Bob Friday Talks: Bytes of Brilliance, Unveiling the AI Canvas. An AI Primer.

0:00 hello and welcome my name is tar redwan

0:03 I'm product marketing director for

0:04 artificial intelligence here at Juniper

0:06 Networks and we're going to have a a

0:08 quick discussion now about the

0:09 fundamentals of

0:11 AI you know when it comes to networking

0:14 AI has become very very popular as of

0:17 late so Bob I want to ask you the first

0:20 question what exactly is AI and what are

0:23 its sailing characteristics yeah TK you

0:26 know what I usually tell people is AI is

0:28 ultimately the next step in the

0:30 evolution of automation was a slight

0:33 twist in what we used to do in the past

0:35 when we automated things it was very

0:36 deterministic we wrote scripts and they

0:39 basically did the same thing day in and

0:41 day out and you really didn't think out

0:42 of it what we're really building with AI

0:45 is really more about cognitive reasoning

0:47 now we're starting to build AI solutions

0:49 that act more like a human and actually

0:51 change Behavior over time it's almost

0:54 like hiring a new intern right you bring

0:56 on that intern you bring on that AI

0:59 assistant the first thing you have to do

1:01 is understand their skills correct you

1:04 know and after you do that you expect

1:06 this person this intern this AI

1:08 assistant to actually get better over

1:10 time as it starts learning so that is a

1:13 fundamental difference between what I

1:14 call AI in the past very deterministic

1:17 ml type of algorithms and now we're

1:19 moving to these much more deep learning

1:21 type of algorithms that really have the

1:23 behavior of cognitive reasoning and

1:25 start to look like a human that's

1:26 actually learning things as you actually

1:28 interact and use it more

1:30 and in our industry when did AI exactly

1:33 become important you know that's an

1:36 interesting question because you look at

1:37 my experience you know back in the 80s

1:40 when I did my masters I was actually

1:41 doing neural networks so when did AI

1:44 really become a big thing if you look at

1:46 Google and the Google searches AIML took

1:49 off around the 2014 time frame and if

1:53 you think about what happened back in

1:54 the 2014 time frame that's when we had a

1:57 convergence of several things coming

1:59 together we had open source code coming

2:02 together we had many more packages

2:04 making it easier to develop these

2:05 algorithms we had these Cloud models

2:07 right AWS Google aure where compute and

2:11 storage got a lot cheaper so I think we

2:13 saw kind of a certain thing happened in

2:16 2014 where we saw AI startup startup

2:19 between a combination of cloud

2:21 technology and really the open source

2:23 Community bringing tools together and

2:25 make it easier to actually build bigger

2:27 Solutions and probably the other big

2:29 thing especially what we saw of chat gbt

2:32 is we found out compared to the models I

2:33 was doing 20 years ago now we have deep

2:36 learning models with 100 200 500 billion

2:40 param weights in them right these models

2:43 are getting on the size of your brain

2:44 now and that's the other thing that's

2:46 happening now in 2014 So lately is the

2:49 models have gotten very large and they

2:51 do much more interesting things let's

2:53 double click a little bit and maybe

2:55 explain in basic terms what exactly is

2:58 ml machine learning I would say machine

3:00 learning has been around for decades

3:02 this is probably things like logistic

3:04 regression decision tree algorithms

3:06 these are algorithms we've been using

3:08 for 20 30 40 years so that's what I call

3:11 machine learning what's really changed

3:14 is really what we call Deep learning

3:16 this is where we're starting to use much

3:18 more complicated models with many more

3:20 layers many more weights this is what

3:22 brought Transformers and chat GPT to

3:24 life is these really deep learning

3:26 models that are continuously learning on

3:28 a lot more data

3:30 and so similar what we''re doing with

3:31 teams in zoom and Marvis now we're

3:33 basically bringing in deep learning

3:35 models and a lot of continuous learning

3:37 into those models and that what brings

3:39 out this neric Behavior where these

3:41 models learn and basically get better

3:44 over

3:45 time now we've heard about these terms

3:49 of unsupervised learning versus

3:50 supervised learning can you kind of

3:52 explain what that all means yeah

3:55 supervised learning is like what they

3:56 did with chat gbt and Transformers right

3:59 they took all the data off the internet

4:02 and they trained a model to predict the

4:05 next word similar what we're doing with

4:07 zoom and teams where we're taking tons

4:10 of Zoom data and building models that

4:12 can actually learn the zoom and teams

4:15 performance that is supervised learning

4:17 where you have a whole bunch of labeled

4:20 data that you're trying to learn to

4:21 predict something unsupervised learning

4:24 is really what we did with location

4:26 where I don't have any label data you

4:29 know and we're trying to learn the path

4:31 loss model for every mobile device in an

4:34 area where I basically have no label

4:37 data but I'm learning to predict that

4:39 path loss model for all devices in your

4:41 network to make a better location

4:43 estimate and just to cover a bit more on

4:46 the AI alphabet soup can you can you

4:48 maybe touch on neural networks and maybe

4:51 deep neural networks yeah these are

4:54 subtly the same you know when you look

4:56 at the original thesis on neural

4:57 networks it was pattern after after your

4:59 brain right if you look at the neurons

5:02 inside of your brain and how you learn

5:04 you're basically analog weights so that

5:07 original neural network model was

5:09 basically building up layers of Weights

5:11 that would basically able to be trained

5:13 to predict something and as I mentioned

5:15 before what has really changed now is

5:18 these deep learning models like

5:21 Transformers that we're using with chat

5:23 gbt this is what we call assistant um

5:27 attention based trans famon models these

5:31 are really very large neural network

5:33 models now with 175 billion different

5:37 weights inside of them so that's the

5:39 subtle difference between neural network

5:41 and what we call Deep learning deep

5:42 neural networks is really around the

5:44 size and the amount of Weights we have

5:46 to

5:47 train since you did mention chat GPT can

5:50 you explain what is an llm yeah these

5:53 large language model is what got

5:54 everyone's attention about a year or so

5:56 ago uh and it's made our lives a lot

5:59 easier easier because I can tell you

6:00 there's a lot fewer AI Skeptics out

6:02 there once they actually use chbt so

6:05 these large language models are based on

6:07 a model called Transformers and tension

6:11 based Transformers and this is where

6:13 they're basically taking tons of data

6:16 they're taking basically everything on

6:17 the internet and they're training these

6:19 models to predict the next word so you

6:23 think of this as a conditional

6:24 probability where I take the last 500

6:26 words to predict the next word and this

6:29 is where these large language models are

6:31 starting to sound like a natural human

6:33 language and even more they're starting

6:35 to look like reasoning so large language

6:37 models are really an amazing thing

6:40 coming out in the networking right now

6:42 both for how they sound like natural

6:44 language humans and where they're

6:45 actually starting to have reasoning

6:47 capabilities and this is going to play

6:48 into helping us troubleshoot networking

6:51 going forward now is this the same as

6:54 generative AI or gen AI yeah I think gen

6:58 Ai and Alm are basically synonms for the

7:01 same concept here whether it's a large

7:03 language model General AI uh large

7:05 language models usually apply to some

7:07 sort of language you can technically

7:09 take these same attention based

7:11 transferer models these gen AI models

7:14 and apply them to other use cases uh an

7:16 example would be instead of using

7:18 language we can use these gen AI models

7:21 to actually translate text to search SQL

7:24 elastic search this will basically

7:27 enable any customers to actually start

7:28 to expl expl their network data and

7:30 replace some of these bi tools making it

7:33 easier for customers and Enterprises to

7:36 explore data in their

7:38 databases I can't help but ask the

7:41 question how has Juniper used or is

7:44 using gen AI today yeah so what we

7:47 announced last year Mobility field day

7:50 was basically integrating gen Ai and

7:52 large language models into our marest

7:55 and the first phase of this is really

7:58 around public documents

8:00 making it much easier for customers and

8:02 support teams to get specific answers

8:05 out of hundreds of thousands of

8:07 documents so this is Phase One of taking

8:10 gen Ai and LM and making it easier to

8:13 get both the answer and the documents

8:16 and the information you're looking

8:18 for now just to quickly Go Back to

8:21 Basics what exactly is model training

8:24 yeah so every model is slightly

8:26 different it depends on what problem

8:28 you're trying to train but you could

8:29 think of every model basically has a set

8:32 of trained AA for supervised learning

8:34 that we talked about so let's take Zoom

8:36 teams for example right I have a bunch

8:39 of Zoom or teams data that I have to

8:41 join with my network feature data and

8:44 then I want to train that model once a

8:46 day once a week and so when we look at

8:49 the training model it's basically the

8:51 problem you're Sig to solve with anomaly

8:53 detection we're basically training

8:55 models every night and every day on the

8:58 last 3 months of data and that's for

9:00 anomaly detection making sure we look

9:02 back in time trying to predict anomalies

9:05 for the next 10 minutes so that is a

9:07 training it's slightly different whe

9:09 you're doing self-supervised or

9:10 supervised but almost all deep learning

9:13 models have some sort of training

9:14 schedule assigned to them thanks again

9:17 for your time Bob that was really

9:18 informative appreciate it it's good to

9:20 be here TK always enjoy talking about AI

9:28 Ops