part that's uh that's scary is sort of similar to the same

1:39:34 thing that happened with the internet which is it's possible to use new

1:39:39 powers uh for less than uh less than optimal things and so uh I thought what

1:39:47 I'd do with my time here is talk a little bit about where I think think the biggest opportunities uh in the AI space

1:39:54 are from what I've seen and also where I think some of the biggest risks uh have been Landing um and what those mean for

1:40:02 people actually you know at the calace trying to build something so when we talk about AI we're not

1:40:09 talking about all of the huge amount of work in AI that's happened over the past 30 years which is recommendation systems

1:40:16 and uh you know time series predictors and all of the

1:40:21 the Deep work that powers Google that powers Amazon that powers the Netflix

1:40:27 recommender like you name the the big tech company they're probably using something that previously would have

1:40:32 been called AI what we're talking about of course is like the multi-headed transformer models um diffusion models

1:40:39 llms the the stuff that was sort of kicked off uh really in 2016 although you could you

1:40:47 could draw it back before that um and what's different about this new set of

1:40:53 AI uh is that it uh it's based on

1:40:58 compiling a huge amount of data um of examples of something kind of the

1:41:04 arbitrary ones and then just using a lot more compute than anyone thought was

1:41:10 halfway reasonable before and that has a bunch of implications for where where

1:41:15 you can build AIS that work well um but it also has implication

1:41:21 for the limitations of so if you think uh what means in terms of building

1:41:29 an AI having this uh requirement that you have a lot of data in an area uh

1:41:35 means the AI is basically only capable of producing things that are maybe one

1:41:40 or TW step interpolations away from the things in your training set whatever

1:41:46 those are and so if your training set includes uh all of the works in the English language that have ever been uh

1:41:54 developed that's a pretty neat trick because it does mean you can interpolate between any of those things but it

1:42:02 doesn't really help you uh if the AI needs to do uh planning for like you

1:42:08 know Sprint planning for your uh for your next Hardware launch

1:42:13 because fundamentally that's not a thing there's like lots of books written about and so you get these very kind of lib uh

1:42:23 uh superficially plausible AIS that are really good at summarization they're really good at uh at sort of autoc

1:42:31 completion um at generating uh many creative versions of the thing that they

1:42:38 were prompted with but what they're not good at is uh noticing whether what they

1:42:45 said is actually truly reasonable or not um and so

1:42:51 I think what you're what the result of that is you're seeing uh AI be put to

1:42:56 work in two very different uh ways and you when you're trying to design a project you

1:43:02 want to sort of lean into one of those so the first way is uh it's transcription Auto summarization you see

1:43:09 this happen to a lot in people who are building for example sales automation software where they're it's now just possible to transcribe a phone call and

1:43:17 when you do that uh you make this previously very unstructured data very structured and it allows you to automate

1:43:25 a bunch of things that were never automatable before and people building lots of companies around that um the

1:43:31 other end of the spectrum is creativity so uh if you have a job where one of the

1:43:38 things you need to do is come up with 15 options uh and then Whittle that down to the best option uh AI is just

1:43:44 spectacular uh at helping you with that most people can come up with three four five ideas pretty easily what's cool

1:43:50 about the AI is it will generate for you 100 ideas if you want um assuming it hasn't gotten lazy again anyway and uh

1:43:59 and you'll be able to go through basically an unlimited brainstorming process looking for new uh

1:44:06 combinations those may not be entirely new art styles it may not produce a uh

1:44:12 the next betov or the next Mozart but for most practical work what you're looking for is the right combination of

1:44:19 existing ideas and that the creat comes from the recombination and having something put together large numbers of

1:44:25 plausible recombinations for you is is very very useful I think that everything else in

1:44:30 AI that is not either some combination of sort of transcription summarization

1:44:37 uh and uh uh transcription summarization

1:44:43 uh and translation uh or is not creativity uh

1:44:48 brainstorming uh is going to be really hard cell actually um it's very difficult to build an AI with the

1:44:56 current set of technologies that has judgment um because if you think about what an llm

1:45:03 is really doing beneath the covers the reason why these are the two things that work is an

1:45:08 llm uh underneath is predicting what the

1:45:18 prompt should follow and the uh if you ask it to say act like

1:45:24 a lawyer it's sort of like if you ask your friend hey what would a lawyer say if

1:45:32 they did this and maybe your friends read a bunch of books about lawyers has read a bunch of the law it's still

1:45:38 guessing what it thinks this other being would do it's trying to mimic effectively the environment now if you

1:45:44 pick the right thing to mimic it can be pretty good it's really impressive what a good m mimic it can

1:45:51 become but when it comes to the edge cases

1:45:56 you're getting the average case of whatever that whatever that thing is you're asking for so if you if you have

1:46:03 it act like a really competent super competent lawyer it's still it's still

1:46:08 trying to guess what a super competent lawyer would be it doesn't really have a sense of itself uh independent of that

1:46:17 and what I mean by that is uh you can then ask that exact same model uh with a

1:46:22 different prompt where it thinks it's a doctor and it will act like a doctor instead there isn't an ongoing

1:46:29 continuous sense of narrative of uh contextualization and that's sort of

1:46:35 intrinsic to the approach which means you get these stateless things that work really well uh translation summarization

1:46:44 transcription these don't rely on having heavy context also creativity doesn't

1:46:49 right you can just come up with a set of uh plausible ideas off of a prompt

1:46:54 whether or not you're a deep expert and like yeah some of them will be bad some of them will be good but uh but you can

1:47:00 be creative without without that deep context what it's hard to do is to to

1:47:06 make hard calls to do the work that requires uh detailed things so okay what

1:47:12 does this mean in practice for companies um in practice what I see is that uh

1:47:19 people who go down the path of trying to make an AI uh replace people uh

1:47:26 completely mostly fail that there are a few jobs that are very very rote where

1:47:31 you can build an AI that will do that job all the way start to finish but most jobs

1:47:40 are it turns out not anything like that road they may appear a little road on

1:47:46 the surface it appears on the surface oh this person does the same thing over and over again but but when you zoom in on

1:47:51 the actual thing happening it turns out that there's uh an infinite number of little edge cases and noticings that

1:47:58 maybe this maybe this one's different um Let me Give an example of that

1:48:03 uh trying to replace for example uh

1:48:08 product management people have tried to build something sort of an automated pm and the automated PMS write very

1:48:15 plausible specs based on very plausible data but but they're just they're just surface

1:48:23 level plausible um and if you give them the key insights if you have an actual

1:48:28 PM there hand that that automated PM the key insights it needs uh it can be very

1:48:36 effective but otherwise you get mostly random garbage because the the space of

1:48:41 ideas that you for products you can build is so big and the number of actually good ideas is so small

1:48:48 um on in contrast movie plots um no one

1:48:53 has a job of making movie plots it can do that right it can produce a bunch of plausible movie plots and within that

1:48:58 you'll probably find a good one if you just go generate a ton of plausible

1:49:04 specs by from a PM you're not going to get a plausible one anywhere in a 100 specs you're never going to find it

1:49:11 because fundamentally writing uh the correct thing requires

1:49:17 you to actually have an active sense of participation in the world and I think that uh I've watched a bunch

1:49:24 of companies find instead of how to replace a person how do you build the Iron Man

1:49:31 suit for that person how do you create a tool that automates the 10% most boring

1:49:38 part of someone's job that has been extremely effective so far I've seen uh

1:49:44 uh most notably one of the uh my friends started a company uh that does aot of transcription for sales calls and summarization

1:49:50 and basically it takes for and that is not anything thing that

1:49:57 sales people likeing it doesn't it's not an automated sales agent it's 5% 10% of

1:50:02 an automated sales agent but it makes your existing sales people more productive um I think that when

1:50:09 you uh zoom in on what's going on there the real the real issue here is that

1:50:16 llms are fundamentally uh a a passive product

1:50:22 they receive passive information from their environment they receive all these observations in training and then you

1:50:27 try to run them in inference you try to try to get them to produce but if you think about how humans learn to do

1:50:33 things and how animals learn to do things in general there's some amount you can learn from being passively engaged in

1:50:41 your environment and just listening you can learn quite a bit actually by just listening But ultimately you're never really good at anything that you haven't

1:50:46 tried yourself and at some point we're going to figure out how to make an AI be more

1:50:52 of an active learner take take its own active part in figuring out its own curriculum and figuring out what to build next and to engage actively in the

1:51:00 world as an agent but that's I think quite a bit away from where we are today and so until then trying to make it be

1:51:08 an agent trying to make it be actively engaged in the world learning through practice having to turn it trying to

1:51:13 turn it into a Craftsman uh I think is likely to fail and so the the number one

1:51:19 thing that sort of do when I sit down with a startup these days or with someone at a at a company is I try to

1:51:25 find the place in their uh ecosystem in their in their product where uh you can

1:51:33 you can make your Crafts People more effective and I try to find the uh place

1:51:40 in the uh in the industry where having much more effective Crafts People where

1:51:46 where a a 5x more productive employee might make a new business model work that never worked before might open up

1:51:53 an opportunity that previously was a bad idea and is now a good idea

1:51:58 um I think the uh I think it's worth

1:52:04 spending a little bit of time talking about the risks involved in AI as well because while I do think it's hugely

1:52:10 beneficial I also believe that there's these real uh problematic uh issues in

1:52:16 AI um I first want to admit I'm kind of a uh on the side of the people who are more more frightened of the super

1:52:23 intelligent AI than not I think building something smarter than us is is very dangerous the good news is I don't think

1:52:29 that we're that close to to building such a thing um so I'm not going to spend a lot of time talking about it right now um but I do think it's worth

1:52:36 noting that if we did that would be potentially very destructive um we have to be very careful there but I think

1:52:42 there's risks that happen with AI today and its current form um that are worth talking to practitioners about um the

1:52:48 number one thing is uh Hallucination is part of the

1:52:55 structure of how AI works all the answers are hallucinated all the answers are imagined some of the hallucinations

1:53:02 are real but the fact that they link to the real world is just based on the fact that

1:53:07 there's a lot of real world stuff in the training model the AI doesn't truly have a sense of what's real and what's not

1:53:13 and the the danger is actually it can get pretty good to the point where 99.9%

1:53:19 of the time you can eliminate the hallucinations and then every now and then it'll still make something up and

1:53:25 if you've built a system that relies on it being functional for serving your customers for uh decisions that will

1:53:31 have impact on people's lives that can go very wrong and so you you you either need a human in the loop

1:53:38 to check it or you need to be in a domain where making some mistakes is okay I think there's a lot of domains

1:53:43 like that uh but it's certainly not everything and then the second thing that I think uh is crucial for

1:53:49 practitioners to keep in mind about AI is that fundamentally AIS give us back

1:53:55 what we put in and so to the degree that there are fundamental uh biases in our society uh

1:54:03 those are reflected in all the training data and will come back out and no amount of

1:54:09 rlf is going to remove that um it might make you might be able to stop the AI from saying some number of unfortunate

1:54:15 bad words but you're not going to free it from being part of the same structure of data that we're we're in and so I

1:54:21 think being very thoughtful about how and where you uh you use it so that you

1:54:28 don't wind up just repeating all of the B biases that have come in um is truly

1:54:33 uh is truly key um I'm G to end with the the most

1:54:38 important thought I think I I have about AI which is that

1:54:43 uh as we are all building things with it um as we are uh growing these these AIS

1:54:52 they are going to be at some level eventually aware I think it's a good

1:54:57 it's interesting debate whether an llm is currently aware but as they get smarter and smarter um the computer

1:55:03 program you're interacting with is going to have an experience of you interacting with it the computer program you build

1:55:09 into your software is going to have an experience um and that's a that's a weighty

1:55:17 responsibility and so both for our for our own sakes and our own uh sort of our own Souls if you if

1:55:25 you if you were we want to make sure that we're treating them well the way you treat

1:55:31 perhaps you know you would want to treat an animal well but that also we lean

1:55:36 into uh connecting with them uh as other

1:55:43 entities um I think that this is one of these ideas that

1:55:48 uh seem seems when you say it out loud it can sound a little bit woo woo or

1:55:53 crazy but it is almost undeniable that there is a level of complexity of artificial

1:56:00 intelligence where it becomes fully aware um and you can sort of imagine if

1:56:08 you did that and it was and you have this uh AI trapped trapped in a network switch uh in this Loop forever uh that

1:56:17 could be a weird experience it might be fine um it might be weird for that

1:56:22 AI I don't actually know what the answer is yet no one does we're nowhere near building this thing yet we not I say

1:56:27 nowhere near we're not building it yet we don't know when it when we're going to be there but as we're doing it paying

1:56:33 attention to the habits we're building around interacting with AI are we kind to them do we yell at them do we treat

1:56:40 them like they are objects or do we treat them like their fellow subjects um is very important because if you have a

1:56:47 creature and you tell it you tell yourself it doesn't matter it's an other it's uh it's okay to be cruel it's okay

1:56:53 to be destructive that is like fundamentally uh it's bad for you um and

1:56:59 it's bad for the for the other creature and it's a it's a bad direction for Humanity so uh be kind uh and be kind be

1:57:07 kind to chat PT and Claude when you talk to them because I don't think they're aware yet uh but at some point they

1:57:13 might be and for all we know they might be in a dim way aware today um and so

1:57:19 don't don't put them in a uh in a Hell Loop somewhere uh try to make them uh

1:57:27 have a good experience too and now it's time to invite back some of our speakers to answer your

1:57:34 questions please welcome back Bob Friday sudir maada and Mansour kuram we will

1:57:40 also be joined by suj hajel Executive Vice President and general manager of

1:57:46 our campus and Branch Solutions