Bob Friday Talks: Generative Horizons, Gen AI and Juniper’s CUEL Breakthroughs

0:00 hello everyone my name is Aku and I'm

0:04 junipers head of AI evangelism today I

0:07 have the honor of being joined by our

0:09 very own Bob Friday who's our chief AI

0:12 officer well hello there Bob cool thanks

0:15 for having me this is a great time to be

0:17 in network and it's exciting to be here

0:18 with you well thank you so much well

0:21 everyone today we are going to learn

0:23 more about geni and what Juniper

0:26 Networks is doing in that field also

0:28 we'll dive deeper into llms as well as

0:31 Microsoft teams as well as zoom and what

0:34 we're doing for our customers in that

0:35 space so what is the power of gen and

0:38 what does this mean for Juniper Network

0:41 yeah cool start with at an industry

0:43 level gen chatti is the next generation

0:47 of user interfaces for a lot of vertical

0:49 including networking it's really what's

0:52 going to allow people to get the data

0:54 out of their Network in a much more

0:55 natural easier way if you look what

0:57 happened in networking we've gone from

0:59 CLI to dashboards chat GPT and LM is

1:02 that next Generation user interface for

1:05 Juniper specifically in Marvis we've

1:07 been doing conversational interfaces

1:09 since 2018 and helping Enterprises

1:12 answer real-time troubleshooting

1:14 questions natural language is going to

1:16 bring that natural human interaction

1:19 into Marvis now making it really easier

1:21 for people to understand real type

1:23 troubleshooting and what we've added

1:25 lately is that public knowledge based

1:28 documentation questions where people now

1:30 answer questions about how to configure

1:32 a switch an AP and a router and making

1:34 it easier to find that information in

1:36 public documents so I love have you re

1:40 um in the M team reor it to put the data

1:43 scientists and the support team side by

1:46 side to work together and I think that

1:48 is such an important partnership for the

1:50 success of the organization on that note

1:53 I would love to learn about how do llms

1:55 Empower your customers so that they

1:57 don't open as many support tickets

2:00 when we started miss that was the

2:02 original vision for Miss is really using

2:04 cloud and AI Ops to reduce the

2:06 operational hassle of managing and

2:08 operating these networks right and with

2:10 llms we're continuing that Journey

2:12 making it easier for customers to get

2:15 questions answered before they open the

2:16 support tickets so with llm we've

2:19 actually added it into Marvis we

2:21 actually are using it in our support

2:23 team now every support ticket that comes

2:25 into Miss right now basically has an LM

2:28 answer and the documents in that ticket

2:30 trying to help our support team get to

2:33 resolution faster and that same

2:35 technology is now being available to our

2:37 customers and that's how we're trying to

2:39 help prevent support tickets both inside

2:41 of our support team and in our customers

2:43 team you know as we've discussed right

2:46 when we move to Cloud support is

2:48 ultimately at the best proxy for our row

2:50 customers you know and helping them

2:52 reduce tickets is basically helping our

2:54 customers the fewer tickets they see is

2:56 the fewer tickets our customers are

2:58 sending them yeah I love love that point

3:00 Bob around how if we reduce the support

3:04 tickets of our customers that is the

3:06 best you know Testament of success yeah

3:09 and what I tell people they always ask

3:10 me you know what's the difference

3:11 between Ai and whitewashing AI the first

3:14 step on that journey is really making

3:16 sure you're using your own cloud a Ops

3:19 in your support team if they're not

3:20 doing that they have not started the

3:22 journey to Cloud a Ops well thanks Bob

3:25 those are really you know interesting

3:27 points and on that note we'd love to

3:28 talk more about the zoom and the team

3:31 Integrations and how they're adding

3:33 value to customers because I know that's

3:35 essentially very important aspect and

3:38 many customers have found a lot of value

3:40 from from zoom and team Integrations

3:42 yeah know this is an interesting one you

3:44 know this started with a big customer

3:45 about three years ago who to us that you

3:48 know Zoom calls were in the top three

3:50 support tickets they were getting right

3:52 and the challenge was how do we help

3:53 customers with their zoom and teams

3:56 because when you think about it zoom and

3:58 teams are what I call the canary and the

3:59 coal mine right if your network has any

4:03 problems it's going to show up in those

4:04 video collaboration apps right and so

4:07 what we've done with zoom and teams is

4:09 basically we've taken Zoom data and

4:12 network feature data and now we build

4:14 models that can accurately predict user

4:17 experience Zoom teams experience and

4:20 once we have these models built we can

4:22 actually use them now to use things like

4:24 shapley to get the visibility on what is

4:27 causing a bad Zoom or teams experience

4:30 all the way at the user level AP level

4:32 and the site level so Canary and the

4:35 coal mine very important usually the top

4:38 three or four support ticket problems

4:40 yeah that sounds amazing Bob I I totally

4:43 agree just being a user of teams and

4:45 zoom myself it is so valuable to have

4:48 like I mean the Wi-Fi is so essential

4:51 and adding AI to the Wi-Fi that's that's

4:54 a big win yeah and you know I think the

4:56 other thing with zoom and teens you know

4:58 I always tell people people you know we

5:00 all have been using machine learning

5:02 algorithms zoom and teams what we're

5:04 doing here is what I call continuous

5:06 learning you know where we're

5:07 continuously training these models with

5:09 megabytes terabytes of data right and

5:12 that is the power when we go from

5:14 machine learning to deep learning you

5:16 know machine learning is what we've been

5:17 using for years to automate these

5:19 networks deep learning is really where

5:21 we're differentiating this is your chat

5:23 PT experience so how does Juniper

5:26 differentiate itself from more

5:27 traditional networking companies

5:30 yeah you know I think this question goes

5:31 all the way back to the origins of mist

5:33 you know when sujan and I left Cisco we

5:36 made a big bet that basically cloud and

5:39 AI were architectural changes in the

5:41 industry right because we knew we were

5:43 not going to beat the big guys with a

5:44 feature so I think when you look what we

5:47 did at Juniper and Miss was really that

5:49 AI native blank sheet of paper story

5:52 where the BET was hey if we're going to

5:53 build a real time AI Network we're going

5:57 to have to start from scratch and

5:58 basically build something something that

5:59 can process tons of data inest tons of

6:02 data from different networking sources

6:04 so I think that is the main

6:05 differentiation is making sure that

6:07 cloud Foundation got built and then on

6:09 top of that AI native Cloud Foundation

6:11 is really building Marvis in the AI Ops

6:14 piece of it I love the the AI native and

6:17 how that that being the foundation and

6:19 Jennifer really differentiating itself

6:22 as having AI at its core and Foundation

6:25 this is big you know cool I think you

6:27 know the other differentiation I think

6:29 would people have a hard time

6:30 appreciating is when you actually go

6:31 into the details of moving from kind of

6:33 a controller based software on Hardware

6:36 into these microservices Cloud

6:38 architectures not everyone fully

6:40 appreciates the difference in the

6:41 software architecture and that's what

6:43 really forced us to actually T start a

6:45 blank sheet of paper because we couldn't

6:47 simply take the software that was on our

6:48 controller and put it in a container

6:50 Cloud you really had to start from

6:52 scratch and start with this software

6:54 architecture built around microservices

6:56 and API and that is the other big

6:59 difference at Juniper is that cloud

7:01 architecture so how do you improve radio

7:04 Resource

7:05 Management yeah so in the Wireless World

7:07 radio Resource Management has been

7:08 around for decades um 20 years ago when

7:11 I was doing Aerospace we were really

7:14 trying to optimize how access points saw

7:16 each other you know this time around

7:19 we're really trying to optimize the user

7:21 experience and that's really trying to

7:22 leverage reinforcement learning I'm

7:24 making sure that we're balancing the

7:26 connectivity between a user and the

7:28 access point you know when you look at

7:30 some of the challenges the challenges

7:32 have changed over the last 20 some years

7:34 you know 20 years ago we had you know

7:36 transmit power and channel basically you

7:38 know we look where Wi-Fi is headed with

7:40 6E and 7 now we have multiple bandwidths

7:43 to deal with we got 20 40 80 160 MHz

7:46 bandwidth and we have more bands now

7:48 right we have 2.4 5 gigs and now we're

7:51 even moving to the six gig bands so

7:53 there's a lot more knobs to adjust in

7:55 this radio reinforcement management uh

7:57 but there's a lot more tools in the

7:58 toolbox now we have reinforcement

8:00 learning that we can apply to it well

8:02 thanks Bob yeah that makes a lot of

8:04 sense on another note one area that I'm

8:08 personally very passionate about is AI

8:10 trust and explainability and

8:13 transparency I would love to learn about

8:15 what are the measures that we're taking

8:17 in that space yeah so as we discussed

8:20 you know as zoom in teams great example

8:23 where we're building models predict user

8:25 experience we're actually using

8:26 something called Mutual information

8:29 right oh Mutual information and now

8:31 we're using shle in the first generation

8:33 it was really around Mutual information

8:35 and trying to understand why a user

8:37 minute was bad now with shapley where

8:39 that are really explaining what network

8:42 feature is actually most likely to

8:44 predict an outcome right and that is

8:47 example of where explain abilities bring

8:49 a lot more granularity to root cause uh

8:52 if we look forward to other use cases

8:55 oops what was that go oh llms so if we

8:57 look forward to what we're doing with

8:59 llms that's almost self-explanatory

9:02 where we're starting to basically use

9:04 what they call retrieval Technologies

9:06 where we're basically retrieving

9:07 documents and basically having chat GPT

9:10 or an open source llm answer questions

9:13 in the context of a document so in that

9:15 case there the customer customer support

9:18 they get an answer but they also get the

9:20 docs from which that answer cames so

9:22 making sure we give domain specific

9:24 answers and minimize the hallucinations

9:26 you see out of

9:27 llms so as a woman been an AI myself and

9:30 a data scientist I've spent the last 10

9:33 plus years working on AI and machine

9:35 learning and now I'm really excited to

9:38 work with Bob Friday and apply Ai and

9:42 machine learning trust and ethics to the

9:44 space of networking because the world

9:47 isn't the world without the internet

9:49 right and applying AI to the Wi-Fi and

9:51 the internet makes the world much more

9:54 connected and we have every connection

9:56 counts for me you know if you look

9:58 what's going on in the industry right

9:59 now you know I tell people when you go

10:01 to your hospital and your doctor for get

10:03 C you're going to want to make sure your

10:04 doctor is using the latest and greatest

10:06 AI when he diagnoses you same with those

10:09 cars you're going to want to make sure

10:10 they use the latest and greatest AI you

10:11 know and in networking it's the same

10:13 you're going to want to make sure when

10:14 you connect to that Network that it's

10:16 using the latest and greatest AI to

10:18 provide you the best experience you know

10:20 and that's why it's exciting times to be

10:21 in networking so it's great to have you

10:23 thanks Bob for joining us today we

10:26 appreciate your time thanks and Co for

10:28 having me and look forward to meting

10:35 more